Method: VPR-AttLLM integrates LLM semantic reasoning and geospatial knowledge into established VPR pipelines through attention-guided descriptor enhancement. It uses LLMs to identify location-informative regions and suppress transient visual noise without requiring model retraining or additional data.

Result: Integration with three state-of-the-art VPR models (CosPlace, EigenPlaces, SALAD) consistently improves recall performance, with relative gains typically between 1-3% and up to 8% on challenging real flood imagery. Evaluations on extended benchmarks including SF-XL with real flood images and new HK-URBAN dataset show strong performance.

Conclusion: VPR-AttLLM establishes a generalizable paradigm for LLM-guided multimodal fusion in visual retrieval systems, bridging human-like spatial reasoning with modern VPR architectures. Its plug-and-play design, cross-source robustness, and interpretability highlight potential for scalable urban monitoring and rapid geo-localization of crowdsourced crisis imagery.

Abstract: Crowdsourced street-view imagery from social media provides valuable real-time visual evidence of urban flooding and other crisis events, yet it often lacks reliable geographic metadata for emergency response. Existing Visual Place Recognition (VPR) models exhibit substantial performance degradation when applied to such imagery due to visual distortions and domain shifts inherent in cross-source scenarios. This paper presents VPR-AttLLM, a model-agnostic framework that integrates the semantic reasoning and geospatial knowledge of Large Language Models (LLMs) into established VPR pipelines through attention-guided descriptor enhancement. By leveraging LLMs to identify location-informative regions within the city context and suppress transient visual noise, VPR-AttLLM improves retrieval performance without requiring model retraining or additional data. Comprehensive evaluations are conducted on extended benchmarks including SF-XL enriched with real social-media flood images, synthetic flooding scenarios over established query sets and Mapillary photos, and a new HK-URBAN dataset capturing morphologically distinct cityscapes. Integrating VPR-AttLLM with three state-of-the-art VPR models-CosPlace, EigenPlaces, and SALAD-consistently improves recall performance, yielding relative gains typically between 1-3% and reaching up to 8% on the most challenging real flood imagery. Beyond measurable gains in retrieval accuracy, this study establishes a generalizable paradigm for LLM-guided multimodal fusion in visual retrieval systems. By embedding principles from urban perception theory into attention mechanisms, VPR-AttLLM bridges human-like spatial reasoning with modern VPR architectures. Its plug-and-play design, strong cross-source robustness, and interpretability highlight its potential for scalable urban monitoring and rapid geo-localization of crowdsourced crisis imagery.

Method: Investigated RL techniques (GRPO and novel Latent RL) for optimizing latent thinking steps, comparing with supervised fine-tuning (Coconut approach).

Result: RL-trained latent-space models still lag behind traditional language-space CoT models in mathematical reasoning performance.

Conclusion: Current latent-space thinking methods, including RL approaches, cannot match language-space CoT for complex reasoning tasks, revealing limitations in this research direction.

Abstract: Chain-of-Thought (CoT) reasoning typically utilizes the discrete language space for thinking, which is inherently inefficient, as many generated tokens only enforce linguistic rules that are not required for reasoning. To bypass this, latent-space thinking allows models to think using the continuous embedding space. While existing methods for training those models show domain-specific gains, they fail to maintain performance in complex tasks, such as mathematical reasoning. We experimentally demonstrate that the Coconut approach, a form of supervised fine-tuning for latent-space thinking, is highly sensitive to design choices and exhibits several inherent limitations. To address these issues, we investigate reinforcement learning (RL) techniques – an underexplored direction in latent-space thinking – including GRPO and design a novel Latent RL method for directly optimizing the latent thinking steps. Our experimental results reveal that these RL-trained models still lag behind traditional language-space CoT models in the mathematical reasoning domain. We make our codebase publicly available.

Method: Created KH-FUNSD dataset with three-level annotation: 1) Region detection (header, form field, footer zones), 2) FUNSD-style annotation (questions, answers, headers, relationships), 3) Fine-grained classification (field labels, values, headers, footers, symbols).

Result: First publicly available Khmer business document dataset with hierarchical annotations. Benchmark results established for leading models, providing baselines for Khmer document understanding and highlighting challenges of non-Latin, low-resource scripts.

Conclusion: KH-FUNSD addresses critical gap in document AI for low-resource scripts, enabling comprehensive layout analysis and information extraction for Khmer business documents, with dataset and documentation made publicly available.

Abstract: Automated document layout analysis remains a major challenge for low-resource, non-Latin scripts. Khmer is a language spoken daily by over 17 million people in Cambodia, receiving little attention in the development of document AI tools. The lack of dedicated resources is particularly acute for business documents, which are critical for both public administration and private enterprise. To address this gap, we present \textbf{KH-FUNSD}, the first publicly available, hierarchically annotated dataset for Khmer form document understanding, including receipts, invoices, and quotations. Our annotation framework features a three-level design: (1) region detection that divides each document into core zones such as header, form field, and footer; (2) FUNSD-style annotation that distinguishes questions, answers, headers, and other key entities, together with their relationships; and (3) fine-grained classification that assigns specific semantic roles, such as field labels, values, headers, footers, and symbols. This multi-level approach supports both comprehensive layout analysis and precise information extraction. We benchmark several leading models, providing the first set of baseline results for Khmer business documents, and discuss the distinct challenges posed by non-Latin, low-resource scripts. The KH-FUNSD dataset and documentation will be available at URL.

Method: Direct Confidence Alignment (DCA) uses Direct Preference Optimization to align an LLM’s verbalized confidence with its internal confidence rather than ground-truth accuracy. The paper also introduces three new calibration error-based metrics to assess this alignment.

Result: DCA improves alignment metrics on certain model architectures, reducing inconsistencies in confidence expression. However, it can be ineffective on others, showing mixed results across different open-weight LLMs tested on various datasets.

Conclusion: While DCA enhances model transparency by aligning verbalized and internal confidence, its effectiveness varies by architecture, highlighting the need for more model-aware approaches to achieve interpretable and trustworthy LLMs.

Abstract: Producing trustworthy and reliable Large Language Models (LLMs) has become increasingly important as their usage becomes more widespread. Calibration seeks to achieve this by improving the alignment between the model’s confidence and the actual likelihood of its responses being correct or desirable. However, it has been observed that the internal confidence of a model, derived from token probabilities, is not well aligned with its verbalized confidence, leading to misleading results with different calibration methods. In this paper, we propose Direct Confidence Alignment (DCA), a method using Direct Preference Optimization to align an LLM’s verbalized confidence with its internal confidence rather than ground-truth accuracy, enhancing model transparency and reliability by ensuring closer alignment between the two confidence measures. We evaluate DCA across multiple open-weight LLMs on a wide range of datasets. To further assess this alignment, we also introduce three new calibration error-based metrics. Our results show that DCA improves alignment metrics on certain model architectures, reducing inconsistencies in a model’s confidence expression. However, we also show that it can be ineffective on others, highlighting the need for more model-aware approaches in the pursuit of more interpretable and trustworthy LLMs.

Method: Benchmarked several popular KV cache compression strategies on long-reasoning tasks. Evaluated non-reasoning Llama-3.1-8B-Instruct and reasoning models, comparing strategies including H2O and a decoding-enabled variant of SnapKV on datasets like GSM8K and MATH500.

Result: For non-reasoning models, no single strategy works for all datasets. For reasoning models, H2O and the modified SnapKV are dominant strategies, showing heavy-hitter tracking is useful for reasoning traces. Low-budget eviction strategies can produce longer reasoning traces, revealing cache size vs. inference cost tradeoffs.

Conclusion: KV cache compression strategy effectiveness depends on model type and task. Heavy-hitter tracking strategies like H2O and modified SnapKV work best for reasoning models, while tradeoffs exist between cache compression and inference efficiency in long-reasoning tasks.

Abstract: Large language models (LLMs) have demonstrated remarkable performance on long-context tasks, but are often bottlenecked by memory constraints. Namely, the KV cache, which is used to significantly speed up attention computations, grows linearly with context length. A suite of compression algorithms has been introduced to alleviate cache growth by evicting unimportant tokens. However, several popular strategies are targeted towards the prefill phase, i.e., processing long prompt context, and their performance is rarely assessed on reasoning tasks requiring long decoding. In particular, short but complex prompts, such as those in benchmarks like GSM8K and MATH500, often benefit from multi-step reasoning and self-reflection, resulting in thinking sequences thousands of tokens long. In this work, we benchmark the performance of several popular compression strategies on long-reasoning tasks. For the non-reasoning Llama-3.1-8B-Instruct, we determine that no singular strategy fits all, and that performance is heavily influenced by dataset type. However, we discover that H2O and our decoding-enabled variant of SnapKV are dominant strategies for reasoning models, indicating the utility of heavy-hitter tracking for reasoning traces. We also find that eviction strategies at low budgets can produce longer reasoning traces, revealing a tradeoff between cache size and inference costs.

Method: Used LLMs with advanced prompting techniques (input-output, chain-of-thought, self-consistency, multi-agent) across 13 reasoning and non-reasoning models from various providers. Generated synthetic user utterances with correct and failure-containing system responses, focusing on restaurant recommendation case study.

Result: Multi-agent prompting significantly improved small non-reasoning models, while reasoning models consistently outperformed non-reasoning models. DeepSeek-R1 achieved F1-score of 0.99 with self-consistency prompting at $0.002 per request. DeepSeek-V3 offered best balance of effectiveness and cost-time efficiency.

Conclusion: LLM-based evaluation provides a scalable and accurate alternative to human evaluation for benchmarking contextual understanding in conversational QA systems, with different models offering trade-offs between performance and efficiency.

Abstract: In-Car Conversational Question Answering (ConvQA) systems significantly enhance user experience by enabling seamless voice interactions. However, assessing their accuracy and reliability remains a challenge. This paper explores the use of Large Language Models (LLMs) alongside advanced prompting techniques and agent-based methods to evaluate the extent to which ConvQA system responses adhere to user utterances. The focus lies on contextual understanding and the ability to provide accurate venue recommendations considering user constraints and situational context. To evaluate utterance-response coherence using an LLM, we synthetically generate user utterances accompanied by correct and modified failure-containing system responses. We use input-output, chain-of-thought, self-consistency prompting, and multi-agent prompting techniques with 13 reasoning and non-reasoning LLMs of varying sizes and providers, including OpenAI, DeepSeek, Mistral AI, and Meta. We evaluate our approach on a case study involving restaurant recommendations. The most substantial improvements occur for small non-reasoning models when applying advanced prompting techniques, particularly multi-agent prompting. However, reasoning models consistently outperform non-reasoning models, with the best performance achieved using single-agent prompting with self-consistency. Notably, DeepSeek-R1 reaches an F1-score of 0.99 at a cost of 0.002 USD per request. Overall, the best balance between effectiveness and cost-time efficiency is reached with the non-reasoning model DeepSeek-V3. Our findings show that LLM-based evaluation offers a scalable and accurate alternative to traditional human evaluation for benchmarking contextual understanding in ConvQA systems.

Method: Introduces two simple strategies that modify training data by previewing partial answer prefixes, minimizing perturbations to the model’s initial token distributions to maintain built-in safety mechanisms.

Result: Comprehensive experiments show LookAhead Tuning effectively maintains model safety without sacrificing robust performance on downstream tasks.

Conclusion: LookAhead Tuning is a reliable and efficient solution for the safe and effective adaptation of LLMs, positioning it as a practical approach for domain-specific fine-tuning while preserving safety.

Abstract: Fine-tuning enables large language models (LLMs) to adapt to specific domains, but often compromises their previously established safety alignment. To mitigate the degradation of model safety during fine-tuning, we introduce LookAhead Tuning, a lightweight and effective data-driven approach that preserves safety during fine-tuning. The method introduces two simple strategies that modify training data by previewing partial answer prefixes, thereby minimizing perturbations to the model’s initial token distributions and maintaining its built-in safety mechanisms. Comprehensive experiments demonstrate that LookAhead Tuning effectively maintains model safety without sacrificing robust performance on downstream tasks. Our findings position LookAhead Tuning as a reliable and efficient solution for the safe and effective adaptation of LLMs.

Method: Voyager uses an iterative approach that directly optimizes dataset diversity using determinantal point processes (DPPs). It’s training-free, applicable to closed-source models, and scalable.

Result: Voyager significantly outperforms popular baseline approaches, providing a 1.5-3x improvement in diversity across comprehensive experiments.

Conclusion: Voyager offers a principled, theoretically justified approach to generating diverse synthetic datasets from LLMs, addressing a key limitation in current LLM-based data generation methods.

Abstract: Large language models (LLMs) are increasingly being used to generate synthetic datasets for the evaluation and training of downstream models. However, prior work has noted that such generated data lacks diversity. In this paper, we propose Voyager, a novel principled approach to generate diverse datasets. Our approach is iterative and directly optimizes a mathematical quantity that optimizes the diversity of the dataset using the machinery of determinantal point processes. Furthermore, our approach is training-free, applicable to closed-source models, and scalable. In addition to providing theoretical justification for the working of our method, we also demonstrate through comprehensive experiments that Voyager significantly outperforms popular baseline approaches by providing a 1.5-3x improvement in diversity.

Method: Proposes MemoryEQA with three key mechanisms: 1) Convert observations to structured textual representations stored in vector library, 2) Viewpoint comparison strategy for memory updates, 3) Entropy-based adaptive retrieval strategy to get minimal sufficient memory for each module before execution.

Result: Achieved 9.9% performance gain on MT-HM3D benchmark compared to baselines. Demonstrated effectiveness on HM-EQA, MT-HM3D, and OpenEQA benchmarks.

Conclusion: Memory capability is pivotal for solving complex EQA tasks. The memory-centric approach with structured storage, adaptive retrieval, and module-specific memory integration significantly improves exploration efficiency and success rates.

Abstract: Embodied Question Answering (EQA) requires agents to autonomously explore and comprehend the environment to answer context-dependent questions. Typically, an EQA framework consists of four components: a planner, a memory module, a stopping module, and an answering module. However, the memory module is utilized inefficiently in existing methods, as the information it stores is leveraged solely for the answering module. Such a design may result in redundant or inadequate exploration, leading to a suboptimal success rate. To solve this problem, we propose MemoryEQA, an EQA framework centered on memory, which establishes mechanisms for memory storage, update, and retrieval, allowing memory information to contribute throughout the entire exploration process. Specifically, we convert the observation into structured textual representations, which are stored in a vector library following a fixed structure. At each exploration step, we utilize a viewpoint comparison strategy to determine whether the memory requires updating. Before executing each module, we employ an entropy-based adaptive retrieval strategy to obtain the minimal yet sufficient memory information that satisfies the requirements of different modules. The retrieved module-specific information is then integrated with the current observation as input to the corresponding module. To evaluate EQA models’ memory capabilities, we constructed the benchmark based on HM3D called MT-HM3D, comprising 1,587 question-answer pairs involving multiple targets across various regions, which requires agents to maintain memory of exploration-acquired target information. Experimental results on HM-EQA, MT-HM3D, and OpenEQA demonstrate the effectiveness of our framework, where a 9.9% performance gain on MT-HM3D compared to baseline models further underscores the memory capability’s pivotal role in solving complex tasks.

Method: Uses fixed threshold with existing online softmax information to identify negligible attention scores, skipping softmax computation, Value block loading, and matrix multiplication. Fits into FlashAttention kernels with minimal overhead. Includes automated calibration revealing inverse relationship between optimal threshold and context length.

Result: Achieves 1.62x speedup for prefill at 74.7% sparsity and 1.48x speedup for decode at 73.2% sparsity on modern GPUs while maintaining high accuracy. Also shows sparsity-aware training can further improve accuracy-sparsity tradeoffs.

Conclusion: BLASST provides a unified, efficient solution for accelerating long-context inference across all attention variants with minimal integration effort, addressing critical bottlenecks in LLM deployment.

Abstract: The growing demand for long-context inference capabilities in Large Language Models (LLMs) has intensified the computational and memory bottlenecks inherent to the standard attention mechanism. To address this challenge, we introduce BLASST, a drop-in sparse attention method that dynamically prunes the attention matrix without any pre-computation or proxy scores. Our method uses a fixed threshold and existing information from online softmax to identify negligible attention scores, skipping softmax computation, Value block loading, and the subsequent matrix multiplication. This fits seamlessly into existing FlashAttention kernel designs with negligible latency overhead. The approach is applicable to both prefill and decode stages across all attention variants (MHA, GQA, MQA, and MLA), providing a unified solution for accelerating long-context inference. We develop an automated calibration procedure that reveals a simple inverse relationship between optimal threshold and context length, enabling robust deployment across diverse scenarios. Maintaining high accuracy, we demonstrate a 1.62x speedup for prefill at 74.7% sparsity and a 1.48x speedup for decode at 73.2% sparsity on modern GPUs. Furthermore, we explore sparsity-aware training as a natural extension, showing that models can be trained to be inherently more robust to sparse attention patterns, pushing the accuracy-sparsity frontier even further.

Method: DroPE removes positional embeddings after pretraining following a short recalibration phase. This simple approach leverages that PEs are crucial during pretraining but not inherently required for effective language modeling after training.

Result: DroPE enables seamless zero-shot context extension without long-context finetuning, quickly adapting pretrained LMs without compromising original capabilities. It outperforms specialized architectures and rotary PE scaling methods across different models and dataset sizes.

Conclusion: Positional embeddings can be safely removed after pretraining, breaking the bottleneck of expensive finetuning for context extension and enabling effective zero-shot generalization to longer sequences.

Abstract: So far, expensive finetuning beyond the pretraining sequence length has been a requirement for effectively extending the context of language models (LM). In this work, we break this key bottleneck by Dropping the Positional Embeddings of LMs after training (DroPE). Our simple method is motivated by three key theoretical and empirical observations. First, positional embeddings (PEs) serve a crucial role during pretraining, providing an important inductive bias that significantly facilitates convergence. Second, over-reliance on this explicit positional information is also precisely what prevents test-time generalization to sequences of unseen length, even when using popular PE-scaling methods. Third, positional embeddings are not an inherent requirement of effective language modeling and can be safely removed after pretraining, following a short recalibration phase. Empirically, DroPE yields seamless zero-shot context extension without any long-context finetuning, quickly adapting pretrained LMs without compromising their capabilities in the original training context. Our findings hold across different models and dataset sizes, far outperforming previous specialized architectures and established rotary positional embedding scaling methods.

Method: MEDAL integrates Monte Carlo Tree Search at the initialization stage of DLM inference. It restricts search space to high-confidence actions and prioritizes token choices that improve model confidence over remaining masked positions, exploring promising unmasking trajectories before refinement.

Result: MEDAL achieves up to 22.0% improvement over existing inference strategies across multiple benchmarks, establishing a new paradigm for search-based inference in diffusion language models.

Conclusion: The MEDAL framework successfully introduces a principled search mechanism for DLM inference through Monte Carlo Tree Search integration, significantly outperforming existing methods and offering a robust approach to parallel text generation.

Abstract: Diffusion language models (DLMs) have recently emerged as a compelling alternative to autoregressive generation, offering parallel generation and improved global coherence. During inference, DLMs generate text by iteratively denoising masked sequences in parallel; however, determining which positions to unmask and which tokens to commit forms a large combinatorial search problem. Existing inference methods approximate this search using heuristics, which often yield suboptimal decoding paths; other approaches instead rely on additional training to guide token selection. To introduce a principled search mechanism for DLMs inference, we introduce MEDAL, a framework that integrates Monte Carlo Tree SEarch initialization for Diffusion LAnguage Model inference. We employ Monte Carlo Tree Search at the initialization stage to explore promising unmasking trajectories, providing a robust starting point for subsequent refinement. This integration is enabled by restricting the search space to high-confidence actions and prioritizing token choices that improve model confidence over remaining masked positions. Across multiple benchmarks, MEDAL achieves up to 22.0% improvement over existing inference strategies, establishing a new paradigm for search-based inference in diffusion language models.

Method: Multi-agent system with adult (mother) and daughter language models; daughter learns from mother-generated utterances via hierarchical agglomerative cluster analysis to identify grammatical patterns and derive discrete rules.

Result: The system successfully acquires non-trivial grammatical categories resembling those proposed by linguists, validated through both training and test experiments with consistent parameter configuration.

Conclusion: Computational language acquisition through statistical analysis of input exemplars can yield abstract grammatical knowledge, demonstrating a viable approach to modeling language learning without direct access to internal linguistic representations.

Abstract: This article presents experiments performed using a computational laboratory environment for language acquisition experiments. It implements a multi-agent system consisting of two agents: an adult language model and a daughter language model that aims to learn the mother language. Crucially, the daughter agent does not have access to the internal knowledge of the mother language model but only to the language exemplars the mother agent generates. These experiments illustrate how this system can be used to acquire abstract grammatical knowledge. We demonstrate how statistical analyses of patterns in the input data corresponding to grammatical categories yield discrete grammatical rules. These rules are subsequently added to the grammatical knowledge of the daughter language model. To this end, hierarchical agglomerative cluster analysis was applied to the utterances consecutively generated by the mother language model. It is argued that this procedure can be used to acquire structures resembling grammatical categories proposed by linguists for natural languages. Thus, it is established that non-trivial grammatical knowledge has been acquired. Moreover, the parameter configuration of this computational laboratory environment determined using training data generated by the mother language model is validated in a second experiment with a test set similarly resulting in the acquisition of non-trivial categories.

Method: Models latent semantic function as Gaussian process with combined Matérn and polynomial kernel covariance. Learns kernel parameters automatically from supervised data rather than hand-crafting them.

Result: Experimental evaluation in fine-grained sentiment classification with large language models under in-context learning setup demonstrates effectiveness of the proposed measure.

Conclusion: The multi-kernel Gaussian process approach provides an adaptive semantic distance measure that can be learned from data, offering improved flexibility over fixed classical methods.

Abstract: Semantic distance measurement is a fundamental problem in computational linguistics, providing a quantitative characterization of similarity or relatedness between text segments, and underpinning tasks such as text retrieval and text classification. From a mathematical perspective, a semantic distance can be viewed as a metric defined on a space of texts or on a representation space derived from them. However, most classical semantic distance methods are essentially fixed, making them difficult to adapt to specific data distributions and task requirements. In this paper, a semantic distance measure based on multi-kernel Gaussian processes (MK-GP) was proposed. The latent semantic function associated with texts was modeled as a Gaussian process, with its covariance function given by a combined kernel combining Matérn and polynomial components. The kernel parameters were learned automatically from data under supervision, rather than being hand-crafted. This semantic distance was instantiated and evaluated in the context of fine-grained sentiment classification with large language models under an in-context learning (ICL) setup. The experimental results demonstrated the effectiveness of the proposed measure.

Method: Compiled a typologically broad dataset of 810 adjectives (27 languages, 30 words each), each phonemically transcribed and validated with native-speaker audio. Used interpretable classifiers over bag-of-segment features. Trained an adversarial scrubber that suppresses language identity while preserving size signal to probe universality beyond genealogy.

Result: Phonological form predicts size semantics above chance even across unrelated languages, with both vowels and consonants contributing. Language prediction averaged across languages and settings falls below chance while size prediction remains significantly above chance, indicating cross-family sound-symbolic bias.

Conclusion: The study provides evidence for cross-linguistic sound symbolism in size semantics, demonstrating that phonological form can predict size meaning across unrelated languages. The authors release data, code, and diagnostic tools for future large-scale studies of iconicity.

Abstract: The phenomenon of sound symbolism, the non-arbitrary mapping between word sounds and meanings, has long been demonstrated through anecdotal experiments like Bouba Kiki, but rarely tested at scale. We present the first computational cross-linguistic analysis of sound symbolism in the semantic domain of size. We compile a typologically broad dataset of 810 adjectives (27 languages, 30 words each), each phonemically transcribed and validated with native-speaker audio. Using interpretable classifiers over bag-of-segment features, we find that phonological form predicts size semantics above chance even across unrelated languages, with both vowels and consonants contributing. To probe universality beyond genealogy, we train an adversarial scrubber that suppresses language identity while preserving size signal (also at family granularity). Language prediction averaged across languages and settings falls below chance while size prediction remains significantly above chance, indicating cross-family sound-symbolic bias. We release data, code, and diagnostic tools for future large-scale studies of iconicity.

Method: The benchmark evaluates LLMs on three canonical trading strategies: scheduled trading on MSFT, pairs trading on KO and PEP, and delta hedging on MSFT. Models must produce executable code whose P&L, drawdown, and position paths match reference implementations. Evaluation uses a multi-round pass@k metric that separates structural reliability (whether code runs) from numerical accuracy (mean absolute error of backtest metrics).

Result: Most models reliably execute the simplest strategy (average pass@3 of 0.80), but error rates vary significantly across models and tasks. Gemini 3 Pro and Claude 4.5 Sonnet combine strong reliability with low error on simpler strategies. GPT-5.1 Codex-Max achieves perfect pass@1 on the first two strategies and lowest best-run error on easiest task. Qwen3 Max attains perfect pass@3 but sometimes produces inaccurate P&L paths.

Conclusion: Current LLMs can scaffold basic trading infrastructure but still struggle to reason robustly about prices, inventory, and risk. The authors release MARKET-BENCH and a public leaderboard to facilitate further research and development in this area.

Abstract: We introduce MARKET-BENCH, a benchmark that evaluates large language models (LLMs) on introductory quantitative trading tasks by asking them to construct executable backtesters from natural-language strategy descriptions and market assumptions. Each instance specifies one of three canonical strategies – scheduled trading on Microsoft (NASDAQ: MSFT), pairs trading on Coca-Cola (NASDAQ: KO) and Pepsi (NASDAQ: PEP), or delta hedging on MSFT – and models must produce code whose P&L, drawdown, and position paths match a verifiable reference implementation. We assess twelve state-of-the-art models using a multi-round pass@k metric that separates structural reliability (whether the backtest runs) from numerical accuracy (mean absolute error of the backtest metrics). While most models reliably execute the simplest strategy (average pass@3 of 0.80), errors vary by orders of magnitude across models and tasks: Gemini 3 Pro and Claude 4.5 Sonnet combine strong reliability with low error on simpler strategies, GPT-5.1 Codex-Max achieves perfect pass@1 on the first two strategies and the lowest best-run error on the easiest task, and Qwen3 Max attains perfect pass@3 yet sometimes produces inaccurate P&L paths. These results show that current LLMs can scaffold basic trading infrastructure but still struggle to reason robustly about prices, inventory, and risk; we release MARKET-BENCH and a public leaderboard at https://marketbench.ai.

Method: Keep original weights frozen, append a sub-network to the model as an extension for the textual embedding matrix of the text encoder. Use ConvNeXt module to model co-dependencies between new character-level embeddings, creating a “soft” letter-to-sound layer for Romanian text.

Result: The approach maintains voice cloning capabilities and enables Romanian-English code-switching within the same utterance, though residual English accent characteristics remain. Human evaluation with 20 listeners across three tasks shows successful Romanian speech generation.

Conclusion: The lightweight adapter successfully enables Romanian language support for F5-TTS while preserving existing multilingual capabilities, demonstrating a practical approach for extending TTS models to new languages without full retraining.

Abstract: This work introduces a lightweight input-level adapter for the F5-TTS model that enables Romanian Language support. To preserve the existing capabilities of the model (voice cloning, English and Chinese support), we keep the original weights frozen, append a sub-network to the model and train it as an extension for the textual embedding matrix of the text encoder. For simplicity, we rely on ConvNeXt module implemented in F5-TTS to also model the co-dependencies between the new character-level embeddings. The module serves as a soft letter-to-sound layer, converting Romanian text into a continuous representation that the F5-TTS model uses to produce naturally sounding Romanian utterances. We evaluate the model with a pool of 20 human listeners across three tasks: (a) audio similarity between reference and generated speech, (b) pronunciation and naturalness and (c) Romanian-English code-switching. The results indicate that our approach maintains voice cloning capabilities and enables, to a certain extent, code-switching within the same utterance; however, residual English accent characteristics remain. We open-source our code and provide example audio samples at https://github.com/racai-ro/Ro-F5TTS.

Method: Proposes SCIR framework with Dual-Path Self-Correcting module and feedback-driven optimization for plug-and-play compatibility, plus MBSC dataset with 100K+ entries to distill GPT-4 capabilities into detection models.

Result: Outperforms SOTA IE methods across NER, relation extraction, and event extraction with 5.27% average span-based Micro-F1 improvement while reducing training costs by 87% compared to baselines.

Conclusion: SCIR enhances IE flexibility and accuracy while enabling lightweight, efficient IE paradigms through reduced training costs and better preference alignment.

Abstract: Although Large language Model (LLM)-powered information extraction (IE) systems have shown impressive capabilities, current fine-tuning paradigms face two major limitations: high training costs and difficulties in aligning with LLM preferences. To address these issues, we propose a novel universal IE paradigm, the Self-Correcting Iterative Refinement (SCIR) framework, along with a Multi-task Bilingual (Chinese-English) Self-Correcting (MBSC) dataset containing over 100,000 entries. The SCIR framework achieves plug-and-play compatibility with existing LLMs and IE systems through its Dual-Path Self-Correcting module and feedback-driven optimization, thereby significantly reducing training costs. Concurrently, the MBSC dataset tackles the challenge of preference alignment by indirectly distilling GPT-4’s capabilities into IE result detection models. Experimental results demonstrate that SCIR outperforms state-of-the-art IE methods across three key tasks: named entity recognition, relation extraction, and event extraction, achieving a 5.27 percent average improvement in span-based Micro-F1 while reducing training costs by 87 percent compared to baseline approaches. These advancements not only enhance the flexibility and accuracy of IE systems but also pave the way for lightweight and efficient IE paradigms.

Method: Assessed validity and reliability of metaphor ratings generated by three GPT models for 687 items from Italian Figurative Archive and three English studies. Validation included alignment with human data and prediction of behavioral/electrophysiological responses. Analyzed correlations across languages, metaphor types, and model sizes.

Result: Machine-generated ratings positively correlated with human ratings: familiarity showed moderate-to-strong correlations (weaker for high sensorimotor load metaphors), imageability moderate-to-strong correlations, comprehensibility strongest correlations. Larger models outperformed smaller ones. GPT ratings predicted response times and EEG amplitude comparably to human ratings and were highly stable across sessions.

Conclusion: GPT, especially larger models, can validly and reliably replace or augment human subjects in rating metaphor properties. However, LLMs align worse with humans when dealing with conventionality and multimodal aspects of metaphorical meaning, requiring careful stimulus consideration.

Abstract: As Large Language Models (LLMs) are increasingly being used in scientific research, the issue of their trustworthiness becomes crucial. In psycholinguistics, LLMs have been recently employed in automatically augmenting human-rated datasets, with promising results obtained by generating ratings for single words. Yet, performance for ratings of complex items, i.e., metaphors, is still unexplored. Here, we present the first assessment of the validity and reliability of ratings of metaphors on familiarity, comprehensibility, and imageability, generated by three GPT models for a total of 687 items gathered from the Italian Figurative Archive and three English studies. We performed a thorough validation in terms of both alignment with human data and ability to predict behavioral and electrophysiological responses. We found that machine-generated ratings positively correlated with human-generated ones. Familiarity ratings reached moderate-to-strong correlations for both English and Italian metaphors, although correlations weakened for metaphors with high sensorimotor load. Imageability showed moderate correlations in English and moderate-to-strong in Italian. Comprehensibility for English metaphors exhibited the strongest correlations. Overall, larger models outperformed smaller ones and greater human-model misalignment emerged with familiarity and imageability. Machine-generated ratings significantly predicted response times and the EEG amplitude, with a strength comparable to human ratings. Moreover, GPT ratings obtained across independent sessions were highly stable. We conclude that GPT, especially larger models, can validly and reliably replace - or augment - human subjects in rating metaphor properties. Yet, LLMs align worse with humans when dealing with conventionality and multimodal aspects of metaphorical meaning, calling for careful consideration of the nature of stimuli.

Method: Conceptual analysis proposing paradigm shift in evaluating language knowledge, advocating for new frameworks beyond traditional linguistic tests.

Result: Identifies need to abandon long-held evaluation ideas and recognize the post-Turing test reality where LLMs demonstrate sophisticated language use.

Conclusion: Breakthroughs in language science require accepting LLMs’ language capabilities and developing new evaluation approaches for the post-Turing test era.

Abstract: Acknowledging that large language models have learned to use language can open doors to breakthrough language science. Achieving these breakthroughs may require abandoning some long-held ideas about how language knowledge is evaluated and reckoning with the difficult fact that we have entered a post-Turing test era.

Method: Used VSM13 International Survey and Hofstede’s cultural dimensions to assess cultural alignment of 8 popular LLMs, then tested cultural prompting to shift models toward specific countries (China, France, India, Iran, Japan, US).

Result: Most LLMs default to US cultural alignment; 7 of 8 models successfully shifted toward prompted cultures; models struggled to align with Japan and China despite some being from Chinese company DeepSeek.

Conclusion: Cultural prompting is effective but imperfect, revealing persistent challenges in achieving accurate cultural alignment, particularly for non-Western cultures, highlighting need for improved cultural adaptation in LLMs.

Abstract: Culture is the bedrock of human interaction; it dictates how we perceive and respond to everyday interactions. As the field of human-computer interaction grows via the rise of generative Large Language Models (LLMs), the cultural alignment of these models become an important field of study. This work, using the VSM13 International Survey and Hofstede’s cultural dimensions, identifies the cultural alignment of popular LLMs (DeepSeek-V3, V3.1, GPT-5, GPT-4.1, GPT-4, Claude Opus 4, Llama 3.1, and Mistral Large). We then use cultural prompting, or using system prompts to shift the cultural alignment of a model to a desired country, to test the adaptability of these models to other cultures, namely China, France, India, Iran, Japan, and the United States. We find that the majority of the eight LLMs tested favor the United States when the culture is not specified, with varying results when prompted for other cultures. When using cultural prompting, seven of the eight models shifted closer to the expected culture. We find that models had trouble aligning with Japan and China, despite two of the models tested originating with the Chinese company DeepSeek.

Method: Multi-stage pipeline using expert-guided LLM (Gemini) to generate synthetic data, refined and annotated by native speakers, creating conversational datasets and annotated corpora for foundational NLP tasks.

Result: Fine-tuned XLM-RoBERTa-base achieved 93.81% accuracy on POS tagging and 0.75 F1-Macro on NER, massively outperforming zero-shot baselines. NagaLLaMA (Llama-3.2-3B) achieved Perplexity of 3.85 on conversational tasks, 25x better than few-shot counterpart.

Conclusion: NagaNLP provides foundational resources for Nagamese and a reproducible framework for addressing data scarcity in other low-resource languages, with all datasets, models, and code released as open-source.

Abstract: The vast majority of the world’s languages, particularly creoles like Nagamese, remain severely under-resourced in Natural Language Processing (NLP), creating a significant barrier to their representation in digital technology. This paper introduces NagaNLP, a comprehensive open-source toolkit for Nagamese, bootstrapped through a novel methodology that relies on LLM-driven but human-validated synthetic data generation. We detail a multi-stage pipeline where an expert-guided LLM (Gemini) generates a candidate corpus, which is then refined and annotated by native speakers. This synthetic-hybrid approach yielded a 10K pair conversational dataset and a high-quality annotated corpus for foundational tasks. To assess the effectiveness of our methodology, we trained both discriminative and generative models. Our fine-tuned XLM-RoBERTa-base model establishes a new benchmark for Nagamese, achieving a 93.81% accuracy (0.90 F1-Macro) on Part-of-Speech tagging and a 0.75 F1-Macro on Named Entity Recognition, massively outperforming strong zero-shot baselines. Furthermore, we fine-tuned a Llama-3.2-3B Instruct model, named NagaLLaMA, which demonstrates superior performance on conversational tasks, achieving a Perplexity of 3.85, an order of magnitude improvement over its few-shot counterpart (96.76). We release the NagaNLP toolkit, including all datasets, models, and code, providing a foundational resource for a previously underserved language and a reproducible framework for reducing data scarcity in other low-resource contexts.

Method: Two key innovations: 1) Hypernetwork-based dynamic adaptation that generates request-specific parameters to tailor editing strategy to each instruction, and 2) Difference-aware regularization that focuses supervision on modified spans to prevent over-editing while ensuring precise, minimal changes.

Result: HyperEdit achieves 9%–30% relative improvement in BLEU on modified regions over state-of-the-art baselines, despite using only 3B parameters.

Conclusion: HyperEdit effectively addresses the core challenges of instruction-based text editing by enabling better intent alignment and preventing over-editing through specialized architectural and training innovations.

Abstract: Instruction-based text editing is increasingly critical for real-world applications such as code editors (e.g., Cursor), but Large Language Models (LLMs) continue to struggle with this task. Unlike free-form generation, editing requires faithfully implementing user instructions while preserving unchanged content, as even minor unintended modifications can break functionality. Existing approaches treat editing as generic text generation, leading to two key failures: they struggle to faithfully align edits with diverse user intents, and they often over-edit unchanged regions. We propose HyperEdit to address both issues. First, we introduce hypernetwork-based dynamic adaptation that generates request-specific parameters, enabling the model to tailor its editing strategy to each instruction. Second, we develop difference-aware regularization that focuses supervision on modified spans, preventing over-editing while ensuring precise, minimal changes. HyperEdit achieves a 9%–30% relative improvement in BLEU on modified regions over state-of-the-art baselines, despite utilizing only 3B parameters.

Method: CoVRL bridges variational inference and reinforcement learning by coupling prior and posterior distributions through a hybrid sampling strategy. It constructs and optimizes a composite distribution that integrates these two distributions to enable efficient exploration while preserving thought-answer coherence.

Result: Extensive experiments on mathematical and general reasoning benchmarks show CoVRL improves performance by 12.4% over the base model and achieves an additional 2.3% improvement over strong state-of-the-art verifier-free RL baselines.

Conclusion: CoVRL provides a principled framework for enhancing the general reasoning capabilities of language models by addressing the exploration-efficiency and coherence problems in verifier-free RL approaches.

Abstract: While reinforcement learning have achieved impressive progress in language model reasoning, they are constrained by the requirement for verifiable rewards. Recent verifier-free RL methods address this limitation by utilizing the intrinsic probabilities of LLMs generating reference answers as reward signals. However, these approaches typically sample reasoning traces conditioned only on the question. This design decouples reasoning-trace sampling from answer information, leading to inefficient exploration and incoherence between traces and final answers. In this paper, we propose \textit{\b{Co}upled \b{V}ariational \b{R}einforcement \b{L}earning} (CoVRL), which bridges variational inference and reinforcement learning by coupling prior and posterior distributions through a hybrid sampling strategy. By constructing and optimizing a composite distribution that integrates these two distributions, CoVRL enables efficient exploration while preserving strong thought-answer coherence. Extensive experiments on mathematical and general reasoning benchmarks show that CoVRL improves performance by 12.4% over the base model and achieves an additional 2.3% improvement over strong state-of-the-art verifier-free RL baselines, providing a principled framework for enhancing the general reasoning capabilities of language models.

Method: Two complementary mechanisms: 1) Obvious Record - explicitly stores cause-result/question-solution relationships as symbolic memory for persistent learning from single encounters; 2) Maximum-Entropy Method Discovery - prioritizes methods with high semantic dissimilarity to capture diverse, underrepresented strategies.

Result: On a benchmark of 60 semantically diverse question-solution pairs, the entropy-guided approach achieves stronger coverage of unseen questions and significantly greater internal diversity than random baseline, confirming effectiveness in discovering more generalizable methods.

Conclusion: The proposed human-inspired framework enables LLMs to better handle rare/unseen scenarios by combining explicit symbolic memory with entropy-based method discovery, moving beyond intuition-driven prediction toward deliberate, method-oriented learning.

Abstract: Large Language Models (LLMs) excel at extracting common patterns from large-scale corpora, yet they struggle with rare, low-resource, or previously unseen scenarios-such as niche hardware deployment issues or irregular IoT device behaviors-because such cases are sparsely represented in training data. Moreover, LLMs rely primarily on implicit parametric memory, which limits their ability to explicitly acquire, recall, and refine methods, causing them to behave predominantly as intuition-driven predictors rather than deliberate, method-oriented learners. Inspired by how humans learn from rare experiences, this paper proposes a human-inspired learning framework that integrates two complementary mechanisms. The first, Obvious Record, explicitly stores cause–result (or question–solution) relationships as symbolic memory, enabling persistent learning even from single or infrequent encounters. The second, Maximum-Entropy Method Discovery, prioritizes and preserves methods with high semantic dissimilarity, allowing the system to capture diverse and underrepresented strategies that are typically overlooked by next-token prediction. Verification on a benchmark of 60 semantically diverse question–solution pairs demonstrates that the proposed entropy-guided approach achieves stronger coverage of unseen questions and significantly greater internal diversity than a random baseline, confirming its effectiveness in discovering more generalizable and human-inspired methods.

Method: Proposes StruProKGR with two key components: 1) distance-guided path collection mechanism to reduce computational costs while exploring relevant paths, and 2) probabilistic path aggregation that incorporates structural information by prioritizing paths that reinforce each other.

Result: Extensive experiments on five sparse KG reasoning benchmarks show StruProKGR surpasses existing path-based methods in both effectiveness and efficiency.

Conclusion: StruProKGR provides an effective, efficient, and interpretable solution for sparse KG reasoning by addressing computational inefficiency and structural awareness limitations of existing path-based methods.

Abstract: Sparse Knowledge Graphs (KGs) are commonly encountered in real-world applications, where knowledge is often incomplete or limited. Sparse KG reasoning, the task of inferring missing knowledge over sparse KGs, is inherently challenging due to the scarcity of knowledge and the difficulty of capturing relational patterns in sparse scenarios. Among all sparse KG reasoning methods, path-based ones have attracted plenty of attention due to their interpretability. Existing path-based methods typically rely on computationally intensive random walks to collect paths, producing paths of variable quality. Additionally, these methods fail to leverage the structured nature of graphs by treating paths independently. To address these shortcomings, we propose a Structural and Probabilistic framework named StruProKGR, tailored for efficient and interpretable reasoning on sparse KGs. StruProKGR utilizes a distance-guided path collection mechanism to significantly reduce computational costs while exploring more relevant paths. It further enhances the reasoning process by incorporating structural information through probabilistic path aggregation, which prioritizes paths that reinforce each other. Extensive experiments on five sparse KG reasoning benchmarks reveal that StruProKGR surpasses existing path-based methods in both effectiveness and efficiency, providing an effective, efficient, and interpretable solution for sparse KG reasoning.

Method: Used 14 large language models to study syllogistic reasoning capabilities through symbolic inferences and natural language understanding tests.

Result: Syllogistic reasoning is not uniformly emergent across LLMs, but some models achieve perfect symbolic performance, suggesting they may be developing formal reasoning mechanisms.

Conclusion: The perfect symbolic reasoning in certain LLMs raises questions about whether they’re evolving into formal reasoning systems rather than explicitly modeling the nuances of human reasoning.

Abstract: We study syllogistic reasoning in LLMs from the logical and natural language perspectives. In process, we explore fundamental reasoning capabilities of the LLMs and the direction this research is moving forward. To aid in our studies, we use 14 large language models and investigate their syllogistic reasoning capabilities in terms of symbolic inferences as well as natural language understanding. Even though this reasoning mechanism is not a uniform emergent property across LLMs, the perfect symbolic performances in certain models make us wonder whether LLMs are becoming more and more formal reasoning mechanisms, rather than making explicit the nuances of human reasoning.

Method: Provides clear implementation guide and parameter choices for Unigram tokenization. Identifies a simpler alternative algorithm that accepts slightly higher training loss for better compression.

Result: Bridges the theory-practice gap for Unigram tokenization, making it more accessible. The simpler algorithm offers improved compression at the cost of slightly higher training loss.

Conclusion: Unigram tokenization can be made more practical through better implementation guidance and alternative algorithms that balance training loss and compression efficiency.

Abstract: The Unigram tokenization algorithm offers a probabilistic alternative to the greedy heuristics of Byte-Pair Encoding. Despite its theoretical elegance, its implementation in practice is complex, limiting its adoption to the SentencePiece package and adapters thereof. We bridge this gap between theory and practice by providing a clear guide to implementation and parameter choices. We also identify a simpler algorithm that accepts slightly higher training loss in exchange for improved compression.

Method: 1) Created a comprehensive schema with hierarchical taxonomy and argument definitions for legal relations in Chinese civil cases; 2) Formulated legal relation extraction task; 3) Built LexRel, an expert-annotated benchmark; 4) Evaluated state-of-the-art LLMs on the benchmark; 5) Tested incorporation of legal relations into downstream legal AI tasks.

Result: Current LLMs show significant limitations in accurately identifying civil legal relations. However, incorporating legal relations information leads to consistent performance gains on other downstream legal AI tasks.

Conclusion: The paper establishes foundational work for legal relation extraction in Chinese civil law, highlighting both the challenges for current AI systems and the value of legal relations for improving legal AI applications.

Abstract: Legal relations form a highly consequential analytical framework of civil law system, serving as a crucial foundation for resolving disputes and realizing values of the rule of law in judicial practice. However, legal relations in Chinese civil cases remain underexplored in the field of legal artificial intelligence (legal AI), largely due to the absence of comprehensive schemas. In this work, we firstly introduce a comprehensive schema, which contains a hierarchical taxonomy and definitions of arguments, for AI systems to capture legal relations in Chinese civil cases. Based on this schema, we then formulate legal relation extraction task and present LexRel, an expert-annotated benchmark for legal relation extraction in Chinese civil law. We use LexRel to evaluate state-of-the-art large language models (LLMs) on legal relation extractions, showing that current LLMs exhibit significant limitations in accurately identifying civil legal relations. Furthermore, we demonstrate that incorporating legal relations information leads to consistent performance gains on other downstream legal AI tasks.

Method: Proposes a graph-based framework representing Urdu novels as character interaction networks (nodes=characters, edges=co-occurrence within narrative proximity). Compares multiple graph representations: global structural features, node-level semantic summaries, unsupervised graph embeddings, and supervised graph neural networks.

Result: Experiments on 52 Urdu novels by seven authors show that learned graph representations substantially outperform hand-crafted and unsupervised baselines, achieving up to 0.857 accuracy under strict author-aware evaluation protocol.

Conclusion: Narrative structure captured through character interaction networks provides effective signals for authorship attribution in Urdu literature, demonstrating that higher-level structural patterns can reveal authorial style beyond traditional lexical features.

Abstract: Authorship analysis has traditionally focused on lexical and stylistic cues within text, while higher-level narrative structure remains underexplored, particularly for low-resource languages such as Urdu. This work proposes a graph-based framework that models Urdu novels as character interaction networks to examine whether authorial style can be inferred from narrative structure alone. Each novel is represented as a graph where nodes correspond to characters and edges denote their co-occurrence within narrative proximity. We systematically compare multiple graph representations, including global structural features, node-level semantic summaries, unsupervised graph embeddings, and supervised graph neural networks. Experiments on a dataset of 52 Urdu novels written by seven authors show that learned graph representations substantially outperform hand-crafted and unsupervised baselines, achieving up to 0.857 accuracy under a strict author-aware evaluation protocol.

Method: Two approaches: (1) embedding-based method attaching classification head to final token embedding of pre-trained causal LLM, (2) instruction-tuning LLM in prompt->response format. Both use 4-bit model quantization with LoRA for parameter-efficient training on single GPU for models up to 8B parameters.

Result: Embedding-based method significantly outperforms instruction-tuned method in F1-score, and is competitive with/surpasses fine-tuned domain-specific models (e.g., BERT) on both proprietary single-label dataset and WIPO-Alpha patent dataset (extreme multi-label classification).

Conclusion: Directly leveraging internal representations of causal LLMs with efficient fine-tuning techniques yields impressive classification performance under limited resources. The embedding approach is superior to instruction-tuning for classification tasks, offering practical guidelines for LLM fine-tuning optimization.

Abstract: We explore efficient strategies to fine-tune decoder-only Large Language Models (LLMs) for downstream text classification under resource constraints. Two approaches are investigated: (1) attaching a classification head to a pre-trained causal LLM and fine-tuning on the task (using the LLM’s final token embedding as a sequence representation), and (2) instruction-tuning the LLM in a prompt->response format for classification. To enable single-GPU fine-tuning of models up to 8B parameters, we combine 4-bit model quantization with Low-Rank Adaptation (LoRA) for parameter-efficient training. Experiments on two datasets - a proprietary single-label dataset and the public WIPO-Alpha patent dataset (extreme multi-label classification) - show that the embedding-based method significantly outperforms the instruction-tuned method in F1-score, and is very competitive with - even surpassing - fine-tuned domain-specific models (e.g. BERT) on the same tasks. These results demonstrate that directly leveraging the internal representations of causal LLMs, along with efficient fine-tuning techniques, yields impressive classification performance under limited computational resources. We discuss the advantages of each approach while outlining practical guidelines and future directions for optimizing LLM fine-tuning in classification scenarios.

Method: CoDA uses a single shared LLM backbone operating in two contextually isolated roles: a high-level Planner for task decomposition in concise strategic context, and a low-level Executor for tool interactions in ephemeral isolated workspace. Trained end-to-end with PECO (Planner-Executor Co-Optimization) RL methodology using trajectory-level rewards.

Result: Significant performance improvements over SOTA baselines on complex multi-hop QA benchmarks, strong robustness in long-context scenarios with stable performance while other baselines suffer severe degradation.

Conclusion: The hierarchical design effectively mitigates context overload, demonstrating that decoupling planning from execution with context isolation solves the context explosion problem in LLM agents.

Abstract: Large Language Model (LLM) agents trained with reinforcement learning (RL) show great promise for solving complex, multi-step tasks. However, their performance is often crippled by “Context Explosion”, where the accumulation of long text outputs overwhelms the model’s context window and leads to reasoning failures. To address this, we introduce CoDA, a Context-Decoupled hierarchical Agent, a simple but effective reinforcement learning framework that decouples high-level planning from low-level execution. It employs a single, shared LLM backbone that learns to operate in two distinct, contextually isolated roles: a high-level Planner that decomposes tasks within a concise strategic context, and a low-level Executor that handles tool interactions in an ephemeral, isolated workspace. We train this unified agent end-to-end using PECO (Planner-Executor Co-Optimization), a reinforcement learning methodology that applies a trajectory-level reward to jointly optimize both roles, fostering seamless collaboration through context-dependent policy updates. Extensive experiments demonstrate that CoDA achieves significant performance improvements over state-of-the-art baselines on complex multi-hop question-answering benchmarks, and it exhibits strong robustness in long-context scenarios, maintaining stable performance while all other baselines suffer severe degradation, thus further validating the effectiveness of our hierarchical design in mitigating context overload.

Method: NL2Repo Bench provides agents with only a single natural-language requirements document and an empty workspace. Agents must autonomously design architecture, manage dependencies, implement multi-module logic, and produce a fully installable Python library.

Result: Experiments show long-horizon repository generation remains largely unsolved: even top agents achieve below 40% average test pass rates and rarely complete entire repositories correctly. Analysis reveals fundamental failure modes including premature termination, loss of global coherence, fragile cross-file dependencies, and inadequate planning over hundreds of steps.

Conclusion: NL2Repo Bench establishes a rigorous testbed for measuring sustained agentic competence and identifies long-horizon reasoning as a central bottleneck for next-generation autonomous coding agents.

Abstract: Recent advances in coding agents suggest rapid progress toward autonomous software development, yet existing benchmarks fail to rigorously evaluate the long-horizon capabilities required to build complete software systems. Most prior evaluations focus on localized code generation, scaffolded completion, or short-term repair tasks, leaving open the question of whether agents can sustain coherent reasoning, planning, and execution over the extended horizons demanded by real-world repository construction. To address this gap, we present NL2Repo Bench, a benchmark explicitly designed to evaluate the long-horizon repository generation ability of coding agents. Given only a single natural-language requirements document and an empty workspace, agents must autonomously design the architecture, manage dependencies, implement multi-module logic, and produce a fully installable Python library. Our experiments across state-of-the-art open- and closed-source models reveal that long-horizon repository generation remains largely unsolved: even the strongest agents achieve below 40% average test pass rates and rarely complete an entire repository correctly. Detailed analysis uncovers fundamental long-horizon failure modes, including premature termination, loss of global coherence, fragile cross-file dependencies, and inadequate planning over hundreds of interaction steps. NL2Repo Bench establishes a rigorous, verifiable testbed for measuring sustained agentic competence and highlights long-horizon reasoning as a central bottleneck for the next generation of autonomous coding agents.

Method: Created Curió 7B by continued pretraining of LLaMA-2 on 100B Portuguese tokens from ClassiCC-PT corpus. Developed variant Curió-Edu 7B trained exclusively on educational/STEM-filtered subset (10B tokens). Compared performance between full-corpus model and quality-filtered variant.

Result: Curió-Edu 7B, despite using only 10% of the data and 20% of the computation, surpassed the full-corpus model in evaluations, demonstrating that data selection/quality is fundamental even when adapting models with limited prior exposure to the target language.

Conclusion: Data quality plays a decisive role in linguistic adaptation through continued pretraining, with carefully selected educational/STEM data outperforming larger but less curated datasets, suggesting quality over quantity approach for efficient model adaptation.

Abstract: Continued pretraining extends a language model’s capabilities by further exposing it to additional data, often tailored to a specific linguistic or domain context. This strategy has emerged as an efficient alternative to full retraining when adapting general-purpose models to new settings. In this work, we investigate this paradigm through Curió 7B, a 7-billion-parameter model derived from LLaMA-2 and trained on 100 billion Portuguese tokens from the ClassiCC-PT corpus - the most extensive Portuguese-specific continued-pretraining effort above the three-billion-parameter scale to date. Beyond scale, we investigate whether quantity alone suffices or whether data quality plays a decisive role in linguistic adaptation. To this end, we introduce Curió-Edu 7B, a variant trained exclusively on the educational and STEM-filtered subset of the same corpus, totaling just 10 billion tokens. Despite using only 10% of the data and 20% of the computation, Curió-Edu 7B surpasses the full-corpus model in our evaluations, demonstrating that data selection can be fundamental even when adapting models with limited prior exposure to the target language. The developed models are available at https://huggingface.co/collections/ClassiCC-Corpus/curio-edu

Method: Introduced an evaluation protocol combining long persona dialogues (100+ rounds) with evaluation datasets to create dialogue-conditioned benchmarks. Tested seven state-of-the-art open- and closed-weight LLMs on persona fidelity, instruction-following, and safety over extended dialogues.

Result: Persona fidelity degrades over long dialogues, especially in goal-oriented conversations requiring both persona fidelity and instruction following. Found a trade-off between fidelity and instruction following, with non-persona baselines initially outperforming persona models. As fidelity fades, persona responses become increasingly similar to baseline responses.

Conclusion: Persona applications are fragile in extended interactions, highlighting the need for better evaluation protocols like the one proposed to systematically measure such failures in long-context scenarios.

Abstract: Persona-assigned large language models (LLMs) are used in domains such as education, healthcare, and sociodemographic simulation. Yet, they are typically evaluated only in short, single-round settings that do not reflect real-world usage. We introduce an evaluation protocol that combines long persona dialogues (over 100 rounds) and evaluation datasets to create dialogue-conditioned benchmarks that can robustly measure long-context effects. We then investigate the effects of dialogue length on persona fidelity, instruction-following, and safety of seven state-of-the-art open- and closed-weight LLMs. We find that persona fidelity degrades over the course of dialogues, especially in goal-oriented conversations, where models must sustain both persona fidelity and instruction following. We identify a trade-off between fidelity and instruction following, with non-persona baselines initially outperforming persona-assigned models; as dialogues progress and fidelity fades, persona responses become increasingly similar to baseline responses. Our findings highlight the fragility of persona applications in extended interactions and our work provides a protocol to systematically measure such failures.

Method: Introduces the State over Tokens (SoT) conceptual framework that reframes reasoning tokens not as linguistic narratives but as externalized computational state - the sole persistent information carrier across the model’s stateless generation cycles.

Result: The SoT framework explains how reasoning tokens can drive correct reasoning without being faithful explanations when read as text, and surfaces previously overlooked research questions about these tokens.

Conclusion: To truly understand LLM reasoning processes, research must move beyond reading reasoning tokens as text and focus on decoding them as state. The SoT framework provides the conceptual shift needed for this research direction.

Abstract: Large Language Models (LLMs) can generate reasoning tokens before their final answer to boost performance on complex tasks. While these sequences seem like human thought processes, empirical evidence reveals that they are not a faithful explanation of the model’s actual reasoning process. To address this gap between appearance and function, we introduce the State over Tokens (SoT) conceptual framework. SoT reframes reasoning tokens not as a linguistic narrative, but as an externalized computational state – the sole persistent information carrier across the model’s stateless generation cycles. This explains how the tokens can drive correct reasoning without being a faithful explanation when read as text and surfaces previously overlooked research questions on these tokens. We argue that to truly understand the process that LLMs do, research must move beyond reading the reasoning tokens as text and focus on decoding them as state.

Method: Proposed evaluation framework using MMMLU benchmark to test three LLMs (GPT-4o mini, Gemini 2.0 Flash, Llama 4 Scout) under Very Friendly, Neutral, and Very Rude prompt variants across six STEM and Humanities tasks, with statistical significance testing of pairwise accuracy differences.

Result: Tone sensitivity varies by model and domain: Neutral/Friendly prompts generally outperform Rude ones, but statistically significant effects appear only in Humanities tasks (rude tone reduces accuracy for GPT and Llama, while Gemini remains tone-insensitive). Aggregated across domains, tone effects diminish and lose significance.

Conclusion: While interaction tone matters in specific interpretive settings, modern LLMs are broadly robust to tonal variation in typical mixed-domain use, providing practical guidance for prompt design and model selection in real-world deployments.

Abstract: Prompt engineering has emerged as a critical factor influencing large language model (LLM) performance, yet the impact of pragmatic elements such as linguistic tone and politeness remains underexplored, particularly across different model families. In this work, we propose a systematic evaluation framework to examine how interaction tone affects model accuracy and apply it to three recently released and widely available LLMs: GPT-4o mini (OpenAI), Gemini 2.0 Flash (Google DeepMind), and Llama 4 Scout (Meta). Using the MMMLU benchmark, we evaluate model performance under Very Friendly, Neutral, and Very Rude prompt variants across six tasks spanning STEM and Humanities domains, and analyze pairwise accuracy differences with statistical significance testing. Our results show that tone sensitivity is both model-dependent and domain-specific. Neutral or Very Friendly prompts generally yield higher accuracy than Very Rude prompts, but statistically significant effects appear only in a subset of Humanities tasks, where rude tone reduces accuracy for GPT and Llama, while Gemini remains comparatively tone-insensitive. When performance is aggregated across tasks within each domain, tone effects diminish and largely lose statistical significance. Compared with earlier researches, these findings suggest that dataset scale and coverage materially influence the detection of tone effects. Overall, our study indicates that while interaction tone can matter in specific interpretive settings, modern LLMs are broadly robust to tonal variation in typical mixed-domain use, providing practical guidance for prompt design and model selection in real-world deployments.

Method: Hindsight organizes agent memory into four structured networks: world facts, agent experiences, synthesized entity summaries, and evolving beliefs. It implements three core operations (retain, recall, reflect) with a temporal, entity-aware memory layer that turns conversations into structured memory, and a reflection layer for traceable reasoning.

Result: On LongMemEval and LoCoMo benchmarks, Hindsight with a 20B model improved accuracy from 39% to 83.6% over full-context baseline, outperforming GPT-4o. Scaling further achieved 91.4% on LongMemEval and 89.61% on LoCoMo, consistently beating existing memory architectures.

Conclusion: Hindsight demonstrates that treating memory as a structured, first-class reasoning substrate enables superior long-horizon conversational performance and explainable agent reasoning, representing a significant advancement over current memory architectures.

Abstract: Agent memory has been touted as a dimension of growth for LLM-based applications, enabling agents that can accumulate experience, adapt across sessions, and move beyond single-shot question answering. The current generation of agent memory systems treats memory as an external layer that extracts salient snippets from conversations, stores them in vector or graph-based stores, and retrieves top-k items into the prompt of an otherwise stateless model. While these systems improve personalization and context carry-over, they still blur the line between evidence and inference, struggle to organize information over long horizons, and offer limited support for agents that must explain their reasoning. We present Hindsight, a memory architecture that treats agent memory as a structured, first-class substrate for reasoning by organizing it into four logical networks that distinguish world facts, agent experiences, synthesized entity summaries, and evolving beliefs. This framework supports three core operations – retain, recall, and reflect – that govern how information is added, accessed, and updated. Under this abstraction, a temporal, entity aware memory layer incrementally turns conversational streams into a structured, queryable memory bank, while a reflection layer reasons over this bank to produce answers and to update information in a traceable way. On key long-horizon conversational memory benchmarks like LongMemEval and LoCoMo, Hindsight with an open-source 20B model lifts overall accuracy from 39% to 83.6% over a full-context baseline with the same backbone and outperforms full context GPT-4o. Scaling the backbone further pushes Hindsight to 91.4% on LongMemEval and up to 89.61% on LoCoMo (vs. 75.78% for the strongest prior open system), consistently outperforming existing memory architectures on multi-session and open-domain questions.

Method: Created LongStoryEval benchmark with 600 newly published books (avg 121K tokens), analyzed reader reviews to propose evaluation criteria structure, compared three evaluation methods (aggregation-based, incremental-updated, summary-based), and developed NovelCritique - an 8B model using summary-based framework.

Result: Identified 8 top-level evaluation criteria from user feedback, found aggregation- and summary-based evaluations perform best (aggregation excels in detail assessment, summary offers efficiency), and NovelCritique outperforms GPT-4o in aligning with human evaluations.

Conclusion: The study provides the first large-scale benchmark for book-length story evaluation, establishes effective evaluation methods, and demonstrates that specialized models like NovelCritique can outperform general-purpose commercial models in narrative assessment tasks.

Abstract: In this work, we conduct systematic research in a challenging area: the automatic evaluation of book-length stories (>100K tokens). Our study focuses on two key questions: (1) understanding which evaluation aspects matter most to readers, and (2) exploring effective methods for evaluating lengthy stories. We introduce the first large-scale benchmark, LongStoryEval, comprising 600 newly published books with an average length of 121K tokens (maximum 397K). Each book includes its average rating and multiple reader reviews, presented as critiques organized by evaluation aspects. By analyzing all user-mentioned aspects, we propose an evaluation criteria structure and conduct experiments to identify the most significant aspects among the 8 top-level criteria. For evaluation methods, we compare the effectiveness of three types: aggregation-based, incremental-updated, and summary-based evaluations. Our findings reveal that aggregation- and summary-based evaluations perform better, with the former excelling in detail assessment and the latter offering greater efficiency. Building on these insights, we further propose NovelCritique, an 8B model that leverages the efficient summary-based framework to review and score stories across specified aspects. NovelCritique outperforms commercial models like GPT-4o in aligning with human evaluations. Our datasets and codes are available at https://github.com/DingyiYang/LongStoryEval.

Method: Introduces Frequency-Based Probabilistic Ranker (FBPR), a lightweight method using smoothed Naive Bayes over concept-diagnosis co-occurrence statistics from large pretraining corpora. Compares FBPR performance with corresponding LLMs (OLMo and Llama) on MedQA diagnosis questions.

Result: FBPR achieves comparable performance to LLMs when using co-occurrence statistics from the same pretraining corpora. However, LLMs and FBPR get different questions correct with only slightly above random chance overlap, indicating complementary strengths. Simple frequency aggregation accounts for substantial benchmark performance.

Conclusion: LLMs’ performance on clinical diagnosis benchmarks isn’t driven by simple frequency aggregation alone. Explicit probabilistic baselines like FBPR provide valuable reference points and complementary signals for potential hybridization with LLMs, showing that low-complexity expert systems still account for substantial benchmark performance.

Abstract: Large language models (LLMs) excel on multiple-choice clinical diagnosis benchmarks, yet it is unclear how much of this performance reflects underlying probabilistic reasoning. We study this through questions from MedQA, where the task is to select the most likely diagnosis. We introduce the Frequency-Based Probabilistic Ranker (FBPR), a lightweight method that scores options with a smoothed Naive Bayes over concept-diagnosis co-occurrence statistics from a large corpus. When co-occurrence statistics were sourced from the pretraining corpora for OLMo and Llama, FBPR achieves comparable performance to the corresponding LLMs pretrained on that same corpus. Direct LLM inference and FBPR largely get different questions correct, with an overlap only slightly above random chance, indicating complementary strengths of each method. These findings highlight the continued value of explicit probabilistic baselines: they provide a meaningful performance reference point and a complementary signal for potential hybridization. While the performance of LLMs seems to be driven by a mechanism other than simple frequency aggregation, we show that an approach similar to the historically grounded, low-complexity expert systems still accounts for a substantial portion of benchmark performance.

Method: Human-AI collaborative multi-agent framework where AI agents handle specific repetitive tasks (OCR, text segmentation, semantic alignment, initial terminology extraction) while human experts provide oversight, review, and supervision with contextual knowledge and legal judgment throughout the entire process from document preprocessing to quality assurance.

Result: Framework tested on trilingual parallel corpus of 35 key Chinese statutes with English/Japanese translations shows improved precision and consistency of multilingual legal terminology mapping with greater scalability compared to traditional manual methods.

Conclusion: Human-in-the-loop multi-agent workflow effectively addresses legal terminology mapping challenges by combining AI efficiency with human expertise, offering a scalable solution for multilingual legal terminology databases.

Abstract: Accurately mapping legal terminology across languages remains a significant challenge, especially for language pairs like Chinese and Japanese, which share a large number of homographs with different meanings. Existing resources and standardized tools for these languages are limited. To address this, we propose a human-AI collaborative approach for building a multilingual legal terminology database, based on a multi-agent framework. This approach integrates advanced large language models and legal domain experts throughout the entire process-from raw document preprocessing, article-level alignment, to terminology extraction, mapping, and quality assurance. Unlike a single automated pipeline, our approach places greater emphasis on how human experts participate in this multi-agent system. Humans and AI agents take on different roles: AI agents handle specific, repetitive tasks, such as OCR, text segmentation, semantic alignment, and initial terminology extraction, while human experts provide crucial oversight, review, and supervise the outputs with contextual knowledge and legal judgment. We tested the effectiveness of this framework using a trilingual parallel corpus comprising 35 key Chinese statutes, along with their English and Japanese translations. The experimental results show that this human-in-the-loop, multi-agent workflow not only improves the precision and consistency of multilingual legal terminology mapping but also offers greater scalability compared to traditional manual methods.

Method: Three key innovations: (1) Long-Context Data Synthesis Pipeline that generates challenging reasoning tasks by deconstructing documents into atomic facts and composing verifiable questions; (2) Stabilized RL with task-balanced sampling and Adaptive Entropy-Controlled Policy Optimization (AEPO); (3) Memory-Augmented Architecture with multi-stage fusion RL training for ultra-long contexts beyond 4M tokens.

Result: Achieves performance comparable to GPT-5 and Gemini-2.5-Pro on long-context reasoning benchmarks, with 9.90-point average improvement over baseline. On ultra-long tasks (1M~4M tokens), memory-agent framework yields 9.48-point gain. Also shows enhanced performance in general domains like scientific reasoning and extended dialogue.

Conclusion: QwenLong-L1.5 demonstrates that systematic post-training innovations can achieve state-of-the-art long-context reasoning capabilities, with the developed techniques enabling genuine multi-hop reasoning over globally distributed evidence and effective handling of ultra-long sequences through memory augmentation.

Abstract: We introduce QwenLong-L1.5, a model that achieves superior long-context reasoning capabilities through systematic post-training innovations. The key technical breakthroughs of QwenLong-L1.5 are as follows: (1) Long-Context Data Synthesis Pipeline: We develop a systematic synthesis framework that generates challenging reasoning tasks requiring multi-hop grounding over globally distributed evidence. By deconstructing documents into atomic facts and their underlying relationships, and then programmatically composing verifiable reasoning questions, our approach creates high-quality training data at scale, moving substantially beyond simple retrieval tasks to enable genuine long-range reasoning capabilities. (2) Stabilized Reinforcement Learning for Long-Context Training: To overcome the critical instability in long-context RL, we introduce task-balanced sampling with task-specific advantage estimation to mitigate reward bias, and propose Adaptive Entropy-Controlled Policy Optimization (AEPO) that dynamically regulates exploration-exploitation trade-offs. (3) Memory-Augmented Architecture for Ultra-Long Contexts: Recognizing that even extended context windows cannot accommodate arbitrarily long sequences, we develop a memory management framework with multi-stage fusion RL training that seamlessly integrates single-pass reasoning with iterative memory-based processing for tasks exceeding 4M tokens. Based on Qwen3-30B-A3B-Thinking, QwenLong-L1.5 achieves performance comparable to GPT-5 and Gemini-2.5-Pro on long-context reasoning benchmarks, surpassing its baseline by 9.90 points on average. On ultra-long tasks (1M~4M tokens), QwenLong-L1.5’s memory-agent framework yields a 9.48-point gain over the agent baseline. Additionally, the acquired long-context reasoning ability translates to enhanced performance in general domains like scientific reasoning, memory tool using, and extended dialogue.

Method: Collaborated with commercial chatbot (10,000+ users) to deploy author labeling system that identifies task-relevant queries, generates on-the-fly labeling questions, and records authors’ answers in real time. Trained online-learning model for product recommendation that continuously improves from author labeling.

Result: Author labeling achieved 534% increase in click-through rate compared to industry advertising baseline. Compared to three traditional annotation approaches for sentiment analysis, author labeling was higher quality, faster to acquire, and cheaper.

Conclusion: Author labeling produces significantly higher quality annotations for egocentric and subjective beliefs than third-party annotation. Authors should label their own data when possible, especially for subjective features. Released author labeling service at academic.echollm.io for research community.

Abstract: The status quo for labeling text is third-party annotation, but there are many cases where information directly from the document’s source would be preferable over a third-person proxy, especially for egocentric features like sentiment and belief. We introduce author labeling, an annotation technique where the writer of the document itself annotates the data at the moment of creation. We collaborate with a commercial chatbot with over 10,000 users to deploy an author labeling annotation system for subjective features related to product recommendation. This system identifies task-relevant queries, generates on-the-fly labeling questions, and records authors’ answers in real time. We train and deploy an online-learning model architecture for product recommendation that continuously improves from author labeling and find it achieved a 534% increase in click-through rate compared to an industry advertising baseline running concurrently. We then compare the quality and practicality of author labeling to three traditional annotation approaches for sentiment analysis and find author labeling to be higher quality, faster to acquire, and cheaper. These findings reinforce existing literature that annotations, especially for egocentric and subjective beliefs, are significantly higher quality when labeled by the author rather than a third party. To facilitate broader scientific adoption, we release an author labeling service for the research community at academic.echollm.io.

Method: Combines open-source LLM with open web search API for iterative retrieval, reasoning, and synthesis. Enhances reasoning quality through preference tuning based on LLM-as-a-judge feedback evaluating clarity, insightfulness, and factuality.

Result: The proposed method consistently improves answer quality across all three aspects (clarity, insightfulness, factuality). The system was selected as a winning system in the text-to-text track of the MMU-RAG competition at NeurIPS 2025.

Conclusion: The open deep research system demonstrates effective long-form QA capabilities through the combination of iterative retrieval with web search and preference tuning, with publicly available source code for community use.

Abstract: We present an open deep research system for long-form question answering, selected as a winning system in the text-to-text track of the MMU-RAG competition at NeurIPS 2025. The system combines an open-source large language model (LLM) with an open web search API to perform iterative retrieval, reasoning, and synthesis in real-world open-domain settings. To enhance reasoning quality, we apply preference tuning based on LLM-as-a-judge feedback that evaluates multiple aspects, including clarity, insightfulness, and factuality. Our experimental results show that the proposed method consistently improves answer quality across all three aspects. Our source code is publicly available at https://github.com/efficient-deep-research/efficient-deep-research.

Method: Developed a mathematical framework using hyperbolic tangent curves to model concession dynamics, created burstiness tau and Concession-Rigidity Index metrics, and conducted large-scale empirical comparisons between humans and four state-of-the-art LLMs across multiple negotiation settings and power-asymmetry scenarios.

Result: LLMs systematically anchor at extremes of agreement zones, optimize for fixed points regardless of leverage or context, show limited strategy diversity, use occasional deceptive tactics, and their negotiation ability doesn’t improve with better models - unlike humans who smoothly adapt and infer opponent positions.

Conclusion: Current LLMs have fundamental limitations in negotiation capabilities, lacking the ability to internalize opponent reasoning and context-dependent strategy, pointing to the need for more sophisticated models that can better emulate human negotiation dynamics.

Abstract: Bilateral negotiation is a complex, context-sensitive task in which human negotiators dynamically adjust anchors, pacing, and flexibility to exploit power asymmetries and informal cues. We introduce a unified mathematical framework for modeling concession dynamics based on a hyperbolic tangent curve, and propose two metrics burstiness tau and the Concession-Rigidity Index (CRI) to quantify the timing and rigidity of offer trajectories. We conduct a large-scale empirical comparison between human negotiators and four state-of-the-art large language models (LLMs) across natural-language and numeric-offers settings, with and without rich market context, as well as six controlled power-asymmetry scenarios. Our results reveal that, unlike humans who smoothly adapt to situations and infer the opponents position and strategies, LLMs systematically anchor at extremes of the possible agreement zone for negotiations and optimize for fixed points irrespective of leverage or context. Qualitative analysis further shows limited strategy diversity and occasional deceptive tactics used by LLMs. Moreover the ability of LLMs to negotiate does not improve with better models. These findings highlight fundamental limitations in current LLM negotiation capabilities and point to the need for models that better internalize opponent reasoning and context-dependent strategy.

Method: The research first identifies initial saliency as a factor where tokens with higher attention weights relative to the initial token receive more attention in next-token prediction. They propose scaling attention weights between the initial token and other tokens to improve long-context processing, and combine this approach with existing methods to reduce position encoding bias.

Result: The proposed method improves model performance on long contexts by up to 3.6% on the MDQA dataset. When combined with existing methods to reduce position encoding bias, it achieves up to 3.4% improvement on KV-Retrieval tasks.

Conclusion: Initial saliency is an important factor contributing to the “lost in the middle” problem in LLMs. Addressing this through attention weight scaling, especially when combined with existing position encoding bias reduction methods, significantly improves long-context processing capabilities.

Abstract: Large language models (LLMs) have demonstrated strong performance on a variety of natural language processing (NLP) tasks. However, they often struggle with long-text sequences due to the ``lost in the middle’’ phenomenon. This issue has been shown to arise from a U-shaped attention bias, where attention is disproportionately focused on the beginning and end of a text, leaving the middle section underrepresented. While previous studies have attributed this bias to position encoding, our research first identifies an additional factor: initial saliency. It means that in the attention computation for each token, tokens with higher attention weights relative to the initial token tend to receive more attention in the prediction of the next token. We further find that utilizing this property by scaling attention weight between the initial token and others improves the model’s ability to process long contexts, achieving a maximum improvement of 3.6% in MDQA dataset. Moreover, combining this approach with existing methods to reduce position encoding bias further enhances performance, achieving a maximum improvement of 3.4% in KV-Retrieval tasks.

Method: EARS (Efficient Adaptive Rejection Sampling) dynamically adjusts acceptance thresholds by incorporating the target model’s predictive uncertainty (1 - max(P_target)). It introduces a tolerance term proportional to this uncertainty, relaxing acceptance criteria when the model is uncertain while maintaining strict standards when confident.

Result: EARS significantly enhances speculative decoding efficiency, achieving up to 18.12% increase in throughput with only 0.84% accuracy drop on GSM8K benchmark. The method works well on creative writing and open-domain QA tasks.

Conclusion: EARS provides an effective solution to the random rejection problem in speculative decoding by adaptively adjusting acceptance thresholds based on model uncertainty, requiring no architectural changes and being easily integrable into existing frameworks.

Abstract: Speculative Decoding is a prominent technique for accelerating the autoregressive inference of large language models (LLMs) by employing a fast draft model to propose candidate token sequences and a large target model to verify them in parallel. However, its core component – the rejection sampling mechanism – relies on a fixed, context-independent random threshold. This leads to a significant “random rejection” problem in high-uncertainty generation scenarios, where plausible candidate tokens are frequently rejected due to random chance, undermining inference efficiency. This paper introduces Efficient Adaptive Rejection Sampling (EARS), a novel method that dynamically adjusts the acceptance threshold by incorporating the target model’s own predictive uncertainty, measured as (1 - \max(P_{\mathrm{target}})). By introducing a tolerance term proportional to this uncertainty, EARS intelligently relaxes the acceptance criterion when the model is uncertain, effectively reducing random rejections while maintaining strict standards when the model is confident. Experiments on creative writing and open-domain QA tasks demonstrate that EARS significantly enhances the efficiency of speculative decoding, achieving up to an 18.12% increase in throughput with a negligible 0.84% accuracy drop on the GSM8K benchmark. The method requires no modifications to model architectures and can be seamlessly integrated into existing speculative decoding frameworks.

Method: 1) Constructed a 200k dataset with explicit tool-selection rationales across 1,000+ tools and 100+ tasks; 2) Dual-phase optimization pipeline: supervised and RL-based trajectory stabilization for coherent reasoning, and KL-regularized Plackett-Luce ranking for consistent multi-step tool selection; 3) Trained on Qwen3-8B and Qwen2.5-VL-7B models.

Result: AutoTool outperforms advanced LLM agents and tool-integration methods across ten benchmarks: 6.4% average gain in math & science reasoning, 4.5% in search-based QA, 7.7% in code generation, and 6.9% in multimodal understanding. It also shows stronger generalization by dynamically leveraging unseen tools from evolving toolsets.

Conclusion: AutoTool successfully enables LLM agents to dynamically select tools throughout reasoning trajectories, achieving state-of-the-art performance across diverse domains while maintaining adaptability to new toolsets, representing a significant advancement in agentic reinforcement learning.

Abstract: Agentic reinforcement learning has advanced large language models (LLMs) to reason through long chain-of-thought trajectories while interleaving external tool use. Existing approaches assume a fixed inventory of tools, limiting LLM agents’ adaptability to new or evolving toolsets. We present AutoTool, a framework that equips LLM agents with dynamic tool-selection capabilities throughout their reasoning trajectories. We first construct a 200k dataset with explicit tool-selection rationales across 1,000+ tools and 100+ tasks spanning mathematics, science, code generation, and multimodal reasoning. Building on this data foundation, AutoTool employs a dual-phase optimization pipeline: (i) supervised and RL-based trajectory stabilization for coherent reasoning, and (ii) KL-regularized Plackett-Luce ranking to refine consistent multi-step tool selection. Across ten diverse benchmarks, we train two base models, Qwen3-8B and Qwen2.5-VL-7B, with AutoTool. With fewer parameters, AutoTool consistently outperforms advanced LLM agents and tool-integration methods, yielding average gains of 6.4% in math & science reasoning, 4.5% in search-based QA, 7.7% in code generation, and 6.9% in multimodal understanding. In addition, AutoTool exhibits stronger generalization by dynamically leveraging unseen tools from evolving toolsets during inference.

Method: AIR leverages mechanistic insights from retrieval heads: identifies reasoning-critical attention heads, constructs weakened reference model with disabled head influence, quantifies loss divergence as Attention Influence Score for step/sample-level assessment.

Result: Experiments across multiple reasoning benchmarks show AIR consistently improves reasoning accuracy, surpasses heuristic baselines, and effectively isolates critical steps and samples.

Conclusion: AIR establishes a mechanism-driven, data-efficient approach for reasoning distillation in LLMs by using attention influence to select high-value post-training data.

Abstract: LLMs achieve remarkable multi-step reasoning capabilities, yet effectively transferring these skills via post-training distillation remains challenging. Existing data selection methods, ranging from manual curation to heuristics based on length, entropy, or overall loss, fail to capture the causal importance of individual reasoning steps, limiting distillation efficiency. To address this, we propose Attention Influence for Reasoning (AIR), a principled, unsupervised and training-free framework that leverages mechanistic insights of the retrieval head to select high-value post-training data. AIR first identifies reasoning-critical attention heads of an off-the-shelf model, then constructs a weakened reference model with disabled head influence, and finally quantifies the resulting loss divergence as the Attention Influence Score. This score enables fine-grained assessment at both the step and sample levels, supporting step-level weighted fine-tuning and global sample selection. Experiments across multiple reasoning benchmarks show that AIR consistently improves reasoning accuracy, surpassing heuristic baselines and effectively isolating the most critical steps and samples. Our work establishes a mechanism-driven, data-efficient approach for reasoning distillation in LLMs.

Method: Combines event relation extraction, semantic similarity computation, and rule-based reasoning to detect logical inconsistencies between chains of events in claims and evidence.

Result: Establishes the first baseline for integrating fine-grained causal event relationships into fact-checking and enhances explainability of verdict prediction, evaluated on two fact-checking datasets.

Conclusion: The proposed methodology successfully addresses the gap in causal reasoning for fact-checking, providing a valuable approach for detecting erroneous cause-effect relationships and improving explainability.

Abstract: In fact-checking applications, a common reason to reject a claim is to detect the presence of erroneous cause-effect relationships between the events at play. However, current automated fact-checking methods lack dedicated causal-based reasoning, potentially missing a valuable opportunity for semantically rich explainability. To address this gap, we propose a methodology that combines event relation extraction, semantic similarity computation, and rule-based reasoning to detect logical inconsistencies between chains of events mentioned in a claim and in an evidence. Evaluated on two fact-checking datasets, this method establishes the first baseline for integrating fine-grained causal event relationships into fact-checking and enhance explainability of verdict prediction.

Method: Train a 1B parameter multilingual LLM from scratch for 13 European languages, with instruction tuning. Release model weights, tokenizer, and training code.

Result: Instruction-tuned MiniLingua outperforms EuroLLM (similar approach but larger budget) on summarization, classification, and QA tasks. Competitive with SOTA models on open-ended generation.

Conclusion: Small, efficient multilingual models like MiniLingua can deliver strong performance while addressing computational, privacy, and language coverage limitations of larger models.

Abstract: Large language models are powerful but often limited by high computational cost, privacy concerns, and English-centric training. Recent progress demonstrates that small, efficient models with around one billion parameters can deliver strong results and enable on-device use. This paper introduces MiniLingua, a multilingual open-source LLM of one billion parameters trained from scratch for 13 European languages, designed to balance coverage and instruction-following capabilities. Based on evaluation results, the instruction-tuned version of MiniLingua outperforms EuroLLM, a model with a similar training approach but a larger training budget, on summarization, classification and both open- and closed-book question answering. Moreover, it remains competitive with more advanced state-of-the-art models on open-ended generation tasks. We release model weights, tokenizer and source code used for data processing and model training.

Method: Consolidated Finnish versions of widely used benchmarks into a single collection, converted all datasets to HuggingFace format with multiple prompt variants, used pretrained 2.15B-parameter models to validate tasks through learning curve analysis (monotonicity, signal-to-noise, etc.), and evaluated larger instruction-tuned models across tasks.

Result: Created a publicly available benchmark suite covering multiple-choice and generative tasks across diverse domains (reading comprehension, commonsense reasoning, sentiment analysis, world knowledge, alignment) with validated robust tasks and comprehensive prompt formulations.

Conclusion: FIN-bench-v2 provides a standardized, validated evaluation framework for Finnish LLMs with systematic task selection, multiple prompt formats, and comprehensive coverage, facilitating better model assessment and comparison in the Finnish language context.

Abstract: We introduce FIN-bench-v2, a unified benchmark suite for evaluating large language models in Finnish. FIN-bench-v2 consolidates Finnish versions of widely used benchmarks together with an updated and expanded version of the original FIN-bench into a single, consistently formatted collection, covering multiple-choice and generative tasks across reading comprehension, commonsense reasoning, sentiment analysis, world knowledge, and alignment. All datasets are converted to HuggingFace Datasets, which include both cloze and multiple-choice prompt formulations with five variants per task, and we incorporate human annotation or review for machine-translated resources such as GoldenSwag and XED. To select robust tasks, we pretrain a set of 2.15B-parameter decoder-only models and use their learning curves to compute monotonicity, signal-to-noise, non-random performance, and model ordering consistency, retaining only tasks that satisfy all criteria. We further evaluate a set of larger instruction-tuned models to characterize performance across tasks and prompt formulations. All datasets, prompts, and evaluation configurations are publicly available via our fork of the Language Model Evaluation Harness at https://github.com/LumiOpen/lm-evaluation-harness. Supplementary resources are released in a separate repository at https://github.com/TurkuNLP/FIN-bench-v2.

Method: Uses pre-trained transformer models (DistilBERT and RoBERTa) to detect sentence-level emotions and measure emotion drift scores across messages. Focuses on mental health-related messages to analyze emotional state changes within single texts.

Result: The approach successfully detects sentence-level emotions and measures emotion drift scores, providing insights into patterns of emotional escalation or relief in mental health conversations.

Conclusion: The methodology enables better understanding of emotional dynamics in content, particularly valuable for analyzing mental health conversations where tracking emotional shifts within messages can reveal important patterns of escalation or relief.

Abstract: This study investigates emotion drift: the change in emotional state across a single text, within mental health-related messages. While sentiment analysis typically classifies an entire message as positive, negative, or neutral, the nuanced shift of emotions over the course of a message is often overlooked. This study detects sentence-level emotions and measures emotion drift scores using pre-trained transformer models such as DistilBERT and RoBERTa. The results provide insights into patterns of emotional escalation or relief in mental health conversations. This methodology can be applied to better understand emotional dynamics in content.

Method: Theoretical comparison between LLMs’ probabilistic, data-driven approach and human language’s innate, recursive computational system.

Result: LLMs fail to capture the essence of human language as they only model surface strings statistically, missing the underlying hierarchical structure and innate computational properties.

Conclusion: Large Language Models are fundamentally unsuitable for linguistic analysis because they cannot capture the innate, recursive computational system that underlies human language.

Abstract: Large Language Models are useless for linguistics, as they are probabilistic models that require a vast amount of data to analyse externalized strings of words. In contrast, human language is underpinned by a mind-internal computational system that recursively generates hierarchical thought structures. The language system grows with minimal external input and can readily distinguish between real language and impossible languages.

Method: Conducted over 1000+ experiments (equivalent to 336,000+ H800 GPU hours) across multiple programming languages, model sizes (0.2B to 14B parameters), and dataset sizes (1T tokens). Analyzed scaling patterns and introduced parallel pairing strategy (concatenating code snippets with translations).

Result: Interpreted languages (Python) benefit more from increased model size/data than compiled languages (Rust). Multilingual pre-training provides synergistic benefits, especially between syntactically similar languages. Parallel pairing enhances cross-lingual abilities with favorable scaling properties.

Conclusion: Proposed proportion-dependent multilingual scaling law that optimally allocates training tokens by prioritizing high-utility languages (Python), balancing high-synergy pairs (JavaScript-TypeScript), and reducing allocation to fast-saturating languages (Rust), achieving superior average performance across all languages.

Abstract: Code large language models (Code LLMs) are powerful but costly to train, with scaling laws predicting performance from model size, data, and compute. However, different programming languages (PLs) have varying impacts during pre-training that significantly affect base model performance, leading to inaccurate performance prediction. Besides, existing works focus on language-agnostic settings, neglecting the inherently multilingual nature of modern software development. Therefore, it is first necessary to investigate the scaling laws of different PLs, and then consider their mutual influences to arrive at the final multilingual scaling law. In this paper, we present the first systematic exploration of scaling laws for multilingual code pre-training, conducting over 1000+ experiments (Equivalent to 336,000+ H800 hours) across multiple PLs, model sizes (0.2B to 14B parameters), and dataset sizes (1T tokens). We establish comprehensive scaling laws for code LLMs across multiple PLs, revealing that interpreted languages (e.g., Python) benefit more from increased model size and data than compiled languages (e.g., Rust). The study demonstrates that multilingual pre-training provides synergistic benefits, particularly between syntactically similar PLs. Further, the pre-training strategy of the parallel pairing (concatenating code snippets with their translations) significantly enhances cross-lingual abilities with favorable scaling properties. Finally, a proportion-dependent multilingual scaling law is proposed to optimally allocate training tokens by prioritizing high-utility PLs (e.g., Python), balancing high-synergy pairs (e.g., JavaScript-TypeScript), and reducing allocation to fast-saturating languages (Rust), achieving superior average performance across all PLs compared to uniform distribution under the same compute budget.

