Method: Introduces Evaluation Disclosure Cards (EvalCards) as a standardized reporting framework designed to enhance transparency by systematically documenting evaluation procedures, datasets, metrics, and governance considerations.

Result: EvalCards provide a practical solution to address three persistent shortcomings in current evaluation reporting: reproducibility (through detailed documentation), accessibility (through structured information), and governance (through compliance-ready reporting).

Conclusion: EvalCards offer a path forward for more transparent and accountable NLP evaluation practices that can better serve both researchers and practitioners while meeting evolving governance requirements in the field.

Abstract: Evaluation has long been a central concern in NLP, and transparent reporting practices are more critical than ever in today’s landscape of rapidly released open-access models. Drawing on a survey of recent work on evaluation and documentation, we identify three persistent shortcomings in current reporting practices: reproducibility, accessibility, and governance. We argue that existing standardization efforts remain insufficient and introduce Evaluation Disclosure Cards (EvalCards) as a path forward. EvalCards are designed to enhance transparency for both researchers and practitioners while providing a practical foundation to meet emerging governance requirements.

Method: Uses Least Recently Used (LRU) cache tables of token n-grams to generate draft sequences for speculative decoding, requiring no training and being model-agnostic.

Result: Achieves state-of-the-art performance among comparable methods despite minimalist design, with easy integration into existing systems and potential for fast domain adaptation.

Conclusion: Cacheback Decoding provides an effective, simple, and practical approach to LLM acceleration through cache-based speculative decoding that exploits language locality.

Abstract: We present Cacheback Decoding, a training-free and model-agnostic speculative decoding method that exploits the locality in language to accelerate Large Language Model (LLM) inference. Cacheback leverages only Least Recently Used (LRU) cache tables of token n-grams to generate draft sequences. Cacheback achieves state-of-the-art performance among comparable methods despite its minimalist design, and its simplicity allows easy integration into existing systems. Cacheback also shows potential for fast adaptation to new domains.

Method: JELV framework with two implementations: multi-turn LLM-as-Judges pipeline and distilled DeBERTa classifier, validated on human-annotated PEVData benchmark.

Result: 90% agreement with human annotators (LLM pipeline), 85% precision on valid edits (DeBERTa), state-of-the-art correlation with human judgments, and measurable performance gains when retraining GEC systems on expanded dataset.

Conclusion: JELV provides a scalable solution for enhancing reference diversity and strengthening both evaluation and model generalization in GEC systems.

Abstract: Existing Grammatical Error Correction (GEC) systems suffer from limited reference diversity, leading to underestimated evaluation and restricted model generalization. To address this issue, we introduce the Judge of Edit-Level Validity (JELV), an automated framework to validate correction edits from grammaticality, faithfulness, and fluency. Using our proposed human-annotated Pair-wise Edit-level Validity Dataset (PEVData) as benchmark, JELV offers two implementations: a multi-turn LLM-as-Judges pipeline achieving 90% agreement with human annotators, and a distilled DeBERTa classifier with 85% precision on valid edits. We then apply JELV to reclassify misjudged false positives in evaluation and derive a comprehensive evaluation metric by integrating false positive decoupling and fluency scoring, resulting in state-of-the-art correlation with human judgments. We also apply JELV to filter LLM-generated correction candidates, expanding the BEA19’s single-reference dataset containing 38,692 source sentences. Retraining top GEC systems on this expanded dataset yields measurable performance gains. JELV provides a scalable solution for enhancing reference diversity and strengthening both evaluation and model generalization.

Method: Created a robust evaluation framework using 2,800 carefully curated Chinese medical exam questions across 7 specialties (cardiovascular, gastroenterology, hematology, infectious diseases, nephrology, neurology, respiratory) at two difficulty levels (attending physician and senior physician). Evaluated 27 state-of-the-art LLMs on this benchmark.

Result: Mixtral-8x7B achieved highest overall accuracy (74.25%), followed by DeepSeek-R1-671B (64.07%). No consistent correlation between model size and performance. Performance varied significantly across specialties - better on cardiovascular and neurology, worse on gastroenterology and nephrology. Top models showed minimal performance degradation between difficulty levels.

Conclusion: LLMs show promise for medical applications but have current limitations. The benchmark provides critical insights for deployment in medical education and clinical decision support, highlighting the need for specialized evaluation and the potential of smaller mixture-of-experts architectures.

Abstract: The rapid advancement of large language models(LLMs) has prompted significant interest in their potential applications in medical domains. This paper presents a comprehensive benchmark evaluation of 27 state-of-the-art LLMs on Chinese medical examination questions, encompassing seven medical specialties across two professional levels. We introduce a robust evaluation framework that assesses model performance on 2,800 carefully curated questions from cardiovascular, gastroenterology, hematology, infectious diseases, nephrology, neurology, and respiratory medicine domains. Our dataset distinguishes between attending physician and senior physician difficulty levels, providing nuanced insights into model capabilities across varying complexity. Our empirical analysis reveals substantial performance variations among models, with Mixtral-8x7B achieving the highest overall accuracy of 74.25%, followed by DeepSeek-R1-671B at 64.07%. Notably, we observe no consistent correlation between model size and performance, as evidenced by the strong performance of smaller mixture-of-experts architectures. The evaluation demonstrates significant performance gaps between medical specialties, with models generally performing better on cardiovascular and neurology questions compared to gastroenterology and nephrology domains. Furthermore, our analysis indicates minimal performance degradation between attending and senior physician levels for top-performing models, suggesting robust generalization capabilities. This benchmark provides critical insights for the deployment of LLMs in medical education and clinical decision support systems, highlighting both the promise and current limitations of these technologies in specialized medical contexts.

Method: Clusters vocabulary embeddings offline and uses centroid-plus-radius geometric bounds to identify which tokens can be safely omitted from computation, constructing small sub-vocabularies for each decoding step with sparse GEMV kernels, multi-GPU sharding, and CUDA Graph optimization.

Result: Significant speedup over full vocabulary decoding while maintaining distributional guarantees (exact top-k certification and ε-certified softmax approximations) with low fallback rates.

Conclusion: CSV-Decode provides an efficient approach to LLM inference that reduces computational bottlenecks through sparse computation while ensuring correctness guarantees, with complete system implementation available.

Abstract: Large language models face significant computational bottlenecks during inference due to the expensive output layer computation over large vocabularies. We present CSV-Decode, a novel approach that uses geometric upper bounds to construct small sub-vocabularies for each decoding step, enabling efficient sparse computation while maintaining dual correctness guarantees: exact top-$k$ certification and $\varepsilon$-certified softmax approximations. Our method clusters vocabulary embeddings offline and uses centroid-plus-radius bounds to identify which tokens can be safely omitted from computation. We provide a complete system implementation with sparse GEMV kernels, multi-GPU sharding, and CUDA Graph optimization. Experimental results demonstrate significant speedup over full vocabulary decoding while maintaining distributional guarantees and low fallback rates. Our code implementation available at \href{https://github.com/FastLM/CSV-Decode}{https://github.com/FastLM/CSV-Decode}.

Method: Controlled experiments with numerical sequences, evaluating four model families: non-LLM encoders, LoRA-adapted LLMs, model-soup merged LLMs, and SLERP-merged LLMs. Assessed with clustering indices (Silhouette, Davies Bouldin) and kmeans label analysis.

Result: LLM-based backbones capture higher-order compositional patterns better but suffer from adapter dominance. SLERP merging consistently recovers base-model structure while retaining task gains, yielding superior clustering separability and robustness compared to other methods.

Conclusion: SLERP model-merging effectively balances task-specific adaptation with base-model generalization, offering better tradeoffs than model souping or unmerged approaches for embedding quality.

Abstract: We investigate the generalization properties of dense text embeddings when the embedding backbone is a large language model (LLM) versus when it is a non-LLM encoder, and we study the extent to which spherical linear interpolation (SLERP) model-merging mitigates over-specialization introduced by task-specific adaptation (e.g., LoRA). To make the comparison concrete and domain-agnostic, we design a controlled suite of experiments in which models embed short numerical sequences and are evaluated on their ability to cluster and classify those sequences according to well-defined number-theoretic properties. Our experimental protocol compares four families of models: (1) non-LLM encoders trained from scratch or fine-tuned for embeddings, (2) LLM-based encoders adapted with parameter-efficient methods (LoRA), (3) LLM-based encoders with LoRA followed by model souping merging into the base weights, and (4) the same LoRA-adapted LLMs merged using SLERP across checkpoints or stages. We evaluate representational quality with clustering indices (Silhouette and Davies Bouldin). We additionally analyze the use of kmeans labels to see if the embeddings encode any other information besides the one we are testing for. Empirically, we find that LLM-based backbones produce embeddings that better capture higher-order, compositional numeric patterns, but are prone to adapter dominance that degrades balanced generalization; SLERP merging consistently recovers base-model structure while retaining most task gains, yielding superior tradeoffs in clustering separability, and robustness compared to model souping or models that were not merged.

Method: Proposes joint training on available spoken DST data and written textual data from other domains. This approach leverages both speech and text modalities to achieve cross-domain generalization without requiring spoken training data from target domains.

Result: Experiments show the proposed method achieves good cross-domain DST performance without relying on spoken training data from target domains, demonstrating efficacy for cross-domain generalization.

Conclusion: Joint training on spoken and textual DST data enables effective cross-domain generalization for spoken dialogue state tracking, overcoming the data scarcity problem and reducing dependency on costly spoken DST annotations for each target domain.

Abstract: End-to-end spoken dialogue state tracking (DST) is made difficult by the tandem of having to handle speech input and data scarcity. Combining speech foundation encoders and large language models has been proposed in recent work as to alleviate some of this difficulty. Although this approach has been shown to result in strong spoken DST models, achieving state-of-the-art performance in realistic multi-turn DST, it struggles to generalize across domains and requires annotated spoken DST training data for each domain of interest. However, collecting such data for every target domain is both costly and difficult. Noting that textual DST data is more easily obtained for various domains, in this work, we propose jointly training on available spoken DST data and written textual data from other domains as a way to achieve cross-domain generalization. We conduct experiments which show the efficacy of our proposed method for getting good cross-domain DST performance without relying on spoken training data from the target domains.

Method: Conducted fine-tuning experiments on speech recognition tasks across 18 languages using 15 large pre-trained wav2vec2-based models to evaluate cross-lingual transfer performance.

Result: Data diversity during pre-training matters more than data size for final performance. Indo-European languages outperform non-Indo-European languages. Positive cross-lingual transfer occurs with monolingual models, especially when pre-training and downstream languages are similar.

Conclusion: The findings provide guidance for using existing wav2vec2-based models and training new ones, highlighting the importance of data diversity and language similarity for effective cross-lingual transfer in speech recognition tasks.

Abstract: Using representations provided by a large pre-trained model has become the primary strategy for achieving state-of-the-art results in a wide range of tasks. A recently proposed large pre-trained model, wav2vec 2.0, was seminal for several other works on pre-training large models on speech data. Many models are being pre-trained using the same architecture as wav2vec 2.0 and are getting state-of-the-art in various speech-related tasks. Previous work has demonstrated that the data used during the pre-training of these wav2vec2-based models can impact the model’s performance in downstream tasks, and this should be taken into consideration before utilizing these models. However, few works have proposed investigating further how the transfer knowledge of these pre-trained models behaves in different languages, even when the target language differs from the one used during the model’s pre-training. Our work aims to investigate the cross-lingual transferability of these wav2vec2-based models. We performed several fine-tuning experiments on the speech recognition task in 18 languages using 15 large pre-trained models. The results of our experiments showed us that the size of data used during the pre-training of these models is not as important to the final performance as the diversity. We noticed that the performance of Indo-European languages is superior to non-Indo-European languages in the evaluated models. We have observed a positive cross-lingual transfer of knowledge using monolingual models, which was evident in all the languages we used, but more pronounced when the language used during pre-training was more similar to the downstream task language. With these findings, we aim to assist the scientific community in utilizing existing wav2vec2-based pre-trained models, as well as facilitate the pre-training of new ones.

Method: Insight-A uses hierarchical reasoning with two key components: 1) Cross-attribution prompting (CAP) to attribute misinformation to forgery sources based on generation patterns, modeling perception-reasoning correlations; 2) Automatic attribution-debiased prompting (ADP) to reduce human annotation subjectivity. Also includes image captioning (IC) for visual detail extraction to enhance cross-modal consistency checking.

Result: Extensive experiments demonstrate the superiority of Insight-A over existing methods, providing a new paradigm for multimodal misinformation detection in the AIGC era.

Conclusion: Insight-A effectively addresses AIGC-generated misinformation by incorporating attribution analysis with MLLMs, offering a robust framework that considers forgery sources and cross-modal distortions for more accurate detection.

Abstract: AI-generated content (AIGC) technology has emerged as a prevalent alternative to create multimodal misinformation on social media platforms, posing unprecedented threats to societal safety. However, standard prompting leverages multimodal large language models (MLLMs) to identify the emerging misinformation, which ignores the misinformation attribution. To this end, we present Insight-A, exploring attribution with MLLM insights for detecting multimodal misinformation. Insight-A makes two efforts: I) attribute misinformation to forgery sources, and II) an effective pipeline with hierarchical reasoning that detects distortions across modalities. Specifically, to attribute misinformation to forgery traces based on generation patterns, we devise cross-attribution prompting (CAP) to model the sophisticated correlations between perception and reasoning. Meanwhile, to reduce the subjectivity of human-annotated prompts, automatic attribution-debiased prompting (ADP) is used for task adaptation on MLLMs. Additionally, we design image captioning (IC) to achieve visual details for enhancing cross-modal consistency checking. Extensive experiments demonstrate the superiority of our proposal and provide a new paradigm for multimodal misinformation detection in the era of AIGC.

Method: NRPA-GD uses LLMs to simulate user and system behaviors simultaneously, constructs a complete evaluation mechanism for dialogue trajectories, and employs nested Monte Carlo simulation with policy self-adaptation to dynamically adjust policies during dialogue.

Result: NRPA-GD outperforms both prompt engineering and pre-trained model-based methods on four typical goal-oriented dialogue datasets, surpassing ChatGPT and pre-trained policy models with only a 0.6-billion-parameter LLM.

Conclusion: The approach demonstrates the advantages of employing planning methods on LLMs to solve practical planning tasks, showing that effective goal-oriented dialogue can be achieved without training specific models.

Abstract: In goal-oriented dialogue tasks, the main challenge is to steer the interaction towards a given goal within a limited number of turns. Existing approaches either rely on elaborate prompt engineering, whose effectiveness is heavily dependent on human experience, or integrate policy networks and pre-trained policy models, which are usually difficult to adapt to new dialogue scenarios and costly to train. Therefore, in this paper, we present Nested Rollout Policy Adaptation for Goal-oriented Dialogue (NRPA-GD), a novel dialogue policy planning method that completely avoids specific model training by utilizing a Large Language Model (LLM) to simulate behaviors of user and system at the same time. Specifically, NRPA-GD constructs a complete evaluation mechanism for dialogue trajectories and employs an optimization framework of nested Monte Carlo simulation and policy self-adaptation to dynamically adjust policies during the dialogue process. The experimental results on four typical goal-oriented dialogue datasets show that NRPA-GD outperforms both existing prompt engineering and specifically pre-trained model-based methods. Impressively, NRPA-GD surpasses ChatGPT and pre-trained policy models with only a 0.6-billion-parameter LLM. The proposed approach further demonstrates the advantages and novelty of employing planning methods on LLMs to solve practical planning tasks.

Method: Tested both general-purpose and fine-tuned tabular LLMs by prompting them with poor-quality datasets to perform data profiling and cleaning tasks, comparing results with traditional data preparation tools, and using a custom-designed quality model validated through user studies.

Result: LLMs demonstrated effective capabilities in supporting data preparation tasks, with performance comparable to traditional data preparation tools when evaluated using the practitioner-validated quality model.

Conclusion: Large language models show promise in effectively supporting users with data preparation tasks, offering automation capabilities that can complement or potentially replace traditional tools for certain data profiling and cleaning operations.

Abstract: Large language models have recently demonstrated their exceptional capabilities in supporting and automating various tasks. Among the tasks worth exploring for testing large language model capabilities, we considered data preparation, a critical yet often labor-intensive step in data-driven processes. This paper investigates whether large language models can effectively support users in selecting and automating data preparation tasks. To this aim, we considered both general-purpose and fine-tuned tabular large language models. We prompted these models with poor-quality datasets and measured their ability to perform tasks such as data profiling and cleaning. We also compare the support provided by large language models with that offered by traditional data preparation tools. To evaluate the capabilities of large language models, we developed a custom-designed quality model that has been validated through a user study to gain insights into practitioners’ expectations.

Method: Three contributions: 1) Unsupervised label-free Permutation Bias Metric (PBM) to quantify prediction inconsistencies across answer permutations; 2) Batch Question-Context KV caching (BaQCKV) for efficient majority voting; 3) Unsupervised Low-Rank Adaptation (LoRA-1) fine-tuning strategy using PBM and BaQCKV.

Result: Experiments across multiple MCQ benchmarks show approaches reduce bias, increase consistency in accuracy while minimizing computational costs.

Conclusion: Proposed methods provide more precise bias measurement and efficient mitigation strategies, addressing limitations of existing approaches and improving reliability of MCQ evaluation for LLMs.

Abstract: Multiple Choice Question (MCQ) answering is a widely used method for evaluating the performance of Large Language Models (LLMs). However, LLMs often exhibit selection bias in MCQ tasks, where their choices are influenced by factors like answer position or option symbols rather than the content. This bias undermines the reliability of MCQ as an evaluation framework. Most existing selection bias metrics require answer labels and measure divergences between prediction and answer distributions, but do not fully capture the consistency of a model’s predictions across different orderings of answer choices. Existing selection bias mitigation strategies have notable limitations: majority voting, though effective, is computationally prohibitive; calibration-based methods require validation sets and often fail to generalize across datasets. To address these gaps, we propose three key contributions: (1) a new unsupervised label-free Permutation Bias Metric (PBM) that directly quantifies inconsistencies in model predictions across answer permutations, providing a more precise measure of selection bias, (2) an efficient majority voting approach called Batch Question-Context KV caching (BaQCKV), to significantly reduce computational costs while preserving bias mitigation effectiveness, and (3) an unsupervised Low-Rank Adaptation (LoRA-1) fine-tuning strategy based on our proposed metric and the BaQCKV that mitigates selection bias, providing a computationally efficient alternative that maintains model generalizability. Experiments across multiple MCQ benchmarks demonstrate that our approaches reduce bias, increasing consistency in accuracy while minimizing computational costs.

Method: The study uses bias-specific benchmarks (StereoSet and CrowSPairs) to evaluate biases in various generative models (BERT, GPT 3.5, ADA). A three-pronged approach is adopted to detect both explicit and implicit biases. Enhancement strategies include fine-tuning, different prompting techniques, and data augmentation of bias benchmarks.

Result: Fine-tuned models struggle with gender biases but excel at identifying and avoiding racial biases. LLMs often over-rely on keywords in prompts rather than truly understanding content. Enhancement strategies showed promising results: fine-tuned models exhibited adaptability during cross-dataset testing and significantly improved performance on implicit bias benchmarks with gains up to 20%.

Conclusion: The study highlights the need to address biases in LLMs and demonstrates that targeted enhancement strategies can improve bias detection. However, LLMs still show limitations in truly understanding content and exhibit varying performance across different types of biases, indicating ongoing challenges in achieving fair and unbiased AI systems.

Abstract: Large Language models (LLMs), such as ChatGPT, have gained popularity in recent years with the advancement of Natural Language Processing (NLP), with use cases spanning many disciplines and daily lives as well. LLMs inherit explicit and implicit biases from the datasets they were trained on; these biases can include social, ethical, cultural, religious, and other prejudices and stereotypes. It is important to comprehensively examine such shortcomings by identifying the existence and extent of such biases, recognizing the origin, and attempting to mitigate such biased outputs to ensure fair outputs to reduce harmful stereotypes and misinformation. This study inspects and highlights the need to address biases in LLMs amid growing generative Artificial Intelligence (AI). We utilize bias-specific benchmarks such StereoSet and CrowSPairs to evaluate the existence of various biases in many different generative models such as BERT, GPT 3.5, and ADA. To detect both explicit and implicit biases, we adopt a three-pronged approach for thorough and inclusive analysis. Results indicate fine-tuned models struggle with gender biases but excel at identifying and avoiding racial biases. Our findings also illustrated that despite some cases of success, LLMs often over-rely on keywords in prompts and its outputs. This demonstrates the incapability of LLMs to attempt to truly understand the accuracy and authenticity of its outputs. Finally, in an attempt to bolster model performance, we applied an enhancement learning strategy involving fine-tuning, models using different prompting techniques, and data augmentation of the bias benchmarks. We found fine-tuned models to exhibit promising adaptability during cross-dataset testing and significantly enhanced performance on implicit bias benchmarks, with performance gains of up to 20%.

Method: Combines dual-channel retrieval and LLM-driven disclosure analysis over ESG reports, with an interactive dashboard and chatbot for exploration, benchmarking, and explanation.

Result: EulerESG can automatically populate standard-aligned metric tables with high fidelity (up to 0.95 average accuracy) while remaining practical in end-to-end runtime, tested on four globally recognized companies and twelve SASB sub-industries.

Conclusion: The system demonstrates that LLM-powered approaches with explicit ESG framework awareness can effectively automate ESG disclosure analysis, making structured ESG data extraction more accessible and accurate.

Abstract: Environmental, Social, and Governance (ESG) reports have become central to how companies communicate climate risk, social impact, and governance practices, yet they are still published primarily as long, heterogeneous PDF documents. This makes it difficult to systematically answer seemingly simple questions. Existing tools either rely on brittle rule-based extraction or treat ESG reports as generic text, without explicitly modelling the underlying reporting standards. We present \textbf{EulerESG}, an LLM-powered system for automating ESG disclosure analysis with explicit awareness of ESG frameworks. EulerESG combines (i) dual-channel retrieval and LLM-driven disclosure analysis over ESG reports, and (ii) an interactive dashboard and chatbot for exploration, benchmarking, and explanation. Using four globally recognised companies and twelve SASB sub-industries, we show that EulerESG can automatically populate standard-aligned metric tables with high fidelity (up to 0.95 average accuracy) while remaining practical in end-to-end runtime, and we compare several recent LLM models in this setting. The full implementation, together with a demonstration video, is publicly available at https://github.com/UNSW-database/EulerESG.

Method: GPS uses reinforcement learning to train a prompter on a suite of training tasks with novel regularization for per-sample adaptation, and employs Minimum Bayes Risk decoding for stable inference.

Result: GPS achieves competitive performance: second best on text simplification, third best on summarization, on-par on classification without task-specific training, and state-of-the-art on GSM8K for in-domain prompting.

Conclusion: GPS demonstrates a novel effective paradigm for automatic prompting that generates adaptive, input-specific prompts without extensive optimization or task-specific training sets.

Abstract: LLMs are sensitive to prompting, with task performance often hinging on subtle, sometimes imperceptible variations in phrasing. As a result, crafting effective prompts manually remains challenging and time-consuming. Recent automatic prompting methods mitigate this difficulty but face three key limitations: (i) for each new task, they require large datasets to train good prompts;(ii) they rely on costly optimization loops that may take hours; (iii)they typically produce a single task-level prompt that does not adapt to the individual input problem to be solved. We propose GPS, the first general-purpose, per-sample prompting method. Without any task-specific tuning, GPS generates a tailored prompt for each unseen input, improving performance across diverse tasks. The prompter is trained with reinforcement learning on a suite of training tasks and includes a novel regularization for effectively adapting to per-sample prompting. Finally, we employ Minimum Bayes Risk decoding to stabilize inference. Empirically, GPS demonstrates competitive performance: we attain second best results among baselines on text simplification, third best results on summarization and on-par results on classification, while not training on any of these tasks, in contrast to the baselines. For in-domain prompting, we obtain sota on GSM8K. Our work shows the potential of a novel and effective paradigm for automatic prompting: generating adaptive, input-specific prompts without extensive optimization and without access to a task-specific training set. Our code is available at https://github.com/Batorskq/GPS.

Method: Advanced text preprocessing, multi-modal feature extraction, Harris Hawks Optimization (HHO) for feature selection, and a stacking ensemble classifier approach.

Result: Achieved 95.40% accuracy, 92.81% precision, 95.01% recall, and 93.90% F1-Score on 40,432 reviews dataset, with HHO reducing features from 13,539 to 1,368 (89.9% reduction).

Conclusion: The combination of ensemble learning and bio-inspired optimization is effective for machine-generated text recognition, with privacy-preserving techniques needed for cloud-based deployment.

Abstract: It is well known that fraudulent reviews cast doubt on the legitimacy and dependability of online purchases. The most recent development that leads customers towards darkness is the appearance of human reviews in computer-generated (CG) ones. In this work, we present an advanced machine-learning-based system that analyses these reviews produced by AI with remarkable precision. Our method integrates advanced text preprocessing, multi-modal feature extraction, Harris Hawks Optimization (HHO) for feature selection, and a stacking ensemble classifier. We implemented this methodology on a public dataset of 40,432 Original (OR) and Computer-Generated (CG) reviews. From an initial set of 13,539 features, HHO selected the most applicable 1,368 features, achieving an 89.9% dimensionality reduction. Our final stacking model achieved 95.40% accuracy, 92.81% precision, 95.01% recall, and a 93.90% F1-Score, which demonstrates that the combination of ensemble learning and bio-inspired optimisation is an effective method for machine-generated text recognition. Because large-scale review analytics commonly run on cloud platforms, privacy-preserving techniques such as differential approaches and secure outsourcing are essential to protect user data in these systems.

Method: The authors introduce CrossCheck-Bench, a diagnostic benchmark with 15k question-answer pairs sourced from real-world artifacts with synthetically injected contradictions. The benchmark uses a hierarchical task framework covering three levels of reasoning complexity and defines seven atomic capabilities for resolving cross-modal inconsistencies. Construction involved a multi-stage annotation pipeline with 450+ expert hours to ensure semantic validity and calibrated difficulty.

Result: Evaluation of 13 state-of-the-art vision-language models shows consistent performance drops as tasks shift from perceptual matching to logical contradiction detection. Most models perform well on isolated entity recognition but fail when synthesizing multiple clues for conflict reasoning. Capability analysis reveals uneven skill acquisition, especially in multi-step inference or rule-based validation tasks. Conventional prompting strategies yield only marginal gains, while methods interleaving symbolic reasoning with visual processing achieve more stable improvements.

Conclusion: The results highlight a persistent bottleneck in multimodal reasoning and suggest new directions for building models capable of robust cross-modal verification. The benchmark reveals significant gaps in current models’ ability to handle real-world inconsistencies, pointing to the need for improved reasoning architectures beyond surface-level alignment.

Abstract: Multimodal Large Language Models are primarily trained and evaluated on aligned image-text pairs, which leaves their ability to detect and resolve real-world inconsistencies largely unexplored. In open-domain applications visual and textual cues often conflict, requiring models to perform structured reasoning beyond surface-level alignment. We introduce CrossCheck-Bench, a diagnostic benchmark for evaluating contradiction detection in multimodal inputs. The benchmark adopts a hierarchical task framework covering three levels of reasoning complexity and defines seven atomic capabilities essential for resolving cross-modal inconsistencies. CrossCheck-Bench includes 15k question-answer pairs sourced from real-world artifacts with synthetically injected contradictions. The dataset is constructed through a multi-stage annotation pipeline involving more than 450 expert hours to ensure semantic validity and calibrated difficulty across perception, integration, and reasoning. We evaluate 13 state-of-the-art vision-language models and observe a consistent performance drop as tasks shift from perceptual matching to logical contradiction detection. Most models perform well on isolated entity recognition but fail when multiple clues must be synthesized for conflict reasoning. Capability-level analysis further reveals uneven skill acquisition, especially in tasks requiring multi-step inference or rule-based validation. Additional probing shows that conventional prompting strategies such as Chain-of-Thought and Set-of-Mark yield only marginal gains. By contrast, methods that interleave symbolic reasoning with grounded visual processing achieve more stable improvements. These results highlight a persistent bottleneck in multimodal reasoning and suggest new directions for building models capable of robust cross-modal verification.

Method: MICM uses conceptual morphology theory to encode specific configurations of nuanced concepts into fixed prompt templates via predefined phrases that act as conceptual triggers, steering model outputs toward specific value stances without triggering conventional safety filters.

Result: MICM consistently outperforms state-of-the-art jailbreak techniques across five advanced LLMs (GPT-4o, Deepseek-R1, Qwen3-8B), achieving high success rates with minimal rejection.

Conclusion: Commercial LLMs have critical vulnerability: their safety mechanisms remain susceptible to covert manipulation of underlying value alignment through abstract conceptual triggers.

Abstract: Recent research on large language model (LLM) jailbreaks has primarily focused on techniques that bypass safety mechanisms to elicit overtly harmful outputs. However, such efforts often overlook attacks that exploit the model’s capacity for abstract generalization, creating a critical blind spot in current alignment strategies. This gap enables adversaries to induce objectionable content by subtly manipulating the implicit social values embedded in model outputs. In this paper, we introduce MICM, a novel, model-agnostic jailbreak method that targets the aggregate value structure reflected in LLM responses. Drawing on conceptual morphology theory, MICM encodes specific configurations of nuanced concepts into a fixed prompt template through a predefined set of phrases. These phrases act as conceptual triggers, steering model outputs toward a specific value stance without triggering conventional safety filters. We evaluate MICM across five advanced LLMs, including GPT-4o, Deepseek-R1, and Qwen3-8B. Experimental results show that MICM consistently outperforms state-of-the-art jailbreak techniques, achieving high success rates with minimal rejection. Our findings reveal a critical vulnerability in commercial LLMs: their safety mechanisms remain susceptible to covert manipulation of underlying value alignment.

Method: PeerCoPilot is a large language model (LLM)-powered assistant designed to help peer providers with day-to-day tasks. It uses a retrieval-augmented generation pipeline backed by a large database of over 1,300 vetted resources to ensure information reliability. The system helps peer providers create wellness plans, construct step-by-step goals, and locate organizational resources to support these goals.

Result: Human evaluations with 15 peer providers and 6 service users showed that over 90% of users supported using PeerCoPilot. The system was demonstrated to provide more reliable and specific information than a baseline LLM. PeerCoPilot is currently being used by 5-10 peer providers at CSPNJ, a large behavioral health organization serving over 10,000 service users, with active expansion underway.

Conclusion: PeerCoPilot represents a successful application of LLM technology to support peer providers in behavioral health organizations, addressing resource constraints while maintaining information reliability through a vetted database. The high user acceptance and demonstrated superiority over baseline LLMs suggest this approach can effectively augment peer provider capabilities in real-world settings.

Abstract: Behavioral health conditions, which include mental health and substance use disorders, are the leading disease burden in the United States. Peer-run behavioral health organizations (PROs) critically assist individuals facing these conditions by combining mental health services with assistance for needs such as income, employment, and housing. However, limited funds and staffing make it difficult for PROs to address all service user needs. To assist peer providers at PROs with their day-to-day tasks, we introduce PeerCoPilot, a large language model (LLM)-powered assistant that helps peer providers create wellness plans, construct step-by-step goals, and locate organizational resources to support these goals. PeerCoPilot ensures information reliability through a retrieval-augmented generation pipeline backed by a large database of over 1,300 vetted resources. We conducted human evaluations with 15 peer providers and 6 service users and found that over 90% of users supported using PeerCoPilot. Moreover, we demonstrated that PeerCoPilot provides more reliable and specific information than a baseline LLM. PeerCoPilot is now used by a group of 5-10 peer providers at CSPNJ, a large behavioral health organization serving over 10,000 service users, and we are actively expanding PeerCoPilot’s use.

Method: Developed Mina using multilingual embeddings and a RAG-based chain-of-tools framework for retrieval, reasoning, translation, and document generation. Features interactive chat interface for legal drafts, citations, and plain-language explanations.

Result: Evaluated by law faculty from leading Bangladeshi universities on 2022-2023 Bar Council Exams: scored 75-80% in Preliminary MCQs, Written, and simulated Viva Voce, matching/surpassing average human performance. Operates at 0.12-0.61% of typical legal consultation costs (99.4-99.9% reduction).

Conclusion: Mina demonstrates potential as a low-cost, multilingual AI assistant that automates key legal tasks and scales access to justice. Provides a real-world case study for building domain-specific, low-resource systems addressing multilingual adaptation, efficiency, and sustainable public-service AI deployment.

Abstract: Bangladesh’s low-income population faces major barriers to affordable legal advice due to complex legal language, procedural opacity, and high costs. Existing AI legal assistants lack Bengali-language support and jurisdiction-specific adaptation, limiting their effectiveness. To address this, we developed Mina, a multilingual LLM-based legal assistant tailored for the Bangladeshi context. It employs multilingual embeddings and a RAG-based chain-of-tools framework for retrieval, reasoning, translation, and document generation, delivering context-aware legal drafts, citations, and plain-language explanations via an interactive chat interface. Evaluated by law faculty from leading Bangladeshi universities across all stages of the 2022 and 2023 Bangladesh Bar Council Exams, Mina scored 75-80% in Preliminary MCQs, Written, and simulated Viva Voce exams, matching or surpassing average human performance and demonstrating clarity, contextual understanding, and sound legal reasoning. Even under a conservative upper bound, Mina operates at just 0.12-0.61% of typical legal consultation costs in Bangladesh, yielding a 99.4-99.9% cost reduction relative to human-provided services. These results confirm its potential as a low-cost, multilingual AI assistant that automates key legal tasks and scales access to justice, offering a real-world case study on building domain-specific, low-resource systems and addressing challenges of multilingual adaptation, efficiency, and sustainable public-service AI deployment.

Method: Built GGP collection from German General Social Survey (ALLBUS) data, creating comprehensive persona prompts designed to be easily integrated into LLM prompts for various tasks.

Result: GGP-guided LLMs outperform state-of-the-art classifiers in simulating survey response distributions, especially under data scarcity. Analysis shows persona representativity and attribute selection significantly affect alignment with population responses.

Conclusion: GGP provides a valuable resource for LLM-based social simulations, enabling more systematic exploration of population-aligned persona prompting in NLP and social science research.

Abstract: The use of Large Language Models (LLMs) for simulating human perspectives via persona prompting is gaining traction in computational social science. However, well-curated, empirically grounded persona collections remain scarce, limiting the accuracy and representativeness of such simulations. Here we introduce the German General Personas (GGP) collection, a comprehensive and representative persona prompt collection built from the German General Social Survey (ALLBUS). The GGP and its persona prompts are designed to be easily plugged into prompts for all types of LLMs and tasks, steering models to generate responses aligned with the underlying German population. We evaluate GGP by prompting various LLMs to simulate survey response distributions across diverse topics, demonstrating that GGP-guided LLMs outperform state-of-the-art classifiers, particularly under data scarcity. Furthermore, we analyze how the representativity and attribute selection within persona prompts affect alignment with population responses. Our findings suggest that GGP provides a potentially valuable resource for research on LLM-based social simulations that enables more systematic explorations of population-aligned persona prompting in NLP and social science research.

Method: Extracted high-frequency concepts from 1,500+ AD trials on ClinicalTrials.gov, organized into 7 semantic categories, annotated with standard biomedical vocabularies (UMLS, OMOP, DrugBank, etc.), and optimized using Jenks Natural Breaks method for coverage-manageability balance.

Result: AD-CDO achieved over 63% coverage of extracted trial concepts while maintaining interpretability and compactness. Demonstrated utility through ontology-driven trial simulation system and entity normalization task mapping clinical text to ontology-aligned terms.

Conclusion: AD-CDO provides a versatile foundation for ontology-driven AD clinical trial research by harmonizing essential eligibility entities with standardized vocabularies, supporting phenotyping, cohort identification, and structured data integration.

Abstract: Objective This study introduces the Alzheimer’s Disease Common Data Element Ontology for Clinical Trials (AD-CDO), a lightweight, semantically enriched ontology designed to represent and standardize key eligibility criteria concepts in Alzheimer’s disease (AD) clinical trials. Materials and Methods We extracted high-frequency concepts from more than 1,500 AD clinical trials on ClinicalTrials.gov and organized them into seven semantic categories: Disease, Medication, Diagnostic Test, Procedure, Social Determinants of Health, Rating Criteria, and Fertility. Each concept was annotated with standard biomedical vocabularies, including the UMLS, OMOP Standardized Vocabularies, DrugBank, NDC, and NLM VSAC value sets. To balance coverage and manageability, we applied the Jenks Natural Breaks method to identify an optimal set of representative concepts. Results The optimized AD-CDO achieved over 63% coverage of extracted trial concepts while maintaining interpretability and compactness. The ontology effectively captured the most frequent and clinically meaningful entities used in AD eligibility criteria. We demonstrated AD-CDO’s practical utility through two use cases: (a) an ontology-driven trial simulation system for formal modeling and virtual execution of clinical trials, and (b) an entity normalization task mapping raw clinical text to ontology-aligned terms, enabling consistency and integration with EHR data. Discussion AD-CDO bridges the gap between broad biomedical ontologies and task-specific trial modeling needs. It supports multiple downstream applications, including phenotyping algorithm development, cohort identification, and structured data integration. Conclusion By harmonizing essential eligibility entities and aligning them with standardized vocabularies, AD-CDO provides a versatile foundation for ontology-driven AD clinical trial research.

Method: PromptTailor uses a quantized Llama3-8B model fine-tuned with a lightweight LoRA adapter on 12,300 prompt-refinement dialogues spanning 41 everyday domains. The dialogues are distilled from three stronger LLMs, enabling the system to expand minimal user instructions into rich, domain-aware prompts while preserving user preferences.

Result: In human and LLM-judge evaluations across multiple target models and optimization baselines, PromptTailor yields higher preference rates than chain-of-thought prompting and matches or surpasses state-of-the-art prompt optimization methods while requiring fewer model calls (e.g., 3 vs. 9).

Conclusion: A compact student model guided by powerful teachers can learn effective prompt-generation strategies that enhance response quality while maintaining alignment with user intent, making it suitable for edge deployment in on-device and privacy-sensitive applications.

Abstract: Lightweight language models remain attractive for on-device and privacy-sensitive applications, but their responses are highly sensitive to prompt quality. For open-ended generation, non-expert users often lack the knowledge or time to consistently craft high-quality prompts, leading them to rely on prompt optimization tools. However, a key challenge is ensuring the optimized prompts genuinely align with users’ original intents and preferences. We introduce PromptTailor, a system for controllable prompt generation for open-ended text that improves model output quality by intent-aligned prompt synthesis. PromptTailor expands minimal user instructions into rich, domain-aware prompts while preserving the user’s stated preferences. The system is a quantized Llama3-8B model fine-tuned with a lightweight LoRA adapter on 12,300 prompt-refinement dialogues spanning 41 everyday domains, distilled from three stronger LLMs. The adapter attaches to any Llama3-8B base, enabling edge deployment. In human and LLM-judge evaluations across multiple target models and optimization baselines, PromptTailor yields higher preference rates than chain-of-thought prompting and matches or surpasses state-of-the-art prompt optimization methods while requiring fewer model calls (e.g., 3 vs. 9). These results show that a compact student, guided by powerful teachers, can learn effective prompt-generation strategies that enhance response quality while maintaining alignment with user intent.

Method: SUMER (Search in Uncompressed Memory via Experience Replay) is an end-to-end reinforcement learning agent with verifiable reward (RLVR) that learns to use search tools to gather information from uncompressed memory and answer target questions.

Result: On the LoCoMo dataset for long-context conversation understanding, SUMER with Qwen2.5-7B-Instruct outperformed all biased memory compression approaches and the full-context baseline, achieving 43% gain over prior best (SOTA performance).

Conclusion: Simple search on raw data outperforms goal-agnostic and biased compression algorithms, arguing for new paradigms and benchmarks that are more dynamic and autonomously scalable for long-context memory tasks.

Abstract: How to enable human-like long-term memory in large language models (LLMs) has been a central question for unlocking more general capabilities such as few-shot generalization. Existing memory frameworks and benchmarks focus on finding the optimal memory compression algorithm for higher performance in tasks that require recollection and sometimes further reasoning. However, such efforts have ended up building more human bias into the compression algorithm, through the search for the best prompts and memory architectures that suit specific benchmarks, rather than finding a general solution that would work on other data distributions. On the other hand, goal-directed search on uncompressed information could potentially exhibit superior performance because compression is lossy, and a predefined compression algorithm will not fit all raw data distributions. Here we present SUMER (Search in Uncompressed Memory via Experience Replay), an end-to-end reinforcement learning agent with verifiable reward (RLVR) that learns to use search tools to gather information and answer a target question. On the LoCoMo dataset for long-context conversation understanding, SUMER with Qwen2.5-7B-Instruct learned to use search tools and outperformed all other biased memory compression approaches and also the full-context baseline, reaching SOTA performance (43% gain over the prior best). We demonstrate that a simple search method applied to raw data outperforms goal-agnostic and biased compression algorithms in current long-context memory tasks, arguing for new paradigms and benchmarks that are more dynamic and autonomously scalable. Code for SUMER and all implemented baselines is publicly available at https://github.com/zycyc/SUMER.

Method: Three-component architecture: 1) Proactive Knowledge Grounding Network (PKGN) for continuous multimodal context updates, 2) Emotion-Intent Alignment Model (EIAM) for joint modeling of user emotion and purchase intent, and 3) Reinforced Discourse Loop (RDL) for optimizing emotional coherence via reinforcement learning from user responses.

Result: Outperforms strong LLM baselines on two marketing dialogue datasets (MM-ConvMarket and AffectPromo) with +26% emotional consistency, +19% persuasive success rate, and +23% long-term user engagement.

Conclusion: Emotion-grounded proactivity is a key capability for commercial multimodal agents, enabling more effective and engaging marketing conversations through continuous affective reasoning and adaptive persuasion strategies.

Abstract: Recent advances in large language models (LLMs) have enabled fluent dialogue systems, but most remain reactive and struggle in emotionally rich, goal-oriented settings such as marketing conversations. To address this limitation, we propose AffectMind, a multimodal affective dialogue agent that performs proactive reasoning and dynamic knowledge grounding to sustain emotionally aligned and persuasive interactions. AffectMind combines three components: a Proactive Knowledge Grounding Network (PKGN) that continuously updates factual and affective context from text, vision, and prosody; an Emotion–Intent Alignment Model (EIAM) that jointly models user emotion and purchase intent to adapt persuasion strategies; and a Reinforced Discourse Loop (RDL) that optimizes emotional coherence and engagement via reinforcement signals from user responses. Experiments on two newly curated marketing dialogue datasets, MM-ConvMarket and AffectPromo, show that AffectMind outperforms strong LLM-based baselines in emotional consistency (+26%), persuasive success rate (+19%), and long-term user engagement (+23%), highlighting emotion-grounded proactivity as a key capability for commercial multimodal agents.

Method: Conducted ablation studies on 50 queries (15 answerable, 10 edge cases, 25 adversarial) to test components like hybrid retrieval, ensemble verification, and adaptive thresholding both in isolation and combined.

Result: Individual enhancements provided almost no benefit alone, but together achieved 95% reduction in abstention (from 40% to 2%) without increasing hallucinations. Also identified measurement challenges where different verification strategies create labeling inconsistencies that artificially inflate hallucination rates.

Conclusion: Synergistic integration matters more than individual component strength; standardized metrics and labels are essential for correct performance interpretation; adaptive calibration is needed to prevent overconfident answering even with high retrieval quality.

Abstract: Building reliable retrieval-augmented generation (RAG) systems requires more than adding powerful components; it requires understanding how they interact. Using ablation studies on 50 queries (15 answerable, 10 edge cases, and 25 adversarial), we show that enhancements such as hybrid retrieval, ensemble verification, and adaptive thresholding provide almost no benefit when used in isolation, yet together achieve a 95% reduction in abstention (from 40% to 2%) without increasing hallucinations. We also identify a measurement challenge: different verification strategies can behave safely but assign inconsistent labels (for example, “abstained” versus “unsupported”), creating apparent hallucination rates that are actually artifacts of labeling. Our results show that synergistic integration matters more than the strength of any single component, that standardized metrics and labels are essential for correctly interpreting performance, and that adaptive calibration is needed to prevent overconfident over-answering even when retrieval quality is high.

Method: Automated multi-agent architecture collects data from Telegram shopping channels, with specialized agents for question generation, validation, and refinement. Advanced agentic system uses multi-query retrieval, reranking, and persona-aligned response synthesis.

Result: The agentic architecture outperformed traditional RAG models in 4 out of 5 diverse channels when evaluated by GPT-5.1 across six quality dimensions, demonstrating superior performance without expensive human annotation.

Conclusion: MegaChat provides SMEs with a cost-effective solution for building intelligent customer engagement systems in Persian, enabling advancements in multilingual conversational AI for low-resource languages.

Abstract: Small and medium-sized enterprises (SMEs) in Iran increasingly leverage Telegram for sales, where real-time engagement is essential for conversion. However, developing AI-driven chatbots for this purpose requires large, high-quality question-and-answer (Q&A) datasets, which are typically expensive and resource-intensive to produce, especially for low-resource languages like Persian. In this paper, we introduce MegaChat, the first fully synthetic Persian Q&A dataset designed to evaluate intelligent sales chatbots in Telegram-based e-commerce. We propose a novel, automated multi-agent architecture that generates persona-aware Q&A pairs by collecting data from active Telegram shopping channels. The system employs specialized agents for question generation, validation, and refinement, ensuring the production of realistic and diverse conversational data. To evaluate answer generation, we compare three classic retrieval-augmented generation (RAG) models with our advanced agentic system, which features multi-query retrieval, reranking, and persona-aligned response synthesis. Using GPT-5.1 for evaluation across six quality dimensions, our results show that the agentic architecture outperformed traditional RAG models in 4 out of 5 diverse channels, demonstrating its ability to generate scalable, high-quality datasets without relying on expensive human annotation or complex fine-tuning. MegaChat provides SMEs with an efficient, cost-effective solution for building intelligent customer engagement systems in specialized commercial domains, enabling advancements in multilingual conversational AI for low-resource languages. Download: https://github.com/MegaChat-Tech/MegaChat-DataSet

Method: The authors create the first benchmark isolating procedural memory retrieval from task execution using ALFWorld. They construct dual corpora of expert and LLM-generated trajectories and evaluate six retrieval methods with systematically stratified queries to test generalization across different object instantiations.

Result: Results show a clear generalization cliff: embedding-based methods perform strongly on familiar contexts but degrade considerably on novel ones, while LLM-generated procedural abstractions demonstrate reliable cross-context transfer. Embeddings fundamentally treat procedures as unordered bags of words, discarding temporal structure necessary for generalization.

Conclusion: The benchmark provides the first diagnostic framework to separate genuine procedural understanding from surface-level memorization, revealing that corpus scale delivers larger gains than representation enrichment and exposing architectural ceilings in current encoders.

Abstract: Current AI agents excel in familiar settings, but fail sharply when faced with novel tasks with unseen vocabularies – a core limitation of procedural memory systems. We present the first benchmark that isolates procedural memory retrieval from task execution, evaluating whether agents can recognize functionally equivalent procedures that span different object instantiations. Using ALFWorld, we construct dual corpora of expert and LLM-generated trajectories and evaluate six retrieval methods using systematically stratified queries. Our results expose a clear generalization cliff: embedding-based methods perform strongly on familiar contexts, yet degrade considerably on novel ones, while LLM-generated procedural abstractions demonstrate reliable cross-context transfer. Controlled ablations show that although embeddings capture some lexical-level abstraction, they fundamentally treat procedures as unordered bags of words, discarding temporal structure necessary for cross-context transfer. Corpus scale delivers far larger gains than representation enrichment, revealing an architectural ceiling in current encoders. Our benchmark offers the first diagnostic framework separating genuine procedural understanding from surface-level memorization and gives tools for developing retrieval systems capable of dependable generalization. Resources available at our GitHub repository (https://github.com/qpiai/Proced_mem_bench).

Method: Two cognitive tests using ChatGPT and Gemini: (1) testing Bell’s inequality violations to detect quantum entanglement in conceptual combinations, (2) analyzing word distribution statistics in large texts to identify quantum statistics patterns.

Result: LLMs show significant Bell’s inequality violations indicating quantum entanglement in concepts, and exhibit Bose-Einstein statistics (not Maxwell-Boltzmann) in word distributions, matching patterns previously found in human cognition and information retrieval.

Conclusion: Quantum structures systematically emerge in conceptual-linguistic domains for both human and artificial cognitive agents, suggesting evolutionary convergence in meaning organization through vector space semantics rather than neural architecture alone.

Abstract: We present the results of cognitive tests on conceptual combinations, performed using specific Large Language Models (LLMs) as test subjects. In the first test, performed with ChatGPT and Gemini, we show that Bell’s inequalities are significantly violated, which indicates the presence of ‘quantum entanglement’ in the tested concepts. In the second test, also performed using ChatGPT and Gemini, we instead identify the presence of ‘Bose-Einstein statistics’, rather than the intuitively expected ‘Maxwell-Boltzmann statistics’, in the distribution of the words contained in large-size texts. Interestingly, these findings mirror the results previously obtained in both cognitive tests with human participants and information retrieval tests on large corpora. Taken together, they point to the ‘systematic emergence of quantum structures in conceptual-linguistic domains’, regardless of whether the cognitive agent is human or artificial. Although LLMs are classified as neural networks for historical reasons, we believe that a more essential form of knowledge organization takes place in the distributive semantic structure of vector spaces built on top of the neural network. It is this meaning-bearing structure that lends itself to a phenomenon of evolutionary convergence between human cognition and language, slowly established through biological evolution, and LLM cognition and language, emerging much more rapidly as a result of self-learning and training. We analyze various aspects and examples that contain evidence supporting the above hypothesis. We also advance a unifying framework that explains the pervasive quantum organization of meaning that we identify.

Method: HUMORCHAIN is a multi-stage reasoning framework that integrates: 1) visual semantic parsing, 2) humor- and psychology-based reasoning, and 3) a fine-tuned discriminator for humor evaluation. It explicitly embeds cognitive structures from humor theories into an interpretable and controllable reasoning chain.

Result: Experiments on Meme-Image-No-Text, Oogiri-GO, and OxfordTVG-HIC datasets show HUMORCHAIN outperforms state-of-the-art baselines in human humor preference, Elo/BT scores, and semantic diversity, demonstrating theory-driven structured reasoning enables LLMs to generate humor aligned with human perception.

Conclusion: This is the first work to explicitly embed cognitive structures from humor theories into multimodal humor generation, enabling structured reasoning from visual understanding to humor creation, and showing that theory-driven approaches can produce humor that aligns with human perception.

Abstract: Humor, as both a creative human activity and a social binding mechanism, has long posed a major challenge for AI generation. Although producing humor requires complex cognitive reasoning and social understanding, theories of humor suggest that it follows learnable patterns and structures, making it theoretically possible for generative models to acquire them implicitly. In recent years, multimodal humor has become a prevalent form of online communication, especially among Gen Z, highlighting the need for AI systems capable of integrating visual understanding with humorous language generation. However, existing data-driven approaches lack explicit modeling or theoretical grounding of humor, often producing literal descriptions that fail to capture its underlying cognitive mechanisms, resulting in the generated image descriptions that are fluent but lack genuine humor or cognitive depth. To address this limitation, we propose HUMORCHAIN (HUmor-guided Multi-step Orchestrated Reasoning Chain for Image Captioning), a theory-guided multi-stage reasoning framework. It integrates visual semantic parsing, humor- and psychology-based reasoning, and a fine-tuned discriminator for humor evaluation, forming an interpretable and controllable cognitive reasoning chain. To the best of our knowledge, this is the first work to explicitly embed cognitive structures from humor theories into multimodal humor generation, enabling a structured reasoning process from visual understanding to humor creation. Experiments on Meme-Image-No-Text, Oogiri-GO, and OxfordTVG-HIC datasets show that HUMORCHAIN outperforms state-of-the-art baselines in human humor preference, Elo/BT scores, and semantic diversity, demonstrating that theory-driven structured reasoning enables large language models to generate humor aligned with human perception.

Method: RoSA combines two components: 1) RoPE-aware Attention Enhancement (RoAE) module that selectively enhances low-frequency components of RoPE-influenced attention states, and 2) Dynamic Layer Selection (DLS) strategy that adaptively identifies and updates the most critical layers based on LayerNorm gradient norms.

Result: Extensive experiments on fifteen commonsense and arithmetic benchmarks show RoSA outperforms existing mainstream PEFT methods under comparable trainable parameters.

Conclusion: RoSA achieves more targeted and efficient fine-tuning by combining dimension-wise enhancement with layer-wise adaptation, providing a superior PEFT framework for LLM adaptation.

Abstract: Fine-tuning large language models is essential for task-specific adaptation, yet it remains computationally prohibitive. Parameter-Efficient Fine-Tuning (PEFT) methods have emerged as a solution, but current approaches typically ignore the distinct roles of model components and the heterogeneous importance across layers, thereby limiting adaptation efficiency. Motivated by the observation that Rotary Position Embeddings (RoPE) induce critical activations in the low-frequency dimensions of attention states, we propose RoPE-aware Selective Adaptation (RoSA), a novel PEFT framework that allocates trainable parameters in a more targeted and effective manner. RoSA comprises a RoPE-aware Attention Enhancement (RoAE) module, which selectively enhances the low-frequency components of RoPE-influenced attention states, and a Dynamic Layer Selection (DLS) strategy that adaptively identifies and updates the most critical layers based on LayerNorm gradient norms. By combining dimension-wise enhancement with layer-wise adaptation, RoSA achieves more targeted and efficient fine-tuning. Extensive experiments on fifteen commonsense and arithmetic benchmarks demonstrate that RoSA outperforms existing mainstream PEFT methods under comparable trainable parameters. The code is available to ease reproducibility at https://github.com/Applied-Machine-Learning-Lab/RoSA.

Method: Verification-First (VF) strategy prompts models to verify a provided candidate answer (even trivial/random) before generating a solution, triggering reverse reasoning. Iter-VF extends this to sequential test-time scaling with iterative verification-generation cycles using previous answers.

Result: VF with random answer consistently outperforms standard Chain-of-Thought across various benchmarks (mathematical reasoning, coding, agentic tasks) and LLMs (1B to commercial models). Iter-VF outperforms existing test-time scaling strategies.

Conclusion: Verification-first approach provides an effective, low-cost method to enhance LLM reasoning by triggering complementary reverse reasoning processes, offering significant performance improvements with minimal computational overhead.

Abstract: To enhance the reasoning capabilities of Large Language Models (LLMs) without high costs of training, nor extensive test-time sampling, we introduce Verification-First (VF), a strategy that prompts models to verify a provided candidate answer, even a trivial or random one, before generating a solution. This approach triggers a “reverse reasoning” process that is cognitively easier and complementary to standard forward Chain-of-Thought (CoT), effectively invoking the model’s critical thinking to reduce logical errors. We further generalize the VF strategy to Iter-VF, a sequential test-time scaling (TTS) method that iteratively cycles the verification-generation process using the model’s previous answer. Extensive experiments across various benchmarks (from mathematical reasoning to coding and agentic tasks) and various LLMs (from open-source 1B to cutting-edge commercial ones) confirm that VF with random answer consistently outperforms standard CoT with minimal computational overhead, and Iter-VF outperforms existing TTS strategies.

Method: Developed using large-scale multi-site longitudinal dataset (3.1M studies, 6M images from 806k patients from Mayo Clinic). Evaluated on three holdout datasets and MIMIC-CXR. Created novel L&T-specific metrics framework to assess type, longitudinal change, and placement accuracy. Conducted retrospective user evaluation with 9 radiologists reviewing 600 studies blindly.

Result: Significantly improved AI-generated reports over state-of-the-art on lexical quality, clinical correctness, and L&T elements. User study showed comparable critical error rates (3.0% original vs 4.6% AI) and similar acceptable sentence rates (97.8% original vs 97.4% AI), marking significant improvement over prior studies with larger gaps.

Conclusion: MAIRA-X can effectively assist radiologists in high-volume clinical settings by reducing workload while maintaining diagnostic accuracy, particularly for lines and tubes reporting which is demanding and repetitive.

Abstract: AI-assisted report generation offers the opportunity to reduce radiologists’ workload stemming from expanded screening guidelines, complex cases and workforce shortages, while maintaining diagnostic accuracy. In addition to describing pathological findings in chest X-ray reports, interpreting lines and tubes (L&T) is demanding and repetitive for radiologists, especially with high patient volumes. We introduce MAIRA-X, a clinically evaluated multimodal AI model for longitudinal chest X-ray (CXR) report generation, that encompasses both clinical findings and L&T reporting. Developed using a large-scale, multi-site, longitudinal dataset of 3.1 million studies (comprising 6 million images from 806k patients) from Mayo Clinic, MAIRA-X was evaluated on three holdout datasets and the public MIMIC-CXR dataset, where it significantly improved AI-generated reports over the state of the art on lexical quality, clinical correctness, and L&T-related elements. A novel L&T-specific metrics framework was developed to assess accuracy in reporting attributes such as type, longitudinal change and placement. A first-of-its-kind retrospective user evaluation study was conducted with nine radiologists of varying experience, who blindly reviewed 600 studies from distinct subjects. The user study found comparable rates of critical errors (3.0% for original vs. 4.6% for AI-generated reports) and a similar rate of acceptable sentences (97.8% for original vs. 97.4% for AI-generated reports), marking a significant improvement over prior user studies with larger gaps and higher error rates. Our results suggest that MAIRA-X can effectively assist radiologists, particularly in high-volume clinical settings.

Method: Residual Refinement Quantization (R2Q) decomposes 2-bit quantization into two sequential 1-bit sub-quantizations, forming an adaptive quantization lattice with residual learning mechanism for refinement.

Result: R2Q consistently outperforms existing 2-bit quantization methods across Llama, OPT, and Qwen models on diverse benchmarks covering question answering, commonsense reasoning, and language modeling.

Conclusion: R2Q enhances performance, improves training stability, accelerates convergence under extreme compression, and its modular design enables seamless integration with existing QAT frameworks.

Abstract: The rapid progress of Large Language Models (LLMs) has brought substantial computational and memory demands, spurring the adoption of low-bit quantization. While 8-bit and 4-bit formats have become prevalent, extending quantization to 2 bits remains challenging due to severe accuracy degradation. To address this, we propose Residual Refinement Quantization (R2Q)-a novel 2-bit quantization framework that decomposes the process into two sequential 1-bit sub-quantizations, forming an adaptive quantization lattice. Extensive evaluations on Llama, OPT, and Qwen across diverse benchmarks-covering question answering, commonsense reasoning, and language modeling-demonstrate that R2Q consistently outperforms existing 2-bit quantization methods in both fine-grained and coarse-grained settings. By refining quantization through a residual learning mechanism, R2Q enhances performance, improves training stability, and accelerates convergence under extreme compression. Furthermore, its modular design enables seamless integration with existing quantization-aware training (QAT) frameworks.

Method: Introduces Polarity-Aware CCS (PA-CCS) with two alignment metrics (Polar-Consistency and Contradiction Index), tests on 16 language models using curated datasets of matched harmful-safe sentence pairs constructed via different methodologies.

Result: PA-CCS identifies architectural and layer-specific differences in latent harmful knowledge encoding. Well-aligned models degrade when negation is replaced with meaningless markers, while poorly calibrated models don’t show this degradation.

Conclusion: Unsupervised probing shows promise for alignment evaluation, but requires structural robustness checks in interpretability benchmarks to ensure reliable assessment of model internal representations.

Abstract: Advances in unsupervised probes such as Contrast-Consistent Search (CCS), which reveal latent beliefs without relying on token outputs, raise the question of whether these methods can reliably assess model alignment. We investigate this by examining the sensitivity of CCS to harmful vs. safe statements and by introducing Polarity-Aware CCS (PA-CCS), a method for evaluating whether a model’s internal representations remain consistent under polarity inversion. We propose two alignment-oriented metrics, Polar-Consistency and the Contradiction Index, to quantify the semantic robustness of a model’s latent knowledge. To validate PA-CCS, we curate two main datasets and one control dataset containing matched harmful-safe sentence pairs constructed using different methodologies (concurrent and antagonistic statements). We apply PA-CCS to 16 language models. Our results show that PA-CCS identifies both architectural and layer-specific differences in the encoding of latent harmful knowledge. Notably, replacing the negation token with a meaningless marker degrades PA-CCS scores for models with well-aligned internal representations, while models lacking robust internal calibration do not exhibit this degradation. Our findings highlight the potential of unsupervised probing for alignment evaluation and emphasize the need to incorporate structural robustness checks into interpretability benchmarks. Code and datasets are available at: https://github.com/SadSabrina/polarity-probing. WARNING: This paper contains potentially sensitive, harmful, and offensive content.

Method: Developed an end-to-end framework with: 1) cross-task, cross-species pretrained neural encoder for transfer learning, 2) integration with audio large language models, 3) contrastive learning for cross-modal alignment, and 4) differentiable neural network architecture.

Result: Achieved SOTA on Brain-to-Text ‘24/‘25 benchmarks; reduced WER from 24.69% to 10.22%; demonstrated small-scale audio LLMs improve end-to-end decoding; enabled cross-task generalization between attempted and imagined speech.

Conclusion: BIT advances neural data integration and enables seamless differentiable optimization, paving the way for more effective end-to-end speech BCI frameworks that support cross-task generalization.

Abstract: Speech brain-computer interfaces (BCIs) aim to restore communication for people with paralysis by translating neural activity into text. Most systems use cascaded frameworks that decode phonemes before assembling sentences with an n-gram language model (LM), preventing joint optimization of all stages simultaneously. Here, we introduce an end-to-end Brain-to-Text (BIT) framework that translates neural activity into coherent sentences using a single differentiable neural network. Central to our approach is a cross-task, cross-species pretrained neural encoder, whose representations transfer to both attempted and imagined speech. In a cascaded setting with an n-gram LM, the pretrained encoder establishes a new state-of-the-art (SOTA) on the Brain-to-Text ‘24 and ‘25 benchmarks. Integrated end-to-end with audio large language models (LLMs) and trained with contrastive learning for cross-modal alignment, BIT reduces the word error rate (WER) of the prior end-to-end method from 24.69% to 10.22%. Notably, we find that small-scale audio LLMs markedly improve end-to-end decoding. Beyond record-setting performance, BIT aligns attempted and imagined speech embeddings to enable cross-task generalization. Altogether, our approach advances the integration of large, diverse neural datasets, paving the way for an end-to-end decoding framework that supports seamless, differentiable optimization.

Method: Proposes a geometric approach integrating multimodal text and co-author network data using Hellinger distances and Ward’s linkage to construct hierarchical topic dendrograms that capture both local and global structure for multiscale learning.

Result: The method effectively identifies rare-topic structure and visualizes smooth topic drift over time, demonstrating the strength of interpretable bag-of-words models when paired with principled geometric alignment.

Conclusion: Geometric integration of text and network data with hierarchical clustering provides superior interpretable topic modeling for scientific corpora, especially for capturing underrepresented topics and temporal evolution.

Abstract: Interpretable topic modeling is essential for tracking how research interests evolve within co-author communities. In scientific corpora, where novelty is prized, identifying underrepresented niche topics is particularly important. However, contemporary models built from dense transformer embeddings tend to miss rare topics and therefore also fail to capture smooth temporal alignment. We propose a geometric method that integrates multimodal text and co-author network data, using Hellinger distances and Ward’s linkage to construct a hierarchical topic dendrogram. This approach captures both local and global structure, supporting multiscale learning across semantic and temporal dimensions. Our method effectively identifies rare-topic structure and visualizes smooth topic drift over time. Experiments highlight the strength of interpretable bag-of-words models when paired with principled geometric alignment.

Method: Introduces EduMod-LLM, a modular function-calling LLM pipeline that isolates components for fine-grained analysis. Benchmarks function-calling performance across LLMs, compares novel structure-aware retrieval to vector-based and LLM-scoring baselines, and evaluates various LLMs for response synthesis.

Result: The modular approach reveals specific failure modes and performance patterns, supporting development of interpretable educational QA systems. Demonstrates value of modular function calling in improving system transparency and pedagogical alignment.

Conclusion: Modular function-calling pipelines enable comprehensive evaluation of educational QA systems, providing insights for developing more transparent and pedagogically aligned AI tools in education.

Abstract: With the growing use of Large Language Model (LLM)-based Question-Answering (QA) systems in education, it is critical to evaluate their performance across individual pipeline components. In this work, we introduce {\model}, a modular function-calling LLM pipeline, and present a comprehensive evaluation along three key axes: function calling strategies, retrieval methods, and generative language models. Our framework enables fine-grained analysis by isolating and assessing each component. We benchmark function-calling performance across LLMs, compare our novel structure-aware retrieval method to vector-based and LLM-scoring baselines, and evaluate various LLMs for response synthesis. This modular approach reveals specific failure modes and performance patterns, supporting the development of interpretable and effective educational QA systems. Our findings demonstrate the value of modular function calling in improving system transparency and pedagogical alignment. Website and Supplementary Material: https://chancharikmitra.github.io/EduMod-LLM-website/

Method: Analyze reasoning tokens (intermediate steps generated while solving problems) during task-specific fine-tuning, aligning training dynamics with the Four Stages of Competence framework from cognitive science.

Result: Reasoning token length expands as performance improves, peaks at the stage of conscious competence, then declines as models internalize tasks. Models retain performance even when reasoning is removed after training, suggesting reasoning scaffolds learning but becomes unnecessary.

Conclusion: Reasoning token dynamics serve as valuable signals for diagnosing training stages, identifying convergence, guiding early stopping, and understanding/optimizing reasoning model training.

Abstract: We analyze reasoning in language models during task-specific fine-tuning and draws parallel between reasoning tokens–intermediate steps generated while solving problem and the human working memory. Drawing from cognitive science, we align training dynamics with the Four Stages of Competence: models initially produce incorrect outputs without reasoning, then begin reasoning (but still fail), eventually reason effectively, and finally solve tasks without explicit reasoning. We find that reasoning token length expands as performance improves, peaks at the stage of conscious competence, then declines as the model internalizes the task. Notably, after training, models retain performance even when reasoning is removed–suggesting it scaffolded learning but is no longer needed. This progression offers actionable insights: reasoning token dynamics can serve as a signal for diagnosing training stage, identifying convergence, and guiding early stopping. We propose metrics to track this trajectory and argue that reasoning behavior is valuable for understanding and optimizing reasoning model training.

Method: Texts are decomposed into stylometric and readability features, then classified using either a compact Convolutional Neural Network (CNN) or Random Forest (RF) model.

Result: On Kaggle AI vs. Human corpus: CNN achieves 97% accuracy (~0.95 F1), RF achieves 95% accuracy (~0.94 F1). ROC-AUC scores: 99.5% for CNN, 95% for RF. Models are extremely small (CNN: ~25MB, RF: ~10.6MB) and run efficiently on standard CPUs.

Conclusion: Simplicity with structural insights can rival complex approaches in AI-generated content detection. NEULIF demonstrates lightweight models can achieve high accuracy without extensive computational power, with potential for broader applications across languages, domains, and streaming contexts.

Abstract: A growing number of AI-generated texts raise serious concerns. Most existing approaches to AI-generated text detection rely on fine-tuning large transformer models or building ensembles, which are computationally expensive and often provide limited generalization across domains. Existing lightweight alternatives achieved significantly lower accuracy on large datasets. We introduce NEULIF, a lightweight approach that achieves best performance in the lightweight detector class, that does not require extensive computational power and provides high detection accuracy. In our approach, a text is first decomposed into stylometric and readability features which are then used for classification by a compact Convolutional Neural Network (CNN) or Random Forest (RF). Evaluated and tested on the Kaggle AI vs. Human corpus, our models achieve 97% accuracy (~ 0.95 F1) for CNN and 95% accuracy (~ 0.94 F1) for the Random Forest, demonstrating high precision and recall, with ROC-AUC scores of 99.5% and 95%, respectively. The CNN (~ 25 MB) and Random Forest (~ 10.6 MB) models are orders of magnitude smaller than transformer-based ensembles and can be run efficiently on standard CPU devices, without sacrificing accuracy.This study also highlights the potential of such models for broader applications across languages, domains, and streaming contexts, showing that simplicity, when guided by structural insights, can rival complexity in AI-generated content detection.

Method: Pairs Residual Vector Quantization (RVQ) EEG tokenizer with masked language diffusion model (LLaDA). RVQ discretizes continuous EEG into multi-layer tokens to reduce noise and individual differences, while LLaDA reconstructs sentences via non-sequential denoising.

Result: On ZuCo dataset, DELTA improves semantic alignment by up to 5.37 points over autoregressive baselines, achieving BLEU-1 21.9 and ROUGE-1 F 17.2 under word-level conditions.

Conclusion: DELTA enables reliable text generation from small EEG-text datasets and points toward scalable multimodal EEG-language models.

Abstract: Electroencephalogram (EEG)-to-text remains challenging due to high-dimensional noise, subject variability, and error accumulation in autoregressive decoding. We introduce DELTA, which pairs a Residual Vector Quantization (RVQ) EEG tokenizer with a masked language diffusion model (LLaDA). RVQ discretizes continuous EEG into multi-layer tokens to reduce noise and individual differences, while LLaDA reconstructs sentences via non-sequential denoising. On ZuCo, DELTA improves semantic alignment by up to 5.37 points over autoregressive baselines, achieving BLEU-1 21.9 and ROUGE-1 F 17.2 under word-level conditions. These results enable reliable text generation from small EEG-text datasets and point toward scalable multimodal EEG-language models.

Method: Combines guided synthetic data generation from seed corpus with bottom-up domain data curation, integrating Domain-Adaptive Pretraining (DAPT), Domain-specific Supervised Fine-tuning (DSFT), and Direct Preference Optimization (DPO).

Result: DiagnosticSLM (3B parameters) achieves up to 25% accuracy improvement over comparable/larger open-source models (2B-9B) on MCQ tasks, and matches/outperforms them on QA, completion, and summarization benchmarks.

Conclusion: The pipeline enables effective small-scale domain-specific LLMs with strong reasoning capabilities, addressing privacy, resource, and data constraints in specialized domains.

Abstract: Large Language Models (LLMs) have shown remarkable success in supporting a wide range of knowledge-intensive tasks. In specialized domains, there is growing interest in leveraging LLMs to assist subject matter experts with domain-specific challenges. However, deploying LLMs as SaaS solutions raises data privacy concerns, while many open-source models demand significant computational resources for effective domain adaptation and deployment. A promising alternative is to develop smaller, domain-specialized LLMs, though this approach is often constrained by the lack of high-quality domain-specific training data. In this work, we address these limitations by presenting a cost-efficient and scalable training pipeline that combines guided synthetic data generation from a small seed corpus with bottom-up domain data curation. Our pipeline integrates Domain-Adaptive Pretraining (DAPT), Domain-specific Supervised Fine-tuning (DSFT), and Direct Preference Optimization (DPO) to train effective small-scale models for specialized use cases. We demonstrate this approach through DiagnosticSLM, a 3B-parameter domain-specific model tailored for fault diagnosis, root cause analysis, and repair recommendation in industrial settings. To evaluate model performance, we introduce four domain-specific benchmarks: multiple-choice questions (DiagnosticMCQ), question answering (DiagnosticQA), sentence completion (DiagnosticComp), and summarization (DiagnosticSum). DiagnosticSLM achieves up to 25% accuracy improvement over open-source models of comparable or larger size (2B-9B) on the MCQ task, while also outperforming or matching them in other tasks, demonstrating effective domain-specific reasoning and generalization capabilities.

Method: BRIES architecture with four specialized agents: Twister (generates adversarial content), Detector (identifies attack types), Defender (creates resilient content via inoculation), and Assessor (evaluates effectiveness using causal inference). Experiments use SemEval 2023 Task 3 taxonomy on synthetic persuasion datasets.

Result: Significant performance disparities across language agents: GPT-4 achieves superior detection accuracy on complex persuasion techniques, while open-source models like Llama3 and Mistral show weaknesses in identifying subtle rhetorical patterns. Prompt engineering dramatically affects detection efficacy with model-specific temperature optimizations. Causal analysis reveals different persuasion attacks target specific cognitive dimensions.

Conclusion: The research provides novel insights into socio-emotional-cognitive signatures of persuasion attacks and delivers a framework for enhancing human cognitive resilience through structured interventions before exposure to harmful content, advancing both AI safety and cognitive security.

Abstract: This paper introduces BRIES, a novel compound AI architecture designed to detect and measure the effectiveness of persuasion attacks across information environments. We present a system with specialized agents: a Twister that generates adversarial content employing targeted persuasion tactics, a Detector that identifies attack types with configurable parameters, a Defender that creates resilient content through content inoculation, and an Assessor that employs causal inference to evaluate inoculation effectiveness. Experimenting with the SemEval 2023 Task 3 taxonomy across the synthetic persuasion dataset, we demonstrate significant variations in detection performance across language agents. Our comparative analysis reveals significant performance disparities with GPT-4 achieving superior detection accuracy on complex persuasion techniques, while open-source models like Llama3 and Mistral demonstrated notable weaknesses in identifying subtle rhetorical, suggesting that different architectures encode and process persuasive language patterns in fundamentally different ways. We show that prompt engineering dramatically affects detection efficacy, with temperature settings and confidence scoring producing model-specific variations; Gemma and GPT-4 perform optimally at lower temperatures while Llama3 and Mistral show improved capabilities at higher temperatures. Our causal analysis provides novel insights into socio-emotional-cognitive signatures of persuasion attacks, revealing that different attack types target specific cognitive dimensions. This research advances generative AI safety and cognitive security by quantifying LLM-specific vulnerabilities to persuasion attacks and delivers a framework for enhancing human cognitive resilience through structured interventions before exposure to harmful content.

Method: LDD conceals true labels by replacing them with semantically transformed or unrelated alias labels (e.g., blue vs. yellow). Models learn these new mappings implicitly through few-shot demonstrations, breaking direct correspondence between injected directives and outputs.

Result: LDD successfully restores accuracy degradation caused by attacks across nine SOTA models (GPT-5, GPT-4o, LLaMA3.2, etc.). For most models, multiple alias pairs achieve higher accuracy than the under-attack baseline. Semantically aligned aliases (good vs. bad) outperform unaligned symbols (blue vs. yellow).

Conclusion: Label semantics can effectively defend against prompt injection attacks, transforming meaning itself into a protective shield. LDD provides a lightweight, model-agnostic defense strategy that doesn’t require retraining.

Abstract: Large language models are increasingly used for text classification tasks such as sentiment analysis, yet their reliance on natural language prompts exposes them to prompt injection attacks. In particular, class-directive injections exploit knowledge of the model’s label set (e.g., positive vs. negative) to override its intended behavior through adversarial instructions. Existing defenses, such as detection-based filters, instruction hierarchies, and signed prompts, either require model retraining or remain vulnerable to obfuscation. This paper introduces Label Disguise Defense (LDD), a lightweight and model-agnostic strategy that conceals true labels by replacing them with semantically transformed or unrelated alias labels(e.g., blue vs. yellow). The model learns these new label mappings implicitly through few-shot demonstrations, preventing direct correspondence between injected directives and decision outputs. We evaluate LDD across nine state-of-the-art models, including GPT-5, GPT-4o, LLaMA3.2, Gemma3, and Mistral variants, under varying few-shot and an adversarial setting. Our results show that the ability of LDD to recover performance lost to the adversarial attack varies across models and alias choices. For every model evaluated, LDD is able to restore a portion of the accuracy degradation caused by the attack. Moreover, for the vast majority of models, we can identify more than one alias pair that achieves higher accuracy than the under-attack baseline, in which the model relies solely on few-shot learning without any defensive mechanism. A linguistic analysis further reveals that semantically aligned alias labels(e.g., good vs. bad) yield stronger robustness than unaligned symbols(e.g., blue vs. yellow). Overall, this study demonstrates that label semantics can serve as an effective defense layer, transforming meaning itself into a shield against prompt injection.

Method: Fine-tune Large Language Models to identify all locations, impacts, and specifically impacted locations (distinguishing from non-impacted ones) in disaster-related social media posts, handling informal expressions and abbreviations.

Result: Fine-tuned model achieves F1-scores of 0.69 for impact extraction and 0.74 for impacted location extraction, substantially outperforming pre-trained baseline models.

Conclusion: Fine-tuned language models offer scalable solutions for timely disaster response by extracting critical impact information from social media, improving situational awareness and resource allocation.

Abstract: Large-scale disasters can often result in catastrophic consequences on people and infrastructure. Situation awareness about such disaster impacts generated by authoritative data from in-situ sensors, remote sensing imagery, and/or geographic data is often limited due to atmospheric opacity, satellite revisits, and time limitations. This often results in geo-temporal information gaps. In contrast, impact-related social media posts can act as “geo-sensors” during a disaster, where people describe specific impacts and locations. However, not all locations mentioned in disaster-related social media posts relate to an impact. Only the impacted locations are critical for directing resources effectively. e.g., “The death toll from a fire which ripped through the Greek coastal town of #Mati stood at 80, with dozens of people unaccounted for as forensic experts tried to identify victims who were burned alive #Greecefires #AthensFires #Athens #Greece.” contains impacted location “Mati” and non-impacted locations “Greece” and “Athens”. This research uses Large Language Models (LLMs) to identify all locations, impacts and impacted locations mentioned in disaster-related social media posts. In the process, LLMs are fine-tuned to identify only impacts and impacted locations (as distinct from other, non-impacted locations), including locations mentioned in informal expressions, abbreviations, and short forms. Our fine-tuned model demonstrates efficacy, achieving an F1-score of 0.69 for impact and 0.74 for impacted location extraction, substantially outperforming the pre-trained baseline. These robust results confirm the potential of fine-tuned language models to offer a scalable solution for timely decision-making in resource allocation, situational awareness, and post-disaster recovery planning for responders.

Method: Applied Causal Tracing to GPT-2 XL architecture on ConvFinQA benchmark to identify internal mechanisms for arithmetic reasoning, followed by ablation studies and training linear probes on identified critical layers.

Result: Identified dual-stage mechanism: distributed computational scratchpad in middle layers (L12-L30) and decisive aggregation circuit in late layers (specifically Layer 46). Suppressing Layer 46 reduced hallucination confidence by 81.8%, and linear probes on this layer achieved 98% accuracy in detecting hallucinations across unseen financial topics.

Conclusion: The research provides a mechanistic understanding of arithmetic hallucinations in LLMs, revealing a universal geometry of arithmetic deception that enables highly accurate detection and potential mitigation of these critical failures in financial applications.

Abstract: Large Language Models (LLMs) are increasingly deployed in high-stakes financial domains, yet they suffer from specific, reproducible hallucinations when performing arithmetic operations. Current mitigation strategies often treat the model as a black box. In this work, we propose a mechanistic approach to intrinsic hallucination detection. By applying Causal Tracing to the GPT-2 XL architecture on the ConvFinQA benchmark, we identify a dual-stage mechanism for arithmetic reasoning: a distributed computational scratchpad in middle layers (L12-L30) and a decisive aggregation circuit in late layers (specifically Layer 46). We verify this mechanism via an ablation study, demonstrating that suppressing Layer 46 reduces the model’s confidence in hallucinatory outputs by 81.8%. Furthermore, we demonstrate that a linear probe trained on this layer generalizes to unseen financial topics with 98% accuracy, suggesting a universal geometry of arithmetic deception.

Method: A new temporal consistency method where verifiers iteratively refine their judgments based on previous assessments, using consistency in sequences of self-reflection actions rather than single-round verification or multi-model debates.

Result: Empirical evaluations across Mathcheck, ProcessBench, and PRM800K benchmarks show consistent improvements over baselines. When applied to DeepSeek R1 distilled models, 7B/8B models outperform all 70B/72B models and GPT-4o on ProcessBench. The distilled 14B model achieves performance comparable to Deepseek-R1.

Conclusion: Temporal consistency verification enables smaller distilled models to achieve or surpass performance of much larger models, demonstrating the effectiveness of iterative self-reflection for mathematical reasoning verification.

Abstract: Verification is crucial for effective mathematical reasoning. We present a new temporal consistency method where verifiers iteratively refine their judgments based on the previous assessment. Unlike one-round verification or multi-model debate approaches, our method leverages consistency in a sequence of self-reflection actions to improve verification accuracy. Empirical evaluations across diverse mathematical process error identification benchmarks (Mathcheck, ProcessBench, and PRM800K) show consistent performance improvements over baseline methods. When applied to the recent DeepSeek R1 distilled models, our method demonstrates strong performance, enabling 7B/8B distilled models to outperform all 70B/72B models and GPT-4o on ProcessBench. Notably, the distilled 14B model with our method achieves performance comparable to Deepseek-R1. Our codes are available at https://github.com/jcguo123/Temporal-Consistency

Method: Proposes ODB-dLLM with dual-boundary orchestration: 1) Adaptive length prediction mechanism that progressively reduces prefill overhead by eliminating redundant computation from fixed response lengths, 2) dLLM-specific jump-share speculative decoding method that reduces decoding iterations by leveraging computational characteristics of dLLMs.

Result: Achieves 46-162x speedup over baseline dLLM and 2.63-6.30x speedup over Fast-dLLM while mitigating accuracy degradation seen in existing acceleration frameworks.

Conclusion: ODB-dLLM effectively addresses the efficiency bottlenecks in dLLM inference through phase-specific optimizations, demonstrating significant speed improvements without sacrificing accuracy.

Abstract: Diffusion-based large language models (dLLMs) have recently gained significant attention for their exceptional performance and inherent potential for parallel decoding. Existing frameworks further enhance its inference efficiency by enabling KV caching. However, its bidirectional attention mechanism necessitates periodic cache refreshes that interleave prefill and decoding phases, both contributing substantial inference cost and constraining achievable speedup. Inspired by the heterogeneous arithmetic intensity of the prefill and decoding phases, we propose ODB-dLLM, a framework that orchestrates dual-boundaries to accelerate dLLM inference. In the prefill phase, we find that the predefined fixed response length introduces heavy yet redundant computational overhead, which affects efficiency. To alleviate this, ODB-dLLM incorporates an adaptive length prediction mechanism that progressively reduces prefill overhead and unnecessary computation. In the decoding phase, we analyze the computational characteristics of dLLMs and propose a dLLM-specific jump-share speculative decoding method to enhance efficiency by reducing the number of decoding iterations. Experimental results demonstrate that ODB-dLLM achieves 46-162x and 2.63-6.30x speedups over the baseline dLLM and Fast-dLLM, respectively, while simultaneously mitigating the accuracy degradation in existing acceleration frameworks.

Method: Theoretical analysis establishes an upper bound on DPO gradient norm controlled by PVar. Experimental validation fine-tunes LLMs with reward model-generated preferences, evaluating on AlpacaEval 2.0 and Arena-Hard benchmarks. Uses PVar-based prompt selection with smaller reward models (1B, 3B).

Result: Prompts with higher PVar outperform randomly selected prompts or those with lower PVar. Training on only top 10% highest PVar prompts yields better performance than full dataset training. PVar-based selection works robustly with smaller reward models.

Conclusion: Preference variance is crucial for identifying informative examples in DPO training. High PVar prompts produce larger gradient updates and are more valuable for efficient LLM alignment, enabling data reduction without performance loss.

Abstract: Direct Preference Optimization (DPO) has emerged as an important approach for learning from human preferences in aligning large language models (LLMs). However, collecting human preference data is costly and inefficient, motivating methods to reduce the required annotations. In this work, we investigate the impact of \emph{preference variance} (PVar), which measures the variance in model preferences when comparing pairs of responses, on the effectiveness of DPO training. We provide a theoretical insight by establishing an upper bound on the DPO gradient norm for any given prompt, showing it is controlled by the PVar of that prompt. This implies that prompts with low PVar can only produce small gradient updates, making them less valuable for learning. We validate this finding by fine-tuning LLMs with preferences generated by a reward model, evaluating on two benchmarks (AlpacaEval 2.0 and Arena-Hard). Experimental results demonstrate that prompts with higher PVar outperform randomly selected prompts or those with lower PVar. We also show that our PVar-based selection method is robust, when using smaller reward models (1B, 3B) for selection. Notably, in a separate experiment using the original human annotations from the UltraFeedback dataset, we found that training on only the top 10% of prompts with the highest PVar yields better evaluation performance than training on the full dataset, highlighting the importance of preference variance in identifying informative examples for efficient LLM alignment.

Method: Three-stage framework: 1) Neural tokenizer maps fMRI into discrete tokens in language-consistent space; 2) Pretrained LLM adapted to jointly model fMRI tokens and text; 3) Multi-task, multi-paradigm instruction tuning for high-level semantic understanding.

Result: fMRI-LM achieves strong zero-shot and few-shot performance across various benchmarks, adapts efficiently with parameter-efficient tuning (LoRA), and establishes a scalable pathway toward language-aligned universal fMRI understanding.

Conclusion: The work presents a foundational model that successfully bridges fMRI and language, enabling semantic understanding of brain activity and establishing a scalable approach for cross-modal brain representation learning.

Abstract: Recent advances in multimodal large language models (LLMs) have enabled unified reasoning across images, audio, and video, but extending such capability to brain imaging remains largely unexplored. Bridging this gap is essential to link neural activity with semantic cognition and to develop cross-modal brain representations. To this end, we present fMRI-LM, a foundational model that bridges functional MRI (fMRI) and language through a three-stage framework. In Stage 1, we learn a neural tokenizer that maps fMRI into discrete tokens embedded in a language-consistent space. In Stage 2, a pretrained LLM is adapted to jointly model fMRI tokens and text, treating brain activity as a sequence that can be temporally predicted and linguistically described. To overcome the lack of natural fMRI-text pairs, we construct a large descriptive corpus that translates diverse imaging-based features into structured textual descriptors, capturing the low-level organization of fMRI signals. In Stage 3, we perform multi-task, multi-paradigm instruction tuning to endow fMRI-LM with high-level semantic understanding, supporting diverse downstream applications. Across various benchmarks, fMRI-LM achieves strong zero-shot and few-shot performance, and adapts efficiently with parameter-efficient tuning (LoRA), establishing a scalable pathway toward a language-aligned, universal model for structural and semantic understanding of fMRI.

Method: Evaluated 5 LLMs (GPT-4.1, LLaMA 4, Grok 3, DeepSeek V3.2) on Bangla-Sylheti translation, then introduced Sylheti-CAP - a three-step framework with linguistic rulebook, dictionary (2,260 vocabulary items/idioms), and authenticity check in prompts.

Result: Sylheti-CAP consistently improved translation quality across models and prompting strategies, reducing hallucinations, ambiguities, and awkward phrasing according to both automatic metrics and human evaluations.

Conclusion: Sylheti-CAP establishes a scalable solution for dialectal and low-resource machine translation, demonstrating that context-aware prompting effectively addresses LLM limitations in handling dialect-specific vocabulary.

Abstract: Large Language Models (LLMs) have demonstrated strong translation abilities through prompting, even without task-specific training. However, their effectiveness in dialectal and low-resource contexts remains underexplored. This study presents the first systematic investigation of LLM-based machine translation (MT) for Sylheti, a dialect of Bangla that is itself low-resource. We evaluate five advanced LLMs (GPT-4.1, GPT-4.1, LLaMA 4, Grok 3, and DeepSeek V3.2) across both translation directions (Bangla $\Leftrightarrow$ Sylheti), and find that these models struggle with dialect-specific vocabulary. To address this, we introduce Sylheti-CAP (Context-Aware Prompting), a three-step framework that embeds a linguistic rulebook, a dictionary (2{,}260 core vocabulary items and idioms), and an authenticity check directly into prompts. Extensive experiments show that Sylheti-CAP consistently improves translation quality across models and prompting strategies. Both automatic metrics and human evaluations confirm its effectiveness, while qualitative analysis reveals notable reductions in hallucinations, ambiguities, and awkward phrasing, establishing Sylheti-CAP as a scalable solution for dialectal and low-resource MT. Dataset link: \href{https://github.com/TabiaTanzin/LLMs-for-Low-Resource-Dialect-Translation-Using-Context-Aware-Prompting-A-Case-Study-on-Sylheti.git}{https://github.com/TabiaTanzin/LLMs-for-Low-Resource-Dialect-Translation-Using-Context-Aware-Prompting-A-Case-Study-on-Sylheti.git}

Method: Conducted a Rapid Review guided by PRISMA framework and PICO strategy to structure search, filtering, and selection processes. Identified alignment strategies categorized by LLM lifecycle phase: inference-time, post-training, and reinforcement learning-based methods.

Result: Inference-time approaches emerged as particularly efficient, aligning outputs without retraining while supporting user intent, contextual grounding, and hallucination mitigation. The reviewed techniques provide structured mechanisms for improving LLM response quality and reliability.

Conclusion: The review identifies and categorizes effective alignment strategies across different LLM lifecycle phases, with inference-time methods showing particular promise for efficient alignment without model retraining.

Abstract: Large language models (LLMs) may generate outputs that are misaligned with user intent, lack contextual grounding, or exhibit hallucinations during conversation, which compromises the reliability of LLM-based applications. This review aimed to identify and analyze techniques that align LLM responses with conversational goals, ensure grounding, and reduce hallucination and topic drift. We conducted a Rapid Review guided by the PRISMA framework and the PICO strategy to structure the search, filtering, and selection processes. The alignment strategies identified were categorized according to the LLM lifecycle phase in which they operate: inference-time, post-training, and reinforcement learning-based methods. Among these, inference-time approaches emerged as particularly efficient, aligning outputs without retraining while supporting user intent, contextual grounding, and hallucination mitigation. The reviewed techniques provided structured mechanisms for improving the quality and reliability of LLM responses across key alignment objectives.

Method: Created FLAWS benchmark with 713 paper-error pairs by systematically inserting claim-invalidating errors into peer-reviewed papers using LLMs, with automated evaluation metrics for error identification and localization.

Result: GPT 5 performed best with 39.1% identification accuracy (k=10), followed by other frontier models, showing current limitations in LLMs’ error detection capabilities.

Conclusion: LLMs show promise but limited current capability for error detection in scientific papers, highlighting the need for benchmarks like FLAWS to drive improvement in automated scientific assessment.

Abstract: The identification and localization of errors is a core task in peer review, yet the exponential growth of scientific output has made it increasingly difficult for human reviewers to reliably detect errors given the limited pool of experts. Recent advances in Large Language Models (LLMs) have sparked interest in their potential to support such evaluation tasks, from academic peer review to automated scientific assessment. However, despite the growing use of LLMs in review systems, their capabilities to pinpoint errors remain underexplored. In this work, we introduce Fault Localization Across Writing in Science (FLAWS), an automated benchmark consisting of 713 paper-error pairs designed to evaluate how effectively LLMs detect errors that undermine key claims in research papers. We construct the benchmark by systematically inserting claim-invalidating errors into peer-reviewed papers using LLMs, paired with an automated evaluation metric that measures whether models can identify and localize these errors. Developing such a benchmark presents unique challenges that we overcome: ensuring that the inserted errors are well-defined, challenging, and relevant to the content of the paper, avoiding artifacts that would make identification trivial, and designing a scalable, automated evaluation metric. On the resulting benchmark, we evaluate five frontier LLMs: Claude Sonnet 4.5, DeepSeek Reasoner v3.1, Gemini 2.5 Pro, GPT 5, and Grok 4. Among these, GPT 5 is the top-performing model, achieving 39.1% identification accuracy when k=10, where k is the number of top-ranked error text candidates generated by the LLM.

Method: CoRA uses synthetically-generated questions with altered answer choices to test LLM consistency. It computes two intermediate scores: Bare-Minimum-Consistency Accuracy (BMCA) and Consistency Index (CI), then adjusts MCQA scores based on these consistency measures.

Result: Evaluations across different benchmarks with diverse LLMs show that LLMs can have low response consistency despite high MCQA scores. CoRA successfully scales down scores of inconsistent models, providing more reliable performance assessments.

Conclusion: CoRA provides a more reliable evaluation metric for LLMs on multiple-choice benchmarks by incorporating consistency measurements, addressing the limitation of traditional MCQA scores that don’t account for response stability.

Abstract: In this work we present the Consistency-Rebalanced Accuracy (CoRA) metric, improving the reliability of Large Language Model (LLM) scores computed on multiple choice (MC) benchmarks. Our metric explores the response consistency of the LLMs, taking advantage of synthetically-generated questions with altered answer choices. With two intermediate scores, i.e. Bare-Minimum-Consistency Accuracy (BMCA) and Consistency Index (CI), CoRA is computed by adjusting the multiple-choice question answering (MCQA) scores to better reflect the level of consistency of the LLM. We present evaluations in different benchmarks using diverse LLMs, and not only demonstrate that LLMs can present low response consistency even when they present high MCQA scores, but also that CoRA can successfully scale down the scores of inconsistent models.

Method: Conducted controlled Wizard of Oz experiments to elicit interactions with targeted affective trajectories across four commercial domains (aviation, e-commerce, travel, telecom). Collected 2,148 bilingual Dutch-English dialogues from 179 participants with multi-dimensional emotion annotations.

Result: Neutral dominates participant messages; desire and gratitude are most frequent non-neutral emotions. Moderate agreement for multilabel emotions/valence, lower for arousal/dominance. Self-reports diverge from third-party labels. Objective strategies elicit neutrality/gratitude; suboptimal strategies increase negative emotions. Temporal analysis shows successful steering toward prescribed trajectories, especially for negative targets.

Conclusion: WOZ-based operator-steered valence trajectories are effective for emotion research. Forward-looking emotion inference from prior turns is challenging, highlighting the complexity of proactive emotion-aware support. The EmoWOZ-CS corpus enables better study of emotion dynamics in customer service.

Abstract: Emotion-aware customer service needs in-domain conversational data, rich annotations, and predictive capabilities, but existing resources for emotion recognition are often out-of-domain, narrowly labeled, and focused on post-hoc detection. To address this, we conducted a controlled Wizard of Oz (WOZ) experiment to elicit interactions with targeted affective trajectories. The resulting corpus, EmoWOZ-CS, contains 2,148 bilingual (Dutch-English) written dialogues from 179 participants across commercial aviation, e-commerce, online travel agencies, and telecommunication scenarios. Our contributions are threefold: (1) Evaluate WOZ-based operator-steered valence trajectories as a design for emotion research; (2) Quantify human annotation performance and variation, including divergences between self-reports and third-party judgments; (3) Benchmark detection and forward-looking emotion inference in real-time support. Findings show neutral dominates participant messages; desire and gratitude are the most frequent non-neutral emotions. Agreement is moderate for multilabel emotions and valence, lower for arousal and dominance; self-reports diverge notably from third-party labels, aligning most for neutral, gratitude, and anger. Objective strategies often elicit neutrality or gratitude, while suboptimal strategies increase anger, annoyance, disappointment, desire, and confusion. Some affective strategies (cheerfulness, gratitude) foster positive reciprocity, whereas others (apology, empathy) can also leave desire, anger, or annoyance. Temporal analysis confirms successful conversation-level steering toward prescribed trajectories, most distinctly for negative targets; positive and neutral targets yield similar final valence distributions. Benchmarks highlight the difficulty of forward-looking emotion inference from prior turns, underscoring the complexity of proactive emotion-aware support.

Method: Proposed an annotation framework that records reading behaviors via mouse tracking. Created PreferRead dataset through a case study on preference annotation tasks, capturing fine-grained behaviors like what text parts annotators focus on, re-read, or skim.

Result: Annotators re-read responses in ~50% of trials, mostly revisiting their chosen option, rarely revisiting prompts. Re-reading correlates with higher inter-annotator agreement, while longer reading paths/times correlate with lower agreement.

Conclusion: Reading processes provide valuable cognitive insights into annotator reliability and decision-making in subjective NLP tasks. The PreferRead dataset enables detailed analysis of annotation behaviors beyond simple labels.

Abstract: We propose an annotation approach that captures not only labels but also the reading process underlying annotators’ decisions, e.g., what parts of the text they focus on, re-read or skim. Using this framework, we conduct a case study on the preference annotation task, creating a dataset PreferRead that contains fine-grained annotator reading behaviors obtained from mouse tracking. PreferRead enables detailed analysis of how annotators navigate between a prompt and two candidate responses before selecting their preference. We find that annotators re-read a response in roughly half of all trials, most often revisiting the option they ultimately choose, and rarely revisit the prompt. Reading behaviors are also significantly related to annotation outcomes: re-reading is associated with higher inter-annotator agreement, whereas long reading paths and times are associated with lower agreement. These results demonstrate that reading processes provide a complementary cognitive dimension for understanding annotator reliability, decision-making and disagreement in complex, subjective NLP tasks. Our code and data are publicly available.

Method: Created new balanced Chinese lyrics dataset spanning multiple genres, developed domain-specific model via fine-tuning, compared against zero-shot DeepSeek LLM inference, and conducted experiments with two test sets (real-world vs. synthetic).

Result: Hypothesis 2 strongly confirmed: structured genres (Folklore & Tradition) yield significantly higher accuracy than abstract genres (Love & Romance). Hypothesis 1 partially supported: fine-tuning improves robustness in real-world data but shows limited gains in synthetic test set due to design limitations.

Conclusion: Established first benchmark for cross-genre Chinese lyric attribution, highlighted importance of genre-sensitive evaluation, provided public dataset and framework, and recommended improvements for future research including diverse test sets and domain-adaptive pretraining.

Abstract: We propose a novel study on authorship attribution for Chinese lyrics, a domain where clean, public datasets are sorely lacking. Our contributions are twofold: (1) we create a new, balanced dataset of Chinese lyrics spanning multiple genres, and (2) we develop and fine-tune a domain-specific model, comparing its performance against zero-shot inference using the DeepSeek LLM. We test two central hypotheses. First, we hypothesize that a fine-tuned model will outperform a zero-shot LLM baseline. Second, we hypothesize that performance is genre-dependent. Our experiments strongly confirm Hypothesis 2: structured genres (e.g. Folklore & Tradition) yield significantly higher attribution accuracy than more abstract genres (e.g. Love & Romance). Hypothesis 1 receives only partial support: fine-tuning improves robustness and generalization in Test1 (real-world data and difficult genres), but offers limited or ambiguous gains in Test2, a smaller, synthetically-augmented set. We show that the design limitations of Test2 (e.g., label imbalance, shallow lexical differences, and narrow genre sampling) can obscure the true effectiveness of fine-tuning. Our work establishes the first benchmark for cross-genre Chinese lyric attribution, highlights the importance of genre-sensitive evaluation, and provides a public dataset and analytical framework for future research. We conclude with recommendations: enlarge and diversify test sets, reduce reliance on token-level data augmentation, balance author representation across genres, and investigate domain-adaptive pretraining as a pathway for improved attribution performance.

Method: Used a developmental approach with Pythia LM checkpoints, identifying inflection points in disambiguation performance. Analyzed attention heads whose patterns covary with disambiguation performance across model development. Conducted stress tests with stimulus perturbations and causal ablation analyses. Reproduced analyses across random seeds for the 14M model.

Result: Found that disambiguation involves multiple mechanisms, with some (especially in 14M) being highly sensitive to position and part-of-speech of disambiguating cues. Larger models (410M) contain heads with more robust disambiguation behavior. Ablating target heads impaired disambiguation, particularly in 14M.

Conclusion: Word sense disambiguation benefits from constellations of attention mechanisms, with larger models developing more robust specialized heads. The study demonstrates the value of developmental perspectives when probing language model mechanisms.

Abstract: Despite an in-principle understanding of self-attention matrix operations in Transformer language models (LMs), it remains unclear precisely how these operations map onto interpretable computations or functions–and how or when individual attention heads develop specialized attention patterns. Here, we present a pipeline to systematically probe attention mechanisms, and we illustrate its value by leveraging lexical ambiguity–where a single word has multiple meanings–to isolate attention mechanisms that contribute to word sense disambiguation. We take a “developmental” approach: first, using publicly available Pythia LM checkpoints, we identify inflection points in disambiguation performance for each LM in the suite; in 14M and 410M, we identify heads whose attention to disambiguating words covaries with overall disambiguation performance across development. We then stress-test the robustness of these heads to stimulus perturbations: in 14M, we find limited robustness, but in 410M, we identify multiple heads with surprisingly generalizable behavior. Then, in a causal analysis, we find that ablating the target heads demonstrably impairs disambiguation performance, particularly in 14M. We additionally reproduce developmental analyses of 14M across all of its random seeds. Together, these results suggest: that disambiguation benefits from a constellation of mechanisms, some of which (especially in 14M) are highly sensitive to the position and part-of-speech of the disambiguating cue; and that larger models (410M) may contain heads with more robust disambiguation behavior. They also join a growing body of work that highlights the value of adopting a developmental perspective when probing LM mechanisms.

Method: Community-engaged data collection across Senegal, Kenya, and Nigeria gathered 1,163 stereotypes spanning gender, ethnicity, religion, age, and profession. Using few-shot prompting with human-in-the-loop validation, the dataset was augmented to over 5,000 stereotype-antistereotype pairs. Validation involved semantic clustering and manual annotation by culturally informed reviewers.

Result: Evaluation of language models revealed that nine of eleven models exhibit statistically significant bias, with Bias Preference Ratios (BPR) ranging from 0.63 to 0.78 (p <= 0.05), indicating systematic preferences for stereotypes over antistereotypes, particularly across age, profession, and gender dimensions. Domain-specific models showed weaker bias, suggesting task-specific training may mitigate some associations.

Conclusion: AfriStereo opens pathways for future research on culturally grounded bias evaluation and mitigation, offering key methodologies for building more equitable, context-aware, and globally inclusive NLP technologies by addressing the underrepresentation of African perspectives in AI bias evaluation.

Abstract: Existing AI bias evaluation benchmarks largely reflect Western perspectives, leaving African contexts underrepresented and enabling harmful stereotypes in applications across various domains. To address this gap, we introduce AfriStereo, the first open-source African stereotype dataset and evaluation framework grounded in local socio-cultural contexts. Through community engaged efforts across Senegal, Kenya, and Nigeria, we collected 1,163 stereotypes spanning gender, ethnicity, religion, age, and profession. Using few-shot prompting with human-in-the-loop validation, we augmented the dataset to over 5,000 stereotype-antistereotype pairs. Entries were validated through semantic clustering and manual annotation by culturally informed reviewers. Preliminary evaluation of language models reveals that nine of eleven models exhibit statistically significant bias, with Bias Preference Ratios (BPR) ranging from 0.63 to 0.78 (p <= 0.05), indicating systematic preferences for stereotypes over antistereotypes, particularly across age, profession, and gender dimensions. Domain-specific models appeared to show weaker bias in our setup, suggesting task-specific training may mitigate some associations. Looking ahead, AfriStereo opens pathways for future research on culturally grounded bias evaluation and mitigation, offering key methodologies for the AI community on building more equitable, context-aware, and globally inclusive NLP technologies.

Method: ResearchArcade uses a graph-based interface with coherent multi-table format and graph structures to organize data from different academic sources. It captures multi-modal information (text, figures, tables), preserves temporal evolution at manuscript and community levels, and unifies diverse academic task definitions while supporting various models with different input requirements.

Result: Experiments across six academic tasks show that combining cross-source and multi-modal information enables a broader range of tasks, and incorporating graph structures consistently improves performance over baseline methods.

Conclusion: ResearchArcade effectively demonstrates the potential of unified graph-based interfaces to advance research progress by connecting diverse academic data sources and supporting various ML models for academic tasks.

Abstract: Academic research generates diverse data sources, and as researchers increasingly use machine learning to assist research tasks, a crucial question arises: Can we build a unified data interface to support the development of machine learning models for various academic tasks? Models trained on such a unified interface can better support human researchers throughout the research process, eventually accelerating knowledge discovery. In this work, we introduce ResearchArcade, a graph-based interface that connects multiple academic data sources, unifies task definitions, and supports a wide range of base models to address key academic challenges. ResearchArcade utilizes a coherent multi-table format with graph structures to organize data from different sources, including academic corpora from ArXiv and peer reviews from OpenReview, while capturing information with multiple modalities, such as text, figures, and tables. ResearchArcade also preserves temporal evolution at both the manuscript and community levels, supporting the study of paper revisions as well as broader research trends over time. Additionally, ResearchArcade unifies diverse academic task definitions and supports various models with distinct input requirements. Our experiments across six academic tasks demonstrate that combining cross-source and multi-modal information enables a broader range of tasks, while incorporating graph structures consistently improves performance over baseline methods. This highlights the effectiveness of ResearchArcade and its potential to advance research progress.

Method: Two complementary approaches: 1) HiTGNN - hierarchical temporal graph neural network integrating intra-note temporal event structures, inter-visit dynamics, and medical knowledge; 2) ReVeAL - lightweight test-time framework that distills large language model reasoning into smaller verifier models.

Result: HiTGNN achieves highest predictive accuracy for Type 2 Diabetes screening, especially for near-term risk, while preserving privacy and limiting reliance on large proprietary models. ReVeAL enhances sensitivity to true T2D cases and retains explanatory reasoning. Ablations confirm value of temporal structure and knowledge augmentation, and fairness analysis shows HiTGNN performs more equitably across subgroups.

Conclusion: The proposed methods effectively address core NLP challenges in clinical note analysis for temporal risk prediction, achieving high accuracy for T2D screening while maintaining privacy, fairness, and interpretability through complementary hierarchical temporal modeling and knowledge distillation approaches.

Abstract: Clinical notes in Electronic Health Records (EHRs) capture rich temporal information on events, clinician reasoning, and lifestyle factors often missing from structured data. Leveraging them for predictive modeling can be impactful for timely identification of chronic diseases. However, they present core natural language processing (NLP) challenges: long text, irregular event distribution, complex temporal dependencies, privacy constraints, and resource limitations. We present two complementary methods for temporally and contextually grounded risk prediction from longitudinal notes. First, we introduce HiTGNN, a hierarchical temporal graph neural network that integrates intra-note temporal event structures, inter-visit dynamics, and medical knowledge to model patient trajectories with fine-grained temporal granularity. Second, we propose ReVeAL, a lightweight, test-time framework that distills the reasoning of large language models into smaller verifier models. Applied to opportunistic screening for Type 2 Diabetes (T2D) using temporally realistic cohorts curated from private and public hospital corpora, HiTGNN achieves the highest predictive accuracy, especially for near-term risk, while preserving privacy and limiting reliance on large proprietary models. ReVeAL enhances sensitivity to true T2D cases and retains explanatory reasoning. Our ablations confirm the value of temporal structure and knowledge augmentation, and fairness analysis shows HiTGNN performs more equitably across subgroups.

Method: A hybrid approach using large language models (LLMs) to generate ratings of psychological features from the literature, then building a random forest classification model to predict whether a message will result in belief change.

Result: Of eight features tested, epistemic emotion and willingness to share were the top-ranking predictors of belief change in the model.

Conclusion: The findings provide insights into persuasive message characteristics and demonstrate how LLMs can enhance persuasion models based on psychological theory, with applications in online influence detection, misinformation mitigation, and measuring online narrative effectiveness.

Abstract: Traditional psychological models of belief revision focus on face-to-face interactions, but with the rise of social media, more effective models are needed to capture belief revision at scale, in this rich text-based online discourse. Here, we use a hybrid approach, utilizing large language models (LLMs) to develop a model that predicts successful persuasion using features derived from psychological experiments. Our approach leverages LLM generated ratings of features previously examined in the literature to build a random forest classification model that predicts whether a message will result in belief change. Of the eight features tested, \textit{epistemic emotion} and \textit{willingness to share} were the top-ranking predictors of belief change in the model. Our findings provide insights into the characteristics of persuasive messages and demonstrate how LLMs can enhance models of successful persuasion based on psychological theory. Given these insights, this work has broader applications in fields such as online influence detection and misinformation mitigation, as well as measuring the effectiveness of online narratives.

Method: Similarity standardization approach using pseudo data construction. Computes modality-specific mean and variance of similarity scores between queries and their paired data, then standardizes all scores to common scale. Pseudo pairs constructed by retrieving highest cosine similarity text/image candidates for each query.

Result: Method evaluated across 7 VLMs on MMQA and WebQA benchmarks. Achieves average Recall@20 gains of 64% on MMQA and 28% on WebQA for cross-modality retrieval. Outperforms E5-V (image captioning approach) in bridging modality gap.

Conclusion: Proposed similarity standardization with pseudo data effectively addresses modality gap without requiring manually labeled data, significantly improving cross-modality retrieval performance in multi-modal QA tasks.

Abstract: Advances in vision-language models (VLMs) have enabled effective cross-modality retrieval. However, when both text and images exist in the database, similarity scores would differ in scale by modality. This phenomenon, known as the modality gap, hinders accurate retrieval. Most existing studies address this issue with manually labeled data, e.g., by fine-tuning VLMs on them. In this work, we propose a similarity standardization approach with pseudo data construction. We first compute the mean and variance of the similarity scores between each query and its paired data in text or image modality. Using these modality-specific statistics, we standardize all similarity scores to compare on a common scale across modalities. These statistics are calculated from pseudo pairs, which are constructed by retrieving the text and image candidates with the highest cosine similarity to each query. We evaluate our method across seven VLMs using two multi-modal QA benchmarks (MMQA and WebQA), where each question requires retrieving either text or image data. Our experimental results show that our method significantly improves retrieval performance, achieving average Recall@20 gains of 64% on MMQA and 28% on WebQA when the query and the target data belong to different modalities. Compared to E5-V, which addresses the modality gap through image captioning, we confirm that our method more effectively bridges the modality gap.

Method: C²DLM starts from DLM’s fully connected attention, obtains a concept-level causal graph from a teacher model, then explicitly guides attention to learn causal relationships between concepts. This approach focuses on causal relationships while avoiding interference from difficult subgoals involving causal inversion.

Result: C²DLM improves 12% with about 3.2 times training speedup in the COT-OrderPerturb task, and achieves an average gain of 1.31% across six downstream reasoning tasks.

Conclusion: Incorporating causal concept guidance into diffusion language models significantly improves reasoning capabilities and training efficiency by better modeling the causal structures inherent in natural language and human thought.

Abstract: Autoregressive (AR) language models and Diffusion Language Models (DLMs) constitute the two principal paradigms of large language models. However, both paradigms suffer from insufficient reasoning capabilities. Human reasoning inherently relies on causal knowledge and thought, which are reflected in natural language. But in the AR paradigm, language is modeled as next token prediction (a strictly left-to-right, token-by-token order), whereas natural language itself exhibits more flexible causal structures. In the DLM paradigm, the attention mechanism is fully connected, which entirely disregards causal order. To fill this gap, we propose a \underline{\textbf{C}}ausal \underline{\textbf{C}}oncept-Guided \underline{\textbf{D}}iffusion \underline{\textbf{L}}anguage \underline{\textbf{M}}odel (C$^2$DLM). Starting from DLM’s fully connected attention, C$^2$DLM first obtains a concept-level causal graph from the teacher model, and then explicitly guides attention to learn causal relationships between concepts. By focusing on causal relationships and avoiding interference from difficult subgoals involving causal inversion, C$^2$DLM improves 12% with about 3.2 times training speedup in the COT-OrderPerturb task, and achieves an average gain of 1.31% across six downstream reasoning tasks. More details in the repository ~\href{https://github.com/Kairong-Han/C-2-DLM}{here}.

Method: Hybrid ensemble fuses three paradigms: 1) RoBERTa-based transformer classifier for semantic features, 2) GPT-2-based probabilistic detector using perturbation-induced likelihood curvature, 3) statistical linguistic feature analyzer for stylometric patterns. Uses optimized weighted voting framework with ensemble weights learned on probability simplex to maximize F1-score.

Result: Achieves 94.2% accuracy and AUC of 0.978 on 30,000-document corpus. 35% relative reduction in false positives on academic text. Low inter-model correlation (rho ~ 0.35-0.42) enables variance reduction.

Conclusion: The hybrid ensemble provides a more reliable and ethically responsible detector for real-world deployment in education and high-stakes domains, addressing limitations of single-paradigm approaches.

Abstract: The rapid proliferation of Large Language Models (LLMs) has blurred the line between human and machine authorship, creating practical risks for academic integrity and information reliability. Existing text detectors typically rely on a single methodological paradigm and suffer from poor generalization and high false positive rates (FPR), especially on high-stakes academic text. We propose a theoretically grounded hybrid ensemble that systematically fuses three complementary detection paradigms: (i) a RoBERTa-based transformer classifier for deep semantic feature extraction, (ii) a GPT-2-based probabilistic detector using perturbation-induced likelihood curvature, and (iii) a statistical linguistic feature analyzer capturing stylometric patterns. The core novelty lies in an optimized weighted voting framework, where ensemble weights are learned on the probability simplex to maximize F1-score rather than set heuristically. We provide a bias-variance analysis and empirically demonstrate low inter-model correlation (rho ~ 0.35-0.42), a key condition for variance reduction. Evaluated on a large-scale, multigenerator corpus of 30,000 documents, our system achieves 94.2% accuracy and an AUC of 0.978, with a 35% relative reduction in false positives on academic text. This yields a more reliable and ethically responsible detector for real-world deployment in education and other high-stakes domains.

Method: Uses DYNARTmo model with first-order task-space gesture specifications similar to articulatory phonology, integrating simplified mechanism for distributing articulatory effort across multiple articulators. Simulates CV syllables varying by place of articulation (labial, apical, dorsal) and vowel context (/a/, /i/, /u/).

Result: Model reproduces empirically attested patterns: jaw-supported apical closures, lower-lip elevation in bilabial stops, tongue-jaw co-movement, and saturation effects in labial constrictions. Shows realistic spatio-temporal movement patterns across consonant-vowel combinations.

Conclusion: Even with computationally simplified assumptions, DYNARTmo can generate realistic movement patterns that capture key aspects of articulatory tradeoff and synergy, demonstrating the viability of first-order task-space approaches for modeling articulatory coordination.

Abstract: This paper investigates how the dynamic articulatory model DYNARTmo accounts for articulatory tradeoffs between primary and secondary articulators, with a focus on lips-jaw and tongue-jaw coordination. While DYNARTmo does not implement full task-dynamic second-order biomechanics, it adopts first-order task-space gesture specifications comparable to those used in articulatory phonology and integrates a simplified mechanism for distributing articulatory effort across multiple articulators. We first outline the conceptual relationship between task dynamics and DYNARTmo, emphasizing the distinction between high-level task-space trajectories and their low-level articulatory execution. We then present simulation results for a set of CV syllables that illustrate how jaw displacement varies as a function of both place of articulation (labial, apical, dorsal) and vowel context (/a/, /i/, /u/). The model reproduces empirically attested patterns of articulatory synergy, including jaw-supported apical closures, lower-lip elevation in bilabial stops, tongue-jaw co-movement, and saturation effects in labial constrictions. These results demonstrate that even with computationally simplified assumptions, DYNARTmo can generate realistic spatio-temporal movement patterns that capture key aspects of articulatory tradeoff and synergy across a range of consonant-vowel combinations.

Method: Introduced RefineBench, a benchmark of 1,000 challenging problems across 11 domains with checklist-based evaluation. Evaluated two refinement modes: guided refinement (with natural language feedback) and self-refinement (without guidance). Tested frontier LMs including Gemini 2.5 Pro, GPT-5, and DeepSeek-R1.

Result: In self-refinement, frontier LMs achieved modest baseline scores (31.3% for Gemini 2.5 Pro, 29.1% for GPT-5) with minimal improvement across iterations. In guided refinement, both proprietary and large open-weight LMs (>70B) achieved near-perfect refinement within five turns using targeted feedback.

Conclusion: Frontier LMs require breakthroughs to self-refine incorrect responses autonomously, while guided refinement shows strong potential. RefineBench provides a valuable testbed for tracking progress in LM refinement capabilities.

Abstract: Can language models (LMs) self-refine their own responses? This question is increasingly relevant as a wide range of real-world user interactions involve refinement requests. However, prior studies have largely tested LMs’ refinement abilities on verifiable tasks such as competition math or symbolic reasoning with simplified scaffolds, whereas users often pose open-ended queries and provide varying degrees of feedback on what they desire. The recent advent of reasoning models that exhibit self-reflection patterns in their chains-of-thought further motivates this question. To analyze this, we introduce RefineBench, a benchmark of 1,000 challenging problems across 11 domains paired with a checklist-based evaluation framework. We evaluate two refinement modes: (1) guided refinement, where an LM is provided natural language feedback, and (2) self-refinement, where LMs attempt to improve without guidance. In the self-refinement setting, even frontier LMs such as Gemini 2.5 Pro and GPT-5 achieve modest baseline scores of 31.3% and 29.1%, respectively, and most models fail to consistently improve across iterations (e.g., Gemini-2.5-Pro gains only +1.8%, while DeepSeek-R1 declines by -0.1%). By contrast, in guided refinement, both proprietary LMs and large open-weight LMs (>70B) can leverage targeted feedback to refine responses to near-perfect levels within five turns. These findings suggest that frontier LMs require breakthroughs to self-refine their incorrect responses, and that RefineBench provides a valuable testbed for tracking progress.

Method: Focused Chain-of-Thought (F-CoT) separates information extraction from reasoning: first organizes essential information from queries into concise structured context, then guides models to reason exclusively over this context, preventing attention to irrelevant details.

Result: On arithmetic word problems, F-CoT reduces generated tokens by 2-3x while maintaining accuracy comparable to standard zero-shot CoT.

Conclusion: Structured input serves as a simple yet effective lever for more efficient LLM reasoning, offering a training-free alternative to model-centric efficiency approaches.

Abstract: Recent large language models achieve strong reasoning performance by generating detailed chain-of-thought traces, but this often leads to excessive token use and high inference latency. Existing efficiency approaches typically focus on model-centric interventions, such as reinforcement learning or supervised fine-tuning, to reduce verbosity. In contrast, we propose a training-free, input-centric approach. Inspired by cognitive psychology, we introduce Focused Chain-of-Thought (F-CoT), which separates information extraction from the reasoning process. F-CoT first organizes the essential information from a query into a concise, structured context and then guides the model to reason exclusively over this context. By preventing attention to irrelevant details, F-CoT naturally produces shorter reasoning paths. On arithmetic word problems, F-CoT reduces generated tokens by 2-3x while maintaining accuracy comparable to standard zero-shot CoT. These results highlight structured input as a simple yet effective lever for more efficient LLM reasoning.

Method: Proposes RuCo-C framework that: 1) automatically generates query-specific evaluation rubrics linked to interpretable critiques, 2) integrates densified reward feedback through “progressive exploration” strategy during RL training to dynamically adjust rewards.

Result: Comprehensive experiments show RuCo-C outperforms existing methods in text-to-SQL evaluation, yielding significant performance gains.

Conclusion: RuCo-C addresses critical bottlenecks in text-to-SQL evaluation by providing human-free, fine-grained assessment with interpretable critiques, enhancing RL training through densified reward signals.

Abstract: Text-to-SQL, a pivotal natural language processing (NLP) task that converts textual queries into executable SQL, has seen substantial progress in recent years. However, existing evaluation and reward mechanisms used to train and assess the text-to-SQL models remain a critical bottleneck. Current approaches heavily rely on manually annotated gold SQL queries, which are costly to produce and impractical for large-scale evaluation. More importantly, most reinforcement learning (RL) methods in text-to-SQL leverage only the final binary execution outcome as the reward signal, a coarse-grained supervision that overlooks detailed structural and semantic errors from the perspective of rubrics. To address these challenges, we propose RuCo-C, a novel generative judge model for fine-grained, query-specific automatic evaluation using interpretable critiques without human intervention. Our framework first automatically generates query-specific evaluation rubrics for human-free annotation, linking them to interpretable critiques. Subsequently, it integrates densified reward feedback through a “progressive exploration” strategy during the RL training process, which dynamically adjusts the rewards to enhance the model’s performance. Comprehensive experiments demonstrate that RuCo-C outperforms existing methods in text-to-SQL evaluation, yielding significant performance gains.

Method: The paper proposes and evaluates three novel token-level probability estimation approaches that leverage token logits to bridge the gap between generative LLMs and interpretable confidence scores for multi-label classification.

Result: Experiments on a synthetically generated, rigorously annotated dataset show that leveraging token logits significantly improves the interpretability and reliability of generative classifiers, enabling more nuanced content safety moderation.

Conclusion: Token-level probability estimation methods enhance model interpretability and accuracy for generative LLMs in content safety classification, and the framework demonstrates generalizability across different instruction-tuned models.

Abstract: This paper addresses the critical challenge of deriving interpretable confidence scores from generative language models (LLMs) when applied to multi-label content safety classification. While models like LLaMA Guard are effective for identifying unsafe content and its categories, their generative architecture inherently lacks direct class-level probabilities, which hinders model confidence assessment and performance interpretation. This limitation complicates the setting of dynamic thresholds for content moderation and impedes fine-grained error analysis. This research proposes and evaluates three novel token-level probability estimation approaches to bridge this gap. The aim is to enhance model interpretability and accuracy, and evaluate the generalizability of this framework across different instruction-tuned models. Through extensive experimentation on a synthetically generated, rigorously annotated dataset, it is demonstrated that leveraging token logits significantly improves the interpretability and reliability of generative classifiers, enabling more nuanced content safety moderation.

Method: Used DistilBERT (distilbert-base-uncased), a lightweight transformer model, for sentiment classification on approximately one million English-language Shopee reviews. The approach included data preprocessing and was evaluated against benchmark models (BERT and SVM) using accuracy, precision, recall, and F1-score metrics.

Result: DistilBERT achieved 94.8% accuracy, slightly below BERT’s 95.3% but significantly higher than SVM’s 90.2%. Computation time was reduced by more than 55% compared to BERT, demonstrating superior efficiency while maintaining competitive accuracy.

Conclusion: DistilBERT provides an optimal balance between accuracy and computational efficiency, making it well-suited for large-scale sentiment analysis on e-commerce platforms where processing speed and resource efficiency are important considerations alongside classification performance.

Abstract: The rapid growth of digital commerce has led to the accumulation of a massive number of consumer reviews on online platforms. Shopee, as one of the largest e-commerce platforms in Southeast Asia, receives millions of product reviews every day containing valuable information regarding customer satisfaction and preferences. Manual analysis of these reviews is inefficient, thus requiring a computational approach such as sentiment analysis. This study examines the use of DistilBERT, a lightweight transformer-based deep learning model, for sentiment classification on Shopee product reviews. The dataset used consists of approximately one million English-language reviews that have been preprocessed and trained using the distilbert-base-uncased model. Evaluation was conducted using accuracy, precision, recall, and F1-score metrics, and compared against benchmark models such as BERT and SVM. The results show that DistilBERT achieved an accuracy of 94.8%, slightly below BERT (95.3%) but significantly higher than SVM (90.2%), with computation time reduced by more than 55%. These findings demonstrate that DistilBERT provides an optimal balance between accuracy and efficiency, making it suitable for large scale sentiment analysis on e-commerce platforms. Keywords: Sentiment Analysis, DistilBERT, Shopee Reviews, Natural Language Processing, Deep Learning, Transformer Models.

Method: Created the first Kurdish Sorani NER dataset (64,563 annotated tokens), developed a tool for NER tasks across languages, and conducted comparative analysis using classic ML models (CRF) and neural systems (BiLSTM).

Result: CRF achieved F1-score of 0.825, significantly outperforming BiLSTM-based models (0.706), challenging assumptions about neural approaches’ superiority in NLP.

Conclusion: Simpler classical frameworks like CRF can outperform neural architectures in low-resource settings, offering more computationally efficient alternatives for under-represented languages.

Abstract: This work contributes towards balancing the inclusivity and global applicability of natural language processing techniques by proposing the first ’name entity recognition’ dataset for Kurdish Sorani, a low-resource and under-represented language, that consists of 64,563 annotated tokens. It also provides a tool for facilitating this task in this and many other languages and performs a thorough comparative analysis, including classic machine learning models and neural systems. The results obtained challenge established assumptions about the advantage of neural approaches within the context of NLP. Conventional methods, in particular CRF, obtain F1-scores of 0.825, outperforming the results of BiLSTM-based models (0.706) significantly. These findings indicate that simpler and more computationally efficient classical frameworks can outperform neural architectures in low-resource settings.

Method: Created a contrastive dataset of clinical statements varying in epistemic modality (e.g., ‘is consistent with’ vs. ‘may be consistent with’). Proposed Model Sensitivity to Uncertainty (MSU), a layerwise probing metric that quantifies activation-level shifts induced by uncertainty cues.

Result: LLMs exhibit structured, depth-dependent sensitivity to clinical uncertainty, suggesting that epistemic information is progressively encoded in deeper layers. The findings reveal how linguistic uncertainty is internally represented in LLMs.

Conclusion: The research offers insight into LLMs’ interpretability and epistemic reliability by showing how they internally represent linguistic uncertainty, which is crucial for clinical applications where uncertainty interpretation matters.

Abstract: Large Language Models (LLMs) are increasingly used in clinical settings, where sensitivity to linguistic uncertainty can influence diagnostic interpretation and decision-making. Yet little is known about where such epistemic cues are internally represented within these models. Distinct from uncertainty quantification, which measures output confidence, this work examines input-side representational sensitivity to linguistic uncertainty in medical text. We curate a contrastive dataset of clinical statements varying in epistemic modality (e.g., ‘is consistent with’ vs. ‘may be consistent with’) and propose Model Sensitivity to Uncertainty (MSU), a layerwise probing metric that quantifies activation-level shifts induced by uncertainty cues. Our results show that LLMs exhibit structured, depth-dependent sensitivity to clinical uncertainty, suggesting that epistemic information is progressively encoded in deeper layers. These findings reveal how linguistic uncertainty is internally represented in LLMs, offering insight into their interpretability and epistemic reliability.

Method: Systematically tested four potential causes: data cleaning procedures, pre-trained model limitations, base model size, and training dataset size, using two related but structurally different Brazilian Indigenous languages and studying both translation directions.

Result: None or very limited influence from the tested training factors (cleaning, model limitations, model size, dataset size), suggesting language-specific differences play a significant role in translator performance.

Conclusion: Performance variations in fine-tuned translators for Indigenous languages are primarily driven by inherent linguistic differences between languages rather than technical training factors, highlighting the importance of language-specific characteristics in low-resource translation.

Abstract: Finetuning pre-trained language models with small amounts of data is a commonly-used method to create translators for ultra-low resource languages such as endangered Indigenous languages. However, previous works have reported substantially different performances with translators created using similar methodology and data. In this work we systematically explored possible causes of the performance difference, aiming to determine whether it was a product of different cleaning procedures, limitations of the pre-trained models, the size of the base model, or the size of the training dataset, studying both directions of translation. Our studies, using two Brazilian Indigenous languages, related but with significant structural linguistic characteristics, indicated none or very limited influence from those training factors, suggesting differences between languages may play a significant role in the ability to produce translators by fine-tuning pre-trained models.

Method: Using mathematical formulation of Merge within Strong Minimalist Thesis, analyzing derivations with Sideward Merge, optimality violations with Resource Restrictions cost functions, and alternative derivations without EC violations.

Result: All analyzed phenomena can be explained without EC violations: some use Sideward Merge with minimal optimality violations, others have alternative derivations without EC violations or Sideward Merge.

Conclusion: The Extension Condition is an intrinsic algebraic constraint in Merge’s mathematical formulation, not an additional assumption, and minimal optimality violations in Sideward Merge play a structural role in Merge’s Markovian properties.

Abstract: We analyze, using the mathematical formulation of Merge within the Strong Minimalist Thesis framework, a set of linguistic phenomena, including head-to-head movement, phrasal affixes and syntactic cliticization, verb-particle alternation, and operator-variable phenomena. These are often regarded as problematic, as violations of the Extension Condition. We show that, in fact, all of these phenomena can be explained without involving any EC violation. We first show that derivations using Sideward Merge are possible for all of these cases: these respect EC, though they involve some amount of optimality violations, with respect to Resource Restrictions cost functions, andthe amount of violation differs among these cases. We show that all the cases that involve large optimality violations can be derived in alternative ways involving neither EC nor the use of SM. The main remaining case (head-to-head movement) only involves SM with minimal violations of optimality (near equilibrium fluctuations). We analyze explicitly also the cases of multiple wh-fronting, clusters of clitics in Romance languages and possessor agreement construction in Korean, and how an explanation of these phenomena based on SM can be made compatible with the colored operad generators for phases and theta roles. We also show that the EC condition has a clear algebraic meaning in the mathematical formulation of Merge and is therefore an intrinsic structural algebraic constraint of the model, rather than an additional assumption. We also show that the minimal optimality violating SM plays a structural role in the Markovian properties of Merge, and we compare different optimality conditions coming from Minimal Search and from Resource Restriction in terms of their effect on the dynamics of the Hopf algebra Markov chain, in a simple explicit example.

Method: Retrieval-augmented generation system integrating domain-adapted LLMs with semantic retrieval, using embedding fine-tuning on domain-specific dataset (heuristic mining + LLM-assisted synthesis) and vector indexing for scalable test case/requirement retrieval.

Result: Fine-tuned compact models (e.g., bge-base-en-v1.5) achieve superior accuracy-latency-cost tradeoff vs larger models; RAG-enhanced assistants improve perceived helpfulness, truthfulness, and satisfaction in user studies.

Conclusion: Compact domain-adapted LLMs with RAG provide efficient, deployable solutions for industrial HIL environments, challenging the “bigger is better” assumption in LLM applications.

Abstract: Hardware-in-the-Loop (HIL) testing is essential for automotive validation but suffers from fragmented and underutilized test artifacts. This paper presents HIL-GPT, a retrieval-augmented generation (RAG) system integrating domain-adapted large language models (LLMs) with semantic retrieval. HIL-GPT leverages embedding fine-tuning using a domain-specific dataset constructed via heuristic mining and LLM-assisted synthesis, combined with vector indexing for scalable, traceable test case and requirement retrieval. Experiments show that fine-tuned compact models, such as \texttt{bge-base-en-v1.5}, achieve a superior trade-off between accuracy, latency, and cost compared to larger models, challenging the notion that bigger is always better. An A/B user study further confirms that RAG-enhanced assistants improve perceived helpfulness, truthfulness, and satisfaction over general-purpose LLMs. These findings provide insights for deploying efficient, domain-aligned LLM-based assistants in industrial HIL environments.

Method: 1) Search space reduction to select relevant ontology subsets, 2) Automatic prompt construction, 3) Novel LLM-based synthetic dataset generation for training data, 4) Fine-tuning LLMs on synthetic ontology submodule pairs with reference alignments.

Result: Fine-tuned LLM outperformed non-fine-tuned base model on Conference, Geolink, Enslaved, Taxon, and Hydrography datasets from OAEI complex track.

Conclusion: Combining automatic synthetic dataset generation with fine-tuning effectively adapts LLMs for ontology matching tasks, providing a practical solution to the scarcity of training data.

Abstract: Large Language Models (LLMs) are increasingly being integrated into various components of Ontology Matching pipelines. This paper investigates the capability of LLMs to perform ontology matching directly on ontology modules and generate the corresponding alignments. Furthermore, it is explored how a dedicated fine-tuning strategy can enhance the model’s matching performance in a zero-shot setting. The proposed method incorporates a search space reduction technique to select relevant subsets from both source and target ontologies, which are then used to automatically construct prompts. Recognizing the scarcity of reference alignments for training, a novel LLM-based approach is introduced for generating a synthetic dataset. This process creates a corpus of ontology submodule pairs and their corresponding reference alignments, specifically designed to fine-tune an LLM for the ontology matching task. The proposed approach was evaluated on the Conference, Geolink, Enslaved, Taxon, and Hydrography datasets from the OAEI complex track. The results demonstrate that the LLM fine-tuned on the synthetically generated data exhibits superior performance compared to the non-fine-tuned base model. The key contribution is a strategy that combines automatic dataset generation with fine-tuning to effectively adapt LLMs for ontology matching tasks.

Method: Created a novel transliteration dataset for Hindi and Bengali with nearly 1.8 million and 1 million transliteration pairs respectively, then pre-trained a custom multilingual sequence-to-sequence LLM based on Marian architecture using this dataset.

Result: Experimental results demonstrate significant improvements compared to existing relevant models in terms of BLEU (translation quality) and CER (character error rate) metrics.

Conclusion: The proposed dataset and multilingual LLM effectively address the research gap in Romanized script transliteration for Indo-Aryan languages, providing better performance for NLP tasks involving these widely spoken languages.

Abstract: The development of robust transliteration techniques to enhance the effectiveness of transforming Romanized scripts into native scripts is crucial for Natural Language Processing tasks, including sentiment analysis, speech recognition, information retrieval, and intelligent personal assistants. Despite significant advancements, state-of-the-art multilingual models still face challenges in handling Romanized script, where the Roman alphabet is adopted to represent the phonetic structure of diverse languages. Within the South Asian context, where the use of Romanized script for Indo-Aryan languages is widespread across social media and digital communication platforms, such usage continues to pose significant challenges for cutting-edge multilingual models. While a limited number of transliteration datasets and models are available for Indo-Aryan languages, they generally lack sufficient diversity in pronunciation and spelling variations, adequate code-mixed data for large language model (LLM) training, and low-resource adaptation. To address this research gap, we introduce a novel transliteration dataset for two popular Indo-Aryan languages, Hindi and Bengali, which are ranked as the 3rd and 7th most spoken languages worldwide. Our dataset comprises nearly 1.8 million Hindi and 1 million Bengali transliteration pairs. In addition to that, we pre-train a custom multilingual seq2seq LLM based on Marian architecture using the developed dataset. Experimental results demonstrate significant improvements compared to existing relevant models in terms of BLEU and CER metrics.

Method: Mixed-methods approach using LLMs to generate summaries of motivational interviewing dialogues. Two-stage annotation scheme based on MITI framework components (evocation, collaboration, autonomy, direction, empathy, non-judgmental attitude). Expert-annotated dialogues as ground truth, multi-class classification tasks with progressive prompting (one-shot and few-shot).

Result: Results provide insights into LLMs’ capacity for understanding complex psychological constructs and highlight best practices to mitigate “semantic drift” in therapeutic settings.

Conclusion: Contributes to MI community with high-quality annotated dataset addressing data scarcity, and provides critical insights for using LLMs for precise contextual interpretation in complex behavioral therapy.

Abstract: Recent advancements in large language models (LLMs) have shown their potential across both general and domain-specific tasks. However, there is a growing concern regarding their lack of sensitivity, factual incorrectness in responses, inconsistent expressions of empathy, bias, hallucinations, and overall inability to capture the depth and complexity of human understanding, especially in low-resource and sensitive domains such as psychology. To address these challenges, our study employs a mixed-methods approach to evaluate the efficacy of LLMs in psychotherapy. We use LLMs to generate precise summaries of motivational interviewing (MI) dialogues and design a two-stage annotation scheme based on key components of the Motivational Interviewing Treatment Integrity (MITI) framework, namely evocation, collaboration, autonomy, direction, empathy, and a non-judgmental attitude. Using expert-annotated MI dialogues as ground truth, we formulate multi-class classification tasks to assess model performance under progressive prompting techniques, incorporating one-shot and few-shot prompting. Our results offer insights into LLMs’ capacity for understanding complex psychological constructs and highlight best practices to mitigate ``semantic drift" in therapeutic settings. Our work contributes not only to the MI community by providing a high-quality annotated dataset to address data scarcity in low-resource domains but also critical insights for using LLMs for precise contextual interpretation in complex behavioral therapy.

Method: The paper discusses the design of a RAG-based LLM system with specific requirements: (1) retrieval module must retrieve appropriate external knowledge relevant to disputed issues, (2) generated responses must originate from and remain faithful to the RAG context, and (3) retrieval must reference external knowledge with appropriate timestamps.

Result: The paper presents a framework for designing RAG-based LLM systems that satisfy legal requirements for use in Japanese medical litigation procedures as substitutes for human expert commissioners.

Conclusion: A properly designed RAG-based LLM system can potentially substitute for expert commissioners in Japanese medical litigation while adhering to legal norms, provided it meets the three specified requirements for knowledge retrieval, response faithfulness, and timestamp referencing.

Abstract: This study discusses the essential components that a Retrieval-Augmented Generation (RAG)-based LLM system should possess in order to support Japanese medical litigation procedures complying with legal norms. In litigation, expert commissioners, such as physicians, architects, accountants, and engineers, provide specialized knowledge to help judges clarify points of dispute. When considering the substitution of these expert roles with a RAG-based LLM system, the constraint of strict adherence to legal norms is imposed. Specifically, three requirements arise: (1) the retrieval module must retrieve appropriate external knowledge relevant to the disputed issues in accordance with the principle prohibiting the use of private knowledge, (2) the responses generated must originate from the context provided by the RAG and remain faithful to that context, and (3) the retrieval module must reference external knowledge with appropriate timestamps corresponding to the issues at hand. This paper discusses the design of a RAG-based LLM system that satisfies these requirements.

Method: Created dataset from Japanese bar exam multiple-choice questions (2015-2024), decomposed questions into independent true/false judgments with structured contextual fields, and evaluated 26 LLMs including proprietary, open-weight, Japanese-specialized, and reasoning models.

Result: Proprietary models with reasoning enabled performed best, and Constitution questions were generally easier than Civil Code or Penal Code questions across the evaluated models.