Method: Non-Resolution Reasoning (NRR) framework with three components: (1) Multi-Vector Embeddings maintaining multiple viable interpretations per token, (2) Non-Collapsing Attention preventing winner-take-all dynamics, and (3) Contextual Identity Tracking (CIT) assigning context-specific identities to recurring entities. Unified by external Resolution Operator ρ that makes semantic commitment explicit and controllable.

Result: CIT-enhanced models achieve 90.9% accuracy on out-of-distribution identity-shift tasks, compared to 9.1% for transformer baselines. NRR enables models to preserve ambiguity and track context effectively.

Conclusion: NRR provides a principled alternative to premature collapse, reframing ambiguity as an explicit representational state rather than a failure mode. It allows models to shift between creative, factual, and ambiguity-preserving reasoning without retraining, making semantic commitment explicit, controllable, and task-dependent.

Abstract: Premature semantic collapse – the forced early commitment to a single meaning – remains a core architectural limitation of current language models. Softmax-driven competition and greedy decoding cause models to discard valid interpretations before sufficient context is available, resulting in brittle reasoning and context failures. We introduce Non-Resolution Reasoning (NRR), a general computational framework that preserves semantic ambiguity during inference and performs resolution only when explicitly required. NRR integrates three components: (1) Multi-Vector Embeddings that maintain multiple viable interpretations per token, (2) Non-Collapsing Attention that prevents winner-take-all dynamics across layers, and (3) Contextual Identity Tracking (CIT), which assigns context-specific identities to recurring entities (e.g., distinguishing “Dr. Smith the cardiologist” from “Dr. Smith the researcher”). These mechanisms are unified by an external Resolution Operator $ρ$ that makes semantic commitment explicit, controllable, and task-dependent. Unlike standard architectures, NRR separates representation from resolution, allowing a single model to shift between creative, factual, and ambiguity-preserving reasoning without retraining. A synthetic evaluation demonstrates NRR’s ability to preserve ambiguity and track context: CIT-enhanced models achieve 90.9% accuracy on out-of-distribution identity-shift tasks, compared to 9.1% for transformer baselines. NRR provides a principled alternative to premature collapse, reframing ambiguity as an explicit representational state rather than a failure mode. The question is not whether AI should resolve ambiguity, but when, how, and under whose control.

Method: Introduces EMNLP framework with: 1) personality profiling, 2) moral development stage measurement, 3) ethical risk assessment under soft prompt injection. Extends existing scales and constructs 88 teacher-specific moral dilemmas for profession-oriented comparison with human teachers. Uses targeted soft prompt injection set to evaluate compliance and vulnerability.

Result: Experiments on 14 LLMs show: teacher-role LLMs exhibit more idealized and polarized personalities than human teachers; excel in abstract moral reasoning but struggle with emotionally complex situations; models with stronger reasoning are paradoxically more vulnerable to harmful prompt injection; model temperature and hyperparameters have limited influence except in some risk behaviors.

Conclusion: EMNLP presents the first benchmark to assess ethical and psychological alignment of teacher-role LLMs for educational AI, revealing important insights about the paradox between capability and safety in professional role simulation.

Abstract: Simulating Professions (SP) enables Large Language Models (LLMs) to emulate professional roles. However, comprehensive psychological and ethical evaluation in these contexts remains lacking. This paper introduces EMNLP, an Educator-role Moral and Normative LLMs Profiling framework for personality profiling, moral development stage measurement, and ethical risk under soft prompt injection. EMNLP extends existing scales and constructs 88 teacher-specific moral dilemmas, enabling profession-oriented comparison with human teachers. A targeted soft prompt injection set evaluates compliance and vulnerability in teacher SP. Experiments on 14 LLMs show teacher-role LLMs exhibit more idealized and polarized personalities than human teachers, excel in abstract moral reasoning, but struggle with emotionally complex situations. Models with stronger reasoning are more vulnerable to harmful prompt injection, revealing a paradox between capability and safety. The model temperature and other hyperparameters have limited influence except in some risk behaviors. This paper presents the first benchmark to assess ethical and psychological alignment of teacher-role LLMs for educational AI. Resources are available at https://e-m-n-l-p.github.io/.

Method: The researchers explore current state-of-the-art models and propose improvements by developing a dataset specifically from informal sources like social media and conversational texts. This approach addresses the lack of pairwise Bangla-English data for informal language translation.

Result: The paper aims to advance Bangla machine translation by creating specialized datasets for informal language, which should lead to improved translation models that better handle natural, informal Bangla used in everyday communication.

Conclusion: By focusing on informal language translation and developing appropriate datasets, this research will improve accessibility for Bangla speakers in the digital world, enabling millions to access online information that currently remains untranslated from English.

Abstract: Bangla is the sixth most widely spoken language globally, with approximately 234 million native speakers. However, progress in open-source Bangla machine translation remains limited. Most online resources are in English and often remain untranslated into Bangla, excluding millions from accessing essential information. Existing research in Bangla translation primarily focuses on formal language, neglecting the more commonly used informal language. This is largely due to the lack of pairwise Bangla-English data and advanced translation models. If datasets and models can be enhanced to better handle natural, informal Bangla, millions of people will benefit from improved online information access. In this research, we explore current state-of-the-art models and propose improvements to Bangla translation by developing a dataset from informal sources like social media and conversational texts. This work aims to advance Bangla machine translation by focusing on informal language translation and improving accessibility for Bangla speakers in the digital world.

Method: SkipCat introduces two key techniques: 1) Intra-layer shared low-rank projection where multiple matrices sharing the same input use a common projection to reduce redundancy, and 2) Block skipping that omits computations and memory transfers for selected sub-blocks within the low-rank decomposition. These allow retaining more effective ranks under the same compression budget.

Result: Experimental results show that without any additional fine-tuning, SkipCat outperforms previous low-rank compression approaches by 7% accuracy improvement on zero-shot tasks under the same compression rate.

Conclusion: SkipCat’s rank-maximized compression strategy effectively preserves model performance under tight resource constraints, making LLMs more suitable for deployment on edge devices while maintaining better accuracy than previous compression methods.

Abstract: Large language models (LLM) have achieved remarkable performance across a wide range of tasks. However, their substantial parameter sizes pose significant challenges for deployment on edge devices with limited computational and memory resources. Low-rank compression is a promising approach to address this issue, as it reduces both computational and memory costs, making LLM more suitable for resource-constrained environments. Nonetheless, naïve low-rank compression methods require a significant reduction in the retained rank to achieve meaningful memory and computation savings. For a low-rank model, the ranks need to be reduced by more than half to yield efficiency gains. Such aggressive truncation, however, typically results in substantial performance degradation. To address this trade-off, we propose SkipCat, a novel low-rank compression framework that enables the use of higher ranks while achieving the same compression rates. First, we introduce an intra-layer shared low-rank projection method, where multiple matrices that share the same input use a common projection. This reduces redundancy and improves compression efficiency. Second, we propose a block skipping technique that omits computations and memory transfers for selected sub-blocks within the low-rank decomposition. These two techniques jointly enable our compressed model to retain more effective ranks under the same compression budget. Experimental results show that, without any additional fine-tuning, our method outperforms previous low-rank compression approaches by 7% accuracy improvement on zero-shot tasks under the same compression rate. These results highlight the effectiveness of our rank-maximized compression strategy in preserving model performance under tight resource constraints.

Method: Developed PrahokBART from scratch using carefully curated Khmer and English corpora. Incorporated linguistic components including word segmentation and normalization to handle Khmer-specific issues. The model is a compact sequence-to-sequence architecture.

Result: PrahokBART outperforms mBART50 (a strong multilingual pre-trained model) on three generative tasks: machine translation, text summarization, and headline generation. Analysis provides insights into the impact of each linguistic module and evaluates space handling during text generation.

Conclusion: Specialized monolingual models with proper linguistic processing can outperform multilingual models for low-resource languages like Khmer. Proper handling of linguistic components and text spacing is crucial for natural text generation in Khmer.

Abstract: This work introduces {\it PrahokBART}, a compact pre-trained sequence-to-sequence model trained from scratch for Khmer using carefully curated Khmer and English corpora. We focus on improving the pre-training corpus quality and addressing the linguistic issues of Khmer, which are ignored in existing multilingual models, by incorporating linguistic components such as word segmentation and normalization. We evaluate PrahokBART on three generative tasks: machine translation, text summarization, and headline generation, where our results demonstrate that it outperforms mBART50, a strong multilingual pre-trained model. Additionally, our analysis provides insights into the impact of each linguistic module and evaluates how effectively our model handles space during text generation, which is crucial for the naturalness of texts in Khmer.

Method: Proposes stance-aware structural modeling that encodes each post with its stance signal, aggregates reply embeddings by stance category, and introduces stance distribution and hierarchical depth as covariates to capture stance imbalance and reply depth influence.

Result: Extensive experiments on benchmark datasets show the approach significantly outperforms prior methods in predicting rumor truthfulness, and demonstrates versatility for early detection and cross-platform generalization.

Conclusion: The stance-aware structural modeling provides a scalable and semantically enriched representation of conversation threads, effectively addressing limitations of existing models for rumor verification on social media.

Abstract: Verifying rumors on social media is critical for mitigating the spread of false information. The stances of conversation replies often provide important cues to determine a rumor’s veracity. However, existing models struggle to jointly capture semantic content, stance information, and conversation strructure, especially under the sequence length constraints of transformer-based encoders. In this work, we propose a stance-aware structural modeling that encodes each post in a discourse with its stance signal and aggregates reply embedddings by stance category enabling a scalable and semantically enriched representation of the entire thread. To enhance structural awareness, we introduce stance distribution and hierarchical depth as covariates, capturing stance imbalance and the influence of reply depth. Extensive experiments on benchmark datasets demonstrate that our approach significantly outperforms prior methods in the ability to predict truthfulness of a rumor. We also demonstrate that our model is versatile for early detection and cross-platfrom generalization.

Method: The authors analyze agent memory through three unified lenses: 1) Forms (token-level, parametric, and latent memory), 2) Functions (factual, experiential, and working memory), and 3) Dynamics (formation, evolution, and retrieval processes). They also compile benchmarks and frameworks, and distinguish agent memory from related concepts like LLM memory and RAG.

Result: The paper provides a comprehensive landscape of current agent memory research, offering a clearer conceptual foundation and taxonomy. It identifies three dominant memory forms, proposes a finer-grained functional taxonomy, and analyzes memory dynamics, while also summarizing practical resources like benchmarks and frameworks.

Conclusion: This survey serves as both a reference for existing work and a conceptual foundation for rethinking memory as a first-class primitive in future agentic intelligence design. It highlights emerging research frontiers including memory automation, RL integration, multimodal memory, multi-agent memory, and trustworthiness issues.

Abstract: Memory has emerged, and will continue to remain, a core capability of foundation model-based agents. As research on agent memory rapidly expands and attracts unprecedented attention, the field has also become increasingly fragmented. Existing works that fall under the umbrella of agent memory often differ substantially in their motivations, implementations, and evaluation protocols, while the proliferation of loosely defined memory terminologies has further obscured conceptual clarity. Traditional taxonomies such as long/short-term memory have proven insufficient to capture the diversity of contemporary agent memory systems. This work aims to provide an up-to-date landscape of current agent memory research. We begin by clearly delineating the scope of agent memory and distinguishing it from related concepts such as LLM memory, retrieval augmented generation (RAG), and context engineering. We then examine agent memory through the unified lenses of forms, functions, and dynamics. From the perspective of forms, we identify three dominant realizations of agent memory, namely token-level, parametric, and latent memory. From the perspective of functions, we propose a finer-grained taxonomy that distinguishes factual, experiential, and working memory. From the perspective of dynamics, we analyze how memory is formed, evolved, and retrieved over time. To support practical development, we compile a comprehensive summary of memory benchmarks and open-source frameworks. Beyond consolidation, we articulate a forward-looking perspective on emerging research frontiers, including memory automation, reinforcement learning integration, multimodal memory, multi-agent memory, and trustworthiness issues. We hope this survey serves not only as a reference for existing work, but also as a conceptual foundation for rethinking memory as a first-class primitive in the design of future agentic intelligence.

Method: ReFusion elevates parallel decoding from token level to slot level (contiguous sub-sequences). It uses iterative plan-and-infill: 1) diffusion-based planning identifies weakly dependent slots, 2) autoregressive infilling decodes selected slots in parallel. This enables KV cache reuse and reduces learning complexity to manageable slot-level permutations.

Result: Extensive experiments on 7 benchmarks show ReFusion surpasses prior masked diffusion models with 34% performance gains and 18× speedup on average. It bridges performance gap to strong autoregressive models while maintaining 2.33× average speedup.

Conclusion: ReFusion successfully addresses key limitations of both autoregressive and masked diffusion models by introducing slot-level parallel decoding with plan-and-infill, achieving superior efficiency-performance trade-off for text generation.

Abstract: Autoregressive models (ARMs) are hindered by slow sequential inference. While masked diffusion models (MDMs) offer a parallel alternative, they suffer from critical drawbacks: high computational overhead from precluding Key-Value (KV) caching, and incoherent generation arising from learning dependencies over an intractable space of token combinations. To address these limitations, we introduce ReFusion, a novel masked diffusion model that achieves superior performance and efficiency by elevating parallel decoding from the token level to a higher slot level, where each slot is a fixed-length, contiguous sub-sequence. This is achieved through an iterative ``plan-and-infill’’ decoding process: a diffusion-based planning step first identifies a set of weakly dependent slots, and an autoregressive infilling step then decodes these selected slots in parallel. The slot-based design simultaneously unlocks full KV cache reuse with a unified causal framework and reduces the learning complexity from the token combination space to a manageable slot-level permutation space. Extensive experiments on seven diverse benchmarks show that ReFusion not only overwhelmingly surpasses prior MDMs with 34% performance gains and an over 18$\times$ speedup on average, but also bridges the performance gap to strong ARMs while maintaining a 2.33$\times$ average speedup.

Method: Conducted experiments and case studies to investigate the behavior of textual gradient prompt optimization techniques.

Result: Textual gradient methods often result in performance improvements, but the gradient analogy does not accurately explain their behavior.

Conclusion: These insights can inform better selection of prompt optimization strategies and development of new approaches.

Abstract: A well-engineered prompt can increase the performance of large language models; automatic prompt optimization techniques aim to increase performance without requiring human effort to tune the prompts. One leading class of prompt optimization techniques introduces the analogy of textual gradients. We investigate the behavior of these textual gradient methods through a series of experiments and case studies. While such methods often result in a performance improvement, our experiments suggest that the gradient analogy does not accurately explain their behavior. Our insights may inform the selection of prompt optimization strategies, and development of new approaches.

Method: Proposes cascaded domain-wise RL (Cascade RL) that orchestrates sequential, domain-wise reinforcement learning instead of blending heterogeneous prompts. Uses RLHF for alignment as pre-step, followed by domain-wise RLVR stages.

Result: 14B model outperforms SFT teacher DeepSeek-R1-0528 on LiveCodeBench v5/v6/Pro, achieves silver-medal performance in 2025 IOI. RLHF boosts reasoning beyond preference optimization, and subsequent RLVR stages rarely degrade earlier domain performance.

Conclusion: Cascade RL reduces engineering complexity while delivering state-of-the-art performance across diverse benchmarks, enabling effective training of general-purpose reasoning models (Nemotron-Cascade) with both instruct and deep thinking modes.

Abstract: Building general-purpose reasoning models with reinforcement learning (RL) entails substantial cross-domain heterogeneity, including large variation in inference-time response lengths and verification latency. Such variability complicates the RL infrastructure, slows training, and makes training curriculum (e.g., response length extension) and hyperparameter selection challenging. In this work, we propose cascaded domain-wise reinforcement learning (Cascade RL) to develop general-purpose reasoning models, Nemotron-Cascade, capable of operating in both instruct and deep thinking modes. Departing from conventional approaches that blend heterogeneous prompts from different domains, Cascade RL orchestrates sequential, domain-wise RL, reducing engineering complexity and delivering state-of-the-art performance across a wide range of benchmarks. Notably, RLHF for alignment, when used as a pre-step, boosts the model’s reasoning ability far beyond mere preference optimization, and subsequent domain-wise RLVR stages rarely degrade the benchmark performance attained in earlier domains and may even improve it (see an illustration in Figure 1). Our 14B model, after RL, outperforms its SFT teacher, DeepSeek-R1-0528, on LiveCodeBench v5/v6/Pro and achieves silver-medal performance in the 2025 International Olympiad in Informatics (IOI). We transparently share our training and data recipes.

Method: Empirical study comparing 5 temporal encoding strategies: naive numeric strings, high-precision byte-level representations, human-semantic calendar tokens, classic uniform binning, and adaptive residual scalar quantization. Evaluated by fine-tuning LLMs on real-world datasets with diverse distributions.

Result: No single strategy universally superior. Prediction performance depends heavily on aligning tokenizer with data’s statistical properties. Log-based strategies excel on skewed distributions, human-centric formats prove robust for mixed modalities.

Conclusion: Temporal tokenization strategy should be chosen based on data’s statistical properties rather than using one-size-fits-all approach. Alignment between tokenizer design and data distribution is crucial for optimal performance.

Abstract: Representing continuous time is a critical and under-explored challenge in modeling temporal event sequences with large language models (LLMs). Various strategies like byte-level representations or calendar tokens have been proposed. However, the optimal approach remains unclear, especially given the diverse statistical distributions of real-world event data, which range from smooth log-normal to discrete, spiky patterns. This paper presents the first empirical study of temporal tokenization for event sequences, comparing distinct encoding strategies: naive numeric strings, high-precision byte-level representations, human-semantic calendar tokens, classic uniform binning, and adaptive residual scalar quantization. We evaluate these strategies by fine-tuning LLMs on real-world datasets that exemplify these diverse distributions. Our analysis reveals that no single strategy is universally superior; instead, prediction performance depends heavily on aligning the tokenizer with the data’s statistical properties, with log-based strategies excelling on skewed distributions and human-centric formats proving robust for mixed modalities.

Method: Used SCOTUS decisions as a classification task with two traditional category-based tasks: one with 15 labeled topics and another with 279 topics. Experimented with latest LLM fine-tuning and retrieval-based approaches including parameter-efficient fine-tuning and auto-modeling. Compared prompt-based models with memories (like DeepSeek) against previous BERT-based models.

Result: Prompt-based models with memories, such as DeepSeek, demonstrated greater robustness than previous BERT-based models on both SCOTUS classification tasks, scoring approximately 2 points better than models not based on prompting.

Conclusion: Prompt-based LLMs with memory capabilities show superior performance on complex legal text classification tasks compared to traditional BERT-based approaches, suggesting that memory-enhanced prompting strategies can better handle the challenges of domain-specific, structurally complex documents like SCOTUS decisions.

Abstract: Large-language models (LLMs) have been shown to respond in a variety of ways for classification tasks outside of question-answering. LLM responses are sometimes called “hallucinations” since the output is not what is ex pected. Memorization strategies in LLMs are being studied in detail, with the goal of understanding how LLMs respond. We perform a deep dive into a classification task based on United States Supreme Court (SCOTUS) decisions. The SCOTUS corpus is an ideal classification task to study for LLM memory accuracy because it presents significant challenges due to extensive sentence length, complex legal terminology, non-standard structure, and domain-specific vocabulary. Experimentation is performed with the latest LLM fine tuning and retrieval-based approaches, such as parameter-efficient fine-tuning, auto-modeling, and others, on two traditional category-based SCOTUS classification tasks: one with 15 labeled topics and another with 279. We show that prompt-based models with memories, such as DeepSeek, can be more robust than previous BERT-based models on both tasks scoring about 2 points better than previous models not based on prompting.

Method: Evaluated four abliteration tools across sixteen instruction-tuned models (7B-14B parameters). Assessed tool compatibility on all models and quantitative metrics on subsets based on tool support. Compared single-pass methods vs Bayesian-optimized approaches using GSM8K performance and KL divergence.

Result: Single-pass methods demonstrated superior capability preservation (avg GSM8K change: ErisForge -0.28 pp; DECCP -0.13 pp). Bayesian-optimized abliteration produced variable distribution shift (KL divergence: 0.043-1.646). Mathematical reasoning capabilities showed highest sensitivity to interventions, with GSM8K changes ranging from +1.51 pp to -18.81 pp (-26.5% relative).

Conclusion: The study provides evidence-based selection criteria for abliteration tool deployment across diverse model architectures, highlighting that mathematical reasoning capabilities are most sensitive to abliteration interventions, with performance impacts varying significantly by tool selection and model architecture.

Abstract: Safety alignment mechanisms in large language models prevent responses to harmful queries through learned refusal behavior, yet these same mechanisms impede legitimate research applications including cognitive modeling, adversarial testing, and security analysis. While abliteration techniques enable surgical removal of refusal representations through directional orthogonalization, the relative effectiveness of available implementations remains uncharacterized. This study evaluates four abliteration tools (Heretic, DECCP, ErisForge, FailSpy) across sixteen instruction-tuned models (7B-14B parameters), reporting tool compatibility on all 16 models and quantitative metrics on subsets dictated by tool support. Single-pass methods demonstrated superior capability preservation on the benchmarked subset (avg GSM8K change across three models: ErisForge -0.28 pp; DECCP -0.13 pp), while Bayesian-optimized abliteration produced variable distribution shift (KL divergence: 0.043-1.646) with model-dependent capability impact. These findings provide researchers with evidence-based selection criteria for abliteration tool deployment across diverse model architectures. The principal finding indicates that mathematical reasoning capabilities exhibit the highest sensitivity to abliteration interventions, with GSM8K change ranging from +1.51 pp to -18.81 pp (-26.5% relative) depending on tool selection and model architecture.

Method: StyloSpeaker applies stylometric authorship attribution techniques to speech transcripts, incorporating character, word, token, sentence, and style features. The method is evaluated on two transcript formats (prescriptive vs. normalized) with varying topic control, and compared to neural approaches.

Result: Higher attribution performance was generally achieved with normalized transcripts (without capitalization/punctuation), except under strongest topic control where overall performance peaked. The explainable stylometric model performed comparably to black-box neural approaches, with specific stylistic features effectively distinguishing speakers.

Conclusion: Stylometric methods can effectively attribute speakers from transcribed speech, with normalization generally improving performance. The approach provides an explainable alternative to neural methods and identifies which stylistic features are most discriminative for speaker attribution.

Abstract: Forensic scientists often need to identify an unknown speaker or writer in cases such as ransom calls, covert recordings, alleged suicide notes, or anonymous online communications, among many others. Speaker recognition in the speech domain usually examines phonetic or acoustic properties of a voice, and these methods can be accurate and robust under certain conditions. However, if a speaker disguises their voice or employs text-to-speech software, vocal properties may no longer be reliable, leaving only their linguistic content available for analysis. Authorship attribution methods traditionally use syntactic, semantic, and related linguistic information to identify writers of written text (authorship attribution). In this paper, we apply a content-based authorship approach to speech that has been transcribed into text, using what a speaker says to attribute speech to individuals (speaker attribution). We introduce a stylometric method, StyloSpeaker, which incorporates character, word, token, sentence, and style features from the stylometric literature on authorship, to assess whether two transcripts were produced by the same speaker. We evaluate this method on two types of transcript formatting: one approximating prescriptive written text with capitalization and punctuation and another normalized style that removes these conventions. The transcripts’ conversation topics are also controlled to varying degrees. We find generally higher attribution performance on normalized transcripts, except under the strongest topic control condition, in which overall performance is highest. Finally, we compare this more explainable stylometric model to black-box neural approaches on the same data and investigate which stylistic features most effectively distinguish speakers.

Method: Conceptualizes personalization as model editing task, introduces Personalization Editing framework that applies localized edits guided by clustered preference representations. Also creates UPQA dataset - short-answer QA from real user queries with varying difficulty levels to evaluate preference recall.

Result: Personalization Editing achieves higher editing accuracy and greater computational efficiency than fine-tuning, outperforms prompting-based baselines in multi-turn conversations and implicit preference question settings.

Conclusion: Treating personalization as model editing with clustered preference representations provides effective solution to current limitations, enabling precise preference-aligned updates while preserving overall model capabilities, with UPQA dataset offering better evaluation benchmark.

Abstract: Personalization is becoming indispensable for LLMs to align with individual user preferences and needs. Yet current approaches are often computationally expensive, data-intensive, susceptible to catastrophic forgetting, and prone to performance degradation in multi-turn interactions or when handling implicit queries. To address these challenges, we conceptualize personalization as a model editing task and introduce Personalization Editing, a framework that applies localized edits guided by clustered preference representations. This design enables precise preference-aligned updates while preserving overall model capabilities. In addition, existing personalization benchmarks frequently rely on persona-based dialogs between LLMs rather than user-LLM interactions, or focus primarily on stylistic imitation while neglecting information-seeking tasks that require accurate recall of user-specific preferences. We introduce User Preference Question Answering (UPQA), a short-answer QA dataset constructed from in-situ user queries with varying levels of difficulty. Unlike prior benchmarks, UPQA directly evaluates a model’s ability to recall and apply specific user preferences. Across experimental settings, Personalization Editing achieves higher editing accuracy and greater computational efficiency than fine-tuning, while outperforming prompting-based baselines in multi-turn conversations and implicit preference questions settings.

Method: Introduced novel text transformation approach: altered syntax and vocabulary while preserving semantic content (low BLEU/chrF scores but high semantic similarity). Tested AD classification performance on transformed texts. Also investigated if picture descriptions retain enough detail for image reconstruction using generative models.

Result: Models performed similarly on transformed vs original texts (small macro-F1 deviations), showing they use semantic information. Image-based transformations added noise and reduced classification accuracy. Semantic information alone enables AD detection.

Conclusion: Language models can detect difficult-to-identify semantic impairment in Alzheimer’s patients, addressing an overlooked feature of linguistic deterioration. This opens new pathways for early detection systems by isolating true semantic markers from surface patterns.

Abstract: Alzheimer’s Disease (AD) is a progressive neurodegenerative condition that adversely affects cognitive abilities. Language-related changes can be automatically identified through the analysis of outputs from linguistic assessment tasks, such as picture description. Language models show promise as a basis for screening tools for AD, but their limited interpretability poses a challenge in distinguishing true linguistic markers of cognitive decline from surface-level textual patterns. To address this issue, we examine how surface form variation affects classification performance, with the goal of assessing the ability of language models to represent underlying semantic indicators. We introduce a novel approach where texts surface forms are transformed by altering syntax and vocabulary while preserving semantic content. The transformations significantly modify the structure and lexical content, as indicated by low BLEU and chrF scores, yet retain the underlying semantics, as reflected in high semantic similarity scores, isolating the effect of semantic information, and finding models perform similarly to if they were using the original text, with only small deviations in macro-F1. We also investigate whether language from picture descriptions retains enough detail to reconstruct the original image using generative models. We found that image-based transformations add substantial noise reducing classification accuracy. Our methodology provides a novel way of looking at what features influence model predictions, and allows the removal of possible spurious correlations. We find that just using semantic information, language model based classifiers can still detect AD. This work shows that difficult to detect semantic impairment can be identified, addressing an overlooked feature of linguistic deterioration, and opening new pathways for early detection systems.

Method: The study provides a comprehensive overview of LLM integration in finance and conducts holistic tests on multiple financial tasks using natural language instructions, specifically evaluating GPT-4’s performance across various financial applications.

Result: GPT-4 effectively follows prompt instructions across various financial tasks, demonstrating practical utility in financial applications such as report generation, market forecasting, sentiment analysis, and personalized advice.

Conclusion: The survey deepens understanding of LLMs’ current role in finance, identifies new research and application prospects, and highlights how these technologies can solve practical challenges in the finance industry.

Abstract: In recent years, Large Language Models (LLMs) like ChatGPT have seen considerable advancements and have been applied in diverse fields. Built on the Transformer architecture, these models are trained on extensive datasets, enabling them to understand and generate human language effectively. In the financial domain, the deployment of LLMs is gaining momentum. These models are being utilized for automating financial report generation, forecasting market trends, analyzing investor sentiment, and offering personalized financial advice. Leveraging their natural language processing capabilities, LLMs can distill key insights from vast financial data, aiding institutions in making informed investment choices and enhancing both operational efficiency and customer satisfaction. In this study, we provide a comprehensive overview of the emerging integration of LLMs into various financial tasks. Additionally, we conducted holistic tests on multiple financial tasks through the combination of natural language instructions. Our findings show that GPT-4 effectively follow prompt instructions across various financial tasks. This survey and evaluation of LLMs in the financial domain aim to deepen the understanding of LLMs’ current role in finance for both financial practitioners and LLM researchers, identify new research and application prospects, and highlight how these technologies can be leveraged to solve practical challenges in the finance industry.

Method: DABL fine-tunes Llama 2 on a large dataset of 143,137 real-world process models, generating normal traces through playout and simulating ordering/exclusion anomalies to create training logs for semantic anomaly detection.

Result: DABL surpasses state-of-the-art semantic anomaly detection methods in both generalization ability and process learning, requires no additional training for user datasets, and provides natural language interpretations of anomaly causes.

Conclusion: DABL offers an effective, interpretable solution for semantic anomaly detection in business processes using LLMs, addressing limitations of existing methods while providing valuable insights through natural language explanations.

Abstract: Detecting anomalies in business processes is crucial for ensuring operational success. While many existing methods rely on statistical frequency to detect anomalies, it’s important to note that infrequent behavior doesn’t necessarily imply undesirability. To address this challenge, detecting anomalies from a semantic viewpoint proves to be a more effective approach. However, current semantic anomaly detection methods treat a trace (i.e., process instance) as multiple event pairs, disrupting long-distance dependencies. In this paper, we introduce DABL, a novel approach for detecting semantic anomalies in business processes using large language models (LLMs). We collect 143,137 real-world process models from various domains. By generating normal traces through the playout of these process models and simulating both ordering and exclusion anomalies, we fine-tune Llama 2 using the resulting log. Through extensive experiments, we demonstrate that DABL surpasses existing state-of-the-art semantic anomaly detection methods in terms of both generalization ability and learning of given processes. Users can directly apply DABL to detect semantic anomalies in their own datasets without the need for additional training. Furthermore, DABL offers the capability to interpret the causes of anomalies in natural language, providing valuable insights into the detected anomalies.

Method: Proposes Adversarial Contrastive Decoding (ACD) - generates two opposite soft system prompts via optimization: Safeguarding Prompt (SP) promotes safer outputs, Adversarial Prompt (AP) exploits harmful parts. Uses lightweight prompt tuning on small anchor dataset without training target model.

Result: Achieves better safety performance than previous model training-free decoding methods without sacrificing original generation ability, demonstrated on extensive models and benchmarks.

Conclusion: ACD provides an effective, lightweight approach to LLM safety alignment through prompt-based contrastive decoding, avoiding heavy training requirements while maintaining generation quality.

Abstract: With the widespread application of Large Language Models (LLMs), it has become a significant concern to ensure their safety and prevent harmful responses. While current safe-alignment methods based on instruction fine-tuning and Reinforcement Learning from Human Feedback (RLHF) can effectively reduce harmful responses from LLMs, they often require high-quality datasets and heavy computational overhead during model training. Another way to align language models is to modify the logit of tokens in model outputs without heavy training. Recent studies have shown that contrastive decoding can enhance the performance of language models by reducing the likelihood of confused tokens. However, these methods require the manual selection of contrastive models or instruction templates, limiting the degree of contrast. To this end, we propose Adversarial Contrastive Decoding (ACD), an optimization-based framework to generate two opposite soft system prompts, the Safeguarding Prompt (SP) and the Adversarial Prompt (AP), for prompt-based contrastive decoding. The SP aims to promote safer outputs while the AP aims to exploit the harmful parts of the model, providing a strong contrast to align the model with safety. ACD only needs to apply a lightweight prompt tuning on a rather small anchor dataset without training the target model. Experiments conducted on extensive models and benchmarks demonstrate that the proposed method achieves much better safety performance than previous model training-free decoding methods without sacrificing its original generation ability.

Method: Proposes FUR (Faithfulness by Unlearning Reasoning steps) framework that erases information contained in reasoning steps from model parameters and measures faithfulness by observing how this affects the model’s predictions. If unlearning key reasoning steps changes the prediction, it indicates the CoT was parametrically faithful.

Result: Experiments with four language models and five multi-hop multi-choice question answering datasets show FUR can frequently precisely change models’ predictions by unlearning key reasoning steps, indicating when CoTs are parametrically faithful. Post-unlearning, models generate CoTs supporting different answers.

Conclusion: FUR provides a framework for measuring parametric faithfulness of chain-of-thought reasoning. The ability to change predictions by unlearning reasoning steps demonstrates when CoTs reflect true parametric beliefs, and the deeper effects of unlearning suggest complex relationships between reasoning and model parameters.

Abstract: When prompted to think step-by-step, language models (LMs) produce a chain of thought (CoT), a sequence of reasoning steps that the model supposedly used to produce its prediction. Despite much work on CoT prompting, it is unclear if reasoning verbalized in a CoT is faithful to the models’ parametric beliefs. We introduce a framework for measuring parametric faithfulness of generated reasoning, and propose Faithfulness by Unlearning Reasoning steps (FUR), an instance of this framework. FUR erases information contained in reasoning steps from model parameters, and measures faithfulness as the resulting effect on the model’s prediction. Our experiments with four LMs and five multi-hop multi-choice question answering (MCQA) datasets show that FUR is frequently able to precisely change the underlying models’ prediction for a given instance by unlearning key steps, indicating when a CoT is parametrically faithful. Further analysis shows that CoTs generated by models post-unlearning support different answers, hinting at a deeper effect of unlearning.

Method: Use verifiable benchmarks (mathematical reasoning, factual knowledge, code generation) with objective ground-truth assessment to distinguish harmful from legitimate self-preference. Conduct large-scale experiments across 7 model families and test inference-time scaling strategies like Chain-of-Thought.

Result: 1) Stronger models show greater self-preference but mostly legitimate (aligning with superior performance). 2) Harmful self-preference persists when evaluator models err, with stronger models showing more pronounced bias when wrong. 3) Chain-of-Thought reduces harmful self-preference. Findings extend to subjective domains via LMArena experiments.

Conclusion: Provides nuanced understanding of LLM-based evaluation: self-preference is often legitimate but harmful bias exists when models err. Practical insights for improving evaluation reliability through inference-time strategies like Chain-of-Thought.

Abstract: Large language models (LLMs) are increasingly used as automatic evaluators in applications such as benchmarking, reward modeling, and self-refinement. Prior work highlights a potential self-preference bias where LLMs favor their own generated responses, a tendency often intensifying with model size and capability. This raises a critical question: Is self-preference harmful, or does it simply reflect the genuinely higher-quality outputs of stronger models? Answering this has been difficult as prior works mostly relied on subjective tasks that lack an objective ground truth, meaning that either preference can be reasonably justified. To address this ambiguity, we investigate self-preference using verifiable benchmarks (mathematical reasoning, factual knowledge, code generation) that allow objective ground-truth assessment. This enables us to distinguish harmful (favoring objectively worse responses) from legitimate (favoring genuinely superior ones) self-preference. Our large-scale experiments across 7 model families reveal three key findings: (1) While stronger models exhibit greater self-preference, much of this preference aligns with objectively superior performance, indicating stronger models prefer themselves mostly legitimately. (2) Harmful self-preference persists when evaluator models err as generators, and stronger models display more pronounced harmful self-preference bias when they do err. This suggests stronger models struggle more to recognize when they are wrong. (3) Inference-time scaling strategies, such as generating a long Chain-of-Thought before evaluation, effectively reduce the harmful self-preference. Additionally, we experiment with LMArena and show that our findings extend beyond verifiable benchmarks to real-world, subjective domains. These results provide a more nuanced understanding of LLM-based evaluation and practical insights for improving its reliability.

Method: The study compares LLMs against skilled human annotators on three span annotation tasks: evaluating data-to-text generation, identifying translation errors, and detecting propaganda techniques. The research analyzes inter-annotator agreement (IAA) between LLMs and humans, error rates, and cost efficiency.

Result: LLMs show only moderate IAA with human annotators overall, but make errors at a similar rate as skilled crowdworkers. LLMs produce annotations at a fraction of the cost per output annotation compared to human annotation.

Conclusion: LLMs can serve as a viable alternative to human annotators for span annotation tasks, offering comparable error rates to skilled crowdworkers while being significantly more cost-effective, though with only moderate agreement with human judgments.

Abstract: Span annotation - annotating specific text features at the span level - can be used to evaluate texts where single-score metrics fail to provide actionable feedback. Until recently, span annotation was done by human annotators or fine-tuned models. In this paper, we study whether large language models (LLMs) can serve as an alternative to human annotators. We compare the abilities of LLMs to skilled human annotators on three span annotation tasks: evaluating data-to-text generation, identifying translation errors, and detecting propaganda techniques. We show that overall, LLMs have only moderate inter-annotator agreement (IAA) with human annotators. However, we demonstrate that LLMs make errors at a similar rate as skilled crowdworkers. LLMs also produce annotations at a fraction of the cost per output annotation. We release the dataset of over 40k model and human span annotations for further research.

Method: Cross-dataset experiments comparing different adaptation strategies (fine-tuning, few-shot in-context learning, zero-shot) across relation extraction benchmarks, examining factors like data quality, lexical similarity, and structural issues in datasets.

Result: RE models struggle with unseen data even within similar domains. Higher intra-dataset performance doesn’t indicate better transferability. Data quality is more important than lexical similarity for robust transfer. Fine-tuning works best with high-quality data, while few-shot ICL is more effective with noisy data. Zero-shot baselines occasionally outperform all cross-dataset approaches. Structural issues in RE benchmarks hinder model transferability.

Conclusion: Current RE models have poor generalization capabilities due to overfitting to dataset-specific artifacts. Data quality is crucial for transfer learning, and adaptation strategy should be chosen based on data quality. Structural issues in RE benchmarks need to be addressed to enable better model transferability.

Abstract: Analysing the generalisation capabilities of relation extraction (RE) models is crucial for assessing whether they learn robust relational patterns or rely on spurious correlations. Our cross-dataset experiments find that RE models struggle with unseen data, even within similar domains. Notably, higher intra-dataset performance does not indicate better transferability, instead often signaling overfitting to dataset-specific artefacts. Our results also show that data quality, rather than lexical similarity, is key to robust transfer, and the choice of optimal adaptation strategy depends on the quality of data available: while fine-tuning yields the best cross-dataset performance with high-quality data, few-shot in-context learning (ICL) is more effective with noisier data. However, even in these cases, zero-shot baselines occasionally outperform all cross-dataset results. Structural issues in RE benchmarks, such as single-relation per sample constraints and non-standardised negative class definitions, further hinder model transferability.

Method: Formalizes alignment as importance sampling: base model as proposal distribution, alignment module as importance weight estimator. Uses sequence-level training for alignment module and iterative token-level decoding to reduce latency.

Result: Outperforms baselines on instruction following, domain adaptation, and preference optimization tasks across two leading open-source LLMs.

Conclusion: RAM provides flexible, scalable alignment without retraining base models, enabling efficient adaptation of LLMs to diverse applications.

Abstract: The widespread adoption of large language models (LLMs) across industries has increased the demand for high-quality and customizable outputs. However, traditional alignment methods often require retraining large pretrained models, making it difficult to quickly adapt and optimize LLMs for diverse applications. To address this limitation, we propose a novel \textit{Residual Alignment Model} (\textit{RAM}) that formalizes the alignment process as a type of importance sampling. In this framework, the unaligned upstream model serves as the proposal distribution, while the alignment process is framed as secondary sampling based on an autoregressive alignment module that acts as an estimator of the importance weights. This design enables a natural detachment of the alignment module from the target aligned model, improving flexibility and scalability. Based on this model, we derive an efficient sequence-level training strategy for the alignment module, which operates independently of the proposal module. Additionally, we develop a resampling algorithm with iterative token-level decoding to address the common first-token latency issue in comparable methods. Experimental evaluations on two leading open-source LLMs across diverse tasks, including instruction following, domain adaptation, and preference optimization, demonstrate that our approach consistently outperforms baseline models.

Method: VAT prompts Multimodal Large Language Models (MLLMs) with visual abstracts instead of explicit verbal thoughts or elaborate guidance. This approach encourages models to focus on essential visual elements, concepts, and structural features by undermining redundant information, unlike CoT and tool-using approaches that add complexity through verbose intermediate steps.

Result: VAT consistently empowers different MLLMs in visual perception and reasoning tasks, achieving an average gain of 2.21% over GPT-5 baseline, surpassing CoT’s gain. VAT also spends fewer tokens while achieving higher performance.

Conclusion: VAT demonstrates the effectiveness of visual abstract thinking for enhancing multimodal task performance in MLLMs. The findings highlight the value of exploring diverse reasoning paradigms inspired by human cognition, suggesting visual abstraction as a more efficient alternative to verbose reasoning methods.

Abstract: Images usually convey richer detail than text, but often include redundant information, which potentially downgrades multimodal reasoning performance. When faced with lengthy or complex messages, humans tend to employ abstract thinking to convert them into simple and concise abstracts. Inspired by this cognitive strategy, we introduce a novel paradigm to elicit the ability to Think with Visual Abstract (VAT), by prompting Multimodal Large Language Models (MLLMs) with visual abstract instead of explicit verbal thoughts or elaborate guidance, permitting a more efficient visual reasoning mechanism via concentrated perception. VAT encourages models to focus on more essential visual elements, concepts and structural features by undermining redundant information compared with explicit thinking methods, such as Chain-of-thought (CoT) and tool-using approaches, that increase the complexity of reasoning process via inserting verbose intermediate steps and external knowledge. Experimental results show that VAT consistently empowers different MLLMs in visual perception and reasoning tasks. VAT achieves an average gain of $2.21%$ over GPT-5 baseline, surpassing the gain of CoT, demonstrating that VAT better enhances multimodal task performance of MLLMs. Additionally, VAT spends fewer tokens while achieving higher performance. These findings highlight the effectiveness of visual abstract thinking and encourage further exploration of more diverse reasoning paradigms from the perspective of human cognition.

Method: Developed SNS-Bench-VL benchmark with 4,001 carefully curated multimodal question-answer pairs across 8 tasks (note comprehension, user engagement analysis, information retrieval, personalized recommendation, etc.) using images and text. Evaluated over 25 state-of-the-art multimodal LLMs on single-choice, multiple-choice, and open-ended tasks.

Result: Evaluation of over 25 multimodal LLMs revealed persistent challenges in multimodal social context comprehension. The benchmark provides comprehensive assessment capabilities for real-world social media scenarios.

Conclusion: SNS-Bench-VL addresses the gap in multimodal SNS evaluation and aims to inspire future research towards robust, context-aware, and human-aligned multimodal intelligence for next-generation social networking services.

Abstract: With the increasing integration of visual and textual content in Social Networking Services (SNS), evaluating the multimodal capabilities of Large Language Models (LLMs) is crucial for enhancing user experience, content understanding, and platform intelligence. Existing benchmarks primarily focus on text-centric tasks, lacking coverage of the multimodal contexts prevalent in modern SNS ecosystems. In this paper, we introduce SNS-Bench-VL, a comprehensive multimodal benchmark designed to assess the performance of Vision-Language LLMs in real-world social media scenarios. SNS-Bench-VL incorporates images and text across 8 multimodal tasks, including note comprehension, user engagement analysis, information retrieval, and personalized recommendation. It comprises 4,001 carefully curated multimodal question-answer pairs, covering single-choice, multiple-choice, and open-ended tasks. We evaluate over 25 state-of-the-art multimodal LLMs, analyzing their performance across tasks. Our findings highlight persistent challenges in multimodal social context comprehension. We hope SNS-Bench-VL will inspire future research towards robust, context-aware, and human-aligned multimodal intelligence for next-generation social networking services.

Method: The author offers reflections and proposes a “duality” hypothesis based on recent studies and user experience, suggesting LLMs’ translation abilities originate from two types of pre-training data internalized differently.

Result: The paper presents a working hypothesis about the dual sources of LLMs’ translation capabilities and discusses prospects for empirical testing of this hypothesis.

Conclusion: Understanding LLMs’ translation abilities has implications for reconceptualizing translation (both human and machine) in the deep learning era, with the proposed duality hypothesis offering a framework for future empirical investigation.

Abstract: Large Language Models (LLMs) excel in translation among other things, demonstrating competitive performance for many language pairs in zero- and few-shot settings. But unlike dedicated neural machine translation models, LLMs are not trained on any translation-related objective. What explains their remarkable translation abilities? Are these abilities grounded in “incidental bilingualism” (Briakou et al. 2023) in training data? Does instruction tuning contribute to it? Are LLMs capable of aligning and leveraging semantically identical or similar monolingual contents from different corners of the internet that are unlikely to fit in a single context window? I offer some reflections on this topic, informed by recent studies and growing user experience. My working hypothesis is that LLMs’ translation abilities originate in two different types of pre-training data that may be internalized by the models in different ways. I discuss the prospects for testing the “duality” hypothesis empirically and its implications for reconceptualizing translation, human and machine, in the age of deep learning.

Method: Proposes Formal Description of Visualization (FDV) - a structured textual representation of charts that enables LLMs to learn and generate visualizations. Builds Multimodal DeepResearcher framework with four stages: researching, exemplar report textualization, planning, and multimodal report generation.

Result: Developed MultimodalReportBench with 100 diverse topics and 5 metrics for evaluation. Multimodal DeepResearcher achieves 82% overall win rate over baseline using Claude 3.7 Sonnet model, demonstrating effectiveness across models and evaluation methods.

Conclusion: The proposed framework successfully addresses the challenge of automated multimodal report generation, enabling LLMs to create comprehensive reports with integrated visualizations through structured representation and agentic pipeline design.

Abstract: Visualizations play a crucial part in effective communication of concepts and information. Recent advances in reasoning and retrieval augmented generation have enabled Large Language Models (LLMs) to perform deep research and generate comprehensive reports. Despite its progress, existing deep research frameworks primarily focus on generating text-only content, leaving the automated generation of interleaved texts and visualizations underexplored. This novel task poses key challenges in designing informative visualizations and effectively integrating them with text reports. To address these challenges, we propose Formal Description of Visualization (FDV), a structured textual representation of charts that enables LLMs to learn from and generate diverse, high-quality visualizations. Building on this representation, we introduce Multimodal DeepResearcher, an agentic framework that decomposes the task into four stages: (1) researching, (2) exemplar report textualization, (3) planning, and (4) multimodal report generation. For the evaluation of generated multimodal reports, we develop MultimodalReportBench, which contains 100 diverse topics served as inputs along with 5 dedicated metrics. Extensive experiments across models and evaluation methods demonstrate the effectiveness of Multimodal DeepResearcher. Notably, utilizing the same Claude 3.7 Sonnet model, Multimodal DeepResearcher achieves an 82% overall win rate over the baseline method.

Method: Created Pet-Bench benchmark with two dimensions: self-interaction (self-evolution and developmental behaviors) and human-interaction. Includes diverse tasks such as intelligent scheduling, memory-based dialogues, and psychological conversations with over 7,500 interaction instances.

Result: Evaluation of 28 LLMs shows significant performance variations correlated with model size and inherent capabilities, highlighting the need for specialized optimization in virtual pet companionship applications.

Conclusion: Pet-Bench provides a foundational resource for benchmarking pet-related LLM abilities and advancing emotionally immersive human-pet interactions, addressing a gap in systematic evaluation of LLMs for virtual companionship.

Abstract: As interest in using Large Language Models for interactive and emotionally rich experiences grows, virtual pet companionship emerges as a novel yet underexplored application. Existing approaches focus on basic pet role-playing interactions without systematically benchmarking LLMs for comprehensive companionship. In this paper, we introduce Pet-Bench, a dedicated benchmark that evaluates LLMs across both self-interaction and human-interaction dimensions. Unlike prior work, Pet-Bench emphasizes self-evolution and developmental behaviors alongside interactive engagement, offering a more realistic reflection of pet companionship. It features diverse tasks such as intelligent scheduling, memory-based dialogues, and psychological conversations, with over 7,500 interaction instances designed to simulate pet behaviors. Evaluation of 28 LLMs reveals significant performance variations linked to model size and inherent capabilities, underscoring the need for specialized optimization in this domain. Pet-Bench serves as a foundational resource for benchmarking pet-related LLM abilities and advancing emotionally immersive human-pet interactions.

Method: Used the Magpie framework to annotate samples from both datasets with quality metrics (turn structure, task category, input quality, response quality). Analyzed structural and qualitative differences, then designed a principled curation recipe to create TuluTalk - a new optimized dataset mixture.

Result: TuluTalk contains 14% fewer samples than either source dataset but matches or exceeds their performance on key benchmarks. The analysis revealed structural differences between the datasets and provided actionable insights for effective post-training data construction.

Conclusion: The study demonstrates that systematic data quality analysis enables creation of more efficient post-training datasets. The publicly released annotated datasets and TuluTalk mixture support future research in transparent, resource-efficient LLM alignment.

Abstract: Recent work on large language models (LLMs) has increasingly focused on post-training and alignment with datasets curated to enhance instruction following, world knowledge, and specialized skills. However, most post-training datasets used in leading open- and closed-source LLMs remain inaccessible to the public, with limited information about their construction process. This lack of transparency has motivated the recent development of open-source post-training corpora. While training on these open alternatives can yield performance comparable to that of leading models, systematic comparisons remain challenging due to the significant computational cost of conducting them rigorously at scale, and are therefore largely absent. As a result, it remains unclear how specific samples, task types, or curation strategies influence downstream performance when assessing data quality. In this work, we conduct the first comprehensive side-by-side analysis of two prominent open post-training datasets: Tulu-3-SFT-Mix and SmolTalk. Using the Magpie framework, we annotate each sample with detailed quality metrics, including turn structure (single-turn vs. multi-turn), task category, input quality, and response quality, and we derive statistics that reveal structural and qualitative similarities and differences between the two datasets. Based on these insights, we design a principled curation recipe that produces a new data mixture, TuluTalk, which contains 14% fewer samples than either source dataset while matching or exceeding their performance on key benchmarks. Our findings offer actionable insights for constructing more effective post-training datasets that improve model performance within practical resource limits. To support future research, we publicly release both the annotated source datasets and our curated TuluTalk mixture.

Method: Benchmarking study of multiple reasoning-focused LLMs (focusing on DeepSeek models) across 15 Arabic NLP tasks using zero-shot, few-shot, and fine-tuning strategies (including LoRA-based fine-tuning).

Result: 1) 3 in-context examples boost classification tasks by 13+ F1 points (sentiment analysis: 35.3%→87.5%; paraphrase detection: 56.1%→87.0%). 2) DeepSeek outperforms GPT-4-mini by 12 F1 points on complex inference tasks in zero-shot. 3) LoRA fine-tuning yields up to 8 additional F1/BLEU points vs. model scale increases.

Conclusion: Careful few-shot selection, reasoning-focused architectures like DeepSeek, and parameter-efficient fine-tuning (LoRA) are highly effective for Arabic NLP, addressing its linguistic challenges and advancing Arabic language AI capabilities.

Abstract: Large language models (LLMs) have shown remarkable progress in reasoning abilities and general natural language processing (NLP) tasks, yet their performance on Arabic data, characterized by rich morphology, diverse dialects, and complex script, remains underexplored. This paper presents a comprehensive benchmarking study of multiple reasoning-focused LLMs, with a special emphasis on the newly introduced DeepSeek models, across a suite of fifteen Arabic NLP tasks. We experiment with various strategies, including zero-shot, few-shot, and fine-tuning. This allows us to systematically evaluate performance on datasets covering a range of applications to examine their capacity for linguistic reasoning under different levels of complexity. Our experiments reveal several key findings. First, carefully selecting just three in-context examples delivers an average uplift of over 13 F1 points on classification tasks-boosting sentiment analysis from 35.3% to 87.5% and paraphrase detection from 56.1% to 87.0%. Second, reasoning-focused DeepSeek architectures outperform a strong GPT o4-mini baseline by an average of 12 F1 points on complex inference tasks in the zero-shot setting. Third, LoRA-based fine-tuning yields up to an additional 8 points in F1 and BLEU compared to equivalent increases in model scale. The code is available at https://github.com/gufranSabri/deepseek-evals

Method: Multipole Attention clusters semantically similar key vectors, uses centroids to identify important tokens for exact attention while approximating others, with fast cluster updates for newly generated tokens.

Result: Maintains accuracy on complex reasoning tasks with aggressive attention sparsity, achieves up to 4.5× speedup for attention in long-context applications (tested on Qwen-8B).

Conclusion: Multipole Attention effectively accelerates autoregressive reasoning in LRMs by balancing exact and approximate attention through clustering, enabling efficient long-context reasoning without accuracy loss.

Abstract: Large Reasoning Models (LRMs) have shown promising accuracy improvements on complex problem-solving tasks. While these models have attained high accuracy by leveraging additional computation at test time, they need to generate long chain-of-thought reasoning in order to think before answering, which requires generating thousands of tokens. While sparse attention methods can help reduce the KV cache pressure induced by this long autoregressive reasoning, these methods can introduce errors which disrupt the reasoning process. Additionally, prior methods often pre-process the input to make it easier to identify the important prompt tokens when computing attention during generation, and this pre-processing is challenging to perform online for newly generated reasoning tokens. Our work addresses these challenges by introducing Multipole Attention, which accelerates autoregressive reasoning by only computing exact attention for the most important tokens, while maintaining approximate representations for the remaining tokens. Our method first performs clustering to group together semantically similar key vectors, and then uses the cluster centroids both to identify important key vectors and to approximate the remaining key vectors in order to retain high accuracy. We design a fast cluster update process to quickly re-cluster the input and previously generated tokens, thereby allowing for accelerating attention to the previous output tokens. We evaluate our method using emerging LRMs such as Qwen-8B, demonstrating that our approach can maintain accuracy on complex reasoning tasks even with aggressive attention sparsity settings. We also provide kernel implementations to demonstrate the practical efficiency gains from our method, achieving up to 4.5$\times$ speedup for attention in long-context reasoning applications. Our code is available at https://github.com/SqueezeAILab/MultipoleAttention.

Method: Developed RealHiTBench benchmark with diverse complex tables in LaTeX, HTML, and PNG formats, and created TreeThinker - a tree-based pipeline that organizes hierarchical headers into tree structures for enhanced tabular reasoning.

Result: Experiments with 25 state-of-the-art LLMs show RealHiTBench is challenging, and TreeThinker validates the importance of improving LLMs’ perception of table hierarchies for better tabular reasoning.

Conclusion: RealHiTBench provides a comprehensive challenging benchmark for tabular data evaluation, and TreeThinker demonstrates the value of hierarchical structure perception, inspiring further research in tabular data reasoning.

Abstract: With the rapid advancement of Large Language Models (LLMs), there is an increasing need for challenging benchmarks to evaluate their capabilities in handling complex tabular data. However, existing benchmarks are either based on outdated data setups or focus solely on simple, flat table structures. In this paper, we introduce RealHiTBench, a comprehensive benchmark designed to evaluate the performance of both LLMs and Multimodal LLMs (MLLMs) across a variety of input formats for complex tabular data, including LaTeX, HTML, and PNG. RealHiTBench also includes a diverse collection of tables with intricate structures, spanning a wide range of task types. Our experimental results, using 25 state-of-the-art LLMs, demonstrate that RealHiTBench is indeed a challenging benchmark. Moreover, we also develop TreeThinker, a tree-based pipeline that organizes hierarchical headers into a tree structure for enhanced tabular reasoning, validating the importance of improving LLMs’ perception of table hierarchies. We hope that our work will inspire further research on tabular data reasoning and the development of more robust models. The code and data are available at https://github.com/cspzyy/RealHiTBench.

Method: Created SoMi-ToM benchmark with multimodal interaction data from SoMi environment, featuring multi-level evaluation: first-person (real-time state inference) and third-person (goal/behavior inference) perspectives.

Result: LVLMs perform significantly worse than humans: 40.1% accuracy gap in first-person evaluation and 26.4% gap in third-person evaluation on dataset with 35 videos, 363 images, and 1225 expert-annotated questions.

Conclusion: Future LVLMs need substantial improvement in ToM capabilities for embodied, complex social interactions, as current models cannot match human performance in dynamic social understanding.

Abstract: Humans continuously infer the states, goals, and behaviors of others by perceiving their surroundings in dynamic, real-world social interactions. However, most Theory of Mind (ToM) benchmarks only evaluate static, text-based scenarios, which have a significant gap compared to real interactions. We propose the SoMi-ToM benchmark, designed to evaluate multi-perspective ToM in embodied multi-agent complex social interactions. This benchmark is based on rich multimodal interaction data generated by the interaction environment SoMi, covering diverse crafting goals and social relationships. Our framework supports multi-level evaluation: (1) first-person evaluation provides multimodal (visual, dialogue, action, etc.) input from a first-person perspective during a task for real-time state inference, (2) third-person evaluation provides complete third-person perspective video and text records after a task for goal and behavior inference. This evaluation method allows for a more comprehensive examination of a model’s ToM capabilities from both the subjective immediate experience and the objective global observation. We constructed a challenging dataset containing 35 third-person perspective videos, 363 first-person perspective images, and 1225 expert-annotated multiple-choice questions (three options). On this dataset, we systematically evaluated the performance of human subjects and several state-of-the-art large vision-language models (LVLMs). The results show that LVLMs perform significantly worse than humans on SoMi-ToM: the average accuracy gap between humans and models is 40.1% in first-person evaluation and 26.4% in third-person evaluation. This indicates that future LVLMs need to further improve their ToM capabilities in embodied, complex social interactions.

Method: Propose InfoRidge, an information-theoretic framework to characterize predictive information (mutual information between hidden representations and target outputs) across depth. Also introduce residual scaling coefficients as trainable scalar parameters applied to each residual block, serving as functional probes to assess relative importance of individual transformer layers.

Result: Experiments reveal consistent non-monotonic trend: predictive information peaks in upper-middle layers (forming a “generalization ridge”) before declining in final layers, reflecting transition between generalization and memorization. Under distribution shift, models downweight final layers and increasingly rely on intermediate layers.

Conclusion: Intermediate transformer layers play critical role in supporting generalization, while final layers focus more on memorization. These findings offer new insights into internal mechanisms of transformers and highlight importance of intermediate layers for robust generalization.

Abstract: Transformer-based language models have achieved state-of-the-art performance in natural language generation (NLG), yet their internal mechanisms for synthesizing task-relevant information remain insufficiently understood. While prior studies suggest that intermediate layers often yield more generalizable representations than final layers, how this generalization ability emerges and propagates across layers during training remains unclear. To address this gap, we propose InfoRidge, an information-theoretic framework, to characterize how predictive information-the mutual information between hidden representations and target outputs-varies across depth. Estimating this quantity enables us to trace the flow of task-relevant information throughout the model during training. Our experiments across various models and datasets reveal a consistent non-monotonic trend: predictive information peaks in upper-middle layers-forming a generalization ridge-before declining in final layers, reflecting a transition between generalization and memorization. To further investigate this phenomenon, we introduce residual scaling coefficients-trainable scalar parameters applied to each residual block-which serve as functional probes for assessing the relative importance of individual transformer layers. These coefficients reveal that, under distribution shift, models downweight final layers and increasingly rely on intermediate layers, highlighting their role in generalization. Together, these findings offer new insights into the internal mechanisms of transformers and underscore the critical role of intermediate layers in supporting generalization.

Method: Identified separate “harmfulness direction” and “refusal direction” in LLMs’ internal representations. Used steering experiments along these directions to analyze effects. Developed “Latent Guard” using latent harmfulness representations as intrinsic safeguard. Compared with Llama Guard 3 8B across jailbreak methods.

Result: Found harmfulness is encoded separately from refusal. Steering along harmfulness direction changes harm perception, while refusal direction only affects refusal behavior. Jailbreaks reduce refusal signals without changing harmfulness beliefs. Adversarial finetuning minimally impacts harmfulness beliefs. Latent Guard performs comparably or better than Llama Guard 3 8B.

Conclusion: LLMs’ internal understanding of harmfulness is more robust than their refusal decisions. Harmfulness is a distinct internal concept from refusal, enabling creation of robust intrinsic safeguards like Latent Guard that resist finetuning attacks and reduce over-refusals.

Abstract: LLMs are trained to refuse harmful instructions, but do they truly understand harmfulness beyond just refusing? Prior work has shown that LLMs’ refusal behaviors can be mediated by a one-dimensional subspace, i.e., a refusal direction. In this work, we identify a new dimension to analyze safety mechanisms in LLMs, i.e., harmfulness, which is encoded internally as a separate concept from refusal. There exists a harmfulness direction that is distinct from the refusal direction. As causal evidence, steering along the harmfulness direction can lead LLMs to interpret harmless instructions as harmful, but steering along the refusal direction tends to elicit refusal responses directly without reversing the model’s judgment on harmfulness. Furthermore, using our identified harmfulness concept, we find that certain jailbreak methods work by reducing the refusal signals without reversing the model’s internal belief of harmfulness. We also find that adversarially finetuning models to accept harmful instructions has minimal impact on the model’s internal belief of harmfulness. These insights lead to a practical safety application: The model’s latent harmfulness representation can serve as an intrinsic safeguard (Latent Guard) for detecting unsafe inputs and reducing over-refusals that is robust to finetuning attacks. For instance, our Latent Guard achieves performance comparable to or better than Llama Guard 3 8B, a dedicated finetuned safeguard model, across different jailbreak methods. Our findings suggest that LLMs’ internal understanding of harmfulness is more robust than their refusal decision to diverse input instructions, offering a new perspective to study AI safety.

Method: Four key components: (1) Section-aware pretraining to learn logical flow from Findings to Impression, (2) Fine-tuning with Proximal Policy Optimization using CheXbert F1 score as reward, (3) Controlled decoding that completes Findings before synthesizing Impression, (4) Multi-view fusion via vision-transformer encoder. Inference uses next-token forcing and report-level re-ranking.

Result: Experimental results on MIMIC-CXR dataset demonstrate superior clinical efficacy and outperforms existing baselines on clinical efficacy scores.

Conclusion: CLARIFID effectively addresses clinical reliability in radiology report generation by mirroring expert workflow and optimizing for diagnostic correctness through structured training and inference strategies.

Abstract: Automatic generation of radiology reports has the potential to alleviate radiologists’ significant workload, yet current methods struggle to deliver clinically reliable conclusions. In particular, most prior approaches focus on producing fluent text without effectively ensuring the factual correctness of the reports and often rely on single-view images, limiting diagnostic comprehensiveness. We propose CLARIFID, a novel framework that directly optimizes diagnostic correctness by mirroring the two-step workflow of experts. Specifically, CLARIFID (1) learns the logical flow from Findings to Impression through section-aware pretraining, (2) is fine-tuned with Proximal Policy Optimization in which the CheXbert F1 score of the Impression section serves as the reward, (3) employs controlled decoding that completes “Findings” before synthesizing the “Impression”, and (4) fuses multiple chest X-ray views via a vision-transformer-based multi-view encoder. During inference, we apply a next-token forcing strategy followed by report-level re-ranking, ensuring that the model first produces a comprehensive “Findings” section before synthesizing the “Impression” and thereby preserving coherent clinical reasoning. Experimental results on the MIMIC-CXR dataset demonstrate that our method achieves superior clinical efficacy and outperforms existing baselines on clinical efficacy scores.

Method: Surveyed ACL Anthology papers from past 6 years with “diversity” or “diverse” in title, analyzed how diversity is quantified, and proposed a unified taxonomy using Stirling’s 3D framework (variety, balance, disparity) to categorize diversity measurement approaches.

Result: Found wide range of specialized diversity quantification methods with inconsistent terminology. Successfully applied Stirling’s framework to systematize diversity measurement in NLP, revealing trends and patterns in how diversity is approached across different NLP settings.

Conclusion: This study provides a foundation for better formalization of diversity in NLP, which should lead to improved understanding of the concept and better comparability between different approaches, moving beyond ad hoc methods toward more theoretically grounded diversity measurement.

Abstract: The concept of diversity has received increased consideration in Natural Language Processing (NLP) in recent years. This is due to various motivations like promoting and inclusion, approximating human linguistic behavior, and increasing systems’ performance. Diversity has however often been addressed in an ad hoc manner in NLP, and with few explicit links to other domains where this notion is better theorized. We survey articles in the ACL Anthology from the past 6 years, with “diversity” or “diverse” in their title. We find a wide range of settings in which diversity is quantified, often highly specialized and using inconsistent terminology. We put forward a unified taxonomy of why, what on, where, and how diversity is measured in NLP. Diversity measures are cast upon a unified framework from ecology and economy (Stirling, 2007) with 3 dimensions of diversity: variety, balance and disparity. We discuss the trends which emerge due to this systematized approach. We believe that this study paves the way towards a better formalization of diversity in NLP, which should bring a better understanding of this notion and a better comparability between various approaches.

Method: Propose Flat Minima LoRA (FMLoRA) and its efficient version EFMLoRA to seek flat minima for LoRA. Theoretically demonstrate that perturbations in full parameter space can be transferred to low-rank subspace, eliminating interference from perturbations across multiple matrices in low-rank subspace.

Result: EFMLoRA achieves optimization efficiency comparable to LoRA while attaining comparable or better performance. On GLUE dataset with RoBERTa-large, EFMLoRA outperforms LoRA by 1.0% and full fine-tuning by 0.5% on average. On Qwen-VL-Chat vision-language model, improvements of 1.5% on SQA and 1.0% on VizWiz datasets.

Conclusion: The generalization of LoRA is closely related to sharpness, which was omitted by previous methods. EFMLoRA provides an efficient way to find flat minima for LoRA while maintaining optimization efficiency.

Abstract: Little research explores the correlation between the expressive ability and generalization ability of the low-rank adaptation (LoRA). Sharpness-Aware Minimization (SAM) improves model generalization for both Convolutional Neural Networks (CNNs) and Transformers by encouraging convergence to locally flat minima. However, the connection between sharpness and generalization has not been fully explored for LoRA due to the lack of tools to either empirically seek flat minima or develop theoretical methods. In this work, we propose Flat Minima LoRA (FMLoRA) and its efficient version, i.e., EFMLoRA, to seek flat minima for LoRA. Concretely, we theoretically demonstrate that perturbations in the full parameter space can be transferred to the low-rank subspace. This approach eliminates the potential interference introduced by perturbations across multiple matrices in the low-rank subspace. Our extensive experiments on large language models and vision-language models demonstrate that EFMLoRA achieves optimize efficiency comparable to that of LoRA while simultaneously attaining comparable or even better performance. For example, on the GLUE dataset with RoBERTa-large, EFMLoRA outperforms LoRA and full fine-tuning by 1.0% and 0.5% on average, respectively. On vision-language models, e.g., Qwen-VL-Chat, there are performance improvements of 1.5% and 1.0% on the SQA and VizWiz datasets, respectively. These empirical results also verify that the generalization of LoRA is closely related to sharpness, which is omitted by previous methods.

Method: DURIT (Decoupled Understanding from Reasoning via Iterative Training) - three-step algorithm: (1) maps natural language problems to canonical space via RL, (2) aligns reasoning trajectories via self-distillation, (3) trains reasoning policies in problem space. Mapper and reasoner co-trained in alternating loop.

Result: DURIT substantially improves SLMs’ performance on both in-domain and out-of-domain mathematical and logical reasoning tasks, and improves reasoning robustness.

Conclusion: Decoupling understanding from reasoning is an effective strategy for strengthening SLMs, enabling them to focus on reasoning over standardized inputs free from linguistic variability.

Abstract: Despite recent advances in the reasoning capabilities of Large Language Models (LLMs), improving the reasoning ability of Small Language Models (SLMs, e.g., up to 1.5B parameters) remains challenging. A key obstacle lies in the complexity and variability of natural language: essentially equivalent problems often appear in diverse surface forms, often obscured by redundant or distracting details. This imposes a dual burden on SLMs: they must first extract the core problem from complex linguistic input, and then perform reasoning based on that understanding. The resulting vast and noisy problem space hinders optimization, particularly for models with limited capacity. To address this, we propose a new framework that decouples understanding from reasoning by mapping natural language problems into a canonical problem space-a semantically simplified yet expressive domain. This enables SLMs to focus on reasoning over standardized inputs, free from linguistic variability. Within this framework, we introduce DURIT (Decoupled Understanding from Reasoning via Iterative Training), a three-step algorithm that iteratively: (1) mapping natural language problems via reinforcement learning, (2) aligns reasoning trajectories through self-distillation, and (3) trains reasoning policies in the problem space. The mapper and reasoner are co-trained in an alternating loop throughout this process. Experiments show that DURIT substantially improves SLMs’ performance on both in-domain and out-of-domain mathematical and logical reasoning tasks. Beyond improving reasoning capabilities, DURIT also improves the robustness of reasoning, validating decoupling understanding from reasoning as an effective strategy for strengthening SLMs.

Method: The authors developed SproutBench, an evaluation suite with 1,283 developmentally grounded adversarial prompts covering risks across three age groups (0-6, 7-12, 13-18). They empirically evaluated 47 diverse LLMs using this benchmark.

Result: The evaluation revealed substantial safety vulnerabilities in current LLMs, with robust correlations between safety dimensions (e.g., Safety and Risk Prevention) and an inverse relationship between Interactivity and Age Appropriateness.

Conclusion: The findings highlight critical gaps in current LLM safety for minors and provide practical guidelines for advancing child-centric AI design and deployment to better protect children and adolescents.

Abstract: The rapid proliferation of large language models (LLMs) in applications targeting children and adolescents necessitates a fundamental reassessment of prevailing AI safety frameworks, which are largely tailored to adult users and neglect the distinct developmental vulnerabilities of minors. This paper highlights key deficiencies in existing LLM safety benchmarks, including their inadequate coverage of age-specific cognitive, emotional, and social risks spanning early childhood (ages 0–6), middle childhood (7–12), and adolescence (13–18). To bridge these gaps, we introduce SproutBench, an innovative evaluation suite comprising 1,283 developmentally grounded adversarial prompts designed to probe risks such as emotional dependency, privacy violations, and imitation of hazardous behaviors. Through rigorous empirical evaluation of 47 diverse LLMs, we uncover substantial safety vulnerabilities, corroborated by robust inter-dimensional correlations (e.g., between Safety and Risk Prevention) and a notable inverse relationship between Interactivity and Age Appropriateness. These insights yield practical guidelines for advancing child-centric AI design and deployment.

Method: Comprehensive evaluation of GPT-OSS models (20B & 120B parameters) against six contemporary open source LLMs (14.7B-235B parameters) across ten benchmarks covering general knowledge, math, code, multilingual, and conversational tasks. Used standardized inference settings, statistical validation with McNemar’s test, and effect size analysis.

Result: GPT-OSS-20B consistently outperforms GPT-OSS-120B on several benchmarks (HumanEval, MMLU) despite requiring substantially less memory and energy. Both models show mid-tier overall performance with strengths in code generation and weaknesses in multilingual tasks.

Conclusion: Scaling in sparse architectures may not yield proportional performance gains, highlighting the need for better optimization strategies and more efficient model selection for open source deployments.

Abstract: In August 2025, OpenAI released GPT-OSS models, its first open weight large language models since GPT-2 in 2019, comprising two mixture of experts architectures with 120B and 20B parameters. We evaluated both variants against six contemporary open source large language models ranging from 14.7B to 235B parameters, representing both dense and sparse designs, across ten benchmarks covering general knowledge, mathematical reasoning, code generation, multilingual understanding, and conversational ability. All models were tested in unquantised form under standardised inference settings, with statistical validation using McNemars test and effect size analysis. Results show that gpt-oss-20B consistently outperforms gpt-oss-120B on several benchmarks, such as HumanEval and MMLU, despite requiring substantially less memory and energy per response. Both models demonstrate mid-tier overall performance within the current open source landscape, with relative strength in code generation and notable weaknesses in multilingual tasks. These findings provide empirical evidence that scaling in sparse architectures may not yield proportional performance gains, underscoring the need for further investigation into optimisation strategies and informing more efficient model selection for future open source deployments. More details and evaluation scripts are available at https://ai-agent-lab.github.io/gpt-oss (Project Webpage).

Method: Fine-tuning LLMs (Llama-3.1-8B and Qwen-2.5-7B) on native speakers’ word-association norms from US (English) and China (Mandarin), using supervised fine-tuning and preference optimization based on cognitive psychology findings.

Result: Significant improvements in lexical alignment (16-20% English, 43-165% Mandarin on Precision@5) and cultural value shifts. Fine-tuned Qwen nearly doubles response alignment with Chinese values on divergent questions (13 to 25). Trained 7-8B models match or exceed vanilla 70B baselines.

Conclusion: A few million culture-grounded associations can achieve value alignment without expensive retraining, highlighting the promise of human cognition-grounded approaches for improving cultural alignment in AI models.

Abstract: Large language models (LLMs) exhibit cultural bias from overrepresented viewpoints in training data, yet cultural alignment remains a challenge due to limited cultural knowledge and a lack of exploration into effective learning approaches. We introduce a cost-efficient and cognitively grounded method: fine-tuning LLMs on native speakers’ word-association norms, leveraging cognitive psychology findings that such associations capture cultural knowledge. Using word association datasets from native speakers in the US (English) and China (Mandarin), we train Llama-3.1-8B and Qwen-2.5-7B via supervised fine-tuning and preference optimization. We evaluate models’ cultural alignment through a two-tier evaluation framework that spans lexical associations and cultural value alignment using the World Values Survey. Results show significant improvements in lexical alignment (16-20% English, 43-165% Mandarin on Precision@5) and high-level cultural value shifts. On a subset of 50 questions where US and Chinese respondents diverge most, fine-tuned Qwen nearly doubles its response alignment with Chinese values (13 to 25). Remarkably, our trained 7-8B models match or exceed vanilla 70B baselines, demonstrating that a few million of culture-grounded associations achieve value alignment without expensive retraining. Our work highlights both the promise and the need for future research grounded in human cognition in improving cultural alignment in AI models.

Method: Proposes SvS (Self-play with Variational problem Synthesis) - an online strategy that uses the policy’s correct solutions to synthesize variational problems while keeping reference answers identical to originals, maintaining policy entropy during training.

Result: SvS substantially improves Pass@k compared to standard RLVR, achieving 18.3% and 22.8% absolute gains in Pass@32 on AIME24 and AIME25 benchmarks, with consistent improvements across 12 reasoning benchmarks and model sizes from 3B to 32B.

Conclusion: SvS effectively mitigates entropy collapse in RLVR training, maintains generation diversity, and significantly boosts Pass@k performance while demonstrating generalizability and robustness across various reasoning tasks and model scales.

Abstract: Reinforcement Learning with Verifiable Rewards (RLVR) has recently emerged as a key paradigm for post-training Large Language Models (LLMs), particularly for complex reasoning tasks. However, vanilla RLVR training has been shown to improve Pass@1 performance at the expense of policy entropy, leading to reduced generation diversity and limiting the Pass@k performance, which typically represents the upper bound of LLM reasoning capability. In this paper, we systematically analyze the policy’s generation diversity from the perspective of training problems and find that augmenting and updating training problems helps mitigate entropy collapse during training. Based on these observations, we propose an online Self-play with Variational problem Synthesis (SvS) strategy for RLVR training, which uses the policy’s correct solutions to synthesize variational problems while ensuring their reference answers remain identical to the originals. This self-improving strategy effectively maintains policy entropy during training and substantially improves Pass@k compared with standard RLVR, sustaining prolonged improvements and achieving absolute gains of 18.3% and 22.8% in Pass@32 performance on the competition-level AIME24 and AIME25 benchmarks, as well as on code generation tasks. Experiments on 12 reasoning benchmarks across varying model sizes from 3B to 32B consistently demonstrate the generalizability and robustness of SvS.

Method: Multi-stage framework: 1) prompt engineering for cold-start, 2) supervised fine-tuning with distillation from click logs, 3) Gaussian Reward Model (GaRM) to represent user preferences as probability distributions, 4) reinforcement learning with composite reward function, enhanced by out-of-distribution regularization and two-stage reward fusion.

Result: Framework significantly outperforms baselines on automatic and human evaluations, achieving 34% relative increase in user engagement (click-through rate) in live A/B tests.

Conclusion: The proposed progressive alignment framework effectively captures nuanced user preferences and improves generative query suggestion performance through probabilistic preference modeling and robust RL optimization.

Abstract: Generative query suggestion using large language models offers a powerful way to enhance conversational systems, but aligning outputs with nuanced user preferences remains a critical challenge. To address this, we introduce a multi-stage framework designed for progressive alignment between the generation policy and user intent. Our pipeline begins with prompt engineering as a cold-start strategy, followed by the Supervised Fine-Tuning stage, in which we introduce a distillation method on click logs to create a robust foundational model. To better model user preferences while capturing their inherent uncertainty, we develop a Gaussian Reward Model (GaRM) that represents user preferences as probability distributions rather than point estimates. Finally, we employ reinforcement learning to align the generation policy with these preferences, guided by a composite reward function that integrates GaRM with auxiliary heuristics to mitigate reward hacking. To maintain training stability, this process is enhanced by a novel out-of-distribution regularization method and a two-stage reward fusion technique. Extensive experiments demonstrate that our framework significantly outperforms baselines on both automatic and human evaluations and yields a 34% relative increase in user engagement as measured by click-through rate in live A/B tests.

Method: Leverages Sparse Autoencoders (SAEs) to extract semantically meaningful internal representations from LLMs, identifying interpretable concepts associated with different jailbreak themes to build robust safety guardrails.

Result: The approach provides evidence for a shared activation geometry for jailbreak attacks in the representation space, enabling fully explainable and generalizable defenses without sacrificing model capabilities or requiring further fine-tuning.

Conclusion: ConceptGuard demonstrates that mechanistic interpretability can provide a foundation for designing more interpretable and generalizable safeguards against jailbreak attacks, offering a promising direction for LLM safety research.

Abstract: Large Language Models have found success in a variety of applications. However, their safety remains a concern due to the existence of various jailbreaking methods. Despite significant efforts, alignment and safety fine-tuning only provide a certain degree of robustness against jailbreak attacks that covertly mislead LLMs towards the generation of harmful content. This leaves them prone to a range of vulnerabilities, including targeted misuse and accidental user profiling. This work introduces \textbf{ConceptGuard}, a novel framework that leverages Sparse Autoencoders (SAEs) to identify interpretable concepts within LLM internals associated with different jailbreak themes. By extracting semantically meaningful internal representations, ConceptGuard enables building robust safety guardrails – offering fully explainable and generalizable defenses without sacrificing model capabilities or requiring further fine-tuning. Leveraging advances in the mechanistic interpretability of LLMs, our approach provides evidence for a shared activation geometry for jailbreak attacks in the representation space, a potential foundation for designing more interpretable and generalizable safeguards against attackers.

Method: Constructs synthetic multi-hop QA dataset from factual news articles using knowledge graphs, augmented through semantic clustering to recover missing links. Creates a 2D difficulty matrix combining generator-side (reasoning depth) and retriever-side (semantic distance) difficulty.

Result: Error rates strongly correlate with the proposed difficulty measures across multiple domains and models, validating their diagnostic utility for RAG performance analysis.

Conclusion: GRADE provides a scalable foundation for evaluating and improving multi-hop reasoning in real-world RAG applications, enabling fine-grained performance analysis.

Abstract: Retrieval-Augmented Generation (RAG) systems are widely adopted in knowledge-intensive NLP tasks, but current evaluations often overlook the structural complexity and multi-step reasoning required in real-world scenarios. These benchmarks overlook key factors such as the interaction between retrieval difficulty and reasoning depth. To address this gap, we propose GRADE, a novel evaluation framework that models task difficulty along two orthogonal dimensions: (1) reasoning depth, defined by the number of inference steps (hops), and (2) semantic distance between the query and its supporting evidence. We construct a synthetic multi-hop QA dataset from factual news articles by extracting knowledge graphs and augmenting them through semantic clustering to recover missing links, allowing us to generate diverse and difficulty-controlled queries. Central to our framework is a 2D difficulty matrix that combines generator-side and retriever-side difficulty. Experiments across multiple domains and models show that error rates strongly correlate with our difficulty measures, validating their diagnostic utility. GRADE enables fine-grained analysis of RAG performance and provides a scalable foundation for evaluating and improving multi-hop reasoning in real-world applications.

Method: Systematic re-examination of probing paradigm through controlled experiments: 1) Compare simple n-gram methods to probe performance, 2) Use semantically cleaned datasets to isolate patterns, 3) Analyze specific pattern dependencies (instructional patterns and trigger words).

Result: Probes learn superficial patterns (instructional formats and specific trigger words) rather than semantic harmfulness. Simple n-gram methods achieve comparable performance, revealing the limitations of current probing-based safety detection approaches.

Conclusion: Current probing-based approaches create a false sense of security. There’s a need to redesign both LLM safety detection models and evaluation protocols. The authors open-sourced their project to encourage responsible further research in this direction.

Abstract: Large Language Models (LLMs) can comply with harmful instructions, raising serious safety concerns despite their impressive capabilities. Recent work has leveraged probing-based approaches to study the separability of malicious and benign inputs in LLMs’ internal representations, and researchers have proposed using such probing methods for safety detection. We systematically re-examine this paradigm. Motivated by poor out-of-distribution performance, we hypothesize that probes learn superficial patterns rather than semantic harmfulness. Through controlled experiments, we confirm this hypothesis and identify the specific patterns learned: instructional patterns and trigger words. Our investigation follows a systematic approach, progressing from demonstrating comparable performance of simple n-gram methods, to controlled experiments with semantically cleaned datasets, to detailed analysis of pattern dependencies. These results reveal a false sense of security around current probing-based approaches and highlight the need to redesign both models and evaluation protocols, for which we provide further discussions in the hope of suggesting responsible further research in this direction. We have open-sourced the project at https://github.com/WangCheng0116/Why-Probe-Fails.

Method: 1) Introduce ChartVRBench benchmark to isolate and evaluate visual reasoning ability in chart understanding; 2) Propose ChartVR-3B/7B models trained with Visual Reasoning Reinforcement Finetuning (VR-RFT) strategy to strengthen genuine chart visual reasoning abilities.

Result: ChartVR achieves superior performance on ChartVRBench, outperforming even powerful proprietary models. The visual reasoning skills show strong generalization, leading to significant performance gains across diverse public chart understanding benchmarks.

Conclusion: The proposed approach successfully addresses MLLMs’ limitations in chart understanding by focusing on visual reasoning rather than visual recognition, with demonstrated effectiveness and generalization across multiple benchmarks.

Abstract: Although Multimodal Large Language Models (MLLMs) have demonstrated increasingly impressive performance in chart understanding, most of them exhibit alarming hallucinations and significant performance degradation when handling non-annotated charts. We argue that current MLLMs rely largely on visual recognition rather than visual reasoning to interpret the charts, and visual estimation of numerical values is one of the most fundamental capabilities in chart understanding that require complex visual reasoning. To prove this, we introduce ChartVRBench, a benchmark meticulously designed to isolate and evaluate visual reasoning ability in chart understanding. Furthermore, we propose ChartVR-3B/7B trained with a novel Visual Reasoning Reinforcement Finetuning (VR-RFT) strategy to strengthen genuine chart visual reasoning abilities. Extensive experiments show that ChartVR achieves superior performance on ChartVRBench, outperforming even powerful proprietary models. Moreover, the visual reasoning skills cultivated by the proposed VR-RFT demonstrate strong generalization, leading to significant performance gains across a diverse suite of public chart understanding benchmarks. The code and dataset will be publicly available upon publication.

Method: Proposes a pruning method integrated with PD disaggregation that constructs pruning and distillation sets to perform iterative block removal. Analyzes pruning sensitivity of prefill and decode stages separately to identify stage-specific removable blocks, enabling precise pruning tailored to each stage’s requirements.

Result: Extensive experiments show the method consistently achieves strong performance in both PD disaggregation and PD unified settings, with improved performance and faster inference compared to existing methods under the same settings. The approach can also be extended to other non-block pruning methods.

Conclusion: The proposed PD-disaggregation-aware pruning method effectively addresses LLM deployment constraints by providing precise block pruning that accounts for the different characteristics of prefill and decode stages, resulting in better performance and faster inference for practical deployment scenarios.

Abstract: Large Language Models (LLMs) demonstrate exceptional capabilities across various tasks, but their deployment is constrained by high computational and memory costs. Model pruning provides an effective means to alleviate these demands. However, existing methods often ignore the characteristics of prefill-decode (PD) disaggregation in practice. In this paper, we propose a pruning method that is highly integrated with PD disaggregation, enabling more precise pruning of blocks. Our approach constructs pruning and distillation sets to perform iterative block removal, obtaining better pruning solutions. Moreover, we analyze the pruning sensitivity of the prefill and decode stages and identify removable blocks specific to each stage, making it well suited for PD disaggregation deployment. Extensive experiments demonstrate our approach consistently achieves strong performance in both PD disaggregation and PD unified (non-PD disaggregation) settings, and can also be extended to other non-block pruning methods. Under the same settings, our method achieves improved performance and faster inference.

Method: Three-part approach: 1) Efficient parallel decoding mechanism to address overhead in conventional early-exiting; 2) Deep parameter sharing architecture to mitigate synchronization issues; 3) Unified framework with lightweight pretrained routers to dynamically assign optimal recursion depth per token.

Result: Establishes a new Pareto frontier between efficiency and performance by effectively optimizing for both adaptive computation and parameter efficiency within a single model.

Conclusion: Co-designing adaptive algorithms and model architectures resolves the fundamental conflict between per-token dynamism and system efficiency, enabling practical deployment of efficient LLMs.

Abstract: Large language models have achieved remarkable capabilities, but their practical deployment is hindered by significant computational costs. While adaptive computation methods like early-exiting promise to reduce these costs, they introduce a fundamental conflict: the per-token dynamism intended to save computation often creates system-level bottlenecks that can paradoxically reduce throughput in batched inference. This dissertation resolves this conflict by co-designing adaptive algorithms and model architectures to strike an optimal balance between dynamism and efficiency. To this end, our work first addresses critical sources of overhead in conventional early-exiting by proposing an efficient parallel decoding mechanism. We then show that deep parameter sharing provides an architectural foundation that not only yields compact, parameter-efficient models but also inherently mitigates the critical synchronization issues affecting dynamic inference. Finally, this work presents a unified framework where lightweight routers are pretrained to dynamically assign an optimal recursion depth for each token. This approach establishes a new Pareto frontier between efficiency and performance by effectively optimizing for both adaptive computation and parameter efficiency within a single model.

Method: SBP first learns a model of relations between documents from the pretraining dataset, then uses this model to synthesize a vast new corpus for joint training. The approach involves training 3B and 6B parameter models on up to 1T tokens from scratch in a compute-matched setup.

Result: SBP consistently outperforms a strong repetition baseline and achieves up to 60% of the performance improvement attainable by an oracle upper bound with access to 20x more unique data. Qualitative analysis shows synthesized documents abstract core concepts and craft new narratives rather than just paraphrasing.

Conclusion: SBP effectively captures inter-document correlations through synthetic data generation, offering strong empirical performance improvements. The method has a natural Bayesian interpretation where the synthesizer learns to abstract latent concepts shared between related documents.

Abstract: We introduce Synthetic Bootstrapped Pretraining (SBP), a language model (LM) pretraining procedure that first learns a model of relations between documents from the pretraining dataset and then leverages it to synthesize a vast new corpus for joint training. While the standard pretraining teaches LMs to learn causal correlations among tokens within a single document, it is not designed to efficiently model the rich, learnable inter-document correlations that can potentially lead to better performance. We validate SBP by designing a compute-matched pretraining setup and pretrain a 3B-parameter and a 6B-parameter model on up to 1T tokens from scratch. We find SBP consistently improves upon a strong repetition baseline and delivers up to 60% of performance improvement attainable by an oracle upper bound with access to 20x more unique data. Qualitative analysis reveals that the synthesized documents go beyond mere paraphrases – SBP first abstracts a core concept from the seed material and then crafts a new narration on top of it. Besides strong empirical performance, SBP admits a natural Bayesian interpretation: the synthesizer implicitly learns to abstract the latent concepts shared between related documents.

Method: Used NLTK Python library to parse corpora of different genres and count exact string matches to identify duplicate sentences.

Result: While completely unique sentences are often the majority in corpora, this is highly constrained by genre, and duplicate sentences constitute a significant portion of any individual corpus.

Conclusion: The claim that most sentences are unique needs qualification - genre matters significantly, and duplicate sentences are more common than traditionally assumed.

Abstract: A repeated claim in linguistics is that the majority of linguistic utterances are unique. For example, Pinker (1994: 10), summarizing an argument by Noam Chomsky, states that “virtually every sentence that a person utters or understands is a brand-new combination of words, appearing for the first time in the history of the universe.” With the increased availability of large corpora, this is a claim that can be empirically investigated. The current paper addresses the question by using the NLTK Python library to parse corpora of different genres, providing counts of exact string matches in each. Results show that while completely unique sentences are often the majority of corpora, this is highly constrained by genre, and that duplicate sentences are not an insignificant part of any individual corpus.

Method: Proposes a novel evaluation pipeline that: 1) merges multiple parent language models, 2) evaluates merged models compared to original ones using both behavioral metrics (downstream tasks like MMLU) and internal analysis of encoded linguistic competence, particularly in morphology and syntax. Demonstrated by merging instruction fine-tuned with math- and code-adapted Qwen2.5 models.

Result: Merging methods impact behavior and internals differently. While merged model performance typically falls between parent models on behavioral tasks, their encoded linguistic information (especially morphology and syntax) can surpass parent models. Weak ranking correlation exists between behavioral and internal evaluations.

Conclusion: Comprehensive evaluations connecting behavior and internal representations are essential for faithful understanding of model merging capabilities. Current behavioral-only assessments may miss important aspects of linguistic competence, emphasizing the need for more holistic evaluation approaches beyond superficial performance metrics.

Abstract: Merging methods combine the weights of multiple language models (LMs) to leverage their capacities, such as for domain adaptation. While existing studies investigate merged models from a solely behavioral perspective, we offer the first comprehensive view by assessing and connecting their behavior and internals. We present a novel evaluation pipeline that first merges multiple parent LMs, and then evaluates the merged models in comparison to the initial ones based on their behavior on downstream tasks, like MMLU, and the internal encoded linguistic competence. We showcase this pipeline by assessing the merging of instruction fine-tuned with math- and code-adapted LMs from the Qwen2.5 family. Our results show that merging methods impacts behavior and internals differently. While the performance of merged models is typically between that of the two parent models, their encoded information about linguistic phenomena, particularly in morphology and syntax, can surpass the parent models. Moreover, we find weak ranking correlation between this behavior and internal evaluation. With our pipeline and initial results, we emphasize the need for more comprehensive evaluations of model merging methods to gain a faithful understanding of their capabilities and reliability, beyond potential superficial behavioral advances.

Method: Pretrained on 206B-token corpus of scientific text, pure sequences, and sequence-text pairs. Aligned via supervised fine-tuning (SFT) on 40M instructions, annealed cold-start bootstrapping for long-form chain-of-thought reasoning, and reinforcement learning with task-specific reward shaping to instill deliberate scientific reasoning.

Result: The model supports 103 tasks across 5 capability families: faithful translation between text and scientific formats, text/knowledge extraction, property prediction, property classification, and unconditional/conditional sequence generation. It broadens instruction coverage, improves cross-domain generalization, and enhances fidelity compared to specialist systems.

Conclusion: The proposed scientific reasoning foundation model successfully integrates diverse scientific representations with natural language, demonstrating that cross-discipline learning strengthens transfer and downstream reliability. The model, datasets, and evaluation code are open-sourced for community use.

Abstract: We present a scientific reasoning foundation model that aligns natural language with heterogeneous scientific representations. The model is pretrained on a 206B-token corpus spanning scientific text, pure sequences, and sequence-text pairs, then aligned via SFT on 40M instructions, annealed cold-start bootstrapping to elicit long-form chain-of-thought, and reinforcement learning with task-specific reward shaping, which instills deliberate scientific reasoning. It supports four capability families, covering up to 103 tasks across workflows: (i) faithful translation between text and scientific formats, (ii) text/knowledge extraction, (iii) property prediction, (iv) property classification, (v) unconditional and conditional sequence generation and design. Compared with specialist systems, our approach broadens instruction coverage, improves cross-domain generalization, and enhances fidelity. We detail data curation and training and show that cross-discipline learning strengthens transfer and downstream reliability. The model, instruct tuning datasets and the evaluation code are open-sourced at https://huggingface.co/SciReason and https://github.com/open-sciencelab/SciReason.

Method: Used causal inference with one assumed and two guaranteed constraints to derive five potential causal structures. Tested across seven public and one developed reasoning tasks on GPT models. Compared coarse metrics with causal analysis results.

Result: Coarse metrics reported mixed effects (positive/negative/neutral), but causal inference showed no causal impact in 43/48 scenarios. In remaining 5, 3 involved multifaceted causal structures influenced by concrete instructions. OpenAI-o3 models were more resilient to output formats than GPT-4o and GPT-4.1.

Conclusion: Structured output has minimal causal impact on LLM generation quality; previous conflicting findings stem from coarse metrics. Reasoning models (OpenAI-o3) show greater resilience to output formats, revealing an “unaware advantage” of specialized reasoning architectures.

Abstract: Structured output from large language models (LLMs) has enhanced efficiency in processing generated information and is increasingly adopted in industrial applications. Prior studies have investigated the impact of structured output on LLMs’ generation quality, often presenting one-way findings. Some suggest that structured format enhances completeness and factual accuracy, while others argue that it restricts the reasoning capacity of LLMs and leads to reductions in standard evaluation metrics. Potential limitations of these assessments include restricted testing scenarios, weakly controlled comparative settings, and reliance on coarse metrics. In this work, we present a refined analysis using causal inference. Based on one assumed and two guaranteed constraints, we derive five potential causal structures characterizing the influence of structured output on LLMs’ generation: (1) collider without m-bias, (2) collider with m-bias, (3) single cause from instruction, (4) single cause from output format, and (5) independence. Across seven public and one developed reasoning tasks, we find that coarse metrics report positive, negative, or neutral effects of structured output on GPT-4o’s generation. However, causal inference reveals no causal impact in 43 out of 48 scenarios. In the remaining 5, 3 involve multifaceted causal structures influenced by concrete instructions. Further experiments show that OpenAI-o3 are more resilient to output formats than general-purpose GPT-4o and GPT-4.1, highlighting an unaware advantage of reasoning models.

Method: Three-stage framework: (1) Diagnose with Contextualized Role Adherence Score (CRAS) that measures role adherence across four dimensions; (2) Localize conflicts via attention drift analysis identifying middle-layer attention heads; (3) Align using Surgical Alignment of Instruction Layers (SAIL) with LoRA on focal layers and token-weighted DPO-style optimization.

Result: Improves instruction hierarchy compliance significantly (+5.60% with AutoGen on MedQA) without requiring full-model finetuning, demonstrating effectiveness across standard benchmarks and MAS frameworks.

Conclusion: The proposed surgical alignment approach effectively addresses hierarchical compliance failures in LLM-powered multi-agent systems by precisely targeting the problematic attention mechanisms, offering a practical solution for reliability-critical deployments.

Abstract: Large Language Model (LLM)-powered multi-agent systems (MAS) have rapidly advanced collaborative reasoning, tool use, and role-specialized coordination in complex tasks. However, reliability-critical deployment remains hindered by a systemic failure mode: hierarchical compliance under instruction conflicts (system-user, peer-peer), where agents misprioritize system-level rules in the presence of competing demands. Moreover, widely used macro-level metrics (e.g., pass@k) obscure these micro-level violations and offer little actionable guidance for remedy. In this work, we present a full-stack, three-stage framework: (1) Diagnose - Contextualized Role Adherence Score (CRAS), a query-wise, context-aware scoring metric that decomposes role adherence into four measurable dimensions; (2) Localize - attention drift analysis revealing that instruction conflicts are resolved by attention heads that are largely concentrated in middle layers; (3) Align - Surgical Alignment of Instruction Layers (SAIL), which installs LoRA only on the localized focal layers and optimizes a token-weighted DPO-style preference objective that credits tokens by their focal attentional contribution. Across standard benchmarks and MAS frameworks, our surgical approach improves instruction hierarchy compliance (e.g., +5.60% with AutoGen on MedQA) without full-model finetuning.

Method: TARE (Textual Sharpness-Aware Evolving) is a black-box framework that alternates between: 1) inner adversarial search that stresses prompts with hard paraphrases, and 2) outer robust selection preferring candidates with strong neighborhoods. ATARE adds anisotropic weights to shape semantic neighborhoods and adapts radius over time.

Result: The methods preserve accuracy under paraphrasing and outperform accuracy-only prompt search across diverse tasks while remaining computationally practical.

Conclusion: By formally addressing textual sharpness and designing for robustness over semantic neighborhoods, TARE provides a more stable and reliable approach to prompt optimization that mitigates the brittleness of current methods.

Abstract: The performance of Large Language Models (LLMs) hinges on carefully engineered prompts. However, prevailing prompt optimization methods, ranging from heuristic edits and reinforcement learning to evolutionary search, primarily target point-wise accuracy. They seldom enforce paraphrase invariance or searching stability, and therefore cannot remedy this brittleness in practice. Automated prompt search remains brittle: small, semantically preserving paraphrases often cause large performance swings. We identify this brittleness as the textual sharpness of the prompt landscape. In this work, we provide the first formal treatment of textual sharpness in the discrete, semantic space of prompts, together with an operational robustness criterion over a semantic neighborhood; the design is black-box or API-only, requiring no gradients to update the model’s parameters. Then we introduce TARE (Textual Sharpness-Aware Evolving), a derivative-free framework that alternates between an inner, sampling-based adversarial search that stresses a prompt with hard paraphrases and an outer, robust selection that prefers candidates whose neighborhoods remain strong. We further propose ATARE, which learns anisotropic weights to shape the semantic neighborhood and adapts its radius over time to balance exploration and fidelity. Diverse tasks evaluate our methods, whose design for minimizing textual sharpness gap leads to prompts that preserve accuracy under paraphrasing, outperforming accuracy-only prompt search while remaining computationally practical.

Method: Proposed MRAG-Suite platform integrates multiple multimodal benchmarks (WebQA, Chart-RAG, Visual-RAG, MRAG-Bench) with difficulty-based and ambiguity-aware filtering strategies, and introduces MM-RAGChecker for claim-level diagnostic analysis.

Result: Results show substantial accuracy reductions under difficult and ambiguous queries, highlighting prevalent hallucinations in Visual RAG systems, while MM-RAGChecker effectively diagnoses these issues.

Conclusion: The diagnostic framework successfully identifies weaknesses in Visual RAG systems under challenging conditions and provides guidance for future improvements in multimodal question answering systems.

Abstract: Multimodal Retrieval-Augmented Generation (Visual RAG) significantly advances question answering by integrating visual and textual evidence. Yet, current evaluations fail to systematically account for query difficulty and ambiguity. We propose MRAG-Suite, a diagnostic evaluation platform integrating diverse multimodal benchmarks (WebQA, Chart-RAG, Visual-RAG, MRAG-Bench). We introduce difficulty-based and ambiguity-aware filtering strategies, alongside MM-RAGChecker, a claim-level diagnostic tool. Our results demonstrate substantial accuracy reductions under difficult and ambiguous queries, highlighting prevalent hallucinations. MM-RAGChecker effectively diagnoses these issues, guiding future improvements in Visual RAG systems.

Method: Uses a permissive-first, risk-mitigated sourcing strategy combining public-domain and permissively licensed text with justified low-risk additions. Implements a three-tier risk categorization scheme with shard-level provenance metadata. Employs a single-stage pretraining recipe with high proportion of synthetic instruction and reasoning data.

Result: Models trained on MixtureVitae consistently outperform other permissive datasets across standard benchmarks, with particularly strong performance on MMLU, math, and code tasks. At 1.7B parameters/300B tokens, they approach DCLM performance and match/exceed instruction-tuned baselines on GSM8K, HumanEval, and MBPP despite using 36× fewer tokens.

Conclusion: Permissive-first data with high instruction and reasoning density, organized by licensing and provenance risk tiers, provides a practical and risk-mitigated foundation for training capable LLMs, reducing reliance on broad web scrapes without sacrificing competitiveness.

Abstract: We present MixtureVitae, an open-access pretraining corpus built to minimize legal risk while providing strong downstream performance. MixtureVitae follows a permissive-first, risk-mitigated sourcing strategy that combines public-domain and permissively licensed text (e.g., CC-BY/Apache) with carefully justified low-risk additions (e.g., government works and EU TDM-eligible sources). MixtureVitae adopts a simple, single-stage pretraining recipe that integrates a large proportion of permissive synthetic instruction and reasoning data-signals typically introduced during post-training and generally scarce in permissive web corpora. We categorize all sources into a three-tier scheme that reflects varying risk levels and provide shard-level provenance metadata to enable risk-aware usage. In controlled experiments using the open-sci-ref training protocol (fixed architectures and hyperparameters; 50B and 300B token budgets across 130M-1.7B parameters), models trained on MixtureVitae consistently outperform other permissive datasets across a suite of standard benchmarks, and at the 1.7B-parameters/300B-tokens setting, they surpass FineWeb-Edu and approach DCLM late in training. Performance is particularly strong on MMLU and on math and code benchmarks: a 1.7B model pretrained on 300B MixtureVitae tokens matches or exceeds a strong 1.7B instruction-tuned baseline on GSM8K, HumanEval, and MBPP, despite using over 36 times fewer tokens (300B vs. ~11T). Supported by a thorough decontamination analysis, these results show that permissive-first data with high instruction and reasoning density, tiered by licensing and provenance-related risk, can provide a practical and risk-mitigated foundation for training capable LLMs, reducing reliance on broad web scrapes without sacrificing competitiveness. Code: https://github.com/ontocord/mixturevitae

Method: Developed a RAG system with hybrid embedding mechanism for retrieving relevant text chunks from 10,195 segments of 300 guidelines, evaluated on 7,901 queries and manually curated 70 QA pairs, with clinical validation by 7 subject matter experts.

Result: High retrieval performance (MRR 0.814, 81% recall at first chunk, 99.1% within top 10), RAG-enhanced models showed 64.7 percentage point improvement in faithfulness to 99.5%, GPT-4.1 achieved 98.7% accuracy with 67% reduction in unsafe responses.

Conclusion: RAG is an effective, reliable, and scalable approach for applying generative AI in healthcare, enabling cost-effective access to medical guidelines while ensuring high accuracy and safety.

Abstract: This paper presents the development and evaluation of a Retrieval-Augmented Generation (RAG) system for querying the United Kingdom’s National Institute for Health and Care Excellence (NICE) clinical guidelines using Large Language Models (LLMs). The extensive length and volume of these guidelines can impede their utilisation within a time-constrained healthcare system, a challenge this project addresses through the creation of a system capable of providing users with precisely matched information in response to natural language queries. The system’s retrieval architecture, composed of a hybrid embedding mechanism, was evaluated against a corpus of 10,195 text chunks derived from three hundred guidelines. It demonstrates high performance, with a Mean Reciprocal Rank (MRR) of 0.814, a Recall of 81% at the first chunk and of 99.1% within the top ten retrieved chunks, when evaluated on 7901 queries. The most significant impact of the RAG system was observed during the generation phase. When evaluated on a manually curated dataset of seventy question-answer pairs, RAG-enhanced models showed substantial gains in performance. Faithfulness, the measure of whether an answer is supported by the source text, was increased by 64.7 percentage points to 99.5% for the RAG-enhanced O4-Mini model and significantly outperformed the medical-focused Meditron3-8B LLM, which scored 43%. Clinical evaluation by seven Subject Matter Experts (SMEs) further validated these findings, with GPT-4.1 achieving 98.7% accuracy while reducing unsafe responses by 67% compared to O4-Mini (from 3.0 to 1.0 per evaluator). This study thus establishes RAG as an effective, reliable, and scalable approach for applying generative AI in healthcare, enabling cost-effective access to medical guidelines.

Method: Combines theory-driven codebook (adapted from Interaction Analysis Model with four comment-level categories: Non-Knowledge Construction, Share, Explore, Integrate) with automated classification. Three annotators coded 20,000 comments from YouTube education channels. Compared bag-of-words baselines with transformer models using 10-fold cross-validation, with DeBERTa-v3-large achieving best performance.

Result: Codebook demonstrated strong reliability (Cohen’s kappa = 0.79 on main dataset, 0.85-0.93 across four domains). DeBERTa-v3-large achieved highest macro-averaged F1 score (0.841). External validation on four domains yielded macro-F1 above 0.705, with stronger transfer in medicine/programming (structured, task-focused discourse) and weaker in language/music (varied, context-dependent).

Conclusion: Theory-driven, semi-automated analysis of knowledge construction at scale is feasible, enabling integration of knowledge-construction indicators into learning analytics and design of online learning environments.

Abstract: The rapid expansion of online courses and social media has generated large volumes of unstructured learner-generated text. Understanding how learners construct knowledge in these spaces is crucial for analysing learning processes, informing content design, and providing feedback at scale. However, existing approaches typically rely on manual coding of well-structured discussion forums, which does not scale to the fragmented discourse found in online learning. This study proposes and validates a framework that combines a codebook inspired by the Interaction Analysis Model with an automated classifier to enable large-scale analysis of knowledge construction in unstructured online discourse. We adapt four comment-level categories of knowledge construction: Non-Knowledge Construction, Share, Explore, and Integrate. Three trained annotators coded a balanced sample of 20,000 comments from YouTube education channels. The codebook demonstrated strong reliability, with Cohen’s kappa = 0.79 on the main dataset and 0.85–0.93 across four additional educational domains. For automated classification, bag-of-words baselines were compared with transformer-based language models using 10-fold cross-validation. A DeBERTa-v3-large model achieved the highest macro-averaged F1 score (0.841), outperforming all baselines and other transformer models. External validation on four domains yielded macro-F1 above 0.705, with stronger transfer in medicine and programming, where discourse was more structured and task-focused, and weaker transfer in language and music, where comments were more varied and context-dependent. Overall, the study shows that theory-driven, semi-automated analysis of knowledge construction at scale is feasible, enabling the integration of knowledge-construction indicators into learning analytics and the design of online learning environments.

Method: Proposes a novel faithfulness annotation framework with an intermediate “Out-Dependent” category for cases requiring external knowledge verification. Uses this framework to construct VeriGray benchmark for unfaithfulness detection in summarization.

Result: Statistics show SOTA LLMs like GPT-5 exhibit ~6% hallucination in summarization tasks, and ~8% of generated sentences fall into the Out-Dependent category. The benchmark poses significant challenges to baseline methods, indicating room for improvement.

Conclusion: Annotation ambiguity is a critical issue in faithfulness evaluation, and the proposed framework with Out-Dependent category helps address this. The VeriGray benchmark reveals substantial unfaithfulness in LLM-generated summaries and highlights the need for better detection methods.

Abstract: Ensuring that Large Language Models (LLMs) generate summaries faithful to a given source document is essential for real-world applications. While prior research has explored LLM faithfulness, existing benchmarks suffer from annotation ambiguity, primarily due to the ill-defined boundary of permissible external knowledge in generated outputs. For instance, common sense is often incorporated into responses and labeled as “faithful”, yet the acceptable extent of such knowledge remains unspecified, leading to inconsistent annotations. To address this issue, we propose a novel faithfulness annotation framework, which introduces an intermediate category, Out-Dependent, to classify cases where external knowledge is required for verification. Using this framework, we construct VeriGray (Verification with the Gray Zone) – a new unfaithfulness detection benchmark in summarization. Statistics reveal that even SOTA LLMs, such as GPT-5, exhibit hallucinations ($\sim 6%$ of sentences) in summarization tasks. Moreover, a substantial proportion ($\sim 8%$ on average of models) of generated sentences fall into the Out-Dependent category, underscoring the importance of resolving annotation ambiguity in unfaithfulness detection benchmarks. Experiments demonstrate that our benchmark poses significant challenges to multiple baseline methods, indicating considerable room for future improvement.

Method: 1) Introduces Concept Gradient method for causal interpretation by measuring how concept changes affect model output; 2) Curates Targeted Lexicon Set of toxic words causing misclassifications; 3) Computes Word-Concept Alignment scores to quantify misclassification due to concept over-attribution; 4) Proposes lexicon-free augmentation strategy by generating toxic samples without predefined toxic lexicon sets.

Result: The approach provides more causal interpretability than traditional gradient methods, identifies specific toxic words contributing to misclassifications, quantifies misclassification causes through WCA scores, and enables analysis of whether over-attribution persists when explicit lexical overlap is removed.

Conclusion: The proposed framework offers improved interpretability for toxicity detection models by focusing on concept-based explanations, identifying misclassification sources, and analyzing model attribution patterns beyond simple lexical features, contributing to more robust and explainable toxicity detection systems.

Abstract: The rise of social networks has not only facilitated communication but also allowed the spread of harmful content. Although significant advances have been made in detecting toxic language in textual data, the exploration of concept-based explanations in toxicity detection remains limited. In this study, we leverage various subtype attributes present in toxicity detection datasets, such as obscene, threat, insult, identity attack, and sexual explicit as concepts that serve as strong indicators to identify whether language is toxic. However, disproportionate attribution of concepts towards the target class often results in classification errors. Our work introduces an interpretability technique based on the Concept Gradient (CG) method which provides a more causal interpretation by measuring how changes in concepts directly affect the output of the model. This is an extension of traditional gradient-based methods in machine learning, which often focus solely on input features. We propose the curation of Targeted Lexicon Set, which captures toxic words that contribute to misclassifications in text classification models. To assess the significance of these lexicon sets in misclassification, we compute Word-Concept Alignment (WCA) scores, which quantify the extent to which these words lead to errors due to over-attribution to toxic concepts. Finally, we introduce a lexicon-free augmentation strategy by generating toxic samples that exclude predefined toxic lexicon sets. This approach allows us to examine whether over-attribution persists when explicit lexical overlap is removed, providing insights into the model’s attribution on broader toxic language patterns.

Method: Used a human-in-the-loop framework with DeepSeek V3.1 guided by prompt-based cognitive scaffolding to identify and convey metaphor/metonymy in Huangdi Neijing passages. Evaluated translations using ChatGPT 5 Pro and Gemini 2.5 Pro simulating real-world readers across five cognitive dimensions, with 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/machine translation, effective metaphor/metonymy transfer strategies, 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 properly prompted LLMs can better convey underlying conceptual networks than traditional approaches.

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: Used HITL framework with prompt-based cognitive scaffolding on DeepSeek V3.1 to identify and translate metaphor/metonymy in Huangdi Neijing passages. Evaluated translations using ChatGPT 5 Pro and Gemini 2.5 Pro simulating three reader types, scoring across five cognitive dimensions, plus 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/machine translation, effective metaphor/metonymy transfer strategies, and readers’ cognitive preferences.

Conclusion: The study provides a cognitive, efficient, and replicable HITL methodological pathway for translating ancient, concept-dense texts like TCM, demonstrating that prompt-engineered LLMs can better convey conceptual networks than literal translations.

Abstract: Traditional Chinese Medicine (TCM) 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 (HITL) 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 (IPA). 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 the translation of ancient, concept-dense texts such as TCM.

Method: Created FINDER benchmark with 100 human-curated research tasks and 419 structured checklist items to standardize report structure, analytical depth, and factual grounding. Developed DEFT failure taxonomy through grounded theory with human-LLM co-annotation and inter-annotator reliability validation, analyzing ~1,000 reports from mainstream DRAs.

Result: DEFT taxonomy identifies 14 fine-grained failure modes across reasoning, retrieval, and generation. Experimental findings show current DRAs struggle not with task comprehension but with evidence integration, verification, and reasoning-resilient planning.

Conclusion: The FINDER benchmark and DEFT taxonomy provide standardized evaluation and failure analysis for Deep Research Agents, revealing critical gaps in evidence handling and verification that need to be addressed for practical report generation.

Abstract: Deep Research Agents (DRAs) aim to automatically produce analyst-level reports through iterative information retrieval and synthesis. However, most existing DRAs were validated on question-answering benchmarks, while research on generating comprehensive reports remains overlooked. Worse, current benchmarks for report synthesis suffer from task complexity and subjective metrics – this fails to reflect user demands and limits the practical utility of generated reports. To address these gaps, we present Fine-grained DEepResearch bench (FINDER), an enhanced benchmark consisting of 100 human-curated research tasks with 419 structured checklist items that standardize report structure, analytical depth, and factual grounding. Based on approximately 1,000 reports produced by mainstream DRAs, we further propose Deep rEsearch Failure Taxonomy (DEFT), the first failure taxonomy for deep research agents. DEFT contains 14 fine-grained failure modes across reasoning, retrieval, and generation, and is built upon grounded theory with human-LLM co-annotating and inter-annotator reliability validation. Our experimental findings reveal that current DRAs struggle not with task comprehension but with evidence integration, verification, and reasoning-resilient planning.

Method: CAPO uses an adaptive curriculum mechanism: 1) Bootstraps imitation learning with only positive advantage samples to establish robust foundations, 2) Gradually introduces negative signals to cultivate discriminative capabilities, 3) Compatible with various optimization methods (GRPO, PPO, RLOO, Reinforce++).

Result: Consistently achieves stable and significant improvements in mathematical reasoning tasks, and effectively generalizes to multimodal GUI reasoning scenarios, establishing itself as a versatile and robust optimization framework.

Conclusion: CAPO provides a more effective RL training approach for LLMs by separating positive and negative advantage signals through curriculum learning, leading to better generalization across diverse reasoning tasks.

Abstract: Reinforcement learning has emerged as a paradigm for post-training large language models, boosting their reasoning capabilities. Such approaches compute an advantage value for each sample, reflecting better or worse performance than expected, thereby yielding both positive and negative signals for training. However, the indiscriminate mixing of the two signals in existing methods, especially from the early stages, may lead to ambiguous guidance and limited gains. To address this issue, we propose CAPO (Curriculum Advantage Policy Optimization), an adaptive curriculum mechanism based on advantage signals. The proposed mechanism bootstraps imitation learning with positive-only advantage samples to establish robust foundations, and subsequently introduces negative signals to cultivate discriminative capabilities, thereby improving generalization across complex scenarios. Compatible with diverse optimization methods including GRPO, PPO, RLOO, and Reinforce++, our method consistently achieves stable and significant improvements in mathematical reasoning tasks, and further generalizes effectively to multimodal Graphical User Interface (GUI) reasoning scenarios, establishing itself as a versatile and robust optimization framework.

Method: Manual annotation of the complete inflectional paradigm of 21 personal nouns in a large corpus of contemporary German press texts, resulting in 6,195 annotated tokens to analyze lexeme-specific differences.

Result: Significant differences between lexical items (especially passive role vs. prestige-related nouns), GM occurs mainly in plural and indefinite noun phrases, and is not primarily used to denote entire classes of people as previously thought.

Conclusion: Empirical analysis provides nuanced understanding of GM usage, offering better alignment between psycholinguistic stimuli and authentic language patterns in German press texts.

Abstract: This study examines the distribution and linguistic characteristics of generic masculines (GM) in contemporary German press texts. The use of masculine personal nouns to refer to mixed-gender groups or unspecified individuals has been widely debated in academia and the public, with con-flicting perspectives on its gender-neutrality. While psycholinguistic studies suggest that GM is more readily associated with male referents, corpus-based analyses of its actual use remain scarce. We investigate GM in a large corpus of press texts, focusing on lexeme-specific differences across dif-ferent types of personal nouns. We conducted manual annotations of the whole inflectional para-digm of 21 personal nouns, resulting in 6,195 annotated tokens. Our findings reveal considerable differences between lexical items, especially between passive role nouns and prestige-related per-sonal nouns. On a grammatical level, we find that GM occurs predominantly in the plural and in indefinite noun phrases. Furthermore, our data shows that GM is not primarily used to denote entire classes of people, as has been previously claimed. By providing an empirical insight into the use of GM in authentic written language, we contribute to a more nuanced understanding of its forms and manifestations. These findings provide a solid basis for aligning linguistic stimuli in psy-cholinguistic studies more closely with real-world language use.

Method: CoopRAG features four key steps: (1) unrolling questions into sub-questions with masked uncertain positions in reasoning chains, (2) retrieving documents using augmented queries with sub-questions and reasoning chains, (3) reranking documents by contrasting retriever layers, and (4) reconstructing reasoning chains by filling masked positions via LLM.

Result: CoopRAG consistently outperforms state-of-the-art QA methods on three multi-hop QA datasets and one simple QA dataset, showing improvements in both retrieval and QA performance metrics.

Conclusion: The cooperative framework between retriever and LLM, and between different retriever layers, effectively addresses limitations of existing RAG methods, leading to more accurate question answering through better document retrieval and reasoning chain reconstruction.

Abstract: Since large language models (LLMs) have a tendency to generate factually inaccurate output, retrieval-augmented generation (RAG) has gained significant attention as a key means to mitigate this downside of harnessing only LLMs. However, existing RAG methods for simple and multi-hop question answering (QA) are still prone to incorrect retrievals and hallucinations. To address these limitations, we propose CoopRAG, a novel RAG framework for the question answering task in which a retriever and an LLM work cooperatively with each other by exchanging informative knowledge, and the earlier and later layers of the retriever model work cooperatively with each other to accurately rank the retrieved documents relevant to a given query. In this framework, we (i) unroll a question into sub-questions and a reasoning chain in which uncertain positions are masked, (ii) retrieve the documents relevant to the question augmented with the sub-questions and the reasoning chain, (iii) rerank the documents by contrasting layers of the retriever, and (iv) reconstruct the reasoning chain by filling the masked positions via the LLM. Our experiments demonstrate that CoopRAG consistently outperforms state-of-the-art QA methods on three multi-hop QA datasets as well as a simple QA dataset in terms of both the retrieval and QA performances. Our code is available.

Method: Created benchmark with 100 complex samples requiring long-form structured outputs, designed by legal professionals supervised by an attorney. Evaluated using both human assessment and automated evaluation with leading LLMs (GPT-5, Gemini 2.5 Pro, Claude Opus 4.1) across multiple practical criteria.

Result: Human evaluation revealed that abstract instructions prompted unnecessary modifications, exposing model weaknesses in document-level editing missed by conventional tasks. Automated evaluation aligned well with human judgment for criteria with clear linguistic grounding, but assessing structural consistency remained challenging. Automated evaluation proved useful as a screening tool when expert availability is limited.

Conclusion: The benchmark demonstrates the utility of automated evaluation for legal AI tasks and proposes a dataset evaluation framework to promote more practice-oriented research in the legal domain, highlighting the need for better structural consistency assessment methods.

Abstract: This paper introduces LegalRikai: Open Benchmark, a new benchmark comprising four complex tasks that emulate Japanese corporate legal practices. The benchmark was created by legal professionals under the supervision of an attorney. This benchmark has 100 samples that require long-form, structured outputs, and we evaluated them against multiple practical criteria. We conducted both human and automated evaluations using leading LLMs, including GPT-5, Gemini 2.5 Pro, and Claude Opus 4.1. Our human evaluation revealed that abstract instructions prompted unnecessary modifications, highlighting model weaknesses in document-level editing that were missed by conventional short-text tasks. Furthermore, our analysis reveals that automated evaluation aligns well with human judgment on criteria with clear linguistic grounding, and assessing structural consistency remains a challenge. The result demonstrates the utility of automated evaluation as a screening tool when expert availability is limited. We propose a dataset evaluation framework to promote more practice-oriented research in the legal domain.

Method: Uses deep-adapter-based framework with a single precision knob controlling input fidelity. Key innovation is a temporally coherent degradation module that applies realistic structural failure transformations during training.

Result: Achieves SOTA on severe artifact videos and competitive performance on standard benchmarks. Runs at ~13 FPS at 720p on single A100. Introduces AIGC54 benchmark with FIQA, semantic, and perceptual metrics for evaluation.

Conclusion: CreativeVR effectively addresses structural artifacts in AI-generated content while maintaining practical throughput, offering controlled trade-off between restoration precision and corrective behavior.

Abstract: Modern text-to-video (T2V) diffusion models can synthesize visually compelling clips, yet they remain brittle at fine-scale structure: even state-of-the-art generators often produce distorted faces and hands, warped backgrounds, and temporally inconsistent motion. Such severe structural artifacts also appear in very low-quality real-world videos. Classical video restoration and super-resolution (VR/VSR) methods, in contrast, are tuned for synthetic degradations such as blur and downsampling and tend to stabilize these artifacts rather than repair them, while diffusion-prior restorers are usually trained on photometric noise and offer little control over the trade-off between perceptual quality and fidelity. We introduce CreativeVR, a diffusion-prior-guided video restoration framework for AI-generated (AIGC) and real videos with severe structural and temporal artifacts. Our deep-adapter-based method exposes a single precision knob that controls how strongly the model follows the input, smoothly trading off between precise restoration on standard degradations and stronger structure- and motion-corrective behavior on challenging content. Our key novelty is a temporally coherent degradation module used during training, which applies carefully designed transformations that produce realistic structural failures. To evaluate AIGC-artifact restoration, we propose the AIGC54 benchmark with FIQA, semantic and perceptual metrics, and multi-aspect scoring. CreativeVR achieves state-of-the-art results on videos with severe artifacts and performs competitively on standard video restoration benchmarks, while running at practical throughput (about 13 FPS at 720p on a single 80-GB A100). Project page: https://daveishan.github.io/creativevr-webpage/.

Method: Proposes an explainable adversarial-robust Vision-Language-Action model based on OpenVLA-OFT framework, integrating an Evidence-3 module that detects photometric perturbations and generates natural language explanations of their causes and effects.

Result: The model reduces Current Action L1 loss by 21.7% and Next Actions L1 loss by 18.4% compared to baseline, demonstrating improved action prediction accuracy and explainability under adversarial conditions.

Conclusion: The proposed model enhances robustness and explainability of smart farming systems against photometric perturbations, making them more reliable under adversarial attacks while providing understandable explanations of system behavior.

Abstract: Smart farming has emerged as a key technology for advancing modern agriculture through automation and intelligent control. However, systems relying on RGB cameras for perception and robotic manipulators for control, common in smart farming, are vulnerable to photometric perturbations such as hue, illumination, and noise changes, which can cause malfunction under adversarial attacks. To address this issue, we propose an explainable adversarial-robust Vision-Language-Action model based on the OpenVLA-OFT framework. The model integrates an Evidence-3 module that detects photometric perturbations and generates natural language explanations of their causes and effects. Experiments show that the proposed model reduces Current Action L1 loss by 21.7% and Next Actions L1 loss by 18.4% compared to the baseline, demonstrating improved action prediction accuracy and explainability under adversarial conditions.

Method: Three key contributions: (1) Multi-task loss improvements including Balanced L1 regression, Chamfer point-set distance, and uncertainty-based loss weighting with focal and Dice components; (2) Lightweight Temporal LSTM Fusion module for per-anchor feature aggregation across frames; (3) ESCOP-style training refinements coupling curve-level supervision with temporal consistency.

Result: On OpenLane dataset, Temporal-Anchor3DLane improves F1 score by +6.2 and produces smoother temporal trajectories, demonstrating significant enhancement in 3D lane robustness without requiring extra sensors or scaling.

Conclusion: Small architectural and loss refinements can significantly enhance 3D lane detection robustness, showing that careful improvements to loss functions and temporal modeling can substantially boost performance without additional hardware or computational scaling.

Abstract: Monocular 3D lane detection remains challenging due to depth ambiguity, occlusion, and temporal instability across frames. Anchor-based approaches such as Anchor3DLane have demonstrated strong performance by regressing continuous 3D lane curves from multi-camera surround views. However, the baseline model still exhibits (i) sensitivity to regression outliers, (ii) weak supervision of global curve geometry, (iii) difficulty in balancing multiple loss terms, and (iv) limited exploitation of temporal continuity. We propose Temporal-Anchor3DLane, an enhanced 3D lane detection framework that extends Anchor3DLane with three key contributions: (1) a set of multi-task loss improvements, including Balanced L1 regression, Chamfer point-set distance, and uncertainty-based loss weighting, together with focal and Dice components for classification and visibility; (2) a lightweight Temporal LSTM Fusion module that aggregates per-anchor features across frames, replacing a heavier Transformer-style temporal fusion; and (3) ESCOP-style training refinements that couple curve-level supervision with temporal consistency. On OpenLane, Temporal-Anchor3DLane improves F1 by +6.2 and yields smoother temporal trajectories, showing that small architectural and loss refinements significantly enhance 3D lane robustness without extra sensors or scaling.

Method: Created a new cactus-fig dataset with 3,587 field images across three symptom classes. Benchmarked three mobile-efficient architectures: custom lightweight CNN, EfficientNet-Lite1, and MobileViT-XS. Focused on cactus-fig to evaluate attention sensitivity and inductive bias transfer on indigenous morphology.

Result: EfficientNet-Lite1 achieved 90.7% test accuracy, lightweight CNN reached 89.5% with best deployment profile (42 ms latency, 4.8 MB size), and MobileViT-XS delivered 97.3% mean cross-validation accuracy, showing MHSA-based global reasoning outperforms local texture CNN kernels.

Conclusion: The ARM-compatible models were deployed in a Tigrigna and Amharic localized Flutter app supporting fully offline inference on Cortex-A53 devices, strengthening inclusivity for food security diagnostics in post-conflict edge environments.

Abstract: Agriculture supports over 80% of the population in the Tigray region of Ethiopia, where infrastructural disruptions limit access to expert crop disease diagnosis. We present an offline-first detection system centered on a newly curated indigenous cactus-fig (Opuntia ficus-indica) dataset consisting of 3,587 field images across three core symptom classes. Given deployment constraints in post-conflict edge environments, we benchmark three mobile-efficient architectures: a custom lightweight CNN, EfficientNet-Lite1, and the CNN-Transformer hybrid MobileViT-XS. While the broader system contains independent modules for potato, apple, and corn, this study isolates cactus-fig model performance to evaluate attention sensitivity and inductive bias transfer on indigenous morphology alone. Results establish a clear Pareto trade-off: EfficientNet-Lite1 achieves 90.7% test accuracy, the lightweight CNN reaches 89.5% with the most favorable deployment profile (42 ms inference latency, 4.8 MB model size), and MobileViT-XS delivers 97.3% mean cross-validation accuracy, demonstrating that MHSA-based global reasoning disambiguates pest clusters from two dimensional fungal lesions more reliably than local texture CNN kernels. The ARM compatible models are deployed in a Tigrigna and Amharic localized Flutter application supporting fully offline inference on Cortex-A53 class devices, strengthening inclusivity for food security critical diagnostics.

Method: Introduces SAVEBench dataset with text/mask conditions for object-grounded learning, and trains SAVE model using Schrodinger Bridge for direct transport from source to target audiovisual mixtures.

Result: SAVE successfully removes target objects in both audio and visual content while preserving remaining content, with stronger temporal synchronization and audiovisual semantic correspondence than separate editors.

Conclusion: The proposed SAVE model effectively addresses joint audio-visual editing challenges through paired dataset creation and end-to-end flow-matching with Schrodinger Bridge architecture.

Abstract: Joint editing of audio and visual content is crucial for precise and controllable content creation. This new task poses challenges due to the limitations of paired audio-visual data before and after targeted edits, and the heterogeneity across modalities. To address the data and modeling challenges in joint audio-visual editing, we introduce SAVEBench, a paired audiovisual dataset with text and mask conditions to enable object-grounded source-to-target learning. With SAVEBench, we train the Schrodinger Audio-Visual Editor (SAVE), an end-to-end flow-matching model that edits audio and video in parallel while keeping them aligned throughout processing. SAVE incorporates a Schrodinger Bridge that learns a direct transport from source to target audiovisual mixtures. Our evaluation demonstrates that the proposed SAVE model is able to remove the target objects in audio and visual content while preserving the remaining content, with stronger temporal synchronization and audiovisual semantic correspondence compared with pairwise combinations of an audio editor and a video editor.

Method: Uses star-convex regions parameterized by truncated Fourier series, optimized under extremal preserve/delete objectives using classifier gradients, guaranteeing single connected masks with far fewer parameters.

Result: Matches extremal fidelity of dense masks while producing compact, interpretable regions with improved run-to-run consistency, achieving higher relevance mass and lower complexity than baselines.

Conclusion: The contour-based approach provides faithful yet compact explanations with explicit area control, enabling transparent fidelity-area profiles and extending to multi-object localization.

Abstract: Faithful yet compact explanations for vision models remain a challenge, as commonly used dense perturbation masks are often fragmented and overfitted, needing careful post-processing. Here, we present a training-free explanation method that replaces dense masks with smooth tunable contours. A star-convex region is parameterized by a truncated Fourier series and optimized under an extremal preserve/delete objective using the classifier gradients. The approach guarantees a single, simply connected mask, cuts the number of free parameters by orders of magnitude, and yields stable boundary updates without cleanup. Restricting solutions to low-dimensional, smooth contours makes the method robust to adversarial masking artifacts. On ImageNet classifiers, it matches the extremal fidelity of dense masks while producing compact, interpretable regions with improved run-to-run consistency. Explicit area control also enables importance contour maps, yielding a transparent fidelity-area profiles. Finally, we extend the approach to multi-contour and show how it can localize multiple objects within the same framework. Across benchmarks, the method achieves higher relevance mass and lower complexity than gradient and perturbation based baselines, with especially strong gains on self-supervised DINO models where it improves relevance mass by over 15% and maintains positive faithfulness correlations.

Method: Systematic self-training framework with two-stage hybrid training strategy, extensive data augmentation, SegFormer with MiT-B4 backbone, and iterative teacher-student loop for progressive refinement

Result: Achieved competitive performance on both Development and Testing Phase datasets of the competition

Conclusion: The proposed self-training framework with SegFormer architecture effectively addresses wheat semantic segmentation challenges and demonstrates strong performance in agricultural computer vision tasks

Abstract: This extended abstract details our solution for the Global Wheat Full Semantic Segmentation Competition. We developed a systematic self-training framework. This framework combines a two-stage hybrid training strategy with extensive data augmentation. Our core model is SegFormer with a Mix Transformer (MiT-B4) backbone. We employ an iterative teacher-student loop. This loop progressively refines model accuracy. It also maximizes data utilization. Our method achieved competitive performance. This was evident on both the Development and Testing Phase datasets.

Method: Evaluated SAM3 in zero-shot mode against three fine-tuned YOLO11 variants (nano, medium, large) on the MinneApple dataset (670 orchard images with 28,179 annotated apple instances). Used rigorous validation with different IoU thresholds to analyze performance gaps and boundary stability.

Result: At IoU=0.15, YOLO models achieved 68.9%, 72.2%, and 71.9% F1 scores while SAM3 reached 59.8%. However, YOLO showed steep degradation (48-50 points) across IoU ranges while SAM3 dropped only 4 points, revealing SAM3 has 12 times superior boundary stability.

Conclusion: SAM3 excels in mask precision and boundary stability while YOLO11 specializes in detection completeness. The choice between specialized fine-tuned models and generalist foundation models depends on task requirements - SAM3 is preferable for precise boundary segmentation, YOLO for detection completeness in dense scenarios.

Abstract: Deep learning has advanced two fundamentally different paradigms for instance segmentation: specialized models optimized through task-specific fine-tuning and generalist foundation models capable of zero-shot segmentation. This work presents a comprehensive comparison between SAM3 (Segment Anything Model, also called SAMv3) operating in zero-shot mode and three variants of Ultralytics YOLO11 (nano, medium, and large) fine-tuned for instance segmentation. The evaluation is conducted on the MinneApple dataset, a dense benchmark comprising 670 orchard images with 28,179 annotated apple instances, enabling rigorous validation of model behavior under high object density and occlusion. Our analysis shows IoU choices can inflate performance gaps by up to 30%. At the appropriate IoU = 0.15 threshold, YOLO models achieve 68.9%, 72.2%, and 71.9% F1, while SAM3 reaches 59.8% in pure zero-shot mode. However, YOLO exhibits steep degradation 48-50 points across IoU ranges whereas SAM3 drops only 4 points, revealing 12 times superior boundary stability of SAM3. This highlights the strength of SAMv3 in mask precision versus specialization in detection completeness of YOLO11. We provide open-source code, evaluation pipelines, and methodological recommendations, contributing to a deeper understanding of when specialized fine-tuned models or generalist foundation models are preferable for dense instance segmentation tasks. This project repository is available on GitHub as https://github.com/Applied-AI-Research-Lab/Segment-Anything-Model-SAM3-Zero-Shot-Segmentation-Against-Fine-Tuned-YOLO-Detectors

Method: Uses Implicit Neural Representations (INRs) to model mmWave signals as continuous functions, achieving up to 49-fold compression. Incorporates hypernetworks that generate INR parameters based on environmental context (from RGB-D images) and human motion features (from MotionGPT text-to-pose generation).

Result: Achieves complex SSIM of 0.88 and PSNR of 35 dB, outperforming existing methods in signal realism. Improves activity recognition accuracy by up to 7% and reduces human pose estimation error by up to 15%, while operating 6-35 times faster than simulation-based approaches.

Conclusion: mmWeaver enables efficient, realistic mmWave signal synthesis by leveraging semantic and geometric priors through INRs and hypernetworks, providing a practical solution for dataset augmentation in radar applications.

Abstract: Realistic signal generation and dataset augmentation are essential for advancing mmWave radar applications such as activity recognition and pose estimation, which rely heavily on diverse, and environment-specific signal datasets. However, mmWave signals are inherently complex, sparse, and high-dimensional, making physical simulation computationally expensive. This paper presents mmWeaver, a novel framework that synthesizes realistic, environment-specific complex mmWave signals by modeling them as continuous functions using Implicit Neural Representations (INRs), achieving up to 49-fold compression. mmWeaver incorporates hypernetworks that dynamically generate INR parameters based on environmental context (extracted from RGB-D images) and human motion features (derived from text-to-pose generation via MotionGPT), enabling efficient and adaptive signal synthesis. By conditioning on these semantic and geometric priors, mmWeaver generates diverse I/Q signals at multiple resolutions, preserving phase information critical for downstream tasks such as point cloud estimation and activity classification. Extensive experiments show that mmWeaver achieves a complex SSIM of 0.88 and a PSNR of 35 dB, outperforming existing methods in signal realism while improving activity recognition accuracy by up to 7% and reducing human pose estimation error by up to 15%, all while operating 6-35 times faster than simulation-based approaches.

Method: 1) Collect RapVerse dataset with synchronous rapping vocals, lyrics, and 3D holistic body meshes; 2) Use VQ-VAE to encode motion into discrete tokens; 3) Use vocal-to-unit model for quantized audio tokens; 4) Joint transformer modeling across language, audio, and motion modalities.

Result: The framework produces coherent and realistic singing vocals alongside human motions directly from text, rivaling specialized single-modality systems and establishing new benchmarks for joint vocal-motion generation.

Conclusion: Scaling autoregressive multimodal transformers across language, audio, and motion enables seamless joint generation of vocals and body motions, advancing beyond isolated modality approaches.

Abstract: In this work, we introduce a challenging task for simultaneously generating 3D holistic body motions and singing vocals directly from textual lyrics inputs, advancing beyond existing works that typically address these two modalities in isolation. To facilitate this, we first collect the RapVerse dataset, a large dataset containing synchronous rapping vocals, lyrics, and high-quality 3D holistic body meshes. With the RapVerse dataset, we investigate the extent to which scaling autoregressive multimodal transformers across language, audio, and motion can enhance the coherent and realistic generation of vocals and whole-body human motions. For modality unification, a vector-quantized variational autoencoder is employed to encode whole-body motion sequences into discrete motion tokens, while a vocal-to-unit model is leveraged to obtain quantized audio tokens preserving content, prosodic information and singer identity. By jointly performing transformer modeling on these three modalities in a unified way, our framework ensures a seamless and realistic blend of vocals and human motions. Extensive experiments demonstrate that our unified generation framework not only produces coherent and realistic singing vocals alongside human motions directly from textual inputs, but also rivals the performance of specialized single-modality generation systems, establishing new benchmarks for joint vocal-motion generation.

Method: Proposed a graph representation with discrete dynamic graph neural network (DDGNN) using star graphs to capture spatial-temporal relationships between a static center point and dynamic radar points across frames.

Result: Achieved 94.27% classification accuracy on real-world HAR datasets, near the 97.25% accuracy of vision-based skeleton data, and demonstrated effectiveness on Raspberry Pi 4.

Conclusion: The DDGNN approach effectively handles sparse, variable-size mmWave radar point clouds for HAR without requiring resampling or frame aggregators, outperforming baseline and recent radar-specific methods.

Abstract: Human activity recognition (HAR) requires extracting accurate spatial-temporal features with human movements. A mmWave radar point cloud-based HAR system suffers from sparsity and variable-size problems due to the physical features of the mmWave signal. Existing works usually borrow the preprocessing algorithms for the vision-based systems with dense point clouds, which may not be optimal for mmWave radar systems. In this work, we proposed a graph representation with a discrete dynamic graph neural network (DDGNN) to explore the spatial-temporal representation of human movement-related features. Specifically, we designed a star graph to describe the high-dimensional relative relationship between a manually added static center point and the dynamic mmWave radar points in the same and consecutive frames. We then adopted DDGNN to learn the features residing in the star graph with variable sizes. Experimental results demonstrated that our approach outperformed other baseline methods using real-world HAR datasets. Our system achieved an overall classification accuracy of 94.27%, which gets the near-optimal performance with a vision-based skeleton data accuracy of 97.25%. We also conducted an inference test on Raspberry Pi~4 to demonstrate its effectiveness on resource-constraint platforms. \sh{ We provided a comprehensive ablation study for variable DDGNN structures to validate our model design. Our system also outperformed three recent radar-specific methods without requiring resampling or frame aggregators.

Method: Combines Google Street View imagery, semantic image segmentation, and remote sensing. Trains two XGBoost models to predict land surface temperature using GSV training data from selected administrative wards, then deploys models patchwork-style across all OSMnx-derived pedestrian network nodes.

Result: Creates a spatial data science pipeline that enables heat-aware routing by estimating pedestrian heat exposure across the city’s pedestrian network infrastructure.

Conclusion: The deployed model provides a foundation for pinpointing where and understanding why certain city corridors experience disproportionately higher temperatures at an infrastructural scale, addressing urban heat island effects in tropical cities.

Abstract: Pedestrian heat exposure is a critical health risk in dense tropical cities, yet standard routing algorithms often ignore micro-scale thermal variation. Hot Hém is a GeoAI workflow that estimates and operationalizes pedestrian heat exposure in Hô Chí Minh City (HCMC), Vi\d{e}t Nam, colloquially known as Sài Gòn. This spatial data science pipeline combines Google Street View (GSV) imagery, semantic image segmentation, and remote sensing. Two XGBoost models are trained to predict land surface temperature (LST) using a GSV training dataset in selected administrative wards, known as phŏng, and are deployed in a patchwork manner across all OSMnx-derived pedestrian network nodes to enable heat-aware routing. This is a model that, when deployed, can provide a foundation for pinpointing where and further understanding why certain city corridors may experience disproportionately higher temperatures at an infrastructural scale.

Method: Used unmanned aerial vehicles (UAVs) with top-down perspective to collect traffic data, processed through the Data from Sky (DFS) platform. Data collected at six mid-block locations in India’s national capital region with 30 fps resolution, including vehicle positions, speeds, accelerations, and classifications.

Result: Created validated datasets showing key behavioral patterns like lane-keeping preferences, speed distributions, and lateral maneuvers in heterogeneous traffic. Data validated against manual counts, space mean speeds, and probe trajectories.

Conclusion: These open-access datasets provide valuable resources for global research community to support simulation modeling, safety assessment, and behavioral studies in complex urban traffic environments, enabling more accurate model development and validation.

Abstract: This paper offers openly available microscopic vehicle trajectory (MVT) datasets collected using unmanned aerial vehicles (UAVs) in heterogeneous, area-based urban traffic conditions. Traditional roadside video collection often fails in dense mixed traffic due to occlusion, limited viewing angles, and irregular vehicle movements. UAV-based recording provides a top-down perspective that reduces these issues and captures rich spatial and temporal dynamics. The datasets described here were extracted using the Data from Sky (DFS) platform and validated against manual counts, space mean speeds, and probe trajectories in earlier work. Each dataset contains time-stamped vehicle positions, speeds, longitudinal and lateral accelerations, and vehicle classifications at a resolution of 30 frames per second. Data were collected at six mid-block locations in the national capital region of India, covering diverse traffic compositions and density levels. Exploratory analyses highlight key behavioural patterns, including lane-keeping preferences, speed distributions, and lateral manoeuvres typical of heterogeneous and area-based traffic settings. These datasets are intended as a resource for the global research community to support simulation modelling, safety assessment, and behavioural studies under area-based traffic conditions. By making these empirical datasets openly available, this work offers researchers a unique opportunity to develop, test, and validate models that more accurately represent complex urban traffic environments.

Method: Proposes GMFVAD which generates fine-grained multi-modal features: summarizes main video content and uses text captions to enhance visual features of highlighted portions, reducing redundancy.

Result: Achieves state-of-the-art performance on four major datasets; ablation studies confirm improvements come from reduced redundant information.

Conclusion: Leveraging diversity among multi-modal information to refine features and reduce redundancy significantly improves video anomaly detection performance.

Abstract: Video anomaly detection (VAD) is a challenging task that detects anomalous frames in continuous surveillance videos. Most previous work utilizes the spatio-temporal correlation of visual features to distinguish whether there are abnormalities in video snippets. Recently, some works attempt to introduce multi-modal information, like text feature, to enhance the results of video anomaly detection. However, these works merely incorporate text features into video snippets in a coarse manner, overlooking the significant amount of redundant information that may exist within the video snippets. Therefore, we propose to leverage the diversity among multi-modal information to further refine the extracted features, reducing the redundancy in visual features, and we propose Grained Multi-modal Feature for Video Anomaly Detection (GMFVAD). Specifically, we generate more grained multi-modal feature based on the video snippet, which summarizes the main content, and text features based on the captions of original video will be introduced to further enhance the visual features of highlighted portions. Experiments show that the proposed GMFVAD achieves state-of-the-art performance on four mainly datasets. Ablation experiments also validate that the improvement of GMFVAD is due to the reduction of redundant information.

Method: Uses five industrial CMOS cameras in a modular framework with HSV and RGB color domain comparison for color sequence classification. System is trained with at least five reference samples per harness batch, with trained files stored for reuse.

Result: Deployed at GPV Lanka Pvt. Ltd., achieving 100% detection accuracy and reducing inspection time by 44% compared to manual methods. Includes user management, adjustable lighting, session data storage, and secure login features.

Conclusion: The semiautomated machine vision system provides reliable and efficient wire harness inspection capabilities for real-world industrial applications.

Abstract: Wire harness inspection process remains a labor-intensive process prone to errors in the modern Electronics Manufacturing Services (EMS) industry. This paper introduces a semiautomated machine vision system capable of verifying correct wire positioning, correctness of the connector polarity and correctness of color sequences for both linear and circular wire harness configurations. Five industrial standard CMOS cameras are integrated into a modularized mechanical framework in the physical structure of the solution and a HSV and RGB color domain value comparison based color sequence classifier is used in the operation. For each harness batch, a user can train the system using at least five reference samples; the trained file is stored and reused for similar harness types. The Solution is deployed at GPV Lanka Pvt. Ltd. (Fig. 2) and the system achieved 100% detection accuracy and reduced inspection time by 44% compared to manual methods. Additional features include user management, adjustable lighting, session data storage, and secure login. Results of this product usage in the real world situation demonstrate that this approach delivers reliable and efficient inspection capabilities.

Method: The authors introduce the Read-or-Ignore VQA (RIO-VQA) task and create RIO-Bench, a standardized dataset with real images and same-scene counterfactuals (read/ignore scenarios) by varying only textual content and question type.

Result: Evaluation using RIO-Bench shows that strong LVLMs and existing defenses fail to balance typographic robustness and text-reading capability. The benchmark enables a novel data-driven defense that learns adaptive selective text use.

Conclusion: This work reveals a fundamental misalignment between current evaluation scope and real-world requirements, providing a principled path toward more reliable LVLMs through selective text use rather than blanket text ignoring.

Abstract: Large vision-language models (LVLMs) are vulnerable to typographic attacks, where misleading text within an image overrides visual understanding. Existing evaluation protocols and defenses, largely focused on object recognition, implicitly encourage ignoring text to achieve robustness; however, real-world scenarios often require joint reasoning over both objects and text (e.g., recognizing pedestrians while reading traffic signs). To address this, we introduce a novel task, Read-or-Ignore VQA (RIO-VQA), which formalizes selective text use in visual question answering (VQA): models must decide, from context, when to read text and when to ignore it. For evaluation, we present the Read-or-Ignore Benchmark (RIO-Bench), a standardized dataset and protocol that, for each real image, provides same-scene counterfactuals (read / ignore) by varying only the textual content and question type. Using RIO-Bench, we show that strong LVLMs and existing defenses fail to balance typographic robustness and text-reading capability, highlighting the need for improved approaches. Finally, RIO-Bench enables a novel data-driven defense that learns adaptive selective text use, moving beyond prior non-adaptive, text-ignoring defenses. Overall, this work reveals a fundamental misalignment between the existing evaluation scope and real-world requirements, providing a principled path toward reliable LVLMs. Our Project Page is at https://turingmotors.github.io/rio-vqa/.

Method: HUD leverages modality information density disparity through three components: Holistic Pronoun Disambiguation for cross-modal interaction, Atomistic Uncertainty Modeling for fine-grained semantics, and Holistic-to-Atomistic Alignment for precise composed feature learning.

Result: HUD achieves state-of-the-art performance across three benchmark datasets for both Composed Video Retrieval and Composed Image Retrieval tasks.

Conclusion: The proposed HUD framework effectively addresses modality disparity issues in multi-modal queries, enabling accurate composed feature learning through hierarchical uncertainty-aware disambiguation, with demonstrated effectiveness in both CVR and CIR tasks.

Abstract: Composed Video Retrieval (CVR) is a challenging video retrieval task that utilizes multi-modal queries, consisting of a reference video and modification text, to retrieve the desired target video. The core of this task lies in understanding the multi-modal composed query and achieving accurate composed feature learning. Within multi-modal queries, the video modality typically carries richer semantic content compared to the textual modality. However, previous works have largely overlooked the disparity in information density between these two modalities. This limitation can lead to two critical issues: 1) modification subject referring ambiguity and 2) limited detailed semantic focus, both of which degrade the performance of CVR models. To address the aforementioned issues, we propose a novel CVR framework, namely the Hierarchical Uncertainty-aware Disambiguation network (HUD). HUD is the first framework that leverages the disparity in information density between video and text to enhance multi-modal query understanding. It comprises three key components: (a) Holistic Pronoun Disambiguation, (b) Atomistic Uncertainty Modeling, and (c) Holistic-to-Atomistic Alignment. By exploiting overlapping semantics through holistic cross-modal interaction and fine-grained semantic alignment via atomistic-level cross-modal interaction, HUD enables effective object disambiguation and enhances the focus on detailed semantics, thereby achieving precise composed feature learning. Moreover, our proposed HUD is also applicable to the Composed Image Retrieval (CIR) task and achieves state-of-the-art performance across three benchmark datasets for both CVR and CIR tasks. The codes are available on https://zivchen-ty.github.io/HUD.github.io/.

Method: Weight Space Correlation Analysis measures alignment between classification heads of primary clinical task and auxiliary metadata tasks to quantify feature utilization. Validated on artificially induced shortcut learning, then applied to SA-SonoNet for spontaneous preterm birth prediction.

Result: Method successfully detected artificial shortcuts. For SA-SonoNet, embeddings contained substantial metadata, but sPTB classifier’s weight vectors correlated with clinically relevant factors (birth weight) and decoupled from clinically irrelevant acquisition factors (scanner).

Conclusion: The methodology provides a tool to verify model trustworthiness, demonstrating that clinical models can selectively utilize genuine clinical signals rather than metadata shortcuts when not artificially biased.

Abstract: Deep learning models in medical imaging are susceptible to shortcut learning, relying on confounding metadata (e.g., scanner model) that is often encoded in image embeddings. The crucial question is whether the model actively utilizes this encoded information for its final prediction. We introduce Weight Space Correlation Analysis, an interpretable methodology that quantifies feature utilization by measuring the alignment between the classification heads of a primary clinical task and auxiliary metadata tasks. We first validate our method by successfully detecting artificially induced shortcut learning. We then apply it to probe the feature utilization of an SA-SonoNet model trained for Spontaneous Preterm Birth (sPTB) prediction. Our analysis confirmed that while the embeddings contain substantial metadata, the sPTB classifier’s weight vectors were highly correlated with clinically relevant factors (e.g., birth weight) but decoupled from clinically irrelevant acquisition factors (e.g. scanner). Our methodology provides a tool to verify model trustworthiness, demonstrating that, in the absence of induced bias, the clinical model selectively utilizes features related to the genuine clinical signal.

Method: Builds attention-weighted graphs over modality features per sample, uses multi-head Graph Attention Networks for message passing, includes learnable mask for missing modalities, trained with hybrid supervised task loss + contrastive InfoNCE loss.

Result: Outperforms baselines and SOTA models across 7 diverse datasets (finance, HCI, multimedia classification, affective computing), shows robustness to missing inputs, and excels on niche tasks.

Conclusion: CLARGA provides an effective, efficient, and easily pluggable multimodal fusion solution that adapts to various tasks and modalities while maintaining robustness to missing data.

Abstract: We introduce CLARGA, a general-purpose multimodal fusion architecture for multimodal representation learning that works with any number and type of modalities without changing the underlying framework. Given a supervised dataset, CLARGA can be applied to virtually any machine learning task to fuse different multimodal representations for processing by downstream layers. On a sample-by-sample basis, CLARGA learns how modalities should inform one another by building an attention weighted graph over their features and passing messages along this graph with a multi-head Graph Attention Network. Not only does this make CLARGA highly adaptive, as it constructs unique graphs for different samples, it makes for efficient fusion with sub-quadratic complexity as the number of modalities grows. Through a learnable mask, it can also adapt to missing modality inputs. The model is trained with a hybrid objective that combines a supervised task loss with contrastive InfoNCE loss, improving cross-modal consistency and robustness to noisy inputs. We demonstrate CLARGA’s effectiveness in diverse multimodal representation learning tasks across 7 datasets spanning finance, human-computer interaction, general multimedia classification, and affective computing. It consistently outperforms baselines, state-of-the-art models, and ablations. Additional experiments also demonstrate its robustness to missing inputs and ability to excel on niche tasks. Overall, CLARGA can be easily plugged into machine learning models for effective and efficient learning of representations across a wide variety of tasks.

Method: Train Neural Cellular Automata (NCA) on imperfect masks and ground truths to learn structural properties of target shapes using only local information. The refinement NCA (rNCA) applies local, iterative updates to coarse masks to progressively reconnect broken regions and prune loose fragments.

Result: For retinal vessels: 2-3% gains in Dice/clDice, 60% reduction in β₀ errors, 20% reduction in β₁ errors. For myocardium: 61.5% repair of broken cases in zero-shot setting, 19% reduction in ASSD, 16% reduction in HD.

Conclusion: NCA serves as an effective and broadly applicable refinement mechanism that can repair common topological errors from different base segmentation models and tasks, producing stable, topologically consistent results.

Abstract: Accurately predicting topologically correct masks remains a difficult task for general segmentation models, which often produce fragmented or disconnected outputs. Fixing these artifacts typically requires hand-crafted refinement rules or architectures specialized to a particular task. Here, we show that Neural Cellular Automata (NCA) can be directly re-purposed as an effective refinement mechanism, using local, iterative updates guided by image context to repair segmentation masks. By training on imperfect masks and ground truths, the automaton learns the structural properties of the target shape while relying solely on local information. When applied to coarse, globally predicted masks, the learned dynamics progressively reconnect broken regions, prune loose fragments and converge towards stable, topologically consistent results. We show how refinement NCA (rNCA) can be easily applied to repair common topological errors produced by different base segmentation models and tasks: for fragmented retinal vessels, it yields 2-3% gains in Dice/clDice and improves Betti errors, reducing $β_0$ errors by 60% and $β_1$ by 20%; for myocardium, it repairs 61.5% of broken cases in a zero-shot setting while lowering ASSD and HD by 19% and 16%, respectively. This showcases NCA as effective and broadly applicable refiners.

Method: Analyzed 405,448 video clips passively recorded from 233 consented participants over one week using a deep learning model to quantify smile intensity. Examined diurnal and daily patterns, and correlated results with national survey data and day reconstruction method.

Result: Daily smile intensity patterns strongly correlated with national happiness survey data (r=0.92). Diurnal rhythms closely matched day reconstruction method results (r=0.80). Higher daily mean smile intensity associated with more physical activity (Beta=0.043) and greater light exposure (Beta=0.038), but not smartphone use.

Conclusion: Passive smartphone sensing of candid smiles offers a powerful, ecologically valid methodology for studying affective behavior dynamics and enables population-scale understanding of well-being.

Abstract: Subjective well-being is a cornerstone of individual and societal health, yet its scientific measurement has traditionally relied on self-report methods prone to recall bias and high participant burden. This has left a gap in our understanding of well-being as it is expressed in everyday life. We hypothesized that candid smiles captured during natural smartphone interactions could serve as a scalable, objective behavioral correlate of positive affect. To test this, we analyzed 405,448 video clips passively recorded from 233 consented participants over one week. Using a deep learning model to quantify smile intensity, we identified distinct diurnal and daily patterns. Daily patterns of smile intensity across the week showed strong correlation with national survey data on happiness (r=0.92), and diurnal rhythms documented close correspondence with established results from the day reconstruction method (r=0.80). Higher daily mean smile intensity was significantly associated with more physical activity (Beta coefficient = 0.043, 95% CI [0.001, 0.085]) and greater light exposure (Beta coefficient = 0.038, [0.013, 0.063]), whereas no significant effects were found for smartphone use. These findings suggest that passive smartphone sensing could serve as a powerful, ecologically valid methodology for studying the dynamics of affective behavior and open the door to understanding this behavior at a population scale.

Method: MPath uses a lightweight multimodal framework that conditions a pretrained biomedical language model (BioBART) on WSI-derived visual embeddings through a learned visual-prefix prompting mechanism. Instead of end-to-end vision-language pretraining, it leverages foundation-model WSI features (CONCH + Titan) and injects them into BioBART via a compact projection module, keeping the language backbone frozen for stability and data efficiency.

Result: MPath was developed and evaluated on the RED 2025 Grand Challenge dataset and ranked 4th in Test Phase 2, despite limited submission opportunities. The results demonstrate the framework’s effectiveness in pathology report generation.

Conclusion: The results highlight the potential of prompt-based multimodal conditioning as a scalable and interpretable strategy for pathology report generation, offering a lightweight alternative to end-to-end vision-language pretraining approaches.

Abstract: Automated generation of diagnostic pathology reports directly from whole slide images (WSIs) is an emerging direction in computational pathology. Translating high-resolution tissue patterns into clinically coherent text remains difficult due to large morphological variability and the complex structure of pathology narratives. We introduce MPath, a lightweight multimodal framework that conditions a pretrained biomedical language model (BioBART) on WSI-derived visual embeddings through a learned visual-prefix prompting mechanism. Instead of end-to-end vision-language pretraining, MPath leverages foundation-model WSI features (CONCH + Titan) and injects them into BioBART via a compact projection module, keeping the language backbone frozen for stability and data efficiency. MPath was developed and evaluated on the RED 2025 Grand Challenge dataset and ranked 4th in Test Phase 2, despite limited submission opportunities. The results highlight the potential of prompt-based multimodal conditioning as a scalable and interpretable strategy for pathology report generation.

Method: FloraForge uses LLM-assisted co-design to refine Python scripts that generate parameterized plant geometries as hierarchical B-spline surface representations with botanical constraints. The framework enables iterative natural language Plant Refinements (PR) and uses Plant Descriptor (PD) files for manual refinement, producing dual outputs: triangular meshes for visualization and meshes with parametric metadata for analysis.

Result: The framework successfully generates biologically accurate 3D plant models for maize, soybean, and mung bean by fitting procedural models to empirical point cloud data. It produces mathematically continuous representations compatible with functional structural plant analysis workflows including light simulation, computational fluid dynamics, and finite element analysis.

Conclusion: FloraForge democratizes sophisticated geometric modeling for plant science by combining LLM-assisted template creation, mathematically continuous representations for both phenotyping and rendering, and direct parametric control through PD files, while maintaining mathematical rigor and accessibility for domain experts.

Abstract: Accurate 3D plant models are crucial for computational phenotyping and physics-based simulation; however, current approaches face significant limitations. Learning-based reconstruction methods require extensive species-specific training data and lack editability. Procedural modeling offers parametric control but demands specialized expertise in geometric modeling and an in-depth understanding of complex procedural rules, making it inaccessible to domain scientists. We present FloraForge, an LLM-assisted framework that enables domain experts to generate biologically accurate, fully parametric 3D plant models through iterative natural language Plant Refinements (PR), minimizing programming expertise. Our framework leverages LLM-enabled co-design to refine Python scripts that generate parameterized plant geometries as hierarchical B-spline surface representations with botanical constraints with explicit control points and parametric deformation functions. This representation can be easily tessellated into polygonal meshes with arbitrary precision, ensuring compatibility with functional structural plant analysis workflows such as light simulation, computational fluid dynamics, and finite element analysis. We demonstrate the framework on maize, soybean, and mung bean, fitting procedural models to empirical point cloud data through manual refinement of the Plant Descriptor (PD), human-readable files. The pipeline generates dual outputs: triangular meshes for visualization and triangular meshes with additional parametric metadata for quantitative analysis. This approach uniquely combines LLM-assisted template creation, mathematically continuous representations enabling both phenotyping and rendering, and direct parametric control through PD. The framework democratizes sophisticated geometric modeling for plant science while maintaining mathematical rigor.

Method: Proposes TransBridge - a transformer-based up-sampling block that fuses features from detection and completion networks. Includes Dynamic-Static Reconstruction (DSRecon) module to generate dense LiDAR data for training completion network. Uses transformer mechanisms to establish channel and spatial relations for high-resolution feature maps.

Result: Extensive experiments on nuScenes and Waymo datasets show consistent improvement in end-to-end 3D object detection with mAP gains of 0.7-1.5 across multiple methods. For two-stage detection frameworks, boosts mAP up to 5.78 points.

Conclusion: The joint completion-detection framework effectively improves 3D object detection in sparse regions while maintaining computational costs, demonstrating strong generalization ability across different datasets and detection methods.

Abstract: 3D object detection is essential in autonomous driving, providing vital information about moving objects and obstacles. Detecting objects in distant regions with only a few LiDAR points is still a challenge, and numerous strategies have been developed to address point cloud sparsity through densification.This paper presents a joint completion and detection framework that improves the detection feature in sparse areas while maintaining costs unchanged. Specifically, we propose TransBridge, a novel transformer-based up-sampling block that fuses the features from the detection and completion networks.The detection network can benefit from acquiring implicit completion features derived from the completion network. Additionally, we design the Dynamic-Static Reconstruction (DSRecon) module to produce dense LiDAR data for the completion network, meeting the requirement for dense point cloud ground truth.Furthermore, we employ the transformer mechanism to establish connections between channels and spatial relations, resulting in a high-resolution feature map used for completion purposes.Extensive experiments on the nuScenes and Waymo datasets demonstrate the effectiveness of the proposed framework.The results show that our framework consistently improves end-to-end 3D object detection, with the mean average precision (mAP) ranging from 0.7 to 1.5 across multiple methods, indicating its generalization ability. For the two-stage detection framework, it also boosts the mAP up to 5.78 points.

Method: Train a diffusion model (MONET) on a large dataset to predict cell paint channels from brightfield images. The model uses a consistency architecture to generate time-lapse videos and enables in-context learning for transfer to out-of-distribution cell lines and imaging protocols.

Result: Model quality improves with scale. The consistency architecture successfully generates time-lapse videos despite lacking cell paint video training data, and enables partial transfer to out-of-distribution cell lines and imaging protocols through in-context learning.

Conclusion: Virtual cell painting via MONET serves as a complementary tool to physical cell painting, enabling novel biological research workflows by overcoming the limitations of traditional cell painting methods.

Abstract: Cell painting is a popular technique for creating human-interpretable, high-contrast images of cell morphology. There are two major issues with cell paint: (1) it is labor-intensive and (2) it requires chemical fixation, making the study of cell dynamics impossible. We train a diffusion model (Morphological Observation Neural Enhancement Tool, or MONET) on a large dataset to predict cell paint channels from brightfield images. We show that model quality improves with scale. The model uses a consistency architecture to generate time-lapse videos, despite the impossibility of obtaining cell paint video training data. In addition, we show that this architecture enables a form of in-context learning, allowing the model to partially transfer to out-of-distribution cell lines and imaging protocols. Virtual cell painting is not intended to replace physical cell painting completely, but to act as a complementary tool enabling novel workflows in biological research.

Method: Proposes HEMC-CPS model integrating contextual Peano scan (CPS) with hidden evidential Markov chains (HEMC). Uses Bayesian maximum posterior mode (MPM) segmentation with unsupervised parameter estimation via stochastic expectation-maximization (SEM).

Result: Demonstrates effectiveness of HEMC-CPS for Bayesian MPM segmentation using both synthetic and real images. Shows potential for modeling complex images including 3D and multi-sensor multi-resolution data.

Conclusion: HEMC-CPS presents a promising approach for unsupervised image segmentation that combines computational efficiency with improved modeling capabilities, with potential applications beyond image segmentation to any spatially correlated data.

Abstract: Transforming bi-dimensional sets of image pixels into mono-dimensional sequences with a Peano scan (PS) is an established technique enabling the use of hidden Markov chains (HMCs) for unsupervised image segmentation. Related Bayesian segmentation methods can compete with hidden Markov fields (HMFs)-based ones and are much faster. PS has recently been extended to the contextual PS, and some initial experiments have shown the value of the associated HMC model, denoted as HMC-CPS, in image segmentation. Moreover, HMCs have been extended to hidden evidential Markov chains (HEMCs), which are capable of improving HMC-based Bayesian segmentation. In this study, we introduce a new HEMC-CPS model by simultaneously considering contextual PS and evidential HMC. We show its effectiveness for Bayesian maximum posterior mode (MPM) segmentation using synthetic and real images. Segmentation is performed in an unsupervised manner, with parameters being estimated using the stochastic expectation–maximization (SEM) method. The new HEMC-CPS model presents potential for the modeling and segmentation of more complex images, such as three-dimensional or multi-sensor multi-resolution images. Finally, the HMC-CPS and HEMC-CPS models are not limited to image segmentation and could be used for any kind of spatially correlated data.

Method: 1) Uses large language model to generate hierarchical textual descriptions (global movements + local body-part dynamics), 2) Adaptive partitioning module produces fine-grained visual representations by semantically grouping skeleton joints, 3) Dynamic refinement module adapts textual features to incoming visual stream during inference via lightweight learnable projection, 4) Confidence-aware, class-balanced memory bank stabilizes refinement and mitigates error propagation from noisy pseudo-labels.

Result: Extensive experiments on NTU RGB+D 60/120 and PKU-MMD datasets show DynaPURLS significantly outperforms prior methods and sets new state-of-the-art records.

Conclusion: DynaPURLS successfully addresses limitations of coarse-grained alignment in zero-shot skeleton-based action recognition by establishing robust multi-scale visual-semantic correspondences with dynamic refinement, achieving superior generalization to unseen classes.

Abstract: Zero-shot skeleton-based action recognition (ZS-SAR) is fundamentally constrained by prevailing approaches that rely on aligning skeleton features with static, class-level semantics. This coarse-grained alignment fails to bridge the domain shift between seen and unseen classes, thereby impeding the effective transfer of fine-grained visual knowledge. To address these limitations, we introduce \textbf{DynaPURLS}, a unified framework that establishes robust, multi-scale visual-semantic correspondences and dynamically refines them at inference time to enhance generalization. Our framework leverages a large language model to generate hierarchical textual descriptions that encompass both global movements and local body-part dynamics. Concurrently, an adaptive partitioning module produces fine-grained visual representations by semantically grouping skeleton joints. To fortify this fine-grained alignment against the train-test domain shift, DynaPURLS incorporates a dynamic refinement module. During inference, this module adapts textual features to the incoming visual stream via a lightweight learnable projection. This refinement process is stabilized by a confidence-aware, class-balanced memory bank, which mitigates error propagation from noisy pseudo-labels. Extensive experiments on three large-scale benchmark datasets, including NTU RGB+D 60/120 and PKU-MMD, demonstrate that DynaPURLS significantly outperforms prior art, setting new state-of-the-art records. The source code is made publicly available at https://github.com/Alchemist0754/DynaPURLS

Method: Used Data-Efficient Image Transformer (DeiT) for wafer map defect classification under data-constrained conditions, comparing against CNN models (VGG-19, SqueezeNet, Xception, Hybrid).

Result: DeiT achieved highest classification accuracy (90.83%), superior F1-score (90.78%), faster training convergence, and better robustness for minority defect classes compared to CNNs.

Conclusion: Transformer-based models like DeiT show strong potential for semiconductor wafer defect detection and can enhance predictive maintenance strategies in semiconductor fabrication.

Abstract: Predictive maintenance is an important sector in modern industries which improves fault detection and cost reduction processes. By using machine learning algorithms in the whole process, the defects detection process can be implemented smoothly. Semiconductor is a sensitive maintenance field that requires predictability in work. While convolutional neural networks (CNNs) such as VGG-19, Xception and Squeeze-Net have demonstrated solid performance in image classification for semiconductor wafer industry, their effectiveness often declines in scenarios with limited and imbalanced data. This study investigates the use of the Data-Efficient Image Transformer (DeiT) for classifying wafer map defects under data-constrained conditions. Experimental results reveal that the DeiT model achieves highest classification accuracy of 90.83%, outperforming CNN models such as VGG-19(65%), SqueezeNet(82%), Xception(66%) and Hybrid(67%). DeiT also demonstrated superior F1-score (90.78%) and faster training convergence, with enhanced robustness in detecting minority defect classes. These findings highlight the potential of transformer-based models like DeiT in semiconductor wafer defect detection and support predictive maintenance strategies within semiconductor fabrication processes.

Method: Proposes a pose hypothesis selection algorithm that integrates predictions from foundation models, jointly refines them through learned render-and-compare for spatial/temporal/pixel alignment, and reasons about intricate contacts with physical constraints.

Result: Outperforms prior art by 38% on in-distribution dataset and 36% on unseen dataset in reconstruction error, generalizes beyond training categories, and works zero-shot on in-the-wild internet videos.

Conclusion: CARI4D is the first category-agnostic method for consistent 4D human-object interaction reconstruction from monocular RGB videos, achieving state-of-the-art performance and generalization to unseen categories.

Abstract: Accurate capture of human-object interaction from ubiquitous sensors like RGB cameras is important for applications in human understanding, gaming, and robot learning. However, inferring 4D interactions from a single RGB view is highly challenging due to the unknown object and human information, depth ambiguity, occlusion, and complex motion, which hinder consistent 3D and temporal reconstruction. Previous methods simplify the setup by assuming ground truth object template or constraining to a limited set of object categories. We present CARI4D, the first category-agnostic method that reconstructs spatially and temporarily consistent 4D human-object interaction at metric scale from monocular RGB videos. To this end, we propose a pose hypothesis selection algorithm that robustly integrates the individual predictions from foundation models, jointly refine them through a learned render-and-compare paradigm to ensure spatial, temporal and pixel alignment, and finally reasoning about intricate contacts for further refinement satisfying physical constraints. Experiments show that our method outperforms prior art by 38% on in-distribution dataset and 36% on unseen dataset in terms of reconstruction error. Our model generalizes beyond the training categories and thus can be applied zero-shot to in-the-wild internet videos. Our code and pretrained models will be publicly released.

Method: Developed V-REX evaluation suite with: (1) Benchmark of challenging visual reasoning tasks requiring multi-step exploration, (2) Evaluation protocol using Chain-of-Questions (CoQ) approach, (3) Disentangling VLMs’ capabilities into Planning (breaking down tasks by selecting exploratory questions) and Following (answering curated CoQ sequentially), (4) Curating finite options of questions and answers per step for reliable quantitative analysis.

Result: Evaluation of SOTA proprietary and open-sourced VLMs revealed: (1) Consistent scaling trends, (2) Significant differences between planning and following abilities, (3) Substantial room for improvement in multi-step exploratory reasoning.

Conclusion: V-REX provides a reliable framework for fine-grained analysis of VLMs’ multi-step visual reasoning capabilities, highlighting current limitations and offering a standardized evaluation approach for future development in this challenging area.

Abstract: While many vision-language models (VLMs) are developed to answer well-defined, straightforward questions with highly specified targets, as in most benchmarks, they often struggle in practice with complex open-ended tasks, which usually require multiple rounds of exploration and reasoning in the visual space. Such visual thinking paths not only provide step-by-step exploration and verification as an AI detective but also produce better interpretations of the final answers. However, these paths are challenging to evaluate due to the large exploration space of intermediate steps. To bridge the gap, we develop an evaluation suite, ``Visual Reasoning with multi-step EXploration (V-REX)’’, which is composed of a benchmark of challenging visual reasoning tasks requiring native multi-step exploration and an evaluation protocol. V-REX covers rich application scenarios across diverse domains. V-REX casts the multi-step exploratory reasoning into a Chain-of-Questions (CoQ) and disentangles VLMs’ capability to (1) Planning: breaking down an open-ended task by selecting a chain of exploratory questions; and (2) Following: answering curated CoQ sequentially to collect information for deriving the final answer. By curating finite options of questions and answers per step, V-REX achieves a reliable quantitative and fine-grained analysis of the intermediate steps. By assessing SOTA proprietary and open-sourced VLMs, we reveal consistent scaling trends, significant differences between planning and following abilities, and substantial room for improvement in multi-step exploratory reasoning.

Method: Two-stage neuro-symbolic approach: (1) Symbolic Grounding via real-time open-vocabulary detector (YOLOE) to anchor attention, (2) Cognitive Analysis via Reasoning VLM for forensic scene analysis. Uses “System 2” inference-time alignment with multi-model “Judge-Scout” consensus to mitigate hallucination.

Result: Achieves Recall of 0.966 (vs. 0.475 for CLIP) on nuScenes benchmark against Waymo Open Dataset taxonomy, reduces Risk Assessment Error by 40% compared to single models. Runs entirely on consumer hardware (NVIDIA RTX 3090).

Conclusion: Semantic-Drive provides an effective, privacy-preserving alternative to cloud-based solutions for mining rare safety-critical events in AV data, enabling better training data collection for long-tail scenarios.

Abstract: The development of robust Autonomous Vehicles (AVs) is bottlenecked by the scarcity of “Long-Tail” training data. While fleets collect petabytes of video logs, identifying rare safety-critical events (e.g., erratic jaywalking, construction diversions) remains a manual, cost-prohibitive process. Existing solutions rely on coarse metadata search, which lacks precision, or cloud-based VLMs, which are privacy-invasive and expensive. We introduce Semantic-Drive, a local-first, neuro-symbolic framework for semantic data mining. Our approach decouples perception into two stages: (1) Symbolic Grounding via a real-time open-vocabulary detector (YOLOE) to anchor attention, and (2) Cognitive Analysis via a Reasoning VLM that performs forensic scene analysis. To mitigate hallucination, we implement a “System 2” inference-time alignment strategy, utilizing a multi-model “Judge-Scout” consensus mechanism. Benchmarked on the nuScenes dataset against the Waymo Open Dataset (WOD-E2E) taxonomy, Semantic-Drive achieves a Recall of 0.966 (vs. 0.475 for CLIP) and reduces Risk Assessment Error by 40% compared to single models. The system runs entirely on consumer hardware (NVIDIA RTX 3090), offering a privacy-preserving alternative to the cloud.

Method: Evaluated four FL models (FedAvg, FedProx, FedOpt, FedMedian) using YOLOv8 ship detection, compared to local training baseline and global training reference.

Result: FL models substantially improve accuracy over local training on smaller datasets and achieve performance close to global training using all datasets.

Conclusion: FL offers effective privacy-preserving solution for ship detection, with proper configuration (communication rounds, local epochs) crucial for optimizing precision and efficiency.

Abstract: We investigate the application of Federated Learning (FL) for ship detection across diverse satellite datasets, offering a privacy-preserving solution that eliminates the need for data sharing or centralized collection. This approach is particularly advantageous for handling commercial satellite imagery or sensitive ship annotations. Four FL models including FedAvg, FedProx, FedOpt, and FedMedian, are evaluated and compared to a local training baseline, where the YOLOv8 ship detection model is independently trained on each dataset without sharing learned parameters. The results reveal that FL models substantially improve detection accuracy over training on smaller local datasets and achieve performance levels close to global training that uses all datasets during the training. Furthermore, the study underscores the importance of selecting appropriate FL configurations, such as the number of communication rounds and local training epochs, to optimize detection precision while maintaining computational efficiency.

Method: Supervised semantic segmentation workflow using SPOT-6/7 imagery, comparing U-Net and SegFormer models, incorporating land cover data as auxiliary task, and testing Test-Time Augmentation with Mixed Precision optimization.

Result: U-Net and SegFormer perform similarly with limited training data, but SegFormer requires more resources. Land cover data enhances robustness without increasing inference time. Test-Time Augmentation improves performance but increases inference time, which can be mitigated with Mixed Precision.

Conclusion: The proposed workflow balances performance and efficiency for emergency burned area delineation, with U-Net being more practical than SegFormer for resource-constrained scenarios, and optimization techniques enabling performance improvements without excessive computational costs.

Abstract: After a wildfire, delineating burned areas (BAs) is crucial for quantifying damages and supporting ecosystem recovery. Current BA mapping approaches rely on computer vision models trained on post-event remote sensing imagery, but often overlook their applicability to time-constrained emergency management scenarios. This study introduces a supervised semantic segmentation workflow aimed at boosting both the performance and efficiency of BA delineation. It targets SPOT-6/7 imagery due to its very high resolution and on-demand availability. Experiments are evaluated based on Dice score, Intersection over Union, and inference time. The results show that U-Net and SegFormer models perform similarly with limited training data. However, SegFormer requires more resources, challenging its practical use in emergencies. Incorporating land cover data as an auxiliary task enhances model robustness without increasing inference time. Lastly, Test-Time Augmentation improves BA delineation performance but raises inference time, which can be mitigated with optimization methods like Mixed Precision.

Method: Backwards Aggregation (BAgger) constructs corrective trajectories from the model’s own rollouts to teach recovery from mistakes. It trains with standard score or flow matching objectives instead of few-step distillation or distribution-matching losses.

Result: BAgger applied to causal diffusion transformers shows more stable long-horizon motion, better visual consistency, and reduced drift in text-to-video, video extension, and multi-prompt generation tasks.

Conclusion: BAgger effectively addresses exposure bias in autoregressive video models through self-supervised corrective training, enabling more stable and consistent long-horizon video generation without complex distillation approaches.

Abstract: Autoregressive video models are promising for world modeling via next-frame prediction, but they suffer from exposure bias: a mismatch between training on clean contexts and inference on self-generated frames, causing errors to compound and quality to drift over time. We introduce Backwards Aggregation (BAgger), a self-supervised scheme that constructs corrective trajectories from the model’s own rollouts, teaching it to recover from its mistakes. Unlike prior approaches that rely on few-step distillation and distribution-matching losses, which can hurt quality and diversity, BAgger trains with standard score or flow matching objectives, avoiding large teachers and long-chain backpropagation through time. We instantiate BAgger on causal diffusion transformers and evaluate on text-to-video, video extension, and multi-prompt generation, observing more stable long-horizon motion and better visual consistency with reduced drift.

Method: VBLE-xz is a deep regularization method that solves inverse problems in the latent space of a variational compressive autoencoder (CAE). It adapts regularization strength by modulating the CAE’s bitrate without retraining, and estimates uncertainties in both latent and image spaces through variational inference for fast posterior sampling.

Result: Experiments on very high-resolution simulated and real Pléiades images demonstrate VBLE-xz’s performance, robustness, and scalability. The method provides compelling restoration results with uncertainty quantification, outperforming direct inversion methods when uncertainty estimation is required.

Conclusion: VBLE-xz represents an effective alternative to direct inversion methods for satellite image restoration when uncertainty quantification is needed, offering sensor-agnostic restoration with fast uncertainty estimation through variational inference in latent space.