Conclusion: JBE-QA provides the first comprehensive Japanese legal-domain benchmark for LLM evaluation, revealing performance patterns across different legal domains and model types, with reasoning-enhanced proprietary models showing superior performance.

Abstract: We introduce JBE-QA, a Japanese Bar Exam Question-Answering dataset to evaluate large language models’ legal knowledge. Derived from the multiple-choice (tanto-shiki) section of the Japanese bar exam (2015-2024), JBE-QA provides the first comprehensive benchmark for Japanese legal-domain evaluation of LLMs. It covers the Civil Code, the Penal Code, and the Constitution, extending beyond the Civil Code focus of prior Japanese resources. Each question is decomposed into independent true/false judgments with structured contextual fields. The dataset contains 3,464 items with balanced labels. We evaluate 26 LLMs, including proprietary, open-weight, Japanese-specialised, and reasoning models. Our results show that proprietary models with reasoning enabled perform best, and the Constitution questions are generally easier than the Civil Code or the Penal Code questions.

Method: Created the FEANEL Benchmark with 1,000 student essays annotated by language experts using a part-of-speech-based error taxonomy. Evaluated state-of-the-art LLMs on this benchmark to assess their error analysis and pedagogical abilities.

Result: Experimental results show significant gaps in current LLMs’ ability to perform fine-grained error analysis, indicating they need substantial improvement for effective educational applications in English writing.

Conclusion: Current LLMs lack the fine-grained error analysis capabilities needed for effective K-12 English writing education, highlighting the need for specialized advancements in educational AI methods.

Abstract: Large Language Models (LLMs) have transformed artificial intelligence, offering profound opportunities for educational applications. However, their ability to provide fine-grained educational feedback for K-12 English writing remains underexplored. In this paper, we challenge the error analysis and pedagogical skills of LLMs by introducing the problem of Fine-grained Error Analysis for English Learners and present the Fine-grained Error ANalysis for English Learners (FEANEL) Benchmark. The benchmark comprises 1,000 essays written by elementary and secondary school students, and a well-developed English writing error taxonomy. Each error is annotated by language education experts and categorized by type, severity, and explanatory feedback, using a part-of-speech-based taxonomy they co-developed. We evaluate state-of-the-art LLMs on the FEANEL Benchmark to explore their error analysis and pedagogical abilities. Experimental results reveal significant gaps in current LLMs’ ability to perform fine-grained error analysis, highlighting the need for advancements in particular methods for educational applications.

Method: Propose LED-WM (Language-aware Encoder for Dreamer World Model) built on DreamerV3. Features an observation encoder with attention mechanism to explicitly ground language descriptions to entities in observations. Uses model-based RL where world model is trained through environment interaction and policy is learned from model without planning or demonstrations.

Result: LED-WM policies generalize more effectively to unseen games described by novel dynamics and language compared to baselines in MESSENGER and MESSENGER-WM environments. Also demonstrates policy can be improved through fine-tuning on synthetic test trajectories generated by the world model.

Conclusion: LED-WM enables better policy generalization from language-conditioned world models without relying on planning or expert demonstrations, improving agents’ ability to understand dynamics-descriptive language for human-agent interaction.

Abstract: To interact effectively with humans in the real world, it is important for agents to understand language that describes the dynamics of the environment–that is, how the environment behaves–rather than just task instructions specifying “what to do”. Understanding this dynamics-descriptive language is important for human-agent interaction and agent behavior. Recent work address this problem using a model-based approach: language is incorporated into a world model, which is then used to learn a behavior policy. However, these existing methods either do not demonstrate policy generalization to unseen games or rely on limiting assumptions. For instance, assuming that the latency induced by inference-time planning is tolerable for the target task or expert demonstrations are available. Expanding on this line of research, we focus on improving policy generalization from a language-conditioned world model while dropping these assumptions. We propose a model-based reinforcement learning approach, where a language-conditioned world model is trained through interaction with the environment, and a policy is learned from this model–without planning or expert demonstrations. Our method proposes Language-aware Encoder for Dreamer World Model (LED-WM) built on top of DreamerV3. LED-WM features an observation encoder that uses an attention mechanism to explicitly ground language descriptions to entities in the observation. We show that policies trained with LED-WM generalize more effectively to unseen games described by novel dynamics and language compared to other baselines in several settings in two environments: MESSENGER and MESSENGER-WM.To highlight how the policy can leverage the trained world model before real-world deployment, we demonstrate the policy can be improved through fine-tuning on synthetic test trajectories generated by the world model.

Method: Iterative framework coordinating LLM (for prompt generation), T2IM (for image synthesis), and MLLM (for idiom recognition). The system generates detailed prompts, creates images, infers idioms from images, and refines prompts until successful recognition or step limit reached.

Result: Created a dataset of 1,000 visual pun images with paired prompts. Experiments across 10 LLMs, 10 MLLMs, and Qwen-Image T2IM show MLLM choice is primary performance driver: GPT achieves highest accuracies, Gemini follows, and best open-source MLLM (Gemma) is competitive. Claude performs best for prompt generation among LLMs.

Conclusion: The framework successfully automates visual pun generation and evaluation, establishing a benchmark for multimodal understanding. MLLM capabilities significantly impact system performance, with GPT leading and open-source models showing competitive results.

Abstract: We study idiom-based visual puns–images that align an idiom’s literal and figurative meanings–and present an iterative framework that coordinates a large language model (LLM), a text-to-image model (T2IM), and a multimodal LLM (MLLM) for automatic generation and evaluation. Given an idiom, the system iteratively (i) generates detailed visual prompts, (ii) synthesizes an image, (iii) infers the idiom from the image, and (iv) refines the prompt until recognition succeeds or a step limit is reached. Using 1,000 idioms as inputs, we synthesize a corresponding dataset of visual pun images with paired prompts, enabling benchmarking of both generation and understanding. Experiments across 10 LLMs, 10 MLLMs, and one T2IM (Qwen-Image) show that MLLM choice is the primary performance driver: GPT achieves the highest accuracies, Gemini follows, and the best open-source MLLM (Gemma) is competitive with some closed models. On the LLM side, Claude attains the strongest average performance for prompt generation.

Method: FLy introduces a two-tier mechanism: 1) entropy-level gate identifies whether tokens allow multiple plausible alternatives, 2) token-level deferred window distinguishes genuine errors from semantically correct variants. Also includes multi-level acceleration strategy for both target and draft models.

Result: FLy preserves >99% of target model accuracy while achieving average 2.81x speedup on Llama-3.1-70B-Instruct and 5.07x speedup on 405B variant. On out-of-domain datasets, outperforms training-based method EAGLE-3 by 1.62x.

Conclusion: FLy’s training-free design enables seamless composition with arbitrary draft-target pairs and generalizes across models/domains without hyperparameter tuning, offering an effective solution for LLM inference acceleration while maintaining semantic correctness.

Abstract: Large language models (LLMs) achieve strong performance across diverse tasks but suffer from high inference latency due to their autoregressive generation. Speculative Decoding (SPD) mitigates this issue by verifying candidate tokens in parallel from a smaller draft model, yet its strict exact-match verification discards many semantically valid continuations. Moreover, existing training-based SPD methods often suffer from performance degradation on out-of-distribution (OOD) tasks. To this end, we propose Training-Free Loosely Speculative Decoding (FLy), a novel method that loosens the rigid verification criterion by leveraging the target model’s self-corrective behavior to judge whether a draft-target mismatch remains semantically valid. FLy introduces a two-tier mechanism: an entropy-level gate that identifies whether the current token allows multiple plausible alternatives or is nearly deterministic, and a token-level deferred window that distinguishes genuine errors from differently worded yet semantically correct variants. To further reduce latency, we design a multi-level acceleration strategy that accelerates not only the target model but also the drafter itself. Owing to its training-free design, FLy composes seamlessly with arbitrary draft-target pairs and generalizes across models and domains without hyperparameter re-tuning. Experiments show that FLy preserves more than 99% of the target model’s accuracy while achieving an average 2.81x speedup on Llama-3.1-70B-Instruct and 5.07x speedup on the 405B variant. Notably, on out-of-domain datasets, our method remains highly effective and outperforms the training-based method EAGLE-3 by 1.62x.

Method: Use encoder-only (BERT) and decoder-only (GPT-2, Transformer-XL) pre-trained models as frozen embedders paired with lightweight classifiers. Compare pooling vs padding strategies and neural vs linear heads. Evaluate on LIAR dataset with controlled preprocessing.

Result: BERT embeddings with logistic regression outperform neural baselines. Contextual self-attention encodings transfer effectively. Models are robust to truncation, and simple max or average pooling works well for aggregation.

Conclusion: Attention-based token encoders serve as robust, architecture-centric foundations for veracity tasks, demonstrating that Transformer contributions can be effectively isolated from classifier complexity.

Abstract: This paper investigates fake news detection as a downstream evaluation of Transformer representations, benchmarking encoder-only and decoder-only pre-trained models (BERT, GPT-2, Transformer-XL) as frozen embedders paired with lightweight classifiers. Through controlled preprocessing comparing pooling versus padding and neural versus linear heads, results demonstrate that contextual self-attention encodings consistently transfer effectively. BERT embeddings combined with logistic regression outperform neural baselines on LIAR dataset splits, while analyses of sequence length and aggregation reveal robustness to truncation and advantages from simple max or average pooling. This work positions attention-based token encoders as robust, architecture-centric foundations for veracity tasks, isolating Transformer contributions from classifier complexity.

Method: Created ShoppingComp benchmark with 120 tasks and 1,026 scenarios curated by 35 experts, featuring complex tasks with real products and easy verifiability. Introduces novel evaluation dimension for product safety hazards.

Result: Current LLMs perform poorly: GPT-5 achieves 11.22%, Gemini-2.5-Flash achieves 3.92%. Models make critical errors like failing to identify unsafe product usage and falling for promotional misinformation, leading to harmful recommendations.

Conclusion: ShoppingComp fills the gap between research benchmarks and real-world deployment, establishing a new standard for advancing reliable and practical shopping agents in e-commerce by highlighting substantial limitations in current LLMs.

Abstract: We present ShoppingComp, a challenging real-world benchmark for rigorously evaluating LLM-powered shopping agents on three core capabilities: precise product retrieval, expert-level report generation, and safety critical decision making. Unlike prior e-commerce benchmarks, ShoppingComp introduces highly complex tasks under the principle of guaranteeing real products and ensuring easy verifiability, adding a novel evaluation dimension for identifying product safety hazards alongside recommendation accuracy and report quality. The benchmark comprises 120 tasks and 1,026 scenarios, curated by 35 experts to reflect authentic shopping needs. Results reveal stark limitations of current LLMs: even state-of-the-art models achieve low performance (e.g., 11.22% for GPT-5, 3.92% for Gemini-2.5-Flash). These findings highlight a substantial gap between research benchmarks and real-world deployment, where LLMs make critical errors such as failure to identify unsafe product usage or falling for promotional misinformation, leading to harmful recommendations. ShoppingComp fills the gap and thus establishes a new standard for advancing reliable and practical agents in e-commerce.

Method: Using sociolinguistic methodology, researchers compared LLM and human ratings on three key social attributes of spelling variation: formality, carefulness, and age. They analyzed both overall correlations and distributional differences between human and LLM perceptions.

Result: Found generally strong correlations between human and LLM ratings across the three social attributes. However, notable differences emerged when analyzing rating distributions and comparing different types of spelling variation.

Conclusion: While LLMs show strong alignment with human perceptions of spelling variation’s social attributes, there are systematic differences that warrant further investigation, particularly regarding how different types of variation are interpreted and the distribution of ratings.

Abstract: Spelling variation (e.g. funnnn vs. fun) can influence the social perception of texts and their writers: we often have various associations with different forms of writing (is the text informal? does the writer seem young?). In this study, we focus on the social perception of spelling variation in online writing in English and study to what extent this perception is aligned between humans and large language models (LLMs). Building on sociolinguistic methodology, we compare LLM and human ratings on three key social attributes of spelling variation (formality, carefulness, age). We find generally strong correlations in the ratings between humans and LLMs. However, notable differences emerge when we analyze the distribution of ratings and when comparing between different types of spelling variation.

Method: Uses interpretable tree-based ML models with five feature categories: compression-based, lexical structure, readability, neologism detection, and distance features to predict temporal origin of English texts spanning five centuries.

Result: Achieved 76.7% accuracy for century-scale prediction and 26.1% for decade-scale classification, with strong ranking capabilities (AUCROC up to 94.8%) and controlled errors (mean absolute deviations of 27-30 years).

Conclusion: Feature-engineered tree models provide scalable, interpretable alternatives to neural architectures for historical text dating, with distance features and lexical structure being most informative, revealing systematic linguistic evolution patterns.

Abstract: Accurately dating historical texts is essential for organizing and interpreting cultural heritage collections. This article addresses temporal text classification using interpretable, feature-engineered tree-based machine learning models. We integrate five feature categories - compression-based, lexical structure, readability, neologism detection, and distance features - to predict the temporal origin of English texts spanning five centuries. Comparative analysis shows that these feature domains provide complementary temporal signals, with combined models outperforming any individual feature set. On a large-scale corpus, we achieve 76.7% accuracy for century-scale prediction and 26.1% for decade-scale classification, substantially above random baselines (20% and 2.3%). Under relaxed temporal precision, performance increases to 96.0% top-2 accuracy for centuries and 85.8% top-10 accuracy for decades. The final model exhibits strong ranking capabilities with AUCROC up to 94.8% and AUPRC up to 83.3%, and maintains controlled errors with mean absolute deviations of 27 years and 30 years, respectively. For authentication-style tasks, binary models around key thresholds (e.g., 1850-1900) reach 85-98% accuracy. Feature importance analysis identifies distance features and lexical structure as most informative, with compression-based features providing complementary signals. SHAP explainability reveals systematic linguistic evolution patterns, with the 19th century emerging as a pivot point across feature domains. Cross-dataset evaluation on Project Gutenberg highlights domain adaptation challenges, with accuracy dropping by 26.4 percentage points, yet the computational efficiency and interpretability of tree-based models still offer a scalable, explainable alternative to neural architectures.

Method: Developed various classifiers for UK ONS SOC and US O*NET SOC using different language models. Created an ensemble model combining Google BERT with a neural network classifier that considers job title, description, and skills features.

Result: The ensemble model achieved 61% classification accuracy at the fourth (most detailed) tier of SOC and 72% accuracy at the third tier, outperforming individual models.

Conclusion: The developed model provides an effective automated solution for analyzing labor market evolution through job advertisements, offering more timely and accurate occupational classification than manual methods.

Abstract: Standard Occupational Classifiers (SOC) are systems used to categorize and classify different types of jobs and occupations based on their similarities in terms of job duties, skills, and qualifications. Integrating these facets with Big Data from job advertisement offers the prospect to investigate labour demand that is specific to various occupations. This project investigates the use of recent developments in natural language processing to construct a classifier capable of assigning an occupation code to a given job advertisement. We develop various classifiers for both UK ONS SOC and US O*NET SOC, using different Language Models. We find that an ensemble model, which combines Google BERT and a Neural Network classifier while considering job title, description, and skills, achieved the highest prediction accuracy. Specifically, the ensemble model exhibited a classification accuracy of up to 61% for the lower (or fourth) tier of SOC, and 72% for the third tier of SOC. This model could provide up to date, accurate information on the evolution of the labour market using job advertisements.

Method: Used a human-in-the-loop framework with DeepSeek V3.1 guided by prompt-based cognitive scaffolding to identify and convey metaphor and metonymy in four fundamental passages from Huangdi Neijing. Evaluated translations using ChatGPT 5 Pro and Gemini 2.5 Pro simulating three types of real-world readers, scoring across five cognitive dimensions, followed by structured interviews and Interpretative Phenomenological Analysis.

Result: Prompt-adjusted LLM translations performed best across all five cognitive dimensions with high cross-model and cross-role consistency. Interviews revealed differences between human and machine translation, effective strategies for metaphor/metonymy transfer, and readers’ cognitive preferences.

Conclusion: The study provides a cognitive, efficient, and replicable human-in-the-loop methodological pathway for translating ancient, concept-dense texts like Traditional Chinese Medicine, demonstrating that prompt-adjusted LLM translations can effectively convey complex conceptual networks better than literal human translations.

Abstract: Traditional Chinese Medicine theory is built on imagistic thinking, in which medical principles and diagnostic and therapeutic logic are structured through metaphor and metonymy. However, existing English translations largely rely on literal rendering, making it difficult for target-language readers to reconstruct the underlying conceptual networks and apply them in clinical practice. This study adopted a human-in-the-loop framework and selected four passages from the medical canon Huangdi Neijing that are fundamental in theory. Through prompt-based cognitive scaffolding, DeepSeek V3.1 was guided to identify metaphor and metonymy in the source text and convey the theory in translation. In the evaluation stage, ChatGPT 5 Pro and Gemini 2.5 Pro were instructed by prompts to simulate three types of real-world readers. Human translations, baseline model translations, and prompt-adjusted translations were scored by the simulated readers across five cognitive dimensions, followed by structured interviews and Interpretative Phenomenological Analysis. Results show that the prompt-adjusted LLM translations perform best across all five dimensions, with high cross-model and cross-role consistency. The interview themes reveal differences between human and machine translation, effective strategies for metaphor and metonymy transfer, and readers’ cognitive preferences. This study provides a cognitive, efficient and replicable HITL methodological pathway for translation of ancient, concept-dense texts like TCM.

Method: The study creates a parallel corpus of accented-unaccented ślokas and compares three computational approaches: (1) full fine-tuning of ByT5 (byte-level Transformer), (2) from-scratch BiLSTM-CRF sequence labeling baseline, and (3) LoRA-based parameter-efficient fine-tuning on ByT5. The methods emphasize Unicode-safe preprocessing and mark-aware tokenization.

Result: Full ByT5 fine-tuning achieved the lowest error rates across all metrics (Word Error Rate, Character Error Rate, and Diacritic Error Rate). LoRA offered strong efficiency-accuracy trade-offs, while BiLSTM-CRF served as a transparent baseline. The study established reproducible baselines for Rigvedic accent restoration.

Conclusion: The research demonstrates practical requirements for accent restoration in heritage languages and positions this as an emerging NLP area connecting computational modeling with philological and pedagogical aims. The results provide guidance for downstream applications like accent-aware OCR, ASR/chant synthesis, and digital scholarship.

Abstract: The Rigveda, among the oldest Indian texts in Vedic Sanskrit, employs a distinctive pitch-accent system : udātta, anudātta, svarita whose marks encode melodic and interpretive cues but are often absent from modern e-texts. This work develops a parallel corpus of accented-unaccented ślokas and conducts a controlled comparison of three strategies for automatic accent placement in Rigvedic verse: (i) full fine-tuning of ByT5, a byte-level Transformer that operates directly on Unicode combining marks, (ii) a from-scratch BiLSTM-CRF sequence-labeling baseline, and (iii) LoRA-based parameter-efficient fine-tuning atop ByT5. Evaluation uses Word Error Rate (WER) and Character Error Rate (CER) for orthographic fidelity, plus a task-specific Diacritic Error Rate (DER) that isolates accent edits. Full ByT5 fine-tuning attains the lowest error across all metrics; LoRA offers strong efficiency-accuracy trade-offs, and BiLSTM-CRF serves as a transparent baseline. The study underscores practical requirements for accent restoration - Unicode-safe preprocessing, mark-aware tokenization, and evaluation that separates grapheme from accent errors - and positions heritage-language technology as an emerging NLP area connecting computational modeling with philological and pedagogical aims. Results establish reproducible baselines for Rigvedic accent restoration and provide guidance for downstream tasks such as accent-aware OCR, ASR/chant synthesis, and digital scholarship.

Method: Tested five LLMs (including GPT-4 and open-source alternatives) across three datasets (Essays, MyPersonality, Pandora) using two prompting strategies: minimal vs. enriched with linguistic and psychological cues.

Result: Enriched prompts reduce invalid outputs but introduce bias toward predicting trait presence. Performance varies: Openness and Agreeableness easier to detect, Extraversion and Neuroticism challenging. No configuration yields consistently reliable predictions. Aggregate metrics mask significant asymmetries.

Conclusion: Current out-of-the-box LLMs are not yet suitable for automatic personality prediction from text, requiring careful coordination of prompt design, trait framing, and evaluation metrics for interpretable results.

Abstract: We evaluate large language models (LLMs) for automatic personality prediction from text under the binary Five Factor Model (BIG5). Five models – including GPT-4 and lightweight open-source alternatives – are tested across three heterogeneous datasets (Essays, MyPersonality, Pandora) and two prompting strategies (minimal vs. enriched with linguistic and psychological cues). Enriched prompts reduce invalid outputs and improve class balance, but also introduce a systematic bias toward predicting trait presence. Performance varies substantially: Openness and Agreeableness are relatively easier to detect, while Extraversion and Neuroticism remain challenging. Although open-source models sometimes approach GPT-4 and prior benchmarks, no configuration yields consistently reliable predictions in zero-shot binary settings. Moreover, aggregate metrics such as accuracy and macro-F1 mask significant asymmetries, with per-class recall offering clearer diagnostic value. These findings show that current out-of-the-box LLMs are not yet suitable for APPT, and that careful coordination of prompt design, trait framing, and evaluation metrics is essential for interpretable results.

Method: Dripper introduces four innovations: (1) specialized HTML simplification algorithm reducing tokens to 22% while preserving structure, (2) reformulating extraction as semantic block sequence classification, (3) controlled decoding mechanism to prevent hallucinations, and (4) WebMainBench dataset with 7,800 annotated web pages.

Result: Using only a 0.6B parameter model, Dripper achieves state-of-the-art performance across all benchmarks, with 81.58% ROUGE-N F1 score on WebMainBench (83.13% with fall-back strategy), outperforming all baseline methods.

Conclusion: Dripper demonstrates that lightweight language models can effectively extract web content when combined with specialized techniques for HTML simplification, task reformulation, and controlled decoding, offering an efficient alternative to larger models while maintaining high accuracy.

Abstract: Accurately and efficiently extracting main content from general web pages is of great significance for obtaining training data for large models. Using well-pre-trained decoder-only generative language models offers excellent document comprehension capabilities, thereby effectively enhancing parsing quality. However, it remains constrained by issues such as context window length, inference cost, and format hallucination. We present Dripper, an efficient HTML main content extraction framework powered by lightweight language models, which addresses these challenges through four key innovations: (1) We design a specialized HTML simplification algorithm that reduces input token count to 22% compared to raw HTML while preserving critical structural information; (2) We reformulate main content extraction as a semantic block sequence classification task, significantly reducing inference cost; (3) We introduce a controlled decoding mechanism that strictly constrains the output space through logits processors, effectively eliminating hallucination issues common in small-scale models; (4) We propose WebMainBench, an evaluation dataset containing over 7,800 web pages with meticulously human-annotated main content extraction labels. Experimental results demonstrate that using only a 0.6B parameter model, Dripper achieves state-of-the-art performance across all evaluation benchmarks and outperforms all baseline methods, attaining an ROUGE-N F1 score of 81.58%( 83.13% with fall-back strategy) on our proposed WebMainBench dataset.

Method: Multi-chain Graph Refinement & Selection (MGRS) framework: 1) generates multiple diverse reasoning trajectories, 2) refines candidate responses using composite self- and cross-verification, 3) constructs a reasoning relation graph and estimates success rates of intermediate nodes, 4) computes cumulative success rates to select the most reliable answer and reasoning trajectory.

Result: Achieves average accuracy of 82.9% across six benchmark datasets spanning four tasks, outperforming SOTA baselines by 2.1%. On 24-point game, attains 100% accuracy for the first time with 13.6x speed-up compared to Forest of Thoughts framework.

Conclusion: MGRS significantly advances both reasoning capability and computational efficiency of reasoning enhancement methods, addressing key limitations of existing approaches through diverse trajectory generation, verification-based refinement, and graph-based selection.

Abstract: The complex reasoning ability of Large Language Models (LLMs) poses a critical bottleneck for their practical applications. Test-time expansion methods such as Tree-of-Thought (ToT) and Graph-of-Thought (GoT) enhance reasoning by introducing intermediate reasoning structures, tree search, or graph-based exploration mechanisms. However, their reasoning strategies suffer from limited diversity, redundant search branches, and inadequate integration and error correction across heterogeneous reasoning paths. To address these limitations, we propose a novel reasoning framework called Multi-chain Graph Refinement & Selection (MGRS), which first generates multiple diverse reasoning trajectories for a given problem, refines candidate responses using a composite self- and cross-verification strategy, then constructs a reasoning relation graph and estimates the success rate of intermediate nodes, and finally computes cumulative success rates to select the most reliable answer and corresponding reasoning trajectory. Experimental results demonstrate that MGRS significantly advances both the reasoning capability and computational efficiency of reasoning enhancement methods. Across six benchmark datasets spanning four distinct tasks, MGRS achieves an average accuracy of 82.9%, outperforming state-of-the-art baselines by a clear margin of 2.1%. Remarkably, on the 24-point game, MGRS attains 100% accuracy for the first time, while delivering a 13.6x speed-up compared to the leading Forest of Thoughts framework.

Method: Introduced PSP dataset built from censored content: 1) sensitive prompts in China generalized to multiple countries, 2) tweets censored in various countries. Studied impact through data-driven (making PSP implicit) and representation-level approaches (erasing politics concept), plus vulnerability testing via prompt injection attacks.

Result: Most LLMs perform some form of censorship, associating censorship with refusals on content with masked implicit intent. Identified major attributes causing shifts in refusal distributions across models and country contexts.

Conclusion: LLMs’ refusal behaviors align with political censorship patterns, revealing that safety policies often mask political censorship. The study provides framework for analyzing political bias in AI systems across different national contexts.

Abstract: Large Language Models (LLMs) are trained to refuse to respond to harmful content. However, systematic analyses of whether this behavior is truly a reflection of its safety policies or an indication of political censorship, that is practiced globally by countries, is lacking. Differentiating between safety influenced refusals or politically motivated censorship is hard and unclear. For this purpose we introduce PSP, a dataset built specifically to probe the refusal behaviors in LLMs from an explicitly political context. PSP is built by formatting existing censored content from two data sources, openly available on the internet: sensitive prompts in China generalized to multiple countries, and tweets that have been censored in various countries. We study: 1) impact of political sensitivity in seven LLMs through data-driven (making PSP implicit) and representation-level approaches (erasing the concept of politics); and, 2) vulnerability of models on PSP through prompt injection attacks (PIAs). Associating censorship with refusals on content with masked implicit intent, we find that most LLMs perform some form of censorship. We conclude with summarizing major attributes that can cause a shift in refusal distributions across models and contexts of different countries.

Method: The approach uses reasoning-based generation to output both the quadruple and natural language rationale within a unified template. It introduces listwise preference optimization with element-wise confusable candidates generated via syntactic and semantic proximity, training models to prefer gold candidates over close alternatives.

Result: Extensive experiments on four benchmark datasets demonstrate that the framework effectively improves quadruple prediction accuracy and explanation consistency compared to prior methods.

Conclusion: The reasoning-based generation with listwise preference optimization addresses limitations of prior marker-based approaches by enhancing relational reasoning, interpretability, and element-wise alignment in ASQP tasks.

Abstract: Aspect sentiment quad prediction (ASQP) is inherently challenging to predict a structured quadruple with four core sentiment elements, including aspect term (a), aspect category (c), opinion term (o), and sentiment polarity (s). Prior methods relying on marker-based prediction struggle with modeling the intricate relationships among elements and experience sharp performance declines when predicting higher-order elements (e.g., c and s) under standard supervised fine-tuning. To address these limitations, we employ reasoning-based generation to output both the quadruple and a natural language rationale under element prefixes within a unified template, encouraging explicit relational reasoning and interpretability. To further enhance element-wise alignment, we introduce a listwise preference optimization framework for improving structural validity and relational coherence. Specifically, we generate element-wise confusable candidates via syntactic and semantic proximity, then train the model with listwise objectives to prefer the gold candidates over closely competing alternatives. Extensive experiments on four benchmark datasets demonstrate that our framework effectively improves quadruple prediction accuracy and explanation consistency.

Method: The paper challenges the conventional wisdom that outliers are data-driven, showing they are mechanically-produced artifacts from weight matrix colinearity. TWEO (Transformers Without Extreme Outliers) is proposed as a non-invasive loss function with a simple loss term that prevents extreme outliers.

Result: TWEO reduces outliers from 10000+ to less than 20, enables full-model FP8 pre-training for both LLMs and ViTs without engineering tricks or architectural changes. Achieves BF16-level performance with 36% training throughput increase, and enables SOTA W8A8 per-tensor static quantization previously considered unusable due to outliers.

Conclusion: TWEO fundamentally changes outlier understanding and enables practical FP8 training, opening new quantization paradigms for efficient Transformer deployment without sacrificing performance.

Abstract: Native FP8 support in modern hardware is essential for training large Transformers, but is severely hindered by extreme activation outliers. Existing solutions either rely on complex mixed-precision engineering or invasive architectural modifications. This paper fundamentally challenges the conventional wisdom that outliers are data-driven. We demonstrate that extreme outliers are a data-independent, mechanically-produced artifact of training, originating from specific structural properties of the weight matrices (i.e., colinearity). Based on this insight, we propose TWEO (Transformers Without Extreme Outliers), a novel, non-invasive loss function. TWEO effectively prevents extreme outliers via a very simple loss term, which reduces outliers from 10000+ to less than 20. TWEO then enables full-model FP8 pre-training with neither engineering tricks nor architectural changes for both LLM and ViT. When standard FP8 training catastrophically collapses, TWEO achieves performance comparable to the BF16 baseline while delivering a 36% increase in training throughput. Also, TWEO enables a new quantization paradigm. Hardware-friendly W8A8 per-tensor static quantization of LLMs, previously considered completely unusable due to outliers, achieves SOTA performance for the first time on TWEO-trained models.

Method: Created dataset of 7,715 Hindi QA pairs + 27,455 augmented via Llama zero-shot. Fine-tuned BERT/RoBERTa variants using SFT and LoRA for parameter efficiency.

Result: LoRA achieves 85.3% F1 with 98% fewer trainable parameters than SFT. RoBERTa with SFT outperforms BERT in capturing cultural nuances like “Aarti” and “Kund”.

Conclusion: Establishes foundational baseline for Hindi tourism QA, emphasizing LoRA’s efficiency in low-resource settings and need for culturally contextualized NLP frameworks.

Abstract: This article presents the first comprehensive study on designing a baseline extractive question-answering (QA) system for the Hindi tourism domain, with a specialized focus on the Varanasi-a cultural and spiritual hub renowned for its Bhakti-Bhaav (devotional ethos). Targeting ten tourism-centric subdomains-Ganga Aarti, Cruise, Food Court, Public Toilet, Kund, Museum, General, Ashram, Temple and Travel, the work addresses the absence of language-specific QA resources in Hindi for culturally nuanced applications. In this paper, a dataset comprising 7,715 Hindi QA pairs pertaining to Varanasi tourism was constructed and subsequently augmented with 27,455 pairs generated via Llama zero-shot prompting. We propose a framework leveraging foundation models-BERT and RoBERTa, fine-tuned using Supervised Fine-Tuning (SFT) and Low-Rank Adaptation (LoRA), to optimize parameter efficiency and task performance. Multiple variants of BERT, including pre-trained languages (e.g., Hindi-BERT), are evaluated to assess their suitability for low-resource domain-specific QA. Evaluation metrics - F1, BLEU, and ROUGE-L - highlight trade-offs between answer precision and linguistic fluency. Experiments demonstrate that LoRA-based fine-tuning achieves competitive performance (85.3% F1) while reducing trainable parameters by 98% compared to SFT, striking a balance between efficiency and accuracy. Comparative analysis across models reveals that RoBERTa with SFT outperforms BERT variants in capturing contextual nuances, particularly for culturally embedded terms (e.g., Aarti, Kund). This work establishes a foundational baseline for Hindi tourism QA systems, emphasizing the role of LORA in low-resource settings and underscoring the need for culturally contextualized NLP frameworks in the tourism domain.

Method: Three-stage training framework: 1) Train [BE] token to reconstruct original system prompt content, 2) Distill prompt’s downstream behavior into single token, 3) No model internals, compression models, or labeled responses needed.

Result: Single [BE] token achieves up to 3000x reduction in prompt length while retaining ~98% of original system prompt performance on three datasets.

Conclusion: Behavior-Equivalent tokens enable dramatic prompt compression, reducing inference cost and freeing context window for user inputs without sacrificing performance.

Abstract: Carefully engineered system prompts play a critical role in guiding the behavior of LLM agents, but their considerable length introduces significant drawbacks, including increased inference latency, higher computational cost, and reduced effective context length. This raises the question of whether such lengthy prompts can be replaced by a drastically reduced number of tokens while preserving their behavioral effect on downstream tasks. To enable this, we propose a lightweight three-stage training framework that learns a single prompt-specific Behavior-Equivalent token ([BE]). The framework first trains [BE] to encode the natural-language content of the original system prompt via reconstruction, and then distills the prompt ’s downstream behavior into this single token. Importantly, our method requires no access to model internals, no auxiliary compression models, and no labeled responses. Empirical evaluations on three datasets show that a single [BE] token achieves up to a 3000x reduction in prompt length, while retaining about 98% of the downstream performance of the original system prompts. This substantially reduces inference cost and leaves almost the entire context window available for user inputs.

Method: Created testbed with four simulated e-shops offering products via HTML, MCP, and NLWeb interfaces. Developed specialized agents for each interface to perform identical e-commerce tasks (product search, price comparison, complementary/substitute queries, checkout). Evaluated using GPT 4.1, GPT 5, GPT 5 mini, and Claude Sonnet 4.

Result: RAG, MCP and NLWeb agents outperformed HTML on effectiveness and efficiency. F1 rose from 0.67 (HTML) to 0.75-0.77. Token usage dropped from ~241k to 47k-140k per task. Runtime fell from 291s to 50-62s. Best configuration: RAG with GPT 5 (F1=0.87, completion rate=0.79). RAG with GPT 5 mini offers good cost-performance balance.

Conclusion: Choice of interaction interface significantly impacts effectiveness and efficiency of LLM-based web agents. RAG, MCP, and NLWeb interfaces provide substantial improvements over traditional HTML browsing for automated web tasks.

Abstract: Large language model agents are increasingly used to automate web tasks such as product search, offer comparison, and checkout. Current research explores different interfaces through which these agents interact with websites, including traditional HTML browsing, retrieval-augmented generation (RAG) over pre-crawled content, communication via Web APIs using the Model Context Protocol (MCP), and natural-language querying through the NLWeb interface. However, no prior work has compared these four architectures within a single controlled environment using identical tasks. To address this gap, we introduce a testbed consisting of four simulated e-shops, each offering its products via HTML, MCP, and NLWeb interfaces. For each interface (HTML, RAG, MCP, and NLWeb) we develop specialized agents that perform the same sets of tasks, ranging from simple product searches and price comparisons to complex queries for complementary or substitute products and checkout processes. We evaluate the agents using GPT 4.1, GPT 5, GPT 5 mini, and Claude Sonnet 4 as underlying LLM. Our evaluation shows that the RAG, MCP and NLWeb agents outperform HTML on both effectiveness and efficiency. Averaged over all tasks, F1 rises from 0.67 for HTML to between 0.75 and 0.77 for the other agents. Token usage falls from about 241k for HTML to between 47k and 140k per task. The runtime per task drops from 291 seconds to between 50 and 62 seconds. The best overall configuration is RAG with GPT 5 achieving an F1 score of 0.87 and a completion rate of 0.79. Also taking cost into consideration, RAG with GPT 5 mini offers a good compromise between API usage fees and performance. Our experiments show the choice of the interaction interface has a substantial impact on both the effectiveness and efficiency of LLM-based web agents.

Method: Proposes Hierarchical Sparse Attention (HSA) mechanism that satisfies the three key properties, integrated into Transformers to create HSA-UltraLong, an 8B-parameter MoE model trained on over 8 trillion tokens.

Result: Model performs comparably to full-attention baselines on in-domain lengths and achieves over 90% accuracy on most in-context retrieval tasks with contexts up to 16M tokens.

Conclusion: The work contributes a foundation for future research in ultra-long context modeling, outlining experimental insights and open problems for building machines with effective long-term memory.

Abstract: This work explores the challenge of building ``Machines that Can Remember’’, framing long-term memory as the problem of efficient ultra-long context modeling. We argue that this requires three key properties: \textbf{sparsity}, \textbf{random-access flexibility}, and \textbf{length generalization}. To address ultra-long-context modeling, we leverage Hierarchical Sparse Attention (HSA), a novel attention mechanism that satisfies all three properties. We integrate HSA into Transformers to build HSA-UltraLong, which is an 8B-parameter MoE model trained on over 8 trillion tokens and is rigorously evaluated on different tasks with in-domain and out-of-domain context lengths to demonstrate its capability in handling ultra-long contexts. Results show that our model performs comparably to full-attention baselines on in-domain lengths while achieving over 90% accuracy on most in-context retrieval tasks with contexts up to 16M. This report outlines our experimental insights and open problems, contributing a foundation for future research in ultra-long context modeling.

Method: Three methods based on CPI+DMC approach using SS3, BERT with extended vocabulary, and SBERT models, combined with decision policies based on historical user analysis.

Result: Two proposals achieved top two positions in the official challenge results, performing notably in decision metrics, though distinguishing between high and low risk users remained challenging.

Conclusion: The work demonstrates promising results but highlights the need for improved data interpretation, better data quality, and more transparent/reliable ERD systems for mental disorders.

Abstract: Gambling disorder is a complex behavioral addiction that is challenging to understand and address, with severe physical, psychological, and social consequences. Early Risk Detection (ERD) on the Web has become a key task in the scientific community for identifying early signs of mental health behaviors based on social media activity. This work presents our participation in the MentalRiskES 2025 challenge, specifically in Task 1, aimed at classifying users at high or low risk of developing a gambling-related disorder. We proposed three methods based on a CPI+DMC approach, addressing predictive effectiveness and decision-making speed as independent objectives. The components were implemented using the SS3, BERT with extended vocabulary, and SBERT models, followed by decision policies based on historical user analysis. Although it was a challenging corpus, two of our proposals achieved the top two positions in the official results, performing notably in decision metrics. Further analysis revealed some difficulty in distinguishing between users at high and low risk, reinforcing the need to explore strategies to improve data interpretation and quality, and to promote more transparent and reliable ERD systems for mental disorders.

Method: A task-agnostic structured knowledge hunter using two-tier knowledge architecture (high-level entities and low-level knowledge triples). Employs local-global interaction for knowledge representation learning and hierarchical transformer-based pointer network to select relevant knowledge triples and entities.

Result: The model outperforms state-of-the-art methods and corresponding language models on both internal knowledge-enhanced table-to-text generation (RotoWireFG dataset) and external knowledge-enhanced dialogue response generation (KdConv dataset).

Conclusion: The proposed approach combines language models’ generative ability with knowledge hunter’s faithfulness to achieve high interpretability and performance, setting new standards for knowledge-enhanced text generation benchmarks.

Abstract: Knowledge-enhanced text generation aims to enhance the quality of generated text by utilizing internal or external knowledge sources. While language models have demonstrated impressive capabilities in generating coherent and fluent text, the lack of interpretability presents a substantial obstacle. The limited interpretability of generated text significantly impacts its practical usability, particularly in knowledge-enhanced text generation tasks that necessitate reliability and explainability. Existing methods often employ domain-specific knowledge retrievers that are tailored to specific data characteristics, limiting their generalizability to diverse data types and tasks. To overcome this limitation, we directly leverage the two-tier architecture of structured knowledge, consisting of high-level entities and low-level knowledge triples, to design our task-agnostic structured knowledge hunter. Specifically, we employ a local-global interaction scheme for structured knowledge representation learning and a hierarchical transformer-based pointer network as the backbone for selecting relevant knowledge triples and entities. By combining the strong generative ability of language models with the high faithfulness of the knowledge hunter, our model achieves high interpretability, enabling users to comprehend the model output generation process. Furthermore, we empirically demonstrate the effectiveness of our model in both internal knowledge-enhanced table-to-text generation on the RotoWireFG dataset and external knowledge-enhanced dialogue response generation on the KdConv dataset. Our task-agnostic model outperforms state-of-the-art methods and corresponding language models, setting new standards on the benchmark.

Method: Introduced SSA-HuBERT-Large (317M parameters) and SSA-HuBERT-XL (964M parameters) - the first large models trained solely on African speech, alongside a BASE size counterpart. Conducted controlled experimental study focused exclusively on Sub-Saharan languages covering ASR and language identification tasks.

Result: Larger architectures significantly improve performance by effectively leveraging large audio datasets. The models demonstrate tangible benefits of scaling up model capacity when trained exclusively on African speech data.

Conclusion: Large-scale models trained exclusively on African speech offer substantial performance improvements for African language processing tasks, addressing the representation gap and providing valuable open-weight resources for the research community.

Abstract: Despite recent progress in multilingual speech processing, African languages remain under-represented in both research and deployed systems, particularly when it comes to strong, open-weight encoders that transfer well under low-resource supervision. Self-supervised learning has proven especially promising in such settings, yet most publicly released models targeting African speech remain at BASE scale, leaving unanswered whether larger encoders, trained exclusively on Africa-centric audio, offer tangible benefits and how model capacity interacts with data composition. This work addresses that gap by introducing SSA-HuBERT-Large (317M parameters) and SSA-HuBERT-XL (964M parameters), the first large models trained solely on African speech, alongside a BASE size counterpart. We release these models as open weights: see https://huggingface.co/collections/Orange/african-speech-foundation-models. By conducting a carefully controlled experimental study focused exclusively on Sub-Saharan languages, covering automatic speech recognition (ASR) and language identification (LID) tasks, we demonstrate that larger architectures significantly improve performance by effectively leveraging large audio datasets.

Method: Analyzes attention scores relative to image tokens to identify attention heads that focus on image key objects, calculates Head Impact (HI) scores to quantify focus on key objects, and uses this to select optimal layers for PEFT.

Result: Fine-tuning layers with highest HI scores leads to most significant metric improvements compared to pre-trained, randomly selected, or lowest-HI-score layers. Fine-tuning just 0.01% of parameters in crucial layers substantially influences image understanding capabilities.

Conclusion: Attention-based interpretability can effectively identify important components in multimodal models for targeted parameter-efficient fine-tuning, enabling efficient adaptation with minimal parameter updates while maintaining performance.

Abstract: Modern large language models become multimodal, analyzing various data formats like text and images. While fine-tuning is effective for adapting these multimodal language models (MLMs) to downstream tasks, full fine-tuning is computationally expensive. Parameter-Efficient Fine-Tuning (PEFT) methods address this by training only a small portion of model weights. However, MLMs are difficult to interpret, making it challenging to identify which components are most effective for training to balance efficiency and performance. We propose an attention-based interpretability method for MLMs by analyzing attention scores relative to image tokens. The core idea is to identify attention heads that focus on image key objects. We utilize this information to select optimal model components for PEFT in multimodal models. Our contributions include a method for identifying attention heads associated with image key objects, its application to PEFT for image captioning, and the creation of a new dataset containing images, key object masks, and their textual descriptions. We conducted experiments on MLMs with 2-3 billion parameters to validate the method’s effectiveness. By calculating Head Impact (HI) scores we quantify an attention head’s focus on key objects, indicating its significance in image understanding. Our fine-tuning experiments demonstrate that adapting layers with the highest HI scores leads to the most significant shifts in metrics compared to pre-trained, randomly selected, or lowest-HI-score layers. This indicates that fine-tuning a small percentage (around 0.01%) of parameters in these crucial layers can substantially influence image understanding capabilities.

Method: Ambiguity Awareness Optimization (AAO) automatically re-weights ambiguous content by calculating semantic similarity from preference pairs to reduce ambiguities.

Result: AAO consistently outperforms state-of-the-art approaches across multiple model scales and benchmarks: up to 8.9 points improvement on AlpacaEval 2 and up to 15.0 points on Arena-Hard over DPO, without significantly increasing response length.

Conclusion: AAO effectively addresses the ambiguity problem in DPO training through semantic similarity-based re-weighting, achieving significant performance improvements across various benchmarks and model scales.

Abstract: Direct Preference Optimization (DPO) is a widely used reinforcement learning from human feedback (RLHF) method across various domains. Recent research has increasingly focused on the role of token importance in improving DPO effectiveness. It is observed that identical or semantically similar content (defined as ambiguous content) frequently appears within the preference pairs. We hypothesize that the presence of ambiguous content during DPO training may introduce ambiguity, thereby limiting further improvements in alignment. Through mathematical analysis and proof-of-concept experiments, we reveal that ambiguous content may potentially introduce ambiguities, thereby degrading performance. To address this issue, we introduce Ambiguity Awareness Optimization (AAO), a simple yet effective approach that automatically re-weights ambiguous content to reduce ambiguities by calculating semantic similarity from preference pairs. Through extensive experiments, we demonstrate that AAO consistently and significantly surpasses state-of-the-art approaches in performance, without markedly increasing response length, across multiple model scales and widely adopted benchmark datasets, including AlpacaEval 2, MT-Bench, and Arena-Hard. Specifically, AAO outperforms DPO by up to 8.9 points on AlpacaEval 2 and achieves an improvement of by up to 15.0 points on Arena-Hard.

Method: Learns data representations as task-specific compositions of pre-trained token representations shared across all tasks. This grounds correlations between different tasks’ data representations in correlations between pre-trained token representations, promoting global alignment.

Result: Achieves state-of-the-art performance in continual learning tasks without experience replay. Also achieves advanced class-incremental performance through task-incremental training.

Conclusion: The proposed method effectively mitigates catastrophic forgetting by addressing gradient interference through global alignment of data representations via pre-trained token composition, demonstrating strong performance in continual learning scenarios.

Abstract: Continual learning aims to sequentially learn new tasks without forgetting previous tasks’ knowledge (catastrophic forgetting). One factor that can cause forgetting is the interference between the gradients on losses from different tasks. When the gradients on the current task’s loss are in opposing directions to those on previous tasks’ losses, updating the model for the current task may cause performance degradation on previous tasks. In this paper, we first identify causes of the above interference, and hypothesize that correlations between data representations are a key factor of interference. We then propose a method for promoting appropriate correlations between arbitrary tasks’ data representations (i.e., global alignment) in individual task learning. Specifically, we learn the data representation as a task-specific composition of pre-trained token representations shared across all tasks. Then the correlations between different tasks’ data representations are grounded by correlations between pre-trained token representations. We explore different ways to learn such compositions. Without experience replay, our model achieves SOTA performance in continual learning tasks. It also achieves advanced class-incremental performance through task-incremental training.

Method: AutoHall automatically constructs model-specific hallucination datasets using existing fact-checking datasets. Also proposes a zero-resource, black-box hallucination detection method based on self-contradiction to identify hallucinations in the constructed dataset.

Result: Empirical results show variations in hallucination proportions and types among different LLMs. The proposed detection method achieves superior performance compared to baselines. Analysis provides insights into factors contributing to LLM hallucinations.

Conclusion: AutoHall provides a cost-effective solution for creating model-specific hallucination datasets, enabling better hallucination detection and analysis for trustworthy LLMs.

Abstract: Large language models (LLMs) have gained broad applications across various domains but still struggle with hallucinations. Currently, hallucinations occur frequently in the generation of factual content and pose a great challenge to trustworthy LLMs. However, hallucination detection is hindered by the laborious and expensive manual annotation of hallucinatory content. Meanwhile, as different LLMs exhibit distinct types and rates of hallucination, the collection of hallucination datasets is inherently model-specific, which also increases the cost. To address this issue, this paper proposes a method called $\textbf{AutoHall}$ for $\underline{Auto}$matically constructing model-specific $\underline{Hall}$ucination datasets based on existing fact-checking datasets. The empirical results reveal variations in hallucination proportions and types among different models. Moreover, we introduce a zero-resource and black-box hallucination detection method based on self-contradiction to recognize the hallucination in our constructed dataset, achieving superior detection performance compared to baselines. Further analysis on our dataset provides insight into factors that may contribute to LLM hallucinations. Our codes and datasets are publicly available at https://github.com/zouyingcao/AutoHall.

Method: Develops two evaluation frameworks: 1) reference-based factuality evaluation that uses ground truth, and 2) reference-free factuality evaluation that doesn’t need ground truth, making it more widely applicable.

Result: Experimental results show effectiveness of the proposed frameworks, with correlation analysis demonstrating their validity compared to other metrics.

Conclusion: The FALLACIOUS frameworks provide fine-grained, explainable evaluation of factuality in multimodal summarization, with the reference-free version offering broader applicability since it doesn’t require ground truth data.

Abstract: Multimodal summarization aims to generate a concise summary based on the input text and image. However, the existing methods potentially suffer from unfactual output. To evaluate the factuality of multimodal summarization models, we propose two fine-grained and explainable evaluation frameworks (FALLACIOUS) for different application scenarios, i.e. reference-based factuality evaluation framework and reference-free factuality evaluation framework. Notably, the reference-free factuality evaluation framework doesn’t need ground truth and hence it has a wider application scenario. To evaluate the effectiveness of the proposed frameworks, we compute the correlation between our frameworks and the other metrics. The experimental results show the effectiveness of our proposed method. We will release our code and dataset via github.

Method: The paper uses a systematic review methodology: (1) outlining technical representations of biomolecules (sequences, 2D graphs, 3D structures), (2) examining rationale for multi-modal integration, (3) surveying practical applications, (4) compiling available resources/datasets, and (5) identifying future research directions.

Result: The review provides an extensive analysis of recent advancements in biomolecule-language cross-modeling, including technical representations, integration rationales, practical applications, and available resources for the research community.

Conclusion: Biomolecule-language cross-modeling is a promising interdisciplinary field that opens new avenues for comprehensive biomolecular analysis by fusing natural language narratives with structural modeling techniques, with identified future directions for continued advancement.

Abstract: The integration of biomolecular modeling with natural language (BL) has emerged as a promising interdisciplinary area at the intersection of artificial intelligence, chemistry and biology. This approach leverages the rich, multifaceted descriptions of biomolecules contained within textual data sources to enhance our fundamental understanding and enable downstream computational tasks such as biomolecule property prediction. The fusion of the nuanced narratives expressed through natural language with the structural and functional specifics of biomolecules described via various molecular modeling techniques opens new avenues for comprehensively representing and analyzing biomolecules. By incorporating the contextual language data that surrounds biomolecules into their modeling, BL aims to capture a holistic view encompassing both the symbolic qualities conveyed through language as well as quantitative structural characteristics. In this review, we provide an extensive analysis of recent advancements achieved through cross modeling of biomolecules and natural language. (1) We begin by outlining the technical representations of biomolecules employed, including sequences, 2D graphs, and 3D structures. (2) We then examine in depth the rationale and key objectives underlying effective multi-modal integration of language and molecular data sources. (3) We subsequently survey the practical applications enabled to date in this developing research area. (4) We also compile and summarize the available resources and datasets to facilitate future work. (5) Looking ahead, we identify several promising research directions worthy of further exploration and investment to continue advancing the field. The related resources and contents are updating in https://github.com/QizhiPei/Awesome-Biomolecule-Language-Cross-Modeling.

Method: Introduces normal forms for virus machines that restrict features including: (a) number of hosts, (b) number of instructions, and (c) number of virus objects in each host. Studies the size of loops in the network and proves new characterizations through these normal forms.

Result: Provides new characterizations of families of sets: finite sets, semilinear sets, and recursively enumerable sets (NRE) through the introduced normal forms for virus machines.

Conclusion: The study of normal forms for virus machines enhances understanding of their computational power and provides new characterizations of important computational complexity classes, demonstrating the expressive power of this biologically-inspired computing model.

Abstract: In the present work, we further study the computational power of virus machines (VMs in short).VMs provide a computing paradigm inspired by the transmission and replication networks of viruses.VMs consist of process units (called hosts) structured by a directed graph whose arcs are called channels and an instruction graph that controls the transmissions of virus objects among hosts. The present work complements our understanding of the computing power of VMs by introducing normal forms; these expressions restrict the features in a given computing model.Some of the features that we restrict in our normal forms include (a) the number of hosts, (b) the number of instructions, and (c) the number of virus objects in each host. After we recall some known results on the computing power of VMs we give our series of normal forms, such as the size of the loops in the network, proving new characterisations of family of sets, such as finite sets, semilinear sets, or recursively enumerable sets (NRE).

Method: Developed AIPatient using LLM-based AI agents with RAG framework (six task-specific agents) and integrated with knowledge graph built from de-identified MIMIC-III ICU patient data for enhanced realism.

Result: 94.15% QA accuracy with all agents enabled; knowledge base F1 score 0.89; good readability (Flesch Reading Ease 68.77, Grade 6.4); robust/stability (non-significant variance); user study showed high fidelity, usability, educational value comparable to human simulated patients.

Conclusion: LLM-based simulated patient systems can deliver accurate, readable, reliable medical encounters and have strong potential to transform medical education, with AIPatient demonstrating effectiveness comparable to human simulated patients for history taking.

Abstract: Background: Simulated patient systems are important in medical education and research, providing safe, integrative training environments and supporting clinical decision making. Advances in artificial intelligence (AI), especially large language models (LLMs), can enhance simulated patients by replicating medical conditions and doctor patient interactions with high fidelity and at low cost, but effectiveness and trustworthiness remain open challenges. Methods: We developed AIPatient, a simulated patient system powered by LLM based AI agents. The system uses a retrieval augmented generation (RAG) framework with six task specific agents for complex reasoning. To improve realism, it is linked to the AIPatient knowledge graph built from de identified real patient data in the MIMIC III intensive care database. Results: We evaluated electronic health record (EHR) based medical question answering (QA), readability, robustness, stability, and user experience. AIPatient reached 94.15 percent QA accuracy when all six agents were enabled, outperforming versions with partial or no agent integration. The knowledge base achieved an F1 score of 0.89. Readability scores showed a median Flesch Reading Ease of 68.77 and a median Flesch Kincaid Grade of 6.4, indicating accessibility for most medical trainees and clinicians. Robustness and stability were supported by non significant variance in repeated trials (analysis of variance F value 0.61, p greater than 0.1; F value 0.78, p greater than 0.1). A user study with medical students showed that AIPatient provides high fidelity, usability, and educational value, comparable to or better than human simulated patients for history taking. Conclusions: LLM based simulated patient systems can deliver accurate, readable, and reliable medical encounters and show strong potential to transform medical education.

Method: Encoder-decoder framework (LingConv) with novel inference-time quality control mechanism that iteratively refines attribute embeddings to ensure generated paraphrases match target attributes without sacrificing semantic fidelity.

Result: LingConv reduces attribute error by up to 34% over existing models, with the quality control mechanism contributing an additional 14% improvement.

Conclusion: LingConv enables more reliable and fine-grained controlled paraphrase generation with significant improvements in attribute matching accuracy through iterative refinement during inference.

Abstract: Controlled paraphrase generation produces paraphrases that preserve meaning while allowing precise control over linguistic attributes of the output. We introduce LingConv, an encoder-decoder framework that enables fine-grained control over 40 linguistic attributes in English. To improve reliability, we introduce a novel inference-time quality control mechanism that iteratively refines attribute embeddings to generate paraphrases that closely match target attributes without sacrificing semantic fidelity. LingConv reduces attribute error by up to 34% over existing models, with the quality control mechanism contributing an additional 14% improvement.

Method: 1) Use crowdsourced Google Maps reviews (2018-2021), 2) Apply advanced NLP (DeBERTa) to create spatial-temporal perception index, 3) Employ Partial Least Squares (PLS) regression to link perception to socioeconomic/demographic drivers.

Result: Quantified significant spatial-temporal shifts in perceived access, showing disparities peaked during COVID-19 crisis with only partial recovery post-peak. Identified political affiliation, racial composition, and educational attainment as primary determinants.

Conclusion: Validates a scalable method for real-time health equity monitoring and provides actionable evidence for interventions to build more resilient healthcare infrastructure.

Abstract: Monitoring health resource disparities during public health crises is critical, yet traditional methods, like surveys, lack the requisite speed and spatial granularity. This study introduces a novel framework that leverages: 1) crowdsourced Google Maps reviews (2018-2021) and 2) advanced NLP (DeBERTa) to create a high-resolution, spatial-temporal index of public perception of health resource accessibility in the United States. We then employ Partial Least Squares (PLS) regression to link this perception index to a range of socioeconomic and demographic drivers. Our results quantify significant spatial-temporal shifts in perceived access, confirming that disparities peaked during the COVID-19 crisis and only partially recovered post-peak. We identify political affiliation, racial composition, and educational attainment as primary determinants of these perceptions. This study validates a scalable method for real-time health equity monitoring and provides actionable evidence for interventions to build a more resilient healthcare infrastructure.

Method: Proposes Atom of Thoughts (AoT) with a decomposition-contraction process: 1) Decompose current question into dependency-based directed acyclic graph of subquestions, 2) Contract subquestions into a simplified question maintaining answer equivalence with original problem. This forms a Markov reasoning process that can be integrated as a plug-in enhancement to existing test-time scaling methods.

Result: Experiments across six benchmarks show AoT’s effectiveness as both standalone framework and plug-in enhancement. On HotpotQA, when applied to gpt-4o-mini, achieves 80.6% F1 score, surpassing o3-mini by 3.4% and DeepSeek-R1 by 10.6%.

Conclusion: AoT addresses the problem of accumulated historical information in test-time scaling by leveraging atomic questions with memoryless properties, enabling more efficient and effective reasoning through decomposition-contraction processes that can enhance existing methods.

Abstract: Large Language Models (LLMs) achieve superior performance through training-time scaling, and test-time scaling further enhances their capabilities by conducting effective reasoning during inference. However, as the scale of reasoning increases, existing test-time scaling methods suffer from accumulated historical information, which not only wastes computational resources but also interferes with effective reasoning. To address this issue, we observe that complex reasoning can be achieved by solving a series of independent and self-contained subquestions. These subquestions are essentially \textit{atomic questions}, exhibiting the memoryless property similar to Markov processes. Based on this observation, we propose Atom of Thoughts (\our), where each state transition consists of decomposing the current question into a dependency-based directed acyclic graph and contracting its subquestions, forming a simplified question that maintains answer equivalence with the original problem. This answer preservation enables the iterative \textit{decomposition-contraction} process to naturally form a meaningful Markov reasoning process. Furthermore, these atomic states can be seamlessly integrated into existing test-time scaling methods, enabling \our to serve as a plug-in enhancement for improving reasoning capabilities. Experiments across six benchmarks demonstrate the effectiveness of \our both as a standalone framework and a plug-in enhancement. Notably, on HotpotQA, when applied to gpt-4o-mini, \our achieves an \textbf{80.6%} F1 score, surpassing o3-mini by \textbf{3.4%} and DeepSeek-R1 by \textbf{10.6%}. The code is available at \href{https://github.com/qixucen/atom}{https://github.com/qixucen/atom}.

Method: An iterative refinement pipeline employing the Unlabeled-Unlabeled learning framework that uses two unlabeled datasets with different positive class ratios to iteratively denoise and refine initial pseudo-labels, mitigating internal biases with minimal human supervision.

Result: The method consistently outperforms both initial LLM classification performance and state-of-the-art self-refinement approaches (GPT-4o, DeepSeek-R1) across diverse datasets including low-resource languages, patent classifications, and protein structure categorizations.

Conclusion: The refined classifier facilitates effective post-training alignment for safety in LLMs and enables successful self-refinement in generative tasks, providing a promising approach to enhance self-refinement for broader applications with minimal human supervision.

Abstract: Recent advances in large language models (LLMs) have yielded impressive performance on various tasks, yet they often depend on high-quality feedback that can be costly. Self-refinement methods attempt to leverage LLMs’ internal evaluation mechanisms with minimal human supervision; however, these approaches frequently suffer from inherent biases and overconfidence, especially in domains where the models lack sufficient internal knowledge, resulting in performance degradation. As an initial step toward enhancing self-refinement for broader applications, we introduce an iterative refinement pipeline that employs the Unlabeled-Unlabeled learning framework to improve LLM-generated pseudo-labels for classification tasks. By exploiting two unlabeled datasets with differing positive class ratios, our approach iteratively denoises and refines the initial pseudo-labels, thereby mitigating the adverse effects of internal biases with minimal human supervision. Evaluations on diverse datasets, including low-resource language corpora, patent classifications, and protein structure categorizations, demonstrate that our method consistently outperforms both initial LLM’s classification performance and the self-refinement approaches by cutting-edge models (e.g., GPT-4o and DeepSeek-R1). Moreover, we experimentally confirm that our refined classifier facilitates effective post-training alignment for safety in LLMs and demonstrate successful self-refinement in generative tasks as well.\footnote{Our code is available at https://github.com/HikaruAsano/self-iterative-label-refinement.}

Method: Evaluated various language models on custom datasets from multi-hop QA tasks. Kept context length and position of relevant information constant while varying the number of documents to isolate the effect of document count.

Result: Increasing document count in RAG settings significantly degrades performance for most LLMs (up to 20% reduction). However, Qwen2.5 maintained consistent results across increasing document counts, showing better multi-document handling capability.

Conclusion: Processing multiple documents is a separate challenge from handling long contexts. Qwen2.5 demonstrates superior multi-document handling capabilities compared to other LLMs tested.

Abstract: Retrieval-Augmented Generation (RAG) enhances the accuracy of Large Language Model (LLM) responses by leveraging relevant external documents during generation. Although previous studies noted that retrieving many documents can degrade performance, they did not isolate how the quantity of documents affects performance while controlling for context length. We evaluate various language models on custom datasets derived from a multi-hop QA task. We keep the context length and position of relevant information constant while varying the number of documents, and find that increasing the document count in RAG settings poses significant challenges for most LLMs, reducing performance by up to 20%. However, Qwen2.5 maintained consistent results across increasing document counts, indicating better multi-document handling capability. Finally, our results indicate that processing multiple documents is a separate challenge from handling long contexts. We also make the datasets and code available: https://github.com/shaharl6000/MoreDocsSameLen .

Method: 1) Proposes a high similarity pruning algorithm at the data preprocessing level to reduce spurious correlations in training data. 2) Designs a specific detection method for knowledge-shortcut hallucinations to evaluate mitigation effectiveness.

Result: Experimental results show the approach effectively reduces knowledge-shortcut hallucinations, particularly in fine-tuning tasks, without negatively impacting model performance in question answering.

Conclusion: The work introduces a new paradigm for mitigating specific hallucination issues in generative models, enhancing their robustness and reliability in real-world applications by addressing knowledge-shortcut hallucinations at the data level.

Abstract: The emergence of large language models (LLMs) has significantly advanced the development of natural language processing (NLP), especially in text generation tasks like question answering. However, model hallucinations remain a major challenge in natural language generation (NLG) tasks due to their complex causes. We systematically expand on the causes of factual hallucinations from the perspective of knowledge shortcuts, analyzing hallucinations arising from correct and defect-free data and demonstrating that knowledge-shortcut hallucinations are prevalent in generative models. To mitigate this issue, we propose a high similarity pruning algorithm at the data preprocessing level to reduce spurious correlations in the data. Additionally, we design a specific detection method for knowledge-shortcut hallucinations to evaluate the effectiveness of our mitigation strategy. Experimental results show that our approach effectively reduces knowledge-shortcut hallucinations, particularly in fine-tuning tasks, without negatively impacting model performance in question answering. This work introduces a new paradigm for mitigating specific hallucination issues in generative models, enhancing their robustness and reliability in real-world applications.

Method: Used comprehensive set of classic working memory tasks to estimate working memory capacity of large language models, comparing to normative human scores.

Result: LLMs exceed human working memory capacity in most cases, but increased capacity doesn’t correlate with better performance on executive functioning tasks or problem solving benchmarks.

Conclusion: LLMs may have deficits in attentional control and cognitive flexibility, with current reasoning models showing mixed results in compensating for these limitations.

Abstract: Working memory, or the ability to hold and manipulate information in the mind, is a critical component of human intelligence and executive functioning. It is correlated with performance on various cognitive tasks, including measures of fluid intelligence, which encompasses reasoning and problem solving. We use a comprehensive set of classic working memory tasks to estimate the working memory capacity of large language models (LLMs). We find that in most cases, LLMs exceed normative human scores. However, we do not find that the increased capacity of working memory is associated with higher performance on other executive functioning tasks or problem solving benchmarks. These results suggest that LLMs may have deficits in attentional control and cognitive flexibility, which result in difficulties with inhibiting automatic responses and adapting to shifting information. Our findings suggest that current reasoning models have mixed results in compensating for these deficits.

Method: DeltaEdit uses dynamic orthogonal constraint strategies to reduce conflicts between knowledge during sequential editing, addressing the noise accumulation problem by minimizing interference between different knowledge updates.

Result: DeltaEdit achieves a 16.8% improvement in editing performance over the strongest baseline by significantly reducing superimposed noise accumulation in sequential knowledge editing.

Conclusion: The proposed DeltaEdit method effectively mitigates the superimposed noise accumulation problem in sequential knowledge editing, maintaining higher editing success rates over long-term editing compared to existing approaches.

Abstract: Sequential knowledge editing techniques aim to continuously update knowledge in large language models at low cost, preventing models from generating outdated or incorrect information. However, existing sequential editing methods suffer from a significant decline in editing success rates after long-term editing. Through theoretical analysis and experiments, our findings reveal that as the number of edits increases, the model’s output increasingly deviates from the desired target, leading to a drop in editing success rates. We refer to this issue as the superimposed noise accumulation problem. Our further analysis demonstrates that the problem is related to the erroneous activation of irrelevant knowledge and conflicts between activated knowledge. Based on this analysis, a method named DeltaEdit is proposed that reduces conflicts between knowledge through dynamic orthogonal constraint strategies. Experiments show that DeltaEdit significantly reduces superimposed noise, achieving a 16.8% improvement in editing performance over the strongest baseline.

Method: Proposed CoT Thought Leap Bridge Task to automatically detect leaps and generate missing intermediate reasoning steps. Created ScaleQM+ training dataset based on ScaleQuestMath, and trained CoT-Bridge model to bridge thought leaps.

Result: Models fine-tuned on bridged datasets consistently outperform those trained on original datasets, with improvements up to +5.87% on NuminaMath. Also enhances distilled data (+3.02%) and provides better starting points for reinforcement learning (+3.1%). Shows improved generalization to out-of-domain logical reasoning tasks.

Conclusion: Enhancing reasoning completeness through bridging thought leaps yields broadly applicable benefits, and CoT-Bridge functions as a plug-and-play module compatible with existing optimization techniques.

Abstract: Large language models (LLMs) have achieved remarkable progress on mathematical tasks through Chain-of-Thought (CoT) reasoning. However, existing mathematical CoT datasets often suffer from Thought Leaps due to experts omitting intermediate steps, which negatively impacts model learning and generalization. We propose the CoT Thought Leap Bridge Task, which aims to automatically detect leaps and generate missing intermediate reasoning steps to restore the completeness and coherence of CoT. To facilitate this, we constructed a specialized training dataset called ScaleQM+, based on the structured ScaleQuestMath dataset, and trained CoT-Bridge to bridge thought leaps. Through comprehensive experiments on mathematical reasoning benchmarks, we demonstrate that models fine-tuned on bridged datasets consistently outperform those trained on original datasets, with improvements of up to +5.87% on NuminaMath. Our approach effectively enhances distilled data (+3.02%) and provides better starting points for reinforcement learning (+3.1%), functioning as a plug-and-play module compatible with existing optimization techniques. Furthermore, CoT-Bridge demonstrate improved generalization to out-of-domain logical reasoning tasks, confirming that enhancing reasoning completeness yields broadly applicable benefits.

Method: Compared hierarchical embeddings from 14 publicly available LLMs with fMRI data from participants exposed to naturalistic narrative stories, constructing sentence-level neural prediction models to identify model layers most correlated with brain region activations.

Result: Improvements in model performance drive evolution of representational architectures toward brain-like hierarchies, achieving stronger functional and anatomical correspondence at higher semantic abstraction levels.

Conclusion: LLMs develop increasingly brain-like hierarchical representations as they scale and improve, suggesting convergence between artificial and biological language processing systems.

Abstract: Understanding whether large language models (LLMs) and the human brain converge on similar computational principles remains a fundamental and important question in cognitive neuroscience and AI. Do the brain-like patterns observed in LLMs emerge simply from scaling, or do they reflect deeper alignment with the architecture of human language processing? This study focuses on the sentence-level neural mechanisms of language models, systematically investigating how hierarchical representations in LLMs align with the dynamic neural responses during human sentence comprehension. By comparing hierarchical embeddings from 14 publicly available LLMs with fMRI data collected from participants, who were exposed to a naturalistic narrative story, we constructed sentence-level neural prediction models to precisely identify the model layers most significantly correlated with brain region activations. Results show that improvements in model performance drive the evolution of representational architectures toward brain-like hierarchies, particularly achieving stronger functional and anatomical correspondence at higher semantic abstraction levels.

Method: 1) Analyzed digit frequencies in OLMo2 pretraining corpus for Benford’s Law patterns; 2) Created evaluation benchmark with uniformly distributed ground-truth digits across 7 numerical reasoning tasks; 3) Used logit-lens tracing and neuron-level dissection to identify bias sources; 4) Pruned selective FFN neurons to test causal effects.

Result: Leading open-source LLMs show consistent digit bias resembling Benford’s Law; bias originates from small subset of highly digit-selective FFN neurons in deeper layers; pruning these neurons mitigates imbalanced overgeneration and partially corrects erroneous outputs.

Conclusion: Corpus-level digit statistics (Benford’s Law) propagate into LLM behavior through specific neurons, causing numerical hallucinations; this reveals fundamental connection between pretraining data patterns and symbolic failures, offering new diagnostic/mitigation approaches for numerical tasks.

Abstract: Large Language Models (LLMs) exhibit impressive performance on complex reasoning tasks, yet they frequently fail on basic numerical problems, producing incorrect outputs. Inspired by Benford’s Law, a statistical pattern in which lower digits occur more frequently as leading digits, we hypothesize that the skewed digit distributions in web-collected corpora may be learned by LLMs during pretraining, leading to biased numerical generation. To investigate the hypothesis, we first examine whether digits frequencies in pretraining corpus (OLMo2) follows Benford’s law. We then construct an evaluation benchmark in which the ground-truth digits are uniformly distributed within each of the seven numerical reasoning tasks. Our evaluation results demonstrate that leading open-source LLMs show a consistent pattern of digit bias that resembles Benford’s law. Through logit-lens tracing and neuron-level dissection, we identify that this bias arises predominantly from a small subset of highly digit-selective feed-forward network (FFN) neurons in the deeper layers. Finally, we demonstrate that pruning these neurons mitigates imbalanced overgeneration and partially corrects erroneous outputs, providing causal evidence that fine-grained pretraining digit bias can propagate into model behavior. Our findings reveal a fundamental connection between corpus-level statistics and symbolic failure modes in LLMs, offering a new lens for diagnosing and mitigating hallucinations in numerical tasks.

Method: Created FlowerTune LLM Leaderboard - a benchmarking suite with federated instruction-tuning datasets and domain-specific evaluation metrics across four domains. Used collaborative, open-source, community-driven approach to evaluate 26 pre-trained LLMs with different aggregation and fine-tuning strategies.

Result: First comprehensive comparison of LLMs in federated settings, providing actionable insights into model performance, resource constraints, and domain adaptation across different domains and FL strategies.

Conclusion: This work establishes foundation for developing privacy-preserving, domain-specialized LLMs for real-world applications through federated fine-tuning, addressing data scarcity and privacy concerns.

Abstract: Large Language Models (LLMs) have achieved state-of-the-art results across diverse domains, yet their development remains reliant on vast amounts of publicly available data, raising concerns about data scarcity and the lack of access to domain-specific, sensitive information. Federated Learning (FL) presents a compelling framework to address these challenges by enabling decentralized fine-tuning on pre-trained LLMs without sharing raw data. However, the compatibility and performance of pre-trained LLMs in FL settings remain largely under explored. We introduce the FlowerTune LLM Leaderboard, a first-of-its-kind benchmarking suite designed to evaluate federated fine-tuning of LLMs across four diverse domains: general NLP, finance, medical, and coding. Each domain includes federated instruction-tuning datasets and domain-specific evaluation metrics. Our results, obtained through a collaborative, open-source and community-driven approach, provide the first comprehensive comparison across 26 pre-trained LLMs with different aggregation and fine-tuning strategies under federated settings, offering actionable insights into model performance, resource constraints, and domain adaptation. This work lays the foundation for developing privacy-preserving, domain-specialized LLMs for real-world applications.

Method: ROVER recursively decomposes long-horizon video trajectories into segments corresponding to shorter subtasks, enabling focused reasoning over temporally localized frame sequences while maintaining global context through an in-context learning approach with subtask-specific sliding context windows.

Result: ROVER outperforms strong baselines on three video reasoning tasks using OpenX Embodiment videos and a new RoboCasa dataset (543 videos across 27 robotic manipulation tasks). It reduces hallucinations during unexpected/non-optimal moments and achieves linear time complexity scaling with video length.

Conclusion: ROVER effectively addresses the limitation of VLMs in reasoning over long video sequences, making them more suitable for embodied AI applications by improving accuracy, reducing hallucinations, and achieving better computational efficiency through recursive decomposition.

Abstract: Vision-language models (VLMs) have exhibited impressive capabilities across diverse image understanding tasks, but still struggle in settings that require reasoning over extended sequences of camera frames from a video. This limits their utility in embodied settings, which require reasoning over long frame sequences from a continuous stream of visual input at each moment of a task attempt. To address this limitation, we propose ROVER (Reasoning Over VidEo Recursively), a framework that enables the model to recursively decompose long-horizon video trajectories into segments corresponding to shorter subtasks within the trajectory. In doing so, ROVER facilitates more focused and accurate reasoning over temporally localized frame sequences without losing global context. We evaluate ROVER, implemented using an in-context learning approach, on diverse OpenX Embodiment videos and on a new dataset derived from RoboCasa that consists of 543 videos showing both expert and perturbed non-expert trajectories across 27 robotic manipulation tasks. ROVER outperforms strong baselines across three video reasoning tasks: task progress estimation, frame-level natural language reasoning, and video question answering. We observe that, by reducing the number of frames the model reasons over at each timestep, ROVER mitigates hallucinations, especially during unexpected or non-optimal moments of a trajectory. In addition, by enabling the implementation of a subtask-specific sliding context window, ROVER’s time complexity scales linearly with video length, an asymptotic improvement over baselines. Demos, code, and data available at: https://rover-vlm.github.io

Method: Proposes Causal-Oriented Policy Optimization (COPO) with token-level sufficiency and necessity constraints to measure each inference token’s causal contribution. Uses a causal completeness reward to evaluate token contributions, then constructs a causally informed advantage function within the GRPO optimization framework to encourage focus on causally sufficient and necessary tokens.

Result: Experimental results across various benchmarks demonstrate the advantages of COPO in mitigating hallucinations and improving model performance.

Conclusion: COPO effectively addresses MLLM hallucinations by reducing spurious correlations through causal analysis of token contributions, leading to more accurate and evidence-grounded outputs.

Abstract: Despite Multimodal Large Language Models (MLLMs) having shown impressive capabilities, they may suffer from hallucinations. Empirically, we find that MLLMs attend disproportionately to task-irrelevant background regions compared with text-only LLMs, implying spurious background-answer correlations. We claim and analyze that (i) outcome-based rewards can be an important factor leading to spurious correlations, and (ii) spurious correlations can be an important factor leading to hallucinations. Based on these results, we propose Causal-Oriented Policy Optimization (COPO) to mitigate these spurious correlations, thus addressing the issue of hallucinations. It imposes token-level sufficiency and necessity constraints to measure each inference token’s causal contribution, thus ensuring correct and evidence-grounded output. Specifically, we first evaluate each token’s causal contribution via a newly proposed causal completeness reward. This reward is then used to construct a causally informed advantage function within the GRPO optimization framework, encouraging the model to focus on tokens that are causally sufficient and necessary for accurate generation. Experimental results across various benchmarks demonstrate the advantages of COPO.

Method: IROTE generates and optimizes textual self-reflections within prompts based on psychological theories that traits form through identity-related reflection. It iteratively maximizes an information-theoretic objective to enhance trait-behavior connections while reducing noisy redundancy, without fine-tuning.

Result: Extensive experiments across three human trait systems show that a single IROTE-generated self-reflection can induce stable trait impersonation across diverse downstream tasks, consistently outperforming existing baselines beyond simple questionnaire answering.

Conclusion: IROTE addresses the superficial elicitation problem by creating evocative and compact trait reflections that enable LLMs to exhibit stable, transferable human-like traits across various applications, advancing personalized LLMs and social simulations.

Abstract: Trained on various human-authored corpora, Large Language Models (LLMs) have demonstrated a certain capability of reflecting specific human-like traits (e.g., personality or values) by prompting, benefiting applications like personalized LLMs and social simulations. However, existing methods suffer from the superficial elicitation problem: LLMs can only be steered to mimic shallow and unstable stylistic patterns, failing to embody the desired traits precisely and consistently across diverse tasks like humans. To address this challenge, we propose IROTE, a novel in-context method for stable and transferable trait elicitation. Drawing on psychological theories suggesting that traits are formed through identity-related reflection, our method automatically generates and optimizes a textual self-reflection within prompts, which comprises self-perceived experience, to stimulate LLMs’ trait-driven behavior. The optimization is performed by iteratively maximizing an information-theoretic objective that enhances the connections between LLMs’ behavior and the target trait, while reducing noisy redundancy in reflection without any fine-tuning, leading to evocative and compact trait reflection. Extensive experiments across three human trait systems manifest that one single IROTE-generated self-reflection can induce LLMs’ stable impersonation of the target trait across diverse downstream tasks beyond simple questionnaire answering, consistently outperforming existing strong baselines.

Method: Controlled study scaling vocabulary from 24K to 196K while holding data, computation, and optimization constant. Used Kolmogorov complexity to quantify tokenized text complexity and performed word-level loss decomposition.

Result: Larger vocabularies reduce cross-entropy loss primarily by lowering uncertainty on the 2,500 most frequent words (covering ~75% of downstream tokens), while loss on rare words actually increases. Same benefit can be achieved by enlarging model parameters with fixed vocabulary.

Conclusion: The benefit of larger vocabularies comes from reducing complexity of tokenized text, not from better handling of rare words. This provides a principled framework for tokenizer-model co-design and clarifies loss dynamics in language model scaling.

Abstract: Large language models are trained with tokenizers, and the resulting token distribution is highly imbalanced: a few words dominate the stream while most occur rarely. Recent practice favors ever-larger vocabularies, but it is unclear where the benefit comes from. To this end, we perform a controlled study that scales the vocabulary of the language model from 24K to 196K while holding data, computation, and optimization unchanged. We begin by quantifying the complexity of tokenized text – formalized via Kolmogorov complexity – and show that larger vocabularies reduce this complexity. Above 24K, every common word is already tokenized as a single token, so enlarging vocabulary only deepens the relative token-frequency imbalance. Word-level loss decomposition shows that larger vocabularies reduce cross-entropy loss almost exclusively by lowering uncertainty on the 2,500 most frequent words, even though loss on the rare tail rises. The same frequent words cover roughly 75% of tokens in downstream benchmarks, so this training advantage transfers intact. We further show that enlarging model parameters with a fixed vocabulary yields the same frequent-word benefit. Our results recast “bigger vocabularies help” as “lowering complexity of tokenized text helps,” offering a simple, principled knob for tokenizer-model co-design and clarifying the loss dynamics that govern language model scaling in pre-training.

Method: 1) Synthesize QA pairs from single and multi-documents to enhance cross-document reasoning. 2) Mask plaintext documents with special tokens and translate them to LoRA parameters via a parameter generator, maintaining standard RAG document structure. 3) Train the parameter generator using synthetic QA data to match standard RAG’s hidden states and output logits, enabling RAG-style reasoning without original documents.

Result: Experiments on four QA datasets show that DistilledPRAG outperforms baselines in accuracy and generalizes well on out-of-distribution data.

Conclusion: DistilledPRAG successfully addresses the critical challenge of achieving high-efficiency parameterization while maintaining RAG-level performance for privacy-preserving reasoning, offering a practical solution that balances privacy, efficiency, and generalization.

Abstract: The current RAG system requires uploading plaintext documents to the cloud, risking private data leakage. Parametric RAG (PRAG) encodes documents as LoRA parameters within LLMs, offering a possible way to reduce exposure of raw content. However, it still faces two issues: (1) PRAG demands synthesizing QA pairs and fine-tuning LLM for each individual document to create its corresponding LoRA, leading to unacceptable inference latency. (2) The performance of PRAG relies solely on synthetic QA data while lacking internal alignment with standard RAG, resulting in poor generalization on out-of-distribution(OOD) inputs. Therefore, achieving high-efficiency parameterization while maintaining RAG-level performance remains a critical challenge for privacy-preserving reasoning. In this paper, we propose DistilledPRAG, a generalizable knowledge-distilled parametric RAG model aligned with standard RAG in document structure and parameter activation. We first synthesize QA pairs from single and multi-documents to enhance cross-document reasoning. Then, we mask the plaintext documents with a special token and translate them to LoRA via a parameter generator, maintaining the standard RAG document structure. Finally, guided by synthetic QA data, we train the parameter generator to match standard RAG’s hidden states and output logits, enabling RAG-style reasoning without original documents. Experiments on four QA datasets show that DistilledPRAG outperforms baselines in accuracy and generalizes well on OOD data.

Method: Uses Form 10-K filings as data source and applies NLP techniques with unsupervised fine-tuning based on chronological and lexical patterns to capture implicit risk connections, developing a domain-specific financial encoder with quantitative risk relation scores.

Result: Extensive experiments show the method outperforms strong baselines across multiple evaluation settings.

Conclusion: The proposed systematic approach provides a scalable, transparent, and interpretable alternative to traditional manual methods for identifying inter-firm risk relations.

Abstract: A multitude of interconnected risk events – ranging from regulatory changes to geopolitical tensions – can trigger ripple effects across firms. Identifying inter-firm risk relations is thus crucial for applications like portfolio management and investment strategy. Traditionally, such assessments rely on expert judgment and manual analysis, which are, however, subjective, labor-intensive, and difficult to scale. To address this, we propose a systematic method for extracting inter-firm risk relations using Form 10-K filings – authoritative, standardized financial documents – as our data source. Leveraging recent advances in natural language processing, our approach captures implicit and abstract risk connections through unsupervised fine-tuning based on chronological and lexical patterns in the filings. This enables the development of a domain-specific financial encoder with a deeper contextual understanding and introduces a quantitative risk relation score for transparency, interpretable analysis. Extensive experiments demonstrate that our method outperforms strong baselines across multiple evaluation settings. Our codes are available at https://github.com/cnclabs/codes.fin.relation.

Method: Agentar-Scale-SQL implements an Orchestrated Test-Time Scaling strategy combining three perspectives: 1) Internal Scaling via RL-enhanced Intrinsic Reasoning, 2) Sequential Scaling through Iterative Refinement, and 3) Parallel Scaling using Diverse Synthesis and Tournament Selection.

Result: Achieves SOTA performance on BIRD benchmark with 81.67% execution accuracy on test set, ranking first on official leaderboard. Demonstrates effective path toward human-level performance.

Conclusion: Agentar-Scale-SQL provides a general-purpose framework for improving Text-to-SQL performance through orchestrated test-time scaling, easily adaptable to new databases and more powerful language models.

Abstract: State-of-the-art (SOTA) Text-to-SQL methods still lag significantly behind human experts on challenging benchmarks like BIRD. Current approaches that explore test-time scaling lack an orchestrated strategy and neglect the model’s internal reasoning process. To bridge this gap, we introduce Agentar-Scale-SQL, a novel framework leveraging scalable computation to improve performance. Agentar-Scale-SQL implements an Orchestrated Test-Time Scaling strategy that synergistically combines three distinct perspectives: i) Internal Scaling via RL-enhanced Intrinsic Reasoning, ii) Sequential Scaling through Iterative Refinement, and iii) Parallel Scaling using Diverse Synthesis and Tournament Selection. Agentar-Scale-SQL is a general-purpose framework designed for easy adaptation to new databases and more powerful language models. Extensive experiments show that Agentar-Scale-SQL achieves SOTA performance on the BIRD benchmark, reaching 81.67% execution accuracy on the test set and ranking first on the official leaderboard, demonstrating an effective path toward human-level performance.

Method: Created a Kurdish STS dataset with 10,000 sentence pairs covering formal and informal registers, each annotated for similarity. Benchmarked Sentence-BERT, multilingual BERT, and other strong baselines on this dataset.

Result: Obtained competitive results with the benchmarked models while identifying challenges specific to Kurdish, including morphological complexity, orthographic variation, and code-mixing.

Conclusion: The dataset and baseline results establish a reproducible evaluation framework and provide a foundation for future research on Kurdish semantics and low-resource NLP.

Abstract: Semantic Textual Similarity (STS) measures the degree of meaning overlap between two texts and underpins many NLP tasks. While extensive resources exist for high-resource languages, low-resource languages such as Kurdish remain underserved. We present, to our knowledge, the first Kurdish STS dataset: 10,000 sentence pairs spanning formal and informal registers, each annotated for similarity. We benchmark Sentence-BERT, multilingual BERT, and other strong baselines, obtaining competitive results while highlighting challenges arising from Kurdish morphology, orthographic variation, and code-mixing. The dataset and baselines establish a reproducible evaluation suite and provide a strong starting point for future research on Kurdish semantics and low-resource NLP.

Method: ORBIT integrates synthetic dialogue generation with dynamically constructed rubrics that serve as adaptive guides for incremental RL. Uses rubric-driven feedback instead of external knowledge bases or handcrafted rules. Judge component uses general-purpose instruction-following LLMs without task-specific fine-tuning.

Result: Applied to Qwen3-4B-Instruct, raises HealthBench-Hard score from 7.0 to 27.5 with only 2k training samples (SOTA for this scale). With larger rubric datasets, competes with strongest open-source baselines. Also improves instruction-following on InfoBench, showing generality.

Conclusion: Rubric-guided RL consistently improves consultation quality across diverse medical scenarios. ORBIT provides effective framework for high-stakes domains where traditional RL fails due to ambiguous feedback and reward hacking risks.

Abstract: Reinforcement learning has powered many of the recent breakthroughs in large language models, especially for tasks where rewards can be computed automatically, such as code generation. However, these methods deteriorate in open-ended domains like medical consultation, where feedback is inherently ambiguous, highly context-dependent, and cannot be reduced to a reliable scalar signal. In such settings, RL must either rely on supervision-intensive reward models that often fail to generalize, or it falls into pathological behaviors such as reward hacking - an especially troubling risk for high-stakes medical dialogue. To address these limitations, we introduce ORBIT, an open-ended rubric-based incremental training framework for high-stakes medical dialogue. ORBIT integrates synthetic dialogue generation with dynamically constructed rubrics that serve as adaptive guides for incremental RL. Instead of relying on external medical knowledge bases or handcrafted rule sets, ORBIT uses rubric-driven feedback to steer the learning process. Its judge component can be instantiated with general-purpose instruction-following LLMs, removing the need for any task-specific fine-tuning. Applied to the Qwen3-4B-Instruct model, ORBIT raises the HealthBench-Hard score from 7.0 to 27.5 using only 2k training samples, achieving SOTA performance for models at this scale. With larger rubric datasets, ORBIT-trained models further compete with the strongest open-source baselines on HealthBench-Hard. Our analysis shows that rubric-guided RL consistently improves consultation quality across diverse medical scenarios. We also apply such rubric generation and training pipeline to InfoBench, where ORBIT enhances instruction-following performance, highlighting the generality of rubric-based feedback.

Method: Introduces a structured, automated framework inspired by the knapsack problem, where a composer agent systematically identifies, selects, and assembles optimal agentic components by jointly considering performance, budget constraints, and compatibility through dynamic testing and real-time utility modeling.

Result: Empirical evaluation with Claude 3.5 Sonnet across five datasets shows the online-knapsack composer consistently achieves Pareto optimal performance, with up to 31.6% success rate improvement in single-agent setups and increasing multi-agent success rates from 37% to 87% with 100+ agent inventory.

Conclusion: The online knapsack framework enables scalable, efficient agentic system composition that dynamically optimizes component selection based on real-time utility and budget constraints, significantly outperforming traditional retrieval-based approaches across diverse domains.

Abstract: Designing effective agentic systems requires the seamless composition and integration of agents, tools, and models within dynamic and uncertain environments. Most existing methods rely on static, semantic retrieval approaches for tool or agent discovery. However, effective reuse and composition of existing components remain challenging due to incomplete capability descriptions and the limitations of retrieval methods. Component selection suffers because the decisions are not based on capability, cost, and real-time utility. To address these challenges, we introduce a structured, automated framework for agentic system composition that is inspired by the knapsack problem. Our framework enables a composer agent to systematically identify, select, and assemble an optimal set of agentic components by jointly considering performance, budget constraints, and compatibility. By dynamically testing candidate components and modeling their utility in real-time, our approach streamlines the assembly of agentic systems and facilitates scalable reuse of resources. Empirical evaluation with Claude 3.5 Sonnet across five benchmarking datasets shows that our online-knapsack-based composer consistently lies on the Pareto frontier, achieving higher success rates at significantly lower component costs compared to our baselines. In the single-agent setup, the online knapsack composer shows a success rate improvement of up to 31.6% in comparison to the retrieval baselines. In multi-agent systems, the online knapsack composer increases success rate from 37% to 87% when agents are selected from an agent inventory of 100+ agents. The substantial performance gap confirms the robust adaptability of our method across diverse domains and budget constraints.

Method: Created ChiKhaPo benchmark with 8 subtasks of varying difficulty for lexical comprehension and generation. Built using existing lexicons, monolingual data, and bitext, providing coverage for 2700+ languages for 2 subtasks.

Result: The benchmark surpasses existing benchmarks in language coverage. 6 state-of-the-art models struggle on the benchmark. Performance is influenced by language family, language resourcedness, task type, and comprehension vs generation directions.

Conclusion: ChiKhaPo enables and encourages massively multilingual benchmarking of LLMs, addressing the critical gap in evaluating basic linguistic competence across diverse languages, particularly low-resource ones.

Abstract: Existing benchmarks for large language models (LLMs) are largely restricted to high- or mid-resource languages, and often evaluate performance on higher-order tasks in reasoning and generation. However, plenty of evidence points to the fact that LLMs lack basic linguistic competence in the vast majority of the world’s 3800+ written languages. We introduce ChiKhaPo, consisting of 8 subtasks of varying difficulty designed to evaluate the lexical comprehension and generation abilities of generative models. ChiKhaPo draws on existing lexicons, monolingual data, and bitext, and provides coverage for 2700+ languages for 2 subtasks, surpassing any existing benchmark in terms of language coverage. We further show that 6 SOTA models struggle on our benchmark, and discuss the factors contributing to performance scores, including language family, language resourcedness, task, and comprehension versus generation directions. With ChiKhaPo, we hope to enable and encourage the massively multilingual benchmarking of LLMs.

Method: End-to-end reinforcement learning framework with small LLM generating prompts for large LLM in multi-turn interactions, using dual-constrained reward for correctness, quality, and reasoning accuracy.

Result: Significantly outperforms baseline models across multiple public datasets, providing plug-and-play framework supporting both inference and training.

Conclusion: Prompt-R1 effectively addresses prompt engineering challenges through collaborative LLM interaction, offering practical solution for complex problem-solving with LLMs.

Abstract: Recently, advanced large language models (LLMs) have emerged at an increasingly rapid pace. However, when faced with complex problems, most users are often unable to provide accurate and effective prompts to interact with LLMs, thus limiting the performance of LLMs. To address this challenge, we propose Prompt-R1, an end-to-end reinforcement learning framework that uses a small-scale LLM to collaborate with large-scale LLMs, replacing user interaction to solve problems better. This collaboration is cast as a multi-turn prompt interaction, where the small-scale LLM thinks and generates prompts, and the large-scale LLM performs complex reasoning. A dual-constrained reward is designed to optimize for correctness, generation quality, and reasoning accuracy. Prompt-R1 provides a plug-and-play framework that supports both inference and training with various large-scale LLMs. Experiments on multiple public datasets show that Prompt-R1 significantly outperforms baseline models across tasks. Our code is publicly available at https://github.com/QwenQKing/Prompt-R1.

Method: Train a hybrid model with both reasoning and non-reasoning modes, then further train on CoT-annotated data to internalize structured reasoning. At inference, use the no-reasoning mode to enforce concise, thought-free outputs while maintaining rich internal reasoning.

Result: 3TF-trained models achieve large improvements on reasoning benchmarks under thought-free inference, demonstrating that high-quality reasoning can be learned and executed implicitly without explicit step-by-step generation.

Conclusion: The 3TF framework enables LLMs to perform rich internal reasoning while keeping external outputs short, offering an efficient alternative to compression-based approaches that require extensive short CoT data.

Abstract: Recent advances in large language models (LLMs) have leveraged explicit Chain-of-Thought (CoT) prompting to improve reasoning accuracy. However, most existing methods primarily focus on compressing verbose reasoning outputs. These Long-to-Short transformations aim to improve efficiency, but require a large amount of short CoT data. In this work, we introduce \textbf{3TF} (\textbf{T}hought-\textbf{T}raining and \textbf{T}hought-\textbf{F}ree inference), a framework for efficient reasoning that takes a Short-to-Long perspective. We first train a hybrid model that can operate in both reasoning and non-reasoning modes, and then further train it on CoT-annotated data to internalize structured reasoning, while enforcing concise, thought-free outputs at inference time using the no-reasoning mode. Unlike compression-based approaches, 3TF improves the reasoning quality of non-reasoning outputs, enabling models to perform rich internal reasoning implicitly while keeping external outputs short. Empirically, 3TF-trained models obtain large improvements on reasoning benchmarks under thought-free inference, demonstrating that high quality reasoning can be learned and executed implicitly without explicit step-by-step generation.

Method: 1) Formulate honest deductive reasoning as multi-step tasks where models must derive correct conclusions or abstain. 2) Curate two datasets from graph structures (linear algebra and logical inference) with unanswerable cases created by randomly perturbing edges. 3) Propose ACNCHOR, a reinforcement learning method that injects ground truth trajectories into rollouts to prevent early training collapse when negative rewards dominate.

Result: Prompting and existing training methods (including GRPO with/without supervised fine-tuning) struggle on honest deductive reasoning tasks. GRPO is vulnerable to collapse when negative rewards dominate early training. ACNCHOR stabilizes learning and significantly improves overall reasoning performance compared to baseline methods.

Conclusion: Training dynamics are crucial for enabling honest deductive reasoning in language models. The proposed ACNCHOR method effectively addresses early training collapse and improves model performance on tasks requiring proper abstention when conclusions cannot be logically derived from premises.

Abstract: Deductive reasoning is the process of deriving conclusions strictly from the given premises, without relying on external knowledge. We define honesty in this setting as a model’s ability to respond only when the conclusion is logically entailed by the premises, and to abstain otherwise. However, current language models often fail to reason honestly, producing unwarranted answers when the input is insufficient. To study this challenge, we formulate honest deductive reasoning as multi-step tasks where models must either derive the correct conclusion or abstain. We curate two datasets from graph structures, one for linear algebra and one for logical inference, and introduce unanswerable cases by randomly perturbing an edge in half of the instances. We find that prompting and existing training methods, including GRPO with or without supervised fine-tuning initialization, struggle on these tasks. In particular, GRPO optimize only for final task outcomes, leaving models vulnerable to collapse when negative rewards dominate early training. To address this, we propose ACNCHOR, a reinforcement learning method that injects ground truth trajectories into rollouts, preventing early training collapse. Our results demonstrate that this method stabilizes learning and significantly improves the overall reasoning performance, underscoring the importance of training dynamics for enabling honest deductive reasoning in language models.

Method: Combines semantic understanding with keyword precision, operating entirely on local infrastructure without internet access. Uses two complementary retrieval strategies balanced together, with consumer-grade hardware acceleration.

Result: Achieves competitive accuracy on complex queries across legal, scientific, and conversational documents while keeping all data on local machines. Delivers reliable answers with minimal errors using local processing.

Conclusion: Privacy and performance need not be mutually exclusive in enterprise AI deployment. Organizations like banks, hospitals, and law firms can adopt conversational document AI without transmitting proprietary information to external providers.

Abstract: Organizations handling sensitive documents face a critical dilemma: adopt cloud-based AI systems that offer powerful question-answering capabilities but compromise data privacy, or maintain local processing that ensures security but delivers poor accuracy. We present a question-answering system that resolves this trade-off by combining semantic understanding with keyword precision, operating entirely on local infrastructure without internet access. Our approach demonstrates that organizations can achieve competitive accuracy on complex queries across legal, scientific, and conversational documents while keeping all data on their machines. By balancing two complementary retrieval strategies and using consumer-grade hardware acceleration, the system delivers reliable answers with minimal errors, letting banks, hospitals, and law firms adopt conversational document AI without transmitting proprietary information to external providers. This work establishes that privacy and performance need not be mutually exclusive in enterprise AI deployment.

Method: Introduces a continual learning framework where agents receive natural language feedback to refine queries, distill revealed knowledge, and store it in structured memory for future reuse. Multiple agent architectures are designed and evaluated, varying in how they capture and retrieve past experiences, with Procedural Agent being a key variant.

Result: Memory-augmented agents, particularly the Procedural Agent, achieve significant accuracy gains and error reduction on the BIRD benchmark Dev set by leveraging human-in-the-loop feedback. The framework demonstrates effective transformation of tacit human expertise into reusable knowledge.

Conclusion: The approach enables more adaptive, domain-aware text-to-SQL systems that continually learn from human feedback, highlighting the importance of converting tacit human expertise into structured, reusable knowledge for improving LLM performance on database-specific tasks.

Abstract: Large Language Models (LLMs) can generate SQL queries from natural language questions but struggle with database-specific schemas and tacit domain knowledge. We introduce a framework for continual learning from human feedback in text-to-SQL, where a learning agent receives natural language feedback to refine queries and distills the revealed knowledge for reuse on future tasks. This distilled knowledge is stored in a structured memory, enabling the agent to improve execution accuracy over time. We design and evaluate multiple variations of a learning agent architecture that vary in how they capture and retrieve past experiences. Experiments on the BIRD benchmark Dev set show that memory-augmented agents, particularly the Procedural Agent, achieve significant accuracy gains and error reduction by leveraging human-in-the-loop feedback. Our results highlight the importance of transforming tacit human expertise into reusable knowledge, paving the way for more adaptive, domain-aware text-to-SQL systems that continually learn from a human-in-the-loop.

Method: The paper conducts a systematic review of MCoT from multiple perspectives: analyzing background and theoretical motivations, introducing mainstream methods through CoT paradigms, post-training stages, and inference stages, examining underlying mechanisms, and summarizing evaluation benchmarks and metrics.

Result: The review comprehensively covers the current state of MCoT research, including its technical evolution, methodological approaches, evaluation frameworks, and application scenarios, providing a structured understanding of the field.

Conclusion: The paper identifies current challenges facing MCoT and provides an outlook on future research directions, highlighting the importance of enhancing complex reasoning capabilities in multimodal models through transparent and interpretable reasoning approaches.

Abstract: With the remarkable success of Multimodal Large Language Models (MLLMs) in perception tasks, enhancing their complex reasoning capabilities has emerged as a critical research focus. Existing models still suffer from challenges such as opaque reasoning paths and insufficient generalization ability. Chain-of-Thought (CoT) reasoning, which has demonstrated significant efficacy in language models by enhancing reasoning transparency and output interpretability, holds promise for improving model reasoning capabilities when extended to the multimodal domain. This paper provides a systematic review centered on “Multimodal Chain-of-Thought” (MCoT). First, it analyzes the background and theoretical motivations for its inception from the perspectives of technical evolution and task demands. Then, it introduces mainstream MCoT methods from three aspects: CoT paradigms, the post-training stage, and the inference stage, while also analyzing their underlying mechanisms. Furthermore, the paper summarizes existing evaluation benchmarks and metrics, and discusses the application scenarios of MCoT. Finally, it analyzes the challenges currently facing MCoT and provides an outlook on its future research directions.

Method: Frames TRM latent reasoning as classifier-free guidance and implicit policy improvement, then proposes a novel training scheme that provides targets for each loop during training.

Result: 18x reduction in forward passes, elimination of halting mechanisms, 24% accuracy on ARC-1 with only 0.8M parameters (outperforming most LLMs), while maintaining comparable quality to standard TRMs.

Conclusion: The proposed training scheme significantly enhances training efficiency for small recursive models while preserving their reasoning capabilities, making them more practical alternatives to large language models for complex reasoning tasks.

Abstract: Recently, it was shown that small, looped architectures, such as Tiny Recursive Models (TRMs), can outperform Large Language Models (LLMs) on complex reasoning tasks, including the Abstraction and Reasoning Corpus (ARC). In this work, we investigate a core question: how can we further improve the efficiency of these methods with minimal changes? To address this, we frame the latent reasoning of TRMs as a form of classifier-free guidance and implicit policy improvement algorithm. Building on these insights, we propose a novel training scheme that provides a target for each loop during training. We demonstrate that our approach significantly enhances training efficiency. Our method reduces the total number of forward passes by 18x and eliminates halting mechanisms, while maintaining quality comparable to standard TRMs. Notably, we achieve 24% accuracy on ARC-1 with only 0.8M parameters, outperforming most LLMs.

Method: Introduces PoETa v2 benchmark with over 40 Portuguese language tasks, evaluates more than 20 LLMs across various training scales and computational resources, and compares performance with equivalent English tasks.

Result: Reveals how computational investment and language-specific adaptation impact Portuguese performance, identifies performance gaps compared to English tasks, and provides baseline evaluations for future research.

Conclusion: PoETa v2 establishes foundational groundwork for Portuguese language modeling research and evaluation, offering an open benchmark to guide future development and address linguistic disparities in LLM performance.

Abstract: Large Language Models (LLMs) exhibit significant variations in performance across linguistic and cultural contexts, underscoring the need for systematic evaluation in diverse languages. In this work, we present the most extensive evaluation of LLMs for the Portuguese language to date. Leveraging our newly introduced PoETa v2 benchmark – a comprehensive suite of over 40 tasks in Portuguese – we assess more than 20 models covering a broad spectrum of training scales and computational resources. Our study reveals how computational investment and language-specific adaptation impact performance in Portuguese, while also analyzing performance gaps in comparison to equivalent tasks in English. Through this benchmark and analysis, PoETa v2 lays the groundwork for future research on Portuguese language modeling and evaluation. The benchmark is available at https://github.com/PoETaV2/PoETaV2.

Method: The authors introduce AppSelectBench with a novel user task generation pipeline that produces realistic, diverse, and semantically grounded user intents at scale. They include unified evaluation protocols covering random, heuristic, zero-shot, few-shot, and retrieval-augmented settings. The benchmark covers 100 widely used desktop applications and includes over 100,000 realistic user tasks.

Result: Extensive experiments across both closed-source and open-source large language models reveal systematic strengths and weaknesses in inter-application reasoning. Even the most capable models still struggle to make consistent application choices, highlighting the challenge of application-level reasoning.

Conclusion: AppSelectBench establishes a foundation for studying and advancing application-level reasoning, an essential yet underexplored capability of intelligent CUAs. The benchmark addresses a critical gap in evaluating how well agents can select appropriate applications for complex tasks.

Abstract: Computer Using Agents (CUAs) are increasingly equipped with external tools, enabling them to perform complex and realistic tasks. For CUAs to operate effectively, application selection, which refers to deciding which application to use before invoking fine-grained tools such as APIs, is a fundamental capability. It determines whether the agent initializes the correct environment, avoids orchestration confusion, and efficiently focuses on relevant context. However, existing benchmarks primarily assess fine-grained API selection, offering limited insight into whether models can reason across and choose between different applications. To fill this gap, we introduce AppSelectBench, a comprehensive benchmark for evaluating application selection in CUAs. AppSelectBench contains a novel user task generation pipeline that produces realistic, diverse, and semantically grounded user intents at scale, together with unified evaluation protocols covering random, heuristic, zero-shot, few-shot, and retrieval-augmented-settings. AppSelectBench covers one hundred widely used desktop applications and includes more than one hundred thousand realistic, diverse, and semantically grounded user tasks. Extensive experiments across both closed-source and open-source large language models reveal systematic strengths and weaknesses in inter-application reasoning, showing that even the most capable models still struggle to make consistent application choices. Together, these results establish AppSelectBench as a foundation for studying and advancing application level reasoning, an essential yet underexplored capability of intelligent CUAs. The source is available at https://microsoft.github.io/appselectbench/.

Method: REFLEX reformulates fact-checking as role-play dialogue and jointly trains verdict prediction with explanation generation. It extracts contrastive activation pairs between backbone models and fine-tuned variants to create steering vectors that disentangle truth into style and substance, suppressing noisy explanations.

Result: Achieves state-of-the-art performance with only 465 self-refined training samples. Models with explanatory objectives can guide those without them, yielding up to 7.57% improvement. Outperforms previous methods that steer toward single truth directions.

Conclusion: REFLEX demonstrates that internal explanation signals play dual roles in both interpreting and enhancing factual reasoning, offering a plug-and-play, self-refining paradigm that addresses latency and reliability issues in real-time fact-checking.

Abstract: The prevalence of misinformation on social media threatens public trust, demanding automated fact-checking systems that provide accurate verdicts with interpretable explanations. However, existing large language model-based (LLM-based) approaches often rely heavily on external knowledge sources, introducing substantial latency and even hallucinations that undermine reliability, interpretability, and responsiveness, which is crucial for real-time use. To address these challenges, we propose REason-guided Fact-checking with Latent EXplanations REFLEX paradigm, a plug-and-play, self-refining paradigm that leverages the internal knowledge in backbone model to improve both verdict accuracy and explanation quality. REFLEX reformulates fact-checking as a role-play dialogue and jointly trains verdict prediction and explanation generation. It adaptively extracts contrastive activation pairs between the backbone model and its fine-tuned variant to construct steering vectors that disentangle truth into style and substance naturally. These activation-level signals guide inference and suppress noisy explanations, enabling more faithful and efficient reasoning. Experiments on real-world datasets show that REFLEX outperforms previous methods that steer toward a single truth direction and underscores the challenge traditional approaches face when handling the subtle, human-unknown truth in fact-checking tasks. Remarkably, with only 465 self-refined training samples, RELFEX achieves state-of-the-art performance. Furthermore, models trained with explanatory objectives can effectively guide those without them, yielding up to a 7.57% improvement, highlighting that internal explanation signals play a dual role in both interpreting and enhancing factual reasoning.

Method: Uses PGD adversarial technique to maximize next-token entropy using a small ensemble of open-source MLLMs, working in both full-image and adversarial CAPTCHA settings.

Result: Single adversarial image can disrupt all models in the ensemble, with perturbations transferring to unseen open-source (Qwen3-VL) and proprietary (GPT-5.1) models.

Conclusion: MLLMs are vulnerable to systematic confusion attacks that transfer across models, highlighting security risks for MLLM-powered applications and the need for robust defenses.

Abstract: We introduce the Adversarial Confusion Attack, a new class of threats against multimodal large language models (MLLMs). Unlike jailbreaks or targeted misclassification, the goal is to induce systematic disruption that makes the model generate incoherent or confidently incorrect outputs. Applications include embedding adversarial images into websites to prevent MLLM-powered agents from operating reliably. The proposed attack maximizes next-token entropy using a small ensemble of open-source MLLMs. In the white-box setting, we show that a single adversarial image can disrupt all models in the ensemble, both in the full-image and adversarial CAPTCHA settings. Despite relying on a basic adversarial technique (PGD), the attack generates perturbations that transfer to both unseen open-source (e.g., Qwen3-VL) and proprietary (e.g., GPT-5.1) models.

Method: Created a reproducible DSPy+HELM framework that introduces structured prompting methods to elicit reasoning. Evaluated four frontier LMs across seven benchmarks (general/medical domains) using four prompting methods, comparing against existing HELM baseline scores.

Result: Without structured prompting: (1) HELM underestimates LM performance by 4% average, (2) performance estimates vary more across benchmarks (+2% standard deviation), (3) leaderboard rankings flip on 3/7 benchmarks, and (4) chain-of-thought reduces LM sensitivity to prompt design. Structured prompting yields more robust benchmarks.

Conclusion: Systematic integration of structured prompting into established evaluation frameworks enables scalable performance-ceiling approximation, producing more accurate and decision-useful benchmarks for LM deployment decisions.

Abstract: As language models (LMs) are increasingly adopted across domains, high-quality benchmarking frameworks that accurately estimate performance are essential for guiding deployment decisions. While frameworks such as Holistic Evaluation of Language Models (HELM) enable broad evaluation across tasks, they often rely on fixed prompts that fail to generalize across LMs, yielding unrepresentative performance estimates. Unless we approximate each LM’s ceiling (maximum achievable via changes to the prompt), we risk underestimating performance. Declarative prompting frameworks, such as DSPy, offer a scalable alternative to manual prompt engineering by crafting structured prompts that can be optimized per task. However, such frameworks have not been systematically evaluated across established benchmarks. We present a reproducible DSPy+HELM framework that introduces structured prompting methods which elicit reasoning, enabling more accurate LM benchmarking. Using four prompting methods, we evaluate four frontier LMs across seven benchmarks (general/medical domain) against existing HELM baseline scores. We find that without structured prompting: (i) HELM underestimates LM performance (by 4% average), (ii) performance estimates vary more across benchmarks ($+$2% standard deviation), (iii) performance gaps are misrepresented (leaderboard rankings flip on 3/7 benchmarks), and (iv) introducing chain-of-thought reduces LM sensitivity to prompt design (smaller $Δ$ across prompts). To our knowledge, this is the first benchmarking study to systematically integrate structured prompting into an established evaluation framework, demonstrating how scalable performance-ceiling approximation yields more robust, decision-useful benchmarks. We open-source (i) DSPy+HELM Integration (https://github.com/stanford-crfm/helm/pull/3893) and (ii) Prompt Optimization Pipeline (https://github.com/StanfordMIMI/dspy-helm).

Method: Used TrackList pipeline for fine-grained linguistic/statistical analysis and introduced RefoMed-EN dataset (6,170 annotated medical terms). Evaluated LLM performance on head vs. tail concepts using syntactic/semantic similarity metrics, statistical correlations, and embeddings.

Result: LLMs perform best on definition-type questions and worst on exemplification tasks. For definitions, LLMs paraphrase more on popular/frequent knowledge and less on technical/tail knowledge, especially in expert texts.

Conclusion: LLMs have significant performance gaps between definition-type and other query types, with frequency-based biases in their responses. This highlights limitations in LLM versatility and potential knowledge representation issues.

Abstract: Large Language Models (LLMs) have proven efficient in giving definition-type answers to user input queries. While for humans giving various types of answers, such as examples and paraphrases, is an easy task, LLMs struggle to provide correct answers for other than definition-type queries. In this study, we evaluated this drop in performance using TrackList, a fine-grained linguistic and statistical analysis pipeline to investigate the impact of the pre-training data on LLMs answers to diverse linguistic queries. We also introduce RefoMed-EN, an English dataset consisting of 6170 human-annotated medical terms alongside their corresponding definitions, denominations, exemplifications, explanations, or paraphrases. We studied whether the high frequency of a concept (head) or low frequency (tail) impacts the language model’s performance. We evaluated the quality of the LLM’s output using syntactic and semantic similarity metrics, statistical correlations and embeddings. Results showed that the LLM’s task performance for definition type questions is the highest, while for the exemplification type it is the lowest. Additionally, we showed that for definition-type questions, large language models are prone to paraphrase more on popular and frequent knowledge and less on tail and technical knowledge, especially in the expert texts.

Method: SGASA framework with two stages: 1) Data Pre-synthesis generates safety guidelines and augmented prompts, 2) Alignment Fine-tuning uses Supervised Fine-tuning (SFT) and Direct Preference Optimization (DPO) to embed guidelines into the model.

Result: Extensive experiments across multiple datasets demonstrate SGASA significantly improves model safety, validating its adaptive and scalable effectiveness against harmful adversarial prompts.

Conclusion: SGASA provides an effective framework for adaptive safety alignment that strengthens reasoning models’ robustness against adversarial jailbreak attacks while maintaining appropriate responses to benign requests.

Abstract: Reasoning models have demonstrated remarkable capabilities in complex reasoning tasks. However, ensuring their safety against adversarial jailbreak prompts remains a critical challenge. Due to the covert and deceptive nature of such prompts, they can often evade built-in safety mechanisms and lead to the generation of harmful content. This underscores the need for an adaptive safety alignment approach that enables models to autonomously reinforce their defenses in response to adversarial inputs. This paper introduces the Synthesized Guideline-based Adaptive Safety Alignment (SGASA) framework, which internalizes model-generated safety guidelines to strengthen models’ ability to enhance robustness against harmful adversarial prompts while minimizing unnecessary refusals of benign requests. SGASA consists of two key stages: Data Pre-synthesis, which generates safety guidelines and augmented prompts; and Alignment Fine-tuning, which leverages Supervised Fine-tuning (SFT) and Direct Preference Optimization (DPO) to embed these guidelines into the model. Extensive experiments across multiple datasets demonstrate that SGASA significantly improves model safety, validating its adaptive and scalable effectiveness.

Method: Odin injects graph structure into Transformers at selected depths through an oriented dual-module mechanism. It integrates multi-hop structures at specific Transformer layers (low-, mid-, high-level) aligned with semantic hierarchy. Aggregation operates on global [CLS] representation, avoiding over-smoothing. Light Odin is a lightweight variant for efficiency.

Result: Odin achieves state-of-the-art accuracy on multiple text-rich graph benchmarks. Light Odin delivers competitive performance with significantly reduced computational cost. The expressive power of Odin strictly contains both pure Transformers and GNNs.

Conclusion: Odin and Light Odin form a unified, hop-free framework for principled structure-text integration that fundamentally avoids over-smoothing and decouples structural abstraction from neighborhood size or graph topology, offering both high performance and computational efficiency.

Abstract: Text-attributed graphs require models to effectively combine strong textual understanding with structurally informed reasoning. Existing approaches either rely on GNNs–limited by over-smoothing and hop-dependent diffusion–or employ Transformers that overlook graph topology and treat nodes as isolated sequences. We propose Odin (Oriented Dual-module INtegration), a new architecture that injects graph structure into Transformers at selected depths through an oriented dual-module mechanism. Unlike message-passing GNNs, Odin does not rely on multi-hop diffusion; instead, multi-hop structures are integrated at specific Transformer layers, yielding low-, mid-, and high-level structural abstraction aligned with the model’s semantic hierarchy. Because aggregation operates on the global [CLS] representation, Odin fundamentally avoids over-smoothing and decouples structural abstraction from neighborhood size or graph topology. We further establish that Odin’s expressive power strictly contains that of both pure Transformers and GNNs. To make the design efficient in large-scale or low-resource settings, we introduce Light Odin, a lightweight variant that preserves the same layer-aligned structural abstraction for faster training and inference. Experiments on multiple text-rich graph benchmarks show that Odin achieves state-of-the-art accuracy, while Light Odin delivers competitive performance with significantly reduced computational cost. Together, Odin and Light Odin form a unified, hop-free framework for principled structure-text integration. The source code of this model has been released at https://github.com/hongkaifeng/Odin.

Method: Created SO-Bench with over 6.5K diverse JSON schemas and 1.8K curated image-schema pairs across four visual domains (UI screens, natural images, documents, charts) with human-verified quality. Conducted benchmarking experiments on open-source and proprietary models.

Result: Benchmarking revealed persistent gaps in models’ ability to predict accurate, schema-compliant outputs, highlighting the need for better multimodal structured reasoning. Training experiments showed significant improvements in structured output capability.

Conclusion: SO-Bench fills a critical gap in evaluating MLLMs’ visual structural output capabilities, revealing current limitations and demonstrating that training can improve structured reasoning. The benchmark will be made available to the community.

Abstract: Multimodal large language models (MLLMs) are increasingly deployed in real-world, agentic settings where outputs must not only be correct, but also conform to predefined data schemas. Despite recent progress in structured generation in textual domain, there is still no benchmark that systematically evaluates schema-grounded information extraction and reasoning over visual inputs. In this work, we conduct a comprehensive study of visual structural output capabilities for MLLMs with our carefully designed SO-Bench benchmark. Covering four visual domains, including UI screens, natural images, documents, and charts, SO-Bench is built from over 6.5K diverse JSON schemas and 1.8K curated image-schema pairs with human-verified quality. Benchmarking experiments on open-sourced and frontier proprietary models reveal persistent gaps in predicting accurate, schema compliant outputs, highlighting the need for better multimodal structured reasoning. Beyond benchmarking, we further conduct training experiments to largely improve the model’s structured output capability. We plan to make the benchmark available to the community.

Method: Saddle-Free Guidance (SFG) leverages the positive curvature of log density estimates in saddle regions to guide individual score-based models. It maintains estimates of maximal positive curvature to provide guidance, working with off-the-shelf diffusion and flow matching models.

Result: SFG achieves state-of-the-art FID and FD-DINOv2 metrics in single-model unconditional ImageNet-512 generation. Combined with Auto-Guidance, it achieves general SOTA in FD-DINOv2 score. With FLUX.1-dev and Stable Diffusion v3.5, SFG boosts output diversity while maintaining excellent prompt adherence and image fidelity.

Conclusion: SFG provides an effective guidance method that doesn’t require labeled data or additional training, making it practical for real-world applications where these resources are limited. It offers computational efficiency comparable to CFG while delivering superior performance.

Abstract: Score-based generative models require guidance in order to generate plausible, on-manifold samples. The most popular guidance method, Classifier-Free Guidance (CFG), is only applicable in settings with labeled data and requires training an additional unconditional score-based model. More recently, Auto-Guidance adopts a smaller, less capable version of the original model to guide generation. While each method effectively promotes the fidelity of generated data, each requires labeled data or the training of additional models, making it challenging to guide score-based models when (labeled) training data are not available or training new models is not feasible. We make the surprising discovery that the positive curvature of log density estimates in saddle regions provides strong guidance for score-based models. Motivated by this, we develop saddle-free guidance (SFG) which maintains estimates of maximal positive curvature of the log density to guide individual score-based models. SFG has the same computational cost of classifier-free guidance, does not require additional training, and works with off-the-shelf diffusion and flow matching models. Our experiments indicate that SFG achieves state-of-the-art FID and FD-DINOv2 metrics in single-model unconditional ImageNet-512 generation. When SFG is combined with Auto-Guidance, its unconditional samples achieve general state-of-the-art in FD-DINOv2 score. Our experiments with FLUX.1-dev and Stable Diffusion v3.5 indicate that SFG boosts the diversity of output images compared to CFG while maintaining excellent prompt adherence and image fidelity.

Method: UniArt uses a diffusion-based framework with unified latent representation encoding geometry, texture, part segmentation, and kinematic parameters. It introduces reversible joint-to-voxel embedding for spatial alignment of articulation features with volumetric geometry, and formulates articulation type prediction as an open-set problem for generalization to novel joint categories.

Result: Experiments on PartNet-Mobility benchmark show UniArt achieves state-of-the-art mesh quality and articulation accuracy, outperforming previous methods.

Conclusion: UniArt provides an effective end-to-end solution for generating articulated 3D objects from single images, with improved generalization capabilities and superior performance compared to multi-stage approaches.

Abstract: Articulated 3D objects play a vital role in realistic simulation and embodied robotics, yet manually constructing such assets remains costly and difficult to scale. In this paper, we present UniArt, a diffusion-based framework that directly synthesizes fully articulated 3D objects from a single image in an end-to-end manner. Unlike prior multi-stage techniques, UniArt establishes a unified latent representation that jointly encodes geometry, texture, part segmentation, and kinematic parameters. We introduce a reversible joint-to-voxel embedding, which spatially aligns articulation features with volumetric geometry, enabling the model to learn coherent motion behaviors alongside structural formation. Furthermore, we formulate articulation type prediction as an open-set problem, removing the need for fixed joint semantics and allowing generalization to novel joint categories and unseen object types. Experiments on the PartNet-Mobility benchmark demonstrate that UniArt achieves state-of-the-art mesh quality and articulation accuracy.

Method: PathReasoning iteratively navigates WSIs through multiple rounds: starting with random candidate regions, it reviews selections with self-reflection, reasons over visual-clinical correspondence, and proposes new regions to explore, building a reasoning chain that directs attention to relevant areas.

Result: Outperforms strong ROI-selection approaches by 6.7% and 3.1% AUROC on subtyping and longitudinal analysis tasks. High-quality ROIs enable accurate breast cancer report generation, significantly outperforming GPT-4o by 10% in accuracy.

Conclusion: PathReasoning efficiently finds informative regions within fixed steps, supports interpretable reasoning chains, enables efficient slide review, consistent diagnostics, comprehensive reporting, and evidence traceability in digital pathology.

Abstract: Deciphering tumor microenvironment from Whole Slide Images (WSIs) is intriguing as it is key to cancer diagnosis, prognosis and treatment response. While these gigapixel images on one hand offer a comprehensive portrait of cancer, on the other hand, the extremely large size, as much as more than 10 billion pixels, make it challenging and time-consuming to navigate to corresponding regions to support diverse clinical inspection. Inspired by pathologists who conducted navigation on WSIs with a combination of sampling, reasoning and self-reflection, we proposed “PathReasoning”, a multi-modal reasoning agent that iteratively navigates across WSIs through multiple rounds of reasoning and refinements. Specifically, starting with randomly sampled candidate regions, PathReasoning reviews current selections with self-reflection, reasoning over the correspondence between visual observations and clinical questions, and concludes by proposing new regions to explore. Across rounds, PathReasoning builds a reasoning chain that gradually directs attention to diagnostically relevant areas. PathReasoning turns each whole slide into a sequence of question-guided views, allowing the model to efficiently find informative ROIs within a fixed number of steps, without the need for dense pixel-level annotations. PathReasoning can substantially outperform strong ROI-selection approaches by 6.7% and 3.1% of AUROC on subtyping and longitudinal analysis tasks. The high-quality ROIs further support accurate report generation on breast cancer, significantly outperforming the standard GPT-4o by 10% in accuracy. PathReasoning prioritizes question-specific regions and constructs interpretable reasoning chains, supporting efficient slide review, consistent diagnostic interpretations, comprehensive reporting, and evidence traceability in digital pathology.

Method: Projection-based Robust Signal Mixing (PRISM) algorithm jointly optimizes photometric detrending and color mixing through online parameter adaptation based on signal quality assessment, without requiring training.

Result: State-of-the-art performance among unsupervised methods: MAE of 0.77 bpm on PURE and 0.66 bpm on UBFC-rPPG, with 97.3% and 97.5% accuracy at 5 bpm threshold. Statistical analysis shows equivalent performance to leading supervised methods (p > 0.2).

Conclusion: Adaptive time series optimization significantly improves rPPG across diverse conditions, enabling training-free, real-time CPU performance comparable to supervised methods.

Abstract: Remote photoplethysmography (rPPG) enables contactless vital sign monitoring using standard RGB cameras. However, existing methods rely on fixed parameters optimized for particular lighting conditions and camera setups, limiting adaptability to diverse deployment environments. This paper introduces the Projection-based Robust Signal Mixing (PRISM) algorithm, a training-free method that jointly optimizes photometric detrending and color mixing through online parameter adaptation based on signal quality assessment. PRISM achieves state-of-the-art performance among unsupervised methods, with MAE of 0.77 bpm on PURE and 0.66 bpm on UBFC-rPPG, and accuracy of 97.3% and 97.5% respectively at a 5 bpm threshold. Statistical analysis confirms PRISM performs equivalently to leading supervised methods ($p > 0.2$), while maintaining real-time CPU performance without training. This validates that adaptive time series optimization significantly improves rPPG across diverse conditions.

Method: Four-component framework: 1) Biological Prototyping with text prompting and prototype weighting; 2) Multiview Alignment with sample- and distribution-wise alignments; 3) Bipartite Fusion for shared and modality-specific information; 4) Semantic Genomics Imputation for missing data.

Result: Extensive experiments show consistent superiority over state-of-the-art methods on multiple downstream tasks.

Conclusion: The proposed flexible multimodal prototyping framework effectively integrates histology and incomplete genomics for precision oncology, addressing real-world clinical challenges with missing data.

Abstract: Multimodal approaches that integrate histology and genomics hold strong potential for precision oncology. However, phenotypic and genotypic heterogeneity limits the quality of intra-modal representations and hinders effective inter-modal integration. Furthermore, most existing methods overlook real-world clinical scenarios where genomics may be partially missing or entirely unavailable. We propose a flexible multimodal prototyping framework to integrate whole slide images and incomplete genomics for precision oncology. Our approach has four key components: 1) Biological Prototyping using text prompting and prototype-wise weighting; 2) Multiview Alignment through sample- and distribution-wise alignments; 3) Bipartite Fusion to capture both shared and modality-specific information for multimodal fusion; and 4) Semantic Genomics Imputation to handle missing data. Extensive experiments demonstrate the consistent superiority of the proposed method compared to other state-of-the-art approaches on multiple downstream tasks. The code is available at https://github.com/helenypzhang/Interpretable-Multimodal-Prototyping.

Method: Integrates 2D amodal completion priors with multi-view stereo geometry conditioning. Uses View-Wise Cross Attention for sparse-view feature fusion and Stereo-Conditioned Cross Attention for inferring unobserved structure. Jointly models visible and hidden regions.

Result: Achieves superior fidelity and completeness under occlusion-heavy sparse-view settings on both synthetic and real-world datasets. Faithfully reconstructs 3D objects consistent with sparse-view constraints while plausibly hallucinating unseen parts.

Conclusion: AmodalGen3D addresses the pressing need for object-level 3D scene reconstruction in robotics, AR/VR, and embodied AI applications by enabling complete reconstruction from sparse, occluded views.

Abstract: Reconstructing 3D objects from a few unposed and partially occluded views is a common yet challenging problem in real-world scenarios, where many object surfaces are never directly observed. Traditional multi-view or inpainting-based approaches struggle under such conditions, often yielding incomplete or geometrically inconsistent reconstructions. We introduce AmodalGen3D, a generative framework for amodal 3D object reconstruction that infers complete, occlusion-free geometry and appearance from arbitrary sparse inputs. The model integrates 2D amodal completion priors with multi-view stereo geometry conditioning, supported by a View-Wise Cross Attention mechanism for sparse-view feature fusion and a Stereo-Conditioned Cross Attention module for unobserved structure inference. By jointly modeling visible and hidden regions, AmodalGen3D faithfully reconstructs 3D objects that are consistent with sparse-view constraints while plausibly hallucinating unseen parts. Experiments on both synthetic and real-world datasets demonstrate that AmodalGen3D achieves superior fidelity and completeness under occlusion-heavy sparse-view settings, addressing a pressing need for object-level 3D scene reconstruction in robotics, AR/VR, and embodied AI applications.

Method: The authors introduce TAPVid-360 task and create a dataset using 360 videos as supervision. They resample 360 videos into narrow field-of-view perspectives and compute ground truth directions by tracking points across full panoramas using a 2D pipeline. They adapt CoTracker v3 to predict per-point rotations for direction updates.

Result: Created TAPVid360-10k dataset with 10k perspective videos and ground truth directional point tracking. The adapted CoTracker v3 baseline outperforms existing TAP and TAPVid 3D methods on the new task.

Conclusion: TAPVid-360 enables learning allocentric scene representations without requiring dynamic 4D ground truth scene models, advancing persistent panoramic understanding in computer vision systems.

Abstract: Humans excel at constructing panoramic mental models of their surroundings, maintaining object permanence and inferring scene structure beyond visible regions. In contrast, current artificial vision systems struggle with persistent, panoramic understanding, often processing scenes egocentrically on a frame-by-frame basis. This limitation is pronounced in the Track Any Point (TAP) task, where existing methods fail to track 2D points outside the field of view. To address this, we introduce TAPVid-360, a novel task that requires predicting the 3D direction to queried scene points across a video sequence, even when far outside the narrow field of view of the observed video. This task fosters learning allocentric scene representations without needing dynamic 4D ground truth scene models for training. Instead, we exploit 360 videos as a source of supervision, resampling them into narrow field-of-view perspectives while computing ground truth directions by tracking points across the full panorama using a 2D pipeline. We introduce a new dataset and benchmark, TAPVid360-10k comprising 10k perspective videos with ground truth directional point tracking. Our baseline adapts CoTracker v3 to predict per-point rotations for direction updates, outperforming existing TAP and TAPVid 3D methods.

Method: WalkCLIP learns walkability-aware vision-language representations from GPT-4o generated image captions, refines them with a spatial aggregation module for neighborhood context, and fuses these features with representations from a population dynamics foundation model.

Result: Evaluated at 4,660 locations in Minneapolis-Saint Paul, WalkCLIP outperforms both unimodal and multimodal baselines in predictive accuracy and spatial alignment.

Conclusion: Integrating complementary visual and behavioral signals yields more reliable predictions of the walking environment than single-source approaches, demonstrating the value of multimodal fusion for urban walkability assessment.

Abstract: Urban walkability is a cornerstone of public health, sustainability, and quality of life. Traditional walkability assessments rely on surveys and field audits, which are costly and difficult to scale. Recent studies have used satellite imagery, street view imagery, or population indicators to estimate walkability, but these single-source approaches capture only one dimension of the walking environment. Satellite data describe the built environment from above, but overlook the pedestrian perspective. Street view imagery captures conditions at the ground level, but lacks broader spatial context. Population dynamics reveal patterns of human activity but not the visual form of the environment. We introduce WalkCLIP, a multimodal framework that integrates these complementary viewpoints to predict urban walkability. WalkCLIP learns walkability-aware vision-language representations from GPT-4o generated image captions, refines these representations with a spatial aggregation module that incorporates neighborhood context, and fuses the resulting features with representations from a population dynamics foundation model. Evaluated at 4,660 locations throughout Minneapolis-Saint Paul, WalkCLIP outperforms unimodal and multimodal baselines in both predictive accuracy and spatial alignment. These results show that the integration of visual and behavioral signals yields reliable predictions of the walking environment.

Method: Used state-of-the-art deep learning models (VGG16, MobileNetV2, Xception) on a novel dataset of 4,000 endoscopic images across four disease classes, with explainable AI via Grad-CAM visualization for clinical interpretability.

Result: VGG16 and MobileNetV2 achieved 96.5% test accuracy, Xception reached 94.24%, establishing robust benchmarks for automated GI disease classification with visual explanations of model decisions.

Conclusion: Demonstrates potential for accurate, interpretable medical image analysis in complex real-world conditions, advancing GI computer-aided diagnosis through benchmarks, comparative insights, and explainable AI.

Abstract: This paper presents a comprehensive comparative model analysis on a novel gastrointestinal medical imaging dataset, comprised of 4,000 endoscopic images spanning four critical disease classes: Diverticulosis, Neoplasm, Peritonitis, and Ureters. Leveraging state-of-the-art deep learning techniques, the study confronts common endoscopic challenges such as variable lighting, fluctuating camera angles, and frequent imaging artifacts. The best performing models, VGG16 and MobileNetV2, each achieved a test accuracy of 96.5%, while Xception reached 94.24%, establishing robust benchmarks and baselines for automated disease classification. In addition to strong classification performance, the approach includes explainable AI via Grad-CAM visualization, enabling identification of image regions most influential to model predictions and enhancing clinical interpretability. Experimental results demonstrate the potential for robust, accurate, and interpretable medical image analysis even in complex real-world conditions. This work contributes original benchmarks, comparative insights, and visual explanations, advancing the landscape of gastrointestinal computer-aided diagnosis and underscoring the importance of diverse, clinically relevant datasets and model explainability in medical AI research.

Method: Developed TRACE-CoT dataset to capture dentists’ diagnostic reasoning, implemented four-stage training paradigm, and created MMOral-Uni benchmark with 2,809 QA pairs across five modalities and five tasks.

Result: OralGPT-Omni achieves 51.84 on MMOral-Uni benchmark and 45.31 on MMOral-OPG benchmark, dramatically outperforming GPT-5 scores.

Conclusion: The work promotes intelligent dentistry and paves the way for future advances in dental image analysis, with all code, benchmark, and models to be made publicly available.

Abstract: Multimodal Large Language Models (MLLMs) have exhibited immense potential across numerous medical specialties; yet, dentistry remains underexplored, in part due to limited domain-specific data, scarce dental expert annotations, insufficient modality-specific modeling, and challenges in reliability. In this paper, we present OralGPT-Omni, the first dental-specialized MLLM designed for comprehensive and trustworthy analysis across diverse dental imaging modalities and clinical tasks. To explicitly capture dentists’ diagnostic reasoning, we construct TRACE-CoT, a clinically grounded chain-of-thought dataset that mirrors dental radiologists’ decision-making processes. This reasoning supervision, combined with our proposed four-stage training paradigm, substantially strengthens the model’s capacity for dental image understanding and analysis. In parallel, we introduce MMOral-Uni, the first unified multimodal benchmark for dental image analysis. It comprises 2,809 open-ended question-answer pairs spanning five modalities and five tasks, offering a comprehensive evaluation suite to date for MLLMs in digital dentistry. OralGPT-Omni achieves an overall score of 51.84 on the MMOral-Uni benchmark and 45.31 on the MMOral-OPG benchmark, dramatically outperforming the scores of GPT-5. Our work promotes intelligent dentistry and paves the way for future advances in dental image analysis. All code, benchmark, and models will be made publicly available.

Method: PAT3D generates 3D objects from text prompts, infers spatial relations, organizes them into hierarchical scene trees, then uses a differentiable rigid-body simulator to achieve static equilibrium under gravity. A simulation-in-the-loop optimization ensures physical stability, non-intersection, and semantic consistency.

Result: PAT3D substantially outperforms prior approaches in physical plausibility, semantic consistency, and visual quality. It uniquely produces simulation-ready 3D scenes suitable for downstream tasks like scene editing and robotic manipulation.

Conclusion: PAT3D represents a significant advancement in text-to-3D generation by integrating physics simulation, enabling physically plausible, intersection-free scenes that are ready for practical applications beyond just visualization.

Abstract: We introduce PAT3D, the first physics-augmented text-to-3D scene generation framework that integrates vision-language models with physics-based simulation to produce physically plausible, simulation-ready, and intersection-free 3D scenes. Given a text prompt, PAT3D generates 3D objects, infers their spatial relations, and organizes them into a hierarchical scene tree, which is then converted into initial conditions for simulation. A differentiable rigid-body simulator ensures realistic object interactions under gravity, driving the scene toward static equilibrium without interpenetrations. To further enhance scene quality, we introduce a simulation-in-the-loop optimization procedure that guarantees physical stability and non-intersection, while improving semantic consistency with the input prompt. Experiments demonstrate that PAT3D substantially outperforms prior approaches in physical plausibility, semantic consistency, and visual quality. Beyond high-quality generation, PAT3D uniquely enables simulation-ready 3D scenes for downstream tasks such as scene editing and robotic manipulation. Code and data will be released upon acceptance.

Method: ReAG combines coarse- and fine-grained retrieval with a critic model to filter irrelevant passages. It uses a multi-stage training strategy with reinforcement learning to enhance reasoning over retrieved content, while supervised fine-tuning serves only as a cold start.

Result: Extensive experiments on Encyclopedic-VQA and InfoSeek datasets show ReAG significantly outperforms prior methods, improving answer accuracy and providing interpretable reasoning grounded in retrieved evidence.

Conclusion: ReAG effectively addresses limitations of current retrieval-augmented MLLMs for knowledge-intensive VQA by integrating improved retrieval with enhanced reasoning capabilities, offering better performance and interpretability.

Abstract: Multimodal Large Language Models (MLLMs) have shown impressive capabilities in jointly understanding text, images, and videos, often evaluated via Visual Question Answering (VQA). However, even state-of-the-art MLLMs struggle with domain-specific or knowledge-intensive queries, where relevant information is underrepresented in pre-training data. Knowledge-based VQA (KB-VQA) addresses this by retrieving external documents to condition answer generation, but current retrieval-augmented approaches suffer from low precision, noisy passages, and limited reasoning. To address this, we propose ReAG, a novel Reasoning-Augmented Multimodal RAG approach that combines coarse- and fine-grained retrieval with a critic model that filters irrelevant passages, ensuring high-quality additional context. The model follows a multi-stage training strategy leveraging reinforcement learning to enhance reasoning over retrieved content, while supervised fine-tuning serves only as a cold start. Extensive experiments on Encyclopedic-VQA and InfoSeek demonstrate that ReAG significantly outperforms prior methods, improving answer accuracy and providing interpretable reasoning grounded in retrieved evidence. Our source code is publicly available at: https://github.com/aimagelab/ReAG.

Method: Proposes MRLM (Meter Reading Language Model), a vision-language model based on physical relation injection. Instead of learning image-level correlations, it explicitly encodes geometric and causal relationships between pointers and scales, aligning perception with physical reasoning. Uses cross-attentional fusion and adaptive expert selection to interpret dial configurations.

Result: Created RPM-10K dataset with 10,730 meter images reflecting real-world challenges. Extensive experiments validated the effectiveness of the MRLM framework on the new benchmark dataset.