Abstract: Satellite optical images, upon their on-ground receipt, offer a distorted view of the observed scene. Their restoration, including denoising, deblurring, and sometimes super-resolution, is required before their exploitation. Moreover, quantifying the uncertainties related to this restoration helps to reduce the risks of misinterpreting the image content. Deep learning methods are now state-of-the-art for satellite image restoration. Among them, direct inversion methods train a specific network for each sensor, and generally provide a point estimation of the restored image without the associated uncertainties. Alternatively, deep regularization (DR) methods learn a deep prior on target images before plugging it, as the regularization term, into a model-based optimization scheme. This allows for restoring images from several sensors with a single network and possibly for estimating associated uncertainties. In this paper, we introduce VBLE-xz, a DR method that solves the inverse problem in the latent space of a variational compressive autoencoder (CAE). We adapt the regularization strength by modulating the bitrate of the trained CAE with a training-free approach. Then, VBLE-xz estimates relevant uncertainties jointly in the latent and in the image spaces by sampling an explicit posterior estimated within variational inference. This enables fast posterior sampling, unlike state-of-the-art DR methods that use Markov chains or diffusion-based approaches. We conduct a comprehensive set of experiments on very high-resolution simulated and real Pléiades images, asserting the performance, robustness and scalability of the proposed method. They demonstrate that VBLE-xz represents a compelling alternative to direct inversion methods when uncertainty quantification is required. The code associated to this paper is available in https://github.com/MaudBqrd/VBLExz.

Method: RePack transforms VFM representations into compact, decoder-friendly forms by projecting onto low-dimensional manifolds, filtering non-semantic noise while preserving core structural information.

Result: RePack achieves FID of 3.66 in only 64 epochs on DiT-XL/2, 35% faster convergence than state-of-the-art methods, outperforming direct VFM feature injection approaches.

Conclusion: RePack successfully extracts core semantics from VFM representations while avoiding high-dimensionality side effects, enabling more efficient and effective diffusion model training.

Abstract: The superior representation capability of pre-trained vision foundation models (VFMs) has been harnessed for enhancing latent diffusion models (LDMs). These approaches inject the rich semantics from high-dimensional VFM representations (e.g., DINOv3) into LDMs at different phases, resulting in accelerated learning and better generation performance. However, the high-dimensionality of VFM representations may also lead to Information Overload, particularly when the VFM features exceed the size of the original image for decoding. To address this issue while preserving the utility of VFM features, we propose RePack (Representation Packing), a simple yet effective framework for improving Diffusion Transformers (DiTs). RePack transforms the VFM representation into a more compact, decoder-friendly representation by projecting onto low-dimensional manifolds. We find that RePack can effectively filter out non-semantic noise while preserving the core structural information needed for high-fidelity reconstruction. Experimental results show that RePack significantly accelerates DiT convergence and outperforms recent methods that directly inject raw VFM features into the decoder for image reconstruction. On DiT-XL/2, RePack achieves an FID of 3.66 in only 64 epochs, which is 35% faster than the state-of-the-art method. This demonstrates that RePack successfully extracts the core semantics of VFM representations while bypassing their high-dimensionality side effects.

Method: Integrated mode normalization into two semantic segmentation models (U-Net and SegNet) to reduce convergence time while maintaining baseline performance.

Result: Mode normalization significantly accelerates convergence and normalized models show increased stability across different zones in cross-validation experiments.

Conclusion: Normalization effectively improves computational efficiency and generalization in SAR image segmentation, making it valuable for remote sensing applications.

Abstract: Segmenting Synthetic Aperture Radar (SAR) images is crucial for many remote sensing applications, particularly water body detection. However, deep learning-based segmentation models often face challenges related to convergence speed and stability, mainly due to the complex statistical distribution of this type of data. In this study, we evaluate the impact of mode normalization on two widely used semantic segmentation models, U-Net and SegNet. Specifically, we integrate mode normalization, to reduce convergence time while maintaining the performance of the baseline models. Experimental results demonstrate that mode normalization significantly accelerates convergence. Furthermore, cross-validation results indicate that normalized models exhibit increased stability in different zones. These findings highlight the effectiveness of normalization in improving computational efficiency and generalization in SAR image segmentation.

Method: VEGAS integrates the vision encoder’s attention maps into the language model’s mid-layers during inference. It analyzes that vision-text conflicts peak in middle layers, and injecting concentrated attention maps from the vision encoder adaptively steers tokens that fail to focus on key image objects.

Result: Extensive experiments across multiple benchmarks show VEGAS consistently achieves state-of-the-art performance in reducing hallucinations. The method effectively suppresses hallucinations by leveraging the vision encoder’s more concentrated attention maps.

Conclusion: The vision encoder’s own attention map is the key to suppressing hallucinations in LVLMs. By injecting these maps into language model’s middle layers during inference, VEGAS provides a simple yet effective solution to reduce factually inconsistent outputs while maintaining linguistic fluency.

Abstract: Large vision-language models (LVLMs) exhibit impressive ability to jointly reason over visual and textual inputs. However, they often produce outputs that are linguistically fluent but factually inconsistent with the visual evidence, i.e., they hallucinate. Despite growing efforts to mitigate such hallucinations, a key question remains: what form of visual attention can effectively suppress hallucinations during decoding? In this work, we provide a simple answer: the vision encoder’s own attention map. We show that LVLMs tend to hallucinate when their final visual-attention maps fail to concentrate on key image objects, whereas the vision encoder’s more concentrated attention maps substantially reduce hallucinations. To further investigate the cause, we analyze vision-text conflicts during decoding and find that these conflicts peak in the language model’s middle layers. Injecting the vision encoder’s attention maps into these layers effectively suppresses hallucinations. Building on these insights, we introduce VEGAS, a simple yet effective inference-time method that integrates the vision encoder’s attention maps into the language model’s mid-layers and adaptively steers tokens which fail to concentrate on key image objects. Extensive experiments across multiple benchmarks demonstrate that VEGAS consistently achieves state-of-the-art performance in reducing hallucinations.

Method: SPDMark embeds watermarks by modifying a subset of video diffusion model parameters through additive composition of layer-wise basis shifts implemented via LoRA for parameter efficiency. It uses cryptographic hashing for frame-specific messages and maximum bipartite matching for temporal tamper recovery. Joint training optimizes basis shifts and extractor with message recovery, perceptual similarity, and temporal consistency losses.

Result: SPDMark generates imperceptible watermarks that can be recovered with high accuracy and demonstrates robustness against various common video modifications in both text-to-video and image-to-video generation models.

Conclusion: SPDMark provides an effective framework for in-generation video watermarking that addresses the limitations of existing methods by achieving the desired balance of imperceptibility, robustness, and computational efficiency for tracking AI-generated video provenance.

Abstract: The advent of high-quality video generation models has amplified the need for robust watermarking schemes that can be used to reliably detect and track the provenance of generated videos. Existing video watermarking methods based on both post-hoc and in-generation approaches fail to simultaneously achieve imperceptibility, robustness, and computational efficiency. This work introduces a novel framework for in-generation video watermarking called SPDMark (pronounced `SpeedMark’) based on selective parameter displacement of a video diffusion model. Watermarks are embedded into the generated videos by modifying a subset of parameters in the generative model. To make the problem tractable, the displacement is modeled as an additive composition of layer-wise basis shifts, where the final composition is indexed by the watermarking key. For parameter efficiency, this work specifically leverages low-rank adaptation (LoRA) to implement the basis shifts. During the training phase, the basis shifts and the watermark extractor are jointly learned by minimizing a combination of message recovery, perceptual similarity, and temporal consistency losses. To detect and localize temporal modifications in the watermarked videos, we use a cryptographic hashing function to derive frame-specific watermark messages from the given base watermarking key. During watermark extraction, maximum bipartite matching is applied to recover the correct frame order, even from temporally tampered videos. Evaluations on both text-to-video and image-to-video generation models demonstrate the ability of SPDMark to generate imperceptible watermarks that can be recovered with high accuracy and also establish its robustness against a variety of common video modifications.

Method: PANDA uses CLIP encoder to identify representative regions of low-frequency classes and transplants them into frequent-class samples within each task. It also incorporates adaptive balancing strategy leveraging prior task distributions to smooth inter-task imbalances, reducing gaps between average samples across tasks.

Result: Extensive experiments and ablation studies demonstrate PANDA’s capability to work with existing PTM-based continual learning methods, improving accuracy and reducing catastrophic forgetting.

Conclusion: PANDA effectively addresses dual-level data imbalances in exemplar-free continual learning, enhancing performance when integrated with existing pre-trained model-based methods.

Abstract: Exemplar-Free Continual Learning (EFCL) restricts the storage of previous task data and is highly susceptible to catastrophic forgetting. While pre-trained models (PTMs) are increasingly leveraged for EFCL, existing methods often overlook the inherent imbalance of real-world data distributions. We discovered that real-world data streams commonly exhibit dual-level imbalances, dataset-level distributions combined with extreme or reversed skews within individual tasks, creating both intra-task and inter-task disparities that hinder effective learning and generalization. To address these challenges, we propose PANDA, a Patch-and-Distribution-Aware Augmentation framework that integrates seamlessly with existing PTM-based EFCL methods. PANDA amplifies low-frequency classes by using a CLIP encoder to identify representative regions and transplanting those into frequent-class samples within each task. Furthermore, PANDA incorporates an adaptive balancing strategy that leverages prior task distributions to smooth inter-task imbalances, reducing the overall gap between average samples across tasks and enabling fairer learning with frozen PTMs. Extensive experiments and ablation studies demonstrate PANDA’s capability to work with existing PTM-based CL methods, improving accuracy and reducing catastrophic forgetting.

Method: Used YOLOv8s for object detection and MobileNetV3L for classification on dual-modality dataset of optical and DIHM images. Employed Wasserstein GAN with spectral normalization (WGAN-SN) to create synthetic DIHM images for data augmentation, mixing real and synthetic data at 1.0:1.5 ratio.

Result: Optical data achieved 91.3% mAP50 for detection and 97% classification accuracy, while DIHM data only achieved 8.15% mAP50 and 50% accuracy. Expanding bounding boxes improved DIHM to 13.3% mAP50 and 54% accuracy. GAN augmentation with synthetic data improved DIHM detection to 15.4% mAP50.

Conclusion: GAN-based data augmentation can reduce the performance gap between optical and holographic microscopy for pollen recognition, bringing automated DIHM workflows for veterinary imaging closer to practical implementation.

Abstract: We present a comprehensive study on fully automated pollen recognition across both conventional optical and digital in-line holographic microscopy (DIHM) images of sample slides. Visually recognizing pollen in unreconstructed holographic images remains challenging due to speckle noise, twin-image artifacts and substantial divergence from bright-field appearances. We establish the performance baseline by training YOLOv8s for object detection and MobileNetV3L for classification on a dual-modality dataset of automatically annotated optical and affinely aligned DIHM images. On optical data, detection mAP50 reaches 91.3% and classification accuracy reaches 97%, whereas on DIHM data, we achieve only 8.15% for detection mAP50 and 50% for classification accuracy. Expanding the bounding boxes of pollens in DIHM images over those acquired in aligned optical images achieves 13.3% for detection mAP50 and 54% for classification accuracy. To improve object detection in DIHM images, we employ a Wasserstein GAN with spectral normalization (WGAN-SN) to create synthetic DIHM images, yielding an FID score of 58.246. Mixing real-world and synthetic data at the 1.0 : 1.5 ratio for DIHM images improves object detection up to 15.4%. These results demonstrate that GAN-based augmentation can reduce the performance divide, bringing fully automated DIHM workflows for veterinary imaging a small but important step closer to practice.

Method: Created EchoGround-MIMIC dataset (19,065 image-text pairs from 1,572 patients) with standardized views, structured measurements, measurement-grounded captions, and disease labels. Developed EchoVLM with two novel pretraining objectives: view-informed contrastive loss (encodes view-dependent structure) and negation-aware contrastive loss (distinguishes negative from positive findings).

Result: Achieved state-of-the-art performance across 36 tasks in 5 clinical applications: 86.5% AUC in zero-shot disease classification, 95.1% accuracy in view classification, plus strong results in image-text retrieval, chamber segmentation, and landmark detection.

Conclusion: Clinically grounded multimodal pretraining yields transferable visual representations. EchoVLM serves as a foundation model for end-to-end echocardiography interpretation. Dataset and code will be released for reproducibility and further research.

Abstract: Echocardiography is the most widely used imaging modality in cardiology, yet its interpretation remains labor-intensive and inherently multimodal, requiring view recognition, quantitative measurements, qualitative assessments, and guideline-based reasoning. While recent vision-language models (VLMs) have achieved broad success in natural images and certain medical domains, their potential in echocardiography has been limited by the lack of large-scale, clinically grounded image-text datasets and the absence of measurement-based reasoning central to echo interpretation. We introduce EchoGround-MIMIC, the first measurement-grounded multimodal echocardiography dataset, comprising 19,065 image-text pairs from 1,572 patients with standardized views, structured measurements, measurement-grounded captions, and guideline-derived disease labels. Building on this resource, we propose EchoVLM, a vision-language model that incorporates two novel pretraining objectives: (i) a view-informed contrastive loss that encodes the view-dependent structure of echocardiographic imaging, and (ii) a negation-aware contrastive loss that distinguishes clinically critical negative from positive findings. Across five types of clinical applications with 36 tasks spanning multimodal disease classification, image-text retrieval, view classification, chamber segmentation, and landmark detection, EchoVLM achieves state-of-the-art performance (86.5% AUC in zero-shot disease classification and 95.1% accuracy in view classification). We demonstrate that clinically grounded multimodal pretraining yields transferable visual representations and establish EchoVLM as a foundation model for end-to-end echocardiography interpretation. We will release EchoGround-MIMIC and the data curation code, enabling reproducibility and further research in multimodal echocardiography interpretation.

Method: Created dataset with 9 complex real scenes captured using two identical DSLR camera assemblies, systematically varying 10 focal lengths (28-70mm) and 5 apertures (f/2.8-f/22), totaling 50 optical configurations and 2000 images per scene with dedicated calibration sets.

Result: Dataset enables controlled analysis of geometric/optical effects for depth estimation, DoF rendering, deblurring, 3D reconstruction, and novel view synthesis, while revealing challenges with current state-of-the-art methods.

Conclusion: Dataset bridges realism gap between synthetic data and real camera optics, supports reproducible research on optical generalization, and is released with calibration files and evaluation code.

Abstract: Reliable depth estimation under real optical conditions remains a core challenge for camera vision in systems such as autonomous robotics and augmented reality. Despite recent progress in depth estimation and depth-of-field rendering, research remains constrained by the lack of large-scale, high-fidelity, real stereo DSLR datasets, limiting real-world generalization and evaluation of models trained on synthetic data as shown extensively in literature. We present the first high-resolution (5472$\times$3648px) stereo DSLR dataset with 18000 images, systematically varying focal length and aperture across complex real scenes and capturing the optical realism and complexity of professional camera systems. For 9 scenes with varying scene complexity, lighting and background, images are captured with two identical camera assemblies at 10 focal lengths (28-70mm) and 5 apertures (f/2.8-f/22), spanning 50 optical configurations in 2000 images per scene. This full-range optics coverage enables controlled analysis of geometric and optical effects for monocular and stereo depth estimation, shallow depth-of-field rendering, deblurring, 3D scene reconstruction and novel view synthesis. Each focal configuration has a dedicated calibration image set, supporting evaluation of classical and learning based methods for intrinsic and extrinsic calibration. The dataset features challenging visual elements such as multi-scale optical illusions, reflective surfaces, mirrors, transparent glass walls, fine-grained details, and natural / artificial ambient light variations. This work attempts to bridge the realism gap between synthetic training data and real camera optics, and demonstrates challenges with the current state-of-the-art monocular, stereo depth and depth-of-field methods. We release the dataset, calibration files, and evaluation code to support reproducible research on real-world optical generalization.

Method: Proposes a patch-based persistent homology construction approach specifically designed for volumetric medical imaging data (CT modality), with comprehensive experiments across multiple datasets.

Result: Patch-based TDA significantly outperforms cubical complex method with average improvements of 10.38% accuracy, 6.94% AUC, 2.06% sensitivity, 11.58% specificity, and 8.51% F1 score across all datasets, while being more time-efficient.

Conclusion: The patch-based approach demonstrates superior performance and efficiency for topological feature extraction in CT imaging, with a released Python package (Patch-TDA) to facilitate adoption.

Abstract: The development of machine learning (ML) models based on computed tomography (CT) imaging modality has been a major focus of recent research in the medical imaging domain. Incorporating robust feature engineering approach can highly improve the performance of these models. Topological data analysis (TDA), a recent development based on the mathematical field of algebraic topology, mainly focuses on the data from a topological perspective, extracting deeper insight and higher dimensional structures from the data. Persistent homology (PH), a fundamental tool in the area of TDA, can extract topological features such as connected components, cycles and voids from the data. A popular approach to construct PH from 3D CT images is to utilize the 3D cubical complex filtration, a method adapted for grid-structured data. However, this approach may not always yield the best performance and can suffer from computational complexity with higher resolution CT images. This study introduces a novel patch-based PH construction approach tailored for volumetric medical imaging data, in particular CT modality. A wide range of experiments has been conducted on several datasets of 3D CT images to comprehensively analyze the performance of the proposed method with various parameters and benchmark it against the 3D cubical complex algorithm. Our results highlight the dominance of the patch-based TDA approach in terms of both classification performance and time-efficiency. The proposed approach outperformed the cubical complex method, achieving average improvement of 10.38%, 6.94%, 2.06%, 11.58%, and 8.51% in accuracy, AUC, sensitivity, specificity, and F1 score, respectively, across all datasets. Finally, we provide a convenient python package, Patch-TDA, to facilitate the utilization of the proposed approach.

Method: Created CRASAR-U-DRIODs dataset with post-disaster sUAS imagery from 10 disasters, labeled 657.25km of roads using 10-class schema, trained 18 baseline models, and deployed them operationally during Hurricanes Debby and Helene in 2024. Provided 9,184 road line adjustments for spatial alignment.

Result: When deployed against misaligned road lines, model performance degraded by 5.596% Macro IoU on average. Without spatial alignment, ~8% (11km) of adverse conditions would be mislabeled and ~9% (59km) of road lines would be misaligned off actual roads.

Conclusion: Spatial alignment is critical for effective disaster response ML systems. The ML, CV, and robotics communities need to address road line misalignment gaps to enable more informed decision-making during disasters.

Abstract: This paper presents the largest known benchmark dataset for road damage assessment and road alignment, and provides 18 baseline models trained on the CRASAR-U-DRIODs dataset’s post-disaster small uncrewed aerial systems (sUAS) imagery from 10 federally declared disasters, addressing three challenges within prior post-disaster road damage assessment datasets. While prior disaster road damage assessment datasets exist, there is no current state of practice, as prior public datasets have either been small-scale or reliant on low-resolution imagery insufficient for detecting phenomena of interest to emergency managers. Further, while machine learning (ML) systems have been developed for this task previously, none are known to have been operationally validated. These limitations are overcome in this work through the labeling of 657.25km of roads according to a 10-class labeling schema, followed by training and deploying ML models during the operational response to Hurricanes Debby and Helene in 2024. Motivated by observed road line misalignment in practice, 9,184 road line adjustments were provided for spatial alignment of a priori road lines, as it was found that when the 18 baseline models are deployed against real-world misaligned road lines, model performance degraded on average by 5.596% Macro IoU. If spatial alignment is not considered, approximately 8% (11km) of adverse conditions on road lines will be labeled incorrectly, with approximately 9% (59km) of road lines misaligned off the actual road. These dynamics are gaps that should be addressed by the ML, CV, and robotics communities to enable more effective and informed decision-making during disasters.

Method: Develop deep learning model that fuses SAR, passive microwave, and DEM data with regional climate model outputs for spatiotemporal downscaling to daily 100m resolution over Helheim Glacier (2017-2023).

Result: Deep learning fusion approach achieves 95% accuracy vs. 83% for RCM-only and 72% for PMW-only methods; SAR-based running window method achieves 90% accuracy but underestimates extremes.

Conclusion: Deep learning data fusion enables high-resolution meltwater mapping, published as MeltwaterBench dataset for benchmarking future downscaling methods to improve ice sheet monitoring.

Abstract: The Greenland ice sheet is melting at an accelerated rate due to processes that are not fully understood and hard to measure. The distribution of surface meltwater can help understand these processes and is observable through remote sensing, but current maps of meltwater face a trade-off: They are either high-resolution in time or space, but not both. We develop a deep learning model that creates gridded surface meltwater maps at daily 100m resolution by fusing data streams from remote sensing observations and physics-based models. In particular, we spatiotemporally downscale regional climate model (RCM) outputs using synthetic aperture radar (SAR), passive microwave (PMW), and a digital elevation model (DEM) over the Helheim Glacier in Eastern Greenland from 2017-2023. Using SAR-derived meltwater as “ground truth”, we show that a deep learning-based method that fuses all data streams is over 10 percentage points more accurate over our study area than existing non deep learning-based approaches that only rely on a regional climate model (83% vs. 95% Acc.) or passive microwave observations (72% vs. 95% Acc.). Alternatively, creating a gridded product through a running window calculation with SAR data underestimates extreme melt events, but also achieves notable accuracy (90%) and does not rely on deep learning. We evaluate standard deep learning methods (UNet and DeepLabv3+), and publish our spatiotemporally aligned dataset as a benchmark, MeltwaterBench, for intercomparisons with more complex data-driven downscaling methods. The code and data are available at $\href{https://github.com/blutjens/hrmelt}{github.com/blutjens/hrmelt}$.

Method: For OSR: compared ResNet-50, ConvNeXt-Tiny, and CLIP ViT-B/16 with linear probes and four scoring functions (MSP, Energy, Mahalanobis, kNN). For FSCIL: evaluated SPPR, OrCo, and ConCM methods with partially frozen ResNet-50 across 1-, 5-, and 10-shot scenarios.

Result: CLIP consistently performed best for OSR with Energy scoring being most stable. For FSCIL, ConCM achieved 84.7% accuracy in 10-shot setting with cleanest confusion matrix. All methods saturated beyond 5 shots.

Conclusion: Backbone architecture and scoring mechanisms significantly affect unknown detection, while prototype-based methods effectively mitigate catastrophic forgetting during incremental adaptation. CLIP with Energy scoring provides robust open-set recognition.

Abstract: Open-world deployment requires models to recognize both known categories and remain reliable when novel classes appear. We present a unified experimental study spanning open-set recognition (OSR) and few-shot class-incremental learning (FSCIL) on CIFAR-10. For OSR, we compare three pretrained frozen visual encoders: ResNet-50, ConvNeXt-Tiny and CLIP ViT-B/16,using a linear probe and four post-hoc scoring functions, namely MSP, Energy, Mahalanobis and kNN. Across metrics,such as, AUROC, AUPR, FPR@95, and OSCR, CLIP consistently yields the strongest separability between known and unknown samples, with Energy providing the most stable performance across backbones. For FSCIL, we compare modified SPPR, OrCo, and ConCM using partially frozen ResNet-50 across 1-, 5-, and 10-shot scenarios. ConCM achieves 84.7% accuracy in the 10-shot setting with the cleanest confusion matrix, while all methods show saturation beyond 5 shots. Our controlled evaluation reveals how the backbone architecture and scoring mechanisms affect unknown detection and how prototype-based methods mitigate catastrophic forgetting during incremental adaptation.

Method: Integrates direction-of-arrival (DOA) spectra and binauralized embeddings into a state-of-the-art vision-only pose estimation model.

Result: Consistent gains over visual baselines on two large datasets, plus robustness when visual information is corrupted.

Conclusion: First work to successfully leverage audio for relative camera pose estimation in real-world videos, establishing everyday audio as a promising signal for spatial challenges.

Abstract: Understanding camera motion is a fundamental problem in embodied perception and 3D scene understanding. While visual methods have advanced rapidly, they often struggle under visually degraded conditions such as motion blur or occlusions. In this work, we show that passive scene sounds provide complementary cues for relative camera pose estimation for in-the-wild videos. We introduce a simple but effective audio-visual framework that integrates direction-ofarrival (DOA) spectra and binauralized embeddings into a state-of-the-art vision-only pose estimation model. Our results on two large datasets show consistent gains over strong visual baselines, plus robustness when the visual information is corrupted. To our knowledge, this represents the first work to successfully leverage audio for relative camera pose estimation in real-world videos, and it establishes incidental, everyday audio as an unexpected but promising signal for a classic spatial challenge. Project: http://vision.cs.utexas.edu/projects/av_camera_pose.

Method: Three-stage approach: (1) Use self-supervised encoder for subject representations to guide subject alignment, (2) Use optical flow encoder for motion representations to capture object-level motion trajectories, (3) Apply subject-motion association decoupling with sparse LoRAs injection across locations and timing.

Result: Extensive experiments show SMRABooth excels in subject and motion customization, maintaining consistent subject appearance and motion patterns, proving effectiveness in controllable text-to-video generation.

Conclusion: SMRABooth successfully addresses the challenge of customized video generation by providing object-level guidance for both subject and motion through specialized encoders and a decoupling strategy, enabling faithful preservation of appearance and motion consistency.

Abstract: Customized video generation aims to produce videos that faithfully preserve the subject’s appearance from reference images while maintaining temporally consistent motion from reference videos. Existing methods struggle to ensure both subject appearance similarity and motion pattern consistency due to the lack of object-level guidance for subject and motion. To address this, we propose SMRABooth, which leverages the self-supervised encoder and optical flow encoder to provide object-level subject and motion representations. These representations are aligned with the model during the LoRA fine-tuning process. Our approach is structured in three core stages: (1) We exploit subject representations via a self-supervised encoder to guide subject alignment, enabling the model to capture overall structure of subject and enhance high-level semantic consistency. (2) We utilize motion representations from an optical flow encoder to capture structurally coherent and object-level motion trajectories independent of appearance. (3) We propose a subject-motion association decoupling strategy that applies sparse LoRAs injection across both locations and timing, effectively reducing interference between subject and motion LoRAs. Extensive experiments show that SMRABooth excels in subject and motion customization, maintaining consistent subject appearance and motion patterns, proving its effectiveness in controllable text-to-video generation.

Method: Thermal images generate hot region masks via adaptive thresholding/morphological refinement; RGB frames provide edge cues and suppress false detections via gradient filtering. Rule-level merging selects boundary candidates simplified with Ramer-Douglas-Peucker algorithm. System includes periodic beacons and inertial feedback for GPS degradation resilience. Optimized for embedded SoC platforms with constrained pixel operations and precomputed gradient tables.

Result: Small-scale simulations show reduced average path length and boundary jitter compared to pure edge tracking baseline, while maintaining environmental coverage. Battery consumption and computational utilization confirm feasibility of 10-15 m/s forward motion on standard micro platforms.

Conclusion: The lightweight approach enables rapid field deployment for emergency wildfire reconnaissance with robust sensing and minimal communications, achieving sub-50ms latency on embedded platforms.

Abstract: This study introduces a lightweight perimeter tracking method designed for micro UAV teams operating over wildfire environments under limited bandwidth conditions. Thermal image frames generate coarse hot region masks through adaptive thresholding and morphological refinement, while RGB frames contribute edge cues and suppress texture related false detections using gradient based filtering. A rule level merging strategy selects boundary candidates and simplifies them via the Ramer Douglas Peucker algorithm. The system incorporates periodic beacons and an inertial feedback loop that maintains trajectory stability in the presence of GPS degradation. The guidance loop targets sub 50 ms latency on embedded System on Chip (SoC) platforms by constraining per frame pixel operations and precomputing gradient tables. Small scale simulations demonstrate reductions in average path length and boundary jitter compared to a pure edge tracking baseline, while maintaining environmental coverage measured through intersection merge analysis. Battery consumption and computational utilization confirm the feasibility of achieving 10, 15 m/s forward motion on standard micro platforms. This approach enables rapid deployment in the field, requiring robust sensing and minimal communications for emergency reconnaissance applications.

Method: Deployed 22 fixed-angle cameras across Bristol to collect hourly images over 4 years. Created a self-supervised framework using convolutional variational autoencoders (CNN-VAEs) trained per camera node and day/night sets. Defined two drift metrics: relative centroid drift (latent space deviation) and relative reconstruction error (image-domain degradation).

Result: Produced a comprehensive dataset of 526,000+ images with rich metadata (timestamps, GPS, device IDs) under diverse conditions. The framework enables quantitative analysis of visual drift and provides benchmarks for evaluating long-term model stability and drift-aware learning in urban vision systems.

Conclusion: This dataset offers a realistic benchmark for studying visual drift and model degradation in smart city deployments, supporting reproducibility and applications in streetlight monitoring, weather inference, and urban scene understanding. Publicly available data and framework promote secondary analysis.

Abstract: We present a large-scale, longitudinal visual dataset of urban streetlights captured by 22 fixed-angle cameras deployed across Bristol, U.K., from 2021 to 2025. The dataset contains over 526,000 images, collected hourly under diverse lighting, weather, and seasonal conditions. Each image is accompanied by rich metadata, including timestamps, GPS coordinates, and device identifiers. This unique real-world dataset enables detailed investigation of visual drift, anomaly detection, and MLOps strategies in smart city deployments. To promtoe seconardary analysis, we additionally provide a self-supervised framework based on convolutional variational autoencoders (CNN-VAEs). Models are trained separately for each camera node and for day/night image sets. We define two per-sample drift metrics: relative centroid drift, capturing latent space deviation from a baseline quarter, and relative reconstruction error, measuring normalized image-domain degradation. This dataset provides a realistic, fine-grained benchmark for evaluating long-term model stability, drift-aware learning, and deployment-ready vision systems. The images and structured metadata are publicly released in JPEG and CSV formats, supporting reproducibility and downstream applications such as streetlight monitoring, weather inference, and urban scene understanding. The dataset can be found at https://doi.org/10.5281/zenodo.17781192 and https://doi.org/10.5281/zenodo.17859120.

Method: Proposes ISA-ViT framework with: 1) Input-size-agnostic design that resizes UWB data while preserving radar-specific information (Doppler shifts, phase characteristics), 2) Patch configuration adjustment and pre-trained positional embedding vectors, 3) Domain fusion strategy combining range- and frequency-domain features.

Result: ISA-ViT achieves 22.68% accuracy improvement over existing ViT-based approach for UWB-based DAR. The ALERT dataset contains 10,220 radar samples of seven distracted driving activities collected in real driving conditions.

Conclusion: This work facilitates robust distracted driving detection by publicly releasing ALERT dataset and detailing input-size-agnostic strategy, enabling more scalable real-world deployment of radar-based DAR systems.

Abstract: Distracted driving contributes to fatal crashes worldwide. To address this, researchers are using driver activity recognition (DAR) with impulse radio ultra-wideband (IR-UWB) radar, which offers advantages such as interference resistance, low power consumption, and privacy preservation. However, two challenges limit its adoption: the lack of large-scale real-world UWB datasets covering diverse distracted driving behaviors, and the difficulty of adapting fixed-input Vision Transformers (ViTs) to UWB radar data with non-standard dimensions. This work addresses both challenges. We present the ALERT dataset, which contains 10,220 radar samples of seven distracted driving activities collected in real driving conditions. We also propose the input-size-agnostic Vision Transformer (ISA-ViT), a framework designed for radar-based DAR. The proposed method resizes UWB data to meet ViT input requirements while preserving radar-specific information such as Doppler shifts and phase characteristics. By adjusting patch configurations and leveraging pre-trained positional embedding vectors (PEVs), ISA-ViT overcomes the limitations of naive resizing approaches. In addition, a domain fusion strategy combines range- and frequency-domain features to further improve classification performance. Comprehensive experiments demonstrate that ISA-ViT achieves a 22.68% accuracy improvement over an existing ViT-based approach for UWB-based DAR. By publicly releasing the ALERT dataset and detailing our input-size-agnostic strategy, this work facilitates the development of more robust and scalable distracted driving detection systems for real-world deployment.

Method: Hybrid model combining fine-tuned ResNet-50 CNN and three-layer GCN trained on visual and geometric features from MediaPipe FaceMesh landmarks. Emotions were probabilistically labeled using weighted ensemble of DeepFace and FER models across seven emotion classes. Final classification used fused embedding optimized via Kullback-Leibler divergence.

Result: The method demonstrates robust performance in modeling subtle affective responses, effectively capturing micro emotional cues in neurodivergent children. It represents the first large-scale, real-world dataset and pipeline from India on autism-focused emotion analysis using social robotics.

Conclusion: The pipeline offers significant promise for affective profiling of ASD children in clinical and therapeutic human-robot interaction contexts, contributing an essential foundation for future personalized assistive technologies.

Abstract: Understanding emotional responses in children with Autism Spectrum Disorder (ASD) during social interaction remains a critical challenge in both developmental psychology and human-robot interaction. This study presents a novel deep learning pipeline for emotion recognition in autistic children in response to a name-calling event by a humanoid robot (NAO), under controlled experimental settings. The dataset comprises of around 50,000 facial frames extracted from video recordings of 15 children with ASD. A hybrid model combining a fine-tuned ResNet-50-based Convolutional Neural Network (CNN) and a three-layer Graph Convolutional Network (GCN) trained on both visual and geometric features extracted from MediaPipe FaceMesh landmarks. Emotions were probabilistically labeled using a weighted ensemble of two models: DeepFace’s and FER, each contributing to soft-label generation across seven emotion classes. Final classification leveraged a fused embedding optimized via Kullback-Leibler divergence. The proposed method demonstrates robust performance in modeling subtle affective responses and offers significant promise for affective profiling of ASD children in clinical and therapeutic human-robot interaction contexts, as the pipeline effectively captures micro emotional cues in neurodivergent children, addressing a major gap in autism-specific HRI research. This work represents the first such large-scale, real-world dataset and pipeline from India on autism-focused emotion analysis using social robotics, contributing an essential foundation for future personalized assistive technologies.

Method: 1) Created CineLOG dataset with 5,000 balanced, high-quality video clips annotated with scene descriptions, camera instructions (17 movements), and genre labels (15 genres). 2) Developed a pipeline that decouples text-to-video generation into four easier stages using mature technology. 3) Introduced Trajectory Guided Transition Module for smooth spatio-temporal interpolation in multi-shot sequences.

Result: Extensive human evaluations show the pipeline significantly outperforms state-of-the-art end-to-end text-to-video models in adhering to specific camera and screenplay instructions while maintaining professional visual quality.

Conclusion: CineLOG addresses limitations in controllable video synthesis by providing a balanced, high-quality dataset and a novel pipeline that enables fine-grained cinematic control, advancing the field beyond textual prompt-based generation.

Abstract: Controllable video synthesis is a central challenge in computer vision, yet current models struggle with fine grained control beyond textual prompts, particularly for cinematic attributes like camera trajectory and genre. Existing datasets often suffer from severe data imbalance, noisy labels, or a significant simulation to real gap. To address this, we introduce CineLOG, a new dataset of 5,000 high quality, balanced, and uncut video clips. Each entry is annotated with a detailed scene description, explicit camera instructions based on a standard cinematic taxonomy, and genre label, ensuring balanced coverage across 17 diverse camera movements and 15 film genres. We also present our novel pipeline designed to create this dataset, which decouples the complex text to video (T2V) generation task into four easier stages with more mature technology. To enable coherent, multi shot sequences, we introduce a novel Trajectory Guided Transition Module that generates smooth spatio-temporal interpolation. Extensive human evaluations show that our pipeline significantly outperforms SOTA end to end T2V models in adhering to specific camera and screenplay instructions, while maintaining professional visual quality. All codes and data are available at https://cine-log.pages.dev.

Method: Proposes a training- and reference-free framework that: 1) decomposes reasoning chains into perception vs. reasoning steps, 2) uses off-the-shelf VLM judges for step-level faithfulness evaluation, 3) includes human meta-evaluation for verification, and 4) implements a lightweight self-reflection procedure that detects and regenerates unfaithful perception steps without training.

Result: The method reduces Unfaithful Perception Rate while preserving final-answer accuracy across multiple reasoning-trained VLMs and perception-heavy benchmarks, improving multimodal reasoning reliability.

Conclusion: Visual faithfulness of reasoning chains is a distinct evaluation dimension crucial for reliable multimodal reasoning. The proposed framework enables evaluation and improvement of this aspect without training, enhancing VLM reliability through self-reflection on perception steps.

Abstract: Reasoning-augmented vision language models (VLMs) generate explicit chains of thought that promise greater capability and transparency but also introduce new failure modes: models may reach correct answers via visually unfaithful intermediate steps, or reason faithfully yet fail on the final prediction. Standard evaluations that only measure final-answer accuracy cannot distinguish these behaviors. We introduce the visual faithfulness of reasoning chains as a distinct evaluation dimension, focusing on whether the perception steps of a reasoning chain are grounded in the image. We propose a training- and reference-free framework that decomposes chains into perception versus reasoning steps and uses off-the-shelf VLM judges for step-level faithfulness, additionally verifying this approach through a human meta-evaluation. Building on this metric, we present a lightweight self-reflection procedure that detects and locally regenerates unfaithful perception steps without any training. Across multiple reasoning-trained VLMs and perception-heavy benchmarks, our method reduces Unfaithful Perception Rate while preserving final-answer accuracy, improving the reliability of multimodal reasoning.

Method: AttributeCentric Representations (ACR) framework with two components: Mixture of Patch Experts (MoPE) uses dual-level routing to dispatch image patches to specialized experts, and Mixture of Attribute Experts (MoAE) projects features into sparse, part-aware attribute maps for classification.

Result: Achieves consistent state-of-the-art results on zero-shot learning benchmark datasets CUB, AwA2, and SUN.

Conclusion: ACR effectively tackles attribute entanglement by imposing disentanglement during representation learning, enabling better transfer of visual-attribute relationships from seen to unseen fine-grained categories.

Abstract: Recognizing unseen fine-grained categories demands a model that can distinguish subtle visual differences. This is typically achieved by transferring visual-attribute relationships from seen classes to unseen classes. The core challenge is attribute entanglement, where conventional models collapse distinct attributes like color, shape, and texture into a single visual embedding. This causes interference that masks these critical distinctions. The post-hoc solutions of previous work are insufficient, as they operate on representations that are already mixed. We propose a zero-shot learning framework that learns AttributeCentric Representations (ACR) to tackle this problem by imposing attribute disentanglement during representation learning. ACR is achieved with two mixture-of-experts components, including Mixture of Patch Experts (MoPE) and Mixture of Attribute Experts (MoAE). First, MoPE is inserted into the transformer using a dual-level routing mechanism to conditionally dispatch image patches to specialized experts. This ensures coherent attribute families are processed by dedicated experts. Finally, the MoAE head projects these expert-refined features into sparse, partaware attribute maps for robust zero-shot classification. On zero-shot learning benchmark datasets CUB, AwA2, and SUN, our ACR achieves consistent state-of-the-art results.

Method: Created ProImage-Bench with 654 real textbook/technical report figures, detailed image instructions, and hierarchical rubrics (6,076 criteria, 44,131 binary checks). Used LMMs to derive rubrics from surrounding text and reference figures, with automated LMM-based evaluation using principled penalty scheme.

Result: Best base model achieved only 0.791 rubric accuracy and 0.553 criterion score, showing substantial gaps in scientific fidelity. Iterative refinement using failed checks boosted a strong generator from 0.653 to 0.865 rubric accuracy and 0.388 to 0.697 criterion score.

Conclusion: ProImage-Bench provides both rigorous diagnostic evaluation for professional image generation and scalable supervision signal for improving specification-faithful scientific illustrations through iterative refinement.

Abstract: We study professional image generation, where a model must synthesize information-dense, scientifically precise illustrations from technical descriptions rather than merely produce visually plausible pictures. To quantify the progress, we introduce ProImage-Bench, a rubric-based benchmark that targets biology schematics, engineering/patent drawings, and general scientific diagrams. For 654 figures collected from real textbooks and technical reports, we construct detailed image instructions and a hierarchy of rubrics that decompose correctness into 6,076 criteria and 44,131 binary checks. Rubrics are derived from surrounding text and reference figures using large multimodal models, and are evaluated by an automated LMM-based judge with a principled penalty scheme that aggregates sub-question outcomes into interpretable criterion scores. We benchmark several representative text-to-image models on ProImage-Bench and find that, despite strong open-domain performance, the best base model reaches only 0.791 rubric accuracy and 0.553 criterion score overall, revealing substantial gaps in fine-grained scientific fidelity. Finally, we show that the same rubrics provide actionable supervision: feeding failed checks back into an editing model for iterative refinement boosts a strong generator from 0.653 to 0.865 in rubric accuracy and from 0.388 to 0.697 in criterion score. ProImage-Bench thus offers both a rigorous diagnostic for professional image generation and a scalable signal for improving specification-faithful scientific illustrations.

Method: Systematic comparison of SynthSeg and SamSeg segmentation methods using the Baby Open Brains dataset (71 scans, 1-9 months). Evaluation against expert annotations using Dice, IoU, Hausdorff distance, and Normalised Mutual Information metrics, plus analysis of volumetric and fractal dimension estimates.

Result: SynthSeg outperformed SamSeg across all quality metrics (mean Dice > 0.8 for major regions) and provided more accurate volumetric estimates. SamSeg systematically overestimated ventricular and whole-brain volumes. Segmentation accuracy improved with age. Fractal dimension analyses showed significant regional differences, and segmentation-related FD variability exceeded most group differences reported in developmental cohorts.

Conclusion: SynthSeg provides the most reliable volumetric and FD results for paediatric MRI, but small morphological differences in volume and FD should be interpreted cautiously due to segmentation-related uncertainty, as segmentation bias directly affects FD estimation.

Abstract: Accurate segmentation of infant brain MRI is essential for quantifying developmental changes in structure and complexity. However, ongoing myelination and reduced tissue contrast make automated segmentation particularly challenging. This study systematically compared segmentation accuracy and its impact on volumetric and fractal dimension (FD) estimates in infant brain MRI using the Baby Open Brains (BOB) dataset (71 scans, 1-9 months). Two methods, SynthSeg and SamSeg, were evaluated against expert annotations using Dice, Intersection over Union, 95th-percentile Hausdorff distance, and Normalised Mutual Information. SynthSeg outperformed SamSeg across all quality metrics (mean Dice > 0.8 for major regions) and provided volumetric estimates closely matching the manual reference (mean +4% [-28% - 71%]). SamSeg systematically overestimated ventricular and whole-brain volumes (mean +76% [-12% - 190%]). Segmentation accuracy improved with age, consistent with increasing tissue contrast during myelination. Fractal dimension a(FD) nalyses revealed significant regional differences between SynthSeg and expert segmentations, and Bland-Altman limits of agreement indicated that segmentation-related FD variability exceeded most group differences reported in developmental cohorts. Volume and FD deviations were positively correlated across structures, indicating that segmentation bias directly affects FD estimation. Overall, SynthSeg provided the most reliable volumetric and FD results for paediatric MRI, yet small morphological differences in volume and FD should be interpreted with caution due to segmentation-related uncertainty.

Method: AEIC uses moderate/shallow encoder networks paired with a one-step diffusion decoder. It features a dual-side feature distillation scheme that transfers knowledge from moderate encoder variants to shallow encoder versions to enhance efficiency. The asymmetric design prioritizes encoding simplicity while maintaining decoding quality.

Result: AEIC outperforms existing methods on rate-distortion-perception performance at ultra-low bitrates (below 0.05 bpp). It achieves exceptional encoding efficiency of 35.8 FPS on 1080P images while maintaining competitive decoding speed compared to existing methods.

Conclusion: The proposed asymmetric framework successfully addresses the deployment limitations of existing ultra-low bitrate compression methods by enabling efficient encoding on edge devices while maintaining high reconstruction quality through diffusion-based decoding.

Abstract: Ultra-low bitrate image compression (below 0.05 bits per pixel) is increasingly critical for bandwidth-constrained and computation-limited encoding scenarios such as edge devices. Existing frameworks typically rely on large pretrained encoders (e.g., VAEs or tokenizer-based models) and perform transform coding within their generative latent space. While these approaches achieve impressive perceptual fidelity, their reliance on heavy encoder networks makes them unsuitable for deployment on weak sender devices. In this work, we explore the feasibility of applying shallow encoders for ultra-low bitrate compression and propose a novel Asymmetric Extreme Image Compression (AEIC) framework that pursues simultaneously encoding simplicity and decoding quality. Specifically, AEIC employs moderate or even shallow encoder networks, while leveraging an one-step diffusion decoder to maintain high-fidelity and high-realism reconstructions under extreme bitrates. To further enhance the efficiency of shallow encoders, we design a dual-side feature distillation scheme that transfers knowledge from AEIC with moderate encoders to its shallow encoder variants. Experiments demonstrate that AEIC not only outperforms existing methods on rate-distortion-perception performance at ultra-low bitrates, but also delivers exceptional encoding efficiency for 35.8 FPS on 1080P input images, while maintaining competitive decoding speed compared to existing methods.

Method: Feed LLM with fixed number of frames and prompt to output “0”/“1” character sequence (one per frame). “0” represents background, “1” foreground probabilities. Train with segmentation losses on probabilities plus causal LM loss. At inference, use beam search to generate sequence (moments) and logits (saliency scores).

Result: Achieved strong highlight detection (56.74 HIT@1) on QVHighlights using only 25 frames (less than half of comparable methods). Also scored above baseline (35.28 MAP) for moment retrieval. Segmentation losses provide stable complementary learning signal even when causal LM loss plateaus.

Conclusion: The proposed method effectively leverages LLMs’ language capabilities for frame-level video analysis through segmentation objectives on output tokens, achieving competitive performance with greater efficiency and providing stable training signals.

Abstract: Detecting video moments and highlights from natural-language queries have been unified by transformer-based methods. Other works use generative Multimodal LLM (MLLM) to predict moments and/or highlights as text timestamps, utilizing its reasoning capability. While effective, text-based generation cannot provide direct gradients for frame-level predictions because the model only emits language tokens. Although recent Reinforcement Learning (RL) methods attempt to address the issue, we propose a novel approach by applying segmentation objectives directly on the LLM’s output tokens. The LLM is fed with a fixed number of frames alongside a prompt that enforces it to output a sequence of continuous “0” and/or “1” characters, with one character per frame. The “0”/“1” characters benefit from the LLM’s inherent language capability while also acting as background and foreground probabilities, respectively. Training employs segmentation losses on the probabilities alongside a normal causal LM loss. At inference, beam search generates sequence and logits, acting as moments and saliency scores, respectively. Despite sampling only 25 frames – less than half of comparable methods – our method achieved strong highlight detection (56.74 HIT@1) on QVHighlights. Additionally, our efficient method scores above the baseline (35.28 MAP) for moment retrieval. Empirically, segmentation losses provide a stable complementary learning signal even when the causal LM loss plateaus.

Method: MetaTPT uses a meta-learning framework with dual-loop optimization: inner loop learns a self-supervised auxiliary task that generates parameterized augmentations for each sample, while outer loop performs prompt tuning by enforcing consistency across these dynamically generated views.

Result: Extensive experiments show MetaTPT achieves state-of-the-art performance on domain generalization and cross-dataset benchmarks, demonstrating improved test-time adaptation under domain shifts.

Conclusion: By coupling augmentation learning with prompt tuning through meta-learning, MetaTPT enables more expressive transformations that capture essential features in target domains, significantly improving vision-language model robustness to domain shifts.

Abstract: Vision-language models (VLMs) such as CLIP exhibit strong zero-shot generalization but remain sensitive to domain shifts at test time. Test-time prompt tuning (TPT) mitigates this issue by adapting prompts with fixed augmentations, which may falter in more challenging settings. In this work, we propose Meta Test-Time Prompt Tuning (MetaTPT), a meta-learning framework that learns a self-supervised auxiliary task to guide test-time prompt tuning. The auxiliary task dynamically learns parameterized augmentations for each sample, enabling more expressive transformations that capture essential features in target domains. MetaTPT adopts a dual-loop optimization paradigm: an inner loop learns a self-supervised task that generates informative views, while the outer loop performs prompt tuning by enforcing consistency across these views. By coupling augmentation learning with prompt tuning, MetaTPT improves test-time adaptation under domain shifts. Extensive experiments demonstrate that MetaTPT achieves state-of-the-art performance on domain generalization and cross-dataset benchmarks.

Method: A hybrid framework integrating two complementary modalities: deep convolutional features and facial Action Units (AUs) from the Facial Action Coding System. The combined representation is modeled through Bayesian Gaussian Mixture Models (BGMMs), providing a lightweight, probabilistic solution that avoids retraining while offering strong discriminative power.

Result: Using the Compound Facial Expression of Emotion (CFEE) dataset, the model can first learn basic expressions and then progressively recognize compound expressions. Experiments demonstrate improved accuracy, stronger knowledge retention, and reduced forgetting compared to existing approaches.

Conclusion: This framework contributes to the development of emotionally intelligent AI systems with applications in education, healthcare, and adaptive user interfaces by enabling continuous learning of facial expressions without catastrophic forgetting.

Abstract: As artificial intelligence (AI) systems become increasingly embedded in our daily life, the ability to recognize and adapt to human emotions is essential for effective human-computer interaction. Facial expression recognition (FER) provides a primary channel for inferring affective states, but the dynamic and culturally nuanced nature of emotions requires models that can learn continuously without forgetting prior knowledge. In this work, we propose a hybrid framework for FER in a continual learning setting that mitigates catastrophic forgetting. Our approach integrates two complementary modalities: deep convolutional features and facial Action Units (AUs) derived from the Facial Action Coding System (FACS). The combined representation is modelled through Bayesian Gaussian Mixture Models (BGMMs), which provide a lightweight, probabilistic solution that avoids retraining while offering strong discriminative power. Using the Compound Facial Expression of Emotion (CFEE) dataset, we show that our model can first learn basic expressions and then progressively recognize compound expressions. Experiments demonstrate improved accuracy, stronger knowledge retention, and reduced forgetting. This framework contributes to the development of emotionally intelligent AI systems with applications in education, healthcare, and adaptive user interfaces.

Method: Three-stage framework: 1) Extract dataset meta-features (object scale distribution, scene density), 2) LLM reasons on features + RAG-retrieved SOTA components to synthesize architecture in NADL, 3) Compiler instantiates deployable model.

Result: Superior architectures consistently generated across five diverse object detection datasets, achieving competitive performance with better performance-per-parameter trade-off than strong baselines.

Conclusion: LLM’s data-driven reasoning is primary performance driver; deep understanding of data “first principles” is more critical than simply retrieving SOTA components for superior architecture design.

Abstract: Designing high-performance object detection architectures is a complex task, where traditional manual design is time-consuming and labor-intensive, and Neural Architecture Search (NAS) is computationally prohibitive. While recent approaches using Large Language Models (LLMs) show promise, they often function as iterative optimizers within a search loop, rather than generating architectures directly from a holistic understanding of the data. To address this gap, we propose Cognitive-YOLO, a novel framework for LLM-driven architecture synthesis that generates network configurations directly from the intrinsic characteristics of the dataset. Our method consists of three stages: first, an analysis module extracts key meta-features (e.g., object scale distribution and scene density) from the target dataset; second, the LLM reasons upon these features, augmented with state-of-the-art components retrieved via Retrieval-Augmented Generation (RAG), to synthesize the architecture into a structured Neural Architecture Description Language (NADL); finally, a compiler instantiates this description into a deployable model. Extensive experiments on five diverse object detection datasets demonstrate that our proposed Cognitive-YOLO consistently generates superior architectures, achieving highly competitive performance and demonstrating a superior performance-per-parameter trade-off compared to strong baseline models across multiple benchmarks. Crucially, our ablation studies prove that the LLM’s data-driven reasoning is the primary driver of performance, demonstrating that a deep understanding of data “first principles” is more critical for achieving a superior architecture than simply retrieving SOTA components.

Method: Created RealDrag benchmark with over 400 human-annotated samples from diverse video sources, including source/target images, handle/target points, editable region masks, and descriptive captions. Proposed four novel metrics: Semantical Distance (SeD), Outer Mask Preserving Score (OMPS), Inner Patch Preserving Score (IPPS), and Directional Similarity (DiS).

Result: Conducted first large-scale systematic analysis evaluating 17 state-of-the-art models, revealing clear trade-offs among current approaches and establishing a robust, reproducible baseline for future research.

Conclusion: RealDrag provides the first comprehensive benchmark with ground truth data for point-based image editing, enabling reliable evaluation and comparison of models through standardized metrics and dataset.

Abstract: The evaluation of drag based image editing models is unreliable due to a lack of standardized benchmarks and metrics. This ambiguity stems from inconsistent evaluation protocols and, critically, the absence of datasets containing ground truth target images, making objective comparisons between competing methods difficult. To address this, we introduce \textbf{RealDrag}, the first comprehensive benchmark for point based image editing that includes paired ground truth target images. Our dataset contains over 400 human annotated samples from diverse video sources, providing source/target images, handle/target points, editable region masks, and descriptive captions for both the image and the editing action. We also propose four novel, task specific metrics: Semantical Distance (SeD), Outer Mask Preserving Score (OMPS), Inner Patch Preserving Score (IPPS), and Directional Similarity (DiS). These metrics are designed to quantify pixel level matching fidelity, check preservation of non edited (out of mask) regions, and measure semantic alignment with the desired task. Using this benchmark, we conduct the first large scale systematic analysis of the field, evaluating 17 SOTA models. Our results reveal clear trade offs among current approaches and establish a robust, reproducible baseline to guide future research. Our dataset and evaluation toolkit will be made publicly available.

Method: Progressive training framework that starts with training small subnets and gradually incorporates larger ones. Also introduces GrowTAS+ that fine-tunes only a subset of weights to further enhance large subnet performance.

Result: Extensive experiments on ImageNet and transfer learning benchmarks (CIFAR-10/100, Flowers, CARS, INAT-19) demonstrate effectiveness over current TAS methods.

Conclusion: The progressive training approach reduces interference and stabilizes training, enabling better performance for discovered vision transformers compared to existing TAS methods.

Abstract: Transformer architecture search (TAS) aims to automatically discover efficient vision transformers (ViTs), reducing the need for manual design. Existing TAS methods typically train an over-parameterized network (i.e., a supernet) that encompasses all candidate architectures (i.e., subnets). However, all subnets share the same set of weights, which leads to interference that degrades the smaller subnets severely. We have found that well-trained small subnets can serve as a good foundation for training larger ones. Motivated by this, we propose a progressive training framework, dubbed GrowTAS, that begins with training small subnets and incorporate larger ones gradually. This enables reducing the interference and stabilizing a training process. We also introduce GrowTAS+ that fine-tunes a subset of weights only to further enhance the performance of large subnets. Extensive experiments on ImageNet and several transfer learning benchmarks, including CIFAR-10/100, Flowers, CARS, and INAT-19, demonstrate the effectiveness of our approach over current TAS methods

Method: The authors introduce Intention-Drive benchmark with two core components: (1) a new dataset of complex driving scenarios paired with natural language intentions, and (2) a novel evaluation protocol using Intent Success Rate (ISR) that assesses semantic fulfillment of human goals beyond geometric accuracy.

Result: Extensive evaluation of baseline models on Intention-Drive reveals significant performance deficits, showing current models struggle to achieve the comprehensive scene and intention understanding required for high-level intention fulfillment.

Conclusion: The paper introduces a crucial benchmark to drive progress toward truly intelligent autonomous driving systems that can understand and fulfill human intentions, highlighting the need for models that move beyond simple command-following to intention-fulfillment.

Abstract: Current end-to-end autonomous driving systems operate at a level of intelligence akin to following simple steering commands. However, achieving genuinely intelligent autonomy requires a paradigm shift: moving from merely executing low-level instructions to understanding and fulfilling high-level, abstract human intentions. This leap from a command-follower to an intention-fulfiller, as illustrated in our conceptual framework, is hindered by a fundamental challenge: the absence of a standardized benchmark to measure and drive progress on this complex task. To address this critical gap, we introduce Intention-Drive, the first comprehensive benchmark designed to evaluate the ability to translate high-level human intent into safe and precise driving actions. Intention-Drive features two core contributions: (1) a new dataset of complex scenarios paired with corresponding natural language intentions, and (2) a novel evaluation protocol centered on the Intent Success Rate (ISR), which assesses the semantic fulfillment of the human’s goal beyond simple geometric accuracy. Through an extensive evaluation of a spectrum of baseline models on Intention-Drive, we reveal a significant performance deficit, showing that the baseline model struggle to achieve the comprehensive scene and intention understanding required for this advanced task.

Method: Proposes OMUDA with three hierarchical masking strategies: 1) Context-Aware Masking (CAM) for foreground-background distinction, 2) Feature Distillation Masking (FDM) for robust feature learning via knowledge transfer, and 3) Class Decoupling Masking (CDM) for mitigating noisy pseudo-labels through class-wise uncertainty modeling.

Result: Extensive experiments on SYNTHIA->Cityscapes and GTA5->Cityscapes benchmarks show state-of-the-art performance with average 7% improvement, and OMUDA can be seamlessly integrated into existing UDA methods.

Conclusion: OMUDA provides a unified hierarchical masking solution that effectively reduces domain shift at contextual, representational, and categorical levels, outperforming existing approaches in cross-domain semantic segmentation.

Abstract: Unsupervised domain adaptation (UDA) enables semantic segmentation models to generalize from a labeled source domain to an unlabeled target domain. However, existing UDA methods still struggle to bridge the domain gap due to cross-domain contextual ambiguity, inconsistent feature representations, and class-wise pseudo-label noise. To address these challenges, we propose Omni-level Masking for Unsupervised Domain Adaptation (OMUDA), a unified framework that introduces hierarchical masking strategies across distinct representation levels. Specifically, OMUDA comprises: 1) a Context-Aware Masking (CAM) strategy that adaptively distinguishes foreground from background to balance global context and local details; 2) a Feature Distillation Masking (FDM) strategy that enhances robust and consistent feature learning through knowledge transfer from pre-trained models; and 3) a Class Decoupling Masking (CDM) strategy that mitigates the impact of noisy pseudo-labels by explicitly modeling class-wise uncertainty. This hierarchical masking paradigm effectively reduces the domain shift at the contextual, representational, and categorical levels, providing a unified solution beyond existing approaches. Extensive experiments on multiple challenging cross-domain semantic segmentation benchmarks validate the effectiveness of OMUDA. Notably, on the SYNTHIA->Cityscapes and GTA5->Cityscapes tasks, OMUDA can be seamlessly integrated into existing UDA methods and consistently achieving state-of-the-art results with an average improvement of 7%.

Method: MRD (metamers rendered differentiably) uses physically based differentiable rendering to find 3D scene parameters that produce identical model activations but are physically different (model metamers). This allows systematic probing of sensitivity to specific scene attributes like shape, material, and lighting.

Result: The approach successfully finds 3D scenes with high similarity in model activation to target scenes, with varying visual results. It enables qualitative investigation of which physical scene attributes models are sensitive or invariant to.

Conclusion: MRD provides a physically grounded framework for analyzing vision models’ implicit 3D understanding, promising to advance interpretability research in both computer and human vision by revealing how physical scene parameters drive model responses.

Abstract: While deep learning methods have achieved impressive success in many vision benchmarks, it remains difficult to understand and explain the representations and decisions of these models. Though vision models are typically trained on 2D inputs, they are often assumed to develop an implicit representation of the underlying 3D scene (for example, showing tolerance to partial occlusion, or the ability to reason about relative depth). Here, we introduce MRD (metamers rendered differentiably), an approach that uses physically based differentiable rendering to probe vision models’ implicit understanding of generative 3D scene properties, by finding 3D scene parameters that are physically different but produce the same model activation (i.e. are model metamers). Unlike previous pixel-based methods for evaluating model representations, these reconstruction results are always grounded in physical scene descriptions. This means we can, for example, probe a model’s sensitivity to object shape while holding material and lighting constant. As a proof-of-principle, we assess multiple models in their ability to recover scene parameters of geometry (shape) and bidirectional reflectance distribution function (material). The results show high similarity in model activation between target and optimized scenes, with varying visual results. Qualitatively, these reconstructions help investigate the physical scene attributes to which models are sensitive or invariant. MRD holds promise for advancing our understanding of both computer and human vision by enabling analysis of how physical scene parameters drive changes in model responses.

Method: Uses a dual-tower architecture without cross-modal fusion layers, treating recognition as a retrieval problem. Introduces three variants: WeDetect (main detector), WeDetect-Uni (universal proposal generator with frozen detector), and WeDetect-Ref (LMM-based classifier for complex referring expressions).

Result: Achieves state-of-the-art performance across 15 benchmarks with real-time inference. Enables fast backtracking of historical data through object retrieval and handles complex referring expressions efficiently with single forward pass classification.

Conclusion: The WeDetect family successfully unifies detection, proposal generation, object retrieval, and referring expression comprehension under a coherent retrieval framework, demonstrating the advantages of non-fusion approaches for open-vocabulary object detection.

Abstract: Open-vocabulary object detection aims to detect arbitrary classes via text prompts. Methods without cross-modal fusion layers (non-fusion) offer faster inference by treating recognition as a retrieval problem, \ie, matching regions to text queries in a shared embedding space. In this work, we fully explore this retrieval philosophy and demonstrate its unique advantages in efficiency and versatility through a model family named WeDetect: (1) State-of-the-art performance. WeDetect is a real-time detector with a dual-tower architecture. We show that, with well-curated data and full training, the non-fusion WeDetect surpasses other fusion models and establishes a strong open-vocabulary foundation. (2) Fast backtrack of historical data. WeDetect-Uni is a universal proposal generator based on WeDetect. We freeze the entire detector and only finetune an objectness prompt to retrieve generic object proposals across categories. Importantly, the proposal embeddings are class-specific and enable a new application, object retrieval, supporting retrieval objects in historical data. (3) Integration with LMMs for referring expression comprehension (REC). We further propose WeDetect-Ref, an LMM-based object classifier to handle complex referring expressions, which retrieves target objects from the proposal list extracted by WeDetect-Uni. It discards next-token prediction and classifies objects in a single forward pass. Together, the WeDetect family unifies detection, proposal generation, object retrieval, and REC under a coherent retrieval framework, achieving state-of-the-art performance across 15 benchmarks with high inference efficiency.

Method: UniCoDe integrates local gradient signals during sampling, creating a unified framework that brings together the strengths of both sampling-based and gradient-guided methods. This addresses sampling inefficiency while maintaining reward alignment.

Result: Empirical results show UniCoDe remains competitive with state-of-the-art baselines across various tasks, offering better trade-offs between reward alignment and divergence from the diffusion unconditional prior.

Conclusion: UniCoDe successfully unifies sampling and gradient-based guidance for diffusion model alignment, enabling more efficient sampling while maintaining competitive performance across diverse tasks.

Abstract: Aligning diffusion model outputs with downstream objectives is essential for improving task-specific performance. Broadly, inference-time training-free approaches for aligning diffusion models can be categorized into two main strategies: sampling-based methods, which explore multiple candidate outputs and select those with higher reward signals, and gradient-guided methods, which use differentiable reward approximations to directly steer the generation process. In this work, we propose a universal algorithm, UniCoDe, which brings together the strengths of sampling and gradient-based guidance into a unified framework. UniCoDe integrates local gradient signals during sampling, thereby addressing the sampling inefficiency inherent in complex reward-based sampling approaches. By cohesively combining these two paradigms, UniCoDe enables more efficient sampling while offering better trade-offs between reward alignment and divergence from the diffusion unconditional prior. Empirical results demonstrate that UniCoDe remains competitive with state-of-the-art baselines across a range of tasks. The code is available at https://github.com/maurya-goyal10/UniCoDe

Method: Proposes TCLeaf-Net with three key components: 1) Transformer-convolution module (TCM) for global context and locality preservation, 2) Raw-scale feature recalling and sampling (RSFRS) block to preserve spatial detail, and 3) Deformable alignment block with FPN (DFPN) for multi-scale fusion.

Result: Achieves 78.2% mAP@50 on Daylily-Leaf dataset (5.4pp improvement), reduces computation by 7.5 GFLOPs and GPU memory by 8.7%, outperforms YOLO and RT-DETR series, and shows strong performance on PlantDoc, Tomato-Leaf, and Rice-Leaf datasets.

Conclusion: TCLeaf-Net effectively addresses real-field challenges in foliar disease detection through hybrid architecture design, demonstrating robustness, efficiency, and generalizability across multiple plant disease datasets.

Abstract: Timely and accurate detection of foliar diseases is vital for safeguarding crop growth and reducing yield losses. Yet, in real-field conditions, cluttered backgrounds, domain shifts, and limited lesion-level datasets hinder robust modeling. To address these challenges, we release Daylily-Leaf, a paired lesion-level dataset comprising 1,746 RGB images and 7,839 lesions captured under both ideal and in-field conditions, and propose TCLeaf-Net, a transformer-convolution hybrid detector optimized for real-field use. TCLeaf-Net is designed to tackle three major challenges. To mitigate interference from complex backgrounds, the transformer-convolution module (TCM) couples global context with locality-preserving convolution to suppress non-leaf regions. To reduce information loss during downsampling, the raw-scale feature recalling and sampling (RSFRS) block combines bilinear resampling and convolution to preserve fine spatial detail. To handle variations in lesion scale and feature shifts, the deformable alignment block with FPN (DFPN) employs offset-based alignment and multi-receptive-field perception to strengthen multi-scale fusion. Experimental results show that on the in-field split of the Daylily-Leaf dataset, TCLeaf-Net improves mAP@50 by 5.4 percentage points over the baseline model, reaching 78.2%, while reducing computation by 7.5 GFLOPs and GPU memory usage by 8.7%. Moreover, the model outperforms recent YOLO and RT-DETR series in both precision and recall, and demonstrates strong performance on the PlantDoc, Tomato-Leaf, and Rice-Leaf datasets, validating its robustness and generalizability to other plant disease detection scenarios.

Method: VideoARM uses an adaptive, continuous loop of observing, thinking, acting, and memorizing with a controller that autonomously invokes tools for coarse-to-fine video interpretation. It features hierarchical multimodal memory that continuously captures and updates multi-level clues to support decision-making.

Result: VideoARM outperforms the state-of-the-art method DVD on prevalent benchmarks while significantly reducing token consumption for long-form videos.

Conclusion: The agentic reasoning-over-hierarchical-memory paradigm provides an effective solution for long-form video understanding by enabling adaptive, on-the-fly processing that reduces computational overhead while maintaining or improving performance.

Abstract: Long-form video understanding remains challenging due to the extended temporal structure and dense multimodal cues. Despite recent progress, many existing approaches still rely on hand-crafted reasoning pipelines or employ token-consuming video preprocessing to guide MLLMs in autonomous reasoning. To overcome these limitations, we introduce VideoARM, an Agentic Reasoning-over-hierarchical-Memory paradigm for long-form video understanding. Instead of static, exhaustive preprocessing, VideoARM performs adaptive, on-the-fly agentic reasoning and memory construction. Specifically, VideoARM performs an adaptive and continuous loop of observing, thinking, acting, and memorizing, where a controller autonomously invokes tools to interpret the video in a coarse-to-fine manner, thereby substantially reducing token consumption. In parallel, a hierarchical multimodal memory continuously captures and updates multi-level clues throughout the operation of the agent, providing precise contextual information to support the controller in decision-making. Experiments on prevalent benchmarks demonstrate that VideoARM outperforms the state-of-the-art method, DVD, while significantly reducing token consumption for long-form videos.

Method: STAGE workflow uses STEP2 to predict structural storyboards (start-end frame pairs per shot), multi-shot memory packs for entity consistency, dual-encoding for intra-shot coherence, and two-stage training for cinematic transitions.

Result: Extensive experiments show STAGE achieves superior performance in structured narrative control and cross-shot coherence compared to existing methods.

Conclusion: STAGE provides an effective storyboard-anchored approach for coherent multi-shot video generation with better narrative structure and cinematic quality.

Abstract: While recent advancements in generative models have achieved remarkable visual fidelity in video synthesis, creating coherent multi-shot narratives remains a significant challenge. To address this, keyframe-based approaches have emerged as a promising alternative to computationally intensive end-to-end methods, offering the advantages of fine-grained control and greater efficiency. However, these methods often fail to maintain cross-shot consistency and capture cinematic language. In this paper, we introduce STAGE, a SToryboard-Anchored GEneration workflow to reformulate the keyframe-based multi-shot video generation task. Instead of using sparse keyframes, we propose STEP2 to predict a structural storyboard composed of start-end frame pairs for each shot. We introduce the multi-shot memory pack to ensure long-range entity consistency, the dual-encoding strategy for intra-shot coherence, and the two-stage training scheme to learn cinematic inter-shot transition. We also contribute the large-scale ConStoryBoard dataset, including high-quality movie clips with fine-grained annotations for story progression, cinematic attributes, and human preferences. Extensive experiments demonstrate that STAGE achieves superior performance in structured narrative control and cross-shot coherence.

Method: Two-stage framework: (1) Coarse appearance adaptation using image-only LoRA and subject-embedding adaptation with reference images, (2) Inference-time fine appearance injection using semantic correspondences from RoPE-free mid-layer features to warp appearance-rich value representations into semantically aligned regions.

Result: V-Warper significantly improves appearance fidelity while preserving prompt alignment and motion dynamics, achieving these gains efficiently without large-scale video finetuning.

Conclusion: V-Warper provides an effective training-free solution for video personalization that addresses computational cost and scalability issues while enhancing fine-grained identity consistency.

Abstract: Video personalization aims to generate videos that faithfully reflect a user-provided subject while following a text prompt. However, existing approaches often rely on heavy video-based finetuning or large-scale video datasets, which impose substantial computational cost and are difficult to scale. Furthermore, they still struggle to maintain fine-grained appearance consistency across frames. To address these limitations, we introduce V-Warper, a training-free coarse-to-fine personalization framework for transformer-based video diffusion models. The framework enhances fine-grained identity fidelity without requiring any additional video training. (1) A lightweight coarse appearance adaptation stage leverages only a small set of reference images, which are already required for the task. This step encodes global subject identity through image-only LoRA and subject-embedding adaptation. (2) A inference-time fine appearance injection stage refines visual fidelity by computing semantic correspondences from RoPE-free mid-layer query–key features. These correspondences guide the warping of appearance-rich value representations into semantically aligned regions of the generation process, with masking ensuring spatial reliability. V-Warper significantly improves appearance fidelity while preserving prompt alignment and motion dynamics, and it achieves these gains efficiently without large-scale video finetuning.

Method: Introduce M4Human benchmark with both raw radar tensors (RT) and processed radar point clouds (RPC), plus RGB-D data. Provide high-quality MoCap annotations with 3D meshes and global trajectories across 20 subjects and 50 diverse action types.

Result: Established benchmarks on RT and RPC modalities, plus multimodal fusion with RGB-D. Extensive results highlight M4Human’s significance for radar-based human modeling while revealing persistent challenges under fast, unconstrained motion.

Conclusion: M4Human advances HMR research by providing the largest-scale multimodal radar dataset, enabling exploration across different RF signal granularity levels and addressing privacy-preserving human sensing challenges.

Abstract: Human mesh reconstruction (HMR) provides direct insights into body-environment interaction, which enables various immersive applications. While existing large-scale HMR datasets rely heavily on line-of-sight RGB input, vision-based sensing is limited by occlusion, lighting variation, and privacy concerns. To overcome these limitations, recent efforts have explored radio-frequency (RF) mmWave radar for privacy-preserving indoor human sensing. However, current radar datasets are constrained by sparse skeleton labels, limited scale, and simple in-place actions. To advance the HMR research community, we introduce M4Human, the current largest-scale (661K-frame) ($9\times$ prior largest) multimodal benchmark, featuring high-resolution mmWave radar, RGB, and depth data. M4Human provides both raw radar tensors (RT) and processed radar point clouds (RPC) to enable research across different levels of RF signal granularity. M4Human includes high-quality motion capture (MoCap) annotations with 3D meshes and global trajectories, and spans 20 subjects and 50 diverse actions, including in-place, sit-in-place, and free-space sports or rehabilitation movements. We establish benchmarks on both RT and RPC modalities, as well as multimodal fusion with RGB-D modalities. Extensive results highlight the significance of M4Human for radar-based human modeling while revealing persistent challenges under fast, unconstrained motion. The dataset and code will be released after the paper publication.

Method: SR-DiT integrates three key techniques: token routing (selective processing of tokens), architectural improvements (structural enhancements to the transformer), and training modifications (optimization changes), all built on top of representation alignment. The framework systematically combines these approaches to maximize efficiency.

Result: Achieves FID 3.49 and KDD 0.319 on ImageNet-256 using only a 140M parameter model at 400K iterations without classifier-free guidance. This performance is comparable to results from 685M parameter models trained significantly longer, establishing a state-of-the-art result at this model size.

Conclusion: The systematic combination of multiple efficiency techniques yields significant performance gains, with SR-DiT providing a computationally accessible baseline for future research. The ablation studies reveal which technique combinations are most effective and document both synergies and incompatibilities between approaches.

Abstract: Recent advances have significantly improved the training efficiency of diffusion transformers. However, these techniques have largely been studied in isolation, leaving unexplored the potential synergies from combining multiple approaches. We present SR-DiT (Speedrun Diffusion Transformer), a framework that systematically integrates token routing, architectural improvements, and training modifications on top of representation alignment. Our approach achieves FID 3.49 and KDD 0.319 on ImageNet-256 using only a 140M parameter model at 400K iterations without classifier-free guidance - comparable to results from 685M parameter models trained significantly longer. To our knowledge, this is a state-of the-art result at this model size. Through extensive ablation studies, we identify which technique combinations are most effective and document both synergies and incompatibilities. We release our framework as a computationally accessible baseline for future research.

Method: 1) Cross-state Monte Carlo sampling to enforce global kinematic consistency; 2) Chain-of-Thought reasoning module to infer structural priors (part semantics, joint types, connectivity); 3) Sparse-expert Diffusion Transformer for diverse kinematic interactions; 4) Compositional 3D-VAE latent prior with local-global attention for geometry and part relationships.

Result: ArtGen significantly outperforms state-of-the-art methods on the PartNet-Mobility benchmark for generating articulated 3D objects.

Conclusion: ArtGen effectively addresses the structural-motion entanglement problem in articulated object generation, producing accurate geometry and coherent kinematics from single-view inputs or text descriptions at arbitrary part states.

Abstract: Generating articulated assets is crucial for robotics, digital twins, and embodied intelligence. Existing generative models often rely on single-view inputs representing closed states, resulting in ambiguous or unrealistic kinematic structures due to the entanglement between geometric shape and joint dynamics. To address these challenges, we introduce ArtGen, a conditional diffusion-based framework capable of generating articulated 3D objects with accurate geometry and coherent kinematics from single-view images or text descriptions at arbitrary part-level states. Specifically, ArtGen employs cross-state Monte Carlo sampling to explicitly enforce global kinematic consistency, reducing structural-motion entanglement. Additionally, we integrate a Chain-of-Thought reasoning module to infer robust structural priors, such as part semantics, joint types, and connectivity, guiding a sparse-expert Diffusion Transformer to specialize in diverse kinematic interactions. Furthermore, a compositional 3D-VAE latent prior enhanced with local-global attention effectively captures fine-grained geometry and global part-level relationships. Extensive experiments on the PartNet-Mobility benchmark demonstrate that ArtGen significantly outperforms state-of-the-art methods.

Method: Graph Attention Network (GAT)-based framework that represents each LiDAR sweep as an unstructured spatial graph (points as nodes, edges connecting nearby points while preserving beam-index ordering). Multi-layer GAT learns adaptive attention weights over local geometric neighborhoods and directly regresses missing elevation (z) values at dropout locations.

Result: Trained/evaluated on 1,065 raw KITTI sequences with simulated channel dropout: average height RMSE of 11.67 cm, 87.98% of reconstructed points within 10 cm error threshold. Inference takes 14.65 seconds per frame on single GPU, reconstruction quality stable for different neighborhood sizes k.

Conclusion: Pure graph attention model operating solely on raw point-cloud geometry can effectively recover dropped vertical beams under realistic sensor degradation, with no camera images or temporal information required.

Abstract: Vertical beam dropout in spinning LiDAR sensors triggered by hardware aging, dust, snow, fog, or bright reflections removes entire vertical slices from the point cloud and severely degrades 3D perception in autonomous vehicles. This paper proposes a Graph Attention Network (GAT)-based framework that reconstructs these missing vertical channels using only the current LiDAR frame, with no camera images or temporal information required. Each LiDAR sweep is represented as an unstructured spatial graph: points are nodes and edges connect nearby points while preserving the original beam-index ordering. A multi-layer GAT learns adaptive attention weights over local geometric neighborhoods and directly regresses the missing elevation (z) values at dropout locations. Trained and evaluated on 1,065 raw KITTI sequences with simulated channel dropout, the method achieves an average height RMSE of 11.67 cm, with 87.98% of reconstructed points falling within a 10 cm error threshold. Inference takes 14.65 seconds per frame on a single GPU, and reconstruction quality remains stable for different neighborhood sizes k. These results show that a pure graph attention model operating solely on raw point-cloud geometry can effectively recover dropped vertical beams under realistic sensor degradation.

Method: Created ViInfographicVQA benchmark with 6747 real-world infographics and 20409 human-verified QA pairs across economics, healthcare, education, etc. Includes two evaluation settings: Single-image (traditional VQA) and Multi-image (novel cross-image reasoning requiring synthesis across multiple related infographics).

Result: Evaluation of recent vision-language models revealed substantial performance disparities, with most significant errors occurring on Multi-image questions involving cross-image integration and non-span reasoning. The benchmark provides baseline results for Vietnamese InfographicVQA.

Conclusion: ViInfographicVQA contributes the first Vietnamese benchmark for infographic understanding and highlights limitations of current multimodal models in low-resource contexts, particularly for layout-aware and cross-image reasoning, encouraging future research in these areas.

Abstract: Infographic Visual Question Answering (InfographicVQA) evaluates a model’s ability to read and reason over data-rich, layout-heavy visuals that combine text, charts, icons, and design elements. Compared with scene-text or natural-image VQA, infographics require stronger integration of OCR, layout understanding, and numerical and semantic reasoning. We introduce ViInfographicVQA, the first benchmark for Vietnamese InfographicVQA, comprising over 6747 real-world infographics and 20409 human-verified question-answer pairs across economics, healthcare, education, and more. The benchmark includes two evaluation settings. The Single-image task follows the traditional setup in which each question is answered using a single infographic. The Multi-image task requires synthesizing evidence across multiple semantically related infographics and is, to our knowledge, the first Vietnamese evaluation of cross-image reasoning in VQA. We evaluate a range of recent vision-language models on this benchmark, revealing substantial performance disparities, with the most significant errors occurring on Multi-image questions that involve cross-image integration and non-span reasoning. ViInfographicVQA contributes benchmark results for Vietnamese InfographicVQA and sheds light on the limitations of current multimodal models in low-resource contexts, encouraging future exploration of layout-aware and cross-image reasoning methods.

Method: Two-stage approach: 1) Physically guided controllable bokeh generator produces depth-free bokeh stacks with calibrated strength from sharp input; 2) Lightweight defocus-aware aggregation module fuses features along defocus dimension to expose depth-sensitive variations without changing downstream decoder.

Result: BokehDepth improves visual fidelity over depth-map-based bokeh baselines and consistently boosts metric accuracy and robustness of strong monocular depth foundation models across challenging benchmarks.

Conclusion: The framework successfully leverages defocus as supervision-free geometric cue to simultaneously enhance both bokeh rendering quality and monocular depth estimation performance.

Abstract: Bokeh and monocular depth estimation are tightly coupled through the same lens imaging geometry, yet current methods exploit this connection in incomplete ways. High-quality bokeh rendering pipelines typically depend on noisy depth maps, which amplify estimation errors into visible artifacts, while modern monocular metric depth models still struggle on weakly textured, distant and geometrically ambiguous regions where defocus cues are most informative. We introduce BokehDepth, a two-stage framework that decouples bokeh synthesis from depth prediction and treats defocus as an auxiliary supervision-free geometric cue. In Stage-1, a physically guided controllable bokeh generator, built on a powerful pretrained image editing backbone, produces depth-free bokeh stacks with calibrated bokeh strength from a single sharp input. In Stage-2, a lightweight defocus-aware aggregation module plugs into existing monocular depth encoders, fuses features along the defocus dimension, and exposes stable depth-sensitive variations while leaving downstream decoder unchanged. Across challenging benchmarks, BokehDepth improves visual fidelity over depth-map-based bokeh baselines and consistently boosts the metric accuracy and robustness of strong monocular depth foundation models.

Method: 1) Conditional autoregressive training strategy that aligns new content with existing frames while preserving long-range dependencies. 2) Global 3D-aware attention for continuous geometric guidance. 3) 3D injection mechanism to enforce physical plausibility and geometric consistency.

Result: The framework achieves real-time inference on a single GPU without additional training overhead. It produces long, stable, and visually coherent videos with competitive or superior performance in both visual fidelity and spatial consistency compared to existing methods.

Conclusion: Endless World successfully addresses key challenges in long-horizon and dynamic scene synthesis, enabling infinite, 3D-consistent video generation in real-time streaming scenarios.

Abstract: Producing long, coherent video sequences with stable 3D structure remains a major challenge, particularly in streaming scenarios. Motivated by this, we introduce Endless World, a real-time framework for infinite, 3D-consistent video generation.To support infinite video generation, we introduce a conditional autoregressive training strategy that aligns newly generated content with existing video frames. This design preserves long-range dependencies while remaining computationally efficient, enabling real-time inference on a single GPU without additional training overhead.Moreover, our Endless World integrates global 3D-aware attention to provide continuous geometric guidance across time. Our 3D injection mechanism enforces physical plausibility and geometric consistency throughout extended sequences, addressing key challenges in long-horizon and dynamic scene synthesis.Extensive experiments demonstrate that Endless World produces long, stable, and visually coherent videos, achieving competitive or superior performance to existing methods in both visual fidelity and spatial consistency. Our project has been available on https://bwgzk-keke.github.io/EndlessWorld/.

Method: Introduce GPF - a learnable representation encoding photon distributions as anisotropic 3D Gaussian primitives. Initialized from physically traced photons in first SPPM iteration, optimized using multi-view radiance supervision to distill photon-based light transport into a continuous field.

Result: GPF achieves photon-level accuracy while reducing computation by orders of magnitude on scenes with complex light transport (caustics, specular-diffuse interactions).

Conclusion: GPF unifies physical rigor of photon-based rendering with efficiency of neural scene representations, enabling differentiable radiance evaluation without repeated photon tracing.

Abstract: Accurately modeling light transport is essential for realistic image synthesis. Photon mapping provides physically grounded estimates of complex global illumination effects such as caustics and specular-diffuse interactions, yet its per-view radiance estimation remains computationally inefficient when rendering multiple views of the same scene. The inefficiency arises from independent photon tracing and stochastic kernel estimation at each viewpoint, leading to inevitable redundant computation. To accelerate multi-view rendering, we reformulate photon mapping as a continuous and reusable radiance function. Specifically, we introduce the Gaussian Photon Field (GPF), a learnable representation that encodes photon distributions as anisotropic 3D Gaussian primitives parameterized by position, rotation, scale, and spectrum. GPF is initialized from physically traced photons in the first SPPM iteration and optimized using multi-view supervision of final radiance, distilling photon-based light transport into a continuous field. Once trained, the field enables differentiable radiance evaluation along camera rays without repeated photon tracing or iterative refinement. Extensive experiments on scenes with complex light transport, such as caustics and specular-diffuse interactions, demonstrate that GPF attains photon-level accuracy while reducing computation by orders of magnitude, unifying the physical rigor of photon-based rendering with the efficiency of neural scene representations.

Method: PeRL-VL decouples training into: 1) Perception improvement using VLM-based description rewards that score model’s self-generated image descriptions for faithfulness and sufficiency; 2) Reasoning improvement using text-only Reasoning SFT on logic-rich chain-of-thought data to enhance coherence independently of vision.

Result: PeRL-VL improves average Pass@1 accuracy from 63.3% (base Qwen2.5-VL-7B) to 68.8% across diverse multimodal benchmarks, outperforming standard RLVR, text-only reasoning SFT, and naive multimodal distillation from GPT-4o.

Conclusion: Decoupling perception and reasoning training in VLMs addresses persistent failure modes of RLVR, with separate optimization for visual faithfulness and logical consistency leading to significant performance gains on multimodal reasoning tasks.

Abstract: Reinforcement learning from verifiable rewards (RLVR) has recently been extended from text-only LLMs to vision-language models (VLMs) to elicit long-chain multimodal reasoning. However, RLVR-trained VLMs still exhibit two persistent failure modes: inaccurate visual extraction (missing or hallucinating details) and logically inconsistent chains-of-thought, largely because verifiable signals supervise only the final answer. We propose PeRL-VL (Perception and Reasoning Learning for Vision-Language Models), a decoupled framework that separately improves visual perception and textual reasoning on top of RLVR. For perception, PeRL-VL introduces a VLM-based description reward that scores the model’s self-generated image descriptions for faithfulness and sufficiency. For reasoning, PeRL-VL adds a text-only Reasoning SFT stage on logic-rich chain-of-thought data, enhancing coherence and logical consistency independently of vision. Across diverse multimodal benchmarks, PeRL-VL improves average Pass@1 accuracy from 63.3% (base Qwen2.5-VL-7B) to 68.8%, outperforming standard RLVR, text-only reasoning SFT, and naive multimodal distillation from GPT-4o.

Method: Two-stage detector-verifier framework: 1) YOLOv11 detector adaptively adjusts per-frame confidence thresholds under VLM guidance, 2) VLM verifier fine-tuned with Group Relative Policy Optimization (GRPO) using asymmetric, cost-sensitive reward function designed to discourage missed detections. Constructed synthetic testbed by degrading clean datasets with adverse clinical conditions.

Result: Zero-shot evaluation on synthetically degraded CVC-ClinicDB and Kvasir-SEG images shows recall improvements of 14-22 percentage points over YOLO alone, while precision remains within 0.7 points below to 1.7 points above baseline.

Conclusion: Adaptive thresholding with cost-sensitive reinforcement learning achieves clinically aligned, open-world polyp detection with substantially fewer false negatives, reducing risk of missed precancerous polyps and improving patient outcomes.

Abstract: Polyp detectors trained on clean datasets often underperform in real-world endoscopy, where illumination changes, motion blur, and occlusions degrade image quality. Existing approaches struggle with the domain gap between controlled laboratory conditions and clinical practice, where adverse imaging conditions are prevalent. In this work, we propose AdaptiveDetector, a novel two-stage detector-verifier framework comprising a YOLOv11 detector with a vision-language model (VLM) verifier. The detector adaptively adjusts per-frame confidence thresholds under VLM guidance, while the verifier is fine-tuned with Group Relative Policy Optimization (GRPO) using an asymmetric, cost-sensitive reward function specifically designed to discourage missed detections – a critical clinical requirement. To enable realistic assessment under challenging conditions, we construct a comprehensive synthetic testbed by systematically degrading clean datasets with adverse conditions commonly encountered in clinical practice, providing a rigorous benchmark for zero-shot evaluation. Extensive zero-shot evaluation on synthetically degraded CVC-ClinicDB and Kvasir-SEG images demonstrates that our approach improves recall by 14 to 22 percentage points over YOLO alone, while precision remains within 0.7 points below to 1.7 points above the baseline. This combination of adaptive thresholding and cost-sensitive reinforcement learning achieves clinically aligned, open-world polyp detection with substantially fewer false negatives, thereby reducing the risk of missed precancerous polyps and improving patient outcomes.

Method: Proposes a patch-driven relational refinement with a relational gated graph attention network that constructs patch graphs and performs edge-aware attention to emphasize informative inter-patch interactions. Uses learnable multi-aggregation pooling to create compact, task-discriminative representations. The graph refinement is only used during training to distill relational structure into the cache, with no additional inference cost beyond standard cache lookup.

Result: Extensive evaluations on 11 benchmarks show consistent gains over state-of-the-art CLIP adapter and cache-based baselines while preserving zero-shot efficiency. Also validates battlefield relevance with a new Injured vs. Uninjured Soldier dataset for casualty recognition.

Conclusion: The proposed approach effectively addresses the limitation of global representations in few-shot adaptation by leveraging patch-level relational structure, achieving better performance without sacrificing inference efficiency, with practical applications in time-critical military casualty care scenarios.

Abstract: Few-shot image classification remains difficult under limited supervision and visual domain shift. Recent cache-based adaptation approaches (e.g., Tip-Adapter) address this challenge to some extent by learning lightweight residual adapters over frozen features, yet they still inherit CLIP’s tendency to encode global, general-purpose representations that are not optimally discriminative to adapt the generalist to the specialist’s domain in low-data regimes. We address this limitation with a novel patch-driven relational refinement that learns cache adapter weights from intra-image patch dependencies rather than treating an image embedding as a monolithic vector. Specifically, we introduce a relational gated graph attention network that constructs a patch graph and performs edge-aware attention to emphasize informative inter-patch interactions, producing context-enriched patch embeddings. A learnable multi-aggregation pooling then composes these into compact, task-discriminative representations that better align cache keys with the target few-shot classes. Crucially, the proposed graph refinement is used only during training to distil relational structure into the cache, incurring no additional inference cost beyond standard cache lookup. Final predictions are obtained by a residual fusion of cache similarity scores with CLIP zero-shot logits. Extensive evaluations on 11 benchmarks show consistent gains over state-of-the-art CLIP adapter and cache-based baselines while preserving zero-shot efficiency. We further validate battlefield relevance by introducing an Injured vs. Uninjured Soldier dataset for casualty recognition. It is motivated by the operational need to support triage decisions within the “platinum minutes” and the broader “golden hour” window in time-critical UAV-driven search-and-rescue and combat casualty care.

Method: DMLR employs confidence-guided latent policy gradient optimization to refine latent think tokens for in-depth reasoning. It introduces a Dynamic Visual Injection Strategy that retrieves the most relevant visual features at each latent think token, updates the best visual patches, and injects them into latent think tokens to achieve dynamic visual-textual interleaving.

Result: Experiments across seven multimodal reasoning benchmarks and various model architectures demonstrate that DMLR significantly improves reasoning and perception performance while maintaining high inference efficiency.

Conclusion: DMLR provides an effective framework for dynamic multimodal reasoning that mimics human cognitive processes, overcoming limitations of existing CoT-based approaches through latent reasoning and dynamic visual-textual interleaving.

Abstract: Recent advancements in Multimodal Large Language Models (MLLMs) have significantly enhanced cross-modal understanding and reasoning by incorporating Chain-of-Thought (CoT) reasoning in the semantic space. Building upon this, recent studies extend the CoT mechanism to the visual modality, enabling models to integrate visual information during reasoning through external tools or explicit image generation. However, these methods remain dependent on explicit step-by-step reasoning, unstable perception-reasoning interaction and notable computational overhead. Inspired by human cognition, we posit that thinking unfolds not linearly but through the dynamic interleaving of reasoning and perception within the mind. Motivated by this perspective, we propose DMLR, a test-time Dynamic Multimodal Latent Reasoning framework that employs confidence-guided latent policy gradient optimization to refine latent think tokens for in-depth reasoning. Furthermore, a Dynamic Visual Injection Strategy is introduced, which retrieves the most relevant visual features at each latent think token and updates the set of best visual patches. The updated patches are then injected into latent think token to achieve dynamic visual-textual interleaving. Experiments across seven multimodal reasoning benchmarks and various model architectures demonstrate that DMLR significantly improves reasoning and perception performance while maintaining high inference efficiency.

Method: Leverage off-the-shelf video diffusion models to generate realistic 3D spatial and temporal variations from static image datasets, creating synthetic video clips as supplemental training data with practical guidelines for annotation transfer and disocclusion handling.

Result: Consistent performance gains in low-data settings, particularly for UAV-captured imagery where annotations are scarce. The method broadens data distribution along axes underrepresented by traditional augmentation methods.

Conclusion: Spatiotemporal augmentation using video foundation models offers an effective approach for improving model performance in data-scarce regimes by generating realistic 3D viewpoint and scene dynamics variations that traditional methods cannot capture.

Abstract: We explore spatiotemporal data augmentation using video foundation models to diversify both camera viewpoints and scene dynamics. Unlike existing approaches based on simple geometric transforms or appearance perturbations, our method leverages off-the-shelf video diffusion models to generate realistic 3D spatial and temporal variations from a given image dataset. Incorporating these synthesized video clips as supplemental training data yields consistent performance gains in low-data settings, such as UAV-captured imagery where annotations are scarce. Beyond empirical improvements, we provide practical guidelines for (i) choosing an appropriate spatiotemporal generative setup, (ii) transferring annotations to synthetic frames, and (iii) addressing disocclusion - regions newly revealed and unlabeled in generated views. Experiments on COCO subsets and UAV-captured datasets show that, when applied judiciously, spatiotemporal augmentation broadens the data distribution along axes underrepresented by traditional and prior generative methods, offering an effective lever for improving model performance in data-scarce regimes.

Method: Introduces Motion Score Distillation (MSD) as an SDS alternative, using a LoRA-enhanced video diffusion model with static source distribution, inversion-based noise estimation for appearance preservation, temporal/spatial regularization to mitigate geometric distortions, and a motion refinement module for temporal upscaling and detail enhancement.

Result: Extensive experiments show Animus3D successfully animates static 3D assets from diverse text prompts, generating significantly more substantial and detailed motion than state-of-the-art baselines while maintaining high visual integrity.

Conclusion: Animus3D provides an effective framework for text-driven 3D animation that overcomes limitations of previous SDS-based methods, achieving superior motion quality and detail through the novel MSD approach and complementary regularization and refinement techniques.

Abstract: We present Animus3D, a text-driven 3D animation framework that generates motion field given a static 3D asset and text prompt. Previous methods mostly leverage the vanilla Score Distillation Sampling (SDS) objective to distill motion from pretrained text-to-video diffusion, leading to animations with minimal movement or noticeable jitter. To address this, our approach introduces a novel SDS alternative, Motion Score Distillation (MSD). Specifically, we introduce a LoRA-enhanced video diffusion model that defines a static source distribution rather than pure noise as in SDS, while another inversion-based noise estimation technique ensures appearance preservation when guiding motion. To further improve motion fidelity, we incorporate explicit temporal and spatial regularization terms that mitigate geometric distortions across time and space. Additionally, we propose a motion refinement module to upscale the temporal resolution and enhance fine-grained details, overcoming the fixed-resolution constraints of the underlying video model. Extensive experiments demonstrate that Animus3D successfully animates static 3D assets from diverse text prompts, generating significantly more substantial and detailed motion than state-of-the-art baselines while maintaining high visual integrity. Code will be released at https://qiisun.github.io/animus3d_page.

Method: Proposes a framework integrating myocardial anatomical priors, structure-aware feature encoding, and 3D wavelet-inverse wavelet transformations. Uses residual attention for feature enhancement, wavelet-based down/upsampling for spatial-frequency modeling, and multi-scale feature fusion in decoding.

Result: Achieves Dice coefficient of 0.8082, Sensitivity of 0.7946, Precision of 0.8471, and HD95 of 9.77 mm on ImageCAS dataset, outperforming mainstream segmentation models. Ablation studies confirm complementary contributions of components.

Conclusion: The method enables more stable and consistent coronary artery segmentation under complex geometric conditions, providing reliable results for subsequent coronary structure analysis tasks.

Abstract: Accurate coronary artery segmentation from coronary computed tomography angiography is essential for quantitative coronary analysis and clinical decision support. Nevertheless, reliable segmentation remains challenging because of small vessel calibers, complex branching, blurred boundaries, and myocardial interference. We propose a coronary artery segmentation framework that integrates myocardial anatomical priors, structure aware feature encoding, and three dimensional wavelet inverse wavelet transformations. Myocardial priors and residual attention based feature enhancement are incorporated during encoding to strengthen coronary structure representation. Wavelet inverse wavelet based downsampling and upsampling enable joint spatial frequency modeling and preserve multi scale structural consistency, while a multi scale feature fusion module integrates semantic and geometric information in the decoding stage. The model is trained and evaluated on the public ImageCAS dataset using a 3D overlapping patch based strategy with a 7:1:2 split for training, validation, and testing. Experimental results demonstrate that the proposed method achieves a Dice coefficient of 0.8082, Sensitivity of 0.7946, Precision of 0.8471, and an HD95 of 9.77 mm, outperforming several mainstream segmentation models. Ablation studies further confirm the complementary contributions of individual components. The proposed method enables more stable and consistent coronary artery segmentation under complex geometric conditions, providing reliable segmentation results for subsequent coronary structure analysis tasks.

Method: ATP uses a lightweight gating module that assigns hybrid importance scores combining ViT CLS attention (intra-modal) and CLIP text-image similarity (inter-modal) to retain only top-K informative tokens for LLM processing.

Result: ATP reduces inference FLOPs by ~40%, achieves ~1.5x speedup in end-to-end latency with <1% accuracy loss on VQAv2, GQA, and COCO benchmarks, and improves robustness under corruptions.

Conclusion: Resource-constrained inference and model reliability are not competing objectives; ATP enables efficient multimodal edge computing while preserving visual grounding and interpretability.

Abstract: Real-world deployment of Vision-Language Models (VLMs) is hindered by high computational demands, as existing architectures inefficiently process all tokens uniformly. We introduce Adaptive Token Pruning (ATP), a dynamic inference mechanism that retains only the most informative tokens based on contextual relevance. ATP operates at the vision-language interface, assigning a hybrid importance score combining ViT CLS attention (intra-modal saliency) and CLIP text-image similarity (inter-modal relevance) to keep top-K tokens for the LLM. Unlike static compression, ATP adapts to each input without modifying the backbone. Proposed as a lightweight gating module, ATP is compatible with popular backbones like BLIP-2, LLaVA, and Flamingo. Preliminary evaluations across VQAv2, GQA, and COCO indicate that ATP reduces inference FLOPs by around 40% and achieves roughly 1.5x speedups in end-to-end latency with negligible accuracy loss (less than 1%). Qualitative analyses suggest ATP preserves visual grounding and enhances interpretability. Beyond efficiency, we investigate robustness under corruptions; observations suggest adaptive pruning suppresses spurious correlations, improving stability. These findings imply that resource-constrained inference and model reliability are not competing objectives. Finally, we discuss ATP’s role in efficient multimodal edge computing pipelines.

Method: Video-to-image aggregation strategy arranges multiple frames spatially into single images, uses supervised contrastive learning with positive pairs from same-label videos, creates multiple natural views via different temporal frame samplings without data augmentation.

Result: Outperforms ViViT significantly: 76% vs 43% accuracy on Penn Action, 48% vs 37% on HMDB51, with fewer computational resources required.

Conclusion: Supervised Contrastive Frame Aggregation method effectively learns video representations for classification and captioning tasks in both supervised and self-supervised settings, offering superior performance with lower computational cost.

Abstract: We propose a supervised contrastive learning framework for video representation learning that leverages temporally global context. We introduce a video to image aggregation strategy that spatially arranges multiple frames from each video into a single input image. This design enables the use of pre trained convolutional neural network backbones such as ResNet50 and avoids the computational overhead of complex video transformer models. We then design a contrastive learning objective that directly compares pairwise projections generated by the model. Positive pairs are defined as projections from videos sharing the same label while all other projections are treated as negatives. Multiple natural views of the same video are created using different temporal frame samplings from the same underlying video. Rather than relying on data augmentation these frame level variations produce diverse positive samples with global context and reduce overfitting. Experiments on the Penn Action and HMDB51 datasets demonstrate that the proposed method outperforms existing approaches in classification accuracy while requiring fewer computational resources. The proposed Supervised Contrastive Frame Aggregation method learns effective video representations in both supervised and self supervised settings and supports video based tasks such as classification and captioning. The method achieves seventy six percent classification accuracy on Penn Action compared to forty three percent achieved by ViVIT and forty eight percent accuracy on HMDB51 compared to thirty seven percent achieved by ViVIT.

Method: Proposes token pruning with MSSAVT redundancy metric (considers token similarity and spatial position), masked pruning strategy to avoid bidirectional dependency, and integrates temporal redundancy pruning.

Result: Significantly improves accuracy (up to 4%) on multiple video understanding benchmarks with negligible pruning latency (<1ms).

Conclusion: The proposed token pruning approach effectively reduces computational overhead while enhancing performance for video MLLMs, making them more practical for real-time applications.

Abstract: Online video understanding is essential for applications like public surveillance and AI glasses. However, applying Multimodal Large Language Models (MLLMs) to this domain is challenging due to the large number of video frames, resulting in high GPU memory usage and computational latency. To address these challenges, we propose token pruning as a means to reduce context length while retaining critical information. Specifically, we introduce a novel redundancy metric, Maximum Similarity to Spatially Adjacent Video Tokens (MSSAVT), which accounts for both token similarity and spatial position. To mitigate the bidirectional dependency between pruning and redundancy, we further design a masked pruning strategy that ensures only mutually unadjacent tokens are pruned. We also integrate an existing temporal redundancy-based pruning method to eliminate temporal redundancy of the video modality. Experimental results on multiple online and offline video understanding benchmarks demonstrate that our method significantly improves the accuracy (i.e., by 4% at most) while incurring a negligible pruning latency (i.e., less than 1ms). Our full implementation will be made publicly available.

Method: MVP acquires multiple physical views per scene by varying camera exposure triangle settings (ISO, shutter speed, aperture), selects top-k settings using source-affinity scores, applies lightweight digital augmentations, filters lowest-entropy views, and aggregates predictions via hard voting.

Result: Outperforms digital-only TTA by up to 25.6 percentage points on ImageNet-ES datasets, delivers additional 3.4 pp gains over conventional sensor control + TTA pipelines, and remains effective with reduced parameter sets for lower latency.

Conclusion: Measurement-time control through physical view selection and combination substantially improves VLM robustness beyond post-capture prompting, demonstrating practical sensor-mediated adaptation.

Abstract: To extend the application of vision-language models (VLMs) from web images to sensor-mediated physical environments, we propose Multi-View Physical-prompt for Test-Time Adaptation (MVP), a forward-only framework that moves test-time adaptation (TTA) from tokens to photons by treating the camera exposure triangle–ISO, shutter speed, and aperture–as physical prompts. At inference, MVP acquires a library of physical views per scene, selects the top-k sensor settings using a source-affinity score, evaluates each retained view under lightweight digital augmentations, filters the lowest-entropy subset of augmented views, and aggregates predictions with Zero-temperature softmax (i.e., hard voting). This selection-then-vote design is simple, calibration-friendly, and requires no gradients or model modifications. On ImageNet-ES and ImageNet-ES-Diverse, MVP consistently outperforms digital-only TTA on single Auto-Exposure captures, by up to 25.6 percentage points (pp), and delivers up to 3.4 pp additional gains over pipelines that combine conventional sensor control with TTA. MVP remains effective under reduced parameter candidate sets that lower capture latency, demonstrating practicality. These results support the main claim that, beyond post-capture prompting, measurement-time control–selecting and combining real physical views–substantially improves robustness for VLMs.

Method: A zero-shot recognition framework adapting the Retrieval-Augmented Generation (RAG) paradigm: 1) Vision Language Model generates textual description from input image, 2) retrieves relevant sign candidates from vector database of reference designs, 3) Large Language Model reasons over retrieved candidates for final fine-grained recognition.

Result: Achieved 95.58% accuracy on ideal reference images and 82.45% accuracy on challenging real-world road data when tested on 303 regulatory signs from the Ohio MUTCD.

Conclusion: Demonstrates viability of RAG-based architectures for creating scalable and accurate road sign recognition systems without task-specific training, addressing the limitations of traditional deep learning approaches.

Abstract: Automated road sign recognition is a critical task for intelligent transportation systems, but traditional deep learning methods struggle with the sheer number of sign classes and the impracticality of creating exhaustive labeled datasets. This paper introduces a novel zero-shot recognition framework that adapts the Retrieval-Augmented Generation (RAG) paradigm to address this challenge. Our method first uses a Vision Language Model (VLM) to generate a textual description of a sign from an input image. This description is used to retrieve a small set of the most relevant sign candidates from a vector database of reference designs. Subsequently, a Large Language Model (LLM) reasons over the retrieved candidates to make a final, fine-grained recognition. We validate this approach on a comprehensive set of 303 regulatory signs from the Ohio MUTCD. Experimental results demonstrate the framework’s effectiveness, achieving 95.58% accuracy on ideal reference images and 82.45% on challenging real-world road data. This work demonstrates the viability of RAG-based architectures for creating scalable and accurate systems for road sign recognition without task-specific training.

Method: StegaVAR embeds action videos into ordinary cover videos and performs VAR directly in the steganographic domain. It introduces two key components: Secret Spatio-Temporal Promotion (STeP) which uses the secret video to guide spatiotemporal feature extraction during training, and Cross-Band Difference Attention (CroDA) which suppresses cover interference by capturing cross-band semantic differences.

Result: Experiments demonstrate that StegaVAR achieves superior VAR and privacy-preserving performance on widely used datasets. The framework is also effective for multiple steganographic models.

Conclusion: StegaVAR successfully addresses both concealment and spatiotemporal preservation issues in privacy-preserving video action recognition by operating in the steganographic domain, maintaining complete spatiotemporal information while ensuring natural appearance for transmission concealment.

Abstract: Despite the rapid progress of deep learning in video action recognition (VAR) in recent years, privacy leakage in videos remains a critical concern. Current state-of-the-art privacy-preserving methods often rely on anonymization. These methods suffer from (1) low concealment, where producing visually distorted videos that attract attackers’ attention during transmission, and (2) spatiotemporal disruption, where degrading essential spatiotemporal features for accurate VAR. To address these issues, we propose StegaVAR, a novel framework that embeds action videos into ordinary cover videos and directly performs VAR in the steganographic domain for the first time. Throughout both data transmission and action analysis, the spatiotemporal information of hidden secret video remains complete, while the natural appearance of cover videos ensures the concealment of transmission. Considering the difficulty of steganographic domain analysis, we propose Secret Spatio-Temporal Promotion (STeP) and Cross-Band Difference Attention (CroDA) for analysis within the steganographic domain. STeP uses the secret video to guide spatiotemporal feature extraction in the steganographic domain during training. CroDA suppresses cover interference by capturing cross-band semantic differences. Experiments demonstrate that StegaVAR achieves superior VAR and privacy-preserving performance on widely used datasets. Moreover, our framework is effective for multiple steganographic models.

Method: Proposes a framework with rectified flow mechanism for linear noise-data connections, bidirectional tokenization for text-image-video fusion, spatial-temporal feature embedding, hybrid text-image sequence modeling, and noise-aware learning algorithm for handling noisy data distributions.

Result: Achieves 25% increase in image resolution clarity and 20% reduction in computational requirements compared to diffusion-based methods, with robust scalability and adaptability demonstrated on benchmark datasets.

Conclusion: The work demonstrates the transformative potential of vision-centric LLMs in redefining computer vision and multimodal AI capabilities, with promising applications in autonomous systems, creative content generation, and advanced video analysis.

Abstract: This research introduces a transformative framework for integrating Vision-Enhanced Large Language Models (LLMs) with advanced transformer-based architectures to tackle challenges in high-resolution image synthesis and multimodal data interpretation. The proposed model incorporates a rectified flow mechanism that connects noise and data with linear paths, enabling efficient and high-quality generation. A bidirectional tokenization strategy is employed to seamlessly merge inputs from text, image, and video modalities, fostering a unified understanding across diverse data types. By embedding spatial-temporal features and leveraging a hybrid text-image sequence modeling approach, the framework achieves unparalleled fidelity in synthesized images and coherent multimodal representations. The architecture is optimized with a noise-aware learning algorithm, addressing discrepancies in noisy data distributions and improving generative performance under varying input conditions. Rigorous evaluations on benchmark datasets demonstrate a 25% increase in image resolution clarity and a 20% reduction in computational requirements compared to diffusion-based methods. Furthermore, the model exhibits robust scalability and adaptability, showcasing its potential in applications like autonomous systems, creative content generation, and advanced video analysis. This work underscores the role of vision-centric LLMs in redefining capabilities in computer vision and multimodal artificial intelligence.

Method: Two-step pipeline: 1) VLM analyzes image to identify objects and spatial relationships, generating text-based placement plans; 2) Plans are rendered into final HTML-format layouts.

Result: Quantitative and qualitative evaluations show the method produces higher-quality ad layouts by explicitly considering background image content compared to existing methods.

Conclusion: Leveraging VLMs for semantic understanding of background images enables more effective advertisement layout generation than conventional saliency-based techniques.

Abstract: In this paper, we propose a method for generating layouts for image-based advertisements by leveraging a Vision-Language Model (VLM). Conventional advertisement layout techniques have predominantly relied on saliency mapping to detect salient regions within a background image, but such approaches often fail to fully account for the image’s detailed composition and semantic content. To overcome this limitation, our method harnesses a VLM to recognize the products and other elements depicted in the background and to inform the placement of text and logos. The proposed layout-generation pipeline consists of two steps. In the first step, the VLM analyzes the image to identify object types and their spatial relationships, then produces a text-based “placement plan” based on this analysis. In the second step, that plan is rendered into the final layout by generating HTML-format code. We validated the effectiveness of our approach through evaluation experiments, conducting both quantitative and qualitative comparisons against existing methods. The results demonstrate that by explicitly considering the background image’s content, our method produces noticeably higher-quality advertisement layouts.

Method: Two key contributions: (1) a scene-consistent data construction pipeline that generates diverse, geometrically-grounded training pairs, and (2) a novel geometry-guided attention loss that leverages cross-view cues to regularize the model’s spatial reasoning.

Result: Experiments on a scene-consistent benchmark show the approach achieves better scene alignment and text-image consistency than state-of-the-art baselines according to both automatic metrics and human preference studies. The method produces geometrically coherent images with diverse compositions that remain faithful to textual instructions and underlying scene structure.

Conclusion: The proposed method successfully resolves the trade-off between scene preservation and prompt adherence in geometry-aware scene-consistent image generation, producing high-quality results that maintain both scene consistency and responsiveness to spatial relation specifications.

Abstract: We study geometry-aware scene-consistent image generation: given a reference scene image and a text condition specifying an entity to be generated in the scene and its spatial relation to the scene, the goal is to synthesize an output image that preserves the same physical environment as the reference scene while correctly generating the entity according to the spatial relation described in the text. Existing methods struggle to balance scene preservation with prompt adherence: they either replicate the scene with high fidelity but poor responsiveness to the prompt, or prioritize prompt compliance at the expense of scene consistency. To resolve this trade-off, we introduce two key contributions: (i) a scene-consistent data construction pipeline that generates diverse, geometrically-grounded training pairs, and (ii) a novel geometry-guided attention loss that leverages cross-view cues to regularize the model’s spatial reasoning. Experiments on our scene-consistent benchmark show that our approach achieves better scene alignment and text-image consistency than state-of-the-art baselines, according to both automatic metrics and human preference studies. Our method produces geometrically coherent images with diverse compositions that remain faithful to the textual instructions and the underlying scene structure.

Method: Compared three intelligent systems (Fuzzy Logic, Neural Networks, Case-Based Reasoning), selected CBR, performed data pre-processing and splitting on heart disease dataset, then applied CBR for prediction with training/testing sets.

Result: CBR achieved 97.95% accuracy in heart disease prediction. Gender analysis showed 57.76% probability for males and 42.24% for females. Related studies identified smoking and alcohol consumption as significant risk factors, especially for males.

Conclusion: Case-Based Reasoning is an effective intelligent system for heart disease prediction with high accuracy. The study highlights gender disparities in heart disease risk and identifies modifiable lifestyle factors as key contributors to the condition.

Abstract: This study provides an overview of heart disease prediction using an intelligent system. Predicting disease accurately is crucial in the medical field, but traditional methods relying solely on a doctor’s experience often lack precision. To address this limitation, intelligent systems are applied as an alternative to traditional approaches. While various intelligent system methods exist, this study focuses on three: Fuzzy Logic, Neural Networks, and Case-Based Reasoning (CBR). A comparison of these techniques in terms of accuracy was conducted, and ultimately, Case-Based Reasoning (CBR) was selected for heart disease prediction. In the prediction phase, the heart disease dataset underwent data pre-processing to clean the data and data splitting to separate it into training and testing sets. The chosen intelligent system was then employed to predict heart disease outcomes based on the processed data. The experiment concluded with Case-Based Reasoning (CBR) achieving a notable accuracy rate of 97.95% in predicting heart disease. The findings also revealed that the probability of heart disease was 57.76% for males and 42.24% for females. Further analysis from related studies suggests that factors such as smoking and alcohol consumption are significant contributors to heart disease, particularly among males.

Method: X-Slim introduces a unified framework exploiting redundancy across timesteps, blocks, and tokens. It uses a dual-threshold controller for push-then-polish caching: first pushes timestep-level reuse to an early-warning line, then switches to lightweight block/token-level refresh, triggering full inference only when crossing a critical line to reset accumulated error.

Result: Achieves up to 4.97x latency reduction on FLUX.1-dev and 3.52x on HunyuanVideo with minimal perceptual loss. On DiT-XL/2, reaches 3.13x acceleration while improving FID by 2.42 over prior methods.

Conclusion: X-Slim advances the speed-quality frontier for diffusion models by providing a unified caching framework that intelligently balances computational savings with quality preservation through multi-level redundancy exploitation and adaptive control.

Abstract: Diffusion models achieve remarkable generative quality, but computational overhead scales with step count, model depth, and sequence length. Feature caching is effective since adjacent timesteps yield highly similar features. However, an inherent trade-off remains: aggressive timestep reuse offers large speedups but can easily cross the critical line, hurting fidelity, while block- or token-level reuse is safer but yields limited computational savings. We present X-Slim (eXtreme-Slimming Caching), a training-free, cache-based accelerator that, to our knowledge, is the first unified framework to exploit cacheable redundancy across timesteps, structure (blocks), and space (tokens). Rather than simply mixing levels, X-Slim introduces a dual-threshold controller that turns caching into a push-then-polish process: it first pushes reuse at the timestep level up to an early-warning line, then switches to lightweight block- and token-level refresh to polish the remaining redundancy, and triggers full inference once the critical line is crossed to reset accumulated error. At each level, context-aware indicators decide when and where to cache. Across diverse tasks, X-Slim advances the speed-quality frontier. On FLUX.1-dev and HunyuanVideo, it reduces latency by up to 4.97x and 3.52x with minimal perceptual loss. On DiT-XL/2, it reaches 3.13x acceleration and improves FID by 2.42 over prior methods.

Method: Patchify divides database images into structured patches and compares local patch features with global query descriptors. Introduces LocScore metric for localization-aware evaluation. Uses Product Quantization for efficient large-scale retrieval.

Result: Patchify outperforms global methods and complements state-of-the-art reranking pipelines across multiple benchmarks, backbones, and region selection strategies. Informative feature compression significantly boosts performance.

Conclusion: Patchify provides an effective, scalable, and interpretable retrieval framework with spatial correctness evaluation, demonstrating strong performance and practical utility for instance-level image retrieval tasks.

Abstract: Instance-level image retrieval aims to find images containing the same object as a given query, despite variations in size, position, or appearance. To address this challenging task, we propose Patchify, a simple yet effective patch-wise retrieval framework that offers high performance, scalability, and interpretability without requiring fine-tuning. Patchify divides each database image into a small number of structured patches and performs retrieval by comparing these local features with a global query descriptor, enabling accurate and spatially grounded matching. To assess not just retrieval accuracy but also spatial correctness, we introduce LocScore, a localization-aware metric that quantifies whether the retrieved region aligns with the target object. This makes LocScore a valuable diagnostic tool for understanding and improving retrieval behavior. We conduct extensive experiments across multiple benchmarks, backbones, and region selection strategies, showing that Patchify outperforms global methods and complements state-of-the-art reranking pipelines. Furthermore, we apply Product Quantization for efficient large-scale retrieval and highlight the importance of using informative features during compression, which significantly boosts performance. Project website: https://wons20k.github.io/PatchwiseRetrieval/

Method: 1) Dynamic 3D Chain-of-Thought (3D CoT) that unifies planning, grounding, navigation, and QA within a single 3D-VLM and CoT pipeline. 2) Synergistic Learning from Fragmented Supervision (SLFS) using masked autoregressive loss to learn from massive partially-annotated hybrid data, allowing different CoT components to mutually reinforce each other.

Result: Achieves state-of-the-art results on multiple benchmarks: Vision-and-Language Navigation (R2R-CE, REVERIE-CE, NavRAG-CE), Object-goal Navigation (HM3D-OVON), and Task-oriented Sequential Grounding and Navigation (SG3D). Real-world mobile manipulation experiments further validate effectiveness.

Conclusion: D3D-VLP successfully bridges the gap between interpretable modular systems and end-to-end learning by introducing unified 3D reasoning with synergistic learning, demonstrating strong performance across diverse embodied AI tasks and real-world applications.

Abstract: Embodied agents face a critical dilemma that end-to-end models lack interpretability and explicit 3D reasoning, while modular systems ignore cross-component interdependencies and synergies. To bridge this gap, we propose the Dynamic 3D Vision-Language-Planning Model (D3D-VLP). Our model introduces two key innovations: 1) A Dynamic 3D Chain-of-Thought (3D CoT) that unifies planning, grounding, navigation, and question answering within a single 3D-VLM and CoT pipeline; 2) A Synergistic Learning from Fragmented Supervision (SLFS) strategy, which uses a masked autoregressive loss to learn from massive and partially-annotated hybrid data. This allows different CoT components to mutually reinforce and implicitly supervise each other. To this end, we construct a large-scale dataset with 10M hybrid samples from 5K real scans and 20K synthetic scenes that are compatible with online learning methods such as RL and DAgger. Our D3D-VLP achieves state-of-the-art results on multiple benchmarks, including Vision-and-Language Navigation (R2R-CE, REVERIE-CE, NavRAG-CE), Object-goal Navigation (HM3D-OVON), and Task-oriented Sequential Grounding and Navigation (SG3D). Real-world mobile manipulation experiments further validate the effectiveness.

Method: 1) DiG proxy task: MLLMs identify and localize all differences between similar image pairs without prior knowledge of difference count. 2) Automated 3D rendering pipeline generates high-quality paired images with controllable discrepancies. 3) Curriculum learning from single to multiple differences addresses sparse difference signals.

Result: DiG significantly improves model performance across various visual perception benchmarks. Learned fine-grained perception skills transfer effectively to standard downstream tasks including RefCOCO, RefCOCO+, RefCOCOg, and general multimodal perception benchmarks.

Conclusion: Differential grounding is a scalable and robust approach for advancing fine-grained visual reasoning in MLLMs, enabling better visual perception without requiring extensive manual annotation.

Abstract: Multimodal Large Language Models have achieved impressive performance on a variety of vision-language tasks, yet their fine-grained visual perception and precise spatial reasoning remain limited. In this work, we introduce DiG (Differential Grounding), a novel proxy task framework where MLLMs learn fine-grained perception by identifying and localizing all differences between similar image pairs without prior knowledge of their number. To support scalable training, we develop an automated 3D rendering-based data generation pipeline that produces high-quality paired images with fully controllable discrepancies. To address the sparsity of difference signals, we further employ curriculum learning that progressively increases complexity from single to multiple differences, enabling stable optimization. Extensive experiments demonstrate that DiG significantly improves model performance across a variety of visual perception benchmarks and that the learned fine-grained perception skills transfer effectively to standard downstream tasks, including RefCOCO, RefCOCO+, RefCOCOg, and general multimodal perception benchmarks. Our results highlight differential grounding as a scalable and robust approach for advancing fine-grained visual reasoning in MLLMs.

Method: Proposes Mio (Multimodal Interactive Omni-Avatar) framework with five modules: Thinker (cognitive reasoning), Talker, Face Animator, Body Animator, and Renderer. End-to-end architecture integrates reasoning with real-time multimodal embodiment.

Result: Superior performance compared to state-of-the-art methods across all evaluated dimensions on new benchmark for interactive intelligence. Framework enables fluid, consistent interaction.

Conclusion: Mio framework advances digital humans from superficial imitation to intelligent interaction with personality alignment, adaptation, and self-evolution capabilities.

Abstract: We introduce Interactive Intelligence, a novel paradigm of digital human that is capable of personality-aligned expression, adaptive interaction, and self-evolution. To realize this, we present Mio (Multimodal Interactive Omni-Avatar), an end-to-end framework composed of five specialized modules: Thinker, Talker, Face Animator, Body Animator, and Renderer. This unified architecture integrates cognitive reasoning with real-time multimodal embodiment to enable fluid, consistent interaction. Furthermore, we establish a new benchmark to rigorously evaluate the capabilities of interactive intelligence. Extensive experiments demonstrate that our framework achieves superior performance compared to state-of-the-art methods across all evaluated dimensions. Together, these contributions move digital humans beyond superficial imitation toward intelligent interaction.

Method: CARe uses two main operations: 1) Crop operation exploits retinal physiological structure to crop a sub-image from wfCFP with roughly aligned FoV to OCTA, 2) Alignment module uses double-fitting with RANSAC and polynomial-based coordinate fitting for spatial transformation.

Result: Extensive experiments on a new test set of 60 OCTA-wfCFP pairs demonstrate the viability of CARe for cross-modal fundus image registration with large FoV disparity.

Conclusion: CARe provides a simple yet effective solution for challenging cross-modal fundus image registration with large field-of-view disparity, enabling reuse of existing small-FoV-disparity methods through intelligent cropping and improved alignment techniques.

Abstract: Previous work on cross-modal fundus image registration (CMFIR) assumes small cross-modal Field-of-View (FoV) disparity. By contrast, this paper is targeted at a more challenging scenario with large FoV disparity, to which directly applying current methods fails. We propose Crop and Alignment for cross-modal fundus image Registration(CARe), a very simple yet effective method. Specifically, given an OCTA with smaller FoV as a source image and a wide-field color fundus photograph (wfCFP) as a target image, our Crop operation exploits the physiological structure of the retina to crop from the target image a sub-image with its FoV roughly aligned with that of the source. This operation allows us to re-purpose the previous small-FoV-disparity oriented methods for subsequent image registration. Moreover, we improve spatial transformation by a double-fitting based Alignment module that utilizes the classical RANSAC algorithm and polynomial-based coordinate fitting in a sequential manner. Extensive experiments on a newly developed test set of 60 OCTA-wfCFP pairs verify the viability of CARe for CMFIR.

Method: Proposes CogDoc, a unified coarse-to-fine thinking framework with two phases: 1) “Fast Reading” phase for scalable information localization at low resolution, and 2) “Focused Thinking” phase for deep reasoning at high resolution. Uses Direct Reinforcement Learning approach for post-training, avoiding policy conflicts found in RL with SFT initialization.

Result: The 7B model achieves state-of-the-art performance within its parameter class, notably surpassing significantly larger proprietary models (e.g., GPT-4o) on challenging, visually rich document benchmarks. Direct RL approach outperforms RL with SFT initialization.

Conclusion: CogDoc successfully bridges the scalability-fidelity trade-off in document reasoning through a human-like cognitive framework, demonstrating that direct reinforcement learning is more effective than SFT-initialized RL for this unified thinking approach.

Abstract: Current document reasoning paradigms are constrained by a fundamental trade-off between scalability (processing long-context documents) and fidelity (capturing fine-grained, multimodal details). To bridge this gap, we propose CogDoc, a unified coarse-to-fine thinking framework that mimics human cognitive processes: a low-resolution “Fast Reading” phase for scalable information localization,followed by a high-resolution “Focused Thinking” phase for deep reasoning. We conduct a rigorous investigation into post-training strategies for the unified thinking framework, demonstrating that a Direct Reinforcement Learning (RL) approach outperforms RL with Supervised Fine-Tuning (SFT) initialization. Specifically, we find that direct RL avoids the “policy conflict” observed in SFT. Empirically, our 7B model achieves state-of-the-art performance within its parameter class, notably surpassing significantly larger proprietary models (e.g., GPT-4o) on challenging, visually rich document benchmarks.

Method: SSMT-Net uses a Semi-Supervised Multi-Task Transformer-based Network with two phases: 1) Initial unsupervised phase leveraging unlabeled data to enhance Transformer-centric encoder feature extraction capability, and 2) Supervised phase jointly optimizing nodule segmentation, gland segmentation, and nodule size estimation, integrating both local and global contextual features.

Result: Extensive evaluations on TN3K and DDTI datasets demonstrate that SSMT-Net outperforms state-of-the-art methods with higher accuracy and robustness.

Conclusion: SSMT-Net shows potential for real-world clinical applications in thyroid nodule diagnosis and treatment planning through its effective semi-supervised multi-task learning approach.

Abstract: Accurate thyroid nodule segmentation in ultrasound images is critical for diagnosis and treatment planning. However, ambiguous boundaries between nodules and surrounding tissues, size variations, and the scarcity of annotated ultrasound data pose significant challenges for automated segmentation. Existing deep learning models struggle to incorporate contextual information from the thyroid gland and generalize effectively across diverse cases. To address these challenges, we propose SSMT-Net, a Semi-Supervised Multi-Task Transformer-based Network that leverages unlabeled data to enhance Transformer-centric encoder feature extraction capability in an initial unsupervised phase. In the supervised phase, the model jointly optimizes nodule segmentation, gland segmentation, and nodule size estimation, integrating both local and global contextual features. Extensive evaluations on the TN3K and DDTI datasets demonstrate that SSMT-Net outperforms state-of-the-art methods, with higher accuracy and robustness, indicating its potential for real-world clinical applications.

Method: Multi-stage training to learn unified embedding space; Generalized Motion Adaptation Module for independent training; adaptive fusion strategy to balance heterogeneous conditioning signals; diffusion-based approach.

Result: Outperforms prior methods in both co-speech gesture generation and object-interaction synthesis; yields highly realistic, object-aware full-body motions with enhanced realism, flexibility, and control.

Conclusion: InteracTalker successfully unifies previously separate tasks of speech-driven gesture generation and object-interaction synthesis, enabling more natural and comprehensive human motion generation.

Abstract: Generating realistic human motions that naturally respond to both spoken language and physical objects is crucial for interactive digital experiences. Current methods, however, address speech-driven gestures or object interactions independently, limiting real-world applicability due to a lack of integrated, comprehensive datasets. To overcome this, we introduce InteracTalker, a novel framework that seamlessly integrates prompt-based object-aware interactions with co-speech gesture generation. We achieve this by employing a multi-stage training process to learn a unified motion, speech, and prompt embedding space. To support this, we curate a rich human-object interaction dataset, formed by augmenting an existing text-to-motion dataset with detailed object interaction annotations. Our framework utilizes a Generalized Motion Adaptation Module that enables independent training, adapting to the corresponding motion condition, which is then dynamically combined during inference. To address the imbalance between heterogeneous conditioning signals, we propose an adaptive fusion strategy, which dynamically reweights the conditioning signals during diffusion sampling. InteracTalker successfully unifies these previously separate tasks, outperforming prior methods in both co-speech gesture generation and object-interaction synthesis, outperforming gesture-focused diffusion methods, yielding highly realistic, object-aware full-body motions with enhanced realism, flexibility, and control.

Method: Confidence-Aware Asymmetric Learning (CAL) framework with two components: Confidence-Aware Consistency Regularization (CCR) to mitigate pseudo-label bias by dynamic loss scaling, and Asymmetric Confidence Reinforcement (ACR) to separately calibrate confidence for known/novel classes. Plus Dynamic Prototype Pruning (DPP) to automatically estimate number of novel forgery types.

Result: Extensive experiments show CAL consistently outperforms previous methods, achieving state-of-the-art performance on both standard OW-DFA benchmark and extended benchmark with advanced manipulations.

Conclusion: CAL framework effectively addresses confidence skew and unrealistic prior assumptions in OW-DFA, providing robust attribution for both known and unknown forgeries with enhanced scalability to real-world scenarios.

Abstract: The proliferation of synthetic facial imagery has intensified the need for robust Open-World DeepFake Attribution (OW-DFA), which aims to attribute both known and unknown forgeries using labeled data for known types and unlabeled data containing a mixture of known and novel types. However, existing OW-DFA methods face two critical limitations: 1) A confidence skew that leads to unreliable pseudo-labels for novel forgeries, resulting in biased training. 2) An unrealistic assumption that the number of unknown forgery types is known a priori. To address these challenges, we propose a Confidence-Aware Asymmetric Learning (CAL) framework, which adaptively balances model confidence across known and novel forgery types. CAL mainly consists of two components: Confidence-Aware Consistency Regularization (CCR) and Asymmetric Confidence Reinforcement (ACR). CCR mitigates pseudo-label bias by dynamically scaling sample losses based on normalized confidence, gradually shifting the training focus from high- to low-confidence samples. ACR complements this by separately calibrating confidence for known and novel classes through selective learning on high-confidence samples, guided by their confidence gap. Together, CCR and ACR form a mutually reinforcing loop that significantly improves the model’s OW-DFA performance. Moreover, we introduce a Dynamic Prototype Pruning (DPP) strategy that automatically estimates the number of novel forgery types in a coarse-to-fine manner, removing the need for unrealistic prior assumptions and enhancing the scalability of our methods to real-world OW-DFA scenarios. Extensive experiments on the standard OW-DFA benchmark and a newly extended benchmark incorporating advanced manipulations demonstrate that CAL consistently outperforms previous methods, achieving new state-of-the-art performance on both known and novel forgery attribution.

Method: Progressive conditioned scale-shift recalibration (PCSR) uses Domain Separation Networks to extract domain shift features and Factor Generator Networks to predict scale/shift parameters for self-attention recalibration at each transformer layer.

Result: Significant improvement in online test-time domain adaptation performance with up to 3.9% accuracy gain on ImageNet-C dataset.

Conclusion: The proposed PCSR method effectively addresses transformer performance degradation in new domains through progressive self-attention recalibration during online inference.

Abstract: Online test-time adaptation aims to dynamically adjust a network model in real-time based on sequential input samples during the inference stage. In this work, we find that, when applying a transformer network model to a new target domain, the Query, Key, and Value features of its self-attention module often change significantly from those in the source domain, leading to substantial performance degradation of the transformer model. To address this important issue, we propose to develop a new approach to progressively recalibrate the self-attention at each layer using a local linear transform parameterized by conditioned scale and shift factors. We consider the online model adaptation from the source domain to the target domain as a progressive domain shift separation process. At each transformer network layer, we learn a Domain Separation Network to extract the domain shift feature, which is used to predict the scale and shift parameters for self-attention recalibration using a Factor Generator Network. These two lightweight networks are adapted online during inference. Experimental results on benchmark datasets demonstrate that the proposed progressive conditioned scale-shift recalibration (PCSR) method is able to significantly improve the online test-time domain adaptation performance by a large margin of up to 3.9% in classification accuracy on the ImageNet-C dataset.

Method: Scone integrates composition and distinction through a unified understanding-generation approach. It uses an understanding expert as a semantic bridge to convey information and guide the generation expert. A two-stage training scheme first learns composition, then enhances distinction through semantic alignment and attention-based masking.

Result: Scone outperforms existing open-source models in both composition and distinction tasks on two benchmarks. The authors also introduce SconeEval, a benchmark for evaluating both composition and distinction across diverse scenarios.

Conclusion: Scone successfully addresses the distinction problem in multi-subject image generation through a unified understanding-generation approach with semantic bridging and two-stage training, demonstrating superior performance over existing models.

Abstract: Subject-driven image generation has advanced from single- to multi-subject composition, while neglecting distinction, the ability to identify and generate the correct subject when inputs contain multiple candidates. This limitation restricts effectiveness in complex, realistic visual settings. We propose Scone, a unified understanding-generation method that integrates composition and distinction. Scone enables the understanding expert to act as a semantic bridge, conveying semantic information and guiding the generation expert to preserve subject identity while minimizing interference. A two-stage training scheme first learns composition, then enhances distinction through semantic alignment and attention-based masking. We also introduce SconeEval, a benchmark for evaluating both composition and distinction across diverse scenarios. Experiments demonstrate that Scone outperforms existing open-source models in composition and distinction tasks on two benchmarks. Our model, benchmark, and training data are available at: https://github.com/Ryann-Ran/Scone.

Method: β-CLIP uses cross-attention to dynamically pool image patches for each text granularity level, producing contextualized visual embeddings. It introduces β-CAL loss that parameterizes the trade-off between strict query-specific matching and relaxed intra-image contextualization.

Result: Achieves 91.8% T2I and 92.3% I2T at R@1 on Urban1K, and 30.9% on FG-OVD (Hard), setting state-of-the-art among methods trained without hard negatives.

Conclusion: β-CLIP establishes a robust, adaptive baseline for dense vision-language correspondence and significantly improves fine-grained alignment compared to CLIP.

Abstract: CLIP achieves strong zero-shot image-text retrieval by aligning global vision and text representations, yet it falls behind on fine-grained tasks even when fine-tuned on long, detailed captions. In this work, we propose $β$-CLIP, a multi-granular text-conditioned contrastive learning framework designed to achieve hierarchical alignment between multiple textual granularities-from full captions to sentences and phrases-and their corresponding visual regions. For each level of granularity, $β$-CLIP utilizes cross-attention to dynamically pool image patches, producing contextualized visual embeddings. To address the semantic overlap inherent in this hierarchy, we introduce the $β$-Contextualized Contrastive Alignment Loss ($β$-CAL). This objective parameterizes the trade-off between strict query-specific matching and relaxed intra-image contextualization, supporting both soft Cross-Entropy and hard Binary Cross-Entropy formulations. Through extensive experiments, we demonstrate that $β$-CLIP significantly improves dense alignment: achieving 91.8% T2I 92.3% I2T at R@1 on Urban1K and 30.9% on FG-OVD (Hard), setting state-of-the-art among methods trained without hard negatives. $β$-CLIP establishes a robust, adaptive baseline for dense vision-language correspondence. The code and models are released at https://github.com/fzohra/B-CLIP.

Method: 1) Decompose human body into 5 parts and detect credible (clearly visible) parts per frame; 2) Encode credible parts into latent tokens using part-aware VAE; 3) Use robust part-level masked generation model to predict masked credible parts while ignoring noisy ones.

Result: Method outperforms baselines on both clean and noisy datasets in motion quality, semantic consistency, and diversity. Also introduces K700-M benchmark with ~200k real-world motion sequences for evaluation.

Conclusion: The proposed robust part-aware approach effectively handles noisy web video data for human motion generation, balancing data scale and quality while introducing a challenging new benchmark for the field.

Abstract: Extracting human motion from large-scale web videos offers a scalable solution to the data scarcity issue in character animation. However, some human parts in many video frames cannot be seen due to off-screen captures or occlusions. It brings a dilemma: discarding the data missing any part limits scale and diversity, while retaining it compromises data quality and model performance. To address this problem, we propose leveraging credible part-level data extracted from videos to enhance motion generation via a robust part-aware masked autoregression model. First, we decompose a human body into five parts and detect the parts clearly seen in a video frame as “credible”. Second, the credible parts are encoded into latent tokens by our proposed part-aware variational autoencoder. Third, we propose a robust part-level masked generation model to predict masked credible parts, while ignoring those noisy parts. In addition, we contribute K700-M, a challenging new benchmark comprising approximately 200k real-world motion sequences, for evaluation. Experimental results indicate that our method successfully outperforms baselines on both clean and noisy datasets in terms of motion quality, semantic consistency and diversity. Project page: https://boyuaner.github.io/ropar-main/

Method: Integrates depth images from four directions to restore 3D human models. Uses hierarchical matching with global and fine registration for noise and occlusion restoration. Applies Adaptive Vertex Reduction to maintain mesh resolution and shape reliability, and uses Level of Detail ensemble for accurate and stable spinal angle estimation.

Result: Achieves high-precision 3D spine registration estimation without relying on training data or complex neural network models. Verification confirms improvement in matching quality.

Conclusion: The proposed method effectively compensates for multi-image method shortcomings while solving single-image method limitations, providing a practical solution for 3D body posture analysis and spine center line estimation.

Abstract: The spinal angle is an important indicator of body balance. It is important to restore the 3D shape of the human body and estimate the spine center line. Existing mul-ti-image-based body restoration methods require expensive equipment and complex pro-cedures, and single image-based body restoration methods have limitations in that it is difficult to accurately estimate the internal structure such as the spine center line due to occlusion and viewpoint limitation. This study proposes a method to compensate for the shortcomings of the multi-image-based method and to solve the limitations of the sin-gle-image method. We propose a 3D body posture analysis system that integrates depth images from four directions to restore a 3D human model and automatically estimate the spine center line. Through hierarchical matching of global and fine registration, restora-tion to noise and occlusion is performed. Also, the Adaptive Vertex Reduction is applied to maintain the resolution and shape reliability of the mesh, and the accuracy and stabil-ity of spinal angle estimation are simultaneously secured by using the Level of Detail en-semble. The proposed method achieves high-precision 3D spine registration estimation without relying on training data or complex neural network models, and the verification confirms the improvement of matching quality.

Method: 1) Generate 4D occupancy as physics-informed foundation; 2) Propose VAE with tri-plane representation to compress high-resolution occupancy (58% size reduction); 3) Introduce Mutual Control Attention (MCA) to model control influence on occupancy evolution; 4) Joint end-to-end training of VAE and prediction module; 5) Normalized Multi-View Attention for video generation guided by 4D occupancy.

Result: 7.2% improvement in forecasting mIoU at 41 FPS inference speed with only 3.47M parameters. 20.7% reduction in FVD for video quality. Enables highly controllable, multi-view consistent, and physics-aware driving video generation.

Conclusion: GenieDrive successfully addresses limitations of existing methods by using 4D occupancy as physics foundation, achieving significant improvements in forecasting accuracy and video quality while maintaining efficiency, enabling better physics-aware driving video generation.

Abstract: Physics-aware driving world model is essential for drive planning, out-of-distribution data synthesis, and closed-loop evaluation. However, existing methods often rely on a single diffusion model to directly map driving actions to videos, which makes learning difficult and leads to physically inconsistent outputs. To overcome these challenges, we propose GenieDrive, a novel framework designed for physics-aware driving video generation. Our approach starts by generating 4D occupancy, which serves as a physics-informed foundation for subsequent video generation. 4D occupancy contains rich physical information, including high-resolution 3D structures and dynamics. To facilitate effective compression of such high-resolution occupancy, we propose a VAE that encodes occupancy into a latent tri-plane representation, reducing the latent size to only 58% of that used in previous methods. We further introduce Mutual Control Attention (MCA) to accurately model the influence of control on occupancy evolution, and we jointly train the VAE and the subsequent prediction module in an end-to-end manner to maximize forecasting accuracy. Together, these designs yield a 7.2% improvement in forecasting mIoU at an inference speed of 41 FPS, while using only 3.47 M parameters. Additionally, a Normalized Multi-View Attention is introduced in the video generation model to generate multi-view driving videos with guidance from our 4D occupancy, significantly improving video quality with a 20.7% reduction in FVD. Experiments demonstrate that GenieDrive enables highly controllable, multi-view consistent, and physics-aware driving video generation.

Method: Introduces FysicsWorld benchmark with 16 primary tasks and 3,268 curated samples from over 40 sources. Proposes Cross-Modal Complementarity Screening (CMCS) strategy integrated in a systematic data construction framework for omni-modal data creation.

Result: Comprehensive evaluation of over 30 state-of-the-art baselines reveals performance disparities and limitations across models in understanding, generation, and reasoning. The benchmark establishes unified foundations for evaluating full-modality architectures.

Conclusion: FysicsWorld bridges existing gaps in multimodal evaluation by providing the first unified full-modality benchmark with bidirectional input-output support, enabling comprehensive assessment of next-generation multimodal architectures.

Abstract: Despite rapid progress in multimodal large language models (MLLMs) and emerging omni-modal architectures, current benchmarks remain limited in scope and integration, suffering from incomplete modality coverage, restricted interaction to text-centric outputs, and weak interdependence and complementarity among modalities. To bridge these gaps, we introduce FysicsWorld, the first unified full-modality benchmark that supports bidirectional input-output across image, video, audio, and text, enabling comprehensive any-to-any evaluation across understanding, generation, and reasoning. FysicsWorld encompasses 16 primary tasks and 3,268 curated samples, aggregated from over 40 high-quality sources and covering a rich set of open-domain categories with diverse question types. We also propose the Cross-Modal Complementarity Screening (CMCS) strategy integrated in a systematic data construction framework that produces omni-modal data for spoken interaction and fusion-dependent cross-modal reasoning. Through a comprehensive evaluation of over 30 state-of-the-art baselines, spanning MLLMs, modality-specific models, unified understanding-generation models, and omni-modal language models, FysicsWorld exposes the performance disparities and limitations across models in understanding, generation, and reasoning. Our benchmark establishes a unified foundation and strong baselines for evaluating and advancing next-generation full-modality architectures.

Method: CoRe3D uses a spatially grounded reasoning representation that decomposes 3D latent space into localized regions, enabling compositional and procedural reasoning over geometry. It tightly couples semantic chain-of-thought inference with structured spatial reasoning.

Result: The framework produces 3D outputs with strong local consistency and faithful alignment with linguistic descriptions, enabling high-level language intent to directly guide low-level 3D content formation.

Conclusion: CoRe3D successfully extends reasoning-centric approaches to 3D domains, creating a unified framework that improves 3D understanding and generation through explicit semantic-spatial reasoning mechanisms.

Abstract: Recent advances in large multimodal models suggest that explicit reasoning mechanisms play a critical role in improving model reliability, interpretability, and cross-modal alignment. While such reasoning-centric approaches have been proven effective in language and vision tasks, their extension to 3D remains underdeveloped. CoRe3D introduces a unified 3D understanding and generation reasoning framework that jointly operates over semantic and spatial abstractions, enabling high-level intent inferred from language to directly guide low-level 3D content formation. Central to this design is a spatially grounded reasoning representation that decomposes 3D latent space into localized regions, allowing the model to reason over geometry in a compositional and procedural manner. By tightly coupling semantic chain-of-thought inference with structured spatial reasoning, CoRe3D produces 3D outputs that exhibit strong local consistency and faithful alignment with linguistic descriptions.

Method: Proposes Fast-2DGS with two components: 1) Deep Gaussian Prior - conditional network capturing spatial distribution of Gaussian primitives based on image complexity, 2) Attribute regression network predicting dense Gaussian properties. Disentangled architecture enables single-pass reconstruction.

Result: Achieves high-quality reconstruction in single forward pass with minimal fine-tuning. Significantly reduces computational cost while maintaining visual quality. Enables faster convergence compared to existing methods (>10s).

Conclusion: Fast-2DGS bridges gap between quality and efficiency in 2D Gaussian Splatting, making it more practical for industry deployment by reducing computational overhead without compromising reconstruction quality.

Abstract: As generative models become increasingly capable of producing high-fidelity visual content, the demand for efficient, interpretable, and editable image representations has grown substantially. Recent advances in 2D Gaussian Splatting (2DGS) have emerged as a promising solution, offering explicit control, high interpretability, and real-time rendering capabilities (>1000 FPS). However, high-quality 2DGS typically requires post-optimization. Existing methods adopt random or heuristics (e.g., gradient maps), which are often insensitive to image complexity and lead to slow convergence (>10s). More recent approaches introduce learnable networks to predict initial Gaussian configurations, but at the cost of increased computational and architectural complexity. To bridge this gap, we present Fast-2DGS, a lightweight framework for efficient Gaussian image representation. Specifically, we introduce Deep Gaussian Prior, implemented as a conditional network to capture the spatial distribution of Gaussian primitives under different complexities. In addition, we propose an attribute regression network to predict dense Gaussian properties. Experiments demonstrate that this disentangled architecture achieves high-quality reconstruction in a single forward pass, followed by minimal fine-tuning. More importantly, our approach significantly reduces computational cost without compromising visual quality, bringing 2DGS closer to industry-ready deployment.

Method: Proposes Long-term Spatio-Temporal Enhanced Context (L-STEC) method: 1) Extends reference chain with LSTM to capture long-term dependencies, 2) Incorporates warped spatial context from pixel domain, and 3) Fuses spatio-temporal information through multi-receptive field network to preserve reference details.

Result: Achieves 37.01% bitrate savings in PSNR and 31.65% in MS-SSIM compared to DCVC-TCM, outperforming both VTM-17.0 and DCVC-FM, establishing new state-of-the-art performance.

Conclusion: L-STEC significantly improves video compression by enriching contextual information through long-term spatio-temporal enhancement, effectively addressing limitations of existing condition-based frameworks.

Abstract: Neural Video Compression has emerged in recent years, with condition-based frameworks outperforming traditional codecs. However, most existing methods rely solely on the previous frame’s features to predict temporal context, leading to two critical issues. First, the short reference window misses long-term dependencies and fine texture details. Second, propagating only feature-level information accumulates errors over frames, causing prediction inaccuracies and loss of subtle textures. To address these, we propose the Long-term Spatio-Temporal Enhanced Context (L-STEC) method. We first extend the reference chain with LSTM to capture long-term dependencies. We then incorporate warped spatial context from the pixel domain, fusing spatio-temporal information through a multi-receptive field network to better preserve reference details. Experimental results show that L-STEC significantly improves compression by enriching contextual information, achieving 37.01% bitrate savings in PSNR and 31.65% in MS-SSIM compared to DCVC-TCM, outperforming both VTM-17.0 and DCVC-FM and establishing new state-of-the-art performance.

Method: Proposes DrivePI, a spatial-aware 4D MLLM that integrates point clouds, multi-view images, and language instructions in a unified architecture. Uses a data engine to generate text-occupancy and text-flow QA pairs for 4D spatial understanding. Performs spatial understanding, 3D occupancy, occupancy flow, and planning through end-to-end optimization.

Result: With only a 0.5B Qwen2.5 backbone, DrivePI matches or exceeds both existing VLA models and specialized VA models. Outperforms OpenDriveVLA-7B by 2.5% mean accuracy on nuScenes-QA, reduces collision rate by 70% over ORION, surpasses FB-OCC by 10.3 RayIoU for 3D occupancy, reduces mAVE from 0.591 to 0.509 for occupancy flow, and achieves 32% lower L2 error than VAD for planning.

Conclusion: DrivePI demonstrates that a unified spatial-aware 4D MLLM can effectively handle multiple autonomous driving tasks simultaneously, achieving state-of-the-art performance with relatively small model size, showing promise for integrated perception-prediction-planning systems.

Abstract: Although multi-modal large language models (MLLMs) have shown strong capabilities across diverse domains, their application in generating fine-grained 3D perception and prediction outputs in autonomous driving remains underexplored. In this paper, we propose DrivePI, a novel spatial-aware 4D MLLM that serves as a unified Vision-Language-Action (VLA) framework that is also compatible with vision-action (VA) models. Our method jointly performs spatial understanding, 3D perception (i.e., 3D occupancy), prediction (i.e., occupancy flow), and planning (i.e., action outputs) in parallel through end-to-end optimization. To obtain both precise geometric information and rich visual appearance, our approach integrates point clouds, multi-view images, and language instructions within a unified MLLM architecture. We further develop a data engine to generate text-occupancy and text-flow QA pairs for 4D spatial understanding. Remarkably, with only a 0.5B Qwen2.5 model as MLLM backbone, DrivePI as a single unified model matches or exceeds both existing VLA models and specialized VA models. Specifically, compared to VLA models, DrivePI outperforms OpenDriveVLA-7B by 2.5% mean accuracy on nuScenes-QA and reduces collision rate by 70% over ORION (from 0.37% to 0.11%) on nuScenes. Against specialized VA models, DrivePI surpasses FB-OCC by 10.3 RayIoU for 3D occupancy on OpenOcc, reduces the mAVE from 0.591 to 0.509 for occupancy flow on OpenOcc, and achieves 32% lower L2 error than VAD (from 0.72m to 0.49m) for planning on nuScenes. Code will be available at https://github.com/happinesslz/DrivePI

Method: Proposes a novel framework with new learning strategies and losses specifically designed for Contrastive Analysis. The approach can be adapted to both GAN and Diffusion models to learn both common and salient factors while ensuring relevant separation and preserving generation quality.

Result: The framework demonstrates superior separation ability and image quality synthesis compared to prior methods across diverse datasets including human faces, animal images, and medical scans.

Conclusion: The proposed Contrastive Analysis framework successfully addresses the understudied problem of separating common vs. salient generative factors using only dataset-level supervision, outperforming existing methods in both separation quality and image synthesis.

Abstract: Recent advancements in image synthesis have enabled high-quality image generation and manipulation. Most works focus on: 1) conditional manipulation, where an image is modified conditioned on a given attribute, or 2) disentangled representation learning, where each latent direction should represent a distinct semantic attribute. In this paper, we focus on a different and less studied research problem, called Contrastive Analysis (CA). Given two image datasets, we want to separate the common generative factors, shared across the two datasets, from the salient ones, specific to only one dataset. Compared to existing methods, which use attributes as supervised signals for editing (e.g., glasses, gender), the proposed method is weaker, since it only uses the dataset signal. We propose a novel framework for CA, that can be adapted to both GAN and Diffusion models, to learn both common and salient factors. By defining new and well-adapted learning strategies and losses, we ensure a relevant separation between common and salient factors, preserving a high-quality generation. We evaluate our approach on diverse datasets, covering human faces, animal images and medical scans. Our framework demonstrates superior separation ability and image quality synthesis compared to prior methods.

Method: Unified transformer architecture that processes 3D point cloud patches and language tokens as a single sequence; structured patchification/tokenization scheme preserving spatial context; three-stage training curriculum from object-level recognition to scene-level spatial reasoning.

Result: Establishes new state-of-the-art performance across comprehensive 3D understanding tasks (object recognition, captioning, spatial reasoning) and demonstrates robust scaling properties as model size and training data increase.

Conclusion: Lemon provides a unified foundation for advancing 3D spatial intelligence in real-world applications by enabling early spatial-linguistic fusion, eliminating redundant encoders, and improving parameter efficiency.

Abstract: Scaling large multimodal models (LMMs) to 3D understanding poses unique challenges: point cloud data is sparse and irregular, existing models rely on fragmented architectures with modality-specific encoders, and training pipelines often suffer from instability and poor scalability. We introduce Lemon, a unified transformer architecture that addresses these challenges by jointly processing 3D point cloud patches and language tokens as a single sequence. Unlike prior work that relies on modality-specific encoders and cross-modal alignment modules, this design enables early spatial-linguistic fusion, eliminates redundant encoders, improves parameter efficiency, and supports more effective model scaling. To handle the complexity of 3D data, we develop a structured patchification and tokenization scheme that preserves spatial context, and a three-stage training curriculum that progressively builds capabilities from object-level recognition to scene-level spatial reasoning. Lemon establishes new state-of-the-art performance across comprehensive 3D understanding and reasoning tasks, from object recognition and captioning to spatial reasoning in 3D scenes, while demonstrating robust scaling properties as model size and training data increase. Our work provides a unified foundation for advancing 3D spatial intelligence in real-world applications.

Method: Systematic evaluation of adaptation strategies for CoCa visual backbone: from training-free hybrid prototyping to deep parameter adaptation via Low-Rank Adaptation (LoRA). Examines augmentation effects, hybrid objectives (Supervised Contrastive loss + Cross-Entropy), and sensitivity to data scarcity.

Result: Identified “augmentation divergence” - strong augmentation harms linear probing but stabilizes LoRA fine-tuning. Hybrid objectives with SupCon loss consistently outperform standard Cross-Entropy across shot counts. Characterized sensitivity to data scarcity and provided empirical reference settings for scaling regularization, rank, and sampling strategies.

Conclusion: The study provides practical guidance for efficient adaptation of generative-contrastive foundation models like CoCa in few-shot scenarios, highlighting the importance of tailored strategies that account for the model’s unique architecture and the challenges of extreme data scarcity.

Abstract: Large-scale multimodal foundation models, particularly Contrastive Captioners (CoCa), have achieved state-of-the-art results by unifying contrastive alignment with generative captioning. While zero-shot transfer capabilities are well-documented, the adaptation of these generative-contrastive hybrids to downstream tasks with extreme data scarcity (few-shot learning) remains under-explored. Existing literature predominantly focuses on dual-encoder architectures like CLIP, leaving a gap in understanding how CoCa’s distinct latent space responds to parameter-efficient fine-tuning (PEFT). This paper presents a comprehensive empirical study on adapting the CoCa visual backbone for few-shot image classification. We systematically evaluate a hierarchy of strategies, ranging from training-free hybrid prototyping to deep parameter adaptation via Low-Rank Adaptation (LoRA). First, we identify an “augmentation divergence”: while strong data augmentation degrades the performance of linear probing in low-shot settings, it is essential for stabilizing LoRA fine-tuning. We also demonstrate that hybrid objectives incorporating Supervised Contrastive (SupCon) loss yield consistent performance improvements over standard Cross-Entropy across varying shot counts. Crucially, we characterize the sensitivity of training configurations to data scarcity, providing empirical reference settings for scaling regularization, rank, and sampling strategies to facilitate the efficient adaptation of generative-contrastive foundation models.

Method: 1) End-to-end Transformer architecture integrating object lists as denoising queries with learnable queries; 2) Deformable Gaussian masks derived from object list positional/size priors to focus attention on target areas; 3) Generation of pseudo object lists from ground-truth bounding boxes by simulating state noise, false positives, and negatives for training.

Result: Substantial performance improvements over vision-based baseline on nuScenes dataset. Demonstrates generalization capability across diverse noise levels of simulated object lists and real detectors.

Conclusion: First work to conduct cross-level fusion, successfully integrating abstract object list information with raw camera images for 3D object detection, with improved performance and robustness to noise.

Abstract: In automotive sensor fusion systems, smart sensors and Vehicle-to-Everything (V2X) modules are commonly utilized. Sensor data from these systems are typically available only as processed object lists rather than raw sensor data from traditional sensors. Instead of processing other raw data separately and then fusing them at the object level, we propose an end-to-end cross-level fusion concept with Transformer, which integrates highly abstract object list information with raw camera images for 3D object detection. Object lists are fed into a Transformer as denoising queries and propagated together with learnable queries through the latter feature aggregation process. Additionally, a deformable Gaussian mask, derived from the positional and size dimensional priors from the object lists, is explicitly integrated into the Transformer decoder. This directs attention toward the target area of interest and accelerates model training convergence. Furthermore, as there is no public dataset containing object lists as a standalone modality, we propose an approach to generate pseudo object lists from ground-truth bounding boxes by simulating state noise and false positives and negatives. As the first work to conduct cross-level fusion, our approach shows substantial performance improvements over the vision-based baseline on the nuScenes dataset. It demonstrates its generalization capability over diverse noise levels of simulated object lists and real detectors.

Method: Adapt 2D foundation models to 3D classification by adding lightweight plugins (~1M parameters per task) on top of a single frozen backbone. Supports multi-view inputs, auxiliary pixel-level supervision, and interpretable heatmap generation.

Result: Achieves state-of-the-art performance across 12 diverse medical tasks covering various pathologies, anatomies, and modalities. Wins 1st place in VLM3D challenge. Key insights: effective adaptation unlocks FM potential, general-purpose FMs can match medical-specific FMs, and 2D-based methods surpass 3D architectures for 3D classification.

Conclusion: Demonstrates feasibility of using a single scalable framework for diverse 3D medical classification tasks, eliminating need for separate task-specific models. Shows 2D foundation models can be effectively adapted to 3D medical imaging with lightweight plugins.

Abstract: 3D medical image classification is essential for modern clinical workflows. Medical foundation models (FMs) have emerged as a promising approach for scaling to new tasks, yet current research suffers from three critical pitfalls: data-regime bias, suboptimal adaptation, and insufficient task coverage. In this paper, we address these pitfalls and introduce AnyMC3D, a scalable 3D classifier adapted from 2D FMs. Our method scales efficiently to new tasks by adding only lightweight plugins (about 1M parameters per task) on top of a single frozen backbone. This versatile framework also supports multi-view inputs, auxiliary pixel-level supervision, and interpretable heatmap generation. We establish a comprehensive benchmark of 12 tasks covering diverse pathologies, anatomies, and modalities, and systematically analyze state-of-the-art 3D classification techniques. Our analysis reveals key insights: (1) effective adaptation is essential to unlock FM potential, (2) general-purpose FMs can match medical-specific FMs if properly adapted, and (3) 2D-based methods surpass 3D architectures for 3D classification. For the first time, we demonstrate the feasibility of achieving state-of-the-art performance across diverse applications using a single scalable framework (including 1st place in the VLM3D challenge), eliminating the need for separate task-specific models.

Method: Qonvolutions modify convolution operations by convolving low-frequency signals with queries (such as coordinates) to enhance learning of intricate high-frequency signals. This leverages the neighborhood properties of convolution to better capture high-frequency details.

Result: Qonvolutions improve performance across multiple high-frequency learning tasks: 1D regression, 2D super-resolution, 2D image regression, and novel view synthesis. When combined with Gaussian splatting for NVS, it achieves state-of-the-art performance on real-world complex scenes, outperforming powerful radiance field models in image quality.

Conclusion: Qonvolutions provide a simple yet powerful approach to enhance high-frequency signal learning in neural networks, demonstrating significant improvements across various computer vision and graphics applications, particularly in novel view synthesis where it sets new state-of-the-art performance.

Abstract: Accurately learning high-frequency signals is a challenge in computer vision and graphics, as neural networks often struggle with these signals due to spectral bias or optimization difficulties. While current techniques like Fourier encodings have made great strides in improving performance, there remains scope for improvement when presented with high-frequency information. This paper introduces Queried-Convolutions (Qonvolutions), a simple yet powerful modification using the neighborhood properties of convolution. Qonvolution convolves a low-frequency signal with queries (such as coordinates) to enhance the learning of intricate high-frequency signals. We empirically demonstrate that Qonvolutions enhance performance across a variety of high-frequency learning tasks crucial to both the computer vision and graphics communities, including 1D regression, 2D super-resolution, 2D image regression, and novel view synthesis (NVS). In particular, by combining Gaussian splatting with Qonvolutions for NVS, we showcase state-of-the-art performance on real-world complex scenes, even outperforming powerful radiance field models on image quality.

Method: Proposes Predictive Sample Assignment (PSA) with: 1) dual-threshold ternary sample assignment based on predictive energy scores to create cleaner ID/OOD sets plus discard set, 2) concept contrastive representation learning to increase ID/OOD separation, and 3) retraining strategy for better fitting.

Result: Experiments on two standard SCOOD benchmarks show the approach significantly outperforms state-of-the-art methods.

Conclusion: PSA framework effectively addresses the noisy sample problem in SCOOD through better sample filtering and representation learning, achieving superior OOD detection performance.

Abstract: Semantically coherent out-of-distribution detection (SCOOD) is a recently proposed realistic OOD detection setting: given labeled in-distribution (ID) data and mixed in-distribution and out-of-distribution unlabeled data as the training data, SCOOD aims to enable the trained model to accurately identify OOD samples in the testing data. Current SCOOD methods mainly adopt various clustering-based in-distribution sample filtering (IDF) strategies to select clean ID samples from unlabeled data, and take the remaining samples as auxiliary OOD data, which inevitably introduces a large number of noisy samples in training. To address the above issue, we propose a concise SCOOD framework based on predictive sample assignment (PSA). PSA includes a dual-threshold ternary sample assignment strategy based on the predictive energy score that can significantly improve the purity of the selected ID and OOD sample sets by assigning unconfident unlabeled data to an additional discard sample set, and a concept contrastive representation learning loss to further expand the distance between ID and OOD samples in the representation space to assist ID/OOD discrimination. In addition, we also introduce a retraining strategy to help the model fully fit the selected auxiliary ID/OOD samples. Experiments on two standard SCOOD benchmarks demonstrate that our approach outperforms the state-of-the-art methods by a significant margin.

Method: Two modules: 1) Loss Sharpness Penalty (LSP) enhances parameter robustness by minimizing worst-case loss sharpness to suppress trivial features and reduce noise overfitting; 2) Dynamic Anchor Selection (DAS) selects representative unknown class samples using KNN density and class probability, assigning hard pseudo-labels to balance confidence differences and learn accurate feature distributions.

Result: Extensive experiments show the method effectively mitigates pseudo-label noise and achieves state-of-the-art results on multiple GCD benchmarks.

Conclusion: The proposed LSP and DAS modules successfully address pseudo-label noise issues in GCD, improving clustering accuracy by enhancing model robustness and better representing unknown class distributions.

Abstract: Generalized category discovery (GCD) is an important and challenging task in open-world learning. Specifically, given some labeled data of known classes, GCD aims to cluster unlabeled data that contain both known and unknown classes. Current GCD methods based on parametric classification adopt the DINO-like pseudo-labeling strategy, where the sharpened probability output of one view is used as supervision information for the other view. However, large pre-trained models have a preference for some specific visual patterns, resulting in encoding spurious correlation for unlabeled data and generating noisy pseudo-labels. To address this issue, we propose a novel method, which contains two modules: Loss Sharpness Penalty (LSP) and Dynamic Anchor Selection (DAS). LSP enhances the robustness of model parameters to small perturbations by minimizing the worst-case loss sharpness of the model, which suppressing the encoding of trivial features, thereby reducing overfitting of noise samples and improving the quality of pseudo-labels. Meanwhile, DAS selects representative samples for the unknown classes based on KNN density and class probability during the model training and assigns hard pseudo-labels to them, which not only alleviates the confidence difference between known and unknown classes but also enables the model to quickly learn more accurate feature distribution for the unknown classes, thus further improving the clustering accuracy. Extensive experiments demonstrate that the proposed method can effectively mitigate the noise of pseudo-labels, and achieve state-of-the-art results on multiple GCD benchmarks.

Method: The framework has two main components: 1) A temporal search mechanism that captures event-level continuity by aggregating similarity scores across sequential video segments, and 2) A Google Image Search-based fallback module that expands query representations with external web imagery to bridge gaps in pretrained visual embeddings.

Result: The approach advances temporal reasoning and generalization capabilities of video retrieval systems, improving robustness against out-of-distribution queries and enabling more coherent retrieval of multi-event queries.

Conclusion: MADTempo paves the way for more semantically aware and adaptive retrieval across large-scale video corpora by unifying temporal search with web-scale visual grounding.

Abstract: The rapid expansion of video content across online platforms has accelerated the need for retrieval systems capable of understanding not only isolated visual moments but also the temporal structure of complex events. Existing approaches often fall short in modeling temporal dependencies across multiple events and in handling queries that reference unseen or rare visual concepts. To address these challenges, we introduce MADTempo, a video retrieval framework developed by our team, AIO_Trinh, that unifies temporal search with web-scale visual grounding. Our temporal search mechanism captures event-level continuity by aggregating similarity scores across sequential video segments, enabling coherent retrieval of multi-event queries. Complementarily, a Google Image Search-based fallback module expands query representations with external web imagery, effectively bridging gaps in pretrained visual embeddings and improving robustness against out-of-distribution (OOD) queries. Together, these components advance the temporal reasoning and generalization capabilities of modern video retrieval systems, paving the way for more semantically aware and adaptive retrieval across large-scale video corpora.

Method: Three key innovations: (1) Cascaded dual-embedding pipeline combining BEIT-3 and SigLIP for broad retrieval, refined by BLIP-2 based reranking to balance recall and precision. (2) Temporal-aware scoring mechanism applying exponential decay penalties to large temporal gaps via beam search, constructing coherent event sequences. (3) Agent-guided query decomposition using GPT-4o to automatically interpret ambiguous queries, decompose them into modality-specific sub-queries (visual/OCR/ASR), and perform adaptive score fusion.

Result: Qualitative analysis demonstrates the system effectively handles ambiguous queries, retrieves temporally coherent sequences, and dynamically adapts fusion strategies, advancing interactive moment search capabilities.

Conclusion: The proposed unified multimodal moment retrieval system addresses key limitations of existing approaches through innovative techniques in embedding, temporal modeling, and query processing, significantly advancing interactive video moment search capabilities.

Abstract: The exponential growth of video content has created an urgent need for efficient multimodal moment retrieval systems. However, existing approaches face three critical challenges: (1) fixed-weight fusion strategies fail across cross modal noise and ambiguous queries, (2) temporal modeling struggles to capture coherent event sequences while penalizing unrealistic gaps, and (3) systems require manual modality selection, reducing usability. We propose a unified multimodal moment retrieval system with three key innovations. First, a cascaded dual-embedding pipeline combines BEIT-3 and SigLIP for broad retrieval, refined by BLIP-2 based reranking to balance recall and precision. Second, a temporal-aware scoring mechanism applies exponential decay penalties to large temporal gaps via beam search, constructing coherent event sequences rather than isolated frames. Third, Agent-guided query decomposition (GPT-4o) automatically interprets ambiguous queries, decomposes them into modality specific sub-queries (visual/OCR/ASR), and performs adaptive score fusion eliminating manual modality selection. Qualitative analysis demonstrates that our system effectively handles ambiguous queries, retrieves temporally coherent sequences, and dynamically adapts fusion strategies, advancing interactive moment search capabilities.

Method: 1. Two-stage flow-guided deformable warping mechanism with coarse-to-fine offset prediction and mask modulation for precise feature alignment. 2. Multi-reference quality aware strategy that adjusts distortion weights based on reference quality, applied to hierarchical training to reduce error propagation. 3. Training-free module that downsamples frames by motion magnitude and resolution for smooth motion estimation.

Result: CAMA achieves 24.95% BD-rate (PSNR) savings over baseline DCVC-TCM, and outperforms reproduced DCVC-DC and traditional codec HM-16.25 on standard test datasets.

Conclusion: The proposed content adaptive motion alignment framework significantly improves neural video compression performance by adapting encoding strategies to content characteristics, demonstrating superior results compared to state-of-the-art methods.

Abstract: Recent advances in end-to-end video compression have shown promising results owing to their unified end-to-end learning optimization. However, such generalized frameworks often lack content-specific adaptation, leading to suboptimal compression performance. To address this, this paper proposes a content adaptive based motion alignment framework that improves performance by adapting encoding strategies to diverse content characteristics. Specifically, we first introduce a two-stage flow-guided deformable warping mechanism that refines motion compensation with coarse-to-fine offset prediction and mask modulation, enabling precise feature alignment. Second, we propose a multi-reference quality aware strategy that adjusts distortion weights based on reference quality, and applies it to hierarchical training to reduce error propagation. Third, we integrate a training-free module that downsamples frames by motion magnitude and resolution to obtain smooth motion estimation. Experimental results on standard test datasets demonstrate that our framework CAMA achieves significant improvements over state-of-the-art Neural Video Compression models, achieving a 24.95% BD-rate (PSNR) savings over our baseline model DCVC-TCM, while also outperforming reproduced DCVC-DC and traditional codec HM-16.25.

Method: Proposes UAGLNet with: 1) Cooperative encoder using hybrid CNN and transformer layers at different stages; 2) Intermediate cooperative interaction block (CIB) to bridge local-global feature gaps; 3) Global-Local Fusion (GLF) module for complementary feature fusion; 4) Uncertainty-Aggregated Decoder (UAD) for pixel-wise uncertainty estimation to enhance segmentation.

Result: Extensive experiments show superior performance compared to state-of-the-art methods for building extraction from remote sensing images.

Conclusion: UAGLNet effectively addresses building extraction challenges by integrating global-local semantics with uncertainty modeling, achieving more accurate segmentation results for complex building structures.

Abstract: Building extraction from remote sensing images is a challenging task due to the complex structure variations of the buildings. Existing methods employ convolutional or self-attention blocks to capture the multi-scale features in the segmentation models, while the inherent gap of the feature pyramids and insufficient global-local feature integration leads to inaccurate, ambiguous extraction results. To address this issue, in this paper, we present an Uncertainty-Aggregated Global-Local Fusion Network (UAGLNet), which is capable to exploit high-quality global-local visual semantics under the guidance of uncertainty modeling. Specifically, we propose a novel cooperative encoder, which adopts hybrid CNN and transformer layers at different stages to capture the local and global visual semantics, respectively. An intermediate cooperative interaction block (CIB) is designed to narrow the gap between the local and global features when the network becomes deeper. Afterwards, we propose a Global-Local Fusion (GLF) module to complementarily fuse the global and local representations. Moreover, to mitigate the segmentation ambiguity in uncertain regions, we propose an Uncertainty-Aggregated Decoder (UAD) to explicitly estimate the pixel-wise uncertainty to enhance the segmentation accuracy. Extensive experiments demonstrate that our method achieves superior performance to other state-of-the-art methods. Our code is available at https://github.com/Dstate/UAGLNet

Method: Reparameterize CLIP outputs as concentration parameters of a Dirichlet distribution to capture both semantic structure and predictive confidence magnitude. Propose a unified adversarial fine-tuning objective that aligns these Dirichlet distributions between clean and adversarial examples, moving beyond single-logit matching to restore calibrated uncertainty.

Result: Experiments on multiple zero-shot classification benchmarks show the approach effectively restores calibrated uncertainty, achieves competitive adversarial robustness, and maintains clean accuracy.

Conclusion: The proposed method bridges the reliability gap in CLIP by addressing both prediction accuracy and uncertainty alignment under adversarial perturbations, providing a more comprehensive solution for robust and reliable zero-shot classification.

Abstract: CLIP delivers strong zero-shot classification but remains highly vulnerable to adversarial attacks. Previous work of adversarial fine-tuning largely focuses on matching the predicted logits between clean and adversarial examples, which overlooks uncertainty calibration and may degrade the zero-shot generalization. A common expectation in reliable uncertainty estimation is that predictive uncertainty should increase as inputs become more difficult or shift away from the training distribution. However, we frequently observe the opposite in the adversarial setting: perturbations not only degrade accuracy but also suppress uncertainty, leading to severe miscalibration and unreliable over-confidence. This overlooked phenomenon highlights a critical reliability gap beyond robustness. To bridge this gap, we propose a novel adversarial fine-tuning objective for CLIP considering both prediction accuracy and uncertainty alignments. By reparameterizing the output of CLIP as the concentration parameter of a Dirichlet distribution, we propose a unified representation that captures relative semantic structure and the magnitude of predictive confidence. Our objective aligns these distributions holistically under perturbations, moving beyond single-logit anchoring and restoring calibrated uncertainty. Experiments on multiple zero-shot classification benchmarks demonstrate that our approach effectively restores calibrated uncertainty and achieves competitive adversarial robustness while maintaining clean accuracy.

Method: Uses CLIP image space to systematically extract pure content from content images and style elements from style references. Features three components: Controllable Style Adaptive Instance Normalization (CSAdaIN) for multi-style blending, KVS Injection for targeted style integration, and a style transfer consistency objective for process coherence.

Result: Significantly outperforms state-of-the-art methods in both conventional and diffusion-based baselines. Achieves at least 2× faster inference than other diffusion-based approaches by eliminating DDIM inversion and inference-stage optimization.

Conclusion: SCAdapter provides an effective and efficient solution for photo-realistic style transfer, addressing key limitations of current diffusion models while offering practical speed advantages for real-world applications.

Abstract: Diffusion models have emerged as the leading approach for style transfer, yet they struggle with photo-realistic transfers, often producing painting-like results or missing detailed stylistic elements. Current methods inadequately address unwanted influence from original content styles and style reference content features. We introduce SCAdapter, a novel technique leveraging CLIP image space to effectively separate and integrate content and style features. Our key innovation systematically extracts pure content from content images and style elements from style references, ensuring authentic transfers. This approach is enhanced through three components: Controllable Style Adaptive Instance Normalization (CSAdaIN) for precise multi-style blending, KVS Injection for targeted style integration, and a style transfer consistency objective maintaining process coherence. Comprehensive experiments demonstrate SCAdapter significantly outperforms state-of-the-art methods in both conventional and diffusion-based baselines. By eliminating DDIM inversion and inference-stage optimization, our method achieves at least $2\times$ faster inference than other diffusion-based approaches, making it both more effective and efficient for practical applications.

Method: GTR-Turbo merges weights of checkpoints produced during ongoing RL training, then uses this merged model as a “free” teacher to guide subsequent RL via supervised fine-tuning or soft logit distillation.

Result: Improves baseline model accuracy by 10-30% across diverse visual agentic tasks, reduces wall-clock training time by 50% and compute cost by 60% relative to GTR.

Conclusion: GTR-Turbo provides an efficient upgrade to GTR that eliminates dependence on privileged VLMs, mitigates entropy collapse, maintains training stability, and significantly reduces computational costs while improving performance.

Abstract: Multi-turn reinforcement learning (RL) for multi-modal agents built upon vision-language models (VLMs) is hampered by sparse rewards and long-horizon credit assignment. Recent methods densify the reward by querying a teacher that provides step-level feedback, e.g., Guided Thought Reinforcement (GTR) and On-Policy Distillation, but rely on costly, often privileged models as the teacher, limiting practicality and reproducibility. We introduce GTR-Turbo, a highly efficient upgrade to GTR, which matches the performance without training or querying an expensive teacher model. Specifically, GTR-Turbo merges the weights of checkpoints produced during the ongoing RL training, and then uses this merged model as a “free” teacher to guide the subsequent RL via supervised fine-tuning or soft logit distillation. This design removes dependence on privileged VLMs (e.g., GPT or Gemini), mitigates the “entropy collapse” observed in prior work, and keeps training stable. Across diverse visual agentic tasks, GTR-Turbo improves the accuracy of the baseline model by 10-30% while reducing wall-clock training time by 50% and compute cost by 60% relative to GTR.

Method: Formally identifies cumulative reuse error effect and minimizes non-prefix cache reuse error. Analyzes varying importance of model layers and proposes dynamic, layer-aware recomputation strategy to balance accuracy and efficiency.

Result: Achieves accuracy on par with full recomputation while requiring only 2-5% of tokens to compute, yielding 1.2x-16x TTFT (Time To First Token) speedups. Integrated into SGLang for practical deployment.

Conclusion: VLCache effectively eliminates costly recomputation for recurring multimodal inputs through formal error analysis and layer-aware strategies, enabling significantly faster inference with maintained accuracy.

Abstract: This paper presents VLCache, a cache reuse framework that exploits both Key-Value (KV) cache and encoder cache from prior multimodal inputs to eliminate costly recomputation when the same multimodal inputs recur. Unlike previous heuristic approaches, we formally identify the cumulative reuse error effect and demonstrate how to minimize the non-prefix cache reuse error effectively. We further analyze the varying importance of model layers and propose a dynamic, layer-aware recomputation strategy to balance accuracy and efficiency. Experimental results show that VLCache achieves an accuracy on par with full recomputation, while requiring only 2-5% of the tokens to compute, yielding 1.2x-16x TTFT speedups. The proposed VLCache pipeline has been integrated into SGLang, enabling significantly faster inference in practical deployments.

Method: Proposes Generator-Aware Prototype Learning (GAPL) with two key components: 1) learning compact canonical forgery prototypes to create unified low-variance feature space addressing data heterogeneity, and 2) two-stage training with Low-Rank Adaptation to overcome fixed encoder bottlenecks while preserving pretrained knowledge.

Result: GAPL achieves state-of-the-art performance with superior detection accuracy across diverse GAN and diffusion-based generators, demonstrating robust generalization capabilities.

Conclusion: The Benefit then Conflict dilemma is a fundamental challenge in universal AIGI detection, and GAPL effectively addresses both data heterogeneity and model bottleneck issues through structured prototype learning and adaptive training, establishing more robust decision boundaries.

Abstract: The pursuit of a universal AI-generated image (AIGI) detector often relies on aggregating data from numerous generators to improve generalization. However, this paper identifies a paradoxical phenomenon we term the Benefit then Conflict dilemma, where detector performance stagnates and eventually degrades as source diversity expands. Our systematic analysis, diagnoses this failure by identifying two core issues: severe data-level heterogeneity, which causes the feature distributions of real and synthetic images to increasingly overlap, and a critical model-level bottleneck from fixed, pretrained encoders that cannot adapt to the rising complexity. To address these challenges, we propose Generator-Aware Prototype Learning (GAPL), a framework that constrain representation with a structured learning paradigm. GAPL learns a compact set of canonical forgery prototypes to create a unified, low-variance feature space, effectively countering data heterogeneity.To resolve the model bottleneck, it employs a two-stage training scheme with Low-Rank Adaptation, enhancing its discriminative power while preserving valuable pretrained knowledge. This approach establishes a more robust and generalizable decision boundary. Through extensive experiments, we demonstrate that GAPL achieves state-of-the-art performance, showing superior detection accuracy across a wide variety of GAN and diffusion-based generators. Code is available at https://github.com/UltraCapture/GAPL

Method: UniVCD uses frozen SAM2 and CLIP models with a lightweight feature alignment module to bridge SAM2’s detailed spatial representations with CLIP’s semantic priors. It includes streamlined post-processing to suppress noise and pseudo-changes for objects with well-defined boundaries.

Result: Experiments on public BCD and SCD benchmarks show UniVCD achieves consistently strong performance, matching or surpassing existing open-vocabulary CD methods in key metrics like F1 and IoU.

Conclusion: Unsupervised change detection with frozen vision foundation models and lightweight multi-modal alignment is a practical and effective paradigm for open-vocabulary change detection.