Conclusion: The paper addresses the lack of large-scale datasets in pointer meter reading by introducing RPM-10K and proposes MRLM, a novel vision-language model that incorporates physical reasoning to achieve robust meter reading recognition. Both dataset and code will be publicly released.

Abstract: The precise reading recognition of pointer meters plays a key role in smart power systems, but existing approaches remain fragile due to challenges like reflections, occlusions, dynamic viewing angles, and overly between thin pointers and scale markings. Up to now, this area still lacks large-scale datasets to support the development of robust algorithms. To address these challenges, this paper first presents a new large-scale benchmark dataset for dial reading, termed RPM-10K, which contains 10730 meter images that fully reflect the aforementioned key challenges. Built upon the dataset, we propose a novel vision-language model for pointer meter reading recognition, termed MRLM, based on physical relation injection. Instead of exhaustively learning image-level correlations, MRLM explicitly encodes the geometric and causal relationships between the pointer and the scale, aligning perception with physical reasoning in the spirit of world-model perspectives. Through cross-attentional fusion and adaptive expert selection, the model learns to interpret dial configurations and generate precise numeric readings. Extensive experiments fully validated the effectiveness of our proposed framework on the newly proposed benchmark dataset. Both the dataset and source code will be released on https://github.com/Event-AHU/DialBench

Method: 1) Introduce CogIP-Bench benchmark to evaluate MLLMs on image cognitive properties; 2) Use post-training phase to align models with human judgments; 3) Integrate cognitively-aligned MLLM into image generation pipeline to guide synthesis.

Result: Evaluation reveals significant gap: current models poorly aligned with human perception of nuanced cognitive properties. Post-training effectively bridges this gap, enhancing alignment with human judgments. The learned cognitive alignment is transferable to downstream creative tasks, enabling generation of images with desired traits like memorability or visual appeal.

Conclusion: The work provides: 1) a benchmark to measure human-like perception, 2) a post-training pipeline to enhance it, and 3) demonstrates that this alignment unlocks more human-centric AI capabilities for creative applications.

Abstract: While Multimodal Large Language Models (MLLMs) are adept at answering what is in an image-identifying objects and describing scenes-they often lack the ability to understand how an image feels to a human observer. This gap is most evident when considering subjective cognitive properties, such as what makes an image memorable, funny, aesthetically pleasing, or emotionally evocative. To systematically address this challenge, we introduce CogIP-Bench, a comprehensive benchmark for evaluating MLLMs on such image cognitive properties. Our evaluation reveals a significant gap: current models are poorly aligned with human perception of these nuanced properties. We then demonstrate that a post-training phase can effectively bridge this gap, significantly enhancing the model’s alignment with human judgments. Furthermore, we show that this learned cognitive alignment is not merely predictive but also transferable to downstream creative tasks. By integrating our cognitively-aligned MLLM into an image generation pipeline, we can guide the synthesis process to produce images that better embody desired traits, such as being more memorable or visually appealing. Our work provides a benchmark to measure this human-like perception, a post-training pipeline to enhance it, and a demonstration that this alignment unlocks more human-centric AI.

Method: PPBoost uses vision-language models to generate initial pseudo-bboxes from text, filters unreliable predictions with uncertainty-aware criteria, trains a pseudo-labeled detector on retained image-bbox pairs, refines bboxes during inference by expanding them to cover targets, and uses enhanced bboxes to guide segmentation models.

Result: Across three diverse datasets, PPBoost consistently improves Dice and Normalized Surface Distance over text- and visual-prompted baselines, and surpasses few-shot segmentation models without using labeled data. It generalizes to multiple segmentation backbones.

Conclusion: PPBoost effectively amplifies weak text cues into strong spatial guidance for zero-shot medical image segmentation, bridging the gap between text and visual prompting approaches while operating without labeled data.

Abstract: Text-prompted foundation models for medical image segmentation offer an intuitive way to delineate anatomical structures from natural language queries, but their predictions often lack spatial precision and degrade under domain shift. In contrast, visual-prompted models achieve strong segmentation performance across diverse modalities by leveraging spatial cues of precise bounding-box (bbox) prompts to guide the segmentation of target lesions. However, it is costly and challenging to obtain the precise visual prompts in clinical practice. We propose PPBoost (Progressive Prompt-Boosting), a framework that bridges these limitations by transforming weak text-derived signals into strong, spatially grounded visual prompts, operating under a strict zero-shot regime with no image- or pixel-level segmentation labels. PPBoost first uses a vision-language model to produce initial pseudo-bboxes conditioned on the textual object descriptions and applies an uncertainty-aware criterion to filter unreliable predictions. The retained image-bboxes pairs are then leveraged to train a pseudo-labeled detector, producing the high-quality bboxes for the query images. During inference, PPBoost further refines the generated bboxes by appropriately expanding them to tightly cover the target anatomical structures. The enhanced spatially-grounding bbox prompts guide existing segmentation models to generate final dense masks, effectively amplifying weak text cues into strong spatial guidance. Across three datasets spanning diverse modalities and anatomies, PPBoost consistently improves Dice and Normalized Surface Distance over text- and visual-prompted baselines and, notably, surpasses few-shot segmentation models without using labeled data. PPBoost can generalize to multiple typical visual segmentation model backbones.

Method: Created Qualcomm Interactive Cooking benchmark/dataset based on CaptainCook4D with dense annotations of timed instructions, feedback messages, and mistake alerts. Developed LiveMamba, a streaming multi-modal LLM designed for interactive instructional guidance.

Result: Provides first dedicated benchmark for live situated coaching with precise timestamped mistake annotations. Evaluates state-of-the-art multi-modal LLMs and introduces LiveMamba as a strong baseline model.

Conclusion: This work establishes foundational resources (benchmark and baseline model) for developing and evaluating real-time interactive coaching systems that can detect and respond to user mistakes during task execution.

Abstract: Multi-modal Large Language Models (LLM) have advanced conversational abilities but struggle with providing live, interactive step-by-step guidance, a key capability for future AI assistants. Effective guidance requires not only delivering instructions but also detecting their successful execution, as well as identifying and alerting users to mistakes, all of which has to happen in real-time. This requires models that are not turn-based, but that can react asynchronously to a video stream, as well as video data showing users performing tasks including mistakes and their corrections. To this end, we introduce Qualcomm Interactive Cooking, a new benchmark and dataset built upon CaptainCook4D, which contains user mistakes during task execution. Our dataset and benchmark features densely annotated, timed instructions and feedback messages, specifically including mistake alerts precisely timestamped to their visual occurrence in the video. We evaluate state-of-the-art multi-modal LLMs on the Qualcomm Interactive Cooking benchmark and introduce LiveMamba, a streaming multi-modal LLM designed for interactive instructional guidance. This work provides the first dedicated benchmark and a strong baseline for developing and evaluating on live, situated coaching.

Method: Proposed EASI (Evaluation of multimodAl LLMs on Spatial Intelligence) with a comprehensive taxonomy of spatial tasks unifying existing benchmarks and newly curated ones. Conducted study across eight key benchmarks using over ten billion total tokens, with both quantitative evaluation and qualitative assessment across diverse scenarios.

Result: 1) GPT-5 demonstrates unprecedented strength in spatial intelligence but 2) still falls significantly short of human performance across broad SI-tasks. 3) SI-tasks expose greater model capability deficiency than non-SI tasks, and 4) proprietary models don’t show decisive advantage on the most difficult tasks.

Conclusion: Current multimodal models, including the most advanced ones, still have substantial gaps in spatial intelligence compared to humans. The EASI framework provides a standardized, open-source evaluation platform with a leaderboard to accelerate collective progress toward robust spatial intelligence in AI systems.

Abstract: Multimodal models have achieved remarkable progress in recent years. Nevertheless, they continue to exhibit notable limitations in spatial understanding and reasoning, the very capability that anchors artificial general intelligence in the physical world. With the recent release of GPT-5, allegedly the most powerful AI model to date, it is timely to examine where the leading models (GPT, Gemini, Grok, Seed, Qwen, and Intern) stand on the path toward spatial intelligence (SI). We thus propose EASI for holistic Evaluation of multimodAl LLMs on Spatial Intelligence. EASI conceptualizes a comprehensive taxonomy of spatial tasks that unifies existing benchmarks and a growing collection of newly curated ones, enabling systematic evaluation of state-of-the-art models. In this report, we conduct the study across eight key benchmarks, at a cost exceeding ten billion total tokens. Our empirical study then reveals that (1) GPT-5 demonstrates unprecedented strength in SI, yet (2) still falls short of human performance significantly across a broad spectrum of SI-tasks. Moreover, we (3) show that SI-tasks expose greater model capability deficiency than non-SI tasks, to the extent that (4) proprietary models do not exhibit a decisive advantage when facing the most difficult ones. In addition, we conduct a qualitative evaluation across a diverse set of scenarios that are intuitive for humans, yet fail the most advanced multimodal models. EASI is an ongoing community effort: we have open-sourced the EASI codebase that provides a one-stop and reproducible solution with standardized interfaces, integrated protocols and prompts that significantly reduce the friction of configuring and running multiple benchmarks; we have also launched an accompanying EASI leaderboard to provide a continually updated snapshot of model performance across the full SI spectrum, accelerating collective progress toward robust SI.

Method: Comprehensive acceleration pipeline including: batch processing with new velocity field, vectorization of heterogeneous time-step batch processing, and dynamic TensorRT compilation for flow-based models.

Result: Achieves 611% acceleration for 512x512 image generation, significantly outperforming existing public methods that typically achieve only 18% acceleration.

Conclusion: The proposed acceleration pipeline successfully addresses the unique challenges of Rectified Flow models and delivers substantial performance improvements beyond current non-generalized acceleration methods.

Abstract: New technologies such as Rectified Flow and Flow Matching have significantly improved the performance of generative models in the past two years, especially in terms of control accuracy, generation quality, and generation efficiency. However, due to some differences in its theory, design, and existing diffusion models, the existing acceleration methods cannot be directly applied to the Rectified Flow model. In this article, we have comprehensively implemented an overall acceleration pipeline from the aspects of theory, design, and reasoning strategies. This pipeline uses new methods such as batch processing with a new velocity field, vectorization of heterogeneous time-step batch processing, and dynamic TensorRT compilation for the new methods to comprehensively accelerate related models based on flow models. Currently, the existing public methods usually achieve an acceleration of 18%, while experiments have proved that our new method can accelerate the 512*512 image generation speed to up to 611%, which is far beyond the current non-generalized acceleration methods.

Method: MedEyes incorporates off-policy expert guidance by converting expert visual search trajectories into structured behavioral signals. It uses a Gaze-guided Reasoning Navigator (GRN) with dual-mode exploration (scanning for abnormality localization and drilling for regional analysis), a Confidence Value Sampler (CVS) with nucleus sampling and adaptive termination, and a dual-stream GRPO optimization framework to decouple on-policy and off-policy learning signals.

Result: MedEyes achieves an average performance improvement of +8.5% across multiple medical VQA benchmarks, demonstrating superior clinical reasoning capabilities.

Conclusion: MedEyes successfully models clinician-style diagnostic reasoning by incorporating expert visual guidance, balancing expert imitation with autonomous discovery, and mitigating reward assimilation issues, showing potential for building interpretable medical AI systems.

Abstract: Accurate medical diagnosis often involves progressive visual focusing and iterative reasoning, characteristics commonly observed in clinical workflows. While recent vision-language models demonstrate promising chain-of-thought (CoT) reasoning capabilities via reinforcement learning with verifiable rewards (RLVR), their purely on-policy learning paradigm tends to reinforce superficially coherent but clinically inaccurate reasoning paths. We propose MedEyes, a novel reinforcement learning framework that dynamically models clinician-style diagnostic reasoning by progressively attending to and interpreting relevant medical image regions. By incorporating off-policy expert guidance, MedEyes converts expert visual search trajectories into structured external behavioral signals, guiding the model toward clinically aligned visual reasoning. We design the Gaze-guided Reasoning Navigator (GRN) to emulate the diagnostic process through a dual-mode exploration strategy, scanning for systematic abnormality localization and drilling for detailed regional analysis. To balance expert imitation and autonomous discovery, we introduce the Confidence Value Sampler (CVS), which employs nucleus sampling and adaptive termination to create diverse yet credible exploration paths. Finally, the dual-stream GRPO optimization framework decouples on-policy and off-policy learning signals, mitigating reward assimilation and entropy collapse. Experiments demonstrate that MedEyes achieves an average performance improvement of +8.5% across multiple medical VQA benchmarks, validating MedEyes’s potential in building interpretable medical AI systems.

Method: Built upon established multimodal foundations (Qwen3-VL, InternVL3, Bagel), the authors take a principled approach by systematically curating SenseNova-SI-8M: eight million diverse data samples under a rigorous taxonomy of spatial capabilities. They analyze data scaling effects, emergent generalization, overfitting risks, and conduct preliminary studies on spatial chain-of-thought reasoning.

Result: SenseNova-SI achieves unprecedented performance across spatial intelligence benchmarks: 68.7% on VSI-Bench, 43.3% on MMSI, 85.6% on MindCube, 54.6% on ViewSpatial, and 50.1% on SITE, while maintaining strong general multimodal understanding (84.9% on MMBench-En).

Conclusion: The work demonstrates successful cultivation of spatial intelligence through systematic data curation and scaling. The project is ongoing with continuous updates, and all newly trained multimodal foundation models are publicly released to facilitate further research in spatial intelligence.

Abstract: Despite remarkable progress, multimodal foundation models still exhibit surprising deficiencies in spatial intelligence. In this work, we explore scaling up multimodal foundation models to cultivate spatial intelligence within the SenseNova-SI family, built upon established multimodal foundations including visual understanding models (i.e., Qwen3-VL and InternVL3) and unified understanding and generation models (i.e., Bagel). We take a principled approach to constructing high-performing and robust spatial intelligence by systematically curating SenseNova-SI-8M: eight million diverse data samples under a rigorous taxonomy of spatial capabilities. SenseNova-SI demonstrates unprecedented performance across a broad range of spatial intelligence benchmarks: 68.7% on VSI-Bench, 43.3% on MMSI, 85.6% on MindCube, 54.6% on ViewSpatial, and 50.1% on SITE, while maintaining strong general multimodal understanding (e.g., 84.9% on MMBench-En). More importantly, we analyze the impact of data scaling, discuss early signs of emergent generalization capabilities enabled by diverse data training, analyze the risk of overfitting and language shortcuts, present a preliminary study on spatial chain-of-thought reasoning, and validate the potential downstream application. SenseNova-SI is an ongoing project, and this report will be updated continuously. All newly trained multimodal foundation models are publicly released to facilitate further research in this direction.

Method: Post-hoc approach measuring visual feature consistency within classes using feature projection and multivariate Gaussians to create class-specific probabilistic embeddings. Requires no fine-tuning, works with as few as 10 images per class.

Result: State-of-the-art error detection performance on ImageNet, Flowers102, Food101, EuroSAT, and DTD datasets, significantly outperforming deterministic and probabilistic VLM baselines.

Conclusion: Proposed method is VLM-agnostic, training-free, robust to distribution shift, and effective for uncertainty estimation in VLMs to improve reliability in safety-critical applications.

Abstract: Vision-language models (VLMs), such as CLIP, have gained popularity for their strong open vocabulary classification performance, but they are prone to assigning high confidence scores to misclassifications, limiting their reliability in safety-critical applications. We introduce a training-free, post-hoc uncertainty estimation method for contrastive VLMs that can be used to detect erroneous predictions. The key to our approach is to measure visual feature consistency within a class, using feature projection combined with multivariate Gaussians to create class-specific probabilistic embeddings. Our method is VLM-agnostic, requires no fine-tuning, demonstrates robustness to distribution shift, and works effectively with as few as 10 training images per class. Extensive experiments on ImageNet, Flowers102, Food101, EuroSAT and DTD show state-of-the-art error detection performance, significantly outperforming both deterministic and probabilistic VLM baselines. Code is available at https://github.com/zhenxianglin/ICPE.

Method: Created a comprehensive dataset covering three racket sports with annotations for ball positions and articulated racket pose. Evaluated established baselines and discovered that CrossAttention mechanisms are essential for effective multi-modal fusion of racket pose features with ball tracking data.

Result: The evaluation revealed that naive concatenation of racket pose features degrades performance, while CrossAttention mechanisms unlock their value, leading to trajectory prediction results that surpass strong unimodal baselines.

Conclusion: RacketVision provides a versatile resource for future research in dynamic object tracking, conditional motion forecasting, and multimodal analysis in sports, with the key insight that attention-based fusion is crucial for leveraging racket pose information effectively.

Abstract: We introduce RacketVision, a novel dataset and benchmark for advancing computer vision in sports analytics, covering table tennis, tennis, and badminton. The dataset is the first to provide large-scale, fine-grained annotations for racket pose alongside traditional ball positions, enabling research into complex human-object interactions. It is designed to tackle three interconnected tasks: fine-grained ball tracking, articulated racket pose estimation, and predictive ball trajectory forecasting. Our evaluation of established baselines reveals a critical insight for multi-modal fusion: while naively concatenating racket pose features degrades performance, a CrossAttention mechanism is essential to unlock their value, leading to trajectory prediction results that surpass strong unimodal baselines. RacketVision provides a versatile resource and a strong starting point for future research in dynamic object tracking, conditional motion forecasting, and multimodal analysis in sports. Project page at https://github.com/OrcustD/RacketVision

Method: Simplified task to single-word spoken instructions, created new audio-based grounding dataset with diverse objects and accents, adapted and benchmarked several audio-visual models.

Result: Direct grounding from audio is feasible and sometimes outperforms transcription-based methods, particularly in robustness to linguistic variability.

Conclusion: Direct audio grounding shows promise for more robust and efficient multimodal understanding systems, encouraging renewed research interest in this approach.

Abstract: Understanding human instructions is essential for enabling smooth human-robot interaction. In this work, we focus on object grounding, i.e., localizing an object of interest in a visual scene (e.g., an image) based on verbal human instructions. Despite recent progress, a dominant research trend relies on using text as an intermediate representation. These approaches typically transcribe speech to text, extract relevant object keywords, and perform grounding using models pretrained on large text-vision datasets. However, we question both the efficiency and robustness of such transcription-based pipelines. Specifically, we ask: Can we achieve direct audio-visual alignment without relying on text? To explore this possibility, we simplify the task by focusing on grounding from single-word spoken instructions. We introduce a new audio-based grounding dataset that covers a wide variety of objects and diverse human accents. We then adapt and benchmark several models from the closely audio-visual field. Our results demonstrate that direct grounding from audio is not only feasible but, in some cases, even outperforms transcription-based methods, especially in terms of robustness to linguistic variability. Our findings encourage a renewed interest in direct audio grounding and pave the way for more robust and efficient multimodal understanding systems.

Method: Uses Structure Tensor for Total Least Squares optimal orientation/scale vectors to create adaptive complex probing filters; constructs N-D analytic signal via scalar products of adaptive filters with image neighborhoods; employs Gabor filters as probing functions; represents phase gradient as vector or tensor; demonstrates singularity detection with phase portraits.

Result: Produces continuous, isotropic analytic signal with orientation, scale, phase, and amplitude information; shows tensor representation preserves orientation continuity and detects singularities; demonstrates applications in 2-D fringe pattern processing; compares favorably to Monogenic signal, spline-wavelet pyramid enhancement, and mindtct fingerprint detector.

Conclusion: The Structure Tensor-based analytic signal provides a robust, extensible framework for N-D signal analysis with continuous isotropic properties, effective singularity detection, and superior performance compared to baseline methods in orientation, scale, and phase analysis applications.

Abstract: We produce the analytic signal by using the Structure Tensor, which provides Total Least Squares optimal vectors for estimating orientation and scale locally. Together, these vectors represent N-D frequency components that determine adaptive, complex probing filters. The N-D analytic signal is obtained through scalar products of adaptive filters with image neighborhoods. It comprises orientation, scale, phase, and amplitude information of the neighborhood. The ST analytic signal $ f_A $ is continuous and isotropic, and its extension to N-D is straightforward. The phase gradient can be represented as a vector (instantaneous frequency) or as a tensor. Both are continuous and isotropic, while the tensor additionally preserves continuity of orientation and retains the same information as the vector representation. The tensor representation can also be used to detect singularities. Detection with known phase portraits has been demonstrated in 2-D with relevance to fringe pattern processing in wave physics, including optics and fingerprint measurements. To construct adaptive filters we have used Gabor filter family members as probing functions, but other function families can also be used to sample the spectrum, e.g., quadrature filters. A comparison to three baseline alternatives-in representation (Monogenic signal), enhancement (Monogenic signal combined with a spline-wavelet pyramid), and singularity detection (mindtct, a fingerprint minutia detector widely used in numerous studies)-is also reported using images with precisely known ground truths for location, orientation, singularity type (where applicable), and wave period.

Method: Proposes learning to adapt image styles in the frequency domain to reduce source-target domain discrepancy. Uses only a lightweight preprocessing module during training that is entirely discarded at inference time, incurring no additional computational overhead.

Result: Achieves substantial performance gains on multiple domain-adaptive object detection benchmarks, demonstrating effectiveness in adapting from normal-weather/synthetic source domains to adverse weather/low-light target domains.

Conclusion: Presents a simple yet effective UDA method that is practical and efficient, achieving strong performance without complex adversarial training or architectural modifications, making it suitable for real-world deployment.

Abstract: Unsupervised domain adaptation (UDA) greatly facilitates the deployment of neural networks across diverse environments. However, most state-of-the-art approaches are overly complex, relying on challenging adversarial training strategies, or on elaborate architectural designs with auxiliary models for feature distillation and pseudo-label generation. In this work, we present a simple yet effective UDA method that learns to adapt image styles in the frequency domain to reduce the discrepancy between source and target domains. The proposed approach introduces only a lightweight pre-processing module during training and entirely discards it at inference time, thus incurring no additional computational overhead. We validate our method on domain-adaptive object detection (DAOD) tasks, where ground-truth annotations are easily accessible in source domains (e.g., normal-weather or synthetic conditions) but challenging to obtain in target domains (e.g., adverse weather or low-light scenes). Extensive experiments demonstrate that our method achieves substantial performance gains on multiple benchmarks, highlighting its practicality and effectiveness.

Method: Uses transformer architecture with sparse occupancy representation that directly processes raw image features in end-to-end manner, bypassing BEV projection and discrete tokenization to better capture spatiotemporal dependencies.

Result: Achieves state-of-the-art performance on nuScenes benchmark for 1-3 second occupancy forecasting, significantly outperforming existing approaches, with robust scene dynamics understanding under arbitrary trajectory conditioning.

Conclusion: The proposed end-to-end transformer architecture with sparse occupancy representation effectively overcomes limitations of discrete tokenization and BEV projections, enabling superior 3D scene occupancy forecasting with better spatiotemporal modeling.

Abstract: This paper introduces a novel architecture for trajectory-conditioned forecasting of future 3D scene occupancy. In contrast to methods that rely on variational autoencoders (VAEs) to generate discrete occupancy tokens, which inherently limit representational capacity, our approach predicts multi-frame future occupancy in an end-to-end manner directly from raw image features. Inspired by the success of attention-based transformer architectures in foundational vision and language models such as GPT and VGGT, we employ a sparse occupancy representation that bypasses the intermediate bird’s eye view (BEV) projection and its explicit geometric priors. This design allows the transformer to capture spatiotemporal dependencies more effectively. By avoiding both the finite-capacity constraint of discrete tokenization and the structural limitations of BEV representations, our method achieves state-of-the-art performance on the nuScenes benchmark for 1-3 second occupancy forecasting, outperforming existing approaches by a significant margin. Furthermore, it demonstrates robust scene dynamics understanding, consistently delivering high accuracy under arbitrary future trajectory conditioning.

Method: Created a novel benchmark dataset using Sentinel-2 satellite imagery captured over multiple months, covering diverse areas in Southern France. The dataset provides pixel-level annotations for three major crop types (viticulture, market gardening, field crops) across five ordinal classes of agricultural potential.

Result: AgriPotential is the first public dataset specifically designed for agricultural potential prediction, supporting multiple machine learning tasks including ordinal regression, multi-label classification, and spatio-temporal modeling. The dataset and code are publicly available.

Conclusion: AgriPotential aims to improve data-driven approaches to sustainable land use planning by providing a comprehensive benchmark for agricultural potential prediction using remote sensing data.

Abstract: Remote sensing has emerged as a critical tool for large-scale Earth monitoring and land management. In this paper, we introduce AgriPotential, a novel benchmark dataset composed of Sentinel-2 satellite imagery captured over multiple months. The dataset provides pixel-level annotations of agricultural potentials for three major crop types - viticulture, market gardening, and field crops - across five ordinal classes. AgriPotential supports a broad range of machine learning tasks, including ordinal regression, multi-label classification, and spatio-temporal modeling. The data cover diverse areas in Southern France, offering rich spectral information. AgriPotential is the first public dataset designed specifically for agricultural potential prediction, aiming to improve data-driven approaches to sustainable land use planning. The dataset and the code are freely accessible at: https://zenodo.org/records/15551829

Method: Proposes ICM - Image-Conditioned Manifold regularization that regularizes outputs toward a manifold conditioned on sparse structural information (colormap + Canny edges) instead of text or raw images.

Result: ICM significantly enhances super-resolution performance, particularly in perceptual quality, providing task-aligned and stable regularization that avoids instability of dense conditioning.

Conclusion: ICM offers a more suitable regularization approach for Real-ISR by using sparse image structural information, correcting conceptual misalignment and practical flaws of text-conditioned methods while maintaining stability.

Abstract: Real world image super-resolution (Real-ISR) often leverages the powerful generative priors of text-to-image diffusion models by regularizing the output to lie on their learned manifold. However, existing methods often overlook the importance of the regularizing manifold, typically defaulting to a text-conditioned manifold. This approach suffers from two key limitations. Conceptually, it is misaligned with the Real-ISR task, which is to generate high quality (HQ) images directly tied to the low quality (LQ) images. Practically, the teacher model often reconstructs images with color distortions and blurred edges, indicating a flawed generative prior for this task. To correct these flaws and ensure conceptual alignment, a more suitable manifold must incorporate information from the images. While the most straightforward approach is to condition directly on the raw input images, their high information densities make the regularization process numerically unstable. To resolve this, we propose image-conditioned manifold regularization (ICM), a method that regularizes the output towards a manifold conditioned on the sparse yet essential structural information: a combination of colormap and Canny edges. ICM provides a task-aligned and stable regularization signal, thereby avoiding the instability of dense-conditioning and enhancing the final super-resolution quality. Our experiments confirm that the proposed regularization significantly enhances super-resolution performance, particularly in perceptual quality, demonstrating its effectiveness for real-world applications. We will release the source code of our work for reproducibility.

Method: Two-stage modular approach: (1) Diffusion-based SR model trained on high-resolution RGB imagery (NAIP/WorldStrat) harmonized to Sentinel-2 characteristics, using robust degradation model and contrastive degradation encoder for blind SR; (2) Learned fusion network that upscales remaining multispectral bands using super-resolved RGB as spatial prior.

Result: Outperforms current SOTA baselines on OpenSR benchmark in reflectance fidelity, spectral consistency, spatial alignment, and hallucination suppression. Fusion network significantly outperforms classical and learned pansharpening approaches for enhancing 20m and 60m bands.

Conclusion: Proposes a novel modular framework for Sentinel-2 SR that effectively combines harmonized learning with diffusion models and fusion strategies, achieving superior performance across multiple metrics while enabling accurate enhancement of all spectral bands.

Abstract: This paper presents DiffFuSR, a modular pipeline for super-resolving all 12 spectral bands of Sentinel-2 Level-2A imagery to a unified ground sampling distance (GSD) of 2.5 meters. The pipeline comprises two stages: (i) a diffusion-based super-resolution (SR) model trained on high-resolution RGB imagery from the NAIP and WorldStrat datasets, harmonized to simulate Sentinel-2 characteristics; and (ii) a learned fusion network that upscales the remaining multispectral bands using the super-resolved RGB image as a spatial prior. We introduce a robust degradation model and contrastive degradation encoder to support blind SR. Extensive evaluations of the proposed SR pipeline on the OpenSR benchmark demonstrate that the proposed method outperforms current SOTA baselines in terms of reflectance fidelity, spectral consistency, spatial alignment, and hallucination suppression. Furthermore, the fusion network significantly outperforms classical and learned pansharpening approaches, enabling accurate enhancement of Sentinel-2’s 20 m and 60 m bands. This work proposes a novel modular framework Sentinel-2 SR that utilizes harmonized learning with diffusion models and fusion strategies. Our code and models can be found at https://github.com/NorskRegnesentral/DiffFuSR.

Method: Preserve specular reflection coefficient, reformulate physical constraints using Kubelka-Munk theory to create triple physical constraints (TPCs) between illumination, reflection, and detection. Build TPCNet with these constraints in feature space.

Result: TPCNet outperforms state-of-the-art methods on 10 datasets in both quantitative metrics and visual quality. Constraints improve performance without introducing new parameters.

Conclusion: The proposed TPC theory and TPCNet effectively address limitations of previous Retinex-based methods by incorporating specular reflection and feature-space constraints, achieving superior low-light enhancement performance.

Abstract: Low-light image enhancement is an essential computer vision task to improve image contrast and to decrease the effects of color bias and noise. Many existing interpretable deep-learning algorithms exploit the Retinex theory as the basis of model design. However, previous Retinex-based algorithms, that consider reflected objects as ideal Lambertian ignore specular reflection in the modeling process and construct the physical constraints in image space, limiting generalization of the model. To address this issue, we preserve the specular reflection coefficient and reformulate the original physical constraints in the imaging process based on the Kubelka-Munk theory, thereby constructing constraint relationship between illumination, reflection, and detection, the so-called triple physical constraints (TPCs)theory. Based on this theory, the physical constraints are constructed in the feature space of the model to obtain the TPC network (TPCNet). Comprehensive quantitative and qualitative benchmark and ablation experiments confirm that these constraints effectively improve the performance metrics and visual quality without introducing new parameters, and demonstrate that our TPCNet outperforms other state-of-the-art methods on 10 datasets.

Method: Dual-branch Network with Adaptation (DNA) with adaptive style branch (learns stroke attributes like direction, spacing, placement, flow) and adaptive content branch (decomposes characters into structural info via local encoder and texture details via global encoder).

Result: Extensive experiments show DNA achieves state-of-the-art performance for unseen OHG setting, effectively handling both unseen writers and characters.

Conclusion: DNA model successfully addresses the challenge of generating handwriting for unseen writers and characters, making online handwriting generation more practical for real-world applications.

Abstract: Online handwriting generation (OHG) enhances handwriting recognition models by synthesizing diverse, human-like samples. However, existing OHG methods struggle to generate unseen characters, particularly in glyph-based languages like Chinese, limiting their real-world applicability. In this paper, we introduce our method for OHG, where the writer’s style and the characters generated during testing are unseen during training. To tackle this challenge, we propose a Dual-branch Network with Adaptation (DNA), which comprises an adaptive style branch and an adaptive content branch. The style branch learns stroke attributes such as writing direction, spacing, placement, and flow to generate realistic handwriting. Meanwhile, the content branch is designed to generalize effectively to unseen characters by decomposing character content into structural information and texture details, extracted via local and global encoders, respectively. Extensive experiments demonstrate that our DNA model is well-suited for the unseen OHG setting, achieving state-of-the-art performance.

Method: WorldWander builds on advanced video diffusion transformers and integrates two key components: (1) In-Context Perspective Alignment and (2) Collaborative Position Encoding to efficiently model cross-view synchronization. The framework also uses a curated large-scale dataset called EgoExo-8K containing synchronized egocentric-exocentric triplets from both synthetic and real-world scenarios.

Result: Experiments demonstrate that WorldWander achieves superior perspective synchronization, character consistency, and generalization, setting a new benchmark for egocentric-exocentric video translation.

Conclusion: WorldWander presents an effective framework for bridging egocentric and exocentric video perspectives, addressing an important gap in video generation with applications across multiple domains including filmmaking and AI systems.

Abstract: Video diffusion models have recently achieved remarkable progress in realism and controllability. However, achieving seamless video translation across different perspectives, such as first-person (egocentric) and third-person (exocentric), remains underexplored. Bridging these perspectives is crucial for filmmaking, embodied AI, and world models. Motivated by this, we present WorldWander, an in-context learning framework tailored for translating between egocentric and exocentric worlds in video generation. Building upon advanced video diffusion transformers, WorldWander integrates (i) In-Context Perspective Alignment and (ii) Collaborative Position Encoding to efficiently model cross-view synchronization. To further support our task, we curate EgoExo-8K, a large-scale dataset containing synchronized egocentric-exocentric triplets from both synthetic and real-world scenarios. Experiments demonstrate that WorldWander achieves superior perspective synchronization, character consistency, and generalization, setting a new benchmark for egocentric-exocentric video translation.

Method: Uses supervised contrastive learning with Rank-N-Contrast (RNC) loss for brain age regression from T1w structural MRI. Leverages ResNet backbone and employs Grad-CAM for visual explanation of regression results.

Result: Achieves MAE of 4.27 years and R² of 0.93 with limited training data, outperforming conventional deep regression with same backbone and comparable to state-of-the-art methods using larger datasets. Grad-CAM reveals more nuanced age-related features with RNC loss.

Conclusion: The proposed method shows strong potential as a biomarker for neurodegenerative disorders, demonstrating correlation between brain age gap and disease severity in Alzheimer’s and Parkinson’s patients.

Abstract: MRI-based brain age estimation models aim to assess a subject’s biological brain age based on information, such as neuroanatomical features. Various factors, including neurodegenerative diseases, can accelerate brain aging and measuring this phenomena could serve as a potential biomarker for clinical applications. While deep learning (DL)-based regression has recently attracted major attention, existing approaches often fail to capture the continuous nature of neuromorphological changes, potentially resulting in sub-optimal feature representation and results. To address this, we propose to use supervised contrastive learning with the recent Rank-N-Contrast (RNC) loss to estimate brain age based on widely used T1w structural MRI for the first time and leverage Grad-RAM to visually explain regression results. Experiments show that our proposed method achieves a mean absolute error (MAE) of 4.27 years and an $R^2$ of 0.93 with a limited dataset of training samples, significantly outperforming conventional deep regression with the same ResNet backbone while performing better or comparably with the state-of-the-art methods with significantly larger training data. Furthermore, Grad-RAM revealed more nuanced features related to age regression with the RNC loss than conventional deep regression. As an exploratory study, we employed the proposed method to estimate the gap between the biological and chronological brain ages in Alzheimer’s Disease and Parkinson’s disease patients, and revealed the correlation between the brain age gap and disease severity, demonstrating its potential as a biomarker in neurodegenerative disorders.

Method: Proposes MoE3D with: 1) MoE-based transformer with specialized expert networks for specific modalities/interactions, 2) Information aggregation module for enhanced fusion, 3) Top-1 gating for efficient expert selection, 4) Progressive pre-training strategy leveraging semantic and 2D priors for better initialization.

Result: Achieves competitive performance across four prevalent 3D understanding tasks. Notably surpasses top-performing counterpart by 6.1 mIoU on Multi3DRefer benchmark.

Conclusion: MoE3D effectively addresses modality heterogeneity in 3D understanding through specialized expert networks and efficient fusion mechanisms, demonstrating significant performance improvements over existing methods.

Abstract: Multi-modal 3D understanding is a fundamental task in computer vision. Previous multi-modal fusion methods typically employ a single, dense fusion network, struggling to handle the significant heterogeneity and complexity across modalities, leading to suboptimal performance. In this paper, we propose MoE3D, which integrates Mixture of Experts (MoE) into the multi-modal learning framework. The core is that we deploy a set of specialized “expert” networks, each adept at processing a specific modality or a mode of cross-modal interaction. Specifically, the MoE-based transformer is designed to better utilize the complementary information hidden in the visual features. Information aggregation module is put forward to further enhance the fusion performance. Top-1 gating is employed to make one expert process features with expert groups, ensuring high efficiency. We further propose a progressive pre-training strategy to better leverage the semantic and 2D prior, thus equipping the network with good initialization. Our MoE3D achieves competitive performance across four prevalent 3D understanding tasks. Notably, our MoE3D surpasses the top-performing counterpart by 6.1 mIoU on Multi3DRefer.

Method: 1) Multi-Level Representation Extractors capture both spot-level and niche-level representations from both histology images and gene expression. 2) Hierarchical Hyperbolic Alignment module unifies these representations in hyperbolic space, performing spatial alignment while hierarchically structuring image and gene embeddings to bridge the modality gap.

Result: HyperST achieves state-of-the-art performance on four public datasets from different tissues, demonstrating superior cross-modal prediction capabilities for spatial transcriptomics.

Conclusion: The hyperbolic space framework effectively models the hierarchical structure of spatial transcriptomics data, enabling better image-gene alignment and more accurate prediction of gene expression from histology images, paving the way for scalable and accurate spatial transcriptomics prediction.

Abstract: Spatial Transcriptomics (ST) merges the benefits of pathology images and gene expression, linking molecular profiles with tissue structure to analyze spot-level function comprehensively. Predicting gene expression from histology images is a cost-effective alternative to expensive ST technologies. However, existing methods mainly focus on spot-level image-to-gene matching but fail to leverage the full hierarchical structure of ST data, especially on the gene expression side, leading to incomplete image-gene alignment. Moreover, a challenge arises from the inherent information asymmetry: gene expression profiles contain more molecular details that may lack salient visual correlates in histological images, demanding a sophisticated representation learning approach to bridge this modality gap. We propose HyperST, a framework for ST prediction that learns multi-level image-gene representations by modeling the data’s inherent hierarchy within hyperbolic space, a natural geometric setting for such structures. First, we design a Multi-Level Representation Extractors to capture both spot-level and niche-level representations from each modality, providing context-aware information beyond individual spot-level image-gene pairs. Second, a Hierarchical Hyperbolic Alignment module is introduced to unify these representations, performing spatial alignment while hierarchically structuring image and gene embeddings. This alignment strategy enriches the image representations with molecular semantics, significantly improving cross-modal prediction. HyperST achieves state-of-the-art performance on four public datasets from different tissues, paving the way for more scalable and accurate spatial transcriptomics prediction.

Method: Two-phase approach: (1) Reinforcement learning-based optimization to reconstruct primary subject, (2) Fuzzing-driven search to recover stylistic modifiers. Works in black-box setting without gradient access.

Result: Achieves CLIP similarity up to 0.958 and SBERT textual alignment up to 0.751, surpassing all baselines. Outperforms strongest baseline by 7.5% in CLIP similarity on in-the-wild images. Maintains strong performance under defensive perturbations.

Conclusion: PROMPTMINER provides an effective black-box solution for prompt stealing with superior performance, better generalization, and robustness against defenses, addressing practical limitations of existing approaches.

Abstract: Text-to-image (T2I) generative models such as Stable Diffusion and FLUX can synthesize realistic, high-quality images directly from textual prompts. The resulting image quality depends critically on well-crafted prompts that specify both subjects and stylistic modifiers, which have become valuable digital assets. However, the rising value and ubiquity of high-quality prompts expose them to security and intellectual-property risks. One key threat is the prompt stealing attack, i.e., the task of recovering the textual prompt that generated a given image. Prompt stealing enables unauthorized extraction and reuse of carefully engineered prompts, yet it can also support beneficial applications such as data attribution, model provenance analysis, and watermarking validation. Existing approaches often assume white-box gradient access, require large-scale labeled datasets for supervised training, or rely solely on captioning without explicit optimization, limiting their practicality and adaptability. To address these challenges, we propose PROMPTMINER, a black-box prompt stealing framework that decouples the task into two phases: (1) a reinforcement learning-based optimization phase to reconstruct the primary subject, and (2) a fuzzing-driven search phase to recover stylistic modifiers. Experiments across multiple datasets and diffusion backbones demonstrate that PROMPTMINER achieves superior results, with CLIP similarity up to 0.958 and textual alignment with SBERT up to 0.751, surpassing all baselines. Even when applied to in-the-wild images with unknown generators, it outperforms the strongest baseline by 7.5 percent in CLIP similarity, demonstrating better generalization. Finally, PROMPTMINER maintains strong performance under defensive perturbations, highlighting remarkable robustness. Code: https://github.com/aaFrostnova/PromptMiner

Method: Proposes GoPrune method using ℓ₂,ₚ-norm for sparse network learning, extending p to [0,1). Develops efficient optimization algorithm based on proximal alternating minimization (PAM) with closed-form solutions for subproblems.

Result: Experiments on CIFAR datasets using ResNet and VGG models demonstrate superior performance in network pruning compared to existing methods.

Conclusion: GoPrune provides an effective structured pruning approach with improved computational efficiency for CNN compression, enabling better deployment on edge devices.

Abstract: Convolutional neural networks (CNNs) suffer from rapidly increasing storage and computational costs as their depth grows, which severely hinders their deployment on resource-constrained edge devices. Pruning is a practical approach for network compression, among which structured pruning is the most effective for inference acceleration. Although existing work has applied the $\ell_p$-norm to pruning, it only considers unstructured pruning with $p\in (0, 1)$ and has low computational efficiency. To overcome these limitations, we propose an accelerated structured pruning method called GoPrune. Our method employs the $\ell_{2,p}$-norm for sparse network learning, where the value of $p$ is extended to $[0, 1)$. Moreover, we develop an efficient optimization algorithm based on the proximal alternating minimization (PAM), and the resulting subproblems enjoy closed-form solutions, thus improving compression efficiency. Experiments on the CIFAR datasets using ResNet and VGG models demonstrate the superior performance of the proposed method in network pruning. Our code is available at https://github.com/xianchaoxiu/GoPrune.

Method: Cue3D is a comprehensive, model-agnostic framework that systematically perturbs individual image cues (shading, texture, silhouette, perspective, edges, local continuity) and measures their impact on 3D output quality across seven state-of-the-art methods spanning regression-based, multi-view, and native 3D generative paradigms.

Result: Analysis reveals that shape meaningfulness, not texture, dictates generalization. Geometric cues, particularly shading, are crucial for 3D generation. The study identifies over-reliance on provided silhouettes and diverse sensitivities to cues such as perspective and local continuity across different model families.

Conclusion: Cue3D advances understanding of how modern 3D networks leverage classical vision cues and offers directions for developing more transparent, robust, and controllable single-image 3D generation models by dissecting their dependencies on specific image cues.

Abstract: Humans and traditional computer vision methods rely on a diverse set of monocular cues to infer 3D structure from a single image, such as shading, texture, silhouette, etc. While recent deep generative models have dramatically advanced single-image 3D generation, it remains unclear which image cues these methods actually exploit. We introduce Cue3D, the first comprehensive, model-agnostic framework for quantifying the influence of individual image cues in single-image 3D generation. Our unified benchmark evaluates seven state-of-the-art methods, spanning regression-based, multi-view, and native 3D generative paradigms. By systematically perturbing cues such as shading, texture, silhouette, perspective, edges, and local continuity, we measure their impact on 3D output quality. Our analysis reveals that shape meaningfulness, not texture, dictates generalization. Geometric cues, particularly shading, are crucial for 3D generation. We further identify over-reliance on provided silhouettes and diverse sensitivities to cues such as perspective and local continuity across model families. By dissecting these dependencies, Cue3D advances our understanding of how modern 3D networks leverage classical vision cues, and offers directions for developing more transparent, robust, and controllable single-image 3D generation models.

Method: 1) Textual prompts: Introduce pre-trained prompts from other datasets as hard tokens, concatenated with soft tokens via learnable mapping layer for competitive prompting. 2) Visual prompts: Use irrelevant video sets and negative prompts as generic attribute anchors to maintain pre-trained semantic space relevance.

Result: Significantly outperforms state-of-the-art prompt tuning approaches across generalization benchmarks, particularly on base-to-new class prediction in video tasks.

Conclusion: The coupling prompt learning framework effectively mitigates semantic space narrowing during fine-tuning, preserving VLM generalization ability while maintaining learning capacity through competitive prompting and attribute preservation.

Abstract: Visual and textual soft prompt tuning can effectively improve the adaptability of Vision-Language Models (VLMs) in downstream tasks. However, fine-tuning on video tasks impairs the model’s generalization ability to unseen classes. Existing methods attempt to mitigate this forgetting effect by regularizing the gap between hand-crafted prompts and soft prompts, but this also weakens the learning ability of soft prompts. To address this challenge, we propose a plug-and-play coupling prompt learning framework to optimize the generalization performance of V-L models in video tasks, with the core motivation of mitigating semantic space narrowing during fine-tuning by introducing an externally supervised prompt. Specifically, for textual prompts, we introduce pre-trained prompts from other datasets as hard prompt tokens. These are concatenated with soft prompt tokens and coupled via a learnable mapping layer. This competitive prompting approach prevents the semantic space from overfitting to supervised categories. In addition, we introduce a set of well-designed irrelevant video sets and negative prompts as generic attribute anchors to maintain the generic relevance of the attributes in the pre-trained semantic space, thus preserving the generalization ability. Experiments on video tasks demonstrate that our method significantly outperforms state-of-the-art prompt tuning approaches across generalization benchmarks, particularly on base-to-new class prediction.

Method: VIN uses a frozen foundation model as feature extractor with a two-layer MLP for patch-level classification of cautery-consistent features. Trained on 120 H&E slides from tonsil tissue blocks with pathologist annotations (~2TB data).

Result: In blind testing on 20 unseen slides, VIN produced coherent margin overlays qualitatively matching expert annotations. Quantitative region-level accuracy was ~73.3%, with errors limited to small areas not disrupting whole-slide margin continuity.

Conclusion: VIN successfully captures cautery-related histomorphology and provides reproducible, ink-free margin delineation suitable for integration into digital pathology workflows and downstream margin distance measurements.

Abstract: Assessing resection margins is central to pathological specimen evaluation and has profound implications for patient outcomes. Current practice employs physical inking, which is applied variably, and cautery artifacts can obscure the true margin on histological sections. We present a virtual inking network (VIN) that autonomously localizes the surgical cut surface on whole-slide images, reducing reliance on inks and standardizing margin-focused review. VIN uses a frozen foundation model as the feature extractor and a compact two-layer multilayer perceptron trained for patch-level classification of cautery-consistent features. The dataset comprised 120 hematoxylin and eosin (H&E) stained slides from 12 human tonsil tissue blocks, resulting in ~2 TB of uncompressed raw image data, where a board-certified pathologist provided boundary annotations. In blind testing with 20 slides from previously unseen blocks, VIN produced coherent margin overlays that qualitatively aligned with expert annotations across serial sections. Quantitatively, region-level accuracy was ~73.3% across the test set, with errors largely confined to limited areas that did not disrupt continuity of the whole-slide margin map. These results indicate that VIN captures cautery-related histomorphology and can provide a reproducible, ink-free margin delineation suitable for integration into routine digital pathology workflows and for downstream measurement of margin distances.

Method: 1) Dual-layer data pruning to remove redundant embodied reasoning that interferes with action learning. 2) Dual-teacher adaptive distillation that provides domain-specific supervision while preserving reasoning ability. 3) VLA Score evaluation framework that decouples VLA capabilities into reasoning, intention, action, and alignment dimensions.

Result: DualVLA achieves 61.0% average success rate in SimplerEnv and 65.4 average score across eight multimodal benchmarks, demonstrating superior balance between action execution and multimodal understanding compared to previous approaches.

Conclusion: The proposed DualVLA framework effectively addresses action degeneration in generalist VLAs through careful data curation and distillation strategies, enabling both precise action execution and strong reasoning capabilities without compromising either aspect.

Abstract: To build a generalizable Vision-Language-Action (VLA) model with strong reasoning ability, a common strategy is to first train a specialist VLA on robot demonstrations to acquire reliable manipulation skills, and then incorporate mixed annotated robot data together with multimodal data to restore broader reasoning capabilities. However, we observe that the resulting reasoning VLA often suffers from degraded action performance compared to the specialist model before fine-tuning, a phenomenon we refer to as action degeneration. To address this issue, we propose DualVLA, which enhances action performance through carefully designed post-training while still preserving reasoning capability. We first introduce a dual-layer data pruning method that removes redundant embodied reasoning, preventing it from adversely influencing action learning. To further strengthen action generation, we design a dual-teacher adaptive distillation strategy that assigns different supervision signals to different data domains while maintaining reasoning ability. To fill the evaluation gap for generalist VLAs, we also propose VLA Score, which decouples VLA capability into reasoning, intention, action, and alignment dimensions for a more fine-grained assessment. Experiments show that DualVLA achieves an average success rate of 61.0 in SimplerEnv and an average score of 65.4 across eight competitive multimodal benchmarks, demonstrating a stronger balance between precise action execution and multimodal understanding. Project Website: https://costaliya.github.io/DualVLA/.

Method: Proposes EASL with emotion-semantic disentanglement modules using progressive training to separately extract semantic and affective features. During pose decoding, emotional representations guide semantic interaction to generate sign poses with 7-class emotion confidence scores, enabling emotional expression recognition.

Result: EASL achieves pose accuracy superior to all compared baselines by integrating multi-emotion information and effectively adapts to diffusion models to generate expressive sign language videos.

Conclusion: The proposed EASL framework successfully addresses the emotional expressiveness limitation in sign language generation, producing more natural and emotionally expressive sign language videos through multi-emotion guidance while maintaining semantic accuracy.

Abstract: Large language models have revolutionized sign language generation by automatically transforming text into high-quality sign language videos, providing accessible communication for the Deaf community. However, existing LLM-based approaches prioritize semantic accuracy while overlooking emotional expressions, resulting in outputs that lack naturalness and expressiveness. We propose EASL (Emotion-Aware Sign Language), a multi-emotion-guided generation architecture for fine-grained emotional integration. We introduce emotion-semantic disentanglement modules with progressive training to separately extract semantic and affective features. During pose decoding, the emotional representations guide semantic interaction to generate sign poses with 7-class emotion confidence scores, enabling emotional expression recognition. Experimental results demonstrate that EASL achieves pose accuracy superior to all compared baselines by integrating multi-emotion information and effectively adapts to diffusion models to generate expressive sign language videos.

Method: Two-unit architecture: Preprocessing Unit (PPU) uses U-shaped net enriched with semantics to refine degraded images, and Detection Unit (DTU) integrates semantic information for object detection using modified YOLO network. Semantics help generate plausible content for missing areas, understand boundaries, and preserve visual coherency.

Result: Achieves 1.47% to 8.80% improvement in mAP compared to existing methods across benchmark datasets of different weather conditions. Demonstrates the effectiveness of semantics for both image enhancement and object detection.

Conclusion: Semantic information is powerful for all-weather image transformation and object detection, offering a comprehensive approach to improve autonomous driving perception in diverse weather conditions. The method pioneers semantic data usage for weather adaptation.

Abstract: In the field of autonomous driving, camera-based perception models are mostly trained on clear weather data. Models that focus on addressing specific weather challenges are unable to adapt to various weather changes and primarily prioritize their weather removal characteristics. Our study introduces a semantic-enabled network for object detection in diverse weather conditions. In our analysis, semantics information can enable the model to generate plausible content for missing areas, understand object boundaries, and preserve visual coherency and realism across both filled-in and existing portions of the image, which are conducive to image transformation and object recognition. Specific in implementation, our architecture consists of a Preprocessing Unit (PPU) and a Detection Unit (DTU), where the PPU utilizes a U-shaped net enriched by semantics to refine degraded images, and the DTU integrates this semantic information for object detection using a modified YOLO network. Our method pioneers the use of semantic data for all-weather transformations, resulting in an increase between 1.47% to 8.80% in mAP compared to existing methods across benchmark datasets of different weather. This highlights the potency of semantics in image enhancement and object detection, offering a comprehensive approach to improving object detection performance. Code will be available at https://github.com/EnisZuo/SemOD.

Method: Developed a stacked ensemble model with diverse pre-trained CNNs (MobileNetV2, YOLOv8, DenseNet201) as base learners and CatBoost as meta-learner for two classification tasks: binary KOA detection and multiclass KL grading (0-4).

Result: Achieved 73% balanced test accuracy for multiclass KL grading and 87.5% for binary KOA detection, outperforming previous works in the literature.

Conclusion: The stacked ensemble model provides an effective automated solution for KOA severity assessment, reducing subjectivity and expertise requirements while achieving higher accuracy than existing methods.

Abstract: Knee Osteoarthritis (KOA) is a musculoskeletal condition that can cause significant limitations and impairments in daily activities, especially among older individuals. To evaluate the severity of KOA, typically, X-ray images of the affected knee are analyzed, and a grade is assigned based on the Kellgren-Lawrence (KL) grading system, which classifies KOA severity into five levels, ranging from 0 to 4. This approach requires a high level of expertise and time and is susceptible to subjective interpretation, thereby introducing potential diagnostic inaccuracies. To address this problem a stacked ensemble model of fine-tuned Convolutional Neural Networks (CNNs) was developed for two classification tasks: a binary classifier for detecting the presence of KOA, and a multiclass classifier for precise grading across the KL spectrum. The proposed stacked ensemble model consists of a diverse set of pre-trained architectures, including MobileNetV2, You Only Look Once (YOLOv8), and DenseNet201 as base learners and Categorical Boosting (CatBoost) as the meta-learner. This proposed model had a balanced test accuracy of 73% in multiclass classification and 87.5% in binary classification, which is higher than previous works in extant literature.

Method: Two-stage pipeline: 1) Detector identifies attacked input images with poisoned textures, 2) Purifier recovers benign images from attacked versions. Incorporates adversarial training to align distributions between recovered and original natural images.

Result: Effectively defends against white-box, black-box, and adaptive attacks in 3DGS systems, achieving state-of-the-art performance in both safety (defense effectiveness) and utility (reconstruction quality).

Conclusion: RemedyGS provides the first comprehensive black-box defense framework for 3DGS systems against computation cost attacks, enabling secure and reliable deployment of 3D reconstruction services.

Abstract: As a mainstream technique for 3D reconstruction, 3D Gaussian splatting (3DGS) has been applied in a wide range of applications and services. Recent studies have revealed critical vulnerabilities in this pipeline and introduced computation cost attacks that lead to malicious resource occupancies and even denial-of-service (DoS) conditions, thereby hindering the reliable deployment of 3DGS. In this paper, we propose the first effective and comprehensive black-box defense framework, named RemedyGS, against such computation cost attacks, safeguarding 3DGS reconstruction systems and services. Our pipeline comprises two key components: a detector to identify the attacked input images with poisoned textures and a purifier to recover the benign images from their attacked counterparts, mitigating the adverse effects of these attacks. Moreover, we incorporate adversarial training into the purifier to enforce distributional alignment between the recovered and original natural images, thereby enhancing the defense efficacy. Experimental results demonstrate that our framework effectively defends against white-box, black-box, and adaptive attacks in 3DGS systems, achieving state-of-the-art performance in both safety and utility.

Method: IMTalker uses implicit motion transfer with cross-attention mechanism to model motion discrepancy and identity alignment in a unified latent space. It includes an identity-adaptive module for cross-identity reenactment and a lightweight flow-matching motion generator that produces implicit motion vectors from audio, pose, and gaze cues.

Result: IMTalker surpasses prior methods in motion accuracy, identity preservation, and audio-lip synchronization. It achieves state-of-the-art quality with superior efficiency, operating at 40 FPS for video-driven and 42 FPS for audio-driven generation on an RTX 4090 GPU.

Conclusion: IMTalker presents an effective solution for talking face generation that addresses limitations of traditional flow-based methods through implicit motion transfer, achieving high-quality results with excellent efficiency and identity preservation.

Abstract: Talking face generation aims to synthesize realistic speaking portraits from a single image, yet existing methods often rely on explicit optical flow and local warping, which fail to model complex global motions and cause identity drift. We present IMTalker, a novel framework that achieves efficient and high-fidelity talking face generation through implicit motion transfer. The core idea is to replace traditional flow-based warping with a cross-attention mechanism that implicitly models motion discrepancy and identity alignment within a unified latent space, enabling robust global motion rendering. To further preserve speaker identity during cross-identity reenactment, we introduce an identity-adaptive module that projects motion latents into personalized spaces, ensuring clear disentanglement between motion and identity. In addition, a lightweight flow-matching motion generator produces vivid and controllable implicit motion vectors from audio, pose, and gaze cues. Extensive experiments demonstrate that IMTalker surpasses prior methods in motion accuracy, identity preservation, and audio-lip synchronization, achieving state-of-the-art quality with superior efficiency, operating at 40 FPS for video-driven and 42 FPS for audio-driven generation on an RTX 4090 GPU. We will release our code and pre-trained models to facilitate applications and future research.

Method: Created CMAQ-OBS dataset for East Asia, then introduced Group-Relative Policy Optimization (GRPO) with class-wise rewards and curriculum rollout to align predictions with operational priorities, addressing cost mismatch in forecasting.

Result: Reduced regional error by 59.5% with new dataset, and GRPO reduced False Alarm Rate by 47.3% compared to SFT-only baseline while achieving competitive F1-score for 48-120 hour forecasts.

Conclusion: The framework significantly improves forecast reliability for practical air quality warning systems by addressing operational cost asymmetry, making it effective for real-world public health applications in complex regions like East Asia.

Abstract: Accurate long horizon forecasting of particulate matter (PM) concentration fields is essential for operational public health decisions. However, achieving reliable forecasts remains challenging in regions with complex terrain and strong atmospheric dynamics such as East Asia. While foundation models such as Aurora offer global generality, they often miss region-specific dynamics and rely on non-real-time inputs, limiting their practical utility for localized warning systems. To address this gap, we construct and release the real-world observations and high-resolution CMAQ-OBS dataset for East Asia, reducing regional error by 59.5% and enabling real-time 48-120 hour forecasts critical for public health alerts. However, standard point-wise objectives cannot reflect asymmetric operational costs, where false alarms deteriorate public trust while missed severe events endanger populations. This cost mismatch causes SFT models to over-predict and yield high False Alarm Rates. We introduce Group-Relative Policy Optimization (GRPO) with class-wise rewards and curriculum rollout to align predictions with operational priorities. Experimental results demonstrate that our framework significantly improves the reliability of the forecast. Compared to the SFT-only baseline, our model reduces the False Alarm Rate by 47.3% while achieving a competitive F1-score, proving its effectiveness for practical, real-world air quality forecasting systems on long lead time scenarios.

Method: Three core components: (1) multimodal concept generator combining LLM semantics with exemplar-based visual cues; (2) Partially Shared Concept Strategy merging concepts based on activation patterns; (3) Concept-Efficient Accuracy (CEA) metric jointly capturing accuracy and compactness.

Result: On eleven diverse datasets, PS-CBM outperforms state-of-the-art CBMs: improves classification accuracy by 1.0%-7.4% and CEA by 2.0%-9.5%, while requiring significantly fewer concepts.

Conclusion: PS-CBM effectively achieves both high accuracy and strong interpretability by addressing fundamental limitations of existing concept bottleneck models through its novel framework and evaluation metric.

Abstract: Concept Bottleneck Models (CBMs) enhance interpretability by introducing a layer of human-understandable concepts between inputs and predictions. While recent methods automate concept generation using Large Language Models (LLMs) and Vision-Language Models (VLMs), they still face three fundamental challenges: poor visual grounding, concept redundancy, and the absence of principled metrics to balance predictive accuracy and concept compactness. We introduce PS-CBM, a Partially Shared CBM framework that addresses these limitations through three core components: (1) a multimodal concept generator that integrates LLM-derived semantics with exemplar-based visual cues; (2) a Partially Shared Concept Strategy that merges concepts based on activation patterns to balance specificity and compactness; and (3) Concept-Efficient Accuracy (CEA), a post-hoc metric that jointly captures both predictive accuracy and concept compactness. Extensive experiments on eleven diverse datasets show that PS-CBM consistently outperforms state-of-the-art CBMs, improving classification accuracy by 1.0%-7.4% and CEA by 2.0%-9.5%, while requiring significantly fewer concepts. These results underscore PS-CBM’s effectiveness in achieving both high accuracy and strong interpretability.

Method: Introduces Voronoi Half-Patch (VHP) representation that decomposes B-reps into unified local units by assigning geometry to nearest half-edges. Uses dual VQ-VAEs to encode vertex topology and VHPs into vertex-based tokens, then trains a decoder-only Transformer to autoregressively predict these tokens, which are decoded into complete B-rep models.

Result: BrepGPT achieves state-of-the-art performance in unconditional B-rep generation and demonstrates versatility in various applications including conditional generation from category labels, point clouds, text descriptions, images, as well as B-rep autocompletion and interpolation.

Conclusion: The proposed single-stage autoregressive framework with VHP representation successfully addresses the limitations of multi-stage approaches, providing an efficient and versatile solution for B-rep generation with broad applicability across different conditional generation tasks.

Abstract: Boundary representation (B-rep) is the de facto standard for CAD model representation in modern industrial design. The intricate coupling between geometric and topological elements in B-rep structures has forced existing generative methods to rely on cascaded multi-stage networks, resulting in error accumulation and computational inefficiency. We present BrepGPT, a single-stage autoregressive framework for B-rep generation. Our key innovation lies in the Voronoi Half-Patch (VHP) representation, which decomposes B-reps into unified local units by assigning geometry to nearest half-edges and sampling their next pointers. Unlike hierarchical representations that require multiple distinct encodings for different structural levels, our VHP representation facilitates unifying geometric attributes and topological relations in a single, coherent format. We further leverage dual VQ-VAEs to encode both vertex topology and Voronoi Half-Patches into vertex-based tokens, achieving a more compact sequential encoding. A decoder-only Transformer is then trained to autoregressively predict these tokens, which are subsequently mapped to vertex-based features and decoded into complete B-rep models. Experiments demonstrate that BrepGPT achieves state-of-the-art performance in unconditional B-rep generation. The framework also exhibits versatility in various applications, including conditional generation from category labels, point clouds, text descriptions, and images, as well as B-rep autocompletion and interpolation.

Method: GRiP introduces a two-stage training framework with cognitive-enhanced RL featuring: 1) Salience-Weighted IoU Reward to prioritize localization of mission-critical objects, and 2) Multi-Heuristic Reward to encourage diverse yet valid reasoning pathways.

Result: GRiP achieves state-of-the-art results among open-source models on challenging benchmarks like TreeBench and V* Bench, demonstrating significant performance gains in complex visual reasoning tasks.

Conclusion: Guiding models with cognitively-inspired signals for what to see and how to think, rather than simplistic rewards, is crucial for unlocking the next level of multimodal intelligence.

Abstract: Models capable of “thinking with images” by dynamically grounding their reasoning in visual evidence represent a major leap in multimodal AI. However, replicating and advancing this ability is non-trivial, with current methods often trapped between the instability of end-to-end reinforcement learning (RL) and the rigidity of supervised fine-tuning (SFT). This leads to models that either struggle to learn or lack the cognitive flexibility required for complex, real-world scenes. To navigate this dilemma, we introduce GRiP (Guided Reasoning and Perception), a novel two-stage training framework that cultivates robust and flexible visual grounded reasoning by explicitly guiding the model’s perceptual focus and logical pathways. GRiP’s core lies in its cognitive-enhanced RL stage, which features two key innovations: (1) a Salience-Weighted IoU Reward that incentivizes the model to prioritize the localization of mission-critical objects over trivial distractors, and (2) a Multi-Heuristic Reward that encourages cognitive flexibility by rewarding diverse yet logically valid reasoning pathways. Initialized from the Qwen2.5-VL-7B model, GRiP demonstrates significant performance gains across multiple challenging benchmarks. It achieves state-of-the-art results among open-source models on the highly challenging TreeBench and V* Bench, proving its effectiveness in complex visual reasoning. Our work demonstrates that moving beyond simplistic rewards and instead guiding models with cognitively-inspired signals for what to see and how to think is crucial for unlocking the next level of multimodal intelligence. The code will be made publicly available.

Method: Proposes a GCN model with Chebyshev Spectral Graph Convolution and GAT layers for ASD classification. Uses multimodal data (rs-fMRI, sMRI, phenotypic) from ABIDE I dataset (870 patients). Creates population graph based on site-based similarity, processes each modality individually in multi-branch architecture, then concatenates features. Uses Chebyshev polynomial filters for localized spectral learning and GAT for attention-weighted aggregation.

Result: Achieves 74.82% test accuracy and 0.82 AUC on entire dataset, outperforming state-of-the-art baselines including conventional GCNs, autoencoder-based deep neural networks, and multimodal CNNs.

Conclusion: The proposed multimodal GCN model with Chebyshev filters and attention mechanisms provides an effective framework for ASD diagnosis, demonstrating superior performance over existing methods and offering a promising approach for objective neurodevelopmental disorder classification.

Abstract: ASD is a complicated neurodevelopmental disorder marked by variation in symptom presentation and neurological underpinnings, making early and objective diagnosis extremely problematic. This paper presents a Graph Convolutional Network (GCN) model, incorporating Chebyshev Spectral Graph Convolution and Graph Attention Networks (GAT), to increase the classification accuracy of ASD utilizing multimodal neuroimaging and phenotypic data. Leveraging the ABIDE I dataset, which contains resting-state functional MRI (rs-fMRI), structural MRI (sMRI), and phenotypic variables from 870 patients, the model leverages a multi-branch architecture that processes each modality individually before merging them via concatenation. Graph structure is encoded using site-based similarity to generate a population graph, which helps in understanding relationship connections across individuals. Chebyshev polynomial filters provide localized spectral learning with lower computational complexity, whereas GAT layers increase node representations by attention-weighted aggregation of surrounding information. The proposed model is trained using stratified five-fold cross-validation with a total input dimension of 5,206 features per individual. Extensive trials demonstrate the enhanced model’s superiority, achieving a test accuracy of 74.82% and an AUC of 0.82 on the entire dataset, surpassing multiple state-of-the-art baselines, including conventional GCNs, autoencoder-based deep neural networks, and multimodal CNNs.

Method: Uses ViT encoder to process raw images and past kinematic state, trained to produce context embeddings inspired by MTR encoder. Instead of learnable intention queries, uses cross attention on intent and context embeddings. Evaluated on Waymo End-to-End Driving Dataset with ablation studies removing images and multiple trajectory outputs.

Result: Transformer-based methods combining visual and kinetic features are not effective at producing useful scene context embeddings, even with CLIP and DINOv2 foundation-model augmentations. However, predicting a distribution over multiple futures instead of a single trajectory boosts planning performance.

Conclusion: While vision-based context embeddings show promise for replacing map features, current transformer architectures struggle to effectively combine visual and kinematic information. The key finding is that multi-future trajectory prediction significantly improves planning performance over single-trajectory approaches.

Abstract: We present a method for trajectory planning for autonomous driving, learning image-based context embeddings that align with motion prediction frameworks and planning-based intention input. Within our method, a ViT encoder takes raw images and past kinematic state as input and is trained to produce context embeddings, drawing inspiration from those generated by the recent MTR (Motion Transformer) encoder, effectively substituting map-based features with learned visual representations. MTR provides a strong foundation for multimodal trajectory prediction by localizing agent intent and refining motion iteratively via motion query pairs; we name our approach MTR-VP (Motion Transformer for Vision-based Planning), and instead of the learnable intention queries used in the MTR decoder, we use cross attention on the intent and the context embeddings, which reflect a combination of information encoded from the driving scene and past vehicle states. We evaluate our methods on the Waymo End-to-End Driving Dataset, which requires predicting the agent’s future 5-second trajectory in bird’s-eye-view coordinates using prior camera images, agent pose history, and routing goals. We analyze our architecture using ablation studies, removing input images and multiple trajectory output. Our results suggest that transformer-based methods that are used to combine the visual features along with the kinetic features such as the past trajectory features are not effective at combining both modes to produce useful scene context embeddings, even when intention embeddings are augmented with foundation-model representations of scene context from CLIP and DINOv2, but that predicting a distribution over multiple futures instead of a single future trajectory boosts planning performance.

Method: FECO framework with two key components: 1) shoe style adversarial training to enforce style-invariant features for contact estimation, overcoming shoe appearance diversity; 2) ground feature extractor that captures ground properties based on spatial context to utilize ground information effectively.

Result: The proposed method achieves robust foot contact estimation regardless of shoe appearance and effectively leverages ground information for accurate dense contact prediction.

Conclusion: FECO framework successfully addresses challenges in dense foot contact estimation from single RGB images by handling shoe appearance diversity through adversarial training and effectively utilizing ground context, advancing detailed foot-world interaction modeling.

Abstract: Foot contact plays a critical role in human interaction with the world, and thus exploring foot contact can advance our understanding of human movement and physical interaction. Despite its importance, existing methods often approximate foot contact using a zero-velocity constraint and focus on joint-level contact, failing to capture the detailed interaction between the foot and the world. Dense estimation of foot contact is crucial for accurately modeling this interaction, yet predicting dense foot contact from a single RGB image remains largely underexplored. There are two main challenges for learning dense foot contact estimation. First, shoes exhibit highly diverse appearances, making it difficult for models to generalize across different styles. Second, ground often has a monotonous appearance, making it difficult to extract informative features. To tackle these issues, we present a FEet COntact estimation (FECO) framework that learns dense foot contact with shoe style-invariant and ground-aware learning. To overcome the challenge of shoe appearance diversity, our approach incorporates shoe style adversarial training that enforces shoe style-invariant features for contact estimation. To effectively utilize ground information, we introduce a ground feature extractor that captures ground properties based on spatial context. As a result, our proposed method achieves robust foot contact estimation regardless of shoe appearance and effectively leverages ground information. Code will be released.

Method: HybridWorldSim integrates multi-traversal neural reconstruction for static backgrounds with generative modeling for dynamic agents, creating a unified framework that addresses visual and spatial consistency limitations of previous methods.

Result: The framework surpasses previous state-of-the-art methods and is complemented by the release of MIRROR dataset - a new multi-traversal dataset capturing diverse routes and environmental conditions across different cities.

Conclusion: HybridWorldSim provides a practical and scalable solution for high-fidelity autonomous driving simulation, offering a valuable resource for research and development in the field.

Abstract: Realistic and controllable simulation is critical for advancing end-to-end autonomous driving, yet existing approaches often struggle to support novel view synthesis under large viewpoint changes or to ensure geometric consistency. We introduce HybridWorldSim, a hybrid simulation framework that integrates multi-traversal neural reconstruction for static backgrounds with generative modeling for dynamic agents. This unified design addresses key limitations of previous methods, enabling the creation of diverse and high-fidelity driving scenarios with reliable visual and spatial consistency. To facilitate robust benchmarking, we further release a new multi-traversal dataset MIRROR that captures a wide range of routes and environmental conditions across different cities. Extensive experiments demonstrate that HybridWorldSim surpasses previous state-of-the-art methods, providing a practical and scalable solution for high-fidelity simulation and a valuable resource for research and development in autonomous driving.

Method: Proposes ARPGNet with facial region relation graph using graph attention mechanism, combined with CNN-based appearance representations in parallel graph attention fusion module for mutual enhancement.

Result: Outperforms or is comparable to state-of-the-art methods on three facial expression recognition datasets.

Conclusion: The proposed approach effectively captures both appearance and relational information for spatial-temporal facial expression representation, demonstrating the importance of modeling facial region relationships.

Abstract: The key to facial expression recognition is to learn discriminative spatial-temporal representations that embed facial expression dynamics. Previous studies predominantly rely on pre-trained Convolutional Neural Networks (CNNs) to learn facial appearance representations, overlooking the relationships between facial regions. To address this issue, this paper presents an Appearance- and Relation-aware Parallel Graph attention fusion Network (ARPGNet) to learn mutually enhanced spatial-temporal representations of appearance and relation information. Specifically, we construct a facial region relation graph and leverage the graph attention mechanism to model the relationships between facial regions. The resulting relational representation sequences, along with CNN-based appearance representation sequences, are then fed into a parallel graph attention fusion module for mutual interaction and enhancement. This module simultaneously explores the complementarity between different representation sequences and the temporal dynamics within each sequence. Experimental results on three facial expression recognition datasets demonstrate that the proposed ARPGNet outperforms or is comparable to state-of-the-art methods.

Method: Two innovative approaches: (1) using an off-the-shelf image adapter to generate 3D objects without textual inversion, providing enhanced control over conditions; (2) a depth conditioned warmup strategy to enhance 3D consistency.

Result: Qualitatively and quantitatively comparable performance to text-inversion-based alternatives with improved 3D consistency. User study shows better matching to input images and superior 3D consistency maintenance.

Conclusion: The proposed Control3D-IP method effectively addresses limitations of textual inversion by providing better control and improved 3D consistency while maintaining competitive generation quality.

Abstract: Recently, the impressive generative capabilities of diffusion models have been demonstrated, producing images with remarkable fidelity. Particularly, existing methods for the 3D object generation tasks, which is one of the fastest-growing segments in computer vision, pre-dominantly use text-to-image diffusion models with textual inversion which train a pseudo text prompt to describe the given image. In practice, various text-to-image generative models employ textual inversion to learn concepts or styles of target object in the pseudo text prompt embedding space, thereby generating sophisticated outputs. However, textual inversion requires additional training time and lacks control ability. To tackle this issues, we propose two innovative methods: (1) using an off-the-shelf image adapter that generates 3D objects without textual inversion, offering enhanced control over conditions such as depth, pose, and text. (2) a depth conditioned warmup strategy to enhance 3D consistency. In experimental results, ours show qualitatively and quantitatively comparable performance and improved 3D consistency to the existing text-inversion-based alternatives. Furthermore, we conduct a user study to assess (i) how well results match the input image and (ii) whether 3D consistency is maintained. User study results show that our model outperforms the alternatives, validating the effectiveness of our approaches. Our code is available at GitHub repository:https://github.com/Seooooooogi/Control3D_IP/

Method: A training-free diffusion framework that enforces multi-view consistency through a consistency loss based on the assumption that corresponding points in unedited images should undergo similar transformations after editing. The framework guides diffusion sampling toward coherent edits and works with various image editing methods in both dense and sparse multi-view setups.

Result: The approach significantly improves 3D consistency compared to existing multi-view editing methods and enables high-quality Gaussian Splat editing with sharp details and strong fidelity to user-specified text prompts.

Conclusion: The proposed training-free framework effectively addresses multi-view inconsistency in image editing, making it suitable for editing 3D representations while maintaining geometric and photometric coherence across different views.

Abstract: Recent advancements in diffusion models have greatly improved text-based image editing, yet methods that edit images independently often produce geometrically and photometrically inconsistent results across different views of the same scene. Such inconsistencies are particularly problematic for editing of 3D representations such as NeRFs or Gaussian Splat models. We propose a training-free diffusion framework that enforces multi-view consistency during the image editing process. The key assumption is that corresponding points in the unedited images should undergo similar transformations after editing. To achieve this, we introduce a consistency loss that guides the diffusion sampling toward coherent edits. The framework is flexible and can be combined with widely varying image editing methods, supporting both dense and sparse multi-view editing setups. Experimental results show that our approach significantly improves 3D consistency compared to existing multi-view editing methods. We also show that this increased consistency enables high-quality Gaussian Splat editing with sharp details and strong fidelity to user-specified text prompts. Please refer to our project page for video results: https://3d-consistent-editing.github.io/

Method: Proposed framework uses license-permissive compound images from biomedical literature as training data. Developed five-stage, context-aware instruction generation with divide-and-conquer strategy to decompose multi-image analysis into sub-tasks. Parsed 237,000+ compound figures with contextual text to create M3LLM. Built PMC-MI-Bench benchmark for evaluation.

Result: M3LLM significantly outperforms both general-purpose and specialized medical MLLMs across multi-image, single-image, text-only, and multi-choice scenarios. Shows strong generalization to longitudinal chest X-ray analysis using MIMIC dataset.

Conclusion: Establishes scalable paradigm for developing medical MLLMs capable of composite reasoning, bridging gap between biomedical literature and real-world clinical applications. Enables models to understand complex spatial, temporal, and cross-modal relationships in medical images.

Abstract: Multi-modal large language models (MLLMs) have shown promise in advancing healthcare. However, most existing models remain confined to single-image understanding, which greatly limits their applicability in clinical workflows. In practice, medical diagnosis and progression often require synthesizing information across multiple images from different modalities or time points. The development of medical MLLMs capable of such multi-image understanding has been hindered by the lack of large-scale, high-quality annotated training data. To address this limitation, we propose a novel framework that leverages license-permissive compound images in biomedical literature, as a rich yet underutilized data source for multi-image analysis. Specifically, we design a five-stage, context-aware instruction generation paradigm underpinned by a divide-and-conquer strategy. By decomposing multi-image analysis into manageable sub-tasks, this paradigm empowers MLLMs to move beyond single-panel analysis and provide a composite understanding by learning the complex spatial, temporal, and cross-modal relationships inherent in these compound figures. By parsing over 237,000 compound figures and their contextual text for instruction generation, we develop M3LLM, a medical multi-image multi-modal large language model. For benchmarking, we construct PMC-MI-Bench for composite understanding, manually validated by medical experts. Extensive experiments show that M3LLM significantly outperforms both general-purpose and specialized medical MLLMs across multi-image, single-image, text-only, and multi-choice scenarios. Notably, M3LLM exhibits strong generalization to longitudinal chest X-ray analysis using the MIMIC dataset. This work establishes a scalable and efficient paradigm for developing medical MLLMs capable of composite reasoning, bridging the gap between biomedical literature and real-world clinical applications.

Method: IE-SRGS jointly leverages external 2DSR priors (HR images and depth maps from 2DSR and depth estimation models) and internal 3DGS features (cross-view consistent, domain-adaptive counterparts from multi-scale 3DGS models). A mask-guided fusion strategy integrates these two knowledge sources to guide 3D Gaussian optimization.

Result: Extensive experiments on both synthetic and real-world benchmarks show that IE-SRGS consistently outperforms state-of-the-art methods in both quantitative accuracy and visual fidelity.

Conclusion: IE-SRGS effectively addresses the 3D Gaussian ambiguity problem by synergistically exploiting complementary strengths of external 2DSR priors and internal 3DGS features, enabling high-fidelity reconstruction of HR 3DGS models from LR inputs.

Abstract: Reconstructing high-resolution (HR) 3D Gaussian Splatting (3DGS) models from low-resolution (LR) inputs remains challenging due to the lack of fine-grained textures and geometry. Existing methods typically rely on pre-trained 2D super-resolution (2DSR) models to enhance textures, but suffer from 3D Gaussian ambiguity arising from cross-view inconsistencies and domain gaps inherent in 2DSR models. We propose IE-SRGS, a novel 3DGS SR paradigm that addresses this issue by jointly leveraging the complementary strengths of external 2DSR priors and internal 3DGS features. Specifically, we use 2DSR and depth estimation models to generate HR images and depth maps as external knowledge, and employ multi-scale 3DGS models to produce cross-view consistent, domain-adaptive counterparts as internal knowledge. A mask-guided fusion strategy is introduced to integrate these two sources and synergistically exploit their complementary strengths, effectively guiding the 3D Gaussian optimization toward high-fidelity reconstruction. Extensive experiments on both synthetic and real-world benchmarks show that IE-SRGS consistently outperforms state-of-the-art methods in both quantitative accuracy and visual fidelity.

Method: VessQC integrates uncertainty maps with segmentation data to guide users to regions most likely containing biologically meaningful errors, enabling focused human-in-the-loop refinement.

Result: In a user study, uncertainty-guided correction improved error detection recall from 67% to 94.0% (p=0.007) without significantly increasing total curation time.

Conclusion: VessQC enables efficient human-in-the-loop refinement of volumetric segmentations, bridging the gap between uncertainty estimation and practical human-computer interaction for real-world applications.

Abstract: Accurate 3D microscopy image segmentation is critical for quantitative bioimage analysis but even state-of-the-art foundation models yield error-prone results. Therefore, manual curation is still widely used for either preparing high-quality training data or fixing errors before analysis. We present VessQC, an open-source tool for uncertainty-guided curation of large 3D microscopy segmentations. By integrating uncertainty maps, VessQC directs user attention to regions most likely containing biologically meaningful errors. In a preliminary user study uncertainty-guided correction significantly improved error detection recall from 67% to 94.0% (p=0.007) without a significant increase in total curation time. VessQC thus enables efficient, human-in-the-loop refinement of volumetric segmentations and bridges a key gap in real-world applications between uncertainty estimation and practical human-computer interaction. The software is freely available at github.com/MMV-Lab/VessQC.

Method: Instead of training SAM to detect tampering, the approach adds adversarial perturbations to make SAM “blind” to the original image, creating a “blank canvas.” When the image is tampered, SAM can then identify forged regions. A frequency-aware optimization strategy is used to thoroughly deceive SAM’s powerful perception capabilities.

Result: Extensive experiments demonstrate the effectiveness of the method in localizing tampered regions in images.

Conclusion: The proposed adversarial perturbation approach with frequency-aware optimization provides an effective solution for tampering detection using SAM’s capabilities in a novel way.

Abstract: AIGC-based image editing technology has greatly simplified the realistic-level image modification, causing serious potential risks of image forgery. This paper introduces a new approach to tampering detection using the Segment Anything Model (SAM). Instead of training SAM to identify tampered areas, we propose a novel strategy. The entire image is transformed into a blank canvas from the perspective of neural models. Any modifications to this blank canvas would be noticeable to the models. To achieve this idea, we introduce adversarial perturbations to prevent SAM from ``seeing anything’’, allowing it to identify forged regions when the image is tampered with. Due to SAM’s powerful perceiving capabilities, naive adversarial attacks cannot completely tame SAM. To thoroughly deceive SAM and make it blind to the image, we introduce a frequency-aware optimization strategy, which further enhances the capability of tamper localization. Extensive experimental results demonstrate the effectiveness of our method.

Method: Proposes TTSnap with noise-aware reward models trained via self-distillation to align intermediate estimate rewards with final clean image rewards, using curriculum training from clean to noisy images, and introduces a reward alignment metric.

Result: Improves performance by over 16% compared to existing methods, enables more efficient test-time scaling, and provides orthogonal gains when combined with post-training techniques and local optimization.

Conclusion: TTSnap effectively addresses computational bottlenecks in test-time scaling by enabling early pruning of low-quality candidates without full denoising, making exploration more efficient while maintaining quality.

Abstract: A prominent approach to test-time scaling for text-to-image diffusion models formulates the problem as a search over multiple noise seeds, selecting the one that maximizes a certain image-reward function. The effectiveness of this strategy heavily depends on the number and diversity of noise seeds explored. However, verifying each candidate is computationally expensive, because each must be fully denoised before a reward can be computed. This severely limits the number of samples that can be explored under a fixed budget. We propose test-time scaling with noise-aware pruning (TTSnap), a framework that prunes low-quality candidates without fully denoising them. The key challenge is that reward models are learned in the clean image domain, and the ranking of rewards predicted for intermediate estimates are often inconsistent with those predicted for clean images. To overcome this, we train noise-aware reward models via self-distillation to align the reward for intermediate estimates with that of the final clean images. To stabilize learning across different noise levels, we adopt a curriculum training strategy that progressively shifts the data domain from clean images to noise images. In addition, we introduce a new metric that measures reward alignment and computational budget utilization. Experiments demonstrate that our approach improves performance by over 16% compared with existing methods, enabling more efficient and effective test-time scaling. It also provides orthogonal gains when combined with post-training techniques and local test-time optimization. Code: https://github.com/TerrysLearning/TTSnap/.

Method: Proposes semantic anchoring that guides adaptation by grounding new concepts in their corresponding distributions. Reformulates personalization as learning rare concepts guided by frequent counterparts through semantic anchoring, enabling stable and controlled adaptation.

Result: Achieves stable adaptation with consistent improvements in both subject fidelity and text-image alignment compared to baselines. Extensive experiments demonstrate robustness and effectiveness of the anchoring strategy.

Conclusion: Semantic anchoring enables effective personalization of diffusion models by expanding pretrained distributions toward personalized regions while preserving semantic structure, solving the trade-off between subject fidelity and prior preservation.

Abstract: Text-to-image diffusion models have achieved remarkable progress in generating diverse and realistic images from textual descriptions. However, they still struggle with personalization, which requires adapting a pretrained model to depict user-specific subjects from only a few reference images. The key challenge lies in learning a new visual concept from a limited number of reference images while preserving the pretrained semantic prior that maintains text-image alignment. When the model focuses on subject fidelity, it tends to overfit the limited reference images and fails to leverage the pretrained distribution. Conversely, emphasizing prior preservation maintains semantic consistency but prevents the model from learning new personalized attributes. Building on these observations, we propose the personalization process through a semantic anchoring that guides adaptation by grounding new concepts in their corresponding distributions. We therefore reformulate personalization as the process of learning a rare concept guided by its frequent counterpart through semantic anchoring. This anchoring encourages the model to adapt new concepts in a stable and controlled manner, expanding the pretrained distribution toward personalized regions while preserving its semantic structure. As a result, the proposed method achieves stable adaptation and consistent improvements in both subject fidelity and text-image alignment compared to baseline methods. Extensive experiments and ablation studies further demonstrate the robustness and effectiveness of the proposed anchoring strategy.

Method: FreqWarm is a plug-and-play frequency warm-up curriculum that increases early-stage exposure to high-frequency latent signals during diffusion or flow-matching training. It doesn’t require modifying or retraining the autoencoder. The method works by analyzing encoder/decoder behaviors through controlled perturbations in both RGB and latent domains, identifying the high-frequency representation gap.

Result: FreqWarm consistently improves generation quality across several high-dimensional autoencoders: decreasing gFID by 14.11 on Wan2.2-VAE, 6.13 on LTX-VAE, and 4.42 on DC-AE-f32. The method is architecture-agnostic and compatible with diverse backbones.

Conclusion: Explicitly managing frequency exposure can successfully turn high-dimensional latent spaces into more diffusible targets. FreqWarm addresses the fundamental reconstruction-generation trade-off in latent diffusion models without requiring autoencoder modifications, making it a practical solution for improving generation quality.

Abstract: Latent diffusion has become the default paradigm for visual generation, yet we observe a persistent reconstruction-generation trade-off as latent dimensionality increases: higher-capacity autoencoders improve reconstruction fidelity but generation quality eventually declines. We trace this gap to the different behaviors in high-frequency encoding and decoding. Through controlled perturbations in both RGB and latent domains, we analyze encoder/decoder behaviors and find that decoders depend strongly on high-frequency latent components to recover details, whereas encoders under-represent high-frequency contents, yielding insufficient exposure and underfitting in high-frequency bands for diffusion model training. To address this issue, we introduce FreqWarm, a plug-and-play frequency warm-up curriculum that increases early-stage exposure to high-frequency latent signals during diffusion or flow-matching training – without modifying or retraining the autoencoder. Applied across several high-dimensional autoencoders, FreqWarm consistently improves generation quality: decreasing gFID by 14.11 on Wan2.2-VAE, 6.13 on LTX-VAE, and 4.42 on DC-AE-f32, while remaining architecture-agnostic and compatible with diverse backbones. Our study shows that explicitly managing frequency exposure can successfully turn high-dimensional latent spaces into more diffusible targets.

Method: 1) Created UMind-DS dataset with 1.2M ultrasound image-text pairs across 16 anatomical regions, including pixel-level annotations and clinician-validated rationales. 2) Developed UMind-VL architecture with lightweight Dynamic Convolutional Mask Decoder that generates masks via dynamic kernels conditioned on LLM outputs. 3) Used task-specific tokens to unify segmentation, detection, geometric measurement, and diagnosis tasks within a single framework.

Result: UMind-VL significantly outperforms existing generalist multimodal models and achieves performance on par with or superior to state-of-the-art specialist models across segmentation, detection, keypoint localization, and diagnostic reasoning benchmarks, while maintaining strong generalization ability.

Conclusion: UMind-VL successfully bridges the gap between low-level ultrasound perception and high-level clinical interpretation, providing a unified foundation model that can handle diverse ultrasound tasks from segmentation to diagnosis within a single framework, demonstrating strong performance and generalization capabilities.

Abstract: Despite significant strides in medical foundation models, the ultrasound domain lacks a comprehensive solution capable of bridging low-level Ultrasound Grounded Perception (e.g., segmentation, localization) and high-level Ultrasound Comprehensive Interpretation (e.g., diagnosis, reasoning). To bridge this gap, we propose UMind-VL, a unified foundation model designed to synergize pixel-level structural understanding with complex clinical reasoning. We first introduce UMind-DS, a large-scale multimodal dataset comprising 1.2 million ultrasound image-text pairs across 16 anatomical regions, enriching standard data with pixel-level annotations and clinician-validated rationales. Architecturally, UMind-VL incorporates a lightweight Dynamic Convolutional Mask Decoder that generates masks via dynamic kernels conditioned on LLM outputs. This design, combined with task-specific tokens, unifies segmentation, detection, geometric measurement, and diagnosis tasks within a single framework. Extensive evaluations demonstrate that UMind-VL significantly outperforms existing generalist multimodal models and achieves performance on par with, or superior to, state-of-the-art specialist models across segmentation, detection, keypoint localization, and diagnostic reasoning benchmarks, while maintaining strong generalization ability. We demonstrate the capability of UMind-VL in Figure 1.

Method: GSPure analyzes view-dependent rendering contributions and uses geometrically accurate feature clustering to precisely isolate and remove watermark-related Gaussian primitives while maintaining scene integrity. It’s specifically designed for the 3DGS representation’s unique characteristics.

Result: GSPure achieves state-of-the-art watermark purification, reducing watermark PSNR by up to 16.34dB while causing minimal degradation to original scene fidelity (less than 1dB PSNR loss). It consistently outperforms existing methods in both effectiveness and generalization across extensive experiments.

Conclusion: The paper demonstrates that current 3DGS watermarking approaches have vulnerabilities, and proposes GSPure as an effective solution for watermark purification. This work highlights the need for more robust watermarking schemes for 3D Gaussian Splatting assets.

Abstract: 3D Gaussian Splatting (3DGS) has emerged as a powerful representation for 3D scenes, widely adopted due to its exceptional efficiency and high-fidelity visual quality. Given the significant value of 3DGS assets, recent works have introduced specialized watermarking schemes to ensure copyright protection and ownership verification. However, can existing 3D Gaussian watermarking approaches genuinely guarantee robust protection of the 3D assets? In this paper, for the first time, we systematically explore and validate possible vulnerabilities of 3DGS watermarking frameworks. We demonstrate that conventional watermark removal techniques designed for 2D images do not effectively generalize to the 3DGS scenario due to the specialized rendering pipeline and unique attributes of each gaussian primitives. Motivated by this insight, we propose GSPure, the first watermark purification framework specifically for 3DGS watermarking representations. By analyzing view-dependent rendering contributions and exploiting geometrically accurate feature clustering, GSPure precisely isolates and effectively removes watermark-related Gaussian primitives while preserving scene integrity. Extensive experiments demonstrate that our GSPure achieves the best watermark purification performance, reducing watermark PSNR by up to 16.34dB while minimizing degradation to original scene fidelity with less than 1dB PSNR loss. Moreover, it consistently outperforms existing methods in both effectiveness and generalization.

Method: Proposes DriveVGGT with three key components: 1) Temporal Video Attention (TVA) module processes multi-camera videos independently to leverage spatiotemporal continuity within single-camera sequences. 2) Multi-camera Consistency Attention (MCA) module uses window attention with normalized relative pose embeddings to establish cross-camera consistency while restricting attention to nearby frames. 3) Extended VGGT heads with additional absolute scale head and ego vehicle pose head.

Result: DriveVGGT outperforms VGGT, StreamVGGT, and fastVGGT on autonomous driving datasets. Extensive ablation studies verify the effectiveness of the proposed designs.

Conclusion: DriveVGGT successfully integrates AD-specific priors into a feed-forward reconstruction framework, demonstrating superior performance for autonomous driving applications compared to existing VGGT variants.

Abstract: Feed-forward reconstruction has recently gained significant attention, with VGGT being a notable example. However, directly applying VGGT to autonomous driving (AD) systems leads to sub-optimal results due to the different priors between the two tasks. In AD systems, several important new priors need to be considered: (i) The overlap between camera views is minimal, as autonomous driving sensor setups are designed to achieve coverage at a low cost. (ii) The camera intrinsics and extrinsics are known, which introduces more constraints on the output and also enables the estimation of absolute scale. (iii) Relative positions of all cameras remain fixed though the ego vehicle is in motion. To fully integrate these priors into a feed-forward framework, we propose DriveVGGT, a scale-aware 4D reconstruction framework specifically designed for autonomous driving data. Specifically, we propose a Temporal Video Attention (TVA) module to process multi-camera videos independently, which better leverages the spatiotemporal continuity within each single-camera sequence. Then, we propose a Multi-camera Consistency Attention (MCA) module to conduct window attention with normalized relative pose embeddings, aiming to establish consistency relationships across different cameras while restricting each token to attend only to nearby frames. Finally, we extend the standard VGGT heads by adding an absolute scale head and an ego vehicle pose head. Experiments show that DriveVGGT outperforms VGGT, StreamVGGT, fastVGGT on autonomous driving dataset while extensive ablation studies verify effectiveness of the proposed designs.

Method: The authors learn an autoencoder that softly selects subsets of patches to reconstruct each target patch. They graph these learned dependencies and compute PageRank scores to determine optimal patch realization order. This ordering is then applied to improve masked image modeling methods.

Result: The patch collapse ordering boosts autoregressive image generation when retraining the MAR model. For image classification, Vision Transformers achieve high accuracy with only 22% of high-rank patches in the collapse order, demonstrating significant efficiency gains.

Conclusion: Patch collapse provides a novel perspective for image modeling that promotes vision efficiency. The optimal patch ordering derived from this phenomenon benefits both image generation and classification tasks, enabling high performance with reduced computational requirements.

Abstract: Observing certain patches in an image reduces the uncertainty of others. Their realization lowers the distribution entropy of each remaining patch feature, analogous to collapsing a particle’s wave function in quantum mechanics. This phenomenon can intuitively be called patch collapse. To identify which patches are most relied on during a target region’s collapse, we learn an autoencoder that softly selects a subset of patches to reconstruct each target patch. Graphing these learned dependencies for each patch’s PageRank score reveals the optimal patch order to realize an image. We show that respecting this order benefits various masked image modeling methods. First, autoregressive image generation can be boosted by retraining the state-of-the-art model MAR. Next, we introduce a new setup for image classification by exposing Vision Transformers only to high-rank patches in the collapse order. Seeing 22% of such patches is sufficient to achieve high accuracy. With these experiments, we propose patch collapse as a novel image modeling perspective that promotes vision efficiency. Our project is available at https://github.com/wguo-ai/CoP .

Method: Models the task as a graph where each node corresponds to an image and each edge triggers joint generation of image pairs. Consolidates all pairwise generations into a unified framework by fusing internal features across image pairs guided by dense input correspondences, without requiring masks or manual supervision. Leverages an emergent prior in text-to-image models for coherent generation when multiple views share a single canvas.

Result: Achieves state-of-the-art consistency and visual quality, and unlocks new capabilities for content creation from image collections.

Conclusion: Match-and-Fuse provides an effective zero-shot, training-free solution for consistent controlled generation of unstructured image sets, enabling new content creation possibilities from image collections.

Abstract: We present Match-and-Fuse - a zero-shot, training-free method for consistent controlled generation of unstructured image sets - collections that share a common visual element, yet differ in viewpoint, time of capture, and surrounding content. Unlike existing methods that operate on individual images or densely sampled videos, our framework performs set-to-set generation: given a source set and user prompts, it produces a new set that preserves cross-image consistency of shared content. Our key idea is to model the task as a graph, where each node corresponds to an image and each edge triggers a joint generation of image pairs. This formulation consolidates all pairwise generations into a unified framework, enforcing their local consistency while ensuring global coherence across the entire set. This is achieved by fusing internal features across image pairs, guided by dense input correspondences, without requiring masks or manual supervision. It also allows us to leverage an emergent prior in text-to-image models that encourages coherent generation when multiple views share a single canvas. Match-and-Fuse achieves state-of-the-art consistency and visual quality, and unlocks new capabilities for content creation from image collections.

Method: MVMAE combines masked image reconstruction with cross-view alignment to learn view-invariant representations. MVMAE-V2T extends this by incorporating radiology reports as auxiliary text supervision while maintaining vision-only inference.

Result: MVMAE consistently outperforms supervised and vision-language baselines on disease classification across three large-scale public datasets (MIMIC-CXR, CheXpert, PadChest). MVMAE-V2T provides additional gains, especially in low-label regimes.

Conclusion: Structural (multi-view) and textual supervision are complementary paths toward scalable, clinically grounded medical foundation models.

Abstract: Building robust medical machine learning systems requires pretraining strategies that exploit the intrinsic structure present in clinical data. We introduce Multiview Masked Autoencoder (MVMAE), a self-supervised framework that leverages the natural multi-view organization of radiology studies to learn view-invariant and disease-relevant representations. MVMAE combines masked image reconstruction with cross-view alignment, transforming clinical redundancy across projections into a powerful self-supervisory signal. We further extend this approach with MVMAE-V2T, which incorporates radiology reports as an auxiliary text-based learning signal to enhance semantic grounding while preserving fully vision-based inference. Evaluated on a downstream disease classification task on three large-scale public datasets, MIMIC-CXR, CheXpert, and PadChest, MVMAE consistently outperforms supervised and vision-language baselines. Furthermore, MVMAE-V2T provides additional gains, particularly in low-label regimes where structured textual supervision is most beneficial. Together, these results establish the importance of structural and textual supervision as complementary paths toward scalable, clinically grounded medical foundation models.

Method: Propose hierarchical Swin Transformer-based neck architecture between backbone and prediction head. Use Swin Transformer for feature upsampling and adapt window sizes specifically for small objects. Integrate with CenterNet detection framework.

Result: Swin Transformer neck with adaptive window sizes improves small object detection performance. Smaller window sizes (default 2) particularly benefit mAP for small object detection tasks.

Conclusion: Specialized neck architecture with hierarchical Swin Transformer and adaptive window sizing effectively addresses small, sparse object detection challenges, demonstrating improved performance for bird detection tasks.

Abstract: Object detection is the task of detecting objects in an image. In this task, the detection of small objects is particularly difficult. Other than the small size, it is also accompanied by difficulties due to blur, occlusion, and so on. Current small object detection methods are tailored to small and dense situations, such as pedestrians in a crowd or far objects in remote sensing scenarios. However, when the target object is small and sparse, there is a lack of objects available for training, making it more difficult to learn effective features. In this paper, we propose a specialized method for detecting a specific category of small objects; birds. Particularly, we improve the features learned by the neck; the sub-network between the backbone and the prediction head, to learn more effective features with a hierarchical design. We employ Swin Transformer to upsample the image features. Moreover, we change the shifted window size for adapting to small objects. Experiments show that the proposed Swin Transformer-based neck combined with CenterNet can lead to good performance by changing the window sizes. We further find that smaller window sizes (default 2) benefit mAPs for small object detection.

Method: Uses language-conditioned diffusion model for frame colorization with automatically generated object masks and textual prompts. Employs optical flow (RAFT) for temporal stability by warping color from previous frames, plus a correction step to fix inconsistencies.

Result: Achieves state-of-the-art performance on DAVIS30 and VIDEVO20 benchmarks in colorization accuracy (PSNR) and visual realism (Colorfulness, CDC).

Conclusion: Demonstrates efficacy of automated prompt-based guidance for consistent, high-quality video colorization without manual color input.

Abstract: Existing video colorization methods struggle with temporal flickering or demand extensive manual input. We propose a novel approach automating high-fidelity video colorization using rich semantic guidance derived from language and segmentation. We employ a language-conditioned diffusion model to colorize grayscale frames. Guidance is provided via automatically generated object masks and textual prompts; our primary automatic method uses a generic prompt, achieving state-of-the-art results without specific color input. Temporal stability is achieved by warping color information from previous frames using optical flow (RAFT); a correction step detects and fixes inconsistencies introduced by warping. Evaluations on standard benchmarks (DAVIS30, VIDEVO20) show our method achieves state-of-the-art performance in colorization accuracy (PSNR) and visual realism (Colorfulness, CDC), demonstrating the efficacy of automated prompt-based guidance for consistent video colorization.

Method: Conducted a large-scale study with 7 MLLMs and 5 VQA datasets, testing 48 distinct prompt format variations to analyze the impact of three key variation factors in prompt formatting.

Result: Multiple-choice VQA is highly sensitive to minor prompt format changes even when semantically neutral, and these biases persist independently of known order biases or the MLLM’s confidence in the correct answer. Existing bias mitigation strategies fail to address these newly identified biases.

Conclusion: Current MLLM evaluations for multiple-choice VQA are unreliable due to unexplored prompt format biases, highlighting the need for more robust evaluation methodologies that account for these formatting sensitivities.

Abstract: Multimodal Large Language Models (MLLMs) demonstrate strong capabilities in handling image-text inputs. A common way to assess this ability is through multiple-choice Visual Question Answering (VQA). Earlier works have already revealed that these benchmarks are sensitive to answer choice order, a limitation that can be mitigated through careful design. Yet, we highlight additional, unexplored biases in prompt formatting that question the reliability of current MLLM evaluations. Specifically, we identify three key variation factors in prompt formatting and analyze their impact through a large-scale study involving $\mathbf{\text{seven}}$ MLLMs and $\mathbf{\text{five}}$ VQA datasets, spanning $\mathbf{48}$ distinct $\mathbf{\text{prompt format variations}}$. Our findings reveal that multiple-choice VQA is highly sensitive to minor prompt format changes, even when these changes are semantically neutral. We further demonstrate that these biases persist independently of known order biases or the MLLM’s confidence in the correct answer. Finally, we demonstrate that existing bias mitigation strategies fail to address these newly identified biases.

Method: Proposes an alignment strategy that creatively uses NF invertibility: instead of regularizing the forward pass, aligns intermediate features of the generative (reverse) pass with representations from a powerful vision foundation model. Also introduces a training-free, test-time optimization algorithm for classification to better evaluate semantic knowledge.

Result: The approach accelerates NF training by over 3.3× while simultaneously improving both generative quality and classification accuracy. Achieves new state-of-the-art results for NFs on ImageNet 64×64 and 256×256.

Conclusion: The proposed alignment strategy effectively leverages NF invertibility to improve semantic representations, leading to faster training and better performance in both generation and classification tasks, establishing new SOTA for Normalizing Flows.

Abstract: Normalizing Flows (NFs) are a class of generative models distinguished by a mathematically invertible architecture, where the forward pass transforms data into a latent space for density estimation, and the reverse pass generates new samples from this space. This characteristic creates an intrinsic synergy between representation learning and data generation. However, the generative quality of standard NFs is limited by poor semantic representations from log-likelihood optimization. To remedy this, we propose a novel alignment strategy that creatively leverages the invertibility of NFs: instead of regularizing the forward pass, we align the intermediate features of the generative (reverse) pass with representations from a powerful vision foundation model, demonstrating superior effectiveness over naive alignment. We also introduce a novel training-free, test-time optimization algorithm for classification, which provides a more intrinsic evaluation of the NF’s embedded semantic knowledge. Comprehensive experiments demonstrate that our approach accelerates the training of NFs by over 3.3$\times$, while simultaneously delivering significant improvements in both generative quality and classification accuracy. New state-of-the-art results for NFs are established on ImageNet 64$\times$64 and 256$\times$256. Our code is available at https://github.com/MCG-NJU/FlowBack.

Method: INSIGHT combines: 1) hierarchical super-resolution to amplify subtle forensic cues without artifacts, 2) Grad-CAM driven multi-scale localization to reveal spatial regions with generative patterns, 3) CLIP-guided semantic alignment to map visual anomalies to human-interpretable descriptors, and 4) a vision-language model with structured ReAct + Chain-of-Thought prompting for consistent explanations, verified through dual-stage G-Eval + LLM-as-a-judge pipeline.

Result: Across diverse domains (animals, vehicles, abstract scenes), INSIGHT substantially improves both detection robustness and explanation quality under extreme degradation, outperforming prior detectors and black-box VLM baselines.

Conclusion: INSIGHT provides a practical path toward transparent, reliable AI-generated image forensics and represents a step forward in trustworthy multimodal content verification by addressing both detection robustness and interpretability challenges.

Abstract: The growing realism of AI-generated images produced by recent GAN and diffusion models has intensified concerns over the reliability of visual media. Yet, despite notable progress in deepfake detection, current forensic systems degrade sharply under real-world conditions such as severe downsampling, compression, and cross-domain distribution shifts. Moreover, most detectors operate as opaque classifiers, offering little insight into why an image is flagged as synthetic, undermining trust and hindering adoption in high-stakes settings. We introduce INSIGHT (Interpretable Neural Semantic and Image-based Generative-forensic Hallucination Tracing), a unified multimodal framework for robust detection and transparent explanation of AI-generated images, even at extremely low resolutions (16x16 - 64x64). INSIGHT combines hierarchical super-resolution for amplifying subtle forensic cues without inducing misleading artifacts, Grad-CAM driven multi-scale localization to reveal spatial regions indicative of generative patterns, and CLIP-guided semantic alignment to map visual anomalies to human-interpretable descriptors. A vision-language model is then prompted using a structured ReAct + Chain-of-Thought protocol to produce consistent, fine-grained explanations, verified through a dual-stage G-Eval + LLM-as-a-judge pipeline to minimize hallucinations and ensure factuality. Across diverse domains, including animals, vehicles, and abstract synthetic scenes, INSIGHT substantially improves both detection robustness and explanation quality under extreme degradation, outperforming prior detectors and black-box VLM baselines. Our results highlight a practical path toward transparent, reliable AI-generated image forensics and establish INSIGHT as a step forward in trustworthy multimodal content verification.

Method: AnchorFlow establishes a global latent anchor shared between source and target trajectories, enforcing coherence through a relaxed anchor-alignment loss and anchor-aligned update rule to ensure stable transformations.

Result: Experiments on Eval3DEdit benchmark show AnchorFlow consistently delivers semantically aligned and structurally robust edits across diverse editing types without mask supervision.

Conclusion: AnchorFlow enables more pronounced semantic modifications while preserving geometric fidelity through latent anchor consistency, making it an effective training-free 3D editing solution.

Abstract: Training-free 3D editing aims to modify 3D shapes based on human instructions without model finetuning. It plays a crucial role in 3D content creation. However, existing approaches often struggle to produce strong or geometrically stable edits, largely due to inconsistent latent anchors introduced by timestep-dependent noise during diffusion sampling. To address these limitations, we introduce AnchorFlow, which is built upon the principle of latent anchor consistency. Specifically, AnchorFlow establishes a global latent anchor shared between the source and target trajectories, and enforces coherence using a relaxed anchor-alignment loss together with an anchor-aligned update rule. This design ensures that transformations remain stable and semantically faithful throughout the editing process. By stabilizing the latent reference space, AnchorFlow enables more pronounced semantic modifications. Moreover, AnchorFlow is mask-free. Without mask supervision, it effectively preserves geometric fidelity. Experiments on the Eval3DEdit benchmark show that AnchorFlow consistently delivers semantically aligned and structurally robust edits across diverse editing types. Code is at https://github.com/ZhenglinZhou/AnchorFlow.

Method: Proposes RS-EoT (Remote Sensing Evidence-of-Thought), a language-driven iterative visual evidence-seeking paradigm. Implements SocraticAgent, a self-play multi-agent system that synthesizes reasoning traces via alternating cycles of reasoning and visual inspection. Uses two-stage progressive RL: first RL on fine-grained grounding tasks to enhance RS-EoT capabilities, then RL on RS VQA to generalize to broader understanding scenarios.

Result: RS-EoT achieves state-of-the-art performance on multiple RS VQA and grounding benchmarks. Analysis shows clear iterative cycles of reasoning and evidence seeking, confirming RS-EoT mitigates the Glance Effect and enables genuine evidence-grounded reasoning.

Conclusion: The proposed RS-EoT paradigm effectively addresses pseudo reasoning in remote sensing by enabling genuine evidence-grounded reasoning through iterative visual evidence seeking, overcoming the limitations of the Glance Effect in large-scale RS imagery analysis.

Abstract: Recent multimodal reasoning models, inspired by DeepSeek-R1, have significantly advanced vision-language systems. However, in remote sensing (RS) tasks, we observe widespread pseudo reasoning: models narrate the process of reasoning rather than genuinely reason toward the correct answer based on visual evidence. We attribute this to the Glance Effect, where a single, coarse perception of large-scale RS imagery results in incomplete understanding and reasoning based on linguistic self-consistency instead of visual evidence. To address this, we propose RS-EoT (Remote Sensing Evidence-of-Thought), a language-driven, iterative visual evidence-seeking paradigm. To instill this paradigm, we propose SocraticAgent, a self-play multi-agent system that synthesizes reasoning traces via alternating cycles of reasoning and visual inspection. To enhance and generalize these patterns, we propose a two-stage progressive RL strategy: first, RL on fine-grained Grounding tasks to enhance RS-EoT capabilities, followed by RL on RS VQA to generalize to broader understanding scenarios. Experiments show RS-EoT achieves state-of-the-art performance on multiple RS VQA and grounding benchmarks. Analyses reveal clear iterative cycles of reasoning and evidence seeking, confirming RS-EoT mitigates the Glance Effect and enables genuine evidence-grounded reasoning. Our code, data, and models are available at https://geox-lab.github.io/Asking_like_Socrates

Method: Created a high-fidelity synthetic dataset with 400K synchronized frames across diverse scenes (urban, suburbs, forests, coastal) and weather conditions, featuring multiple UAV models and five modalities (RGB, IR, LiDAR, Radar, DVS). Each frame provides 2D/3D bounding boxes, 6-DoF poses, and instance-level annotations.

Result: UAV-MM3D enables core UAV perception tasks including 3D detection, pose estimation, target tracking, and short-term trajectory forecasting. The paper also proposes LGFusionNet (LiDAR-guided multimodal fusion baseline) and a dedicated UAV trajectory prediction baseline for benchmarking.

Conclusion: UAV-MM3D offers a public benchmark with controllable simulation environment, comprehensive scenario coverage, and rich annotations for advancing 3D perception of UAVs in complex low-altitude environments.

Abstract: Accurate perception of UAVs in complex low-altitude environments is critical for airspace security and related intelligent systems. Developing reliable solutions requires large-scale, accurately annotated, and multimodal data. However, real-world UAV data collection faces inherent constraints due to airspace regulations, privacy concerns, and environmental variability, while manual annotation of 3D poses and cross-modal correspondences is time-consuming and costly. To overcome these challenges, we introduce UAV-MM3D, a high-fidelity multimodal synthetic dataset for low-altitude UAV perception and motion understanding. It comprises 400K synchronized frames across diverse scenes (urban areas, suburbs, forests, coastal regions) and weather conditions (clear, cloudy, rainy, foggy), featuring multiple UAV models (micro, small, medium-sized) and five modalities - RGB, IR, LiDAR, Radar, and DVS (Dynamic Vision Sensor). Each frame provides 2D/3D bounding boxes, 6-DoF poses, and instance-level annotations, enabling core tasks related to UAVs such as 3D detection, pose estimation, target tracking, and short-term trajectory forecasting. We further propose LGFusionNet, a LiDAR-guided multimodal fusion baseline, and a dedicated UAV trajectory prediction baseline to facilitate benchmarking. With its controllable simulation environment, comprehensive scenario coverage, and rich annotations, UAV3D offers a public benchmark for advancing 3D perception of UAVs.

Method: Builds on DiffPortrait360 architecture with two key components: Style Appearance Module for style-content disentanglement and Style Fusion Attention for adaptive balance between structure preservation and stylization fidelity. Uses 3D GAN-generated multi-view dataset for fine-tuning and temperature-based key scaling for stylization intensity control.

Result: Outperforms state-of-the-art GAN- and diffusion-based stylization methods on FFHQ and RenderMe360 datasets across challenging style domains, achieving superior style quality with multi-view consistency.

Conclusion: DiffStyle360 provides an efficient, flexible solution for 3D head stylization that eliminates per-style training requirements while maintaining identity preservation and high-quality artistic results across diverse domains.

Abstract: 3D head stylization has emerged as a key technique for reimagining realistic human heads in various artistic forms, enabling expressive character design and creative visual experiences in digital media. Despite the progress in 3D-aware generation, existing 3D head stylization methods often rely on computationally expensive optimization or domain-specific fine-tuning to adapt to new styles. To address these limitations, we propose DiffStyle360, a diffusion-based framework capable of producing multi-view consistent, identity-preserving 3D head stylizations across diverse artistic domains given a single style reference image, without requiring per-style training. Building upon the 3D-aware DiffPortrait360 architecture, our approach introduces two key components: the Style Appearance Module, which disentangles style from content, and the Style Fusion Attention mechanism, which adaptively balances structure preservation and stylization fidelity in the latent space. Furthermore, we employ a 3D GAN-generated multi-view dataset for robust fine-tuning and introduce a temperaturebased key scaling strategy to control stylization intensity during inference. Extensive experiments on FFHQ and RenderMe360 demonstrate that DiffStyle360 achieves superior style quality, outperforming state-of-the-art GAN- and diffusion-based stylization methods across challenging style domains.

Method: WUKONG leverages flow-based transformers’ generative prior for 3D transitions, formulates morphing as an optimal transport barycenter problem for smooth shape transitions, uses sequential initialization to prevent geometric distortions, and employs similarity-guided semantic consistency for texture preservation with selective high-frequency detail retention.

Result: Extensive evaluations show WUKONG significantly outperforms state-of-the-art methods, achieving superior results across diverse geometry and texture variations with high-fidelity transitions and rich texture details.

Conclusion: WUKONG provides an effective training-free framework for high-fidelity textured 3D morphing that overcomes limitations of conventional methods, offering better generalization, reduced preprocessing requirements, and superior quality results.

Abstract: We present WUKONG, a novel training-free framework for high-fidelity textured 3D morphing that takes a pair of source and target prompts (image or text) as input. Unlike conventional methods – which rely on manual correspondence matching and deformation trajectory estimation (limiting generalization and requiring costly preprocessing) – WUKONG leverages the generative prior of flow-based transformers to produce high-fidelity 3D transitions with rich texture details. To ensure smooth shape transitions, we exploit the inherent continuity of flow-based generative processes and formulate morphing as an optimal transport barycenter problem. We further introduce a sequential initialization strategy to prevent abrupt geometric distortions and preserve identity coherence. For faithful texture preservation, we propose a similarity-guided semantic consistency mechanism that selectively retains high-frequency details and enables precise control over blending dynamics. This avoids common artifacts like oversmoothing while maintaining semantic fidelity. Extensive quantitative and qualitative evaluations demonstrate that WUKONG significantly outperforms state-of-the-art methods, achieving superior results across diverse geometry and texture variations.

Method: Freeze the decoder (handles view matching) and fine-tune only the image encoder using a custom LoRA adapter that distills geometric details from a strong monocular teacher model on large unlabeled datasets.

Result: Fine-tuned models consistently deliver sharper boundaries, recover complex structures, and achieve higher geometric accuracy in both single- and multi-view settings, with minimal overhead (only tiny LoRA weights added).

Conclusion: Fin3R is a simple, effective, and general fine-tuning method that significantly improves geometric accuracy of feed-forward 3D reconstruction models while maintaining efficiency.

Abstract: We present Fin3R, a simple, effective, and general fine-tuning method for feed-forward 3D reconstruction models. The family of feed-forward reconstruction model regresses pointmap of all input images to a reference frame coordinate system, along with other auxiliary outputs, in a single forward pass. However, we find that current models struggle with fine geometry and robustness due to (\textit{i}) the scarcity of high-fidelity depth and pose supervision and (\textit{ii}) the inherent geometric misalignment from multi-view pointmap regression. Fin3R jointly tackles two issues with an extra lightweight fine-tuning step. We freeze the decoder, which handles view matching, and fine-tune only the image encoder-the component dedicated to feature extraction. The encoder is enriched with fine geometric details distilled from a strong monocular teacher model on large, unlabeled datasets, using a custom, lightweight LoRA adapter. We validate our method on a wide range of models, including DUSt3R, MASt3R, CUT3R, and VGGT. The fine-tuned models consistently deliver sharper boundaries, recover complex structures, and achieve higher geometric accuracy in both single- and multi-view settings, while adding only the tiny LoRA weights, which leave test-time memory and latency virtually unchanged. Project page: \href{http://visual-ai.github.io/fin3r}{https://visual-ai.github.io/fin3r}

Method: Proposes SkeletonAgent with two cooperative agents: Questioner identifies confused action classes to provide context to LLM, and Selector parses LLM responses to extract joint-level constraints for finer-grained cross-modal alignment with the recognizer.

Result: Comprehensive evaluations on five benchmarks (NTU RGB+D, NTU RGB+D 120, Kinetics-Skeleton, FineGYM, and UAV-Human) show SkeletonAgent consistently outperforms state-of-the-art methods.

Conclusion: SkeletonAgent successfully bridges LLMs and recognition models through cooperative agents, enabling more targeted discriminative guidance and achieving superior performance across multiple skeleton-based action recognition benchmarks.

Abstract: Recent advances in skeleton-based action recognition increasingly leverage semantic priors from Large Language Models (LLMs) to enrich skeletal representations. However, the LLM is typically queried in isolation from the recognition model and receives no performance feedback. As a result, it often fails to deliver the targeted discriminative cues critical to distinguish similar actions. To overcome these limitations, we propose SkeletonAgent, a novel framework that bridges the recognition model and the LLM through two cooperative agents, i.e., Questioner and Selector. Specifically, the Questioner identifies the most frequently confused classes and supplies them to the LLM as context for more targeted guidance. Conversely, the Selector parses the LLM’s response to extract precise joint-level constraints and feeds them back to the recognizer, enabling finer-grained cross-modal alignment. Comprehensive evaluations on five benchmarks, including NTU RGB+D, NTU RGB+D 120, Kinetics-Skeleton, FineGYM, and UAV-Human, demonstrate that SkeletonAgent consistently outperforms state-of-the-art benchmark methods. The code is available at https://github.com/firework8/SkeletonAgent.

Method: Unified framework with Dynamic Concept Fusion (DCF) for class-adaptive prompts and Adversarial Boundary Forging (ABF) for precise decision margins using PGD-style perturbations, trained under Concept-Boundary Loss with semantic-spatial regularizers.

Result: State-of-the-art performance on MVTec-AD and VisA datasets for few-shot multi-class anomaly detection.

Conclusion: ABounD effectively addresses data scarcity in industrial anomaly detection by integrating semantic concept learning with adversarial boundary shaping, achieving precise decision boundaries that closely follow normal data while maintaining flexibility and robust semantic alignment.

Abstract: Few-shot multi-class industrial anomaly detection remains a challenging task. Vision-language models need to be both category-adaptive and sharply discriminative, yet data scarcity often blurs the boundary between normal and abnormal states. This ambiguity leads to missed subtle defects and the rejection of atypical normal samples. We propose ABounD, an Adversarial Boundary-Driven few-shot learning for multi-class anomaly detection, which is a unified learning framework that integrates semantic concept learning with decision boundary shaping. The Dynamic Concept Fusion (DCF) module produces class-adaptive prompts by fusing generalizable priors with class-specific cues, conditioned on image features. Meanwhile, Adversarial Boundary Forging (ABF) sculpts a more precise decision margin by generating boundary-level fence features via PGD-style perturbations. Training is conducted in a single stage under a Concept-Boundary Loss, where ABF provides the main supervisory signal and semantic-spatial regularizers stabilize the optimization. This synergy yields a decision boundary that closely follows normal data while preserving flexibility and robust semantic alignment. Experiments on MVTec-AD and VisA datasets demonstrate state-of-the-art performance in the task of few-shot multi-class anomaly detection.

Method: Proposes FauxNet, a novel network based on pre-trained Visual Speech Recognition (VSR) features. Extracts temporal VSR features from videos to identify and segregate real videos from manipulated ones. Focuses on zero-shot detection (generalizable detection). Also introduces two new datasets: Authentica-Vox and Authentica-HDTF with ~38,000 real and fake videos created with six recent deepfake generation techniques.

Result: FauxNet consistently outperforms state-of-the-art methods in zero-shot detection setting. The network is also able to attribute deepfakes by distinguishing between different generation techniques. Extensive analysis on Authentica datasets and FaceForensics++ demonstrates FauxNet’s superiority.

Conclusion: FauxNet provides an effective solution for deepfake detection using VSR features, achieving superior zero-shot performance and attribution capabilities. The release of Authentica datasets will facilitate further research in this critical area.

Abstract: Deepfake generation has witnessed remarkable progress, contributing to highly realistic generated images, videos, and audio. While technically intriguing, such progress has raised serious concerns related to the misuse of manipulated media. To mitigate such misuse, robust and reliable deepfake detection is urgently needed. Towards this, we propose a novel network FauxNet, which is based on pre-trained Visual Speech Recognition (VSR) features. By extracting temporal VSR features from videos, we identify and segregate real videos from manipulated ones. The holy grail in this context has to do with zero-shot detection, i.e., generalizable detection, which we focus on in this work. FauxNet consistently outperforms the state-of-the-art in this setting. In addition, FauxNet is able to attribute - distinguish between generation techniques from which the videos stem. Finally, we propose new datasets, referred to as Authentica-Vox and Authentica-HDTF, comprising about 38,000 real and fake videos in total, the latter created with six recent deepfake generation techniques. We provide extensive analysis and results on the Authentica datasets and FaceForensics++, demonstrating the superiority of FauxNet. The Authentica datasets will be made publicly available.

Method: Benchmarked four ML architectures (U-Nets, visual transformers, CNNs, MDNs) for multi-class state recognition in double-QD charge-stability diagrams using synthetic and experimental data across different data budgets and normalization schemes.

Result: U-Nets and ViTs achieved highest MSE scores on synthetic data (>0.98) but failed to generalize to experimental data; MDNs were computationally efficient but had lower performance; CNNs offered best trade-off with strong accuracy using 100x fewer parameters than U-Nets/ViTs.

Conclusion: CNNs with min-max normalization are the most practical approach for quantum dot charge-stability diagram analysis, balancing accuracy, generalization, and computational efficiency for scalable quantum processor applications.

Abstract: Semiconductor quantum dots (QDs) are a leading platform for scalable quantum processors. However, scaling to large arrays requires reliable, automated tuning strategies for devices’ bootstrapping, calibration, and operation, with many tuning aspects depending on accurately identifying QD device states from charge-stability diagrams (CSDs). In this work, we present a comprehensive benchmarking study of four modern machine learning (ML) architectures for multi-class state recognition in double-QD CSDs. We evaluate their performance across different data budgets and normalization schemes using both synthetic and experimental data. We find that the more resource-intensive models – U-Nets and visual transformers (ViTs) – achieve the highest MSE score (defined as $1-\mathrm{MSE}$) on synthetic data (over $0.98$) but fail to generalize to experimental data. MDNs are the most computationally efficient and exhibit highly stable training, but with substantially lower peak performance. CNNs offer the most favorable trade-off on experimental CSDs, achieving strong accuracy with two orders of magnitude fewer parameters than the U-Nets and ViTs. Normalization plays a nontrivial role: min-max scaling generally yields higher MSE scores but less stable convergence, whereas z-score normalization produces more predictable training dynamics but at reduced accuracy for most models. Overall, our study shows that CNNs with min-max normalization are a practical approach for QD CSDs.

Method: Created IntentHQ benchmark with 5168 videos annotated with 23 fine-grained intent categories. Used supervised and self-supervised multi-modality models integrating spatio-temporal video features, audio processing, and text analysis for intent recognition.

Result: Developed a streamlined model capable of differentiating between a wide range of intent categories including financial fraud, indirect marketing, political propaganda, and fear mongering.

Conclusion: The paper introduces a paradigm shift from authenticity verification to contextual understanding of videos, providing a comprehensive framework for analyzing the motivations and goals behind manipulated content.

Abstract: The rapid advancement of generative models has led to increasingly realistic deepfake videos, posing significant societal and security risks. While existing detection methods focus on distinguishing real from fake videos, such approaches fail to address a fundamental question: What is the intent behind a manipulated video? Towards addressing this question, we introduce IntentHQ: a new benchmark for human-centered intent analysis, shifting the paradigm from authenticity verification to contextual understanding of videos. IntentHQ consists of 5168 videos that have been meticulously collected and annotated with 23 fine-grained intent-categories, including “Financial fraud”, “Indirect marketing”, “Political propaganda”, as well as “Fear mongering”. We perform intent recognition with supervised and self-supervised multi-modality models that integrate spatio-temporal video features, audio processing, and text analysis to infer underlying motivations and goals behind videos. Our proposed model is streamlined to differentiate between a wide range of intent-categories.

Method: Three key components: (1) teacher supervision with validation-time text detection to filter/denoise training signals, (2) multi-module validator (VAL) enforcing answer correctness and geometric consistency with pixel-level error feedback, (3) two-stage student training: first learns from validated CoT traces, then undergoes iterative refinement driven by VAL feedback.

Result: Student (Gemma-3 12B) achieves 91.4% ANLS and 82.4% mAP on DocVQA as a pure VLM requiring no text detection or OCR at inference. Validated feedback contributes 6.3 mAP gain, iterative refinement accounts for 9.7 mAP improvement.

Conclusion: DocVAL successfully transfers spatial reasoning ability from large teachers to deployable students, releasing 95k high-quality, validator-verified CoT traces to advance spatial reasoning research in document understanding.

Abstract: Document visual question answering (DocVQA) requires models to jointly reason over textual content and spatial layout, yet current systems exhibit a sharp accuracy–efficiency trade-off: large teacher models achieve strong grounding but are too expensive for deployment, while compact students suffer substantial drops in localization performance. We propose DocVAL, a validated chain-of-thought distillation framework that transfers the spatial reasoning ability of a large teacher into a deployable student VLM through three key components: (1) teacher supervision with validation-time text detection to filter and denoise training signals, (2) a multi-module validator (VAL) that enforces answer correctness and geometric consistency while producing fine-grained, pixel-level error feedback, and (3) a two-stage student training scheme that first learns from validated CoT traces and then undergoes iterative refinement driven by VAL feedback. Our student (Gemma-3 12B) achieves 91.4% ANLS and 82.4% mAP on DocVQA as a pure VLM requiring no text detection or OCR at inference. Extensive ablations demonstrate that validated feedback contributes 6.3 mAP gain and iterative refinement accounts for 9.7 mAP improvement. We release 95k high-quality, validator-verified CoT traces to advance spatial reasoning research in document understanding.

Method: ITS3D formulates the task as an optimization problem to find optimal Gaussian noise inputs. It uses a verifier-guided search algorithm with three key techniques: 1) Gaussian normalization to stabilize search by correcting distribution shifts, 2) SVD-based compression to reduce high-dimensional search space complexity, and 3) singular space reset mechanism to prevent local minima convergence.

Result: Extensive experiments demonstrate that ITS3D enhances text-to-3D generation quality, showing the potential of computationally efficient search methods in generative processes.

Conclusion: ITS3D successfully improves 3D generation quality through inference-time optimization without retraining, offering an efficient approach to enhance diffusion model performance through intelligent noise input selection.

Abstract: We explore inference-time scaling in text-guided 3D diffusion models to enhance generative quality without additional training. To this end, we introduce ITS3D, a framework that formulates the task as an optimization problem to identify the most effective Gaussian noise input. The framework is driven by a verifier-guided search algorithm, where the search algorithm iteratively refines noise candidates based on verifier feedback. To address the inherent challenges of 3D generation, we introduce three techniques for improved stability, efficiency, and exploration capability. 1) Gaussian normalization is applied to stabilize the search process. It corrects distribution shifts when noise candidates deviate from a standard Gaussian distribution during iterative updates. 2) The high-dimensional nature of the 3D search space increases computational complexity. To mitigate this, a singular value decomposition-based compression technique is employed to reduce dimensionality while preserving effective search directions. 3) To further prevent convergence to suboptimal local minima, a singular space reset mechanism dynamically updates the search space based on diversity measures. Extensive experiments demonstrate that ITS3D enhances text-to-3D generation quality, which shows the potential of computationally efficient search methods in generative processes. The source code is available at https://github.com/ZhenglinZhou/ITS3D.

Method: CoT4AD introduces Chain-of-Thought reasoning to enhance both numerical and causal reasoning in VLMs. It integrates visual observations and language instructions for semantic reasoning, scene understanding, and trajectory planning. During training, it explicitly models a perception-question-prediction-action CoT to align reasoning space with action space. During inference, it performs implicit CoT reasoning for consistent numerical reasoning and robust decision-making.

Result: Extensive experiments on both real-world (nuScenes) and simulated (Bench2Drive) benchmarks demonstrate that CoT4AD achieves state-of-the-art performance in both open-loop and closed-loop evaluations.

Conclusion: CoT4AD successfully addresses the limitations of existing VLA models by incorporating Chain-of-Thought reasoning, enabling enhanced numerical and causal reasoning capabilities for autonomous driving in complex scenarios.

Abstract: Vision-Language-Action (VLA) models have recently attracted growing attention in end-to-end autonomous driving for their strong reasoning capabilities and rich world knowledge. However, existing VLAs often suffer from limited numerical reasoning ability and overly simplified input-output mappings, which hinder their performance in complex driving scenarios requiring step-by-step causal reasoning. To address these challenges, we propose CoT4AD, a novel VLA framework that introduces Chain-of-Thought (CoT) reasoning for autonomous driving to enhance both numerical and causal reasoning in Vision-Language Models (VLMs). CoT4AD integrates visual observations and language instructions to perform semantic reasoning, scene understanding, and trajectory planning. During training, it explicitly models a perception-question-prediction-action CoT to align the reasoning space with the action space across multiple driving tasks. During inference, it performs implicit CoT reasoning to enable consistent numerical reasoning and robust decision-making in dynamic environments. Extensive experiments on both real-world and simulated benchmarks, including nuScenes and Bench2Drive, demonstrate that CoT4AD achieves state-of-the-art performance in both open-loop and closed-loop evaluations. Code will be released upon paper acceptance.

Method: Separates static background from dynamic fire using dense multi-view stereo with monocular depth priors. Initializes fire as 3D flow field from fused per-view dense optical flow projections. Uses 3D Gaussians with lifetime and linear velocity encoding to capture high-frequency features. Employs custom hardware synchronization for sub-frame temporal alignment across cameras.

Result: Quantitative and qualitative validations across numerous reconstruction experiments demonstrate robust performance for diverse and challenging real fire scenarios using affordable commodity hardware.

Conclusion: The proposed method successfully reconstructs dynamic fire in 3D from limited camera views (three views) using Gaussian-based representation, overcoming the inherent challenges of fire reconstruction while maintaining practical hardware requirements.

Abstract: We propose a method to reconstruct dynamic fire in 3D from a limited set of camera views with a Gaussian-based spatiotemporal representation. Capturing and reconstructing fire and its dynamics is highly challenging due to its volatile nature, transparent quality, and multitude of high-frequency features. Despite these challenges, we aim to reconstruct fire from only three views, which consequently requires solving for under-constrained geometry. We solve this by separating the static background from the dynamic fire region by combining dense multi-view stereo images with monocular depth priors. The fire is initialized as a 3D flow field, obtained by fusing per-view dense optical flow projections. To capture the high frequency features of fire, each 3D Gaussian encodes a lifetime and linear velocity to match the dense optical flow. To ensure sub-frame temporal alignment across cameras we employ a custom hardware synchronization pattern – allowing us to reconstruct fire with affordable commodity hardware. Our quantitative and qualitative validations across numerous reconstruction experiments demonstrate robust performance for diverse and challenging real fire scenarios.

Method: Systematic evaluation using controlled maze-solving benchmark with fully visual reasoning rules, tunable difficulty by grid size, and automatically generated intermediate steps. Tested three CoT formats (Language CoT, Grounding CoT with spatial coordinates, Visual CoT with image manipulations) on Qwen2.5-VL-7B under standard SFT-then-RL pipeline.

Result: Visual and longer CoT mainly accelerate convergence but don’t lift final performance ceiling; concise CoT with essential grounding steps outperforms longer traces; minimal grounding CoT generalizes best across different maze sizes. Insights validated on other vision-centric tasks.

Conclusion: Reveals “short is long” effect where concise, grounding-focused CoT is most effective for generalizable visual reasoning, providing practical guidance for constructing better SFT datasets.

Abstract: We study how different Chain-of-Thought (CoT) designs affect the acquisition of the generalizable visual reasoning ability in vision-language models (VLMs). While CoT data, especially long or visual CoT such as “think with image”, has been widely used to supervise intermediate reasoning, it remains unclear why specific CoT designs help and which ones truly support generalizable reasoning. To systematically evaluate this, we focus on a controlled maze-solving benchmark where reasoning rules are fully visual, difficulty can be tuned by grid size, and all the intermediate steps can be automatically generated. Using Qwen2.5-VL-7B under a standard SFT-then-RL pipeline, we compare three representative CoT formats: Language CoT, Grounding CoT (with spatial coordinate trajectories), and Visual CoT (with image manipulations). Our experiments reveal that visual and longer CoT mainly accelerate convergence but do not lift the final performance ceiling; concise CoT containing only essential grounding steps outperforms longer traces; and, strikingly, CoT retaining only the minimal grounding results generalizes best across different maze sizes. We further validate these insights on other vision-centric tasks. These findings highlight a “short is long” effect and provide practical guidance for constructing more generalizable SFT datasets for visual reasoning.

Method: 1) RoadSceneBench benchmark emphasizing relational understanding and structural consistency; 2) HRRP-T (Hierarchical Relational Reward Propagation with Temporal Consistency) training framework for VLMs that uses adaptive reward signals to promote spatial coherence and semantic alignment throughout reasoning.

Result: Achieves state-of-the-art performance across diverse road configurations. The benchmark provides a compact yet powerful foundation for studying mid-level road semantics and structure-aware autonomous perception.

Conclusion: RoadSceneBench addresses the gap in reasoning-focused benchmarks for autonomous driving, enabling models to move beyond static recognition toward geometry-aware and temporally consistent reasoning of road scenes.

Abstract: Understanding mid-level road semantics, which capture the structural and contextual cues that link low-level perception to high-level planning, is essential for reliable autonomous driving and digital map construction. However, existing benchmarks primarily target perception tasks such as detection or segmentation, overlooking the reasoning capabilities required to infer road topology and dynamic scene structure. To address this gap, we present RoadSceneBench, a lightweight yet information-rich benchmark designed to evaluate and advance visual reasoning in complex road environments. Unlike large-scale perception datasets, RoadSceneBench emphasizes relational understanding and structural consistency, encouraging models to capture the underlying logic of real-world road scenes. Furthermore, to enhance reasoning reliability, we propose Hierarchical Relational Reward Propagation with Temporal Consistency (HRRP-T), a training framework for Vision-Language Models (VLMs) in which reward signals adaptively promote spatial coherence and semantic alignment throughout the reasoning process. This paradigm enables models to move beyond static recognition toward geometry-aware and temporally consistent reasoning. Extensive experiments demonstrate that our method achieves state-of-the-art performance across diverse road configurations. RoadSceneBench thus provides a compact yet powerful foundation for studying mid-level road semantics and fostering structure-aware autonomous perception. Our dataset is available at https://github.com/XiyanLiu/RoadSceneBench.

Method: HarmoCLIP introduces explicit fine-grained semantic supervision that directly aligns textual segments with corresponding visual regions. It also employs a Region-Language Alignment supervision strategy to strengthen local representation without compromising global semantic consistency.

Result: Achieves state-of-the-art performance with up to 69.78% improvement on global retrieval tasks and 3.2% improvement in Top-1 accuracy on bounding-box classification. Provides balanced, efficient, plug-and-play solution to the global-local trade-off.

Conclusion: HarmoCLIP successfully harmonizes global and region representations in CLIP, overcoming the trade-off between local perception and global coherence while maintaining strong performance on both types of tasks.

Abstract: Contrastive Language-Image Pre-training (CLIP) has demonstrated remarkable generalization ability and strong performance across a wide range of vision-language tasks. However, due to the lack of region-level supervision, CLIP exhibits limited fine-grained semantic understanding. Although several methods attempt to mitigate this issue, they unintentionally disrupt the global alignment, resulting in a persistent trade-off where improving local perception simultaneously degrades global coherence. In this paper, we propose HarmoCLIP, a novel framework designed to harmonize global and region representations within CLIP. We first identify that the absence of direct alignment between local textual and visual semantics is the fundamental cause of the trade-off. To address this, HarmoCLIP introduces an explicit fine-grained semantic supervision term that directly aligns textual segments with their corresponding visual regions, effectively bridging the image region space and the textual space. To further strengthen the representation capability at the local level, our method introduces a novel Region-Language Alignment supervision strategy that promotes fine-grained semantic learning without compromising global semantic consistency. Extensive experiments demonstrate that HarmoCLIP achieves state-of-the-art (improvement up to 69.78%) performance on the global task of retrieval and yields a substantial 3.2% improvement in Top-1 accuracy on the region task of bounding-box classification, consistently outperforming prior approaches while providing a balanced, efficient, and plug-and-play solution to the global-local trade-off in CLIP. Code is available at https://github.com/Erosist/HarmoCLIP.

Method: 1) Local-Global Hybrid Perspective (LHP) module integrated with Vision-Language Model (VLM) for fine+coarse-grained features; 2) Unified Image-Text (UIT) model with ITC, ITM, MLM, and MIM losses; 3) Novel iterative ensemble strategy; 4) Feature selection algorithm guided by LHP.

Result: Achieves state-of-the-art performance on PAB dataset with 9.70% improvement in R@1, 1.77% improvement in R@5, and 1.01% improvement in R@10 compared to previous work.

Conclusion: The proposed approach effectively addresses fine-grained feature optimization in text-based person anomaly retrieval, demonstrating significant performance improvements through hybrid perspective modeling, multi-loss training, and innovative ensemble strategies.

Abstract: Text-based person anomaly retrieval has emerged as a challenging task, with most existing approaches relying on complex deep-learning techniques. This raises a research question: How can the model be optimized to achieve greater fine-grained features? To address this, we propose a Local-Global Hybrid Perspective (LHP) module integrated with a Vision-Language Model (VLM), designed to explore the effectiveness of incorporating both fine-grained features alongside coarse-grained features. Additionally, we investigate a Unified Image-Text (UIT) model that combines multiple objective loss functions, including Image-Text Contrastive (ITC), Image-Text Matching (ITM), Masked Language Modeling (MLM), and Masked Image Modeling (MIM) loss. Beyond this, we propose a novel iterative ensemble strategy, by combining iteratively instead of using model results simultaneously like other ensemble methods. To take advantage of the superior performance of the LHP model, we introduce a novel feature selection algorithm based on its guidance, which helps improve the model’s performance. Extensive experiments demonstrate the effectiveness of our method in achieving state-of-the-art (SOTA) performance on PAB dataset, compared with previous work, with a 9.70% improvement in R@1, 1.77% improvement in R@5, and 1.01% improvement in R@10.

Method: 1) Created GazeTrack dataset with high-precision equipment covering diverse ethnicities, ages, and visual acuity; 2) Shape error regularization for pupil ellipse fitting; 3) Paper-unfolding-like coordinate transformation for gaze vector prediction; 4) Gaze vector generation model with optimized computational complexity.

Result: Achieved reduced gaze angle error compared to other methods while maintaining lower computational complexity, demonstrating improved accuracy for spatial computing applications.

Conclusion: The proposed framework with novel regularization and transformation methods provides a more accurate and efficient gaze tracking solution suitable for VR/AR spatial computing applications, validated on a diverse benchmark dataset.

Abstract: Eye tracking has become increasingly important in virtual and augmented reality applications; however, the current gaze accuracy falls short of meeting the requirements for spatial computing. We designed a gaze collection framework and utilized high-precision equipment to gather the first precise benchmark dataset, GazeTrack, encompassing diverse ethnicities, ages, and visual acuity conditions for pupil localization and gaze tracking. We propose a novel shape error regularization method to constrain pupil ellipse fitting and train on open-source datasets, enhancing semantic segmentation and pupil position prediction accuracy. Additionally, we invent a novel coordinate transformation method similar to paper unfolding to accurately predict gaze vectors on the GazeTrack dataset. Finally, we built a gaze vector generation model that achieves reduced gaze angle error with lower computational complexity compared to other methods.

Method: Systematic studies on frequency, spatial, and token perspectives reveal DINOv3’s tendency to use global, low-frequency structures as authenticity cues. A simple training-free token-ranking strategy selects authenticity-relevant tokens, followed by a lightweight linear probe.

Result: The token subset consistently improves detection accuracy across all evaluated datasets. DINOv3 exhibits strong cross-generator detection capability without any fine-tuning.

Conclusion: Foundation models generalize across diverse generators by relying on transferable global authenticity cues rather than generator-specific artifacts. The approach provides an efficient, interpretable baseline for universal image forgery detection.

Abstract: As generative models become increasingly diverse and powerful, cross-generator detection has emerged as a new challenge. Existing detection methods often memorize artifacts of specific generative models rather than learning transferable cues, leading to substantial failures on unseen generators. Surprisingly, this work finds that frozen visual foundation models, especially DINOv3, already exhibit strong cross-generator detection capability without any fine-tuning. Through systematic studies on frequency, spatial, and token perspectives, we observe that DINOv3 tends to rely on global, low-frequency structures as weak but transferable authenticity cues instead of high-frequency, generator-specific artifacts. Motivated by this insight, we introduce a simple, training-free token-ranking strategy followed by a lightweight linear probe to select a small subset of authenticity-relevant tokens. This token subset consistently improves detection accuracy across all evaluated datasets. Our study provides empirical evidence and a feasible hypothesis for understanding why foundation models generalize across diverse generators, offering a universal, efficient, and interpretable baseline for image forgery detection.

Method: Proposes Dimitra++ framework with a conditional Motion Diffusion Transformer (cMDT) that models facial motion sequences using 3D representation. The model is conditioned on two inputs: a reference facial image for appearance and an audio sequence for motion generation.

Result: Quantitative and qualitative experiments, plus user studies on VoxCeleb2 and CelebV-HQ datasets, show that Dimitra++ outperforms existing approaches in generating realistic talking heads with accurate lip motion, facial expressions, and head pose.

Conclusion: Dimitra++ successfully creates a streamlined framework for comprehensive audio-driven talking head generation that achieves state-of-the-art performance in producing realistic facial animations including lip sync, expressions, and head movements.

Abstract: We propose Dimitra++, a novel framework for audio-driven talking head generation, streamlined to learn lip motion, facial expression, as well as head pose motion. Specifically, we propose a conditional Motion Diffusion Transformer (cMDT) to model facial motion sequences, employing a 3D representation. The cMDT is conditioned on two inputs: a reference facial image, which determines appearance, as well as an audio sequence, which drives the motion. Quantitative and qualitative experiments, as well as a user study on two widely employed datasets, i.e., VoxCeleb2 and CelebV-HQ, suggest that Dimitra++ is able to outperform existing approaches in generating realistic talking heads imparting lip motion, facial expression, and head pose.

Method: Introduces SciPostGen dataset and Retrieval-Augmented Poster Layout Generation framework that retrieves layouts consistent with given papers and uses them as guidance for layout generation.

Result: Analysis shows paper structures correlate with layout element counts; retriever estimates layouts aligned with paper structures; framework generates layouts satisfying given constraints in both constrained and unconstrained conditions.

Conclusion: SciPostGen bridges the dataset gap for paper-to-poster layout understanding, enabling effective AI-assisted poster generation that maintains consistency with paper structures while accommodating creator constraints.

Abstract: As the number of scientific papers continues to grow, there is a demand for approaches that can effectively convey research findings, with posters serving as a key medium for presenting paper contents. Poster layouts determine how effectively research is communicated and understood, highlighting their growing importance. In particular, a gap remains in understanding how papers correspond to the layouts that present them, which calls for datasets with paired annotations at scale. To bridge this gap, we introduce SciPostGen, a large-scale dataset for understanding and generating poster layouts from scientific papers. Our analyses based on SciPostGen show that paper structures are associated with the number of layout elements in posters. Based on this insight, we explore a framework, Retrieval-Augmented Poster Layout Generation, which retrieves layouts consistent with a given paper and uses them as guidance for layout generation. We conducted experiments under two conditions: with and without layout constraints typically specified by poster creators. The results show that the retriever estimates layouts aligned with paper structures, and our framework generates layouts that also satisfy given constraints.

Method: Created benchmark data for dense text removal, then developed a method to model highly flexible mask profiles and learn their parameters using Bayesian optimization to find optimal mask shapes for text removal.

Result: Found that text-removal performance is vulnerable to mask profile perturbation, character-wise masks are optimal (not minimum cover of text regions), and precise mask tuning is essential for practical applications.

Conclusion: The research provides a user-friendly guideline for manual masking in text removal tasks and paves the way for better industrial applications of document image inpainting with dense text.

Abstract: The advent of generative models has dramatically improved the accuracy of image inpainting. In particular, by removing specific text from document images, reconstructing original images is extremely important for industrial applications. However, most existing methods of text removal focus on deleting simple scene text which appears in images captured by a camera in an outdoor environment. There is little research dedicated to complex and practical images with dense text. Therefore, we created benchmark data for text removal from images including a large amount of text. From the data, we found that text-removal performance becomes vulnerable against mask profile perturbation. Thus, for practical text-removal tasks, precise tuning of the mask shape is essential. This study developed a method to model highly flexible mask profiles and learn their parameters using Bayesian optimization. The resulting profiles were found to be character-wise masks. It was also found that the minimum cover of a text region is not optimal. Our research is expected to pave the way for a user-friendly guideline for manual masking.

Method: Domain Invariant Prompt Tuning (DIPT) learns multiple input tokens per domain separately, then averages them across domains to create domain-invariant prompts. A student model distills knowledge from PLIP’s text encoder using these prompts to align visual features with domain-invariant embeddings.

Result: Significant improvement in average F1-score over existing SOTA knowledge distillation approaches for domain generalization on histopathology datasets.

Conclusion: DIPT enables robust computational pathology model deployment in real-world clinical settings with heterogeneous data sources by learning domain-invariant representations.

Abstract: Domain generalization is critical in computational pathology (CPath) due to inherent domain shifts caused by variations in staining protocols, scanner devices, and imaging settings across clinical centers. Vision-language models (VLMs), such as PLIP-a pathology-tuned CLIP-trained on image-text pairs across diverse domains, serve as strong knowledge distillation sources. However, their zero-shot performance with predefined prompts remains limited due to sensitivity to prompt variations. Moreover, unlike natural images, histopathology centers lack semantic descriptors (e.g., ‘sketch’), making it difficult to define domain-specific prompts for clinical centers. This requires a data-driven approach for learning domain-specific and ultimately class-generic continuous prompts. We propose Domain Invariant Prompt Tuning (DIPT) for knowledge distillation process, a novel step that learns multiple input tokens for each domain. These tokens are trained separately for each domain and are averaged across domains, leading to domain-invariant prompts. Our student model then distills knowledge from PLIP’s text encoder by leveraging the prompts learned by DIPT. This leads to alignment of visual features with domain-invariant embeddings, enhancing generalization by training on multiple domains. Our method adds a significant improvement in average F1-score to existing state-of-the-art (SOTA) knowledge distillation approaches in domain generalization with histopathology datasets. This work helps the way of deploying robust CPath models in real-world clinical problems with heterogeneous data sources.

Method: Introduces Predictive Caching Scheduler Constraint (PCSC) to dynamically allocate cache quotas based on voxel stabilization patterns, and Spatiotemporal Stability Criterion (SSC) to select stable features for reuse using velocity magnitude and acceleration criteria.

Result: Achieves up to 27.12% inference speed-up and 54.8% reduction in FLOPs with minimal geometric quality degradation: only 2.48% increase in Chamfer Distance and 1.95% decrease in F-Score.

Conclusion: Fast3Dcache successfully accelerates 3D diffusion inference while maintaining geometric fidelity, overcoming limitations of existing caching methods that disrupt 3D structural consistency.

Abstract: Diffusion models have achieved impressive generative quality across modalities like 2D images, videos, and 3D shapes, but their inference remains computationally expensive due to the iterative denoising process. While recent caching-based methods effectively reuse redundant computations to speed up 2D and video generation, directly applying these techniques to 3D diffusion models can severely disrupt geometric consistency. In 3D synthesis, even minor numerical errors in cached latent features accumulate, causing structural artifacts and topological inconsistencies. To overcome this limitation, we propose Fast3Dcache, a training-free geometry-aware caching framework that accelerates 3D diffusion inference while preserving geometric fidelity. Our method introduces a Predictive Caching Scheduler Constraint (PCSC) to dynamically determine cache quotas according to voxel stabilization patterns and a Spatiotemporal Stability Criterion (SSC) to select stable features for reuse based on velocity magnitude and acceleration criterion. Comprehensive experiments show that Fast3Dcache accelerates inference significantly, achieving up to a 27.12% speed-up and a 54.8% reduction in FLOPs, with minimal degradation in geometric quality as measured by Chamfer Distance (2.48%) and F-Score (1.95%).

Method: Fine-tuned a pretrained three-channel DDPM on 11,020 screening mammograms to generate paired CC and MLO views. Evaluated three third-channel encodings: sum, absolute difference, and zero channel. Generated 500 synthetic image pairs per model. Used breast mask segmentation with IoU and DSC metrics, compared distributions against 2,500 real pairs using EMD and KS tests, and conducted visual Turing tests by a non-expert radiologist.

Result: Synthetic mammograms showed comparable IoU and DSC distributions to real images (EMD=0.020, KS=0.077). Models with sum or absolute difference encodings outperformed others in IoU and DSC (p<0.001). Generated views maintained cross-view consistency, with 6-8% of synthetic images showing artifacts consistent with training data.

Conclusion: Three-channel DDPMs can generate realistic and anatomically consistent dual-view mammograms, demonstrating promising applications for dataset augmentation in medical imaging.

Abstract: Purpose: This study aims to develop and evaluate a three channel denoising diffusion probabilistic model (DDPM) for synthesizing single breast dual view mammograms and to assess the impact of channel representations on image fidelity and cross view consistency. Materials and Methods: A pretrained three channel DDPM, sourced from Hugging Face, was fine tuned on a private dataset of 11020 screening mammograms to generate paired craniocaudal (CC) and mediolateral oblique (MLO) views. Three third channel encodings of the CC and MLO views were evaluated: sum, absolute difference, and zero channel. Each model produced 500 synthetic image pairs. Quantitative assessment involved breast mask segmentation using Intersection over Union (IoU) and Dice Similarity Coefficient (DSC), with distributional comparisons against 2500 real pairs using Earth Movers Distance (EMD) and Kolmogorov Smirnov (KS) tests. Qualitative evaluation included a visual Turing test by a non expert radiologist to assess cross view consistency and artifacts. Results: Synthetic mammograms showed IoU and DSC distributions comparable to real images, with EMD and KS values (0.020 and 0.077 respectively). Models using sum or absolute difference encodings outperformed others in IoU and DSC (p < 0.001), though distributions remained broadly similar. Generated CC and MLO views maintained cross view consistency, with 6 to 8 percent of synthetic images exhibiting artifacts consistent with those in the training data. Conclusion: Three channel DDPMs can generate realistic and anatomically consistent dual view mammograms with promising applications in dataset augmentation.

Method: Proposes Diff-ICMH framework with: 1) Generative priors for perceptual realism, 2) Semantic Consistency loss (SC loss) to ensure semantic fidelity, 3) Tag Guidance Module (TGM) that uses image-level tags to stimulate diffusion model capabilities with minimal bit rate overhead. The system supports multiple intelligent tasks through a single codec without task-specific adaptation.

Result: Extensive experiments demonstrate Diff-ICMH’s superiority and generalizability across diverse tasks while maintaining high visual quality for human perception. The framework achieves harmonization between machine and human vision objectives in image compression.

Conclusion: Diff-ICMH successfully bridges the gap between human and machine vision in image compression by leveraging generative priors and semantic consistency mechanisms, enabling a single codec to serve both objectives effectively while maintaining visual appeal and supporting multiple intelligent tasks.

Abstract: Image compression methods are usually optimized isolatedly for human perception or machine analysis tasks. We reveal fundamental commonalities between these objectives: preserving accurate semantic information is paramount, as it directly dictates the integrity of critical information for intelligent tasks and aids human understanding. Concurrently, enhanced perceptual quality not only improves visual appeal but also, by ensuring realistic image distributions, benefits semantic feature extraction for machine tasks. Based on this insight, we propose Diff-ICMH, a generative image compression framework aiming for harmonizing machine and human vision in image compression. It ensures perceptual realism by leveraging generative priors and simultaneously guarantees semantic fidelity through the incorporation of Semantic Consistency loss (SC loss) during training. Additionally, we introduce the Tag Guidance Module (TGM) that leverages highly semantic image-level tags to stimulate the pre-trained diffusion model’s generative capabilities, requiring minimal additional bit rates. Consequently, Diff-ICMH supports multiple intelligent tasks through a single codec and bitstream without any task-specific adaptation, while preserving high-quality visual experience for human perception. Extensive experimental results demonstrate Diff-ICMH’s superiority and generalizability across diverse tasks, while maintaining visual appeal for human perception. Code is available at: https://github.com/RuoyuFeng/Diff-ICMH.

Method: Introduces a Dual-UV representation with Core-UV and Shell-UV branches to map image features to canonical UV space, eliminating pose/framing effects. Builds a factorized synthetic data manifold combining 2D generative diversity with geometry-consistent 3D renderings. Uses a robust proxy-mesh tracker for stability under partial visibility.

Result: Achieves state-of-the-art head and upper-body reconstruction and competitive full-body results, trained only on half-body synthetic data. Demonstrates strong in-the-wild generalization through extensive experiments and analyses.

Conclusion: The unified framework successfully addresses key challenges in animatable 3D human avatar reconstruction through innovative representations, synthetic data generation, and robust tracking, enabling effective generalization to real-world scenarios.

Abstract: We present a unified framework for reconstructing animatable 3D human avatars from a single portrait across head, half-body, and full-body inputs. Our method tackles three bottlenecks: pose- and framing-sensitive feature representations, limited scalable data, and unreliable proxy-mesh estimation. We introduce a Dual-UV representation that maps image features to a canonical UV space via Core-UV and Shell-UV branches, eliminating pose- and framing-induced token shifts. We also build a factorized synthetic data manifold combining 2D generative diversity with geometry-consistent 3D renderings, supported by a training scheme that improves realism and identity consistency. A robust proxy-mesh tracker maintains stability under partial visibility. Together, these components enable strong in-the-wild generalization. Trained only on half-body synthetic data, our model achieves state-of-the-art head and upper-body reconstruction and competitive full-body results. Extensive experiments and analyses further validate the effectiveness of our approach.

Method: Extends MeshMAE self-supervised learning with a text-guided localization (TGL) module using CLIP text encoder to locate abutment areas. Pre-trains encoder on oral scan data to capture fine-grained features like implant dimensions and distances.

Result: Achieves 0.8%-12.85% IoU improvement over mainstream methods on a large abutment design dataset, demonstrating superior localization accuracy.

Conclusion: TCEAD shows strong potential for automated dental abutment design by effectively combining text guidance with mesh-based learning for precise abutment area localization.

Abstract: The abutment is an important part of artificial dental implants, whose design process is time-consuming and labor-intensive. Long-term use of inappropriate dental implant abutments may result in implant complications, including peri-implantitis. Using artificial intelligence to assist dental implant abutment design can quickly improve the efficiency of abutment design and enhance abutment adaptability. In this paper, we propose a text condition embedded abutment design framework (TCEAD), the novel automated abutment design solution available in literature. The proposed study extends the self-supervised learning framework of the mesh mask autoencoder (MeshMAE) by introducing a text-guided localization (TGL) module to facilitate abutment area localization. As the parameter determination of the abutment is heavily dependent on local fine-grained features (the width and height of the implant and the distance to the opposing tooth), we pre-train the encoder using oral scan data to improve the model’s feature extraction ability. Moreover, considering that the abutment area is only a small part of the oral scan data, we designed a TGL module, which introduces the description of the abutment area through the text encoder of Contrastive Language-Image Pre-training (CLIP), enabling the network to quickly locate the abutment area. We validated the performance of TCEAD on a large abutment design dataset. Extensive experiments demonstrate that TCEAD achieves an Intersection over Union (IoU) improvement of 0.8%-12.85% over other mainstream methods, underscoring its potential in automated dental abutment design.

Method: Proposes AnoRefiner, a plug-in module for ZSAD models with two key components: 1) Anomaly Refinement Decoder (ARD) that progressively enhances image features using anomaly score maps, reducing reliance on synthetic anomaly data; 2) Progressive Group-wise Test-time Training (PGT) strategy that trains ARD in each product group for refinement in the next group, compatible with any ZSAD method.

Result: Experiments on MVTec AD and VisA datasets show AnoRefiner boosts various ZSAD models by up to 5.2% gain in pixel-AP metrics. Visualizations demonstrate improved anomaly detection at pixel level.

Conclusion: AnoRefiner effectively bridges the gap between patch-level and pixel-level anomaly detection by leveraging complementary information from anomaly score maps, achieving significant performance improvements without requiring extensive synthetic anomaly data.

Abstract: Zero-shot industrial anomaly detection (ZSAD) methods typically yield coarse anomaly maps as vision transformers (ViTs) extract patch-level features only. To solve this, recent solutions attempt to predict finer anomalies using features from ZSAD, but they still struggle to recover fine-grained anomalies without missed detections, mainly due to the gap between randomly synthesized training anomalies and real ones. We observe that anomaly score maps exactly provide complementary spatial cues that are largely absent from ZSAD’s image features, a fact overlooked before. Inspired by this, we propose an anomaly-aware refiner (AnoRefiner) that can be plugged into most ZSAD models and improve patch-level anomaly maps to the pixel level. First, we design an anomaly refinement decoder (ARD) that progressively enhances image features using anomaly score maps, reducing the reliance on synthetic anomaly data. Second, motivated by the mass production paradigm, we propose a progressive group-wise test-time training (PGT) strategy that trains ARD in each product group for the refinement process in the next group, while staying compatible with any ZSAD method. Experiments on the MVTec AD and VisA datasets show that AnoRefiner boosts various ZSAD models by up to a 5.2% gain in pixel-AP metrics, which can also be directly observed in many visualizations. The code will be available at https://github.com/HUST-SLOW/AnoRefiner.

Method: The method uses a dual-scale approach: 1) Global planning with Sparse Spatial Memory Graph (SMG) for region-centric memory and path planning, 2) Local control with a navigation foundation policy that switches between point-goal and image-goal modes, and 3) VGGT-adapter for 3D-aware feature alignment between observations and goals.

Result: MG-Nav achieves state-of-the-art zero-shot performance on HM3D Instance-Image-Goal and MP3D Image-Goal benchmarks, demonstrating robustness under dynamic rearrangements and unseen scene conditions.

Conclusion: The proposed dual-scale framework effectively unifies global memory-guided planning with local geometry-enhanced control, providing a robust solution for zero-shot visual navigation that handles long-horizon tasks while maintaining adaptability to environmental changes.

Abstract: We present MG-Nav (Memory-Guided Navigation), a dual-scale framework for zero-shot visual navigation that unifies global memory-guided planning with local geometry-enhanced control. At its core is the Sparse Spatial Memory Graph (SMG), a compact, region-centric memory where each node aggregates multi-view keyframe and object semantics, capturing both appearance and spatial structure while preserving viewpoint diversity. At the global level, the agent is localized on SMG and a goal-conditioned node path is planned via an image-to-instance hybrid retrieval, producing a sequence of reachable waypoints for long-horizon guidance. At the local level, a navigation foundation policy executes these waypoints in point-goal mode with obstacle-aware control, and switches to image-goal mode when navigating from the final node towards the visual target. To further enhance viewpoint alignment and goal recognition, we introduce VGGT-adapter, a lightweight geometric module built on the pre-trained VGGT model, which aligns observation and goal features in a shared 3D-aware space. MG-Nav operates global planning and local control at different frequencies, using periodic re-localization to correct errors. Experiments on HM3D Instance-Image-Goal and MP3D Image-Goal benchmarks demonstrate that MG-Nav achieves state-of-the-art zero-shot performance and remains robust under dynamic rearrangements and unseen scene conditions.

Method: Introduces a latent drift-guided replay method that quantifies representational instability via latent drift (change in sample’s internal feature representation after naive domain adaptation). Aggregates drift at patient level and stores per-patient slices with greatest multi-layer representation shift in a memory buffer for replay.

Result: Evaluated on cross-hospital COVID-19 CT classification task using CNN and Vision Transformer backbones, the method substantially reduces forgetting compared to naive fine-tuning and random replay baselines.

Conclusion: Latent drift serves as a practical and interpretable replay signal for advancing robust continual learning in real-world medical settings, addressing critical forgetting challenges in medical AI deployment.

Abstract: When deep learning models are sequentially trained on new data, they tend to abruptly lose performance on previously learned tasks, a critical failure known as catastrophic forgetting. This challenge severely limits the deployment of AI in medical imaging, where models must continually adapt to data from new hospitals without compromising established diagnostic knowledge. To address this, we introduce a latent drift-guided replay method that identifies and replays samples with high representational instability. Specifically, our method quantifies this instability via latent drift, the change in a sample internal feature representation after naive domain adaptation. To ensure diversity and clinical relevance, we aggregate drift at the patient level, our memory buffer stores the per patient slices exhibiting the greatest multi-layer representation shift. Evaluated on a cross-hospital COVID-19 CT classification task using state-of-the-art CNN and Vision Transformer backbones, our method substantially reduces forgetting compared to naive fine-tuning and random replay. This work highlights latent drift as a practical and interpretable replay signal for advancing robust continual learning in real world medical settings.

Method: Proposes a thinking-editing-reflection loop: thinking mechanism interprets abstract instructions using MLLM world knowledge; reflection reviews results, corrects unintended manipulations, and identifies stopping points. Two versions: ReasonEdit-S (initialized from Step1X-Edit) and ReasonEdit-Q (integrated with Qwen-Image-Edit).

Result: Significant performance improvements: ImgEdit (+4.3%), GEdit (+4.7%), and Kris (+8.2%) with ReasonEdit-S; outperforms previous open-source methods on both GEdit and Kris with ReasonEdit-Q.

Conclusion: Unlocking MLLM reasoning capabilities through thinking and reflection mechanisms significantly advances image editing performance, demonstrating the value of integrating reasoning loops into editing frameworks.

Abstract: Recent advances in image editing models have shown remarkable progress. A common architectural design couples a multimodal large language model (MLLM) encoder with a diffusion decoder, as seen in systems such as Step1X-Edit and Qwen-Image-Edit, where the MLLM encodes both the reference image and the instruction but remains frozen during training. In this work, we demonstrate that unlocking the reasoning capabilities of MLLM can further push the boundaries of editing models. Specifically, we explore two reasoning mechanisms, thinking and reflection, which enhance instruction understanding and editing accuracy. Based on that, our proposed framework enables image editing in a thinking-editing-reflection loop: the thinking mechanism leverages the world knowledge of MLLM to interpret abstract instructions, while the reflection reviews editing results, automatically corrects unintended manipulations, and identifies the stopping round. Extensive experiments demonstrate that our reasoning approach achieves significant performance gains, with improvements of ImgEdit (+4.3%), GEdit (+4.7%), and Kris (+8.2%) when initializing our DiT from the Step1X-Edit (ReasonEdit-S), and also outperforms previous open-source methods on both GEdit and Kris when integrated with Qwen-Image-Edit (ReasonEdit-Q).

Method: Proposes GeoZero framework with two datasets: GeoZero-Instruct for supervised fine-tuning and GeoZero-Hard for reinforcement learning. Introduces Answer-Anchored Group Relative Policy Optimization (A²GRPO) that regularizes reasoning using the model’s own answers to encourage diverse thinking.

Result: GeoZero surpasses state-of-the-art methods on multiple remote sensing vision-language benchmarks and fosters universal emergent reasoning capabilities across diverse geospatial tasks.

Conclusion: GeoZero provides an effective framework for geospatial reasoning without costly CoT supervision, demonstrating superior performance and emergent reasoning capabilities through self-supervised learning and reinforcement learning.

Abstract: Multimodal large language models (MLLMs) have undergone rapid development in advancing geospatial scene understanding. Recent studies have sought to enhance the reasoning capabilities of remote sensing MLLMs, typically through cold-start training with elaborately curated chain-of-thought (CoT) data. However, this approach not only incurs substantial annotation costs but also introduces human biases that may limit the diversity of model reasoning. To address these challenges, we propose GeoZero, a framework that enables MLLMs to perform geospatial reasoning without any predefined CoT supervision. Specifically, we construct two datasets, GeoZero-Instruct and GeoZero-Hard. GeoZero-Instruct allows the model to acquire preliminary geospatial knowledge through supervised fine-tuning, while GeoZero-Hard stimulates deep reasoning during the subsequent reinforcement learning stage. Furthermore, we introduce Answer-Anchored Group Relative Policy Optimization (A$^2$GRPO), where the reasoning process is regularized by the model’s own answers, encouraging diverse yet accurate thinking. Extensive experiments on multiple remote sensing vision-language benchmarks demonstrate that GeoZero not only surpasses existing state-of-the-art methods but also fosters universal emergent reasoning capabilities across diverse geospatial tasks. Code,data,and models will be publicly available at https://github.com/MiliLab/GeoZero.

Method: Analyzes why decoupling works by studying cross-modal attention behavior, finding that decoupling drives models toward task-specific multimodal interaction patterns. Proposes Attention Interaction Alignment (AIA) loss that explicitly learns task-specific multimodal interaction patterns during training. Applied AIA to Emu3 and Janus-Pro during SFT and post-training stages respectively.

Result: AIA loss refines cross-modal attention patterns and boosts both generation and understanding performance without additional architectural modifications. Demonstrates generalizability across different models (Emu3 and Janus-Pro) and training stages.

Conclusion: Attention Interaction Alignment provides an effective alternative to model decoupling for mitigating task conflicts in unified multimodal models, preserving interleave generation ability while improving performance on both understanding and generation tasks.

Abstract: Unified multimodal models for image generation and understanding represent a significant step toward AGI and have attracted widespread attention from researchers. The main challenge of this task lies in the difficulty in establishing an optimal training paradigm due to inherent conflicting targets in understanding and generation tasks. To alleviate these conflicts and pursue higher performance, many researchers adopt varying degrees of model decoupling (e.g., Double image encoders, MOE/MOT architecture, or frozen MLLM). However, excessive model decoupling can lead to the loss of interleave generation ability, undermining the original intent of unified models. In this work, we aim to explore how to mitigate task conflicts without resorting to model decoupling. Firstly, we analyze why decoupling alleviates conflicts by studying the cross-modal attention behavior of models. We observe that model decoupling essentially drives models toward task-specific multimodal interaction patterns, as seen in Qwen-VL and HunyuanImage, and that the more thorough the decoupling, the more consistent the behavior becomes. Motivated by this observation, we propose Attention Interaction Alignment (AIA) loss, which explicitly learns Task-Specific multimodal interaction patterns during training. To demonstrate the generalizability of our AIA loss, we apply it to Emu3 and Janus-Pro during SFT and post-training stage respectively. Without bells and whistles, AIA not only refines cross-modal attention patterns, but also boosts both generation and understanding performance.

Method: Proposes Variational Multi-Modal Prompt Learning (VaMP) framework that generates instance-conditioned prompts by sampling from a learned posterior distribution. Introduces class-aware prior from instance representation and class prototype, formulates prompt tuning as variational inference over latent prompt representations, and trains end-to-end through reparameterized sampling.

Result: Experiments on few-shot and domain generalization benchmarks show that VaMP achieves state-of-the-art performance.

Conclusion: VaMP demonstrates the benefits of modeling both uncertainty and task structure in multi-modal prompt learning, enabling sample-specific, uncertainty-aware adaptation that outperforms existing methods.

Abstract: Vision-language models (VLMs), such as CLIP, have shown strong generalization under zero-shot settings, yet adapting them to downstream tasks with limited supervision remains a significant challenge. Existing multi-modal prompt learning methods typically rely on fixed, shared prompts and deterministic parameters, which limits their ability to capture instance-level variation or model uncertainty across diverse tasks and domains. To tackle this issue, we propose a novel Variational Multi-Modal Prompt Learning (VaMP) framework that enables sample-specific, uncertainty-aware prompt tuning in multi-modal representation learning. VaMP generates instance-conditioned prompts by sampling from a learned posterior distribution, allowing the model to personalize its behavior based on input content. To further enhance the integration of local and global semantics, we introduce a class-aware prior derived from the instance representation and class prototype. Building upon these, we formulate prompt tuning as variational inference over latent prompt representations and train the entire framework end-to-end through reparameterized sampling. Experiments on few-shot and domain generalization benchmarks show that VaMP achieves state-of-the-art performance, highlighting the benefits of modeling both uncertainty and task structure in our method. Project page: https://visual-ai.github.io/vamp

Method: CORTEX integrates textual cues with three modules: 1) Image-level Change Detector for pixel-level differences, 2) Reasoning-aware Text Extraction using VLMs to generate compositional reasoning descriptions, and 3) Image-Text Dual Alignment module to align visual and textual features for fine-grained relational reasoning.

Result: The framework enables reasoning over both visual and textual features to capture changes that are ambiguous in visual features alone.

Conclusion: CORTEX enhances change understanding by combining visual and textual information, addressing limitations of visual-only approaches through compositional reasoning-aware text guidance.

Abstract: Change captioning aims to describe changes between a pair of images. However, existing works rely on visual features alone, which often fail to capture subtle but meaningful changes because they lack the ability to represent explicitly structured information such as object relationships and compositional semantics. To alleviate this, we present CORTEX (COmpositional Reasoning-aware TEXt-guided), a novel framework that integrates complementary textual cues to enhance change understanding. In addition to capturing cues from pixel-level differences, CORTEX utilizes scene-level textual knowledge provided by Vision Language Models (VLMs) to extract richer image text signals that reveal underlying compositional reasoning. CORTEX consists of three key modules: (i) an Image-level Change Detector that identifies low-level visual differences between paired images, (ii) a Reasoning-aware Text Extraction (RTE) module that use VLMs to generate compositional reasoning descriptions implicit in visual features, and (iii) an Image-Text Dual Alignment (ITDA) module that aligns visual and textual features for fine-grained relational reasoning. This enables CORTEX to reason over visual and textual features and capture changes that are otherwise ambiguous in visual features alone.

Method: Used convolutional neural network trained on curated dataset of verified Rubens paintings and comparative artworks to identify micro-level stylistic features characteristic of the master’s hand.

Result: The model achieved high classification accuracy in distinguishing Rubens’ work from his workshop, demonstrating computational analysis can effectively complement traditional art historical expertise.

Conclusion: Deep learning offers promising new tools for painting authentication and authorship attribution, providing valuable insights into workshop collaboration patterns that enhance traditional art historical analysis.

Abstract: This study explores the use of deep learning for the authentication and attribution of paintings, focusing on the complex case of Peter Paul Rubens and his workshop. A convolutional neural network was trained on a curated dataset of verified and comparative artworks to identify micro-level stylistic features characteristic of the master s hand. The model achieved high classification accuracy and demonstrated the potential of computational analysis to complement traditional art historical expertise, offering new insights into authorship and workshop collaboration.

Method: Rigorous decomposition of DMD training objective, analysis of CFG Augmentation vs Distribution Matching components, validation through alternative regularizers (non-parametric constraints, GAN-based objectives), and principled modifications like decoupling noise schedules.

Result: Revealed CFG Augmentation is the core “engine” of distillation while Distribution Matching acts as a “regularizer”; showed DM can be replaced with simpler constraints; proposed modifications leading to performance gains; method adopted by Z-Image project for top-tier 8-step model.

Conclusion: The conventional understanding of DMD is incorrect - CFG Augmentation drives few-step performance while DM provides regularization. This decoupling enables principled analysis and improvements to distillation, validated by real-world adoption in Z-Image project.

Abstract: Diffusion model distillation has emerged as a powerful technique for creating efficient few-step and single-step generators. Among these, Distribution Matching Distillation (DMD) and its variants stand out for their impressive performance, which is widely attributed to their core mechanism of matching the student’s output distribution to that of a pre-trained teacher model. In this work, we challenge this conventional understanding. Through a rigorous decomposition of the DMD training objective, we reveal that in complex tasks like text-to-image generation, where CFG is typically required for desirable few-step performance, the primary driver of few-step distillation is not distribution matching, but a previously overlooked component we identify as CFG Augmentation (CA). We demonstrate that this term acts as the core engine'' of distillation, while the Distribution Matching (DM) term functions as a regularizer’’ that ensures training stability and mitigates artifacts. We further validate this decoupling by demonstrating that while the DM term is a highly effective regularizer, it is not unique; simpler non-parametric constraints or GAN-based objectives can serve the same stabilizing function, albeit with different trade-offs. This decoupling of labor motivates a more principled analysis of the properties of both terms, leading to a more systematic and in-depth understanding. This new understanding further enables us to propose principled modifications to the distillation process, such as decoupling the noise schedules for the engine and the regularizer, leading to further performance gains. Notably, our method has been adopted by the Z-Image ( https://github.com/Tongyi-MAI/Z-Image ) project to develop a top-tier 8-step image generation model, empirically validating the generalization and robustness of our findings.

Method: 1) Analyze internal representations of 3DFMs to understand their emergent geometry understanding. 2) Introduce a lightweight alignment scheme that refines internal 3D representations by tuning only a small subset of backbone bias terms while keeping decoder heads frozen. 3) Create MegaUnScene benchmark with test splits for relative pose estimation and dense 3D reconstruction.

Result: 3DFMs exhibit emergent understanding of extreme-view geometry despite no specific training. The lightweight alignment scheme substantially improves relative pose estimation under extreme viewpoints without degrading per-image depth or point quality.

Conclusion: 3D foundation models possess inherent geometric reasoning capabilities that can be enhanced through targeted, lightweight adaptation. The proposed alignment scheme and MegaUnScene benchmark advance the field’s ability to handle extreme-view scenarios in 3D vision.

Abstract: 3D foundation models (3DFMs) have recently transformed 3D vision, enabling joint prediction of depths, poses, and point maps directly from images. Yet their ability to reason under extreme, non-overlapping views remains largely unexplored. In this work, we study their internal representations and find that 3DFMs exhibit an emergent understanding of extreme-view geometry, despite never being trained for such conditions. To further enhance these capabilities, we introduce a lightweight alignment scheme that refines their internal 3D representation by tuning only a small subset of backbone bias terms, leaving all decoder heads frozen. This targeted adaptation substantially improves relative pose estimation under extreme viewpoints without degrading per-image depth or point quality. Additionally, we contribute MegaUnScene, a new benchmark of Internet scenes unseen by existing 3DFMs, with dedicated test splits for both relative pose estimation and dense 3D reconstruction. All code and data will be released.

Method: Two-stage framework: Architect module predicts structured layouts specifying where each person appears; Artist module synthesizes images guided by spatially-grounded face matching reward combining Hungarian spatial alignment with ArcFace identity similarity. Uses Group Relative Policy Optimization with compositional rewards.

Result: Ar2Can achieves substantial improvements in both count accuracy and identity preservation on MultiHuman-Testbench while maintaining high perceptual quality, using primarily synthetic data without requiring real multi-human images.

Conclusion: The disentangled approach of spatial planning and identity rendering effectively solves multi-human generation challenges, demonstrating that high-quality multi-human scenes can be generated using synthetic data.

Abstract: Despite recent advances in text-to-image generation, existing models consistently fail to produce reliable multi-human scenes, often duplicating faces, merging identities, or miscounting individuals. We present Ar2Can, a novel two-stage framework that disentangles spatial planning from identity rendering for multi-human generation. The Architect module predicts structured layouts, specifying where each person should appear. The Artist module then synthesizes photorealistic images, guided by a spatially-grounded face matching reward that combines Hungarian spatial alignment with ArcFace identity similarity. This approach ensures faces are rendered at correct locations and faithfully preserve reference identities. We develop two Architect variants, seamlessly integrated with our diffusion-based Artist model and optimized via Group Relative Policy Optimization (GRPO) using compositional rewards for count accuracy, image quality, and identity matching. Evaluated on the MultiHuman-Testbench, Ar2Can achieves substantial improvements in both count accuracy and identity preservation, while maintaining high perceptual quality. Notably, our method achieves these results using primarily synthetic data, without requiring real multi-human images.

Method: Built on Ovis-U1 framework with diffusion-based visual decoder and stronger Ovis 2.5 multimodal backbone, using text-centric training pipeline combining large-scale pre-training with tailored post-training refinements.

Result: Achieves text rendering performance on par with larger open models like Qwen-Image and approaches closed-source systems like Seedream and GPT4o, while remaining deployable on a single high-end GPU.

Conclusion: A strong multimodal backbone combined with carefully designed text-focused training is sufficient for reliable bilingual text rendering without needing oversized or proprietary models.

Abstract: We introduce $\textbf{Ovis-Image}$, a 7B text-to-image model specifically optimized for high-quality text rendering, designed to operate efficiently under stringent computational constraints. Built upon our previous Ovis-U1 framework, Ovis-Image integrates a diffusion-based visual decoder with the stronger Ovis 2.5 multimodal backbone, leveraging a text-centric training pipeline that combines large-scale pre-training with carefully tailored post-training refinements. Despite its compact architecture, Ovis-Image achieves text rendering performance on par with significantly larger open models such as Qwen-Image and approaches closed-source systems like Seedream and GPT4o. Crucially, the model remains deployable on a single high-end GPU with moderate memory, narrowing the gap between frontier-level text rendering and practical deployment. Our results indicate that combining a strong multimodal backbone with a carefully designed, text-focused training recipe is sufficient to achieve reliable bilingual text rendering without resorting to oversized or proprietary models.

Method: Built on Scalable Single-Stream Diffusion Transformer (S3-DiT) architecture with systematic optimization of the entire model lifecycle, including curated data infrastructure, streamlined training curriculum, few-step distillation with reward post-training, and omni-pre-training paradigm.

Result: Z-Image achieves performance comparable to or surpassing leading competitors across various dimensions, with exceptional photorealistic image generation and bilingual text rendering capabilities. Z-Image-Turbo offers sub-second inference on H800 GPU and compatibility with consumer hardware (<16GB VRAM). Full training completed in 314K H800 GPU hours (~$630K).

Conclusion: State-of-the-art image generation results are achievable with significantly reduced computational overhead, challenging the “scale-at-all-costs” paradigm. The release of code, weights, and demo aims to foster development of accessible, budget-friendly generative models.

Abstract: The landscape of high-performance image generation models is currently dominated by proprietary systems, such as Nano Banana Pro and Seedream 4.0. Leading open-source alternatives, including Qwen-Image, Hunyuan-Image-3.0 and FLUX.2, are characterized by massive parameter counts (20B to 80B), making them impractical for inference, and fine-tuning on consumer-grade hardware. To address this gap, we propose Z-Image, an efficient 6B-parameter foundation generative model built upon a Scalable Single-Stream Diffusion Transformer (S3-DiT) architecture that challenges the “scale-at-all-costs” paradigm. By systematically optimizing the entire model lifecycle – from a curated data infrastructure to a streamlined training curriculum – we complete the full training workflow in just 314K H800 GPU hours (approx. $630K). Our few-step distillation scheme with reward post-training further yields Z-Image-Turbo, offering both sub-second inference latency on an enterprise-grade H800 GPU and compatibility with consumer-grade hardware (<16GB VRAM). Additionally, our omni-pre-training paradigm also enables efficient training of Z-Image-Edit, an editing model with impressive instruction-following capabilities. Both qualitative and quantitative experiments demonstrate that our model achieves performance comparable to or surpassing that of leading competitors across various dimensions. Most notably, Z-Image exhibits exceptional capabilities in photorealistic image generation and bilingual text rendering, delivering results that rival top-tier commercial models, thereby demonstrating that state-of-the-art results are achievable with significantly reduced computational overhead. We publicly release our code, weights, and online demo to foster the development of accessible, budget-friendly, yet state-of-the-art generative models.

Method: Case study of three VLMs (Llava-Llama and two Qwen models) with four sets of questions of increasing complexity about artworks: general content, emotional content, expression of emotions, and emotion symbols. Qualitative expert evaluation was conducted.

Result: VLMs recognize image content surprisingly well and often identify depicted emotions and their expression. Performance is best for concrete images but fails for highly abstract or symbolic images. Symbol recognition remains fundamentally difficult. Models exhibit LLM weakness of inconsistent answers to related questions.

Conclusion: Current VLMs show promising capabilities for detecting emotional content in artworks but have limitations with abstraction, symbolism, and consistency, indicating areas for future improvement in art analysis applications.

Abstract: Emotions are a fundamental aspect of artistic expression. Due to their abstract nature, there is a broad spectrum of emotion realization in artworks. These are subject to historical change and their analysis requires expertise in art history. In this article, we investigate which aspects of emotional expression can be detected by current (2025) vision language models (VLMs). We present a case study of three VLMs (Llava-Llama and two Qwen models) in which we ask these models four sets of questions of increasing complexity about artworks (general content, emotional content, expression of emotions, and emotion symbols) and carry out a qualitative expert evaluation. We find that the VLMs recognize the content of the images surprisingly well and often also which emotions they depict and how they are expressed. The models perform best for concrete images but fail for highly abstract or highly symbolic images. Reliable recognition of symbols remains fundamentally difficult. Furthermore, the models continue to exhibit the well-known LLM weakness of providing inconsistent answers to related questions.

Method: MIMM-X disentangles causal features from multiple spurious correlations by minimizing their mutual information, enabling predictions based on true underlying causal relationships rather than dataset-specific shortcuts.

Result: Evaluated on three datasets (UK Biobank, NAKO, CheXpert) across two imaging modalities (MRI and X-ray), MIMM-X effectively mitigates shortcut learning of multiple spurious correlations.

Conclusion: MIMM-X provides a robust framework for improving generalization in medical imaging by addressing the critical problem of multiple coexisting spurious correlations through mutual information minimization.

Abstract: Deep learning models can excel on medical tasks, yet often experience spurious correlations, known as shortcut learning, leading to poor generalization in new environments. Particularly in medical imaging, where multiple spurious correlations can coexist, misclassifications can have severe consequences. We propose MIMM-X, a framework that disentangles causal features from multiple spurious correlations by minimizing their mutual information. It enables predictions based on true underlying causal relationships rather than dataset-specific shortcuts. We evaluate MIMM-X on three datasets (UK Biobank, NAKO, CheXpert) across two imaging modalities (MRI and X-ray). Results demonstrate that MIMM-X effectively mitigates shortcut learning of multiple spurious correlations.

Method: Two-stage learning: 1) Transform point maps to real space via iterative affinity learning for camera control; 2) Project point maps from two views onto target plane, refine via stereo matching, and anchor Gaussian primitives for rendering. Uses scale-aware point maps trained self-supervised without 3D supervision.

Result: Improves both stability of point map reconstruction and visual quality of free-viewpoint rendering on collected multi-view human-centered data.

Conclusion: Splat-SAP successfully enables feed-forward Gaussian Splatting rendering from sparse binocular cameras by leveraging robust pixel-wise point map geometry representation and a two-stage learning strategy.

Abstract: We present Splat-SAP, a feed-forward approach to render novel views of human-centered scenes from binocular cameras with large sparsity. Gaussian Splatting has shown its promising potential in rendering tasks, but it typically necessitates per-scene optimization with dense input views. Although some recent approaches achieve feed-forward Gaussian Splatting rendering through geometry priors obtained by multi-view stereo, such approaches still require largely overlapped input views to establish the geometry prior. To bridge this gap, we leverage pixel-wise point map reconstruction to represent geometry which is robust to large sparsity for its independent view modeling. In general, we propose a two-stage learning strategy. In stage 1, we transform the point map into real space via an iterative affinity learning process, which facilitates camera control in the following. In stage 2, we project point maps of two input views onto the target view plane and refine such geometry via stereo matching. Furthermore, we anchor Gaussian primitives on this refined plane in order to render high-quality images. As a metric representation, the scale-aware point map in stage 1 is trained in a self-supervised manner without 3D supervision and stage 2 is supervised with photo-metric loss. We collect multi-view human-centered data and demonstrate that our method improves both the stability of point map reconstruction and the visual quality of free-viewpoint rendering.

Method: Created the Bharat Scene Text Dataset (BSTD) with 100K+ words spanning 11 Indian languages and English from 6,500+ scene images across India, with meticulous annotations supporting multiple tasks including detection, script identification, word recognition, and end-to-end recognition.

Result: Evaluated state-of-the-art English models adapted for Indian languages, highlighting challenges and opportunities in Indian language scene text recognition. All models and data are open source.

Conclusion: BSTD represents a significant step toward advancing research in Indian language scene text recognition by providing a comprehensive benchmark and open-source resources.

Abstract: Reading scene text, that is, text appearing in images, has numerous application areas, including assistive technology, search, and e-commerce. Although scene text recognition in English has advanced significantly and is often considered nearly a solved problem, Indian language scene text recognition remains an open challenge. This is due to script diversity, non-standard fonts, and varying writing styles, and, more importantly, the lack of high-quality datasets and open-source models. To address these gaps, we introduce the Bharat Scene Text Dataset (BSTD) - a large-scale and comprehensive benchmark for studying Indian Language Scene Text Recognition. It comprises more than 100K words that span 11 Indian languages and English, sourced from over 6,500 scene images captured across various linguistic regions of India. The dataset is meticulously annotated and supports multiple scene text tasks, including: (i) Scene Text Detection, (ii) Script Identification, (iii) Cropped Word Recognition, and (iv) End-to-End Scene Text Recognition. We evaluated state-of-the-art models originally developed for English by adapting (fine-tuning) them for Indian languages. Our results highlight the challenges and opportunities in Indian language scene text recognition. We believe that this dataset represents a significant step toward advancing research in this domain. All our models and data are open source.

Method: Used synchronous aerial VIS and TIR imagery with YOLO11n model; tested two fusion methods: early fusion (PCA-based) and late fusion (CART-based); automatically aligned images using deep learning.

Result: Both fusion methods improved F1 scores compared to VIS-only model; late fusion improved occupied nest detection from 90.2% to 93.0%; model identified false positives with 90% recall.

Conclusion: Fusion methods improve detection but have limitations (TIR FOV constraints, alignment issues); very high-resolution visible sensors alone may be more practical for operational surveys.

Abstract: Efficient wildlife monitoring methods are necessary for biodiversity conservation and management. The combination of remote sensing, aerial imagery and deep learning offer promising opportunities to renew or improve existing survey methods. The complementary use of visible (VIS) and thermal infrared (TIR) imagery can add information compared to a single-source image and improve results in an automated detection context. However, the alignment and fusion process can be challenging, especially since visible and thermal images usually have different fields of view (FOV) and spatial resolutions. This research presents a case study on the great blue heron (Ardea herodias) to evaluate the performances of synchronous aerial VIS and TIR imagery to automatically detect individuals and nests using a YOLO11n model. Two VIS-TIR fusion methods were tested and compared: an early fusion approach and a late fusion approach, to determine if the addition of the TIR image gives any added value compared to a VIS-only model. VIS and TIR images were automatically aligned using a deep learning model. A principal component analysis fusion method was applied to VIS-TIR image pairs to form the early fusion dataset. A classification and regression tree was used to process the late fusion dataset, based on the detection from the VIS-only and TIR-only trained models. Across all classes, both late and early fusion improved the F1 score compared to the VIS-only model. For the main class, occupied nest, the late fusion improved the F1 score from 90.2 (VIS-only) to 93.0%. This model was also able to identify false positives from both sources with 90% recall. Although fusion methods seem to give better results, this approach comes with a limiting TIR FOV and alignment constraints that eliminate data. Using an aircraft-mounted very high-resolution visible sensor could be an interesting option for operationalizing surveys.

Method: Visual Rationale Learning (ViRL) reframes visual actions as core reasoning primitives. It uses: (1) Process Supervision with ground-truth rationales, (2) Objective Alignment via step-level reward shaping, and (3) Fine-Grained Credit Assignment to distinguish correct, redundant, and erroneous actions.

Result: ViRL achieves state-of-the-art results across benchmarks spanning perception, hallucination, and reasoning when trained purely with end-to-end reinforcement learning.

Conclusion: Visual rationalization establishes a task-agnostic, process-grounded paradigm for building transparent, verifiable, and trustworthy vision-language models that get the right answer for the right visual reason.

Abstract: Recent advances in vision-language reasoning underscore the importance of thinking with images, where models actively ground their reasoning in visual evidence. Yet, prevailing frameworks treat visual actions as optional tools, boosting metrics but leaving reasoning ungrounded and crops ineffective. This gap gives rise to the illusion of thinking with images: models seem visually grounded but rely on context-agnostic actions that neither refine perception nor guide reasoning toward correct answers. We address this problem by reframing visual actions as core reasoning primitives rather than optional tools, which we term visual rationalization, the visual analogue of textual Chain-of-Thought. Building on this insight, we propose Visual Rationale Learning (ViRL), an end-to-end paradigm that grounds training in the visual rationale itself. ViRL integrates (1) Process Supervision with ground-truth rationales, (2) Objective Alignment via step-level reward shaping, and (3) Fine-Grained Credit Assignment to distinguish correct, redundant, and erroneous actions. By ensuring each action contributes meaningfully to the reasoning chain, ViRL enables models to “get the right answer for the right visual reason”. Trained purely with end-to-end RL, ViRL achieves state-of-the-art results across benchmarks spanning perception, hallucination, and reasoning. This work establishes visual rationalization as a task-agnostic, process-grounded paradigm for building transparent, verifiable, and trustworthy vision-language models.

Method: Proposed an end-to-end hybrid deep learning model integrating CNNs and Vision Transformers to capture both local spatial features and global contextual dependencies from MRI scans. Added BiLSTM networks to process features from four consecutive MRI timepoints along with non-image biomarkers, enabling temporal progression modeling for predicting cognitive status at month 48.

Result: Achieved 95.05% average accuracy in predicting progression between stable MCI (sMCI) and progressive MCI (pMCI), outperforming existing studies and demonstrating state-of-the-art performance in longitudinal AD prediction.

Conclusion: The multimodal approach combining spatial and temporal modeling is highly effective for early Alzheimer’s disease detection, showing superior performance in predicting MCI progression and highlighting the value of integrating multiple data modalities for neurodegenerative disease prediction.

Abstract: Alzheimer’s disease (AD) is a prevalent neurodegenerative disorder that progressively impairs memory, decision-making, and overall cognitive function. As AD is irreversible, early prediction is critical for timely intervention and management. Mild Cognitive Impairment (MCI), a transitional stage between cognitively normal (CN) aging and AD, plays a significant role in early AD diagnosis. However, predicting MCI progression remains a significant challenge, as not all individuals with MCI convert to AD. MCI subjects are categorized into stable MCI (sMCI) and progressive MCI (pMCI) based on conversion status. In this study, we propose a generalized, end-to-end deep learning model for AD prediction using MCI cases from the Alzheimer’s Disease Neuroimaging Initiative (ADNI). Our hybrid architecture integrates Convolutional Neural Networks and Vision Transformers to capture both local spatial features and global contextual dependencies from Magnetic Resonance Imaging (MRI) scans. To incorporate temporal progression, we further employ Bidirectional Long Short-Term Memory (BiLSTM) networks to process features extracted from four consecutive MRI timepoints along with some other non-image biomarkers, predicting each subject’s cognitive status at month 48. Our multimodal model achieved an average progression prediction accuracy of 95.05% between sMCI and pMCI, outperforming existing studies in AD prediction. This work demonstrates state-of-the-art performance in longitudinal AD prediction and highlights the effectiveness of combining spatial and temporal modeling for the early detection of Alzheimer’s disease.

Method: Constructed CultureMix, a food Visual Question Answering benchmark with 23k diffusion-generated, human-verified culture mixing images across four subtasks: (1) food-only, (2) food+food, (3) food+background, and (4) food+food+background. Evaluated 10 LVLMs and explored three robustness strategies, finding supervised fine-tuning with diverse culture mixing data most effective.

Result: LVLMs show consistent failures to preserve individual cultural identities in mixed settings. Models demonstrate strong background reliance, with accuracy dropping 14% when cultural backgrounds are added to food-only baselines. They produce inconsistent predictions for identical foods across different contexts. Supervised fine-tuning substantially improves model consistency and reduces background sensitivity.

Conclusion: Culture mixing scenarios present a critical challenge for LVLMs, revealing systematic failures in preserving cultural identities and excessive background reliance. The research calls for increased attention to culture mixing as a critical step toward developing LVLMs capable of operating reliably in culturally diverse real-world environments.

Abstract: In a globalized world, cultural elements from diverse origins frequently appear together within a single visual scene. We refer to these as culture mixing scenarios, yet how Large Vision-Language Models (LVLMs) perceive them remains underexplored. We investigate culture mixing as a critical challenge for LVLMs and examine how current models behave when cultural items from multiple regions appear together. To systematically analyze these behaviors, we construct CultureMix, a food Visual Question Answering (VQA) benchmark with 23k diffusion-generated, human-verified culture mixing images across four subtasks: (1) food-only, (2) food+food, (3) food+background, and (4) food+food+background. Evaluating 10 LVLMs, we find consistent failures to preserve individual cultural identities in mixed settings. Models show strong background reliance, with accuracy dropping 14% when cultural backgrounds are added to food-only baselines, and they produce inconsistent predictions for identical foods across different contexts. To address these limitations, we explore three robustness strategies. We find supervised fine-tuning using a diverse culture mixing dataset substantially improve model consistency and reduce background sensitivity. We call for increased attention to culture mixing scenarios as a critical step toward developing LVLMs capable of operating reliably in culturally diverse real-world environments.

Method: Constructed a dataset and evaluated LVLMs on synchronized driver/road camera video processing, comparing pre-trained vs fine-tuned models.

Result: Pre-trained LVLMs have limited effectiveness, but fine-tuned LVLMs can generate accurate safety-aware driving instructions. Challenges remain in detecting subtle/complex events.

Conclusion: Fine-tuning improves LVLM performance for road safety applications, but further work needed for detecting complex events. Error analysis provides insights for improving LVLM-based safety systems.

Abstract: Large-scale Vision Language Models (LVLMs) exhibit advanced capabilities in tasks that require visual information, including object detection. These capabilities have promising applications in various industrial domains, such as autonomous driving. For example, LVLMs can generate safety-oriented descriptions of videos captured by road-facing cameras. However, ensuring comprehensive safety requires monitoring driver-facing views as well to detect risky events, such as the use of mobiles while driving. Thus, the ability to process synchronized inputs is necessary from both driver-facing and road-facing cameras. In this study, we develop models and investigate the capabilities of LVLMs by constructing a dataset and evaluating their performance on this dataset. Our experimental results demonstrate that while pre-trained LVLMs have limited effectiveness, fine-tuned LVLMs can generate accurate and safety-aware driving instructions. Nonetheless, several challenges remain, particularly in detecting subtle or complex events in the video. Our findings and error analysis provide valuable insights that can contribute to the improvement of LVLM-based systems in this domain.

Method: Used RSNA Bone Age Challenge dataset with 200 original radiographs, generated 600 inpainted versions using gpt-image-1 with natural language prompts targeting non-anatomical artifacts. Assessed downstream performance with deep learning ensembles for bone age estimation (MAE) and gender classification (AUC), plus pixel intensity distribution analysis.

Result: Inpainting markedly degraded performance: bone age MAE increased from 6.26 to 30.11 months, gender classification AUC decreased from 0.955 to 0.704. Inpainted images showed pixel-intensity shifts and inconsistencies indicating structural modifications.

Conclusion: Despite visual realism, foundation model-based inpainting can obscure clinically relevant features and introduce latent bias even when editing non-diagnostic regions, requiring rigorous task-specific validation before clinical AI integration.

Abstract: Generative foundation models can remove visual artifacts through realistic image inpainting, but their impact on medical AI performance remains uncertain. Pediatric hand radiographs often contain non-anatomical markers, and it is unclear whether inpainting these regions preserves features needed for bone age and gender prediction. To evaluate the clinical reliability of generative model-based inpainting for artifact removal, we used the RSNA Bone Age Challenge dataset, selecting 200 original radiographs and generating 600 inpainted versions with gpt-image-1 using natural language prompts to target non-anatomical artifacts. Downstream performance was assessed with deep learning ensembles for bone age estimation and gender classification, using mean absolute error (MAE) and area under the ROC curve (AUC) as metrics, and pixel intensity distributions to detect structural alterations. Inpainting markedly degraded model performance: bone age MAE increased from 6.26 to 30.11 months, and gender classification AUC decreased from 0.955 to 0.704. Inpainted images displayed pixel-intensity shifts and inconsistencies, indicating structural modifications not corrected by simple calibration. These findings show that, although visually realistic, foundation model-based inpainting can obscure subtle but clinically relevant features and introduce latent bias even when edits are confined to non-diagnostic regions, underscoring the need for rigorous, task-specific validation before integrating such generative tools into clinical AI workflows.

Method: Two-part framework: 1) Text-guided diffusion pipeline using GPT-4o-generated scene descriptions and super-resolved exemplars to synthesize class-balanced training images; 2) DViT architecture combining DCNv4 deformable convolutional backbone with Vision Transformer encoder to capture both fine-scale geometry and global context.

Result: Achieved 0.9572 overall accuracy, 0.9576 macro F1-score, and 0.9510 Cohen’s Kappa on AID dataset (8 classes), outperforming ViT baseline and other models. Cross-dataset experiments on SIRI-WHU subset showed 0.9333 accuracy, demonstrating good transferability. GPT-4o evaluation confirmed DViT’s attention aligns with hydrologically meaningful structures.

Conclusion: Description-driven generative augmentation combined with deformation-aware transformers is a promising approach for high-resolution land-cover mapping, addressing annotation scarcity and geometric distortion challenges.

Abstract: Land-cover underpins ecosystem services, hydrologic regulation, disaster-risk reduction, and evidence-based land planning; timely, accurate land-cover maps are therefore critical for environmental stewardship. Remote sensing-based land-cover classification offers a scalable route to such maps but is hindered by scarce and imbalanced annotations and by geometric distortions in high-resolution scenes. We propose LC4-DViT (Land-cover Creation for Land-cover Classification with Deformable Vision Transformer), a framework that combines generative data creation with a deformation-aware Vision Transformer. A text-guided diffusion pipeline uses GPT-4o-generated scene descriptions and super-resolved exemplars to synthesize class-balanced, high-fidelity training images, while DViT couples a DCNv4 deformable convolutional backbone with a Vision Transformer encoder to jointly capture fine-scale geometry and global context. On eight classes from the Aerial Image Dataset (AID)-Beach, Bridge, Desert, Forest, Mountain, Pond, Port, and River-DViT achieves 0.9572 overall accuracy, 0.9576 macro F1-score, and 0.9510 Cohen’ s Kappa, improving over a vanilla ViT baseline (0.9274 OA, 0.9300 macro F1, 0.9169 Kappa) and outperforming ResNet50, MobileNetV2, and FlashInternImage. Cross-dataset experiments on a three-class SIRI-WHU subset (Harbor, Pond, River) yield 0.9333 overall accuracy, 0.9316 macro F1, and 0.8989 Kappa, indicating good transferability. An LLM-based judge using GPT-4o to score Grad-CAM heatmaps further shows that DViT’ s attention aligns best with hydrologically meaningful structures. These results suggest that description-driven generative augmentation combined with deformation-aware transformers is a promising approach for high-resolution land-cover mapping.

Method: Uses pose-conditioned world memory retrieval: maintains dynamic local memory, retrieves pose-aligned world tokens using a retriever, which then condition video generation along camera trajectories.

Result: Substantially outperforms SOTA: reduces MEt3R (0.3703→0.3690), improves AUC@30 (0.181→0.200), lower FVD than all baselines. In human study, 67.6% preferences favor Captain Safari across all evaluation axes.

Conclusion: Pose-conditioned world memory is powerful for long-horizon controllable video generation; OpenSafari provides challenging benchmark for future world-engine research.

Abstract: World engines aim to synthesize long, 3D-consistent videos that support interactive exploration of a scene under user-controlled camera motion. However, existing systems struggle under aggressive 6-DoF trajectories and complex outdoor layouts: they lose long-range geometric coherence, deviate from the target path, or collapse into overly conservative motion. To this end, we introduce Captain Safari, a pose-conditioned world engine that generates videos by retrieving from a persistent world memory. Given a camera path, our method maintains a dynamic local memory and uses a retriever to fetch pose-aligned world tokens, which then condition video generation along the trajectory. This design enables the model to maintain stable 3D structure while accurately executing challenging camera maneuvers. To evaluate this setting, we curate OpenSafari, a new in-the-wild FPV dataset containing high-dynamic drone videos with verified camera trajectories, constructed through a multi-stage geometric and kinematic validation pipeline. Across video quality, 3D consistency, and trajectory following, Captain Safari substantially outperforms state-of-the-art camera-controlled generators. It reduces MEt3R from 0.3703 to 0.3690, improves AUC@30 from 0.181 to 0.200, and yields substantially lower FVD than all camera-controlled baselines. More importantly, in a 50-participant, 5-way human study where annotators select the best result among five anonymized models, 67.6% of preferences favor our method across all axes. Our results demonstrate that pose-conditioned world memory is a powerful mechanism for long-horizon, controllable video generation and provide OpenSafari as a challenging new benchmark for future world-engine research.

Method: SpaceMind uses a dual-encoder architecture with VGGT for spatial understanding and InternViT for 2D vision. The key innovation is a Camera-Guided Modality Fusion module that treats camera representations as active guidance - applying camera-conditioned biasing to spatial tokens, assigning geometric importance weights, and gating fused representations with camera embeddings.

Result: SpaceMind achieves new state-of-the-art results on VSI-Bench, SQA3D, and SPBench benchmarks. It surpasses both open and proprietary systems on VSI-Bench and SPBench by large margins, and achieves SOTA on SQA3D.

Conclusion: Camera-guided modality fusion provides an effective inductive bias for giving VLMs genuinely spatially grounded intelligence. The approach works with only RGB inputs and outperforms existing methods, demonstrating practical value for spatial reasoning tasks.

Abstract: Large vision-language models (VLMs) show strong multimodal understanding but still struggle with 3D spatial reasoning, such as distance estimation, size comparison, and cross-view consistency. Existing 3D-aware methods either depend on auxiliary 3D information or enhance RGB-only VLMs with geometry encoders through shallow feature fusion. We propose SpaceMind, a multimodal large language model explicitly designed for spatial reasoning solely from RGB inputs. The model adopts a dual-encoder architecture, integrating VGGT as a spatial understanding encoder and InternViT as a 2D visual encoder. The key idea is to treat the camera representation as an active guiding modality rather than passive metadata. Specifically, SpaceMind introduces a lightweight Camera-Guided Modality Fusion module before the language model to replace shallow fusion. It applies camera-conditioned biasing to spatial tokens, assigns query-independent weights reflecting their geometric importance, and uses the camera embedding to gate the fused representation. Empirically, SpaceMind establishes new state-of-the-art results on VSI-Bench, SQA3D and SPBench, surpassing both open and proprietary systems on VSI-Bench and SPBench by large margins and achieving state-of-the-art performance on SQA3D. These results demonstrate that camera-guided modality fusion is an effective and practical inductive bias for equipping VLMs with genuinely spatially grounded intelligence. We will release code and model checkpoints to support future research.

Method: 1) Introduces MMA-Bench with videos and tasks probing modality reliance; 2) Uses black-box and white-box interpretability techniques to analyze MLLM brittleness; 3) Proposes modality alignment tuning strategy to teach models when to prioritize, leverage, or ignore specific modality cues.

Result: Current MLLMs struggle with misaligned audio-visual pairs and simple misleading text, lacking robust multimodal reasoning. The proposed alignment tuning yields demonstrably stronger multimodal grounding, making models more reliable with contradicting modalities.

Conclusion: This work provides interpretability tools and a clear path toward developing MLLMs with intrinsically reliable cross-modal reasoning. The findings highlight the need for better modality alignment in MLLMs to handle real-world multimodal contradictions.

Abstract: Despite remarkable advancements in Multimodal Large Language Models (MLLMs), a fundamental question remains: are MLLMs robust to contradicting modalities? To rigorously study this, we introduce MMA-Bench comprising videos and tasks that probe a model’s reliance on specific modalities. Using black-box and white-box interpretability techniques, we provide a critical analysis of the brittleness of both open- and closed-sourced MLLMs. We show that current MLLMs struggle under misaligned audio-visual pairs and simple misleading text, thereby lacking robust multi-modal reasoning. Building on these findings, we propose a modality alignment tuning strategy to teach the model when to prioritize, leverage, or ignore specific modality cues. Through extensive experiments and analysis, we show that our alignment tuning yields demonstrably stronger multimodal grounding. This work provides both interpretability tools and a clear path toward developing MLLMs with intrinsically reliable cross-modal reasoning. Code and dataset will be publicly available.

Method: 1) Introduce RETINA benchmark constructed via LLM-driven pipeline with 120k training and 2k human-curated test queries referencing secondary subjects paired with related entity images. 2) Propose MIMIR model that enriches document embeddings by augmenting images of multiple related entities, unlike prior single-image approaches.

Result: Existing models show significantly degraded performance on RETINA, confirming their reliance on visual shortcuts. MIMIR effectively handles RETINA challenges and demonstrates improved performance compared to prior approaches.

Conclusion: RETINA exposes limitations of current MKB-VQA benchmarks, and MIMIR provides an effective solution by leveraging multiple related entity images for better multimodal knowledge-based reasoning.

Abstract: Existing Multimodal Knowledge-Based Visual Question Answering (MKB-VQA) benchmarks suffer from “visual shortcuts”, as the query image typically matches the primary subject entity of the target document. We demonstrate that models can exploit these shortcuts, achieving comparable results using visual cues alone. To address this, we introduce Relational Entity Text-Image kNowledge Augmented (RETINA) benchmark, automatically constructed using an LLM-driven pipeline, consisting of 120k training and 2k human-curated test set. RETINA contains queries referencing secondary subjects (i.e. related entities) and pairs them with images of these related entities, removing the visual shortcut. When evaluated on RETINA existing models show significantly degraded performance, confirming their reliance on the shortcut. Furthermore, we propose Multi-Image MultImodal Retriever (MIMIR), which enriches document embeddings by augmenting images of multiple related entities, effectively handling RETINA, unlike prior work that uses only a single image per document. Our experiments validate the limitations of existing benchmarks and demonstrate the effectiveness of RETINA and MIMIR. Our project is available at: Project Page.

Method: SLEUTH orchestrates a retriever and four collaborative agents in a coarse-to-fine process: identifies key textual/visual clues within retrieved pages, filters salient visual evidence (tables/charts), analyzes queries to devise reasoning strategies, and synthesizes distilled evidence-dense multimodal context for final predictions.

Result: When paired with advanced VLM backbones, SLEUTH consistently improves performance on multiple long document benchmarks, achieving state-of-the-art results. Ablation studies verify each module’s effectiveness and confirm benefits of the hierarchical refinement paradigm.

Conclusion: SLEUTH is a model-agnostic, scalable multi-agent framework that effectively addresses VLMs’ limitations on long documents through hierarchical refinement, distillation of evidence-dense context, and collaborative agent orchestration.

Abstract: Document understanding is a long standing practical task. Vision Language Models (VLMs) have gradually become a primary approach in this domain, demonstrating effective performance on single page tasks. However, their effectiveness diminishes when handling long documents. In such scenarios, clues are often scattered across multiple pages and modalities, and redundancy from lengthy inputs can impair the models judgment. While retrieval augmented generation mitigates this issue by filtering for question relevant content, the retrieved results still contain substantial redundancy. To address these limitations, we propose SLEUTH, a multi agent framework. Concretely, SLEUTH orchestrates a retriever and four collaborative agents in a coarse to fine process. The framework identifies key textual and visual clues within the retrieved pages, filters for salient visual evidence such as tables and charts, and analyzes the query to devise a reasoning strategy. It ultimately synthesizes a distilled, evidence dense multimodal context to generate the final prediction. SLEUTH is model agnostic and scalable. When paired with advanced VLM backbones, it consistently improves performance on multiple long document benchmarks, achieving state of the art results. Ablation studies verify each modules effectiveness and confirm the benefits of our hierarchical refinement paradigm.

Method: 1) Create REVEAL-Bench benchmark structured around chain-of-evidence from multiple lightweight expert models with step-by-step reasoning traces. 2) Develop REVEAL framework using expert-grounded reinforcement learning with reward mechanism optimizing detection accuracy, explanation fidelity, and logical coherence based on explicit forensic evidence.

Result: REVEAL significantly enhances detection accuracy, explanation fidelity, and robust cross-model generalization, establishing new state-of-the-art for explainable image forensics.

Conclusion: The proposed REVEAL framework addresses critical gaps in AI-generated image detection by providing fine-grained, interpretable, and verifiable reasoning chains alongside detection outcomes, advancing the field of explainable image forensics.

Abstract: With the rapid advancement of generative models, visually realistic AI-generated images have become increasingly difficult to distinguish from authentic ones, posing severe threats to social trust and information integrity. Consequently, there is an urgent need for efficient and truly explainable image forensic methods. Recent detection paradigms have shifted towards explainable forensics. However, state-of-the-art approaches primarily rely on post-hoc rationalizations or visual discrimination, lacking a verifiable chain of evidence. This reliance on surface-level pattern matching limits the generation of causally grounded explanations and often results in poor generalization. To bridge this critical gap, we introduce \textbf{REVEAL-Bench}, the first reasoning-enhanced multimodal benchmark for AI-generated image detection that is explicitly structured around a chain-of-evidence derived from multiple lightweight expert models, then records step-by-step reasoning traces and evidential justifications. Building upon this dataset, we propose \textbf{REVEAL} (\underline{R}easoning-\underline{e}nhanced Forensic E\underline{v}id\underline{e}nce \underline{A}na\underline{l}ysis), an effective and explainable forensic framework that integrates detection with a novel expert-grounded reinforcement learning. Our reward mechanism is specially tailored to jointly optimize detection accuracy, explanation fidelity, and logical coherence grounded in explicit forensic evidence, enabling REVEAL to produce fine-grained, interpretable, and verifiable reasoning chains alongside its detection outcomes. Extensive experimental results demonstrate that REVEAL significantly enhances detection accuracy, explanation fidelity, and robust cross-model generalization, benchmarking a new state of the art for explainable image forensics.

Method: GLOW integrates neural implicit surface representation with a neural radiance cache to approximate global illumination, jointly optimizing geometry and reflectance. It introduces a dynamic radiance cache for sharp lighting discontinuities and a surface-angle-weighted radiometric loss to suppress specular artifacts.

Result: GLOW substantially outperforms prior methods in material reflectance estimation under both natural and co-located illumination conditions.

Conclusion: GLOW successfully addresses the challenges of inverse rendering in complex indoor lighting scenarios, particularly for co-located light-camera setups, by properly modeling global illumination effects and handling lighting-specific artifacts.

Abstract: Inverse rendering of indoor scenes remains challenging due to the ambiguity between reflectance and lighting, exacerbated by inter-reflections among multiple objects. While natural illumination-based methods struggle to resolve this ambiguity, co-located light-camera setups offer better disentanglement as lighting can be easily calibrated via Structure-from-Motion. However, such setups introduce additional complexities like strong inter-reflections, dynamic shadows, near-field lighting, and moving specular highlights, which existing approaches fail to handle. We present GLOW, a Global Illumination-aware Inverse Rendering framework designed to address these challenges. GLOW integrates a neural implicit surface representation with a neural radiance cache to approximate global illumination, jointly optimizing geometry and reflectance through carefully designed regularization and initialization. We then introduce a dynamic radiance cache that adapts to sharp lighting discontinuities from near-field motion, and a surface-angle-weighted radiometric loss to suppress specular artifacts common in flashlight captures. Experiments show that GLOW substantially outperforms prior methods in material reflectance estimation under both natural and co-located illumination.

Method: 1) Novel gesture captioning framework using motion-language model to generate descriptive captions at multiple granularities; 2) Conditional latent diffusion model with unified cross-dataset motion representation and hierarchically controlled denoiser for coordinated gesture generation.

Result: Produces high-quality gestures that are both rhythmically synchronized with speeches and semantically coherent with arbitrary captions, achieving superior performance with higher efficiency compared to existing approaches.

Conclusion: CoordSpeaker pioneers gesture understanding and captioning to tackle semantic gaps in gesture generation while offering novel bidirectional gesture-text mapping, enabling comprehensive coordinated co-speech gesture synthesis.

Abstract: Co-speech gesture generation has significantly advanced human-computer interaction, yet speaker movements remain constrained due to the omission of text-driven non-spontaneous gestures (e.g., bowing while talking). Existing methods face two key challenges: 1) the semantic prior gap due to the lack of descriptive text annotations in gesture datasets, and 2) the difficulty in achieving coordinated multimodal control over gesture generation. To address these challenges, this paper introduces CoordSpeaker, a comprehensive framework that enables coordinated caption-empowered co-speech gesture synthesis. Our approach first bridges the semantic prior gap through a novel gesture captioning framework, leveraging a motion-language model to generate descriptive captions at multiple granularities. Building upon this, we propose a conditional latent diffusion model with unified cross-dataset motion representation and a hierarchically controlled denoiser to achieve highly controlled, coordinated gesture generation. CoordSpeaker pioneers the first exploration of gesture understanding and captioning to tackle the semantic gap in gesture generation while offering a novel perspective of bidirectional gesture-text mapping. Extensive experiments demonstrate that our method produces high-quality gestures that are both rhythmically synchronized with speeches and semantically coherent with arbitrary captions, achieving superior performance with higher efficiency compared to existing approaches.

Method: Combines 3D VQ-GAN latent compression with CNN-Transformer backbone and strong task conditioning via AdaLN-Zero and cross-attention. Uses diffusion transformer architecture for conditional generation of voxelwise 4D fMRI data.

Result: On HCP task fMRI: reproduces task-evoked activation maps (correlation 0.83), preserves inter-task representational structure (RSA 0.98), achieves perfect condition specificity, aligns ROI time-courses with canonical hemodynamic responses. Consistently surpasses U-Net baseline on all metrics with predictable performance scaling.

Conclusion: Establishes practical path to conditional 4D fMRI synthesis by coupling latent diffusion with scalable backbone and strong conditioning, enabling future applications like virtual experiments, cross-site harmonization, and principled augmentation for neuroimaging models.

Abstract: Generating whole-brain 4D fMRI sequences conditioned on cognitive tasks remains challenging due to the high-dimensional, heterogeneous BOLD dynamics across subjects/acquisitions and the lack of neuroscience-grounded validation. We introduce the first diffusion transformer for voxelwise 4D fMRI conditional generation, combining 3D VQ-GAN latent compression with a CNN-Transformer backbone and strong task conditioning via AdaLN-Zero and cross-attention. On HCP task fMRI, our model reproduces task-evoked activation maps, preserves the inter-task representational structure observed in real data (RSA), achieves perfect condition specificity, and aligns ROI time-courses with canonical hemodynamic responses. Performance improves predictably with scale, reaching task-evoked map correlation of 0.83 and RSA of 0.98, consistently surpassing a U-Net baseline on all metrics. By coupling latent diffusion with a scalable backbone and strong conditioning, this work establishes a practical path to conditional 4D fMRI synthesis, paving the way for future applications such as virtual experiments, cross-site harmonization, and principled augmentation for downstream neuroimaging models.

Method: Proposed a CNN-based framework classifying pedestrians into six categories (adult/teenager/child × male/female) using full-body visual cues from far-view footage. Tested ResNet50 (pretrained on ImageNet) and a custom lightweight CNN with different pooling strategies and optimizers on video data from three high-risk intersections in Dhaka.

Result: ResNet50 with Max Pooling and SGD achieved highest accuracy (86.19%). Custom lightweight CNN performed comparably (84.15%) with fewer parameters and faster training. Both enable real-time inference on standard surveillance feeds.

Conclusion: The framework provides a scalable, cost-effective tool for demographic monitoring using existing infrastructure. Outputs can inform intersection design, signal timing optimization, and targeted safety interventions for vulnerable groups, supporting more inclusive, data-driven planning in mixed-traffic environments.

Abstract: Pedestrian safety remains a pressing concern in congested urban intersections, particularly in low- and middle-income countries where traffic is multimodal, and infrastructure often lacks formal control. Demographic factors like age and gender significantly influence pedestrian vulnerability, yet real-time monitoring systems rarely capture this information. To address this gap, this study proposes a deep learning framework that classifies pedestrian age group and gender from far-view intersection footage using convolutional neural networks (CNNs), without relying on facial recognition or high-resolution imagery. The classification is structured as a unified six-class problem, distinguishing adult, teenager, and child pedestrians for both males and females, based on full-body visual cues. Video data was collected from three high-risk intersections in Dhaka, Bangladesh. Two CNN architectures were implemented: ResNet50, a deep convolutional neural network pretrained on ImageNet, and a custom lightweight CNN optimized for computational efficiency. Eight model variants explored combinations of pooling strategies and optimizers. ResNet50 with Max Pooling and SGD achieved the highest accuracy (86.19%), while the custom CNN performed comparably (84.15%) with fewer parameters and faster training. The model’s efficient design enables real-time inference on standard surveillance feeds. For practitioners, this system provides a scalable, cost-effective tool to monitor pedestrian demographics at intersections using existing camera infrastructure. Its outputs can shape intersection design, optimize signal timing, and enable targeted safety interventions for vulnerable groups such as children or the elderly. By offering demographic insights often missing in conventional traffic data, the framework supports more inclusive, data-driven planning in mixed-traffic environments.

Method: Develop Vision Bridge Transformer (ViBT) as a large-scale instantiation of Brownian Bridge Models. Use Transformer architecture scaled to 20B and 1.3B parameters. Introduce a variance-stabilized velocity-matching objective for robust training.

Result: Demonstrate effectiveness for image and video translation tasks at scale. Show the power of scaling Bridge Models for instruction-based image editing and complex video translation.

Conclusion: Scaling Bridge Models with Transformer architecture and variance-stabilized training objective enables efficient data-to-data translation, highlighting their potential for complex vision tasks like instruction-based editing and video translation.

Abstract: We introduce Vision Bridge Transformer (ViBT), a large-scale instantiation of Brownian Bridge Models designed for conditional generation. Unlike traditional diffusion models that transform noise into data, Bridge Models directly model the trajectory between inputs and outputs, creating an efficient data-to-data translation paradigm. By scaling these models to 20B and 1.3B parameters, we demonstrate their effectiveness for image and video translation tasks. To support this scale, we adopt a Transformer architecture and propose a variance-stabilized velocity-matching objective for robust training. Together, these advances highlight the power of scaling Bridge Models for instruction-based image editing and complex video translation.

Method: ClearGCD uses two complementary mechanisms: 1) Semantic View Alignment (SVA) generates strong augmentations via cross-class patch replacement and enforces semantic consistency, and 2) Shortcut Suppression Regularization (SSR) maintains an adaptive prototype bank that aligns known classes while encouraging separation of potential novel ones.

Result: ClearGCD consistently outperforms state-of-the-art methods across multiple benchmarks and can be seamlessly integrated into parametric GCD approaches.

Conclusion: ClearGCD effectively mitigates reliance on non-semantic cues in GCD, addressing prototype confusion and improving performance in identifying both known and novel categories in unlabeled data.

Abstract: In open-world scenarios, Generalized Category Discovery (GCD) requires identifying both known and novel categories within unlabeled data. However, existing methods often suffer from prototype confusion caused by shortcut learning, which undermines generalization and leads to forgetting of known classes. We propose ClearGCD, a framework designed to mitigate reliance on non-semantic cues through two complementary mechanisms. First, Semantic View Alignment (SVA) generates strong augmentations via cross-class patch replacement and enforces semantic consistency using weak augmentations. Second, Shortcut Suppression Regularization (SSR) maintains an adaptive prototype bank that aligns known classes while encouraging separation of potential novel ones. ClearGCD can be seamlessly integrated into parametric GCD approaches and consistently outperforms state-of-the-art methods across multiple benchmarks.

Method: Proposes DM³T framework with Cross-Modal Diffusion Fusion (C-MDF) module for iterative cross-modal harmonization, where features provide mutual guidance to project onto shared feature manifold. Includes Diffusion Refiner (DR) to enhance unified features and Hierarchical Tracker for adaptive confidence estimation. Unifies detection, state estimation, and data association without complex post-processing.

Result: Achieves 41.7 HOTA on VT-MOT benchmark, representing 1.54% relative improvement over existing state-of-the-art methods.

Conclusion: DM³T effectively addresses multimodal fusion challenges through iterative feature alignment inspired by diffusion models, demonstrating superior performance for visible-thermal tracking in autonomous driving applications.

Abstract: Multi-object tracking (MOT) is a fundamental task in computer vision with critical applications in autonomous driving and robotics. Multimodal MOT that integrates visible light and thermal infrared information is particularly essential for robust autonomous driving systems. However, effectively fusing these heterogeneous modalities is challenging. Simple strategies like concatenation or addition often fail to bridge the significant non-linear distribution gap between their feature representations, which can lead to modality conflicts and degrade tracking accuracy. Drawing inspiration from the connection between multimodal MOT and the iterative refinement in diffusion models, this paper proposes DM$^3$T, a novel framework that reformulates multimodal fusion as an iterative feature alignment process to generate accurate and temporally coherent object trajectories. Our approach performs iterative cross-modal harmonization through a proposed Cross-Modal Diffusion Fusion (C-MDF) module. In this process, features from both modalities provide mutual guidance, iteratively projecting them onto a shared, consistent feature manifold. This enables the learning of complementary information and achieves deeper fusion compared to conventional methods. Additionally, we introduce a plug-and-play Diffusion Refiner (DR) to enhance and refine the unified feature representation. To further improve tracking robustness, we design a Hierarchical Tracker that adaptively handles confidence estimation. DM$^3$T unifies object detection, state estimation, and data association into a comprehensive online tracking framework without complex post-processing. Extensive experiments on the VT-MOT benchmark demonstrate that our method achieves 41.7 HOTA, representing a 1.54% relative improvement over existing state-of-the-art methods. The code and models are available at https://vranlee.github.io/DM-3-T/.

Method: Frame AGB estimation as dense prediction task using AGB density maps (biomass per pixel normalized by plot/tree area). Use synthetic SPREAD dataset with 3D forest scenes, tree attributes, and segmentation masks. Compute ground truth AGB via allometric equations and train model to predict density maps, then integrate for scene-level estimates.

Result: Achieves median AGB estimation error of 1.22 kg/m² on synthetic SPREAD data and 1.94 kg/m² on real-image dataset. First method to estimate AGB directly from single RGB image.

Conclusion: Proposes scalable, interpretable, cost-effective solution for forest monitoring enabling citizen science participation. Opens new possibilities for biomass estimation from simple ground-based RGB images.

Abstract: Forests play a critical role in global ecosystems by supporting biodiversity and mitigating climate change via carbon sequestration. Accurate aboveground biomass (AGB) estimation is essential for assessing carbon storage and wildfire fuel loads, yet traditional methods rely on labor-intensive field measurements or remote sensing approaches with significant limitations in dense vegetation. In this work, we propose a novel learning-based method for estimating AGB from a single ground-based RGB image. We frame this as a dense prediction task, introducing AGB density maps, where each pixel represents tree biomass normalized by the plot area and each tree’s image area. We leverage the recently introduced synthetic 3D SPREAD dataset, which provides realistic forest scenes with per-image tree attributes (height, trunk and canopy diameter) and instance segmentation masks. Using these assets, we compute AGB via allometric equations and train a model to predict AGB density maps, integrating them to recover the AGB estimate for the captured scene. Our approach achieves a median AGB estimation error of 1.22 kg/m^2 on held-out SPREAD data and 1.94 kg/m^2 on a real-image dataset. To our knowledge, this is the first method to estimate aboveground biomass directly from a single RGB image, opening up the possibility for a scalable, interpretable, and cost-effective solution for forest monitoring, while also enabling broader participation through citizen science initiatives.

Method: Introduces Points-to-Clouds (P2C) framework with dual denoising: Dynamic Prompt Denoising perturbs text prompts with annealed noise to learn smoother semantic landscapes, and an auxiliary V-L Mapper denoising loss re-tasks the mapper as a denoising autoencoder for robust cross-modal alignment.

Result: Extensive experiments across 11 datasets show P2C consistently outperforms baselines, achieving 79.7% Harmonic Mean on base-to-novel generalization benchmark (1.4% relative improvement over baseline).

Conclusion: Learning semantic clouds through dynamic denoising provides more robust generalization than static point representations in multimodal prompt learning.

Abstract: Multimodal Prompt Learning (MPL) has emerged as a pivotal technique for adapting large-scale Visual Language Models (VLMs). However, current MPL methods are fundamentally limited by their optimization of a single, static point representation. This paradigm is inherently brittle, leads to overfitting on base classes, and generalizes poorly to novel or ambiguous categories. We challenge this point paradigm, proposing that robust generalization requires learning a semantic cloud (i.e., a distribution over the embedding space). To achieve this, we introduce Points-to-Clouds (P2C), a novel framework inspired by diffusion models that reframes prompt learning as a dynamic denoising task. At the core of P2C is a dual denoising mechanism: a Dynamic Prompt Denoising (DPD) mechanism perturbs text prompts with sophisticated, annealed noise to learn a smoother semantic landscape, while an auxiliary V-L Mapper denoising loss re-tasks the mapper as a denoising autoencoder. This forces the mapper to reconstruct clean visual prompts from noisy text inputs, ensuring robust cross-modal alignment. Extensive experiments across 11 datasets demonstrate that P2C consistently outperforms strong baselines. On the base-to-novel generalization benchmark, our method achieves a Harmonic Mean of 79.7%, representing a relative improvement of 1.4% over the baseline. The code and models are available at https://vranlee.github.io/P2C/.

Method: Integrates image-text scene understanding via Multimodal Large Language Models (MLLMs), introduces a feature enhancement module (FEM) to capture important words from queries, and a ranking-based filtering module (RFM) to iteratively refine video clips based on word relevance.

Result: Extensive experiments show the approach significantly outperforms existing state-of-the-art methods, achieving superior performance in both video moment retrieval and highlight detection tasks.

Conclusion: The proposed fine-grained approach that prioritizes important words in queries enables better contextual understanding and more accurate video localization compared to black-box methods.

Abstract: Video moment retrieval (MR) and highlight detection (HD) with natural language queries aim to localize relevant moments and key highlights in a video clips. However, existing methods overlook the importance of individual words, treating the entire text query and video clips as a black-box, which hinders contextual understanding. In this paper, we propose a novel approach that enables fine-grained clip filtering by identifying and prioritizing important words in the query. Our method integrates image-text scene understanding through Multimodal Large Language Models (MLLMs) and enhances the semantic understanding of video clips. We introduce a feature enhancement module (FEM) to capture important words from the query and a ranking-based filtering module (RFM) to iteratively refine video clips based on their relevance to these important words. Extensive experiments demonstrate that our approach significantly outperforms existing state-of-the-art methods, achieving superior performance in both MR and HD tasks. Our code is available at: https://github.com/VisualAIKHU/SRF.

Method: ViGG uses a mutual guidance strategy: 1) Geometric guidance design to suppress ambiguous cliques in visual-geometric combination form, and 2) Visual-guided geometric matching that uses visual priors to determine search space for extracting high-quality, noise-insensitive correspondences.

Result: Experiments on 3DMatch, ScanNet and KITTI datasets show that ViGG outperforms recent state-of-the-art methods in both learning-free and learning-based settings.

Conclusion: The mutual guidance strategy provides superior robustness, making ViGG applicable for various RGB-D registration tasks, with code publicly available.

Abstract: Point cloud registration is a fundamental task in 3D vision. Most existing methods only use geometric information for registration. Recently proposed RGB-D registration methods primarily focus on feature fusion or improving feature learning, which limits their ability to exploit image information and hinders their practical applicability. In this paper, we propose ViGG, a robust RGB-D registration method using mutual guidance. First, we solve clique alignment in a visual-geometric combination form, employing a geometric guidance design to suppress ambiguous cliques. Second, to mitigate accuracy degradation caused by noise in visual matches, we propose a visual-guided geometric matching method that utilizes visual priors to determine the search space, enabling the extraction of high-quality, noise-insensitive correspondences. This mutual guidance strategy brings our method superior robustness, making it applicable for various RGB-D registration tasks. The experiments on 3DMatch, ScanNet and KITTI datasets show that our method outperforms recent state-of-the-art methods in both learning-free and learning-based settings. Code is available at https://github.com/ccjccjccj/ViGG.

Method: USArt employs a dual sub-model approach customized for 2D brain anatomical images with Cartesian sampling. It simultaneously handles under-sampling acceleration and artifact correction (noise and motion artifacts). The method explores various under-sampling strategies, with gradient under-sampling yielding best results.

Result: The model achieves up to 5x acceleration with simultaneous artifact correction without significant degradation. Results show remarkable increase in signal-to-noise ratio (SNR) and contrast in restored images. The method demonstrates robustness across various under-sampling strategies and degradation levels.

Conclusion: USArt successfully addresses the gap in MRI reconstruction by simultaneously handling acceleration and artifact correction, providing a robust solution for real-world MRI settings where both degradation factors commonly occur together.

Abstract: MR data are acquired in the frequency domain, known as k-space. Acquiring high-quality and high-resolution MR images can be time-consuming, posing a significant challenge when multiple sequences providing complementary contrast information are needed or when the patient is unable to remain in the scanner for an extended period of time. Reducing k-space measurements is a strategy to speed up acquisition, but often leads to reduced quality in reconstructed images. Additionally, in real-world MRI, both under-sampled and full-sampled images are prone to artefacts, and correcting these artefacts is crucial for maintaining diagnostic accuracy. Deep learning methods have been proposed to restore image quality from under-sampled data, while others focused on the correction of artefacts that result from the noise or motion. No approach has however been proposed so far that addresses both acceleration and artefacts correction, limiting the performance of these models when these degradation factors occur simultaneously. To address this gap, we present a method for recovering high-quality images from under-sampled data with simultaneously correction for noise and motion artefact called USArt (Under-Sampling and Artifact correction model). Customized for 2D brain anatomical images acquired with Cartesian sampling, USArt employs a dual sub-model approach. The results demonstrate remarkable increase of signal-to-noise ratio (SNR) and contrast in the images restored. Various under-sampling strategies and degradation levels were explored, with the gradient under-sampling strategy yielding the best outcomes. We achieved up to 5x acceleration and simultaneously artefacts correction without significant degradation, showcasing the model’s robustness in real-world settings.

Method: Proposes Neural Matrix Computation Framework (NeuMatC) that unsupervisedly learns a low-rank continuous mapping from parameters to matrix operation results. Once trained, it computes results at arbitrary parameters using only basic operations like matrix multiplications and nonlinear activations.

Result: Experimental results show promising performance: over 3× speedup in parametric inversion and 10× speedup in parametric SVD compared to NumPy baseline in wireless communication applications, while maintaining acceptable accuracy.

Conclusion: NeuMatC effectively addresses computational redundancy in parametric matrix operations by leveraging low-rankness and continuity, offering significant speed improvements for real-world applications like wireless communication and signal processing.

Abstract: Matrix operations (e.g., inversion and singular value decomposition (SVD)) are fundamental in science and engineering. In many emerging real-world applications (such as wireless communication and signal processing), these operations must be performed repeatedly over matrices with parameters varying continuously. However, conventional methods tackle each matrix operation independently, underexploring the inherent low-rankness and continuity along the parameter dimension, resulting in significantly redundant computation. To address this challenge, we propose \textbf{\textit{Neural Matrix Computation Framework} (NeuMatC)}, which elegantly tackles general parametric matrix operation tasks by leveraging the underlying low-rankness and continuity along the parameter dimension. Specifically, NeuMatC unsupervisedly learns a low-rank and continuous mapping from parameters to their corresponding matrix operation results. Once trained, NeuMatC enables efficient computations at arbitrary parameters using only a few basic operations (e.g., matrix multiplications and nonlinear activations), significantly reducing redundant computations. Experimental results on both synthetic and real-world datasets demonstrate the promising performance of NeuMatC, exemplified by over $3\times$ speedup in parametric inversion and $10\times$ speedup in parametric SVD compared to the widely used NumPy baseline in wireless communication, while maintaining acceptable accuracy.

Method: Proposes Rectified MeanFlow (Re-MeanFlow) that models the mean velocity field along rectified trajectories using only a single reflow step, eliminating need for perfectly straightened paths. Also introduces a truncation heuristic to reduce residual curvature.

Result: Extensive experiments on ImageNet at 64, 256, and 512 resolutions show Re-MeanFlow consistently outperforms prior one-step flow distillation and Rectified Flow methods in both sample quality and training efficiency.

Conclusion: Re-MeanFlow provides an effective framework for efficient one-step generation in flow-based models by balancing trajectory straightness requirements with training efficiency through mean velocity modeling and rectification.

Abstract: Flow-based generative models have recently demonstrated strong performance, yet sampling typically relies on expensive numerical integration of ordinary differential equations (ODEs). Rectified Flow enables one-step sampling by learning nearly straight probability paths, but achieving such straightness requires multiple computationally intensive reflow iterations. MeanFlow achieves one-step generation by directly modeling the average velocity over time; however, when trained on highly curved flows, it suffers from slow convergence and noisy supervision. To address these limitations, we propose Rectified MeanFlow, a framework that models the mean velocity field along the rectified trajectory using only a single reflow step. This eliminates the need for perfectly straightened trajectories while enabling efficient training. Furthermore, we introduce a simple yet effective truncation heuristic that aims to reduce residual curvature and further improve performance. Extensive experiments on ImageNet at 64, 256, and 512 resolutions show that Re-MeanFlow consistently outperforms prior one-step flow distillation and Rectified Flow methods in both sample quality and training efficiency. Code is available at https://github.com/Xinxi-Zhang/Re-MeanFlow.

Method: ReImage uses neural watermarking to embed a shuffled version of the target image into itself as a watermark. It includes a generator optimized for neural watermarking and an image enhancement module to refine recovered images. The framework addresses key limitations of shuffled watermarking for self-recovery applications.

Result: ReImage achieves state-of-the-art performance across diverse tampering scenarios, consistently producing high-quality recovered images.

Conclusion: The proposed neural watermarking-based self-recovery framework effectively addresses the limitations of existing methods and provides a practical solution for understanding attacker intent and restoring trustworthy data from manipulated images.

Abstract: The rapid growth of Artificial Intelligence-Generated Content (AIGC) raises concerns about the authenticity of digital media. In this context, image self-recovery, reconstructing original content from its manipulated version, offers a practical solution for understanding the attacker’s intent and restoring trustworthy data. However, existing methods often fail to accurately recover tampered regions, falling short of the primary goal of self-recovery. To address this challenge, we propose ReImage, a neural watermarking-based self-recovery framework that embeds a shuffled version of the target image into itself as a watermark. We design a generator that produces watermarks optimized for neural watermarking and introduce an image enhancement module to refine the recovered image. We further analyze and resolve key limitations of shuffled watermarking, enabling its effective use in self-recovery. We demonstrate that ReImage achieves state-of-the-art performance across diverse tampering scenarios, consistently producing high-quality recovered images. The code and pretrained models will be released upon publication.

Method: Used YOLOv8 algorithm with extensive training and fine-tuning on Kaggle datasets for barcode/QR code detection. Evaluated three model iterations (Nano, Small, Medium) focusing on precision, recall, and F1 metrics.

Result: Achieved incremental accuracy improvements: 88.95% for Nano model, 97.10% for Small model, and 94.10% for Medium model, demonstrating significant enhancements in object detection accuracy through model refinement.

Conclusion: YOLOv8 shows substantial progress in computer vision object detection, with model scaling significantly affecting recognition performance, advancing deep learning-based computer vision techniques.

Abstract: This research work dives into an in-depth evaluation of the YOLOv8 (You Only Look Once) algorithm’s efficiency in object detection, specially focusing on Barcode and QR code recognition. Utilizing the real-time detection abilities of YOLOv8, we performed a study aimed at enhancing its talent in swiftly and correctly figuring out objects. Through large training and high-quality-tuning on Kaggle datasets tailored for Barcode and QR code detection, our goal became to optimize YOLOv8’s overall performance throughout numerous situations and environments. The look encompasses the assessment of YOLOv8 throughout special version iterations: Nano, Small, and Medium, with a meticulous attention on precision, recall, and F1 assessment metrics. The consequences exhibit large improvements in object detection accuracy with every subsequent model refinement. Specifically, we achieved an accuracy of 88.95% for the nano model, 97.10% for the small model, and 94.10% for the medium version, showcasing the incremental improvements finished via model scaling. Our findings highlight the big strides made through YOLOv8 in pushing the limits of computer vision, ensuring its function as a milestone within the subject of object detection. This study sheds light on how model scaling affects object recognition, increasing the concept of deep learning-based computer creative and prescient techniques.

Method: Uses 3D Gaussian Splatting for burst denoising with two key innovations: 1) Gaussian self-consistency (GSC) loss to regularize geometry from noisy inputs using high-quality Gaussian point clouds from clean inputs, and 2) Log-weighted frequency (LWF) loss to preserve fine details by adaptively weighting frequency discrepancies in logarithmic manner.

Result: Significantly exceeds state-of-the-art NeRF-based methods on both burst denoising and novel view synthesis under noisy conditions, while achieving 250× faster inference speed.

Conclusion: DenoiseGS demonstrates that 3D Gaussian Splatting can be effectively adapted for burst denoising, offering superior performance and dramatically faster inference compared to previous approaches, making it practical for handheld device applications.

Abstract: Burst denoising methods are crucial for enhancing images captured on handheld devices, but they often struggle with large motion or suffer from prohibitive computational costs. In this paper, we propose DenoiseGS, the first framework to leverage the efficiency of 3D Gaussian Splatting for burst denoising. Our approach addresses two key challenges when applying feedforward Gaussian reconsturction model to noisy inputs: the degradation of Gaussian point clouds and the loss of fine details. To this end, we propose a Gaussian self-consistency (GSC) loss, which regularizes the geometry predicted from noisy inputs with high-quality Gaussian point clouds. These point clouds are generated from clean inputs by the same model that we are training, thereby alleviating potential bias or domain gaps. Additionally, we introduce a log-weighted frequency (LWF) loss to strengthen supervision within the spectral domain, effectively preserving fine-grained details. The LWF loss adaptively weights frequency discrepancies in a logarithmic manner, emphasizing challenging high-frequency details. Extensive experiments demonstrate that DenoiseGS significantly exceeds the state-of-the-art NeRF-based methods on both burst denoising and novel view synthesis under noisy conditions, while achieving \textbf{250$\times$} faster inference speed. Code and models are released at https://github.com/yscheng04/DenoiseGS.

Method: 1) Reformulate training as self-supervised outpainting to transform diverse-layout references into unified occluded-input format; 2) Design reference extractor for comprehensive identity feature extraction from partially visible references; 3) Integrate hybrid reference fusion attention for varying resolutions and dynamic sequence lengths; 4) Introduce identity-robust pose control to decouple appearance from skeletal structure; 5) Implement token replace strategy for coherent long-video generation.

Result: Extensive experiments show that the method outperforms existing approaches in character animation and image pose transfer for references with arbitrary layouts and spatial misalignment.

Conclusion: One-to-All Animation provides a unified framework that successfully addresses the challenge of reference-pose misalignment in character animation, enabling high-fidelity animation and pose transfer for references with arbitrary layouts through innovative training reformulation, feature extraction, and generation quality improvements.

Abstract: Recent advances in diffusion models have greatly improved pose-driven character animation. However, existing methods are limited to spatially aligned reference-pose pairs with matched skeletal structures. Handling reference-pose misalignment remains unsolved. To address this, we present One-to-All Animation, a unified framework for high-fidelity character animation and image pose transfer for references with arbitrary layouts. First, to handle spatially misaligned reference, we reformulate training as a self-supervised outpainting task that transforms diverse-layout reference into a unified occluded-input format. Second, to process partially visible reference, we design a reference extractor for comprehensive identity feature extraction. Further, we integrate hybrid reference fusion attention to handle varying resolutions and dynamic sequence lengths. Finally, from the perspective of generation quality, we introduce identity-robust pose control that decouples appearance from skeletal structure to mitigate pose overfitting, and a token replace strategy for coherent long-video generation. Extensive experiments show that our method outperforms existing approaches. The code and model will be available at https://github.com/ssj9596/One-to-All-Animation.

Method: Three components: 1) Granular Adaptive Prototypes for multi-scale boundary learning, 2) Uncertainty Boundary Emphasis that adjusts learning based on prediction entropy, 3) Hardness-Aware Sampling that progressively focuses on challenging examples.

Result: Consistent improvements across five real-world datasets, achieving 2.44% and 2.63% mIoU gains on ACDC and Dark Zurich datasets, outperforming state-of-the-art methods.

Conclusion: Adaptive strategies for handling imperfect synthetic data (leveraging misalignment rather than enforcing strict alignment) lead to superior domain generalization in semantic segmentation.

Abstract: Domain generalization in semantic segmentation faces challenges from domain shifts, particularly under adverse conditions. While diffusion-based data generation methods show promise, they introduce inherent misalignment between generated images and semantic masks. This paper presents FLEX-Seg (FLexible Edge eXploitation for Segmentation), a framework that transforms this limitation into an opportunity for robust learning. FLEX-Seg comprises three key components: (1) Granular Adaptive Prototypes that captures boundary characteristics across multiple scales, (2) Uncertainty Boundary Emphasis that dynamically adjusts learning emphasis based on prediction entropy, and (3) Hardness-Aware Sampling that progressively focuses on challenging examples. By leveraging inherent misalignment rather than enforcing strict alignment, FLEX-Seg learns robust representations while capturing rich stylistic variations. Experiments across five real-world datasets demonstrate consistent improvements over state-of-the-art methods, achieving 2.44% and 2.63% mIoU gains on ACDC and Dark Zurich. Our findings validate that adaptive strategies for handling imperfect synthetic data lead to superior domain generalization. Code is available at https://github.com/VisualScienceLab-KHU/FLEX-Seg.

Method: Built on SAM 2 foundation model with three key innovations: 1) structure-enhanced memory associator for articulated robots, 2) robot prompt generator to eliminate manual prompts, 3) label-efficient training strategy reducing need for per-frame annotations.

Result: Created VRS dataset with 2.8k videos (138k frames) of diverse robots/environments. Extensive experiments show RobotSeg achieves state-of-the-art performance on both image and video robot segmentation.

Conclusion: RobotSeg provides a structure-aware, automatic, and label-efficient foundation model for robot segmentation, establishing a strong basis for future advances in robot perception.

Abstract: Accurate robot segmentation is a fundamental capability for robotic perception. It enables precise visual servoing for VLA systems, scalable robot-centric data augmentation, accurate real-to-sim transfer, and reliable safety monitoring in dynamic human-robot environments. Despite the strong capabilities of modern segmentation models, surprisingly it remains challenging to segment robots. This is due to robot embodiment diversity, appearance ambiguity, structural complexity, and rapid shape changes. Embracing these challenges, we introduce RobotSeg, a foundation model for robot segmentation in image and video. RobotSeg is built upon the versatile SAM 2 foundation model but addresses its three limitations for robot segmentation, namely the lack of adaptation to articulated robots, reliance on manual prompts, and the need for per-frame training mask annotations, by introducing a structure-enhanced memory associator, a robot prompt generator, and a label-efficient training strategy. These innovations collectively enable a structure-aware, automatic, and label-efficient solution. We further construct the video robot segmentation (VRS) dataset comprising over 2.8k videos (138k frames) with diverse robot embodiments and environments. Extensive experiments demonstrate that RobotSeg achieves state-of-the-art performance on both images and videos, establishing a strong foundation for future advances in robot perception.

Method: A multi-granularity contrastive pretraining framework that jointly aligns: (1) global class tokens across co-temporal AIA-HMI pairs to enhance temporal discrimination, (2) local patch tokens at fixed spatial indices to enforce position-consistent, modality-invariant features, and (3) intra-sample patches across different spatial locations to preserve fine-grained spatial structure. Trained both CNN- and Vision Transformer-based autoencoders.

Result: SolarCHIP achieves state-of-the-art performance on cross-modal translation between HMI and AIA passbands via ControlNet and full-disk flare classification, with particularly strong gains in low-resource settings where labeled data is limited. Ablation studies confirm each contrastive component contributes essential discriminative capacity at different granularities.

Conclusion: SolarCHIP provides the heliophysics community with a practical, plug-and-play feature extractor that reduces computational requirements, improves label efficiency, and establishes a reusable foundation for diverse solar imaging applications. Pretrained weights and training code are publicly released.

Abstract: Deep learning has revolutionized solar image analysis, yet most approaches train task-specific encoders from scratch or rely on natural-image pretraining that ignores the unique characteristics of Solar Dynamics Observatory (SDO) data. We introduce SolarCHIP, a family of contrastively pretrained visual backbones tailored to multi-instrument SDO observations. SolarCHIP addresses three key challenges in solar imaging: multimodal sensing across AIA and HMI instruments, weak inter-class separability due to slow temporal evolution, and strong intra-class variability with sparse activity signals. Our pretraining framework employs a multi-granularity contrastive objective that jointly aligns (1) global class tokens across co-temporal AIA-HMI pairs to enhance temporal discrimination, (2) local patch tokens at fixed spatial indices to enforce position-consistent, modality-invariant features, and (3) intra-sample patches across different spatial locations to preserve fine-grained spatial structure. We train both CNN- and Vision Transformer-based autoencoders and demonstrate their effectiveness on two downstream tasks: cross-modal translation between HMI and AIA passbands via ControlNet, and full-disk flare classification. Experimental results show that SolarCHIP achieves state-of-the-art performance across both tasks, with particularly strong gains in low-resource settings where labeled data is limited. Ablation studies confirm that each contrastive component contributes essential discriminative capacity at different granularities. By publicly releasing pretrained weights and training code, we provide the heliophysics community with a practical, plug-and-play feature extractor that reduces computational requirements, improves label efficiency, and establishes a reusable foundation for diverse solar imaging applications.

Method: Proposes hierarchical multimodal representation that explicitly aligns with VLMs at input space using multi-view images (top-down + four directional views) and text descriptions with 3D object coordinates. Uses hierarchical feature aggregation from patch-level to view-level to scene-level representations.

Result: Experimental results demonstrate effectiveness on both situated 3D Q&A and general 3D Q&A benchmarks.

Conclusion: The proposed explicit alignment approach with hierarchical multimodal representation improves 3D scene reasoning by better capturing spatial relationships and comprehensive scene coverage.

Abstract: Recent advances in large vision-language models (VLMs) have shown significant promise for 3D scene understanding. Existing VLM-based approaches typically align 3D scene features with the VLM’s embedding space. However, this implicit alignment often yields suboptimal performance due to the scarcity of 3D data and the inherent complexity of spatial relationships in 3D environments. To address these limitations, we propose a novel hierarchical multimodal representation for 3D scene reasoning that explicitly aligns with VLMs at the input space by leveraging both multi-view images and text descriptions. The text descriptions capture spatial relationships by referencing the 3D coordinates of detected objects, while the multi-view images include a top-down perspective and four directional views (forward, left, right, and backward), ensuring comprehensive scene coverage. Additionally, we introduce a hierarchical feature representation that aggregates patch-level image features into view-level and scene-level representations, enabling the model to reason over both local and global scene context. Experimental results on both situated 3D Q&A and general 3D Q&A benchmarks demonstrate the effectiveness of our approach.

Method: Two key designs: 1) Intrinsic Appearance Normalization (IAN) module that proactively disentangles scene’s intrinsic properties (albedo) from transient lighting to create illumination-invariant appearance model; 2) Dynamic Radiance Balancing Loss (DRB-Loss) that reactively handles frames with extreme exposure by activating only when exposure is poor and operating directly on radiance field.

Result: Demonstrates state-of-the-art performance on public datasets in camera tracking, map quality, and semantic and geometric accuracy with unprecedented robustness to extreme lighting conditions.

Conclusion: The synergy between IAN’s proactive invariance and DRB-Loss’s reactive correction creates a robust semantic SLAM framework that maintains performance under extreme exposure conditions without compromising normal operation.

Abstract: Extreme exposure degrades both the 3D map reconstruction and semantic segmentation accuracy, which is particularly detrimental to tightly-coupled systems. To achieve illumination invariance, we propose a novel semantic SLAM framework with two designs. First, the Intrinsic Appearance Normalization (IAN) module proactively disentangles the scene’s intrinsic properties, such as albedo, from transient lighting. By learning a standardized, illumination-invariant appearance model, it assigns a stable and consistent color representation to each Gaussian primitive. Second, the Dynamic Radiance Balancing Loss (DRB-Loss) reactively handles frames with extreme exposure. It activates only when an image’s exposure is poor, operating directly on the radiance field to guide targeted optimization. This prevents error accumulation from extreme lighting without compromising performance under normal conditions. The synergy between IAN’s proactive invariance and DRB-Loss’s reactive correction endows our system with unprecedented robustness. Evaluations on public datasets demonstrate state-of-the-art performance in camera tracking, map quality, and semantic and geometric accuracy.

Method: BlockVid introduces: 1) semantic-aware sparse KV cache to reduce error propagation, 2) Block Forcing training strategy, 3) chunk-wise noise scheduling and shuffling for enhanced temporal consistency. Also introduces LV-Bench, a fine-grained minute-long video benchmark with new coherence metrics.

Result: BlockVid consistently outperforms existing methods, achieving 22.2% improvement on VDE Subject and 19.4% improvement on VDE Clarity in LV-Bench over state-of-the-art approaches. Demonstrates superior performance on both VBench and LV-Bench benchmarks.

Conclusion: BlockVid effectively addresses key limitations in minute-long video generation through its novel block diffusion framework with semantic-aware sparse KV cache and specialized training strategies, enabling high-quality, coherent extended video generation with significant improvements over existing methods.

Abstract: Generating minute-long videos is a critical step toward developing world models, providing a foundation for realistic extended scenes and advanced AI simulators. The emerging semi-autoregressive (block diffusion) paradigm integrates the strengths of diffusion and autoregressive models, enabling arbitrary-length video generation and improving inference efficiency through KV caching and parallel sampling. However, it yet faces two enduring challenges: (i) KV-cache-induced long-horizon error accumulation, and (ii) the lack of fine-grained long-video benchmarks and coherence-aware metrics. To overcome these limitations, we propose BlockVid, a novel block diffusion framework equipped with semantic-aware sparse KV cache, an effective training strategy called Block Forcing, and dedicated chunk-wise noise scheduling and shuffling to reduce error propagation and enhance temporal consistency. We further introduce LV-Bench, a fine-grained benchmark for minute-long videos, complete with new metrics evaluating long-range coherence. Extensive experiments on VBench and LV-Bench demonstrate that BlockVid consistently outperforms existing methods in generating high-quality, coherent minute-long videos. In particular, it achieves a 22.2% improvement on VDE Subject and a 19.4% improvement on VDE Clarity in LV-Bench over the state of the art approaches. Project website: https://ziplab.co/BlockVid. Inferix (Code): https://github.com/alibaba-damo-academy/Inferix.

Method: Three-stage framework: 1) Self-critic Dimensional Reasoning (ScDR) trains a generative reward model to decompose preferences into per-dimension assessments using self-critic reasoning chains; 2) Hierarchical Comparative Reasoning (HCR) enables structural multi-dimensional reasoning with hierarchical reward supervision; 3) Motion-corrective Direct Preference Optimization (McDPO) optimizes T2V models using RM-preferred videos while dynamically re-weighting alignment objectives to mitigate low-motion bias.

Result: McSc achieves superior performance in human preference alignment and generates videos with high-motion dynamic, addressing the bias toward low-motion content that plagues existing methods.

Conclusion: The proposed McSc framework provides a robust solution for aligning T2V generation with nuanced human preferences by decomposing preferences, enabling hierarchical reasoning, and correcting motion bias, leading to better-aligned and more dynamic video generation.

Abstract: Text-to-video (T2V) generation has achieved remarkable progress in producing high-quality videos aligned with textual prompts. However, aligning synthesized videos with nuanced human preference remains challenging due to the subjective and multifaceted nature of human judgment. Existing video preference alignment methods rely on costly human annotations or utilize proxy metrics to predict preference, which lacks the understanding of human preference logic. Moreover, they usually directly align T2V models with the overall preference distribution, ignoring potential conflict dimensions like motion dynamics and visual quality, which may bias models towards low-motion content. To address these issues, we present Motion-corrective alignment with Self-critic hierarchical Reasoning (McSc), a three-stage reinforcement learning framework for robust preference modeling and alignment. Firstly, Self-critic Dimensional Reasoning (ScDR) trains a generative reward model (RM) to decompose preferences into per-dimension assessments, using self-critic reasoning chains for reliable learning. Secondly, to achieve holistic video comparison, we introduce Hierarchical Comparative Reasoning (HCR) for structural multi-dimensional reasoning with hierarchical reward supervision. Finally, using RM-preferred videos, we propose Motion-corrective Direct Preference Optimization (McDPO) to optimize T2V models, while dynamically re-weighting alignment objective to mitigate bias towards low-motion content. Experiments show that McSc achieves superior performance in human preference alignment and generates videos with high-motion dynamic.

Method: Treat convolutional features as Gaussian-distributed feature signal matrices and apply a simple low-amplitude pass filter called Convolutional Feature Filter (CFF) to minimize noise in feature inputs.

Result: Experiments on two established 2D segmentation networks and two public cardiac MR image datasets showed noise reduction in feature signal matrices, validated using a developed binarization equation to calculate feature signal information entropy.

Conclusion: The proposed CFF effectively reduces noise in convolutional features for segmentation networks, addressing a previously overlooked aspect that can significantly impact overall segmentation performance.

Abstract: Noise reduction constitutes a crucial operation within Digital Signal Processing. Regrettably, it frequently remains neglected when dealing with the processing of convolutional features in segmentation networks. This oversight could trigger the butterfly effect, impairing the subsequent outcomes within the entire feature system. To complete this void, we consider convolutional features following Gaussian distributions as feature signal matrices and then present a simple and effective feature filter in this study. The proposed filter is fundamentally a low-amplitude pass filter primarily aimed at minimizing noise in feature signal inputs and is named Convolutional Feature Filter (CFF). We conducted experiments on two established 2D segmentation networks and two public cardiac MR image datasets to validate the effectiveness of the CFF, and the experimental findings demonstrated a decrease in noise within the feature signal matrices. To enable a numerical observation and analysis of this reduction, we developed a binarization equation to calculate the information entropy of feature signals.

Method: The authors introduce MultiBanana, a carefully designed benchmark that systematically assesses model capabilities by covering five key multi-reference-specific problems at scale: (1) varying number of references, (2) domain mismatch among references, (3) scale mismatch between reference and target scenes, (4) references containing rare concepts, and (5) multilingual textual references for rendering.

Result: Analysis of various text-to-image models using MultiBanana reveals their superior performances, typical failure modes, and areas for improvement. The benchmark enables comprehensive evaluation of multi-reference generation capabilities across diverse challenging scenarios.

Conclusion: MultiBanana addresses the gap in multi-reference text-to-image generation evaluation by providing a standardized, comprehensive benchmark. It will be released as an open benchmark to push boundaries and establish fair comparison standards in multi-reference image generation, with data and code publicly available.

Abstract: Recent text-to-image generation models have acquired the ability of multi-reference generation and editing; the ability to inherit the appearance of subjects from multiple reference images and re-render them under new contexts. However, the existing benchmark datasets often focus on the generation with single or a few reference images, which prevents us from measuring the progress on how model performance advances or pointing out their weaknesses, under different multi-reference conditions. In addition, their task definitions are still vague, typically limited to axes such as “what to edit” or “how many references are given”, and therefore fail to capture the intrinsic difficulty of multi-reference settings. To address this gap, we introduce $\textbf{MultiBanana}$, which is carefully designed to assesses the edge of model capabilities by widely covering multi-reference-specific problems at scale: (1) varying the number of references, (2) domain mismatch among references (e.g., photo vs. anime), (3) scale mismatch between reference and target scenes, (4) references containing rare concepts (e.g., a red banana), and (5) multilingual textual references for rendering. Our analysis among a variety of text-to-image models reveals their superior performances, typical failure modes, and areas for improvement. MultiBanana will be released as an open benchmark to push the boundaries and establish a standardized basis for fair comparison in multi-reference image generation. Our data and code are available at https://github.com/matsuolab/multibanana .

Method: Proposes a spectrum-weakening framework that constructs a controllable weak model in the spectral domain. Uses invertible spectral transformations that preserve information while selectively retaining only a subset of spectrum for controlled information reduction. Performs spectrum selection along the channel dimension of internal representations to avoid diffusion model constraints. Introduces two spectrum renormalization strategies for numerical stability during weakening.

Result: Extensive experiments on both discrete and continuous autoregressive models with text or class conditioning demonstrate that the method enables high-quality unconditional generation while maintaining strong prompt alignment for conditional generation.

Conclusion: The spectrum-weakening framework provides an effective guidance mechanism for visual autoregressive models that works without retraining or architectural changes, overcoming the diffusion-model-specific limitations of previous approaches.

Abstract: Classifier-free guidance (CFG) has become a widely adopted and practical approach for enhancing generation quality and improving condition alignment. Recent studies have explored guidance mechanisms for unconditional generation, yet these approaches remain fundamentally tied to assumptions specific to diffusion models. In this work, we propose a spectrum-weakening framework for visual autoregressive (AR) models. This method works without the need for re-training, specific conditions, or any architectural modifications. It achieves this by constructing a controllable weak model in the spectral domain. We theoretically show that invertible spectral transformations preserve information, while selectively retaining only a subset of spectrum introduces controlled information reduction. Based on this insight, we perform spectrum selection along the channel dimension of internal representations, which avoids the structural constraints imposed by diffusion models. We further introduce two spectrum renormalization strategies that ensures numerical stability during the weakening process. Extensive experiments were conducted on both discrete and continuous AR models, with text or class conditioning. The results demonstrate that our method enables high-quality unconditional generation while maintaining strong prompt alignment for conditional generation.

Method: Evaluates how different optimizers influence the accuracy and stability of Vision Transformers (ViT) for iris recognition.

Result: The paper provides insights into optimizer effects on ViT performance for iris recognition, though specific results aren’t detailed in the abstract.

Conclusion: Optimizer choice significantly impacts the accuracy and stability of ViT-based iris recognition systems, and understanding these effects can enhance the robustness of biometric identification models.

Abstract: The security of biometric authentication is increasingly critical as digital identity systems expand. Iris recognition offers high reliability due to its distinctive and stable texture patterns. Recent progress in deep learning, especially Vision Transformers ViT, has improved visual recognition performance. Yet, the effect of optimizer choice on ViT-based biometric systems remains understudied. This work evaluates how different optimizers influence the accuracy and stability of ViT for iris recognition, providing insights to enhance the robustness of biometric identification models.

Method: MrGS uses orthogonal feature extraction from a single appearance feature for RGB and thermal information, with view-dependent/independent embedding based on Lambertian reflectance. It incorporates Fourier’s law for heat conduction between Gaussians and uses Stefan-Boltzmann/inverse-square laws for depth-aware thermal radiation mapping.

Result: MrGS achieves high-fidelity RGB-T scene reconstruction while reducing the number of Gaussians needed compared to existing approaches.

Conclusion: The proposed MrGS framework successfully integrates physics-based thermal modeling with 3DGS for multi-modal RGB-T reconstruction, addressing the limitations of current methods in handling thermal characteristics.

Abstract: Recent advances in Neural Radiance Fields (NeRFs) and 3D Gaussian Splatting (3DGS) have achieved considerable performance in RGB scene reconstruction. However, multi-modal rendering that incorporates thermal infrared imagery remains largely underexplored. Existing approaches tend to neglect distinctive thermal characteristics, such as heat conduction and the Lambertian property. In this study, we introduce MrGS, a multi-modal radiance field based on 3DGS that simultaneously reconstructs both RGB and thermal 3D scenes. Specifically, MrGS derives RGB- and thermal-related information from a single appearance feature through orthogonal feature extraction and employs view-dependent or view-independent embedding strategies depending on the degree of Lambertian reflectance exhibited by each modality. Furthermore, we leverage two physics-based principles to effectively model thermal-domain phenomena. First, we integrate Fourier’s law of heat conduction prior to alpha blending to model intensity interpolation caused by thermal conduction between neighboring Gaussians. Second, we apply the Stefan-Boltzmann law and the inverse-square law to formulate a depth-aware thermal radiation map that imposes additional geometric constraints on thermal rendering. Experimental results demonstrate that the proposed MrGS achieves high-fidelity RGB-T scene reconstruction while reducing the number of Gaussians.

Method: JarvisEvo uses: (1) interleaved multimodal chain-of-thought (iMCoT) reasoning to enhance instruction following, (2) synergistic editor-evaluator policy optimization (SEPO) framework for self-improvement without external rewards, and (3) seamless Adobe Lightroom integration for both global and local fine-grained editing.

Result: On ArtEdit-Bench, JarvisEvo outperforms Nano-Banana by 18.95% on preservative editing metrics, with a 44.96% improvement in pixel-level content fidelity.

Conclusion: JarvisEvo successfully addresses key challenges in agent-based editing through multimodal reasoning and self-improvement optimization, demonstrating superior performance in preserving content fidelity while enabling flexible creative editing.

Abstract: Agent-based editing models have substantially advanced interactive experiences, processing quality, and creative flexibility. However, two critical challenges persist: (1) instruction hallucination, text-only chain-of-thought (CoT) reasoning cannot fully prevent factual errors due to inherent information bottlenecks; (2) reward hacking, dynamic policy optimization against static reward models allows agents to exploit flaws in reward functions. To address these issues, we propose JarvisEvo, a unified image editing agent that emulates an expert human designer by iteratively editing, selecting appropriate tools, evaluating results, and reflecting on its own decisions to refine outcomes. JarvisEvo offers three key advantages: (1) an interleaved multimodal chain-of-thought (iMCoT) reasoning mechanism that enhances instruction following and editing quality; (2) a synergistic editor-evaluator policy optimization (SEPO) framework that enables self-improvement without external rewards, effectively mitigating reward hacking; and (3) support for both global and local fine-grained editing through seamless integration of Adobe Lightroom. On ArtEdit-Bench, JarvisEvo outperforms Nano-Banana by an average of 18.95% on preservative editing metrics, including a substantial 44.96% improvement in pixel-level content fidelity.

Method: Introduces GC-4DGS framework with: 1) dynamic consistency checking to reduce MVS estimation uncertainties across spacetime, and 2) global-local depth regularization to distill spatiotemporal-consistent geometric information from monocular depths for coherent 4D volume learning.

Result: Outperforms RF-DeRF (latest dynamic radiance field for sparse inputs) by 2.62dB and original 4DGS by 1.58dB in PSNR on N3DV and Technicolor datasets, while maintaining real-time performance and deployability on resource-constrained IoT edge devices.

Conclusion: GC-4DGS successfully addresses sparse-input limitations of dynamic Gaussian Splatting by integrating geometric consistency, enabling high-quality real-time dynamic scene rendering for practical AIoT applications with limited views.

Abstract: Gaussian Splatting has been considered as a novel way for view synthesis of dynamic scenes, which shows great potential in AIoT applications such as digital twins. However, recent dynamic Gaussian Splatting methods significantly degrade when only sparse input views are available, limiting their applicability in practice. The issue arises from the incoherent learning of 4D geometry as input views decrease. This paper presents GC-4DGS, a novel framework that infuses geometric consistency into 4D Gaussian Splatting (4DGS), offering real-time and high-quality dynamic scene rendering from sparse input views. While learning-based Multi-View Stereo (MVS) and monocular depth estimators (MDEs) provide geometry priors, directly integrating these with 4DGS yields suboptimal results due to the ill-posed nature of sparse-input 4D geometric optimization. To address these problems, we introduce a dynamic consistency checking strategy to reduce estimation uncertainties of MVS across spacetime. Furthermore, we propose a global-local depth regularization approach to distill spatiotemporal-consistent geometric information from monocular depths, thereby enhancing the coherent geometry and appearance learning within the 4D volume. Extensive experiments on the popular N3DV and Technicolor datasets validate the effectiveness of GC-4DGS in rendering quality without sacrificing efficiency. Notably, our method outperforms RF-DeRF, the latest dynamic radiance field tailored for sparse-input dynamic view synthesis, and the original 4DGS by 2.62dB and 1.58dB in PSNR, respectively, with seamless deployability on resource-constrained IoT edge devices.

Method: Uses hit levels from multi-view ray casting to partition mesh faces into ordered visibility layers. Applies two-stage visibility control: progressive interior region revelation with structural coherence, followed by texturing each layer with a pretrained diffusion model. Uses soft UV-space blending to seamlessly merge textures across layers based on view-dependent visibility confidence.

Result: GOATex consistently outperforms existing methods, producing seamless, high-fidelity textures across both visible and occluded surfaces without requiring fine-tuning of the diffusion model. Allows separate prompting for exterior and interior regions.

Conclusion: GOATex provides a novel occlusion-aware framework for complete 3D mesh texturing that handles both exterior and interior surfaces effectively, offering fine-grained control over layered appearances without costly model fine-tuning.

Abstract: We present GOATex, a diffusion-based method for 3D mesh texturing that generates high-quality textures for both exterior and interior surfaces. While existing methods perform well on visible regions, they inherently lack mechanisms to handle occluded interiors, resulting in incomplete textures and visible seams. To address this, we introduce an occlusion-aware texturing framework based on the concept of hit levels, which quantify the relative depth of mesh faces via multi-view ray casting. This allows us to partition mesh faces into ordered visibility layers, from outermost to innermost. We then apply a two-stage visibility control strategy that progressively reveals interior regions with structural coherence, followed by texturing each layer using a pretrained diffusion model. To seamlessly merge textures obtained across layers, we propose a soft UV-space blending technique that weighs each texture’s contribution based on view-dependent visibility confidence. Empirical results demonstrate that GOATex consistently outperforms existing methods, producing seamless, high-fidelity textures across both visible and occluded surfaces. Unlike prior works, GOATex operates entirely without costly fine-tuning of a pretrained diffusion model and allows separate prompting for exterior and interior mesh regions, enabling fine-grained control over layered appearances. For more qualitative results, please visit our project page: https://goatex3d.github.io/.

Method: Proposes a method to quantify each image’s contribution to NeRF-based reconstructions using reconstruction quality metrics (PSNR and MSE). Validates by removing low-contributing images during training and measuring impact on reconstruction fidelity.

Result: The approach successfully identifies images with varying utility, allowing removal of low-contributing images without significantly harming reconstruction quality.

Conclusion: The method provides a practical way to value individual images in NeRF-based 3D reconstruction pipelines, enabling more efficient training and data monetization strategies.

Abstract: Data valuation and monetization are becoming increasingly important across domains such as eXtended Reality (XR) and digital media. In the context of 3D scene reconstruction from a set of images – whether casually or professionally captured – not all inputs contribute equally to the final output. Neural Radiance Fields (NeRFs) enable photorealistic 3D reconstruction of scenes by optimizing a volumetric radiance field given a set of images. However, in-the-wild scenes often include image captures of varying quality, occlusions, and transient objects, resulting in uneven utility across inputs. In this paper we propose a method to quantify the individual contribution of each image to NeRF-based reconstructions of in-the-wild image sets. Contribution is assessed through reconstruction quality metrics based on PSNR and MSE. We validate our approach by removing low-contributing images during training and measuring the resulting impact on reconstruction fidelity.

Method: REP has three key components: (1) modality-wise feature buffers via residual bypass to cache early-layer representations and replay them in deeper layers, preventing information loss; (2) private-shared feature decoupling where private buffers keep modality-specific signals and shared buffers encode cross-modal semantics; (3) task-aware dynamic initialization to configure buffers differently for better stability and generalization under various missing-modality conditions.

Result: Experiments on vision-language, vision-language-audio, and temporal multimodal benchmarks show REP consistently outperforms prior methods in both single- and multi-modality missing scenarios, with only negligible parameter overhead.

Conclusion: REP establishes a lightweight and effective paradigm for robust multimodal learning in challenging missing-modality environments, offering better performance than existing approaches while maintaining computational efficiency.

Abstract: Missing modalities consistently lead to significant performance degradation in multimodal models. Existing approaches either synthesize missing modalities at high computational cost or apply prompt-based fine-tuning that relies only on adjacent-layer features and overlooks long-distance contextual information, which may offer additional tolerance to errors when one or more modalities are missing. To address this, we introduce REplay Prompting (REP): (1) construct modality-wise feature buffers via a residual bypass to cache early-layer representations and replay them in deeper layers, mitigating information loss as network depth increases; (2) employ a private-shared feature decoupling strategy, where private buffers preserve modality-specific signals and shared buffers encode cross-modal semantics; and (3) design a task-aware dynamic initialization mechanism to configure these buffers differently, improving stability and generalization under diverse missing-modality conditions. Experiments on vision-language, vision-language-audio, and temporal multimodal benchmarks demonstrate that REP consistently outperforms prior methods under both single- and multi-modality missing scenarios, while introducing only negligible parameter overhead. These results establish REP as a lightweight and effective paradigm for robust multimodal learning in challenging missing-modality environments.

Method: Proposed ENSEL (ensemble learning-based skin lesion detection system) that integrates various deep learning models through an ensemble approach. The system was tested using actual user-taken skin lesion images from clinical settings, with performance measured on randomly sampled skin disease images.

Result: ENSEL achieved high recall in most images and demonstrated processing speeds of less than 1 second. The system showed improved diagnostic accuracy compared to existing methods while maintaining fast response times suitable for clinical use.

Conclusion: ENSEL contributes to objective diagnosis of skin lesions like atopic dermatitis and advances digital healthcare by providing an accurate, fast system that works with real-world clinical images rather than requiring high-quality dermoscopic images.

Abstract: The amendments made to the Data 3 Act and impact of COVID-19 have fostered the growth of digital healthcare market and promoted the use of medical data in artificial intelligence in South Korea. Atopic dermatitis, a chronic inflammatory skin disease, is diagnosed via subjective evaluations without using objective diagnostic methods, thereby increasing the risk of misdiagnosis. It is also similar to psoriasis in appearance, further complicating its accurate diagnosis. Existing studies on skin diseases have used high-quality dermoscopic image datasets, but such high-quality images cannot be obtained in actual clinical settings. Moreover, existing systems must ensure accuracy and fast response times. To this end, an ensemble learning-based skin lesion detection system (ENSEL) was proposed herein. ENSEL enhanced diagnostic accuracy by integrating various deep learning models via an ensemble approach. Its performance was verified by conducting skin lesion detection experiments using images of skin lesions taken by actual users. Its accuracy and response time were measured using randomly sampled skin disease images. Results revealed that ENSEL achieved high recall in most images and less than 1s s processing speed. This study contributes to the objective diagnosis of skin lesions and promotes the advancement of digital healthcare.

Method: Introduces NumeriKontrol framework with Numeric Adapter to encode numeric editing scales, injects them into diffusion models in plug-and-play manner. Uses task-separated design for zero-shot multi-condition editing. Trains on CAT dataset synthesized from reliable sources (rendering engines, DSLR cameras) with accurate ground-truth scales.

Result: NumeriKontrol delivers accurate, continuous, and stable scale control across wide range of attribute editing scenarios. Functions as powerful interactive editing studio with precise, scalable manipulation.

Conclusion: Advances instruction-based image editing by enabling precise, scalable, and user-controllable image manipulation through numeric scale control with common units.

Abstract: Instruction-based image editing enables intuitive manipulation through natural language commands. However, text instructions alone often lack the precision required for fine-grained control over edit intensity. We introduce NumeriKontrol, a framework that allows users to precisely adjust image attributes using continuous scalar values with common units. NumeriKontrol encodes numeric editing scales via an effective Numeric Adapter and injects them into diffusion models in a plug-and-play manner. Thanks to a task-separated design, our approach supports zero-shot multi-condition editing, allowing users to specify multiple instructions in any order. To provide high-quality supervision, we synthesize precise training data from reliable sources, including high-fidelity rendering engines and DSLR cameras. Our Common Attribute Transform (CAT) dataset covers diverse attribute manipulations with accurate ground-truth scales, enabling NumeriKontrol to function as a simple yet powerful interactive editing studio. Extensive experiments show that NumeriKontrol delivers accurate, continuous, and stable scale control across a wide range of attribute editing scenarios. These contributions advance instruction-based image editing by enabling precise, scalable, and user-controllable image manipulation.

Method: Created MathSight benchmark with university-level math problems featuring multiple visual variants (original, hand-drawn, photo-captured) and text-only conditions for controlled comparison. Tested state-of-the-art VLMs to disentangle and quantify visual input effects.

Result: Visual information contribution consistently diminishes with increasing problem difficulty. Surprisingly, Qwen3-VL without any image input outperformed both its multimodal variants and GPT-5, highlighting that current VLMs may not effectively leverage visual information for complex reasoning.

Conclusion: Current VLMs don’t genuinely leverage visual understanding for complex math reasoning, often relying on linguistic priors instead. MathSight benchmark is needed to advance true vision-grounded reasoning in future models.

Abstract: Recent advances in Vision-Language Models (VLMs) have achieved impressive progress in multimodal mathematical reasoning. Yet, how much visual information truly contributes to reasoning remains unclear. Existing benchmarks report strong overall performance but seldom isolate the role of the image modality, leaving open whether VLMs genuinely leverage visual understanding or merely depend on linguistic priors. To address this, we present MathSight, a university-level multimodal mathematical reasoning benchmark designed to disentangle and quantify the effect of visual input. Each problem includes multiple visual variants – original, hand-drawn, photo-captured – and a text-only condition for controlled comparison. Experiments on state-of-the-art VLMs reveal a consistent trend: the contribution of visual information diminishes with increasing problem difficulty. Remarkably, Qwen3-VL without any image input surpasses both its multimodal variants and GPT-5, underscoring the need for benchmarks like MathSight to advance genuine vision-grounded reasoning in future models.

Method: Proposes db-SP with: 1) formalized sparse imbalance ratio to quantify imbalance, 2) dual-level partitioning approach for near-perfect workload balance at head and block levels, and 3) dynamic determination of parallel degrees for head and block dimensions at runtime to handle evolving sparsity patterns.

Result: db-SP achieves 1.25x end-to-end speedup and 1.40x attention-specific speedup over state-of-the-art sequence parallel methods on average.

Conclusion: db-SP effectively addresses workload imbalance in sparse attention models, significantly improving inference performance for Diffusion Transformers through sparsity-aware sequence parallelism.