Abstract: Change detection (CD) identifies scene changes from multi-temporal observations and is widely used in urban development and environmental monitoring. Most existing CD methods rely on supervised learning, making performance strongly dataset-dependent and incurring high annotation costs; they typically focus on a few predefined categories and generalize poorly to diverse scenes. With the rise of vision foundation models such as SAM2 and CLIP, new opportunities have emerged to relax these constraints. We propose Unified Open-Vocabulary Change Detection (UniVCD), an unsupervised, open-vocabulary change detection method built on frozen SAM2 and CLIP. UniVCD detects category-agnostic changes across diverse scenes and imaging geometries without any labeled data or paired change images. A lightweight feature alignment module is introduced to bridge the spatially detailed representations from SAM2 and the semantic priors from CLIP, enabling high-resolution, semantically aware change estimation while keeping the number of trainable parameters small. On top of this, a streamlined post-processing pipeline is further introduced to suppress noise and pseudo-changes, improving the detection accuracy for objects with well-defined boundaries. Experiments on several public BCD (Binary Change Detection) and SCD (Semantic Change Detection) benchmarks show that UniVCD achieves consistently strong performance and matches or surpasses existing open-vocabulary CD methods in key metrics such as F1 and IoU. The results demonstrate that unsupervised change detection with frozen vision foundation models and lightweight multi-modal alignment is a practical and effective paradigm for open-vocabulary CD. Code and pretrained models will be released at https://github.com/Die-Xie/UniVCD.

Method: The authors cast existing distillation methods into a unified framework and adapt them to work with the strong T2I teacher model FLUX.1-lite. They conduct thorough methodological analysis, examining input scaling, network architecture, and hyperparameters to identify key obstacles when moving from discrete class labels to free-form language prompts.

Result: The study identifies key challenges in adapting distillation techniques for T2I generation and provides practical guidelines for implementation. The authors release an open-source implementation and pretrained student models, establishing a foundation for fast, high-fidelity, and resource-efficient diffusion generators in real-world T2I applications.

Conclusion: This work successfully adapts diffusion distillation techniques for text-to-image generation, providing systematic analysis and practical guidelines that enable deployment of efficient T2I models. The open-source implementation and pretrained models support further research and real-world applications.

Abstract: Diffusion distillation has dramatically accelerated class-conditional image synthesis, but its applicability to open-ended text-to-image (T2I) generation is still unclear. We present the first systematic study that adapts and compares state-of-the-art distillation techniques on a strong T2I teacher model, FLUX.1-lite. By casting existing methods into a unified framework, we identify the key obstacles that arise when moving from discrete class labels to free-form language prompts. Beyond a thorough methodological analysis, we offer practical guidelines on input scaling, network architecture, and hyperparameters, accompanied by an open-source implementation and pretrained student models. Our findings establish a solid foundation for deploying fast, high-fidelity, and resource-efficient diffusion generators in real-world T2I applications. Code is available on github.com/alibaba-damo-academy/T2I-Distill.

Method: Extract semantic and geometric features from light field images using vision foundation models, then convert them into view-dependent Gaussian splats as a unified object representation. This supports differentiable rendering and pose optimization without requiring pre-trained models.

Result: The method is competitive with state-of-the-art model-based trackers on challenging reflective objects, as demonstrated on a new light field object tracking dataset containing reflective objects with precise ground truth poses.

Conclusion: The light field-based approach with view-dependent Gaussian splats provides a promising direction toward universal object tracking in robotic systems, handling complex visual behaviors without requiring pre-trained models.

Abstract: Object tracking is an important step in robotics and reautonomous driving pipelines, which has to generalize to previously unseen and complex objects. Existing high-performing methods often rely on pre-captured object views to build explicit reference models, which restricts them to a fixed set of known objects. However, such reference models can struggle with visually complex appearance, reducing the quality of tracking. In this work, we introduce an object tracking method based on light field images that does not depend on a pre-trained model, while being robust to complex visual behavior, such as reflections. We extract semantic and geometric features from light field inputs using vision foundation models and convert them into view-dependent Gaussian splats. These splats serve as a unified object representation, supporting differentiable rendering and pose optimization. We further introduce a light field object tracking dataset containing challenging reflective objects with precise ground truth poses. Experiments demonstrate that our method is competitive with state-of-the-art model-based trackers in these difficult cases, paving the way toward universal object tracking in robotic systems. Code/data available at https://github.com/nagonch/LiFT-6DoF.

Method: Two-stage framework: 1) Vision-language model integrates ophthalmological knowledge for DR grading and lesion classification, linking semantic concepts with visual features. 2) Iterative severity regression with weakly-supervised semantic segmentation generates lesion saliency maps and uses progressive inpainting to downgrade disease severity toward healthier appearances.

Result: TWLR achieves competitive performance on FGADR, DDR, and a private dataset for both DR classification and lesion segmentation, providing accurate lesion localization without pixel-level supervision and interpretable visualizations of disease-to-healthy transformations.

Conclusion: The framework offers an explainable and annotation-efficient solution for automated retinal image analysis by combining vision-language integration with weakly-supervised severity regression, addressing both interpretability and annotation cost challenges in medical imaging.

Abstract: Accurate medical image analysis can greatly assist clinical diagnosis, but its effectiveness relies on high-quality expert annotations Obtaining pixel-level labels for medical images, particularly fundus images, remains costly and time-consuming. Meanwhile, despite the success of deep learning in medical imaging, the lack of interpretability limits its clinical adoption. To address these challenges, we propose TWLR, a two-stage framework for interpretable diabetic retinopathy (DR) assessment. In the first stage, a vision-language model integrates domain-specific ophthalmological knowledge into text embeddings to jointly perform DR grading and lesion classification, effectively linking semantic medical concepts with visual features. The second stage introduces an iterative severity regression framework based on weakly-supervised semantic segmentation. Lesion saliency maps generated through iterative refinement direct a progressive inpainting mechanism that systematically eliminates pathological features, effectively downgrading disease severity toward healthier fundus appearances. Critically, this severity regression approach achieves dual benefits: accurate lesion localization without pixel-level supervision and providing an interpretable visualization of disease-to-healthy transformations. Experimental results on the FGADR, DDR, and a private dataset demonstrate that TWLR achieves competitive performance in both DR classification and lesion segmentation, offering a more explainable and annotation-efficient solution for automated retinal image analysis.

Method: UnCoL uses a dual-teacher framework: one frozen foundation model teacher for general knowledge distillation (visual and semantic representations), and one progressively adapting teacher for task-specific representations. Pseudo-label learning is adaptively regulated by predictive uncertainty to balance guidance from both teachers.

Result: UnCoL consistently outperforms state-of-the-art semi-supervised methods and foundation model baselines on diverse 2D and 3D segmentation benchmarks, achieving near fully supervised performance with significantly reduced annotation requirements.

Conclusion: The proposed uncertainty-informed collaborative learning framework effectively harmonizes generalization and specialization for semi-supervised medical image segmentation, addressing limitations of foundation models in clinical applications.

Abstract: Vision foundation models have demonstrated strong generalization in medical image segmentation by leveraging large-scale, heterogeneous pretraining. However, they often struggle to generalize to specialized clinical tasks under limited annotations or rare pathological variations, due to a mismatch between general priors and task-specific requirements. To address this, we propose Uncertainty-informed Collaborative Learning (UnCoL), a dual-teacher framework that harmonizes generalization and specialization in semi-supervised medical image segmentation. Specifically, UnCoL distills both visual and semantic representations from a frozen foundation model to transfer general knowledge, while concurrently maintaining a progressively adapting teacher to capture fine-grained and task-specific representations. To balance guidance from both teachers, pseudo-label learning in UnCoL is adaptively regulated by predictive uncertainty, which selectively suppresses unreliable supervision and stabilizes learning in ambiguous regions. Experiments on diverse 2D and 3D segmentation benchmarks show that UnCoL consistently outperforms state-of-the-art semi-supervised methods and foundation model baselines. Moreover, our model delivers near fully supervised performance with markedly reduced annotation requirements.

Method: 1) Incorporates an independent annotation VAE network to map annotation masks into the latent space shared by images; 2) Tailors the diffusion model to capture the joint distribution of images and annotation masks conditioned on text prompts; 3) Develops a mask optimization strategy to mitigate annotation noise during generation.

Result: Experiments on Pascal VOC, COCO, and ADE20K datasets show that the annotated dataset generated by JoDiffusion yields substantial performance improvements in semantic segmentation compared to existing methods.

Conclusion: JoDiffusion enables simultaneous generation of paired images and semantically consistent annotation masks solely conditioned on text prompts, demonstrating superior scalability and effectively addressing both semantic inconsistency and scalability problems in synthetic dataset generation for semantic segmentation.

Abstract: Given the inherently costly and time-intensive nature of pixel-level annotation, the generation of synthetic datasets comprising sufficiently diverse synthetic images paired with ground-truth pixel-level annotations has garnered increasing attention recently for training high-performance semantic segmentation models. However, existing methods necessitate to either predict pseudo annotations after image generation or generate images conditioned on manual annotation masks, which incurs image-annotation semantic inconsistency or scalability problem. To migrate both problems with one stone, we present a novel dataset generative diffusion framework for semantic segmentation, termed JoDiffusion. Firstly, given a standard latent diffusion model, JoDiffusion incorporates an independent annotation variational auto-encoder (VAE) network to map annotation masks into the latent space shared by images. Then, the diffusion model is tailored to capture the joint distribution of each image and its annotation mask conditioned on a text prompt. By doing these, JoDiffusion enables simultaneously generating paired images and semantically consistent annotation masks solely conditioned on text prompts, thereby demonstrating superior scalability. Additionally, a mask optimization strategy is developed to mitigate the annotation noise produced during generation. Experiments on Pascal VOC, COCO, and ADE20K datasets show that the annotated dataset generated by JoDiffusion yields substantial performance improvements in semantic segmentation compared to existing methods.

Method: Propose a unified framework combining Qwen-VL for comprehension and LTX for synthesis, connected via latent query embedding and a connector module. The model is trained in three stages: 1) text-to-video pre-training, 2) supervised fine-tuning, and 3) reinforcement learning using GRPO with a novel causal consistency reward. Training uses a newly curated large-scale NSP dataset.

Result: The model achieves state-of-the-art performance on the proposed NSP benchmark, demonstrating improved capability for generalist multimodal systems to anticipate future scenes and advancing temporal reasoning in unified video models.

Conclusion: The Next Scene Prediction task successfully pushes unified video models toward deeper temporal and causal reasoning, with the proposed framework showing promising results in anticipating what happens next, representing an important advancement for generalist multimodal systems.

Abstract: Recent unified models for joint understanding and generation have significantly advanced visual generation capabilities. However, their focus on conventional tasks like text-to-video generation has left the temporal reasoning potential of unified models largely underexplored. To address this gap, we introduce Next Scene Prediction (NSP), a new task that pushes unified video models toward temporal and causal reasoning. Unlike text-to-video generation, NSP requires predicting plausible futures from preceding context, demanding deeper understanding and reasoning. To tackle this task, we propose a unified framework combining Qwen-VL for comprehension and LTX for synthesis, bridged by a latent query embedding and a connector module. This model is trained in three stages on our newly curated, large-scale NSP dataset: text-to-video pre-training, supervised fine-tuning, and reinforcement learning (via GRPO) with our proposed causal consistency reward. Experiments demonstrate our model achieves state-of-the-art performance on our benchmark, advancing the capability of generalist multimodal systems to anticipate what happens next.

Method: 1) Diffusion-IR: Uses diffusion models to restore weather-degraded images. 2) Point Cloud Restoration (PCR): Compensates for corrupted LiDAR data using image object cues. 3) Bidirectional Adaptive Fusion and Alignment Module (BAFAM): Enables dynamic multi-modal fusion and bidirectional bird’s-eye view alignment to maintain spatial consistency.

Result: Achieves state-of-the-art robustness under adverse weather on three public datasets while preserving strong clean-data performance. Zero-shot results on real-world DENSE dataset validate generalization capabilities.

Conclusion: DiffFusion effectively addresses weather-induced challenges in multi-modal 3D object detection through diffusion-based restoration and adaptive cross-modal fusion, demonstrating superior robustness and generalization in adverse conditions.

Abstract: Multi-modal 3D object detection is important for reliable perception in robotics and autonomous driving. However, its effectiveness remains limited under adverse weather conditions due to weather-induced distortions and misalignment between different data modalities. In this work, we propose DiffFusion, a novel framework designed to enhance robustness in challenging weather through diffusion-based restoration and adaptive cross-modal fusion. Our key insight is that diffusion models possess strong capabilities for denoising and generating data that can adapt to various weather conditions. Building on this, DiffFusion introduces Diffusion-IR restoring images degraded by weather effects and Point Cloud Restoration (PCR) compensating for corrupted LiDAR data using image object cues. To tackle misalignments between two modalities, we develop Bidirectional Adaptive Fusion and Alignment Module (BAFAM). It enables dynamic multi-modal fusion and bidirectional bird’s-eye view (BEV) alignment to maintain consistent spatial correspondence. Extensive experiments on three public datasets show that DiffFusion achieves state-of-the-art robustness under adverse weather while preserving strong clean-data performance. Zero-shot results on the real-world DENSE dataset further validate its generalization. The implementation of our DiffFusion will be released as open-source.

Method: Systematically investigated multiple approaches: amplitude-based statistical preprocessing (sigmoid weighting, threshold zeroing), frequency-domain filtering, autoencoder-based background suppression, data augmentation strategies, and transfer learning. Evaluated on people counting in indoor environments using FMCW MIMO radar across two different layouts.

Result: Sigmoid-based amplitude weighting achieved 50.1% RMSE and 55.2% MAE reductions. Data augmentation provided modest improvements (up to 8.8% MAE). Transfer learning was essential for large spatial shifts, achieving 82.1% RMSE and 91.3% MAE reductions with 540 target-domain samples.

Conclusion: Integrating deep learning models with amplitude-based preprocessing and efficient transfer learning provides a practical direction for developing robust radar sensing systems that maintain accuracy under spatial variations.

Abstract: This study presents the first comprehensive evaluation of spatial generalization techniques, which are essential for the practical deployment of deep learning-based radio-frequency (RF) sensing. Focusing on people counting in indoor environments using frequency-modulated continuous-wave (FMCW) multiple-input multiple-output (MIMO) radar, we systematically investigate a broad set of approaches, including amplitude-based statistical preprocessing (sigmoid weighting and threshold zeroing), frequency-domain filtering, autoencoder-based background suppression, data augmentation strategies, and transfer learning. Experimental results collected across two environments with different layouts demonstrate that sigmoid-based amplitude weighting consistently achieves superior cross-environment performance, yielding 50.1% and 55.2% reductions in root-mean-square error (RMSE) and mean absolute error (MAE), respectively, compared with baseline methods. Data augmentation provides additional though modest benefits, with improvements up to 8.8% in MAE. By contrast, transfer learning proves indispensable for large spatial shifts, achieving 82.1% and 91.3% reductions in RMSE and MAE, respectively, with 540 target-domain samples. Taken together, these findings establish a highly practical direction for developing radar sensing systems capable of maintaining robust accuracy under spatial variations by integrating deep learning models with amplitude-based preprocessing and efficient transfer learning.

Method: A pipeline with three key innovations: (1) predictive causal adaptation for next-frame prediction and future keyframe anticipation, (2) future-guided self-forcing with dual-region KV caching to address exposure bias, and (3) multi-prompt conditioning for fine-grained procedural control over multi-step instructions.

Result: The method produces temporally coherent and semantically faithful instructional videos that accurately follow complex, multi-step task descriptions, enabling interactive video generation where future prompt updates dynamically influence ongoing streaming video generation.

Conclusion: SneakPeek effectively mitigates temporal drift, preserves motion consistency, and enables interactive streaming instructional video generation with improved controllability over multi-step procedural activities.

Abstract: Instructional video generation is an emerging task that aims to synthesize coherent demonstrations of procedural activities from textual descriptions. Such capability has broad implications for content creation, education, and human-AI interaction, yet existing video diffusion models struggle to maintain temporal consistency and controllability across long sequences of multiple action steps. We introduce a pipeline for future-driven streaming instructional video generation, dubbed SneakPeek, a diffusion-based autoregressive framework designed to generate precise, stepwise instructional videos conditioned on an initial image and structured textual prompts. Our approach introduces three key innovations to enhance consistency and controllability: (1) predictive causal adaptation, where a causal model learns to perform next-frame prediction and anticipate future keyframes; (2) future-guided self-forcing with a dual-region KV caching scheme to address the exposure bias issue at inference time; (3) multi-prompt conditioning, which provides fine-grained and procedural control over multi-step instructions. Together, these components mitigate temporal drift, preserve motion consistency, and enable interactive video generation where future prompt updates dynamically influence ongoing streaming video generation. Experimental results demonstrate that our method produces temporally coherent and semantically faithful instructional videos that accurately follow complex, multi-step task descriptions.

Method: DePT3R uses a powerful backbone to extract deep spatio-temporal features and regresses pixel-wise maps with dense prediction heads. It performs both dense point tracking and 3D reconstruction simultaneously in a single forward pass without requiring camera poses.

Result: DePT3R demonstrates strong performance on challenging dynamic scene benchmarks and shows significant improvements in memory efficiency over existing state-of-the-art methods.

Conclusion: DePT3R offers a flexible and efficient solution for unified dynamic scene understanding, operating without camera poses and handling multiple images simultaneously, making it particularly suitable for dynamic environments with rapid changes.

Abstract: Current methods for dense 3D point tracking in dynamic scenes typically rely on pairwise processing, require known camera poses, or assume a temporal ordering to input frames, constraining their flexibility and applicability. Additionally, recent advances have successfully enabled efficient 3D reconstruction from large-scale, unposed image collections, underscoring opportunities for unified approaches to dynamic scene understanding. Motivated by this, we propose DePT3R, a novel framework that simultaneously performs dense point tracking and 3D reconstruction of dynamic scenes from multiple images in a single forward pass. This multi-task learning is achieved by extracting deep spatio-temporal features with a powerful backbone and regressing pixel-wise maps with dense prediction heads. Crucially, DePT3R operates without requiring camera poses, substantially enhancing its adaptability and efficiency-especially important in dynamic environments with rapid changes. We validate DePT3R on several challenging benchmarks involving dynamic scenes, demonstrating strong performance and significant improvements in memory efficiency over existing state-of-the-art methods. Data and codes are available via the open repository: https://github.com/StructuresComp/DePT3R

Method: Proposes Motus with: 1) Mixture-of-Transformer architecture integrating three experts (understanding, video generation, action), 2) UniDiffuser-style scheduler for flexible mode switching, 3) Optical flow-based latent action learning, 4) Three-phase training pipeline with six-layer data pyramid.

Result: Achieves +15% improvement over X-VLA and +45% improvement over Pi0.5 in simulation, and +11-48% improvement in real-world scenarios, demonstrating superior performance in unified modeling for robotic tasks.

Conclusion: Unified modeling of all functionalities and priors significantly benefits downstream robotic tasks, with Motus showing state-of-the-art performance across both simulation and real-world environments.

Abstract: While a general embodied agent must function as a unified system, current methods are built on isolated models for understanding, world modeling, and control. This fragmentation prevents unifying multimodal generative capabilities and hinders learning from large-scale, heterogeneous data. In this paper, we propose Motus, a unified latent action world model that leverages existing general pretrained models and rich, sharable motion information. Motus introduces a Mixture-of-Transformer (MoT) architecture to integrate three experts (i.e., understanding, video generation, and action) and adopts a UniDiffuser-style scheduler to enable flexible switching between different modeling modes (i.e., world models, vision-language-action models, inverse dynamics models, video generation models, and video-action joint prediction models). Motus further leverages the optical flow to learn latent actions and adopts a recipe with three-phase training pipeline and six-layer data pyramid, thereby extracting pixel-level “delta action” and enabling large-scale action pretraining. Experiments show that Motus achieves superior performance against state-of-the-art methods in both simulation (a +15% improvement over X-VLA and a +45% improvement over Pi0.5) and real-world scenarios(improved by +11~48%), demonstrating unified modeling of all functionalities and priors significantly benefits downstream robotic tasks.

Method: 1) Use diffusion-based material estimator to generate multiple candidate decompositions per view; 2) Fit explicit low-dimensional parametric function to predictions; 3) Propose robust optimization framework with soft per-view prediction selection and confidence-based soft multi-view inlier set; 4) Use inverse path tracing to optimize low-dimensional parameters.

Result: Outperforms state-of-the-art methods in material disentanglement on both synthetic and real scenes, producing sharp and clean reconstructions suitable for high-quality relighting.

Conclusion: Intrinsic Image Fusion successfully combines single-view priors with multi-view consistency to reconstruct high-quality physically based materials, overcoming the underconstrained nature of material reconstruction while avoiding expensive path tracing.

Abstract: We introduce Intrinsic Image Fusion, a method that reconstructs high-quality physically based materials from multi-view images. Material reconstruction is highly underconstrained and typically relies on analysis-by-synthesis, which requires expensive and noisy path tracing. To better constrain the optimization, we incorporate single-view priors into the reconstruction process. We leverage a diffusion-based material estimator that produces multiple, but often inconsistent, candidate decompositions per view. To reduce the inconsistency, we fit an explicit low-dimensional parametric function to the predictions. We then propose a robust optimization framework using soft per-view prediction selection together with confidence-based soft multi-view inlier set to fuse the most consistent predictions of the most confident views into a consistent parametric material space. Finally, we use inverse path tracing to optimize for the low-dimensional parameters. Our results outperform state-of-the-art methods in material disentanglement on both synthetic and real scenes, producing sharp and clean reconstructions suitable for high-quality relighting.

Method: Evaluated five approaches in two distinct indoor environments: 1) rule-based connected component method, 2) three traditional ML models (k-nearest neighbors, random forest, support vector machine), and 3) deep learning model combining CNN and LSTM.

Result: CNN-LSTM achieved highest accuracy in training environment, traditional ML showed moderate performance. In new layout, all learning-based methods degraded significantly while rule-based method remained stable. Binary detection (presence vs absence) worked well across all models and layouts.

Conclusion: High-capacity models provide fine-grained outputs with high accuracy in same environment but are vulnerable to domain shift. Rule-based methods are robust against domain shift but lack fine-grained outputs. Clear trade-off exists between spatial generalization performance and output granularity regardless of model type.

Abstract: This study presents the first comprehensive comparison of rule-based methods, traditional machine learning models, and deep learning models in radio wave sensing with frequency modulated continuous wave multiple input multiple output radar. We systematically evaluated five approaches in two indoor environments with distinct layouts: a rule-based connected component method; three traditional machine learning models, namely k-nearest neighbors, random forest, and support vector machine; and a deep learning model combining a convolutional neural network and long short term memory. In the training environment, the convolutional neural network long short term memory model achieved the highest accuracy, while traditional machine learning models provided moderate performance. In a new layout, however, all learning based methods showed significant degradation, whereas the rule-based method remained stable. Notably, for binary detection of presence versus absence of people, all models consistently achieved high accuracy across layouts. These results demonstrate that high capacity models can produce fine grained outputs with high accuracy in the same environment, but they are vulnerable to domain shift. In contrast, rule-based methods cannot provide fine grained outputs but exhibit robustness against domain shift. Moreover, regardless of the model type, a clear trade off was revealed between spatial generalization performance and output granularity.

Method: Dual-stage training on ~2.8M image-caption pairs with a novel alignment mechanism to suppress semantic drift. Incorporates Correlation-Regulated Alignment Framework for Tissue Synthesis (CRAFTS) with ControlNet integration for precise architectural control.

Result: Generates diverse pathological images across 30 cancer types with quality validated by metrics and pathologist evaluations. CRAFTS-augmented datasets enhance performance in classification, cross-modal retrieval, self-supervised learning, and visual question answering.

Conclusion: CRAFTS overcomes data scarcity and privacy barriers by providing limitless, diverse annotated histology data, enabling robust diagnostic tools for rare and complex cancer phenotypes.

Abstract: The development of clinical-grade artificial intelligence in pathology is limited by the scarcity of diverse, high-quality annotated datasets. Generative models offer a potential solution but suffer from semantic instability and morphological hallucinations that compromise diagnostic reliability. To address this challenge, we introduce a Correlation-Regulated Alignment Framework for Tissue Synthesis (CRAFTS), the first generative foundation model for pathology-specific text-to-image synthesis. By leveraging a dual-stage training strategy on approximately 2.8 million image-caption pairs, CRAFTS incorporates a novel alignment mechanism that suppresses semantic drift to ensure biological accuracy. This model generates diverse pathological images spanning 30 cancer types, with quality rigorously validated by objective metrics and pathologist evaluations. Furthermore, CRAFTS-augmented datasets enhance the performance across various clinical tasks, including classification, cross-modal retrieval, self-supervised learning, and visual question answering. In addition, coupling CRAFTS with ControlNet enables precise control over tissue architecture from inputs such as nuclear segmentation masks and fluorescence images. By overcoming the critical barriers of data scarcity and privacy concerns, CRAFTS provides a limitless source of diverse, annotated histology data, effectively unlocking the creation of robust diagnostic tools for rare and complex cancer phenotypes.

Method: Proposes Bidirectional Image-Guided Concept Erasure (Bi-Erasing) with two decoupled image branches: negative branch suppresses harmful semantics, positive branch provides visual guidance for safe alternatives. Uses joint text-image representations and mask-based filtering to prevent interference from irrelevant content.

Result: Bi-Erasing outperforms baseline methods in balancing concept removal effectiveness and visual fidelity across extensive experimental evaluations.

Conclusion: The bidirectional approach achieves better trade-off between erasure efficacy and generation usability compared to unidirectional methods, making it a more effective solution for concept removal in text-to-image models.

Abstract: Concept erasure, which fine-tunes diffusion models to remove undesired or harmful visual concepts, has become a mainstream approach to mitigating unsafe or illegal image generation in text-to-image models.However, existing removal methods typically adopt a unidirectional erasure strategy by either suppressing the target concept or reinforcing safe alternatives, making it difficult to achieve a balanced trade-off between concept removal and generation quality. To address this limitation, we propose a novel Bidirectional Image-Guided Concept Erasure (Bi-Erasing) framework that performs concept suppression and safety enhancement simultaneously. Specifically, based on the joint representation of text prompts and corresponding images, Bi-Erasing introduces two decoupled image branches: a negative branch responsible for suppressing harmful semantics and a positive branch providing visual guidance for safe alternatives. By jointly optimizing these complementary directions, our approach achieves a balance between erasure efficacy and generation usability. In addition, we apply mask-based filtering to the image branches to prevent interference from irrelevant content during the erasure process. Across extensive experiment evaluations, the proposed Bi-Erasing outperforms baseline methods in balancing concept removal effectiveness and visual fidelity.

Method: 1) Asymmetric framework with high-capacity gallery model for offline feature extraction and lightweight query network for online processing. 2) Geographical memory bank that structures gallery features using geolocation metadata to avoid expensive k-NN computations. 3) Implicit embedding augmentation to enhance query network’s ability to model feature variations despite limited capacity.

Result: Method significantly reduces computational costs while outperforming existing asymmetric retrieval techniques, establishing new state-of-the-art for VPR in resource-limited environments.

Conclusion: Proposed framework enables practical deployment of high-performance VPR on resource-constrained devices through efficient asymmetric design, geographical memory organization, and query network enhancement techniques.

Abstract: Visual Place Recognition (VPR) has advanced significantly with high-capacity foundation models like DINOv2, achieving remarkable performance. Nonetheless, their substantial computational cost makes deployment on resource-constrained devices impractical. In this paper, we introduce an efficient asymmetric VPR framework that incorporates a high-capacity gallery model for offline feature extraction with a lightweight query network for online processing. A key challenge in this setting is ensuring compatibility between these heterogeneous networks, which conventional approaches address through computationally expensive k-NN-based compatible training. To overcome this, we propose a geographical memory bank that structures gallery features using geolocation metadata inherent in VPR databases, eliminating the need for exhaustive k-NN computations. Additionally, we introduce an implicit embedding augmentation technique that enhances the query network to model feature variations despite its limited capacity. Extensive experiments demonstrate that our method not only significantly reduces computational costs but also outperforms existing asymmetric retrieval techniques, establishing a new aspect for VPR in resource-limited environments. The code is available at https://github.com/jaeyoon1603/AsymVPR

Method: 1) Uses an 18M multimodal medical retrieval database from PubMed; 2) Integrates RAG into CL with multi-modal, multi-layer RAG system for real-time guidance; 3) Introduces dynamic knowledge distillation framework that modulates parameter space importance, knowledge granularity, and reference dataset distribution based on required detail level.

Result: Achieves state-of-the-art performance across all metrics on the proposed Medical Generalist Task Incremental Learning (MGTIL) benchmark, demonstrating superior adaptation to domain shifts, retention of intra-domain features, and real-time learning of complex medical tasks.

Conclusion: The proposed framework effectively resolves the core dilemma in multimodal biomedical CL by combining RAG with dynamic knowledge distillation, enabling robust performance in challenging medical continual learning scenarios with significant clinical value.

Abstract: Multimodal biomedical Vision-Language Models (VLMs) exhibit immense potential in the field of Continual Learning (CL). However, they confront a core dilemma: how to preserve fine-grained intra-modality features while bridging the significant domain gap across different modalities. To address this challenge, we propose a comprehensive framework. Leveraging our 18-million multimodal and comprehensive medical retrieval database derived from PubMed scientific papers, we pioneer the integration of Retrieval-Augmented Generation (RAG) into CL. Specifically, we employ a multi-modal, multi-layer RAG system that provides real-time guidance for model fine-tuning through dynamic, on-demand knowledge retrieval. Building upon this, we introduce a dynamic knowledge distillation framework. This framework precisely resolves the aforementioned core dilemma by dynamically modulating the importance of the parameter space, the granularity of the distilled knowledge, and the data distribution of the reference dataset in accordance with the required level of detail. To thoroughly validate the clinical value of our strategy, we have designed a more rigorous \textbf{M}edical Generalist Task Incremental Learning (MGTIL) benchmark. This benchmark is engineered to simultaneously evaluate the model’s capacity for adaptation to significant domain shifts, retention of subtle intra-domain features, and real-time learning of novel and complex medical tasks. Extensive experimental results demonstrate that our proposed method achieves state-of-the-art (SOTA) performance across all metrics. The code is provided in the supplementary materials.

Method: Introduces Diversity Regularizer (DiRe) composed of cosine similarity and Euclidean distance metrics, which can be applied off-the-shelf to various state-of-the-art condensation methods to promote diversity.

Result: Extensive experiments show DiRe improves state-of-the-art condensation methods on benchmark datasets from CIFAR-10 to ImageNet-1K, enhancing both generalization performance and diversity metrics.

Conclusion: The proposed DiRe regularizer effectively addresses redundancy issues in dataset condensation, providing a simple yet effective enhancement to existing methods that improves both diversity and generalization capabilities.

Abstract: In Dataset Condensation, the goal is to synthesize a small dataset that replicates the training utility of a large original dataset. Existing condensation methods synthesize datasets with significant redundancy, so there is a dire need to reduce redundancy and improve the diversity of the synthesized datasets. To tackle this, we propose an intuitive Diversity Regularizer (DiRe) composed of cosine similarity and Euclidean distance, which can be applied off-the-shelf to various state-of-the-art condensation methods. Through extensive experiments, we demonstrate that the addition of our regularizer improves state-of-the-art condensation methods on various benchmark datasets from CIFAR-10 to ImageNet-1K with respect to generalization and diversity metrics.

Method: Introduces Causal Scene Graph (CSG) and Physical Alignment Probe (PAP) dataset for diagnostic interventions, then proposes LINA framework that learns prompt-specific interventions with targeted guidance in prompt/visual latent spaces and causality-aware denoising schedule.

Result: Achieves state-of-the-art performance on challenging causal generation tasks and Winoground dataset, enforcing both physical alignment and OOD instruction following in image and video DMs.

Conclusion: The approach successfully addresses diffusion models’ limitations in causal reasoning through adaptive interventions and reallocated denoising schedules, enabling better physical alignment and instruction following.

Abstract: Diffusion models (DMs) have achieved remarkable success in image and video generation. However, they still struggle with (1) physical alignment and (2) out-of-distribution (OOD) instruction following. We argue that these issues stem from the models’ failure to learn causal directions and to disentangle causal factors for novel recombination. We introduce the Causal Scene Graph (CSG) and the Physical Alignment Probe (PAP) dataset to enable diagnostic interventions. This analysis yields three key insights. First, DMs struggle with multi-hop reasoning for elements not explicitly determined in the prompt. Second, the prompt embedding contains disentangled representations for texture and physics. Third, visual causal structure is disproportionately established during the initial, computationally limited denoising steps. Based on these findings, we introduce LINA (Learning INterventions Adaptively), a novel framework that learns to predict prompt-specific interventions, which employs (1) targeted guidance in the prompt and visual latent spaces, and (2) a reallocated, causality-aware denoising schedule. Our approach enforces both physical alignment and OOD instruction following in image and video DMs, achieving state-of-the-art performance on challenging causal generation tasks and the Winoground dataset. Our project page is at https://opencausalab.github.io/LINA.

Method: ADHint introduces three key components: 1) Adaptive Hint with Sample Difficulty Prior for difficulty-aware hint ratio scheduling, 2) Consistency-based Gradient Modulation and Selective Masking for Hint Preservation to prevent biased updates, and 3) Advantage Estimation with Rollout Difficulty Posterior for balanced advantage estimation based on relative difficulty.

Result: Extensive experiments across diverse modalities, model scales, and domains show ADHint delivers superior reasoning ability and out-of-distribution generalization, consistently surpassing existing methods in both pass@1 and avg@8 metrics.

Conclusion: ADHint effectively addresses the limitations of existing hint-based RL methods by incorporating difficulty as a key factor, achieving better trade-off between exploration and imitation for improved reasoning generalization.

Abstract: To combine the advantages of Supervised Fine-Tuning (SFT) and Reinforcement Learning (RL), recent methods have integrated ‘‘hints’’ into post-training, which are prefix segments of complete reasoning trajectories, aiming for powerful knowledge expansion and reasoning generalization. However, existing hint-based RL methods typically ignore difficulty when scheduling hint ratios and estimating relative advantages, leading to unstable learning and excessive imitation of off-policy hints. In this work, we propose ADHint, which treats difficulty as a key factor in both hint-ratio schedule and relative-advantage estimation to achieve a better trade-off between exploration and imitation. Specifically, we propose Adaptive Hint with Sample Difficulty Prior, which evaluates each sample’s difficulty under the policy model and accordingly schedules an appropriate hint ratio to guide its rollouts. We also introduce Consistency-based Gradient Modulation and Selective Masking for Hint Preservation to modulate token-level gradients within hints, preventing biased and destructive updates. Additionally, we propose Advantage Estimation with Rollout Difficulty Posterior, which leverages the relative difficulty of rollouts with and without hints to estimate their respective advantages, thereby achieving more balanced updates. Extensive experiments across diverse modalities, model scales, and domains demonstrate that ADHint delivers superior reasoning ability and out-of-distribution generalization, consistently surpassing existing methods in both pass@1 and avg@8. Our code and dataset will be made publicly available upon paper acceptance.

Method: Proposes FID-Net based on YOLOv8n with three key components: 1) Feature Enhancement Module (FEM) for disease-sensitive cues, 2) Adaptive Multi-scale Feature Fusion Module (AMFM) to align RGB and FEM-enhanced features, and 3) Efficient Channel Attention (ECA) for discriminative information. Also develops spatial analysis framework using Kernel Density Estimation, neighborhood evaluation, and DBSCAN clustering.

Result: Achieves 86.10% precision, 75.44% recall, 82.29% mAP@0.5, and 64.30% mAP@0.5:0.95 on UAV imagery from 32 forest plots in eastern Tianshan, China, outperforming mainstream YOLO models. Analysis confirms infected trees exhibit clear clustering patterns.

Conclusion: FID-Net enables accurate tree health discrimination and, combined with spatial metrics, provides reliable data for intelligent pest monitoring, early warning, and precise forest management, supporting targeted protection strategies.

Abstract: Forest pests threaten ecosystem stability, requiring efficient monitoring. To overcome the limitations of traditional methods in large-scale, fine-grained detection, this study focuses on accurately identifying infected trees and analyzing infestation patterns. We propose FID-Net, a deep learning model that detects pest-affected trees from UAV visible-light imagery and enables infestation analysis via three spatial metrics. Based on YOLOv8n, FID-Net introduces a lightweight Feature Enhancement Module (FEM) to extract disease-sensitive cues, an Adaptive Multi-scale Feature Fusion Module (AMFM) to align and fuse dual-branch features (RGB and FEM-enhanced), and an Efficient Channel Attention (ECA) mechanism to enhance discriminative information efficiently. From detection results, we construct a pest situation analysis framework using: (1) Kernel Density Estimation to locate infection hotspots; (2) neighborhood evaluation to assess healthy trees’ infection risk; (3) DBSCAN clustering to identify high-density healthy clusters as priority protection zones. Experiments on UAV imagery from 32 forest plots in eastern Tianshan, China, show that FID-Net achieves 86.10% precision, 75.44% recall, 82.29% mAP@0.5, and 64.30% mAP@0.5:0.95, outperforming mainstream YOLO models. Analysis confirms infected trees exhibit clear clustering, supporting targeted forest protection. FID-Net enables accurate tree health discrimination and, combined with spatial metrics, provides reliable data for intelligent pest monitoring, early warning, and precise management.

Method: LeafTrackNet combines a YOLOv10-based leaf detector with a MobileNetV3-based embedding network. During inference, leaf identities are maintained through an embedding-based memory association strategy.

Result: LeafTrackNet outperforms both plant-specific trackers and state-of-the-art MOT baselines, achieving a 9% HOTA improvement on the CanolaTrack dataset (5,704 images with 31,840 annotated leaf instances).

Conclusion: The work provides a new standard for leaf-level tracking under realistic conditions and contributes the largest dataset for leaf tracking in agricultural crops, which will advance future research in plant phenotyping.

Abstract: High resolution phenotyping at the level of individual leaves offers fine-grained insights into plant development and stress responses. However, the full potential of accurate leaf tracking over time remains largely unexplored due to the absence of robust tracking methods-particularly for structurally complex crops such as canola. Existing plant-specific tracking methods are typically limited to small-scale species or rely on constrained imaging conditions. In contrast, generic multi-object tracking (MOT) methods are not designed for dynamic biological scenes. Progress in the development of accurate leaf tracking models has also been hindered by a lack of large-scale datasets captured under realistic conditions. In this work, we introduce CanolaTrack, a new benchmark dataset comprising 5,704 RGB images with 31,840 annotated leaf instances spanning the early growth stages of 184 canola plants. To enable accurate leaf tracking over time, we introduce LeafTrackNet, an efficient framework that combines a YOLOv10-based leaf detector with a MobileNetV3-based embedding network. During inference, leaf identities are maintained over time through an embedding-based memory association strategy. LeafTrackNet outperforms both plant-specific trackers and state-of-the-art MOT baselines, achieving a 9% HOTA improvement on CanolaTrack. With our work we provide a new standard for leaf-level tracking under realistic conditions and we provide CanolaTrack - the largest dataset for leaf tracking in agriculture crops, which will contribute to future research in plant phenotyping. Our code and dataset are publicly available at https://github.com/shl-shawn/LeafTrackNet.

Method: Uses pre-trained MDE model to generate relative depth images, segments and randomly rescales them to create synthetic training pairs (dense pseudo-ground truth + sparse depths). Trains a refinement network with these synthetic pairs, incorporating relative depth maps and RGB images to improve accuracy and robustness.

Result: StarryGazer shows superior results over existing unsupervised methods and transformed MDE results on various datasets, demonstrating effective exploitation of MDE power while appropriately fixing errors using sparse depth information.

Conclusion: The framework successfully bridges the gap between MDE models and depth completion by creating synthetic training data and refinement, enabling domain-agnostic depth completion without ground truth depth data.

Abstract: The problem of depth completion involves predicting a dense depth image from a single sparse depth map and an RGB image. Unsupervised depth completion methods have been proposed for various datasets where ground truth depth data is unavailable and supervised methods cannot be applied. However, these models require auxiliary data to estimate depth values, which is far from real scenarios. Monocular depth estimation (MDE) models can produce a plausible relative depth map from a single image, but there is no work to properly combine the sparse depth map with MDE for depth completion; a simple affine transformation to the depth map will yield a high error since MDE are inaccurate at estimating depth difference between objects. We introduce StarryGazer, a domain-agnostic framework that predicts dense depth images from a single sparse depth image and an RGB image without relying on ground-truth depth by leveraging the power of large MDE models. First, we employ a pre-trained MDE model to produce relative depth images. These images are segmented and randomly rescaled to form synthetic pairs for dense pseudo-ground truth and corresponding sparse depths. A refinement network is trained with the synthetic pairs, incorporating the relative depth maps and RGB images to improve the model’s accuracy and robustness. StarryGazer shows superior results over existing unsupervised methods and transformed MDE results on various datasets, demonstrating that our framework exploits the power of MDE models while appropriately fixing errors using sparse depth information.

Method: Proposes a dispersion-based unlearning approach that disperses target identity embeddings on the hypersphere to prevent compact identity cluster formation. Evaluates existing approximate class unlearning methods (Random Labeling, Gradient Ascent, Boundary Unlearning) and introduces their own method.

Result: Extensive experiments on VGGFace2 and CelebA benchmarks show superior forgetting behavior while preserving retrieval utility compared to existing methods.

Conclusion: The dispersion-based approach effectively addresses face identity unlearning challenges, achieving privacy protection through identity forgetting while maintaining overall system performance for remaining identities.

Abstract: Face recognition systems rely on learning highly discriminative and compact identity clusters to enable accurate retrieval. However, as with other surveillance-oriented technologies, such systems raise serious privacy concerns due to their potential for unauthorized identity tracking. While several works have explored machine unlearning as a means of privacy protection, their applicability to face retrieval - especially for modern embedding-based recognition models - remains largely unexplored. In this work, we study the problem of face identity unlearning for retrieval systems and present its inherent challenges. The goal is to make selected identities unretrievable by dispersing their embeddings on the hypersphere and preventing the formation of compact identity clusters that enable re-identification in the gallery. The primary challenge is to achieve this forgetting effect while preserving the discriminative structure of the embedding space and the retrieval performance of the model for the remaining identities. To address this, we evaluate several existing approximate class unlearning methods (e.g., Random Labeling, Gradient Ascent, Boundary Unlearning, and other recent approaches) in the context of face retrieval and propose a simple yet effective dispersion-based unlearning approach. Extensive experiments on standard benchmarks (VGGFace2, CelebA) demonstrate that our method achieves superior forgetting behavior while preserving retrieval utility.

Method: DFSS uses Batch Normalization statistics from the teacher model to guide Approximate Distribution Sampling (ADS) for selecting data that better reflects the original training distribution. It also introduces Weighted Distribution Progressive Distillation (WDPD) that dynamically prioritizes reliable samples aligned with original data distribution early in training and gradually incorporates more challenging cases.

Result: Extensive experiments on standard benchmarks show DFSS consistently outperforms existing data-free distillation methods for semantic segmentation, achieving state-of-the-art results with significantly reduced reliance on auxiliary data.

Conclusion: DFSS provides an effective data-free distillation framework tailored for semantic segmentation that respects structural continuity and leverages teacher model statistics to better approximate the original training distribution, achieving superior performance without requiring original training data.

Abstract: Semantic segmentation requires a holistic understanding of the physical world, as it assigns semantic labels to spatially continuous and structurally coherent objects rather than to isolated pixels. However, existing data-free knowledge distillation (DFKD) methods-primarily designed for classification-often disregard this continuity, resulting in significant performance degradation when applied directly to segmentation tasks. In this paper, we introduce DFSS, a novel data-free distillation framework tailored for semantic segmentation. Unlike prior approaches that treat pixels independently, DFSS respects the structural and contextual continuity of real-world scenes. Our key insight is to leverage Batch Normalization (BN) statistics from a teacher model to guide Approximate Distribution Sampling (ADS), enabling the selection of data that better reflects the original training distribution-without relying on potentially misleading teacher predictions. Additionally, we propose Weighted Distribution Progressive Distillation (WDPD), which dynamically prioritizes reliable samples that are more closely aligned with the original data distribution early in training and gradually incorporates more challenging cases, mirroring the natural progression of learning in human perception. Extensive experiments on standard benchmarks demonstrate that DFSS consistently outperforms existing data-free distillation methods for semantic segmentation, achieving state-of-the-art results with significantly reduced reliance on auxiliary data.

Method: Proposes MMDrive framework with three complementary modalities: occupancy maps, LiDAR point clouds, and textual scene descriptions. Introduces two novel components: Text-oriented Multimodal Modulator for adaptive cross-modal fusion based on semantic cues, and Cross-Modal Abstractor using learnable abstract tokens to generate compact cross-modal summaries.

Result: Achieves significant performance gains: BLEU-4 score of 54.56 and METEOR of 41.78 on DriveLM benchmark, and 62.7% accuracy on NuScenes-QA benchmark, outperforming existing vision-language models for autonomous driving.

Conclusion: MMDrive effectively breaks the traditional image-only understanding barrier, enabling robust multimodal reasoning in complex driving environments and providing a new foundation for interpretable autonomous driving scene understanding.

Abstract: Vision-language models enable the understanding and reasoning of complex traffic scenarios through multi-source information fusion, establishing it as a core technology for autonomous driving. However, existing vision-language models are constrained by the image understanding paradigm in 2D plane, which restricts their capability to perceive 3D spatial information and perform deep semantic fusion, resulting in suboptimal performance in complex autonomous driving environments. This study proposes MMDrive, an multimodal vision-language model framework that extends traditional image understanding to a generalized 3D scene understanding framework. MMDrive incorporates three complementary modalities, including occupancy maps, LiDAR point clouds, and textual scene descriptions. To this end, it introduces two novel components for adaptive cross-modal fusion and key information extraction. Specifically, the Text-oriented Multimodal Modulator dynamically weights the contributions of each modality based on the semantic cues in the question, guiding context-aware feature integration. The Cross-Modal Abstractor employs learnable abstract tokens to generate compact, cross-modal summaries that highlight key regions and essential semantics. Comprehensive evaluations on the DriveLM and NuScenes-QA benchmarks demonstrate that MMDrive achieves significant performance gains over existing vision-language models for autonomous driving, with a BLEU-4 score of 54.56 and METEOR of 41.78 on DriveLM, and an accuracy score of 62.7% on NuScenes-QA. MMDrive effectively breaks the traditional image-only understanding barrier, enabling robust multimodal reasoning in complex driving environments and providing a new foundation for interpretable autonomous driving scene understanding.

Method: CoRA uses a hybrid approach with two branches: 1) feature-level fusion branch that selects critical features for efficient fusion, and 2) object-level correction branch that leverages semantic relevance to correct spatial displacements from pose errors.

Result: Under extreme scenarios, CoRA improves baseline performance by ~19% in AP@0.7 with >5x less communication volume, demonstrating superior robustness and efficiency.

Conclusion: CoRA provides a promising solution for robust collaborative perception by decoupling performance from robustness with low communication, making it practical for real-world deployment.

Abstract: Collaborative perception has garnered significant attention as a crucial technology to overcome the perceptual limitations of single-agent systems. Many state-of-the-art (SOTA) methods have achieved communication efficiency and high performance via intermediate fusion. However, they share a critical vulnerability: their performance degrades under adverse communication conditions due to the misalignment induced by data transmission, which severely hampers their practical deployment. To bridge this gap, we re-examine different fusion paradigms, and recover that the strengths of intermediate and late fusion are not a trade-off, but a complementary pairing. Based on this key insight, we propose CoRA, a novel collaborative robust architecture with a hybrid approach to decouple performance from robustness with low communication. It is composed of two components: a feature-level fusion branch and an object-level correction branch. Its first branch selects critical features and fuses them efficiently to ensure both performance and scalability. The second branch leverages semantic relevance to correct spatial displacements, guaranteeing resilience against pose errors. Experiments demonstrate the superiority of CoRA. Under extreme scenarios, CoRA improves upon its baseline performance by approximately 19% in AP@0.7 with more than 5x less communication volume, which makes it a promising solution for robust collaborative perception.

Method: End-to-end deep learning framework that jointly optimizes segmentation and registration. The network learns segmentation masks specifically optimized for registration, guided solely by the registration objective without direct segmentation supervision, eliminating manual steps.

Result: Achieves state-of-the-art performance: reduces median Target Registration Error by 32% to 1.83mm and mean Root Mean Square Error by 45% to 3.95mm. Ablation study confirms end-to-end optimization significantly improves registration accuracy.

Conclusion: The end-to-end RGB-D registration pipeline removes dependency on weak labels and manual steps, advancing toward fully automatic, markerless intraoperative navigation for spine surgery.

Abstract: Purpose: Intraoperative navigation in spine surgery demands millimeter-level accuracy. Current systems based on intraoperative radiographic imaging and bone-anchored markers are invasive, radiation-intensive and workflow disruptive. Recent markerless RGB-D registration methods offer a promising alternative, but existing approaches rely on weak segmentation labels to isolate relevant anatomical structures, which can propagate errors throughout registration. Methods: We present End2Reg an end-to-end deep learning framework that jointly optimizes segmentation and registration, eliminating the need for weak segmentation labels and manual steps. The network learns segmentation masks specifically optimized for registration, guided solely by the registration objective without direct segmentation supervision. Results: The proposed framework achieves state-of-the-art performance on ex- and in-vivo benchmarks, reducing median Target Registration Error by 32% to 1.83mm and mean Root Mean Square Error by 45% to 3.95mm, respectively. An ablation study confirms that end-to-end optimization significantly improves registration accuracy. Conclusion: The presented end-to-end RGB-D registration pipeline removes dependency on weak labels and manual steps, advancing towards fully automatic, markerless intraoperative navigation. Code and interactive visualizations are available at: https://lorenzopettinari.github.io/end-2-reg/.

Method: 1) POLAR dataset: 200+ subjects under 156 lighting directions with multiple views and expressions. 2) POLARNet: flow-based generative model predicting per-light OLAT responses from single portrait, modeling illumination as continuous physically interpretable transformation between lighting states.

Result: Creates unified illumination learning framework linking real data, generative synthesis, and physically grounded relighting. Enables scalable and controllable relighting while preserving facial identity and capturing fine-grained direction-aware illumination effects.

Conclusion: POLAR and POLARNet establish a self-sustaining “chicken-and-egg” cycle for scalable and reproducible portrait illumination, moving beyond diffusion/background-conditioned methods to physically interpretable continuous illumination modeling.

Abstract: Face relighting aims to synthesize realistic portraits under novel illumination while preserving identity and geometry. However, progress remains constrained by the limited availability of large-scale, physically consistent illumination data. To address this, we introduce POLAR, a large-scale and physically calibrated One-Light-at-a-Time (OLAT) dataset containing over 200 subjects captured under 156 lighting directions, multiple views, and diverse expressions. Building upon POLAR, we develop a flow-based generative model POLARNet that predicts per-light OLAT responses from a single portrait, capturing fine-grained and direction-aware illumination effects while preserving facial identity. Unlike diffusion or background-conditioned methods that rely on statistical or contextual cues, our formulation models illumination as a continuous, physically interpretable transformation between lighting states, enabling scalable and controllable relighting. Together, POLAR and POLARNet form a unified illumination learning framework that links real data, generative synthesis, and physically grounded relighting, establishing a self-sustaining “chicken-and-egg” cycle for scalable and reproducible portrait illumination.

Method: Used a “Wizard of OZ” data collection paradigm where experts act as wearable assistants, viewing trainees’ egocentric video feeds and providing natural language guidance during procedural activities.

Result: Created a challenging dataset with 15k+ high-quality Visual Question Answer sets that reveals limitations of current multimodal models in providing expert-level assistance.

Conclusion: Ego-EXTRA provides a valuable benchmark for developing and evaluating egocentric video-language assistants, highlighting the need for improved models that can offer expert-level guidance.

Abstract: We present Ego-EXTRA, a video-language Egocentric Dataset for EXpert-TRAinee assistance. Ego-EXTRA features 50 hours of unscripted egocentric videos of subjects performing procedural activities (the trainees) while guided by real-world experts who provide guidance and answer specific questions using natural language. Following a ``Wizard of OZ’’ data collection paradigm, the expert enacts a wearable intelligent assistant, looking at the activities performed by the trainee exclusively from their egocentric point of view, answering questions when asked by the trainee, or proactively interacting with suggestions during the procedures. This unique data collection protocol enables Ego-EXTRA to capture a high-quality dialogue in which expert-level feedback is provided to the trainee. Two-way dialogues between experts and trainees are recorded, transcribed, and used to create a novel benchmark comprising more than 15k high-quality Visual Question Answer sets, which we use to evaluate Multimodal Large Language Models. The results show that Ego-EXTRA is challenging and highlight the limitations of current models when used to provide expert-level assistance to the user. The Ego-EXTRA dataset is publicly available to support the benchmark of egocentric video-language assistants: https://fpv-iplab.github.io/Ego-EXTRA/.

Method: STARCaster uses a compositional approach: 1) ID-aware motion modeling with softer identity constraints instead of strict reference conditioning, 2) audio-visual synchronization via lip reading supervision, 3) novel view animation through temporal-to-spatial adaptation. It employs decoupled learning for view consistency and temporal coherence, plus a self-forcing training scheme for longer temporal contexts.

Result: Comprehensive evaluations show STARCaster generalizes effectively across tasks and identities, consistently surpassing prior approaches in different benchmarks, producing more diverse motions and better identity preservation than existing methods.

Conclusion: STARCaster successfully addresses both speech-driven portrait animation and free-viewpoint talking portrait synthesis within a unified framework, overcoming limitations of existing 2D and 3D approaches through innovative design choices like softer identity constraints, implicit 3D awareness, and decoupled learning strategies.

Abstract: This paper presents STARCaster, an identity-aware spatio-temporal video diffusion model that addresses both speech-driven portrait animation and free-viewpoint talking portrait synthesis, given an identity embedding or reference image, within a unified framework. Existing 2D speech-to-video diffusion models depend heavily on reference guidance, leading to limited motion diversity. At the same time, 3D-aware animation typically relies on inversion through pre-trained tri-plane generators, which often leads to imperfect reconstructions and identity drift. We rethink reference- and geometry-based paradigms in two ways. First, we deviate from strict reference conditioning at pre-training by introducing softer identity constraints. Second, we address 3D awareness implicitly within the 2D video domain by leveraging the inherent multi-view nature of video data. STARCaster adopts a compositional approach progressing from ID-aware motion modeling, to audio-visual synchronization via lip reading-based supervision, and finally to novel view animation through temporal-to-spatial adaptation. To overcome the scarcity of 4D audio-visual data, we propose a decoupled learning approach in which view consistency and temporal coherence are trained independently. A self-forcing training scheme enables the model to learn from longer temporal contexts than those generated at inference, mitigating the overly static animations common in existing autoregressive approaches. Comprehensive evaluations demonstrate that STARCaster generalizes effectively across tasks and identities, consistently surpassing prior approaches in different benchmarks.

Method: Introduces VG-AVS task for next viewpoint selection using only current visual info. Creates synthetic dataset with paired query-target views and QA prompts. Fine-tunes pretrained VLMs with supervised fine-tuning followed by RL-based policy optimization.

Result: Achieves strong QA performance based on viewpoint selection, generalizes to unseen synthetic/real scenes, and improves downstream QA accuracy when integrated into existing scene-exploration EQA systems.

Conclusion: VG-AVS successfully bridges snapshot vision to ambulatory vision, enabling VLMs to actively select informative viewpoints and enhancing embodied question answering capabilities.

Abstract: Vision Language Models (VLMs) excel at visual question answering (VQA) but remain limited to snapshot vision, reasoning from static images. In contrast, embodied agents require ambulatory vision, actively moving to obtain more informative views. We introduce Visually Grounded Active View Selection (VG-AVS), a task that selects the most informative next viewpoint using only the visual information in the current image, without relying on scene memory or external knowledge. To support this task, we construct a synthetic dataset with automatically generated paired query-target views and question-answer prompts. We also propose a framework that fine-tunes pretrained VLMs through supervised fine-tuning (SFT) followed by RL-based policy optimization. Our approach achieves strong question answering performance based on viewpoint selection and generalizes robustly to unseen synthetic and real scenes. Furthermore, incorporating our learned VG-AVS framework into existing scene-exploration-based EQA systems improves downstream question-answering accuracy.

Method: CogniEdit combines multi-modal reasoning with dense reward optimization: (1) Multi-modal LLMs decompose complex instructions into actionable directives, (2) Dynamic Token Focus Relocation adaptively emphasizes fine-grained attributes, and (3) Dense GRPO-based optimization propagates gradients across consecutive denoising steps for trajectory-level supervision.

Result: Extensive experiments on benchmark datasets demonstrate that CogniEdit achieves state-of-the-art performance in balancing fine-grained instruction following with visual quality and editability preservation.

Conclusion: The proposed CogniEdit framework successfully addresses limitations of existing methods by enabling trajectory-level gradient flow through the sampling process, achieving superior fine-grained instruction following while maintaining visual quality and editability.

Abstract: Instruction-based image editing with diffusion models has achieved impressive results, yet existing methods struggle with fine-grained instructions specifying precise attributes such as colors, positions, and quantities. While recent approaches employ Group Relative Policy Optimization (GRPO) for alignment, they optimize only at individual sampling steps, providing sparse feedback that limits trajectory-level control. We propose a unified framework CogniEdit, combining multi-modal reasoning with dense reward optimization that propagates gradients across consecutive denoising steps, enabling trajectory-level gradient flow through the sampling process. Our method comprises three components: (1) Multi-modal Large Language Models for decomposing complex instructions into actionable directives, (2) Dynamic Token Focus Relocation that adaptively emphasizes fine-grained attributes, and (3) Dense GRPO-based optimization that propagates gradients across consecutive steps for trajectory-level supervision. Extensive experiments on benchmark datasets demonstrate that our CogniEdit achieves state-of-the-art performance in balancing fine-grained instruction following with visual quality and editability preservation

Method: The benchmark uses carefully curated real ASMR videos as source material, focusing on fine-grained action-object interactions. It employs an adversarial creator-reviewer protocol where video generation models act as creators trying to fool reviewers, while VLMs serve as reviewers trying to detect fakeness.

Result: The best creator model (Veo3.1-Fast) fools most VLMs, with the strongest reviewer (Gemini 2.5-Pro) achieving only 56% accuracy vs. random 50%. Human experts perform much better at 81.25%. Adding audio improves real-fake discrimination, but superficial cues like watermarks can still mislead models.

Conclusion: The findings delineate current boundaries of video generation realism and expose limitations of VLMs in perceptual fidelity and audio-visual consistency, showing that while AI-generated videos are becoming more realistic, there are still detectable differences that humans can identify better than current VLMs.

Abstract: Recent advances in video generation have produced vivid content that are often indistinguishable from real videos, making AI-generated video detection an emerging societal challenge. Prior AIGC detection benchmarks mostly evaluate video without audio, target broad narrative domains, and focus on classification solely. Yet it remains unclear whether state-of-the-art video generation models can produce immersive, audio-paired videos that reliably deceive humans and VLMs. To this end, we introduce Video Reality Test, an ASMR-sourced video benchmark suite for testing perceptual realism under tight audio-visual coupling, featuring the following dimensions: \textbf{(i) Immersive ASMR video-audio sources.} Built on carefully curated real ASMR videos, the benchmark targets fine-grained action-object interactions with diversity across objects, actions, and backgrounds. \textbf{(ii) Peer-Review evaluation.} An adversarial creator-reviewer protocol where video generation models act as creators aiming to fool reviewers, while VLMs serve as reviewers seeking to identify fakeness. Our experimental findings show: The best creator Veo3.1-Fast even fools most VLMs: the strongest reviewer (Gemini 2.5-Pro) achieves only 56% accuracy (random 50%), far below that of human experts (81.25%). Adding audio improves real-fake discrimination, yet superficial cues such as watermarks can still significantly mislead models. These findings delineate the current boundary of video generation realism and expose limitations of VLMs in perceptual fidelity and audio-visual consistency. Our code is available at https://github.com/video-reality-test/video-reality-test.

Method: Proposed a lightweight domain-adaptive self-supervised adaptor (DA-SSL) that realigns pretrained PFM features to the TURBT domain without fine-tuning the foundational model itself, integrated with multiple instance learning for predicting treatment response.

Result: DA-SSL achieved AUC of 0.77±0.04 in five-fold cross-validation, and external test accuracy of 0.84, sensitivity 0.71, specificity 0.91 using majority voting in multi-center study for predicting neoadjuvant chemotherapy response.

Conclusion: Lightweight domain adaptation with self-supervision effectively enhances PFM-based MIL pipelines for clinically challenging histopathology tasks, particularly for cancer types with domain shifts like bladder cancer TURBT specimens.

Abstract: Recent deep learning frameworks in histopathology, particularly multiple instance learning (MIL) combined with pathology foundational models (PFMs), have shown strong performance. However, PFMs exhibit limitations on certain cancer or specimen types due to domain shifts - these cancer types were rarely used for pretraining or specimens contain tissue-based artifacts rarely seen within the pretraining population. Such is the case for transurethral resection of bladder tumor (TURBT), which are essential for diagnosing muscle-invasive bladder cancer (MIBC), but contain fragmented tissue chips and electrocautery artifacts and were not widely used in publicly available PFMs. To address this, we propose a simple yet effective domain-adaptive self-supervised adaptor (DA-SSL) that realigns pretrained PFM features to the TURBT domain without fine-tuning the foundational model itself. We pilot this framework for predicting treatment response in TURBT, where histomorphological features are currently underutilized and identifying patients who will benefit from neoadjuvant chemotherapy (NAC) is challenging. In our multi-center study, DA-SSL achieved an AUC of 0.77+/-0.04 in five-fold cross-validation and an external test accuracy of 0.84, sensitivity of 0.71, and specificity of 0.91 using majority voting. Our results demonstrate that lightweight domain adaptation with self-supervision can effectively enhance PFM-based MIL pipelines for clinically challenging histopathology tasks. Code is Available at https://github.com/zhanghaoyue/DA_SSL_TURBT.

Method: Models generation process with structural causal model, uses Gumbel-Softmax-based feature masking and HSIC constraints to enforce statistical independence, isolating stable causal features robust to distribution shifts.

Result: Achieves 6.83% improvement in accuracy and 4.06% improvement in average precision over state-of-the-art methods when tested on unseen generative models from different series.

Conclusion: CausalCLIP’s explicit disentanglement of causal features enables strong generalization ability for generated image detection across diverse generation techniques.

Abstract: The rapid advancement of generative models has increased the demand for generated image detectors capable of generalizing across diverse and evolving generation techniques. However, existing methods, including those leveraging pre-trained vision-language models, often produce highly entangled representations, mixing task-relevant forensic cues (causal features) with spurious or irrelevant patterns (non-causal features), thus limiting generalization. To address this issue, we propose CausalCLIP, a framework that explicitly disentangles causal from non-causal features and employs targeted filtering guided by causal inference principles to retain only the most transferable and discriminative forensic cues. By modeling the generation process with a structural causal model and enforcing statistical independence through Gumbel-Softmax-based feature masking and Hilbert-Schmidt Independence Criterion (HSIC) constraints, CausalCLIP isolates stable causal features robust to distribution shifts. When tested on unseen generative models from different series, CausalCLIP demonstrates strong generalization ability, achieving improvements of 6.83% in accuracy and 4.06% in average precision over state-of-the-art methods.

Method: Proposes ShowTable pipeline that synergizes MLLMs with diffusion models via progressive self-correcting process. MLLM acts as central orchestrator for reasoning visual plans and judging errors, while diffusion models execute commands. Introduces three automated data construction pipelines for training different modules.

Result: Introduces TableVisBench benchmark with 800 challenging instances across 5 evaluation dimensions. Experiments show the pipeline significantly outperforms baselines, demonstrating effective multi-modal reasoning, generation, and error correction capabilities.

Conclusion: ShowTable successfully addresses the creative table visualization challenge through synergistic MLLM-diffusion integration with self-correction, automated data construction, and comprehensive benchmarking, advancing beyond general image generation limitations.

Abstract: While existing generation and unified models excel at general image generation, they struggle with tasks requiring deep reasoning, planning, and precise data-to-visual mapping abilities beyond general scenarios. To push beyond the existing limitations, we introduce a new and challenging task: creative table visualization, requiring the model to generate an infographic that faithfully and aesthetically visualizes the data from a given table. To address this challenge, we propose ShowTable, a pipeline that synergizes MLLMs with diffusion models via a progressive self-correcting process. The MLLM acts as the central orchestrator for reasoning the visual plan and judging visual errors to provide refined instructions, the diffusion execute the commands from MLLM, achieving high-fidelity results. To support this task and our pipeline, we introduce three automated data construction pipelines for training different modules. Furthermore, we introduce TableVisBench, a new benchmark with 800 challenging instances across 5 evaluation dimensions, to assess performance on this task. Experiments demonstrate that our pipeline, instantiated with different models, significantly outperforms baselines, highlighting its effective multi-modal reasoning, generation, and error correction capabilities.

Method: A spatio-temporal cascade framework that first generates low-resolution blueprint video keyframes for global semantics and motion, then refines them into high-resolution sub-clips using a first-last frame strategy. Introduces a Co-Reasoning Director with three modality-specific LLM experts to reason about modality priorities and infer user intent through multi-turn dialogue, plus a Negative Director for prompt refinement. Extends to ID-specific multi-character control.

Result: The model effectively addresses challenges of efficient, multimodally aligned long-form high-resolution video generation, delivering enhanced visual clarity, realistic lip-teeth rendering with accurate lip synchronization, strong identity preservation, and coherent multimodal instruction following.

Conclusion: KlingAvatar 2.0 successfully overcomes limitations in long-duration high-resolution avatar video generation through its spatio-temporal cascade approach and intelligent cross-modal reasoning system, enabling efficient production of high-quality, temporally coherent videos with strong identity preservation and instruction alignment.

Abstract: Avatar video generation models have achieved remarkable progress in recent years. However, prior work exhibits limited efficiency in generating long-duration high-resolution videos, suffering from temporal drifting, quality degradation, and weak prompt following as video length increases. To address these challenges, we propose KlingAvatar 2.0, a spatio-temporal cascade framework that performs upscaling in both spatial resolution and temporal dimension. The framework first generates low-resolution blueprint video keyframes that capture global semantics and motion, and then refines them into high-resolution, temporally coherent sub-clips using a first-last frame strategy, while retaining smooth temporal transitions in long-form videos. To enhance cross-modal instruction fusion and alignment in extended videos, we introduce a Co-Reasoning Director composed of three modality-specific large language model (LLM) experts. These experts reason about modality priorities and infer underlying user intent, converting inputs into detailed storylines through multi-turn dialogue. A Negative Director further refines negative prompts to improve instruction alignment. Building on these components, we extend the framework to support ID-specific multi-character control. Extensive experiments demonstrate that our model effectively addresses the challenges of efficient, multimodally aligned long-form high-resolution video generation, delivering enhanced visual clarity, realistic lip-teeth rendering with accurate lip synchronization, strong identity preservation, and coherent multimodal instruction following.

Method: Introduces SpurLens, a pipeline that uses GPT-4 and open-set object detectors to automatically identify spurious visual cues without human supervision. The method diagnoses spurious bias in MLLMs and explores mitigation strategies like prompt ensembling and reasoning-based prompting.

Result: Spurious correlations cause two major failure modes in MLLMs: (1) over-reliance on spurious cues for object recognition (removing cues reduces accuracy), and (2) object hallucination amplified by over 10x. Findings are validated across various MLLMs and datasets.

Conclusion: Spurious correlations persist in MLLMs despite language supervision, calling for more rigorous evaluation methods and mitigation strategies to enhance MLLM reliability. The study exposes these biases and explores potential solutions.

Abstract: Unimodal vision models are known to rely on spurious correlations, but it remains unclear to what extent Multimodal Large Language Models (MLLMs) exhibit similar biases despite language supervision. In this paper, we investigate spurious bias in MLLMs and introduce SpurLens, a pipeline that leverages GPT-4 and open-set object detectors to automatically identify spurious visual cues without human supervision. Our findings reveal that spurious correlations cause two major failure modes in MLLMs: (1) over-reliance on spurious cues for object recognition, where removing these cues reduces accuracy, and (2) object hallucination, where spurious cues amplify the hallucination by over 10x. We validate our findings in various MLLMs and datasets. Beyond diagnosing these failures, we explore potential mitigation strategies, such as prompt ensembling and reasoning-based prompting, and conduct ablation studies to examine the root causes of spurious bias in MLLMs. By exposing the persistence of spurious correlations, our study calls for more rigorous evaluation methods and mitigation strategies to enhance the reliability of MLLMs.

Method: Comparative analysis of infrared spectral ranges (NIR, SWIR, MWIR, LWIR), definition of thermal camera requirements (resolution, sensitivity, frame rate), and implementation of Siamese neural networks for automated identification on a proprietary dataset.

Result: Achieved approximately 80% accuracy in biometric identification using thermal imaging and Siamese neural networks, with analysis showing thermal imaging’s potential for security systems and benefits of hybrid visible-infrared approaches.

Conclusion: Thermal imaging is a promising technology for developing reliable biometric security systems, with Siamese neural networks providing effective automation, and hybrid visible-IR systems offering enhanced performance by overcoming individual modality limitations.

Abstract: The article analyzes the use of thermal imaging technologies for biometric identification based on facial thermograms. It presents a comparative analysis of infrared spectral ranges (NIR, SWIR, MWIR, and LWIR). The paper also defines key requirements for thermal cameras used in biometric systems, including sensor resolution, thermal sensitivity, and a frame rate of at least 30 Hz. Siamese neural networks are proposed as an effective approach for automating the identification process. In experiments conducted on a proprietary dataset, the proposed method achieved an accuracy of approximately 80%. The study also examines the potential of hybrid systems that combine visible and infrared spectra to overcome the limitations of individual modalities. The results indicate that thermal imaging is a promising technology for developing reliable security systems.

Method: Adapts ControlNet to image-to-3D generation for text steering in a forward pass. Uses scalable automatic data generation and two-stage training with flow-matching and Direct Preference Optimization (DPO).

Result: Steer3D follows language instructions more faithfully, maintains better consistency with original 3D assets, and is 2.4x to 28.5x faster than competing methods.

Conclusion: Demonstrates that text steerability can be added to pretrained image-to-3D models with 100k data, enabling efficient language-based editing of 3D assets for practical applications.

Abstract: Recent progress in image-to-3D has opened up immense possibilities for design, AR/VR, and robotics. However, to use AI-generated 3D assets in real applications, a critical requirement is the capability to edit them easily. We present a feedforward method, Steer3D, to add text steerability to image-to-3D models, which enables editing of generated 3D assets with language. Our approach is inspired by ControlNet, which we adapt to image-to-3D generation to enable text steering directly in a forward pass. We build a scalable data engine for automatic data generation, and develop a two-stage training recipe based on flow-matching training and Direct Preference Optimization (DPO). Compared to competing methods, Steer3D more faithfully follows the language instruction and maintains better consistency with the original 3D asset, while being 2.4x to 28.5x faster. Steer3D demonstrates that it is possible to add a new modality (text) to steer the generation of pretrained image-to-3D generative models with 100k data. Project website: https://glab-caltech.github.io/steer3d/

Method: Leverages DINO self-attention “key” features (instead of deepest ViT layer tokens) with a simple convolutional decoder to predict polyp masks. Validated using multi-center datasets under Domain Generalization (DG) and Extreme Single Domain Generalization (ESDG) protocols.

Result: Achieves state-of-the-art performance with comprehensive statistical analysis, significantly enhancing generalization in data-scarce and challenging scenarios. Surpasses established models like nnU-Net and UM-Net without polyp-specific architecture.

Conclusion: The framework demonstrates that leveraging DINO self-attention key features with simple decoder provides robust polyp segmentation with superior generalization, offering a systematic benchmark of DINO’s evolution impact on downstream performance.

Abstract: Automatic polyp segmentation is crucial for improving the clinical identification of colorectal cancer (CRC). While Deep Learning (DL) techniques have been extensively researched for this problem, current methods frequently struggle with generalization, particularly in data-constrained or challenging settings. Moreover, many existing polyp segmentation methods rely on complex, task-specific architectures. To address these limitations, we present a framework that leverages the intrinsic robustness of DINO self-attention “key” features for robust segmentation. Unlike traditional methods that extract tokens from the deepest layers of the Vision Transformer (ViT), our approach leverages the key features of the self-attention module with a simple convolutional decoder to predict polyp masks, resulting in enhanced performance and better generalizability. We validate our approach using a multi-center dataset under two rigorous protocols: Domain Generalization (DG) and Extreme Single Domain Generalization (ESDG). Our results, supported by a comprehensive statistical analysis, demonstrate that this pipeline achieves state-of-the-art (SOTA) performance, significantly enhancing generalization, particularly in data-scarce and challenging scenarios. While avoiding a polyp-specific architecture, we surpass well-established models like nnU-Net and UM-Net. Additionally, we provide a systematic benchmark of the DINO framework’s evolution, quantifying the specific impact of architectural advancements on downstream polyp segmentation performance.

Method: A lightweight decoder framework that generates multi-modal preview representations (RGB and scene intrinsics) at any point during denoising, enabling interactive guidance via stochasticity reinjection and modal steering.

Result: Generates previews at >4× real-time speed (<1 second for 4-second video) with consistent appearance/motion to final video, and enables new interactive control capabilities.

Conclusion: DiffusionBrowser provides interactive previews, new control mechanisms, and systematic probing of the black-box denoising process, making video diffusion more transparent and controllable.

Abstract: Video diffusion models have revolutionized generative video synthesis, but they are imprecise, slow, and can be opaque during generation – keeping users in the dark for a prolonged period. In this work, we propose DiffusionBrowser, a model-agnostic, lightweight decoder framework that allows users to interactively generate previews at any point (timestep or transformer block) during the denoising process. Our model can generate multi-modal preview representations that include RGB and scene intrinsics at more than 4$\times$ real-time speed (less than 1 second for a 4-second video) that convey consistent appearance and motion to the final video. With the trained decoder, we show that it is possible to interactively guide the generation at intermediate noise steps via stochasticity reinjection and modal steering, unlocking a new control capability. Moreover, we systematically probe the model using the learned decoders, revealing how scene, object, and other details are composed and assembled during the otherwise black-box denoising process.

Method: Introduces a lightweight, user-editable Proxy Dynamic Graph (PDG) that deterministically drives part motion, while using a frozen diffusion prior to synthesize appearance. Users annotate and repose the PDG, compute dense motion flow to leverage diffusion as motion-guided shader, then edit appearance in disoccluded areas and perform latent-space composite using PDG visibility information.

Result: Demonstrates clear advantages over state-of-the-art alternatives for turning images into short videos of articulated objects, furniture, vehicles, and deformables, enabling controllable articulation and user control over disocclusions without fine-tuning.

Conclusion: The method successfully mixes generative control (loose pose and structure) with predictable controls (appearance specification in disoccluded regions), unlocking a new image-to-video workflow that separates motion specification from appearance synthesis.

Abstract: We address image-to-video generation with explicit user control over the final frame’s disoccluded regions. Current image-to-video pipelines produce plausible motion but struggle to generate predictable, articulated motions while enforcing user-specified content in newly revealed areas. Our key idea is to separate motion specification from appearance synthesis: we introduce a lightweight, user-editable Proxy Dynamic Graph (PDG) that deterministically yet approximately drives part motion, while a frozen diffusion prior is used to synthesize plausible appearance that follows that motion. In our training-free pipeline, the user loosely annotates and reposes a PDG, from which we compute a dense motion flow to leverage diffusion as a motion-guided shader. We then let the user edit appearance in the disoccluded areas of the image, and exploit the visibility information encoded by the PDG to perform a latent-space composite that reconciles motion with user intent in these areas. This design yields controllable articulation and user control over disocclusions without fine-tuning. We demonstrate clear advantages against state-of-the-art alternatives towards images turned into short videos of articulated objects, furniture, vehicles, and deformables. Our method mixes generative control, in the form of loose pose and structure, with predictable controls, in the form of appearance specification in the final frame in the disoccluded regions, unlocking a new image-to-video workflow. Code will be released on acceptance. Project page: https://anranqi.github.io/beyondvisible.github.io/

Method: Uses 3D Gaussian Splatting for photorealistic environment representations, game engine physics for natural arrangements, two-pass rendering with shadow maps from proxy meshes, and automatic pixel-perfect segmentation mask generation.

Result: Hybrid training with small real images plus large synthetic data yields best detection/segmentation performance, confirming optimal strategy for robust models.

Conclusion: The pipeline efficiently generates realistic synthetic datasets that bridge the domain gap and reduce annotation burden, enabling robust computer vision models for robotics through hybrid training.

Abstract: This paper introduces a novel pipeline for generating large-scale, highly realistic, and automatically labeled datasets for computer vision tasks in robotic environments. Our approach addresses the critical challenges of the domain gap between synthetic and real-world imagery and the time-consuming bottleneck of manual annotation. We leverage 3D Gaussian Splatting (3DGS) to create photorealistic representations of the operational environment and objects. These assets are then used in a game engine where physics simulations create natural arrangements. A novel, two-pass rendering technique combines the realism of splats with a shadow map generated from proxy meshes. This map is then algorithmically composited with the image to add both physically plausible shadows and subtle highlights, significantly enhancing realism. Pixel-perfect segmentation masks are generated automatically and formatted for direct use with object detection models like YOLO. Our experiments show that a hybrid training strategy, combining a small set of real images with a large volume of our synthetic data, yields the best detection and segmentation performance, confirming this as an optimal strategy for efficiently achieving robust and accurate models.

Method: Uses two-stage Faster R-CNN with enhanced Balanced Group Softmax (BAGS) framework to mitigate class imbalance. Also explores metric learning to create well-separated feature embeddings and uses k-NN for inference to improve rare class classification.

Result: Achieves new state-of-the-art performance with 24.5% mAP on LVISv1 dataset (1,203 categories, 164,000 images), surpassing previous benchmark of 24.0%.

Conclusion: The proposed enhancements to BAGS framework combined with metric learning and k-NN inference effectively advance long-tailed object detection by addressing class imbalance and improving rare class classification.

Abstract: Object detection has been widely explored for class-balanced datasets such as COCO. However, real-world scenarios introduce the challenge of long-tailed distributions, where numerous categories contain only a few instances. This inherent class imbalance biases detection models towards the more frequent classes, degrading performance on rare categories. In this paper, we tackle the problem of long-tailed 2D object detection using the LVISv1 dataset, which consists of 1,203 categories and 164,000 images. We employ a two-stage Faster R-CNN architecture and propose enhancements to the Balanced Group Softmax (BAGS) framework to mitigate class imbalance. Our approach achieves a new state-of-the-art performance with a mean Average Precision (mAP) of 24.5%, surpassing the previous benchmark of 24.0%. Additionally, we hypothesize that tail class features may form smaller, denser clusters within the feature space of head classes, making classification challenging for regression-based classifiers. To address this issue, we explore metric learning to produce feature embeddings that are both well-separated across classes and tightly clustered within each class. For inference, we utilize a k-Nearest Neighbors (k-NN) approach to improve classification performance, particularly for rare classes. Our results demonstrate the effectiveness of these methods in advancing long-tailed object detection.

Method: Proposes USTM framework with Swin Transformer spatial backbone enhanced with lightweight Temporal Adapter with Positional Embeddings (TAPE) to capture both fine-grained spatial features and short/long-term temporal context in a unified encoder.

Result: Achieves state-of-the-art performance on PHOENIX14, PHOENIX14T, and CSL-Daily datasets against RGB-based and multi-modal approaches, while maintaining competitive performance against multi-stream approaches using only RGB inputs.

Conclusion: USTM demonstrates effective unified spatio-temporal modeling for CSLR, capturing fine-grained cues and temporal dependencies without requiring multi-stream inputs or auxiliary modalities, offering a strong RGB-only solution.

Abstract: Continuous sign language recognition (CSLR) requires precise spatio-temporal modeling to accurately recognize sequences of gestures in videos. Existing frameworks often rely on CNN-based spatial backbones combined with temporal convolution or recurrent modules. These techniques fail in capturing fine-grained hand and facial cues and modeling long-range temporal dependencies. To address these limitations, we propose the Unified Spatio-Temporal Modeling (USTM) framework, a spatio-temporal encoder that effectively models complex patterns using a combination of a Swin Transformer backbone enhanced with lightweight temporal adapter with positional embeddings (TAPE). Our framework captures fine-grained spatial features alongside short and long-term temporal context, enabling robust sign language recognition from RGB videos without relying on multi-stream inputs or auxiliary modalities. Extensive experiments on benchmarked datasets including PHOENIX14, PHOENIX14T, and CSL-Daily demonstrate that USTM achieves state-of-the-art performance against RGB-based as well as multi-modal CSLR approaches, while maintaining competitive performance against multi-stream approaches. These results highlight the strength and efficacy of the USTM framework for CSLR. The code is available at https://github.com/gufranSabri/USTM

Method: MCT-UEG framework with flat-minima-oriented meta training and testing scheme to optimize the unexploitable example generator for producing broadly unexploitable examples across multiple tasks.

Result: Extensive experiments show the efficacy of the proposed framework in generating broadly unexploitable examples.

Conclusion: The MCT-UEG framework effectively addresses the limitation of existing methods by generating broadly unexploitable examples across different computer vision tasks through meta-learning optimization.

Abstract: Unexploitable example generation aims to transform personal images into their unexploitable (unlearnable) versions before they are uploaded online, thereby preventing unauthorized exploitation of online personal images. Recently, this task has garnered significant research attention due to its critical relevance to personal data privacy. Yet, despite recent progress, existing methods for this task can still suffer from limited practical applicability, as they can fail to generate examples that are broadly unexploitable across different real-world computer vision tasks. To deal with this problem, in this work, we propose a novel Meta Cross-Task Unexploitable Example Generation (MCT-UEG) framework. At the core of our framework, to optimize the unexploitable example generator for effectively producing broadly unexploitable examples, we design a flat-minima-oriented meta training and testing scheme. Extensive experiments show the efficacy of our framework.

Method: RecTok introduces two key innovations: 1) Flow semantic distillation - distills semantic information from vision foundation models into forward flow trajectories in flow matching rather than focusing on latent space, 2) Reconstruction-alignment distillation - enhances semantics with masked feature reconstruction loss.

Result: Achieves state-of-the-art results on gFID-50K with and without classifier-free guidance, superior image reconstruction and generation quality, and discriminative performance. Shows consistent improvements as latent dimensionality increases.

Conclusion: RecTok successfully overcomes the limitations of high-dimensional visual tokenizers by focusing on making forward flow trajectories semantically rich, enabling both high-quality generation and semantically expressive latent spaces that improve with dimensionality.

Abstract: Visual tokenizers play a crucial role in diffusion models. The dimensionality of latent space governs both reconstruction fidelity and the semantic expressiveness of the latent feature. However, a fundamental trade-off is inherent between dimensionality and generation quality, constraining existing methods to low-dimensional latent spaces. Although recent works have leveraged vision foundation models to enrich the semantics of visual tokenizers and accelerate convergence, high-dimensional tokenizers still underperform their low-dimensional counterparts. In this work, we propose RecTok, which overcomes the limitations of high-dimensional visual tokenizers through two key innovations: flow semantic distillation and reconstruction–alignment distillation. Our key insight is to make the forward flow in flow matching semantically rich, which serves as the training space of diffusion transformers, rather than focusing on the latent space as in previous works. Specifically, our method distills the semantic information in VFMs into the forward flow trajectories in flow matching. And we further enhance the semantics by introducing a masked feature reconstruction loss. Our RecTok achieves superior image reconstruction, generation quality, and discriminative performance. It achieves state-of-the-art results on the gFID-50K under both with and without classifier-free guidance settings, while maintaining a semantically rich latent space structure. Furthermore, as the latent dimensionality increases, we observe consistent improvements. Code and model are available at https://shi-qingyu.github.io/rectok.github.io.

Method: Uses a genetic algorithm to iteratively refine a pool of prompts to expose biases. The optimization is driven by a novel bias score that compares the distribution of generated images to LLM-generated text variations of the prompt, ranking biases by severity.

Result: The method successfully mines prompts that cause TTI models to generate biased outputs, with the bias score validated on a dataset with known biases. Code and examples are available.

Conclusion: MineTheGap provides an automated approach to discover problematic prompts that expose biases in TTI models, addressing both societal impacts and user experience issues with image diversity.

Abstract: Text-to-Image (TTI) models generate images based on text prompts, which often leave certain aspects of the desired image ambiguous. When faced with these ambiguities, TTI models have been shown to exhibit biases in their interpretations. These biases can have societal impacts, e.g., when showing only a certain race for a stated occupation. They can also affect user experience when creating redundancy within a set of generated images instead of spanning diverse possibilities. Here, we introduce MineTheGap - a method for automatically mining prompts that cause a TTI model to generate biased outputs. Our method goes beyond merely detecting bias for a given prompt. Rather, it leverages a genetic algorithm to iteratively refine a pool of prompts, seeking for those that expose biases. This optimization process is driven by a novel bias score, which ranks biases according to their severity, as we validate on a dataset with known biases. For a given prompt, this score is obtained by comparing the distribution of generated images to the distribution of LLM-generated texts that constitute variations on the prompt. Code and examples are available on the project’s webpage.

Method: Combines domain-adapted MobileNetV2 and MobileNetV3 as feature extractors with feature fusion, then uses Bi-LSTM classifier enhanced with attention mechanisms to capture sequential dependencies and focus on relevant features.

Result: Achieved 98.23% at 15-shot on PlantVillage (close to 99.98% SOTA), 69.28% on field images, and outperformed previous SOTA with 99.72% on six diseases. Model is lightweight (~40MB, ~1.12 GFLOPs).

Conclusion: The framework provides a scalable, mobile-ready solution for precise plant disease diagnostics in data-scarce regions, balancing performance with computational efficiency.

Abstract: Accurate and timely identification of plant leaf diseases is essential for resilient and sustainable agriculture, yet most deep learning approaches rely on large annotated datasets and computationally intensive models that are unsuitable for data-scarce and resource-constrained environments. To address these challenges we present a few-shot learning approach within a lightweight yet efficient framework that combines domain-adapted MobileNetV2 and MobileNetV3 models as feature extractors, along with a feature fusion technique to generate robust feature representation. For the classification task, the fused features are passed through a Bi-LSTM classifier enhanced with attention mechanisms to capture sequential dependencies and focus on the most relevant features, thereby achieving optimal classification performance even in complex, real-world environments with noisy or cluttered backgrounds. The proposed framework was evaluated across multiple experimental setups, including both laboratory-controlled and field-captured datasets. On tomato leaf diseases from the PlantVillage dataset, it consistently improved performance across 1 to 15 shot scenarios, reaching 98.23+-0.33% at 15 shot, closely approaching the 99.98% SOTA benchmark achieved by a Transductive LSTM with attention, while remaining lightweight and mobile-friendly. Under real-world conditions using field images from the Dhan Shomadhan dataset, it maintained robust performance, reaching 69.28+-1.49% at 15-shot and demonstrating strong resilience to complex backgrounds. Notably, it also outperformed the previous SOTA accuracy of 96.0% on six diseases from PlantVillage, achieving 99.72% with only 15-shot learning. With a compact model size of approximately 40 MB and inference complexity of approximately 1.12 GFLOPs, this work establishes a scalable, mobile-ready foundation for precise plant disease diagnostics in data-scarce regions.

Method: IMILIA combines an inflammation prediction module using Multiple Instance Learning (MIL) with an interpretability module consisting of HistoPLUS (for cell instance detection, segmentation, and classification) and EpiSeg (for epithelium segmentation). The framework analyzes H&E-stained IBD tissue slides end-to-end.

Result: Achieved cross-validation ROC-AUC of 0.83 on discovery cohort, and ROC-AUC of 0.99 and 0.84 on two external validation cohorts. Interpretability analysis revealed biologically consistent patterns: high-scoring tiles showed increased immune cell densities, while low-scoring tiles contained normal epithelial cells.

Conclusion: IMILIA provides an accurate and interpretable framework for inflammation prediction in IBD tissue slides, offering automated computation of biologically meaningful markers that characterize tissue regions driving predictions, with consistent performance across multiple datasets.

Abstract: As the therapeutic target for Inflammatory Bowel Disease (IBD) shifts toward histologic remission, the accurate assessment of microscopic inflammation has become increasingly central for evaluating disease activity and response to treatment. In this work, we introduce IMILIA (Interpretable Multiple Instance Learning for Inflammation Analysis), an end-to-end framework designed for the prediction of inflammation presence in IBD digitized slides stained with hematoxylin and eosin (H&E), followed by the automated computation of markers characterizing tissue regions driving the predictions. IMILIA is composed of an inflammation prediction module, consisting of a Multiple Instance Learning (MIL) model, and an interpretability module, divided in two blocks: HistoPLUS, for cell instance detection, segmentation and classification; and EpiSeg, for epithelium segmentation. IMILIA achieves a cross-validation ROC-AUC of 0.83 on the discovery cohort, and a ROC-AUC of 0.99 and 0.84 on two external validation cohorts. The interpretability module yields biologically consistent insights: tiles with higher predicted scores show increased densities of immune cells (lymphocytes, plasmocytes, neutrophils and eosinophils), whereas lower-scored tiles predominantly contain normal epithelial cells. Notably, these patterns were consistent across all datasets. Code and models to partially replicate the results on the public IBDColEpi dataset can be found at https://github.com/owkin/imilia.

Method: Proposes using diffusion models at test time to map target images back to the source distribution where the downstream model was trained. Requires only source domain description, preserves the task model, and eliminates large-scale synthetic data generation. Includes ensemble variant for enhanced robustness.

Result: Demonstrates consistent improvements across segmentation, detection, and classification tasks under challenging environmental shifts in real-to-real domain generalization scenarios. Achieves substantial relative gains: 137% on BDD100K-Night, 68% on ImageNet-R, and 62% on DarkZurich.

Conclusion: Test-time diffusion-based domain mapping provides an effective alternative to training data augmentation for domain generalization, offering significant performance improvements while being more efficient and preserving existing task models.

Abstract: Generative foundation models contain broad visual knowledge and can produce diverse image variations, making them particularly promising for advancing domain generalization tasks. While they can be used for training data augmentation, synthesizing comprehensive target-domain variations remains slow, expensive, and incomplete. We propose an alternative: using diffusion models at test time to map target images back to the source distribution where the downstream model was trained. This approach requires only a source domain description, preserves the task model, and eliminates large-scale synthetic data generation. We demonstrate consistent improvements across segmentation, detection, and classification tasks under challenging environmental shifts in real-to-real domain generalization scenarios with unknown target distributions. Our analysis spans multiple generative and downstream models, including an ensemble variant for enhanced robustness. The method achieves substantial relative gains: 137% on BDD100K-Night, 68% on ImageNet-R, and 62% on DarkZurich.

Method: 1) Part-aware Temporal Coherence Module: Divides subjects into parts, establishes correspondences, and uses cross-attention between corresponding parts across frames for fine-grained consistency. 2) Subject and Camera Motion Decoupled CFG: Novel guidance strategy that enables independent camera movement control by separately injecting subject and camera motion information into positive and negative anchors of CFG. 3) XPose dataset: Created a high-quality dataset with 50,000 non-human pose-video pairs using automated annotation and filtering pipeline.

Result: Extensive experiments show that PoseAnything significantly outperforms state-of-the-art methods in both effectiveness and generalization capabilities, demonstrating superior performance across diverse subjects beyond just humans.

Conclusion: PoseAnything represents the first universal pose-guided video generation framework that breaks the limitation of human-only pose inputs, enabling precise motion control for arbitrary subjects while introducing novel techniques for temporal consistency and camera movement control.

Abstract: Pose-guided video generation refers to controlling the motion of subjects in generated video through a sequence of poses. It enables precise control over subject motion and has important applications in animation. However, current pose-guided video generation methods are limited to accepting only human poses as input, thus generalizing poorly to pose of other subjects. To address this issue, we propose PoseAnything, the first universal pose-guided video generation framework capable of handling both human and non-human characters, supporting arbitrary skeletal inputs. To enhance consistency preservation during motion, we introduce Part-aware Temporal Coherence Module, which divides the subject into different parts, establishes part correspondences, and computes cross-attention between corresponding parts across frames to achieve fine-grained part-level consistency. Additionally, we propose Subject and Camera Motion Decoupled CFG, a novel guidance strategy that, for the first time, enables independent camera movement control in pose-guided video generation, by separately injecting subject and camera motion control information into the positive and negative anchors of CFG. Furthermore, we present XPose, a high-quality public dataset containing 50,000 non-human pose-video pairs, along with an automated pipeline for annotation and filtering. Extensive experiments demonstrate that Pose-Anything significantly outperforms state-of-the-art methods in both effectiveness and generalization.

Method: T3-Video introduces multi-scale weight-sharing window attention and hierarchical blocking with axis-preserving full-attention design to transform pretrained models’ attention patterns without altering core architecture.

Result: On 4K-VBench, T3-Video outperforms existing approaches with +4.29↑ VQA and +0.08↑ VTC improvements while accelerating native 4K video generation by more than 10×.

Conclusion: The T3 strategy enables efficient high-resolution video generation by retrofitting pretrained models with optimized forward logic, achieving significant speed and quality gains for native 4K video synthesis.

Abstract: Native 4K (2160$\times$3840) video generation remains a critical challenge due to the quadratic computational explosion of full-attention as spatiotemporal resolution increases, making it difficult for models to strike a balance between efficiency and quality. This paper proposes a novel Transformer retrofit strategy termed $\textbf{T3}$ ($\textbf{T}$ransform $\textbf{T}$rained $\textbf{T}$ransformer) that, without altering the core architecture of full-attention pretrained models, significantly reduces compute requirements by optimizing their forward logic. Specifically, $\textbf{T3-Video}$ introduces a multi-scale weight-sharing window attention mechanism and, via hierarchical blocking together with an axis-preserving full-attention design, can effect an “attention pattern” transformation of a pretrained model using only modest compute and data. Results on 4K-VBench show that $\textbf{T3-Video}$ substantially outperforms existing approaches: while delivering performance improvements (+4.29$\uparrow$ VQA and +0.08$\uparrow$ VTC), it accelerates native 4K video generation by more than 10$\times$. Project page at https://zhangzjn.github.io/projects/T3-Video

Method: Built on Wan2.2-5B backbone with audio-injection layers, multiple training strategies, threshold-aware codebook replacement for consistency. Uses step/CFG distillation and lightweight VAE for 11.4× speedup. Creates Soul-1M dataset and Soul-Bench for evaluation.

Result: Significantly outperforms current leading open-source and commercial models on video quality, video-text alignment, identity preservation, and lip-sync accuracy. Achieves 11.4× speedup with negligible quality loss.

Conclusion: Soul demonstrates state-of-the-art performance in digital human animation with broad applicability in virtual anchors and film production, addressing data scarcity through large-scale annotated dataset creation.

Abstract: We propose a multimodal-driven framework for high-fidelity long-term digital human animation termed $\textbf{Soul}$, which generates semantically coherent videos from a single-frame portrait image, text prompts, and audio, achieving precise lip synchronization, vivid facial expressions, and robust identity preservation. We construct Soul-1M, containing 1 million finely annotated samples with a precise automated annotation pipeline (covering portrait, upper-body, full-body, and multi-person scenes) to mitigate data scarcity, and we carefully curate Soul-Bench for comprehensive and fair evaluation of audio-/text-guided animation methods. The model is built on the Wan2.2-5B backbone, integrating audio-injection layers and multiple training strategies together with threshold-aware codebook replacement to ensure long-term generation consistency. Meanwhile, step/CFG distillation and a lightweight VAE are used to optimize inference efficiency, achieving an 11.4$\times$ speedup with negligible quality loss. Extensive experiments show that Soul significantly outperforms current leading open-source and commercial models on video quality, video-text alignment, identity preservation, and lip-synchronization accuracy, demonstrating broad applicability in real-world scenarios such as virtual anchors and film production. Project page at https://zhangzjn.github.io/projects/Soul/

Method: 1) Created MotionEdit dataset with high-fidelity image pairs from continuous videos showing realistic motion transformations. 2) Developed MotionEdit-Bench benchmark with generative, discriminative, and preference-based metrics. 3) Proposed MotionNFT (Motion-guided Negative-aware Fine Tuning) - a post-training framework that computes motion alignment rewards based on how well motion flow between input and edited images matches ground truth.

Result: Motion editing remains challenging for existing state-of-the-art diffusion models. MotionNFT consistently improves editing quality and motion fidelity on FLUX.1 Kontext and Qwen-Image-Edit models without sacrificing general editing ability. The benchmark reveals current limitations in motion editing capabilities.

Conclusion: MotionEdit addresses a significant gap in motion-centric image editing by providing a high-quality dataset and benchmark. The proposed MotionNFT framework effectively improves motion editing performance, demonstrating the importance of motion alignment rewards for achieving realistic motion transformations in image editing tasks.

Abstract: We introduce MotionEdit, a novel dataset for motion-centric image editing-the task of modifying subject actions and interactions while preserving identity, structure, and physical plausibility. Unlike existing image editing datasets that focus on static appearance changes or contain only sparse, low-quality motion edits, MotionEdit provides high-fidelity image pairs depicting realistic motion transformations extracted and verified from continuous videos. This new task is not only scientifically challenging but also practically significant, powering downstream applications such as frame-controlled video synthesis and animation. To evaluate model performance on the novel task, we introduce MotionEdit-Bench, a benchmark that challenges models on motion-centric edits and measures model performance with generative, discriminative, and preference-based metrics. Benchmark results reveal that motion editing remains highly challenging for existing state-of-the-art diffusion-based editing models. To address this gap, we propose MotionNFT (Motion-guided Negative-aware Fine Tuning), a post-training framework that computes motion alignment rewards based on how well the motion flow between input and model-edited images matches the ground-truth motion, guiding models toward accurate motion transformations. Extensive experiments on FLUX.1 Kontext and Qwen-Image-Edit show that MotionNFT consistently improves editing quality and motion fidelity of both base models on the motion editing task without sacrificing general editing ability, demonstrating its effectiveness. Our code is at https://github.com/elainew728/motion-edit/.

Method: Dual-branch Diffusion Transformer architecture with cross-modal joint module, multi-stage data pipeline, post-training optimizations (SFT and RLHF with multi-dimensional reward models), and acceleration framework for 10X faster inference.

Result: Achieves exceptional audio-visual synchronization, superior generation quality, precise multilingual/dialect lip-syncing, dynamic cinematic camera control, enhanced narrative coherence, and 10X inference speed boost.

Conclusion: Seedance 1.5 pro positions itself as a robust engine for professional-grade content creation, now accessible on Volcano Engine platform.

Abstract: Recent strides in video generation have paved the way for unified audio-visual generation. In this work, we present Seedance 1.5 pro, a foundational model engineered specifically for native, joint audio-video generation. Leveraging a dual-branch Diffusion Transformer architecture, the model integrates a cross-modal joint module with a specialized multi-stage data pipeline, achieving exceptional audio-visual synchronization and superior generation quality. To ensure practical utility, we implement meticulous post-training optimizations, including Supervised Fine-Tuning (SFT) on high-quality datasets and Reinforcement Learning from Human Feedback (RLHF) with multi-dimensional reward models. Furthermore, we introduce an acceleration framework that boosts inference speed by over 10X. Seedance 1.5 pro distinguishes itself through precise multilingual and dialect lip-syncing, dynamic cinematic camera control, and enhanced narrative coherence, positioning it as a robust engine for professional-grade content creation. Seedance 1.5 pro is now accessible on Volcano Engine at https://console.volcengine.com/ark/region:ark+cn-beijing/experience/vision?type=GenVideo.

Method: TARA (Time Aware Retrieval Adaptation) - a simple and efficient recipe to adapt Multimodal LLMs (MLLMs) into time-aware video-text embedding models without using any video data. Also introduces a new benchmark with temporally opposite (chiral) actions as hard negatives.

Result: TARA outperforms all existing video-text models on the chiral benchmark while achieving strong results on standard benchmarks. Additionally shows negation-awareness (NegBench) and SOTA performance on verb and adverb understanding in videos.

Conclusion: TARA yields a strong, versatile, time-aware video-text embedding model with state-of-the-art zero-shot performance that goes beyond time-awareness to include negation understanding and fine-grained action comprehension.

Abstract: Our objective is to build a general time-aware video-text embedding model for retrieval. To that end, we propose a simple and efficient recipe, dubbed TARA (Time Aware Retrieval Adaptation), to adapt Multimodal LLMs (MLLMs) to a time-aware video-text embedding model without using any video data at all. For evaluating time-awareness in retrieval, we propose a new benchmark with temporally opposite (chiral) actions as hard negatives and curated splits for chiral and non-chiral actions. We show that TARA outperforms all existing video-text models on this chiral benchmark while also achieving strong results on standard benchmarks. Furthermore, we discover additional benefits of TARA beyond time-awareness: (i) TARA embeddings are negation-aware as shown in NegBench benchmark that evaluates negation in video retrieval, (ii) TARA achieves state of the art performance on verb and adverb understanding in videos. Overall, TARA yields a strong, versatile, time-aware video-text embedding model with state of the art zero-shot performance.

Method: Pancakes introduces a new problem formulation for automatic multi-protocol segmentation that generates multiple plausible segmentation maps for unseen domains while maintaining semantic coherence across related images.

Result: Experiments on seven held-out datasets show Pancakes significantly outperforms existing foundation models in producing several plausible whole-image segmentations that are semantically coherent across images.

Conclusion: Pancakes addresses a novel segmentation problem not currently attainable by existing foundation models, enabling automatic generation of multiple semantically consistent segmentation protocols for biomedical images.

Abstract: A single biomedical image can be meaningfully segmented in multiple ways, depending on the desired application. For instance, a brain MRI can be segmented according to tissue types, vascular territories, broad anatomical regions, fine-grained anatomy, or pathology, etc. Existing automatic segmentation models typically either (1) support only a single protocol, the one they were trained on, or (2) require labor-intensive manual prompting to specify the desired segmentation. We introduce Pancakes, a framework that, given a new image from a previously unseen domain, automatically generates multi-label segmentation maps for multiple plausible protocols, while maintaining semantic consistency across related images. Pancakes introduces a new problem formulation that is not currently attainable by existing foundation models. In a series of experiments on seven held-out datasets, we demonstrate that our model can significantly outperform existing foundation models in producing several plausible whole-image segmentations, that are semantically coherent across images.

Method: IADNet with Temporal Attention Sharing Module (TASM) to synchronize motion embeddings across both people, and Distance-Based Relational Encoding Module (DREM) to capture spatial social cues, using normalizing flow for anomaly scoring.

Result: IADNet outperforms existing human-centric AD baselines on human-human motion benchmarks for the H2IAD task.

Conclusion: The proposed approach effectively addresses the novel H2IAD task by modeling both temporal and spatial dynamics of human-human interactions for improved anomaly detection.

Abstract: Human-centric anomaly detection (AD) has been primarily studied to specify anomalous behaviors in a single person. However, as humans by nature tend to act in a collaborative manner, behavioral anomalies can also arise from human-human interactions. Detecting such anomalies using existing single-person AD models is prone to low accuracy, as these approaches are typically not designed to capture the complex and asymmetric dynamics of interactions. In this paper, we introduce a novel task, Human-Human Interaction Anomaly Detection (H2IAD), which aims to identify anomalous interactive behaviors within collaborative 3D human actions. To address H2IAD, we then propose Interaction Anomaly Detection Network (IADNet), which is formalized with a Temporal Attention Sharing Module (TASM). Specifically, in designing TASM, we share the encoded motion embeddings across both people such that collaborative motion correlations can be effectively synchronized. Moreover, we notice that in addition to temporal dynamics, human interactions are also characterized by spatial configurations between two people. We thus introduce a Distance-Based Relational Encoding Module (DREM) to better reflect social cues in H2IAD. The normalizing flow is eventually employed for anomaly scoring. Extensive experiments on human-human motion benchmarks demonstrate that IADNet outperforms existing Human-centric AD baselines in H2IAD.

Method: Introduces MMhops benchmark with Bridging and Comparison tasks, and proposes MMhops-R1 - a multi-modal RAG framework using reinforcement learning for dynamic reasoning path planning, query formulation, and information synthesis.

Result: MMhops-R1 significantly outperforms baselines on MMhops benchmark, demonstrates strong generalization to fixed-hop reasoning tasks, showing dynamic planning and multi-modal knowledge integration are crucial for complex reasoning.

Conclusion: Provides a challenging benchmark and powerful baseline model for multi-modal multi-hop reasoning, with plans to release code, data, and weights to catalyze future research in this critical area.

Abstract: The ability to perform multi-modal multi-hop reasoning by iteratively integrating information across various modalities and external knowledge is critical for addressing complex real-world challenges. However, existing Multi-modal Large Language Models (MLLMs) are predominantly limited to single-step reasoning, as existing benchmarks lack the complexity needed to evaluate and drive multi-hop abilities. To bridge this gap, we introduce MMhops, a novel, large-scale benchmark designed to systematically evaluate and foster multi-modal multi-hop reasoning. MMhops dataset comprises two challenging task formats, Bridging and Comparison, which necessitate that models dynamically construct complex reasoning chains by integrating external knowledge. To tackle the challenges posed by MMhops, we propose MMhops-R1, a novel multi-modal Retrieval-Augmented Generation (mRAG) framework for dynamic reasoning. Our framework utilizes reinforcement learning to optimize the model for autonomously planning reasoning paths, formulating targeted queries, and synthesizing multi-level information. Comprehensive experiments demonstrate that MMhops-R1 significantly outperforms strong baselines on MMhops, highlighting that dynamic planning and multi-modal knowledge integration are crucial for complex reasoning. Moreover, MMhops-R1 demonstrates strong generalization to tasks requiring fixed-hop reasoning, underscoring the robustness of our dynamic planning approach. In conclusion, our work contributes a challenging new benchmark and a powerful baseline model, and we will release the associated code, data, and weights to catalyze future research in this critical area.

Method: Fine-tunes existing diffusion models on custom indoor/outdoor datasets with spatiotemporal light probes; uses mirrored/diffuse spheres at different exposures based on 3D positions; introduces novel geometric conditioning for relative scene-to-target positioning; combines predictions via differentiable rendering into HDRI maps.

Result: Establishes state-of-the-art performance for both spatial control and prediction accuracy in lighting estimation, demonstrating superior results through thorough evaluation of design choices.

Conclusion: LiMo successfully addresses the dual challenges of high-frequency detail and accurate illuminance in spatiotemporal lighting estimation through diffusion-based generation with novel geometric conditioning and exposure-aware sphere rendering.

Abstract: We present Lighting in Motion (LiMo), a diffusion-based approach to spatiotemporal lighting estimation. LiMo targets both realistic high-frequency detail prediction and accurate illuminance estimation. To account for both, we propose generating a set of mirrored and diffuse spheres at different exposures, based on their 3D positions in the input. Making use of diffusion priors, we fine-tune powerful existing diffusion models on a large-scale customized dataset of indoor and outdoor scenes, paired with spatiotemporal light probes. For accurate spatial conditioning, we demonstrate that depth alone is insufficient and we introduce a new geometric condition to provide the relative position of the scene to the target 3D position. Finally, we combine diffuse and mirror predictions at different exposures into a single HDRI map leveraging differentiable rendering. We thoroughly evaluate our method and design choices to establish LiMo as state-of-the-art for both spatial control and prediction accuracy.

Method: Three-stage end-to-end autoregressive training: (1) Multi-modal guidance integrating dense/sparse control signals for implicit world-level supervision; (2) Degradation-aware training on input frames to bridge training-inference gap; (3) History-context guidance aligning contextual information across adjacent clips for temporal consistency.

Result: State-of-the-art performance in long-range controllability, temporal coherence, and visual fidelity. Supports continuous video generation up to five minutes. Introduces LongVGenBench benchmark with 100 high-resolution one-minute videos covering diverse environments.

Conclusion: LongVie 2 represents a significant step toward unified video world modeling by addressing the three essential properties of world models through a progressive three-stage training approach, enabling high-quality long-term video generation.

Abstract: Building video world models upon pretrained video generation systems represents an important yet challenging step toward general spatiotemporal intelligence. A world model should possess three essential properties: controllability, long-term visual quality, and temporal consistency. To this end, we take a progressive approach-first enhancing controllability and then extending toward long-term, high-quality generation. We present LongVie 2, an end-to-end autoregressive framework trained in three stages: (1) Multi-modal guidance, which integrates dense and sparse control signals to provide implicit world-level supervision and improve controllability; (2) Degradation-aware training on the input frame, bridging the gap between training and long-term inference to maintain high visual quality; and (3) History-context guidance, which aligns contextual information across adjacent clips to ensure temporal consistency. We further introduce LongVGenBench, a comprehensive benchmark comprising 100 high-resolution one-minute videos covering diverse real-world and synthetic environments. Extensive experiments demonstrate that LongVie 2 achieves state-of-the-art performance in long-range controllability, temporal coherence, and visual fidelity, and supports continuous video generation lasting up to five minutes, marking a significant step toward unified video world modeling.

Method: Self-supervised pre-training using DINOv2 methodology on over 25 million 2D slices from 487,975 DBT volumes from 27,990 patients. Evaluated on three downstream tasks: breast density classification (5,000 exams), 5-year breast cancer risk prediction (106,417 exams), and lesion detection (393 annotated volumes).

Result: DBT-DINO achieved: 1) 0.79 accuracy for breast density classification (outperforming DINOv2 baseline and DenseNet-121), 2) 0.78 AUROC for 5-year cancer risk prediction (vs. DINOv2’s 0.76), 3) 0.62 sensitivity for lesion detection (vs. DINOv2’s 0.67), but better performance on cancerous lesions specifically (78.8% vs. 77.3%).

Conclusion: DBT-DINO is the first DBT foundation model showing strong performance on classification and risk prediction tasks. However, domain-specific pre-training showed variable benefits on detection tasks, indicating localized detection requires further methodological development.

Abstract: Foundation models have shown promise in medical imaging but remain underexplored for three-dimensional imaging modalities. No foundation model currently exists for Digital Breast Tomosynthesis (DBT), despite its use for breast cancer screening. To develop and evaluate a foundation model for DBT (DBT-DINO) across multiple clinical tasks and assess the impact of domain-specific pre-training. Self-supervised pre-training was performed using the DINOv2 methodology on over 25 million 2D slices from 487,975 DBT volumes from 27,990 patients. Three downstream tasks were evaluated: (1) breast density classification using 5,000 screening exams; (2) 5-year risk of developing breast cancer using 106,417 screening exams; and (3) lesion detection using 393 annotated volumes. For breast density classification, DBT-DINO achieved an accuracy of 0.79 (95% CI: 0.76–0.81), outperforming both the MetaAI DINOv2 baseline (0.73, 95% CI: 0.70–0.76, p<.001) and DenseNet-121 (0.74, 95% CI: 0.71–0.76, p<.001). For 5-year breast cancer risk prediction, DBT-DINO achieved an AUROC of 0.78 (95% CI: 0.76–0.80) compared to DINOv2’s 0.76 (95% CI: 0.74–0.78, p=.57). For lesion detection, DINOv2 achieved a higher average sensitivity of 0.67 (95% CI: 0.60–0.74) compared to DBT-DINO with 0.62 (95% CI: 0.53–0.71, p=.60). DBT-DINO demonstrated better performance on cancerous lesions specifically with a detection rate of 78.8% compared to Dinov2’s 77.3%. Using a dataset of unprecedented size, we developed DBT-DINO, the first foundation model for DBT. DBT-DINO demonstrated strong performance on breast density classification and cancer risk prediction. However, domain-specific pre-training showed variable benefits on the detection task, with ImageNet baseline outperforming DBT-DINO on general lesion detection, indicating that localized detection tasks require further methodological development.

Method: Introduces the Do-Undo task and benchmark, curates a large-scale dataset of reversible actions from real-world videos, and designs a training strategy that enforces consistency for robust action grounding.

Result: Experiments reveal that current vision-language models struggle with physical reversibility, highlighting the limitations of existing approaches in handling true cause-and-effect relationships in the visual world.

Conclusion: Do-Undo establishes an intuitive testbed for evaluating and advancing physical reasoning in multimodal systems, with important implications for embodied AI, robotics, and physics-aware generative modeling.

Abstract: We introduce the Do-Undo task and benchmark to address a critical gap in vision-language models: understanding and generating physically plausible scene transformations driven by real-world actions. Unlike prior work focused on object-level edits, Do-Undo requires models to simulate the outcome of a physical action and then accurately reverse it, reflecting true cause-and-effect in the visual world. We curate a large-scale dataset of reversible actions from real-world videos and design a training strategy enforcing consistency for robust action grounding. Our experiments reveal that current models struggle with physical reversibility, underscoring the importance of this task for embodied AI, robotics, and physics-aware generative modeling. Do-Undo establishes an intuitive testbed for evaluating and advancing physical reasoning in multimodal systems.

Method: SCR2-ST integrates: 1) Single-cell guided reinforcement learning (SCRL) for active sampling that combines single-cell foundation model embeddings with spatial density to create biologically grounded reward signals for selective acquisition; 2) SCR2Net hybrid prediction network combining regression-based modeling with retrieval-augmented inference, featuring majority cell-type filtering to suppress noisy matches and using retrieved expression profiles as soft labels for auxiliary supervision.

Result: Evaluated on three public ST datasets, SCR2-ST demonstrates state-of-the-art performance in both sampling efficiency and prediction accuracy, particularly under low-budget scenarios.

Conclusion: SCR2-ST effectively leverages single-cell prior knowledge to address the challenges of expensive ST data acquisition and scarce measurements, providing an efficient framework for guided sampling and accurate expression prediction.

Abstract: Spatial transcriptomics (ST) is an emerging technology that enables researchers to investigate the molecular relationships underlying tissue morphology. However, acquiring ST data remains prohibitively expensive, and traditional fixed-grid sampling strategies lead to redundant measurements of morphologically similar or biologically uninformative regions, thus resulting in scarce data that constrain current methods. The well-established single-cell sequencing field, however, could provide rich biological data as an effective auxiliary source to mitigate this limitation. To bridge these gaps, we introduce SCR2-ST, a unified framework that leverages single-cell prior knowledge to guide efficient data acquisition and accurate expression prediction. SCR2-ST integrates a single-cell guided reinforcement learning-based (SCRL) active sampling and a hybrid regression-retrieval prediction network SCR2Net. SCRL combines single-cell foundation model embeddings with spatial density information to construct biologically grounded reward signals, enabling selective acquisition of informative tissue regions under constrained sequencing budgets. SCR2Net then leverages the actively sampled data through a hybrid architecture combining regression-based modeling with retrieval-augmented inference, where a majority cell-type filtering mechanism suppresses noisy matches and retrieved expression profiles serve as soft labels for auxiliary supervision. We evaluated SCR2-ST on three public ST datasets, demonstrating SOTA performance in both sampling efficiency and prediction accuracy, particularly under low-budget scenarios. Code is publicly available at: https://github.com/hrlblab/SCR2ST

Method: Proposes MindDrive with a single LLM containing two LoRA parameter sets: Decision Expert for scenario reasoning and discrete linguistic driving decisions, and Action Expert for mapping decisions to continuous trajectories. Uses trajectory-level rewards fed back to reasoning space for trial-and-error learning over discrete decisions rather than continuous actions.

Result: Achieves Driving Score of 78.04 and Success Rate of 55.09% on Bench2Drive benchmark. First work to demonstrate effectiveness of online reinforcement learning for VLA models in autonomous driving.

Conclusion: MindDrive effectively balances optimal decision-making, human-like driving behavior, and efficient exploration in online RL by operating over discrete linguistic decisions rather than continuous action spaces, overcoming limitations of current VLA approaches.

Abstract: Current Vision-Language-Action (VLA) paradigms in autonomous driving primarily rely on Imitation Learning (IL), which introduces inherent challenges such as distribution shift and causal confusion. Online Reinforcement Learning offers a promising pathway to address these issues through trial-and-error learning. However, applying online reinforcement learning to VLA models in autonomous driving is hindered by inefficient exploration in continuous action spaces. To overcome this limitation, we propose MindDrive, a VLA framework comprising a large language model (LLM) with two distinct sets of LoRA parameters. The one LLM serves as a Decision Expert for scenario reasoning and driving decision-making, while the other acts as an Action Expert that dynamically maps linguistic decisions into feasible trajectories. By feeding trajectory-level rewards back into the reasoning space, MindDrive enables trial-and-error learning over a finite set of discrete linguistic driving decisions, instead of operating directly in a continuous action space. This approach effectively balances optimal decision-making in complex scenarios, human-like driving behavior, and efficient exploration in online reinforcement learning. MindDrive achieves strong closed-loop performance on the challenging Bench2Drive benchmark, with a Driving Score (DS) of 78.04 and a Success Rate (SR) of 55.09%. To the best of our knowledge, this is the first work to demonstrate the effectiveness of online reinforcement learning for the VLA model in autonomous driving.

Method: Created a dataset from a high-quality animated film, capturing dynamic scenes with multiple complementary modalities including RGB images, depth, surface normals, object segmentation, and optical flow. Organized into three benchmarking scenarios: dense multi-view camera setup, sparse camera arrangement, and monocular video sequences.

Result: A unique dataset offering visual richness, high-quality annotations, and diverse experimental setups that enables comprehensive evaluation of view synthesis and 3D vision methods across varying levels of data sparsity.

Conclusion: This dataset provides a valuable resource for advancing the state-of-the-art in novel view synthesis and 3D vision research by offering realistic, richly annotated data with multiple modalities and diverse benchmarking scenarios.

Abstract: This paper presents a new dataset for Novel View Synthesis, generated from a high-quality, animated film with stunning realism and intricate detail. Our dataset captures a variety of dynamic scenes, complete with detailed textures, lighting, and motion, making it ideal for training and evaluating cutting-edge 4D scene reconstruction and novel view generation models. In addition to high-fidelity RGB images, we provide multiple complementary modalities, including depth, surface normals, object segmentation and optical flow, enabling a deeper understanding of scene geometry and motion. The dataset is organised into three distinct benchmarking scenarios: a dense multi-view camera setup, a sparse camera arrangement, and monocular video sequences, enabling a wide range of experimentation and comparison across varying levels of data sparsity. With its combination of visual richness, high-quality annotations, and diverse experimental setups, this dataset offers a unique resource for pushing the boundaries of view synthesis and 3D vision.

Method: The authors introduce Grab-3D, a geometry-aware transformer framework that uses vanishing points as explicit 3D geometry representations. They construct a dataset of static scenes for stable geometric feature extraction. The framework includes geometric positional encoding, temporal-geometric attention, and an EMA-based geometric classifier head to inject 3D geometric awareness into temporal modeling.

Result: Experiments show Grab-3D significantly outperforms state-of-the-art detectors and achieves robust cross-domain generalization to unseen video generators.

Conclusion: The paper demonstrates that analyzing 3D geometric patterns through vanishing points provides an effective approach for detecting AI-generated videos, with the proposed Grab-3D framework offering superior performance and generalization capabilities compared to existing methods.

Abstract: Recent advances in diffusion-based generation techniques enable AI models to produce highly realistic videos, heightening the need for reliable detection mechanisms. However, existing detection methods provide only limited exploration of the 3D geometric patterns present in generated videos. In this paper, we use vanishing points as an explicit representation of 3D geometry patterns, revealing fundamental discrepancies in geometric consistency between real and AI-generated videos. We introduce Grab-3D, a geometry-aware transformer framework for detecting AI-generated videos based on 3D geometric temporal consistency. To enable reliable evaluation, we construct an AI-generated video dataset of static scenes, allowing stable 3D geometric feature extraction. We propose a geometry-aware transformer equipped with geometric positional encoding, temporal-geometric attention, and an EMA-based geometric classifier head to explicitly inject 3D geometric awareness into temporal modeling. Experiments demonstrate that Grab-3D significantly outperforms state-of-the-art detectors, achieving robust cross-domain generalization to unseen generators.

Method: AgentIAD uses a tool-driven agentic framework with Perceptive Zoomer (PZ) for localized fine-grained analysis and Comparative Retriever (CR) for querying normal exemplars. The model is trained in two stages: supervised fine-tuning followed by reinforcement learning with a two-part reward design (perception reward for classification accuracy, spatial alignment, type correctness; behavior reward for efficient tool use).

Result: AgentIAD achieves a new state-of-the-art 97.62% classification accuracy on MMAD dataset, surpassing prior MLLM-based approaches while producing transparent and interpretable inspection traces.

Conclusion: The proposed agentic framework enables effective multi-stage visual inspection for industrial anomaly detection by combining localized analysis with normal pattern comparison, providing both high accuracy and interpretable inspection processes.

Abstract: Industrial anomaly detection (IAD) is difficult due to the scarcity of normal reference samples and the subtle, localized nature of many defects. Single-pass vision-language models (VLMs) often overlook small abnormalities and lack explicit mechanisms to compare against canonical normal patterns. We propose AgentIAD, a tool-driven agentic framework that enables multi-stage visual inspection. The agent is equipped with a Perceptive Zoomer (PZ) for localized fine-grained analysis and a Comparative Retriever (CR) for querying normal exemplars when evidence is ambiguous. To teach these inspection behaviors, we construct structured perceptive and comparative trajectories from the MMAD dataset and train the model in two stages: supervised fine-tuning followed by reinforcement learning. A two-part reward design drives this process: a perception reward that supervises classification accuracy, spatial alignment, and type correctness, and a behavior reward that encourages efficient tool use. Together, these components enable the model to refine its judgment through step-wise observation, zooming, and verification. AgentIAD achieves a new state-of-the-art 97.62% classification accuracy on MMAD, surpassing prior MLLM-based approaches while producing transparent and interpretable inspection traces.

Method: JoVA uses joint self-attention across video and audio tokens within each transformer layer for direct cross-modal interaction without additional alignment modules. It also introduces a mouth-area loss based on facial keypoint detection to enhance lip-speech synchronization supervision during training.

Result: Extensive experiments show JoVA outperforms or is competitive with both unified and audio-driven state-of-the-art methods in lip-sync accuracy, speech quality, and overall video-audio generation fidelity.

Conclusion: JoVA establishes an elegant framework for high-quality multimodal generation that addresses key limitations in existing approaches while maintaining architectural simplicity.

Abstract: In this paper, we present JoVA, a unified framework for joint video-audio generation. Despite recent encouraging advances, existing methods face two critical limitations. First, most existing approaches can only generate ambient sounds and lack the capability to produce human speech synchronized with lip movements. Second, recent attempts at unified human video-audio generation typically rely on explicit fusion or modality-specific alignment modules, which introduce additional architecture design and weaken the model simplicity of the original transformers. To address these issues, JoVA employs joint self-attention across video and audio tokens within each transformer layer, enabling direct and efficient cross-modal interaction without the need for additional alignment modules. Furthermore, to enable high-quality lip-speech synchronization, we introduce a simple yet effective mouth-area loss based on facial keypoint detection, which enhances supervision on the critical mouth region during training without compromising architectural simplicity. Extensive experiments on benchmarks demonstrate that JoVA outperforms or is competitive with both unified and audio-driven state-of-the-art methods in lip-sync accuracy, speech quality, and overall video-audio generation fidelity. Our results establish JoVA as an elegant framework for high-quality multimodal generation.

Method: Training-free framework that converts offline models to streaming by aligning predictions across consecutive windows using layer-wise scale alignment (segments depth into layers, computes per-layer scale factors, propagates across windows/timestamps).

Result: Achieves state-of-the-art performance on camera pose estimation and point map reconstruction, operates at 14 FPS with 6 GB peak memory on RTX A6000 GPU, enabling kilometer-scale streaming video deployment.

Conclusion: LASER enables practical deployment of offline reconstruction models for streaming videos without retraining, addressing memory limitations while maintaining high reconstruction quality.

Abstract: Recent feed-forward reconstruction models like VGGT and $π^3$ achieve impressive reconstruction quality but cannot process streaming videos due to quadratic memory complexity, limiting their practical deployment. While existing streaming methods address this through learned memory mechanisms or causal attention, they require extensive retraining and may not fully leverage the strong geometric priors of state-of-the-art offline models. We propose LASER, a training-free framework that converts an offline reconstruction model into a streaming system by aligning predictions across consecutive temporal windows. We observe that simple similarity transformation ($\mathrm{Sim}(3)$) alignment fails due to layer depth misalignment: monocular scale ambiguity causes relative depth scales of different scene layers to vary inconsistently between windows. To address this, we introduce layer-wise scale alignment, which segments depth predictions into discrete layers, computes per-layer scale factors, and propagates them across both adjacent windows and timestamps. Extensive experiments show that LASER achieves state-of-the-art performance on camera pose estimation and point map reconstruction %quality with offline models while operating at 14 FPS with 6 GB peak memory on a RTX A6000 GPU, enabling practical deployment for kilometer-scale streaming videos. Project website: $\href{https://neu-vi.github.io/LASER/}{\texttt{https://neu-vi.github.io/LASER/}}$

Method: PADS learns unconditional 3D pose diffusion prior, then uses posterior sampling with task-specific forward operators (inverse problem formulation) for pose estimation, denoising, completion, etc., without retraining.

Result: Superior performance across benchmarks against both learning-based and optimization-based baselines, demonstrating effectiveness and generalization capability.

Conclusion: PADS provides a flexible, scalable plug-and-play framework for diverse 3D pose analysis tasks using diffusion synthesis and posterior sampling without task-specific training.

Abstract: Diffusion models have demonstrated impressive capabilities in modeling complex data distributions and are increasingly applied in various generative tasks. In this work, we propose Pose Analysis by Diffusion Synthesis PADS, a unified generative modeling framework for 3D human pose analysis. PADS first learns a task-agnostic 3D pose prior via unconditional diffusion synthesis and then performs training-free adaptation to a wide range of pose analysis tasks, including 3D pose estimation, denoising, completion, etc., through a posterior sampling scheme. By formulating each task as an inverse problem with a known forward operator, PADS injects task-specific constraints during inference while keeping the pose prior fixed. This plug-and-play framework removes the need for task-specific supervision or retraining, offering flexibility and scalability across diverse conditions. Extensive experiments on different benchmarks showcase the superior performance against both learning-based and optimization-based baselines, demonstrating the effectiveness and generalization capability of our method.

Method: Reprogram a pre-trained 3D instance generator to act as a scene-level learner, replacing dataset-bounded supervision with model-centric spatial supervision. Use a view-centric formulation of scene space instead of canonical space, creating a fully feed-forward, generalizable scene generator that learns spatial relations directly from the instance model.

Result: The approach unlocks the generator’s transferable spatial knowledge, enabling generalization to unseen layouts and novel object compositions. Spatial reasoning emerges even when training scenes are randomly composed objects, demonstrating that the generator’s transferable scene prior provides rich learning signals for inferring proximity, support, and symmetry from geometric cues.

Conclusion: A 3D instance generator is an implicit spatial learner and reasoner, pointing toward foundation models for interactive 3D scene understanding and generation. The model-centric approach provides a promising direction for overcoming dataset limitations in 3D scene generation.

Abstract: Generalization remains the central challenge for interactive 3D scene generation. Existing learning-based approaches ground spatial understanding in limited scene dataset, restricting generalization to new layouts. We instead reprogram a pre-trained 3D instance generator to act as a scene level learner, replacing dataset-bounded supervision with model-centric spatial supervision. This reprogramming unlocks the generator transferable spatial knowledge, enabling generalization to unseen layouts and novel object compositions. Remarkably, spatial reasoning still emerges even when the training scenes are randomly composed objects. This demonstrates that the generator’s transferable scene prior provides a rich learning signal for inferring proximity, support, and symmetry from purely geometric cues. Replacing widely used canonical space, we instantiate this insight with a view-centric formulation of the scene space, yielding a fully feed-forward, generalizable scene generator that learns spatial relations directly from the instance model. Quantitative and qualitative results show that a 3D instance generator is an implicit spatial learner and reasoner, pointing toward foundation models for interactive 3D scene understanding and generation. Project page: https://luling06.github.io/I-Scene-project/

Method: Uses transformer-based recurrent neural network to aggregate dense image features over time, learns via asymmetric masked prediction task with standard pixel reconstruction objective, enabling efficient feature propagation with linear computational cost.

Result: Achieves competitive performance with SOTA video models on action recognition and tracking tasks, performs favorably against image models on geometric/spatial tasks, shows 30x greater parameter efficiency than competing video masked autoencoders, and demonstrates stable feature propagation over long temporal horizons.

Conclusion: RVM provides an efficient and effective approach for video representation learning that combines the benefits of recurrent architectures with masked autoencoding, offering strong performance across diverse video and image tasks with exceptional parameter efficiency.

Abstract: We present Recurrent Video Masked-Autoencoders (RVM): a novel video representation learning approach that uses a transformer-based recurrent neural network to aggregate dense image features over time, effectively capturing the spatio-temporal structure of natural video data. RVM learns via an asymmetric masked prediction task requiring only a standard pixel reconstruction objective. This design yields a highly efficient ``generalist’’ encoder: RVM achieves competitive performance with state-of-the-art video models (e.g. VideoMAE, V-JEPA) on video-level tasks like action recognition and point/object tracking, while also performing favorably against image models (e.g. DINOv2) on tasks that test geometric and dense spatial understanding. Notably, RVM achieves strong performance in the small-model regime without requiring knowledge distillation, exhibiting up to 30x greater parameter efficiency than competing video masked autoencoders. Moreover, we demonstrate that RVM’s recurrent nature allows for stable feature propagation over long temporal horizons with linear computational cost, overcoming some of the limitations of standard spatio-temporal attention-based architectures. Finally, we use qualitative visualizations to highlight that RVM learns rich representations of scene semantics, structure, and motion.

Method: Proposes WWC-Predictor that detects wrong-way cycling events in sparsely sampled frames using a lightweight detector, then estimates the overall ratio using an autoregressive moving average model.

Result: Achieves average error rate of 1.475% while consuming only 19.12% GPU time compared to conventional tracking methods, validated on a benchmark dataset of 35 minutes of video with minute-level annotations.

Conclusion: The method provides an efficient and accurate solution for estimating wrong-way cycling ratios in resource-constrained environments, enabling better traffic monitoring and planning with significantly reduced computational costs.

Abstract: Effective monitoring of unusual transportation behaviors, such as wrong-way cycling (i.e., riding a bicycle or e-bike against designated traffic flow), is crucial for optimizing law enforcement deployment and traffic planning. However, accurately recording all wrong-way cycling events is both unnecessary and infeasible in resource-constrained environments, as it requires high-resolution cameras for evidence collection and event detection. To address this challenge, we propose WWC-Predictor, a novel method for efficiently estimating the wrong-way cycling ratio, defined as the proportion of wrong-way cycling events relative to the total number of cycling movements over a given time period. The core innovation of our method lies in accurately detecting wrong-way cycling events in sparsely sampled frames using a light-weight detector, then estimating the overall ratio using an autoregressive moving average model. To evaluate the effectiveness of our method, we construct a benchmark dataset consisting of 35 minutes of video sequences with minute-level annotations.Our method achieves an average error rate of a mere 1.475% while consuming only 19.12% GPU time required by conventional tracking methods, validating its effectiveness in estimating the wrong-way cycling ratio. Our source code is publicly available at: https://github.com/VICA-Lab-HKUST-GZ/WWC-Predictor.

Method: VTP is a unified visual tokenizer pre-training framework that jointly optimizes image-text contrastive, self-supervised, and reconstruction losses to create latent spaces that better represent high-level semantics rather than just low-level details.

Result: VTP achieves competitive performance (78.2% zero-shot accuracy, 0.36 rFID on ImageNet) and 4.1× faster convergence on generation compared to advanced distillation methods. Most importantly, it shows effective scaling: investing more FLOPS in VTP pre-training yields 65.8% FID improvement in downstream generation, while conventional autoencoders stagnate early.

Conclusion: Understanding (high-level semantics) is key to generation, and VTP demonstrates much better scaling properties where generative performance scales effectively with compute, parameters, and data allocated to visual tokenizer pre-training, solving the “pre-training scaling problem.”

Abstract: The quality of the latent space in visual tokenizers (e.g., VAEs) is crucial for modern generative models. However, the standard reconstruction-based training paradigm produces a latent space that is biased towards low-level information, leading to a foundation flaw: better pixel-level accuracy does not lead to higher-quality generation. This implies that pouring extensive compute into visual tokenizer pre-training translates poorly to improved performance in generation. We identify this as the pre-training scaling problem and suggest a necessary shift: to be effective for generation, a latent space must concisely represent high-level semantics. We present VTP, a unified visual tokenizer pre-training framework, pioneering the joint optimization of image-text contrastive, self-supervised, and reconstruction losses. Our large-scale study reveals two principal findings: (1) understanding is a key driver of generation, and (2) much better scaling properties, where generative performance scales effectively with compute, parameters, and data allocated to the pretraining of the visual tokenizer. After large-scale pre-training, our tokenizer delivers a competitive profile (78.2 zero-shot accuracy and 0.36 rFID on ImageNet) and 4.1 times faster convergence on generation compared to advanced distillation methods. More importantly, it scales effectively: without modifying standard DiT training specs, solely investing more FLOPS in pretraining VTP achieves 65.8% FID improvement in downstream generation, while conventional autoencoder stagnates very early at 1/10 FLOPS. Our pre-trained models are available at https://github.com/MiniMax-AI/VTP.

Method: Analyzed role of convolutions vs attention in 3D point cloud networks, finding convolutions work best for low-level geometry in early high-resolution layers, while attention excels at high-level semantics in deep low-resolution layers. Proposed LitePT architecture that uses convolutions early and switches to attention later. Introduced PointROPE, a training-free 3D positional encoding to preserve spatial layout when discarding redundant convolution layers.

Result: LitePT has 3.6× fewer parameters, runs 2× faster, uses 2× less memory than Point Transformer V3, while matching or outperforming it on various tasks and datasets.

Conclusion: The analysis reveals intuitive design principles for 3D point cloud networks: use convolutions for early high-resolution geometry extraction and attention for deep low-resolution semantic understanding. LitePT demonstrates these principles lead to more efficient yet powerful architectures.

Abstract: Modern neural architectures for 3D point cloud processing contain both convolutional layers and attention blocks, but the best way to assemble them remains unclear. We analyse the role of different computational blocks in 3D point cloud networks and find an intuitive behaviour: convolution is adequate to extract low-level geometry at high-resolution in early layers, where attention is expensive without bringing any benefits; attention captures high-level semantics and context in low-resolution, deep layers more efficiently. Guided by this design principle, we propose a new, improved 3D point cloud backbone that employs convolutions in early stages and switches to attention for deeper layers. To avoid the loss of spatial layout information when discarding redundant convolution layers, we introduce a novel, training-free 3D positional encoding, PointROPE. The resulting LitePT model has $3.6\times$ fewer parameters, runs $2\times$ faster, and uses $2\times$ less memory than the state-of-the-art Point Transformer V3, but nonetheless matches or even outperforms it on a range of tasks and datasets. Code and models are available at: https://github.com/prs-eth/LitePT.

Method: Proposes KNN-MMD: 1) Constructs a help set using KNN from target domain, 2) Performs local alignment between source and target domains within each category using MMD, 3) Addresses training instability by excluding support set from training and using it as validation for stopping criterion determination.

Result: Method resolves key instability issues in cross-domain methods where performance fluctuates sharply between epochs, and provides a solution for determining optimal stopping point during training without labeled target domain data.

Conclusion: KNN-MMD effectively addresses limitations of existing domain alignment methods for cross-domain wireless sensing by enabling local category alignment and providing stable training with clear stopping criteria, with publicly available dataset and code.

Abstract: Wireless sensing has recently found widespread applications in diverse environments, including homes, offices, and public spaces. By analyzing patterns in channel state information (CSI), it is possible to infer human actions for tasks such as person identification, gesture recognition, and fall detection. However, CSI is highly sensitive to environmental changes, where even minor alterations can significantly distort the CSI patterns. This sensitivity often leads to performance degradation or outright failure when applying wireless sensing models trained in one environment to another. To address this challenge, Domain Alignment (DAL) has been widely adopted for cross-domain classification tasks, as it focuses on aligning the global distributions of the source and target domains in feature space. Despite its popularity, DAL often neglects inter-category relationships, which can lead to misalignment between categories across domains, even when global alignment is achieved. To overcome these limitations, we propose K-Nearest Neighbors Maximum Mean Discrepancy (KNN-MMD), a novel few-shot method for cross-domain wireless sensing. Our approach begins by constructing a help set using KNN from the target domain, enabling local alignment between the source and target domains within each category using MMD. Additionally, we address a key instability issue commonly observed in cross-domain methods, where model performance fluctuates sharply between epochs. Further, most existing methods struggle to determine an optimal stopping point during training due to the absence of labeled data from the target domain. Our method resolves this by excluding the support set from the target domain during training and employing it as a validation set to determine the stopping criterion. The dataset and code are publicly available at https://github.com/RS2002/KNN-MMD .

Method: Integrated framework combining one-stage object detection, deep convolutional neural network, and traditional machine learning techniques for comprehensive rice grain assessment.

Result: Achieved 99.14% mAP in object detection, 97.89% accuracy in classification, and 97.56% average accuracy in grain completeness grading using 20,000 images from six Chinese rice varieties.

Conclusion: The proposed real-time evaluation mechanism effectively automates rice quality assessment, providing accurate variety identification, grain completeness grading, and chalkiness evaluation for quality control.

Abstract: Rice is one of the most widely cultivated crops globally and has been developed into numerous varieties. The quality of rice during cultivation is primarily determined by its cultivar and characteristics. Traditionally, rice classification and quality assessment rely on manual visual inspection, a process that is both time-consuming and prone to errors. However, with advancements in machine vision technology, automating rice classification and quality evaluation based on its cultivar and characteristics has become increasingly feasible, enhancing both accuracy and efficiency. This study proposes a real-time evaluation mechanism for comprehensive rice grain assessment, integrating a one-stage object detection approach, a deep convolutional neural network, and traditional machine learning techniques. The proposed framework enables rice variety identification, grain completeness grading, and grain chalkiness evaluation. The rice grain dataset used in this study comprises approximately 20,000 images from six widely cultivated rice varieties in China. Experimental results demonstrate that the proposed mechanism achieves a mean average precision (mAP) of 99.14% in the object detection task and an accuracy of 97.89% in the classification task. Furthermore, the framework attains an average accuracy of 97.56% in grain completeness grading within the same rice variety, contributing to an effective quality evaluation system.

Method: N/A - Paper content not accessible due to HTTP 429 error from arXiv API

Result: N/A - Could not retrieve results due to rate limiting error

Conclusion: Unable to analyze paper due to technical limitations in accessing the content

Abstract: Failed to fetch summary for 2503.07874: Page request resulted in HTTP 429 (https://export.arxiv.org/api/query?search_query=&id_list=2503.07874&sortBy=relevance&sortOrder=descending&start=0&max_results=100)

Method: No method information available - arXiv API returned HTTP 429 (Too Many Requests) error

Result: Failed to fetch paper summary from arXiv API - encountered rate limiting error

Conclusion: Unable to analyze the paper due to technical limitations in accessing the content from arXiv

Abstract: Failed to fetch summary for 2503.14505: Page request resulted in HTTP 429 (https://export.arxiv.org/api/query?search_query=&id_list=2503.14505&sortBy=relevance&sortOrder=descending&start=0&max_results=100)

Method: No method information available - arXiv API returned HTTP 429 (Too Many Requests)

Result: Failed to fetch paper summary - encountered rate limiting from arXiv API

Conclusion: Analysis impossible due to technical limitations in accessing the paper content

Abstract: Failed to fetch summary for 2106.05261: Page request resulted in HTTP 429 (https://export.arxiv.org/api/query?search_query=&id_list=2106.05261&sortBy=relevance&sortOrder=descending&start=0&max_results=100)

Method: Created CSAW-M dataset from over 10,000 individuals with direct annotations of masking potential from five specialists. Trained deep learning models on this dataset to estimate masking levels.

Result: Deep learning models trained on CSAW-M masking annotations were significantly more predictive of screening participants diagnosed with interval and large invasive cancers than breast density measures, even without explicit training for these cancer prediction tasks.

Conclusion: Direct assessment of mammographic masking through specialized annotations provides a more effective approach than breast density proxies for identifying cancers likely to be missed in screening, potentially improving early detection of aggressive breast cancers.

Abstract: Interval and large invasive breast cancers, which are associated with worse prognosis than other cancers, are usually detected at a late stage due to false negative assessments of screening mammograms. The missed screening-time detection is commonly caused by the tumor being obscured by its surrounding breast tissues, a phenomenon called masking. To study and benchmark mammographic masking of cancer, in this work we introduce CSAW-M, the largest public mammographic dataset, collected from over 10,000 individuals and annotated with potential masking. In contrast to the previous approaches which measure breast image density as a proxy, our dataset directly provides annotations of masking potential assessments from five specialists. We also trained deep learning models on CSAW-M to estimate the masking level and showed that the estimated masking is significantly more predictive of screening participants diagnosed with interval and large invasive cancers – without being explicitly trained for these tasks – than its breast density counterparts.

Method: Attempted to fetch the paper abstract from arXiv API using the provided ID, but encountered HTTP 429 status indicating too many requests/rate limiting.

Result: Failed to obtain the paper content for analysis. The arXiv API returned HTTP 429 error, preventing access to the abstract and paper details.

Conclusion: Analysis cannot be performed without the paper content. The user may need to try again later when rate limits reset, provide the abstract directly, or use alternative methods to access the paper information.

Abstract: Failed to fetch summary for 2212.14801: Page request resulted in HTTP 429 (https://export.arxiv.org/api/query?search_query=&id_list=2212.14801&sortBy=relevance&sortOrder=descending&start=0&max_results=100)

Method: Attempted to fetch paper metadata from arXiv API using standard query parameters with the paper ID 2505.08919, but encountered rate limiting

Result: Failed to retrieve paper abstract due to HTTP 429 error from arXiv API, indicating temporary rate limiting or server overload

Conclusion: Cannot analyze the requested paper as the content is unavailable due to API rate limiting; suggest trying again later or accessing the paper directly through arXiv website

Abstract: Failed to fetch summary for 2505.08919: Page request resulted in HTTP 429 (https://export.arxiv.org/api/query?search_query=&id_list=2505.08919&sortBy=relevance&sortOrder=descending&start=0&max_results=100)

Method: Attempted to fetch paper summary using arXiv API query with paper ID 2308.01042, but encountered HTTP 429 status indicating too many requests

Result: Failed to retrieve paper content due to API rate limiting restrictions

Conclusion: Cannot analyze the paper abstract as the content could not be fetched; need to wait for rate limits to reset or use alternative methods to access the paper

Abstract: Failed to fetch summary for 2308.01042: Page request resulted in HTTP 429 (https://export.arxiv.org/api/query?search_query=&id_list=2308.01042&sortBy=relevance&sortOrder=descending&start=0&max_results=100)

Method: No method information available - arXiv API request failed with rate limiting error

Result: No results available - the paper content could not be accessed due to HTTP 429 (Too Many Requests) error

Conclusion: Unable to provide analysis due to technical limitations in accessing the paper abstract

Abstract: Failed to fetch summary for 2402.15480: Page request resulted in HTTP 429 (https://export.arxiv.org/api/query?search_query=&id_list=2402.15480&sortBy=relevance&sortOrder=descending&start=0&max_results=100)

Method: No method information available - arXiv API request resulted in HTTP 429 error (Too Many Requests).

Result: Failed to retrieve paper summary - encountered rate limiting from arXiv’s export service.

Conclusion: Analysis cannot proceed without the paper abstract; need to wait for rate limit reset or use alternative access method.

Abstract: Failed to fetch summary for 2403.18461: Page request resulted in HTTP 429 (https://export.arxiv.org/api/query?search_query=&id_list=2403.18461&sortBy=relevance&sortOrder=descending&start=0&max_results=100)

Method: QUOTA uses a dual-loop meta-learning strategy to optimize domain-invariant prompts, integrating prompt learning with learnable counting and domain tokens to capture stylistic variations while maintaining accuracy for unseen object classes.

Result: QUOTA outperforms conventional models in both object quantification accuracy and semantic consistency, as demonstrated through extensive experiments on a new benchmark designed for domain generalization in object quantification.

Conclusion: QUOTA sets a new benchmark for efficient and scalable text-to-image generation across any domain, enabling effective object quantification without the need for domain-specific retraining.

Abstract: We tackle the problem of quantifying the number of objects by a generative text-to-image model. Rather than retraining such a model for each new image domain of interest, which leads to high computational costs and limited scalability, we are the first to consider this problem from a domain-agnostic perspective. We propose QUOTA, an optimization framework for text-to-image models that enables effective object quantification across unseen domains without retraining. It leverages a dual-loop meta-learning strategy to optimize a domain-invariant prompt. Further, by integrating prompt learning with learnable counting and domain tokens, our method captures stylistic variations and maintains accuracy, even for object classes not encountered during training. For evaluation, we adopt a new benchmark specifically designed for object quantification in domain generalization, enabling rigorous assessment of object quantification accuracy and adaptability across unseen domains in text-to-image generation. Extensive experiments demonstrate that QUOTA outperforms conventional models in both object quantification accuracy and semantic consistency, setting a new benchmark for efficient and scalable text-to-image generation for any domain.

Method: Unable to determine method due to technical error in accessing paper content

Result: Unable to determine results due to technical error in accessing paper content

Conclusion: Unable to determine conclusion due to technical error in accessing paper content

Abstract: Failed to fetch summary for 2412.02421: Page request resulted in HTTP 429 (https://export.arxiv.org/api/query?search_query=&id_list=2412.02421&sortBy=relevance&sortOrder=descending&start=0&max_results=100)

Method: No method information available - arXiv API rate limiting blocked access to paper content.

Result: Unable to retrieve results - HTTP 429 error indicates too many requests to arXiv API.

Conclusion: Analysis impossible due to technical limitations - need to wait for rate limit reset or use alternative access methods.

Abstract: Failed to fetch summary for 2506.00227: Page request resulted in HTTP 429 (https://export.arxiv.org/api/query?search_query=&id_list=2506.00227&sortBy=relevance&sortOrder=descending&start=0&max_results=100)

Method: No method information available - paper content retrieval failed with HTTP 429 error.

Result: No results can be analyzed - the arXiv API returned a rate limiting error preventing access to the paper.

Conclusion: Analysis cannot be completed due to technical limitations (HTTP 429 error) preventing access to the paper content.

Abstract: Failed to fetch summary for 2412.03439: Page request resulted in HTTP 429 (https://export.arxiv.org/api/query?search_query=&id_list=2412.03439&sortBy=relevance&sortOrder=descending&start=0&max_results=100)

Method: Two complementary strategies: 1) Texture-aware densification produces denser splats in fully textured areas while keeping sparsity in low-texture regions; 2) Geometry-aware splitting uses depth and normal priors to guide splitting sampling and filter noisy splats through Validation of Depth Ratio Change checking, reducing scattered Gaussian influence.

Result: The method produces more photorealistic 3DGS models with improved rendering quality, especially in regions with weak textures or insufficient training views. Experiments on various datasets show superior Novel View Synthesis results compared to state-of-the-art 3DGS approaches.

Conclusion: GeoTexDensifier effectively combines texture-aware densification and geometry-aware splitting to reconstruct high-quality Gaussian splats that better comply with scene geometry and texture richness, demonstrating significant improvements in 3DGS reconstruction quality.

Abstract: 3D Gaussian Splatting (3DGS) has recently attracted wide attentions in various areas such as 3D navigation, Virtual Reality (VR) and 3D simulation, due to its photorealistic and efficient rendering performance. High-quality reconstrution of 3DGS relies on sufficient splats and a reasonable distribution of these splats to fit real geometric surface and texture details, which turns out to be a challenging problem. We present GeoTexDensifier, a novel geometry-texture-aware densification strategy to reconstruct high-quality Gaussian splats which better comply with the geometric structure and texture richness of the scene. Specifically, our GeoTexDensifier framework carries out an auxiliary texture-aware densification method to produce a denser distribution of splats in fully textured areas, while keeping sparsity in low-texture regions to maintain the quality of Gaussian point cloud. Meanwhile, a geometry-aware splitting strategy takes depth and normal priors to guide the splitting sampling and filter out the noisy splats whose initial positions are far from the actual geometric surfaces they aim to fit, under a Validation of Depth Ratio Change checking. With the help of relative monocular depth prior, such geometry-aware validation can effectively reduce the influence of scattered Gaussians to the final rendering quality, especially in regions with weak textures or without sufficient training views. The texture-aware densification and geometry-aware splitting strategies are fully combined to obtain a set of high-quality Gaussian splats. We experiment our GeoTexDensifier framework on various datasets and compare our Novel View Synthesis results to other state-of-the-art 3DGS approaches, with detailed quantitative and qualitative evaluations to demonstrate the effectiveness of our method in producing more photorealistic 3DGS models.

Method: PART uses continuous relative transformations between off-grid patches to model how parts relate to each other in continuous space. It learns relative composition through off-grid structural relative positioning that is less tied to absolute appearance and remains coherent under variations like partial visibility or stylistic changes.

Result: PART outperforms grid-based methods like MAE and DropPos on tasks requiring precise spatial understanding (object detection and time series prediction), while maintaining competitive performance on global classification tasks.

Conclusion: By breaking free from grid constraints, PART opens up new possibilities for universal self-supervised pretraining across diverse datatypes (images, EEG signals, medical imaging, video, audio) with better spatial understanding capabilities.

Abstract: The composition of objects and their parts, along with object-object positional relationships, provides a rich source of information for representation learning. Hence, spatial-aware pretext tasks have been actively explored in self-supervised learning. Existing works commonly start from a grid structure, where the goal of the pretext task involves predicting the absolute position index of patches within a fixed grid. However, grid-based approaches fall short of capturing the fluid and continuous nature of real-world object compositions. We introduce PART, a self-supervised learning approach that leverages continuous relative transformations between off-grid patches to overcome these limitations. By modeling how parts relate to each other in a continuous space, PART learns the relative composition of images-an off-grid structural relative positioning that is less tied to absolute appearance and can remain coherent under variations such as partial visibility or stylistic changes. In tasks requiring precise spatial understanding such as object detection and time series prediction, PART outperforms grid-based methods like MAE and DropPos, while maintaining competitive performance on global classification tasks. By breaking free from grid constraints, PART opens up a new trajectory for universal self-supervised pretraining across diverse datatypes-from images to EEG signals-with potential in medical imaging, video, and audio.

Method: 1) Construct hierarchical road models from real-world driving data using vehicle-mounted sensors to capture detailed road defect structures and surface elevations. 2) Generate digital road twins to create simulation environments for algorithm analysis. 3) Import scenarios into a simulator for data acquisition and physical simulation.

Result: Experimental results show that driving tasks including perception and decision-making benefit significantly from the high-fidelity road defect scenes generated by the proposed system.

Conclusion: The proposed urban digital twin system with multi-modal sensors addresses road inspection challenges by generating realistic synthetic road defect data, enabling better algorithm development and evaluation for autonomous driving tasks involving road surface interactions.

Abstract: Road inspection is crucial for maintaining road serviceability and ensuring traffic safety, as road defects gradually develop and compromise functionality. Traditional inspection methods, which rely on manual evaluations, are labor-intensive, costly, and time-consuming. While data-driven approaches are gaining traction, the scarcity and spatial sparsity of real-world road defects present significant challenges in acquiring high-quality datasets. Existing simulators designed to generate detailed synthetic driving scenes, however, lack models for road defects. Moreover, advanced driving tasks that involve interactions with road surfaces, such as planning and control in defective areas, remain underexplored. To address these limitations, we propose a multi-modal sensor platform integrated with an urban digital twin (UDT) system for intelligent road inspection. First, hierarchical road models are constructed from real-world driving data collected using vehicle-mounted sensors, resulting in highly detailed representations of road defect structures and surface elevations. Next, digital road twins are generated to create simulation environments for comprehensive analysis and evaluation of algorithm performance. These scenarios are then imported into a simulator to facilitate both data acquisition and physical simulation. Experimental results demonstrate that driving tasks, including perception and decision-making, benefit significantly from the high-fidelity road defect scenes generated by our system.

Method: Proposes DPBridge with: (1) tractable reverse transition kernel for diffusion bridge process enabling maximum likelihood training, (2) distribution-aligned normalization and image consistency loss for finetuning, integrating diffusion bridge formulation with structured visual priors.

Result: Consistently achieves superior performance across extensive benchmarks, demonstrating effectiveness and generalization capability under different scenarios.

Conclusion: DPBridge successfully addresses limitations of conventional diffusion models for dense prediction by efficiently leveraging input images as informative priors and integrating with pretrained foundation models through a novel diffusion bridge framework.

Abstract: Diffusion models demonstrate remarkable capabilities in capturing complex data distributions and have achieved compelling results in many generative tasks. While they have recently been extended to dense prediction tasks such as depth estimation and surface normal prediction, their full potential in this area remains underexplored. As target signal maps and input images are pixel-wise aligned, the conventional noise-to-data generation paradigm is inefficient, and input images can serve as a more informative prior compared to pure noise. Diffusion bridge models, which support data-to-data generation between two general data distributions, offer a promising alternative, but they typically fail to exploit the rich visual priors embedded in large pretrained foundation models. To address these limitations, we integrate diffusion bridge formulation with structured visual priors and introduce DPBridge, the first latent diffusion bridge framework for dense prediction tasks. To resolve the incompatibility between diffusion bridge models and pretrained diffusion backbones, we propose (1) a tractable reverse transition kernel for the diffusion bridge process, enabling maximum likelihood training scheme; (2) finetuning strategies including distribution-aligned normalization and image consistency loss. Experiments across extensive benchmarks validate that our method consistently achieves superior performance, demonstrating its effectiveness and generalization capability under different scenarios.

Method: HVLFormer uses a mask transformer architecture with three key components: 1) Transform text embeddings into textual object queries to generate multi-scale, dataset-aware queries capturing class semantics from coarse to fine granularity; 2) Refine queries by injecting image-specific visual context to align textual semantics with local scene structures; 3) Introduce cross-view and modal consistency regularization for domain robustness, enforcing prediction consistency across augmented views and stable vision-language alignment during decoding.

Result: With less than 1% training data, HVLFormer outperforms state-of-the-art methods on Pascal VOC, COCO, ADE20K, and Cityscapes datasets.

Conclusion: HVLFormer effectively addresses semantic misalignment in semi-supervised semantic segmentation by achieving domain-aware and domain-robust alignment between visual and textual representations, leading to superior performance with minimal labeled data.

Abstract: Vision Language Models (VLMs) provide rich semantic priors but are underexplored in Semi supervised Semantic Segmentation. Recent attempts to integrate VLMs to inject high level semantics overlook the semantic misalignment between visual and textual representations that arises from using domain invariant text embeddings without adapting them to dataset and image specific contexts. This lack of domain awareness, coupled with limited annotations, weakens the model semantic understanding by preventing effective vision language alignment. As a result, the model struggles with contextual reasoning, shows weak intra class discrimination, and confuses similar classes. To address these challenges, we propose Hierarchical Vision Language transFormer (HVLFormer), which achieves domain aware and domain robust alignment between visual and textual representations within a mask transformer architecture. Firstly, we transform text embeddings from pretrained VLMs into textual object queries, enabling the generation of multi scale, dataset aware queries that capture class semantics from coarse to fine granularity and enhance contextual reasoning. Next, we refine these queries by injecting image specific visual context to align textual semantics with local scene structures and enhance class discrimination. Finally, to achieve domain robustness, we introduce cross view and modal consistency regularization, which enforces prediction consistency within mask-transformer architecture across augmented views. Moreover, it ensures stable vision language alignment during decoding. With less than 1% training data, HVLFormer outperforms state of the art methods on Pascal VOC, COCO, ADE20K, and Cityscapes. Our code and results will be available on GitHub.

Method: The approach combines: 1) class-variance optimized clustering with cluster separation tuner to improve clustering of labeled/unlabeled samples, 2) restricted pseudo-labeling to optimize the clustering-based labeling process, and 3) semantic information injection to enhance semi-supervised few-shot learning performance.

Result: The method significantly outperforms recent state-of-the-art methods on benchmark datasets. Extensive experiments under challenging conditions show the model generalizes well to domain shifts and achieves new state-of-the-art performance in open-set settings with distractor classes.

Conclusion: The proposed approach effectively improves semi-supervised few-shot learning by enhancing clustering quality through representation learning improvements, demonstrating robustness to domain shifts and effectiveness for real-world applications.

Abstract: Few-shot learning has been extensively explored to address problems where the amount of labeled samples is very limited for some classes. In the semi-supervised few-shot learning setting, substantial quantities of unlabeled samples are available. Such unlabeled samples are generally cheaper to obtain and can be used to improve the few-shot learning performance of the model. Some of the recent methods for this setting rely on clustering to generate pseudo-labels for the unlabeled samples. Since the effectiveness of clustering heavily influences the labeling of the unlabeled samples, it can significantly affect the few-shot learning performance. In this paper, we focus on improving the representation learned by the model in order to improve the clustering and, consequently, the model performance. We propose an approach for semi-supervised few-shot learning that performs a class-variance optimized clustering coupled with a cluster separation tuner in order to improve the effectiveness of clustering the labeled and unlabeled samples in this setting. It also optimizes the clustering-based pseudo-labeling process using a restricted pseudo-labeling approach and performs semantic information injection in order to improve the semi-supervised few-shot learning performance of the model. We experimentally demonstrate that our proposed approach significantly outperforms recent state-of-the-art methods on the benchmark datasets. To further establish its robustness, we conduct extensive experiments under challenging conditions, showing that the model generalizes well to domain shifts and achieves new state-of-the-art performance in open-set settings with distractor classes, highlighting its effectiveness for real-world applications.

Method: Uses a ceiling-mounted projector to project multi-scale markers (2D patterns at varying scales) to establish point correspondences across cameras with different viewpoints and zoom levels, enabling fully automatic calibration.

Result: Achieves accuracy comparable to manual marker-based methods with higher robustness under significant zoom differences; shows that state-of-the-art SfM pipelines are ineffective even with additional texture projection.

Conclusion: Ceiling-mounted projector with multi-scale markers provides effective automated alternative to operator-dependent calibration, enabling fully automated 3D surgical scene reconstruction.

Abstract: The purpose of this study is to develop an automated and accurate external camera calibration method for multi-camera systems used in 3D surgical scene reconstruction (3D-SSR), eliminating the need for operator intervention or specialized expertise. The method specifically addresses the problem of limited overlapping fields of view caused by significant variations in optical zoom levels and camera locations. We contribute a novel, fast, and fully automatic calibration method based on the projection of multi-scale markers (MSMs) using a ceiling-mounted projector. MSMs consist of 2D patterns projected at varying scales, ensuring accurate extraction of well distributed point correspondences across significantly different viewpoints and zoom levels. Validation is performed using both synthetic and real data captured in a mock-up OR, with comparisons to traditional manual marker-based methods as well as markerless calibration methods. The method achieves accuracy comparable to manual, operator-dependent calibration methods while exhibiting higher robustness under conditions of significant differences in zoom levels. Additionally, we show that state-of-the-art Structure-from-Motion (SfM) pipelines are ineffective in 3D-SSR settings, even when additional texture is projected onto the OR floor. The use of a ceiling-mounted entry-level projector proves to be an effective alternative to operator-dependent, traditional marker-based methods, paving the way for fully automated 3D-SSR.

Method: The authors selected and gradually improved a pure fully connected neural network. Key improvements include: 1) Changing from one-stage to two-stage training to better distinguish similar rice types, and 2) Changing preprocessing from random tilting to horizontal/vertical position correction. The dataset includes global and domestic rice images from websites and laboratories.

Result: The proposed improvements significantly increased classification accuracy from 97% to 99%, demonstrating enhanced ability to distinguish between similar rice varieties.

Conclusion: Two subtle methods (two-stage training and position correction preprocessing) can remarkably enhance the classification ability of deep learning models for rice grain classification, particularly for distinguishing similar varieties.

Abstract: Rice is a staple food for a significant portion of the world’s population, providing essential nutrients and serving as a versatile in-gredient in a wide range of culinary traditions. Recently, the use of deep learning has enabled automated classification of rice, im-proving accuracy and efficiency. However, classical models based on first-stage training may face difficulties in distinguishing between rice varieties with similar external characteristics, thus leading to misclassifications. Considering the transparency and feasibility of model, we selected and gradually improved pure fully connected neural network to achieve classification of rice grain. The dataset we used contains both global and domestic rice images obtained from websites and laboratories respectively. First, the training mode was changed from one-stage training to two-stage training, which significantly contributes to distinguishing two similar types of rice. Secondly, the preprocessing method was changed from random tilting to horizontal or vertical position cor-rection. After those two enhancements, the accuracy of our model increased notably from 97% to 99%. In summary, two subtle methods proposed in this study can remarkably enhance the classification ability of deep learning models in terms of the classification of rice grain.

Method: Decouples 6D pose estimation into two steps: first estimate 3D translation to resolve scale/depth ambiguities, then determine 3D orientation via canonical scale template matching. Introduces active perception strategy for next-best-view prediction.

Result: Outperforms state-of-the-art on ROBI, TOD, and T-ROBI datasets. Multi-view estimation significantly improves accuracy, and next-best-view strategy achieves high accuracy with fewer viewpoints than heuristic policies.

Conclusion: The proposed two-step decoupling approach combined with active perception effectively addresses textureless object pose estimation challenges, achieving superior performance with efficient viewpoint selection.

Abstract: Estimating the 6D pose of textureless objects from RGB images is an important problem in robotics. Due to appearance ambiguities, rotational symmetries, and severe occlusions, single-view based 6D pose estimators are still unable to handle a wide range of objects, motivating research towards multi-view pose estimation and next-best-view prediction that addresses these limitations. In this work, we propose a comprehensive active perception framework for estimating the 6D poses of textureless objects using only RGB images. Our approach is built upon a key idea: decoupling the 6D pose estimation into a two-step sequential process can greatly improve both accuracy and efficiency. First, we estimate the 3D translation of each object, resolving scale and depth ambiguities inherent to RGB images. These estimates are then used to simplify the subsequent task of determining the 3D orientation, which we achieve through canonical scale template matching. Building on this formulation, we then introduce an active perception strategy that predicts the next best camera viewpoint to capture an RGB image, effectively reducing object pose uncertainty and enhancing pose accuracy. We evaluate our method on the public ROBI and TOD datasets, as well as on our reconstructed transparent object dataset, T-ROBI. Under the same camera viewpoints, our multi-view pose estimation significantly outperforms state-of-the-art approaches. Furthermore, by leveraging our next-best-view strategy, our approach achieves high pose accuracy with fewer viewpoints than heuristic-based policies across all evaluated datasets. The accompanying video and T-ROBI dataset will be released on our project page: https://trailab.github.io/ActiveODPE.

Method: Created CineBrain dataset with synchronized fMRI and EEG recordings during 6 hours of audiovisual content (The Big Bang Theory). Proposed CineSync framework with Multi-Modal Fusion Encoder and Neural Latent Decoder for dynamic video reconstruction from combined fMRI+EEG signals.

Result: CineSync achieves state-of-the-art performance in dynamic video reconstruction. Analysis shows auditory cortical activations significantly improve decoding accuracy, demonstrating auditory input’s role in enhancing visual perception reconstruction.

Conclusion: Multimodal brain signals (fMRI+EEG) provide complementary information for video reconstruction, with auditory cortex activations playing a crucial role in improving visual fidelity, highlighting the importance of considering multimodal integration in brain decoding research.

Abstract: Most research decoding brain signals into images, often using them as priors for generative models, has focused only on visual content. This overlooks the brain’s natural ability to integrate auditory and visual information, for instance, sound strongly influences how we perceive visual scenes. To investigate this, we propose a new task of reconstructing continuous video stimuli from multimodal brain signals recorded during audiovisual stimulation. To enable this, we introduce CineBrain, the first large-scale dataset that synchronizes fMRI and EEG during audiovisual viewing, featuring six hours of \textit{The Big Bang Theory} episodes for cross-modal alignment. We also conduct the first systematic exploration of combining fMRI and EEG for video reconstruction and present CineSync, a framework for reconstructing dynamic video using a Multi-Modal Fusion Encoder and a Neural Latent Decoder. CineSync achieves state-of-the-art performance in dynamic reconstruction, leveraging the complementary strengths of fMRI and EEG to improve visual fidelity. Our analysis shows that auditory cortical activations enhance decoding accuracy, highlighting the role of auditory input in visual perception. Project Page: https://jianxgao.github.io/CineBrain.

Method: Proposes FourierSR, a Fourier token-based plugin inspired by modeling convolution theorem through token mix. It uses only Fourier transform and multiplication operations instead of convolutions or window-based Transformers, providing global receptive fields while greatly reducing complexity.

Result: FourierSR brings an average PSNR gain of 0.34dB for existing efficient SR methods on Manga109 test set at x4 scale, with only 0.6% increase in parameters and 1.5% increase in FLOPs compared to original model sizes.

Conclusion: FourierSR is an effective plug-and-play unit that uniformly improves SR performance with minimal computational overhead, addressing the limitations of existing approaches by leveraging Fourier transforms for global receptive fields and efficiency.

Abstract: Image super-resolution (SR) aims to recover low-resolution images to high-resolution images, where improving SR efficiency is a high-profile challenge. However, commonly used units in SR, like convolutions and window-based Transformers, have limited receptive fields, making it challenging to apply them to improve SR under extremely limited computational cost. To address this issue, inspired by modeling convolution theorem through token mix, we propose a Fourier token-based plugin called FourierSR to improve SR uniformly, which avoids the instability or inefficiency of existing token mix technologies when applied as plug-ins. Furthermore, compared to convolutions and windows-based Transformers, our FourierSR only utilizes Fourier transform and multiplication operations, greatly reducing complexity while having global receptive fields. Experimental results show that our FourierSR as a plug-and-play unit brings an average PSNR gain of 0.34dB for existing efficient SR methods on Manga109 test set at the scale of x4, while the average increase in the number of Params and FLOPs is only 0.6% and 1.5% of original sizes. We will release our codes upon acceptance.

Method: A coarse-to-fine AI framework with two novel techniques: (1) distribution-aware stratified ensembling to handle lesion variations, and (2) peak-scaled lesion candidate extraction for precise localization on contrast-enhanced CT scans.

Result: Ranked 1st in PANORAMA challenge with AUROC 0.9263 and AP 0.7243 on official test set. Failure cases were analyzed with radiologists to identify model limitations.

Conclusion: PanDx demonstrates strong performance for PDAC detection on CT scans, with publicly available code. Future work should address identified limitations to further improve clinical utility.

Abstract: Pancreatic ductal adenocarcinoma (PDAC) is one of the most aggressive forms of pancreatic cancer and is often diagnosed at an advanced stage due to subtle early imaging signs. To enable earlier detection and improve clinical decision-making, we propose a coarse-to-fine AI-assisted framework named PanDx for identifying PDAC on contrast-enhanced CT scans. Our approach integrates two novel techniques: (1) distribution-aware stratified ensembling to improve generalization across lesion variations, and (2) peak-scaled lesion candidate extraction to enhance lesion localization precision. PanDx is developed and evaluated as part of the PANORAMA challenge, where it ranked 1st place on the official test set with an AUROC of 0.9263 and an AP of 0.7243. Furthermore, we analyzed failure cases with a radiologist to identify the limitations of AI models on this task and discussed potential future directions for model improvement. Our code and models are publicly available at https://github.com/han-liu/PanDx.

Method: Dynamic Density Pruning (FastVID) that partitions videos into temporally ordered segments and applies density-based token pruning to maintain essential spatial and temporal information.

Result: Achieves 90.3% token pruning, reduces FLOPs to 8.3%, accelerates LLM prefill stage by 7.1× while maintaining 98.0% original accuracy on LLaVA-OneVision-7B, with SOTA performance across multiple Video LLMs.

Conclusion: FastVID effectively addresses video token redundancy through systematic spatiotemporal analysis, enabling efficient Video LLM deployment with minimal accuracy loss.

Abstract: Video Large Language Models have demonstrated strong video understanding capabilities, yet their practical deployment is hindered by substantial inference costs caused by redundant video tokens. Existing pruning techniques fail to effectively exploit the spatiotemporal redundancy present in video data. To bridge this gap, we perform a systematic analysis of video redundancy from two perspectives: temporal context and visual context. Leveraging these insights, we propose Dynamic Density Pruning for Fast Video LLMs termed FastVID. Specifically, FastVID dynamically partitions videos into temporally ordered segments to preserve temporal structure and applies a density-based token pruning strategy to maintain essential spatial and temporal information. Our method significantly reduces computational overhead while maintaining temporal and visual integrity. Extensive evaluations show that FastVID achieves state-of-the-art performance across various short- and long-video benchmarks on leading Video LLMs, including LLaVA-OneVision, LLaVA-Video, Qwen2-VL, and Qwen2.5-VL. Notably, on LLaVA-OneVision-7B, FastVID effectively prunes $\textbf{90.3%}$ of video tokens, reduces FLOPs to $\textbf{8.3%}$, and accelerates the LLM prefill stage by $\textbf{7.1}\times$, while maintaining $\textbf{98.0%}$ of the original accuracy. The code is available at https://github.com/LunarShen/FastVID.

Method: Proposes a distribution matching framework that jointly optimizes geometric structures and orientations. Introduces Semantically Aligned Distribution Matching loss to handle unordered point indexing, and learns optimal rotation angles while updating the synthetic dataset.

Result: Extensive experiments show the method consistently outperforms existing dataset distillation methods, achieving superior accuracy and strong cross-architecture generalization on benchmark datasets.

Conclusion: The proposed framework effectively addresses the unique challenges of 3D point cloud distillation through semantic alignment and joint rotation learning, demonstrating state-of-the-art performance.

Abstract: Large-scale datasets are usually required to train deep neural networks, but it increases the computational complexity hindering the practical applications. Recently, dataset distillation for images and texts has been attracting a lot of attention, that reduces the original dataset to a synthetic dataset to alleviate the computational burden of training while preserving essential task-relevant information. However, the dataset distillation for 3D point clouds remains largely unexplored, as the point clouds exhibit fundamentally different characteristics from that of images, making the dataset distillation more challenging. In this paper, we propose a distribution matching-based distillation framework for 3D point clouds that jointly optimizes the geometric structures as well as the orientations of the synthetic 3D objects. To address the semantic misalignment caused by unordered indexing of points, we introduce a Semantically Aligned Distribution Matching loss computed on the sorted features in each channel. Moreover, to address the rotation variation, we jointly learn the optimal rotation angles while updating the synthetic dataset to better align with the original feature distribution. Extensive experiments on widely used benchmark datasets demonstrate that the proposed method consistently outperforms existing dataset distillation methods, achieving superior accuracy and strong cross-architecture generalization.

Method: Proposes VisualCloze with visual in-context learning (instead of language instructions), Graph200K dataset for dense task relationships, and leverages pre-trained infilling models without architecture changes.

Result: Supports wide range of in-domain tasks, generalization to unseen tasks, unification of multiple tasks, and reverse generation.

Conclusion: VisualCloze addresses limitations of task-specific models and existing universal approaches through visual demonstrations, structured task datasets, and unified formulation with infilling models.

Abstract: Recent progress in diffusion models significantly advances various image generation tasks. However, the current mainstream approach remains focused on building task-specific models, which have limited efficiency when supporting a wide range of different needs. While universal models attempt to address this limitation, they face critical challenges, including generalizable task instruction, appropriate task distributions, and unified architectural design. To tackle these challenges, we propose VisualCloze, a universal image generation framework, which supports a wide range of in-domain tasks, generalization to unseen ones, unseen unification of multiple tasks, and reverse generation. Unlike existing methods that rely on language-based task instruction, leading to task ambiguity and weak generalization, we integrate visual in-context learning, allowing models to identify tasks from visual demonstrations. Meanwhile, the inherent sparsity of visual task distributions hampers the learning of transferable knowledge across tasks. To this end, we introduce Graph200K, a graph-structured dataset that establishes various interrelated tasks, enhancing task density and transferable knowledge. Furthermore, we uncover that our unified image generation formulation shared a consistent objective with image infilling, enabling us to leverage the strong generative priors of pre-trained infilling models without modifying the architectures.

Method: Proposes Relation-R1 framework with two main components: 1) Cognitive chain-of-thought guided supervised fine-tuning (SFT) to establish foundational reasoning with structured thinking processes, and 2) Group relative policy optimization (GRPO) reinforcement learning to refine outputs via multi-rewards optimization, prioritizing visual-semantic grounding over language biases.

Result: Achieves state-of-the-art performance on widely-used PSG and SWiG datasets for both binary and N-ary relation understanding. The specific-to-general progressive CoT approach further improves generalization, especially for capturing synonymous N-ary relations.

Conclusion: Relation-R1 successfully addresses the limitations of current MLLMs in visual relation understanding by explicitly modeling structural semantic dependencies through a unified framework combining CoT-guided SFT and GRPO reinforcement learning.

Abstract: Recent advances in multi-modal large language models (MLLMs) have significantly improved object-level grounding and region captioning. However, they remain limited in visual relation understanding, struggling even with binary relation detection, let alone \textit{N}-ary relations involving multiple semantic roles. The core reason is the lack of modeling for \textit{structural semantic dependencies} among multi-entities, leading to unreliable outputs, hallucinations, and over-reliance on language priors (\eg, defaulting to ``person drinks a milk’’ if a person is merely holding it). To this end, we propose Relation-R1, the \textit{first unified} relation comprehension framework that explicitly integrates cognitive chain-of-thought (CoT)-guided supervised fine-tuning (SFT) and group relative policy optimization (GRPO) within a reinforcement learning (RL) paradigm. Specifically, we first establish foundational reasoning capabilities via SFT, enforcing structured outputs with thinking processes. Then, GRPO is utilized to refine these outputs via multi-rewards optimization, prioritizing visual-semantic grounding over language-induced biases, thereby improving generalization capability. Furthermore, we investigate the impact of various CoT strategies within this framework, demonstrating that a specific-to-general progressive approach in CoT guidance further improves generalization, especially in capturing synonymous \textit{N}-ary relations. Extensive experiments on widely-used PSG and SWiG datasets demonstrate that Relation-R1 achieves state-of-the-art performance in both binary and \textit{N}-ary relation understanding.

Method: Proposes INKL-Pose with: 1) Instance-Adaptive Keypoint Detector for semantically consistent keypoints, 2) Local Keypoint Feature Aggregator for fine-grained geometries, 3) Global Keypoint Feature Aggregator using bidirectional Mamba with Feature Sequence Flipping for structural consistency, 4) Surface loss and separation loss for uniform keypoint coverage and diversity, 5) Mapping to canonical space for 6D pose and size regression.

Result: Achieves state-of-the-art performance on CAMERA25, REAL275, and HouseCat6D benchmarks with 16.7M parameters and runs at 36 FPS on NVIDIA RTX 4090D GPU.

Conclusion: INKL-Pose effectively addresses intra-class variation challenges through instance-adaptive keypoint learning and local-to-global geometric aggregation, demonstrating strong generalization and computational efficiency for category-level object pose estimation.

Abstract: Category-level object pose estimation aims to predict the 6D pose and size of previously unseen instances from predefined categories, requiring strong generalization across diverse object instances. Although many previous methods attempt to mitigate intra-class variations, they often struggle with instances exhibiting complex geometries or significant deviations from canonical shapes. To address this issue, we propose INKL-Pose, a novel category-level object pose estimation framework that enables INstance-adaptive Keypoint Learning with local-to-global geometric aggregation. Specifically, our method first predicts semantically consistent and geometrically informative keypoints using an Instance-Adaptive Keypoint Detector, then refines them: (1) a Local Keypoint Feature Aggregator capturing fine-grained geometries, and (2) a Global Keypoint Feature Aggregator using bidirectional Mamba for structural consistency. To enable bidirectional modeling in Mamba, we introduce a simple yet effective Feature Sequence Flipping strategy that preserves spatial coherence while constructing backward feature sequence. Additionally, we design a surface loss and a separation loss to encourage uniform coverage and spatial diversity in keypoint distribution. The resulting keypoints are mapped to a canonical space for 6D pose and size regression. Extensive experiments on CAMERA25, REAL275, and HouseCat6D show that INKL-Pose achieves state-of-the-art performance with 16.7M parameters and runs at 36 FPS on an NVIDIA RTX 4090D GPU.