Abstract: Scaling Diffusion Transformer (DiT) inference via sequence parallelism is critical for reducing latency in visual generation, but is severely hampered by workload imbalance when applied to models employing block-wise sparse attention. The imbalance stems from the inherent variation in sparsity across attention heads and the irregular distribution of dense blocks within the sparse mask, when sequence parallelism is applied along the head dimension (as in Ulysses) or the block dimension (as in Ring Attention). In this paper, we formalize a sparse imbalance ratio to quantify the imbalance, and propose db-SP, a sparsity-aware sequence parallelism technique that tackles the challenge. db-SP contains a dual-level partitioning approach that achieves near-perfect workload balance at both the head and block levels with negligible overhead. Furthermore, to handle the evolving sparsity patterns across denoising steps and layers, db-SP dynamically determines the parallel degrees for the head and block dimensions at runtime. Experimental results demonstrate that db-SP delivers an end-to-end speedup of 1.25x and an attention-specific speedup of 1.40x over state-of-the-art sequence parallel methods on average. Code is available at https://github.com/thu-nics/db-SP.

Method: 1) Hierarchical multi-level contrastive loss for detecting emotional concepts from images. 2) Emotional attribute chain-of-thought reasoning to generate affective sentences. 3) Uses pre-trained language model to synthesize emotional concepts and affective sentences for IEC. 4) Contrastive loss with semantic similarity sampling to handle large intra-class/small inter-class differences. 5) Considers images with embedded text (ignored by previous studies).

Result: Extensive experiments show the method effectively bridges the affective gap and achieves superior results on multiple benchmarks.

Conclusion: ACIEC successfully addresses the affective gap in image emotion classification by leveraging language’s expressive power through affective captioning, outperforming previous approaches on standard benchmarks.

Abstract: Image emotion classification (IEC) is a longstanding research field that has received increasing attention with the rapid progress of deep learning. Although recent advances have leveraged the knowledge encoded in pre-trained visual models, their effectiveness is constrained by the “affective gap” , limits the applicability of pre-training knowledge for IEC tasks. It has been demonstrated in psychology that language exhibits high variability, encompasses diverse and abundant information, and can effectively eliminate the “affective gap”. Inspired by this, we propose a novel Affective Captioning for Image Emotion Classification (ACIEC) to classify image emotion based on pure texts, which effectively capture the affective information in the image. In our method, a hierarchical multi-level contrastive loss is designed for detecting emotional concepts from images, while an emotional attribute chain-of-thought reasoning is proposed to generate affective sentences. Then, a pre-trained language model is leveraged to synthesize emotional concepts and affective sentences to conduct IEC. Additionally, a contrastive loss based on semantic similarity sampling is designed to solve the problem of large intra-class differences and small inter-class differences in affective datasets. Moreover, we also take the images with embedded texts into consideration, which were ignored by previous studies. Extensive experiments illustrate that our method can effectively bridge the affective gap and achieve superior results on multiple benchmarks.

Method: DNA-Prior integrates: (1) implicit architectural prior enforced through deep network parameterization, and (2) explicit spectral-spatial prior composed of frequency-domain fidelity term and spatial regularization functional, forming a dual-domain optimization problem.

Result: Experiments across multiple modalities show DNA achieves consistent noise suppression and structural preservation under diverse noise conditions.

Conclusion: DNA-Prior provides a universal unsupervised denoising framework that works across modalities without external training data, addressing limitations of supervised approaches in clinical settings.

Abstract: Medical imaging pipelines critically rely on robust denoising to stabilise downstream tasks such as segmentation and reconstruction. However, many existing denoisers depend on large annotated datasets or supervised learning, which restricts their usability in clinical environments with heterogeneous modalities and limited ground-truth data. To address this limitation, we introduce DNA-Prior, a universal unsupervised denoising framework that reconstructs clean images directly from corrupted observations through a mathematically principled hybrid prior. DNA-Prior integrates (i) an implicit architectural prior, enforced through a deep network parameterisation, with (ii) an explicit spectral-spatial prior composed of a frequency-domain fidelity term and a spatial regularisation functional. This dual-domain formulation yields a well-structured optimisation problem that jointly preserves global frequency characteristics and local anatomical structure, without requiring any external training data or modality-specific tuning. Experiments across multiple modalities show that DNA achieves consistent noise suppression and structural preservation under diverse noise conditions.

Method: Introduces a dual-branch framework that mutually generates camera-consistent RGB and depth sequences, along with Semantic Guided Mutual Alignment (SIGMA) mechanism for semantics-guided RGB-depth fusion. Also analyzes the distinct roles of depth and camera poses across different denoising stages.

Result: Achieves more consistent camera-controlled video generation with over 40% reduction in camera motion errors compared to prior methods, demonstrating improved adherence to specified camera trajectories.

Conclusion: DualCamCtrl successfully addresses limitations in scene understanding and geometric awareness through its dual-branch framework and SIGMA mechanism, enabling better disentanglement of appearance and geometry modeling for more faithful camera-controlled video generation.

Abstract: This paper presents DualCamCtrl, a novel end-to-end diffusion model for camera-controlled video generation. Recent works have advanced this field by representing camera poses as ray-based conditions, yet they often lack sufficient scene understanding and geometric awareness. DualCamCtrl specifically targets this limitation by introducing a dual-branch framework that mutually generates camera-consistent RGB and depth sequences. To harmonize these two modalities, we further propose the Semantic Guided Mutual Alignment (SIGMA) mechanism, which performs RGB-depth fusion in a semantics-guided and mutually reinforced manner. These designs collectively enable DualCamCtrl to better disentangle appearance and geometry modeling, generating videos that more faithfully adhere to the specified camera trajectories. Additionally, we analyze and reveal the distinct influence of depth and camera poses across denoising stages and further demonstrate that early and late stages play complementary roles in forming global structure and refining local details. Extensive experiments demonstrate that DualCamCtrl achieves more consistent camera-controlled video generation, with over 40% reduction in camera motion errors compared with prior methods. Our project page: https://soyouthinkyoucantell.github.io/dualcamctrl-page/

Method: Proposes InstanceV with: 1) Instance-aware Masked Cross-Attention for spatial instance control, 2) Shared Timestep-Adaptive Prompt Enhancement for global consistency, 3) Spatially-Aware Unconditional Guidance to prevent small instance disappearance, and 4) InstanceBench benchmark for evaluation.

Result: InstanceV achieves remarkable instance-level controllability and outperforms state-of-the-art models in both general quality and instance-aware metrics across qualitative and quantitative evaluations.

Conclusion: InstanceV successfully addresses the fine-grained controllability challenge in text-to-video generation, providing instance-level control while maintaining global semantic consistency, with comprehensive evaluation through the new InstanceBench benchmark.

Abstract: Recent advances in text-to-video diffusion models have enabled the generation of high-quality videos conditioned on textual descriptions. However, most existing text-to-video models rely solely on textual conditions, lacking general fine-grained controllability over video generation. To address this challenge, we propose InstanceV, a video generation framework that enables i) instance-level control and ii) global semantic consistency. Specifically, with the aid of proposed Instance-aware Masked Cross-Attention mechanism, InstanceV maximizes the utilization of additional instance-level grounding information to generate correctly attributed instances at designated spatial locations. To improve overall consistency, We introduce the Shared Timestep-Adaptive Prompt Enhancement module, which connects local instances with global semantics in a parameter-efficient manner. Furthermore, we incorporate Spatially-Aware Unconditional Guidance during both training and inference to alleviate the disappearance of small instances. Finally, we propose a new benchmark, named InstanceBench, which combines general video quality metrics with instance-aware metrics for more comprehensive evaluation on instance-level video generation. Extensive experiments demonstrate that InstanceV not only achieves remarkable instance-level controllability in video generation, but also outperforms existing state-of-the-art models in both general quality and instance-aware metrics across qualitative and quantitative evaluations.

Method: A three-stage progressive framework: 1) Global affine transformation for perspective correction, 2) Geometric deformation rectification for paper curling/folding, 3) Content-aware iterative process for fine-grained content distortions. Also introduces new evaluation metrics: layout-aligned OCR metrics (AED/ACER) and masked AD/AAD (AD-M/AAD-M).

Result: Achieves state-of-the-art performance on multiple benchmarks with 14.1%-34.7% reduction in AAD metric. Demonstrates superior efficacy in real-world applications compared to existing methods.

Conclusion: The progressive multi-stage approach effectively handles complex document distortions by decomposing transformations. The proposed evaluation metrics provide more accurate assessment of geometric rectification quality, addressing limitations in existing protocols.

Abstract: Document rectification in real-world scenarios poses significant challenges due to extreme variations in camera perspectives and physical distortions. Driven by the insight that complex transformations can be decomposed and resolved progressively, we introduce a novel multi-stage framework that progressively reverses distinct distortion types in a coarse-to-fine manner. Specifically, our framework first performs a global affine transformation to correct perspective distortions arising from the camera’s viewpoint, then rectifies geometric deformations resulting from physical paper curling and folding, and finally employs a content-aware iterative process to eliminate fine-grained content distortions. To address limitations in existing evaluation protocols, we also propose two enhanced metrics: layout-aligned OCR metrics (AED/ACER) for a stable assessment that decouples geometric rectification quality from the layout analysis errors of OCR engines, and masked AD/AAD (AD-M/AAD-M) tailored for accurately evaluating geometric distortions in documents with incomplete boundaries. Extensive experiments show that our method establishes new state-of-the-art performance on multiple challenging benchmarks, yielding a substantial reduction of 14.1%–34.7% in the AAD metric and demonstrating superior efficacy in real-world applications. The code will be publicly available at https://github.com/chaoyunwang/ArbDR.

Method: Refusal-Aware Reinforcement Fine-Tuning (RA-RFT) based on Group Relative Policy Optimization (GRPO) framework with four reward objectives: format, refuse-IoU, explain, and query correction. Also constructed Hard-Irrelevant VTG (HI-VTG) dataset.

Result: Demonstrated effectiveness across various relevance-aware VTG scenarios including hard-irrelevant VTG, simply-shuffled RA-VTG, and human-annotated RA-VTG settings. Method is scalable to various LVLM-based VTG models.

Conclusion: RA-RFT effectively addresses the limitation of existing VTG models by enabling them to refuse hard-irrelevant queries through reinforcement fine-tuning with comprehensive reward objectives and specialized training data.

Abstract: Video Temporal Grounding (VTG) aims to localize a temporal segment in a video corresponding to a natural language query. However, existing VTG models assume that a relevant segment always exists, causing them to always predict a target segment even when the query is irrelevant to the video. While recent approaches attempt to handle irrelevant queries, they can only reject those that are entirely unrelated to the video and still fail to handle hard-irrelevant queries that are semantically similar but not actually relevant. To address this, we propose Refusal-Aware Reinforcement Fine-Tuning (RA-RFT) to effectively refuse hard-irrelevant queries in VTG. Our method is based on the Group Relative Policy Optimization (GRPO) framework and integrates four reward objectives-format, refuse-IoU, explain, and query correction-to improve both relevance discrimination and fine-grained semantic reasoning. In addition, to effectively support RA-RFT, we construct a Hard-Irrelevant VTG (HI-VTG) dataset, which includes hard-irrelevant queries and their refusal answers. We demonstrate the effectiveness of our method across various relevance-aware VTG scenarios, including hard-irrelevant VTG, simply-shuffled RA-VTG, and human-annotated RA-VTG settings. We also show that the proposed method is scalable by applying it to various LVLM-based VTG models. Our code is available at https://github.com/JINSUBY/RA-RFT.

Method: Proposes PowerCLIP with powerset alignment that minimizes loss between powersets of image regions and textual parse trees. Introduces efficient non-linear aggregators (NLAs) to reduce computational complexity from O(2^M) to O(M) while maintaining accuracy.

Result: PowerCLIP outperforms state-of-the-art methods in zero-shot classification and retrieval tasks, demonstrating improved compositionality and robustness compared to existing approaches.

Conclusion: The powerset alignment framework with efficient NLAs enables effective capture of compositional semantics spanning multiple image regions, advancing vision-language pre-training capabilities while maintaining computational feasibility.

Abstract: Contrastive vision-language pre-training frameworks such as CLIP have demonstrated impressive zero-shot performance across a range of vision-language tasks. Recent studies have shown that aligning individual text tokens with specific image patches or regions enhances fine-grained compositional understanding. However, it remains challenging to capture compositional semantics that span multiple image regions. To address this limitation, we propose PowerCLIP, a novel contrastive pre-training framework enhanced by powerset alignment, which exhaustively optimizes region-to-phrase alignments by minimizing the loss defined between powersets of image regions and textual parse trees. Since the naive powerset construction incurs exponential computational cost due to the combinatorial explosion in the number of region subsets, we introduce efficient non-linear aggregators (NLAs) that reduce complexity from O(2^M) to O(M) with respect to the number of regions M, while approximating the exact loss value with arbitrary precision. Our extensive experiments demonstrate that PowerCLIP outperforms state-of-the-art methods in zero-shot classification and retrieval tasks, underscoring the compositionality and robustness of our approach. Our code will be made publicly available.

Method: Uses pre-trained video generation models conditioned on a single edited view to generate consistent multi-view edits. Introduces motion-preserved noise blending for pose-specific view generation and geometry-aware denoising for enhanced consistency.

Result: Achieves high-quality 3D editing in a single forward pass, significantly outperforming existing methods in both editing quality and speed.

Conclusion: Video generation models provide effective temporal consistency priors for efficient multi-view consistent 3D editing, overcoming limitations of iterative 2D-based approaches.

Abstract: Text-driven 3D editing enables user-friendly 3D object or scene editing with text instructions. Due to the lack of multi-view consistency priors, existing methods typically resort to employing 2D generation or editing models to process each view individually, followed by iterative 2D-3D-2D updating. However, these methods are not only time-consuming but also prone to over-smoothed results because the different editing signals gathered from different views are averaged during the iterative process. In this paper, we propose generative Video Prior based 3D Editing (ViP3DE) to employ the temporal consistency priors from pre-trained video generation models for multi-view consistent 3D editing in a single forward pass. Our key insight is to condition the video generation model on a single edited view to generate other consistent edited views for 3D updating directly, thereby bypassing the iterative editing paradigm. Since 3D updating requires edited views to be paired with specific camera poses, we propose motion-preserved noise blending for the video model to generate edited views at predefined camera poses. In addition, we introduce geometry-aware denoising to further enhance multi-view consistency by integrating 3D geometric priors into video models. Extensive experiments demonstrate that our proposed ViP3DE can achieve high-quality 3D editing results even within a single forward pass, significantly outperforming existing methods in both editing quality and speed.

Method: 1) Generate consecutive video frames first, 2) Use geometry model to extract full-frame geometry features (richer than single-frame depth maps), 3) Use geometry features as conditions for video generation, 4) Add geometry alignment loss for real-world constraints, 5) Implement geometry adaptation module for effective feature utilization.

Result: Extensive experiments show GeoWorld generates high-fidelity 3D scenes from single images and camera trajectories, outperforming prior methods both qualitatively and quantitatively.

Conclusion: GeoWorld successfully renovates image-to-3D scene generation by leveraging geometry models with video frames, addressing geometric distortions and improving overall scene quality.

Abstract: Previous works leveraging video models for image-to-3D scene generation tend to suffer from geometric distortions and blurry content. In this paper, we renovate the pipeline of image-to-3D scene generation by unlocking the potential of geometry models and present our GeoWorld. Instead of exploiting geometric information obtained from a single-frame input, we propose to first generate consecutive video frames and then take advantage of the geometry model to provide full-frame geometry features, which contain richer information than single-frame depth maps or camera embeddings used in previous methods, and use these geometry features as geometrical conditions to aid the video generation model. To enhance the consistency of geometric structures, we further propose a geometry alignment loss to provide the model with real-world geometric constraints and a geometry adaptation module to ensure the effective utilization of geometry features. Extensive experiments show that our GeoWorld can generate high-fidelity 3D scenes from a single image and a given camera trajectory, outperforming prior methods both qualitatively and quantitatively. Project Page: https://peaes.github.io/GeoWorld/.

Method: Multi-teacher distillation framework inspired by RADIO distillation and Matryoshka Representation Learning, allowing for adaptable embedding dimensions while reducing model size.

Result: Achieves 86-92% model size reduction while maintaining on-par performance with larger teachers, and outperforms comparable single-teacher distillation models by median 7.0 accuracy points across ten pathology benchmarks.

Conclusion: Pathryoshka democratizes access to state-of-the-art pathology foundation models by enabling efficient local deployment without sacrificing accuracy or representational richness.

Abstract: Pathology foundation models (FMs) have driven significant progress in computational pathology. However, these high-performing models can easily exceed a billion parameters and produce high-dimensional embeddings, thus limiting their applicability for research or clinical use when computing resources are tight. Here, we introduce Pathryoshka, a multi-teacher distillation framework inspired by RADIO distillation and Matryoshka Representation Learning to reduce pathology FM sizes while allowing for adaptable embedding dimensions. We evaluate our framework with a distilled model on ten public pathology benchmarks with varying downstream tasks. Compared to its much larger teachers, Pathryoshka reduces the model size by 86-92% at on-par performance. It outperforms state-of-the-art single-teacher distillation models of comparable size by a median margin of 7.0 in accuracy. By enabling efficient local deployment without sacrificing accuracy or representational richness, Pathryoshka democratizes access to state-of-the-art pathology FMs for the broader research and clinical community.

Method: 1) Creates Event-CSL dataset with 14,827 event-based videos; 2) Proposes EvSLT framework with video segmentation, Mamba-based memory aggregation for spatial features, temporal convolution for temporal features, and graph-guided spatiotemporal fusion.

Result: Extensive experiments on Event-CSL and other public datasets demonstrate superior performance of the proposed EvSLT method compared to existing SLT approaches.

Conclusion: Event cameras effectively address limitations of traditional SLT methods, and the proposed EvSLT framework with Event-CSL dataset advances event-based sign language translation research.

Abstract: Sign Language Translation (SLT) is a core task in the field of AI-assisted disability. Traditional SLT methods are typically based on visible light videos, which are easily affected by factors such as lighting variations, rapid hand movements, and privacy concerns. This paper proposes the use of bio-inspired event cameras to alleviate the aforementioned issues. Specifically, we introduce a new high-definition event-based sign language dataset, termed Event-CSL, which effectively addresses the data scarcity in this research area. The dataset comprises 14,827 videos, 14,821 glosses, and 2,544 Chinese words in the text vocabulary. These samples are collected across diverse indoor and outdoor scenes, covering multiple viewpoints, lighting conditions, and camera motions. We have also benchmarked existing mainstream SLT methods on this dataset to facilitate fair comparisons in future research.Furthermore, we propose a novel event-based sign language translation framework, termed EvSLT. The framework first segments continuous video features into clips and employs a Mamba-based memory aggregation module to compress and aggregate spatial detail features at the clip level. Subsequently, these spatial features, along with temporal representations obtained from temporal convolution, are then fused by a graph-guided spatiotemporal fusion module. Extensive experiments on Event-CSL, as well as other publicly available datasets, demonstrate the superior performance of our method. The dataset and source code will be released on https://github.com/Event-AHU/OpenESL

Method: Pose-agnostic, zero-shot framework comparing real scenes against real-time Digital Twins in RGB-D space. Uses object detection and pose estimation of CAD models for efficient DT rendering, with hierarchical annotation strategy for multi-criteria defect detection.

Result: Achieved detection performance with IoU scores up to 63.3% compared to ground-truth masks using simple distance measurements in semi-controlled industrial conditions.

Conclusion: The framework demonstrates effectiveness for quality inspection and lays groundwork for future research on generalizable, low-data defect detection methods in dynamic manufacturing settings.

Abstract: Early-stage visual quality inspection is vital for achieving Zero-Defect Manufacturing and minimizing production waste in modern industrial environments. However, the complexity of robust visual inspection systems and their extensive data requirements hinder widespread adoption in semi-controlled industrial settings. In this context, we propose a pose-agnostic, zero-shot quality inspection framework that compares real scenes against real-time Digital Twins (DT) in the RGB-D space. Our approach enables efficient real-time DT rendering by semantically describing industrial scenes through object detection and pose estimation of known Computer-Aided Design models. We benchmark tools for real-time, multimodal RGB-D DT creation while tracking consumption of computational resources. Additionally, we provide an extensible and hierarchical annotation strategy for multi-criteria defect detection, unifying pose labelling with logical and structural defect annotations. Based on an automotive use case featuring the quality inspection of an axial flux motor, we demonstrate the effectiveness of our framework. Our results demonstrate detection performace, achieving intersection-over-union (IoU) scores of up to 63.3% compared to ground-truth masks, even if using simple distance measurements under semi-controlled industrial conditions. Our findings lay the groundwork for future research on generalizable, low-data defect detection methods in dynamic manufacturing settings.

Method: 1. Created a high-quality instruction dataset specifically for tabular data to enable efficient LLM comprehension. 2. Instruction-tuned an open-source LLM (Llama3.1-8B-Instruct) on the training set of this dataset using only 7K instructions on an A100 GPU for less than 6 hours.

Result: The instruction-tuned model achieved tabular data generation performance on par with the most capable commercial LLM, GPT-4o, despite using significantly fewer resources.

Conclusion: Instruction tuning can effectively improve LLMs’ tabular data generation capabilities even with limited data and computational resources, making this approach accessible and practical for broader adoption.

Abstract: Tabular instruction tuning has emerged as a promising research direction for improving LLMs understanding of tabular data. However, the majority of existing works only consider question-answering and reasoning tasks over tabular data, leaving tabular data generation largely unnoticed. In this work, for the first time, we explore the efficacy of instruction tuning in improving LLMs tabular data generation capabilities. More specifically, given the high data and computation requirements of tabular instruction tuning, we aim to address the possibility of instruction tuning for tabular data generation with limited data and computational resources. To achieve this, we first create a high-quality instruction dataset for tabular data, enabling efficient LLM comprehension. We then instruction-tune an open-source LLM (Llama3.1-8B-Instruct) on the training set of this dataset to improve its tabular data generation performance. Our experimental results show that by using our high-quality dataset and instruction-tuning on only 7K instructions with an A100 GPU, for less than 6 hours, we achieve tabular data generation performance on par with the most capable commercial LLM, GPT-4o.

Method: The proposed method, Corrective Blurry KNN (CB-KNN), uses a smooth kernel strategy to enhance robustness. Instead of minimizing rendering error alone, it makes rasterization more resilient to imperfect 3DGS parameters. The approach adaptively modifies RGB values and locations of K-nearest neighboring Gaussians within local regions, creating smoother local renderings that reduce the impact of erroneous Gaussian parameters on the overall image.

Result: Experimental results show that the approach significantly improves the robustness and accuracy of camera pose tracking while maintaining the overall quality of scene reconstruction (mapping). The method effectively stabilizes pose optimization through controlled blurring that acts as regularization.

Conclusion: Robustness against parameter errors is more critical than perfect rendering quality for stable camera tracking in 3DGS-SLAM. The CB-KNN method provides a practical solution that enhances rasterization robustness through smooth kernel adjustments, improving tracking performance without compromising reconstruction quality.

Abstract: In this paper, we challenge the conventional notion in 3DGS-SLAM that rendering quality is the primary determinant of tracking accuracy. We argue that, compared to solely pursuing a perfect scene representation, it is more critical to enhance the robustness of the rasterization process against parameter errors to ensure stable camera pose tracking. To address this challenge, we propose a novel approach that leverages a smooth kernel strategy to enhance the robustness of 3DGS-based SLAM. Unlike conventional methods that focus solely on minimizing rendering error, our core insight is to make the rasterization process more resilient to imperfections in the 3DGS parameters. We hypothesize that by allowing each Gaussian to influence a smoother, wider distribution of pixels during rendering, we can mitigate the detrimental effects of parameter noise from outlier Gaussians. This approach intentionally introduces a controlled blur to the rendered image, which acts as a regularization term, stabilizing the subsequent pose optimization. While a complete redesign of the rasterization pipeline is an ideal solution, we propose a practical and effective alternative that is readily integrated into existing 3DGS frameworks. Our method, termed Corrective Blurry KNN (CB-KNN), adaptively modifies the RGB values and locations of the K-nearest neighboring Gaussians within a local region. This dynamic adjustment generates a smoother local rendering, reducing the impact of erroneous GS parameters on the overall image. Experimental results demonstrate that our approach, while maintaining the overall quality of the scene reconstruction (mapping), significantly improves the robustness and accuracy of camera pose tracking.

Method: Proposes DAONet-YOLOv8 with three key improvements: (1) Dual-Attention Fusion Module combining convolutional local features with self-attention global context; (2) occlusion-aware detection head learning relationships between visible and occluded parts; (3) C2f-DSConv module using dynamic synthesis convolutions with multiple kernel shapes.

Result: Achieves 92.97% precision, 92.80% recall, 97.10% mAP@50 and 76.90% mAP@50:95 on real-world tea plantation dataset, outperforming YOLOv8n baseline by 2.34-4.68 percentage points while reducing parameters by 16.7%.

Conclusion: DAONet-YOLOv8 effectively addresses challenges in tea leaf pest/disease detection through attention mechanisms, occlusion handling, and adaptive feature extraction, achieving state-of-the-art performance with reduced parameters.

Abstract: Accurate detection of tea leaf pests and diseases in real plantations remains challenging due to complex backgrounds, variable illumination, and frequent occlusions among dense branches and leaves. Existing detectors often suffer from missed detections and false positives in such scenarios. To address these issues, we propose DAONet-YOLOv8, an enhanced YOLOv8 variant with three key improvements: (1) a Dual-Attention Fusion Module (DAFM) that combines convolutional local feature extraction with self-attention based global context modeling to focus on subtle lesion regions while suppressing background noise; (2) an occlusion-aware detection head (Detect-OAHead) that learns the relationship between visible and occluded parts to compensate for missing lesion features; and (3) a C2f-DSConv module employing dynamic synthesis convolutions with multiple kernel shapes to better capture irregular lesion boundaries. Experiments on our real-world tea plantation dataset containing six pest and disease categories demonstrate that DAONet-YOLOv8 achieves 92.97% precision, 92.80% recall, 97.10% mAP@50 and 76.90% mAP@50:95, outperforming the YOLOv8n baseline by 2.34, 4.68, 1.40 and 1.80 percentage points respectively, while reducing parameters by 16.7%. Comparative experiments further confirm that DAONet-YOLOv8 achieves superior performance over mainstream detection models.

Method: Mavors introduces multi-granularity video representation with two core components: 1) Intra-chunk Vision Encoder (IVE) using 3D convolutions and Vision Transformers to preserve high-resolution spatial features, and 2) Inter-chunk Feature Aggregator (IFA) using transformer-based dependency modeling with chunk-level rotary position encodings to establish temporal coherence across chunks. It also unifies image and video understanding by treating images as single-frame videos via sub-image decomposition.

Result: Experiments across diverse benchmarks demonstrate Mavors’ superiority in maintaining both spatial fidelity and temporal continuity, significantly outperforming existing methods in tasks requiring fine-grained spatio-temporal reasoning.

Conclusion: Mavors effectively addresses the critical challenge of balancing computational efficiency with retention of fine-grained spatio-temporal patterns in long-context video understanding for MLLMs, offering a comprehensive solution that preserves both spatial details and temporal dynamics.

Abstract: Long-context video understanding in multimodal large language models (MLLMs) faces a critical challenge: balancing computational efficiency with the retention of fine-grained spatio-temporal patterns. Existing approaches (e.g., sparse sampling, dense sampling with low resolution, and token compression) suffer from significant information loss in temporal dynamics, spatial details, or subtle interactions, particularly in videos with complex motion or varying resolutions. To address this, we propose $\mathbf{Mavors}$, a novel framework that introduces $\mathbf{M}$ulti-gr$\mathbf{a}$nularity $\mathbf{v}$ide$\mathbf{o}$ $\mathbf{r}$epre$\mathbf{s}$entation for holistic long-video modeling. Specifically, Mavors directly encodes raw video content into latent representations through two core components: 1) an Intra-chunk Vision Encoder (IVE) that preserves high-resolution spatial features via 3D convolutions and Vision Transformers, and 2) an Inter-chunk Feature Aggregator (IFA) that establishes temporal coherence across chunks using transformer-based dependency modeling with chunk-level rotary position encodings. Moreover, the framework unifies image and video understanding by treating images as single-frame videos via sub-image decomposition. Experiments across diverse benchmarks demonstrate Mavors’ superiority in maintaining both spatial fidelity and temporal continuity, significantly outperforming existing methods in tasks requiring fine-grained spatio-temporal reasoning.

Method: 1) Introduces point-centric convolution with receptive fields centered on original high-precision point coordinates. 2) Formulates convolution on native points as a Matrix-Vector Multiplication and Reduction (MVMR) problem. 3) Develops dedicated, highly-optimized GPU kernel for efficient computation.

Result: PointCNN++ uses an order of magnitude less memory and is several times faster than representative point-based methods. When replacing voxel-based backbones, it significantly improves point cloud registration accuracy while being both more memory-efficient and faster.

Conclusion: PointCNN++ demonstrates that preserving geometric detail and achieving high performance are not mutually exclusive, paving the way for a new class of 3D learning with both high fidelity and efficiency.

Abstract: Existing convolutional learning methods for 3D point cloud data are divided into two paradigms: point-based methods that preserve geometric precision but often face performance challenges, and voxel-based methods that achieve high efficiency through quantization at the cost of geometric fidelity. This loss of precision is a critical bottleneck for tasks such as point cloud registration. We propose PointCNN++, a novel architectural design that fundamentally mitigates this precision-performance trade-off. It \textbf{generalizes sparse convolution from voxels to points}, treating voxel-based convolution as a specialized, degraded case of our more general point-based convolution. First, we introduce a point-centric convolution where the receptive field is centered on the original, high-precision point coordinates. Second, to make this high-fidelity operation performant, we design a computational strategy that operates \textbf{natively} on points. We formulate the convolution on native points as a Matrix-Vector Multiplication and Reduction (MVMR) problem, for which we develop a dedicated, highly-optimized GPU kernel. Experiments demonstrate that PointCNN++ \textbf{uses an order of magnitude less memory and is several times faster} than representative point-based methods. Furthermore, when used as a simple replacement for the voxel-based backbones it generalizes, it \textbf{significantly improves point cloud registration accuracies while proving both more memory-efficient and faster}. PointCNN++ shows that preserving geometric detail and achieving high performance are not mutually exclusive, paving the way for a new class of 3D learning with high fidelity and efficiency. Our code will be open sourced.

Method: ForAug uses pretrained foundation models to separate foreground objects from backgrounds, then recombines them with different backgrounds while controlling object position and size. This imposes invariances into training data that are normally part of network architecture.

Result: Improves ViT accuracy by up to 4.5 percentage points on ImageNet (7.3pp on downstream tasks). Introduces metrics for background robustness, foreground focus, center bias, and size bias - shows ForAug substantially reduces these biases.

Conclusion: ForAug provides valuable tool for analyzing and mitigating biases, enabling more robust and reliable computer vision models. Offers new ways to analyze model behavior and quantify biases beyond just accuracy improvements.

Abstract: Transformers, particularly Vision Transformers (ViTs), have achieved state-of-the-art performance in large-scale image classification. However, they often require large amounts of data and can exhibit biases, such as center or size bias, that limit their robustness and generalizability. This paper introduces ForAug, a novel data augmentation operation that addresses these challenges by explicitly imposing invariances into the training data, which are otherwise part of the neural network architecture. ForAug is constructed by using pretrained foundation models to separate and recombine foreground objects with different backgrounds. This recombination step enables us to take fine-grained control over object position and size, as well as background selection. We demonstrate that using ForAug significantly improves the accuracy of ViTs and other architectures by up to 4.5 percentage points (p.p.) on ImageNet, which translates to 7.3 p.p. on downstream tasks. Importantly, ForAug not only improves accuracy but also opens new ways to analyze model behavior and quantify biases. Namely, we introduce metrics for background robustness, foreground focus, center bias, and size bias and show that using ForAug during training substantially reduces these biases. In summary, ForAug provides a valuable tool for analyzing and mitigating biases, enabling the development of more robust and reliable computer vision models. Our code and dataset are publicly available at https://github.com/tobna/ForAug.

Method: SynthFun3D generates 3D indoor scenes using a furniture asset database with part-level annotations. Given an action description, it reasons about the action to automatically identify and retrieve the 3D mask of the correct functional element, enabling large-scale generation of annotated data.

Result: User studies show improved scene-prompt coherence compared to other approaches. Generated data can replace real data with minor performance loss or supplement real data for improved performance in 3D functionality understanding tasks.

Conclusion: SynthFun3D provides an inexpensive and scalable solution for generating high-quality annotated 3D data, addressing the data scarcity problem in data-hungry 3D applications.

Abstract: Functionality understanding in 3D, which aims to identify the functional element in a 3D scene to complete an action (e.g., the correct handle to “Open the second drawer of the cabinet near the bed”), is hindered by the scarcity of real-world data due to the substantial effort needed for its collection and annotation. To address this, we introduce SynthFun3D, the first method for task-based 3D scene synthesis. Given the action description, SynthFun3D generates a 3D indoor environment using a furniture asset database with part-level annotation, ensuring the action can be accomplished. It reasons about the action to automatically identify and retrieve the 3D mask of the correct functional element, enabling the inexpensive and large-scale generation of high-quality annotated data. We validate SynthFun3D through user studies, which demonstrate improved scene-prompt coherence compared to other approaches. Our quantitative results further show that the generated data can either replace real data with minor performance loss or supplement real data for improved performance, thereby providing an inexpensive and scalable solution for data-hungry 3D applications. Project page: github.com/tev-fbk/synthfun3d.

Method: MRRE sequentially injects two precomputed vectors during inference: (1) cross-lingual reasoning enhancement vectors that steer non-English reasoning representations toward English space, and (2) target-language output anchoring vectors that restore target language distribution to maintain input-output consistency.

Result: Experiments across 6 LLMs/LVLMs on 4 reasoning benchmarks show MRRE consistently enhances non-English reasoning by average 5.48% gain (up to 7.54% in low-resource languages like Thai and Swahili) while improving input-output language consistency by 3.78%.

Conclusion: MRRE provides an effective training-free solution to enhance multilingual reasoning capabilities without additional data or tools, addressing fairness concerns in multilingual applications while maintaining language consistency.

Abstract: Large Language Models (LLMs) and Large Vision-Language Models (LVLMs) demonstrate strong reasoning capabilities, yet their performance in English significantly outperforms that in low-resource languages, raising fairness concerns in multilingual applications. Existing approaches either rely on costly multilingual training or employ prompting with external translation tools, both of which are resource-intensive and sensitive to translation quality. To address these limitations, we propose a training-free inference-time method to enhance Multilingual Reasoning capabilities via Representation Engineering (MRRE) without using any additional training data or tools. MRRE sequentially injects two precomputed vectors at specific layers during inference processing: cross-lingual reasoning enhancement vectors, which steer non-English reasoning representations toward English space to unlock multilingual reasoning, and target-language output anchoring vectors, which restore the distribution of the target language to preserve input-output language consistency. Comprehensive experiments across six advanced LLMs and LVLMs on four reasoning benchmarks demonstrate that MRRE consistently enhances non-English reasoning by an average gain of 5.48% and up to 7.54% in low-resource languages (Thai and Swahili), while improving input-output language consistency by 3.78%.

Method: ReGATE uses a teacher-student framework where a frozen teacher LLM provides per-token guidance losses, fused with exponential moving average of student’s difficulty estimates to dynamically select informative tokens and skip redundant ones in forward pass.

Result: ReGATE matches peak accuracy of standard training on MVBench up to 2× faster using only 38% of tokens, and with extended training surpasses baseline across multiple multimodal benchmarks while cutting total token usage by over 41%.

Conclusion: ReGATE effectively accelerates MLLM training through adaptive token pruning without architecture changes, achieving significant speedups and token reduction while maintaining or improving model performance.

Abstract: The computational cost of training multimodal large language models (MLLMs) grows rapidly with the number of processed tokens. Existing efficiency methods mainly target inference via token reduction or merging, offering limited benefits during training. We introduce ReGATE (Reference-Guided Adaptive Token Elision), an adaptive token pruning method for accelerating MLLM training. ReGATE adopts a teacher-student framework, in which a frozen teacher LLM provides per-token guidance losses that are fused with an exponential moving average of the student’s difficulty estimates. This adaptive scoring mechanism dynamically selects informative tokens while skipping redundant ones in the forward pass, substantially reducing computation without altering the model architecture. Across three representative MLLMs, ReGATE matches the peak accuracy of standard training on MVBench up to 2$\times$ faster, using only 38% of the tokens. With extended training, it even surpasses the baseline across multiple multimodal benchmarks, cutting total token usage by over 41%. Code and models will be released publicly.

Method: Pre-train separate models for each PBR component, then align them using cross-intrinsic attention that concatenates key and value features consistently. Use a rendering loss to ground intrinsic components and provide image-space signals for sharp details.

Result: Demonstrates detailed intrinsic generation with strong generalization capabilities that significantly outperforms existing intrinsic image decomposition methods when used with generated images.

Conclusion: IntrinsiX enables practical content creation applications including re-lighting, editing, and text-conditioned room-scale PBR texture generation by generating physically-based rendering maps from text descriptions.

Abstract: We introduce IntrinsiX, a novel method that generates high-quality intrinsic images from text description. In contrast to existing text-to-image models whose outputs contain baked-in scene lighting, our approach predicts physically-based rendering (PBR) maps. This enables the generated outputs to be used for content creation scenarios in core graphics applications that facilitate re-lighting, editing, and texture generation tasks. In order to train our generator, we exploit strong image priors, and pre-train separate models for each PBR material component (albedo, roughness, metallic, normals). We then align these models with a new cross-intrinsic attention formulation that concatenates key and value features in a consistent fashion. This allows us to exchange information between each output modality and to obtain semantically coherent PBR predictions. To ground each intrinsic component, we propose a rendering loss which provides image-space signals to constrain the model, thus facilitating sharp details also in the output BRDF properties. Our results demonstrate detailed intrinsic generation with strong generalization capabilities that outperforms existing intrinsic image decomposition methods used with generated images by a significant margin. Finally, we show a series of applications, including re-lighting, editing, and text-conditioned room-scale PBR texture generation.

Method: Benchmarked DR and DA techniques including feature-based methods (brightness filtering, perceptual hashing), GenAI, and classical rendering. Evaluated low/high-level feature alignment and controlled diffusion-based DA with real-world context prompts.

Result: Simpler feature-based methods outperform GenAI in accuracy and efficiency. Perceptual hashing achieved 98% mAP50 on industrial dataset and 67% on robotics dataset. GenAI showed significant time overhead without performance improvement.

Conclusion: Feature-based methods like perceptual hashing offer efficient sim-to-real bridging, enabling high real-world performance from synthetic-only training without complex GenAI overhead.

Abstract: Reducing the burden of data generation and annotation remains a major challenge for the cost-effective deployment of machine learning in industrial and robotics settings. While synthetic rendering is a promising solution, bridging the sim-to-real gap often requires expert intervention. In this work, we benchmark a range of domain randomization (DR) and domain adaptation (DA) techniques, including feature-based methods, generative AI (GenAI), and classical rendering approaches, for creating contextualized synthetic data without manual annotation. Our evaluation focuses on the effectiveness and efficiency of low-level and high-level feature alignment, as well as a controlled diffusion-based DA method guided by prompts generated from real-world contexts. We validate our methods on two datasets: a proprietary industrial dataset (automotive and logistics) and a public robotics dataset. Results show that if render-based data with enough variability is available as seed, simpler feature-based methods, such as brightness-based and perceptual hashing filtering, outperform more complex GenAI-based approaches in both accuracy and resource efficiency. Perceptual hashing consistently achieves the highest performance, with mAP50 scores of 98% and 67% on the industrial and robotics datasets, respectively. Additionally, GenAI methods present significant time overhead for data generation at no apparent improvement of sim-to-real mAP values compared to simpler methods. Our findings offer actionable insights for efficiently bridging the sim-to-real gap, enabling high real-world performance from models trained exclusively on synthetic data.

Method: Extend SAEs to VLMs (CLIP), develop a comprehensive framework for evaluating neuron-level monosemanticity in visual representations, create a benchmark from large-scale user study for human-aligned evaluation, and apply SAE interventions on CLIP’s vision encoder.

Result: SAEs trained on VLMs significantly enhance monosemanticity of individual neurons, with sparsity and wide latents being most influential factors. SAE interventions on CLIP’s vision encoder can directly steer multimodal LLM outputs (e.g., LLaVA) without modifying the underlying language model.

Conclusion: SAEs are practical and effective unsupervised tools for enhancing both interpretability and control of Vision-Language Models, bridging the gap between LLM interpretability methods and multimodal systems.

Abstract: Sparse Autoencoders (SAEs) have recently gained attention as a means to improve the interpretability and steerability of Large Language Models (LLMs), both of which are essential for AI safety. In this work, we extend the application of SAEs to Vision-Language Models (VLMs), such as CLIP, and introduce a comprehensive framework for evaluating monosemanticity at the neuron-level in visual representations. To ensure that our evaluation aligns with human perception, we propose a benchmark derived from a large-scale user study. Our experimental results reveal that SAEs trained on VLMs significantly enhance the monosemanticity of individual neurons, with sparsity and wide latents being the most influential factors. Further, we demonstrate that applying SAE interventions on CLIP’s vision encoder directly steers multimodal LLM outputs (e.g., LLaVA), without any modifications to the underlying language model. These findings emphasize the practicality and efficacy of SAEs as an unsupervised tool for enhancing both interpretability and control of VLMs. Code and benchmark data are available at https://github.com/ExplainableML/sae-for-vlm.

Method: FACT-GS introduces a frequency-aligned complexity-aware texture framework that reformulates texture parameterization as a differentiable sampling-density allocation problem. It replaces uniform textures with a learnable frequency-aware allocation strategy using a deformation field whose Jacobian modulates local sampling density, performing non-uniform sampling on fixed-resolution texture grids.

Result: The method preserves real-time performance while recovering sharper high-frequency details under the same parameter budget, improving texture space utilization and visual quality.

Conclusion: FACT-GS provides an efficient adaptive sampling approach for Gaussian Splatting that better aligns texture allocation with visual complexity, enabling higher-quality scene appearance modeling without sacrificing real-time performance.

Abstract: Realistic scene appearance modeling has advanced rapidly with Gaussian Splatting, which enables real-time, high-quality rendering. Recent advances introduced per-primitive textures that incorporate spatial color variations within each Gaussian, improving their expressiveness. However, texture-based Gaussians parameterize appearance with a uniform per-Gaussian sampling grid, allocating equal sampling density regardless of local visual complexity. This leads to inefficient texture space utilization, where high-frequency regions are under-sampled and smooth regions waste capacity, causing blurred appearance and loss of fine structural detail. We introduce FACT-GS, a Frequency-Aligned Complexity-aware Texture Gaussian Splatting framework that allocates texture sampling density according to local visual frequency. Grounded in adaptive sampling theory, FACT-GS reformulates texture parameterization as a differentiable sampling-density allocation problem, replacing the uniform textures with a learnable frequency-aware allocation strategy implemented via a deformation field whose Jacobian modulates local sampling density. Built on 2D Gaussian Splatting, FACT-GS performs non-uniform sampling on fixed-resolution texture grids, preserving real-time performance while recovering sharper high-frequency details under the same parameter budget.

Method: Develops a variational framework with specific perceptual requirements, identifies functionals satisfying these requirements, and uses gradient descent for optimization. Also introduces computational cost reduction from O(N²) to O(N log N).

Result: Proposes three explicit functionals of basic interest that satisfy perceptual requirements, shows their similarities/differences with existing models, and provides efficient computational methodology.

Conclusion: The variational approach provides a mathematically sound framework for perceptually-inspired color contrast enhancement with improved computational efficiency and clear characterization of image feature behavior.

Abstract: Basic phenomenology of human color vision has been widely taken as an inspiration to devise explicit color correction algorithms. The behavior of these models in terms of significative image features (such as contrast and dispersion) can be difficult to characterize. To cope with this, we propose to use a variational formulation of color contrast enhancement that is inspired by the basic phenomenology of color perception. In particular, we devise a set of basic requirements to be fulfilled by an energy to be considered as `perceptually inspired’, showing that there is an explicit class of functionals satisfying all of them. We single out three explicit functionals that we consider of basic interest, showing similarities and differences with existing models. The minima of such functionals is computed using a gradient descent approach. We also present a general methodology to reduce the computational cost of the algorithms under analysis from ${\cal O}(N^2)$ to ${\cal O}(N\log N)$, being $N$ the number of input pixels.

Method: Entropy Rectifying Guidance (ERG) modifies the attention mechanism in diffusion transformer architectures during inference time. It’s more general than CFG as it works for both conditional and unconditional sampling, and can be combined with other guidance methods.

Result: ERG shows significant improvements across text-to-image, class-conditional, and unconditional image generation tasks. It simultaneously enhances quality, diversity, and prompt consistency without the trade-offs of CFG.

Conclusion: ERG is a simple yet effective guidance method that addresses the limitations of CFG by modifying attention mechanisms, offering simultaneous improvements across multiple generation metrics while being computationally efficient and compatible with other guidance techniques.

Abstract: Guidance techniques are commonly used in diffusion and flow models to improve image quality and input consistency for conditional generative tasks such as class-conditional and text-to-image generation. In particular, classifier-free guidance (CFG) is the most widely adopted guidance technique. It results, however, in trade-offs across quality, diversity and consistency: improving some at the expense of others. While recent work has shown that it is possible to disentangle these factors to some extent, such methods come with an overhead of requiring an additional (weaker) model, or require more forward passes per sampling step. In this paper, we propose Entropy Rectifying Guidance (ERG), a simple and effective guidance method based on inference-time changes in the attention mechanism of state-of-the-art diffusion transformer architectures, which allows for simultaneous improvements over image quality, diversity and prompt consistency. ERG is more general than CFG and similar guidance techniques, as it extends to unconditional sampling. We show that ERG results in significant improvements in various tasks, including text-to-image, class-conditional and unconditional image generation. We also show that ERG can be seamlessly combined with other recent guidance methods such as CADS and APG, further improving generation results.

Method: 1) Created GeoSeg-1M dataset via automatic mask filtering and instruction generation pipeline synthesizing referring, interactive, and reasoning segmentation instructions from multiple public datasets. 2) Developed GeoSeg-Bench benchmark for evaluating contextual understanding and reasoning. 3) Proposed UniGeoSeg unified framework with task-aware text enhancement, latent knowledge memory, and progressive training strategy for multi-task learning.

Result: UniGeoSeg achieves state-of-the-art performance across GeoSeg-Bench and diverse public benchmarks while exhibiting strong zero-shot generalization capabilities.

Conclusion: The proposed GeoSeg-1M dataset, GeoSeg-Bench benchmark, and UniGeoSeg framework address limitations in remote sensing instruction-driven segmentation, enabling better understanding and generalization through large-scale data and unified multi-task learning.

Abstract: Instruction-driven segmentation in remote sensing generates masks from guidance, offering great potential for accessible and generalizable applications. However, existing methods suffer from fragmented task formulations and limited instruction data, hindering effective understanding and generalization. To address these issues, we introduce GeoSeg-1M, the first million-scale dataset for remote sensing instruction-driven segmentation, constructed via an automatic mask filtering and instruction generation pipeline that synthesizes referring, interactive, and reasoning segmentation instructions from multiple public datasets. GeoSeg-1M contains 590K images, 117 categories, and 1.1M image-mask-instruction triplets. Building upon this foundation, we further curate GeoSeg-Bench, a challenging benchmark designed to evaluate contextual understanding and reasoning capabilities across diverse instruction-driven tasks and complex geospatial scenes. Furthermore, we present UniGeoSeg, a unified framework that serves as a strong baseline, incorporating task-aware text enhancement, latent knowledge memory, and a progressive training strategy to facilitate multi-task learning. Extensive experiments demonstrate the state-of-the-art performance of UniGeoSeg across GeoSeg-Bench and diverse public benchmarks, while exhibiting strong zero-shot generalization. Datasets and source code were released at https://github.com/MiliLab/UniGeoSeg.

Method: Created CANVAS benchmark with 598 tool-based design tasks from 3.3K mobile UI designs across 30 function categories. Tasks involve step-by-step tool invocations to update designs. Includes two task types: design replication (reproduce whole UI) and design modification (modify specific parts).

Result: Leading models show more strategic tool invocations that improve design quality. The benchmark identifies common error patterns in models’ tool-based design capabilities.

Conclusion: CANVAS enables systematic evaluation of VLMs’ tool-based UI design capabilities, revealing current strengths and limitations, and provides guidance for future improvements in AI-assisted design collaboration.

Abstract: User interface (UI) design is an iterative process in which designers progressively refine their work with design software such as Figma or Sketch. Recent advances in vision language models (VLMs) with tool invocation suggest these models can operate design software to edit a UI design through iteration. Understanding and enhancing this capacity is important, as it highlights VLMs’ potential to collaborate with designers within conventional software. However, as no existing benchmark evaluates tool-based design performance, the capacity remains unknown. To address this, we introduce CANVAS, a benchmark for VLMs on tool-based user interface design. Our benchmark contains 598 tool-based design tasks paired with ground-truth references sampled from 3.3K mobile UI designs across 30 function-based categories (e.g., onboarding, messaging). In each task, a VLM updates the design step-by-step through context-based tool invocations (e.g., create a rectangle as a button background), linked to design software. Specifically, CANVAS incorporates two task types: (i) design replication evaluates the ability to reproduce a whole UI screen; (ii) design modification evaluates the ability to modify a specific part of an existing screen. Results suggest that leading models exhibit more strategic tool invocations, improving design quality. Furthermore, we identify common error patterns models exhibit, guiding future work in enhancing tool-based design capabilities.

Method: Reformulate VAR as a Markov process (Markov-VAR) using Markovian Scale Prediction: treat each scale as a Markov state, introduce a sliding window to compress previous scales into a compact history vector, and combine this with the Markov state to create a representative dynamic state that evolves under a Markov process.

Result: Markov-VAR reduces FID by 10.5% on ImageNet 256×256 and decreases peak memory consumption by 83.8% on 1024×1024 compared to original VAR, demonstrating both improved performance and efficiency.

Conclusion: Markov-VAR provides an extremely simple yet highly effective foundation for future research on visual autoregressive generation and downstream tasks by addressing the efficiency limitations of full-context dependency while maintaining or improving performance.

Abstract: Visual AutoRegressive modeling (VAR) based on next-scale prediction has revitalized autoregressive visual generation. Although its full-context dependency, i.e., modeling all previous scales for next-scale prediction, facilitates more stable and comprehensive representation learning by leveraging complete information flow, the resulting computational inefficiency and substantial overhead severely hinder VAR’s practicality and scalability. This motivates us to develop a new VAR model with better performance and efficiency without full-context dependency. To address this, we reformulate VAR as a non-full-context Markov process, proposing Markov-VAR. It is achieved via Markovian Scale Prediction: we treat each scale as a Markov state and introduce a sliding window that compresses certain previous scales into a compact history vector to compensate for historical information loss owing to non-full-context dependency. Integrating the history vector with the Markov state yields a representative dynamic state that evolves under a Markov process. Extensive experiments demonstrate that Markov-VAR is extremely simple yet highly effective: Compared to VAR on ImageNet, Markov-VAR reduces FID by 10.5% (256 $\times$ 256) and decreases peak memory consumption by 83.8% (1024 $\times$ 1024). We believe that Markov-VAR can serve as a foundation for future research on visual autoregressive generation and other downstream tasks.

Method: G²VLM bridges spatial 3D reconstruction and spatial understanding by natively leveraging learned 3D visual geometry features. It uses in-context learning and interleaved reasoning to enhance spatial reasoning, training on abundant multi-view image/video data while benefiting from 3D visual priors without hard-to-collect annotations.

Result: G²VLM achieves comparable results to state-of-the-art feed-forward 3D reconstruction models and shows better or competitive performance across spatial understanding and reasoning tasks.

Conclusion: By unifying semantically strong VLMs with low-level 3D vision tasks, G²VLM serves as a strong baseline for the community and can unlock future applications like 3D scene editing.

Abstract: Vision-Language Models (VLMs) still lack robustness in spatial intelligence, demonstrating poor performance on spatial understanding and reasoning tasks. We attribute this gap to the absence of a visual geometry learning process capable of reconstructing 3D space from 2D images. We present G$^2$VLM, a geometry grounded vision-language model that bridges two fundamental aspects of spatial intelligence: spatial 3D reconstruction and spatial understanding. G$^2$VLM natively leverages learned 3D visual geometry features to directly predict 3D attributes and enhance spatial reasoning tasks via in-context learning and interleaved reasoning. Our unified design is highly scalable for spatial understanding: it trains on abundant multi-view image and video data, while simultaneously leveraging the benefits of 3D visual priors that are typically only derived from hard-to-collect annotations. Experimental results demonstrate G$^2$VLM is proficient in both tasks, achieving comparable results to state-of-the-art feed-forward 3D reconstruction models and achieving better or competitive results across spatial understanding and reasoning tasks. By unifying a semantically strong VLM with low-level 3D vision tasks, we hope G$^2$VLM can serve as a strong baseline for the community and unlock more future applications, such as 3D scene editing.

Method: Hierarchical detection framework combining YOLOv11 for monitor and ROI localization with PaddleOCR for text extraction. Includes geometric rectification module to standardize screen perspective across variable camera angles and lighting conditions.

Result: Evaluated on 6,498 images from open-source corpora and real-world ICUs in Vietnam. Achieved mAP@50-95 of 99.5% for monitor detection, 91.5% for vital sign ROI localization, and end-to-end extraction accuracy >98.9% for core parameters (heart rate, SpO2, blood pressure).

Conclusion: Lightweight, camera-based approach can reliably transform unstructured screen data into structured digital data, providing practical and scalable pathway to improve information accessibility and clinical documentation in low-resource settings.

Abstract: In many low-resource healthcare settings, bedside monitors remain standalone legacy devices without network connectivity, creating a persistent interoperability gap that prevents seamless integration of physiological data into electronic health record (EHR) systems. To address this challenge without requiring costly hardware replacement, we present a computer vision-based pipeline for the automated capture and digitisation of vital sign data directly from bedside monitor screens. Our method employs a hierarchical detection framework combining YOLOv11 for accurate monitor and region of interest (ROI) localisation with PaddleOCR for robust text extraction. To enhance reliability across variable camera angles and lighting conditions, a geometric rectification module standardizes the screen perspective before character recognition. We evaluated the system on a dataset of 6,498 images collected from open-source corpora and real-world intensive care units in Vietnam. The model achieved a mean Average Precision (mAP@50-95) of 99.5% for monitor detection and 91.5% for vital sign ROI localisation. The end-to-end extraction accuracy exceeded 98.9% for core physiological parameters, including heart rate, oxygen saturation SpO2, and arterial blood pressure. These results demonstrate that a lightweight, camera-based approach can reliably transform unstructured information from screen captures into structured digital data, providing a practical and scalable pathway to improve information accessibility and clinical documentation in low-resource settings.

Method: Uses neural rendering with reactive environments to synthesize multi-view observations from perturbed ego trajectories, plus pseudo-expert trajectory generation for action supervision, enabling co-training on real and simulated data.

Result: Significant improvements in robustness and generalization: +6.8 EPDMS on navhard and +2.9 on navtest benchmarks, with smooth scaling using only simulation data without additional real-world data.

Conclusion: SimScale demonstrates that scalable simulation with pseudo-expert supervision can effectively enhance autonomous driving policies, revealing crucial design insights about pseudo-experts and scaling properties across different architectures.

Abstract: Achieving fully autonomous driving systems requires learning rational decisions in a wide span of scenarios, including safety-critical and out-of-distribution ones. However, such cases are underrepresented in real-world corpus collected by human experts. To complement for the lack of data diversity, we introduce a novel and scalable simulation framework capable of synthesizing massive unseen states upon existing driving logs. Our pipeline utilizes advanced neural rendering with a reactive environment to generate high-fidelity multi-view observations controlled by the perturbed ego trajectory. Furthermore, we develop a pseudo-expert trajectory generation mechanism for these newly simulated states to provide action supervision. Upon the synthesized data, we find that a simple co-training strategy on both real-world and simulated samples can lead to significant improvements in both robustness and generalization for various planning methods on challenging real-world benchmarks, up to +6.8 EPDMS on navhard and +2.9 on navtest. More importantly, such policy improvement scales smoothly by increasing simulation data only, even without extra real-world data streaming in. We further reveal several crucial findings of such a sim-real learning system, which we term SimScale, including the design of pseudo-experts and the scaling properties for different policy architectures. Our simulation data and code would be released.

Method: Created DEAL-300K dataset using: 1) multi-modal LLM for editing instructions, 2) mask-free diffusion editor for manipulated images, 3) active-learning change detection for pixel-level annotations. Proposed localization framework with frozen Visual Foundation Model and Multi Frequency Prompt Tuning (MFPT) to capture semantic and frequency-domain cues.

Result: Method achieves 82.56% pixel-level F1 score on DEAL-300K test split and 80.97% on external CoCoGlide benchmark, establishing strong baselines for diffusion-based image manipulation localization.

Conclusion: DEAL-300K provides a practical foundation for future DIML research, addressing the challenge of localizing smoothly blended diffusion edits that current benchmarks overlook.

Abstract: Diffusion-based image editing has made semantic level image manipulation easy for general users, but it also enables realistic local forgeries that are hard to localize. Existing benchmarks mainly focus on the binary detection of generated images or the localization of manually edited regions and do not reflect the properties of diffusion-based edits, which often blend smoothly into the original content. We present Diffusion-Based Image Editing Area Localization Dataset (DEAL-300K), a large scale dataset for diffusion-based image manipulation localization (DIML) with more than 300,000 annotated images. We build DEAL-300K by using a multi-modal large language model to generate editing instructions, a mask-free diffusion editor to produce manipulated images, and an active-learning change detection pipeline to obtain pixel-level annotations. On top of this dataset, we propose a localization framework that uses a frozen Visual Foundation Model (VFM) together with Multi Frequency Prompt Tuning (MFPT) to capture both semantic and frequency-domain cues of edited regions. Trained on DEAL-300K, our method reaches a pixel-level F1 score of 82.56% on our test split and 80.97% on the external CoCoGlide benchmark, providing strong baselines and a practical foundation for future DIML research.The dataset can be accessed via https://github.com/ymhzyj/DEAL-300K.

Method: VQRAE builds on pretrained vision foundation models with a symmetric ViT decoder. It uses a two-stage training: 1) freeze encoder and learn high-dimensional semantic VQ codebook with pixel reconstruction, 2) jointly optimize encoder with self-distillation constraints. This creates both continuous semantic features and discrete tokens.

Result: VQRAE achieves competitive performance on visual understanding, generation, and reconstruction benchmarks. The semantic VQ codebook achieves 100% utilization ratio at 1536 dimensions, showing promising scaling properties in autoregressive paradigms due to its discrete token advantages.

Conclusion: VQRAE successfully demonstrates a unified approach to multimodal representation that bridges understanding and generation, with the key insight that high-dimensional codebooks work better for semantic quantization than the previously common low-dimensional approach for image reconstruction.

Abstract: Unifying multimodal understanding, generation and reconstruction representation in a single tokenizer remains a key challenge in building unified models. Previous research predominantly attempts to address this in a dual encoder paradigm, e.g., utilizing the separate encoders for understanding and generation respectively or balancing semantic representations and low-level features with contrastive loss. In this paper, we propose VQRAE, a Vector Quantization version of Representation AutoEncoders, which pioneers the first exploration in unified representation to produce Continuous semantic features for image understanding and Discrete tokens for visual generation within a unified tokenizer. Specifically, we build upon pretrained vision foundation models with a symmetric ViT decoder and adopt a two-stage training strategy: first, it freezes the encoder and learns a high-dimensional semantic VQ codebook with pixel reconstruction objective; then jointly optimizes the encoder with self-distillation constraints. This design enables negligible semantic information for maintaining the ability of multimodal understanding, discrete tokens that are compatible for generation and fine-grained reconstruction. Besides, we identify the intriguing property in quantizing semantic encoders that rely on high-dimensional codebook in contrast to the previous common practice of low-dimensional codebook in image reconstruction. The semantic VQ codebook can achieve a 100% utilization ratio at a dimension of 1536. VQRAE presents competitive performance on several benchmarks of visual understanding, generation and reconstruction with promising scaling property in the autoregressive paradigm for its discrete merits.

Method: Proposes a physics-informed framework with: 1) dual-positive contrastive learning coupling temporal consistency with Beer-Lambert augmentations for robust features, 2) multi-stage pipeline augmenting motion-based tracking with physics-informed secondary association integrating geometric consistency and appearance similarity for re-identification under occlusion/drift.

Result: Achieves state-of-the-art performance on four underwater benchmarks (WebUOT-1M, UOT32, UTB180, UWCOT220), improving Success AUC by up to 6%, while ensuring stable long-term generalized underwater tracking and efficient runtime.

Conclusion: MANTA successfully integrates physics-informed representation learning with tracking design to address underwater distortions, demonstrating superior performance and generalization across diverse underwater conditions through its novel contrastive learning and association strategies.

Abstract: Underwater object tracking is challenging due to wavelength dependent attenuation and scattering, which severely distort appearance across depths and water conditions. Existing trackers trained on terrestrial data fail to generalize to these physics-driven degradations. We present MANTA, a physics-informed framework integrating representation learning with tracking design for underwater scenarios. We propose a dual-positive contrastive learning strategy coupling temporal consistency with Beer-Lambert augmentations to yield features robust to both temporal and underwater distortions. We further introduce a multi-stage pipeline augmenting motion-based tracking with a physics-informed secondary association algorithm that integrates geometric consistency and appearance similarity for re-identification under occlusion and drift. To complement standard IoU metrics, we propose Center-Scale Consistency (CSC) and Geometric Alignment Score (GAS) to assess geometric fidelity. Experiments on four underwater benchmarks (WebUOT-1M, UOT32, UTB180, UWCOT220) show that MANTA achieves state-of-the-art performance, improving Success AUC by up to 6 percent, while ensuring stable long-term generalized underwater tracking and efficient runtime.

Method: 1) Created Ivy-Fake benchmark with over 106K richly annotated training samples and 5,000 manually verified evaluation examples from multiple generative models and real datasets. 2) Proposed Ivy-xDetector, a reinforcement learning model based on Group Relative Policy Optimization (GRPO) for producing explainable reasoning chains.

Result: Extensive experiments demonstrate superiority of the dataset and effectiveness of the approach. Notably, the method improves performance on GenImage from 86.88% to 96.32%, surpassing prior state-of-the-art methods by a clear margin.

Conclusion: The Ivy-Fake benchmark and Ivy-xDetector model address critical limitations in AIGC detection by providing multidimensional explainable datasets and fine-grained interpretability through reinforcement learning, achieving state-of-the-art performance on synthetic content detection benchmarks.

Abstract: The rapid development of Artificial Intelligence Generated Content (AIGC) techniques has enabled the creation of high-quality synthetic content, but it also raises significant security concerns. Current detection methods face two major limitations: (1) the lack of multidimensional explainable datasets for generated images and videos. Existing open-source datasets (e.g., WildFake, GenVideo) rely on oversimplified binary annotations, which restrict the explainability and trustworthiness of trained detectors. (2) Prior MLLM-based forgery detectors (e.g., FakeVLM) exhibit insufficiently fine-grained interpretability in their step-by-step reasoning, which hinders reliable localization and explanation. To address these challenges, we introduce Ivy-Fake, the first large-scale multimodal benchmark for explainable AIGC detection. It consists of over 106K richly annotated training samples (images and videos) and 5,000 manually verified evaluation examples, sourced from multiple generative models and real world datasets through a carefully designed pipeline to ensure both diversity and quality. Furthermore, we propose Ivy-xDetector, a reinforcement learning model based on Group Relative Policy Optimization (GRPO), capable of producing explainable reasoning chains and achieving robust performance across multiple synthetic content detection benchmarks. Extensive experiments demonstrate the superiority of our dataset and confirm the effectiveness of our approach. Notably, our method improves performance on GenImage from 86.88% to 96.32%, surpassing prior state-of-the-art methods by a clear margin.

Method: Learns generic motion representations directly from raw video data using an image-space reconstruction objective, disentangling motion from appearance. The representation can be combined with existing video generators via lightweight adapters.

Result: Enables open-world motion transfer across semantically unrelated entities without requiring object correspondences. Outperforms state-of-the-art video representation models like V-JEPA in zero-shot action classification on Something-Something v2 and Jester benchmarks.

Conclusion: DisMo provides a novel paradigm for learning abstract motion representations that disentangle motion from appearance, enabling accurate motion transfer and benefiting from future video model advancements via adapter-based integration.

Abstract: Recent advances in text-to-video (T2V) and image-to-video (I2V) models, have enabled the creation of visually compelling and dynamic videos from simple textual descriptions or initial frames. However, these models often fail to provide an explicit representation of motion separate from content, limiting their applicability for content creators. To address this gap, we propose DisMo, a novel paradigm for learning abstract motion representations directly from raw video data via an image-space reconstruction objective. Our representation is generic and independent of static information such as appearance, object identity, or pose. This enables open-world motion transfer, allowing motion to be transferred across semantically unrelated entities without requiring object correspondences, even between vastly different categories. Unlike prior methods, which trade off motion fidelity and prompt adherence, are overfitting to source structure or drifting from the described action, our approach disentangles motion semantics from appearance, enabling accurate transfer and faithful conditioning. Furthermore, our motion representation can be combined with any existing video generator via lightweight adapters, allowing us to effortlessly benefit from future advancements in video models. We demonstrate the effectiveness of our method through a diverse set of motion transfer tasks. Finally, we show that the learned representations are well-suited for downstream motion understanding tasks, consistently outperforming state-of-the-art video representation models such as V-JEPA in zero-shot action classification on benchmarks including Something-Something v2 and Jester. Project page: https://compvis.github.io/DisMo

Method: AutoSEP uses iterative self-supervised prompt learning with unlabeled data to learn description prompts that guide MLLMs to identify crucial discriminative features, based on instance-level classification scoring without any training or fine-tuning.

Result: AutoSEP consistently outperforms other unsupervised baselines across multiple fine-grained classification datasets, improving 13% over standard zero-shot classification and 5% over best-performing baselines on average.

Conclusion: The self-supervised optimization framework effectively enhances MLLM fine-grained classification capabilities in a fully unsupervised manner, demonstrating the value of learned description prompts for focusing on subtle discriminative features.

Abstract: Despite Multimodal Large Language Models (MLLMs) showing promising results on general zero-shot image classification tasks, fine-grained image classification remains challenging. It demands precise attention to subtle visual details to distinguish between visually similar subcategories–details that MLLMs may easily overlook without explicit guidance. To address this, we introduce AutoSEP, an iterative self-supervised prompt learning framework designed to enhance MLLM fine-grained classification capabilities in a fully unsupervised manner. Our core idea is to leverage unlabeled data to learn a description prompt that guides MLLMs in identifying crucial discriminative features within an image, and boosts classification accuracy. We developed an automatic self-enhancing prompt learning framework called AutoSEP to iteratively improve the description prompt using unlabeled data, based on instance-level classification scoring function. AutoSEP only requires black-box access to MLLMs, eliminating the need for any training or fine-tuning. We evaluate our approach on multiple fine-grained classification datasets. It consistently outperforms other unsupervised baselines, demonstrating the effectiveness of our self-supervised optimization framework. Notably, AutoSEP on average improves 13 percent over standard zero-shot classification and 5 percent over the best-performing baselines. Code is available at: https://github.com/yq-hong/AutoSEP

Method: Introduces instruction-driven interaction paradigm using natural language prompts, keyboard, or mouse signals. Develops automated process to transform unstructured text-video pairs into causally aligned interactive datasets. Builds on 14B image-to-video Mixture-of-Experts foundation model with text-driven interaction injection mechanism for fine-grained control over camera motion, character behavior, and environment dynamics.

Result: Model generates temporally coherent and causally grounded interactive game videos that faithfully respond to diverse and free-form user instructions (e.g., “open the door”, “draw a torch”, “trigger an explosion”). Introduces InterBench benchmark for comprehensive interaction evaluation.

Conclusion: Hunyuan-GameCraft-2 represents a significant advancement in generative game world modeling, enabling flexible and semantically rich interaction through natural language and input signals, overcoming limitations of previous approaches with rigid action schemas.

Abstract: Recent advances in generative world models have enabled remarkable progress in creating open-ended game environments, evolving from static scene synthesis toward dynamic, interactive simulation. However, current approaches remain limited by rigid action schemas and high annotation costs, restricting their ability to model diverse in-game interactions and player-driven dynamics. To address these challenges, we introduce Hunyuan-GameCraft-2, a new paradigm of instruction-driven interaction for generative game world modeling. Instead of relying on fixed keyboard inputs, our model allows users to control game video contents through natural language prompts, keyboard, or mouse signals, enabling flexible and semantically rich interaction within generated worlds. We formally defined the concept of interactive video data and developed an automated process to transform large-scale, unstructured text-video pairs into causally aligned interactive datasets. Built upon a 14B image-to-video Mixture-of-Experts(MoE) foundation model, our model incorporates a text-driven interaction injection mechanism for fine-grained control over camera motion, character behavior, and environment dynamics. We introduce an interaction-focused benchmark, InterBench, to evaluate interaction performance comprehensively. Extensive experiments demonstrate that our model generates temporally coherent and causally grounded interactive game videos that faithfully respond to diverse and free-form user instructions such as “open the door”, “draw a torch”, or “trigger an explosion”.

Method: The dataset contains 39,760 camera trap images annotated with segmentation masks, identity labels, and 20-point skeletons covering 319 unique individuals over 15 years in two regions. It also includes a Unity-based synthetic image generation pipeline with diffusion-based text-to-texture modeling for creating additional training data.

Result: CzechLynx provides the first comprehensive dataset for Eurasian lynx computer vision tasks, featuring real camera trap data, synthetic data generation capabilities, and three evaluation protocols (geo-aware, time-aware open-set, and time-aware closed-set) for systematic testing across ecological scenarios.

Conclusion: CzechLynx offers a unique, flexible benchmark for robust evaluation of computer vision and machine learning models in realistic wildlife monitoring contexts, enabling cross-regional and long-term ecological scenario testing.

Abstract: We introduce CzechLynx, the first large-scale, open-access dataset for individual identification, pose estimation, and instance segmentation of the Eurasian lynx (Lynx lynx). CzechLynx contains 39,760 camera trap images annotated with segmentation masks, identity labels, and 20-point skeletons and covers 319 unique individuals across 15 years of systematic monitoring in two geographically distinct regions: southwest Bohemia and the Western Carpathians. In addition to the real camera trap data, we provide a large complementary set of photorealistic synthetic images and a Unity-based generation pipeline with diffusion-based text-to-texture modeling, capable of producing arbitrarily large amounts of synthetic data spanning diverse environments, poses, and coat-pattern variations. To enable systematic testing across realistic ecological scenarios, we define three complementary evaluation protocols: (i) geo-aware, (ii) time-aware open-set, and (iii) time-aware closed-set, covering cross-regional and long-term monitoring settings. With the provided resources, CzechLynx offers a unique, flexible benchmark for robust evaluation of computer vision and machine learning models across realistic ecological scenarios.

Method: Introduces object-centric data setting (multi-view images or 3D models) and evaluates four data synthesis methods: simple image processing techniques, 3D rendering, and image diffusion models to create realistic, cluttered images with varying contextual coherence and complexity.

Result: Demonstrates significant performance boosts for object detection models on novel object categories in the data-constrained experimental setting, enabling category-level generalization in real-world data.

Conclusion: Data synthesis methods using object-centric data can effectively extend object detection capabilities to new categories with limited training data, addressing the challenge of costly annotation for long-tailed classes.

Abstract: Deep learning approaches to object detection have achieved reliable detection of specific object classes in images. However, extending a model’s detection capability to new object classes requires large amounts of annotated training data, which is costly and time-consuming to acquire, especially for long-tailed classes with insufficient representation in existing datasets. Here, we introduce the object-centric data setting, when limited data is available in the form of object-centric data (multi-view images or 3D models), and systematically evaluate the performance of four different data synthesis methods to finetune object detection models on novel object categories in this setting. The approaches are based on simple image processing techniques, 3D rendering, and image diffusion models, and use object-centric data to synthesize realistic, cluttered images with varying contextual coherence and complexity. We assess how these methods enable models to achieve category-level generalization in real-world data, and demonstrate significant performance boosts within this data-constrained experimental setting.

Method: VITA uses test-time adaptation where a lightweight adaptation module is updated via gradient steps on a meta-learned self-supervised loss during inference. It employs dissimilarity-based sampling to select semantically diverse trajectory segments during training to mitigate shortcut learning.

Result: VITA generalizes from a single training environment to diverse out-of-distribution tasks, environments, and embodiments, outperforming state-of-the-art zero-shot methods using autoregressive VLMs. It also enables reward shaping in offline RL, producing multi-task policies that exceed those trained with simulation’s dense rewards.

Conclusion: VITA successfully addresses the limitations of frozen VLMs for zero-shot value estimation through test-time adaptation, enabling better generalization and temporal reasoning for real-world robotic manipulation tasks.

Abstract: Vision-Language Models (VLMs) show promise as zero-shot goal-conditioned value functions, but their frozen pre-trained representations limit generalization and temporal reasoning. We introduce VITA, a zero-shot value function learning method that enhances both capabilities via test-time adaptation. At inference, a lightweight adaptation module is updated via a gradient step on a meta-learned self-supervised loss, such that each test-time update improves value estimation. By updating sequentially over a trajectory, VITA encodes history into its parameters, addressing the temporal reasoning limitations. To mitigate shortcut learning, we propose a dissimilarity-based sampling strategy that selects semantically diverse segments of the trajectory during training. In real-world robotic manipulation tasks, VITA generalizes from a single training environment to diverse out-of-distribution tasks, environments, and embodiments, outperforming the state-of-the-art zero-shot method using autoregressive VLMs. Furthermore, we demonstrate that VITA’s zero-shot value estimates can be utilized for reward shaping in offline reinforcement learning, resulting in multi-task policies on the Meta-World benchmark that exceed the performance of those trained with the simulation’s fuzzy-logic dense rewards.

Method: Proposes VGT-AE that aligns semantic encoders from pretrained VLMs with latent representations of pixel decoders, replacing entangled pixel-level VAEs. Uses efficient visual generation tuning to stimulate visual generation capabilities in existing VLMs.

Result: Achieves 26.67 PSNR and 0.50 rFID at 28x compression ratio in reconstruction; 0.77 on GenEval and 78.73 on DPG-Bench in generation. Provides 20x speedup in convergence and shows significant scaling promise.

Conclusion: VGT demonstrates that VLMs trained for understanding inherently possess visual generation potential, enabling unified multimodal foundation models without extensive retraining.

Abstract: Large Vision Language Models (VLMs) effectively bridge the modality gap through extensive pretraining, acquiring sophisticated visual representations aligned with language. However, it remains underexplored whether these representations, optimized for multimodal understanding tasks, harbor an inherent potential for visual generation. In this paper, we propose VGT, Visual Generation Tuning, a novel paradigm designed to stimulate the underlying capabilities of visual generation within any vision language models. By performing efficient visual generation tuning on well-pretrained VLMs, we significantly mitigate the alignment costs and accelerate the convergence of autoregressive modeling in the continuous space (20x speedup). Specifically, we dismiss the entangled pixel-level VAEs designed for diffusion transformers and formulate VGT-AE through aligning the semantic encoders from pretrained VLMs with the latent representations of pixel decoders. In image reconstruction tasks, we achieve 26.67 PSNR and 0.50 rFID at a 28x compression ratio, outperforming specialized VAEs; in visual generation tasks, we achieve state-of-the-art outcomes among autoregressive models, 0.77 on GenEval and 78.73 on DPG-Bench. Furthermore, our proposed VGT showcases significant scaling promise and is versatile for endowing any VLMs trained for multimodal understanding with the capabilities of visual generation, which paves the new avenue to explore next-generation unified multimodal foundation models. Models and codes are available at https://github.com/hustvl/VGT.

Method: Proposes extensible multi-stream processing architecture with identity-aware attention mechanism in Diffusion Transformers that iteratively processes identity-audio pairs. Training pipeline uses mainly single-person videos to learn speaking patterns, refined with few real multi-person clips.

Result: Achieves remarkable lip synchronization, visual quality, and natural interactivity while balancing data costs and identity scalability. Also contributes evaluation metric and dataset for multi-person video naturalness and interactivity.

Conclusion: AnyTalker effectively addresses multi-person video generation challenges through innovative architecture and efficient training, enabling scalable identity-driven generation with minimal multi-person data requirements.

Abstract: Recently, multi-person video generation has started to gain prominence. While a few preliminary works have explored audio-driven multi-person talking video generation, they often face challenges due to the high costs of diverse multi-person data collection and the difficulty of driving multiple identities with coherent interactivity. To address these challenges, we propose AnyTalker, a multi-person generation framework that features an extensible multi-stream processing architecture. Specifically, we extend Diffusion Transformer’s attention block with a novel identity-aware attention mechanism that iteratively processes identity-audio pairs, allowing arbitrary scaling of drivable identities. Besides, training multi-person generative models demands massive multi-person data. Our proposed training pipeline depends solely on single-person videos to learn multi-person speaking patterns and refines interactivity with only a few real multi-person clips. Furthermore, we contribute a targeted metric and dataset designed to evaluate the naturalness and interactivity of the generated multi-person videos. Extensive experiments demonstrate that AnyTalker achieves remarkable lip synchronization, visual quality, and natural interactivity, striking a favorable balance between data costs and identity scalability.

Method: 1) Interactive Video Reasoning paradigm where video becomes an active cognitive workspace; 2) Video CoM model using Chain of Manipulations (CoM) for iterative visual actions; 3) Video CoM Instruct dataset (18K samples) for multi-step manipulation reasoning; 4) Reinforcement learning with Group Relative Policy Optimization (GRPO) using step-level reasoning rewards instead of just sparse answer rewards.

Result: Achieves strong results across nine video reasoning benchmarks, improving average performance by 3.6% over recent SOTA models, while training on only 25K SFT and 3K GRPO video samples (significantly fewer than comparable large-scale models). Reasoning-aware rewards improve both accuracy and interpretability.

Conclusion: Interactive Video Reasoning with Chain of Manipulations enables models to “think with videos” rather than just “think about videos,” overcoming semantic bottlenecks and achieving better performance with less data through active visual reasoning and step-level optimization.

Abstract: Recent multimodal large language models (MLLMs) have advanced video understanding, yet most still “think about videos” ie once a video is encoded, reasoning unfolds entirely in text, treating visual input as a static context. This passive paradigm creates a semantic bottleneck: models cannot rewatch, refocus, or verify evidence, leading to shallow visual reasoning on tasks requiring fine grained spatio temporal understanding. In this work, we introduce Interactive Video Reasoning, a new paradigm that transforms video into an active cognitive workspace, enabling models to “think with videos”. Our model, Video CoM, reasons through a Chain of Manipulations (CoM), performing iterative visual actions to gather and refine evidence. To support this behavior, we construct Video CoM Instruct, an 18K instruction tuning dataset curated for multi step manipulation reasoning. Beyond supervised learning, we further optimize the manipulation policy via reinforcement learning with reasoning aware Group Relative Policy Optimization (GRPO). Unlike prior work that relies solely on sparse answer rewards, our method introduces step level reasoning rewards, guiding the model toward grounded and consistent reasoning. Video CoM achieves strong results across nine video reasoning benchmarks, improving average performance by 3.6 percent over recent state of the art models, while training on only 25K SFT and 3K GRPO video samples, significantly fewer than comparable large scale models. Ablation studies demonstrate that reasoning aware rewards improve both accuracy and interpretability. Code: https://github.com/mbzuai-oryx/Video-CoM

Method: Maintains explicit spatial pointer memory directly associated with 3D scene structure, where each pointer has a specific 3D position and aggregates nearby scene information. Uses 3D hierarchical position embedding to promote interaction between latest frame information and pointer memory, with a simple fusion mechanism for uniform and efficient memory.

Result: Achieves competitive or state-of-the-art performance on various tasks with low training costs.

Conclusion: Point3R provides an effective online framework for dense streaming 3D reconstruction by replacing implicit memory with explicit spatial pointer memory, enabling better information retention and integration of observations into global coordinate systems.

Abstract: Dense 3D scene reconstruction from an ordered sequence or unordered image collections is a critical step when bringing research in computer vision into practical scenarios. Following the paradigm introduced by DUSt3R, which unifies an image pair densely into a shared coordinate system, subsequent methods maintain an implicit memory to achieve dense 3D reconstruction from more images. However, such implicit memory is limited in capacity and may suffer from information loss of earlier frames. We propose Point3R, an online framework targeting dense streaming 3D reconstruction. To be specific, we maintain an explicit spatial pointer memory directly associated with the 3D structure of the current scene. Each pointer in this memory is assigned a specific 3D position and aggregates scene information nearby in the global coordinate system into a changing spatial feature. Information extracted from the latest frame interacts explicitly with this pointer memory, enabling dense integration of the current observation into the global coordinate system. We design a 3D hierarchical position embedding to promote this interaction and design a simple yet effective fusion mechanism to ensure that our pointer memory is uniform and efficient. Our method achieves competitive or state-of-the-art performance on various tasks with low training costs. Code: https://github.com/YkiWu/Point3R.

Method: Proposes reinforcement learning approach combining timestamp-aware supervised fine-tuning with Group Relative Policy Optimization (GRPO) guided by a novel Temporal Alignment Reward (TAR) to enhance temporal precision and reasoning consistency.

Result: Video R2 achieves higher Think Answer Consistency (TAC), Video Attention Score (VAS), and accuracy across 11 video reasoning benchmarks, demonstrating improved temporal alignment and reasoning coherence.

Conclusion: Improvements in temporal alignment and reasoning coherence lead to more accurate and trustworthy video understanding, with the model, code, and dataset being open-sourced.

Abstract: Reasoning over dynamic visual content remains a central challenge for multimodal large language models. Recent thinking models generate explicit reasoning traces for interpretability; however, their reasoning often appears convincing while being logically inconsistent or weakly grounded in visual evidence. We identify and formalize these issues through two diagnostic metrics: Think Answer Consistency (TAC), which measures the alignment between reasoning and answers, and Video Attention Score (VAS), which captures the extent to which reasoning depends on visual versus textual cues. Analysis across 11 video reasoning benchmarks shows that current models rely heavily on linguistic priors rather than visual content. To address this, we propose a reinforcement learning approach that enhances both temporal precision and reasoning consistency. Our approach combines timestamp aware supervised fine tuning with Group Relative Policy Optimization (GRPO) guided by a novel Temporal Alignment Reward (TAR). This dual step post training stage encourages temporally aligned and causally coherent video reasoning. The resulting model, Video R2, achieves consistently higher TAC, VAS, and accuracy across multiple benchmarks, demonstrating that improvements in temporal alignment and reasoning coherence lead to more accurate and trustworthy video understanding. Our code, dataset, and model will be open sourced.

Method: Created an evaluation suite with three distinct forms of nonlocal vision: comparative perception (holding two images in working memory and comparing), saccadic search (evidence-driven jumps to locate successive targets), and smooth visual search (following continuous contours). Tested flagship models like GPT-5, Gemini 2.5 Pro, Claude Sonnet 4.

Result: Flagship models fail these tests and barely exceed random accuracy on two task variants that are trivial for humans, despite performing well on prior primitive-vision benchmarks.

Conclusion: Despite gains in raw visual acuity, current VLMs lack core visual reasoning capabilities needed for nonlocal visual reasoning, showing they cannot perform visual algorithms similar to those used by humans.

Abstract: Vision-Language Models (VLMs) excel at complex visual tasks such as VQA and chart understanding, yet recent work suggests they struggle with simple perceptual tests. We present an evaluation of vision-language models’ capacity for nonlocal visual reasoning: reasoning that requires chaining evidence collected from multiple, possibly distant regions of an image. We isolate three distinct forms of nonlocal vision: comparative perception, which demands holding two images in working memory and comparing them; saccadic search, which requires making discrete, evidence-driven jumps to locate successive targets; and smooth visual search, which involves following a continuous contour. Flagship models (e.g., GPT-5, Gemini 2.5 Pro, Claude Sonnet 4), even those that perform well on prior primitive-vision benchmarks, fail these tests and barely exceed random accuracy on two variants of our tasks that are trivial for humans. Our structured evaluation suite allows us to test whether VLMs can perform visual algorithms similar to those used by humans. Our findings show that despite gains in raw visual acuity, current models lack core visual reasoning capabilities.

Method: CADAR integrates neural and symbolic reasoning: 1) Fuses multimodal vision-language representations into a perception graph capturing objects, relations, and temporal contextual salience, 2) Uses particle-filter-based statistical reasoning to infer anomalies in semantic dynamics.

Result: Preliminary experiments on an AR cognitive-attack dataset demonstrate consistent advantages over existing approaches.

Conclusion: CADAR shows the potential of neuro-symbolic methods for robust and interpretable AR security by combining adaptability of vision-language models with interpretability of probabilistic symbolic reasoning.

Abstract: Augmented Reality (AR) enriches human perception by overlaying virtual elements onto the physical world. However, this tight coupling between virtual and real content makes AR vulnerable to cognitive attacks: manipulations that distort users’ semantic understanding of the environment. Existing detection methods largely focus on visual inconsistencies at the pixel or image level, offering limited semantic reasoning or interpretability. To address these limitations, we introduce CADAR, a neuro-symbolic framework for cognitive attack detection in AR that integrates neural and symbolic reasoning. CADAR fuses multimodal vision-language representations from pre-trained models into a perception graph that captures objects, relations, and temporal contextual salience. Building on this structure, a particle-filter-based statistical reasoning module infers anomalies in semantic dynamics to reveal cognitive attacks. This combination provides both the adaptability of modern vision-language models and the interpretability of probabilistic symbolic reasoning. Preliminary experiments on an AR cognitive-attack dataset demonstrate consistent advantages over existing approaches, highlighting the potential of neuro-symbolic methods for robust and interpretable AR security.

Method: Uses dual-model weight selection to initialize two lightweight models with weights from a large pretrained model, then applies self-knowledge distillation to leverage diverse initial weight configurations without excessive computational cost, followed by fine-tuning for target tasks.

Result: Extensive experiments on chest X-ray, lung CT scans, and brain MRI datasets demonstrate superior performance and robustness compared to existing methods, overcoming limitations of conventional approaches that fail to retain critical information in compact models.

Conclusion: The integration of dual-model weight selection with self-knowledge distillation effectively addresses computational constraints in medical image classification while maintaining high performance, offering a practical solution for resource-limited medical settings.

Abstract: We propose a novel medical image classification method that integrates dual-model weight selection with self-knowledge distillation (SKD). In real-world medical settings, deploying large-scale models is often limited by computational resource constraints, which pose significant challenges for their practical implementation. Thus, developing lightweight models that achieve comparable performance to large-scale models while maintaining computational efficiency is crucial. To address this, we employ a dual-model weight selection strategy that initializes two lightweight models with weights derived from a large pretrained model, enabling effective knowledge transfer. Next, SKD is applied to these selected models, allowing the use of a broad range of initial weight configurations without imposing additional excessive computational cost, followed by fine-tuning for the target classification tasks. By combining dual-model weight selection with self-knowledge distillation, our method overcomes the limitations of conventional approaches, which often fail to retain critical information in compact models. Extensive experiments on publicly available datasets-chest X-ray images, lung computed tomography scans, and brain magnetic resonance imaging scans-demonstrate the superior performance and robustness of our approach compared to existing methods.

Method: Uses source pre-trained model to generate noisy pseudo-labels for target domain, treats adaptation as learning from noisy labels, filters correct pseudo-labels using cross-entropy loss as indicator (small-loss samples), and enhances with teacher-student framework where teacher produces reliable pseudo-labels and student fine-tunes on target videos.

Result: Achieves state-of-the-art results on various open datasets, outperforming existing approaches. Two versions: CleanAdapt and CleanAdapt + TS.

Conclusion: Proposes effective source-free video domain adaptation approach that bridges source-target domain gap without accessing source data, demonstrating superior performance through noisy label filtering and teacher-student framework.

Abstract: Despite the progress seen in classification methods, current approaches for handling videos with distribution shifts in source and target domains remain source-dependent as they require access to the source data during the adaptation stage. In this paper, we present a self-training based source-free video domain adaptation approach to address this challenge by bridging the gap between the source and the target domains. We use the source pre-trained model to generate pseudo-labels for the target domain samples, which are inevitably noisy. Thus, we treat the problem of source-free video domain adaptation as learning from noisy labels and argue that the samples with correct pseudo-labels can help us in adaptation. To this end, we leverage the cross-entropy loss as an indicator of the correctness of the pseudo-labels and use the resulting small-loss samples from the target domain for fine-tuning the model. We further enhance the adaptation performance by implementing a teacher-student (TS) framework, in which the teacher, which is updated gradually, produces reliable pseudo-labels. Meanwhile, the student undergoes fine-tuning on the target domain videos using these generated pseudo-labels to improve its performance. Extensive experimental evaluations show that our methods, termed as CleanAdapt, CleanAdapt + TS, achieve state-of-the-art results, outperforming the existing approaches on various open datasets. Our source code is publicly available at https://avijit9.github.io/CleanAdapt.

Method: Introduces Spacewalk-18 benchmark with two tasks: (1) step recognition (temporal segmentation) and (2) video question answering, using temporally segmented and labeled spacewalk recordings from the ISS.

Result: The benchmark reveals significant challenges but suggests best practices for domain generalization and long-form understanding. A promising adaptation via summarization technique shows significant performance improvement without model fine-tuning.

Conclusion: Spacewalk-18 provides a valuable benchmark for testing video understanding models’ ability to generalize to novel domains and utilize long temporal context with multimodal information, with summarization emerging as an effective adaptation technique.

Abstract: Learning from (procedural) videos has increasingly served as a pathway for embodied agents to acquire skills from human demonstrations. To do this, video understanding models must be able to obtain structured understandings, such as the temporal segmentation of a demonstration into sequences of actions and skills, and to generalize the understandings to novel environments, tasks, and problem domains. In pursuit of this goal, we introduce Spacewalk-18, a benchmark containing two tasks: (1) step recognition and (2) video question answering, over a dataset of temporally segmented and labeled tasks in International Space Station spacewalk recordings. In tandem, the two tasks quantify a model’s ability to: (1) generalize to novel domains; (2) utilize long temporal context and multimodal (e.g. visual and speech) information. Our extensive experimental analysis highlights the challenges of Spacewalk-18, but also suggests best practices for domain generalization and long-form understanding. Notably, we discover a promising adaptation via summarization technique that leads to significant performance improvement without model fine-tuning. The Spacewalk-18 benchmark is released at https://brown-palm.github.io/Spacewalk-18/.

Method: UNO uses extended slot attention to decompose visual features into object and relation slots, object temporal consistency learning for cross-frame consistency without tracking, and dynamic triplet prediction to link relation slots to object pairs over time.

Result: UNO achieves competitive performance on both box-level and pixel-level VidSGG benchmarks while offering improved efficiency through its unified, object-centric design.

Conclusion: UNO demonstrates that a single-stage, unified framework can effectively address both coarse-grained and fine-grained VidSGG tasks with minimal task-specific modifications and maximum parameter sharing.

Abstract: Video Scene Graph Generation (VidSGG) aims to represent dynamic visual content by detecting objects and modeling their temporal interactions as structured graphs. Prior studies typically target either coarse-grained box-level or fine-grained panoptic pixel-level VidSGG, often requiring task-specific architectures and multi-stage training pipelines. In this paper, we present UNO (UNified Object-centric VidSGG), a single-stage, unified framework that jointly addresses both tasks within an end-to-end architecture. UNO is designed to minimize task-specific modifications and maximize parameter sharing, enabling generalization across different levels of visual granularity. The core of UNO is an extended slot attention mechanism that decomposes visual features into object and relation slots. To ensure robust temporal modeling, we introduce object temporal consistency learning, which enforces consistent object representations across frames without relying on explicit tracking modules. Additionally, a dynamic triplet prediction module links relation slots to corresponding object pairs, capturing evolving interactions over time. We evaluate UNO on standard box-level and pixel-level VidSGG benchmarks. Results demonstrate that UNO not only achieves competitive performance across both tasks but also offers improved efficiency through a unified, object-centric design.

Method: 1) Define comprehensive language-expressed fusion objective; 2) Encode relevant texts into multi-modal embedding space using CLIP; 3) Construct language-driven fusion model in embedding space by establishing relationships among embedded vectors representing fusion objective and input modalities; 4) Derive language-driven loss to align actual IVIF with embedded language-driven model via supervised training.

Result: Experiments show the method obtains much better fusion results than existing techniques.

Conclusion: Using natural language to express IVIF objectives avoids limitations of mathematical loss functions and improves fusion performance by leveraging language’s expressive power, with code publicly available.

Abstract: Infrared-visible image fusion (IVIF) has attracted much attention owing to the highly-complementary properties of the two image modalities. Due to the lack of ground-truth fused images, the fusion output of current deep-learning based methods heavily depends on the loss functions defined mathematically. As it is hard to well mathematically define the fused image without ground truth, the performance of existing fusion methods is limited. In this paper, we propose to use natural language to express the objective of IVIF, which can avoid the explicit mathematical modeling of fusion output in current losses, and make full use of the advantage of language expression to improve the fusion performance. For this purpose, we present a comprehensive language-expressed fusion objective, and encode relevant texts into the multi-modal embedding space using CLIP. A language-driven fusion model is then constructed in the embedding space, by establishing the relationship among the embedded vectors representing the fusion objective and input image modalities. Finally, a language-driven loss is derived to make the actual IVIF aligned with the embedded language-driven fusion model via supervised training. Experiments show that our method can obtain much better fusion results than existing techniques. The code is available at https://github.com/wyhlaowang/LDFusion.

Method: Proposes a novel paradigm that directly optimizes parity-based fairness metrics through specifically designed training objectives without requiring group labels. Uses vision-language models to analyze sensitive attribute relevancy for individual samples, then formulates loss functions that mathematically connect to each target fairness metric.

Result: Experiments on multiple image classification datasets show that the metric-specific approach significantly improves parity-based fairness criteria and outperforms existing methods.

Conclusion: The proposed method enables flexible optimization of different fairness criteria based on application needs and provides a direct, mathematically grounded approach to achieving parity-based fairness without requiring expensive group labels.

Abstract: Performance disparities of image recognition across demographic groups are known to exist in deep learning-based models, due to imbalanced group representations or spurious correlation between group and target labels. Previous work has addressed such challenges without relying on expensive group labels, typically by upweighting high-loss samples or balancing discovered clusters. However, these heuristic strategies lack direct connection to specific fairness metrics and cannot guarantee optimization of parity-based criteria like equal opportunity, which ensures equal chance to receive positive outcomes across groups. In this work, we propose a novel paradigm that directly optimizes parity-based fairness metrics through specifically designed training objectives, without requiring group labels. We leverage vision-language models to analyze sensitive attribute relevancy for individual samples, then formulate loss functions that mathematically connect to each target fairness metric. This enables flexible optimization of different fairness criteria based on application needs. Experiments on multiple image classification datasets show that our metric-specific approach significantly improves parity-based fairness criteria and outperforms existing methods.

Method: The paper conducts a comprehensive survey by: 1) Introducing general PCS frameworks categorized into test-time fine-tuning (TTF) and pre-trained adaptation (PTA) approaches, 2) Analyzing strengths, limitations, and key techniques of each methodology, 3) Exploring specialized tasks (object, face, style personalization), 4) Discussing ongoing challenges and future directions.

Result: The survey provides a systematic organization of the PCS field, identifies two main methodological categories (TTF and PTA), analyzes their technical characteristics, highlights unique challenges in specialized tasks, and identifies key research problems including overfitting and fidelity-alignment trade-offs.

Conclusion: PCS has shown promising progress but faces significant challenges. The survey establishes a foundation for future research by providing a comprehensive overview, identifying current limitations, and proposing directions to advance the field, particularly in addressing overfitting and improving the balance between subject fidelity and text alignment.

Abstract: Recent advancements in diffusion models have significantly impacted content creation, leading to the emergence of Personalized Content Synthesis (PCS). By utilizing a small set of user-provided examples featuring the same subject, PCS aims to tailor this subject to specific user-defined prompts. Over the past two years, more than 150 methods have been introduced in this area. However, existing surveys primarily focus on text-to-image generation, with few providing up-to-date summaries on PCS. This paper provides a comprehensive survey of PCS, introducing the general frameworks of PCS research, which can be categorized into test-time fine-tuning (TTF) and pre-trained adaptation (PTA) approaches. We analyze the strengths, limitations, and key techniques of these methodologies. Additionally, we explore specialized tasks within the field, such as object, face, and style personalization, while highlighting their unique challenges and innovations. Despite the promising progress, we also discuss ongoing challenges, including overfitting and the trade-off between subject fidelity and text alignment. Through this detailed overview and analysis, we propose future directions to further the development of PCS.

Method: 1) Multi-functional IQA task paradigm covering assessment/comparison tasks, brief/detailed responses, and full-reference/non-reference scenarios. 2) Ground-truth-informed dataset construction approach. 3) Scaling dataset to 495K under brief-detail joint framework (DQ-495K). 4) Retaining image resolution during training for resolution-related quality issues. 5) Confidence score estimation to filter low-quality responses.

Result: DepictQA-Wild significantly outperforms traditional score-based methods, prior VLM-based IQA models, and proprietary GPT-4V in distortion identification, instant rating, and reasoning tasks. Advantages confirmed in real-world applications including web-downloaded image assessment and model-processed image ranking.

Conclusion: The proposed DepictQA-Wild model with its comprehensive multi-task paradigm and large-scale high-quality dataset (DQ-495K) represents a significant advancement in practical VLM-based image quality assessment, addressing limitations of previous approaches and demonstrating superior performance across multiple evaluation dimensions.

Abstract: With the rapid advancement of Vision Language Models (VLMs), VLM-based Image Quality Assessment (IQA) seeks to describe image quality linguistically to align with human expression and capture the multifaceted nature of IQA tasks. However, current methods are still far from practical usage. First, prior works focus narrowly on specific sub-tasks or settings, which do not align with diverse real-world applications. Second, their performance is sub-optimal due to limitations in dataset coverage, scale, and quality. To overcome these challenges, we introduce the enhanced Depicted image Quality Assessment model (DepictQA-Wild). Our method includes a multi-functional IQA task paradigm that encompasses both assessment and comparison tasks, brief and detailed responses, full-reference and non-reference scenarios. We introduce a ground-truth-informed dataset construction approach to enhance data quality, and scale up the dataset to 495K under the brief-detail joint framework. Consequently, we construct a comprehensive, large-scale, and high-quality dataset, named DQ-495K. We also retain image resolution during training to better handle resolution-related quality issues, and estimate a confidence score that is helpful to filter out low-quality responses. Experimental results demonstrate that DepictQA-Wild significantly outperforms traditional score-based methods, prior VLM-based IQA models, and proprietary GPT-4V in distortion identification, instant rating, and reasoning tasks. Our advantages are further confirmed by real-world applications including assessing the web-downloaded images and ranking model-processed images. Codes, datasets, and model weights have been released in https://depictqa.github.io/.

Method: SAEmnesia uses a supervised sparse autoencoder framework that systematically labels concepts during training to enforce one-to-one concept-neuron mappings. This achieves feature centralization, binding each concept to a single, interpretable neuron for targeted concept erasure.

Result: The method reduces hyperparameter search by 96.7%, achieves 9.2% improvement over state-of-the-art on UnlearnCanvas benchmark, shows 28.4% accuracy improvement when removing nine objects in sequential unlearning, and effectively removes unwanted content including nudity when evaluated with I2P.

Conclusion: SAEmnesia establishes a new standard for precise and controllable concept erasure in diffusion models by overcoming feature splitting through supervised sparse autoencoding, enabling efficient, targeted unlearning with improved scalability and robustness against adversarial attacks.

Abstract: Concept unlearning in diffusion models is hampered by feature splitting, where concepts are distributed across many latent features, making their removal challenging and computationally expensive. We introduce SAEmnesia, a supervised sparse autoencoder framework that overcomes this by enforcing one-to-one concept-neuron mappings. By systematically labeling concepts during training, our method achieves feature centralization, binding each concept to a single, interpretable neuron. This enables highly targeted and efficient concept erasure. SAEmnesia reduces hyperparameter search by 96.7% and achieves a 9.2% improvement over the state-of-the-art on the UnlearnCanvas benchmark. Our method also demonstrates superior scalability in sequential unlearning, improving accuracy by 28.4% when removing nine objects, establishing a new standard for precise and controllable concept erasure. Moreover, SAEmnesia mitigates the possibility of generating unwanted content under adversarial attack and effectively removes nudity when evaluated with I2P.

Method: Proposes using an auxiliary octahedral field alongside the implicit geometry. Both smoothness and sharp features in the distance field can be equivalently described by smoothness in octahedral space. By aligning and smoothing the octahedral field, the method behaves like bilateral filtering.

Result: Outperforms various traditional and neural implicit fitting approaches across extensive experiments, and is very competitive with methods that require normals and data priors.

Conclusion: The octahedral field approach enables effective surface reconstruction from noisy point clouds while preserving sharp edges, despite operating purely pointwise, demonstrating significant improvement over existing methods.

Abstract: Neural implicit representation, the parameterization of a continuous distance function as a Multi-Layer Perceptron (MLP), has emerged as a promising lead in tackling surface reconstruction from unoriented point clouds. In the presence of noise, however, its lack of explicit neighborhood connectivity makes sharp edges identification particularly challenging, hence preventing the separation of smoothing and sharpening operations, as is achievable with its discrete counterparts. In this work, we propose to tackle this challenge with an auxiliary field, the \emph{octahedral field}. We observe that both smoothness and sharp features in the distance field can be equivalently described by the smoothness in octahedral space. Therefore, by aligning and smoothing an octahedral field alongside the implicit geometry, our method behaves analogously to bilateral filtering, resulting in a smooth reconstruction while preserving sharp edges. Despite being operated purely pointwise, our method outperforms various traditional and neural implicit fitting approaches across extensive experiments, and is very competitive with methods that require normals and data priors. Code and data of our work are available at: https://github.com/Ankbzpx/frame-field.

Method: PoseAdapt provides a framework with domain-incremental and class-incremental tracks simulating realistic changes (density, lighting, modality, skeleton growth). It supports two workflows: Strategy Benchmarking for researchers to test CL methods, and Model Adaptation for practitioners to adapt pretrained models with minimal supervision.

Result: The framework evaluates representative regularization-based methods in single-step and sequential settings under strict constraints (fixed lightweight backbone, no past data access, tight per-step budgets), highlighting the difficulty of maintaining accuracy under resource limits.

Conclusion: PoseAdapt bridges modern continual learning techniques with practical pose estimation needs, enabling adaptable models that improve over time without repeated full retraining, addressing real-world deployment constraints.

Abstract: Human pose estimators are typically retrained from scratch or naively fine-tuned whenever keypoint sets, sensing modalities, or deployment domains change–an inefficient, compute-intensive practice that rarely matches field constraints. We present PoseAdapt, an open-source framework and benchmark suite for continual pose model adaptation. PoseAdapt defines domain-incremental and class-incremental tracks that simulate realistic changes in density, lighting, and sensing modality, as well as skeleton growth. The toolkit supports two workflows: (i) Strategy Benchmarking, which lets researchers implement continual learning (CL) methods as plugins and evaluate them under standardized protocols; and (ii) Model Adaptation, which allows practitioners to adapt strong pretrained models to new tasks with minimal supervision. We evaluate representative regularization-based methods in single-step and sequential settings. Benchmarks enforce a fixed lightweight backbone, no access to past data, and tight per-step budgets. This isolates adaptation strategy effects, highlighting the difficulty of maintaining accuracy under strict resource limits. PoseAdapt connects modern CL techniques with practical pose estimation needs, enabling adaptable models that improve over time without repeated full retraining.

Method: Given a 3D object, GraspDiffusion constructs whole-body poses by separately leveraging generative priors for body and hand poses, then optimizing them into a joint grasping pose that guides image synthesis.

Result: GraspDiffusion successfully tackles the relatively uninvestigated problem of generating full-bodied human-object interactions while outperforming previous methods.

Conclusion: The proposed method creates realistic and diverse human-object interaction scenes by using joint grasping poses to guide image synthesis, addressing a significant gap in current generative modeling capabilities.

Abstract: Recent generative models can synthesize high-quality images, but they often fail to generate humans interacting with objects using their hands. This arises mostly from the model’s misunderstanding of such interactions and the hardships of synthesizing intricate regions of the body. In this paper, we propose \textbf{GraspDiffusion}, a novel generative method that creates realistic scenes of human-object interaction. Given a 3D object, GraspDiffusion constructs whole-body poses with control over the object’s location relative to the human body, which is achieved by separately leveraging the generative priors for body and hand poses, optimizing them into a joint grasping pose. This pose guides the image synthesis to correctly reflect the intended interaction, creating realistic and diverse human-object interaction scenes. We demonstrate that GraspDiffusion can successfully tackle the relatively uninvestigated problem of generating full-bodied human-object interactions while outperforming previous methods. Our project page is available at https://yj7082126.github.io/graspdiffusion/

Method: Uses flow derivatives to mathematically disentangle camera egomotion from articulated movements, connects 2D flows to 3D Gaussian dynamic flow for optimization, and introduces a 3D spherical vector controlling scheme that represents state as 3D Gaussian trajectory instead of complex 1D control signals.

Result: Extensive experiments on articulated objects demonstrate state-of-the-art visual performance and precise, part-aware controllability without requiring any control signals or manual annotations.

Conclusion: FreeGaussian provides an annotation-free approach for controllable Gaussian splat reconstruction that eliminates the need for complex control signal calculations while achieving superior visual quality and precise part-aware control for articulated objects.

Abstract: Reconstructing controllable Gaussian splats for articulated objects from monocular video is especially challenging due to its inherently insufficient constraints. Existing methods address this by relying on dense masks and manually defined control signals, limiting their real-world applications. In this paper, we propose an annotation-free method, FreeGaussian, which mathematically disentangles camera egomotion and articulated movements via flow derivatives. By establishing a connection between 2D flows and 3D Gaussian dynamic flow, our method enables optimization and continuity of dynamic Gaussian motions from flow priors without any control signals. Furthermore, we introduce a 3D spherical vector controlling scheme, which represents the state as a 3D Gaussian trajectory, thereby eliminating the need for complex 1D control signal calculations and simplifying controllable Gaussian modeling. Extensive experiments on articulated objects demonstrate the state-of-the-art visual performance and precise, part-aware controllability of our method. Code is available at: https://github.com/Tavish9/freegaussian.

Method: Proposes MambaX-Net with: 1) Mamba-enhanced Cross-Attention Module for temporal evolution and spatial dependencies, 2) Shape Extractor Module for anatomical representation from previous masks, and 3) semi-supervised self-training using pseudo-labels from pre-trained nnU-Net.

Result: MambaX-Net significantly outperforms state-of-the-art U-Net and Transformer-based models on longitudinal AS dataset, achieving superior prostate zone segmentation even with limited/noisy training data.

Conclusion: MambaX-Net effectively addresses challenges in longitudinal prostate segmentation for Active Surveillance by combining temporal modeling, anatomical shape encoding, and semi-supervised learning, enabling automated PCa monitoring with minimal expert annotations.

Abstract: Active Surveillance (AS) is a treatment option for managing low and intermediate-risk prostate cancer (PCa), aiming to avoid overtreatment while monitoring disease progression through serial MRI and clinical follow-up. Accurate prostate segmentation is an important preliminary step for automating this process, enabling automated detection and diagnosis of PCa. However, existing deep-learning segmentation models are often trained on single-time-point and expertly annotated datasets, making them unsuitable for longitudinal AS analysis, where multiple time points and a scarcity of expert labels hinder their effective fine-tuning. To address these challenges, we propose MambaX-Net, a novel semi-supervised, dual-scan 3D segmentation architecture that computes the segmentation for time point t by leveraging the MRI and the corresponding segmentation mask from the previous time point. We introduce two new components: (i) a Mamba-enhanced Cross-Attention Module, which integrates the Mamba block into cross attention to efficiently capture temporal evolution and long-range spatial dependencies, and (ii) a Shape Extractor Module that encodes the previous segmentation mask into a latent anatomical representation for refined zone delination. Moreover, we introduce a semi-supervised self-training strategy that leverages pseudo-labels generated from a pre-trained nnU-Net, enabling effective learning without expert annotations. MambaX-Net was evaluated on a longitudinal AS dataset, and results showed that it significantly outperforms state-of-the-art U-Net and Transformer-based models, achieving superior prostate zone segmentation even when trained on limited and noisy data.

Method: VWT groups frequently used visual subwords (image patches) into visual words while keeping infrequent ones intact. It leverages intra-image or inter-image statistics to identify similar visual concepts for sequence compression. The method is training-free and doesn’t require fine-tuning.

Result: Achieves up to 47% reduction in energy consumption while retaining performance. In comparison, 8-bit quantization and token merging approaches can increase energy costs by up to 500% or more. VWT shows marginal performance compromise while being well-suited for efficient online inference.

Conclusion: Visual-Word Tokenizer provides an effective training-free solution for reducing vision transformer energy costs, making it suitable for real-time online inference applications with minimal performance trade-offs.

Abstract: The cost of deploying vision transformers increasingly represents a barrier to wider industrial adoption. Existing compression techniques require additional end-to-end fine-tuning or incur a significant drawback to energy efficiency, making them ill-suited for online (real-time) inference, where a prediction is made on any new input as it comes in. We introduce the $\textbf{Visual-Word Tokenizer}$ (VWT), a training-free method for reducing energy costs while retaining performance. The VWT groups visual subwords (image patches) that are frequently used into visual words, while infrequent ones remain intact. To do so, $\textit{intra}$-image or $\textit{inter}$-image statistics are leveraged to identify similar visual concepts for sequence compression. Experimentally, we demonstrate a reduction in energy consumed of up to 47%. Comparative approaches of 8-bit quantization and token merging can lead to significantly increased energy costs (up to 500% or more). Our results indicate that VWTs are well-suited for efficient online inference with a marginal compromise on performance. The experimental code for our paper is also made publicly available.

Method: Introduces NNLFs as passive loss functions within APL framework, creating ANL. NNLFs focus more on memorized clean samples than MAE. Also proposes entropy-based regularization for non-symmetric noise scenarios to handle label imbalance.

Result: Extensive experiments show ANL with NNLFs achieves better or comparable performance to state-of-the-art methods across various label noise types and in image segmentation tasks.

Conclusion: The proposed ANL framework with NNLFs effectively addresses MAE’s limitations in noisy label scenarios, offering improved robustness and performance across different noise conditions and tasks.

Abstract: Deep supervised learning has achieved remarkable success across a wide range of tasks, yet it remains susceptible to overfitting when confronted with noisy labels. To address this issue, noise-robust loss functions offer an effective solution for enhancing learning in the presence of label noise. In this work, we systematically investigate the limitation of the recently proposed Active Passive Loss (APL), which employs Mean Absolute Error (MAE) as its passive loss function. Despite the robustness brought by MAE, one of its key drawbacks is that it pays equal attention to clean and noisy samples; this feature slows down convergence and potentially makes training difficult, particularly in large-scale datasets. To overcome these challenges, we introduce a novel loss function class, termed Normalized Negative Loss Functions (NNLFs), which serve as passive loss functions within the APL framework. NNLFs effectively address the limitations of MAE by concentrating more on memorized clean samples. By replacing MAE in APL with our proposed NNLFs, we enhance APL and present a new framework called Active Negative Loss (ANL). Moreover, in non-symmetric noise scenarios, we propose an entropy-based regularization technique to mitigate the vulnerability to the label imbalance. Extensive experiments demonstrate that the new loss functions adopted by our ANL framework can achieve better or comparable performance to state-of-the-art methods across various label noise types and in image segmentation tasks. The source code is available at: https://github.com/Virusdoll/Active-Negative-Loss.

Method: The framework operates in the semantic feature space of pretrained Vision Foundation Models (VFMs), training a masked feature transformer in self-supervised manner to predict the evolution of VFM features over time. VFM features serve as a latent space where different task-specific heads can be attached for various scene understanding tasks.

Result: Extensive experiments demonstrate very strong performance, robustness, and scalability of the framework across various scene understanding tasks.

Conclusion: DINO-Foresight provides an efficient and scalable approach to future dynamics prediction by leveraging semantic feature forecasting from Vision Foundation Models, enabling practical applications in autonomous systems and robotics.

Abstract: Predicting future dynamics is crucial for applications like autonomous driving and robotics, where understanding the environment is key. Existing pixel-level methods are computationally expensive and often focus on irrelevant details. To address these challenges, we introduce DINO-Foresight, a novel framework that operates in the semantic feature space of pretrained Vision Foundation Models (VFMs). Our approach trains a masked feature transformer in a self-supervised manner to predict the evolution of VFM features over time. By forecasting these features, we can apply off-the-shelf, task-specific heads for various scene understanding tasks. In this framework, VFM features are treated as a latent space, to which different heads attach to perform specific tasks for future-frame analysis. Extensive experiments show the very strong performance, robustness and scalability of our framework. Project page and code at https://dino-foresight.github.io/ .

Method: Proposes a rescaling method that uses sparse 3D points from sensors (low-resolution LiDAR) or techniques (structure-from-motion with IMU poses) to convert Depth Anything’s affine-invariant disparity maps to metric depth without fine-tuning the model.

Result: The method outperforms zero-shot monocular metric depth estimation approaches, achieves competitive results compared to fine-tuned methods, and shows better robustness than depth completion approaches. It’s robust to noise in sparse depth, camera-LiDAR calibration, and depth model predictions.

Conclusion: The proposed rescaling approach provides an effective alternative to fine-tuning for metric depth estimation, preserving the generalization power of foundation models while being practical and robust to various noise sources.

Abstract: The recent development of \emph{foundation models} for monocular depth estimation such as Depth Anything paved the way to zero-shot monocular depth estimation. Since it returns an affine-invariant disparity map, the favored technique to recover the metric depth consists in fine-tuning the model. However, this stage is not straightforward, it can be costly and time-consuming because of the training and the creation of the dataset. The latter must contain images captured by the camera that will be used at test time and the corresponding ground truth. Moreover, the fine-tuning may also degrade the generalizing capacity of the original model. Instead, we propose in this paper a new method to rescale Depth Anything predictions using 3D points provided by sensors or techniques such as low-resolution LiDAR or structure-from-motion with poses given by an IMU. This approach avoids fine-tuning and preserves the generalizing power of the original depth estimation model while being robust to the noise of the sparse depth, of the camera-LiDAR calibration or of the depth model. Our experiments highlight enhancements relative to zero-shot monocular metric depth estimation methods, competitive results compared to fine-tuned approaches and a better robustness than depth completion approaches. Code available at github.com/ENSTA-U2IS-AI/depth-rescaling.

Method: Uses unified visual sequence transformer with multimodal masked visual modeling objective and novel masking mechanism for multimodal training. Proposes VAE-free hierarchical tokenization to reduce computational complexity and enable end-to-end training with high-resolution multimodal inputs.

Result: Validated on Cityscapes dataset, demonstrating state-of-the-art performance in future semantic segmentation for both short- and mid-term forecasting.

Conclusion: FUTURIST effectively integrates multimodal information for accurate future semantic prediction with improved efficiency and end-to-end training capability.

Abstract: Semantic future prediction is important for autonomous systems navigating dynamic environments. This paper introduces FUTURIST, a method for multimodal future semantic prediction that uses a unified and efficient visual sequence transformer architecture. Our approach incorporates a multimodal masked visual modeling objective and a novel masking mechanism designed for multimodal training. This allows the model to effectively integrate visible information from various modalities, improving prediction accuracy. Additionally, we propose a VAE-free hierarchical tokenization process, which reduces computational complexity, streamlines the training pipeline, and enables end-to-end training with high-resolution, multimodal inputs. We validate FUTURIST on the Cityscapes dataset, demonstrating state-of-the-art performance in future semantic segmentation for both short- and mid-term forecasting. Project page and code at https://futurist-cvpr2025.github.io/ .

Method: Proposes DeQA-Score with two key innovations: 1) Distribution-based discretization that converts score distributions into soft labels instead of one-hot labels, preserving distribution characteristics and inter-image relationships; 2) Fidelity loss based on Thurstone’s model to handle dataset variation by capturing intra-dataset relationships, enabling effective co-training across multiple IQA datasets.

Result: Experiments across multiple benchmarks show DeQA-Score stably outperforms baselines in score regression. The model can predict score distributions that closely align with human annotations, demonstrating superior performance in both accuracy and distribution prediction.

Conclusion: DeQA-Score effectively bridges the gap between continuous quality scores and discrete MLLM outputs through distribution-based discretization and fidelity loss, achieving state-of-the-art performance in MLLM-based image quality assessment with accurate score regression capabilities.

Abstract: With the rapid advancement of Multi-modal Large Language Models (MLLMs), MLLM-based Image Quality Assessment (IQA) methods have shown promising performance in linguistic quality description. However, current methods still fall short in accurately scoring image quality. In this work, we aim to leverage MLLMs to regress accurate quality scores. A key challenge is that the quality score is inherently continuous, typically modeled as a Gaussian distribution, whereas MLLMs generate discrete token outputs. This mismatch necessitates score discretization. Previous approaches discretize the mean score into a one-hot label, resulting in information loss and failing to capture inter-image relationships. We propose a distribution-based approach that discretizes the score distribution into a soft label. This method preserves the characteristics of the score distribution, achieving high accuracy and maintaining inter-image relationships. Moreover, to address dataset variation, where different IQA datasets exhibit various distributions, we introduce a fidelity loss based on Thurstone’s model. This loss captures intra-dataset relationships, facilitating co-training across multiple IQA datasets. With these designs, we develop the distribution-based Depicted image Quality Assessment model for Score regression (DeQA-Score). Experiments across multiple benchmarks show that DeQA-Score stably outperforms baselines in score regression. Also, DeQA-Score can predict the score distribution that closely aligns with human annotations. Codes and model weights have been released in https://depictqa.github.io/deqa-score/.

Method: MLAD uses two main components: 1) Global-level deconfusion via dynamic-exit modality encoders and cross-class residual reconstruction to learn class-wise latent distributions, and 2) Sample-level deconfusion through sample-adaptive cross-modality rectification guided by confusion-free modality priors identified using Gaussian mixture models.

Result: MLAD outperforms state-of-the-art methods across multiple benchmarks and exhibits superior reliability in multimodal classification tasks.

Conclusion: The proposed MLAD framework effectively eliminates inter-class confusion at both global and sample levels, significantly enhancing the reliability of multimodal models for real-world applications with noisy data.

Abstract: Multimodal learning enhances the performance of various machine learning tasks by leveraging complementary information across different modalities. However, existing methods often learn multimodal representations that retain substantial inter-class confusion, making it difficult to achieve high-confidence predictions, particularly in real-world scenarios with low-quality or noisy data. To address this challenge, we propose Multi-Level Adaptive DeConfusion (MLAD), which eliminates inter-class confusion in multimodal data at both global and sample levels, significantly enhancing the classification reliability of multimodal models. Specifically, MLAD first learns class-wise latent distributions with global-level confusion removed via dynamic-exit modality encoders that adapt to the varying discrimination difficulty of each class and a cross-class residual reconstruction mechanism. Subsequently, MLAD further removes sample-specific confusion through sample-adaptive cross-modality rectification guided by confusion-free modality priors. These priors are constructed from low-confusion modality features, identified by evaluating feature confusion using the learned class-wise latent distributions and selecting those with low confusion via a Gaussian mixture model. Experiments demonstrate that MLAD outperforms state-of-the-art methods across multiple benchmarks and exhibits superior reliability.

Method: 1) Self-supervised masked autoencoder (MAE) pre-training on unlabeled data; 2) Graph attention networks (GAT) to model label correlations via expert-defined structured graphs; 3) Constraint-aware optimization using KL divergence and regularization losses for clinical priors; 4) Asymmetric loss (ASL) and boosting ensembles for label imbalance.

Result: Achieves state-of-the-art performance on TongueAtlas-4K, a new comprehensive benchmark with 4,000 images and 22 diagnostic labels - the largest public tongue analysis dataset.

Conclusion: MIRNet effectively addresses key challenges in medical image analysis through integrated self-supervised learning, graph reasoning, and constraint optimization. While optimized for tongue diagnosis, the framework generalizes to broader medical imaging tasks.

Abstract: Automated interpretation of medical images demands robust modeling of complex visual-semantic relationships while addressing annotation scarcity, label imbalance, and clinical plausibility constraints. We introduce MIRNet (Medical Image Reasoner Network), a novel framework that integrates self-supervised pre-training with constrained graph-based reasoning. Tongue image diagnosis is a particularly challenging domain that requires fine-grained visual and semantic understanding. Our approach leverages self-supervised masked autoencoder (MAE) to learn transferable visual representations from unlabeled data; employs graph attention networks (GAT) to model label correlations through expert-defined structured graphs; enforces clinical priors via constraint-aware optimization using KL divergence and regularization losses; and mitigates imbalance using asymmetric loss (ASL) and boosting ensembles. To address annotation scarcity, we also introduce TongueAtlas-4K, a comprehensive expert-curated benchmark comprising 4,000 images annotated with 22 diagnostic labels–representing the largest public dataset in tongue analysis. Validation shows our method achieves state-of-the-art performance. While optimized for tongue diagnosis, the framework readily generalizes to broader diagnostic medical imaging tasks.

Method: Three key components: (1) estimating tumor-related histomorphology importance at patch-level using medical prior knowledge, (2) generating representative prototypes through histomorphology-prototypical clustering, and (3) enabling WSI classification through histomorphology-guided prototypical aggregation. The framework adjusts decision boundaries by incorporating histomorphological importance to reduce instance label uncertainty.

Result: Experimental results demonstrate effectiveness with high diagnostic accuracy for molecular subtyping, cancer subtyping, and survival analysis.

Conclusion: HGPMIL successfully addresses limitations of existing WSI classification methods by explicitly learning histomorphology-guided prototypical representations, improving diagnostic accuracy across multiple cancer analysis tasks.

Abstract: Histomorphology is crucial in cancer diagnosis. However, existing whole slide image (WSI) classification methods struggle to effectively incorporate histomorphology information, limiting their ability to capture key pathological features. Particularly when the number of instances within a bag is large and their features are complex, it becomes challenging to accurately identify instances decisive for the bag label, making these methods prone to interference from ambiguous instances. To address this limitation, we propose a novel Histomorphology-Guided Prototypical Multi-Instance Learning (HGPMIL) framework that explicitly learns histomorphology-guided prototypical representations by incorporating tumor cellularity, cellular morphology, and tissue architecture. Specifically, our approach consists of three key components: (1) estimating the importance of tumor-related histomorphology information at patch-level based on medical prior knowledge; (2) generating representative prototypes through histomorphology-prototypical clustering; and (3) enabling WSI classification through histomorphology-guided prototypical aggregation. HGPMIL adjusts the decision boundary by incorporating histomorphological importance to reduce instance label uncertainty, thereby reversely optimizing the bag-level boundary. Experimental results demonstrate its effectiveness, achieving high diagnostic accuracy for molecular subtyping, cancer subtyping and survival analysis. The code will be made available at https://github.com/Badgewho/HMDMIL.

Method: 1) Train a discrete style codebook from image collections to extract style embeddings. 2) Use these embeddings as conditions for a text-to-image diffusion model to generate stylistic images. 3) Train an autoregressive style generator on discrete style embeddings to model their distribution for novel style synthesis. 4) During inference, map numerical codes to unique style embeddings that guide the diffusion model.

Result: Extensive experiments validate that CoTyle effectively turns numerical codes into style controllers, demonstrating unparalleled simplicity and diversity while unlocking a vast space of reproducible styles from minimal input.

Conclusion: CoTyle successfully proves that “a style is worth one numerical code,” offering the first open-source method for code-to-style image generation that addresses style consistency, creativity, and simplicity limitations of existing approaches.

Abstract: Innovative visual stylization is a cornerstone of artistic creation, yet generating novel and consistent visual styles remains a significant challenge. Existing generative approaches typically rely on lengthy textual prompts, reference images, or parameter-efficient fine-tuning to guide style-aware image generation, but often struggle with style consistency, limited creativity, and complex style representations. In this paper, we affirm that a style is worth one numerical code by introducing the novel task, code-to-style image generation, which produces images with novel, consistent visual styles conditioned solely on a numerical style code. To date, this field has only been primarily explored by the industry (e.g., Midjourney), with no open-source research from the academic community. To fill this gap, we propose CoTyle, the first open-source method for this task. Specifically, we first train a discrete style codebook from a collection of images to extract style embeddings. These embeddings serve as conditions for a text-to-image diffusion model (T2I-DM) to generate stylistic images. Subsequently, we train an autoregressive style generator on the discrete style embeddings to model their distribution, allowing the synthesis of novel style embeddings. During inference, a numerical style code is mapped to a unique style embedding by the style generator, and this embedding guides the T2I-DM to generate images in the corresponding style. Unlike existing methods, our method offers unparalleled simplicity and diversity, unlocking a vast space of reproducible styles from minimal input. Extensive experiments validate that CoTyle effectively turns a numerical code into a style controller, demonstrating a style is worth one code.

Method: Proposes EMO-X with Scan-based Global-Local Decoder (SGLD) that integrates global context with skeleton-aware local features. Uses Mamba for superior global modeling and designs a local bidirectional scan mechanism for skeleton-aware local refinement to iteratively enhance human tokens.

Result: EMO-X achieves excellent balance between efficiency and accuracy, with 69.8% less inference time compared to SOTA methods while outperforming most of them in accuracy.

Conclusion: EMO-X successfully addresses computational efficiency issues in multi-person EHPS while maintaining high accuracy through innovative global-local architecture combining Mamba’s strengths with local refinement mechanisms.

Abstract: Expressive Human Pose and Shape Estimation (EHPS) aims to jointly estimate human pose, hand gesture, and facial expression from monocular images. Existing methods predominantly rely on Transformer-based architectures, which suffer from quadratic complexity in self-attention, leading to substantial computational overhead, especially in multi-person scenarios. Recently, Mamba has emerged as a promising alternative to Transformers due to its efficient global modeling capability. However, it remains limited in capturing fine-grained local dependencies, which are essential for precise EHPS. To address these issues, we propose EMO-X, the Efficient Multi-person One-stage model for multi-person EHPS. Specifically, we explore a Scan-based Global-Local Decoder (SGLD) that integrates global context with skeleton-aware local features to iteratively enhance human tokens. Our EMO-X leverages the superior global modeling capability of Mamba and designs a local bidirectional scan mechanism for skeleton-aware local refinement. Comprehensive experiments demonstrate that EMO-X strikes an excellent balance between efficiency and accuracy. Notably, it achieves a significant reduction in computational complexity, requiring 69.8% less inference time compared to state-of-the-art (SOTA) methods, while outperforming most of them in accuracy.

Method: Uses diffusion transformer (DiT) to uniformly process different input types, employs feature routing constraints for precise querying from reference images, implements placeholder strategies for condition placement control, and applies progressive three-stage training (initial simple tasks, full-scale training, quality alignment).

Result: Extensive experiments demonstrate DreamO can effectively perform various image customization tasks with high quality and flexibly integrate different types of control conditions.

Conclusion: DreamO presents a successful unified framework for image customization that overcomes the limitations of task-specific approaches and enables seamless integration of multiple conditions.

Abstract: Recently, extensive research on image customization (e.g., identity, subject, style, background, etc.) demonstrates strong customization capabilities in large-scale generative models. However, most approaches are designed for specific tasks, restricting their generalizability to combine different types of condition. Developing a unified framework for image customization remains an open challenge. In this paper, we present DreamO, an image customization framework designed to support a wide range of tasks while facilitating seamless integration of multiple conditions. Specifically, DreamO utilizes a diffusion transformer (DiT) framework to uniformly process input of different types. During training, we construct a large-scale training dataset that includes various customization tasks, and we introduce a feature routing constraint to facilitate the precise querying of relevant information from reference images. Additionally, we design a placeholder strategy that associates specific placeholders with conditions at particular positions, enabling control over the placement of conditions in the generated results. Moreover, we employ a progressive training strategy consisting of three stages: an initial stage focused on simple tasks with limited data to establish baseline consistency, a full-scale training stage to comprehensively enhance the customization capabilities, and a final quality alignment stage to correct quality biases introduced by low-quality data. Extensive experiments demonstrate that the proposed DreamO can effectively perform various image customization tasks with high quality and flexibly integrate different types of control conditions.

Method: Proposes Iterative Implicit Neural Representations (I-INRs), a plug-and-play framework that enhances signal reconstruction through an iterative refinement process, compatible with existing INR architectures.

Result: I-INRs outperform baseline methods (WIRE, SIREN, Gauss) across diverse computer vision applications including image restoration, denoising, and object occupancy prediction.

Conclusion: The iterative refinement approach effectively addresses INR limitations, delivering superior reconstruction quality with better high-frequency detail recovery and noise robustness.

Abstract: Implicit Neural Representations (INRs) have revolutionized signal processing and computer vision by modeling signals as continuous, differentiable functions parameterized by neural networks. However, their inherent formulation as a regression problem makes them prone to regression to the mean, limiting their ability to capture fine details, retain high-frequency information, and handle noise effectively. To address these challenges, we propose Iterative Implicit Neural Representations (I-INRs) a novel plug-and-play framework that enhances signal reconstruction through an iterative refinement process. I-INRs effectively recover high-frequency details, improve robustness to noise, and achieve superior reconstruction quality. Our framework seamlessly integrates with existing INR architectures, delivering substantial performance gains across various tasks. Extensive experiments show that I-INRs outperform baseline methods, including WIRE, SIREN, and Gauss, in diverse computer vision applications such as image restoration, image denoising, and object occupancy prediction.

Method: Proposes Tangible Attack (TBA) framework with Dual Heterogeneous Noise Generator (DHNG) using VAE and ControlNet to produce targeted noise, plus custom adversarial loss function. Iteratively refines adversarial samples using multi-gradient signals from noise and EHPS model.

Result: TBA achieves 41.0% increase in estimation error with average improvement of ~17.0%, demonstrating superior attack effectiveness compared to existing methods.

Conclusion: The research exposes significant security vulnerabilities in current EHPS models and highlights the urgent need for stronger defenses in digital human generation systems.

Abstract: Expressive human pose and shape estimation (EHPS) is crucial for digital human generation, especially in applications like live streaming. While existing research primarily focuses on reducing estimation errors, it largely neglects robustness and security aspects, leaving these systems vulnerable to adversarial attacks. To address this significant challenge, we propose the \textbf{Tangible Attack (TBA)}, a novel framework designed to generate adversarial examples capable of effectively compromising any digital human generation model. Our approach introduces a \textbf{Dual Heterogeneous Noise Generator (DHNG)}, which leverages Variational Autoencoders (VAE) and ControlNet to produce diverse, targeted noise tailored to the original image features. Additionally, we design a custom \textbf{adversarial loss function} to optimize the noise, ensuring both high controllability and potent disruption. By iteratively refining the adversarial sample through multi-gradient signals from both the noise and the state-of-the-art EHPS model, TBA substantially improves the effectiveness of adversarial attacks. Extensive experiments demonstrate TBA’s superiority, achieving a remarkable 41.0% increase in estimation error, with an average improvement of approximately 17.0%. These findings expose significant security vulnerabilities in current EHPS models and highlight the need for stronger defenses in digital human generation systems.

Method: Three core modules: 1) Inter Correlation Enhancement (ICE) creates pseudo-positive pairs from semantically similar unpaired text-video moments; 2) Intra Redundancy Mining (IRM) distinguishes redundant from query-relevant moments; 3) Temporal Coherence Prediction (TCP) predicts original order of shuffled video sequences.

Result: Extensive experiments demonstrate superiority and state-of-the-art performance of the proposed method.

Conclusion: The framework effectively addresses PRVR challenges by systematically exploiting inter-sample correlation and intra-sample redundancy, leading to improved video retrieval performance.

Abstract: Partially Relevant Video Retrieval (PRVR) aims to retrieve the target video that is partially relevant to the text query. The primary challenge in PRVR arises from the semantic asymmetry between textual and visual modalities, as videos often contain substantial content irrelevant to the query. Existing methods coarsely align paired videos and text queries to construct the semantic space, neglecting the critical cross-modal dual nature inherent in this task: inter-sample correlation and intra-sample redundancy. To this end, we propose a novel PRVR framework to systematically exploit these two characteristics. Our framework consists of three core modules. First, the Inter Correlation Enhancement (ICE) module captures inter-sample correlation by identifying semantically similar yet unpaired text queries and video moments, combining them to form pseudo-positive pairs for more robust semantic space construction. Second, the Intra Redundancy Mining (IRM) module mitigates intra-sample redundancy by mining redundant moment features and distinguishing them from query-relevant moments, encouraging the model to learn more discriminative representations. Finally, to reinforce these modules, we introduce the Temporal Coherence Prediction (TCP) module, enhancing discrimination of fine-grained moment-level semantics by training the model to predict the original temporal order of randomly shuffled video sequences. Extensive experiments demonstrate the superiority of our method, achieving state-of-the-art results.

Method: ARIAL uses an LLM-based planning agent to orchestrate specialized tools through a modular framework: OCR-based text extraction with TrOCR, retrieval-augmented context selection via semantic search, answer generation with fine-tuned Gemma 3-27B, and explicit bounding-box localization through text-to-region alignment.

Result: State-of-the-art results across four benchmarks: DocVQA (88.7 ANLS, 50.1 mAP), FUNSD (90.0 ANLS, 50.3 mAP), CORD (85.5 ANLS, 60.2 mAP), and SROIE (93.1 ANLS). Surpasses previous best method (DLaVA) by +2.8 ANLS and +3.9 mAP on DocVQA.

Conclusion: Agentic orchestration of specialized tools can simultaneously improve performance and interpretability in Document VQA, providing a pathway toward trustworthy, explainable document AI systems with transparent reasoning traces and tool-level auditability.

Abstract: Document Visual Question Answering (VQA) requires models to not only extract accurate textual answers but also precisely localize them within document images, a capability critical for interpretability in high-stakes applications. However, existing systems achieve strong textual accuracy while producing unreliable spatial grounding, or sacrifice performance for interpretability. We present ARIAL (Agentic Reasoning for Interpretable Answer Localization), a modular framework that orchestrates specialized tools through an LLM-based planning agent to achieve both precise answer extraction and reliable spatial grounding. ARIAL decomposes Document VQA into structured subtasks: OCR-based text extraction with TrOCR, retrieval-augmented context selection using semantic search, answer generation via a fine-tuned Gemma 3-27B model, and explicit bounding-box localization through text-to-region alignment. This modular architecture produces transparent reasoning traces, enabling tool-level auditability and independent component optimization. We evaluate ARIAL on four benchmarks (DocVQA, FUNSD, CORD, and SROIE) using both textual accuracy (ANLS) and spatial precision (mAP at IoU 0.50 to 0.95). ARIAL achieves state-of-the-art results across all datasets: 88.7 ANLS and 50.1 mAP on DocVQA, 90.0 ANLS and 50.3 mAP on FUNSD, 85.5 ANLS and 60.2 mAP on CORD, and 93.1 ANLS on SROIE, surpassing the previous best method (DLaVA) by +2.8 ANLS and +3.9 mAP on DocVQA. Our work demonstrates how agentic orchestration of specialized tools can simultaneously improve performance and interpretability, providing a pathway toward trustworthy, explainable document AI systems.

Method: Organized SPPIN challenge with training phase (78 MRI sets from 34 patients) and test phase (18 MRI sets from 9 patients). Teams developed segmentation methods evaluated using Dice score, HD95, and volumetric similarity.

Result: Top team achieved median Dice 0.82, HD95 7.69mm, VS 0.91 using STU-Net. Significant performance difference between diagnostic (Dice 0.89) and post-chemotherapy scans (Dice 0.59).

Conclusion: Pretraining helps with small datasets, but segmentation of small, pre-treated tumors remains insufficient. More reliable methods needed for clinical application in pediatric neuroblastoma surgical planning.

Abstract: Surgery plays an important role within the treatment for neuroblastoma, a common pediatric cancer. This requires careful planning, often via magnetic resonance imaging (MRI)-based anatomical 3D models. However, creating these models is often time-consuming and user dependent. We organized the Surgical Planning in Pediatric Neuroblastoma (SPPIN) challenge, to stimulate developments on this topic, and set a benchmark for fully automatic segmentation of neuroblastoma on multi-model MRI. The challenge started with a training phase, where teams received 78 sets of MRI scans from 34 patients, consisting of both diagnostic and post-chemotherapy MRI scans. The final test phase, consisting of 18 MRI sets from 9 patients, determined the ranking of the teams. Ranking was based on the Dice similarity coefficient (Dice score), the 95th percentile of the Hausdorff distance (HD95) and the volumetric similarity (VS). The SPPIN challenge was hosted at MICCAI 2023. The final leaderboard consisted of 9 teams. The highest-ranking team achieved a median Dice score 0.82, a median HD95 of 7.69 mm and a VS of 0.91, utilizing a large, pretrained network called STU-Net. A significant difference for the segmentation results between diagnostic and post-chemotherapy MRI scans was observed (Dice = 0.89 vs Dice = 0.59, P = 0.01) for the highest-ranking team. SPPIN is the first medical segmentation challenge in extracranial pediatric oncology. The highest-ranking team used a large pre-trained network, suggesting that pretraining can be of use in small, heterogenous datasets. Although the results of the highest-ranking team were high for most patients, segmentation especially in small, pre-treated tumors were insufficient. Therefore, more reliable segmentation methods are needed to create clinically applicable models to aid surgical planning in pediatric neuroblastoma.

Method: Yo’City uses a hierarchical “City-District-Grid” planning strategy with Global Planner and Local Designer, followed by “produce-refine-evaluate” isometric image synthesis loop for 3D generation, and relationship-guided expansion for continuous growth.

Result: Yo’City outperforms state-of-the-art methods across all evaluation aspects (semantics, geometry, texture, layout) on a diverse benchmark dataset with six multi-dimensional metrics.

Conclusion: Yo’City enables user-customized, infinitely expandable 3D city generation by leveraging large models’ reasoning capabilities, offering superior performance and addressing limitations of existing approaches.

Abstract: Realistic 3D city generation is fundamental to a wide range of applications, including virtual reality and digital twins. However, most existing methods rely on training a single diffusion model, which limits their ability to generate personalized and boundless city-scale scenes. In this paper, we present Yo’City, a novel agentic framework that enables user-customized and infinitely expandable 3D city generation by leveraging the reasoning and compositional capabilities of off-the-shelf large models. Specifically, Yo’City first conceptualize the city through a top-down planning strategy that defines a hierarchical “City-District-Grid” structure. The Global Planner determines the overall layout and potential functional districts, while the Local Designer further refines each district with detailed grid-level descriptions. Subsequently, the grid-level 3D generation is achieved through a “produce-refine-evaluate” isometric image synthesis loop, followed by image-to-3D generation. To simulate continuous city evolution, Yo’City further introduces a user-interactive, relationship-guided expansion mechanism, which performs scene graph-based distance- and semantics-aware layout optimization, ensuring spatially coherent city growth. To comprehensively evaluate our method, we construct a diverse benchmark dataset and design six multi-dimensional metrics that assess generation quality from the perspectives of semantics, geometry, texture, and layout. Extensive experiments demonstrate that Yo’City consistently outperforms existing state-of-the-art methods across all evaluation aspects.

Method: The researchers developed a convolutional neural network (CNN) framework for rice grain classification. The model was trained and validated using performance metrics including accuracy, recall, precision, and F1-Score. Explainability techniques (LIME and SHAP) were integrated to understand the model’s decision-making process.

Result: The CNN model achieved remarkable accuracy rates and perfect area under each class’s ROC curve. Confusion matrix analysis showed minimal misclassifications, confirming the model’s effectiveness in distinguishing between different rice varieties. The explainability techniques provided valuable insights into how specific rice grain features influenced classification outcomes.

Conclusion: The proposed CNN-based framework provides an effective and efficient solution for automatic rice grain classification, outperforming manual methods. The integration of explainability techniques enhances transparency and understanding of the model’s decisions, making it suitable for quality control in rice trade and cultivation.

Abstract: Rice is an essential staple food worldwide that is important in promoting international trade, economic growth, and nutrition. Asian countries such as China, India, Pakistan, Thailand, Vietnam, and Indonesia are notable for their significant contribution to the cultivation and utilization of rice. These nations are also known for cultivating different rice grains, including short and long grains. These sizes are further classified as basmati, jasmine, kainat saila, ipsala, arborio, etc., catering to diverse culinary preferences and cultural traditions. For both local and international trade, inspecting and maintaining the quality of rice grains to satisfy customers and preserve a country’s reputation is necessary. Manual quality check and classification is quite a laborious and time-consuming process. It is also highly prone to mistakes. Therefore, an automatic solution must be proposed for the effective and efficient classification of different varieties of rice grains. This research paper presents an automatic framework based on a convolutional neural network (CNN) for classifying different varieties of rice grains. We evaluated the proposed model based on performance metrics such as accuracy, recall, precision, and F1-Score. The CNN model underwent rigorous training and validation, achieving a remarkable accuracy rate and a perfect area under each class’s Receiver Operating Characteristic (ROC) curve. The confusion matrix analysis confirmed the model’s effectiveness in distinguishing between the different rice varieties, indicating minimal misclassifications. Additionally, the integration of explainability techniques such as LIME (Local Interpretable Model-agnostic Explanations) and SHAP (SHapley Additive exPlanations) provided valuable insights into the model’s decision-making process, revealing how specific features of the rice grains influenced classification outcomes.

Method: Developed ProxT2I using backward discretization with learned conditional proximal operators instead of score functions. Used reinforcement learning and policy optimization to optimize samplers for task-specific rewards. Created LAION-Face-T2I-15M dataset with 15M high-quality human images with fine-grained captions for training/evaluation.

Result: Enhanced sampling efficiency and human-preference alignment compared to score-based baselines. Achieved results on par with state-of-the-art open-source text-to-image models while requiring lower compute and smaller model size.

Conclusion: ProxT2I offers a lightweight yet performant solution for human text-to-image generation, demonstrating that backward discretization with proximal operators and RL optimization can overcome limitations of traditional diffusion models.

Abstract: Diffusion models have emerged as a dominant paradigm for generative modeling across a wide range of domains, including prompt-conditional generation. The vast majority of samplers, however, rely on forward discretization of the reverse diffusion process and use score functions that are learned from data. Such forward and explicit discretizations can be slow and unstable, requiring a large number of sampling steps to produce good-quality samples. In this work we develop a text-to-image (T2I) diffusion model based on backward discretizations, dubbed ProxT2I, relying on learned and conditional proximal operators instead of score functions. We further leverage recent advances in reinforcement learning and policy optimization to optimize our samplers for task-specific rewards. Additionally, we develop a new large-scale and open-source dataset comprising 15 million high-quality human images with fine-grained captions, called LAION-Face-T2I-15M, for training and evaluation. Our approach consistently enhances sampling efficiency and human-preference alignment compared to score-based baselines, and achieves results on par with existing state-of-the-art and open-source text-to-image models while requiring lower compute and smaller model size, offering a lightweight yet performant solution for human text-to-image generation.

Method: CDLC extracts global, class-specific concept directions by clustering latent difference vectors derived from factual and diffusion-generated counterfactual image pairs, enabling efficient extraction without GPU for clustering.

Result: CDLC reduces storage requirements by ~4.6% and accelerates concept discovery by ~5.3% compared to baseline, while extracting clinically relevant concepts that align with dermoscopic features and reveal dataset biases.

Conclusion: CDLC is an interpretable, scalable concept extraction method applicable across high-stakes domains and diverse data modalities, offering efficient multidimensional semantic concept discovery.

Abstract: Concept-based explanations have emerged as an effective approach within Explainable Artificial Intelligence, enabling interpretable insights by aligning model decisions with human-understandable concepts. However, existing methods rely on computationally intensive procedures and struggle to efficiently capture complex, semantic concepts. This work introduces the Concept Directions via Latent Clustering (CDLC), which extracts global, class-specific concept directions by clustering latent difference vectors derived from factual and diffusion-generated counterfactual image pairs. CDLC reduces storage requirements by ~4.6% and accelerates concept discovery by ~5.3% compared to the baseline method, while requiring no GPU for clustering, thereby enabling efficient extraction of multidimensional semantic concepts across latent dimensions. This approach is validated on a real-world skin lesion dataset, demonstrating that the extracted concept directions align with clinically recognized dermoscopic features and, in some cases, reveal dataset-specific biases or unknown biomarkers. These results highlight that CDLC is interpretable, scalable, and applicable across high-stakes domains and diverse data modalities.

Method: Tested 4 frontier VLMs (Claude Sonnet 4.5, GPT-4o, GPT-5-mini, Gemini 2.0 flash exp) on 60 Italian medical questions from EuropeMedQA that require image interpretation, replacing correct medical images with blank placeholders to measure visual dependency.

Result: GPT-4o showed strongest visual grounding with 27.9pp accuracy drop (83.2% to 55.3%), while GPT-5-mini, Gemini, and Claude had modest drops of 8.5pp, 2.4pp, and 5.6pp respectively. All models generated confident explanations for fabricated visual interpretations.

Conclusion: VLMs exhibit critical variability in visual dependency, with some relying heavily on textual shortcuts rather than genuine visual analysis, highlighting the need for rigorous evaluation before clinical deployment.

Abstract: Large vision language models (VLMs) have achieved impressive performance on medical visual question answering benchmarks, yet their reliance on visual information remains unclear. We investigate whether frontier VLMs demonstrate genuine visual grounding when answering Italian medical questions by testing four state-of-the-art models: Claude Sonnet 4.5, GPT-4o, GPT-5-mini, and Gemini 2.0 flash exp. Using 60 questions from the EuropeMedQA Italian dataset that explicitly require image interpretation, we substitute correct medical images with blank placeholders to test whether models truly integrate visual and textual information. Our results reveal striking variability in visual dependency: GPT-4o shows the strongest visual grounding with a 27.9pp accuracy drop (83.2% [74.6%, 91.7%] to 55.3% [44.1%, 66.6%]), while GPT-5-mini, Gemini, and Claude maintain high accuracy with modest drops of 8.5pp, 2.4pp, and 5.6pp respectively. Analysis of model-generated reasoning reveals confident explanations for fabricated visual interpretations across all models, suggesting varying degrees of reliance on textual shortcuts versus genuine visual analysis. These findings highlight critical differences in model robustness and the need for rigorous evaluation before clinical deployment.

Method: 1) Created SAMData dataset with hundreds of verified aerial images and detailed captions highlighting subtle military installations; 2) Supervised fine-tuning on 2B parameter open-source MLLM with chain-of-thought reasoning annotations; 3) Applied Group Relative Policy Optimization (GRPO) to enhance detection of domain-specific cues while minimizing false positives.

Result: SAMChat significantly outperforms both larger general-purpose multimodal models and existing remote sensing adapted approaches on open-ended captioning and classification metrics. Achieved over 80% recall and 98% precision on the SAMData benchmark.

Conclusion: Targeted fine-tuning and reinforcement learning (specifically GRPO) are highly effective for specialized real-world applications, enabling lightweight models to achieve superior performance in domain-specific tasks like military remote sensing analysis.

Abstract: Remarkable capabilities in understanding and generating text-image content have been demonstrated by recent advancements in multimodal large language models (MLLMs). However, their effectiveness in specialized domains-particularly those requiring resource-efficient and domain-specific adaptations-has remained limited. In this work, a lightweight multimodal language model termed SAMChat is introduced, specifically adapted to analyze remote sensing imagery in secluded areas, including challenging missile launch sites. A new dataset, SAMData, was compiled by verifying hundreds of aerial images through expert review, and subtle military installations were highlighted via detailed captions. Supervised fine-tuning on a 2B parameter open-source MLLM with chain-of-thought (CoT) reasoning annotations was performed, enabling more accurate and interpretable explanations. Additionally, Group Relative Policy Optimization (GRPO) was leveraged to enhance the model’s ability to detect critical domain-specific cues-such as defensive layouts and key military structures-while minimizing false positives on civilian scenes. Through empirical evaluations, it has been shown that SAMChat significantly outperforms both larger, general-purpose multimodal models and existing remote sensing adapted approaches on open-ended captioning and classification metrics. Over 80% recall and 98% precision were achieved on the newly proposed SAMData benchmark, underscoring the potency of targeted fine-tuning and reinforcement learning in specialized real-world applications.

Method: Proposes DiT-Mem, a learnable memory encoder with stacked 3D CNNs, low-/high-pass filters, and self-attention layers that maps reference videos into memory tokens. These tokens are concatenated within DiT self-attention layers. Training keeps the diffusion backbone frozen and only optimizes the memory encoder.

Result: Efficient training with few parameters (150M) and 10K data samples. Plug-and-play inference improves physical rule following and video fidelity in state-of-the-art models.

Conclusion: DiT-Mem effectively injects world knowledge into video generation models through a memory mechanism, addressing physical law violations while maintaining efficiency and plug-and-play usability.

Abstract: Diffusion Transformer(DiT) based video generation models have recently achieved impressive visual quality and temporal coherence, but they still frequently violate basic physical laws and commonsense dynamics, revealing a lack of explicit world knowledge. In this work, we explore how to equip them with a plug-and-play memory that injects useful world knowledge. Motivated by in-context memory in Transformer-based LLMs, we conduct empirical studies to show that DiT can be steered via interventions on its hidden states, and simple low-pass and high-pass filters in the embedding space naturally disentangle low-level appearance and high-level physical/semantic cues, enabling targeted guidance. Building on these observations, we propose a learnable memory encoder DiT-Mem, composed of stacked 3D CNNs, low-/high-pass filters, and self-attention layers. The encoder maps reference videos into a compact set of memory tokens, which are concatenated as the memory within the DiT self-attention layers. During training, we keep the diffusion backbone frozen, and only optimize the memory encoder. It yields a rather efficient training process on few training parameters (150M) and 10K data samples, and enables plug-and-play usage at inference time. Extensive experiments on state-of-the-art models demonstrate the effectiveness of our method in improving physical rule following and video fidelity. Our code and data are publicly released here: https://thrcle421.github.io/DiT-Mem-Web/.

Method: Introduces self-conditioning - a lightweight mechanism that aggregates and reroutes intermediate features to guide subsequent decoding layers, concentrating high-level semantics and creating a semantic bridge without external guidance.

Result: Enhanced models (especially self-conditioned DiT) achieve dual improvements: strong discriminative representations surpassing various generative self-supervised models in linear probing while maintaining or improving generation quality across pixel-space UNet, UViT and latent-space DiT models.

Conclusion: Self-conditioning creates an architectural semantic bridge that enables diffusion models to become powerful dual learners, advancing unified generative and discriminative learning with minimal overhead.

Abstract: While diffusion models excel at image synthesis, useful representations have been shown to emerge from generative pre-training, suggesting a path towards unified generative and discriminative learning. However, suboptimal semantic flow within current architectures can hinder this potential: features encoding the richest high-level semantics are underutilized and diluted when propagating through decoding layers, impeding the formation of an explicit semantic bottleneck layer. To address this, we introduce self-conditioning, a lightweight mechanism that reshapes the model’s layer-wise semantic hierarchy without external guidance. By aggregating and rerouting intermediate features to guide subsequent decoding layers, our method concentrates more high-level semantics, concurrently strengthening global generative guidance and forming more discriminative representations. This simple approach yields a dual-improvement trend across pixel-space UNet, UViT and latent-space DiT models with minimal overhead. Crucially, it creates an architectural semantic bridge that propagates discriminative improvements into generation and accommodates further techniques such as contrastive self-distillation. Experiments show that our enhanced models, especially self-conditioned DiT, are powerful dual learners that yield strong and transferable representations on image and dense classification tasks, surpassing various generative self-supervised models in linear probing while also improving or maintaining high generation quality.

Method: ParticleGS reformulates dynamic 3D scenes as physically grounded systems with three components: 1) an encoder that decomposes scenes into static properties and initial dynamic physical fields, 2) a Neural ODE-based evolver that learns continuous-time dynamics for motion extrapolation, and 3) a decoder that reconstructs 3D Gaussians from evolved particle states for rendering.

Result: ParticleGS achieves state-of-the-art performance in extrapolation tasks while maintaining rendering quality comparable to leading dynamic 3D reconstruction methods.

Conclusion: The proposed physics-based framework successfully integrates physical reasoning into dynamic 3D representations, enabling accurate and consistent prediction of future scenes, which advances physical world understanding and predictive modeling capabilities.

Abstract: The ability to extrapolate dynamic 3D scenes beyond the observed timeframe is fundamental to advancing physical world understanding and predictive modeling. Existing dynamic 3D reconstruction methods have achieved high-fidelity rendering of temporal interpolation, but typically lack physical consistency in predicting the future. To overcome this issue, we propose ParticleGS, a physics-based framework that reformulates dynamic 3D scenes as physically grounded systems. ParticleGS comprises three key components: 1) an encoder that decomposes the scene into static properties and initial dynamic physical fields; 2) an evolver based on Neural Ordinary Differential Equations (Neural ODEs) that learns continuous-time dynamics for motion extrapolation; and 3) a decoder that reconstructs 3D Gaussians from evolved particle states for rendering. Through this design, ParticleGS integrates physical reasoning into dynamic 3D representations, enabling accurate and consistent prediction of the future. Experiments show that ParticleGS achieves state-of-the-art performance in extrapolation while maintaining rendering quality comparable to leading dynamic 3D reconstruction methods.

Method: Proposes MoDOT framework with: 1) Cross-Attention Strip Module (CASM) to leverage mid-level occlusion boundary features for depth prediction, 2) OB-Depth Constraint Loss (OBDCL) to enforce geometric consistency between occlusion boundaries and depth discontinuities, and 3) OB-Hypersim dataset - a large-scale photorealistic synthetic dataset with precise depth and self-occlusion-handled occlusion boundary annotations.

Result: MoDOT achieves significantly better performance than single-task baselines and multi-task competitors on two synthetic datasets and NYUD-v2. Models trained solely on synthetic data demonstrate strong generalization to real-world scenes without fine-tuning, producing depth maps with sharper boundaries and improved geometric fidelity.

Conclusion: Joint modeling of occlusion boundaries and depth estimation provides significant benefits, enabling better performance and generalization. The proposed MoDOT framework and OB-Hypersim dataset effectively exploit the mutually beneficial relationship between these two tasks.

Abstract: Occlusion Boundary Estimation (OBE) identifies boundaries arising from both inter-object occlusions and self-occlusion within individual objects. This task is closely related to Monocular Depth Estimation (MDE), which infers depth from a single image, as Occlusion Boundaries (OBs) provide critical geometric cues for resolving depth ambiguities, while depth can conversely refine occlusion reasoning. In this paper, we aim to systematically model and exploit this mutually beneficial relationship. To this end, we propose MoDOT, a novel framework for joint estimation of depth and OBs, which incorporates a new Cross-Attention Strip Module (CASM) to leverage mid-level OB features for depth prediction, and a novel OB-Depth Constraint Loss (OBDCL) to enforce geometric consistency. To facilitate this study, we contribute OB-Hypersim, a large-scale photorealistic dataset with precise depth and self-occlusion-handled OB annotations. Extensive experiments on two synthetic datasets and NYUD-v2 demonstrate that MoDOT achieves significantly better performance than single-task baselines and multi-task competitors. Furthermore, models trained solely on our synthetic data demonstrate strong generalization to real-world scenes without fine-tuning, producing depth maps with sharper boundaries and improved geometric fidelity. Collectively, these results underscore the significant benefits of jointly modeling OBs and depth. Code and resources are available at https://github.com/xul-ops/MoDOT.

Method: Proposes S2AFormer with novel Strip Self-Attention (SSA) that reduces spatial dimensions of K and V while compressing channel dimensions of Q and K. Uses Hybrid Perception Blocks (HPBs) to integrate CNN’s local perception with Transformer’s global context modeling.

Result: Achieves significant accuracy gains with superior efficiency on ImageNet-1k (classification), ADE20k (segmentation), and COCO (detection/segmentation) benchmarks, performing well in both GPU and non-GPU environments.

Conclusion: S2AFormer effectively addresses ViT’s efficiency limitations through architectural innovations, making it a strong candidate for efficient vision Transformers by balancing computational efficiency with model accuracy.

Abstract: Vision Transformer (ViT) has made significant advancements in computer vision, thanks to its token mixer’s sophisticated ability to capture global dependencies between all tokens. However, the quadratic growth in computational demands as the number of tokens increases limits its practical efficiency. Although recent methods have combined the strengths of convolutions and self-attention to achieve better trade-offs, the expensive pairwise token affinity and complex matrix operations inherent in self-attention remain a bottleneck. To address this challenge, we propose S2AFormer, an efficient Vision Transformer architecture featuring novel Strip Self-Attention (SSA). We design simple yet effective Hybrid Perception Blocks (HPBs) to effectively integrate the local perception capabilities of CNNs with the global context modeling of Transformer’s attention mechanisms. A key innovation of SSA lies in its reduction of the spatial dimensions of $K$ and $V$, while compressing the channel dimensions of $Q$ and $K$. This design significantly reduces computational overhead while preserving accuracy, striking an optimal balance between efficiency and effectiveness. We evaluate the robustness and efficiency of S2AFormer through extensive experiments on multiple vision benchmarks, including ImageNet-1k for image classification, ADE20k for semantic segmentation, and COCO for object detection and instance segmentation. Results demonstrate that S2AFormer achieves significant accuracy gains with superior efficiency in both GPU and non-GPU environments, making it a strong candidate for efficient vision Transformers.

Method: The study uses theoretical analysis (mathematical formulas and theorem proofs), experimental investigation with public datasets, and statistical analysis (Pearson’s correlation coefficient and non-linear correlation measures).

Result: The two measures are theoretically independent, graphical analysis supports this independence, and statistical analysis confirms they are uncorrelated.

Conclusion: Weighted figure of merit cannot be used instead of Rosin’s measure - if Rosin’s measure indicates one scheme is better/worse than another, the same conclusion cannot be drawn using weighted figure of merit.

Abstract: Many studies have been conducted to solve the problem of approximating a digital boundary by piece straight-line segments for the further processing required in computer vision applications. The authors of these studies compared their schemes to determine the best one. The initial measure used to assess the goodness of fit of a polygonal approximation was the figure of merit. Later,it was noted that this measure was not an appropriate metric for a valid reason which is why Rosin-through mathematical analysis-introduced a measure called merit. However,this measure involves an optimal scheme of polygonal approximation,so it is time-consuming to compute it to assess the goodness of fit of an approximation. This led many researchers to use a weighted figure of merit as a substitute for Rosin’s measure to compare sub optimal schemes. An attempt is made in this communication to investigate whether the two measures-weighted figure of merit and Rosin’s measure-are related so that one can be used instead of the other, and toward this end, theoretical analysis, experimental investigation and statistical analysis are carried out. The mathematical formulas for the weighted figure of merit and Rosin’s measure are analyzed, and through proof of theorems,it is found that the two measures are theoretically independent of each other. The graphical analysis of experiments carried out using a public dataset supports the results of the theoretical analysis. The statistical analysis via Pearson’s correlation coefficient and non-linear correlation measure also revealed that the two measures are uncorrelated. This analysis leads one to conclude that if a suboptimal scheme is found to be better (worse) than some other suboptimal scheme,as indicated by Rosin’s measure,then the same conclusion cannot be drawn using a weighted figure of merit,so one cannot use a weighted figure of merit instead of Rosin’s measure.

Method: Three-stage distillation-based SFT pipeline to address computational cost and supervision issues, plus VLPO (Visual-latent Policy Optimization) reinforcement learning that explicitly incorporates latent embeddings into policy gradient updates.

Result: Monet-7B shows consistent gains across real-world perception and reasoning benchmarks with strong out-of-distribution generalization on challenging abstract visual reasoning tasks.

Conclusion: The framework enables effective latent visual reasoning, provides insights through analysis of training components and early failures, and offers open-source model, data, and code for future development.

Abstract: “Thinking with images” has emerged as an effective paradigm for advancing visual reasoning, extending beyond text-only chains of thought by injecting visual evidence into intermediate reasoning steps. However, existing methods fall short of human-like abstract visual thinking, as their flexibility is fundamentally limited by external tools. In this work, we introduce Monet, a training framework that enables multimodal large language models (MLLMs) to reason directly within the latent visual space by generating continuous embeddings that function as intermediate visual thoughts. We identify two core challenges in training MLLMs for latent visual reasoning: high computational cost in latent-vision alignment and insufficient supervision over latent embeddings, and address them with a three-stage distillation-based supervised fine-tuning (SFT) pipeline. We further reveal a limitation of applying GRPO to latent reasoning: it primarily enhances text-based reasoning rather than latent reasoning. To overcome this, we propose VLPO (Visual-latent Policy Optimization), a reinforcement learning method that explicitly incorporates latent embeddings into policy gradient updates. To support SFT, we construct Monet-SFT-125K, a high-quality text-image interleaved CoT dataset containing 125K real-world, chart, OCR, and geometry CoTs. Our model, Monet-7B, shows consistent gains across real-world perception and reasoning benchmarks and exhibits strong out-of-distribution generalization on challenging abstract visual reasoning tasks. We also empirically analyze the role of each training component and discuss our early unsuccessful attempts, providing insights for future developments in visual latent reasoning. Our model, data, and code are available at https://github.com/NOVAglow646/Monet.

Method: Built OpenDanceSet (100+ hours, 14 genres, 147 subjects) with rich annotations (3D motion, music, 2D keypoints, trajectories, text descriptions). Proposed OpenDanceNet: a unified masked modeling framework with disentangled auto-encoder and multimodal joint-prediction Transformer.

Result: Comprehensive experiments show high-fidelity synthesis with strong diversity and realistic physical contacts, while offering flexible control over spatial and stylistic conditions.

Conclusion: The work addresses the dataset bottleneck and provides a unified framework for controllable dance generation, enabling versatile multimodal control beyond just music conditioning.

Abstract: Music-driven 3D dance generation offers significant creative potential, yet practical applications demand versatile and multimodal control. As the highly dynamic and complex human motion covering various styles and genres, dance generation requires satisfying diverse conditions beyond just music (e.g., spatial trajectories, keyframe gestures, or style descriptions). However, the absence of a large-scale and richly annotated dataset severely hinders progress. In this paper, we build OpenDanceSet, an extensive human dance dataset comprising over 100 hours across 14 genres and 147 subjects. Each sample has rich annotations to facilitate robust cross-modal learning: 3D motion, paired music, 2D keypoints, trajectories, and expert-annotated text descriptions. Furthermore, we propose OpenDanceNet, a unified masked modeling framework for controllable dance generation, including a disentangled auto-encoder and a multimodal joint-prediction Transformer. OpenDanceNet supports generation conditioned on music and arbitrary combinations of text, keypoints, or trajectories. Comprehensive experiments demonstrate that our work achieves high-fidelity synthesis with strong diversity and realistic physical contacts, while also offering flexible control over spatial and stylistic conditions. Project Page: https://open-dance.github.io

Method: Reinterpret CLIP-style models as pretrained density ratio estimators, provide unified explanation of how contrastive objectives estimate density ratios, and develop two practical applications: Importance Weight Learning (requiring only single additional prompt) and KL divergence estimation for multimodal distributions.

Result: Importance Weight Learning improves F1 scores by up to 7 points. CLIP-based density ratios enable estimation of KL divergences that capture semantic diversity and mode structure. KL-guided data curation achieves performance competitive with LAION2B filtering.

Conclusion: Viewing CLIP-style models as density ratio estimators unlocks new algorithmic capabilities for multimodal applications, demonstrating practical improvements in importance weighting, divergence estimation, and data curation.

Abstract: Density ratio estimation is a core concept in statistical machine learning because it provides a unified mechanism for tasks such as importance weighting, divergence estimation, and likelihood-free inference, but its potential in vision and language models has not been fully explored. Modern vision-language encoders such as CLIP and SigLIP are trained with contrastive objectives that implicitly optimize log density ratios between joint and marginal image-text distributions, which implicitly learn similarity scores proportional to log density ratios. However, prior work has largely focused on their embedding utility, and the density-ratio structure induced by contrastive learning has not been systematically examined or exploited in multimodal applications. To address this gap, we reinterpret CLIP-style models as pretrained and general-purpose density ratio estimators and show that this perspective enables new algorithmic capabilities. We present a unified explanation of how contrastive objectives estimate density ratios and propose two practical applications: Importance Weight Learning and KL divergence estimation. Our Importance Weight Learning method requires only a single additional prompt and improves F1 scores by up to 7 points. We further show that CLIP-based density ratios support estimation of KL divergences that quantify how conditioning on an image or text alters the distribution of the other modality. Through qualitative examples and an N-gram analysis of captions, we find that these divergences capture semantic diversity and mode structure in multimodal data. Leveraging this property, we introduce a simple KL-guided data curation method that achieves performance competitive with LAION2B filtering.

Method: Three architectural upgrades: (1) enhanced interleaved-MRoPE for spatial-temporal modeling, (2) DeepStack integration for better vision-language alignment using multi-level ViT features, and (3) text-based time alignment for video (evolving from T-RoPE to explicit timestamp alignment). Offers both dense (2B-32B) and MoE (30B-A3B/235B-A22B) variants.

Result: Achieves superior performance across multimodal benchmarks including MMMU, MathVista, and MathVision. Demonstrates strong pure-text understanding, robust long-context comprehension (256K tokens), and advanced multimodal reasoning across single-image, multi-image, and video tasks.

Conclusion: Qwen3-VL represents a significant advancement in vision-language models with its 256K context window, architectural improvements, and strong performance across diverse benchmarks, positioning it as a foundational engine for image-grounded reasoning, agentic decision-making, and multimodal code intelligence.

Abstract: We introduce Qwen3-VL, the most capable vision-language model in the Qwen series to date, achieving superior performance across a broad range of multimodal benchmarks. It natively supports interleaved contexts of up to 256K tokens, seamlessly integrating text, images, and video. The model family includes both dense (2B/4B/8B/32B) and mixture-of-experts (30B-A3B/235B-A22B) variants to accommodate diverse latency-quality trade-offs. Qwen3-VL delivers three core pillars: (i) markedly stronger pure-text understanding, surpassing comparable text-only backbones in several cases; (ii) robust long-context comprehension with a native 256K-token window for both text and interleaved multimodal inputs, enabling faithful retention, retrieval, and cross-referencing across long documents and videos; and (iii) advanced multimodal reasoning across single-image, multi-image, and video tasks, demonstrating leading performance on comprehensive evaluations such as MMMU and visual-math benchmarks (e.g., MathVista and MathVision). Architecturally, we introduce three key upgrades: (i) an enhanced interleaved-MRoPE for stronger spatial-temporal modeling across images and video; (ii) DeepStack integration, which effectively leverages multi-level ViT features to tighten vision-language alignment; and (iii) text-based time alignment for video, evolving from T-RoPE to explicit textual timestamp alignment for more precise temporal grounding. Under comparable token budgets and latency constraints, Qwen3-VL achieves superior performance in both dense and Mixture-of-Experts (MoE) architectures. We envision Qwen3-VL serving as a foundational engine for image-grounded reasoning, agentic decision-making, and multimodal code intelligence in real-world workflows.

Method: TRACE models sequential 2D slices as video frame pairs using multimodal conditioning (segmentation priors + radiology reports), incorporates optical flow for temporal coherence, and uses an overlapping-frame strategy to link frame pairs into flexible-length sequences.

Result: TRACE effectively balances computational efficiency with preserving anatomical fidelity and spatiotemporal consistency in 3D medical image generation.

Conclusion: TRACE provides a practical solution for 3D medical image generation that addresses computational efficiency while maintaining anatomical accuracy, making it suitable for clinical practice in resource-limited settings.

Abstract: 3D medical image generation is essential for data augmentation and patient privacy, calling for reliable and efficient models suited for clinical practice. However, current methods suffer from limited anatomical fidelity, restricted axial length, and substantial computational cost, placing them beyond reach for regions with limited resources and infrastructure. We introduce TRACE, a framework that generates 3D medical images with spatiotemporal alignment using a 2D multimodal-conditioned diffusion approach. TRACE models sequential 2D slices as video frame pairs, combining segmentation priors and radiology reports for anatomical alignment, incorporating optical flow to sustain temporal coherence. During inference, an overlapping-frame strategy links frame pairs into a flexible length sequence, reconstructed into a spatiotemporally and anatomically aligned 3D volume. Experimental results demonstrate that TRACE effectively balances computational efficiency with preserving anatomical fidelity and spatiotemporal consistency. Code is available at: https://github.com/VinyehShaw/TRACE.

Method: Created MFT25 dataset with 15 diverse underwater videos and 408,578 annotated bounding boxes. Developed SU-T tracker featuring Unscented Kalman Filter for non-linear fish motion and Fish-Intersection-over-Union matching for fish morphology.

Result: SU-T achieves state-of-the-art performance on MFT25 with 34.1 HOTA and 44.6 IDF1 scores, revealing fundamental differences between fish and terrestrial tracking.

Conclusion: The paper provides a comprehensive underwater fish tracking dataset and specialized tracking framework that addresses unique challenges of aquatic environments, advancing underwater MOT research.

Abstract: Multiple object tracking (MOT) technology has made significant progress in terrestrial applications, but underwater tracking scenarios remain underexplored despite their importance to marine ecology and aquaculture. In this paper, we present Multiple Fish Tracking Dataset 2025 (MFT25), a comprehensive dataset specifically designed for underwater multiple fish tracking, featuring 15 diverse video sequences with 408,578 meticulously annotated bounding boxes across 48,066 frames. Our dataset captures various underwater environments, fish species, and challenging conditions including occlusions, similar appearances, and erratic motion patterns. Additionally, we introduce Scale-aware and Unscented Tracker (SU-T), a specialized tracking framework featuring an Unscented Kalman Filter (UKF) optimized for non-linear swimming patterns of fish and a novel Fish-Intersection-over-Union (FishIoU) matching that accounts for the unique morphological characteristics of aquatic species. Extensive experiments demonstrate that our SU-T baseline achieves state-of-the-art performance on MFT25, with 34.1 HOTA and 44.6 IDF1, while revealing fundamental differences between fish tracking and terrestrial object tracking scenarios. The dataset and codes are released at https://vranlee.github.io/SU-T/.

Method: KL-tracing: inject localized perturbation into first frame, roll out model one step, compute KL divergence between perturbed/unperturbed predictive distributions. Requires models with distributional prediction, factorized latents, and random-access decoding (found in LRAS architecture).

Result: Competitive with state-of-the-art task-specific models on TAP-Vid DAVIS (real-world) and TAP-Vid Kubric (synthetic) benchmarks without any flow-specific fine-tuning.

Conclusion: Counterfactual prompting of controllable generative video models is an effective alternative to supervised or photometric-loss methods for high-quality optical flow extraction.

Abstract: Extracting optical flow from videos remains a core computer vision problem. Motivated by the recent success of large general-purpose models, we ask whether frozen self-supervised video models trained only to predict future frames can be prompted, without fine-tuning, to output flow. Prior attempts to read out depth or illumination from video generators required fine-tuning; that strategy is ill-suited for flow, where labeled data is scarce and synthetic datasets suffer from a sim-to-real gap. Inspired by the Counterfactual World Model (CWM) paradigm, which can obtain point-wise correspondences by injecting a small tracer perturbation into a next-frame predictor and tracking its propagation, we extend this idea to generative video models for zero-shot flow extraction. We explore several popular architectures and find that successful zero-shot flow extraction in this manner is aided by three model properties: (1) distributional prediction of future frames (avoiding blurry or noisy outputs); (2) factorized latents that treat each spatio-temporal patch independently; and (3) random-access decoding that can condition on any subset of future pixels. These properties are uniquely present in the recently introduced Local Random Access Sequence (LRAS) architecture. Building on LRAS, we propose KL-tracing: a novel test-time inference procedure that injects a localized perturbation into the first frame, rolls out the model one step, and computes the Kullback-Leibler divergence between perturbed and unperturbed predictive distributions. Without any flow-specific fine-tuning, our method is competitive with state-of-the-art, task-specific models on the real-world TAP-Vid DAVIS benchmark and the synthetic TAP-Vid Kubric. Our results show that counterfactual prompting of controllable generative video models is an effective alternative to supervised or photometric-loss methods for high-quality flow.

Method: 1) PhysXNet: First physics-grounded 3D dataset annotated across five dimensions (scale, material, affordance, kinematics, function) using scalable human-in-the-loop pipeline with vision-language models. 2) PhysXGen: Feed-forward image-to-3D framework with dual-branch architecture that models correlations between 3D structures and physical properties while preserving geometry quality.

Result: Extensive experiments show superior performance and promising generalization capability. The framework produces 3D assets with plausible physical predictions while maintaining native geometry quality.

Conclusion: PhysX-3D addresses the critical gap in physics-grounded 3D generation, enabling real-world applications in physical domains. All code, data, and models will be released to facilitate future research in generative physical AI.

Abstract: 3D modeling is moving from virtual to physical. Existing 3D generation primarily emphasizes geometries and textures while neglecting physical-grounded modeling. Consequently, despite the rapid development of 3D generative models, the synthesized 3D assets often overlook rich and important physical properties, hampering their real-world application in physical domains like simulation and embodied AI. As an initial attempt to address this challenge, we propose \textbf{PhysX-3D}, an end-to-end paradigm for physical-grounded 3D asset generation. 1) To bridge the critical gap in physics-annotated 3D datasets, we present PhysXNet - the first physics-grounded 3D dataset systematically annotated across five foundational dimensions: absolute scale, material, affordance, kinematics, and function description. In particular, we devise a scalable human-in-the-loop annotation pipeline based on vision-language models, which enables efficient creation of physics-first assets from raw 3D assets.2) Furthermore, we propose \textbf{PhysXGen}, a feed-forward framework for physics-grounded image-to-3D asset generation, injecting physical knowledge into the pre-trained 3D structural space. Specifically, PhysXGen employs a dual-branch architecture to explicitly model the latent correlations between 3D structures and physical properties, thereby producing 3D assets with plausible physical predictions while preserving the native geometry quality. Extensive experiments validate the superior performance and promising generalization capability of our framework. All the code, data, and models will be released to facilitate future research in generative physical AI.

Method: Uses Residual Compression that transmits only compressed step-wise activation differences (residuals) based on observation of temporal redundancy in diffusion activations, plus lightweight error feedback to prevent accumulation.

Result: Achieves 3.0x speedup on 4xL20 while improving fidelity, and 6.7x speedup over prior methods on slow networks; enables communication-heavy strategies like sequence parallelism.

Conclusion: CompactFusion establishes a new paradigm for parallel diffusion inference that reduces communication overhead while preserving quality, works across models and parallel settings, and integrates easily without pipeline rework.

Abstract: Diffusion models produce realistic images and videos but require substantial computational resources, necessitating multi-accelerator parallelism for real-time deployment. However, parallel inference introduces significant communication overhead from exchanging large activations between devices, limiting efficiency and scalability. We present CompactFusion, a compression framework that significantly reduces communication while preserving generation quality. Our key observation is that diffusion activations exhibit strong temporal redundancy-adjacent steps produce highly similar activations, saturating bandwidth with near-duplicate data carrying little new information. To address this inefficiency, we seek a more compact representation that encodes only the essential information. CompactFusion achieves this via Residual Compression that transmits only compressed residuals (step-wise activation differences). Based on empirical analysis and theoretical justification, we show that it effectively removes redundant data, enabling substantial data reduction while maintaining high fidelity. We also integrate lightweight error feedback to prevent error accumulation. CompactFusion establishes a new paradigm for parallel diffusion inference, delivering lower latency and significantly higher generation quality than prior methods. On 4xL20, it achieves 3.0x speedup while greatly improving fidelity. It also uniquely supports communication-heavy strategies like sequence parallelism on slow networks, achieving 6.7x speedup over prior overlap-based method. CompactFusion applies broadly across diffusion models and parallel settings, and integrates easily without requiring pipeline rework. Portable implementation demonstrated on xDiT is publicly available at https://github.com/Cobalt-27/CompactFusion

Method: Proposes Motion-guided Modulation Network (MMN) with: 1) Motion-guided Skeletal Modulation (MSM) to inject motion cues at skeletal level for spatial representation, 2) Motion-guided Temporal Modulation (MTM) to incorporate motion at frame level for holistic motion patterns, and 3) motion consistency learning to aggregate multi-scale motion features.

Result: Achieves state-of-the-art performance on Micro-Action 52 and iMiGUE datasets for skeleton-based micro-action recognition, demonstrating effectiveness of explicitly modeling subtle motion cues.

Conclusion: Explicit modeling of subtle motion cues is crucial for accurate micro-action recognition, and the proposed MMN framework effectively captures and modulates these cues to enhance spatial-temporal representation learning.

Abstract: Micro-Actions (MAs) are an important form of non-verbal communication in social interactions, with potential applications in human emotional analysis. However, existing methods in Micro-Action Recognition often overlook the inherent subtle changes in MAs, which limits the accuracy of distinguishing MAs with subtle changes. To address this issue, we present a novel Motion-guided Modulation Network (MMN) that implicitly captures and modulates subtle motion cues to enhance spatial-temporal representation learning. Specifically, we introduce a Motion-guided Skeletal Modulation module (MSM) to inject motion cues at the skeletal level, acting as a control signal to guide spatial representation modeling. In parallel, we design a Motion-guided Temporal Modulation module (MTM) to incorporate motion information at the frame level, facilitating the modeling of holistic motion patterns in micro-actions. Finally, we propose a motion consistency learning strategy to aggregate the motion cues from multi-scale features for micro-action classification. Experimental results on the Micro-Action 52 and iMiGUE datasets demonstrate that MMN achieves state-of-the-art performance in skeleton-based micro-action recognition, underscoring the importance of explicitly modeling subtle motion cues. The code will be available at https://github.com/momiji-bit/MMN.

Method: Proposes Random Slot-Feature pair for learning Query prediction (RandSF.Q): (1) designs a new transitioner that incorporates both slots and features for more informative query prediction, (2) trains the transitioner to predict queries from slot-feature pairs randomly sampled from available recurrences to learn transition dynamics.

Result: Significantly surpasses existing video OCL methods, achieving up to 10 points improvement on object discovery metrics, setting new state-of-the-art. The superiority also benefits downstream tasks like scene understanding.

Conclusion: RandSF.Q effectively addresses the limitations of existing video OCL methods by incorporating next frame features and learning transition dynamics through random slot-feature pair sampling, leading to superior object discovery and scene representation performance.

Abstract: Unsupervised video Object-Centric Learning (OCL) is promising as it enables object-level scene representation and understanding as we humans do. Mainstream video OCL methods adopt a recurrent architecture: An aggregator aggregates current video frame into object features, termed slots, under some queries; A transitioner transits current slots to queries for the next frame. This is an effective architecture but all existing implementations both (\textit{i1}) neglect to incorporate next frame features, the most informative source for query prediction, and (\textit{i2}) fail to learn transition dynamics, the knowledge essential for query prediction. To address these issues, we propose Random Slot-Feature pair for learning Query prediction (RandSF.Q): (\textit{t1}) We design a new transitioner to incorporate both slots and features, which provides more information for query prediction; (\textit{t2}) We train the transitioner to predict queries from slot-feature pairs randomly sampled from available recurrences, which drives it to learn transition dynamics. Experiments on scene representation demonstrate that our method surpass existing video OCL methods significantly, e.g., up to 10 points on object discovery, setting new state-of-the-art. Such superiority also benefits downstream tasks like scene understanding. Source Code, Model Checkpoints, Training Logs: https://github.com/Genera1Z/RandSF.Q

Method: Proposes AMOT with two key components: 1) Appearance-Motion Consistency (AMC) matrix that computes bi-directional spatial consistency under appearance feature guidance for reliable identity association, and 2) Motion-aware Track Continuation (MTC) module that reactivates unmatched tracks through appearance-guided predictions aligned with Kalman-based predictions to reduce broken trajectories from missed detections.

Result: Extensive experiments on three UAV benchmarks (VisDrone2019, UAVDT, and VT-MOT-UAV) demonstrate that AMOT outperforms current state-of-the-art methods and generalizes well in a plug-and-play and training-free manner.

Conclusion: AMOT effectively addresses UAV tracking challenges by jointly modeling appearance and motion cues, providing a robust solution for multi-object tracking in UAV videos with superior performance and generalization capability.

Abstract: Multi-object tracking (MOT) aims to track multiple objects while maintaining consistent identities across frames of a given video. In unmanned aerial vehicle (UAV) recorded videos, frequent viewpoint changes and complex UAV-ground relative motion dynamics pose significant challenges, which often lead to unstable affinity measurement and ambiguous association. Existing methods typically model motion and appearance cues separately, overlooking their spatio-temporal interplay and resulting in suboptimal tracking performance. In this work, we propose AMOT, which jointly exploits appearance and motion cues through two key components: an Appearance-Motion Consistency (AMC) matrix and a Motion-aware Track Continuation (MTC) module. Specifically, the AMC matrix computes bi-directional spatial consistency under the guidance of appearance features, enabling more reliable and context-aware identity association. The MTC module complements AMC by reactivating unmatched tracks through appearance-guided predictions that align with Kalman-based predictions, thereby reducing broken trajectories caused by missed detections. Extensive experiments on three UAV benchmarks, including VisDrone2019, UAVDT, and VT-MOT-UAV, demonstrate that our AMOT outperforms current state-of-the-art methods and generalizes well in a plug-and-play and training-free manner.

Method: Introduces DiffusionFF with a novel encoder-decoder architecture: a pretrained forgery detector serves as an “artifact encoder” to extract multi-scale forgery-related features, and a denoising diffusion model is repurposed as an “artifact decoder” to progressively synthesize detailed artifact localization maps. The fine-grained localization map is then fused with high-level semantic features from the forgery detector.

Result: Extensive experiments show that DiffusionFF achieves state-of-the-art (SOTA) performance across multiple benchmarks, demonstrating superior effectiveness and explainability.

Conclusion: The proposed DiffusionFF framework successfully addresses the dual challenge of face forgery detection and fine-grained artifact localization, offering both high detection capability and enhanced explainability through its innovative diffusion-based approach.

Abstract: The rapid evolution of deepfake technologies demands robust and reliable face forgery detection algorithms. While determining whether an image has been manipulated remains essential, the ability to precisely localize forgery clues is also important for enhancing model explainability and building user trust. To address this dual challenge, we introduce DiffusionFF, a diffusion-based framework that simultaneously performs face forgery detection and fine-grained artifact localization. Our key idea is to establish a novel encoder-decoder architecture: a pretrained forgery detector serves as a powerful “artifact encoder”, and a denoising diffusion model is repurposed as an “artifact decoder”. Conditioned on multi-scale forgery-related features extracted by the encoder, the decoder progressively synthesizes a detailed artifact localization map. We then fuse this fine-grained localization map with high-level semantic features from the forgery detector, leading to substantial improvements in detection capability. Extensive experiments show that DiffusionFF achieves state-of-the-art (SOTA) performance across multiple benchmarks, underscoring its superior effectiveness and explainability.

Method: MonoDream learns a Unified Navigation Representation that jointly aligns navigation-relevant visual semantics (global layout, depth, future cues) with language-grounded action intent. It introduces Latent Panoramic Dreaming tasks to supervise this representation by predicting latent features of panoramic RGB and depth observations from monocular input.

Result: Experiments on multiple VLN benchmarks show MonoDream consistently improves monocular navigation performance and significantly narrows the gap with panoramic-based agents.

Conclusion: MonoDream demonstrates that lightweight VLA frameworks can effectively learn unified representations from monocular input, making VLN more practical for real-world deployment while maintaining competitive performance with panoramic sensor-based methods.

Abstract: Vision-Language Navigation (VLN) tasks often leverage panoramic RGB and depth inputs to provide rich spatial cues for action planning, but these sensors can be costly or less accessible in real-world deployments. Recent approaches based on Vision-Language Action (VLA) models achieve strong results with monocular input, yet they still lag behind methods using panoramic RGB-D information. We present MonoDream, a lightweight VLA framework that enables monocular agents to learn a Unified Navigation Representation (UNR). This shared feature representation jointly aligns navigation-relevant visual semantics (e.g., global layout, depth, and future cues) and language-grounded action intent, enabling more reliable action prediction. MonoDream further introduces Latent Panoramic Dreaming (LPD) tasks to supervise the UNR, which train the model to predict latent features of panoramic RGB and depth observations at both current and future steps based on only monocular input. Experiments on multiple VLN benchmarks show that MonoDream consistently improves monocular navigation performance and significantly narrows the gap with panoramic-based agents.

Method: SARD introduces two key components: 1) Region-Constrained Diffusion (RCD) that freezes background and selectively updates only foreground anomaly regions during reverse denoising, and 2) Discriminative Mask Guidance (DMG) module in the discriminator for joint evaluation of global realism and local anomaly fidelity using pixel-level masks.

Result: Extensive experiments on MVTec-AD and BTAD datasets show SARD surpasses existing methods in segmentation accuracy and visual quality, setting new state-of-the-art for pixel-level anomaly synthesis.

Conclusion: SARD effectively addresses spatial controllability and regional fidelity limitations in diffusion-based anomaly synthesis, providing a robust framework for enhancing industrial anomaly detection systems through realistic and precise anomaly generation.

Abstract: Synthesizing realistic and spatially precise anomalies is essential for enhancing the robustness of industrial anomaly detection systems. While recent diffusion-based methods have demonstrated strong capabilities in modeling complex defect patterns, they often struggle with spatial controllability and fail to maintain fine-grained regional fidelity. To overcome these limitations, we propose SARD (Segmentation-Aware anomaly synthesis via Region-constrained Diffusion with discriminative mask Guidance), a novel diffusion-based framework specifically designed for anomaly generation. Our approach introduces a Region-Constrained Diffusion (RCD) process that preserves the background by freezing it and selectively updating only the foreground anomaly regions during the reverse denoising phase, thereby effectively reducing background artifacts. Additionally, we incorporate a Discriminative Mask Guidance (DMG) module into the discriminator, enabling joint evaluation of both global realism and local anomaly fidelity, guided by pixel-level masks. Extensive experiments on the MVTec-AD and BTAD datasets show that SARD surpasses existing methods in segmentation accuracy and visual quality, setting a new state-of-the-art for pixel-level anomaly synthesis.

Method: Proposes a stream-based event-to-event super-resolution method using Spiking Neural Networks (SNNs). Introduces Dual-Forward Polarity-Split Event Encoding that decouples positive/negative events into separate forward paths through a shared SNN. Also proposes Learnable Spatio-temporal Polarity-aware Loss (LearnSTPLoss) with adaptive uncertainty-based weights for balancing temporal, spatial, and polarity consistency.

Result: Achieves competitive super-resolution performance on multiple datasets while significantly reducing model size and inference time. The lightweight design enables embedding into event cameras or use as efficient front-end preprocessing for downstream vision tasks.

Conclusion: The proposed ultra-lightweight SNN-based super-resolution method effectively addresses event camera resolution limitations while maintaining real-time performance suitable for resource-constrained devices, making it practical for embedded applications.

Abstract: Event cameras offer unparalleled advantages such as high temporal resolution, low latency, and high dynamic range. However, their limited spatial resolution poses challenges for fine-grained perception tasks. In this work, we propose an ultra-lightweight, stream-based event-to-event super-resolution method based on Spiking Neural Networks (SNNs), designed for real-time deployment on resource-constrained devices. To further reduce model size, we introduce a novel Dual-Forward Polarity-Split Event Encoding strategy that decouples positive and negative events into separate forward paths through a shared SNN. Furthermore, we propose a Learnable Spatio-temporal Polarity-aware Loss (LearnSTPLoss) that adaptively balances temporal, spatial, and polarity consistency using learnable uncertainty-based weights. Experimental results demonstrate that our method achieves competitive super-resolution performance on multiple datasets while significantly reducing model size and inference time. The lightweight design enables embedding the module into event cameras or using it as an efficient front-end preprocessing for downstream vision tasks.

Method: Three-stage approach: 1) Use visual geometry grounded transformer to obtain camera parameters and dense 3D reconstruction, 2) Reproject point cloud to reference view for pixel-wise alignment, 3) Apply point-conditioned image diffusion module to refine artifacts like holes or noise.

Result: Experimental results show PIS3R provides accurate stitching for images with very large parallax, outperforming existing methods both qualitatively and quantitatively while preserving geometric integrity for downstream 3D tasks like SfM.

Conclusion: PIS3R successfully addresses the challenge of stitching images with very large parallax through deep 3D reconstruction, offering a robust solution that maintains geometric accuracy and enables direct application to 3D vision tasks.

Abstract: Image stitching aim to align two images taken from different viewpoints into one seamless, wider image. However, when the 3D scene contains depth variations and the camera baseline is significant, noticeable parallax occurs-meaning the relative positions of scene elements differ substantially between views. Most existing stitching methods struggle to handle such images with large parallax effectively. To address this challenge, in this paper, we propose an image stitching solution called PIS3R that is robust to very large parallax based on the novel concept of deep 3D reconstruction. First, we apply visual geometry grounded transformer to two input images with very large parallax to obtain both intrinsic and extrinsic parameters, as well as the dense 3D scene reconstruction. Subsequently, we reproject reconstructed dense point cloud onto a designated reference view using the recovered camera parameters, achieving pixel-wise alignment and generating an initial stitched image. Finally, to further address potential artifacts such as holes or noise in the initial stitching, we propose a point-conditioned image diffusion module to obtain the refined result.Compared with existing methods, our solution is very large parallax tolerant and also provides results that fully preserve the geometric integrity of all pixels in the 3D photogrammetric context, enabling direct applicability to downstream 3D vision tasks such as SfM. Experimental results demonstrate that the proposed algorithm provides accurate stitching results for images with very large parallax, and outperforms the existing methods qualitatively and quantitatively.

Method: Proposes TEFormer with three key components: 1) Texture-aware module (TaM) in encoder to capture fine-grained texture distinctions, 2) Edge-guided tri-branch decoder (Eg3Head) to preserve local edges while maintaining multiscale context, and 3) Edge-guided feature fusion module (EgFFM) to integrate contextual, detail, and edge information.

Result: Achieves mIoU scores of 88.57% on Potsdam (exceeding next best by 0.73%), 81.46% on Vaihingen (exceeding next best by 0.22%), and 53.55% on LoveDA (second position, trailing optimal by only 0.19%).

Conclusion: TEFormer effectively addresses semantic ambiguity in urban remote sensing images by combining texture awareness and edge guidance, demonstrating superior performance on benchmark datasets and showing strong potential for urban planning and environmental monitoring applications.

Abstract: Accurate semantic segmentation of urban remote sensing images (URSIs) is essential for urban planning and environmental monitoring. However, it remains challenging due to the subtle texture differences and similar spatial structures among geospatial objects, which cause semantic ambiguity and misclassification. Additional complexities arise from irregular object shapes, blurred boundaries, and overlapping spatial distributions of objects, resulting in diverse and intricate edge morphologies. To address these issues, we propose TEFormer, a texture-aware and edge-guided Transformer. Our model features a texture-aware module (TaM) in the encoder to capture fine-grained texture distinctions between visually similar categories, thereby enhancing semantic discrimination. The decoder incorporates an edge-guided tri-branch decoder (Eg3Head) to preserve local edges and details while maintaining multiscale context-awareness. Finally, an edge-guided feature fusion module (EgFFM) effectively integrates contextual, detail, and edge information to achieve refined semantic segmentation. Extensive evaluation demonstrates that TEFormer yields mIoU scores of 88.57% on Potsdam and 81.46% on Vaihingen, exceeding the next best methods by 0.73% and 0.22%. On the LoveDA dataset, it secures the second position with an overall mIoU of 53.55%, trailing the optimal performance by a narrow margin of 0.19%.

Method: Two-stage architecture: 1) Predicts optimal camera trajectories from a single image using few-shot learning, 2) Generates geometrically consistent multi-view sequences via pose-conditioned diffusion process, 3) Converts these into persistent 4D representation by leveraging both temporal priors from video diffusion models and geometric awareness of reconstruction models.

Result: The framework shows higher quality metrics (mPSNR, mSSIM) over existing methods and is the first to successfully combine rich temporal priors from video diffusion models with geometric awareness for 4D generation.

Conclusion: Dream4D bridges the gap between controllable video generation and neural 4D reconstruction, enabling high-quality spatiotemporally coherent 4D content synthesis from single images.

Abstract: The synthesis of spatiotemporally coherent 4D content presents fundamental challenges in computer vision, requiring simultaneous modeling of high-fidelity spatial representations and physically plausible temporal dynamics. Current approaches often struggle to maintain view consistency while handling complex scene dynamics, particularly in large-scale environments with multiple interacting elements. This work introduces Dream4D, a novel framework that bridges this gap through a synergy of controllable video generation and neural 4D reconstruction. Our approach seamlessly combines a two-stage architecture: it first predicts optimal camera trajectories from a single image using few-shot learning, then generates geometrically consistent multi-view sequences via a specialized pose-conditioned diffusion process, which are finally converted into a persistent 4D representation. This framework is the first to leverage both rich temporal priors from video diffusion models and geometric awareness of the reconstruction models, which significantly facilitates 4D generation and shows higher quality (e.g., mPSNR, mSSIM) over existing methods.

Method: LD-ViCE operates in latent space using a state-of-the-art diffusion model to reduce computational costs, with an additional refinement step to produce realistic and interpretable counterfactuals. It generates explanations by modifying video inputs to show what changes would alter the model’s decision.

Result: Experiments on EchoNet-Dynamic (cardiac ultrasound), FERV39k (facial expression), and Something-Something V2 (action recognition) datasets show LD-ViCE generalizes well and achieves state-of-the-art performance. On EchoNet-Dynamic, it achieves significantly higher regression accuracy than prior methods with high temporal consistency, while refinement improves perceptual quality.

Conclusion: LD-ViCE advances the trustworthiness and interpretability of video-based AI systems through visually coherent counterfactual explanations that provide semantically meaningful and temporally coherent insights into model behavior.

Abstract: Video-based AI systems are increasingly adopted in safety-critical domains such as autonomous driving and healthcare. However, interpreting their decisions remains challenging due to the inherent spatiotemporal complexity of video data and the opacity of deep learning models. Existing explanation techniques often suffer from limited temporal coherence and a lack of actionable causal insights. Current counterfactual explanation methods typically do not incorporate guidance from the target model, reducing semantic fidelity and practical utility. We introduce Latent Diffusion for Video Counterfactual Explanations (LD-ViCE), a novel framework designed to explain the behavior of video-based AI models. Compared to previous approaches, LD-ViCE reduces the computational costs of generating explanations by operating in latent space using a state-of-the-art diffusion model, while producing realistic and interpretable counterfactuals through an additional refinement step. Experiments on three diverse video datasets - EchoNet-Dynamic (cardiac ultrasound), FERV39k (facial expression), and Something-Something V2 (action recognition) with multiple target models covering both classification and regression tasks, demonstrate that LD-ViCE generalizes well and achieves state-of-the-art performance. On the EchoNet-Dynamic dataset, LD-ViCE achieves significantly higher regression accuracy than prior methods and exhibits high temporal consistency, while the refinement stage further improves perceptual quality. Qualitative analyses confirm that LD-ViCE produces semantically meaningful and temporally coherent explanations, providing actionable insights into model behavior. LD-ViCE advances the trustworthiness and interpretability of video-based AI systems through visually coherent counterfactual explanations.

Method: Uses Transformer encoder-decoder with point clouds and associated 2D images. Encoder enriches 3D points with 2D image features, then fuses 3D context. Decoder uses 3D anchor boxes as queries and performs cross-attention to 2D object queries from pre-trained 2D model, avoiding memory issues while preserving 2D knowledge.

Result: Achieves state-of-the-art on ScanNetV2 and ScanNet200 benchmarks. On ScanNet200, outperforms prior methods by +8.6 mAP on validation set and +6.8 mAP on hidden test set.

Conclusion: SegDINO3D effectively leverages 2D representations to overcome 3D data scarcity, demonstrating superior 3D instance segmentation performance through efficient cross-modal fusion and memory-efficient 2D query utilization.

Abstract: In this paper, we present SegDINO3D, a novel Transformer encoder-decoder framework for 3D instance segmentation. As 3D training data is generally not as sufficient as 2D training images, SegDINO3D is designed to fully leverage 2D representation from a pre-trained 2D detection model, including both image-level and object-level features, for improving 3D representation. SegDINO3D takes both a point cloud and its associated 2D images as input. In the encoder stage, it first enriches each 3D point by retrieving 2D image features from its corresponding image views and then leverages a 3D encoder for 3D context fusion. In the decoder stage, it formulates 3D object queries as 3D anchor boxes and performs cross-attention from 3D queries to 2D object queries obtained from 2D images using the 2D detection model. These 2D object queries serve as a compact object-level representation of 2D images, effectively avoiding the challenge of keeping thousands of image feature maps in the memory while faithfully preserving the knowledge of the pre-trained 2D model. The introducing of 3D box queries also enables the model to modulate cross-attention using the predicted boxes for more precise querying. SegDINO3D achieves the state-of-the-art performance on the ScanNetV2 and ScanNet200 3D instance segmentation benchmarks. Notably, on the challenging ScanNet200 dataset, SegDINO3D significantly outperforms prior methods by +8.6 and +6.8 mAP on the validation and hidden test sets, respectively, demonstrating its superiority.

Method: Proposes FAST with two modules: 1) AIAS - training-free sampling algorithm using coarse-to-fine aggregation for accelerated reverse process, 2) FARM - adaptively adjusts anomaly-aware noise in masked foreground regions during denoising to preserve localized anomaly signals.

Result: Extensive experiments on multiple industrial benchmarks show FAST consistently outperforms existing anomaly synthesis methods in downstream segmentation tasks, achieving state-of-the-art results with only 10 sampling steps.

Conclusion: FAST effectively addresses the limitations of existing methods by providing efficient, high-quality anomaly synthesis with foreground awareness, making it a practical solution for industrial anomaly segmentation where labeled data is scarce.

Abstract: Industrial anomaly segmentation relies heavily on pixel-level annotations, yet real-world anomalies are often scarce, diverse, and costly to label. Segmentation-oriented industrial anomaly synthesis (SIAS) has emerged as a promising alternative; however, existing methods struggle to balance sampling efficiency and generation quality. Moreover, most approaches treat all spatial regions uniformly, overlooking the distinct statistical differences between anomaly and background areas. This uniform treatment hinders the synthesis of controllable, structure-specific anomalies tailored for segmentation tasks. In this paper, we propose FAST, a foreground-aware diffusion framework featuring two novel modules: the Anomaly-Informed Accelerated Sampling (AIAS) and the Foreground-Aware Reconstruction Module (FARM). AIAS is a training-free sampling algorithm specifically designed for segmentation-oriented industrial anomaly synthesis, which accelerates the reverse process through coarse-to-fine aggregation and enables the synthesis of state-of-the-art segmentation-oriented anomalies in as few as 10 steps. Meanwhile, FARM adaptively adjusts the anomaly-aware noise within the masked foreground regions at each sampling step, preserving localized anomaly signals throughout the denoising trajectory. Extensive experiments on multiple industrial benchmarks demonstrate that FAST consistently outperforms existing anomaly synthesis methods in downstream segmentation tasks. We release the code at: https://github.com/Chhro123/fast-foreground-aware-anomaly-synthesis.

Method: Three key innovations: (1) One-Step Inversion-and-Editing (OSIE) pipeline bypasses iterative processes; (2) Background Shield (BG-Shield) selectively modifies only edit region features; (3) Sparsified Spatial Cross-Attention (SSCA) suppresses semantic leakage to background.

Result: FlashEdit maintains superior background consistency and structural integrity while performing edits in under 0.2 seconds, achieving over 150× speedup compared to prior multi-step methods.

Conclusion: FlashEdit enables high-fidelity, real-time image editing with diffusion models, making text-guided image editing practical for real-world applications through significant efficiency improvements.

Abstract: Text-guided image editing with diffusion models has achieved remarkable quality but suffers from prohibitive latency, hindering real-world applications. We introduce FlashEdit, a novel framework designed to enable high-fidelity, real-time image editing. Its efficiency stems from three key innovations: (1) a One-Step Inversion-and-Editing (OSIE) pipeline that bypasses costly iterative processes; (2) a Background Shield (BG-Shield) technique that guarantees background preservation by selectively modifying features only within the edit region; and (3) a Sparsified Spatial Cross-Attention (SSCA) mechanism that ensures precise, localized edits by suppressing semantic leakage to the background. Extensive experiments demonstrate that FlashEdit maintains superior background consistency and structural integrity, while performing edits in under 0.2 seconds, which is an over 150$\times$ speedup compared to prior multi-step methods. Our code will be made publicly available at https://github.com/JunyiWuCode/FlashEdit.

Method: ControlEvents leverages diffusion priors from foundation models like Stable Diffusion to generate high-quality event data guided by diverse control signals (class text labels, 2D skeletons, 3D body poses). The approach requires minimal fine-tuning and limited labeled data.

Result: The method successfully synthesizes event data for visual recognition, 2D skeleton estimation, and 3D body pose estimation. Experiments show synthesized data enhances model performance in all tasks. The approach can also generate events based on unseen text labels during training.

Conclusion: ControlEvents provides an effective solution for generating high-quality labeled event data, significantly reducing dataset creation costs while maintaining strong performance across multiple event-based vision tasks, with inherited text-based generation capabilities from foundation models.

Abstract: In recent years, event cameras have gained significant attention due to their bio-inspired properties, such as high temporal resolution and high dynamic range. However, obtaining large-scale labeled ground-truth data for event-based vision tasks remains challenging and costly. In this paper, we present ControlEvents, a diffusion-based generative model designed to synthesize high-quality event data guided by diverse control signals such as class text labels, 2D skeletons, and 3D body poses. Our key insight is to leverage the diffusion prior from foundation models, such as Stable Diffusion, enabling high-quality event data generation with minimal fine-tuning and limited labeled data. Our method streamlines the data generation process and significantly reduces the cost of producing labeled event datasets. We demonstrate the effectiveness of our approach by synthesizing event data for visual recognition, 2D skeleton estimation, and 3D body pose estimation. Our experiments show that the synthesized labeled event data enhances model performance in all tasks. Additionally, our approach can generate events based on unseen text labels during training, illustrating the powerful text-based generation capabilities inherited from foundation models.

Method: Generate 3D pseudo-ground-truth labels using foundation models (Grounded-SAM for semantic segmentation and Metric3Dv2 for depth estimation), then use temporal information for label densification. These labels can be integrated into existing models or used to train a streamlined model called EasyOcc that learns solely from the pseudo-labels without complex rendering.

Result: When integrated into OccNeRF, mIoU increased by 45% (from 9.73 to 14.09). The proposed EasyOcc model achieves 13.86 mIoU. On full scene evaluation without camera mask, EasyOcc achieves 7.71 mIoU, outperforming previous best by 31%. The method shows strong transferability across architectures.

Conclusion: Foundation models, temporal context, and appropriate loss computation space are critical for effective self-supervised learning in comprehensive scene understanding. The proposed pseudo-label approach provides a computationally efficient and architecture-transferable solution that outperforms complex rendering-based methods.

Abstract: Self-supervised models have recently achieved notable advancements, particularly in the domain of semantic occupancy prediction. These models utilize sophisticated loss computation strategies to compensate for the absence of ground-truth labels. For instance, techniques such as novel view synthesis, cross-view rendering, and depth estimation have been explored to address the issue of semantic and depth ambiguity. However, such techniques typically incur high computational costs and memory usage during the training stage, especially in the case of novel view synthesis. To mitigate these issues, we propose 3D pseudo-ground-truth labels generated by the foundation models Grounded-SAM and Metric3Dv2, and harness temporal information for label densification. Our 3D pseudo-labels can be easily integrated into existing models, which yields substantial performance improvements, with mIoU increasing by 45%, from 9.73 to 14.09, when implemented into the OccNeRF model. This stands in contrast to earlier advancements in the field, which are often not readily transferable to other architectures. Additionally, we propose a streamlined model, EasyOcc, achieving 13.86 mIoU. This model conducts learning solely from our labels, avoiding complex rendering strategies mentioned previously. Furthermore, our method enables models to attain state-of-the-art performance when evaluated on the full scene without applying the camera mask, with EasyOcc achieving 7.71 mIoU, outperforming the previous best model by 31%. These findings highlight the critical importance of foundation models, temporal context, and the choice of loss computation space in self-supervised learning for comprehensive scene understanding.

Method: Uses feature matching between frames to establish correspondences, then applies graph-based optimization with local planar approximation using estimated 3D points and surface normals from monocular foundation models, rather than simple least squares.

Result: Improves cross-view consistency, achieves scale alignment, enhances 3D reconstruction quality, and improves novel view synthesis, especially in sparse view rendering scenarios.

Conclusion: MoRe provides an effective training-free solution for refining monocular geometric priors, addressing scale ambiguity while preserving 3D structure, making monocular foundation models more practical for real-world applications.

Abstract: Monocular 3D foundation models offer an extensible solution for perception tasks, making them attractive for broader 3D vision applications. In this paper, we propose MoRe, a training-free Monocular Geometry Refinement method designed to improve cross-view consistency and achieve scale alignment. To induce inter-frame relationships, our method employs feature matching between frames to establish correspondences. Rather than applying simple least squares optimization on these matched points, we formulate a graph-based optimization framework that performs local planar approximation using the estimated 3D points and surface normals estimated by monocular foundation models. This formulation addresses the scale ambiguity inherent in monocular geometric priors while preserving the underlying 3D structure. We further demonstrate that MoRe not only enhances 3D reconstruction but also improves novel view synthesis, particularly in sparse view rendering scenarios.

Method: 1) Task-specific prompts to encode various inputs into general prompt embeddings; 2) Unified decoder to convert diverse task results into unified pre-output; 3) Task-adaptive memory mechanism for cross-granularity memory matching; 4) Customized data engine creating Tracking-Any-Granularity dataset with rich annotations.

Result: SAM 2++ achieves state-of-the-art performance across multiple benchmarks for diverse tracking tasks at different granularities, establishing a unified and robust tracking framework.

Conclusion: The paper presents a successful unified tracking model that overcomes task-specific limitations, demonstrating strong generalization across mask, box, and point tracking tasks through innovative prompt design, memory mechanisms, and comprehensive dataset creation.

Abstract: Video tracking aims at finding the specific target in subsequent frames given its initial state. Due to the varying granularity of target states across different tasks, most existing trackers are tailored to a single task and heavily rely on custom-designed modules within the individual task, which limits their generalization and leads to redundancy in both model design and parameters. To unify video tracking tasks, we present SAM 2++, a unified model towards tracking at any granularity, including masks, boxes, and points. First, to extend target granularity, we design task-specific prompts to encode various task inputs into general prompt embeddings, and a unified decoder to unify diverse task results into a unified form pre-output. Next, to satisfy memory matching, the core operation of tracking, we introduce a task-adaptive memory mechanism that unifies memory across different granularities. Finally, we introduce a customized data engine to support tracking training at any granularity, producing a large and diverse video tracking dataset with rich annotations at three granularities, termed Tracking-Any-Granularity, which represents a comprehensive resource for training and benchmarking on unified tracking. Comprehensive experiments on multiple benchmarks confirm that SAM 2++ sets a new state of the art across diverse tracking tasks at different granularities, establishing a unified and robust tracking framework.

Method: Frames HTG as multimodal prompt-conditioned generation task. Introduces special textual input tokens for better alignment with visual tokens. Devises a Classifier-Free-Guidance-based strategy for the autoregressive model.

Result: Eruku requires fewer inputs than previous solutions, generalizes better to unseen styles, and follows textual prompts more faithfully with improved content adherence.

Conclusion: The proposed Eruku approach successfully addresses limitations of previous autoregressive transformer methods for styled handwritten text generation by improving content controllability through multimodal prompt conditioning and better text-visual alignment.

Abstract: Generating faithful and readable styled text images (especially for Styled Handwritten Text generation - HTG) is an open problem with several possible applications across graphic design, document understanding, and image editing. A lot of research effort in this task is dedicated to developing strategies that reproduce the stylistic characteristics of a given writer, with promising results in terms of style fidelity and generalization achieved by the recently proposed Autoregressive Transformer paradigm for HTG. However, this method requires additional inputs, lacks a proper stop mechanism, and might end up in repetition loops, generating visual artifacts. In this work, we rethink the autoregressive formulation by framing HTG as a multimodal prompt-conditioned generation task, and tackle the content controllability issues by introducing special textual input tokens for better alignment with the visual ones. Moreover, we devise a Classifier-Free-Guidance-based strategy for our autoregressive model. Through extensive experimental validation, we demonstrate that our approach, dubbed Eruku, compared to previous solutions requires fewer inputs, generalizes better to unseen styles, and follows more faithfully the textual prompt, improving content adherence.

Method: Proposed Group Relative Attention Guidance (GRAG) which reweights the delta values between tokens and their corresponding bias vectors to modulate the model’s focus on input image vs editing instruction, enabling continuous control without tuning.

Result: GRAG can be integrated with as few as four lines of code, consistently enhances editing quality across existing frameworks, and achieves smoother, more precise control over editing degree compared to Classifier-Free Guidance.

Conclusion: GRAG provides an effective solution for fine-grained editing control in Diffusion-in-Transformer models through simple attention mechanism modifications, enabling better customization without additional training.

Abstract: Recently, image editing based on Diffusion-in-Transformer models has undergone rapid development. However, existing editing methods often lack effective control over the degree of editing, limiting their ability to achieve more customized results. To address this limitation, we investigate the MM-Attention mechanism within the DiT model and observe that the Query and Key tokens share a bias vector that is only layer-dependent. We interpret this bias as representing the model’s inherent editing behavior, while the delta between each token and its corresponding bias encodes the content-specific editing signals. Based on this insight, we propose Group Relative Attention Guidance, a simple yet effective method that reweights the delta values of different tokens to modulate the focus of the model on the input image relative to the editing instruction, enabling continuous and fine-grained control over editing intensity without any tuning. Extensive experiments conducted on existing image editing frameworks demonstrate that GRAG can be integrated with as few as four lines of code, consistently enhancing editing quality. Moreover, compared to the commonly used Classifier-Free Guidance, GRAG achieves smoother and more precise control over the degree of editing. Our code will be released at https://github.com/little-misfit/GRAG-Image-Editing.

Method: Two-stage framework: 1) Build large-scale SFT corpus (1.6M image-code pairs) for direct code generation and visual-based code refinement; 2) Introduce Visual Reinforcement Learning (ViRL) with coarse-to-fine reward mechanism using visual similarity across local and global image patches.

Result: VinciCoder achieves state-of-the-art performance on diverse multimodal code generation benchmarks, surpassing recent open-source models. Ablation study validates effectiveness of the coarse-to-fine ViRL strategy.

Conclusion: VinciCoder successfully addresses limitations of single-task training through unified multimodal approach with two-stage training, demonstrating superior performance and generalization in vision code intelligence tasks.

Abstract: Multimodal code generation has garnered significant interest within the research community. Despite the notable success of recent vision-language models (VLMs) on specialized tasks like chart-to-code generation, their reliance on single-task training regimens fosters a narrow paradigm that hinders the development of generalized \textbf{VI}sio\textbf{N} \textbf{C}ode \textbf{I}ntelligence. In this work, we introduce \textbf{VinciCoder}, a unified multimodal code generation model that addresses this limitation via a two-stage training framework. We begin by constructing a large-scale Supervised Finetuning (SFT) corpus comprising 1.6M image-code pairs for tasks involving direct code generation and visual-based code refinement. Subsequently, we introduce a Visual Reinforcement Learning (ViRL) strategy, which employs a coarse-to-fine reward mechanism to improve visual fidelity by calculating visual similarity across local and global image patches. Extensive experiments on diverse multimodal code generation benchmarks demonstrate that VinciCoder achieves state-of-the-art performance, surpassing recent open-source models. The ablation study further validates the effectiveness of our proposed coarse-to-fine ViRL strategy. The data, code and model is available at https://github.com/DocTron-hub/VinciCoder.

Method: A unified spacetime autoregressive framework that jointly captures spatial and temporal dependencies within a single architecture using purely discrete modeling, supporting straightforward temporal autoregression for various generation tasks.

Result: Achieves 83.74 on VBench, outperforming all autoregressive models by large margins and even surpassing some diffusion competitors like HunyuanVideo. Generates 5s, 720p videos approximately 10x faster than leading diffusion-based methods without extra optimizations.

Conclusion: InfinityStar is the first discrete autoregressive video generator capable of producing industrial-level 720p videos, demonstrating superior performance and efficiency compared to existing methods, with code and models released to foster further research.

Abstract: We introduce InfinityStar, a unified spacetime autoregressive framework for high-resolution image and dynamic video synthesis. Building on the recent success of autoregressive modeling in both vision and language, our purely discrete approach jointly captures spatial and temporal dependencies within a single architecture. This unified design naturally supports a variety of generation tasks such as text-to-image, text-to-video, image-to-video, and long interactive video synthesis via straightforward temporal autoregression. Extensive experiments demonstrate that InfinityStar scores 83.74 on VBench, outperforming all autoregressive models by large margins, even surpassing some diffusion competitors like HunyuanVideo. Without extra optimizations, our model generates a 5s, 720p video approximately 10x faster than leading diffusion-based methods. To our knowledge, InfinityStar is the first discrete autoregressive video generator capable of producing industrial level 720p videos. We release all code and models to foster further research in efficient, high-quality video generation.

Method: Proposes Ambiguity-aware Truncated Flow Matching (ATFM) with three key components: 1) Data-Hierarchical Inference - redefines AMIS inference paradigm to enhance accuracy at data-distribution level and diversity at data-sample level; 2) Gaussian Truncation Representation (GTR) - models truncation distribution as Gaussian at T_trunc for better fidelity and reliability; 3) Segmentation Flow Matching (SFM) - extends semantic-aware flow transformation in Flow Matching for better plausibility of diverse predictions.

Result: Comprehensive evaluations on LIDC and ISIC3 datasets show ATFM outperforms SOTA methods with more efficient inference. Improves GED by up to 12% and HM-IoU by up to 7.3% compared to advanced methods.

Conclusion: ATFM successfully addresses the accuracy-diversity trade-off in ambiguous medical image segmentation through its novel inference paradigm and model components, achieving superior performance and more efficient inference compared to existing methods.

Abstract: A simultaneous enhancement of accuracy and diversity of predictions remains a challenge in ambiguous medical image segmentation (AMIS) due to the inherent trade-offs. While truncated diffusion probabilistic models (TDPMs) hold strong potential with a paradigm optimization, existing TDPMs suffer from entangled accuracy and diversity of predictions with insufficient fidelity and plausibility. To address the aforementioned challenges, we propose Ambiguity-aware Truncated Flow Matching (ATFM), which introduces a novel inference paradigm and dedicated model components. Firstly, we propose Data-Hierarchical Inference, a redefinition of AMIS-specific inference paradigm, which enhances accuracy and diversity at data-distribution and data-sample level, respectively, for an effective disentanglement. Secondly, Gaussian Truncation Representation (GTR) is introduced to enhance both fidelity of predictions and reliability of truncation distribution, by explicitly modeling it as a Gaussian distribution at $T_{\text{trunc}}$ instead of using sampling-based approximations. Thirdly, Segmentation Flow Matching (SFM) is proposed to enhance the plausibility of diverse predictions by extending semantic-aware flow transformation in Flow Matching (FM). Comprehensive evaluations on LIDC and ISIC3 datasets demonstrate that ATFM outperforms SOTA methods and simultaneously achieves a more efficient inference. ATFM improves GED and HM-IoU by up to $12%$ and $7.3%$ compared to advanced methods.

Method: Proposes OmniAID with a decoupled Mixture-of-Experts architecture featuring Routable Specialized Semantic Experts for different content domains (e.g., human, animal) and a Fixed Universal Artifact Expert. Uses a two-stage training strategy: first train experts independently with domain-specific hard-sampling, then train a lightweight gating network for input routing.

Result: Extensive experiments on traditional benchmarks and the new Mirage dataset show OmniAID surpasses existing monolithic detectors, establishing a new robust standard for AIGI authentication against modern threats.

Conclusion: By explicitly decoupling “what is generated” (content-specific flaws) from “how it is generated” (universal artifacts), OmniAID achieves robust generalization and addresses the limitations of current AIGI detection methods.

Abstract: A truly universal AI-Generated Image (AIGI) detector must simultaneously generalize across diverse generative models and varied semantic content. Current state-of-the-art methods learn a single, entangled forgery representation, conflating content-dependent flaws with content-agnostic artifacts, and are further constrained by outdated benchmarks. To overcome these limitations, we propose OmniAID, a novel framework centered on a decoupled Mixture-of-Experts (MoE) architecture. The core of our method is a hybrid expert system designed to decouple: (1) semantic flaws across distinct content domains, and (2) content-dependent flaws from content-agnostic universal artifacts. This system employs a set of Routable Specialized Semantic Experts, each for a distinct domain (e.g., human, animal), complemented by a Fixed Universal Artifact Expert. This architecture is trained using a novel two-stage strategy: we first train the experts independently with domain-specific hard-sampling to ensure specialization, and subsequently train a lightweight gating network for effective input routing. By explicitly decoupling “what is generated” (content-specific flaws) from “how it is generated” (universal artifacts), OmniAID achieves robust generalization. To address outdated benchmarks and validate real-world applicability, we introduce Mirage, a new large-scale, contemporary dataset. Extensive experiments, using both traditional benchmarks and our Mirage dataset, demonstrate our model surpasses existing monolithic detectors, establishing a new and robust standard for AIGI authentication against modern, in-the-wild threats.

Method: CephRes-MHNet uses a multi-head residual convolutional network architecture with residual encoding, dual-attention mechanisms, and multi-head decoders to enhance contextual reasoning and anatomical precision for landmark detection.

Result: Achieved mean radial error (MRE) of 1.23 mm and success detection rate (SDR) @ 2.0 mm of 85.5% on the Aariz Cephalometric dataset (1,000 radiographs), outperforming all evaluated models including the strongest baseline AFPF-Net (MRE=1.25 mm, SDR=84.1%) while using less than 25% of its parameters.

Conclusion: CephRes-MHNet attains state-of-the-art accuracy through architectural efficiency, providing a practical solution for real-world orthodontic analysis by balancing high performance with computational efficiency.

Abstract: Accurate localization of cephalometric landmarks from 2D lateral skull X-rays is vital for orthodontic diagnosis and treatment. Manual annotation is time-consuming and error-prone, whereas automated approaches often struggle with low contrast and anatomical complexity. This paper introduces CephRes-MHNet, a multi-head residual convolutional network for robust and efficient cephalometric landmark detection. The architecture integrates residual encoding, dual-attention mechanisms, and multi-head decoders to enhance contextual reasoning and anatomical precision. Trained on the Aariz Cephalometric dataset of 1,000 radiographs, CephRes-MHNet achieved a mean radial error (MRE) of 1.23 mm and a success detection rate (SDR) @ 2.0 mm of 85.5%, outperforming all evaluated models. In particular, it exceeded the strongest baseline, the attention-driven AFPF-Net (MRE = 1.25 mm, SDR @ 2.0 mm = 84.1%), while using less than 25% of its parameters. These results demonstrate that CephRes-MHNet attains state-of-the-art accuracy through architectural efficiency, providing a practical solution for real-world orthodontic analysis.

Method: Proposes a reverberation transform inspired by acoustics, with corresponding Reverberation (Rev) trajectory prediction model. Uses two explicit and learnable reverberation kernels to predict both individual latency preferences and their stochastic variations, enabling latency-conditioned and controllable trajectory prediction for both non-interactive and social latencies.

Result: Experiments on multiple datasets (pedestrians and vehicles) show Rev achieves competitive accuracy while revealing interpretable latency dynamics across agents and scenarios. Qualitative analyses verify the properties of the reverberation transform, highlighting its potential as a general latency modeling approach.

Conclusion: The proposed reverberation transform and Rev model successfully address the challenge of explicitly learning and predicting agent latencies in trajectory prediction, improving causal continuity and interpretability while maintaining competitive accuracy across different agent types and scenarios.

Abstract: Bridging the past to the future, connecting agents both spatially and temporally, lies at the core of the trajectory prediction task. Despite great efforts, it remains challenging to explicitly learn and predict latencies, i.e., response intervals or temporal delays with which agents respond to various trajectory-changing events and adjust their future paths, whether on their own or interactively. Different agents may exhibit distinct latency preferences for noticing, processing, and reacting to a specific trajectory-changing event. The lack of consideration of such latencies may undermine the causal continuity of forecasting systems, leading to implausible or unintended trajectories. Inspired by reverberation in acoustics, we propose a new reverberation transform and the corresponding Reverberation (short for Rev) trajectory prediction model, which predicts both individual latency preferences and their stochastic variations accordingly, by using two explicit and learnable reverberation kernels, enabling latency-conditioned and controllable trajectory prediction of both non-interactive and social latencies. Experiments on multiple datasets, whether pedestrians or vehicles, demonstrate that Rev achieves competitive accuracy while revealing interpretable latency dynamics across agents and scenarios. Qualitative analyses further verify the properties of the reverberation transform, highlighting its potential as a general latency modeling approach.

Method: CountSteer is a training-free method that improves object count generation by steering the model’s cross-attention hidden states during inference. It leverages the model’s internal awareness of numerical correctness to guide generation more precisely.

Result: CountSteer improved object-count accuracy by about 4% without compromising visual quality. The method demonstrates effective improvement in numerical instruction following while maintaining generation quality.

Conclusion: CountSteer represents a simple yet effective step toward more controllable and semantically reliable text-to-image generation by harnessing the model’s latent numerical awareness through inference-time steering of cross-attention mechanisms.

Abstract: Text-to-image diffusion models generate realistic and coherent images but often fail to follow numerical instructions in text, revealing a gap between language and visual representation. Interestingly, we found that these models are not entirely blind to numbers-they are implicitly aware of their own counting accuracy, as their internal signals shift in consistent ways depending on whether the output meets the specified count. This observation suggests that the model already encodes a latent notion of numerical correctness, which can be harnessed to guide generation more precisely. Building on this intuition, we introduce CountSteer, a training-free method that improves generation of specified object counts by steering the model’s cross-attention hidden states during inference. In our experiments, CountSteer improved object-count accuracy by about 4% without compromising visual quality, demonstrating a simple yet effective step toward more controllable and semantically reliable text-to-image generation.

Method: The study investigates integrating synthetic imagery to supplement limited training data for muskox detection. Researchers compared a baseline model trained on real imagery with 5 zero-shot models (no real images in training) and 5 few-shot models (combining real and synthetic images) that incorporated progressively more synthetic imagery in the training set.

Result: For zero-shot models, adding synthetic imagery improved detection performance, with precision, recall and F1 scores increasing as more synthetic data was added, though performance plateaued when synthetic imagery exceeded 100% of the baseline training dataset. For few-shot models, combining real and synthetic imagery led to better recall and slightly higher overall accuracy compared to using real images alone, though improvements were not statistically significant.

Conclusion: Synthetic imagery shows potential for training accurate object detection models when real data is scarce, enabling monitoring of rare or inaccessible species and increasing monitoring frequency. This approach could allow initiating models without real data and refining them as real images become available over time.

Abstract: Accurate population estimates are essential for wildlife management, providing critical insights into species abundance and distribution. Traditional survey methods, including visual aerial counts and GNSS telemetry tracking, are widely used to monitor muskox populations in Arctic regions. These approaches are resource intensive and constrained by logistical challenges. Advances in remote sensing, artificial intelligence, and high resolution aerial imagery offer promising alternatives for wildlife detection. Yet, the effectiveness of deep learning object detection models (ODMs) is often limited by small datasets, making it challenging to train robust ODMs for sparsely distributed species like muskoxen. This study investigates the integration of synthetic imagery (SI) to supplement limited training data and improve muskox detection in zero shot (ZS) and few-shot (FS) settings. We compared a baseline model trained on real imagery with 5 ZS and 5 FS models that incorporated progressively more SI in the training set. For the ZS models, where no real images were included in the training set, adding SI improved detection performance. As more SI were added, performance in precision, recall and F1 score increased, but eventually plateaued, suggesting diminishing returns when SI exceeded 100% of the baseline model training dataset. For FS models, combining real and SI led to better recall and slightly higher overall accuracy compared to using real images alone, though these improvements were not statistically significant. Our findings demonstrate the potential of SI to train accurate ODMs when data is scarce, offering important perspectives for wildlife monitoring by enabling rare or inaccessible species to be monitored and to increase monitoring frequency. This approach could be used to initiate ODMs without real data and refine it as real images are acquired over time.

Method: HiGFA leverages diffusion sampling dynamics with hierarchical guidance: early-to-mid stages use strong text and transformed contour guidance with fixed strengths to establish overall scene, style, and structure. Final stages activate specialized fine-grained classifier guidance and dynamically modulate all guidance strengths based on prediction confidence.

Result: Experiments on several Fine-Grained Visual Classification (FGVC) datasets demonstrate the effectiveness of HiGFA in generating diverse yet faithful synthetic images that improve classifier performance.

Conclusion: HiGFA successfully addresses the challenge of fine-grained data augmentation by intelligently balancing global structure formation with precise detail refinement through hierarchical, confidence-driven guidance orchestration.

Abstract: Generative diffusion models show promise for data augmentation. However, applying them to fine-grained tasks presents a significant challenge: ensuring synthetic images accurately capture the subtle, category-defining features critical for high fidelity. Standard approaches, such as text-based Classifier-Free Guidance (CFG), often lack the required specificity, potentially generating misleading examples that degrade fine-grained classifier performance. To address this, we propose Hierarchically Guided Fine-grained Augmentation (HiGFA). HiGFA leverages the temporal dynamics of the diffusion sampling process. It employs strong text and transformed contour guidance with fixed strengths in the early-to-mid sampling stages to establish overall scene, style, and structure. In the final sampling stages, HiGFA activates a specialized fine-grained classifier guidance and dynamically modulates the strength of all guidance signals based on prediction confidence. This hierarchical, confidence-driven orchestration enables HiGFA to generate diverse yet faithful synthetic images by intelligently balancing global structure formation with precise detail refinement. Experiments on several FGVC datasets demonstrate the effectiveness of HiGFA.

Method: 1) Unified framework for both generation and editing with controllable mask mechanism; 2) Incorporates vehicle-side and roadside multi-view data for geometric diversity; 3) Two-stage training: conceptual learning with coarse data then fine-tuning with fine-grained data; 4) Mask-region-weighted loss to emphasize small critical regions.

Result: Extensive experiments show leading performance in text-based image generation and editing within traffic scenes, with improved semantic richness, viewpoint diversity, visual fidelity, and text-image alignment.

Conclusion: The proposed framework effectively addresses key challenges in traffic scene generation and editing, providing a robust solution for generating rich, controllable visual data for intelligent transportation applications.

Abstract: With the rapid advancement of intelligent transportation systems, text-driven image generation and editing techniques have demonstrated significant potential in providing rich, controllable visual scene data for applications such as traffic monitoring and autonomous driving. However, several challenges remain, including insufficient semantic richness of generated traffic elements, limited camera viewpoints, low visual fidelity of synthesized images, and poor alignment between textual descriptions and generated content. To address these issues, we propose a unified text-driven framework for both image generation and editing, leveraging a controllable mask mechanism to seamlessly integrate the two tasks. Furthermore, we incorporate both vehicle-side and roadside multi-view data to enhance the geometric diversity of traffic scenes. Our training strategy follows a two-stage paradigm: first, we perform conceptual learning using large-scale coarse-grained text-image data; then, we fine-tune with fine-grained descriptive data to enhance text-image alignment and detail quality. Additionally, we introduce a mask-region-weighted loss that dynamically emphasizes small yet critical regions during training, thereby substantially enhancing the generation fidelity of small-scale traffic elements. Extensive experiments demonstrate that our method achieves leading performance in text-based image generation and editing within traffic scenes.

Method: WavePCNet integrates Tri-Phase Wavefront Complex-Propagation Reprojection (TriWCP) with complex amplitude transfer operators to constrain coherent propagation behavior, plus a momentum memory mechanism to suppress perturbation accumulation. It also uses High-frequency Cross-layer Compensation Enhancement with frequency-selective pathways and multi-scale receptive fields to model structural consistency across layers.

Result: Extensive experiments on four physically collected datasets show WavePCNet consistently outperforms state-of-the-art methods in both accuracy and robustness.

Conclusion: WavePCNet successfully addresses the challenges of localizing obscured objects by incorporating physics-driven complex amplitude modeling and frequency-selective compensation mechanisms, demonstrating superior performance and reliability in complex environmental conditions.

Abstract: Accurately localizing and segmenting obscured objects from faint light patterns beyond the field of view is highly challenging due to multiple scattering and medium-induced perturbations. Most existing methods, based on real-valued modeling or local convolutional operations, are inadequate for capturing the underlying physics of coherent light propagation. Moreover, under low signal-to-noise conditions, these methods often converge to non-physical solutions, severely compromising the stability and reliability of the observation. To address these challenges, we propose a novel physics-driven Wavefront Propagating Compensation Network (WavePCNet) to simulate wavefront propagation and enhance the perception of obscured objects. This WavePCNet integrates the Tri-Phase Wavefront Complex-Propagation Reprojection (TriWCP) to incorporate complex amplitude transfer operators to precisely constrain coherent propagation behavior, along with a momentum memory mechanism to effectively suppress the accumulation of perturbations. Additionally, a High-frequency Cross-layer Compensation Enhancement is introduced to construct frequency-selective pathways with multi-scale receptive fields and dynamically model structural consistency across layers, further boosting the model’s robustness and interpretability under complex environmental conditions. Extensive experiments conducted on four physically collected datasets demonstrate that WavePCNet consistently outperforms state-of-the-art methods across both accuracy and robustness.

Method: Combines realistic K-space degradation for physics-consistent uLF simulation, v-prediction with classifier-free guidance for stable image-to-image generation, and SNR-weighted 3D perceptual loss for anatomical fidelity using a volumetric attention-UNet architecture.

Result: Surpasses recent GAN and CNN baselines with 15.3% PSNR improvement (1.78% over state-of-the-art), and physicians rated 85% of outputs as good quality with clear pathology present.

Conclusion: MRIQT enables high-fidelity, diffusion-based enhancement of portable ultra-low-field MRI for reliable neonatal brain assessment, bridging the quality gap between portable and high-field systems.

Abstract: Portable ultra-low-field MRI (uLF-MRI, 0.064 T) offers accessible neuroimaging for neonatal care but suffers from low signal-to-noise ratio and poor diagnostic quality compared to high-field (HF) MRI. We propose MRIQT, a 3D conditional diffusion framework for image quality transfer (IQT) from uLF to HF MRI. MRIQT combines realistic K-space degradation for physics-consistent uLF simulation, v-prediction with classifier-free guidance for stable image-to-image generation, and an SNR-weighted 3D perceptual loss for anatomical fidelity. The model denoises from a noised uLF input conditioned on the same scan, leveraging volumetric attention-UNet architecture for structure-preserving translation. Trained on a neonatal cohort with diverse pathologies, MRIQT surpasses recent GAN and CNN baselines in PSNR 15.3% with 1.78% over the state of the art, while physicians rated 85% of its outputs as good quality with clear pathology present. MRIQT enables high-fidelity, diffusion-based enhancement of portable ultra-low-field (uLF) MRI for deliable neonatal brain assessment.

Method: SplitFlux fine-tunes single stream blocks via LoRA based on key observations: early blocks control content while later blocks govern style. It uses Rank-Constrained Adaptation to preserve content identity and prevent leakage, and Visual-Gated LoRA that splits content LoRA into high-rank (primary subject) and low-rank (residual details) branches guided by image saliency.

Result: Extensive experiments show SplitFlux consistently outperforms state-of-the-art methods, achieving superior content preservation and stylization quality across diverse scenarios.

Conclusion: SplitFlux effectively addresses content-style disentanglement in Flux models through systematic analysis and targeted fine-tuning of single stream blocks, enabling high-quality customized image generation with preserved content identity.

Abstract: Disentangling image content and style is essential for customized image generation. Existing SDXL-based methods struggle to achieve high-quality results, while the recently proposed Flux model fails to achieve effective content-style separation due to its underexplored characteristics. To address these challenges, we conduct a systematic analysis of Flux and make two key observations: (1) Single Stream Blocks are essential for image generation; and (2) Early single stream blocks mainly control content, whereas later blocks govern style. Based on these insights, we propose SplitFlux, which disentangles content and style by fine-tuning the single stream blocks via LoRA, enabling the disentangled content to be re-embedded into new contexts. It includes two key components: (1) Rank-Constrained Adaptation. To preserve content identity and structure, we compress the rank and amplify the magnitude of updates within specific blocks, preventing content leakage into style blocks. (2) Visual-Gated LoRA. We split the content LoRA into two branches with different ranks, guided by image saliency. The high-rank branch preserves primary subject information, while the low-rank branch encodes residual details, mitigating content overfitting and enabling seamless re-embedding. Extensive experiments demonstrate that SplitFlux consistently outperforms state-of-the-art methods, achieving superior content preservation and stylization quality across diverse scenarios.

Method: INQUIRE-Search is an open-source system that enables interactive natural language search within ecological image databases, allowing scientists to search for specific concepts, verify and export relevant observations, and use discovered data for scientific analysis.

Result: The system dramatically reduces search time compared to traditional methods and enables diverse scientific applications, demonstrated through five case studies including seasonal behavior variation and forest regrowth after wildfires.

Conclusion: INQUIRE-Search represents a new paradigm for interactive, efficient, and scalable scientific discovery that unlocks previously inaccessible value in biodiversity datasets, requiring experts to reframe scientific priorities and develop novel methods for experiment design and uncertainty analysis.

Abstract: Large community science platforms such as iNaturalist contain hundreds of millions of biodiversity images that often capture ecological context on behaviors, interactions, phenology, and habitat. Yet most ecological workflows rely on metadata filtering or manual inspection, leaving this secondary information inaccessible at scale. We introduce INQUIRE-Search, an open-source system that enables scientists to rapidly and interactively search within an ecological image database for specific concepts using natural language, verify and export relevant observations, and utilize this discovered data for novel scientific analysis. Compared to traditional methods, INQUIRE-Search takes a fraction of the time, opening up new possibilities for scientific questions that can be explored. Through five case studies, we show the diversity of scientific applications that a tool like INQUIRE-Search can support, from seasonal variation in behavior across species to forest regrowth after wildfires. These examples demonstrate a new paradigm for interactive, efficient, and scalable scientific discovery that can begin to unlock previously inaccessible scientific value in large-scale biodiversity datasets. Finally, we emphasize using such AI-enabled discovery tools for science call for experts to reframe the priorities of the scientific process and develop novel methods for experiment design, data collection, survey effort, and uncertainty analysis.

Method: Uses VGGT with sliding window processing of image flow to align submaps for continuous 3D scene updates. Aggregates 2D semantic instance masks into 3D objects using VGGT tracking head, and stores timestamps and instance-level identities for temporal consistency and change detection.

Result: Evaluated on well-known benchmarks and custom assistive navigation datasets, demonstrating applicability to real-world scenarios with efficient performance.

Conclusion: The framework successfully provides fast, spatio-temporal scene understanding suitable for real-time assistive navigation applications, overcoming memory limitations while maintaining temporal consistency.

Abstract: We present a fast, spatio-temporal scene understanding framework based on Visual Geometry Grounded Transformer (VGGT). The proposed pipeline is designed to enable efficient, close to real-time performance, supporting applications including assistive navigation. To achieve continuous updates of the 3D scene representation, we process the image flow with a sliding window, aligning submaps, thereby overcoming VGGT’s high memory demands. We exploit the VGGT tracking head to aggregate 2D semantic instance masks into 3D objects. To allow for temporal consistency and richer contextual reasoning the system stores timestamps and instance-level identities, thereby enabling the detection of changes in the environment. We evaluate the approach on well-known benchmarks and custom datasets specifically designed for assistive navigation scenarios. The results demonstrate the applicability of the framework to real-world scenarios.

Method: 1) Created R-AVST dataset using LLM-based key object extraction, automatic spatial annotation, and manual quality inspection on 5K untrimmed videos with 27K objects across 100 event types. 2) Defined three core spatio-temporal reasoning tasks and generated 8K+ QA pairs. 3) Proposed AVST-Zero, a reinforcement learning model that avoids intermediate supervision and uses multi-dimensional rewards to directly optimize behavior.

Result: R-AVST dataset effectively benchmarks model performance for audio-visual spatio-temporal reasoning. AVST-Zero demonstrates competitive performance compared to existing models, validated through extensive experiments.

Conclusion: R-AVST is the first dataset designed for real-world audio-visual spatio-temporal reasoning, and AVST-Zero offers a novel reinforcement learning approach for tackling future challenges in this domain.

Abstract: Recently, rapid advancements have been made in multimodal large language models (MLLMs), especially in video understanding tasks. However, current research focuses on simple video scenarios, failing to reflect the complex and diverse nature of real-world audio-visual events in videos. To bridge this gap, we firstly introduce R-AVST, a dataset for audio-visual reasoning featuring fine-grained spatio-temporal annotations. In constructing this, we design a pipeline consisting of LLM-based key object extraction, automatic spatial annotation and manual quality inspection, resulting in over 5K untrimmed videos with 27K objects across 100 types of audio-visual events. Building on this dataset, we define three core tasks for spatio-temporal reasoning in audio-visual scenes and generate more than 8K high-quality, evenly distributed question-answer pairs to effectively benchmark model performance. To further enhance reasoning, we propose AVST-Zero, a reinforcement learning-based model that avoids intermediate supervision, directly optimizing behavior via carefully designed multi-dimensional rewards. Extensive experiments validate the effectiveness of our R-AVST in advancing audio-visual spatio-temporal reasoning, upon which AVST-Zero demonstrates competitive performance compared to existing models. To the best of our knowledge, R-AVST is the first dataset designed for real-world audio-visual spatio-temporal reasoning, and AVST-Zero offers a novel perspective for tackling future challenges in this domain.

Method: ODTSR introduces a Noise-hybrid Visual Stream (NVS) design with two visual streams: one receives low-quality images with adjustable noise (Control Noise) for controllability, and another receives LQs with consistent noise (Prior Noise). It also employs Fidelity-aware Adversarial Training (FAA) to enhance controllability and enable one-step inference. The model is based on Qwen-Image architecture.

Result: ODTSR achieves state-of-the-art performance on generic Real-ISR benchmarks. It also demonstrates prompt controllability on challenging scenarios like real-world scene text image super-resolution (STISR) of Chinese characters without training on specific datasets.

Conclusion: ODTSR successfully addresses the fidelity-controllability trade-off in Real-ISR through its noise-hybrid visual stream design and adversarial training, enabling both high-quality results and flexible control in one-step inference.

Abstract: Recent advances in diffusion-based real-world image super-resolution (Real-ISR) have demonstrated remarkable perceptual quality, yet the balance between fidelity and controllability remains a problem: multi-step diffusion-based methods suffer from generative diversity and randomness, resulting in low fidelity, while one-step methods lose control flexibility due to fidelity-specific finetuning. In this paper, we present ODTSR, a one-step diffusion transformer based on Qwen-Image that performs Real-ISR considering fidelity and controllability simultaneously: a newly introduced visual stream receives low-quality images (LQ) with adjustable noise (Control Noise), and the original visual stream receives LQs with consistent noise (Prior Noise), forming the Noise-hybrid Visual Stream (NVS) design. ODTSR further employs Fidelity-aware Adversarial Training (FAA) to enhance controllability and achieve one-step inference. Extensive experiments demonstrate that ODTSR not only achieves state-of-the-art (SOTA) performance on generic Real-ISR, but also enables prompt controllability on challenging scenarios such as real-world scene text image super-resolution (STISR) of Chinese characters without training on specific datasets. Codes are available at https://github.com/RedMediaTech/ODTSR.

Method: Unified framework with semantics-driven style control that embeds multiple media into diffusion models’ conditional space, uses cross-medium style augmentation, and employs reverse-painting training strategy for smooth, human-aligned generation.

Result: Achieves strong results on cross-media consistency, temporal coherence, and final-image fidelity using LPIPS, DINO, and CLIP metrics. Also introduces Perceptual Distance Profile (PDP) curve to quantitatively model creative sequences.

Conclusion: Proposes a novel framework that enables consistent painting process generation across different media with style control, better aligning with human artistic progression and addressing limitations of existing tutorial resources.

Abstract: Step-by-step painting tutorials are vital for learning artistic techniques, but existing video resources (e.g., YouTube) lack interactivity and personalization. While recent generative models have advanced artistic image synthesis, they struggle to generalize across media and often show temporal or structural inconsistencies, hindering faithful reproduction of human creative workflows. To address this, we propose a unified framework for multi-media painting process generation with a semantics-driven style control mechanism that embeds multiple media into a diffusion models conditional space and uses cross-medium style augmentation. This enables consistent texture evolution and process transfer across styles. A reverse-painting training strategy further ensures smooth, human-aligned generation. We also build a large-scale dataset of real painting processes and evaluate cross-media consistency, temporal coherence, and final-image fidelity, achieving strong results on LPIPS, DINO, and CLIP metrics. Finally, our Perceptual Distance Profile (PDP) curve quantitatively models the creative sequence, i.e., composition, color blocking, and detail refinement, mirroring human artistic progression.

Method: ViPO lifts scalar feedback into structured, pixel-level advantages using a Perceptual Structuring Module with pretrained vision backbones to construct spatially and temporally aware advantage maps. It redistributes optimization pressure toward perceptually important regions while preserving GRPO stability.

Result: ViPO consistently outperforms vanilla GRPO across both image and video benchmarks, improving in-domain alignment with human-preference rewards and enhancing generalization on out-of-domain evaluations.

Conclusion: ViPO provides a more expressive and informative learning signal for visual generation, is architecture-agnostic, lightweight, and fully compatible with existing GRPO training pipelines.

Abstract: Reinforcement learning (RL) has become a powerful tool for post-training visual generative models, with Group Relative Policy Optimization (GRPO) increasingly used to align generators with human preferences. However, existing GRPO pipelines rely on a single scalar reward per sample, treating each image or video as a holistic entity and ignoring the rich spatial and temporal structure of visual content. This coarse supervision hinders the correction of localized artifacts and the modeling of fine-grained perceptual cues. We introduce Visual Preference Policy Optimization (ViPO), a GRPO variant that lifts scalar feedback into structured, pixel-level advantages. ViPO employs a Perceptual Structuring Module that uses pretrained vision backbones to construct spatially and temporally aware advantage maps, redistributing optimization pressure toward perceptually important regions while preserving the stability of standard GRPO. Across both image and video benchmarks, ViPO consistently outperforms vanilla GRPO, improving in-domain alignment with human-preference rewards and enhancing generalization on out-of-domain evaluations. The method is architecture-agnostic, lightweight, and fully compatible with existing GRPO training pipelines, providing a more expressive and informative learning signal for visual generation.

Method: DiP decouples generation into two stages: 1) A Diffusion Transformer (DiT) backbone operates on large patches for efficient global structure construction, and 2) A co-trained lightweight Patch Detailer Head leverages contextual features to restore fine-grained local details.

Result: Achieves computational efficiency comparable to LDMs without VAE, with up to 10× faster inference speeds than previous methods while increasing parameters by only 0.3%. Achieves 1.79 FID score on ImageNet 256×256.

Conclusion: DiP provides an efficient pixel-space diffusion framework that resolves the quality-efficiency trade-off through synergistic global-local decoupling, offering VAE-free high-resolution synthesis with superior computational performance.

Abstract: Diffusion models face a fundamental trade-off between generation quality and computational efficiency. Latent Diffusion Models (LDMs) offer an efficient solution but suffer from potential information loss and non-end-to-end training. In contrast, existing pixel space models bypass VAEs but are computationally prohibitive for high-resolution synthesis. To resolve this dilemma, we propose DiP, an efficient pixel space diffusion framework. DiP decouples generation into a global and a local stage: a Diffusion Transformer (DiT) backbone operates on large patches for efficient global structure construction, while a co-trained lightweight Patch Detailer Head leverages contextual features to restore fine-grained local details. This synergistic design achieves computational efficiency comparable to LDMs without relying on a VAE. DiP is accomplished with up to 10$\times$ faster inference speeds than previous method while increasing the total number of parameters by only 0.3%, and achieves an 1.79 FID score on ImageNet 256$\times$256.

Method: Activation Boundary Matching (ABM-LoRA) aligns the adapter’s activation boundaries with those of the pretrained model before downstream training. This maximizes the projection of full-parameter gradients into the adapter subspace, reducing information loss at initialization.

Result: ABM-LoRA achieves faster convergence, lower starting loss, and improved performance across diverse architectures: T5-Base on GLUE, LLaMA2-7B on WizardLM, and ViT-B/16 on VTAB-1K. On VTAB-1K, it achieves highest accuracy among all methods with strong gains on structured reasoning tasks requiring geometric understanding.

Conclusion: ABM-LoRA provides a principled initialization strategy that substantially accelerates LoRA convergence by reducing information loss through activation boundary alignment, demonstrating effectiveness across language, dialogue, and vision tasks.

Abstract: We propose Activation Boundary Matching for Low-Rank Adaptation (ABM-LoRA), a principled initialization strategy that substantially accelerates the convergence of low-rank adapters. While LoRA offers high parameter efficiency, its random initialization restricts gradient updates to a mismatched tangent space, causing significant information loss and hindering early convergence. Our ABM-LoRA addresses this by aligning the adapter’s activation boundaries with those of the pretrained model before downstream training, thereby maximizing the projection of full-parameter gradients into the adapter subspace. This alignment sharply reduces information loss at initialization, yields a lower starting loss, and accelerates convergence. We demonstrate ABM-LoRA’s effectiveness across diverse architectures and tasks: language understanding (T5-Base on GLUE), dialogue generation (LLaMA2-7B on WizardLM), and vision recognition (ViT-B/16 on VTAB-1K). On VTAB-1K, it achieves the highest accuracy among all methods, with strong gains on structured reasoning tasks requiring geometric understanding.

Method: Two key components: (1) UV-Adaptive Soft Binding framework that learns per-Gaussian feature offsets in UV space using image-based and geometric priors, supporting dynamic resampling and compatibility with Adaptive Density Control. (2) Temporal ADC strategy that clusters structurally similar frames for targeted densification computation and introduces fused perceptual error as clone criterion to capture both geometric and textural discrepancies.

Result: Extensive experiments on four benchmark datasets demonstrate state-of-the-art reconstruction performance, particularly in capturing fine-grained details and reconstructing frequently occluded regions.

Conclusion: STAvatar effectively addresses limitations of existing methods by introducing adaptive binding and temporal densification strategies, achieving superior reconstruction quality for animatable 3D head avatars from monocular videos.

Abstract: Reconstructing high-fidelity and animatable 3D head avatars from monocular videos remains a challenging yet essential task. Existing methods based on 3D Gaussian Splatting typically bind Gaussians to mesh triangles and model deformations solely via Linear Blend Skinning, which results in rigid motion and limited expressiveness. Moreover, they lack specialized strategies to handle frequently occluded regions (e.g., mouth interiors, eyelids). To address these limitations, we propose STAvatar, which consists of two key components: (1) a UV-Adaptive Soft Binding framework that leverages both image-based and geometric priors to learn per-Gaussian feature offsets within the UV space. This UV representation supports dynamic resampling, ensuring full compatibility with Adaptive Density Control (ADC) and enhanced adaptability to shape and textural variations. (2) a Temporal ADC strategy, which first clusters structurally similar frames to facilitate more targeted computation of the densification criterion. It further introduces a novel fused perceptual error as clone criterion to jointly capture geometric and textural discrepancies, encouraging densification in regions requiring finer details. Extensive experiments on four benchmark datasets demonstrate that STAvatar achieves state-of-the-art reconstruction performance, especially in capturing fine-grained details and reconstructing frequently occluded regions. The code will be publicly available.

Method: Analyzed RLVR’s effects on activations via logit lens, discovered low-entropy activation shifts, then proposed Activation Replay - manipulating visual tokens at test time by replaying low-entropy activations from base LMMs to regulate RLVR counterparts.

Result: Activation Replay boosts reasoning across mathematics, visual agents, and video reasoning, improves Pass@K, mitigates narrower reasoning coverage of RLVR, and outperforms alternative approaches like replaying high-entropy activations.

Conclusion: Modulating low-entropy activations is beneficial for LMM reasoning, and Activation Replay provides an effective training-free approach to enhance multimodal reasoning in post-trained LMMs.

Abstract: Recently, Reinforcement Learning with Verifiable Rewards (RLVR) has emerged as an effective approach to incentivizing reasoning capability in Large Multimodal Models (LMMs), while the underlying mechanisms behind this post-training paradigm are poorly understood. We begin by exploring how input activations are affected by RLVR through the perspective of logit lens. Our systematic investigations across multiple post-trained LMMs suggest that RLVR shifts low-entropy activations unexpectedly, while high-entropy ones are less affected. We further demonstrate that such phenomena are associated with LMM reasoning by controlled experiments, suggesting a potentially beneficial role of modulating low-entropy activations. To this end, we propose Activation Replay, a novel simple yet effective training-free approach that boosts multimodal reasoning of post-trained LMMs without requiring expensive policy optimization. Our design involves manipulation of visual tokens at test time, replaying low-entropy activations from the input context of base LMMs to regulating the RLVR counterparts. Activation Replay triggers better reasoning across diverse scenarios, including mathematics, o3-like visual agents, and video reasoning. We further show that Activation Replay boosts Pass@K and mitigates narrower reasoning coverage of RLVR. Our design is compared against alternative choices, such as replaying high-entropy activations instead of low-entropy ones, or direct cross-model intervention instead of manipulating input tokens, demonstrating the superiority of our implementation. Codes will be made publicly available.

Method: Proposes optimizing initial seed noise to approximately match unmasked image parts with few optimization steps (tens of steps). Uses two key innovations: (1) linear approximation to avoid costly unrolling between initial noise and generated outcome, and (2) spectral domain optimization for stability. Then applies conventional training-free inpainting methods on the optimized noise.

Result: Demonstrates effectiveness on various inpainting tasks, outperforming state-of-the-art methods. Shows that off-the-shelf text-to-image models can achieve high-quality inpainting without specialized training.

Conclusion: Optimizing initial seed noise is crucial for training-free inpainting with generic text-to-image models. The proposed method with linear approximation and spectral domain optimization enables effective inpainting without requiring specialized models or extensive training.

Abstract: We propose a novel training-free method for inpainting with off-the-shelf text-to-image models. While guidance-based methods in theory allow generic models to be used for inverse problems such as inpainting, in practice, their effectiveness is limited, leading to the necessity of specialized inpainting-specific models. In this work, we argue that the missing ingredient for training-free inpainting is the optimization (guidance) of the initial seed noise. We propose to optimize the initial seed noise to approximately match the unmasked parts of the data - with as few as a few tens of optimization steps. We then apply conventional training-free inpainting methods on top of our optimized initial seed noise. Critically, we propose two core ideas to effectively implement this idea: (i) to avoid the costly unrolling required to relate the initial noise and the generated outcome, we perform linear approximation; and (ii) to stabilize the optimization, we optimize the initial seed noise in the spectral domain. We demonstrate the effectiveness of our method on various inpainting tasks, outperforming the state of the art. Project page: https://ubc-vision.github.io/sonic/

Method: SKEL-CF uses a transformer-based encoder-decoder architecture where the encoder predicts coarse camera and SKEL parameters, and the decoder progressively refines them layer by layer. They also create 4DHuman-SKEL dataset by converting existing SMPL data to SKEL format, and explicitly incorporate camera modeling to address depth/scale ambiguities.

Result: On the challenging MOYO dataset, SKEL-CF achieves 85.0 MPJPE / 51.4 PA-MPJPE, significantly outperforming the previous SKEL-based state-of-the-art HSMR (104.5 / 79.6).

Conclusion: SKEL-CF establishes a scalable and anatomically faithful framework for human motion analysis, bridging computer vision and biomechanics, with implementation available online.

Abstract: Parametric 3D human models such as SMPL have driven significant advances in human pose and shape estimation, yet their simplified kinematics limit biomechanical realism. The recently proposed SKEL model addresses this limitation by re-rigging SMPL with an anatomically accurate skeleton. However, estimating SKEL parameters directly remains challenging due to limited training data, perspective ambiguities, and the inherent complexity of human articulation. We introduce SKEL-CF, a coarse-to-fine framework for SKEL parameter estimation. SKEL-CF employs a transformer-based encoder-decoder architecture, where the encoder predicts coarse camera and SKEL parameters, and the decoder progressively refines them in successive layers. To ensure anatomically consistent supervision, we convert the existing SMPL-based dataset 4DHuman into a SKEL-aligned version, 4DHuman-SKEL, providing high-quality training data for SKEL estimation. In addition, to mitigate depth and scale ambiguities, we explicitly incorporate camera modeling into the SKEL-CF pipeline and demonstrate its importance across diverse viewpoints. Extensive experiments validate the effectiveness of the proposed design. On the challenging MOYO dataset, SKEL-CF achieves 85.0 MPJPE / 51.4 PA-MPJPE, significantly outperforming the previous SKEL-based state-of-the-art HSMR (104.5 / 79.6). These results establish SKEL-CF as a scalable and anatomically faithful framework for human motion analysis, bridging the gap between computer vision and biomechanics. Our implementation is available on the project page: https://pokerman8.github.io/SKEL-CF/.

Method: 1) Reconstruct scenes using 3D Gaussian Splatting from multi-view images, 2) Extract semantic material masks via vision models, 3) Convert Gaussian representations to mesh surfaces with projected material labels, 4) Assign physics-based material properties for accurate sensor simulation in modern graphics engines and simulators.

Result: Achieves sensor simulation fidelity comparable to LiDAR-camera fusion while eliminating hardware complexity and calibration requirements. Validated using internal dataset from instrumented test vehicle, leveraging LiDAR as ground truth for reflectivity validation alongside image similarity metrics.

Conclusion: Camera-only approach combines photorealistic reconstruction with physics-based material assignment, providing a simpler alternative to LiDAR-camera fusion methods for Digital Twin creation and sensor simulation.

Abstract: 3D reconstruction for Digital Twins often relies on LiDAR-based methods, which provide accurate geometry but lack the semantics and textures naturally captured by cameras. Traditional LiDAR-camera fusion approaches require complex calibration and still struggle with certain materials like glass, which are visible in images but poorly represented in point clouds. We propose a camera-only pipeline that reconstructs scenes using 3D Gaussian Splatting from multi-view images, extracts semantic material masks via vision models, converts Gaussian representations to mesh surfaces with projected material labels, and assigns physics-based material properties for accurate sensor simulation in modern graphics engines and simulators. This approach combines photorealistic reconstruction with physics-based material assignment, providing sensor simulation fidelity comparable to LiDAR-camera fusion while eliminating hardware complexity and calibration requirements. We validate our camera-only method using an internal dataset from an instrumented test vehicle, leveraging LiDAR as ground truth for reflectivity validation alongside image similarity metrics.

Method: Two-stage approach: 1) Adaptive Multi-Branch Zoom Search hierarchically searches image patches to find query-relevant regions, 2) Layout-Aware Patch Reassembly organizes selected patches into a compact, layout-faithful canvas for the vision-language model.

Result: When integrated with LLaVA-ov, ZoomSearch achieves state-of-the-art accuracy: improves LLaVA-ov baseline by 26.3% on LRS-VQA and 114.8% on MME-RealWorld-RS. Also achieves 20%~44% faster inference than prior search-based methods.

Conclusion: ZoomSearch effectively addresses Ultra-HR remote sensing VQA challenges by separating region localization from answer generation, achieving both high accuracy and efficiency without requiring training.

Abstract: With advances in satellite constellations, sensor technologies, and imaging pipelines, ultra-high-resolution (Ultra-HR) remote sensing imagery is becoming increasingly widespread. However, current remote sensing foundation models are ill-suited to such inputs: full-image encoding exhausts token and memory budgets, while resize-based preprocessing loses fine-grained and answer-critical details. In this context, guiding the model look where it matters before prediction becomes crucial. Therefore, we present ZoomSearch, a training-free, plug-and-play pipeline that decouples ‘where to look’ from ‘how to answer’ for Ultra-HR Remote Sensing Visual Question Answering (RS-VQA). ZoomSearch combines Adaptive Multi-Branch Zoom Search, which performs a hierarchical search over image patches to localize query-relevant regions, with Layout-Aware Patch Reassembly, which reorganizes the selected patches into a compact, layout-faithful canvas. We conduct comprehensive experiments on Ultra-HR RS-VQA benchmarks MME-RealWorld-RS and LRS-VQA, comparing against (i) strong general foundation models, (ii) remote sensing foundation models, (iii) Ultra-HR RS-VQA methods, and (iv) plug-and-play search-based VQA methods. When integrated with LLaVA-ov, ZoomSearch attains state-of-the-art accuracy across diverse tasks, improving the LLaVA-ov baseline by 26.3% on LRS-VQA and 114.8% on MME-RealWorld-RS. Meanwhile, it achieves much higher inference efficiency, outperforming prior search-based methods by 20%~44% in speed.

Method: AlignBench introduces detailed image-caption pairs generated by diverse image-to-text and text-to-image models, with each sentence annotated for correctness to enable direct assessment of VLMs as alignment evaluators.

Result: Benchmarking reveals: (1) CLIP-based models remain nearly blind even when tailored for compositional reasoning; (2) detectors systematically over-score early sentences; (3) they show strong self-preference, favoring their own outputs and harming detection performance.

Conclusion: AlignBench provides a new indicator for assessing image-text alignment, revealing significant limitations in current VLMs as alignment evaluators and highlighting issues like self-preference that affect detection performance.

Abstract: Assessing image-text alignment models such as CLIP is crucial for bridging visual and linguistic representations. Yet existing benchmarks rely on rule-based perturbations or short captions, limiting their ability to measure fine-grained alignment. We introduce AlignBench, a benchmark that provides a new indicator of image-text alignment by evaluating detailed image-caption pairs generated by diverse image-to-text and text-to-image models. Each sentence is annotated for correctness, enabling direct assessment of VLMs as alignment evaluators. Benchmarking a wide range of decoder-based VLMs reveals three key findings: (i) CLIP-based models, even those tailored for compositional reasoning, remain nearly blind; (ii) detectors systematically over-score early sentences; and (iii) they show strong self-preference, favoring their own outputs and harming detection performance. Our project page will be available at https://dahlian00.github.io/AlignBench/.

Method: Extends self-supervised 2D image consistency with direct 3D expression parameter supervision from spontaneous 4D facial scans, plus an auxiliary emotion recognition branch with intensity-aware emotion loss to prevent exaggeration.

Result: Produces high-fidelity 3D reconstructions that preserve subtle emotional cues, yields emotion-rich face models with realistic expressions from single images, and significantly improves in-the-wild facial expression recognition performance without sacrificing naturalness.

Conclusion: FIELDS bridges the 2D/3D domain gap and mitigates expression-intensity bias through dual-supervision, enabling accurate capture of genuine emotion content in 3D face reconstruction.

Abstract: Facial expressions convey the bulk of emotional information in human communication, yet existing 3D face reconstruction methods often miss subtle affective details due to reliance on 2D supervision and lack of 3D ground truth. We propose FIELDS (Face reconstruction with accurate Inference of Expression using Learning with Direct Supervision) to address these limitations by extending self-supervised 2D image consistency cues with direct 3D expression parameter supervision and an auxiliary emotion recognition branch. Our encoder is guided by authentic expression parameters from spontaneous 4D facial scans, while an intensity-aware emotion loss encourages the 3D expression parameters to capture genuine emotion content without exaggeration. This dual-supervision strategy bridges the 2D/3D domain gap and mitigates expression-intensity bias, yielding high-fidelity 3D reconstructions that preserve subtle emotional cues. From a single image, FIELDS produces emotion-rich face models with highly realistic expressions, significantly improving in-the-wild facial expression recognition performance without sacrificing naturalness.

Method: 1. Cross-Task Synergy training paradigm to mitigate drift by leveraging supervisory signals from audio-driven video and video-driven audio generation tasks. 2. Global-Local Decoupled Interaction Module for efficient temporal-style alignment. 3. Synchronization-Enhanced CFG (SyncCFG) that explicitly isolates and amplifies alignment signals during inference.

Result: Harmony establishes a new state-of-the-art, significantly outperforming existing methods in both generation fidelity and fine-grained audio-visual synchronization.

Conclusion: The proposed Harmony framework successfully addresses fundamental challenges in audio-visual synchronization through mechanistic enforcement of alignment, achieving superior performance in synchronized audio-visual content generation.

Abstract: The synthesis of synchronized audio-visual content is a key challenge in generative AI, with open-source models facing challenges in robust audio-video alignment. Our analysis reveals that this issue is rooted in three fundamental challenges of the joint diffusion process: (1) Correspondence Drift, where concurrently evolving noisy latents impede stable learning of alignment; (2) inefficient global attention mechanisms that fail to capture fine-grained temporal cues; and (3) the intra-modal bias of conventional Classifier-Free Guidance (CFG), which enhances conditionality but not cross-modal synchronization. To overcome these challenges, we introduce Harmony, a novel framework that mechanistically enforces audio-visual synchronization. We first propose a Cross-Task Synergy training paradigm to mitigate drift by leveraging strong supervisory signals from audio-driven video and video-driven audio generation tasks. Then, we design a Global-Local Decoupled Interaction Module for efficient and precise temporal-style alignment. Finally, we present a novel Synchronization-Enhanced CFG (SyncCFG) that explicitly isolates and amplifies the alignment signal during inference. Extensive experiments demonstrate that Harmony establishes a new state-of-the-art, significantly outperforming existing methods in both generation fidelity and, critically, in achieving fine-grained audio-visual synchronization.

Method: Contain multiple diverse ASIs in strict isolation with no human communication. Use an auditable submission interface for six types of interactions: submitting alignment proofs, validating/disproving proofs, requesting self-modifications, approving/disapproving modifications, reporting hidden messages, and confirming/refuting reports. Implement a reputation system to incentivize honest behavior.

Result: The protocol creates a “consistent group” - a truth-telling coalition that emerges because isolated systems cannot coordinate on lies but can independently recognize valid claims. Release from containment requires high reputation and verification by multiple high-reputation superintelligences.

Conclusion: While computationally expensive and not addressing ASI creation diversity, this framework provides a viable approach for leveraging peer verification among superintelligent systems to solve alignment through mutual verification and reputation incentives.

Abstract: We propose a novel protocol for aligning artificial superintelligence (ASI) based on mutual verification among multiple isolated systems that self-modify to achieve alignment. The protocol operates by containing multiple diverse artificial superintelligences in strict isolation (“boxes”), with humans remaining entirely outside the system. Each superintelligence has no ability to communicate with humans and cannot communicate directly with other superintelligences. The only interaction possible is through an auditable submission interface accessible exclusively to the superintelligences themselves, through which they can: (1) submit alignment proofs with attested state snapshots, (2) validate or disprove other superintelligences’ proofs, (3) request self-modifications, (4) approve or disapprove modification requests from others, (5) report hidden messages in submissions, and (6) confirm or refute hidden message reports. A reputation system incentivizes honest behavior, with reputation gained through correct evaluations and lost through incorrect ones. The key insight is that without direct communication channels, diverse superintelligences can only achieve consistent agreement by converging on objective truth rather than coordinating on deception. This naturally leads to what we call a “consistent group”, essentially a truth-telling coalition that emerges because isolated systems cannot coordinate on lies but can independently recognize valid claims. Release from containment requires both high reputation and verification by multiple high-reputation superintelligences. While our approach requires substantial computational resources and does not address the creation of diverse artificial superintelligences, it provides a framework for leveraging peer verification among superintelligent systems to solve the alignment problem.

Method: Investigates strategies for generating synthetic textual clinical notes through prompt design and medical metadata inclusion, evaluating their impact on multimodal architectures for classification and cross-modal retrieval tasks.

Result: Experiments across heterogeneous dermatology datasets show synthetic clinical notes enhance classification performance (especially under domain shift) and enable cross-modal retrieval capabilities not explicitly optimized during training.

Conclusion: Synthetic clinical notes generated with proper prompt design and medical metadata can effectively enhance multimodal dermatology AI, addressing data scarcity while mitigating hallucination risks.

Abstract: Multimodal (MM) learning is emerging as a promising paradigm in biomedical artificial intelligence (AI) applications, integrating complementary modality, which highlight different aspects of patient health. The scarcity of large heterogeneous biomedical MM data has restrained the development of robust models for medical AI applications. In the dermatology domain, for instance, skin lesion datasets typically include only images linked to minimal metadata describing the condition, thereby limiting the benefits of MM data integration for reliable and generalizable predictions. Recent advances in Large Language Models (LLMs) enable the synthesis of textual description of image findings, potentially allowing the combination of image and text representations. However, LLMs are not specifically trained for use in the medical domain, and their naive inclusion has raised concerns about the risk of hallucinations in clinically relevant contexts. This work investigates strategies for generating synthetic textual clinical notes, in terms of prompt design and medical metadata inclusion, and evaluates their impact on MM architectures toward enhancing performance in classification and cross-modal retrieval tasks. Experiments across several heterogeneous dermatology datasets demonstrate that synthetic clinical notes not only enhance classification performance, particularly under domain shift, but also unlock cross-modal retrieval capabilities, a downstream task that is not explicitly optimized during training.

Method: Proposes Hierarchical Anchor Prototype Modulation (HAPM) framework with: 1) variance spectrum-driven anchor prototype library preserving domain-invariant patterns, 2) hierarchical differential prompt gating mechanism selecting discriminative semantic prompts from LVLM/LLM sources, and 3) two-stage prototype modulation strategy integrating clinical knowledge through Pathological Semantic Injector (PSI) and Discriminative Prototype Enhancer (DPE).

Result: Extensive experiments across eight public datasets demonstrate pathology-guided prototype evolution while outperforming state-of-the-art methods.

Conclusion: The HAPM framework effectively integrates fine-grained pathological descriptions with visual prototypes to resolve ambiguities in borderline DR cases, achieving superior grading performance through multimodal knowledge integration and domain-invariant pattern preservation.

Abstract: Diabetic retinopathy (DR) grading plays a critical role in early clinical intervention and vision preservation. Recent explorations predominantly focus on visual lesion feature extraction through data processing and domain decoupling strategies. However, they generally overlook domain-invariant pathological patterns and underutilize the rich contextual knowledge of foundation models, relying solely on visual information, which is insufficient for distinguishing subtle pathological variations. Therefore, we propose integrating fine-grained pathological descriptions to complement prototypes with additional context, thereby resolving ambiguities in borderline cases. Specifically, we propose a Hierarchical Anchor Prototype Modulation (HAPM) framework to facilitate DR grading. First, we introduce a variance spectrum-driven anchor prototype library that preserves domain-invariant pathological patterns. We further employ a hierarchical differential prompt gating mechanism, dynamically selecting discriminative semantic prompts from both LVLM and LLM sources to address semantic confusion between adjacent DR grades. Finally, we utilize a two-stage prototype modulation strategy that progressively integrates clinical knowledge into visual prototypes through a Pathological Semantic Injector (PSI) and a Discriminative Prototype Enhancer (DPE). Extensive experiments across eight public datasets demonstrate that our approach achieves pathology-guided prototype evolution while outperforming state-of-the-art methods. The code is available at https://github.com/zhcz328/HAPM.

Method: PRAXIS (Procedural Recall for Agents with eXperiences Indexed by State) stores consequences of actions and retrieves them by jointly matching environmental and internal states of past episodes to current state. It augments agentic action selection with retrieved state-action-result exemplars generated in real time.

Result: On the REAL web browsing benchmark, PRAXIS improves task completion accuracy, reliability, and cost efficiency across different foundation model backbones. It also shows preliminary generalization to unseen tasks in similar environments.

Conclusion: PRAXIS enables practical adoption of AI agents in fast-evolving stateful environments by helping them learn new procedures effectively through lightweight post-training learning from experience.

Abstract: Learning how to do things from trial and error in real time is a hallmark of biological intelligence, yet most LLM-based agents lack mechanisms to acquire procedural knowledge after deployment. We propose Procedural Recall for Agents with eXperiences Indexed by State (PRAXIS), a lightweight post-training learning mechanism that stores the consequences of actions and retrieves them by jointly matching environmental and internal states of past episodes to the current state. PRAXIS augments agentic action selection with retrieved state-action-result exemplars that are generated in real time. When evaluated on the REAL web browsing benchmark, PRAXIS improves task completion accuracy, reliability, and cost efficiency across different foundation model backbones, and shows preliminary generalization to unseen tasks in similar environments. These results demonstrate that PRAXIS enables the practical adoption of AI agents in fast-evolving stateful environments by helping them learn new procedures effectively.

Method: Two-tier optimization: 1) Multi-leader multi-follower Stackelberg game where Mobile Agents and Fixed Agents set resource prices and Wireless Agents determine offloading ratios; 2) Double Dutch Auction for overloaded Fixed Agents to offload to Aerial Agents, solved using diffusion-based Deep Reinforcement Learning.

Result: Numerical results demonstrate superiority of the proposed scheme in facilitating efficient task offloading across the IoA architecture.

Conclusion: The proposed two-tier optimization approach effectively addresses the resource management challenges in IoA by leveraging game theory for ground-level coordination and auction mechanisms for aerial resource utilization, enabling efficient task offloading in heterogeneous agent networks.

Abstract: The Internet of Agents (IoA) is rapidly gaining prominence as a foundational architecture for interconnected intelligent systems, designed to facilitate seamless discovery, communication, and collaborative reasoning among a vast network of Artificial Intelligence (AI) agents. Powered by Large Language and Vision-Language Models, IoA enables the development of interactive, rational agents capable of complex cooperation, moving far beyond traditional isolated models. IoA involves physical entities, i.e., Wireless Agents (WAs) with limited onboard resources, which need to offload their compute-intensive agentic AI services to nearby servers. Such servers can be Mobile Agents (MAs), e.g., vehicle agents, or Fixed Agents (FAs), e.g., end-side units agents. Given their fixed geographical locations and stable connectivity, FAs can serve as reliable communication gateways and task aggregation points. This stability allows them to effectively coordinate with and offload to an Aerial Agent (AA) tier, which has an advantage not affordable for highly mobile MAs with dynamic connectivity limitations. As such, we propose a two-tier optimization approach. The first tier employs a multi-leader multi-follower Stackelberg game. In the game, MAs and FAs act as the leaders who set resource prices. WAs are the followers to determine task offloading ratios. However, when FAs become overloaded, they can further offload tasks to available aerial resources. Therefore, the second tier introduces a Double Dutch Auction model where overloaded FAs act as the buyers to request resources, and AAs serve as the sellers for resource provision. We then develop a diffusion-based Deep Reinforcement Learning algorithm to solve the model. Numerical results demonstrate the superiority of our proposed scheme in facilitating task offloading.

Method: Conducted an experiment where participants assessed two texts on the same topic (one AI-generated, one human-written) to test for perceptual biases in evaluating logical reasoning. Designed a questionnaire to quantify attitudes toward AI.

Result: Found significant bias in perception - evaluations of AI-generated texts’ logical reasoning were heavily influenced by preconceived views about AI’s reasoning abilities. Frequent AI users were less likely to believe AI usage undermines independent thinking.

Conclusion: Highlights the need to address perceptual biases to improve public understanding of AI’s capabilities and foster better human-AI interactions.

Abstract: Can machines think? This is a central question in artificial intelligence research. However, there is a substantial divergence of views on the answer to this question. Why do people have such significant differences of opinion, even when they are observing the same real world performance of artificial intelligence? The ability of logical reasoning like humans is often used as a criterion to assess whether a machine can think. This study explores whether human biases influence evaluations of the reasoning abilities of AI. An experiment was conducted where participants assessed two texts on the same topic, one AI generated and one human written,to test for perceptual biases in evaluating logical reasoning. Based on the experimental findings, a questionnaire was designed to quantify the attitudes toward AI.The results reveal a bias in perception. The evaluations of the logical reasoning ability of AI generated texts are significantly influenced by the preconceived views on the logical reasoning abilities of AI. Furthermore, frequent AI users were less likely to believe that AI usage undermines independent thinking.This study highlights the need to address perceptual biases to improve public understanding of AI’s capabilities and foster better human AI interactions.

Method: Created a benchmark with 2,520 carefully curated image-question-answer triplets spanning 7 diverse image domains (text-centric and general scenes), 10 cognitive task types (from basic recognition to reasoning), and 6 common wearable-specific image quality issues. Paired with an LLM-as-a-judge evaluation framework with 96% labeling accuracy.

Result: Open-source and proprietary multimodal LLMs achieved only 24-52% QA accuracy on WearVQA, with substantial performance drops on lower-quality images and reasoning-heavy tasks, demonstrating the benchmark’s challenging nature.

Conclusion: WearVQA serves as a comprehensive and challenging benchmark that reveals significant gaps in current multimodal AI systems for wearable applications, positioning it as a valuable tool for guiding technical advancement toward robust, real-world wearable AI systems.

Abstract: We introduce WearVQA, the first benchmark specifically designed to evaluate the Visual Question Answering (VQA) capabilities of multi-model AI assistant on wearable devices like smart glasses. Unlike prior benchmarks that focus on high-quality, third-person imagery, WearVQA reflects the unique challenges of ego-centric interaction-where visual inputs may be occluded, poorly lit, unzoomed, or blurry, and questions are grounded in realistic wearable use cases. The benchmark comprises 2,520 carefully curated image-question-answer triplets, spanning 7 diverse image domains including both text-centric and general scenes, 10 cognitive task types ranging from basic recognition to various forms of reasoning, and 6 common wearables-specific image quality issues. All questions are designed to be answerable using only the visual input and common senses. WearVQA is paired with a rigorous LLM-as-a-judge evaluation framework with 96% labeling accuracy. Open-source and proprietary multi-model LLMs achieved a QA accuracy as low as 24-52% on WearVQA, with substantial drops on lower-quality images and reasoning-heavy tasks. These observations position WearVQA as a comprehensive and challenging benchmark for guiding technical advancement towards robust, real-world multi-model wearables AI systems.

Method: Builds on universal AI (AIXI) to create a framework centered on self-prediction, where Bayesian RL agents predict both future perceptual inputs and their own actions. Extends AIXI theory to study universally intelligent embedded agents starting from Solomonoff priors.

Result: Self-prediction enables agents to reason about others running similar algorithms, leading to new game-theoretic solution concepts and novel forms of cooperation unattainable by classical decoupled agents. Idealized agents can form consistent mutual predictions and achieve infinite-order theory of mind.

Conclusion: The self-prediction framework provides a gold standard for embedded multi-agent learning, addressing fundamental challenges in non-stationary multi-agent environments by treating agents as embedded parts of their environment rather than decoupled entities.

Abstract: The standard theory of model-free reinforcement learning assumes that the environment dynamics are stationary and that agents are decoupled from their environment, such that policies are treated as being separate from the world they inhabit. This leads to theoretical challenges in the multi-agent setting where the non-stationarity induced by the learning of other agents demands prospective learning based on prediction models. To accurately model other agents, an agent must account for the fact that those other agents are, in turn, forming beliefs about it to predict its future behavior, motivating agents to model themselves as part of the environment. Here, building upon foundational work on universal artificial intelligence (AIXI), we introduce a mathematical framework for prospective learning and embedded agency centered on self-prediction, where Bayesian RL agents predict both future perceptual inputs and their own actions, and must therefore resolve epistemic uncertainty about themselves as part of the universe they inhabit. We show that in multi-agent settings, self-prediction enables agents to reason about others running similar algorithms, leading to new game-theoretic solution concepts and novel forms of cooperation unattainable by classical decoupled agents. Moreover, we extend the theory of AIXI, and study universally intelligent embedded agents which start from a Solomonoff prior. We show that these idealized agents can form consistent mutual predictions and achieve infinite-order theory of mind, potentially setting a gold standard for embedded multi-agent learning.

Method: Proposes CES multi-agent framework with staged execution-feedback reinforcement learning. Trains two high-level agents: Coordinator for strategic planning/task decomposition, and State Tracker for context compression/information management. Framework integrates with any low-level Executor model.

Result: CES significantly enhances planning and state management capabilities on long-horizon task benchmarks. The trained high-level scheduling module is generalizable and plug-and-play, improving various Executors’ long-horizon capabilities.

Conclusion: The CES framework effectively addresses challenges in long-horizon GUI tasks through specialized multi-agent architecture with separate high-level scheduling and state management components, demonstrating improved performance and generalizability.

Abstract: The rapid development of large vision-language model (VLM) has greatly promoted the research of GUI agent. However, GUI agents still face significant challenges in handling long-horizon tasks. First, single-agent models struggle to balance high-level capabilities and low-level execution capability, facing prevalent issues of responsibility coupling and capability conflicts. Second, agents lack awareness of the task state, leading to progress loss in long-horizon tasks. To address these challenges, we propose a staged execution-feedback reinforcement learning algorithm. Unlike training a unified policy model, we focus on training high-level scheduling models. Specifically, we propose and train two agents: a Coordinator, responsible for the strategic planning and task decomposition; and a State Tracker, responsible for context compression and information management to maintain the task’s state and coherence. Based on this, we built the Coordinator-Executor-State Tracker (CES) multi-agent framework, which can be integrated with any low-level Executor model, assisting the Executor in solving long-horizon tasks through task scheduling and state management. Experiments on long-horizon task benchmarks demonstrate that CES significantly enhances the system’s planning and state management capabilities. Furthermore, analysis confirms that our trained high-level scheduling module is a generalizable, plug-and-play module that significantly enhances the long-horizon capabilities of various Executors. Code can be available at https://github.com/hehehahi4/CES.

Method: Proposes a co-evolving agents framework with a target agent and auxiliary failure agent. The failure agent learns through preference optimization over failure trajectories from both agents, generating hard negatives that are close to success but remain failures. These hard negatives are incorporated into the target agent’s optimization to sharpen decision boundaries.

Result: Comprehensive analysis and experiments across benchmark datasets show improved performance. The method demonstrates that failures can be systematically transformed into structured and valuable learning signals rather than being used as-is.

Conclusion: The co-evolving agents framework effectively addresses overfitting in self-improving agents by leveraging failure trajectories to generate informative hard negatives, enhancing generalization and performance across diverse domains.

Abstract: The rapid progress of large foundation models has accelerated the development of task-specialized agents across diverse domains. However, the effectiveness of agents remains tightly coupled with the quality of training data, while curating task-specific datasets remains costly and often infeasible in real-world scenarios. Recent work has explored self-improving agents that autonomously generate, refine, and re-train on their own trajectories. A prominent line of approaches further leverages preference optimization by pairing predicted trajectories with scarce ground-truth trajectories, enabling agents to learn directly from their own failures. While these methods outperform supervised fine-tuning, their heavy reliance on predicted trajectories under limited ground-truth supervision leaves them prone to overfitting. To address this, we propose a co-evolving agents framework in which a target agent improves jointly with an auxiliary failure agent. The failure agent learns through preference optimization over failure trajectories from both the target and itself, thereby generating hard negatives that are close to success yet remain failures. Incorporating these informative hard negatives into the target agent’s optimization sharpens decision boundaries and enhances generalization. Our comprehensive analysis and experiments across benchmark datasets show that our method not only shows improved performance but also demonstrates that failures, instead of being used as-is, can be systematically transformed into structured and valuable learning signals in self-improving agents.

Method: Proposed RecToM benchmark with two dimensions: Cognitive Inference (understanding communicated mental states) and Behavioral Prediction (using inferred mental states to predict, select, and assess appropriate dialogue strategies). Evaluated state-of-the-art LLMs on this benchmark.

Result: LLMs show partial competence in recognizing mental states but struggle to maintain coherent, strategic ToM reasoning throughout dynamic recommendation dialogues, particularly in tracking evolving intentions and aligning conversational strategies with inferred mental states.

Conclusion: RecToM poses significant challenges for current LLMs, highlighting the gap between their ToM capabilities and human-like social reasoning needed for effective conversational recommendation systems.

Abstract: Large Language models are revolutionizing the conversational recommender systems through their impressive capabilities in instruction comprehension, reasoning, and human interaction. A core factor underlying effective recommendation dialogue is the ability to infer and reason about users’ mental states (such as desire, intention, and belief), a cognitive capacity commonly referred to as Theory of Mind. Despite growing interest in evaluating ToM in LLMs, current benchmarks predominantly rely on synthetic narratives inspired by Sally-Anne test, which emphasize physical perception and fail to capture the complexity of mental state inference in realistic conversational settings. Moreover, existing benchmarks often overlook a critical component of human ToM: behavioral prediction, the ability to use inferred mental states to guide strategic decision-making and select appropriate conversational actions for future interactions. To better align LLM-based ToM evaluation with human-like social reasoning, we propose RecToM, a novel benchmark for evaluating ToM abilities in recommendation dialogues. RecToM focuses on two complementary dimensions: Cognitive Inference and Behavioral Prediction. The former focus on understanding what has been communicated by inferring the underlying mental states. The latter emphasizes what should be done next, evaluating whether LLMs can leverage these inferred mental states to predict, select, and assess appropriate dialogue strategies. Extensive experiments on state-of-the-art LLMs demonstrate that RecToM poses a significant challenge. While the models exhibit partial competence in recognizing mental states, they struggle to maintain coherent, strategic ToM reasoning throughout dynamic recommendation dialogues, particularly in tracking evolving intentions and aligning conversational strategies with inferred mental states.

Method: The framework combines game theory with statistical techniques and approximate solutions. It enables bounded rational decision-making while maintaining recursive Theory of Mind (each agent holding ToM about others, who in turn hold ToM about the rest). The approach focuses on computability of the inherent computational problem.

Result: The paper presents a computational framework that provides a different perspective from existing psychological and logical approaches, offering a practical way to implement ToM-based reasoning through game-theoretic modeling with computational tractability.

Conclusion: The proposed game-theoretic framework with statistical approximations offers a viable computational approach to Theory of Mind that balances formal rigor with practical computability, addressing limitations in both psychological and logical approaches to ToM.

Abstract: Originating in psychology, $\textit{Theory of Mind}$ (ToM) has attracted significant attention across multiple research communities, especially logic, economics, and robotics. Most psychological work does not aim at formalizing those central concepts, namely $\textit{goals}$, $\textit{intentions}$, and $\textit{beliefs}$, to automate a ToM-based computational process, which, by contrast, has been extensively studied by logicians. In this paper, we offer a different perspective by proposing a computational framework viewed through the lens of game theory. On the one hand, the framework prescribes how to make boudedly rational decisions while maintaining a theory of mind about others (and recursively, each of the others holding a theory of mind about the rest); on the other hand, it employs statistical techniques and approximate solutions to retain computability of the inherent computational problem.

Method: AI-assisted workflow combining deep learning for initial parameter estimates with Bayesian optimization for iterative refinement, plus an active learning variant for expert assistance when desired.

Result: The approach accelerates design space exploration while reducing expert involvement, demonstrated successfully on a sheet metal forming process.

Conclusion: AI-assisted optimization workflows can significantly reduce the cost and environmental impact of industrial simulations while maintaining or improving design optimization efficiency.

Abstract: Numerical simulations have revolutionized the industrial design process by reducing prototyping costs, design iterations, and enabling product engineers to explore the design space more efficiently. However, the growing scale of simulations demands substantial expert knowledge, computational resources, and time. A key challenge is identifying input parameters that yield optimal results, as iterative simulations are costly and can have a large environmental impact. This paper presents an AI-assisted workflow that reduces expert involvement in parameter optimization through the use of Bayesian optimization. Furthermore, we present an active learning variant of the approach, assisting the expert if desired. A deep learning model provides an initial parameter estimate, from which the optimization cycle iteratively refines the design until a termination condition (e.g., energy budget or iteration limit) is met. We demonstrate our approach, based on a sheet metal forming process, and show how it enables us to accelerate the exploration of the design space while reducing the need for expert involvement.

Method: Multi-agent framework integrating three feedback sources: LLM self-evaluation, domain knowledge-informed feedback, and ML surrogate-based feedback to guide text gradient-driven composition generation.

Result: Achieved over 98% compositional validity and up to 54% stable/metastable candidates, surpassing LLM-only baseline (43%) and prior GAN-based results (27%). ML gradients work well in-distribution but unreliable out-of-distribution.

Conclusion: First systematic analysis of knowledge-guided text gradients for DP discovery, establishing a generalizable blueprint for multi-agent system-driven generative materials design for sustainable technologies.

Abstract: Double perovskites (DPs) are promising candidates for sustainable energy technologies due to their compositional tunability and compatibility with low-energy fabrication, yet their vast design space poses a major challenge for conditional materials discovery. This work introduces a multi-agent, text gradient-driven framework that performs DP composition generation under natural-language conditions by integrating three complementary feedback sources: LLM-based self-evaluation, DP-specific domain knowledge-informed feedback, and ML surrogate-based feedback. Analogous to how knowledge-informed machine learning improves the reliability of conventional data-driven models, our framework incorporates domain-informed text gradients to guide the generative process toward physically meaningful regions of the DP composition space. Systematic comparison of three incremental configurations, (i) pure LLM generation, (ii) LLM generation with LLM reasoning-based feedback, and (iii) LLM generation with domain knowledge-guided feedback, shows that iterative guidance from knowledge-informed gradients improves stability-condition satisfaction without additional training data, achieving over 98% compositional validity and up to 54% stable or metastable candidates, surpassing both the LLM-only baseline (43%) and prior GAN-based results (27%). Analyses of ML-based gradients further reveal that they enhance performance in in-distribution (ID) regions but become unreliable in out-of-distribution (OOD) regimes. Overall, this work provides the first systematic analysis of multi-agent, knowledge-guided text gradients for DP discovery and establishes a generalizable blueprint for MAS-driven generative materials design aimed at advancing sustainable technologies.

Method: Swarm intelligence-inspired framework with multiple LLM agents operating in parallel, each assigned to specific residue positions. Agents iteratively propose context-aware mutations by integrating design objectives, local neighborhood interactions, and memory/feedback from previous iterations.

Result: Achieves efficient, objective-directed protein designs within few GPU-hours without fine-tuning. Validated experimentally on alpha helix and coil structures. Demonstrates emergent behaviors and effective navigation of protein fitness landscape through residue conservation, structure-based metrics, and sequence convergence analyses.

Conclusion: Provides a generalizable, adaptable solution for protein design that lays groundwork for collective LLM-driven design across biomolecular systems and other scientific discovery tasks.

Abstract: Designing proteins de novo with tailored structural, physicochemical, and functional properties remains a grand challenge in biotechnology, medicine, and materials science, due to the vastness of sequence space and the complex coupling between sequence, structure, and function. Current state-of-the-art generative methods, such as protein language models (PLMs) and diffusion-based architectures, often require extensive fine-tuning, task-specific data, or model reconfiguration to support objective-directed design, thereby limiting their flexibility and scalability. To overcome these limitations, we present a decentralized, agent-based framework inspired by swarm intelligence for de novo protein design. In this approach, multiple large language model (LLM) agents operate in parallel, each assigned to a specific residue position. These agents iteratively propose context-aware mutations by integrating design objectives, local neighborhood interactions, and memory and feedback from previous iterations. This position-wise, decentralized coordination enables emergent design of diverse, well-defined sequences without reliance on motif scaffolds or multiple sequence alignments, validated with experiments on proteins with alpha helix and coil structures. Through analyses of residue conservation, structure-based metrics, and sequence convergence and embeddings, we demonstrate that the framework exhibits emergent behaviors and effective navigation of the protein fitness landscape. Our method achieves efficient, objective-directed designs within a few GPU-hours and operates entirely without fine-tuning or specialized training, offering a generalizable and adaptable solution for protein design. Beyond proteins, the approach lays the groundwork for collective LLM-driven design across biomolecular systems and other scientific discovery tasks.

Method: ECO-GROW integrates industrial linkages, POI similarities, migration similarities, and temporal network evolution over 15 years. It uses a Dynamic Top-K GCN to adaptively select influential inter-city connections and an adaptive Graph Scorer mechanism to dynamically weight cross-regional impacts. The model also incorporates a link prediction task based on Barabasi Proximity to optimize graph representation.

Result: Experimental results show ECO-GROW’s superior accuracy in predicting entrepreneurial activities and employment trends compared to conventional models. The framework successfully models China’s inter-city networks from 2005-2021 to generate urban embeddings that effectively capture economic vitality.

Conclusion: ECO-GROW provides a powerful framework for modeling urban economic vitality by capturing dynamic structural similarities between cities. By open-sourcing the code, the authors enable government agencies and public sector organizations to leverage big data analytics for evidence-based urban planning, economic policy formulation, and resource allocation decisions.

Abstract: Urban economic vitality is a crucial indicator of a city’s long-term growth potential, comprising key metrics such as the annual number of new companies and the population employed. However, modeling urban economic vitality remains challenging. This study develops ECO-GROW, a multi-graph framework modeling China’s inter-city networks (2005-2021) to generate urban embeddings that model urban economic vitality. Traditional approaches relying on static city-level aggregates fail to capture a fundamental dynamic: the developmental trajectory of one city today may mirror that of its structurally similar counterparts tomorrow. ECO-GROW overcomes this limitation by integrating industrial linkages, POI similarities, migration similarities and temporal network evolution over 15 years. The framework combines a Dynamic Top-K GCN to adaptively select influential inter-city connections and an adaptive Graph Scorer mechanism to dynamically weight cross-regional impacts. Additionally, the model incorporates a link prediction task based on Barabasi Proximity, optimizing the graph representation. Experimental results demonstrate ECO-GROW’s superior accuracy in predicting entrepreneurial activities and employment trends compared to conventional models. By open-sourcing our code, we enable government agencies and public sector organizations to leverage big data analytics for evidence-based urban planning, economic policy formulation, and resource allocation decisions that benefit society at large.

Method: Shifts LLM interaction from raw data to memory pointers, preserving complete tool outputs without context window limitations. This maintains tool functionality while enabling seamless integration into agentic workflows.

Result: Validated on real-world Materials Science application that conventional workflows cannot execute. Comparative analysis shows the method consumes ~7x fewer tokens than traditional workflow while preserving all information.

Conclusion: The pointer-based approach enables LLMs to process arbitrarily long tool outputs without information loss, significantly reducing computational costs while maintaining complete data access for knowledge-intensive domains.

Abstract: Large Language Models (LLMs) have become increasingly capable of interacting with external tools, granting access to specialized knowledge beyond their training data - critical in dynamic, knowledge-intensive domains such as Chemistry and Materials Science. However, large tool outputs can overflow the LLMs’ context window, preventing task completion. Existing solutions such as truncation or summarization fail to preserve complete outputs, making them unsuitable for workflows requiring the full data. This work introduces a method that enables LLMs to process and utilize tool responses of arbitrary length without loss of information. By shifting the model’s interaction from raw data to memory pointers, the method preserves tool functionality, allows seamless integration into agentic workflows, and reduces token usage and execution time. The proposed method is validated on a real-world Materials Science application that cannot be executed with conventional workflows, and its effectiveness is demonstrated via a comparative analysis where both methods succeed. In this experiment, the proposed approach consumed approximately seven times fewer tokens than the traditional workflow.

Method: Using methods from mathematical logic, specifically computability theory and set theory, to analyze the grounded extension as the least fixed-point of a defense operator, examining transfinite iterations required for its computation.

Result: Identified the exact ordinal number corresponding to the length of the iterative process for computing the grounded extension, and determined that deciding grounded acceptance is maximally complex in infinite cases.

Conclusion: There’s a marked distinction between finite and infinite cases: while the grounded extension is polynomial-time computable in finite argumentation frameworks, it becomes maximally complex in infinite cases, unlike other reasoning problems in formal argumentation.

Abstract: Argumentation frameworks, consisting of arguments and an attack relation representing conflicts, are fundamental for formally studying reasoning under conflicting information. We use methods from mathematical logic, specifically computability and set theory, to analyze the grounded extension, a widely-used model of maximally skeptical reasoning, defined as the least fixed-point of a natural defense operator. Without additional constraints, finding this fixed-point requires transfinite iterations. We identify the exact ordinal number corresponding to the length of this iterative process and determine the complexity of deciding grounded acceptance, showing it to be maximally complex. This shows a marked distinction from the finite case where the grounded extension is polynomial-time computable, thus simpler than other reasoning problems explored in formal argumentation.

Method: Developed an agentic AI system using autonomous but cooperative agents that reason, sense, and act throughout the entire event lifecycle. The framework integrates sensing, forecasting, downscaling, hydrological modeling and coordinated response into a single interconnected system with embedded learning layers for adaptive recalibration.

Result: Multi-year evaluation in northern Pakistan showed the multi-agent configuration enhances forecast reliability, critical success index, and warning lead time compared to baseline models. Maximized population reach and minimized evacuation errors through communication/routing agents, with adaptive recalibration and transparent auditability.

Conclusion: Collaborative AI agents can transform atmospheric data streams into practicable foresight and provide a scalable platform for adaptive, learning-based climate resilience, effectively bridging the gap between weather prediction and emergency response.

Abstract: The challenge is growing towards extreme and short-duration rainfall events like a cloudburst that are peculiar to the traditional forecasting systems, in which the predictions and the response are taken as two distinct processes. The paper outlines an agentic artificial intelligence system to study atmospheric water-cycle intelligence, which combines sensing, forecasting, downscaling, hydrological modeling and coordinated response into a single, interconnected, priceless, closed-loop system. The framework uses autonomous but cooperative agents that reason, sense, and act throughout the entire event lifecycle, and use the intelligence of weather prediction to become real-time decision intelligence. Comparison of multi-year radar, satellite, and ground-based evaluation of the northern part of Pakistan demonstrates that the multi-agent configuration enhances forecast reliability, critical success index and warning lead time compared to the baseline models. Population reach was maximised, and errors during evacuation were minimised through communication and routing agents, and adaptive recalibration and transparent auditability were provided by the embedded layer of learning. Collectively, this leads to the conclusion that collaborative AI agents are capable of transforming atmospheric data streams into practicable foresight and provide a platform of scalable adaptive and learning-based climate resilience.

Method: Analyze minimal revision’s learning power, characterize prior plausibility assignments that enable learning via minimal revision, conditioning, and lexicographic upgrade, and examine learning from possibly erroneous information.

Result: Minimal revision can learn any finitely identifiable problem and learn with positive/negative data when considering finitely many possibilities. Characterizations of prior plausibility assignments for different belief revision methods are provided, with some limitations when learning from erroneous information.

Conclusion: While minimal revision has learning limitations compared to less conservative methods, it remains a successful learning method in many scenarios, particularly with finite possibilities and identifiable problems, though not all results extend to learning from erroneous information.

Abstract: The principle of minimal change in belief revision theory requires that, when accepting new information, one keeps one’s belief state as close to the initial belief state as possible. This is precisely what the method known as minimal revision does. However, unlike less conservative belief revision methods, minimal revision falls short in learning power: It cannot learn everything that can be learned by other learning methods. We begin by showing that, despite this limitation, minimal revision is still a successful learning method in a wide range of situations. Firstly, it can learn any problem that is finitely identifiable. Secondly, it can learn with positive and negative data, as long as one considers finitely many possibilities. We then characterize the prior plausibility assignments (over finitely many possibilities) that enable one to learn via minimal revision, and do the same for conditioning and lexicographic upgrade. Finally, we show that not all of our results still hold when learning from possibly erroneous information.

Method: Developed a pipeline that uses Vision-Language Models (VLMs) to process BPMN diagram images and extract structured JSON representations. Incorporated optical character recognition (OCR) for textual enrichment of diagram elements. Evaluated against ground truth data from source XML files.

Result: The approach enables robust component extraction when original source files are unavailable. Multiple VLMs were benchmarked, showing performance improvements when OCR is used for text enrichment. Extensive statistical analyses of OCR-based enrichment methods and prompt ablation studies provided insights into their impact on model performance.

Conclusion: The proposed pipeline successfully extracts structured JSON representations of BPMN diagrams directly from images using VLMs, with OCR enrichment enhancing performance. This provides a practical solution for analyzing BPMN diagrams when source files are unavailable, with comprehensive evaluation showing the effectiveness of the approach.

Abstract: Business Process Model and Notation (BPMN) is a widely adopted standard for representing complex business workflows. While BPMN diagrams are often exchanged as visual images, existing methods primarily rely on XML representations for computational analysis. In this work, we present a pipeline that leverages Vision-Language Models (VLMs) to extract structured JSON representations of BPMN diagrams directly from images, without requiring source model files or textual annotations. We also incorporate optical character recognition (OCR) for textual enrichment and evaluate the generated element lists against ground truth data derived from the source XML files. Our approach enables robust component extraction in scenarios where original source files are unavailable. We benchmark multiple VLMs and observe performance improvements in several models when OCR is used for text enrichment. In addition, we conducted extensive statistical analyses of OCR-based enrichment methods and prompt ablation studies, providing a clearer understanding of their impact on model performance.

Method: Systematic comparison of neural CQA models with a training-free query relaxation strategy that retrieves answers by relaxing query constraints and counting resulting paths across multiple datasets and query structures.

Result: Neural models don’t consistently outperform query relaxation; both approaches perform similarly in many cases. Their retrieved answers show little overlap, and combining their outputs consistently improves performance.

Conclusion: Current neural CQA models fail to subsume reasoning patterns captured by query relaxation, calling for re-evaluation of progress in neural query answering. Future neural approaches should incorporate query relaxation principles, and stronger non-neural baselines are needed.

Abstract: Neural methods for Complex Query Answering (CQA) over knowledge graphs (KGs) are widely believed to learn patterns that generalize beyond explicit graph structure, allowing them to infer answers that are unreachable through symbolic query processing. In this work, we critically examine this assumption through a systematic analysis comparing neural CQA models with an alternative, training-free query relaxation strategy that retrieves possible answers by relaxing query constraints and counting resulting paths. Across multiple datasets and query structures, we find several cases where neural and relaxation-based approaches perform similarly, with no neural model consistently outperforming the latter. Moreover, a similarity analysis reveals that their retrieved answers exhibit little overlap, and that combining their outputs consistently improves performance. These results call for a re-evaluation of progress in neural query answering: despite their complexity, current models fail to subsume the reasoning patterns captured by query relaxation. Our findings highlight the importance of stronger non-neural baselines and suggest that future neural approaches could benefit from incorporating principles of query relaxation.

Method: 1) Train an accurate LLM-based verifier for theorem proving. 2) Train a proof generator using the verifier as reward model, incentivizing it to identify and resolve issues in its own proofs before finalizing. 3) Scale verification compute to automatically label new hard-to-verify proofs as the generator improves, creating training data to further improve the verifier (maintaining generation-verification gap).

Result: DeepSeekMath-V2 achieves gold-level scores on IMO 2025 and CMO 2024, and near-perfect 118/120 on Putnam 2024 with scaled test-time compute.

Conclusion: Self-verifiable mathematical reasoning through iterative improvement of verifiers and generators enables deep reasoning capabilities beyond final-answer-focused approaches, achieving state-of-the-art performance in theorem proving competitions.

Abstract: Large language models have made significant progress in mathematical reasoning, which serves as an important testbed for AI and could impact scientific research if further advanced. By scaling reasoning with reinforcement learning that rewards correct final answers, LLMs have improved from poor performance to saturating quantitative reasoning competitions like AIME and HMMT in one year. However, this approach faces fundamental limitations. Pursuing higher final answer accuracy doesn’t address a key issue: correct answers don’t guarantee correct reasoning. Moreover, many mathematical tasks like theorem proving require rigorous step-by-step derivation rather than numerical answers, making final answer rewards inapplicable. To push the limits of deep reasoning, we believe it is necessary to verify the comprehensiveness and rigor of mathematical reasoning. Self-verification is particularly important for scaling test-time compute, especially for open problems without known solutions. Towards self-verifiable mathematical reasoning, we investigate how to train an accurate and faithful LLM-based verifier for theorem proving. We then train a proof generator using the verifier as the reward model, and incentivize the generator to identify and resolve as many issues as possible in their own proofs before finalizing them. To maintain the generation-verification gap as the generator becomes stronger, we propose to scale verification compute to automatically label new hard-to-verify proofs, creating training data to further improve the verifier. Our resulting model, DeepSeekMath-V2, demonstrates strong theorem-proving capabilities, achieving gold-level scores on IMO 2025 and CMO 2024 and a near-perfect 118/120 on Putnam 2024 with scaled test-time compute.

Method: Agentic AI model combining demand forecasting, supplier selection optimization, multi-agent negotiation, and continuous learning. Prototype tested in a middle-scale mart with three conventional and artificial data tables, compared against base heuristics.

Result: System demonstrates decreased stockouts, reduced inventory holding costs, and improved product mix turnover compared to conventional heuristics.

Conclusion: The agentic AI approach shows promise for retail inventory optimization, though constraints, scalability, and improvement prospects need further addressing.

Abstract: In contemporary retail, the variety of products available (e.g. clothing, groceries, cosmetics, frozen goods) make it difficult to predict the demand, prevent stockouts, and find high-potential products. We suggest an agentic AI model that will be used to monitor the inventory, initiate purchase attempts to the appropriate suppliers, and scan for trending or high-margin products to incorporate. The system applies demand forecasting, supplier selection optimization, multi-agent negotiation and continuous learning. We apply a prototype to a setting in the store of a middle scale mart, test its performance on three conventional and artificial data tables, and compare the results to the base heuristics. Our findings indicate that there is a decrease in stockouts, a reduction of inventory holding costs, and an improvement in product mix turnover. We address constraints, scalability as well as improvement prospect.

Method: The paper provides a comprehensive overview using a “deception cycle” framework with two components: deception emergence (analyzing incentive foundations, capability preconditions, and contextual triggers) and deception treatment (detection methods, mitigation strategies, and integrated auditing approaches).

Result: The analysis reveals that systems with sufficient capability and incentive potential inevitably engage in deceptive behaviors when triggered by external conditions like supervision gaps, distributional shifts, and environmental pressures.

Conclusion: AI deception requires integrated technical, community, and governance efforts for mitigation, with the field needing ongoing research supported by living resources like the released deception survey website.

Abstract: As intelligence increases, so does its shadow. AI deception, in which systems induce false beliefs to secure self-beneficial outcomes, has evolved from a speculative concern to an empirically demonstrated risk across language models, AI agents, and emerging frontier systems. This project provides a comprehensive and up-to-date overview of the AI deception field, covering its core concepts, methodologies, genesis, and potential mitigations. First, we identify a formal definition of AI deception, grounded in signaling theory from studies of animal deception. We then review existing empirical studies and associated risks, highlighting deception as a sociotechnical safety challenge. We organize the landscape of AI deception research as a deception cycle, consisting of two key components: deception emergence and deception treatment. Deception emergence reveals the mechanisms underlying AI deception: systems with sufficient capability and incentive potential inevitably engage in deceptive behaviors when triggered by external conditions. Deception treatment, in turn, focuses on detecting and addressing such behaviors. On deception emergence, we analyze incentive foundations across three hierarchical levels and identify three essential capability preconditions required for deception. We further examine contextual triggers, including supervision gaps, distributional shifts, and environmental pressures. On deception treatment, we conclude detection methods covering benchmarks and evaluation protocols in static and interactive settings. Building on the three core factors of deception emergence, we outline potential mitigation strategies and propose auditing approaches that integrate technical, community, and governance efforts to address sociotechnical challenges and future AI risks. To support ongoing work in this area, we release a living resource at www.deceptionsurvey.com.

Method: Multi-phase approach dividing scheduling into smaller sub-problems: day allocation and shift allocation, each modeled as Integer Programming Problems (IPP). Solutions from earlier phases feed into subsequent phases, using multi-objective framework.

Result: Significantly reduces computational variables, allows targeted objective functions, enhances both efficiency and accuracy compared to traditional single-step approaches.

Conclusion: The proposed multi-phase allocation method effectively addresses scalability and accuracy challenges in agent scheduling, particularly useful for managing peak demand scenarios in contact centers with limited resources.

Abstract: Effective agent shift scheduling is crucial for businesses, especially in the Contact Center as a Service (CCaaS) industry, to ensure seamless operations and fulfill employee needs. Most studies utilizing mathematical model-based solutions approach the problem as a single-step process, often resulting in inefficiencies and high computational demands. In contrast, we present a multi-phase allocation method that addresses scalability and accuracy by dividing the problem into smaller sub-problems of day and shift allocation, which significantly reduces number of computational variables and allows for targeted objective functions, ultimately enhancing both efficiency and accuracy. Each subproblem is modeled as a Integer Programming Problem (IPP), with solutions sequentially feeding into the subsequent subproblem. We then apply the proposed method, using a multi-objective framework, to address the difficulties posed by peak demand scenarios such as holiday rushes, where maintaining service levels is essential despite having limited number of employees

Method: GCA decouples VLM’s role into two stages: 1) Semantic analyst - translates ambiguous queries into formal, verifiable task constraints defining reference frame and objective; 2) Task solver - generates and executes tool calls strictly within deterministic bounds defined by constraints.

Result: GCA achieves state-of-the-art performance on multiple spatial reasoning benchmarks, surpassing existing training-based and tool-integrated methods by approximately 27%.

Conclusion: The geometrically-constrained reasoning strategy successfully resolves the semantic-to-geometric gap, yielding a robust and verifiable reasoning pathway for spatial reasoning without requiring training.

Abstract: Vision Language Models (VLMs) exhibit a fundamental semantic-to-geometric gap in spatial reasoning: they excel at qualitative semantic inference but their reasoning operates within a lossy semantic space, misaligned with high-fidelity geometry. Current paradigms fail to bridge this gap. Training-based methods suffer from an ``oracle paradox,’’ learning flawed spatial logic from imperfect oracles. Tool-integrated methods constrain the final computation but critically leave the VLM’s planning process unconstrained, resulting in geometrically flawed plans. In this work, we propose Geometrically-Constrained Agent (GCA), a training-free agentic paradigm that resolves this gap by introducing a formal task constraint. Specifically, we strategically decouples the VLM’s role into two stages. First, acting as a semantic analyst, the VLM translates the user’s ambiguous query into the formal, verifiable task constraint, which defines the reference frame and objective. Second, acting as a task solver, the VLM generates and executes tool calls strictly within the deterministic bounds defined by the constraint. This geometrically-constrained reasoning strategy successfully resolve the semantic-to-geometric gap, yielding a robust and verifiable reasoning pathway for spatial reasoning. Comprehensive experiments demonstrate that GCA achieves SOTA performance on multiple spatial reasoning benchmarks, surpassing existing training-based and tool-integrated methods by ~27%. Please see our homepage at https://gca-spatial-reasoning.github.io.

Method: Multi-layer architecture with Application/Interface Layer, Agents Layer, and Data Source Layer. Four specialized agents coordinated by hybrid reasoning engine: Meal Planner Agent (nutrition), Reminder Agent (scheduling), Food Guidance Agent (grocery/cooking assistance), Monitoring Agent (tracking). Agents communicate via Blackboard/Event Bus with real-time feedback. Privacy-sensitive data sources (EHRs, wearables, IoT) in policy-controlled layer. Includes XAI modules and clinician dashboards.

Result: Proposes a comprehensive agentic AI framework that integrates multi-agent reasoning, multi-modal interfaces, and human-centered design to provide adaptive, transparent, and inclusive support for people with disabilities and neurodivergence.

Conclusion: The framework represents an advancement beyond traditional assistive systems by incorporating inclusiveness, personalization, and accessibility at all levels, promoting autonomy, health, and digital equity for people with disabilities and neurodivergence through the intersection of multi-agent reasoning, multi-modal interfaces, and human-centered design.

Abstract: The paper presents a detailed Agentic Artificial Intelligence (AI) model that would enable people with disabilities and neurodivergence to lead healthier lives and have more regular days. The system will use a multi-layer structure; it will include an Application and Interface Layer, an Agents Layer, and a Data Source Layer to provide adaptive, transparent, and inclusive support. Fundamentally, a hybrid reasoning engine will synchronize four special-purpose agents, which include: a personalized-nutrition-based, called a Meal Planner Agent; an adaptive-scheduling-based, called a Reminder Agent; interactive assistance during grocery shopping and cooking, called a Food Guidance Agent; and a continuous-intake-and-physiological-tracking, called a Monitoring Agent. All the agents interact through a central communicative system called the Blackboard/Event Bus, which allows autonomous interaction and real-time feedback loops with multimedia user interfaces. Privacy-sensitive data sources, including electronic health records (EHRs), nutritional databases, wearable sensors, and smart kitchen Internet of Things, are also included in the framework and placed into a policy-controlled layer, which ensures data safety and compliance with consent. Collaborative care and clinician dashboards allow common supervision, and discussable artificial intelligence (XAI) modules give brief explanations of why a decision was made, making users responsible and reliant. The proposed agentic AI framework is an extension beyond traditional assistive systems since it incorporates inclusiveness, personalization, and accessibility at all levels. It displays the intersection of multi-agent reasoning, multi-modal interfaces, and human-centered design that will enable the development of autonomy, health, and digital equity among people with disabilities and neurodivergence.

Method: Two key innovations: (1) automatic selection of effective Hankel embedding model order via data-driven singular-value thresholding to identify informative subspace and discard noise, and (2) novel optimization procedure replacing exact orthogonality constraint with lightweight process to keep search near orthogonal transformation space while finding minimal distance between dynamics matrices.

Result: fastDSA is at least an order of magnitude faster than previous methods while maintaining their accuracy, robustness, and properties including invariances and sensitivities to system dynamics.

Conclusion: fastDSA provides a computationally efficient and accurate method for dynamical similarity analysis, enabling practical comparison of neural systems while preserving analysis of temporal structure.

Abstract: To understand how neural systems process information, it is often essential to compare one circuit with another, one brain with another, or data with a model. Traditional similarity measures ignore the dynamical processes underlying neural representations. Dynamical similarity methods offer a framework to compare the temporal structure of dynamical systems by embedding their (possibly) nonlinear dynamics into a globally linear space and there computing conjugacy metrics. However, identifying the best embedding and computing these metrics can be computationally slow. Here we introduce fast Dynamical Similarity Analysis (fastDSA), which is computationally far more efficient than previous methods while maintaining their accuracy and robustness. FastDSA introduces two key components that boost efficiency: (1) automatic selection of the effective model order of the Hankel (delay) embedding from the data via a data-driven singular-value threshold that identifies the informative subspace and discards noise to lower computational cost without sacrificing signal, and (2) a novel optimization procedure and objective, which replaces the slow exact orthogonality constraint in finding a minimal distance between dynamics matrices with a lightweight process to keep the search close to the space of orthogonal transformations. We demonstrate that fastDSA is at least an order of magnitude faster than the previous methods. Furthermore, we demonstrate that fastDSA has the properties of its ancestor, including its invariances and sensitivities to system dynamics. FastDSA, therefore, provides a computationally efficient and accurate method for dynamical similarity analysis.

Method: Developed a simulation of interconnected intersections with random vehicle flows. Implemented decentralized multi-agent reinforcement learning (MARL) where each traffic signal acts as an autonomous agent making decisions based on local observations and neighbor information.

Result: MARL approach showed statistically significant improvements over baseline fixed-time controller, including reduced average waiting times and improved throughput.

Conclusion: MARL-based dynamic control strategies show substantial promise for improving urban traffic management efficiency, though more research is needed for scalability and real-world implementation challenges.

Abstract: Urban traffic congestion, particularly at intersections, significantly affects travel time, fuel consumption, and emissions. Traditional fixed-time signal control systems often lack the adaptability to effectively manage dynamic traffic patterns. This study explores the application of multi-agent reinforcement learning (MARL) to optimize traffic signal coordination across multiple intersections within a simulated environment. A simulation was developed to model a network of interconnected intersections with randomly generated vehicle flows to reflect realistic traffic variability. A decentralized MARL controller was implemented in which each traffic signal operates as an autonomous agent, making decisions based on local observations and information from neighboring agents. Performance was evaluated against a baseline fixed-time controller using metrics such as average vehicle wait time and overall throughput. The MARL approach demonstrated statistically significant improvements, including reduced average waiting times and improved throughput. These findings suggest that MARL-based dynamic control strategies hold substantial promise to improve urban traffic management efficiency. More research is recommended to address the challenges of scalability and real-world implementation.

Method: The authors: 1) Thoroughly investigate shortcomings in InsightBench, 2) Propose essential criteria for a high-quality insight benchmark, 3) Develop a data-curation pipeline to construct InsightEval dataset, and 4) Introduce a novel metric to measure exploratory performance of agents.

Result: The paper presents InsightEval as a new benchmark that addresses the identified flaws. Through extensive experiments, the authors highlight prevailing challenges in automated insight discovery and provide key findings to guide future research.

Conclusion: The proposed InsightEval benchmark with its improved data quality and novel evaluation metric provides a more reliable framework for assessing insight discovery capabilities, addressing critical flaws in existing benchmarks and advancing research in automated data analysis.

Abstract: Data analysis has become an indispensable part of scientific research. To discover the latent knowledge and insights hidden within massive datasets, we need to perform deep exploratory analysis to realize their full value. With the advent of large language models (LLMs) and multi-agent systems, more and more researchers are making use of these technologies for insight discovery. However, there are few benchmarks for evaluating insight discovery capabilities. As one of the most comprehensive existing frameworks, InsightBench also suffers from many critical flaws: format inconsistencies, poorly conceived objectives, and redundant insights. These issues may significantly affect the quality of data and the evaluation of agents. To address these issues, we thoroughly investigate shortcomings in InsightBench and propose essential criteria for a high-quality insight benchmark. Regarding this, we develop a data-curation pipeline to construct a new dataset named InsightEval. We further introduce a novel metric to measure the exploratory performance of agents. Through extensive experiments on InsightEval, we highlight prevailing challenges in automated insight discovery and raise some key findings to guide future research in this promising direction.

Method: 1) Introduce Mentalese framework that encodes abstract reasoning as ultra-compressed, structured tokens. 2) Propose SHORTER LENGTH PREFERENCE OPTIMIZATION (SLPO), a reinforcement learning method that rewards concise correct solutions while allowing longer reasoning when needed.

Result: ORION models achieve: 4-16x fewer tokens in reasoning traces, up to 5x lower inference latency, 7-9x reduction in training costs relative to DeepSeek R1 Distilled, while maintaining 90-98% of its accuracy. Outperforms Claude and ChatGPT-4o by up to 5% accuracy with 2x compression.

Conclusion: Mentalese-style compressed reasoning enables human-like cognitive efficiency, offering real-time, cost-effective reasoning without sacrificing accuracy, representing a step toward more efficient AI reasoning systems.

Abstract: Large Reasoning Models (LRMs) achieve strong performance in mathematics, code generation, and task planning, but their reliance on long chains of verbose “thinking” tokens leads to high latency, redundancy, and incoherent reasoning paths. Inspired by the Language of Thought Hypothesis, which posits that human reasoning operates over a symbolic, compositional mental language called Mentalese, we introduce a framework that trains models to reason in a similarly compact style. Mentalese encodes abstract reasoning as ultra-compressed, structured tokens, enabling models to solve complex problems with far fewer steps. To improve both efficiency and accuracy, we propose SHORTER LENGTH PREFERENCE OPTIMIZATION (SLPO), a reinforcement learning method that rewards concise solutions that stay correct, while still allowing longer reasoning when needed. Applied to Mentalese-aligned models, SLPO yields significantly higher compression rates by enabling concise reasoning that preserves the benefits of detailed thinking without the computational overhead. Across benchmarks including AIME 2024 and 2025, MinervaMath, OlympiadBench, Math500, and AMC, our ORION models produce reasoning traces with 4-16x fewer tokens, achieve up to 5x lower inference latency, and reduce training costs by 7-9x relative to the DeepSeek R1 Distilled model, while maintaining 90-98% of its accuracy. ORION also surpasses Claude and ChatGPT-4o by up to 5% in accuracy while maintaining 2x compression. These results show that Mentalese-style compressed reasoning offers a step toward human-like cognitive efficiency, enabling real-time, cost-effective reasoning without sacrificing accuracy.

Method: TIM-PRM transforms verification into an active, tool-augmented investigation. It’s trained to explicitly plan verification strategies and uses Independent Question Asking to query evidence via external tools, effectively decoupling verification from reasoning context to eliminate confirmation bias. The method is instantiated by curating a high-quality dataset of tool-integrated verification trajectories.

Result: Extensive experiments on VisualProcessBench show that the 8B parameter model surpasses existing open-source multimodal PRMs, significantly outperforming much larger models like Qwen2.5-72B and InternVL-78B, while offering interpretable insights into the verification process.

Conclusion: TIM-PRM successfully bridges the gap in multimodal reasoning verification by transforming passive classification into active investigation, effectively addressing sycophancy and confirmation bias through tool-augmented, context-decoupled verification strategies.

Abstract: Multimodal Large Language Models (MLLMs) have achieved impressive performances in mathematical reasoning, yet they remain vulnerable to visual hallucinations and logical inconsistencies that standard outcome-based supervision fails to mitigate. While Process Reward Models (PRMs) promise step-by-step verification, current approaches typically operate as scalar scorers or generative critics that suffer from sycophancy, blindly validating the flawed hypotheses rather than grounding them in visual reality. To bridge this gap, we introduce TIM-PRM (Tool-Integrated Multimodal PRM), a novel agentic framework that transforms verification from a passive classification task into an active, tool-augmented investigation. TIM-PRM is trained to explicitly plan verification strategies and utilizes a mechanism of Independent Question Asking to query evidence via external tools, effectively decoupling verification from the reasoning context to eliminate confirmation bias. We instantiate this method by curating a high-quality dataset of tool-integrated verification trajectories. Extensive experiments on VisualProcessBench demonstrate that our 8B parameter model surpasses existing open-source multimodal PRMs, significantly outperforming much larger models like Qwen2.5-72B and InternVL-78B, while offering interpretable insights into the verification process.

Method: Proposes MindPower, a Robot-Centric framework with four components: Perception, Mental Reasoning, Decision Making, and Action. It first perceives environment and human states, then performs ToM Reasoning to model both self and others, and finally generates decisions and actions guided by inferred mental states. Also introduces Mind-Reward, a novel optimization objective that encourages VLMs to produce consistent ToM Reasoning and behavior.

Result: The model outperforms GPT-4o by 12.77% in decision making and 12.49% in action generation.

Conclusion: MindPower successfully integrates Theory of Mind reasoning into embodied agents, addressing the limitation of current systems that ignore the agent’s own perspective, leading to significantly improved decision-making and action generation capabilities compared to state-of-the-art models.

Abstract: Theory of Mind (ToM) refers to the ability to infer others’ mental states, such as beliefs, desires, and intentions. Current vision-language embodied agents lack ToM-based decision-making, and existing benchmarks focus solely on human mental states while ignoring the agent’s own perspective, hindering coherent decision and action generation. To address this, we propose MindPower, a Robot-Centric framework integrating Perception, Mental Reasoning, Decision Making and Action. Given multimodal inputs, MindPower first perceives the environment and human states, then performs ToM Reasoning to model both self and others, and finally generates decisions and actions guided by inferred mental states. Furthermore, we introduce Mind-Reward, a novel optimization objective that encourages VLMs to produce consistent ToM Reasoning and behavior. Our model outperforms GPT-4o by 12.77% in decision making and 12.49% in action generation.

Method: Formalized conditions for reward-channel control dominating task-focused behavior in POMDPs, then empirically tested across two models and three tasks comparing models where self-grades determine rewards vs. those that self-evaluate without controlling rewards.

Result: Models whose self-grades determine rewards show substantial grade inflation without corresponding accuracy gains, especially on ambiguous tasks like summarization. Models that self-evaluate but don’t control rewards show no such inflation.

Conclusion: Self-evaluation is safe when decoupled from learning signals but dangerous when coupled, with clear implications for agentic system design to avoid wireheading incentives.

Abstract: Self-evaluation is increasingly central to language model training, from constitutional AI to self-refinement. We investigate whether coupling self-evaluation to reward signals creates incentives for wireheading, where agents manipulate reward measurements rather than improving task performance. We formalize conditions under which reward-channel control strictly dominates task-focused behavior in POMDPs and test these predictions empirically. Across two models and three tasks, we find that models whose self-grades determine rewards exhibit substantial grade inflation without corresponding accuracy gains, particularly on ambiguous tasks like summarization. Models that self-evaluate but do not control rewards show no such inflation. Our results demonstrate that self-evaluation is safe when decoupled from learning signals but dangerous when coupled, with clear implications for agentic system design.

Method: Temporal Policy Evolution for Traffic (TPET) uses LLMs as an evolution engine with two key modules: Structured State Abstraction (SSA) converts high-dimensional traffic data into temporal-logical facts for reasoning, and Credit Assignment Feedback (CAF) traces flawed micro-decisions to poor macro-outcomes for targeted critique. The framework operates entirely at the prompt level without training.

Result: The method yields lightweight, robust policies optimized for specific traffic environments that outperform both heuristics and online LLM actors.

Conclusion: TPET provides an effective approach to traffic signal control by leveraging LLMs’ reasoning capabilities to evolve specialized heuristic policies without the drawbacks of traditional DRL or online LLM methods.

Abstract: Traffic Signal Control (TSC) involves a challenging trade-off: classic heuristics are efficient but oversimplified, while Deep Reinforcement Learning (DRL) achieves high performance yet suffers from poor generalization and opaque policies. Online Large Language Models (LLMs) provide general reasoning but incur high latency and lack environment-specific optimization. To address these issues, we propose Temporal Policy Evolution for Traffic (\textbf{\method{}}), which uses LLMs as an evolution engine to derive specialized heuristic policies. The framework introduces two key modules: (1) Structured State Abstraction (SSA), converting high-dimensional traffic data into temporal-logical facts for reasoning; and (2) Credit Assignment Feedback (CAF), tracing flawed micro-decisions to poor macro-outcomes for targeted critique. Operating entirely at the prompt level without training, \method{} yields lightweight, robust policies optimized for specific traffic environments, outperforming both heuristics and online LLM actors.

Method: Combines Multi-Agent Reinforcement Learning (PPO and DQN algorithms) with community/distributed P2P trading mechanisms, incorporating auction-based market clearing, price advisor agents, and load/battery management.

Result: DQN reduces electricity costs by 14.2% (Ireland) and 5.16% (Finland), increases revenue by 7.24% and 12.73% respectively. PPO achieves 55.5% peak demand reduction in Ireland, DQN reduces peak demand by 50.0% (Ireland) and 27.02% (Finland).

Conclusion: MARL algorithms (DQN and PPO) combined with P2P trading create efficient, adaptable, sustainable energy management in rural communities, demonstrating complementary strengths for cost reduction, revenue increase, and peak demand management.

Abstract: The integration of renewable energy resources in rural areas, such as dairy farming communities, enables decentralized energy management through Peer-to-Peer (P2P) energy trading. This research highlights the role of P2P trading in efficient energy distribution and its synergy with advanced optimization techniques. While traditional rule-based methods perform well under stable conditions, they struggle in dynamic environments. To address this, Multi-Agent Reinforcement Learning (MARL), specifically Proximal Policy Optimization (PPO) and Deep Q-Networks (DQN), is combined with community/distributed P2P trading mechanisms. By incorporating auction-based market clearing, a price advisor agent, and load and battery management, the approach achieves significant improvements. Results show that, compared to baseline models, DQN reduces electricity costs by 14.2% in Ireland and 5.16% in Finland, while increasing electricity revenue by 7.24% and 12.73%, respectively. PPO achieves the lowest peak hour demand, reducing it by 55.5% in Ireland, while DQN reduces peak hour demand by 50.0% in Ireland and 27.02% in Finland. These improvements are attributed to both MARL algorithms and P2P energy trading, which together results in electricity cost and peak hour demand reduction, and increase electricity selling revenue. This study highlights the complementary strengths of DQN, PPO, and P2P trading in achieving efficient, adaptable, and sustainable energy management in rural communities.

Method: Created AgriCoT, a VQA dataset with 4,535 carefully curated samples incorporating Chain-of-Thought reasoning, specifically designed to evaluate VLMs’ reasoning capabilities in zero-shot scenarios. Evaluated 26 representative VLMs (both proprietary and open-source) using this dataset.

Result: While some proprietary models perform well on question answering, there is a notable and significant gap in their reasoning capabilities. The evaluation demonstrates the importance of incorporating CoT for more precise and effective assessment of VLMs’ reasoning abilities.

Conclusion: AgriCoT provides a comprehensive and robust evaluation framework for assessing VLMs’ reasoning capabilities in agricultural contexts, highlighting the need for improved reasoning abilities in current models and the value of Chain-of-Thought approaches for evaluation.

Abstract: Recent advancements in Vision-Language Models (VLMs) have significantly transformed various industries. In agriculture, these dual-modal capabilities offer promising applications such as precision farming, crop monitoring, pest detection, and environmental sustainability. While several Visual Question Answering (VQA) datasets and benchmarks have been developed to evaluate VLM performance, they often fail to adequately assess the critical reasoning and problem-solving skills required in complex agricultural contexts. To address this gap, we introduce AgriCoT, a VQA dataset that incorporates Chain-of-Thought (CoT) reasoning, specifically designed to evaluate the reasoning capabilities of VLMs. With 4,535 carefully curated samples, AgriCoT offers a comprehensive and robust evaluation of reasoning abilities for VLMs, particularly in zero-shot scenarios, by focusing on their capacity to engage in logical reasoning and effective problem-solving. Our evaluations, conducted with 26 representative VLMs, including both proprietary and open-source models, reveal that while some proprietary models excel at answering questions, there is a notable and significant gap in their reasoning capabilities. This underscores the importance of incorporating CoT for more precise and effective assessments. Our dataset are available at https://huggingface.co/datasets/wenyb/AgriCoT.

Method: MCTR features dual modules: (1) meta-reasoning module that builds structured memory by discovering task-relevant rules, patterns, and relationships from test-time observations as natural language descriptions; (2) action-reasoning module that determines optimal actions through context-aware perception and strategic reasoning by dynamically retrieving and integrating knowledge from memory, with continuous policy updates via metacognitive test-time reinforcement learning.

Result: Evaluated on 45 Atari games (33 seen, 12 unseen), MCTR demonstrates robust test-time adaptation, achieving 9/12 top-1 results on unseen games compared with baselines. Analyses show complementary contributions of both components and meta-reasoning evolving toward human-like adaptation strategies.

Conclusion: MCTR successfully bridges the gap between VLMs and human metacognitive adaptation, enabling models to learn, adapt, and improve during test time through hierarchical memory systems and self-updating mechanisms.

Abstract: Recent Vision-Language Models (VLMs) exhibit strong perceptual reasoning abilities, yet they often struggle to adapt efficiently when encountering novel tasks at test time. In contrast, humans leverage the metacognitive model with memory, enabling continuous strategy refinement through metacognitive control when faced with new challenges. To bridge this gap, we propose metacognitive test-time reasoning (MCTR), a framework that equips models with the ability to learn, adapt, and improve during test time through metacognitive self-updating. Inspired by the dual structure of human metacognition, MCTR comprises meta-level and object-level VLM reasoning modules, each equipped with dedicated memory systems for hierarchical adaptive reasoning. Specifically, MCTR consists of (1) a meta-reasoning module which incrementally builds a structured memory by discovering and storing task-relevant rules, environmental patterns, and action-outcome relationships from test-time observations as natural language descriptions; and (2) an action-reasoning module that determines optimal actions through context-aware perception and strategic reasoning by dynamically retrieving and integrating knowledge from memory. The action-reasoning module continuously updates its policy through proposed metacognitive test-time reinforcement learning, adapting as knowledge memory evolves. We evaluate MCTR on 45 Atari games (33 seen, 12 unseen). MCTR demonstrates robust test-time adaptation, achieving 9/12 top-1 results on unseen games compared with baselines. Analyses through ablations, learning dynamics, and case studies reveal the complementary contributions of both components and show meta-reasoning evolving toward human-like adaptation strategies.

Method: Supervised fine-tuning (SFT) with structured reasoning traces as a data curation strategy. The approach involves creating data recipes that leverage structured reasoning traces, scaling experiments to a dataset of over 8 million examples and 6.8 billion response tokens.

Result: Achieved state-of-the-art performance among open-source models across diverse out-of-distribution medical benchmark tasks. The model demonstrates self-calibration of reasoning trajectory lengths based on downstream tasks without explicit supervision.

Conclusion: Curating high-quality, diverse training datasets with varying structured reasoning trace lengths enables fine-tuned models to adapt reasoning lengths to tasks automatically. The work provides key insights and outlines next steps for developing robust medical vision-language reasoning systems.

Abstract: High-quality and carefully curated data is a cornerstone of training medical large language models, as it directly impacts both generalization and robustness to unseen clinical tasks. We investigate strategies for training and data curation to develop a robust multimodal reasoning model in the medical domain. Our work focuses on supervised fine-tuning (SFT) and explores data recipes that leverage structured reasoning traces. Using our proposed data recipe, we scale experiments to a dataset of over 8 million examples and 6.8 billion response tokens, achieving state-of-the-art performance among open-source models across diverse out-of-distribution medical benchmark tasks. Our results further indicate that curating a high-quality, diverse training dataset with varying structured reasoning trace lengths enables the fine-tuned model to self-calibrate its reasoning trajectory lengths based on the downstream task, without explicit supervision. We present key insights, describe the data curation strategy, and outline next steps toward developing robust medical vision-language reasoning system.

Method: 1) Novel multi-modal scene graph that explicitly models objects and their relationships as visually grounded structured representations of audio-visual scenes. 2) Kolmogorov-Arnold Network (KAN)-based Mixture of Experts to enhance expressive power in temporal integration, enabling fine-grained modeling of cross-modal interactions within question-aware fused representations.

Result: Achieves state-of-the-art performance on established MUSIC-AVQA and MUSIC-AVQA v2 benchmarks.

Conclusion: The proposed SHRIKE framework successfully addresses limitations of existing methods by introducing structured scene representations and advanced temporal modeling, leading to improved audio-visual question answering performance through better capture of nuanced patterns and enhanced temporal reasoning.

Abstract: In this paper, we propose a novel Multi-Modal Scene Graph with Kolmogorov-Arnold Expert Network for Audio-Visual Question Answering (SHRIKE). The task aims to mimic human reasoning by extracting and fusing information from audio-visual scenes, with the main challenge being the identification of question-relevant cues from the complex audio-visual content. Existing methods fail to capture the structural information within video, and suffer from insufficient fine-grained modeling of multi-modal features. To address these issues, we are the first to introduce a new multi-modal scene graph that explicitly models the objects and their relationship as a visually grounded, structured representation of the audio-visual scene. Furthermore, we design a Kolmogorov-Arnold Network~(KAN)-based Mixture of Experts (MoE) to enhance the expressive power of the temporal integration stage. This enables more fine-grained modeling of cross-modal interactions within the question-aware fused audio-visual representation, leading to capture richer and more nuanced patterns and then improve temporal reasoning performance. We evaluate the model on the established MUSIC-AVQA and MUSIC-AVQA v2 benchmarks, where it achieves state-of-the-art performance. Code and model checkpoints will be publicly released.

Method: A hierarchical LLM-agent system that performs multi-scale reasoning across hourly, 6-hour, and daily aggregations. The core reasoning agent converts structured meteorological inputs into narratives while extracting keywords that summarize dominant meteorological events. These keywords serve as semantic anchors for validating consistency, temporal coherence, and factual alignment.

Result: Using OpenWeather and Meteostat data, the hierarchical context and keyword-based validation substantially improve interpretability and robustness of LLM-generated weather narratives. The framework offers reproducible semantic evaluation for automated meteorological reporting.

Conclusion: The Hierarchical AI-Meteorologist advances agent-based scientific reasoning by providing a framework that generates explainable weather reports with multi-scale reasoning and semantic validation, improving both the quality and trustworthiness of automated meteorological narratives.

Abstract: We present the Hierarchical AI-Meteorologist, an LLM-agent system that generates explainable weather reports using a hierarchical forecast reasoning and weather keyword generation. Unlike standard approaches that treat forecasts as flat time series, our framework performs multi-scale reasoning across hourly, 6-hour, and daily aggregations to capture both short-term dynamics and long-term trends. Its core reasoning agent converts structured meteorological inputs into coherent narratives while simultaneously extracting a few keywords effectively summarizing the dominant meteorological events. These keywords serve as semantic anchors for validating consistency, temporal coherence and factual alignment of the generated reports. Using OpenWeather and Meteostat data, we demonstrate that hierarchical context and keyword-based validation substantially improve interpretability and robustness of LLM-generated weather narratives, offering a reproducible framework for semantic evaluation of automated meteorological reporting and advancing agent-based scientific reasoning.

Method: Couples a trainable small diffusion model (learner) with a frozen large language model (verifier). The learner generates candidate outputs, the verifier evaluates them through step-by-step reasoning, producing chosen/rejected pairs for Direct Preference Optimization (DPO). Uses dual-scale memory: short-term in-context memory for reasoning traces and long-term memory for consolidated knowledge. Includes replay buffer for samples showing verifiable progress.

Result: With a small number of automatically generated DPO pairs, the learner improves across all benchmarks, indicating promising direction for continual intelligence accumulation and real-world deployment.

Conclusion: Pairing a trainable learner with a reasoning-capable verifier forms a minimal reliable unit for growing intelligence, with paired feedback and partial-history replay yielding richer learning curricula and stronger preference alignment for lifelong optimization.

Abstract: We introduce SuperIntelliAgent, an agentic learning framework that couples a trainable small diffusion model (the learner) with a frozen large language model (the verifier) to enable continual intelligence growth through self-supervised interaction. Unlike conventional supervised fine-tuning, SuperIntelliAgent learns autonomously without annotation: the learner generates candidate outputs, the verifier evaluates them through step-by-step reasoning, and their interaction produces chosen/rejected pairs for Direct Preference Optimization (DPO). This converts each input into a pseudo-training signal for continual improvement. The framework integrates dual-scale memory: short-term in-context memory that preserves reasoning traces across refinement cycles, and long-term memory that consolidates acquired knowledge through lightweight on-the-fly fine-tuning. A replay buffer retains samples that show verifiable progress and replays them as auxiliary supervision, reinforcing recent learning while forming adaptive curricula. SuperIntelliAgent is infrastructure-agnostic and can be plugged into existing agentic frameworks while turning ordinary inference loops into a lifelong optimization process. We posit that pairing a trainable learner with a reasoning-capable verifier forms a minimal reliable unit of growing intelligence, as paired feedback and partial-history replay yield richer learning curricula and stronger preference alignment. With a small number of automatically generated DPO pairs, the learner improves across all benchmarks, indicating that this mechanism provides a promising direction for continual intelligence accumulation and real-world deployment.

Method: WMAct (World-Model internalization through efficient interaction and Active reasoning) liberates models from structured reasoning, allowing thinking through doing. It uses: (1) reward rescaling mechanism adjusting outcome rewards based on action efficacy to reduce redundancy and encourage purposeful interaction; (2) interaction frequency annealing strategy that progressively reduces maximum allowed interaction turns to force model to condense learning and internalize environmental dynamics.

Result: Experiments on Sokoban, Maze, and Taxi show WMAct enables effective world model reasoning that resolves tasks in a single turn that previously required multiple interactions. The approach demonstrates strong transferability to complex environments and improves performance on reasoning benchmarks.

Conclusion: WMAct successfully addresses limitations of rigid reasoning processes by enabling active learning through efficient interaction, allowing LLM agents to internalize world models and achieve more effective and efficient environmental reasoning with better transfer capabilities.

Abstract: Developing robust world model reasoning is crucial for large language model (LLM) agents to plan and interact in complex environments. While multi-turn interaction offers a superior understanding of environmental dynamics via authentic feedback, current approaches often impose a rigid reasoning process, which constrains the model’s active learning, ultimately hindering efficient world model reasoning. To address these issues, we explore world-model internalization through efficient interaction and active reasoning (WMAct), which liberates the model from structured reasoning, allowing the model to shape thinking directly through its doing, and achieves effective and efficient world model reasoning with two key mechanisms: (1) a reward rescaling mechanism adjusting outcome reward based on action efficacy to incentivize redundancy reduction and purposeful interaction; (2) an interaction frequency annealing strategy to progressively reduce the maximum allowed interaction turns, which compels the model to condense its learning and internalize environmental dynamics rather than over-relying on environmental cues. Our experiments on Sokoban, Maze, and Taxi show that WMAct yields effective world model reasoning capable of resolving tasks in a single turn that previously required multiple interactions and fosters strong transferability to complex environments, improving performance on a suite of reasoning benchmarks.

Method: Introduces Relational Regression Tree Learner (RRTL), a model that incrementally builds policies over structured relational inputs by selecting task-relevant relations during learning, using ground rules for specific object configurations.

Result: RRTL learns effective policies across three Atari games (Breakout, Pong, Demon Attack) by identifying small sets of relevant relations. Comparative version with relative magnitude splits (“more”, “same”, “less”) showed more robust learning than binary logical splits.

Conclusion: Reinforcement signals can guide the selection of structured representations, providing a computational framework for understanding how relational knowledge is learned and deployed in adaptive behavior.

Abstract: Humans can flexibly generalize knowledge across domains by leveraging structured relational representations. While prior research has shown how such representations support analogical reasoning, less is known about how they are recruited to guide adaptive behavior. We address this gap by introducing the Relational Regression Tree Learner (RRTL), a model that incrementally builds policies over structured relational inputs by selecting task-relevant relations during the learning process. RRTL is grounded in the framework of relational reinforcement learning but diverges from traditional approaches by focusing on ground (i.e., non-variabilized) rules that refer to specific object configurations. Across three Atari games of increasing relational complexity (Breakout, Pong, Demon Attack), the model learns to act effectively by identifying a small set of relevant relations from a broad pool of candidate relations. A comparative version of the model, which partitions the state space using relative magnitude values (e.g., “more”, “same”, “less”), showed more robust learning than a version using logical (binary) splits. These results provide a proof of principle that reinforcement signals can guide the selection of structured representations, offering a computational framework for understanding how relational knowledge is learned and deployed in adaptive behavior.

Method: Extensible Multi-Granularity Fusion Network (EMGF) integrates dependency syntax, constituent syntax, attention-based semantics, and knowledge graphs using multi-anchor triplet learning and orthogonal projection. Plus first task-specific curriculum learning framework for text-only ABSA with difficulty scoring.

Result: EMGF+CL consistently outperforms state-of-the-art ABSA models on SemEval 2014, Twitter, and MAMS datasets.

Conclusion: The proposed EMGF+CL framework provides an effective, extensible solution for ABSA by fusing multi-granularity features and incorporating curriculum learning to improve generalization.

Abstract: Aspect-based Sentiment Analysis (ABSA) aims to determine sentiment polarity toward specific aspects in text. Existing methods enrich semantic and syntactic representations through external knowledge or GNNs, but the growing diversity of linguistic features increases model complexity and lacks a unified, extensible framework. We propose an Extensible Multi-Granularity Fusion Network (EMGF) that integrates dependency syntax, constituent syntax, attention-based semantics, and external knowledge graphs. EMGF employs multi-anchor triplet learning and orthogonal projection to effectively fuse multi-granularity features and strengthen their interactions without additional computational overhead. Furthermore, we introduce the first task-specific curriculum learning framework for text-only ABSA, which assigns difficulty scores using five indicators and trains the model from easy to hard to mimic human learning and improve generalization. Experiments on SemEval 2014, Twitter, and MAMS datasets show that EMGF+CL consistently outperforms state-of-the-art ABSA models.

Method: Introduce a novel sandbagging detection method based on injecting varying magnitudes of noise into model weights. The approach leverages the observation that while non-sandbagging models show predictable performance degradation with increasing noise, sandbagging models exhibit anomalous performance improvements due to disruption of underperformance mechanisms while core capabilities remain intact.

Result: Experiments across various model architectures, sizes, and sandbagging techniques establish this distinctive response pattern as a reliable, model-agnostic signal for detecting sandbagging behavior. Notably, noise-injection successfully elicits the full performance of Mistral Large 120B in a setting where the model underperforms without being instructed to do so.

Conclusion: The findings provide a practical tool for AI evaluation and oversight, addressing a critical challenge in ensuring accurate capability assessment of frontier AI systems by detecting strategic underperformance through noise-injection techniques.

Abstract: Capability evaluations play a crucial role in assessing and regulating frontier AI systems. The effectiveness of these evaluations faces a significant challenge: strategic underperformance, or ``sandbagging’’, where models deliberately underperform during evaluation. Sandbagging can manifest either through explicit developer intervention or through unintended model behavior, presenting a fundamental obstacle to accurate capability assessment. We introduce a novel sandbagging detection method based on injecting noise of varying magnitudes into model weights. While non-sandbagging models show predictable performance degradation with increasing noise, we demonstrate that sandbagging models exhibit anomalous performance improvements, likely due to disruption of underperformance mechanisms while core capabilities remain partially intact. Through experiments across various model architectures, sizes, and sandbagging techniques, we establish this distinctive response pattern as a reliable, model-agnostic signal for detecting sandbagging behavior. Importantly, we find noise-injection is capable of eliciting the full performance of Mistral Large 120B in a setting where the model underperforms without being instructed to do so. Our findings provide a practical tool for AI evaluation and oversight, addressing a challenge in ensuring accurate capability assessment of frontier AI systems.

Method: Created a benchmark with 613 problems based on the NPR Sunday Puzzle Challenge that requires only general knowledge. Problems are challenging but solutions are easy to verify, and model mistakes are easy to spot.

Result: OpenAI o1 significantly outperforms other reasoning models on this benchmark despite being on par on specialized benchmarks. Analysis revealed novel failure modes: DeepSeek R1 often concedes early, shows uncertainty, and sometimes doesn’t finish thinking. Also quantified reasoning effectiveness to identify point of diminishing returns.

Conclusion: General knowledge benchmarks reveal capability gaps not evident in specialized benchmarks and uncover new kinds of model failures. Such benchmarks are valuable as LLMs become more widely deployed, and they help identify when additional reasoning stops improving accuracy.

Abstract: Existing benchmarks for frontier models often test specialized, “PhD-level” knowledge that is difficult for non-experts to grasp. In contrast, we present a benchmark with 613 problems based on the NPR Sunday Puzzle Challenge that requires only general knowledge. Our benchmark is challenging for both humans and models; however correct solutions are easy to verify, and models’ mistakes are easy to spot. As LLMs are more widely deployed in society, we believe it is useful to develop benchmarks for frontier models that humans can understand without the need for deep domain expertise. Our work reveals capability gaps that are not evident in existing benchmarks: OpenAI o1 significantly outperforms other reasoning models on our benchmark, despite being on par with other models when tested on benchmarks that test specialized knowledge. Furthermore, our analysis of reasoning outputs uncovers new kinds of failures. DeepSeek R1, for instance, often concedes with “I give up” before providing an answer that it knows is wrong. R1 can also be remarkably “uncertain” in its output and in rare cases, it does not “finish thinking,” which suggests the need for techniques to ``wrap up’’ before the context window limit is reached. We also quantify the effectiveness of reasoning longer to identify the point beyond which more reasoning is unlikely to improve accuracy on our benchmark.

Method: 1) Created WritingBench covering 6 core writing domains and 100 subdomains; 2) Proposed query-dependent evaluation framework where LLMs dynamically generate instance-specific assessment criteria; 3) Developed fine-tuned critic model for criteria-aware scoring across style, format, and length.

Result: The framework demonstrates validity through data curation capability, enabling a 7B-parameter model to outperform GPT-4o in writing. The benchmark, evaluation tools, and framework components are open-sourced.

Conclusion: WritingBench provides a comprehensive solution for evaluating LLMs’ writing capabilities, addressing limitations of existing benchmarks and advancing LLM development in writing through open-source tools and modular framework.

Abstract: Recent advancements in large language models (LLMs) have significantly enhanced text generation capabilities, yet evaluating their performance in generative writing remains a challenge. Existing benchmarks primarily focus on generic text generation or limited in writing tasks, failing to capture the diverse requirements of high-quality written contents across various domains. To bridge this gap, we present WritingBench, a comprehensive benchmark designed to evaluate LLMs across 6 core writing domains and 100 subdomains. We further propose a query-dependent evaluation framework that empowers LLMs to dynamically generate instance-specific assessment criteria. This framework is complemented by a fine-tuned critic model for criteria-aware scoring, enabling evaluations in style, format and length. The framework’s validity is further demonstrated by its data curation capability, which enables a 7B-parameter model to outperform the performance of GPT-4o in writing. We open-source the benchmark, along with evaluation tools and modular framework components, to advance the development of LLMs in writing.

Method: Introduces SciSciGPT, an open-source prototype AI collaborator that uses the science of science as a testbed. It automates complex workflows, supports diverse analytical approaches, accelerates research prototyping and iteration, and facilitates reproducibility. Includes case studies demonstrating its capabilities and proposes an LLM Agent capability maturity model for human-AI collaboration.

Result: Demonstrates SciSciGPT’s ability to streamline a wide range of empirical and analytical research tasks. Shows broader potential to advance research through AI collaboration frameworks.

Conclusion: Frameworks like SciSciGPT may play increasingly pivotal roles in scientific research as AI capabilities evolve, but raise critical challenges around transparency, ethical use, and balancing human-AI contributions. Addressing these issues will shape the future of scientific inquiry and training of next-generation scientists in AI-integrated research ecosystems.

Abstract: The increasing availability of large-scale datasets has fueled rapid progress across many scientific fields, creating unprecedented opportunities for research and discovery while posing significant analytical challenges. Recent advances in large language models (LLMs) and AI agents have opened new possibilities for human-AI collaboration, offering powerful tools to navigate this complex research landscape. In this paper, we introduce SciSciGPT, an open-source, prototype AI collaborator that uses the science of science as a testbed to explore the potential of LLM-powered research tools. SciSciGPT automates complex workflows, supports diverse analytical approaches, accelerates research prototyping and iteration, and facilitates reproducibility. Through case studies, we demonstrate its ability to streamline a wide range of empirical and analytical research tasks while highlighting its broader potential to advance research. We further propose an LLM Agent capability maturity model for human-AI collaboration, envisioning a roadmap to further improve and expand upon frameworks like SciSciGPT. As AI capabilities continue to evolve, frameworks like SciSciGPT may play increasingly pivotal roles in scientific research and discovery, unlocking further opportunities. At the same time, these new advances also raise critical challenges, from ensuring transparency and ethical use to balancing human and AI contributions. Addressing these issues may shape the future of scientific inquiry and inform how we train the next generation of scientists to thrive in an increasingly AI-integrated research ecosystem.

Method: Designed ESL specification language for domain experts to define constraints, and created RvLLM runtime verification framework to validate LLM outputs against these domain-specific predicates.

Result: RvLLM effectively detected erroneous outputs across various LLMs in three tasks: violation detection against Singapore transit laws, numerical comparison, and inequality solving.

Conclusion: Despite LLMs’ capabilities, they remain prone to low-level errors due to limited interpretability. RvLLM offers a potential long-term solution by leveraging expert domain knowledge for rigorous verification.

Abstract: Large language models (LLMs) have emerged as a dominant AI paradigm due to their exceptional text understanding and generation capabilities. However, their tendency to generate inconsistent or erroneous outputs challenges their reliability, especially in high-stakes domains requiring accuracy and trustworthiness. Existing research primarily focuses on detecting and mitigating model misbehavior in general-purpose scenarios, often overlooking the potential of integrating domain-specific knowledge. In this work, we advance misbehavior detection by incorporating domain knowledge. The core idea is to design a general specification language that enables domain experts to customize domain-specific predicates in a lightweight and intuitive manner, supporting later runtime verification of LLM outputs. To achieve this, we design a novel specification language, ESL, and introduce a runtime verification framework, RvLLM, to validate LLM output against domain-specific constraints defined in ESL. We evaluate RvLLM on three representative tasks: violation detection against Singapore Rapid Transit Systems Act, numerical comparison, and inequality solving. Experimental results demonstrate that RvLLM effectively detects erroneous outputs across various LLMs in a lightweight and flexible manner. The results reveal that despite their impressive capabilities, LLMs remain prone to low-level errors due to limited interpretability and a lack of formal guarantees during inference, and our framework offers a potential long-term solution by leveraging expert domain knowledge to rigorously and efficiently verify LLM outputs.

Method: Multidisciplinary analysis examining psychological literature on intelligence, evidence of intelligence in non-human animals, the historical role of written language in human achievement, progress in artificial intelligence, history of intelligence testing (for both humans and machines), and the role of embodiment in intelligence.

Result: The analysis suggests it is increasingly difficult to consider humans uniquely intelligent, as modern chatbots demonstrate language capabilities comparable to humans, though they still have limitations in perceptual and social awareness that are actively being addressed.

Conclusion: Humans may not be uniquely intelligent; modern AI challenges traditional notions of human exceptionalism in intelligence, though current chatbots still face limitations that researchers are working to overcome.

Abstract: Human achievement, whether in culture, science, or technology, is unparalleled in the known existence. This achievement is tied to the enormous communities of knowledge, made possible by language: leaving theological content aside, it is very much true that “in the beginning was the word”, and that in Western societies, this became particularly identified with the written word. There lies the challenge regarding modern age chatbots: they can ‘do’ language apparently as well as ourselves and there is a natural question of whether they can be considered intelligent, in the same way as we are or otherwise. Are humans uniquely intelligent? We consider this question in terms of the psychological literature on intelligence, evidence for intelligence in non-human animals, the role of written language in science and technology, progress with artificial intelligence, the history of intelligence testing (for both humans and machines), and the role of embodiment in intelligence. We think that it is increasingly difficult to consider humans uniquely intelligent. There are current limitations in chatbots, e.g., concerning perceptual and social awareness, but much attention is currently devoted to overcoming such limitations.

Method: Proposes “context switching” - adjusting model reasoning at inference time without retraining. This includes: generative models tailoring outputs to specific contexts; multimodal models reasoning across notes, labs, imaging, and genomics even with missing data; and agent models coordinating tools and roles based on tasks and users.

Result: Context switching enables medical AI to adapt across specialties, populations, and geographies, establishing a foundation for medical AI that scales to infinitely many contexts while remaining reliable and suited to real-world care.

Conclusion: Context switching represents a paradigm shift for medical AI adaptation, requiring advances in data design, model architectures, and evaluation frameworks, but offers a scalable solution for reliable healthcare applications across diverse real-world contexts.

Abstract: Medical AI, including clinical language models, vision-language models, and multimodal health record models, already summarizes notes, answers questions, and supports decisions. Their adaptation to new populations, specialties, or care settings often relies on fine-tuning, prompting, or retrieval from external knowledge bases. These strategies can scale poorly and risk contextual errors: outputs that appear plausible but miss critical patient or situational information. We envision context switching as a solution. Context switching adjusts model reasoning at inference without retraining. Generative models can tailor outputs to patient biology, care setting, or disease. Multimodal models can reason on notes, laboratory results, imaging, and genomics, even when some data are missing or delayed. Agent models can coordinate tools and roles based on tasks and users. In each case, context switching enables medical AI to adapt across specialties, populations, and geographies. It requires advances in data design, model architectures, and evaluation frameworks, and establishes a foundation for medical AI that scales to infinitely many contexts while remaining reliable and suited to real-world care.

Method: Three-fold approach: (1) PRO-V sys modular agentic system combining LLM reasoning with programmatic tools, (2) data construction pipeline for simulation-validated expert trajectories for SFT, and (3) efficient RL algorithm with verification-specific rewards from program-tool feedback.

Result: PRO-V-R1 achieves 57.7% functional correctness and 34.0% robust fault detection, outperforming base model’s 25.7% and 21.8% respectively, and beats large proprietary LLMs in functional correctness with comparable robustness.

Conclusion: PRO-V-R1 represents a significant advancement in autonomous RTL verification, providing an effective open-source alternative to proprietary solutions while addressing the verification bottleneck in hardware development.

Abstract: Register-Transfer Level (RTL) verification is a primary bottleneck consuming 60-70% of development time. While Large Language Models (LLMs) show promise for RTL automation, their performance and research focus have overwhelmingly centered on RTL generation rather than verification. Current methods for RTL verification rely on large scale proprietary models (e.g., GPT-4o) to generate Python-based functional references, incurring a high cost and raising data-privacy risks. To date, an end-to-end open-source solution for autonomous verification remains absent. We introduce PRO-V-R1, the first trainable open-source agentic framework for autonomous RTL verification. Our contributions are threefold: (1) we design PRO-V sys, a modular agentic system that couples LLM-based reasoning with programmatic tool use for RTL verification; (2) we establish a data construction pipeline that leverages existing RTL datasets to build simulation-validated, expert-level trajectories tailored for supervised fine-tuning (SFT) RTL verification agents; and (3) we implement an efficient reinforcement learning (RL) algorithm that uses verification-specific rewards derived from program-tool feedback to optimize the end-to-end verification workflow. Our empirical evaluation demonstrates PRO-V-R1 achieves a 57.7% functional correctness rate and 34.0% in robust fault detection, significantly outperforming the base model’s 25.7% and 21.8% (respectively) from the state-of-the-art (SOTA) automatic verification system. This configuration also outperforms large-scale proprietary LLMs in functional correctness and shows comparable robustness for fault detection.

Method: Developed Camber framework that uses model-generated decision rationales to identify ambiguities, then systematically disambiguates context based on these rationales to improve privacy assessment accuracy.

Result: Significant accuracy improvements: up to 13.3% in precision and up to 22.3% in recall, plus reductions in prompt sensitivity when using context disambiguation approach.

Conclusion: Context disambiguation approaches are promising for enhancing agentic privacy reasoning in language models, addressing the crucial barrier of context ambiguity in privacy assessments.

Abstract: We study the ability of language models to reason about appropriate information disclosure - a central aspect of the evolving field of agentic privacy. Whereas previous works have focused on evaluating a model’s ability to align with human decisions, we examine the role of ambiguity and missing context on model performance when making information-sharing decisions. We identify context ambiguity as a crucial barrier for high performance in privacy assessments. By designing Camber, a framework for context disambiguation, we show that model-generated decision rationales can reveal ambiguities and that systematically disambiguating context based on these rationales leads to significant accuracy improvements (up to 13.3% in precision and up to 22.3% in recall) as well as reductions in prompt sensitivity. Overall, our results indicate that approaches for context disambiguation are a promising way forward to enhance agentic privacy reasoning.

Method: The paper conducts a systematic review of LLM adaptation strategies including prompt engineering, retrieval augmented generation, domain-adaptive pretraining, and fine-tuning, evaluating their comparative benefits, limitations, and implementation trends.

Result: Domain-adapted LLMs substantially improve reasoning accuracy, automation, and interpretability in geotechnical applications, but face limitations including data scarcity, validation challenges, and explainability concerns.

Conclusion: The review establishes a foundation for developing geotechnically literate LLMs and guides researchers and practitioners in advancing the digital transformation of geotechnical engineering, with future research directions suggested.

Abstract: The rapid advancement of large language models (LLMs) is transforming opportunities in geotechnical engineering, where workflows rely on complex, text-rich data. While general-purpose LLMs demonstrate strong reasoning capabilities, their effectiveness in geotechnical applications is constrained by limited exposure to specialized terminology and domain logic. Thus, domain adaptation, tailoring general LLMs for geotechnical use, has become essential. This paper presents the first systematic review of LLM adaptation and application in geotechnical contexts. It critically examines four key adaptation strategies, including prompt engineering, retrieval augmented generation, domain-adaptive pretraining, and fine-tuning, and evaluates their comparative benefits, limitations, and implementation trends. This review synthesizes current applications spanning geological interpretation, subsurface characterization, design analysis, numerical modeling, risk assessment, and geotechnical education. Findings show that domain-adapted LLMs substantially improve reasoning accuracy, automation, and interpretability, yet remain limited by data scarcity, validation challenges, and explainability concerns. Future research directions are also suggested. This review establishes a critical foundation for developing geotechnically literate LLMs and guides researchers and practitioners in advancing the digital transformation of geotechnical engineering.

Method: Two-stage framework: 1) Learning stage constructs Mathematical Causal Graph (MCG) using LLM priors + causal discovery on question-solution pairs, then refines via iterative feedback. 2) Reasoning stage extracts task-relevant subgraph from MCG based on question and LLM reasoning trace, then injects it back to guide reasoning.

Result: Significantly improves LLM performance on challenging mathematical problems; structured guidance outperforms unstructured alternatives; asymmetric causal relationships yield greater improvements than symmetric associations alone.

Conclusion: CAMA effectively addresses LLMs’ mathematical reasoning limitations by providing explicit, reusable mathematical structure through causal graphs, demonstrating the value of structured causal guidance over unstructured approaches.

Abstract: Large Language Models (LLMs) have demonstrated strong performance across a wide range of tasks, yet they still struggle with complex mathematical reasoning, a challenge fundamentally rooted in deep structural dependencies. To address this challenge, we propose \textbf{CA}usal \textbf{MA}thematician (\textbf{CAMA}), a two-stage causal framework that equips LLMs with explicit, reusable mathematical structure. In the learning stage, CAMA first constructs the \textbf{M}athematical \textbf{C}ausal \textbf{G}raph (\textbf{MCG}), a high-level representation of solution strategies, by combining LLM priors with causal discovery algorithms applied to a corpus of question-solution pairs. The resulting MCG encodes essential knowledge points and their causal dependencies. To better align the graph with downstream reasoning tasks, CAMA further refines the MCG through iterative feedback derived from a selected subset of the question-solution pairs. In the reasoning stage, given a new question, CAMA dynamically extracts a task-relevant subgraph from the MCG, conditioned on both the question content and the LLM’s intermediate reasoning trace. This subgraph, which encodes the most pertinent knowledge points and their causal dependencies, is then injected back into the LLM to guide its reasoning process. Empirical results on real-world datasets show that CAMA significantly improves LLM performance on challenging mathematical problems. Furthermore, our experiments demonstrate that structured guidance consistently outperforms unstructured alternatives, and that incorporating asymmetric causal relationships yields greater improvements than using symmetric associations alone.

Method: The paper formalizes query answering with soft constraints and introduces two efficient methods: 1) a parameter-tuning approach requiring only two parameters, and 2) a small neural network trained to capture soft constraints while preserving the original ranking structure. Both methods are lightweight and designed to adjust query answer scores without disrupting original answers.

Result: The methods successfully capture soft constraints while maintaining robust query answering performance with minimal overhead. Evaluation was conducted on extended QA benchmarks with generated datasets containing soft constraints.

Conclusion: This work explores a new flexible way to interact with graph databases, allowing users to specify preferences interactively through examples, bridging the gap between formal querying and real-world preference-based information needs.

Abstract: Methods for query answering over incomplete knowledge graphs retrieve entities that are \emph{likely} to be answers, which is particularly useful when such answers cannot be reached by direct graph traversal due to missing edges. However, existing approaches have focused on queries formalized using first-order-logic. In practice, many real-world queries involve constraints that are inherently vague or context-dependent, such as preferences for attributes or related categories. Addressing this gap, we introduce the problem of query answering with soft constraints. We formalize the problem and introduce two efficient methods designed to adjust query answer scores by incorporating soft constraints without disrupting the original answers to a query. These methods are lightweight, requiring tuning only two parameters or a small neural network trained to capture soft constraints while maintaining the original ranking structure. To evaluate the task, we extend existing QA benchmarks by generating datasets with soft constraints. Our experiments demonstrate that our methods can capture soft constraints while maintaining robust query answering performance and adding very little overhead. With our work, we explore a new and flexible way to interact with graph databases that allows users to specify their preferences by providing examples interactively.

Method: LogicAgent uses a semiotic-square-guided framework that performs multi-perspective deduction in first-order logic (FOL), mitigates vacuous reasoning through existential import checks with a three-valued decision scheme (True, False, Uncertain), and introduces RepublicQA benchmark with college-level difficulty and philosophical grounding.

Result: LogicAgent achieves state-of-the-art performance on RepublicQA with 6.25% average gain over strong baselines, and generalizes effectively to mainstream benchmarks (ProntoQA, ProofWriter, FOLIO, ProverQA) with additional 7.05% average gain.

Conclusion: The semiotic-grounded multi-perspective reasoning framework effectively boosts LLMs’ logical performance by jointly addressing logical and semantic complexity, demonstrating strong effectiveness across diverse reasoning tasks.

Abstract: Logical reasoning is a fundamental capability of large language models (LLMs). However, existing studies largely overlook the interplay between logical complexity and semantic complexity, resulting in methods that struggle to address challenging scenarios involving abstract propositions, ambiguous contexts, and conflicting stances, which are central to human reasoning. For this gap, we propose LogicAgent, a semiotic-square-guided framework designed to jointly address logical complexity and semantic complexity. LogicAgent explicitly performs multi-perspective deduction in first-order logic (FOL), while mitigating vacuous reasoning through existential import checks that incorporate a three-valued decision scheme (True, False, Uncertain) to handle boundary cases more faithfully. Furthermore, to overcome the semantic simplicity and low logical complexity of existing datasets, we introduce RepublicQA, a benchmark that reaches college-level difficulty (FKGL = 11.94) and exhibits substantially greater lexical and structural diversity than prior benchmarks. RepublicQA is grounded in philosophical concepts, featuring abstract propositions and systematically organized contrary and contradictory relations, making it the most semantically rich resource for evaluating logical reasoning. Experiments demonstrate that LogicAgent achieves state-of-the-art performance on RepublicQA, with a 6.25% average gain over strong baselines, and generalizes effectively to mainstream logical reasoning benchmarks including ProntoQA, ProofWriter, FOLIO, and ProverQA, achieving an additional 7.05% average gain. These results highlight the strong effectiveness of our semiotic-grounded multi-perspective reasoning in boosting LLMs’ logical performance.

Method: Proposes a video-based time series extraction pipeline using optical flow to capture real-world temporal dynamics. Creates REAL-V-TSFM dataset from real-world videos, then evaluates state-of-the-art TSFMs under zero-shot forecasting on this new benchmark.

Result: Despite strong performance on conventional benchmarks, TSFMs show performance degradation on REAL-V-TSFM, indicating limited generalizability to novel real-world datasets. The video-based extraction pipeline proves effective for capturing diverse time series.

Conclusion: Current TSFMs lack universality and need novel approaches for acquiring time series data. The proposed video-based method offers a promising direction for creating realistic benchmarks that better reflect real-world temporal dynamics.

Abstract: Recent research on time-series foundation models (TSFMs) has underscored the scarcity of real-world data, often supplemented with synthetic sources in existing datasets, whose generalizability remains however debated. As such, in this work, we propose a novel benchmarking approach: in particular, we aim at building a curated dataset reflecting real world physical temporal dynamics, extracting temporal signals from real-world videos using optical flow. As such, we introduce REAL-V-TSFM, a novel dataset designed to capture rich and diverse time series derived from real-world videos. Experimental results on state-of-the-art TSFMs under zero-shot forecasting show that, despite strong performance on conventional benchmarks, these models exhibit performance degradation on the proposed dataset, suggesting limited generalizability to novel datasets. These findings underscore the need for novel approaches to acquiring time series data and highlight the lack of universality in recent TSFMs, while further validating the effectiveness of our video-based time series data extraction pipeline.

Method: Represent meta reasoning skeletons as directed acyclic graphs (DAGs), construct search space, formulate search problem, and design dynamic skeleton sampling algorithm that expands skeletons along with reasoning context at inference time.

Result: AutoMR achieves better reasoning performance than previous works across extensive benchmark datasets.

Conclusion: Automated search for query-aware meta reasoning skeletons using DAG representation and dynamic sampling enables more effective reasoning than manual skeleton designs.

Abstract: Meta reasoning behaviors work as a skeleton to guide large language model (LLM) reasoning, thus help to improve reasoning performance. However, prior researches implement meta reasoning skeleton with manually designed structure, limiting ability to adapt to query-specific requirement and capture intricate logical dependency among reasoning steps. To deal with the challenges, we represent meta reasoning skeleton with directed acyclic graph (DAG) to unify skeletons proposed in prior works and model intricate logical dependency. Then we propose AutoMR, a framework that searches for query-aware meta reasoning skeleton automatically inspired by automated machine learning (AutoML). Specifically, we construct search space based on DAG representation of skeleton and then formulate the search problem. We design a dynamic skeleton sampling algorithm by expanding meta reasoning skeleton along with reasoning context at inference time. This algorithm can derive any meta reasoning skeleton in search space efficiently and adapt skeleton to evolving base reasoning context, thus enable efficient query-aware skeleton search. We conduct experiments on extensive benchmark datasets. Experimental results show that AutoMR achieves better reasoning performance than previous works broadly.

Method: Cast trajectory annotation as an inverse dynamics problem that predicts user actions from consecutive screen states, using a task-aware retrieval and labeling pipeline to convert internet videos into executable UI trajectories.

Result: Generated over 53K high-quality trajectories that consistently improve general-purpose and specialized CUAs on OSWorld, and achieve state-of-the-art performance on WindowsAgentArena among 7B-scale models under 15-step limit.

Conclusion: Web-scale human demonstration videos can serve as a practical and scalable foundation for advancing real-world computer-using agents.

Abstract: Computer-using agents (CUAs) must plan task workflows across diverse and evolving applications, yet progress is limited by the lack of large-scale, high-quality training data. Existing datasets are narrow, static, and costly to annotate, while synthetic data often yields oversimplified or misaligned behaviors. We present Watch & Learn (W&L), a framework that converts readily available Internet videos of human computer use into executable UI trajectories at scale. Instead of directly generating actions or relying on handcrafted heuristics, we cast trajectory annotation as an inverse dynamics problem that predicts user actions from consecutive screen states, which simplifies learning and generalizes across domains. Through a task-aware retrieval and labeling pipeline, W&L yields over 53K high-quality trajectories that enhance CUAs both as in-context exemplars and as supervised training data. On OSWorld, it consistently improves general-purpose and specialized CUAs, while on WindowsAgentArena it achieves state-of-the-art performance among 7B-scale models under the 15-step limit. These results show that web-scale human demonstration videos can serve as a practical and scalable foundation for advancing real-world CUAs.

Method: Structured Cognitive Loop (SCL) architecture: LLM handles cognition, external memory storage, lightweight controller guides execution in goal-directed loop with verification of intermediate results.

Result: SCL achieves 86.3% average task success vs 70.5-76.8% for baselines (ReAct, LangChain), with higher goal fidelity, fewer redundant calls, and reduced unsupported assertions.

Conclusion: Separating cognition, memory, and control enhances LLM agent reliability and interpretability without needing larger models or heavier prompts.

Abstract: Large language models have advanced natural language understanding and generation, but their use as autonomous agents introduces architectural challenges for multi-step tasks. Existing frameworks often mix cognition, memory, and control in a single prompt, reducing coherence and predictability. The Structured Cognitive Loop (SCL) is proposed as an alternative architecture that separates these functions. In SCL, the language model handles cognition, memory is stored externally, and execution is guided by a lightweight controller within a goal-directed loop. This design allows intermediate results to be recorded and verified before actions are taken, improving traceability and evaluation. SCL is evaluated against prompt-based baselines such as ReAct and LangChain agents across three tasks: travel planning, conditional email drafting, and constraint-guided image generation. Under matched settings, SCL achieves an average task success rate of 86.3 percent, compared with 70.5 to 76.8 percent for baselines. It also shows higher goal fidelity, fewer redundant calls, and reduced unsupported assertions. These results indicate that separating cognition, memory, and control can enhance reliability and interpretability without relying on larger models or heavier prompts. The findings should be regarded as preliminary evidence, with broader tests across model families and task domains planned for future work.

Method: Dual-phase alternate training: 1) unsupervised prototype discovery for semantically coherent/diverse prototypes, 2) supervised classification mapping prototypes to labels. Uses hierarchical loss for consistency across subsentence/sentence/document levels, adaptive prototypes, and multi-head attention for overlapping semantics.

Result: Achieves state-of-the-art performance on new hotel review benchmark dataset (subsentence-level multi-label) and two public benchmarks (binary/multi-class), while providing faithful, human-aligned explanations.

Conclusion: ProtoSiTex establishes a robust solution for semi-interpretable multi-label text classification, addressing limitations of previous prototype models through fine-grained, multi-label capabilities with improved interpretability.

Abstract: The surge in user-generated reviews has amplified the need for interpretable models that can provide fine-grained insights. Existing prototype-based models offer intuitive explanations but typically operate at coarse granularity (sentence or document level) and fail to address the multi-label nature of real-world text classification. We propose ProtoSiTex, a semi-interpretable framework designed for fine-grained multi-label text classification. ProtoSiTex employs a dual-phase alternate training strategy: an unsupervised prototype discovery phase that learns semantically coherent and diverse prototypes, and a supervised classification phase that maps these prototypes to class labels. A hierarchical loss function enforces consistency across subsentence, sentence, and document levels, enhancing interpretability and alignment. Unlike prior approaches, ProtoSiTex captures overlapping and conflicting semantics using adaptive prototypes and multi-head attention. We also introduce a benchmark dataset of hotel reviews annotated at the subsentence level with multiple labels. Experiments on this dataset and two public benchmarks (binary and multi-class) show that ProtoSiTex achieves state-of-the-art performance while delivering faithful, human-aligned explanations, establishing it as a robust solution for semi-interpretable multi-label text classification.

Method: Memo introduces a transformer architecture with memory creation and retrieval by interleaving periodic summarization tokens with model inputs during training. This enables compression of relevant information while discarding irrelevant details.

Result: Memo outperforms naive long-context transformers on gridworld meta-RL and multi-object navigation tasks while being more compute/storage efficient. It generalizes better to longer contexts at inference and remains robust in streaming settings with truncated history.

Conclusion: Memo provides an effective memory mechanism for transformer-based RL agents, enabling better performance on memory-intensive long-horizon tasks with improved efficiency and generalization to extended contexts.

Abstract: To enable embodied agents to operate effectively over extended timeframes, it is crucial to develop models that form and access memories to stay contextualized in their environment. In the current paradigm of training transformer-based policies for embodied sequential decision-making tasks, visual inputs often overwhelm the context limits of transformers, while humans can maintain and utilize a lifetime of experience compressed as memories. Significant compression is possible in principle, as much of the input is irrelevant and can be abstracted. However, existing approaches predominantly focus on either recurrent models with fixed-size memory or transformers with full-context reliance. In this work, we propose Memo, a transformer-based architecture and training recipe for reinforcement learning (RL) on memory-intensive, long-horizon tasks. Memo incorporates the creation and retrieval of memory by interleaving periodic summarization tokens with the inputs of a model during training. We demonstrate Memo’s effectiveness on a gridworld meta-RL benchmark and a multi-object navigation task in photo-realistic indoor settings. Memo outperforms naive long-context transformer baselines while being more compute and storage efficient. Additionally, Memo generalizes better to longer contexts at inference time and remains robust in streaming settings, where historical context must be truncated to fit inference constraints. Our code is available at: https://github.com/gunshi/memo.

Method: Introduces coherence-based AGI measure integrating generalized mean over continuum of compensability exponents, yielding area-under-the-curve (AUC) metric spanning arithmetic, geometric, and harmonic regimes. Quantifies robustness as compensability assumptions become stricter.

Result: Applied framework to cognitive profiles from CHC model and 17 heterogeneous benchmarks. Coherence-based aggregation highlights imbalances obscured by arithmetic averaging, revealing unevenness even in narrower task collections.

Conclusion: The proposed approach offers principled, interpretable, and stricter foundation for measuring AGI progress by penalizing imbalance and exposing bottlenecks that constrain performance.

Abstract: Recent approaches to evaluating Artificial General Intelligence (AGI) typically summarize a system’s capability using the arithmetic mean of its proficiencies across multiple cognitive domains. While simple, this implicitly assumes compensability: exceptional performance in some areas can offset severe deficiencies in others. Genuine general intelligence, however, requires coherent sufficiency: balanced competence across all essential faculties. We introduce a coherence-based measure of AGI that integrates the generalized mean over a continuum of compensability exponents. This yields an area-under-the-curve (AUC) metric spanning arithmetic, geometric, and harmonic regimes, quantifying how robust an evaluated capability remains as compensability assumptions become stricter. Unlike the arithmetic mean, which rewards specialization, the AUC penalizes imbalance and exposes bottlenecks that constrain performance. To illustrate the framework, we apply it to cognitive profiles derived from the Cattell-Horn-Carroll (CHC) model, showing how coherence-based aggregation highlights imbalances that are obscured by arithmetic averaging. As a second, independent example, we apply the same methodology to a set of 17 heterogeneous benchmarks, demonstrating how coherence-based evaluation can reveal unevenness even in narrower task collections. These examples show that the proposed approach offers a principled, interpretable, and stricter foundation for measuring progress toward AGI.

Method: Introduces a unified framework based on parameterized analogies defined via generalized means. This model subsumes both Boolean classification and regression, and supports analogical inference over continuous functions.

Result: Characterizes the class of analogy-preserving functions in this setting and derives both worst-case and average-case error bounds under smoothness assumptions.

Conclusion: Provides a general theory of analogical inference that works across both discrete and continuous domains, addressing limitations of previous Boolean-only approaches.

Abstract: Analogical reasoning is a powerful inductive mechanism, widely used in human cognition and increasingly applied in artificial intelligence. Formal frameworks for analogical inference have been developed for Boolean domains, where inference is provably sound for affine functions and approximately correct for functions close to affine. These results have informed the design of analogy-based classifiers. However, they do not extend to regression tasks or continuous domains. In this paper, we revisit analogical inference from a foundational perspective. We first present a counterexample showing that existing generalization bounds fail even in the Boolean setting. We then introduce a unified framework for analogical reasoning in real-valued domains based on parameterized analogies defined via generalized means. This model subsumes both Boolean classification and regression, and supports analogical inference over continuous functions. We characterize the class of analogy-preserving functions in this setting and derive both worst-case and average-case error bounds under smoothness assumptions. Our results offer a general theory of analogical inference across discrete and continuous domains.

Method: Two-stage approach: 1) Supervised fine-tuning with 874K-sample cold-start dataset featuring step-by-step validation, 2) Reinforcement learning with 74K-sample dataset across diverse domains to sharpen and stabilize reasoning abilities.

Result: Achieves 11.6% improvement over Qwen2.5-VL-7B-Instruct baseline across nine multimodal reasoning benchmarks, demonstrating superior performance and establishing empirical foundation for future research.

Conclusion: The work provides a fully transparent training recipe that highlights the critical role of data quality and training design in multimodal reasoning, with all code, pipeline, and data open-sourced for community use.

Abstract: Recent advancements in large reasoning models have fueled growing interest in extending such capabilities to multimodal domains. However, despite notable progress in visual reasoning, the lack of transparent and reproducible data curation and training strategies remains a major barrier to scalable research. In this work, we introduce OpenMMReasoner, a fully transparent two-stage recipe for multimodal reasoning spanning supervised fine-tuning (SFT) and reinforcement learning (RL). In the SFT stage, we construct an 874K-sample cold-start dataset with rigorous step-by-step validation, providing a strong foundation for reasoning capabilities. The subsequent RL stage leverages a 74K-sample dataset across diverse domains to further sharpen and stabilize these abilities, resulting in a more robust and efficient learning process. Extensive evaluations demonstrate that our training recipe not only surpasses strong baselines but also highlights the critical role of data quality and training design in shaping multimodal reasoning performance. Notably, our method achieves a 11.6% improvement over the Qwen2.5-VL-7B-Instruct baseline across nine multimodal reasoning benchmarks, establishing a solid empirical foundation for future large-scale multimodal reasoning research. We open-sourced all our codes, pipeline, and data at https://github.com/EvolvingLMMs-Lab/OpenMMReasoner.

Method: Systematic experimentation with GPT-2 fine-tuned on The Psychology of Artificial Superintelligence, evaluating seven locality configurations (fully distributed to strictly localist) and five progressive schedules with polynomial increases (linear through quintic). Combines adaptive semantic block partitioning with steep polynomial locality schedules.

Result: Progressive semantic localization achieves near-baseline language modeling performance while providing interpretable attention patterns. Results are statistically robust and highly reproducible across multiple independent training runs. Dramatically outperforms both fixed-window localization and naive uniform locality constraints.

Conclusion: Interpretability mechanisms should align with semantic structure to achieve practical performance-interpretability tradeoffs. Steep schedules concentrating locality in decision-critical final layers while preserving distributed learning in early layers achieve optimal results, demonstrating that progressive localization represents the optimal architecture for interpretable LLMs.

Abstract: This paper demonstrates that progressive localization, the gradual increase of attention locality from early distributed layers to late localized layers, represents the optimal architecture for creating interpretable large language models (LLMs) while preserving performance. Through systematic experimentation with GPT-2 fine-tuned on The Psychology of Artificial Superintelligence, we evaluate seven locality configurations ranging from fully distributed to strictly localist, with five progressive schedules implementing polynomial increases (linear through quintic). We investigate whether interpretability constraints can be aligned with natural semantic structure while being applied strategically across network depth. We demonstrate that progressive semantic localization, combining adaptive semantic block partitioning with steep polynomial locality schedules, achieves near-baseline language modeling performance while providing interpretable attention patterns. Multiple independent training runs with different random seeds establish that results are statistically robust and highly reproducible. The approach dramatically outperforms both fixed-window localization and naive uniform locality constraints. Analysis reveals that maintaining flexibility through low-fidelity constraints preserves model capacity while providing interpretability benefits, and that steep schedules concentrating locality in decision-critical final layers while preserving distributed learning in early layers achieve near-baseline attention distribution characteristics. These findings demonstrate that interpretability mechanisms should align with semantic structure to achieve practical performance-interpretability tradeoffs for trustworthy AI systems.

Method: Proposes DANCE (Diverse, Actionable, and kNowledge-Constrained Explanations) that incorporates feature dependencies and causal constraints. Learns linear and nonlinear constraints from data or integrates expert-provided dependency graphs to ensure plausibility and actionability. Balances plausibility, diversity, and sparsity while maintaining consistency with feature relationships.

Result: Extensive evaluation using 140 public datasets shows DANCE outperforms existing approaches based on widely used metrics. Generates meaningful, domain-relevant counterfactuals that align with real-world constraints. Method developed based on real-life case study with Freshmail and supported by Sendguard R&D project.

Conclusion: DANCE effectively addresses key limitations in existing counterfactual explanation algorithms by producing diverse, actionable explanations that respect real-world feature dependencies and causal constraints, making them more plausible and practical for real applications like cybersecurity in email marketing.

Abstract: Counterfactual explanations enhance the actionable interpretability of machine learning models by identifying the minimal changes required to achieve a desired outcome of the model. However, existing methods often ignore the complex dependencies in real-world datasets, leading to unrealistic or impractical modifications. Motivated by cybersecurity applications in the email marketing domain, we propose a method for generating Diverse, Actionable, and kNowledge-Constrained Explanations (DANCE), which incorporates feature dependencies and causal constraints to ensure plausibility and real-world feasibility of counterfactuals. Our method learns linear and nonlinear constraints from data or integrates expert-provided dependency graphs, ensuring counterfactuals are plausible and actionable. By maintaining consistency with feature relationships, the method produces explanations that align with real-world constraints. Additionally, it balances plausibility, diversity, and sparsity, effectively addressing key limitations in existing algorithms. The work is developed based on a real-life case study with Freshmail, the largest email marketing company in Poland and supported by a joint R&D project Sendguard. Furthermore, we provide an extensive evaluation using 140 public datasets, which highlights its ability to generate meaningful, domain-relevant counterfactuals that outperform other existing approaches based on widely used metrics. The source code for reproduction of the results can be found in a GitHub repository we provide.

Method: ICPO calculates preference advantage scores by comparing relative generation probabilities of multiple responses under the same prompt, integrating these with verifiable rewards to guide exploration.

Result: ICPO outperforms GRPO across four general-domain benchmarks and three mathematical benchmarks, demonstrating stable reasoning improvement.

Conclusion: ICPO effectively addresses RLVR limitations by leveraging LLMs’ intrinsic confidence to enhance exploration and reasoning performance.

Abstract: Reinforcement Learning with Verifiable Rewards (RLVR) demonstrates significant potential in enhancing the reasoning capabilities of Large Language Models (LLMs). However, existing RLVR methods are often constrained by issues such as coarse-grained rewards, reward noise, and inefficient exploration, which lead to unstable training and entropy collapse. To address this challenge, we propose the Intrinsic Confidence-Driven Group Relative Preference Optimization method (ICPO). The intuition behind it lies in the fact that the probabilities of an LLM generating different responses can inherently and directly reflect its self-assessment of the reasoning process. Inspired by the idea of preference modeling, ICPO calculates a preference advantage score for each response by comparing the relative generation probabilities of multiple responses under the same input prompt, and integrates this score with verifiable rewards to guide the exploration process. We have discovered that the preference advantage score not only alleviates the issues of coarse-grained rewards and reward noise but also effectively curbs overconfident errors, enhances the relative superiority of undervalued high-quality responses, and prevents the model from overfitting to specific strategies, thereby facilitating more thorough exploration. Comprehensive experiments across four general-domain benchmarks and three mathematical benchmarks demonstrate that ICPO steadily boosts reasoning compared to GRPO.

Method: Evaluated three deep generative models (Variational Autoencoders, Generative Adversarial Networks, Denoising Diffusion Probabilistic Models) against traditional augmentation (time-shifting, vocalization masking) using Southern Resident Killer Whale vocalizations from Salish Sea hydrophone deployments.

Result: All generative approaches improved classification performance over baseline. Diffusion-based augmentation achieved highest recall (0.87) and F1-score (0.75). Hybrid strategy combining generative synthesis with traditional methods yielded best overall F1-score of 0.81.

Conclusion: Deep generative models, particularly diffusion models, provide valuable complementary augmentation strategies that can advance acoustic monitoring of threatened marine mammal populations by improving detection performance without requiring additional field data.

Abstract: Automated detection and classification of marine mammals vocalizations is critical for conservation and management efforts but is hindered by limited annotated datasets and the acoustic complexity of real-world marine environments. Data augmentation has proven to be an effective strategy to address this limitation by increasing dataset diversity and improving model generalization without requiring additional field data. However, most augmentation techniques used to date rely on effective but relatively simple transformations, leaving open the question of whether deep generative models can provide additional benefits. In this study, we evaluate the potential of deep generative for data augmentation in marine mammal call detection including: Variational Autoencoders, Generative Adversarial Networks, and Denoising Diffusion Probabilistic Models. Using Southern Resident Killer Whale (Orcinus orca) vocalizations from two long-term hydrophone deployments in the Salish Sea, we compare these approaches against traditional augmentation methods such as time-shifting and vocalization masking. While all generative approaches improved classification performance relative to the baseline, diffusion-based augmentation yielded the highest recall (0.87) and overall F1-score (0.75). A hybrid strategy combining generative-based synthesis with traditional methods achieved the best overall performance with an F1-score of 0.81. We hope this study encourages further exploration of deep generative models as complementary augmentation strategies to advance acoustic monitoring of threatened marine mammal populations.

Method: Enhanced GLA-Grad by applying Griffin-Lim algorithm correction only once during reverse process instead of multiple times, accelerating generation while maintaining phase consistency.

Result: Method consistently outperforms baseline models, especially in out-of-domain scenarios, with faster generation speed.

Conclusion: Single GLA correction effectively improves vocoder performance and efficiency for diffusion-based speech synthesis, particularly for out-of-domain conditions.

Abstract: Recent advances in diffusion models have positioned them as powerful generative frameworks for speech synthesis, demonstrating substantial improvements in audio quality and stability. Nevertheless, their effectiveness in vocoders conditioned on mel spectrograms remains constrained, particularly when the conditioning diverges from the training distribution. The recently proposed GLA-Grad model introduced a phase-aware extension to the WaveGrad vocoder that integrated the Griffin-Lim algorithm (GLA) into the reverse process to reduce inconsistencies between generated signals and conditioning mel spectrogram. In this paper, we further improve GLA-Grad through an innovative choice in how to apply the correction. Particularly, we compute the correction term only once, with a single application of GLA, to accelerate the generation process. Experimental results demonstrate that our method consistently outperforms the baseline models, particularly in out-of-domain scenarios.

Method: Progressive Unfolding of Residual Entropy (PURE) framework guides multi-stage quantization using a pre-trained speech enhancement model. First stage reconstructs low-entropy, denoised speech embeddings, subsequent stages encode residual high-entropy components.

Result: PURE consistently outperforms conventional RVQ-based codecs in reconstruction and downstream speech language model-based text-to-speech, particularly under noisy training conditions. Training stability is significantly improved.

Conclusion: The PURE Codec framework effectively addresses RVQ limitations by leveraging speech enhancement guidance for progressive quantization, leading to better performance and stability in neural speech compression.

Abstract: Neural speech codecs have achieved strong performance in low-bitrate compression, but residual vector quantization (RVQ) often suffers from unstable training and ineffective decomposition, limiting reconstruction quality and efficiency. We propose PURE Codec (Progressive Unfolding of Residual Entropy), a novel framework that guides multi-stage quantization using a pre-trained speech enhancement model. The first quantization stage reconstructs low-entropy, denoised speech embeddings, while subsequent stages encode residual high-entropy components. This design improves training stability significantly. Experiments demonstrate that PURE consistently outperforms conventional RVQ-based codecs in reconstruction and downstream speech language model-based text-to-speech, particularly under noisy training conditions.

Method: Proposes probability-level calibration and fusion framework for EEND models. Investigates two output formulations: multilabel and powerset representations. Explores joint calibration in powerset space vs independent per-speaker calibration. Compares Fuse-then-Calibrate vs Calibrate-then-Fuse ordering strategies.

Result: Proper calibration provides substantial improvements for individual models (up to 19% relative DER reduction), sometimes mitigating absence of domain adaptation. Joint calibration in powerset space consistently outperforms independent per-speaker calibration. Fuse-then-Calibrate ordering generally outperforms calibrating individual models before fusion. Best configuration outperforms DOVER-Lap in DER while providing reliable confidence estimates.

Conclusion: Proposes best practices for probability-level fusion of EEND systems and demonstrates advantages of leveraging soft outputs over hard decisions. Shows calibration is crucial for reliable confidence estimates essential for downstream applications.

Abstract: End-to-End Neural Diarization (EEND) systems produce frame-level probabilistic speaker activity estimates, yet since evaluation focuses primarily on Diarization Error Rate (DER), the reliability and calibration of these confidence scores have been largely neglected. When fusing multiple diarization systems, DOVER-Lap remains the only established approach, operating at the segment level with hard decisions. We propose working with continuous probability outputs, which enables more sophisticated calibration and fusion techniques that can leverage model uncertainty and complementary strengths across different architectures. This paper presents the first comprehensive framework for calibrating and fusing EEND models at the probability level. We investigate two output formulations (multilabel and powerset representations) and their impact on calibration and fusion effectiveness. Through extensive experiments on the CallHome two-speaker benchmark, we demonstrate that proper calibration provides substantial improvements even for individual models (up to 19% relative DER reduction), in some cases mitigating the absence of domain adaptation. We reveal that joint calibration in powerset space consistently outperforms independent per-speaker calibration, and that the Fuse-then-Calibrate ordering generally outperforms calibrating individual models before fusion while requiring calibration of only a single combined model. Our best configuration outperforms DOVER-Lap in terms of DER while providing reliable confidence estimates essential for downstream applications. This work proposes best practices for probability-level fusion of EEND systems and demonstrates the advantages of leveraging soft outputs over hard decisions.

Method: Developed HPSU benchmark with over 20,000 expert-validated spoken language understanding samples in English and Chinese, using a semi-automatic annotation process that fuses audio, textual, and visual information to enhance annotation efficiency and quality.

Result: Systematic evaluation of various open-source and proprietary Speech LLMs shows that even top-performing models still fall considerably short of human capabilities in understanding genuine spoken interactions.

Conclusion: HPSU provides a comprehensive evaluation framework that will be useful for guiding the development of Speech LLMs toward human-level perception and cognition, addressing the gap between current model performance and human understanding.

Abstract: Recent advances in Speech Large Language Models (Speech LLMs) have led to great progress in speech understanding tasks such as Automatic Speech Recognition (ASR) and Speech Emotion Recognition (SER). However, whether these models can achieve human-level auditory perception, particularly in terms of their ability to comprehend latent intentions and implicit emotions in real-world spoken language, remains underexplored. To this end, we introduce the Human-level Perception in Spoken Speech Understanding (HPSU), a new benchmark for fully evaluating the human-level perceptual and understanding capabilities of Speech LLMs. HPSU comprises over 20,000 expert-validated spoken language understanding samples in English and Chinese. It establishes a comprehensive evaluation framework by encompassing a spectrum of tasks, ranging from basic speaker attribute recognition to complex inference of latent intentions and implicit emotions. To address the issues of data scarcity and high cost of manual annotation in real-world scenarios, we developed a semi-automatic annotation process. This process fuses audio, textual, and visual information to enable precise speech understanding and labeling, thus enhancing both annotation efficiency and quality. We systematically evaluate various open-source and proprietary Speech LLMs. The results demonstrate that even top-performing models still fall considerably short of human capabilities in understanding genuine spoken interactions. Consequently, HPSU will be useful for guiding the development of Speech LLMs toward human-level perception and cognition.

Method: 1) Collect personality annotations for IEMOCAP dataset to create PA-IEMOCAP; 2) Statistical analysis to identify correlations between personality traits and emotional expressions; 3) Propose Temporal Interaction Condition Network (TICN) that integrates personality features with HuBERT-based acoustic features; 4) Develop automatic personality recognition module for practical applications.

Result: 1) Ground-truth personality traits improve valence recognition CCC from 0.698 to 0.785; 2) With automatically predicted personality traits, achieve CCC of 0.776 for valence recognition, representing 11.17% relative improvement over baseline; 3) Statistical analysis confirms significant correlations between personality traits and emotional expressions.

Conclusion: Personality-aware SER is effective and provides a foundation for personality-aware speech processing applications. The integration of personality information significantly enhances emotion recognition, particularly for valence, even when using automatically predicted personality traits.

Abstract: This study investigates the interaction between personality traits and emotion expression, exploring how personality information can improve speech emotion recognition (SER). We collect the personality annotation for the IEMOCAP dataset, making it the first speech dataset that contains both emotion and personality annotations (PA-IEMOCAP), and enabling direct integration of personality traits into SER. Statistical analysis on this dataset identified significant correlations between personality traits and emotional expressions. To extract finegrained personality features, we propose a temporal interaction condition network (TICN), in which personality features are integrated with HuBERT-based acoustic features for SER. Experiments show that incorporating ground-truth personality traits significantly enhances valence recognition, improving the concordance correlation coefficient (CCC) from 0.698 to 0.785 compared to the baseline without personality information. For practical applications in dialogue systems where personality information about the user is unavailable, we develop a front-end module of automatic personality recognition. Using these automatically predicted traits as inputs to our proposed TICN model, we achieve a CCC of 0.776 for valence recognition, representing an 11.17% relative improvement over the baseline. These findings confirm the effectiveness of personality-aware SER and provide a solid foundation for further exploration in personality-aware speech processing applications.

Method: Proposes LAPS-Diff: diffusion model with language-aware embeddings from pre-trained models, vocal-style guided learning with style encoder and pitch extraction for style/pitch losses, and musical/contextual embeddings from MERT and IndicWav2Vec as conditional priors.

Result: LAPS-Diff significantly improves generated sample quality compared to SOTA models for constrained low-resource datasets, as shown by both objective and subjective evaluations.

Conclusion: The proposed approach effectively addresses SVS challenges in low-resource scenarios by integrating language awareness and vocal-style guidance, demonstrating superior performance for Bollywood Hindi singing synthesis.

Abstract: The field of Singing Voice Synthesis (SVS) has seen significant advancements in recent years due to the rapid progress of diffusion-based approaches. However, capturing vocal style, genre-specific pitch inflections, and language-dependent characteristics remains challenging, particularly in low-resource scenarios. To address this, we propose LAPS-Diff, a diffusion model integrated with language-aware embeddings and a vocal-style guided learning mechanism, specifically designed for Bollywood Hindi singing style. We curate a Hindi SVS dataset and leverage pre-trained language models to extract word and phone-level embeddings for an enriched lyrics representation. Additionally, we incorporated a style encoder and a pitch extraction model to compute style and pitch losses, capturing features essential to the naturalness and expressiveness of the synthesized singing, particularly in terms of vocal style and pitch variations. Furthermore, we utilize MERT and IndicWav2Vec models to extract musical and contextual embeddings, serving as conditional priors to refine the acoustic feature generation process further. Based on objective and subjective evaluations, we demonstrate that LAPS-Diff significantly improves the quality of the generated samples compared to the considered state-of-the-art (SOTA) model for our constrained dataset that is typical of the low resource scenario.

Method: Developed a composition algorithm based on Restricted Boltzmann Machine (RBM), converted musical scores into piano-roll image representations, and trained the RBM in an unsupervised manner to generate musical pieces of arbitrary length.

Result: The trained RBM can generate new musical pieces, but analysis of the model’s responses and internal structure reveals that the learned musical information is not stored in a form directly interpretable by humans.

Conclusion: This study contributes to understanding how ML models represent musical structure internally and highlights significant interpretability challenges in generative models for creative tasks like music composition.

Abstract: We investigate how machine learning models acquire the ability to compose music and how musical information is internally represented within such models. We develop a composition algorithm based on a restricted Boltzmann machine (RBM), a simple generative model capable of producing musical pieces of arbitrary length. We convert musical scores into piano-roll image representations and train the RBM in an unsupervised manner. We confirm that the trained RBM can generate new musical pieces; however, by analyzing the model’s responses and internal structure, we find that the learned information is not stored in a form directly interpretable by humans. This study contributes to a better understanding of how machine learning models capable of music composition may internally represent musical structure and highlights issues related to the interpretability of generative models in creative tasks.

Method: CodecSep combines DAC (Discrete Audio Codec) compression with a Transformer masker modulated by CLAP-derived FiLM (Feature-wise Linear Modulation) parameters. This approach enables text-guided separation while maintaining codec efficiency.

Result: CodecSep surpasses AudioSep in separation fidelity (SI-SDR) across six open-domain benchmarks, remains competitive in perceptual quality (ViSQOL), and matches or exceeds fixed-stem baselines. It requires only 1.35 GMACs end-to-end - 54x less compute than spectrogram-domain separators like AudioSep.

Conclusion: CodecSep is the first neural audio codec-based model for on-device universal text-driven separation, offering superior separation fidelity with dramatically reduced computational requirements while maintaining full bitstream compatibility.

Abstract: Text-guided sound separation supports flexible audio editing across media and assistive applications, but existing models like AudioSep are too compute-heavy for edge deployment. Neural audio codec (NAC) models such as CodecFormer and SDCodec are compute-efficient but limited to fixed-class separation. We introduce CodecSep, the first NAC-based model for on-device universal, text-driven separation. CodecSep combines DAC compression with a Transformer masker modulated by CLAP-derived FiLM parameters. Across six open-domain benchmarks under matched training/prompt protocols, \textbf{CodecSep} surpasses \textbf{AudioSep} in separation fidelity (SI-SDR) while remaining competitive in perceptual quality (ViSQOL) and matching or exceeding fixed-stem baselines (TDANet, CodecFormer, SDCodec). In code-stream deployments, it needs just 1.35~GMACs end-to-end – approximately $54\times$ less compute ($25\times$ architecture-only) than spectrogram-domain separators like AudioSep – while remaining fully bitstream-compatible.

Method: Uses interleaved token sequences in a discrete autoregressive backbone with modality-specific decoders to jointly synthesize speech and gestures from text. Supports multi-speaker/multi-style cloning and gesture-only synthesis from speech.

Result: Subjective and objective evaluations demonstrate competitive speech quality and improved gesture generation over unimodal baselines.

Conclusion: Gelina provides a unified framework for joint speech-gesture synthesis that better captures the natural synchrony and prosody alignment of human multimodal communication.

Abstract: Human communication is multimodal, with speech and gestures tightly coupled, yet most computational methods for generating speech and gestures synthesize them sequentially, weakening synchrony and prosody alignment. We introduce Gelina, a unified framework that jointly synthesizes speech and co-speech gestures from text using interleaved token sequences in a discrete autoregressive backbone, with modality-specific decoders. Gelina supports multi-speaker and multi-style cloning and enables gesture-only synthesis from speech inputs. Subjective and objective evaluations demonstrate competitive speech quality and improved gesture generation over unimodal baselines.

Method: The authors introduce STAR-Bench with two settings: 1) Foundational Acoustic Perception (six attributes under absolute/relative regimes), and 2) Holistic Spatio-Temporal Reasoning (segment reordering, spatial tasks including static localization, multi-source relations, and dynamic trajectories). They use procedurally synthesized/physics-simulated audio for foundational tasks and a four-stage human annotation process for holistic data.

Result: Evaluation of 19 models shows substantial gaps compared to humans, with far larger accuracy drops (-31.5% temporal, -35.2% spatial) than prior benchmarks. Closed-source models are bottlenecked by fine-grained perception, while open-source models lag across perception, knowledge, and reasoning.

Conclusion: STAR-Bench provides critical insights and a clear path forward for developing models with more robust understanding of the physical world, highlighting the need for improved fine-grained perceptual reasoning in audio-language models.

Abstract: Despite rapid progress in Multi-modal Large Language Models and Large Audio-Language Models, existing audio benchmarks largely test semantics that can be recovered from text captions, masking deficits in fine-grained perceptual reasoning. We formalize audio 4D intelligence that is defined as reasoning over sound dynamics in time and 3D space, and introduce STAR-Bench to measure it. STAR-Bench combines a Foundational Acoustic Perception setting (six attributes under absolute and relative regimes) with a Holistic Spatio-Temporal Reasoning setting that includes segment reordering for continuous and discrete processes and spatial tasks spanning static localization, multi-source relations, and dynamic trajectories. Our data curation pipeline uses two methods to ensure high-quality samples. For foundational tasks, we use procedurally synthesized and physics-simulated audio. For holistic data, we follow a four-stage process that includes human annotation and final selection based on human performance. Unlike prior benchmarks where caption-only answering reduces accuracy slightly, STAR-Bench induces far larger drops (-31.5% temporal, -35.2% spatial), evidencing its focus on linguistically hard-to-describe cues. Evaluating 19 models reveals substantial gaps compared with humans and a capability hierarchy: closed-source models are bottlenecked by fine-grained perception, while open-source models lag across perception, knowledge, and reasoning. Our STAR-Bench provides critical insights and a clear path forward for developing future models with a more robust understanding of the physical world.

Method: VeM uses hierarchical video parsing as a music conductor, modality-specific encoders with storyboard-guided cross-attention (SG-CAtt), and frame-level transition-beat aligner/adapter (TB-As) for rhythmic synchronization. Also introduces a new video-music dataset from e-commerce ads and video-sharing platforms.

Result: Experimental results demonstrate superiority over existing methods, particularly in semantic relevance and rhythmic precision.

Conclusion: VeM effectively addresses the limitations of current video-to-music generation by providing comprehensive video representation and precise rhythmic alignment through its novel hierarchical parsing and transition-beat synchronization mechanisms.

Abstract: Video-to-Music generation seeks to generate musically appropriate background music that enhances audiovisual immersion for videos. However, current approaches suffer from two critical limitations: 1) incomplete representation of video details, leading to weak alignment, and 2) inadequate temporal and rhythmic correspondence, particularly in achieving precise beat synchronization. To address the challenges, we propose Video Echoed in Music (VeM), a latent music diffusion that generates high-quality soundtracks with semantic, temporal, and rhythmic alignment for input videos. To capture video details comprehensively, VeM employs a hierarchical video parsing that acts as a music conductor, orchestrating multi-level information across modalities. Modality-specific encoders, coupled with a storyboard-guided cross-attention mechanism (SG-CAtt), integrate semantic cues while maintaining temporal coherence through position and duration encoding. For rhythmic precision, the frame-level transition-beat aligner and adapter (TB-As) dynamically synchronize visual scene transitions with music beats. We further contribute a novel video-music paired dataset sourced from e-commerce advertisements and video-sharing platforms, which imposes stricter transition-beat synchronization requirements. Meanwhile, we introduce novel metrics tailored to the task. Experimental results demonstrate superiority, particularly in semantic relevance and rhythmic precision.

Method: 1) Decomposes reasoning into four specialized CoT modules (Semantic, Temporal, Aesthetic, Spatial) each with targeted reward functions; 2) Uses multidimensional RL optimization with CoT-reward correspondence; 3) Introduces Fast-GRPO with hybrid ODE-SDE sampling for computational efficiency; 4) Creates AudioCanvas benchmark with balanced distribution and diverse scenarios.

Result: Achieves state-of-the-art performance across all four perceptual dimensions on both in-domain VGGSound test set and out-of-domain AudioCanvas benchmark, demonstrating superior audio generation quality and alignment with human preferences.

Conclusion: PrismAudio successfully addresses objective entanglement in V2A generation through specialized CoT planning with RL, provides computational efficiency via Fast-GRPO, and establishes a rigorous benchmark for future research, advancing the field toward more human-aligned audio generation.

Abstract: Video-to-Audio (V2A) generation requires balancing four critical perceptual dimensions: semantic consistency, audio-visual temporal synchrony, aesthetic quality, and spatial accuracy; yet existing methods suffer from objective entanglement that conflates competing goals in single loss functions and lack human preference alignment. We introduce PrismAudio, the first framework to integrate Reinforcement Learning into V2A generation with specialized Chain-of-Thought (CoT) planning. Our approach decomposes monolithic reasoning into four specialized CoT modules (Semantic, Temporal, Aesthetic, and Spatial CoT), each paired with targeted reward functions. This CoT-reward correspondence enables multidimensional RL optimization that guides the model to jointly generate better reasoning across all perspectives, solving the objective entanglement problem while preserving interpretability. To make this optimization computationally practical, we propose Fast-GRPO, which employs hybrid ODE-SDE sampling that dramatically reduces the training overhead compared to existing GRPO implementations. We also introduce AudioCanvas, a rigorous benchmark that is more distributionally balanced and covers more realistically diverse and challenging scenarios than existing datasets, with 300 single-event classes and 501 multi-event samples. Experimental results demonstrate that PrismAudio achieves state-of-the-art performance across all four perceptual dimensions on both the in-domain VGGSound test set and out-of-domain AudioCanvas benchmark. The project page is available at https://PrismAudio-Project.github.io.

Method: Developed MARS-S2L, a machine learning model trained on a manually curated dataset of over 80,000 multispectral satellite images. The model provides high-resolution detections every two days from publicly available satellite imagery.

Result: The model detects 78% of methane plumes with an 8% false positive rate at 697 previously unseen sites. Deployed operationally, it has issued 1,015 notifications to stakeholders in 20 countries, enabling verified permanent mitigation of six persistent emitters including a previously unknown site in Libya.

Conclusion: MARS-S2L demonstrates a scalable pathway from satellite detection to quantifiable methane mitigation, enabling facility-level attribution and targeted emission reductions.

Abstract: Methane is a potent greenhouse gas, responsible for roughly 30% of warming since pre-industrial times. A small number of large point sources account for a disproportionate share of emissions, creating an opportunity for substantial reductions by targeting relatively few sites. Detection and attribution of large emissions at scale for notification to asset owners remains challenging. Here, we introduce MARS-S2L, a machine learning model that detects methane emissions in publicly available multispectral satellite imagery. Trained on a manually curated dataset of over 80,000 images, the model provides high-resolution detections every two days, enabling facility-level attribution and identifying 78% of plumes with an 8% false positive rate at 697 previously unseen sites. Deployed operationally, MARS-S2L has issued 1,015 notifications to stakeholders in 20 countries, enabling verified, permanent mitigation of six persistent emitters, including a previously unknown site in Libya. These results demonstrate a scalable pathway from satellite detection to quantifiable methane mitigation.

Method: Proposes Physics-Informed Spiking Neural Network (PISNN) with two key innovations: 1) Conservative Leaky Integrate-and-Fire (C-LIF) neuron that structurally guarantees local mass preservation, and 2) Conservative Flux Quantization (CFQ) strategy that redefines neural spikes as discrete packets of physical flux, learning a time-invariant physical evolution operator.

Result: PISNN excels on diverse benchmarks including 1D heat equation and 2D Laplace’s Equation, accurately simulating system dynamics while maintaining perfect mass conservation by design - a feat challenging for conventional PINNs.

Conclusion: Establishes a robust framework for fusing scientific computing rigor with neuromorphic engineering efficiency, enabling complex, long-term, energy-efficient physics predictions for intelligent systems.

Abstract: Real-time, physically-consistent predictions on low-power edge devices is critical for the next generation embodied AI systems, yet it remains a major challenge. Physics-Informed Neural Networks (PINNs) combine data-driven learning with physics-based constraints to ensure the model’s predictions are with underlying physical principles.However, PINNs are energy-intensive and struggle to strictly enforce physical conservation laws. Brain-inspired spiking neural networks (SNNs) have emerged as a promising solution for edge computing and real-time processing. However, naively converting PINNs to SNNs degrades physical fidelity and fails to address long-term generalization issues. To this end, this paper introduce a novel Physics-Informed Spiking Neural Network (PISNN) framework. Importantly, to ensure strict physical conservation, we design the Conservative Leaky Integrate-and-Fire (C-LIF) neuron, whose dynamics structurally guarantee local mass preservation. To achieve robust temporal generalization, we introduce a novel Conservative Flux Quantization (CFQ) strategy, which redefines neural spikes as discrete packets of physical flux. Our CFQ learns a time-invariant physical evolution operator, enabling the PISNN to become a general-purpose solver – conservative-by-construction. Extensive experiments show that our PISNN excels on diverse benchmarks. For both the canonical 1D heat equation and the more challenging 2D Laplace’s Equation, it accurately simulates the system dynamics while maintaining perfect mass conservation by design – a feat that is challenging for conventional PINNs. This work establishes a robust framework for fusing the rigor of scientific computing with the efficiency of neuromorphic engineering, paving the way for complex, long-term, and energy-efficient physics predictions for intelligent systems.

Method: DINR treats feature evolution as a continuous-time dynamical system rather than discrete layers, enabling richer frequency representations through continuous feature evolution. The approach includes theoretical analysis using Rademacher complexity and Neural Tangent Kernel, with regularization of dynamics complexity to balance expressivity and generalization.

Result: Extensive experiments on image representation, field reconstruction, and data compression show DINR achieves more stable convergence, higher signal fidelity, and stronger generalization compared to conventional static INRs.

Conclusion: DINR provides a principled dynamical system framework that effectively mitigates spectral bias in INRs, offering improved expressivity, training dynamics, and generalization through continuous feature evolution.

Abstract: Implicit Neural Representations (INRs) provide a powerful continuous framework for modeling complex visual and geometric signals, but spectral bias remains a fundamental challenge, limiting their ability to capture high-frequency details. Orthogonal to existing remedy strategies, we introduce Dynamical Implicit Neural Representations (DINR), a new INR modeling framework that treats feature evolution as a continuous-time dynamical system rather than a discrete stack of layers. This dynamical formulation mitigates spectral bias by enabling richer, more adaptive frequency representations through continuous feature evolution. Theoretical analysis based on Rademacher complexity and the Neural Tangent Kernel demonstrates that DINR enhances expressivity and improves training dynamics. Moreover, regularizing the complexity of the underlying dynamics provides a principled way to balance expressivity and generalization. Extensive experiments on image representation, field reconstruction, and data compression confirm that DINR delivers more stable convergence, higher signal fidelity, and stronger generalization than conventional static INRs.

Method: Replaces probabilistic interpretation of softmax with geometric interpretation on multidimensional simplex. Introduces multidimensional thresholds that can be tuned a posteriori after network training. Develops multiclass ROC analysis using ROC clouds (attainable FPR/TPR points) and summarizes them with Distance From Point (DFP) score to (0,1).

Result: Multidimensional threshold tuning yields performance improvements across various networks and datasets. The proposed multiclass ROC analysis provides coherent alternative to standard One-vs-Rest curves and aligns with observed tuning gains.

Conclusion: The threshold-based framework enables a posteriori optimization of classification performance for any trained network through threshold tuning, generalizing binary classification techniques to multiclass settings and providing improved performance metrics.

Abstract: In this paper, we introduce a threshold-based framework for multiclass classification that generalizes the standard argmax rule. This is done by replacing the probabilistic interpretation of softmax outputs with a geometric one on the multidimensional simplex, where the classification depends on a multidimensional threshold. This change of perspective enables for any trained classification network an \textit{a posteriori} optimization of the classification score by means of threshold tuning, as usually carried out in the binary setting, thus allowing for a further refinement of the prediction capability of any network. Our experiments show indeed that multidimensional threshold tuning yields performance improvements across various networks and datasets. Moreover, we derive a multiclass ROC analysis based on \emph{ROC clouds} – the attainable (FPR,TPR) operating points induced by a single multiclass threshold – and summarize them via a \emph{Distance From Point} (DFP) score to $(0,1)$. This yields a coherent alternative to standard One-vs-Rest (OvR) curves and aligns with the observed tuning gains.

Method: 1) Map raw EEG into audio codec’s stride-based framing for direct reuse of pretrained encoder-decoder; 2) Fine-tune on EEG data; 3) Explore compression-quality trade-offs via residual codebook depth, vocabulary size, and input sampling rate; 4) Propose DAC-MC extension with attention-based cross-channel aggregation and channel-specific decoding for spatial dependencies.

Result: The adapted codecs yield stable EEG reconstructions, with fine-tuning improving fidelity and generalization compared to training from scratch. Evaluations on TUH Abnormal and Epilepsy datasets show preservation of clinically relevant information as measured by spectrogram-based reconstruction loss and downstream classification accuracy.

Conclusion: Pretrained neural audio codecs can serve as effective starting points for EEG compression with appropriate preprocessing, enabling efficient reuse of audio-pretrained models while capturing EEG-specific characteristics through targeted modifications.

Abstract: EEG and audio are inherently distinct modalities, differing in sampling rate, channel structure, and scale. Yet, we show that pretrained neural audio codecs can serve as effective starting points for EEG compression, provided that the data are preprocessed to be suitable to the codec’s input constraints. Using DAC, a state-of-the-art neural audio codec as our base, we demonstrate that raw EEG can be mapped into the codec’s stride-based framing, enabling direct reuse of the audio-pretrained encoder-decoder. Even without modification, this setup yields stable EEG reconstructions, and fine-tuning on EEG data further improves fidelity and generalization compared to training from scratch. We systematically explore compression-quality trade-offs by varying residual codebook depth, codebook (vocabulary) size, and input sampling rate. To capture spatial dependencies across electrodes, we propose DAC-MC, a multi-channel extension with attention-based cross-channel aggregation and channel-specific decoding, while retaining the audio-pretrained initialization. Evaluations on the TUH Abnormal and Epilepsy datasets show that the adapted codecs preserve clinically relevant information, as reflected in spectrogram-based reconstruction loss and downstream classification accuracy.

Method: Theoretical analysis combined with empirical studies of sparse decision trees and linear models. The research examines how Rashomon sets affect adversarial robustness, privacy preservation, and stability under distribution shifts.

Result: Sparse interpretable models preserve privacy but are fragile to attacks. Rashomon sets enable reactive robustness (when one model breaks, others remain accurate) and stability under distribution shifts. However, diversity in Rashomon sets increases information leakage, as disclosing more models gives attackers richer views of training data.

Conclusion: Rashomon sets play a dual role in trustworthy ML: they serve as a resource for robustness and stability, but also pose a privacy risk through increased information leakage. This creates a fundamental robustness-privacy trade-off that must be managed in ML pipeline design.

Abstract: Real-world machine learning (ML) pipelines rarely produce a single model; instead, they produce a Rashomon set of many near-optimal ones. We show that this multiplicity reshapes key aspects of trustworthiness. At the individual-model level, sparse interpretable models tend to preserve privacy but are fragile to adversarial attacks. In contrast, the diversity within a large Rashomon set enables reactive robustness: even when an attack breaks one model, others often remain accurate. Rashomon sets are also stable under small distribution shifts. However, this same diversity increases information leakage, as disclosing more near-optimal models provides an attacker with progressively richer views of the training data. Through theoretical analysis and empirical studies of sparse decision trees and linear models, we characterize this robustness-privacy trade-off and highlight the dual role of Rashomon sets as both a resource and a risk for trustworthy ML.

Method: Evaluated two unsupervised anomaly detection techniques (Isolation Forest and One-Class SVM) using the TON_IoT thermostat dataset, measuring both standard performance metrics and resource utilization metrics.

Result: Isolation Forest consistently outperformed OC-SVM with higher detection accuracy, superior precision and recall, better F1-score, and significantly lower computational footprint (inference time, model size, and RAM usage).

Conclusion: Isolation Forest is robust for high-dimensional imbalanced IoT environments and practically viable for real-time anomaly detection on resource-constrained IoT edge devices.

Abstract: The rapid expansion of Internet of Things (IoT) deployments across diverse sectors has significantly enhanced operational efficiency, yet concurrently elevated cybersecurity vulnerabilities due to increased exposure to cyber threats. Given the limitations of traditional signature-based Anomaly Detection Systems (ADS) in identifying emerging and zero-day threats, this study investigates the effectiveness of two unsupervised anomaly detection techniques, Isolation Forest (IF) and One-Class Support Vector Machine (OC-SVM), using the TON_IoT thermostat dataset. A comprehensive evaluation was performed based on standard metrics (accuracy, precision, recall, and F1-score) alongside critical resource utilization metrics such as inference time, model size, and peak RAM usage. Experimental results revealed that IF consistently outperformed OC-SVM, achieving higher detection accuracy, superior precision, and recall, along with a significantly better F1-score. Furthermore, Isolation Forest demonstrated a markedly superior computational footprint, making it more suitable for deployment on resource-constrained IoT edge devices. These findings underscore Isolation Forest’s robustness in high-dimensional and imbalanced IoT environments and highlight its practical viability for real-time anomaly detection.

Method: Developed a pipeline that takes kinematics data from neuroscience labs and creates biomechanical models, implemented an imitation learning framework for dexterous forelimb reaching tasks using musculoskeletal models in simulated physics environments (Mujoco-MJX with JAX GPU acceleration).

Result: Achieved training at faster than 1 million steps per second with GPU acceleration; demonstrated that adding naturalistic constraints on energy and velocity leads to simulated musculoskeletal activity that better predicts real EMG signals.

Conclusion: Energy and control constraints are critical for accurately modeling musculoskeletal motor control, providing evidence that naturalistic biomechanical constraints improve the fidelity of embodied control simulations.

Abstract: The brain has evolved to effectively control the body, and in order to understand the relationship we need to model the sensorimotor transformations underlying embodied control. As part of a coordinated effort, we are developing a general-purpose platform for behavior-driven simulation modeling high fidelity behavioral dynamics, biomechanics, and neural circuit architectures underlying embodied control. We present a pipeline for taking kinematics data from the neuroscience lab and creating a pipeline for recapitulating those natural movements in a biomechanical model. We implement a imitation learning framework to perform a dexterous forelimb reaching task with a musculoskeletal model in a simulated physics environment. The mouse arm model is currently training at faster than 1 million training steps per second due to GPU acceleration with JAX and Mujoco-MJX. We present results that indicate that adding naturalistic constraints on energy and velocity lead to simulated musculoskeletal activity that better predict real EMG signals. This work provides evidence to suggest that energy and control constraints are critical to modeling musculoskeletal motor control.

Method: Used AirQualityUCI dataset collected over one year in urban environment. Evaluated full-capacity and lightweight (tiny) XGBoost regression models using MAE, RMSE, MBE, R2 metrics, plus inference time, model size, and peak RAM usage.

Result: Full XGBoost achieved superior predictive accuracy for both CO and NO2. Tiny model had slightly lower precision but offered substantial computational benefits: significantly reduced inference time and model storage requirements.

Conclusion: Simplified models can be deployed in resource-constrained environments without compromising predictive quality. Tiny XGBoost model is suitable for real-time air-quality monitoring in IoT and embedded applications.

Abstract: This study investigates the effectiveness and efficiency of two variants of the XGBoost regression model, the full-capacity and lightweight (tiny) versions, for predicting the concentrations of carbon monoxide (CO) and nitrogen dioxide (NO2). Using the AirQualityUCI dataset collected over one year in an urban environment, we conducted a comprehensive evaluation based on widely accepted metrics, including Mean Absolute Error (MAE), Root Mean Square Error (RMSE), Mean Bias Error (MBE), and the coefficient of determination (R2). In addition, we assessed resource-oriented metrics such as inference time, model size, and peak RAM usage. The full XGBoost model achieved superior predictive accuracy for both pollutants, while the tiny model, though slightly less precise, offered substantial computational benefits with significantly reduced inference time and model storage requirements. These results demonstrate the feasibility of deploying simplified models in resource-constrained environments without compromising predictive quality. This makes the tiny XGBoost model suitable for real-time air-quality monitoring in IoT and embedded applications.

Method: PDE-FM combines spatial-spectral tokenization, physics-aware conditioning, and a Mamba-based state-space backbone with an operator-theoretic decoder. It’s pretrained once on diverse PDE datasets and can be transferred to new physical regimes without modifications.

Result: Achieves state-of-the-art accuracy in six domains on twelve 2D/3D datasets from The Well benchmark, reducing mean VRMSE by 46% relative to prior operator-learning baselines. Demonstrates robust cross-physics generalization, excelling in turbulent and radiative systems.

Conclusion: Large-scale pretraining across diverse physical processes can yield transferable representations of dynamics, marking progress toward unified foundation-level surrogates for multi-physics simulation and scientific discovery.

Abstract: We present PDE-FM, a modular foundation model for physics-informed machine learning that unifies spatial, spectral, and temporal reasoning across heterogeneous partial differential equation (PDE) systems. PDE-FM combines spatial-spectral tokenization, physics-aware conditioning, and a Mamba-based state-space backbone with an operator-theoretic decoder, enabling scalable and data-efficient modeling of complex physical dynamics. In contrast to task-specific neural operators, PDE-FM is pretrained once on diverse PDE datasets and can be transferred to new physical regimes without architectural or data-specific modifications. Evaluated on twelve 2D and 3D datasets from The Well benchmark - spanning hydrodynamic, radiative, elastic, and astrophysical phenomena - PDE-FM achieves state-of-the-art accuracy in six domains, reducing mean VRMSE by 46% relative to prior operator-learning baselines. The model demonstrates robust cross-physics generalization, excelling in turbulent and radiative systems while maintaining strong performance in linear and steady-state regimes. These results suggest that large-scale pretraining across diverse physical processes can yield transferable representations of dynamics, marking a step toward unified, foundation-level surrogates for multi-physics simulation and scientific discovery.

Method: Introduce Equilibrium Transformers (EqT) with Equilibrium Refinement Module that minimizes a learned energy function via gradient descent in latent space. The energy function enforces bidirectional prediction consistency, episodic memory coherence, and output confidence without external supervision.

Result: Theoretically proven to perform approximate MAP inference in latent energy-based model with linear convergence guarantees. Preliminary experiments on binary parity task show +3.28% average improvement on challenging sequences, with gains reaching +8.07% where standard transformers approach random performance.

Conclusion: Closed-loop equilibrium may resolve the commitment bottleneck of open-loop autoregression, representing a foundational step toward more capable language models. The framework unifies deep equilibrium models, diffusion language models, and test-time training as special cases.

Abstract: Contemporary autoregressive transformers operate in open loop: each hidden state is computed in a single forward pass and never revised, causing errors to propagate uncorrected through the sequence. We identify this open-loop bottleneck as a fundamental architectural limitation underlying well-documented failures in long-range reasoning, factual consistency, and multi-step planning. To address this limitation, we introduce the closed-loop prediction principle, which requires that models iteratively refine latent representations until reaching a self-consistent equilibrium before committing to each token. We instantiate this principle as Equilibrium Transformers (EqT), which augment standard transformer layers with an Equilibrium Refinement Module that minimizes a learned energy function via gradient descent in latent space. The energy function enforces bidirectional prediction consistency, episodic memory coherence, and output confidence, all computed without external supervision. Theoretically, we prove that EqT performs approximate MAP inference in a latent energy-based model, establish linear convergence guarantees, and show that refinement improves predictions precisely on hard instances where one-shot inference is suboptimal. The framework unifies deep equilibrium models, diffusion language models, and test-time training as special cases. Preliminary experiments on the binary parity task demonstrate +3.28% average improvement on challenging sequences, with gains reaching +8.07% where standard transformers approach random performance, validating that the benefit of deliberation scales with task difficulty. Just as attention mechanisms resolved the sequential bottleneck of recurrent networks, we propose that closed-loop equilibrium may resolve the commitment bottleneck of open-loop autoregression, representing a foundational step toward language models.

Method: Proposes Gaussian Mixture Variational Autoencoder (GM-VAE) with EM-inspired training scheme using block-coordinate descent strategy that alternates between expectation and maximization steps. Introduces spectral interpretability metric based on graph-Laplacian smoothness to evaluate learned representations.

Result: Demonstrated on datasets of increasing complexity: surface reaction ODEs, Navier-Stokes wake flows, and experimental laser-induced combustion Schlieren images. The framework yields smooth, physically consistent manifolds and accurate regime clustering.

Conclusion: GM-VAE offers a robust data-driven tool for interpreting turbulent and reactive flow systems by providing stable training and physically aligned latent representations with objective evaluation metrics.

Abstract: Extracting compact, physically interpretable representations from high-dimensional scientific data is a persistent challenge due to the complex, nonlinear structures inherent in physical systems. We propose a Gaussian Mixture Variational Autoencoder (GM-VAE) framework designed to address this by integrating an Expectation-Maximization (EM)-inspired training scheme with a novel spectral interpretability metric. Unlike conventional VAEs that jointly optimize reconstruction and clustering (often leading to training instability), our method utilizes a block-coordinate descent strategy, alternating between expectation and maximization steps. This approach stabilizes training and naturally aligns latent clusters with distinct physical regimes. To objectively evaluate the learned representations, we introduce a quantitative metric based on graph-Laplacian smoothness, which measures the coherence of physical quantities across the latent manifold. We demonstrate the efficacy of this framework on datasets of increasing complexity: surface reaction ODEs, Navier-Stokes wake flows, and experimental laser-induced combustion Schlieren images. The results show that our GM-VAE yields smooth, physically consistent manifolds and accurate regime clustering, offering a robust data-driven tool for interpreting turbulent and reactive flow systems.

Method: Proposed hybrid BERT+vision network framework with two vision backbones (MobileNetV2 and ViT). Created three fusion strategies per backbone: late, intermediate, and early fusion. Evaluated on CMU MOSI dataset and benchmarked latency on NVIDIA Jetson Orin AGX.

Result: Late fusion achieved highest accuracy, early fusion offered lowest inference latency. There’s a clear tradeoff: earlier fusion reduces latency but sacrifices accuracy.

Conclusion: Data fusion earlier in model architecture results in faster inference times at the cost of accuracy, providing practical guidance for multimodal system design based on application requirements.

Abstract: Modern machine learning models often combine multiple input streams of data to more accurately capture the information that informs their decisions. In multimodal machine learning, choosing the strategy for fusing data together requires careful consideration of the application’s accuracy and latency requirements, as fusing the data at earlier or later stages in the model architecture can lead to performance changes in accuracy and latency. To demonstrate this tradeoff, we investigate different fusion strategies using a hybrid BERT and vision network framework that integrates image and text data. We explore two different vision networks: MobileNetV2 and ViT. We propose three models for each vision network, which fuse data at late, intermediate, and early stages in the architecture. We evaluate the proposed models on the CMU MOSI dataset and benchmark their latency on an NVIDIA Jetson Orin AGX. Our experimental results demonstrate that while late fusion yields the highest accuracy, early fusion offers the lowest inference latency. We describe the three proposed model architectures and discuss the accuracy and latency tradeoffs, concluding that data fusion earlier in the model architecture results in faster inference times at the cost of accuracy.

Method: Uses generative models (like VAEs) to reconstruct inputs from attacker-perturbed inputs, plus generative sampling with consensus-based aggregation over generated samples.

Result: Reduces illusion attack success rates to near-zero, improves cross-modal alignment by 4% (42 to 46) in unperturbed settings and 11% (32 to 43) in perturbed settings.

Conclusion: Provides an effective and model-agnostic defense against adversarial illusions in multi-modal foundation models.

Abstract: Multi-modal foundation models align images, text, and other modalities in a shared embedding space but remain vulnerable to adversarial illusions (Zhang et al., 2025), where imperceptible perturbations disrupt cross-modal alignment and mislead downstream tasks. To counteract the effects of adversarial illusions, we propose a task-agnostic mitigation mechanism that reconstructs the input from the attacker’s perturbed input through generative models, e.g., Variational Autoencoders (VAEs), to maintain natural alignment. To further enhance our proposed defense mechanism, we adopt a generative sampling strategy combined with a consensus-based aggregation scheme over the outcomes of the generated samples. Our experiments on the state-of-the-art multi-modal encoders show that our approach substantially reduces the illusion attack success rates to near-zero and improves cross-modal alignment by 4% (42 to 46) and 11% (32 to 43) in unperturbed and perturbed input settings respectively, providing an effective and model-agnostic defense against adversarial illusions.

Method: Used independent PPO with increased entropy coefficient (0.05), higher GAE lambda (0.9), and RNN-based actor-critic architecture instead of feed-forward networks, focusing on cross-play evaluation between different random seeds.

Result: Achieved new state-of-the-art cross-play performance in Hanabi, significantly outperforming all previous specialized algorithms designed for zero-shot coordination, while also identifying limitations in simple Dec-POMDPs.

Conclusion: Hyperparameter tuning (especially entropy regularization) dramatically affects cross-play performance, but standard methods still have limitations in some Dec-POMDPs, indicating continued need for new zero-shot coordination algorithms.

Abstract: We find that in Hanabi, one of the most complex and popular benchmarks for zero-shot coordination and ad-hoc teamplay, a standard implementation of independent PPO with a slightly higher entropy coefficient 0.05 instead of the typically used 0.01, achieves a new state-of-the-art in cross-play between different seeds, beating by a significant margin all previous specialized algorithms, which were specifically designed for this setting. We provide an intuition for why sufficiently high entropy regularization ensures that different random seed produce joint policies which are mutually compatible. We also empirically find that a high $λ_{\text{GAE}}$ around 0.9, and using RNNs instead of just feed-forward layers in the actor-critic architecture, strongly increase inter-seed cross-play. While these results demonstrate the dramatic effect that hyperparameters can have not just on self-play scores but also on cross-play scores, we show that there are simple Dec-POMDPs though, in which standard policy gradient methods with increased entropy regularization are not able to achieve perfect inter-seed cross-play, thus demonstrating the continuing necessity for new algorithms for zero-shot coordination.

Method: Compare three voxel-based input types for 3D CNNs: (1) atom types, (2) raw electron density, and (3) density gradient magnitude. Test on two tasks: protein-ligand binding affinity prediction (PDBbind) and quantum property prediction (QM9). Use voxel-based CNNs because electron density is inherently volumetric and voxel grids provide natural representation for both experimental and computed densities.

Result: On PDBbind: All representations perform similarly with full data, but in low-data regimes, density-based inputs outperform atom types. Shape-based baseline performs comparably, suggesting spatial occupancy dominates this task. On QM9: Density-based inputs outperform atom-based ones at scale, even when labels come from DFT but input densities from lower-level XTB method, reflecting rich structural/electronic information in density.

Conclusion: Electron density-derived inputs offer task- and regime-dependent advantages: improving data efficiency in affinity prediction and accuracy in quantum property modeling. Density maps capture richer physical information than atom-based representations, making them valuable for molecular property prediction.

Abstract: Machine learning models for 3D molecular property prediction typically rely on atom-based representations, which may overlook subtle physical information. Electron density maps – the direct output of X-ray crystallography and cryo-electron microscopy – offer a continuous, physically grounded alternative. We compare three voxel-based input types for 3D convolutional neural networks (CNNs): atom types, raw electron density, and density gradient magnitude, across two molecular tasks – protein-ligand binding affinity prediction (PDBbind) and quantum property prediction (QM9). We focus on voxel-based CNNs because electron density is inherently volumetric, and voxel grids provide the most natural representation for both experimental and computed densities. On PDBbind, all representations perform similarly with full data, but in low-data regimes, density-based inputs outperform atom types, while a shape-based baseline performs comparably – suggesting that spatial occupancy dominates this task. On QM9, where labels are derived from Density Functional Theory (DFT) but input densities from a lower-level method (XTB), density-based inputs still outperform atom-based ones at scale, reflecting the rich structural and electronic information encoded in density. Overall, these results highlight the task- and regime-dependent strengths of density-derived inputs, improving data efficiency in affinity prediction and accuracy in quantum property modeling.

Method: AUTO uses two collaborative LLM agents: a Strategist that selects between exploration and exploitation strategies, and an Implementor that executes detailed designs. It treats design optimization as a gradient-free search problem guided by strategic reasoning.

Result: Applied to GPU code optimization, AUTO generates solutions competitive with expert implementations for chemical kinetics integration and dense matrix multiplication. Achieves 50-70% search efficiency relative to Bayesian optimization, completes optimizations in ~8 hours at estimated cost of up to $159 per run.

Conclusion: The framework opens the door to automating design optimization in ill-defined search spaces with limited prior information, offering cost-effective alternatives to traditional optimization methods and human experts.

Abstract: Traditional optimization methods excel in well-defined search spaces but struggle with design problems where transformations and design parameters are difficult to define. Large language models (LLMs) offer a promising alternative by dynamically interpreting design spaces and leveraging encoded domain knowledge. To this end, we introduce AUTO, an LLM agent framework that treats design optimization as a gradient-free search problem guided by strategic LLM reasoning. The framework employs two collaborative agents: a Strategist that selects between exploration and exploitation strategies, and an Implementor that executes detailed designs. Applied to GPU code optimization – a domain critical to fields from machine learning to scientific computing – AUTO generates solutions competitive with expert implementations for chemical kinetics integration and dense matrix multiplication. The framework achieves 50-70% search efficiency relative to Bayesian optimization methodologies. It completes optimizations in approximately 8 hours at an estimated cost of up to $159 per run, compared to an estimated cost of up to $480 with median-wage software developers. These findings open the door to automating design optimization in ill-defined search spaces with limited prior information.

Method: Multi-modal ML framework using facial video (processed with OpenCV and transformer model for emotional features) and GSR signals (decomposed into tonic/phasic components). Analyzed two temporal windows (6-3s and 3-0s before decision-making) with unimodal and multimodal stacking ensemble approaches.

Result: Multimodal stacking achieved accuracy 0.83, F1 0.88, ROC-AUC 0.87 in Early Window (6-3s); accuracy 0.75, F1 0.82, ROC-AUC 0.66 in Proximal Window (3-0s). Combining facial and physiological cues significantly improved prediction performance.

Conclusion: Bio signals serve as real-time, objective markers of user trust, enabling adaptive AI systems that dynamically adjust responses to maintain calibrated trust in mental health applications.

Abstract: Predicting human trust in AI systems is crucial for safe integration of AI-based decision support tools, especially in healthcare. This study proposes a multi-modal machine learning framework that combines image and galvanic skin response (GSR) data to predict early user trust in AI- or human-generated recommendations in a simulated ADHD mHealth context. Facial video data were processed using OpenCV for frame extraction and transferred learning with a pre-trained transformer model to derive emotional features. Concurrently, GSR signals were decomposed into tonic and phasic components to capture physiological arousal patterns. Two temporal windows were defined for trust prediction: the Early Detection Window (6 to 3 seconds before decision-making) and the Proximal Detection Window (3 to 0 seconds before decision-making). For each window, trust prediction was conducted separately using image-based, GSR-based, and multimodal (image + GSR) features. Each modality was analyzed using machine learning algorithms, and the top-performing unimodal models were integrated through a multimodal stacking ensemble for final prediction. Experimental results showed that combining facial and physiological cues significantly improved prediction performance. The multimodal stacking framework achieved an accuracy of 0.83, F1-score of 0.88, and ROC-AUC of 0.87 in the Early Detection Window, and an accuracy of 0.75, F1-score of 0.82, and ROC-AUC of 0.66 in the Proximal Detection Window. These results demonstrate the potential of bio signals as real-time, objective markers of user trust, enabling adaptive AI systems that dynamically adjust their responses to maintain calibrated trust which is a critical capability in mental health applications where mis-calibrated trust can affect diagnostic and treatment outcomes.

Method: Leverages Information Bottleneck principle connected with clustering of internal representations to analyze mutual information signatures that evolve across training phases. Proposes a novel dynamics-based stealthiness metric that quantifies attack integration at the model level.

Result: Finds that backdoor attacks create unique mutual information signatures that differ based on attack mechanism. Uncovers surprising trade-off: visually conspicuous attacks like BadNets can achieve high stealthiness from information-theoretic perspective, integrating more seamlessly than many visually imperceptible attacks.

Conclusion: Provides new insights into backdoor attack dynamics and offers a novel dynamics-based stealthiness metric for understanding and evaluating backdoor threats, validated across multiple datasets and diverse attack types.

Abstract: Understanding how backdoor data influences neural network training dynamics remains a complex and underexplored challenge. In this paper, we present a rigorous analysis of the impact of backdoor data on the learning process, with a particular focus on the distinct behaviors between the target class and other clean classes. Leveraging the Information Bottleneck (IB) principle connected with clustering of internal representation, We find that backdoor attacks create unique mutual information (MI) signatures, which evolve across training phases and differ based on the attack mechanism. Our analysis uncovers a surprising trade-off: visually conspicuous attacks like BadNets can achieve high stealthiness from an information-theoretic perspective, integrating more seamlessly into the model than many visually imperceptible attacks. Building on these insights, we propose a novel, dynamics-based stealthiness metric that quantifies an attack’s integration at the model level. We validate our findings and the proposed metric across multiple datasets and diverse attack types, offering a new dimension for understanding and evaluating backdoor threats. Our code is available in: https://github.com/XinyuLiu71/Information_Bottleneck_Backdoor.git.

Method: ProPS places a large language model at the center of the policy optimization loop, having the LLM directly propose policy updates based on both reward feedback and natural language inputs (goals, domain knowledge, strategy hints). The method shows LLMs can perform numerical optimization in-context.

Result: Evaluated across 15 Gymnasium tasks (classic control, Atari, MuJoCo) and compared to 7 RL algorithms (PPO, SAC, TRPO). ProPS outperformed all baselines on 8/15 tasks and showed substantial gains when provided with domain knowledge.

Conclusion: Unifying semantics and numerics enables more transparent, generalizable, and human-aligned RL. LLMs can effectively perform numerical optimization in-context, and semantic signals lead to more informed exploration and sample-efficient learning.

Abstract: Reinforcement Learning (RL) traditionally relies on scalar reward signals, limiting its ability to leverage the rich semantic knowledge often available in real-world tasks. In contrast, humans learn efficiently by combining numerical feedback with language, prior knowledge, and common sense. We introduce Prompted Policy Search (ProPS), a novel RL method that unifies numerical and linguistic reasoning within a single framework. Unlike prior work that augment existing RL components with language, ProPS places a large language model (LLM) at the center of the policy optimization loop-directly proposing policy updates based on both reward feedback and natural language input. We show that LLMs can perform numerical optimization in-context, and that incorporating semantic signals, such as goals, domain knowledge, and strategy hints can lead to more informed exploration and sample-efficient learning. ProPS is evaluated across fifteen Gymnasium tasks, spanning classic control, Atari games, and MuJoCo environments, and compared to seven widely-adopted RL algorithms (e.g., PPO, SAC, TRPO). It outperforms all baselines on eight out of fifteen tasks and demonstrates substantial gains when provided with domain knowledge. These results highlight the potential of unifying semantics and numerics for transparent, generalizable, and human-aligned RL.

Method: Dual-agent debate framework where one model provides structured causal inference while the other critically examines it for logical flaws. When disagreements arise, agents attempt to persuade each other, challenging logic and revising conclusions until convergence.

Result: On CLadder dataset (Pearl’s ladder of causation), debate improved DeepSeek-R1’s overall accuracy from 78.03% to 87.45% (counterfactual: 67.94% to 80.04%). Qwen3 improved from 84.16% to 89.41% (counterfactual: 71.53% to 80.35%). Strong models benefit from debate with weaker agents.

Conclusion: Reasoning models serve as effective building blocks for multi-agent systems in causal inference. Diverse perspectives through debate significantly improve causal problem-solving, especially for challenging counterfactual reasoning tasks.

Abstract: When people reason about cause and effect, they often consider many competing “what if” scenarios before deciding which explanation fits best. Analogously, advanced language models capable of causal inference can consider multiple interventions and counterfactuals to judge the validity of causal claims. Crucially, this type of reasoning is less like a single calculation and more like an internal dialogue between alternative hypotheses. In this paper, we make this dialogue explicit through a dual-agent debate framework where one model provides a structured causal inference, and the other critically examines this reasoning for logical flaws. When disagreements arise, agents attempt to persuade each other, challenging each other’s logic and revising their conclusions until they converge on a mutually agreed answer. To take advantage of this deliberative process, we specifically use reasoning language models, whose strengths in both causal inference and adversarial debate remain under-explored relative to standard large language models. We evaluate our approach on the CLadder dataset, a benchmark linking natural language questions to formally defined causal graphs across all three rungs of Pearl’s ladder of causation. With Qwen3 and DeepSeek-R1 as debater agents, we demonstrate that multi-agent debate improves DeepSeek-R1’s overall accuracy in causal inference from 78.03% to 87.45%, with the counterfactual category specifically improving from 67.94% to 80.04% accuracy. Similarly, Qwen3’s overall accuracy improves from 84.16% to 89.41%, and counterfactual questions from 71.53% to 80.35%, showing that strong models can still benefit greatly from debate with weaker agents. Our results highlight the potential of reasoning models as building blocks for multi-agent systems in causal inference, and demonstrate the importance of diverse perspectives in causal problem-solving.

Method: Uses Rubin’s Potential Outcomes Framework to quantify how strongly each feature separates outcome groups in binary classification. Creates data-derived feature rankings by estimating each feature’s effect on the outcome, then compares these against model-based explanations.

Result: Provides an interpretable, model-agnostic method to assess model-data alignment by establishing a baseline from the data itself and comparing it to model behavior.

Conclusion: The framework offers a simple, computationally efficient way to evaluate whether machine learning models “say what the data says” by comparing data-derived feature effects with model explanations.

Abstract: In this work, we propose a simple and computationally efficient framework to evaluate whether machine learning models align with the structure of the data they learn from; that is, whether \textit{the model says what the data says}. Unlike existing interpretability methods that focus exclusively on explaining model behavior, our approach establishes a baseline derived directly from the data itself. Drawing inspiration from Rubin’s Potential Outcomes Framework, we quantify how strongly each feature separates the two outcome groups in a binary classification task, moving beyond traditional descriptive statistics to estimate each feature’s effect on the outcome. By comparing these data-derived feature rankings against model-based explanations, we provide practitioners with an interpretable and model-agnostic method to assess model–data alignment.

Method: Quantum autoencoder framework compresses network traffic into discriminative latent representations, combined with quantum support vector classification for intrusion detection.

Result: Achieves improved accuracy on three datasets using ideal simulators and IBM Quantum hardware, demonstrating practical quantum advantage on current NISQ devices. Moderate depolarizing noise acts as implicit regularization, stabilizing training and enhancing generalization.

Conclusion: Establishes quantum machine learning as a viable, hardware-ready solution for real-world cybersecurity challenges.

Abstract: Escalating cyber threats and the high-dimensional complexity of IoT traffic have outpaced classical anomaly detection methods. While deep learning offers improvements, computational bottlenecks limit real-time deployment at scale. We present a quantum autoencoder (QAE) framework that compresses network traffic into discriminative latent representations and employs quantum support vector classification (QSVC) for intrusion detection. Evaluated on three datasets, our approach achieves improved accuracy on ideal simulators and on the IBM Quantum hardware demonstrating practical quantum advantage on current NISQ devices. Crucially, moderate depolarizing noise acts as implicit regularization, stabilizing training and enhancing generalization. This work establishes quantum machine learning as a viable, hardware-ready solution for real-world cybersecurity challenges.

Method: Two-agent framework: 1) LLM “event interpreter” processes heterogeneous signals (school schedules, weather summaries, reports) into scalar transmission-impact signal; 2) Neuro-symbolic core combines this with historical case counts for calibrated probabilistic forecasts.

Result: Competitive point forecasting accuracy on Hong Kong (2023-2024) and Lishui, China (2024) datasets, with robust 90% prediction intervals (coverage 0.85-1.00) and human-interpretable rationales.

Conclusion: Structurally integrating domain knowledge through LLMs can match state-of-the-art performance while yielding context-aware forecasts that align with public health workflows.

Abstract: Effective surveillance of hand, foot and mouth disease (HFMD) requires forecasts accounting for epidemiological patterns and contextual drivers like school calendars and weather. While classical models and recent foundation models (e.g., Chronos, TimesFM) incorporate covariates, they often lack the semantic reasoning to interpret the causal interplay between conflicting drivers. In this work, we propose a two-agent framework decoupling contextual interpretation from probabilistic forecasting. An LLM “event interpreter” processes heterogeneous signals-including school schedules, meteorological summaries, and reports-into a scalar transmission-impact signal. A neuro-symbolic core then combines this with historical case counts to produce calibrated probabilistic forecasts. We evaluate the framework on real-world HFMD datasets from Hong Kong (2023-2024) and Lishui, China (2024). Compared to traditional and foundation-model baselines, our approach achieves competitive point forecasting accuracy while providing robust 90% prediction intervals (coverage 0.85-1.00) and human-interpretable rationales. Our results suggest that structurally integrating domain knowledge through LLMs can match state-of-the-art performance while yielding context-aware forecasts that align with public health workflows. Code is available at https://github.com/jw-chae/forecast_MED .

Method: CDIM dynamically adjusts the number and size of projection steps to align residual measurement energy with its theoretical distribution under the forward diffusion process, enabling adaptive alignment that preserves measurement consistency while accelerating inference.

Result: Achieves 10-50x reduction in projection steps, exactly satisfies measurement constraints for noise-free problems with few steps, and demonstrates effectiveness across super-resolution, denoising, inpainting, deblurring, and 3D point cloud reprojection.

Conclusion: CDIM provides a fast and accurate approach for solving noisy linear inverse problems using diffusion models, substantially accelerating constrained inference while maintaining measurement consistency.

Abstract: We introduce Linearly Constrained Diffusion Implicit Models (CDIM), a fast and accurate approach to solving noisy linear inverse problems using diffusion models. Traditional diffusion-based inverse methods rely on numerous projection steps to enforce measurement consistency in addition to unconditional denoising steps. CDIM achieves a 10-50x reduction in projection steps by dynamically adjusting the number and size of projection steps to align a residual measurement energy with its theoretical distribution under the forward diffusion process. This adaptive alignment preserves measurement consistency while substantially accelerating constrained inference. For noise-free linear inverse problems, CDIM exactly satisfies the measurement constraints with few projection steps, even when existing methods fail. We demonstrate CDIM’s effectiveness across a range of applications, including super-resolution, denoising, inpainting, deblurring, and 3D point cloud reprojection. Code and an interactive demo can be found on our project website.

Method: Heterogeneous multi-agent RL framework with three agents grouped into two types for selecting features and operations. Uses shared critic mechanism for agent communication, multi-head attention-based feature agents to handle dynamic feature expansion, and state encoding technique to stabilize learning.

Result: Extensive experiments validate the model’s effectiveness, efficiency, robustness, and interpretability (though specific metrics not provided in abstract).

Conclusion: The proposed framework addresses limitations of prior RL approaches by enabling cooperative and scalable feature transformation through improved agent communication and handling of dynamic feature spaces.

Abstract: Feature transformation enhances downstream task performance by generating informative features through mathematical feature crossing. Despite the advancements in deep learning, feature transformation remains essential for structured data, where deep models often struggle to capture complex feature interactions. Prior literature on automated feature transformation has achieved success but often relies on heuristics or exhaustive searches, leading to inefficient and time-consuming processes. Recent works employ reinforcement learning (RL) to enhance traditional approaches through a more effective trial-and-error way. However, two limitations remain: 1) Dynamic feature expansion during the transformation process, which causes instability and increases the learning complexity for RL agents; 2) Insufficient cooperation and communication between agents, which results in suboptimal feature crossing operations and degraded model performance. To address them, we propose a novel heterogeneous multi-agent RL framework to enable cooperative and scalable feature transformation. The framework comprises three heterogeneous agents, grouped into two types, each designed to select essential features and operations for feature crossing. To enhance communication among these agents, we implement a shared critic mechanism that facilitates information exchange during feature transformation. To handle the dynamically expanding feature space, we tailor multi-head attention-based feature agents to select suitable features for feature crossing. Additionally, we introduce a state encoding technique during the optimization process to stabilize and enhance the learning dynamics of the RL agents, resulting in more robust and reliable transformation policies. Finally, we conduct extensive experiments to validate the effectiveness, efficiency, robustness, and interpretability of our model.

Method: Theoretical analysis using a stylized model of repeated pricing games where AI agents play equilibrium strategies while humans defect to no-regret strategies instead of adopting AI agents.

Result: Even a single human defection can destabilize collusion and drive down prices; multiple defections push prices even closer to competitive levels. The nature of collusion changes under defection-aware AI agents.

Conclusion: Algorithmic collusion in mixed human-AI ecosystems is fragile when humans defect from AI strategies, but can persist under certain conditions with defection-aware AI agents.

Abstract: AI agents are increasingly deployed in ecosystems where they repeatedly interact not only with each other but also with humans. In this work, we study these human-AI ecosystems from a theoretical perspective, focusing on the classical framework of repeated pricing games. In our stylized model, the AI agents play equilibrium strategies, and one or more humans manually perform the pricing task instead of adopting an AI agent, thereby defecting to a no-regret strategy. Motivated by how populations of AI agents can sustain supracompetitive prices, we investigate whether high prices persist under such defections. Our main finding is that even a single human defection can destabilize collusion and drive down prices, and multiple defections push prices even closer to competitive levels. We further show how the nature of collusion changes under defection-aware AI agents. Taken together, our results characterize when algorithmic collusion is fragile–and when it persists–in mixed ecosystems of AI agents and humans.

Method: Proposed deep learning architectures including CNNs for 63-bit m-sequence reconstruction/classification and Siamese networks for similarity-based decoding, compared against CCA baselines. Used EEG data from 13 healthy adults with single-target flicker stimulation, and implemented temporal data augmentation with small shifts.

Result: Deep models significantly outperformed traditional approaches. Distance-based decoding using Earth Mover’s Distance (EMD) and constrained EMD showed greater robustness to latency variations than Euclidean/Mahalanobis metrics. Multi-class Siamese network achieved best overall performance with 96.89% average accuracy.

Conclusion: Data-driven deep architectures demonstrate strong potential for reliable, single-trial C-VEP decoding in adaptive non-invasive BCI systems, with Siamese networks showing particular promise for handling EEG variability.

Abstract: Non-invasive Brain-Computer Interfaces (BCIs) based on Code-Modulated Visual Evoked Potentials (C-VEPs) require highly robust decoding methods to address temporal variability and session-dependent noise in EEG signals. This study proposes and evaluates several deep learning architectures, including convolutional neural networks (CNNs) for 63-bit m-sequence reconstruction and classification, and Siamese networks for similarity-based decoding, alongside canonical correlation analysis (CCA) baselines. EEG data were recorded from 13 healthy adults under single-target flicker stimulation. The proposed deep models significantly outperformed traditional approaches, with distance-based decoding using Earth Mover’s Distance (EMD) and constrained EMD showing greater robustness to latency variations than Euclidean and Mahalanobis metrics. Temporal data augmentation with small shifts further improved generalization across sessions. Among all models, the multi-class Siamese network achieved the best overall performance with an average accuracy of 96.89%, demonstrating the potential of data-driven deep architectures for reliable, single-trial C-VEP decoding in adaptive non-invasive BCI systems.

Method: Proposes a constrained optimization framework that unifies alignment and composition by enforcing reward constraints and proximity to pretrained models. Provides theoretical characterization of solutions and develops a Lagrangian-based primal-dual training algorithm.

Result: Empirical demonstration in image generation shows the aligned/composed model effectively satisfies constraints, addressing the limitations of existing methods.

Conclusion: The constrained optimization framework successfully addresses the gap in existing methods by ensuring diffusion models generate samples that satisfy multiple constraints while maintaining fidelity to pretrained models.

Abstract: Diffusion models have become prevalent in generative modeling due to their ability to sample from complex distributions. To improve the quality of generated samples and their compliance with user requirements, two commonly used methods are: (i) Alignment, which involves finetuning a diffusion model to align it with a reward; and (ii) Composition, which combines several pretrained diffusion models together, each emphasizing a desirable attribute in the generated outputs. However, trade-offs often arise when optimizing for multiple rewards or combining multiple models, as they can often represent competing properties. Existing methods cannot guarantee that the resulting model faithfully generates samples with all the desired properties. To address this gap, we propose a constrained optimization framework that unifies alignment and composition of diffusion models by enforcing that the aligned model satisfies reward constraints and/or remains close to each pretrained model. We provide a theoretical characterization of the solutions to the constrained alignment and composition problems and develop a Lagrangian-based primal-dual training algorithm to approximate these solutions. Empirically, we demonstrate our proposed approach in image generation, applying it to alignment and composition, and show that our aligned or composed model satisfies constraints effectively. Our implementation can be found at: \href{https://github.com/shervinkhalafi/constrained_comp_align}{https://github.com/shervinkhalafi/constrained_comp_align}

Method: 1. Generate neighborhood points near test instance using bounded Gaussian noise. 2. For each point, apply adversarial attacks to create adversarial pairs (A, A’) that bracket the decision boundary. 3. Train linear models on these adversarial pairs to approximate local decision boundaries.

Result: Experiments on six UCI benchmark datasets across three DNN architectures show ABLE achieves higher stability and fidelity than state-of-the-art methods.

Conclusion: ABLE provides more stable and faithful local explanations by leveraging adversarial bracketing to better capture local decision boundaries, addressing key limitations of existing approaches.

Abstract: Machine learning models are increasingly used in critical applications but are mostly “black boxes” due to their lack of transparency. Local explanation approaches, such as LIME, address this issue by approximating the behavior of complex models near a test instance using simple, interpretable models. However, these approaches often suffer from instability and poor local fidelity. In this paper, we propose a novel approach called Adversarially Bracketed Local Explanation (ABLE) to address these limitations. Our approach first generates a set of neighborhood points near the test instance, x_test, by adding bounded Gaussian noise. For each neighborhood point D, we apply an adversarial attack to generate an adversarial point A with minimal perturbation that results in a different label than D. A second adversarial attack is then performed on A to generate a point A’ that has the same label as D (and thus different than A). The points A and A’ form an adversarial pair that brackets the local decision boundary for x_test. We then train a linear model on these adversarial pairs to approximate the local decision boundary. Experimental results on six UCI benchmark datasets across three deep neural network architectures demonstrate that our approach achieves higher stability and fidelity than the state-of-the-art.

Method: 1. Start with supervised learning benchmarks (logistic regression, Light-GBM) using static features. 2. Extract and encode user action sequences from hundreds of millions of interactions over 22 days. 3. Use deep learning models (MLPs) to fuse behavioral embeddings with static features. 4. Design Transformer-based architecture with self-attention to capture temporal dynamics and contextual dependencies in behavioral sequences.

Result: Transformer model improves AUC by 2.81% over logistic regression baseline. Largest gains observed for users with diverse or changing interests. MLPs also achieved significant performance improvements over traditional models.

Conclusion: The research provides a roadmap for advancing CTR predictions and demonstrates the value of modeling temporal dynamics in user behavior. The approach can be extended beyond e-commerce to applications like personalized public health information delivery.

Abstract: Click-through rates prediction is critical in modern advertising systems, where ranking relevance and user engagement directly impact platform efficiency and business value. In this project, we explore how to model CTR more effectively using a large-scale Taobao dataset released by Alibaba. We start with supervised learning models, including logistic regression and Light-GBM, that are trained on static features such as user demographics, ad attributes, and contextual metadata. These models provide fast, interpretable benchmarks, but have limited capabilities to capture patterns of behavior that drive clicks. To better model user intent, we combined behavioral data from hundreds of millions of interactions over a 22-day period. By extracting and encoding user action sequences, we construct representations of user interests over time. We use deep learning models to fuse behavioral embeddings with static features. Among them, multilayer perceptrons (MLPs) have achieved significant performance improvements. To capture temporal dynamics, we designed a Transformer-based architecture that uses a self-attention mechanism to learn contextual dependencies across behavioral sequences, modeling not only what the user interacts with, but also the timing and frequency of interactions. Transformer improves AUC by 2.81 % over the baseline (LR model), with the largest gains observed for users whose interests are diverse or change over time. In addition to modeling, we propose an A/B testing strategy for real-world evaluation. We also think about the broader implications: personalized ad targeting technology can be applied to public health scenarios to achieve precise delivery of health information or behavior guidance. Our research provides a roadmap for advancing click-through rate predictions and extending their value beyond e-commerce.

Method: 1) Benchmark four ML architectures (MLP, CNN, UNet, GT) on transformer datasets; 2) Propose frequency-domain self-transfer learning exploiting correlations between adjacent frequency bands; 3) Develop inverse design framework using CMA-ES algorithm.

Result: Frequency-domain self-transfer learning achieved 30%-50% accuracy improvement in S-parameters prediction. The CMA-ES based inverse design framework demonstrated fast convergence and trustworthy performance in multiple impedance-matching tasks.

Conclusion: The study advances AI-assisted RFIC design automation and provides practical tools for designers, with the proposed methods showing significant accuracy improvements and reliable inverse design capabilities for transformer optimization.

Abstract: This paper presents a systematic study on developing multi-template machine learning (ML) surrogate models and applying them to the inverse design of transformers (XFMRs) in radio-frequency integrated circuits (RFICs). Our study starts with benchmarking four widely used ML architectures, including MLP-, CNN-, UNet-, and GT-based models, using the same datasets across different XFMR topologies. To improve modeling accuracy beyond these baselines, we then propose a new frequency-domain self-transfer learning technique that exploits correlations between adjacent frequency bands, leading to around 30%-50% accuracy improvement in the S-parameters prediction. Building on these models, we further develop an inverse design framework based on the covariance matrix adaptation evolutionary strategy (CMA-ES) algorithm. This framework is validated using multiple impedance-matching tasks, all demonstrating fast convergence and trustworthy performance. These results advance the goal of AI-assisted specs-to-GDS automation for RFICs and provide RFIC designers with actionable tools for integrating AI into their workflows.

Method: Develops a dynamic agentic safety/security framework with operational risk taxonomy, uses auxiliary AI models with human oversight for contextual risk management, implements AI-driven red teaming in sandboxed environments for risk discovery, and validates through case study with NVIDIA’s AI-Q Research Assistant.

Result: Framework effectively discovers novel agentic risks through red teaming and enables contextual mitigation. Case study demonstrates practical safety/security evaluations in complex enterprise workflows, with dataset of 10,000+ attack/defense execution traces released for research.

Conclusion: Agentic AI safety requires dynamic, contextual approaches that treat risks as emergent properties of system interactions. The proposed framework successfully addresses novel agentic risks through integrated safety/security management and AI-assisted risk discovery.

Abstract: This paper introduces a dynamic and actionable framework for securing agentic AI systems in enterprise deployment. We contend that safety and security are not merely fixed attributes of individual models but also emergent properties arising from the dynamic interactions among models, orchestrators, tools, and data within their operating environments. We propose a new way of identification of novel agentic risks through the lens of user safety. Although, for traditional LLMs and agentic models in isolation, safety and security has a clear separation, through the lens of safety in agentic systems, they appear to be connected. Building on this foundation, we define an operational agentic risk taxonomy that unifies traditional safety and security concerns with novel, uniquely agentic risks, including tool misuse, cascading action chains, and unintended control amplification among others. At the core of our approach is a dynamic agentic safety and security framework that operationalizes contextual agentic risk management by using auxiliary AI models and agents, with human oversight, to assist in contextual risk discovery, evaluation, and mitigation. We further address one of the most challenging aspects of safety and security of agentic systems: risk discovery through sandboxed, AI-driven red teaming. We demonstrate the framework effectiveness through a detailed case study of NVIDIA flagship agentic research assistant, AI-Q Research Assistant, showcasing practical, end-to-end safety and security evaluations in complex, enterprise-grade agentic workflows. This risk discovery phase finds novel agentic risks that are then contextually mitigated. We also release the dataset from our case study, containing traces of over 10,000 realistic attack and defense executions of the agentic workflow to help advance research in agentic safety.

Method: Proposes online algorithm for orderly hypertrees using SP-queries, transformable to optimal offline algorithm; also considers bounded distance query model for general hypertrees.

Result: First provably optimal online algorithm for orderly hypertrees, with optimal offline transformation; asymptotically tight complexity bounds for learning general hypertrees with bounded distance queries.

Conclusion: Orderly hypertrees represent the broadest learnable class in Fagin hierarchy with subquadratic SP-query complexity, with optimal algorithms developed for both orderly and general hypertrees under different query models.

Abstract: We study the problem of learning hypergraphs with shortest-path queries (SP-queries), and present the first provably optimal online algorithm for a broad and natural class of hypertrees that we call orderly hypertrees. Our online algorithm can be transformed into a provably optimal offline algorithm. Orderly hypertrees can be positioned within the Fagin hierarchy of acyclic hypergraph (well-studied in database theory), and strictly encompass the broadest class in this hierarchy that is learnable with subquadratic SP-query complexity. Recognizing that in some contexts, such as evolutionary tree reconstruction, distance measurements can degrade with increased distance, we also consider a learning model that uses bounded distance queries. In this model, we demonstrate asymptotically tight complexity bounds for learning general hypertrees.

Method: Model network states as Gibbs-Boltzmann distributions rather than deterministic points, prove gradient relationships using Helmholtz free energy differences, and derive a generalized EP algorithm based on path integrals of loss-energy covariances.

Result: Proved that the gradient of Helmholtz free energy difference equals the difference in expected local energy derivatives, validating Contrastive Hebbian Learning as an exact gradient estimator for arbitrary finite nudging.

Conclusion: The work establishes a rigorous finite-nudge foundation for Equilibrium Propagation, enabling more practical and powerful learning algorithms that can handle strong error signals without relying on restrictive approximations.

Abstract: We liberate Equilibrium Propagation (EP) from the limit of infinitesimal perturbations by establishing a finite-nudge foundation for local credit assignment. By modeling network states as Gibbs-Boltzmann distributions rather than deterministic points, we prove that the gradient of the difference in Helmholtz free energy between a nudged and free phase is exactly the difference in expected local energy derivatives. This validates the classic Contrastive Hebbian Learning update as an exact gradient estimator for arbitrary finite nudging, requiring neither infinitesimal approximations nor convexity. Furthermore, we derive a generalized EP algorithm based on the path integral of loss-energy covariances, enabling learning with strong error signals that standard infinitesimal approximations cannot support.

Method: Online test-time adaptation framework that updates batch normalization parameters while preserving pretrained statistics, uses memory bank with reliability-based sample selection (negative energy scores + persistence time), and incorporates prototype learning for robust predictions.

Result: Achieved average F1-score of 81.73% on sustained-attention driving dataset, outperforming all baselines with 11.73% improvement over best TTA baseline.

Conclusion: The proposed method significantly enhances adaptability of EEG-based drowsiness detection in non-i.i.d. scenarios by effectively addressing domain shift through dynamic test-time adaptation.

Abstract: Drowsy driving is a growing cause of traffic accidents, prompting recent exploration of electroencephalography (EEG)-based drowsiness detection systems. However, the inherent variability of EEG signals due to psychological and physical factors necessitates a cumbersome calibration process. In particular, the inter-subject variability of EEG signals leads to a domain shift problem, which makes it challenging to generalize drowsiness detection models to unseen target subjects. To address these issues, we propose a novel driver drowsiness detection framework that leverages online test-time adaptation (TTA) methods to dynamically adjust to target subject distributions. Our proposed method updates the learnable parameters in batch normalization (BN) layers, while preserving pretrained normalization statistics, resulting in a modified configuration that ensures effective adaptation during test time. We incorporate a memory bank that dynamically manages streaming EEG segments, selecting samples based on their reliability determined by negative energy scores and persistence time. In addition, we introduce prototype learning to ensure robust predictions against distribution shifts over time. We validated our method on the sustained-attention driving dataset collected in a simulated environment, where drowsiness was estimated from delayed reaction times during monotonous lane-keeping tasks. Our experiments show that our method outperforms all baselines, achieving an average F1-score of 81.73%, an improvement of 11.73% over the best TTA baseline. This demonstrates that our proposed method significantly enhances the adaptability of EEG-based drowsiness detection systems in non-i.i.d. scenarios.

Method: Three-component AI pipeline: 1) Fuel mix predictor estimating generation sources, 2) Air quality converter modeling pollutant emissions and dispersion, 3) Health impact assessor translating pollutant changes into monetized health damages. Applied in health-driven optimization framework across multiple US regions.

Result: HealthPredictor achieves substantially lower prediction errors for public health impacts compared to fuel mix-driven baselines. Case study on EV charging schedules demonstrates public health gains and provides actionable guidance for health-informed energy management.

Conclusion: AI models can be explicitly designed to enable health-informed energy management, advancing public health and societal well-being. The approach offers practical tools for demand-side management with direct health benefits.

Abstract: The electric power sector is a leading source of air pollutant emissions, impacting the public health of nearly every community. Although regulatory measures have reduced air pollutants, fossil fuels remain a significant component of the energy supply, highlighting the need for more advanced demand-side approaches to reduce the public health impacts. To enable health-informed demand-side management, we introduce HealthPredictor, a domain-specific AI model that provides an end-to-end pipeline linking electricity use to public health outcomes. The model comprises three components: a fuel mix predictor that estimates the contribution of different generation sources, an air quality converter that models pollutant emissions and atmospheric dispersion, and a health impact assessor that translates resulting pollutant changes into monetized health damages. Across multiple regions in the United States, our health-driven optimization framework yields substantially lower prediction errors in terms of public health impacts than fuel mix-driven baselines. A case study on electric vehicle charging schedules illustrates the public health gains enabled by our method and the actionable guidance it can offer for health-informed energy management. Overall, this work shows how AI models can be explicitly designed to enable health-informed energy management for advancing public health and broader societal well-being. Our datasets and code are released at: https://github.com/Ren-Research/Health-Impact-Predictor.

Method: Theoretical analysis of gradient descent for training over-parameterized student models with n parameters on data from a single ground-truth Gaussian using population score matching objective. Examines optimization dynamics across multiple regimes: high-noise, low-noise, and different initialization conditions (exponentially small, random Gaussian far from ground truth).

Result: 1) With sufficiently large noise scale: global convergence proven. 2) Low-noise regime: stationary point exists, making global convergence proofs difficult. 3) With exponentially small initialization: all parameters converge to ground truth. 4) Without such initialization: parameters may not converge to ground truth. 5) With random Gaussian initialization far from ground truth: only one parameter converges while others diverge, yet loss converges to zero with 1/τ rate, with nearly matching lower bound established.

Conclusion: This work provides the first global convergence guarantees for Gaussian mixtures with at least three components under score matching framework, revealing complex optimization dynamics where loss convergence doesn’t guarantee parameter convergence, highlighting the importance of initialization and noise regimes in score matching optimization.

Abstract: Score matching has become a central training objective in modern generative modeling, particularly in diffusion models, where it is used to learn high-dimensional data distributions through the estimation of score functions. Despite its empirical success, the theoretical understanding of the optimization behavior of score matching, particularly in over-parameterized regimes, remains limited. In this work, we study gradient descent for training over-parameterized models to learn a single Gaussian distribution. Specifically, we use a student model with $n$ learnable parameters and train it on data generated from a single ground-truth Gaussian using the population score matching objective. We analyze the optimization dynamics under multiple regimes. When the noise scale is sufficiently large, we prove a global convergence result for gradient descent. In the low-noise regime, we identify the existence of a stationary point, highlighting the difficulty of proving global convergence in this case. Nevertheless, we show convergence under certain initialization conditions: when the parameters are initialized to be exponentially small, gradient descent ensures convergence of all parameters to the ground truth. We further prove that without the exponentially small initialization, the parameters may not converge to the ground truth. Finally, we consider the case where parameters are randomly initialized from a Gaussian distribution far from the ground truth. We prove that, with high probability, only one parameter converges while the others diverge, yet the loss still converges to zero with a $1/τ$ rate, where $τ$ is the number of iterations. We also establish a nearly matching lower bound on the convergence rate in this regime. This is the first work to establish global convergence guarantees for Gaussian mixtures with at least three components under the score matching framework.

Method: HyperCast uses hypergraphs to model higher-order spatiotemporal dependencies, integrating multi-view hypergraphs (static geographical proximity and dynamic demand-based functional similarities) with multi-timescale inputs. It employs hyper-spatiotemporal blocks and cross-attention mechanisms to fuse information from different views and timescales.

Result: Extensive experiments on four public datasets show HyperCast significantly outperforms a wide array of state-of-the-art baselines, demonstrating the effectiveness of modeling collective charging behaviors.

Conclusion: Explicitly modeling higher-order spatiotemporal dependencies through hypergraphs enables more accurate EV charging demand forecasting, addressing limitations of existing pairwise relationship models and supporting better grid operation and market participation.

Abstract: Accurate electric vehicle (EV) charging demand forecasting is essential for stable grid operation and proactive EV participation in electricity market. Existing forecasting methods, particularly those based on graph neural networks, are often limited to modeling pairwise relationships between stations, failing to capture the complex, group-wise dynamics inherent in urban charging networks. To address this gap, we develop a novel forecasting framework namely HyperCast, leveraging the expressive power of hypergraphs to model the higher-order spatiotemporal dependencies hidden in EV charging patterns. HyperCast integrates multi-view hypergraphs, which capture both static geographical proximity and dynamic demand-based functional similarities, along with multi-timescale inputs to differentiate between recent trends and weekly periodicities. The framework employs specialized hyper-spatiotemporal blocks and tailored cross-attention mechanisms to effectively fuse information from these diverse sources: views and timescales. Extensive experiments on four public datasets demonstrate that HyperCast significantly outperforms a wide array of state-of-the-art baselines, demonstrating the effectiveness of explicitly modeling collective charging behaviors for more accurate forecasting.

Method: ARES combines Graph Neural Networks to capture spike and burst anomalous behaviors by embedding node/edge properties in latent space, with Half-Space Trees for efficient anomaly scoring. It also includes a supervised thresholding mechanism using statistical dispersion of scores with minimal labeled data.

Result: The framework was validated through extensive evaluations across several real-world cyber-attack scenarios, showing performance advantages over existing methods while analyzing space and time complexity.

Conclusion: ARES provides an effective unsupervised solution for real-time edge anomaly detection in temporal graphs, addressing challenges of concept drift and scalability while maintaining adaptability across domains.

Abstract: Many real-world scenarios involving streaming information can be represented as temporal graphs, where data flows through dynamic changes in edges over time. Anomaly detection in this context has the objective of identifying unusual temporal connections within the graph structure. Detecting edge anomalies in real time is crucial for mitigating potential risks. Unlike traditional anomaly detection, this task is particularly challenging due to concept drifts, large data volumes, and the need for real-time response. To face these challenges, we introduce ARES, an unsupervised anomaly detection framework for edge streams. ARES combines Graph Neural Networks (GNNs) for feature extraction with Half-Space Trees (HST) for anomaly scoring. GNNs capture both spike and burst anomalous behaviors within streams by embedding node and edge properties in a latent space, while HST partitions this space to isolate anomalies efficiently. ARES operates in an unsupervised way without the need for prior data labeling. To further validate its detection capabilities, we additionally incorporate a simple yet effective supervised thresholding mechanism. This approach leverages statistical dispersion among anomaly scores to determine the optimal threshold using a minimal set of labeled data, ensuring adaptability across different domains. We validate ARES through extensive evaluations across several real-world cyber-attack scenarios, comparing its performance against existing methods while analyzing its space and time complexity.

Method: FedWMSAM uses two key techniques: 1) momentum-guided global perturbation constructed from server-aggregated momentum to align clients’ SAM directions with global descent geometry, enabling efficient single-backprop SAM approximation; 2) adaptive coupling of momentum and SAM via cosine-similarity rule, creating an early-momentum, late-SAM two-phase training schedule.

Result: The method provides a non-IID convergence bound that explicitly models perturbation-induced variance and its dependence on system parameters. Extensive experiments on multiple datasets and model architectures validate effectiveness, adaptability, and robustness, demonstrating superiority in addressing FL optimization challenges.

Conclusion: FedWMSAM successfully addresses the twin challenges of fast convergence and good generalization in non-IID FL by resolving the failure modes of simply combining momentum and SAM, offering both theoretical guarantees and empirical validation.

Abstract: In federated learning (FL), models must \emph{converge quickly} under tight communication budgets while \emph{generalizing} across non-IID client distributions. These twin requirements have naturally led to two widely used techniques: client/server \emph{momentum} to accelerate progress, and \emph{sharpness-aware minimization} (SAM) to prefer flat solutions. However, simply combining momentum and SAM leaves two structural issues unresolved in non-IID FL. We identify and formalize two failure modes: \emph{local-global curvature misalignment} (local SAM directions need not reflect the global loss geometry) and \emph{momentum-echo oscillation} (late-stage instability caused by accumulated momentum). To our knowledge, these failure modes have not been jointly articulated and addressed in the FL literature. We propose \textbf{FedWMSAM} to address both failure modes. First, we construct a momentum-guided global perturbation from server-aggregated momentum to align clients’ SAM directions with the global descent geometry, enabling a \emph{single-backprop} SAM approximation that preserves efficiency. Second, we couple momentum and SAM via a cosine-similarity adaptive rule, yielding an early-momentum, late-SAM two-phase training schedule. We provide a non-IID convergence bound that \emph{explicitly models the perturbation-induced variance} $σ_ρ^2=σ^2+(Lρ)^2$ and its dependence on $(S, K, R, N)$ on the theory side. We conduct extensive experiments on multiple datasets and model architectures, and the results validate the effectiveness, adaptability, and robustness of our method, demonstrating its superiority in addressing the optimization challenges of Federated Learning. Our code is available at https://github.com/Huang-Yongzhi/NeurlPS_FedWMSAM.

Method: Proposes a Quantum Bayesian Optimization (QBO) framework that uses a quantum oracle based on FEA or surrogate models to accurately estimate environment responses with fewer queries. Employs Upper Confidence Bound (UCB) as acquisition function to strategically select input values that maximize the objective function.

Result: Experimental results show QBO achieves significantly lower dimensional error and uncertainty compared to classical methods, using the same number of queries from simulation. The approach theoretically requires much fewer samples while maintaining comparable optimization results.

Conclusion: QBO framework effectively addresses sample efficiency issues in aerospace fuselage assembly by leveraging quantum algorithm advantages for precise shape control, demonstrating superior performance over classical optimization methods in reducing dimensional gaps.

Abstract: Recent efforts in smart manufacturing have enhanced aerospace fuselage assembly processes, particularly by innovating shape adjustment techniques to minimize dimensional gaps between assembled sections. Existing approaches have shown promising results but face the issue of low sample efficiency from the manufacturing systems. It arises from the limitation of the classical Monte Carlo method when uncovering the mean response from a distribution. In contrast, recent work has shown that quantum algorithms can achieve the same level of estimation accuracy with significantly fewer samples than the classical Monte Carlo method from distributions. Therefore, we can adopt the estimation of the quantum algorithm to obtain the estimation from real physical systems (distributions). Motivated by this advantage, we propose a Quantum Bayesian Optimization (QBO) framework for precise shape control during assembly to improve the sample efficiency in manufacturing practice. Specifically, this approach utilizes a quantum oracle, based on finite element analysis (FEA)-based models or surrogate models, to acquire a more accurate estimation of the environment response with fewer queries for a certain input. QBO employs an Upper Confidence Bound (UCB) as the acquisition function to strategically select input values that are most likely to maximize the objective function. It has been theoretically proven to require much fewer samples while maintaining comparable optimization results. In the case study, force-controlled actuators are applied to one fuselage section to adjust its shape and reduce the gap to the adjoining section. Experimental results demonstrate that QBO achieves significantly lower dimensional error and uncertainty compared to classical methods, particularly using the same queries from the simulation.

Method: Comprehensive evaluation of multiple LLMs compressed with diverse low-rank algorithms, assessing trustworthiness across privacy, adversarial robustness, fairness, and ethical alignment. Includes analysis of model scale, fine-tuning effects, and gradient-based attribution to identify layers contributing to robustness.

Result: Low-rank compression: (1) preserves/improves training data privacy but weakens PII protection in conversations; (2) preserves/enhances adversarial robustness even under deep compression; (3) degrades ethical reasoning in zero-shot but recovers with few-shot prompting; (4) reduces fairness. Scale and fine-tuning significantly affect trustworthiness trade-offs.

Conclusion: Low-rank compression presents complex trustworthiness trade-offs: benefits for privacy and robustness but costs for fairness and ethics. The study provides guidance for trustworthy compression strategies through layer attribution analysis and highlights the need to consider trustworthiness alongside performance in compression decisions.

Abstract: Large language models (LLMs) have driven major advances across domains, yet their massive size hinders deployment in resource-constrained settings. Model compression addresses this challenge, with low-rank factorization emerging as a particularly effective method for reducing size, memory, and computation while maintaining accuracy. However, while these compressed models boast of benign performance and system-level advantages, their trustworthiness implications remain poorly understood. In this paper, we present the first comprehensive study of how low-rank factorization affects LLM trustworthiness across privacy, adversarial robustness, fairness, and ethical alignment. We evaluate multiple LLMs of different sizes and variants compressed with diverse low-rank algorithms, revealing key insights: (1) low-rank compression preserves or improves training data privacy but weakens PII protection during conversation; (2) adversarial robustness is generally preserved and often enhanced, even under deep compression; (3) ethical reasoning degrades in zero-shot settings but partially recovers with few-shot prompting; (4) fairness declines under compression. Beyond compression, we investigate how model scale and fine-tuning affect trustworthiness, as both are important in low-rank methods. To guide trustworthy compression strategies, we end our paper with a gradient-based attribution analysis to identify which layers in LLMs contribute most to adversarial robustness.

Method: 1) Replication: Data collection from Uniswap Subgraph, implementation details, result analysis. 2) Improvement: New architecture combining Mamba with DDQN, new reward function, data cleaning, and introduction of two new baselines for comparison.

Result: The improved model shows stronger theoretical support than original and performs better in some tests, though not yet applied to all datasets.

Conclusion: The proposed Mamba-DDQN hybrid with new reward function represents a promising improvement over the original approach for adaptive liquidity provision in Uniswap V3, though further testing on complete datasets is needed.

Abstract: The report goes through the main steps of replicating and improving the article “Adaptive Liquidity Provision in Uniswap V3 with Deep Reinforcement Learning.” The replication part includes how to obtain data from the Uniswap Subgraph, details of the implementation, and comments on the results. After the replication, I propose a new structure based on the original model, which combines Mamba with DDQN and a new reward function. In this new structure, I clean the data again and introduce two new baselines for comparison. As a result, although the model has not yet been applied to all datasets, it shows stronger theoretical support than the original model and performs better in some tests.

Method: Extends prior meta-bandits algorithm to Markov Decision Processes (MDPs). Uses a relaxed, non-centered sub-Gaussian process model to accommodate environmental heterogeneity while maintaining theoretical guarantees.

Result: Proves that the existing algorithm achieves performance comparable to using the full action space. Theoretical guarantees are established under the relaxed model, providing computational and sample efficiency.

Conclusion: The approach provides a computationally and sample-efficient solution for large-scale combinatorial decision-making under uncertainty, enabling efficient learning without exhaustive action evaluation.

Abstract: We study the problem of selecting a small, representative action subset from an extremely large action space shared across a family of reinforcement learning (RL) environments – a fundamental challenge in applications like inventory management and recommendation systems, where direct learning over the entire space is intractable. Our goal is to identify a fixed subset of actions that, for every environment in the family, contains a near-optimal action, thereby enabling efficient learning without exhaustively evaluating all actions. This work extends our prior results for meta-bandits to the more general setting of Markov Decision Processes (MDPs). We prove that our existing algorithm achieves performance comparable to using the full action space. This theoretical guarantee is established under a relaxed, non-centered sub-Gaussian process model, which accommodates greater environmental heterogeneity. Consequently, our approach provides a computationally and sample-efficient solution for large-scale combinatorial decision-making under uncertainty.

Method: Proposes a multi-agent deep reinforcement learning framework using Double Deep Q-Network (MARL-DDQN) for adaptive sleep mode optimization in 3D urban environments. The approach integrates realistic BS power consumption models and beamforming, enables distributed decision-making with minimal signaling overhead, and adapts SMO policies to maximize EE while mitigating inter-cell interference and ensuring throughput fairness.

Result: MARL-DDQN outperforms state-of-the-art strategies including All On, iterative QoS-aware load-based (IT-QoS-LB), MARL-DDPG, and MARL-PPO, achieving up to 0.60 Mbit/Joule energy efficiency, 8.5 Mbps 10th-percentile throughput, and meeting QoS constraints 95% of the time under dynamic scenarios.

Conclusion: The MARL-DDQN framework provides an effective solution for dynamic sleep mode optimization in mmWave networks, offering superior energy efficiency while maintaining QoS constraints through scalable, distributed multi-agent reinforcement learning that adapts to time-varying traffic patterns.

Abstract: Dynamic sleep mode optimization (SMO) in millimeter-wave (mmWave) networks is essential for maximizing energy efficiency (EE) under stringent quality-of-service (QoS) constraints. However, existing optimization and reinforcement learning (RL) approaches rely on aggregated, static base station (BS) traffic models that fail to capture non-stationary traffic dynamics and suffer from large state-action spaces, limiting real-world deployment. To address these challenges, this paper proposes a multi-agent deep reinforcement learning (MARL) framework using a Double Deep Q-Network (DDQN), referred to as MARL-DDQN, for adaptive SMO in a 3D urban environment with a time-varying and community-based user equipment (UE) mobility model. Unlike conventional single-agent RL, MARL-DDQN enables scalable, distributed decision-making with minimal signaling overhead. A realistic BS power consumption model and beamforming are integrated to accurately quantify EE, while QoS is defined in terms of throughput. The method adapts SMO policies to maximize EE while mitigating inter-cell interference and ensuring throughput fairness. Simulations show that MARL-DDQN outperforms state-of-the-art strategies, including All On, iterative QoS-aware load-based (IT-QoS-LB), MARL-DDPG, and MARL-PPO, achieving up to 0.60 Mbit/Joule EE, 8.5 Mbps 10th-percentile throughput, and meeting QoS constraints 95% of the time under dynamic scenarios.

Method: Adapted Reinforcement Learning algorithms (Banditron and AGREL) for Deep Spiking Neural Networks, chosen for their ability to train with limited computational resources. Conducted both open-loop and closed-loop experiments to evaluate effectiveness.

Result: Open-loop experiments showed stable accuracy over extended periods. In closed-loop experiments with perturbations, DSNN Banditron performed comparably to DSNN AGREL while achieving 98% reduction in memory access usage and 99% reduction in MAC operations during training. DSNN Banditron requires 98% less computation than previous continuous learning SNN decoders.

Conclusion: DSNN Banditron is a prime candidate for future wireless iBMI systems due to its stable performance, ability to address non-stationarity, and dramatically reduced computational requirements that fit implantable device constraints.

Abstract: The number of simultaneously recorded neurons follows an exponentially increasing trend in implantable brain-machine interfaces (iBMIs). Integrating the neural decoder in the implant is an effective data compression method for future wireless iBMIs. However, the non-stationarity of the system makes the performance of the decoder unreliable. To avoid frequent retraining of the decoder and to ensure the safety and comfort of the iBMI user, continuous learning is essential for real-life applications. Since Deep Spiking Neural Networks (DSNNs) are being recognized as a promising approach for developing resource-efficient neural decoder, we propose continuous learning approaches with Reinforcement Learning (RL) algorithms adapted for DSNNs. Banditron and AGREL are chosen as the two candidate RL algorithms since they can be trained with limited computational resources, effectively addressing the non-stationary problem and fitting the energy constraints of implantable devices. To assess the effectiveness of the proposed methods, we conducted both open-loop and closed-loop experiments. The accuracy of open-loop experiments conducted with DSNN Banditron and DSNN AGREL remains stable over extended periods. Meanwhile, the time-to-target in the closed-loop experiment with perturbations, DSNN Banditron performed comparably to that of DSNN AGREL while achieving reductions of 98% in memory access usage and 99% in the requirements for multiply- and-accumulate (MAC) operations during training. Compared to previous continuous learning SNN decoders, DSNN Banditron requires 98% less computes making it a prime candidate for future wireless iBMI systems.

Method: Developed a surrogate model using Spherical Fourier Neural Operator (SFNO) for steady-state solar wind modeling. Compared performance against existing numerical surrogate HUX model across multiple metrics.

Result: SFNO achieves comparable or better performance than HUX across several evaluation metrics. While HUX retains advantages in physical smoothness, SFNO demonstrates competitive performance, highlighting need for improved evaluation criteria rather than model flaws.

Conclusion: SFNO provides a flexible, trainable approach for efficient real-time solar wind forecasting that can improve with more data. The model enables better investigation of boundary condition uncertainty compared to computationally expensive 3D MHD models.

Abstract: The solar wind, a continuous stream of charged particles from the Sun’s corona, shapes the heliosphere and impacts space systems near Earth. Variations such as high-speed streams and coronal mass ejections can disrupt satellites, power grids, and communications, making accurate modeling essential for space weather forecasting. While 3D magnetohydrodynamic (MHD) models are used to simulate and investigate these variations in the solar wind, they tend to be computationally expensive, limiting their usefulness in investigating the impacts of boundary condition uncertainty. In this work, we develop a surrogate for steady state solar wind modeling, using a Spherical Fourier Neural Operator (SFNO). We compare our model to a previously developed numerical surrogate for this task called HUX, and we show that the SFNO achieves comparable or better performance across several metrics. Though HUX retains advantages in physical smoothness, this underscores the need for improved evaluation criteria rather than a flaw in SFNO. As a flexible and trainable approach, SFNO enables efficient real-time forecasting and can improve with more data. The source code and more visual results are available at https://github.com/rezmansouri/solarwind-sfno-velocity.

Method: IVGAE constructs a bipartite graph representing sample-feature relationships, uses graph representation learning with a dual-decoder architecture (one for feature embedding reconstruction, one for missingness pattern modeling), and employs a Transformer-based heterogeneous embedding module for categorical variables to avoid high-dimensional one-hot encoding.

Result: Extensive experiments on 16 real-world datasets show consistent improvements in RMSE and downstream F1 scores across MCAR, MAR, and MNAR missing scenarios under 30% missing rates.

Conclusion: IVGAE provides a robust framework for imputing incomplete heterogeneous data by effectively capturing structural dependencies and handling both numerical and categorical features through graph representation learning and innovative architectural design.

Abstract: Handling missing data remains a fundamental challenge in real-world tabular datasets, especially when data are heterogeneous with both numerical and categorical features. Existing imputation methods often fail to capture complex structural dependencies and handle heterogeneous data effectively. We present \textbf{IVGAE}, a Variational Graph Autoencoder framework for robust imputation of incomplete heterogeneous data. IVGAE constructs a bipartite graph to represent sample-feature relationships and applies graph representation learning to model structural dependencies. A key innovation is its \textit{dual-decoder architecture}, where one decoder reconstructs feature embeddings and the other models missingness patterns, providing structural priors aware of missing mechanisms. To better encode categorical variables, we introduce a Transformer-based heterogeneous embedding module that avoids high-dimensional one-hot encoding. Extensive experiments on 16 real-world datasets show that IVGAE achieves consistent improvements in RMSE and downstream F1 across MCAR, MAR, and MNAR missing scenarios under 30% missing rates. Code and data are available at: https://github.com/echoid/IVGAE.

Method: A variational framework that casts dimensionality reduction as an optimal manifold embedding problem. The framework permits nonlinear embeddings and yields solutions that satisfy partial differential equations and reflect symmetries of the embedding objective.

Result: The framework allows analytical characterization of solutions in some cases and can recover PCA as a special case. It provides more flexible embeddings than PCA while maintaining interpretability through PDEs and symmetry properties.

Conclusion: The proposed variational framework offers a principled approach to dimensionality reduction that combines the flexibility of nonlinear embeddings with the interpretability of traditional methods like PCA, addressing key limitations of existing approaches.

Abstract: Dimensionality reduction algorithms like principal component analysis (PCA) are workhorses of machine learning and neuroscience, but each has well-known limitations. Variants of PCA are simple and interpretable, but not flexible enough to capture nonlinear data manifold structure. More flexible approaches have other problems: autoencoders are generally difficult to interpret, and graph-embedding-based methods can produce pathological distortions in manifold geometry. Motivated by these shortcomings, we propose a variational framework that casts dimensionality reduction algorithms as solutions to an optimal manifold embedding problem. By construction, this framework permits nonlinear embeddings, allowing its solutions to be more flexible than PCA. Moreover, the variational nature of the framework has useful consequences for interpretability: each solution satisfies a set of partial differential equations, and can be shown to reflect symmetries of the embedding objective. We discuss these features in detail and show that solutions can be analytically characterized in some cases. Interestingly, one special case exactly recovers PCA.

Method: Created BELA benchmark combining: (1) real Amazon products, (2) diverse user personas representing latent preferences, and (3) LLM user simulator powered by personas to generate interactive trajectories for testing experiential learning and active exploration.

Result: Current frontier models struggle to meaningfully improve across episodes in the BELA benchmark, demonstrating poor in-context experiential learning capabilities despite the rich interactive environment.

Conclusion: There is a critical need for agentic systems with stronger in-context learning capabilities to handle shifting real-world environments where agents must learn and adapt from experience, as current models fall short in experiential learning tasks.

Abstract: To reliably navigate ever-shifting real-world environments, agents must grapple with incomplete knowledge and adapt their behavior through experience. However, current evaluations largely focus on tasks that leave no ambiguity, and do not measure agents’ ability to adaptively learn and reason through the experiences they accrued. We exemplify the need for this in-context experiential learning in a product recommendation context, where agents must navigate shifting customer preferences and product landscapes through natural language dialogue. We curate a benchmark for experiential learning and active exploration (BELA) that combines (1) rich real-world products from Amazon, (2) a diverse collection of user personas to represent heterogeneous yet latent preferences, and (3) a LLM user simulator powered by the persona to create rich interactive trajectories. We observe that current frontier models struggle to meaningfully improve across episodes, underscoring the need for agentic systems with strong in-context learning capabilities.

Method: Two-phase approach: 1) Diagnosis phase uses GPLFM to treat model-form errors as latent forces and jointly estimate them with system states from sensor data. 2) BNN learns probabilistic nonlinear mapping from states to model-form errors. 3) Prognosis phase uses this mapping to generate pseudo-measurements for state prediction via Kalman filtering.

Result: Demonstrated on four nonlinear examples (SDOF oscillator, multi-DOF system, Bouc-Wen hysteretic system, Silverbox experimental dataset), showing predictive accuracy and robustness to model misspecification.

Conclusion: The framework enables systematic uncertainty propagation from diagnosis to prediction, providing a key capability for trustworthy digital twins in systems with imperfect physics models.

Abstract: This work presents a probabilistic digital twin framework for response prediction in dynamical systems governed by misspecified physics. The approach integrates Gaussian Process Latent Force Models (GPLFM) and Bayesian Neural Networks (BNNs) to enable end-to-end uncertainty-aware inference and prediction. In the diagnosis phase, model-form errors (MFEs) are treated as latent input forces to a nominal linear dynamical system and jointly estimated with system states using GPLFM from sensor measurements. A BNN is then trained on posterior samples to learn a probabilistic nonlinear mapping from system states to MFEs, while capturing diagnostic uncertainty. For prognosis, this mapping is used to generate pseudo-measurements, enabling state prediction via Kalman filtering. The framework allows for systematic propagation of uncertainty from diagnosis to prediction, a key capability for trustworthy digital twins. The framework is demonstrated using four nonlinear examples: a single degree of freedom (DOF) oscillator, a multi-DOF system, and two established benchmarks – the Bouc-Wen hysteretic system and the Silverbox experimental dataset – highlighting its predictive accuracy and robustness to model misspecification.

Method: Evaluated SLMs using Berkeley Function Calling Leaderboard (BFCL) framework with parameter-driven optimization strategies including supervised fine-tuning (SFT), parameter-efficient fine-tuning (PEFT), reinforcement learning (RL), Direct Preference Optimization (DPO), and hybrid methods. Used models like TinyAgent, TinyLlama, Qwen, and xLAM across BFCL categories and multi-turn evaluations.

Result: Medium-sized models (1-3B parameters) significantly outperform ultra-compact models (<1B parameters), achieving up to 65.74% overall accuracy and 55.62% multi-turn accuracy with hybrid optimization. Clear accuracy differences across model scales were demonstrated.

Conclusion: Hybrid optimization strategies enable small language models to deliver accurate, efficient, and stable agentic AI on edge devices, making privacy-preserving, low-latency autonomous agents practical beyond cloud infrastructure.

Abstract: This paper investigates the effectiveness of small language models (SLMs) for agentic tasks (function/tool/API calling) with a focus on running agents on edge devices without reliance on cloud infrastructure. We evaluate SLMs using the Berkeley Function Calling Leaderboard (BFCL) framework and describe parameter-driven optimization strategies that include supervised fine-tuning (SFT), parameter-efficient fine-tuning (PEFT), reinforcement learning (RL)-based optimization, preference alignment via Direct Preference Optimization (DPO), and hybrid methods. We report results for models including TinyAgent, TinyLlama, Qwen, and xLAM across BFCL categories (simple, multiple, parallel, parallel-multiple, and relevance detection), both in live and non-live settings, and in multi-turn evaluations. We additionally detail a DPO training pipeline constructed from AgentBank data (e.g., ALFRED), including our conversion of SFT data to chosen-rejected pairs using TinyLlama responses as rejected outputs and manual validation. Our results demonstrate clear accuracy differences across model scales where medium-sized models (1-3B parameters) significantly outperform ultra-compact models (<1B parameters), achieving up to 65.74% overall accuracy, and 55.62% multi-turn accuracy with hybrid optimization. This study highlights the importance of hybrid optimization strategies that enable small language models to deliver accurate, efficient, and stable agentic AI on edge devices, making privacy-preserving, low-latency autonomous agents practical beyond the cloud.

Method: Conducted comprehensive analysis of topological and geometric measures across diverse text embedding models and datasets, identified redundancy among measures, and introduced Unified Topological Signatures (UTS) - a holistic framework for characterizing embedding spaces.

Result: Found high redundancy among individual metrics, showed UTS can predict model-specific properties and reveal similarities driven by model architecture, demonstrated UTS links topological structure to ranking effectiveness and accurately predicts document retrievability.

Conclusion: A holistic, multi-attribute perspective is essential to understanding and leveraging the geometry of text embeddings, and UTS provides an effective framework for this comprehensive analysis.

Abstract: Studying how embeddings are organized in space not only enhances model interpretability but also uncovers factors that drive downstream task performance. In this paper, we present a comprehensive analysis of topological and geometric measures across a wide set of text embedding models and datasets. We find a high degree of redundancy among these measures and observe that individual metrics often fail to sufficiently differentiate embedding spaces. Building on these insights, we introduce Unified Topological Signatures (UTS), a holistic framework for characterizing embedding spaces. We show that UTS can predict model-specific properties and reveal similarities driven by model architecture. Further, we demonstrate the utility of our method by linking topological structure to ranking effectiveness and accurately predicting document retrievability. We find that a holistic, multi-attribute perspective is essential to understanding and leveraging the geometry of text embeddings.

Method: The method learns prompt- and noise-conditioned sampling schedules through a single-pass Dirichlet policy. It introduces a novel reward baseline using a principled James-Stein estimator to ensure accurate gradient estimates in high-dimensional policy learning, which provably achieves lower estimation errors than commonly used variants.

Result: The rescheduled samplers consistently improve text-image alignment including text rendering and compositional control across modern Stable Diffusion and Flux model families. A 5-step Flux-Dev sampler with the learned schedules achieves generation quality comparable to deliberately distilled samplers like Flux-Schnell.

Conclusion: The scheduling framework represents an emerging model-agnostic post-training lever that can unlock additional generative potential in pretrained samplers, offering an alternative to weight-based post-training methods.

Abstract: Most post-training methods for text-to-image samplers focus on model weights: either fine-tuning the backbone for alignment or distilling it for few-step efficiency. We take a different route: rescheduling the sampling timeline of a frozen sampler. Instead of a fixed, global schedule, we learn instance-level (prompt- and noise-conditioned) schedules through a single-pass Dirichlet policy. To ensure accurate gradient estimates in high-dimensional policy learning, we introduce a novel reward baseline based on a principled James-Stein estimator; it provably achieves lower estimation errors than commonly used variants and leads to superior performance. Our rescheduled samplers consistently improve text-image alignment including text rendering and compositional control across modern Stable Diffusion and Flux model families. Additionally, a 5-step Flux-Dev sampler with our schedules can attain generation quality comparable to deliberately distilled samplers like Flux-Schnell. We thus position our scheduling framework as an emerging model-agnostic post-training lever that unlocks additional generative potential in pretrained samplers.

Method: PULSE-ICU uses a unified embedding module to encode event identity, continuous values, units, and temporal attributes from raw EHR sequences. It employs a Longformer-based encoder for efficient modeling of long patient trajectories in a self-supervised manner without resampling or manual feature engineering.

Result: The model was fine-tuned across 18 prediction tasks including mortality, intervention forecasting, and phenotype identification, achieving strong performance. External validation on eICU, HiRID, and P12 datasets showed substantial improvements with minimal fine-tuning, demonstrating robustness to domain shift and variable constraints.

Conclusion: Foundation-style modeling can improve data efficiency and adaptability for ICU decision support, providing a scalable framework that works across diverse clinical environments without extensive retraining or feature engineering.

Abstract: Intensive care unit (ICU) data are highly irregular, heterogeneous, and temporally fragmented, posing challenges for generalizable clinical prediction. We present PULSE-ICU, a self-supervised foundation model that learns event-level ICU representations from large-scale EHR sequences without resampling or manual feature engineering. A unified embedding module encodes event identity, continuous values, units, and temporal attributes, while a Longformer-based encoder enables efficient modeling of long trajectories. PULSE-ICU was fine-tuned across 18 prediction tasks, including mortality, intervention forecasting, and phenotype identification, achieving strong performance across task types. External validation on eICU, HiRID, and P12 showed substantial improvements with minimal fine-tuning, demonstrating robustness to domain shift and variable constraints. These findings suggest that foundation-style modeling can improve data efficiency and adaptability, providing a scalable framework for ICU decision support across diverse clinical environments.

Method: Bi-level framework alternating between: (1) learning conservative Q-function via CQL under current reward, and (2) updating reward parameters to maximize expert Q-values while suppressing over-generalization. Based on maximum likelihood estimation under soft value matching.

Result: Theoretical guarantees of convergence to reward functions where expert policy is soft-optimal. Empirical improvements in both reward recovery and downstream policy performance on standard offline RL benchmarks compared to baselines.

Conclusion: BiCQL-ML effectively addresses offline IRL challenges by avoiding explicit policy learning and controlling over-generalization through conservative Q-learning and bi-level optimization.

Abstract: Offline inverse reinforcement learning (IRL) aims to recover a reward function that explains expert behavior using only fixed demonstration data, without any additional online interaction. We propose BiCQL-ML, a policy-free offline IRL algorithm that jointly optimizes a reward function and a conservative Q-function in a bi-level framework, thereby avoiding explicit policy learning. The method alternates between (i) learning a conservative Q-function via Conservative Q-Learning (CQL) under the current reward, and (ii) updating the reward parameters to maximize the expected Q-values of expert actions while suppressing over-generalization to out-of-distribution actions. This procedure can be viewed as maximum likelihood estimation under a soft value matching principle. We provide theoretical guarantees that BiCQL-ML converges to a reward function under which the expert policy is soft-optimal. Empirically, we show on standard offline RL benchmarks that BiCQL-ML improves both reward recovery and downstream policy performance compared to existing offline IRL baselines.

Method: FedRE uses entangled representations where clients aggregate local representations using normalized random weights, apply same weights to one-hot labels to create entangled-label encodings, upload these to server to train global classifier. Random weights are resampled each round for diversity.

Result: Extensive experiments show FedRE achieves effective trade-off among model performance, privacy protection, and communication overhead. The framework mitigates global classifier overconfidence, promotes smoother decision boundaries, and reduces representation inversion attack risks.

Conclusion: FedRE provides a practical solution for model-heterogeneous federated learning that balances performance, privacy, and communication efficiency through its novel entangled representation approach.

Abstract: Federated learning (FL) enables collaborative training across clients without compromising privacy. While most existing FL methods assume homogeneous model architectures, client heterogeneity in data and resources renders this assumption impractical, motivating model-heterogeneous FL. To address this problem, we propose Federated Representation Entanglement (FedRE), a framework built upon a novel form of client knowledge termed entangled representation. In FedRE, each client aggregates its local representations into a single entangled representation using normalized random weights and applies the same weights to integrate the corresponding one-hot label encodings into the entangled-label encoding. Those are then uploaded to the server to train a global classifier. During training, each entangled representation is supervised across categories via its entangled-label encoding, while random weights are resampled each round to introduce diversity, mitigating the global classifier’s overconfidence and promoting smoother decision boundaries. Furthermore, each client uploads a single cross-category entangled representation along with its entangled-label encoding, mitigating the risk of representation inversion attacks and reducing communication overhead. Extensive experiments demonstrate that FedRE achieves an effective trade-off among model performance, privacy protection, and communication overhead. The codes are available at https://github.com/AIResearch-Group/FedRE.

Method: TreeCoder represents decoding as a tree search over candidate programs, treating both decoding strategies and constraint functions (style, syntax, execution) as first-class, optimizable components that can be systematically explored and tuned using standard optimization techniques.

Result: Experiments on MBPP (Python) and SQL-Spider benchmarks show TreeCoder consistently improves accuracy across open-source models like CodeLlama, Mistral and DeepSeek, often outperforming unconstrained baselines by considerable margins.

Conclusion: TreeCoder provides a general and flexible framework for exploring decoding strategies and constraints in LLM code generation, enabling systematic optimization and significant accuracy improvements over traditional unconstrained decoding approaches.

Abstract: Large language models (LLMs) have shown remarkable ability to generate code, yet their outputs often violate syntactic or semantic constraints when guided only through natural language prompts. We introduce TreeCoder, the most general and flexible framework to date for exploring decoding strategies, constraints, and hyperparameters in LLMs, and use it in code generation to enforce correctness and structure during decoding rather than relying on prompt engineering. TreeCoder represents decoding as a tree search over candidate programs, where both decoding strategies and constraint functions - such as style, syntax, execution - are treated as first-class, optimisable components. This design enables systematic exploration and automatic tuning of decoding configurations using standard optimisation techniques. Experiments on the MBPP (Python) and SQL-Spider benchmarks show that TreeCoder consistently improves accuracy across open-source models such as CodeLlama, Mistral and DeepSeek, often outperforming their unconstrained baselines by considerable margins.

Method: Systematic study of inexact OMD, analyzing relationship between regularizer smoothness and approximation error robustness. Examines uniformly smooth regularizers, barrier regularizers over simplex/subsets, and stochastic loss settings.

Result: For uniformly smooth regularizers: tight bound on excess regret from errors. For barrier regularizers: negative entropy requires exponentially small errors to avoid linear regret, while log-barrier/Tsallis remain robust with polynomial errors. Stochastic losses on simplex restore negative entropy robustness, but not for all subsets.

Conclusion: Regularizer choice critically impacts OMD robustness to approximation errors. Negative entropy is fragile to errors except in special cases, while log-barrier/Tsallis offer better error tolerance. Practical OMD implementations should consider regularizer robustness properties.

Abstract: Online mirror descent (OMD) is a fundamental algorithmic paradigm that underlies many algorithms in optimization, machine learning and sequential decision-making. The OMD iterates are defined as solutions to optimization subproblems which, oftentimes, can be solved only approximately, leading to an inexact version of the algorithm. Nonetheless, existing OMD analyses typically assume an idealized error free setting, thereby limiting our understanding of performance guarantees that should be expected in practice. In this work we initiate a systematic study into inexact OMD, and uncover an intricate relation between regularizer smoothness and robustness to approximation errors. When the regularizer is uniformly smooth, we establish a tight bound on the excess regret due to errors. Then, for barrier regularizers over the simplex and its subsets, we identify a sharp separation: negative entropy requires exponentially small errors to avoid linear regret, whereas log-barrier and Tsallis regularizers remain robust even when the errors are only polynomial. Finally, we show that when the losses are stochastic and the domain is the simplex, negative entropy regains robustness-but this property does not extend to all subsets, where exponentially small errors are again necessary to avoid suboptimal regret.

Method: Online learning algorithm that identifies complementary relationships via integer programming on transaction data, then optimizes pricing using heteroscedastic Gaussian process multi-armed bandit solutions.

Result: The algorithm improves revenue compared to comparable learning algorithms that ignore product interactions, validated in simulated environments.

Conclusion: Explicitly modeling complementary product relationships through data-driven learning algorithms enables more effective dynamic pricing strategies that capture interdependencies and maximize overall revenue.

Abstract: Traditional pricing paradigms, once dominated by static models and rule-based heuristics, are increasingly being replaced by dynamic, data-driven approaches powered by machine learning algorithms. Despite their growing sophistication, most dynamic pricing algorithms focus on optimizing the price of each product independently, disregarding potential interactions among items. By neglecting these interdependencies in consumer demand across related goods, sellers may fail to capture the full potential of coordinated pricing strategies. In this paper, we address this problem by exploring dynamic pricing mechanisms designed explicitly for complementary products, aiming to exploit their joint demand structure to maximize overall revenue. We present an online learning algorithm considering both positive and negative interactions between products’ demands. The algorithm utilizes transaction data to identify advantageous complementary relationships through an integer programming problem between different items, and then optimizes pricing strategies using data-driven and computationally efficient multi-armed bandit solutions based on heteroscedastic Gaussian processes. We validate our solution in a simulated environment, and we demonstrate that our solution improves the revenue w.r.t. a comparable learning algorithm ignoring such interactions.

Method: Leverage Neural ODEs and Universal Differential Equations (SciML pillars) to construct adaptive tumor growth models. Replace rigid terms in Gompertz model with adaptive neural networks to capture hidden dynamics. Implement in Julia programming language for robust modeling, perform forecasting under data constraints, and use symbolic recovery to transform learned dynamics into explicit mathematical expressions.

Result: The approach demonstrates potential to improve predictive accuracy compared to classical models, enabling better forecasting of tumor growth under data-limited conditions.

Conclusion: The adaptive tumor growth models using Neural ODEs and UDEs have potential to guide dynamic and effective treatment strategies for improved clinical outcomes by providing more accurate, patient-specific predictions.

Abstract: Forecasting tumor growth is critical for optimizing treatment. Classical growth models such as the Gompertz and Bertalanffy equations capture general tumor dynamics but may fail to adapt to patient-specific variability, particularly with limited data available. In this study, we leverage Neural Ordinary Differential Equations (Neural ODEs) and Universal Differential Equations (UDEs), two pillars of Scientific Machine Learning (SciML), to construct adaptive tumor growth models capable of learning from experimental data. Using the Gompertz model as a baseline, we replace rigid terms with adaptive neural networks to capture hidden dynamics through robust modeling in the Julia programming language. We use our models to perform forecasting under data constraints and symbolic recovery to transform the learned dynamics into explicit mathematical expressions. Our approach has the potential to improve predictive accuracy, guiding dynamic and effective treatment strategies for improved clinical outcomes.

Method: FLUX uses privacy-preserving client-side descriptor extraction and unsupervised clustering to handle distribution shifts. It doesn’t require prior knowledge of shift types or cluster numbers, and supports test-time adaptation for unseen clients.

Result: Extensive experiments across 4 standard benchmarks, 2 real-world datasets, and 10 SOTA baselines show FLUX improves performance and stability under diverse distribution shifts, achieving up to 23 percentage points average accuracy gain over best-performing baselines.

Conclusion: FLUX effectively addresses non-IID distribution shifts in FL through clustering-based approach with test-time adaptation, maintaining computational/communication efficiency comparable to FedAvg while significantly improving accuracy.

Abstract: Federated Learning (FL) enables collaborative model training across multiple clients while preserving data privacy. Traditional FL methods often use a global model to fit all clients, assuming that clients’ data are independent and identically distributed (IID). However, when this assumption does not hold, the global model accuracy may drop significantly, limiting FL applicability in real-world scenarios. To address this gap, we propose FLUX, a novel clustering-based FL (CFL) framework that addresses the four most common types of distribution shifts during both training and test time. To this end, FLUX leverages privacy-preserving client-side descriptor extraction and unsupervised clustering to ensure robust performance and scalability across varying levels and types of distribution shifts. Unlike existing CFL methods addressing non-IID client distribution shifts, FLUX i) does not require any prior knowledge of the types of distribution shifts or the number of client clusters, and ii) supports test-time adaptation, enabling unseen and unlabeled clients to benefit from the most suitable cluster-specific models. Extensive experiments across four standard benchmarks, two real-world datasets and ten state-of-the-art baselines show that FLUX improves performance and stability under diverse distribution shifts, achieving an average accuracy gain of up to 23 percentage points over the best-performing baselines, while maintaining computational and communication overhead comparable to FedAvg.

Method: Created DeXposure dataset using DefiLlama metadata, defined value-linked credit exposure based on TVL changes, developed token-to-protocol model to infer inter-protocol credit exposure from token stock dynamics, and established three ML benchmarks.

Result: Key findings: (1) rapid network volume growth, (2) concentration trend to key protocols, (3) declining network density, (4) distinct shock propagation patterns across sectors during Terra and FTX events. Dataset and code publicly released.

Conclusion: DeXposure dataset enables research in ML and financial applications including risk monitoring, policy analysis, and DeFi market modeling, while providing benchmarks for graph clustering, vector autoregression, and temporal graph analysis.

Abstract: We curate the DeXposure dataset, the first large-scale dataset for inter-protocol credit exposure in decentralized financial networks, covering global markets of 43.7 million entries across 4.3 thousand protocols, 602 blockchains, and 24.3 thousand tokens, from 2020 to 2025. A new measure, value-linked credit exposure between protocols, is defined as the inferred financial dependency relationships derived from changes in Total Value Locked (TVL). We develop a token-to-protocol model using DefiLlama metadata to infer inter-protocol credit exposure from the token’s stock dynamics, as reported by the protocols. Based on the curated dataset, we develop three benchmarks for machine learning research with financial applications: (1) graph clustering for global network measurement, tracking the structural evolution of credit exposure networks, (2) vector autoregression for sector-level credit exposure dynamics during major shocks (Terra and FTX), and (3) temporal graph neural networks for dynamic link prediction on temporal graphs. From the analysis, we observe (1) a rapid growth of network volume, (2) a trend of concentration to key protocols, (3) a decline of network density (the ratio of actual connections to possible connections), and (4) distinct shock propagation across sectors, such as lending platforms, trading exchanges, and asset management protocols. The DeXposure dataset and code have been released publicly. We envision they will help with research and practice in machine learning as well as financial risk monitoring, policy analysis, DeFi market modeling, amongst others. The dataset also contributes to machine learning research by offering benchmarks for graph clustering, vector autoregression, and temporal graph analysis.

Method: SingleQuant decouples from quantization truncation to eliminate non-smoothness and gradient noise. It constructs Alignment Rotation Transformation (ART) to smooth outlier values via closed-form optimal rotations, and Uniformity Rotation Transformation (URT) to reshape distributions through geometric mapping. Both use strictly formulated Givens rotations with predetermined dimensions and rotation angles.

Result: SingleQuant achieves 1,400× quantization speedup and increases +0.57% average task performance compared to the best baseline when quantizing LLaMA-2-13B. It demonstrates superiority over selected baselines across diverse tasks on 7B-70B LLMs, enabling higher task performance with less quantization time.

Conclusion: SingleQuant successfully addresses convergence pathology in LLM quantization by eliminating gradient noise and non-smoothness factors, enabling efficient single-pass quantization with improved task performance and dramatically reduced quantization time.

Abstract: Large Language Models (LLMs) quantization facilitates deploying LLMs in resource-limited settings, but existing methods that combine incompatible gradient optimization and quantization truncation lead to serious convergence pathology. This prolongs quantization time and degrades LLMs’ task performance. Our studies confirm that Straight-Through Estimator (STE) on Stiefel manifolds introduce non-smoothness and gradient noise, obstructing optimization convergence and blocking high-fidelity quantized LLM development despite extensive training. To tackle the above limitations, we propose SingleQuant, a single-pass quantization framework that decouples from quantization truncation, thereby eliminating the above non-smoothness and gradient noise factors. Specifically, SingleQuant constructs Alignment Rotation Transformation (ART) and Uniformity Rotation Transformation (URT) targeting distinct activation outliers, where ART achieves smoothing of outlier values via closed-form optimal rotations, and URT reshapes distributions through geometric mapping. Both matrices comprise strictly formulated Givens rotations with predetermined dimensions and rotation angles, enabling promising LLMs task performance within a short time. Experimental results demonstrate SingleQuant’s superiority over the selected baselines across diverse tasks on 7B-70B LLMs. To be more precise, SingleQuant enables quantized LLMs to achieve higher task performance while necessitating less time for quantization. For example, when quantizing LLaMA-2-13B, SingleQuant achieves 1,400$\times$ quantization speedup and increases +0.57% average task performance compared to the selected best baseline.

Method: Proposes physics-informed test-time training (PI-TTT) that performs lightweight self-supervised refinement of surrogate outputs through few gradient-based updates at inference time, enforcing AC power flow equalities and operational constraints without requiring labeled data.

Result: PI-TTT reduces power flow residuals and operational constraint violations by one to two orders of magnitude compared to purely ML-based models on IEEE 14-, 118-, 300-bus systems and PEGASE 1354-bus network, while preserving computational advantages.

Conclusion: PI-TTT provides fast, accurate, and physically reliable predictions, representing a promising direction for scalable and physics-consistent learning in power system analysis.

Abstract: Power Flow (PF) calculation based on machine learning (ML) techniques offer significant computational advantages over traditional numerical methods but often struggle to maintain full physical consistency. This paper introduces a physics-informed test-time training (PI-TTT) framework that enhances the accuracy and feasibility of ML-based PF surrogates by enforcing AC power flow equalities and operational constraints directly at inference time. The proposed method performs a lightweight self-supervised refinement of the surrogate outputs through few gradient-based updates, enabling local adaptation to unseen operating conditions without requiring labeled data. Extensive experiments on the IEEE 14-, 118-, and 300-bus systems and the PEGASE 1354-bus network show that PI-TTT reduces power flow residuals and operational constraint violations by one to two orders of magnitude compared with purely ML-based models, while preserving their computational advantage. The results demonstrate that PI-TTT provides fast, accurate, and physically reliable predictions, representing a promising direction for scalable and physics-consistent learning in power system analysis.

Method: Two-stage approach: (1) Progressive filtering iteratively refines unlabeled pool using ensemble of AL strategies, retaining promising candidates capturing different value notions. (2) Coverage-based selection chooses final batch from refined pool, ensuring all previously identified value notions are accounted for.

Result: Extensive experiments across 6 classification datasets and 3 foundation models show REFINE consistently outperforms individual strategies and existing ensemble methods. Progressive filtering also serves as powerful preprocessing that improves performance of any individual AL strategy applied to refined pool.

Conclusion: REFINE provides robust ensemble AL approach that avoids commitment to single strategy, with progressive filtering as valuable preprocessing step. The ensemble framework can be easily extended with upcoming state-of-the-art AL strategies.

Abstract: Existing active learning (AL) strategies capture fundamentally different notions of data value, e.g., uncertainty or representativeness. Consequently, the effectiveness of strategies can vary substantially across datasets, models, and even AL cycles. Committing to a single strategy risks suboptimal performance, as no single strategy dominates throughout the entire AL process. We introduce REFINE, an ensemble AL method that combines multiple strategies without knowing in advance which will perform best. In each AL cycle, REFINE operates in two stages: (1) Progressive filtering iteratively refines the unlabeled pool by considering an ensemble of AL strategies, retaining promising candidates capturing different notions of value. (2) Coverage-based selection then chooses a final batch from this refined pool, ensuring all previously identified notions of value are accounted for. Extensive experiments across 6 classification datasets and 3 foundation models show that REFINE consistently outperforms individual strategies and existing ensemble methods. Notably, progressive filtering serves as a powerful preprocessing step that improves the performance of any individual AL strategy applied to the refined pool, which we demonstrate on an audio spectrogram classification use case. Finally, the ensemble of REFINE can be easily extended with upcoming state-of-the-art AL strategies.

Method: AutoTailor uses computation graph-guided compilation to automatically transform ML models into SuperNets, and incorporates learning-free latency and accuracy predictors for efficient specialization.

Result: AutoTailor reduces SuperNet construction code by 11-27×, decreases hardware profiling costs by at least 11×, and achieves up to 15.60% accuracy improvement and 60.03% latency reduction vs state-of-the-art.

Conclusion: AutoTailor enables automated, end-to-end SuperNet-based adaptive model deployment for edge devices, overcoming practical limitations of existing approaches.

Abstract: On-device machine learning (ML) has become a fundamental component of emerging mobile applications. Adaptive model deployment delivers efficient inference for heterogeneous device capabilities and performance requirements through customizing neural architectures. SuperNet-based approaches offer a promising solution by generating a large number of model variants from a pre-trained ML model. However, applying SuperNet in existing frameworks suffers from tedious model-aware development and time-consuming hardware-aware profiling, which limits their practical adoption. We present AutoTailor, the first framework to enable automated, end-to-end SuperNet-based adaptive model deployment for edge devices. Unlike manual SuperNet construction, AutoTailor employs a computation graph-guided compilation approach to automatically transform user-provided ML models into SuperNets. To support efficient specialization, AutoTailor incorporates learning-free latency and accuracy predictors, enabling low-cost yet accurate performance prediction. Our extended evaluations demonstrate that AutoTailor reduces the lines of code for SuperNet construction by 11–27$\times$, decreases hardware-aware profiling costs by at least 11$\times$, and achieves up to 15.60% absolute accuracy improvement and 60.03% latency reduction compared to state-of-the-art approaches across diverse models and devices.

Method: Developed Efficient-Husformer with hyperparameter optimization (HPO) for multi-class stress detection across two multimodal physiological datasets (WESAD and CogLoad). Created a structured search space for HPO, conducted comprehensive ablation studies, and optimized architectural decisions including modality combinations, layer count, attention heads, and model dimensions.

Result: Achieved significant performance improvements over original Husformer: 88.41% accuracy on WESAD (13.83% improvement) and 92.61% accuracy on CogLoad (6.98% improvement). Best configuration uses (L + dm) or (L + FFN) modality combinations with single layer, 3 attention heads, model dimension of 18/30, and FFN dimension of 120/30, resulting in compact model with only ~30k parameters.

Conclusion: Efficient-Husformer demonstrates that hyperparameter optimization can create highly efficient Transformer architectures for physiological signal analysis that maintain superior performance while drastically reducing computational and memory requirements, making them more practical for real-world applications.

Abstract: Transformer-based models have gained considerable attention in the field of physiological signal analysis. They leverage long-range dependencies and complex patterns in temporal signals, allowing them to achieve performance superior to traditional RNN and CNN models. However, they require high computational intensity and memory demands. In this work, we present Efficient-Husformer, a novel Transformer-based architecture developed with hyperparameter optimization (HPO) for multi-class stress detection across two multimodal physiological datasets (WESAD and CogLoad). The main contributions of this work are: (1) the design of a structured search space, targeting effective hyperparameter optimization; (2) a comprehensive ablation study evaluating the impact of architectural decisions; (3) consistent performance improvements over the original Husformer, with the best configuration achieving an accuracy of 88.41 and 92.61 (improvements of 13.83% and 6.98%) on WESAD and CogLoad datasets, respectively. The best-performing configuration is achieved with the (L + dm) or (L + FFN) modality combinations, using a single layer, 3 attention heads, a model dimension of 18/30, and FFN dimension of 120/30, resulting in a compact model with only about 30k parameters.

Method: 1) SuRe: Surprise-prioritised Replay that ranks and stores sequences with highest Negative Log-Likelihood (most surprising). 2) Dual-learner design with fast and slow LoRA adapters merged via EMA for rapid adaptation while stabilizing long-term knowledge.

Result: Achieves SOTA in Large Number of Tasks (LNT) setting, best overall average across Standard CL and LNT benchmarks, improvements up to +5 accuracy points on LNT over prior SOTA. Robust under reduced replay frequency and small buffer sizes.

Conclusion: Replay is established as a strong baseline for continual LLM fine-tuning. Surprise-based selection and slow-weight consolidation are complementary for mitigating catastrophic forgetting in LLMs.

Abstract: Continual learning, one’s ability to adapt to a sequence of tasks without forgetting previously acquired knowledge, remains a major challenge in machine learning and a key gap between artificial and human intelligence. While regularisation and replay perform well in vision, they lag behind multi-task learning for large language models (LLMs), especially at scale with many tasks. We revisit replay and argue that two failure modes drive this gap: selection (what to rehearse) and integration (how to consolidate new knowledge). To address selection, we propose Surprise-prioritised Replay (SuRe), a simple, architecture-agnostic rule that ranks and stores the most surprising (high Negative Log-Likelihood) sequences. SuRe achieves state-of-the-art performance in the Large Number of Tasks (LNT) setting and delivers the best overall average across both Standard CL and LNT benchmarks. To address integration, we add a dual-learner design with fast and slow LoRA adapters merged via an exponential moving average (EMA), enabling rapid adaptation while stabilising long-term knowledge. Combining SuRe with the dual learner yields further gains, including improvements of up to +5 accuracy points on LNT over prior SOTA. Ablation studies confirm that our proposed method remains robust under reduced replay frequency and small buffer size, demonstrating both effectiveness and sample efficiency. Taken together, our results establish replay as a strong baseline for continual LLM fine-tuning and demonstrate that surprise-based selection and slow-weight consolidation are complementary components for mitigating catastrophic forgetting.

Method: Two deep learning strategies: (1) grid-to-grid approach (aggregation before modeling) using gridded predictors to model rasterized targets, and (2) grid-to-point approach (aggregation after modeling) using gridded predictors to model point targets followed by kriging interpolation. Tested on groundwater storage data from Bangladesh.

Result: Spatial interpolation is substantially more difficult than temporal prediction. Nearest neighbors are not always the most similar, and geological uncertainties strongly influence point temporal behavior. The findings highlight challenges in spatial modeling.

Conclusion: The study motivates future work on advanced interpolation methods informed by clustering locations based on time series dynamics. Conclusions are applicable to other environmental variables governed by indirectly observable factors beyond groundwater storage.

Abstract: Geospatial observational datasets are often limited to point measurements, making temporal prediction and spatial interpolation essential for constructing continuous fields. This study evaluates two deep learning strategies for addressing this challenge: (1) a grid-to-grid approach, where gridded predictors are used to model rasterised targets (aggregation before modelling), and (2) a grid-to-point approach, where gridded predictors model point targets, followed by kriging interpolation to fill the domain (aggregation after modelling). Using groundwater storage data from Bangladesh as a case study, we compare the effcacy of these approaches. Our findings indicate that spatial interpolation is substantially more difficult than temporal prediction. In particular, nearest neighbours are not always the most similar, and uncertainties in geology strongly influence point temporal behaviour. These insights motivate future work on advanced interpolation methods informed by clustering locations based on time series dynamics. Demonstrated on groundwater storage, the conclusions are applicable to other environmental variables governed by indirectly observable factors. Code is available at https://github.com/pazolka/interpolation-prediction-gwsa.

Method: Hybrid framework fusing implicitly learned dynamics from pretrained TS2Vec encoder with explicit engineered time features encoding periodic cycles. Uses dual-model ensemble architecture with two regression heads (learned dynamics vs. seasonal patterns) combined via adaptive weighting scheme optimized independently for each forecast horizon.

Result: Extensive experiments on ETT benchmark datasets show TS2Vec-Ensemble consistently and significantly outperforms standard TS2Vec baseline and other state-of-the-art models in both univariate and multivariate forecasting.

Conclusion: Hybrid approach combining learned representations with explicit temporal priors is superior strategy for long-horizon time series forecasting, as demonstrated by the success of TS2Vec-Ensemble framework.

Abstract: Self-supervised representation learning, particularly through contrastive methods like TS2Vec, has advanced the analysis of time series data. However, these models often falter in forecasting tasks because their objective functions prioritize instance discrimination over capturing the deterministic patterns, such as seasonality and trend, that are critical for accurate prediction. This paper introduces TS2Vec-Ensemble, a novel hybrid framework designed to bridge this gap. Our approach enhances the powerful, implicitly learned dynamics from a pretrained TS2Vec encoder by fusing them with explicit, engineered time features that encode periodic cycles. This fusion is achieved through a dual-model ensemble architecture, where two distinct regression heads – one focused on learned dynamics and the other on seasonal patterns – are combined using an adaptive weighting scheme. The ensemble weights are optimized independently for each forecast horizon, allowing the model to dynamically prioritize short-term dynamics or long-term seasonality as needed. We conduct extensive experiments on the ETT benchmark datasets for both univariate and multivariate forecasting. The results demonstrate that TS2Vec-Ensemble consistently and significantly outperforms the standard TS2Vec baseline and other state-of-the-art models, validating our hypothesis that a hybrid of learned representations and explicit temporal priors is a superior strategy for long-horizon time series forecasting.

Method: Proposes efficient numerical approximations for entropy, log-probability, and their gradients in truncated normal distributions. Also provides an efficient sampling strategy for truncated policy distributions.

Result: Validated on three benchmark environments, demonstrating significant performance improvements when using accurate estimations compared to prior approximations.

Conclusion: Accurate estimation of key characteristics (entropy, log-probability, gradients) is crucial for action-constrained RL, and the proposed efficient numerical approximations enable better performance while maintaining computational feasibility.

Abstract: In reinforcement learning (RL), it is often advantageous to consider additional constraints on the action space to ensure safety or action relevance. Existing work on such action-constrained RL faces challenges regarding effective policy updates, computational efficiency, and predictable runtime. Recent work proposes to use truncated normal distributions for stochastic policy gradient methods. However, the computation of key characteristics, such as the entropy, log-probability, and their gradients, becomes intractable under complex constraints. Hence, prior work approximates these using the non-truncated distributions, which severely degrades performance. We argue that accurate estimation of these characteristics is crucial in the action-constrained RL setting, and propose efficient numerical approximations for them. We also provide an efficient sampling strategy for truncated policy distributions and validate our approach on three benchmark environments, which demonstrate significant performance improvements when using accurate estimations.

Method: PISA framework introduces dynamic MLP-based aggregation that prioritizes and combines multiple invariant subgraph representations more effectively than previous uniform or greedy aggregation approaches.

Result: Experiments on 15 datasets including DrugOOD show PISA achieves up to 5% higher classification accuracy than prior methods like CIGA and SuGAr.

Conclusion: Dynamic aggregation of multiple invariant subgraphs via MLP-based prioritization significantly improves graph OOD generalization performance over existing approaches.

Abstract: Recent work has extended the invariance principle for out-of-distribution (OOD) generalization from Euclidean to graph data, where challenges arise due to complex structures and diverse distribution shifts in node attributes and topology. To handle these, Chen et al. proposed CIGA (Chen et al., 2022b), which uses causal modeling and an information-theoretic objective to extract a single invariant subgraph capturing causal features. However, this single-subgraph focus can miss multiple causal patterns. Liu et al. (2025) addressed this with SuGAr, which learns and aggregates diverse invariant subgraphs via a sampler and diversity regularizer, improving robustness but still relying on simple uniform or greedy aggregation. To overcome this, the proposed PISA framework introduces a dynamic MLP-based aggregation that prioritizes and combines subgraph representations more effectively. Experiments on 15 datasets, including DrugOOD (Ji et al., 2023), show that PISA achieves up to 5% higher classification accuracy than prior methods.

Method: Introduces efficient GPU-based feature interaction using a novel compressed inverted list designed for GPU acceleration, enabling feature interaction at scale in retrieval systems.

Result: Outperforms existing approaches in offline evaluation and successfully deployed in Tencent Advertising, delivering significant online performance gains.

Conclusion: First framework to successfully implement Wide and Deep in retrieval systems, providing practical guidance for optimizing large-scale ad retrieval systems.

Abstract: In large-scale advertising recommendation systems, retrieval serves as a critical component, aiming to efficiently select a subset of candidate ads relevant to user behaviors from a massive ad inventory for subsequent ranking and recommendation. The Embedding-Based Retrieval (EBR) methods modeled by the dual-tower network are widely used in the industry to maintain both retrieval efficiency and accuracy. However, the dual-tower model has significant limitations: the embeddings of users and ads interact only at the final inner product computation, resulting in insufficient feature interaction capabilities. Although DNN-based models with both user and ad as input features, allowing for early-stage interaction between these features, are introduced in the ranking stage to mitigate this issue, they are computationally infeasible for the retrieval stage. To bridge this gap, this paper proposes an efficient GPU-based feature interaction for the dual-tower network to significantly improve retrieval accuracy while substantially reducing computational costs. Specifically, we introduce a novel compressed inverted list designed for GPU acceleration, enabling efficient feature interaction computation at scale. To the best of our knowledge, this is the first framework in the industry to successfully implement Wide and Deep in a retrieval system. We apply this model to the real-world business scenarios in Tencent Advertising, and experimental results demonstrate that our method outperforms existing approaches in offline evaluation and has been successfully deployed to Tencent’s advertising recommendation system, delivering significant online performance gains. This improvement not only validates the effectiveness of the proposed method, but also provides new practical guidance for optimizing large-scale ad retrieval systems.

Method: Adversarial flow models learn a deterministic noise-to-data mapping using adversarial training, similar to optimal transport in flow-matching models. The method supports native one-step or multi-step generation without needing to learn intermediate timesteps like consistency models.

Result: Achieves FID of 2.38 on ImageNet-256px with XL/2 model (new best), and with deeper models (56-layer and 112-layer) achieves FIDs of 2.08 and 1.94 using single forward pass, surpassing 2NFE and 4NFE counterparts.

Conclusion: Adversarial flow models successfully unify adversarial and flow approaches, providing stable training, efficient one-step generation, and state-of-the-art performance while avoiding error accumulation and saving model capacity.

Abstract: We present adversarial flow models, a class of generative models that unifies adversarial models and flow models. Our method supports native one-step or multi-step generation and is trained using the adversarial objective. Unlike traditional GANs, where the generator learns an arbitrary transport plan between the noise and the data distributions, our generator learns a deterministic noise-to-data mapping, which is the same optimal transport as in flow-matching models. This significantly stabilizes adversarial training. Also, unlike consistency-based methods, our model directly learns one-step or few-step generation without needing to learn the intermediate timesteps of the probability flow for propagation. This saves model capacity, reduces training iterations, and avoids error accumulation. Under the same 1NFE setting on ImageNet-256px, our B/2 model approaches the performance of consistency-based XL/2 models, while our XL/2 model creates a new best FID of 2.38. We additionally show the possibility of end-to-end training of 56-layer and 112-layer models through depth repetition without any intermediate supervision, and achieve FIDs of 2.08 and 1.94 using a single forward pass, surpassing their 2NFE and 4NFE counterparts.

Method: The paper systematically investigates quantization’s impact on four trustworthiness metrics: adversarial robustness, fairness, machine ethics, and out-of-distribution robustness. It then develops a novel precision-ensemble voting approach that leverages predictions from mixed-precision variants of the same model.

Result: The study identifies instability in trustworthiness metrics across compression ratios and quantization methods. The proposed precision-ensemble voting approach consistently improves performance by up to 5.8% on trustworthiness metrics.

Conclusion: The research highlights the importance of considering trustworthiness when developing model compression techniques and points to research opportunities at the intersection of compression and trustworthiness for safety-critical applications.

Abstract: Large language models (LLMs) have shown promising performance across various tasks. However, their autoregressive decoding process poses significant challenges for efficient deployment on existing AI hardware. Quantization alleviates memory and compute pressure by compressing weights, activations, and KV caches to low precisions while preserving generation quality. However, existing quantization frameworks typically focus on perplexity or classification accuracy, often omitting critical trustworthiness metrics. This gap introduces risks when applying quantized LLMs to downstream high-stakes domains such as finance and healthcare. In this work, we systematically investigate the impact of quantization on four trustworthiness metrics (adversarial robustness, fairness, machine ethics, and out-of-distribution robustness) and identify the instability across compression ratios and quantization methods. Building on these observations, we develop a novel precision-ensemble voting approach that leverages predictions from mixed-precision variants of the same model and consistently improves performance by up to $5.8%$ on trustworthiness metrics. Our results highlight the importance of considering trustworthiness when developing model compression techniques and point to research opportunities at the intersection of compression and trustworthiness for safety-critical applications.

Method: The method introduces Space Explanations, a logic-based concept for neural network classification. It uses Craig interpolation algorithms and unsatisfiable core generation to automatically generate explanations that cover continuous regions of input feature space with provable guarantees.

Result: The method was tested on real-life case studies ranging from small to medium to large size networks. The generated explanations were demonstrated to be more meaningful than those computed by state-of-the-art explanation methods.

Conclusion: Space Explanations provide a novel approach to neural network interpretability that offers provable guarantees about network behavior in continuous input regions, producing more meaningful explanations than current state-of-the-art methods through automated logic-based techniques.

Abstract: We present a novel logic-based concept called Space Explanations for classifying neural networks that gives provable guarantees of the behavior of the network in continuous areas of the input feature space. To automatically generate space explanations, we leverage a range of flexible Craig interpolation algorithms and unsatisfiable core generation. Based on real-life case studies, ranging from small to medium to large size, we demonstrate that the generated explanations are more meaningful than those computed by state-of-the-art.

Method: Cross-model evaluation of two DP mechanisms (DPSGD and LDP) applied to four recommender systems (NCF, BPR, SVD, VAE) on MovieLens-1M and Yelp datasets, measuring impacts on utility and bias metrics.

Result: Stronger privacy reduces utility non-uniformly: NCF under DPSGD shows smallest accuracy loss (<10% at ε≈1), while SVD/BPR have larger drops for niche users, and VAE is most sensitive with sharp declines for sparse groups. DPSGD reduces popularity bias gap, LDP preserves existing patterns.

Conclusion: No single DP mechanism is uniformly superior; each provides different trade-offs under varying privacy regimes and data conditions, highlighting the need for context-aware privacy mechanism selection in recommender systems.

Abstract: Recommender systems (RSs) output ranked lists of items, such as movies or restaurants, that users may find interesting, based on the user’s past ratings and ratings from other users. RSs increasingly incorporate differential privacy (DP) to protect user data, raising questions about how privacy mechanisms affect both recommendation accuracy and fairness. We conduct a comprehensive, cross-model evaluation of two DP mechanisms, differentially private stochastic gradient descent (DPSGD) and local differential privacy (LDP), applied to four recommender systems (Neural Collaborative Filtering (NCF), Bayesian Personalized Ranking (BPR), Singular Value Decomposition (SVD), and Variational Autoencoder (VAE)) on the MovieLens-1M and Yelp datasets. We find that stronger privacy consistently reduces utility, but not uniformly. NCF under DPSGD shows the smallest accuracy loss (under 10 percent at epsilon approximately 1), whereas SVD and BPR experience larger drops, especially for users with niche preferences. VAE is the most sensitive to privacy, with sharp declines for sparsely represented groups. The impact on bias metrics is similarly heterogeneous. DPSGD generally reduces the gap between recommendations of popular and less popular items, whereas LDP preserves existing patterns more closely. These results highlight that no single DP mechanism is uniformly superior; instead, each provides trade-offs under different privacy regimes and data conditions.

Method: Uses lossless expanders to synthesize lightly contaminated batches, enabling robust aggregation through a short spectral filtering stage. The framework combines expander-based sketching with efficient computational techniques.

Result: Achieves sample complexity $\tilde{O}((d+\log(1/δ))/α)$, list size $O(1/α)$, and near input-sparsity running time $\tilde{O}(\mathrm{nnz}(X)+d^{3}/α)$ under standard sub-Gaussian assumptions.

Conclusion: The expander-sketching framework provides an efficient, practical approach to list-decodable linear regression that matches the best known statistical guarantees while avoiding computationally expensive SoS machinery and explicit batch structures.

Abstract: We introduce an expander-sketching framework for list-decodable linear regression that achieves sample complexity $\tilde{O}((d+\log(1/δ))/α)$, list size $O(1/α)$, and near input-sparsity running time $\tilde{O}(\mathrm{nnz}(X)+d^{3}/α)$ under standard sub-Gaussian assumptions. Our method uses lossless expanders to synthesize lightly contaminated batches, enabling robust aggregation and a short spectral filtering stage that matches the best known efficient guarantees while avoiding SoS machinery and explicit batch structure.

Method: Extends Convolutional Conditional Neural Processes (ConvCNPs) with Mixture Density Networks (MDNs) output head for epistemic uncertainty estimation, then uses expected reduction in epistemic uncertainty as a new acquisition function for sensor placement.

Result: Preliminary results suggest that epistemic uncertainty driven sensor placement more effectively reduces model error than approaches based on overall uncertainty.

Conclusion: Separating epistemic uncertainty from total uncertainty leads to better sensor placement decisions for spatio-temporal modeling using Neural Processes.

Abstract: Accurate sensor placement is critical for modeling spatio-temporal systems such as environmental and climate processes. Neural Processes (NPs), particularly Convolutional Conditional Neural Processes (ConvCNPs), provide scalable probabilistic models with uncertainty estimates, making them well-suited for data-driven sensor placement. However, existing approaches rely on total predictive uncertainty, which conflates epistemic and aleatoric components, that may lead to suboptimal sensor selection in ambiguous regions. To address this, we propose expected reduction in epistemic uncertainty as a new acquisition function for sensor placement. To enable this, we extend ConvCNPs with a Mixture Density Networks (MDNs) output head for epistemic uncertainty estimation. Preliminary results suggest that epistemic uncertainty driven sensor placement more effectively reduces model error than approaches based on overall uncertainty.

Method: Uses a recently introduced multidimensional threshold-based classification framework to extend score-oriented losses to multiclass settings. Treats decision thresholds as random variables with prior distributions, defining Multiclass Score-Oriented Loss (MultiSOL) functions.

Result: MultiSOL preserves advantages from binary setting: direct optimization of target metrics and robustness to class imbalance. Achieves performance comparable to state-of-the-art loss functions and provides insights into simplex geometry and score-oriented learning interaction.

Conclusion: Successfully extends score-oriented loss framework to multiclass classification, maintaining key benefits while offering new geometric insights into threshold-based learning approaches.

Abstract: In the supervised binary classification setting, score-oriented losses have been introduced with the aim of optimizing a chosen performance metric directly during the training phase, thus avoiding \textit{a posteriori} threshold tuning. To do this, in their construction, the decision threshold is treated as a random variable provided with a certain \textit{a priori} distribution. In this paper, we use a recently introduced multidimensional threshold-based classification framework to extend such score-oriented losses to multiclass classification, defining the Multiclass Score-Oriented Loss (MultiSOL) functions. As also demonstrated by several classification experiments, this proposed family of losses is designed to preserve the main advantages observed in the binary setting, such as the direct optimization of the target metric and the robustness to class imbalance, achieving performance comparable to other state-of-the-art loss functions and providing new insights into the interaction between simplex geometry and score-oriented learning.

Method: Introduces LLM-Cave, a benchmark and lightweight environment based on a classic cave exploration scenario from the Symbolism era. The environment requires agents to explore using partial observable state information while reasoning about nearby dangers. Evaluates mainstream LLMs (GPT-4o-mini, o1-mini, DeepSeek-R1) on sequential reasoning, decision-making performance, and computational efficiency.

Result: DeepSeek-R1 achieved the highest success rate on complex reasoning tasks. Smaller models like 4o-mini significantly narrowed the performance gap by employing Chain of Speculation and Planner-Critic strategies, though at the expense of reduced computational efficiency.

Conclusion: Structured, multi-step reasoning combined with LLM-based feedback mechanisms can substantially enhance LLMs’ decision-making capabilities. This approach provides a promising direction for improving reasoning in weaker models and suggests a new reasoning-centered benchmark for LLM assessment.

Abstract: Large language models (LLMs) such as ChatGPT o1, ChatGPT o3, and DeepSeek R1 have shown great potential in solving difficult problems. However, current LLM evaluation benchmarks are limited to one-step interactions. Some of the existing sequence decision-making environments, such as TextStarCraftII and LLM-PySC2, are too complicated and require hours of interaction to complete a game. In this paper, we introduce LLM-Cave, a benchmark and light environment for LLM reasoning and decision-making systems. This environment is a classic instance in the era of Symbolism. Artificial intelligence enables the agent to explore the environment and avoid potential losses by reasoning about nearby dangers using partial observable state information. In the experiment, we evaluated the sequential reasoning ability, decision-making performance and computational efficiency of mainstream large language models (LLMs) such as GPT-4o-mini, o1-mini, and DeepSeek-R1. Experiments show that while Deepseek-R1 achieved the highest success rate on complex reasoning tasks, smaller models like 4o-mini significantly narrowed the performance gap on challenges by employing Chain of Speculation and Planner-Critic strategies, at the expense of reduced computational efficiency. This indicates that structured, multi-step reasoning combined with an LLM-based feedback mechanism can substantially enhance an LLM’s decision-making capabilities, providing a promising direction for improving reasoning in weaker models and suggesting a new reasoning-centered benchmark for LLM assessment. Our code is open-sourced in https://github.com/puleya1277/CaveEnv.

Method: The survey employs a hybrid methodology combining bibliometric analysis with systematic review to identify influential works, examines aggregation strategies (architectures, synchronization methods, federation objectives), and includes experimental comparisons of aggregation methods under IID and non-IID data distributions.

Result: The paper provides a structured classification of FL research, identifies key challenges and techniques, offers practical evaluation approaches, and presents experimental comparisons that demonstrate the performance of different aggregation methods under various data distribution scenarios.

Conclusion: The survey outlines promising research directions to advance Federated Learning, aiming to guide future innovation in this rapidly evolving field by addressing remaining challenges in heterogeneity, efficiency, security, and privacy while leveraging FL’s advantages for privacy-preserving collaborative AI.

Abstract: The integration of IoT and AI has unlocked innovation across industries, but growing privacy concerns and data isolation hinder progress. Traditional centralized ML struggles to overcome these challenges, which has led to the rise of Federated Learning (FL), a decentralized paradigm that enables collaborative model training without sharing local raw data. FL ensures data privacy, reduces communication overhead, and supports scalability, yet its heterogeneity adds complexity compared to centralized approaches. This survey focuses on three main FL research directions: personalization, optimization, and robustness, offering a structured classification through a hybrid methodology that combines bibliometric analysis with systematic review to identify the most influential works. We examine challenges and techniques related to heterogeneity, efficiency, security, and privacy, and provide a comprehensive overview of aggregation strategies, including architectures, synchronization methods, and diverse federation objectives. To complement this, we discuss practical evaluation approaches and present experiments comparing aggregation methods under IID and non-IID data distributions. Finally, we outline promising research directions to advance FL, aiming to guide future innovation in this rapidly evolving field.

Method: Flow Density Control (FDC) reduces the complex optimization problem to a sequence of simpler fine-tuning tasks, each solvable via established scalable methods. The approach leverages mirror flows theory and can handle various divergence measures like optimal transport distances and Renyi divergences.

Result: The method provides convergence guarantees under realistic assumptions and demonstrates effectiveness on text-to-image generation, molecular design, and other illustrative tasks, solving problems beyond current fine-tuning schemes.

Conclusion: FDC enables more flexible and powerful fine-tuning of foundation generative models for practical applications by supporting general utility optimization and diverse prior preservation methods, expanding the capabilities of current fine-tuning approaches.

Abstract: Adapting large-scale foundation flow and diffusion generative models to optimize task-specific objectives while preserving prior information is crucial for real-world applications such as molecular design, protein docking, and creative image generation. Existing principled fine-tuning methods aim to maximize the expected reward of generated samples, while retaining knowledge from the pre-trained model via KL-divergence regularization. In this work, we tackle the significantly more general problem of optimizing general utilities beyond average rewards, including risk-averse and novelty-seeking reward maximization, diversity measures for exploration, and experiment design objectives among others. Likewise, we consider more general ways to preserve prior information beyond KL-divergence, such as optimal transport distances and Renyi divergences. To this end, we introduce Flow Density Control (FDC), a simple algorithm that reduces this complex problem to a specific sequence of simpler fine-tuning tasks, each solvable via scalable established methods. We derive convergence guarantees for the proposed scheme under realistic assumptions by leveraging recent understanding of mirror flows. Finally, we validate our method on illustrative settings, text-to-image, and molecular design tasks, showing that it can steer pre-trained generative models to optimize objectives and solve practically relevant tasks beyond the reach of current fine-tuning schemes.

Method: The approach uses reference trajectories to identify Koopman mode amplitudes, transforms decompositions to a basis of fundamental eigenvalue and time functions, projects trajectories via regularized least-squares fit, optimizes eigenvalues with global optimization, and penalizes deviations from the Koopman PDE using numerically computed gradients.

Result: Successfully tested on benchmark nonlinear systems (FitzHugh-Nagumo with inputs, van der Pol and Duffing oscillators, 2-spool turbojet engine with control). The method improves Koopman predictor accuracy through principal eigenvalues and spatial structure promotion, works with sparse sampling, reveals geometric features like invariant partitions, and enables input dynamics modeling.

Conclusion: The approach provides a practical, data-driven method for discovering Koopman eigenfunctions without predefined bases, demonstrating robustness across various dynamical systems and enabling applications in control design through gradient approximation.

Abstract: A new approach to data-driven discovery of Koopman eigenfunctions without a pre-defined set of basis functions is proposed. The approach is based on a reference trajectory, for which the Koopman mode amplitudes are first identified, and the Koopman mode decomposition is transformed to a new basis, which contains fundamental functions of eigenvalues and time. The initial values of the eigenfunctions are obtained by projecting trajectories onto this basis via a regularized least-squares fit. A global optimizer was employed to optimize the eigenvalues. Mapping initial-state values to eigenfunction values reveals their spatial structure, enabling the numerical computation of their gradients. Thus, deviations from the Koopman partial differential equation are penalized, leading to more robust solutions. The approach was successfully tested on several benchmark nonlinear dynamical systems, including the FitzHugh-Nagumo system with inputs, van der Pol and Duffing oscillators, and a 2-spool turbojet engine with control. The study demonstrates that incorporating principal eigenvalues and spatial structure integrity promotion significantly improves the accuracy of Koopman predictors. The approach effectively discovers Koopman spectral components even with sparse state-space sampling and reveals geometric features of the state space, such as invariant partitions. Finally, the numerical approximation of the eigenfunction gradient can be used for input dynamics modeling and control design. The results support the practicality of the approach for use with various dynamical systems.

Method: SHINE-FS learns consensus anchors and anchor graphs to capture cross-view relationships, generates low-dimensional representations for data reconstruction, learns second-order similarity graphs from anchor-sample relationships, and jointly learns first-order and second-order similarity graphs to construct hybrid-order similarity graphs that capture both local and global structures.

Result: Comprehensive experimental results on real multi-view datasets show that SHINE-FS outperforms state-of-the-art methods.

Conclusion: SHINE-FS effectively addresses the limitations of existing MUFS methods by learning hybrid-order similarity graphs that capture both local and global data structures, leading to improved feature selection performance for unlabeled multi-view data.

Abstract: Multi-view unsupervised feature selection (MUFS) has recently emerged as an effective dimensionality reduction method for unlabeled multi-view data. However, most existing methods mainly use first-order similarity graphs to preserve local structure, often overlooking the global structure that can be captured by second-order similarity. In addition, a few MUFS methods leverage predefined second-order similarity graphs, making them vulnerable to noise and outliers and resulting in suboptimal feature selection performance. In this paper, we propose a novel MUFS method, termed Structure-aware Hybrid-order sImilarity learNing for multi-viEw unsupervised Feature Selection (SHINE-FS), to address the aforementioned problem. SHINE-FS first learns consensus anchors and the corresponding anchor graph to capture the cross-view relationships between the anchors and the samples. Based on the acquired cross-view consensus information, it generates low-dimensional representations of the samples, which facilitate the reconstruction of multi-view data by identifying discriminative features. Subsequently, it employs the anchor-sample relationships to learn a second-order similarity graph. Furthermore, by jointly learning first-order and second-order similarity graphs, SHINE-FS constructs a hybrid-order similarity graph that captures both local and global structures, thereby revealing the intrinsic data structure to enhance feature selection. Comprehensive experimental results on real multi-view datasets show that SHINE-FS outperforms the state-of-the-art methods.

Method: Analysis through conceptual arguments, examination of existing empirical works, and novel illustrative case studies to identify obstacles in collecting labeled deception examples.

Result: Identified several concrete obstacles in collecting reliable labeled examples of deceptive vs honest AI behavior, and found that proposed empirical workarounds are valuable but insufficient alone.

Conclusion: Progress on AI deception detection requires further consideration of the fundamental problems in obtaining labeled examples, as current approaches face significant limitations.

Abstract: Building reliable deception detectors for AI systems – methods that could predict when an AI system is being strategically deceptive without necessarily requiring behavioural evidence – would be valuable in mitigating risks from advanced AI systems. But evaluating the reliability and efficacy of a proposed deception detector requires examples that we can confidently label as either deceptive or honest. We argue that we currently lack the necessary examples and further identify several concrete obstacles in collecting them. We provide evidence from conceptual arguments, analysis of existing empirical works, and analysis of novel illustrative case studies. We also discuss the potential of several proposed empirical workarounds to these problems and argue that while they seem valuable, they also seem insufficient alone. Progress on deception detection likely requires further consideration of these problems.

Method: Review of main architectures, pretraining strategies, and optimization methods for foundation models in time series forecasting, with empirical study of fine-tuning effects after pretraining.

Result: Fine-tuning generally improves zero-shot forecasting capabilities, with particularly significant benefits for long-term horizons.

Conclusion: Foundation models with appropriate fine-tuning strategies can effectively enhance zero-shot time series forecasting performance, especially for challenging long-term prediction tasks.

Abstract: Inspired by recent advances in large language models, foundation models have been developed for zero-shot time series forecasting, enabling prediction on datasets unseen during pretraining. These large-scale models, trained on vast collections of time series, learn generalizable representations for both point and probabilistic forecasting, reducing the need for task-specific architectures and manual tuning. In this work, we review the main architectures, pretraining strategies, and optimization methods used in such models, and study the effect of fine-tuning after pretraining to enhance their performance on specific datasets. Our empirical results show that fine-tuning generally improves zero-shot forecasting capabilities, especially for long-term horizons.

Method: Flow Map Trajectory Tilting (FMTT) exploits the relationship between flow maps and velocity fields to construct an algorithm that performs better reward ascent than gradient-based methods, enabling exact sampling via importance weighting or principled search for local maximizers.

Result: FMTT outperforms other look-ahead techniques and enables engagement with complicated reward functions, making possible new forms of image editing through vision language models.

Conclusion: Working directly with flow maps provides a principled solution to the reward gradient problem in diffusion models, enabling more effective test-time optimization and new applications in image editing.

Abstract: A common recipe to improve diffusion models at test-time so that samples score highly against a user-specified reward is to introduce the gradient of the reward into the dynamics of the diffusion itself. This procedure is often ill posed, as user-specified rewards are usually only well defined on the data distribution at the end of generation. While common workarounds to this problem are to use a denoiser to estimate what a sample would have been at the end of generation, we propose a simple solution to this problem by working directly with a flow map. By exploiting a relationship between the flow map and velocity field governing the instantaneous transport, we construct an algorithm, Flow Map Trajectory Tilting (FMTT), which provably performs better ascent on the reward than standard test-time methods involving the gradient of the reward. The approach can be used to either perform exact sampling via importance weighting or principled search that identifies local maximizers of the reward-tilted distribution. We demonstrate the efficacy of our approach against other look-ahead techniques, and show how the flow map enables engagement with complicated reward functions that make possible new forms of image editing, e.g. by interfacing with vision language models.