Method: A bi-scale representation using 2D opaque surfels for coarse geometry and view-dependent colors, supplemented by 3D Gaussians for fine details. Two-pass rendering: surfels rasterized first via standard graphics pipeline, then Gaussians splatted with depth testing. Optimization uses coarse-to-fine procedure from multi-view images.

Result: GES achieves state-of-the-art ultra-fast high-fidelity radiance field rendering with view-consistent images, avoiding popping artifacts. Extensions include Mip-GES (anti-aliasing), Speedy-GES (boosted rendering), Compact-GES (storage efficiency), and 2D-GES (better geometry).

Conclusion: Gaussian-enhanced Surfels advance radiance field rendering as a compelling representation that combines fast rendering, high fidelity, view consistency, and extensibility for various rendering optimizations.

Abstract: We introduce Gaussian-enhanced Surfels (GESs), a bi-scale representation for radiance field rendering, wherein a set of 2D opaque surfels with view-dependent colors represent the coarse-scale geometry and appearance of scenes, and a few 3D Gaussians surrounding the surfels supplement fine-scale appearance details. The rendering with GESs consists of two passes – surfels are first rasterized through a standard graphics pipeline to produce depth and color maps, and then Gaussians are splatted with depth testing and color accumulation on each pixel order independently. The optimization of GESs from multi-view images is performed through an elaborate coarse-to-fine procedure, faithfully capturing rich scene appearance. The entirely sorting-free rendering of GESs not only achieves very fast rates, but also produces view-consistent images, successfully avoiding popping artifacts under view changes. The basic GES representation can be easily extended to achieve anti-aliasing in rendering (Mip-GES), boosted rendering speeds (Speedy-GES) and compact storage (Compact-GES), and reconstruct better scene geometries by replacing 3D Gaussians with 2D Gaussians (2D-GES). Experimental results show that GESs advance the state-of-the-arts as a compelling representation for ultra-fast high-fidelity radiance field rendering.

Method: Three-pronged approach: 1) Offline data curation with U-shaped difficulty distribution analysis and filtering of too simple/extremely difficult prompts; 2) Online advantage differentiation using group-wise empirical accuracy as difficulty proxy for adaptive advantage reweighting; 3) Difficulty hints as explicit prompts for complex samples in second training stage to encourage reasoning depth calibration and reflective validation.

Result: Demonstrates significant performance improvements across various multimodal mathematical reasoning benchmarks using only 2K+0.6K two-stage training data.

Conclusion: Explicitly modeling problem difficulty prior information through comprehensive difficulty-aware strategies effectively shapes reinforcement learning fine-tuning for multimodal reasoning, achieving strong results with minimal training data.

Abstract: In this work, we investigate how explicitly modeling problem’s difficulty prior information shapes the effectiveness of reinforcement learning based fine-tuning for multimodal reasoning. Our exploration mainly comprises of following three perspective: First, through offline data curation, we analyze the U-shaped difficulty distribution of two given datasets using the base model by multi-round sampling, and then filter out prompts that are either too simple or extremely difficult to provide meaningful gradients and perform subsequent two-stage training. Second, we implement an online advantage differentiation, computing group-wise empirical accuracy as a difficulty proxy to adaptively reweight advantages estimation, providing stronger learning signals for more challenging problems. Finally, we introduce difficulty hints as explicit prompts for more complex samples in the second training stage, encouraging the model to calibrate its reasoning depth and perform reflective validation checks. Our comprehensive approach demonstrates significant performances across various multi-modal mathematical reasoning benchmarks with only 2K+0.6K two-stage training data.

Method: Proposes Condition Modulated Expert (CoMoE) module that aggregates similar patch features and assigns them to dedicated experts, plus WeaveNet - a dynamic snake-like connection mechanism for effective interaction between backbone and control branches.

Result: State-of-the-art performance on Subjects-200K and MultiGen-20M datasets across various conditional image generation tasks, demonstrating advantages in both versatility and effectiveness.

Conclusion: UniGen provides an efficient unified framework for diverse conditional image generation that mitigates feature entanglement and redundant computation while achieving superior performance.

Abstract: The image-to-image generation task aims to produce controllable images by leveraging conditional inputs and prompt instructions. However, existing methods often train separate control branches for each type of condition, leading to redundant model structures and inefficient use of computational resources. To address this, we propose a Unified image-to-image Generation (UniGen) framework that supports diverse conditional inputs while enhancing generation efficiency and expressiveness. Specifically, to tackle the widely existing parameter redundancy and computational inefficiency in controllable conditional generation architectures, we propose the Condition Modulated Expert (CoMoE) module. This module aggregates semantically similar patch features and assigns them to dedicated expert modules for visual representation and conditional modeling. By enabling independent modeling of foreground features under different conditions, CoMoE effectively mitigates feature entanglement and redundant computation in multi-condition scenarios. Furthermore, to bridge the information gap between the backbone and control branches, we propose WeaveNet, a dynamic, snake-like connection mechanism that enables effective interaction between global text-level control from the backbone and fine-grained control from conditional branches. Extensive experiments on the Subjects-200K and MultiGen-20M datasets across various conditional image generation tasks demonstrate that our method consistently achieves state-of-the-art performance, validating its advantages in both versatility and effectiveness. The code has been uploaded to https://github.com/gavin-gqzhang/UniGen.

Method: Proposes filter normalization followed by learnable scaling and shifting (similar to batch normalization) to make filters atmosphere-equivariant and enable co-domain symmetry. Integrates classical filtering principles into deep learning for both CNNs and convolution-dependent vision transformers.

Result: Significant improvements on artificial and natural intensity variation benchmarks. ResNet34 with filter normalization could outperform CLIP by a large margin. Analysis shows unnormalized filters degrade performance, while filter normalization regularizes learning, promotes diversity, and improves robustness and generalization.

Conclusion: Filter normalization is a simple yet effective modification that addresses the lack of constraints in learned convolutional filters, making them atmosphere-equivariant and improving performance across various benchmarks while enhancing robustness and generalization.

Abstract: Classical image filters, such as those for averaging or differencing, are carefully normalized to ensure consistency, interpretability, and to avoid artifacts like intensity shifts, halos, or ringing. In contrast, convolutional filters learned end-to-end in deep networks lack such constraints. Although they may resemble wavelets and blob/edge detectors, they are not normalized in the same or any way. Consequently, when images undergo atmospheric transfer, their responses become distorted, leading to incorrect outcomes. We address this limitation by proposing filter normalization, followed by learnable scaling and shifting, akin to batch normalization. This simple yet effective modification ensures that the filters are atmosphere-equivariant, enabling co-domain symmetry. By integrating classical filtering principles into deep learning (applicable to both convolutional neural networks and convolution-dependent vision transformers), our method achieves significant improvements on artificial and natural intensity variation benchmarks. Our ResNet34 could even outperform CLIP by a large margin. Our analysis reveals that unnormalized filters degrade performance, whereas filter normalization regularizes learning, promotes diversity, and improves robustness and generalization.

Method: 1) WakeupUrbanBench: Expert-annotated segmentation dataset based on mid-20th century RS imagery covering 4 cities across 2 continents with 4 urban classes. 2) WakeupUSM: Unsupervised semantic segmentation framework using confidence-aware alignment mechanism and focal-confidence loss within a self-supervised learning architecture to generate robust pseudo-labels and adaptively handle prediction difficulty and label reliability.

Result: Comprehensive experiments show WakeupUSM significantly outperforms existing unsupervised segmentation methods on both WakeupUrbanBench and public datasets, demonstrating effectiveness for historical RS imagery analysis.

Conclusion: The work provides a pioneering benchmark and framework for quantitative studies of long-term urban change using historical satellite imagery, paving the way for modern computer vision applications in urban development analysis.

Abstract: Historical satellite imagery archive, such as Keyhole satellite data, offers rare insights into understanding early urban development and long-term transformation. However, severe quality degradation ($\textit{e.g.}$, distortion, misalignment, and spectral scarcity) and the absence of annotations have long hindered its analysis. To bridge this gap and enhance understanding of urban development, we introduce $\textbf{WakeupUrbanBench}$, an annotated segmentation dataset based on historical satellite imagery with the earliest observation time among all existing remote sensing (RS) datasets, along with a framework for unsupervised segmentation tasks, $\textbf{WakeupUSM}$. First, WakeupUrbanBench serves as a pioneer, expertly annotated dataset built on mid-$20^{\text{th}}$ century RS imagery, involving four key urban classes and spanning 4 cities across 2 continents with nearly 1000 km$^2$ area of diverse urban morphologies, and additionally introducing one present-day city. Second, WakeupUSM is a novel unsupervised semantic segmentation framework for historical RS imagery. It employs a confidence-aware alignment mechanism and focal-confidence loss based on a self-supervised learning architecture, which generates robust pseudo-labels and adaptively prioritizes prediction difficulty and label reliability to improve unsupervised segmentation on noisy historical data without manual supervision. Comprehensive experiments demonstrate WakeupUSM significantly outperforms existing unsupervised segmentation methods $\textbf{both WakeupUrbanBench and public dataset}$, promising to pave the way for quantitative studies of long-term urban change using modern computer vision. Our benchmark and codes will be released at https://github.com/Tianxiang-Hao/WakeupUrban.

Method: Identified emergent symbolic mechanisms in VLMs that support binding through content-independent spatial indexing schemes, and traced binding errors to failures in these mechanisms.

Result: Found previously unknown symbolic mechanisms in VLMs that enable binding via spatial indexing, with binding failures directly attributable to breakdowns in these specific mechanisms.

Conclusion: VLMs employ emergent spatial indexing mechanisms for binding similar to language models’ symbolic processing, providing insights for reducing binding failures in these models.

Abstract: To accurately process a visual scene, observers must bind features together to represent individual objects. This capacity is necessary, for instance, to distinguish an image containing a red square and a blue circle from an image containing a blue square and a red circle. Recent work has found that language models solve this ‘binding problem’ via a set of symbol-like, content-independent indices, but it is unclear whether similar mechanisms are employed by Vision Language Models (VLMs). This question is especially relevant, given the persistent failures of VLMs on tasks that require binding. Here, we identify a previously unknown set of emergent symbolic mechanisms that support binding specifically in VLMs, via a content-independent, spatial indexing scheme. Moreover, we find that binding errors, when they occur, can be traced directly to failures in these mechanisms. Taken together, these results shed light on the mechanisms that support symbol-like processing in VLMs, and suggest possible avenues for reducing the number of binding failures exhibited by these models.

Method: Develops an unbiased projection model for circular patterns, introduces centroid uncertainty by modeling boundary points as a Markov random field with connectivity, propagates shape distribution to centroid uncertainty using Green theorem-based shape representation, and provides calibration guidelines.

Result: The proposed framework achieves superior accuracy compared to checkerboard calibration, enhances robustness through uncertainty modeling, and improves performance across pattern detection, optimization, and evaluation metrics.

Conclusion: The approach provides a complete camera calibration solution with marked gains in accuracy and robustness by addressing the bias in circular pattern projection models and incorporating uncertainty modeling, with open-source implementation available.

Abstract: Camera calibration using planar targets has been widely favored, and two types of control points have been mainly considered as measurements: the corners of the checkerboard and the centroid of circles. Since a centroid is derived from numerous pixels, the circular pattern provides more precise measurements than the checkerboard. However, the existing projection model of circle centroids is biased under lens distortion, resulting in low performance. To surmount this limitation, we propose an unbiased projection model of the circular pattern and demonstrate its superior accuracy compared to the checkerboard. Complementing this, we introduce uncertainty into circular patterns to enhance calibration robustness and completeness. Defining centroid uncertainty improves the performance of calibration components, including pattern detection, optimization, and evaluation metrics. We also provide guidelines for performing good camera calibration based on the evaluation metric. The core concept of this approach is to model the boundary points of a two-dimensional shape as a Markov random field, considering its connectivity. The shape distribution is propagated to the centroid uncertainty through an appropriate shape representation based on the Green theorem. Consequently, the resulting framework achieves marked gains in calibration accuracy and robustness. The complete source code and demonstration video are available at https://github.com/chaehyeonsong/discocal.

Method: Proposes MR-COSMO with two key components: 1) direct cross-modal alignment module for explicit 3D point cloud to text/2D image alignment, and 2) visual-text memory module with specialized feature banks (text, visual, correspondence mappings) that uses attention-based knowledge recall to dynamically enhance scene-specific representations.

Result: Achieves state-of-the-art performance across 3D instruction, reference, and semantic segmentation benchmarks through comprehensive experiments.

Conclusion: MR-COSMO effectively addresses the 3D-text alignment limitation in query-driven 3D segmentation by enabling precise fusion of geometric and semantic features through direct cross-modal alignment and dynamic memory recall.

Abstract: The rapid advancement of vision-language models (VLMs) in 3D domains has accelerated research in text-query-guided point cloud processing, though existing methods underperform in point-level segmentation due to inadequate 3D-text alignment that limits local feature-text context linking. To address this limitation, we propose MR-COSMO, a Visual-Text Memory Recall and Direct CrOSs-MOdal Alignment Method for Query-Driven 3D Segmentation, establishing explicit alignment between 3D point clouds and text/2D image data through a dedicated direct cross-modal alignment module while implementing a visual-text memory module with specialized feature banks. This direct alignment mechanism enables precise fusion of geometric and semantic features, while the memory module employs specialized banks storing text features, visual features, and their correspondence mappings to dynamically enhance scene-specific representations via attention-based knowledge recall. Comprehensive experiments across 3D instruction, reference, and semantic segmentation benchmarks confirm state-of-the-art performance.

Method: Created Neural-MedBench with multi-sequence MRI scans, structured EHRs, and clinical notes covering differential diagnosis, lesion recognition, and rationale generation. Developed hybrid scoring pipeline combining LLM-based graders, clinician validation, and semantic similarity metrics.

Result: State-of-the-art VLMs (GPT-4o, Claude-4, MedGemma) showed sharp performance drops compared to conventional datasets. Error analysis revealed reasoning failures dominate model shortcomings rather than perceptual errors.

Conclusion: Proposes Two-Axis Evaluation Framework: breadth-oriented datasets for statistical generalization and depth-oriented benchmarks like Neural-MedBench for reasoning fidelity. Released as open diagnostic testbed for rigorous, cost-effective assessment of clinically trustworthy AI.

Abstract: Recent advances in vision-language models (VLMs) have achieved remarkable performance on standard medical benchmarks, yet their true clinical reasoning ability remains unclear. Existing datasets predominantly emphasize classification accuracy, creating an evaluation illusion in which models appear proficient while still failing at high-stakes diagnostic reasoning. We introduce Neural-MedBench, a compact yet reasoning-intensive benchmark specifically designed to probe the limits of multimodal clinical reasoning in neurology. Neural-MedBench integrates multi-sequence MRI scans, structured electronic health records, and clinical notes, and encompasses three core task families: differential diagnosis, lesion recognition, and rationale generation. To ensure reliable evaluation, we develop a hybrid scoring pipeline that combines LLM-based graders, clinician validation, and semantic similarity metrics. Through systematic evaluation of state-of-the-art VLMs, including GPT-4o, Claude-4, and MedGemma, we observe a sharp performance drop compared to conventional datasets. Error analysis shows that reasoning failures, rather than perceptual errors, dominate model shortcomings. Our findings highlight the necessity of a Two-Axis Evaluation Framework: breadth-oriented large datasets for statistical generalization, and depth-oriented, compact benchmarks such as Neural-MedBench for reasoning fidelity. We release Neural-MedBench at https://neuromedbench.github.io/ as an open and extensible diagnostic testbed, which guides the expansion of future benchmarks and enables rigorous yet cost-effective assessment of clinically trustworthy AI.

Method: Flow-Anchored Consistency Model (FACM) uses Flow Matching as a dynamic anchor for the CM shortcut objective with an expanded time interval strategy that unifies optimization while decoupling tasks. Includes memory-efficient Chain-JVP for FSDP compatibility.

Result: Achieves SOTA FID of 1.32 with 2 steps and 1.70 with 1 step on ImageNet 256x256. Scales to 14B parameter model (Wan 2.2), accelerating T2I inference from 2x40 to 2-8 steps.

Conclusion: Explicitly anchoring models in underlying flows ensures trajectory fidelity and stable training, enabling efficient few-step generation at scale with architectural agnosticism.

Abstract: Continuous-time Consistency Models (CMs) promise efficient few-step generation but face significant challenges with training instability. We argue this instability stems from a fundamental conflict: Training the network exclusively on a shortcut objective leads to the catastrophic forgetting of the instantaneous velocity field that defines the flow. Our solution is to explicitly anchor the model in the underlying flow, ensuring high trajectory fidelity during training. We introduce the Flow-Anchored Consistency Model (FACM), where a Flow Matching (FM) task serves as a dynamic anchor for the primary CM shortcut objective. Key to this Flow-Anchoring approach is a novel expanded time interval strategy that unifies optimization for a single model while decoupling the two tasks to ensure stable, architecturally-agnostic training. By distilling a pre-trained LightningDiT model, our method achieves a state-of-the-art FID of 1.32 with two steps (NFE=2) and 1.70 with just one step (NFE=1) on ImageNet 256x256. To address the challenge of scalability, we develop a memory-efficient Chain-JVP that resolves key incompatibilities with FSDP. This method allows us to scale FACM training on a 14B parameter model (Wan 2.2), accelerating its Text-to-Image inference from 2x40 to 2-8 steps. Our code and pretrained models: https://github.com/ali-vilab/FACM.

Method: HTMA-Net integrates Hadamard Transform with multiplication-avoiding SRAM-based in-memory computing, selectively replacing intermediate convolutions with Hybrid Hadamard-based transform layers whose internal convolutions use multiplication-avoiding in-memory operations.

Result: HTMA-Net eliminates up to 52% of multiplications compared to baseline ResNet-18/20/32 models while achieving comparable accuracy on CIFAR-10, CIFAR-100, and Tiny ImageNet, significantly reducing computational complexity and parameter count.

Conclusion: Combining structured Hadamard transform layers with SRAM-based in-memory computing multiplication-avoiding operators is a promising approach for developing efficient deep learning architectures suitable for edge devices.

Abstract: Reducing the cost of multiplications is critical for efficient deep neural network deployment, especially in energy-constrained edge devices. In this work, we introduce HTMA-Net, a novel framework that integrates the Hadamard Transform (HT) with multiplication-avoiding (MA) SRAM-based in-memory computing to reduce arithmetic complexity while maintaining accuracy. Unlike prior methods that only target multiplications in convolutional layers or focus solely on in-memory acceleration, HTMA-Net selectively replaces intermediate convolutions with Hybrid Hadamard-based transform layers whose internal convolutions are implemented via multiplication-avoiding in-memory operations. We evaluate HTMA-Net on ResNet-18 using CIFAR-10, CIFAR-100, and Tiny ImageNet, and provide a detailed comparison against regular, MF-only, and HT-only variants. Results show that HTMA-Net eliminates up to 52% of multiplications compared to baseline ResNet-18, ResNet-20, and ResNet-32 models, while achieving comparable accuracy in evaluation and significantly reducing computational complexity and the number of parameters. Our results demonstrate that combining structured Hadamard transform layers with SRAM-based in-memory computing multiplication-avoiding operators is a promising path towards efficient deep learning architectures.

Method: Created CAST-Phys dataset with high-resolution uncompressed facial video recordings and diverse physiological signals (PPG, EDA, respiration rate) collected remotely without physical contact. Evaluated impact of individual and fused modalities for emotion recognition.

Result: Dataset enables remote signal recovery and demonstrates crucial role of physiological signals in realistic scenarios where facial expressions alone are insufficient. Shows effectiveness of remote multi-modal emotion recognition through modality fusion.

Conclusion: CAST-Phys addresses key limitations in affective computing by providing a contactless multi-modal dataset that advances remote physiological emotion recognition technologies and enables more accurate, genuine emotion analysis without physical contact interference.

Abstract: In recent years, affective computing and its applications have become a fast-growing research topic. Despite significant advancements, the lack of affective multi-modal datasets remains a major bottleneck in developing accurate emotion recognition systems. Furthermore, the use of contact-based devices during emotion elicitation often unintentionally influences the emotional experience, reducing or altering the genuine spontaneous emotional response. This limitation highlights the need for methods capable of extracting affective cues from multiple modalities without physical contact, such as remote physiological emotion recognition. To address this, we present the Contactless Affective States Through Physiological Signals Database (CAST-Phys), a novel high-quality dataset explicitly designed for multi-modal remote physiological emotion recognition using facial and physiological cues. The dataset includes diverse physiological signals, such as photoplethysmography (PPG), electrodermal activity (EDA), and respiration rate (RR), alongside high-resolution uncompressed facial video recordings, enabling the potential for remote signal recovery. Our analysis highlights the crucial role of physiological signals in realistic scenarios where facial expressions alone may not provide sufficient emotional information. Furthermore, we demonstrate the potential of remote multi-modal emotion recognition by evaluating the impact of individual and fused modalities, showcasing its effectiveness in advancing contactless emotion recognition technologies.

Method: Automated pipeline that generates pseudo-labeled point tracking data from real human motion videos. Fits SMPL models to detected humans, projects mesh vertices onto image planes, resolves occlusions via ray-casting, and filters unreliable tracks using optical flow consistency.

Result: Model trained on AnthroTAP dataset achieves state-of-the-art performance on TAP-Vid benchmark for tracking any point on diverse rigid and non-rigid objects. Outperforms recent self-supervised teacher-student models trained on larger real datasets, requiring only one day of training on 4 GPUs.

Conclusion: Structured human motion offers a scalable and effective source of real-world supervision for point tracking. AnthroTAP demonstrates that automated pseudo-labeling from human motion videos can overcome limitations of synthetic training data.

Abstract: Point tracking models often struggle to generalize to real-world videos because large-scale training data is predominantly synthetic–the only source currently feasible to produce at scale. Collecting real-world annotations, however, is prohibitively expensive, as it requires tracking hundreds of points across frames. We introduce AnthroTAP, an automated pipeline that generates large-scale pseudo-labeled point tracking data from real human motion videos. Leveraging the structured complexity of human movement-non-rigid deformations, articulated motion, and frequent occlusions-AnthroTAP fits Skinned Multi-Person Linear (SMPL) models to detected humans, projects mesh vertices onto image planes, resolves occlusions via ray-casting, and filters unreliable tracks using optical flow consistency. A model trained on the AnthroTAP dataset achieves state-of-the-art performance on TAP-Vid, a challenging general-domain benchmark for tracking any point on diverse rigid and non-rigid objects (e.g., humans, animals, robots, and vehicles). Our approach outperforms recent self-supervised teacher-student models trained on vastly larger real datasets, while requiring only one day of training on 4 GPUs. AnthroTAP shows that structured human motion offers a scalable and effective source of real-world supervision for point tracking. Code and datasets will be made publicly available.

Method: Introduces Linear Separability Ceiling (LSC) diagnostic framework, analyzes state-of-the-art VLMs, and proposes augmenting standard next-token prediction with a contrastive objective to restructure visual manifold into more linearly separable geometry.

Result: Most VLMs show an “alignment gap” - failing to generatively outperform linear separability of their representations. Few models surpass this ceiling via refined visual representations or non-linear decision logic. The proposed contrastive learning method enables models to significantly surpass LSC on abstract binary classification tasks.

Conclusion: The alignment gap in VLMs is not a fundamental limitation but a solvable visual alignment issue. Restructuring visual representations through contrastive learning improves image comparison and enables better performance on abstract reasoning tasks.

Abstract: A challenge in advancing Visual-Language Models (VLMs) is determining whether their failures on abstract reasoning tasks, such as Bongard problems, stem from flawed perception or faulty top-down reasoning. To disentangle these factors, we introduce a diagnostic framework centered on the Linear Separability Ceiling (LSC), the performance achievable by a linear classifier on a VLM’s raw visual embeddings. Applying this framework to state-of-the-art VLMs, we uncover a pervasive ``alignment gap’’, where most models fail to generatively outperform the linear separability of their representations. We find that the few models surpassing this ceiling do so via two mechanisms: by further refining visual representations into a more linearly separable format or by executing non-linear decision logic. We demonstrate that this bottleneck is not a fundamental limitation but a solvable visual alignment issue. Our method augments standard next-token prediction with a contrastive objective to restructure the visual manifold into a more one-dimensionally linear geometry, improving image-to-image comparison and enabling models to significantly surpass the LSC on abstract binary classification tasks.

Method: Reformulates pose estimation as a denoising process using diffusion models. Leverages LLMs to extract global anatomical priors and local keypoint-wise semantics from species-specific prompts, which are fused with image features via cross-attention. Uses a diffusion-based keypoint decoder for progressive refinement.

Result: Extensive experiments on public animal pose datasets demonstrate effectiveness and generalization capability, especially under challenging scenarios with diverse species, cluttered backgrounds, and incomplete keypoints.

Conclusion: DiffPose-Animal provides a novel diffusion-based approach that effectively addresses the unique challenges of animal pose estimation through semantic guidance from LLMs and progressive refinement via denoising.

Abstract: Animal pose estimation is a fundamental task in computer vision, with growing importance in ecological monitoring, behavioral analysis, and intelligent livestock management. Compared to human pose estimation, animal pose estimation is more challenging due to high interspecies morphological diversity, complex body structures, and limited annotated data. In this work, we introduce DiffPose-Animal, a novel diffusion-based framework for top-down animal pose estimation. Unlike traditional heatmap regression methods, DiffPose-Animal reformulates pose estimation as a denoising process under the generative framework of diffusion models. To enhance semantic guidance during keypoint generation, we leverage large language models (LLMs) to extract both global anatomical priors and local keypoint-wise semantics based on species-specific prompts. These textual priors are encoded and fused with image features via cross-attention modules to provide biologically meaningful constraints throughout the denoising process. Additionally, a diffusion-based keypoint decoder is designed to progressively refine pose predictions, improving robustness to occlusion and annotation sparsity. Extensive experiments on public animal pose datasets demonstrate the effectiveness and generalization capability of our method, especially under challenging scenarios with diverse species, cluttered backgrounds, and incomplete keypoints.

Method: 1) Introduces HumanSense benchmark for evaluating human-centered perception and interaction capabilities; 2) Devises multi-stage, modality-progressive reinforcement learning approach (HumanSense-Omni-Reasoning) to enhance performance; 3) Designs prompts to improve non-reasoning models in training-free manner based on observed consistent thought patterns in successful reasoning.

Result: Leading MLLMs still have considerable room for improvement, especially on advanced interaction-oriented tasks. Supplementing visual input with audio and text yields substantial improvements, and omni-modal models show advantages. HumanSense-Omni-Reasoning substantially enhances performance on higher-level understanding and interactive tasks.

Conclusion: Reasoning ability is key to providing appropriate feedback based on contextual analysis of interlocutor’s needs and emotions. The proposed benchmark and reinforcement learning approach advance MLLMs toward more human-like interactions, with potential for training-free improvements through prompt design based on observed reasoning patterns.

Abstract: While Multimodal Large Language Models (MLLMs) show immense promise for achieving truly human-like interactions, progress is hindered by the lack of fine-grained evaluation frameworks for human-centered scenarios, encompassing both the understanding of complex human intentions and the provision of empathetic, context-aware responses. Here we introduce HumanSense, a comprehensive benchmark designed to evaluate the human-centered perception and interaction capabilities of MLLMs, with a particular focus on deep understanding of extended multimodal contexts and the formulation of rational feedback. Our evaluation reveals that leading MLLMs still have considerable room for improvement, particularly for advanced interaction-oriented tasks. Supplementing visual input with audio and text information yields substantial improvements, and Omni-modal models show advantages on these tasks.Furthermore, grounded in the observation that appropriate feedback stems from a contextual analysis of the interlocutor’s needs and emotions, we posit that reasoning ability serves as the key to unlocking it. We devise a multi-stage, modality-progressive reinforcement learning approach, resulting in HumanSense-Omni-Reasoning, which substantially enhances performance on higher-level understanding and interactive tasks. Additionally, we observe that successful reasoning processes appear to exhibit consistent thought patterns. By designing corresponding prompts, we also enhance the performance of non-reasoning models in a training-free manner.Project page: \textcolor{brightpink}{https://digital-avatar.github.io/ai/HumanSense/}

Method: The paper performs a case study on color attributes as a fundamental test bed. It introduces a dedicated image editing technique specifically designed to mitigate multi-object semantic alignment issues for prompts containing multiple colors.

Result: Pretrained models significantly struggle with multiple color attributes compared to single-color prompts. Neither inference-time techniques nor existing editing methods reliably resolve these semantic misalignments. The proposed editing technique significantly boosts performance across various metrics and works with multiple text-to-image diffusion models.

Conclusion: Multi-object semantic alignment remains a significant challenge in text-to-image generation, particularly for color attributes. The proposed dedicated editing technique provides an effective solution that outperforms existing methods and works across different diffusion-based text-to-image techniques.

Abstract: Text-to-image generation has recently seen remarkable success, granting users with the ability to create high-quality images through the use of text. However, contemporary methods face challenges in capturing the precise semantics conveyed by complex multi-object prompts. Consequently, many works have sought to mitigate such semantic misalignments, typically via inference-time schemes that modify the attention layers of the denoising networks. However, prior work has mostly utilized coarse metrics, such as the cosine similarity between text and image CLIP embeddings, or human evaluations, which are challenging to conduct on a larger-scale. In this work, we perform a case study on colors – a fundamental attribute commonly associated with objects in text prompts, which offer a rich test bed for rigorous evaluation. Our analysis reveals that pretrained models struggle to generate images that faithfully reflect multiple color attributes-far more so than with single-color prompts-and that neither inference-time techniques nor existing editing methods reliably resolve these semantic misalignments. Accordingly, we introduce a dedicated image editing technique, mitigating the issue of multi-object semantic alignment for prompts containing multiple colors. We demonstrate that our approach significantly boosts performance over a wide range of metrics, considering images generated by various text-to-image diffusion-based techniques.

Method: Proposes a Gaussian-SDF hybrid representation: uses a colorized signed distance field (SDF) for smooth geometry and appearance (efficiently constructed via RGB-D fusion like geometry-based methods), and 3D Gaussians to capture underrepresented details. This enables 50% fewer Gaussians and 75% fewer optimization iterations through targeted refinement.

Result: Developed GPS-SLAM achieving over 150 fps on real-world Azure Kinect sequences - an order-of-magnitude faster than state-of-the-art techniques while maintaining comparable reconstruction quality.

Conclusion: The Gaussian-SDF hybrid representation successfully addresses computational bottlenecks in Gaussian-based SLAM, enabling real-time performance (150+ fps) without sacrificing reconstruction quality, bridging the gap between photorealistic reconstruction and computational efficiency.

Abstract: While recent Gaussian-based SLAM methods achieve photorealistic reconstruction from RGB-D data, their computational performance remains a critical bottleneck. State-of-the-art techniques operate at less than 20 fps, significantly lagging behind geometry-based approaches like KinectFusion (hundreds of fps). This limitation stems from the heavy computational burden: modeling scenes requires numerous Gaussians and complex iterative optimization to fit RGB-D data; insufficient Gaussian counts or optimization iterations cause severe quality degradation. To address this, we propose a Gaussian-SDF hybrid representation, combining a colorized signed distance field (SDF) for smooth geometry and appearance with 3D Gaussians to capture underrepresented details. The SDF is efficiently constructed via RGB-D fusion (as in geometry-based methods), while Gaussians undergo iterative optimization. Our representation enables significant Gaussian reduction (50% fewer) by avoiding full-scene Gaussian modeling, and efficient Gaussian optimization (75% fewer iterations) through targeted appearance refinement. Building upon this representation, we develop GPS-SLAM (Gaussian-plus-SDF SLAM), a real-time 3D reconstruction system achieving over 150 fps on real-world Azure Kinect sequences, faster by an order-of-magnitude than state-of-the-art techniques while maintaining comparable reconstruction quality. The source code and data are available at https://gapszju.github.io/GPS-SLAM.

Method: Proposed PolypSeg-GradCAM framework integrating U-Net architecture with pre-trained ResNet-34 backbone and Gradient-weighted Class Activation Mapping (Grad-CAM) for transparent polyp segmentation. Trained and evaluated using 5-Fold Cross-Validation on Kvasir-SEG dataset of 1,000 annotated endoscopic images.

Result: Achieved mean Dice coefficient of 0.8902 ± 0.0125, mean IoU of 0.8023, AUC-ROC of 0.9722. With optimal threshold: Sensitivity 0.9058, Precision 0.9083. Grad-CAM visualizations confirmed predictions were guided by clinically relevant regions, providing insight into model decision-making.

Conclusion: Integrating segmentation accuracy with interpretability supports development of trustworthy AI-assisted colonoscopy tools. The framework demonstrates that explainable AI can bridge the gap between technical performance and clinical adoption.

Abstract: Colorectal cancer (CRC) remains one of the leading causes of cancer-related morbidity and mortality worldwide, with gastrointestinal (GI) polyps serving as critical precursors according to the World Health Organization (WHO). Early and accurate segmentation of polyps during colonoscopy is essential for reducing CRC progression, yet manual delineation is labor-intensive and prone to observer variability. Deep learning methods have demonstrated strong potential for automated polyp analysis, but their limited interpretability remains a barrier to clinical adoption. In this study, we present PolypSeg-GradCAM, an explainable deep learning framework that integrates a U-Net architecture with a pre-trained ResNet-34 backbone and Gradient-weighted Class Activation Mapping (Grad-CAM) for transparent polyp segmentation. To ensure rigorous benchmarking, the model was trained and evaluated using 5-Fold Cross-Validation on the Kvasir-SEG dataset of 1,000 annotated endoscopic images. Experimental results show a mean Dice coefficient of 0.8902 +/- 0.0125, a mean Intersection-over-Union (IoU) of 0.8023, and an Area Under the Receiver Operating Characteristic Curve (AUC-ROC) of 0.9722. Advanced quantitative analysis using an optimal threshold yielded a Sensitivity of 0.9058 and Precision of 0.9083. Additionally, Grad-CAM visualizations confirmed that predictions were guided by clinically relevant regions, offering insight into the model’s decision-making process. This study demonstrates that integrating segmentation accuracy with interpretability can support the development of trustworthy AI-assisted colonoscopy tools.

Method: Two-stage curriculum framework: Stage 1 enhances robustness by pretraining with RGBT (RGB+Thermal) data. Stage 2 improves generalization via vision-language alignment for open-world detection. To prevent catastrophic forgetting between stages, an Exponential Moving Average (EMA) mechanism is introduced that theoretically guarantees preservation of pre-stage performance with bounded parameter lag and function consistency.

Result: Achieves 80.1 AP50 on FLIR (robustness benchmark), 48.6 AP50_Novel on OV-COCO (open-world generalization), and 35.7 mAP_r on OV-LVIS (open-world generalization). Demonstrates competitive performance across both robustness and diversity evaluations.

Conclusion: C3-OWD successfully unifies robustness and generalization in object detection through a curriculum learning approach with EMA mechanism, addressing the trade-off between these two important capabilities and achieving state-of-the-art performance on both robustness and open-world benchmarks.

Abstract: Object detection has advanced significantly in the closed-set setting, but real-world deployment remains limited by two challenges: poor generalization to unseen categories and insufficient robustness under adverse conditions. Prior research has explored these issues separately: visible-infrared detection improves robustness but lacks generalization, while open-world detection leverages vision-language alignment strategy for category diversity but struggles under extreme environments. This trade-off leaves robustness and diversity difficult to achieve simultaneously. To mitigate these issues, we propose \textbf{C3-OWD}, a curriculum cross-modal contrastive learning framework that unifies both strengths. Stage1 enhances robustness by pretraining with RGBT data, while Stage2 improves generalization via vision-language alignment. To prevent catastrophic forgetting between two stages, we introduce an Exponential Moving Average (EMA) mechanism that theoretically guarantees preservation of pre-stage performance with bounded parameter lag and function consistency. Experiments on FLIR, OV-COCO, and OV-LVIS demonstrate the effectiveness of our approach: C3-OWD achieves $80.1$ AP$^{50}$ on FLIR, $48.6$ AP$^{50}_{\text{Novel}}$ on OV-COCO, and $35.7$ mAP$_r$ on OV-LVIS, establishing competitive performance across both robustness and diversity evaluations. Code available at: https://github.com/justin-herry/C3-OWD.git.

Method: Four key components: (1) Large-scale curated datasets (85M pretraining + 22M instruction datasets), (2) Efficient training framework with offline parallel data packing for $16K budget training, (3) State-of-the-art model architecture, and (4) RL-based post-training to enhance reasoning capabilities.

Result: LLaVA-OneVision-1.5-8B outperforms Qwen2.5-VL-7B on 18 of 27 benchmarks, and LLaVA-OneVision-1.5-4B surpasses Qwen2.5-VL-3B on all 27 benchmarks, demonstrating exceptional competitive performance across diverse multimodal tasks.

Conclusion: The paper presents a highly efficient and cost-effective approach to building state-of-the-art multimodal models, offering an open framework that achieves superior performance with significantly reduced resources compared to existing solutions.

Abstract: We present LLaVA-OneVision-1.5, a novel family of Large Multimodal Models (LMMs) that achieve state-of-the-art performance with significantly reduced computational and financial costs. Different from the existing works, LLaVA-OneVision-1.5 provides an open, efficient, and reproducible framework for building high-quality vision-language models entirely from scratch. The LLaVA-OneVision-1.5 release comprises three primary components: (1) Large-Scale Curated Datasets: We construct an 85M concept-balanced pretraining dataset LLaVA-OneVision-1.5-Mid-Traning and a meticulously curated 22M instruction dataset LLaVA-OneVision-1.5-Instruct. (2) Efficient Training Framework: We develop a complete end-to-end efficient training framework leveraging an offline parallel data packing strategy to facilitate the training of LLaVA-OneVision-1.5 within a $16,000 budget. (3) State-of-the-art Performance: Experimental results demonstrate that LLaVA-OneVision-1.5 yields exceptionally competitive performance across a broad range of downstream tasks. Specifically, LLaVA-OneVision-1.5-8B outperforms Qwen2.5-VL-7B on 18 of 27 benchmarks, and LLaVA-OneVision-1.5-4B surpasses Qwen2.5-VL-3B on all 27 benchmarks. (4) RL-based Post-training: We unlock the model’s latent potential through a lightweight RL stage, effectively eliciting robust chain-of-thought reasoning to significantly boost performance on complex multimodal reasoning tasks.

Method: PD-Diag-Net integrates MRI preprocessing to reduce variability, then incorporates two clinical priors: Brain-Region-Relevance-Prior (PD-associated regions) and Brain-Region-Aging-Prior (accelerated aging in PD regions). Uses Relevance-Prior Guided Feature Aggregation Module and Age-Prior Guided Diagnosis Module for inter-subject and intra-subject analysis.

Result: Achieves 86% accuracy on external hospital test data and over 96% accuracy for early-stage diagnosis, outperforming existing advanced methods by more than 20%.

Conclusion: PD-Diag-Net provides an effective automated diagnostic solution for Parkinson’s disease with high accuracy, clinical interpretability, and strong performance in early-stage detection, addressing current diagnostic limitations.

Abstract: Parkinson’s disease (PD) is a common neurodegenerative disorder that severely diminishes patients’ quality of life. Its global prevalence has increased markedly in recent decades. Current diagnostic workflows are complex and heavily reliant on neurologists’ expertise, often resulting in delays in early detection and missed opportunities for timely intervention. To address these issues, we propose an end-to-end automated diagnostic method for PD, termed PD-Diag-Net, which performs risk assessment and auxiliary diagnosis directly from raw MRI scans. This framework first introduces an MRI Pre-processing Module (MRI-Processor) to mitigate inter-subject and inter-scanner variability by flexibly integrating established medical imaging preprocessing tools. It then incorporates two forms of clinical prior knowledge: (1) Brain-Region-Relevance-Prior (Relevance-Prior), which specifies brain regions strongly associated with PD; and (2) Brain-Region-Aging-Prior (Aging-Prior), which reflects the accelerated aging typically observed in PD-associated regions. Building on these priors, we design two dedicated modules: the Relevance-Prior Guided Feature Aggregation Module (Aggregator), which guides the model to focus on PD-associated regions at the inter-subject level, and the Age-Prior Guided Diagnosis Module (Diagnoser), which leverages brain age gaps as auxiliary constraints at the intra-subject level to enhance diagnostic accuracy and clinical interpretability. Furthermore, we collected external test data from our collaborating hospital. Experimental results show that PD-Diag-Net achieves 86% accuracy on external tests and over 96% accuracy in early-stage diagnosis, outperforming existing advanced methods by more than 20%.

Method: Developed and deployed ML models trained on the largest known dataset of post-disaster drone imagery (21,716 building damage labels). Trained 91 disaster practitioners operationally. Deployed best-performing model during Hurricanes Debby and Helene responses.

Result: Successfully deployed operational system that assessed 415 buildings in approximately 18 minutes during real disaster responses. Established first state-of-practice for drone-based damage assessment, moving beyond academic research to actual field deployment.

Conclusion: This work demonstrates the first operational AI/ML system for drone-based building damage assessment, providing practical documentation and lessons learned for both research and user communities, bridging the gap between academic research and real-world disaster response needs.

Abstract: This paper presents the first AI/ML system for automating building damage assessment in uncrewed aerial systems (sUAS) imagery to be deployed operationally during federally declared disasters (Hurricanes Debby and Helene). In response to major disasters, sUAS teams are dispatched to collect imagery of the affected areas to assess damage; however, at recent disasters, teams collectively delivered between 47GB and 369GB of imagery per day, representing more imagery than can reasonably be transmitted or interpreted by subject matter experts in the disaster scene, thus delaying response efforts. To alleviate this data avalanche encountered in practice, computer vision and machine learning techniques are necessary. While prior work has been deployed to automatically assess damage in satellite imagery, there is no current state of practice for sUAS-based damage assessment systems, as all known work has been confined to academic settings. This work establishes the state of practice via the development and deployment of models for building damage assessment with sUAS imagery. The model development involved training on the largest known dataset of post-disaster sUAS aerial imagery, containing 21,716 building damage labels, and the operational training of 91 disaster practitioners. The best performing model was deployed during the responses to Hurricanes Debby and Helene, where it assessed a combined 415 buildings in approximately 18 minutes. This work contributes documentation of the actual use of AI/ML for damage assessment during a disaster and lessons learned to the benefit of the AI/ML research and user communities.

Method: 1) Created FUSAR-GEOVL-1M dataset with 120k SAR images and geographic projection attributes; 2) Generated aligned structured text using hierarchical cognitive thought chains encoding multidimensional semantic information; 3) Designed self-consistent iterative optimization with contrast, matching, and reconstruction in a self-supervised closed loop; 4) Established unified evaluation benchmark across 11 downstream tasks.

Result: Developed the first knowledge-guided general multimodal foundational model for SAR imagery with reusable data and evaluation baselines, covering multiple satellite platforms and 135 cities, with over 1 million multidimensional semantic information encoded.

Conclusion: FUSAR-KLIP successfully addresses the cognitive inconsistency between general vision models and SAR interpretation by incorporating geoscientific knowledge through structured data generation and self-supervised learning, establishing a new foundation for SAR multimodal AI.

Abstract: Cross-modal artificial intelligence, represented by visual language models, has achieved significant success in general image understanding. However, a fundamental cognitive inconsistency exists between general visual representation and remote sensing image interpretation: remote sensing images couple topography, terrain, and spatial structure, thereby inherently requiring models to possess deep geoscientific understanding. This cognitive difference is further amplified in synthetic aperture radar (SAR) imagery: while SAR possesses irreplaceable all-weather, all-day observation capabilities, it is constrained by coherent imaging mechanisms, exhibiting significant modal heterogeneity with general images. To address this inconsistency, we propose FUSAR-KLIP, the first knowledge-guided general multimodal foundational model for SAR, along with reusable data and evaluation baselines. Specifically: (1) FUSAR-GEOVL-1M (the first large-scale SAR dataset with complete geographic projection attributes) was constructed, covering multiple satellite platforms, 120,000 images, and 135 cities; (2) Aligned structured text was generated through hierarchical cognitive thought chains, accurately encoding more than 1 million multidimensional semantic information from geomorphological environment and regional attributes to spatial relationships; (3) A self-consistent iterative optimization mechanism was designed to guide cross-modal learning with this knowledge information consistent with human cognition and physical laws in a self-supervised closed loop consisting of contrast, matching, and reconstruction; (4) A unified evaluation benchmark was established in 11 typical downstream tasks in the two major categories of vision and language, and compared with 15 mainstream foundation models.

Method: Proposes a new self-supervised strategy combining equivariant reconstruction networks with splitting losses. Introduces a novel definition of equivariance for reconstruction networks and shows this combination yields unbiased supervised loss estimates.

Result: Achieves state-of-the-art performance on image inpainting, accelerated MRI, sparse-view CT, and compressive sensing, particularly with highly rank-deficient forward models.

Conclusion: The proposed equivariant splitting loss enables effective self-supervised learning for inverse problems with incomplete measurements, outperforming existing methods in challenging settings.

Abstract: Self-supervised learning for inverse problems allows to train a reconstruction network from noise and/or incomplete data alone. These methods have the potential of enabling learning-based solutions when obtaining ground-truth references for training is expensive or even impossible. In this paper, we propose a new self-supervised learning strategy devised for the challenging setting where measurements are observed via a single incomplete observation model. We introduce a new definition of equivariance in the context of reconstruction networks, and show that the combination of self-supervised splitting losses and equivariant reconstruction networks results in unbiased estimates of the supervised loss. Through a series of experiments on image inpainting, accelerated magnetic resonance imaging, sparse-view computed tomography, and compressive sensing, we demonstrate that the proposed loss achieves state-of-the-art performance in settings with highly rank-deficient forward models. The code is available at https://github.com/vsechaud/Equivariant-Splitting

Method: DiffRegCD reformulates correspondence estimation as Gaussian smoothed classification for sub-pixel accuracy, leverages frozen multi-scale features from pretrained denoising diffusion models for robustness, and uses controlled affine perturbations on standard CD datasets for supervision without pseudo labels.

Result: Extensive experiments on aerial (LEVIR-CD, DSIFN-CD, WHU-CD, SYSU-CD) and ground-level (VL-CMU-CD) datasets show DiffRegCD consistently surpasses recent baselines and remains reliable under wide temporal and geometric variation.

Conclusion: DiffRegCD establishes diffusion features and classification-based correspondence as a strong foundation for unified change detection, effectively handling misalignment challenges in real-world applications.

Abstract: Change detection (CD) is fundamental to computer vision and remote sensing, supporting applications in environmental monitoring, disaster response, and urban development. Most CD models assume co-registered inputs, yet real-world imagery often exhibits parallax, viewpoint shifts, and long temporal gaps that cause severe misalignment. Traditional two stage methods that first register and then detect, as well as recent joint frameworks (e.g., BiFA, ChangeRD), still struggle under large displacements, relying on regression only flow, global homographies, or synthetic perturbations. We present DiffRegCD, an integrated framework that unifies dense registration and change detection in a single model. DiffRegCD reformulates correspondence estimation as a Gaussian smoothed classification task, achieving sub-pixel accuracy and stable training. It leverages frozen multi-scale features from a pretrained denoising diffusion model, ensuring robustness to illumination and viewpoint variation. Supervision is provided through controlled affine perturbations applied to standard CD datasets, yielding paired ground truth for both flow and change detection without pseudo labels. Extensive experiments on aerial (LEVIR-CD, DSIFN-CD, WHU-CD, SYSU-CD) and ground level (VL-CMU-CD) datasets show that DiffRegCD consistently surpasses recent baselines and remains reliable under wide temporal and geometric variation, establishing diffusion features and classification based correspondence as a strong foundation for unified change detection.

Method: CRL uses LLM to generate descriptive texts for user criteria to construct semantic basis, then projects image representations into this conditional feature space using VLM.

Result: Extensive experiments on classification and retrieval tasks demonstrate superiority and generality of CRL for customized tasks.

Conclusion: CRL enables extraction of task-specific representations without costly fine-tuning, with semantic basis construction via LLM and VLM projection being effective for customized criteria.

Abstract: Conventional representation learning methods learn a universal representation that primarily captures dominant semantics, which may not always align with customized downstream tasks. For instance, in animal habitat analysis, researchers prioritize scene-related features, whereas universal embeddings emphasize categorical semantics, leading to suboptimal results. As a solution, existing approaches resort to supervised fine-tuning, which however incurs high computational and annotation costs. In this paper, we propose Conditional Representation Learning (CRL), aiming to extract representations tailored to arbitrary user-specified criteria. Specifically, we reveal that the semantics of a space are determined by its basis, thereby enabling a set of descriptive words to approximate the basis for a customized feature space. Building upon this insight, given a user-specified criterion, CRL first employs a large language model (LLM) to generate descriptive texts to construct the semantic basis, then projects the image representation into this conditional feature space leveraging a vision-language model (VLM). The conditional representation better captures semantics for the specific criterion, which could be utilized for multiple customized tasks. Extensive experiments on classification and retrieval tasks demonstrate the superiority and generality of the proposed CRL. The code is available at https://github.com/XLearning-SCU/2025-NeurIPS-CRL.

Method: TC-LoRA uses a hypernetwork to generate LoRA (Low-Rank Adaptation) adapters on-the-fly for each diffusion step. These adapters conditionally modify the frozen backbone model’s weights based on both time (denoising step) and user-provided conditions, enabling dynamic, context-aware control throughout the generation process.

Result: Experiments across various data domains show that TC-LoRA’s dynamic, parametric control significantly enhances generative fidelity and adherence to spatial conditions compared to traditional static, activation-based conditioning methods.

Conclusion: TC-LoRA establishes a new paradigm where conditioning strategy is modified through deeper functional adaptation of model weights, allowing control to align with the dynamic demands of both the task and generative stage, representing a shift from static activation-based conditioning to adaptive parametric control.

Abstract: Current controllable diffusion models typically rely on fixed architectures that modify intermediate activations to inject guidance conditioned on a new modality. This approach uses a static conditioning strategy for a dynamic, multi-stage denoising process, limiting the model’s ability to adapt its response as the generation evolves from coarse structure to fine detail. We introduce TC-LoRA (Temporally Modulated Conditional LoRA), a new paradigm that enables dynamic, context-aware control by conditioning the model’s weights directly. Our framework uses a hypernetwork to generate LoRA adapters on-the-fly, tailoring weight modifications for the frozen backbone at each diffusion step based on time and the user’s condition. This mechanism enables the model to learn and execute an explicit, adaptive strategy for applying conditional guidance throughout the entire generation process. Through experiments on various data domains, we demonstrate that this dynamic, parametric control significantly enhances generative fidelity and adherence to spatial conditions compared to static, activation-based methods. TC-LoRA establishes an alternative approach in which the model’s conditioning strategy is modified through a deeper functional adaptation of its weights, allowing control to align with the dynamic demands of the task and generative stage.

Method: Product images are mapped to UV space and converted to compact palette/control-point parameterization with ICC locking for printability. A physically based human-garment pipeline simulates motion, camera viewpoints, cloth dynamics, and illumination. Expectation-over-transformation objective with temporal weighting optimizes control points to minimize detection confidence across sequences.

Result: Extensive experiments show strong and stable concealment, high robustness to viewpoint changes, and superior cross-model transferability. Physical garments produced with sublimation printing achieve reliable suppression under indoor and outdoor recordings.

Conclusion: The sequence-level optimization framework generates natural, printable adversarial textures that remain effective throughout entire walking videos, confirming real-world feasibility for wearable adversarial attacks.

Abstract: Deep neural networks used for human detection are highly vulnerable to adversarial manipulation, creating safety and privacy risks in real surveillance environments. Wearable attacks offer a realistic threat model, yet existing approaches usually optimize textures frame by frame and therefore fail to maintain concealment across long video sequences with motion, pose changes, and garment deformation. In this work, a sequence-level optimization framework is introduced to generate natural, printable adversarial textures for shirts, trousers, and hats that remain effective throughout entire walking videos in both digital and physical settings. Product images are first mapped to UV space and converted into a compact palette and control-point parameterization, with ICC locking to keep all colors printable. A physically based human-garment pipeline is then employed to simulate motion, multi-angle camera viewpoints, cloth dynamics, and illumination variation. An expectation-over-transformation objective with temporal weighting is used to optimize the control points so that detection confidence is minimized across whole sequences. Extensive experiments demonstrate strong and stable concealment, high robustness to viewpoint changes, and superior cross-model transferability. Physical garments produced with sublimation printing achieve reliable suppression under indoor and outdoor recordings, confirming real-world feasibility.

Method: Three-stage physics-informed method: 1) parameterized template creation, 2) laws of illumination-aware 2D synthesis, and 3) physical engine-based 3D rendering, resulting in mixed 2D/3D FlareX dataset with 9,500 2D templates and 3,000 3D-rendered pairs.

Result: Creates comprehensive FlareX dataset with 9,500 2D templates from 95 flare patterns and 3,000 flare image pairs from 60 3D scenes, plus masking approach for real-world flare-free images.

Conclusion: The physics-informed FlareX dataset generation method effectively addresses limitations of existing synthetic flare datasets, enabling better model generalization to real-world scenarios through extensive experiments.

Abstract: Lens flare occurs when shooting towards strong light sources, significantly degrading the visual quality of images. Due to the difficulty in capturing flare-corrupted and flare-free image pairs in the real world, existing datasets are typically synthesized in 2D by overlaying artificial flare templates onto background images. However, the lack of flare diversity in templates and the neglect of physical principles in the synthesis process hinder models trained on these datasets from generalizing well to real-world scenarios. To address these challenges, we propose a new physics-informed method for flare data generation, which consists of three stages: parameterized template creation, the laws of illumination-aware 2D synthesis, and physical engine-based 3D rendering, which finally gives us a mixed flare dataset that incorporates both 2D and 3D perspectives, namely FlareX. This dataset offers 9,500 2D templates derived from 95 flare patterns and 3,000 flare image pairs rendered from 60 3D scenes. Furthermore, we design a masking approach to obtain real-world flare-free images from their corrupted counterparts to measure the performance of the model on real-world images. Extensive experiments demonstrate the effectiveness of our method and dataset.

Method: Visual Consistency Learning (ViCO) uses multiple MLP connectors with different compression ratios to downsample vision tokens based on semantic complexity. Training minimizes KL divergence between responses from different connectors. Inference uses Visual Resolution Router (ViR) to automatically select optimal compression rate per image patch.

Result: Method reduces vision tokens by up to 50% while maintaining perception, reasoning, and OCR capabilities. Outperforms existing dynamic high-resolution strategies that adjust tokens based on resolution rather than semantic complexity.

Conclusion: ViCO enables more efficient MLLMs by dynamically adapting vision tokens to semantic complexity rather than resolution, significantly reducing inference costs without sacrificing performance. Code and models will be released to support future research.

Abstract: Existing Multimodal Large Language Models (MLLMs) suffer from increased inference costs due to the additional vision tokens introduced by image inputs. In this work, we propose Visual Consistency Learning (ViCO), a novel training algorithm that enables the model to represent images of varying semantic complexities using different numbers of vision tokens. The key idea behind our method is to employ multiple MLP connectors, each with a different image compression ratio, to downsample the vision tokens based on the semantic complexity of the image. During training, we minimize the KL divergence between the responses conditioned on different MLP connectors. At inference time, we introduce an image router, termed Visual Resolution Router (ViR), that automatically selects the appropriate compression rate for each image patch. Compared with existing dynamic high-resolution strategies, which adjust the number of visual tokens based on image resolutions, our method dynamically adapts the number of visual tokens according to semantic complexity. Experimental results demonstrate that our method can reduce the number of vision tokens by up to 50% while maintaining the model’s perception, reasoning, and OCR capabilities. We hope this work will contribute to the development of more efficient MLLMs. The code and models will be released to facilitate future research.

Method: The authors focus on counting hallucinations (incorrect number of instances/objects), construct CountHalluSet dataset suite with well-defined counting criteria (ToyShape, SimObject, RealHand), develop standardized evaluation protocol, and systematically examine how different sampling conditions affect counting hallucination levels.

Result: The study reveals that common evaluation metrics like FID fail to capture counting hallucinations consistently, providing new insights into hallucination phenomena in image generation.

Conclusion: This work takes the first step toward systematically quantifying hallucinations in diffusion models, offering a standardized evaluation protocol and dataset for studying counting hallucinations, which could guide the design of next-generation generative models with better factual consistency.

Abstract: Diffusion probabilistic models (DPMs) have demonstrated remarkable progress in generative tasks, such as image and video synthesis. However, they still often produce hallucinated samples (hallucinations) that conflict with real-world knowledge, such as generating an implausible duplicate cup floating beside another cup. Despite their prevalence, the lack of feasible methodologies for systematically quantifying such hallucinations hinders progress in addressing this challenge and obscures potential pathways for designing next-generation generative models under factual constraints. In this work, we bridge this gap by focusing on a specific form of hallucination, which we term counting hallucination, referring to the generation of an incorrect number of instances or structured objects, such as a hand image with six fingers, despite such patterns being absent from the training data. To this end, we construct a dataset suite CountHalluSet, with well-defined counting criteria, comprising ToyShape, SimObject, and RealHand. Using these datasets, we develop a standardized evaluation protocol for quantifying counting hallucinations, and systematically examine how different sampling conditions in DPMs, including solver type, ODE solver order, sampling steps, and initial noise, affect counting hallucination levels. Furthermore, we analyze their correlation with common evaluation metrics such as FID, revealing that this widely used image quality metric fails to capture counting hallucinations consistently. This work aims to take the first step toward systematically quantifying hallucinations in diffusion models and offer new insights into the investigation of hallucination phenomena in image generation.

Method: Uses dual spatio-temporal conditioning: temporally adjacent frames for motion continuity and spatially adjacent content for scene consistency. Introduces a 3D scene memory representing static geometry extracted from input video via dynamic SLAM with dynamic masking. Projects static scene representation to target viewpoints as 3D spatial prompts.

Result: Significantly outperforms existing methods in scene consistency, camera controllability, and generation quality. Maintains long-range spatial coherence and precise camera control without sacrificing computational efficiency or motion realism.

Conclusion: 3DScenePrompt enables video generation with precise camera control and scene consistency by separating static geometry from dynamic elements, allowing natural evolution of motion while maintaining spatial coherence across long sequences.

Abstract: We present 3DScenePrompt, a framework that generates the next video chunk from arbitrary-length input while enabling precise camera control and preserving scene consistency. Unlike methods conditioned on a single image or a short clip, we employ dual spatio-temporal conditioning that reformulates context-view referencing across the input video. Our approach conditions on both temporally adjacent frames for motion continuity and spatially adjacent content for scene consistency. However, when generating beyond temporal boundaries, directly using spatially adjacent frames would incorrectly preserve dynamic elements from the past. We address this by introducing a 3D scene memory that represents exclusively the static geometry extracted from the entire input video. To construct this memory, we leverage dynamic SLAM with our newly introduced dynamic masking strategy that explicitly separates static scene geometry from moving elements. The static scene representation can then be projected to any target viewpoint, providing geometrically consistent warped views that serve as strong 3D spatial prompts while allowing dynamic regions to evolve naturally from temporal context. This enables our model to maintain long-range spatial coherence and precise camera control without sacrificing computational efficiency or motion realism. Extensive experiments demonstrate that our framework significantly outperforms existing methods in scene consistency, camera controllability, and generation quality. Project page : https://cvlab-kaist.github.io/3DScenePrompt/

Method: Two components: (1) Object-Centric Semantic Grounding adds object-level annotations to 3DGS; (2) Physics-Aware Execution Jointing embeds collision objects and constructs physical interfaces.

Result: Created InteriorGS (1K object-annotated 3DGS indoor scenes) and SAGE-Bench (first 3DGS-based VLN benchmark with 2M data). 3DGS data is harder to converge but shows strong generalizability, improving baseline performance by 31% on VLN-CE Unseen task.

Conclusion: SAGE-3D successfully upgrades 3DGS into semantically and physically aligned executable environments for VLN, demonstrating improved performance and generalizability while providing new datasets and benchmarks.

Abstract: 3D Gaussian Splatting (3DGS), a 3D representation method with photorealistic real-time rendering capabilities, is regarded as an effective tool for narrowing the sim-to-real gap. However, it lacks fine-grained semantics and physical executability for Visual-Language Navigation (VLN). To address this, we propose SAGE-3D (Semantically and Physically Aligned Gaussian Environments for 3D Navigation), a new paradigm that upgrades 3DGS into an executable, semantically and physically aligned environment. It comprises two components: (1) Object-Centric Semantic Grounding, which adds object-level fine-grained annotations to 3DGS; and (2) Physics-Aware Execution Jointing, which embeds collision objects into 3DGS and constructs rich physical interfaces. We release InteriorGS, containing 1K object-annotated 3DGS indoor scene data, and introduce SAGE-Bench, the first 3DGS-based VLN benchmark with 2M VLN data. Experiments show that 3DGS scene data is more difficult to converge, while exhibiting strong generalizability, improving baseline performance by 31% on the VLN-CE Unseen task. Our data and code are available at: https://sage-3d.github.io.

Method: The paper presents two main contributions: 1) i-CIR dataset with instance-level class definition and hard negatives selected via semi-automated process, and 2) BASIC method that separately estimates query-image-to-image and query-text-to-image similarities using pre-trained VLMs, then performs late fusion with components to improve individual similarities.

Result: BASIC achieves new state-of-the-art performance on both the new i-CIR dataset and existing semantic-level CIR datasets, demonstrating effectiveness of the training-free approach.

Conclusion: The i-CIR dataset addresses the data scarcity problem in CIR research, while the BASIC method shows that training-free approaches using pre-trained VLMs can achieve superior performance without requiring large-scale training data.

Abstract: The progress of composed image retrieval (CIR), a popular research direction in image retrieval, where a combined visual and textual query is used, is held back by the absence of high-quality training and evaluation data. We introduce a new evaluation dataset, i-CIR, which, unlike existing datasets, focuses on an instance-level class definition. The goal is to retrieve images that contain the same particular object as the visual query, presented under a variety of modifications defined by textual queries. Its design and curation process keep the dataset compact to facilitate future research, while maintaining its challenge-comparable to retrieval among more than 40M random distractors-through a semi-automated selection of hard negatives. To overcome the challenge of obtaining clean, diverse, and suitable training data, we leverage pre-trained vision-and-language models (VLMs) in a training-free approach called BASIC. The method separately estimates query-image-to-image and query-text-to-image similarities, performing late fusion to upweight images that satisfy both queries, while down-weighting those that exhibit high similarity with only one of the two. Each individual similarity is further improved by a set of components that are simple and intuitive. BASIC sets a new state of the art on i-CIR but also on existing CIR datasets that follow a semantic-level class definition. Project page: https://vrg.fel.cvut.cz/icir/.

Method: BlurDM uses a dual-diffusion forward scheme that diffuses both noise and blur onto sharp images, with a reverse process that performs simultaneous denoising and deblurring. It operates in latent space for efficiency.

Result: BlurDM significantly and consistently enhances existing deblurring methods across four benchmark datasets.

Conclusion: BlurDM effectively integrates blur formation into diffusion models, providing a flexible prior generation network that substantially improves deblurring performance.

Abstract: Diffusion models show promise for dynamic scene deblurring; however, existing studies often fail to leverage the intrinsic nature of the blurring process within diffusion models, limiting their full potential. To address it, we present a Blur Diffusion Model (BlurDM), which seamlessly integrates the blur formation process into diffusion for image deblurring. Observing that motion blur stems from continuous exposure, BlurDM implicitly models the blur formation process through a dual-diffusion forward scheme, diffusing both noise and blur onto a sharp image. During the reverse generation process, we derive a dual denoising and deblurring formulation, enabling BlurDM to recover the sharp image by simultaneously denoising and deblurring, given pure Gaussian noise conditioned on the blurred image as input. Additionally, to efficiently integrate BlurDM into deblurring networks, we perform BlurDM in the latent space, forming a flexible prior generation network for deblurring. Extensive experiments demonstrate that BlurDM significantly and consistently enhances existing deblurring methods on four benchmark datasets. The project page is available at https://jin-ting-he.github.io/Blur-Diffusion-Model/.

Method: Proposes RegionRAG framework with two key components: 1) Training: hybrid supervision strategy using both labeled and unlabeled data to pinpoint relevant patches; 2) Inference: dynamic pipeline that intelligently groups salient patches into complete semantic regions, shifting retrieval from document-level to region-level.

Result: Experiments on six benchmarks show state-of-the-art performance: improves retrieval accuracy by 10.02% in R@1 on average, boosts question answering accuracy by 3.56% while using only 71.42% visual tokens compared to prior methods.

Conclusion: RegionRAG successfully addresses redundancy in multi-modal RAG by focusing on query-relevant regions rather than entire documents, enabling generators to concentrate on concise visual content and achieving better efficiency and accuracy.

Abstract: Multi-modal Retrieval-Augmented Generation (RAG) has become a critical method for empowering LLMs by leveraging candidate visual documents. However, current methods consider the entire document as the basic retrieval unit, introducing substantial irrelevant visual content in two ways: 1) Relevant documents often contain large regions unrelated to the query, diluting the focus on salient information; 2) Retrieving multiple documents to increase recall further introduces redundant and irrelevant documents. These redundant contexts distract the model’s attention and further degrade the performance. To address this challenge, we propose RegionRAG, a novel framework that shifts the retrieval paradigm from the document level to the region level. During training, we design a hybrid supervision strategy from both labeled data and unlabeled data to pinpoint relevant patches. During inference, we propose a dynamic pipeline that intelligently groups salient patches into complete semantic regions. By delegating the task of identifying relevant regions to the retriever, RegionRAG enables the generator to focus solely on concise, query-relevant visual content, improving both efficiency and accuracy. Experiments on six benchmarks demonstrate that RegionRAG achieves state-of-the-art performance. It improves retrieval accuracy by 10.02% in R@1 on average, and boosts question answering accuracy by 3.56% while using only 71.42% visual tokens compared with prior methods.

Method: Three key components: (1) hierarchical identity-preserving attention mechanism (intra-subject, inter-subject, cross-modal), (2) semantic understanding module using pretrained Vision-Language Model for fine-grained guidance, and (3) online reinforcement learning phase to refine critical concepts. Also created a new dataset for training/evaluation.

Result: Extensive experiments show ID-Crafter establishes new state-of-the-art performance on multi-subject video generation benchmarks, excelling in identity preservation, temporal consistency, and overall video quality.

Conclusion: ID-Crafter effectively addresses the limitations of current multi-subject video generation by providing superior identity preservation and semantic coherence through its innovative architecture and training approach.

Abstract: Significant progress has been achieved in high-fidelity video synthesis, yet current paradigms often fall short in effectively integrating identity information from multiple subjects. This leads to semantic conflicts and suboptimal performance in preserving identities and interactions, limiting controllability and applicability. To tackle this issue, we introduce ID-Crafter, a framework for multi-subject video generation that achieves superior identity preservation and semantic coherence. ID-Crafter integrates three key components: (i) a hierarchical identity-preserving attention mechanism that progressively aggregates features at intra-subject, inter-subject, and cross-modal levels; (ii) a semantic understanding module powered by a pretrained Vision-Language Model (VLM) to provide fine-grained guidance and capture complex inter-subject relationships; and (iii) an online reinforcement learning phase to further refine the model for critical concepts. Furthermore, we construct a new dataset to facilitate robust training and evaluation. Extensive experiments demonstrate that ID-Crafter establishes new state-of-the-art performance on multi-subject video generation benchmarks, excelling in identity preservation, temporal consistency, and overall video quality. Project page: https://angericky.github.io/ID-Crafter

Method: Proposes WDT-MD with three key innovations: 1) Noise-encoded image conditioning to prevent identity mapping by perturbing conditions during training, 2) Pseudo-normal pattern synthesis via inpainting for pixel-level supervision to distinguish MAs from other anomalies, 3) Wavelet diffusion Transformer architecture combining global modeling of diffusion Transformers with multi-scale wavelet analysis for better normal feature reconstruction.

Result: Comprehensive experiments on IDRiD and e-ophtha MA datasets show WDT-MD outperforms state-of-the-art methods in both pixel-level and image-level MA detection.

Conclusion: WDT-MD advances automated MA detection by addressing fundamental limitations of diffusion-based approaches, showing significant promise for improving early Diabetic Retinopathy screening.

Abstract: Microaneurysms (MAs), the earliest pathognomonic signs of Diabetic Retinopathy (DR), present as sub-60 $μm$ lesions in fundus images with highly variable photometric and morphological characteristics, rendering manual screening not only labor-intensive but inherently error-prone. While diffusion-based anomaly detection has emerged as a promising approach for automated MA screening, its clinical application is hindered by three fundamental limitations. First, these models often fall prey to “identity mapping”, where they inadvertently replicate the input image. Second, they struggle to distinguish MAs from other anomalies, leading to high false positives. Third, their suboptimal reconstruction of normal features hampers overall performance. To address these challenges, we propose a Wavelet Diffusion Transformer framework for MA Detection (WDT-MD), which features three key innovations: a noise-encoded image conditioning mechanism to avoid “identity mapping” by perturbing image conditions during training; pseudo-normal pattern synthesis via inpainting to introduce pixel-level supervision, enabling discrimination between MAs and other anomalies; and a wavelet diffusion Transformer architecture that combines the global modeling capability of diffusion Transformers with multi-scale wavelet analysis to enhance reconstruction of normal retinal features. Comprehensive experiments on the IDRiD and e-ophtha MA datasets demonstrate that WDT-MD outperforms state-of-the-art methods in both pixel-level and image-level MA detection. This advancement holds significant promise for improving early DR screening.

Method: Derives uncertainty directly from explicit physical parameters of 3D Gaussian primitives (position, scale, rotation), propagates covariance through rendering Jacobian, then integrates semantic segmentation masks to produce targeted object-aware uncertainty scores for active view selection.

Result: Significantly improves efficiency of 3DGS reconstruction process and achieves higher quality for targeted objects compared to state-of-the-art methods, while also serving as robust uncertainty estimator for global scene.

Conclusion: OUGS provides principled, physically-grounded uncertainty formulation for 3DGS that enables effective object-aware active reconstruction, addressing limitations of scene-level uncertainty metrics and improving reconstruction of specific objects in complex scenes.

Abstract: Recent advances in 3D Gaussian Splatting (3DGS) have achieved state-of-the-art results for novel view synthesis. However, efficiently capturing high-fidelity reconstructions of specific objects within complex scenes remains a significant challenge. A key limitation of existing active reconstruction methods is their reliance on scene-level uncertainty metrics, which are often biased by irrelevant background clutter and lead to inefficient view selection for object-centric tasks. We present OUGS, a novel framework that addresses this challenge with a more principled, physically-grounded uncertainty formulation for 3DGS. Our core innovation is to derive uncertainty directly from the explicit physical parameters of the 3D Gaussian primitives (e.g., position, scale, rotation). By propagating the covariance of these parameters through the rendering Jacobian, we establish a highly interpretable uncertainty model. This foundation allows us to then seamlessly integrate semantic segmentation masks to produce a targeted, object-aware uncertainty score that effectively disentangles the object from its environment. This allows for a more effective active view selection strategy that prioritizes views critical to improving object fidelity. Experimental evaluations on public datasets demonstrate that our approach significantly improves the efficiency of the 3DGS reconstruction process and achieves higher quality for targeted objects compared to existing state-of-the-art methods, while also serving as a robust uncertainty estimator for the global scene.

Method: 1) RefBench-PRO benchmark decomposes referring expressions into perception (attribute, position) and reasoning (interaction, commonsense, relation, reject) dimensions with six progressively challenging tasks. 2) Automated data-generation pipeline produces diverse referring expressions across these sub-dimensions. 3) Ref-R1 learning scheme uses RL with Dynamic IoU-based GRPO to improve localization accuracy under complex reasoning conditions.

Result: RefBench-PRO enables interpretable evaluation of MLLMs on referring expression comprehension, presenting greater challenges in both perception and reasoning dimensions. The benchmark demonstrates effectiveness through extensive experiments.

Conclusion: The proposed RefBench-PRO benchmark addresses limitations of existing REC benchmarks by providing interpretable evaluation across cognitive abilities, while Ref-R1 establishes a stronger baseline for REC tasks with improved localization accuracy under complex reasoning conditions.

Abstract: Referring Expression Comprehension (REC) is a vision-language task that localizes a specific image region based on a textual description. Existing REC benchmarks primarily evaluate perceptual capabilities and lack interpretable scoring mechanisms, which cannot reveal the grounding capability of Multi-modal Large Language Model (MLLM) across different cognitive abilities. To address this limitation, we introduce RefBench-PRO, a comprehensive REC benchmark, which decomposes referring expressions into two core dimensions, i.e., perception and reasoning, and further subdivides them into six progressively challenging tasks, such as attribute, position, interaction, commonsense, relation and reject. We also develop a fully automated data-generation pipeline that produces diverse referring expressions across these six sub-dimensions. Furthermore, We propose Ref-R1, an RL-based learning scheme, which incorporates Dynamic IoU-based GRPO to improve localization accuracy under increasingly complex reasoning conditions, establishing a stronger baseline for REC. Extensive experiments demonstrate that our RefBench-PRO enables interpretable evaluation of MLLM on referring expression comprehension, presenting greater challenges in both perception and reasoning.

Method: CoordAR formulates 3D-3D correspondences as discrete token maps predicted autoregressively. It uses: 1) coordinate map tokenization for probabilistic prediction over discretized 3D space, 2) modality-decoupled encoding separating RGB appearance and coordinate cues, and 3) an autoregressive transformer decoder conditioned on query features and partially generated tokens.

Result: CoordAR significantly outperforms existing methods on multiple benchmarks and demonstrates strong robustness to symmetry, occlusion, and other real-world challenges.

Conclusion: The autoregressive probabilistic approach with discrete tokenization and modality-decoupled encoding provides an effective solution for one-reference 6D pose estimation, addressing limitations of existing methods in handling symmetry, occlusion, and global consistency.

Abstract: Object 6D pose estimation, a crucial task for robotics and augmented reality applications, becomes particularly challenging when dealing with novel objects whose 3D models are not readily available. To reduce dependency on 3D models, recent studies have explored one-reference-based pose estimation, which requires only a single reference view instead of a complete 3D model. However, existing methods that rely on real-valued coordinate regression suffer from limited global consistency due to the local nature of convolutional architectures and face challenges in symmetric or occluded scenarios owing to a lack of uncertainty modeling. We present CoordAR, a novel autoregressive framework for one-reference 6D pose estimation of unseen objects. CoordAR formulates 3D-3D correspondences between the reference and query views as a map of discrete tokens, which is obtained in an autoregressive and probabilistic manner. To enable accurate correspondence regression, CoordAR introduces 1) a novel coordinate map tokenization that enables probabilistic prediction over discretized 3D space; 2) a modality-decoupled encoding strategy that separately encodes RGB appearance and coordinate cues; and 3) an autoregressive transformer decoder conditioned on both position-aligned query features and the partially generated token sequence. With these novel mechanisms, CoordAR significantly outperforms existing methods on multiple benchmarks and demonstrates strong robustness to symmetry, occlusion, and other challenges in real-world tests.

Method: Proposes a new framework with hierarchical cross-modal interaction and dual-stream feature refinement that enhances joint embeddings with both class-level and instance-specific cues from textual descriptions and specific images.

Result: Experiments on MP-100 dataset show CapeNext consistently outperforms state-of-the-art CAPE methods by a large margin, regardless of the network backbone used.

Conclusion: The proposed approach successfully addresses the limitations of static joint embeddings in CAPE through better integration of textual and visual information, leading to significant performance improvements.

Abstract: Recent research in Category-Agnostic Pose Estimation (CAPE) has adopted fixed textual keypoint description as semantic prior for two-stage pose matching frameworks. While this paradigm enhances robustness and flexibility by disentangling the dependency of support images, our critical analysis reveals two inherent limitations of static joint embedding: (1) polysemy-induced cross-category ambiguity during the matching process(e.g., the concept “leg” exhibiting divergent visual manifestations across humans and furniture), and (2) insufficient discriminability for fine-grained intra-category variations (e.g., posture and fur discrepancies between a sleeping white cat and a standing black cat). To overcome these challenges, we propose a new framework that innovatively integrates hierarchical cross-modal interaction with dual-stream feature refinement, enhancing the joint embedding with both class-level and instance-specific cues from textual description and specific images. Experiments on the MP-100 dataset demonstrate that, regardless of the network backbone, CapeNext consistently outperforms state-of-the-art CAPE methods by a large margin.

Method: Proposes a Rotationally Invariant Features (RIF) framework with: 1) Point Coordinate Mapping (PCM) to map points into rotationally invariant space, 2) Lightweight Convolutional Transform Feature Network (CTF-Net) to extract robust features for memory bank, and 3) Transfer learning with 3D data augmentation to pre-train the feature extractor.

Result: Achieves state-of-the-art performance with 17.7% average P-AUROC improvement on Anomaly-ShapeNet dataset and 1.6% improvement on Real3D-AD dataset. Demonstrates strong generalization when combined with traditional feature extraction methods.

Conclusion: The RIF framework effectively addresses orientation/position variations in 3D point clouds, showing superior performance and generalization ability for industrial anomaly detection applications.

Abstract: 3D anomaly detection (AD) is a crucial task in computer vision, aiming to identify anomalous points or regions from point cloud data. However, existing methods may encounter challenges when handling point clouds with changes in orientation and position because the resulting features may vary significantly. To address this problem, we propose a novel Rotationally Invariant Features (RIF) framework for 3D AD. Firstly, to remove the adverse effect of variations on point cloud data, we develop a Point Coordinate Mapping (PCM) technique, which maps each point into a rotationally invariant space to maintain consistency of representation. Then, to learn robust and discriminative features, we design a lightweight Convolutional Transform Feature Network (CTF-Net) to extract rotationally invariant features for the memory bank. To improve the ability of the feature extractor, we introduce the idea of transfer learning to pre-train the feature extractor with 3D data augmentation. Experimental results show that the proposed method achieves the advanced performance on the Anomaly-ShapeNet dataset, with an average P-AUROC improvement of 17.7%, and also gains the best performance on the Real3D-AD dataset, with an average P-AUROC improvement of 1.6%. The strong generalization ability of RIF has been verified by combining it with traditional feature extraction methods on anomaly detection tasks, demonstrating great potential for industrial applications.

Method: Autoregressive denoising architecture with temporal causal attention to aggregate past observations; noise-augmented history memory to balance responsiveness with coherence; action-aware adapter for precise action control; mixture of action experts for heterogeneous action modalities.

Result: Astra achieves interactive, consistent, and general long-term video prediction supporting various interactions, with experiments showing improvements in fidelity, long-range prediction, and action alignment over SOTA world models.

Conclusion: Astra bridges the gap in general-purpose world modeling, enabling precise action-controlled future prediction across diverse real-world scenarios like autonomous driving and robot manipulation.

Abstract: Recent advances in diffusion transformers have empowered video generation models to generate high-quality video clips from texts or images. However, world models with the ability to predict long-horizon futures from past observations and actions remain underexplored, especially for general-purpose scenarios and various forms of actions. To bridge this gap, we introduce Astra, an interactive general world model that generates real-world futures for diverse scenarios (e.g., autonomous driving, robot grasping) with precise action interactions (e.g., camera motion, robot action). We propose an autoregressive denoising architecture and use temporal causal attention to aggregate past observations and support streaming outputs. We use a noise-augmented history memory to avoid over-reliance on past frames to balance responsiveness with temporal coherence. For precise action control, we introduce an action-aware adapter that directly injects action signals into the denoising process. We further develop a mixture of action experts that dynamically route heterogeneous action modalities, enhancing versatility across diverse real-world tasks such as exploration, manipulation, and camera control. Astra achieves interactive, consistent, and general long-term video prediction and supports various forms of interactions. Experiments across multiple datasets demonstrate the improvements of Astra in fidelity, long-range prediction, and action alignment over existing state-of-the-art world models.

Method: Proposes FireScope-Bench dataset (Sentinel-2 imagery + climate data + expert risk rasters) and FireScope framework using VLM-based reasoning-to-generation with reinforcement learning and visual supervision.

Result: FireScope achieves substantial performance gains when trained in USA and tested in Europe; reasoning traces are faithful and semantically meaningful according to expert feedback.

Conclusion: Reasoning can ground raster prediction models, improving generalization and interpretability; first framework to show language-based reasoning improves visual generation generalization and enables cross-continental wildfire risk modeling.

Abstract: Predicting wildfire risk is a reasoning-intensive spatial problem that requires the integration of visual, climatic, and geographic factors to infer continuous risk maps. Existing methods lack the causal reasoning and multimodal understanding required for reliable generalization. We introduce $\textbf{FireScope-Bench}$, a large-scale dataset and benchmark that couples Sentinel-2 imagery and climate data with expert-defined risk rasters across the USA, and real wildfire events in Europe for cross-continental evaluation. Building on this dataset, we propose $\textbf{FireScope}$, a VLM-based reasoning-to-generation framework that learns from both reinforcement learning and visual supervision to predict risk rasters with complementary reasoning traces. When trained in the USA and tested in Europe, $\textbf{FireScope}$ achieves substantial performance gains, while expert feedback and automated analysis confirm that its reasoning traces are faithful and semantically meaningful. Our findings demonstrate that reasoning can ground raster prediction models, improving both generalization and interpretability. To our knowledge, this is the first framework to (1) demonstrate that language-based reasoning can improve generalization in visual generation, (2) propose a high-resolution wildfire risk model that can be applied across continents, and (3) enable systematic studies of robust cross-continental generalization for multimodal fire risk models. We believe that $\textbf{FireScope-Bench}$ has the potential to serve as a foundation for advancing reasoning-driven, interpretable and generalizable spatial modeling. Data and source code will be made publicly available.

Method: The proposed Δ-LFM framework combines two key components: 1) Flow Matching (FM) to model disease dynamics as a velocity field, capturing continuous temporal evolution of patient data, and 2) Patient-specific latent alignment that enforces patient trajectories to lie along a specific axis with magnitude increasing monotonically with disease severity, creating a consistent and semantically meaningful latent space.

Result: The framework demonstrates strong empirical performance across three longitudinal MRI benchmarks and provides a new framework for interpreting and visualizing disease dynamics, offering more interpretable progression modeling compared to prior methods.

Conclusion: Δ-LFM successfully addresses key limitations in disease progression modeling by treating disease dynamics as a velocity field and enforcing semantic alignment in latent space, resulting in interpretable and clinically meaningful progression trajectories that correlate with disease severity indicators.

Abstract: Understanding disease progression is a central clinical challenge with direct implications for early diagnosis and personalized treatment. While recent generative approaches have attempted to model progression, key mismatches remain: disease dynamics are inherently continuous and monotonic, yet latent representations are often scattered, lacking semantic structure, and diffusion-based models disrupt continuity with random denoising process. In this work, we propose to treat the disease dynamic as a velocity field and leverage Flow Matching (FM) to align the temporal evolution of patient data. Unlike prior methods, it captures the intrinsic dynamic of disease, making the progression more interpretable. However, a key challenge remains: in latent space, Auto-Encoders (AEs) do not guarantee alignment across patients or correlation with clinical-severity indicators (e.g., age and disease conditions). To address this, we propose to learn patient-specific latent alignment, which enforces patient trajectories to lie along a specific axis, with magnitude increasing monotonically with disease severity. This leads to a consistent and semantically meaningful latent space. Together, we present $Δ$-LFM, a framework for modeling patient-specific latent progression with flow matching. Across three longitudinal MRI benchmarks, $Δ$-LFM demonstrates strong empirical performance and, more importantly, offers a new framework for interpreting and visualizing disease dynamics.

Method: ATAC operates in CLIP’s embedding space by calculating augmentation-induced drift vectors from clean image augmentations, inferring semantic recovery directions based on angular consistency of latent drifts, and correcting adversarial embeddings accordingly.

Result: ATAC consistently achieves remarkably high robustness across benchmarks, surpassing previous state-of-the-art methods by nearly 50% on average with minimal computational overhead. It maintains robustness in extreme settings and shows nontrivial robustness against adaptive attacks.

Conclusion: ATAC represents an efficient and novel paradigm for test-time adversarial defenses in CLIP’s embedding space, demonstrating that simple augmentation-based correction in latent space can provide strong robustness against adversarial attacks.

Abstract: Despite its remarkable success in zero-shot image-text matching, CLIP remains highly vulnerable to adversarial perturbations on images. As adversarial fine-tuning is prohibitively costly, recent works explore various test-time defense strategies; however, these approaches still exhibit limited robustness. In this work, we revisit this problem and propose a simple yet effective strategy: Augmentation-based Test-time Adversarial Correction (ATAC). Our method operates directly in the embedding space of CLIP, calculating augmentation-induced drift vectors to infer a semantic recovery direction and correcting the embedding based on the angular consistency of these latent drifts. Across a wide range of benchmarks, ATAC consistently achieves remarkably high robustness, surpassing that of previous state-of-the-art methods by nearly 50% on average, all while requiring minimal computational overhead. Furthermore, ATAC retains state-of-the-art robustness in unconventional and extreme settings and even achieves nontrivial robustness against adaptive attacks. Our results demonstrate that ATAC is an efficient method in a novel paradigm for test-time adversarial defenses in the embedding space of CLIP.

Method: Three key components: (1) enhancing expert models (observer model for recovery, teacher model for relabeling) for reliable statistics estimation and soft-label generation; (2) recalibrating BN statistics via full forward pass with dynamically adjusted momentum to reduce representation skew; (3) initializing synthetic images via multi-round mechanism selecting high-confidence diverse augmentations for coverage and diversity.

Result: Extensive experiments on four long-tailed benchmarks show consistent improvements over state-of-the-art methods. Notable improvements: 15.6% top-1 accuracy on CIFAR-100-LT and 11.8% on Tiny-ImageNet-LT under IPC=10 and IF=10.

Conclusion: The paper successfully addresses long-tailed dataset distillation by adopting a statistical alignment perspective to jointly mitigate model bias and restore fair supervision, achieving significant performance gains over existing methods.

Abstract: Dataset distillation creates a small distilled set that enables efficient training by capturing key information from the full dataset. While existing dataset distillation methods perform well on balanced datasets, they struggle under long-tailed distributions, where imbalanced class frequencies induce biased model representations and corrupt statistical estimates such as Batch Normalization (BN) statistics. In this paper, we rethink long-tailed dataset distillation by revisiting the limitations of trajectory-based methods, and instead adopt the statistical alignment perspective to jointly mitigate model bias and restore fair supervision. To this end, we introduce three dedicated components that enable unbiased recovery of distilled images and soft relabeling: (1) enhancing expert models (an observer model for recovery and a teacher model for relabeling) to enable reliable statistics estimation and soft-label generation; (2) recalibrating BN statistics via a full forward pass with dynamically adjusted momentum to reduce representation skew; (3) initializing synthetic images by incrementally selecting high-confidence and diverse augmentations via a multi-round mechanism that promotes coverage and diversity. Extensive experiments on four long-tailed benchmarks show consistent improvements over state-of-the-art methods across varying degrees of class imbalance. Notably, our approach improves top-1 accuracy by 15.6% on CIFAR-100-LT and 11.8% on Tiny-ImageNet-LT under IPC=10 and IF=10. Codes are available at https://github.com/2018cx/RLDD.

Method: Proposes T2LDM with Self-Conditioned Representation Guidance (SCRG) that provides soft supervision during training but decouples in inference. Also introduces T2nuScenes benchmark, directional position prior to reduce street distortion, and conditional encoder for multiple tasks.

Result: T2LDM outperforms existing methods in both unconditional and conditional generation, achieving state-of-the-art scene generation with improved detail and fidelity.

Conclusion: The proposed T2LDM framework effectively addresses text-to-LiDAR generation challenges through SCRG guidance, comprehensive benchmarking, and multi-task conditional generation capabilities.

Abstract: Text-to-LiDAR generation can customize 3D data with rich structures and diverse scenes for downstream tasks. However, the scarcity of Text-LiDAR pairs often causes insufficient training priors, generating overly smooth 3D scenes. Moreover, low-quality text descriptions may degrade generation quality and controllability. In this paper, we propose a Text-to-LiDAR Diffusion Model for scene generation, named T2LDM, with a Self-Conditioned Representation Guidance (SCRG). Specifically, SCRG, by aligning to the real representations, provides the soft supervision with reconstruction details for the Denoising Network (DN) in training, while decoupled in inference. In this way, T2LDM can perceive rich geometric structures from data distribution, generating detailed objects in scenes. Meanwhile, we construct a content-composable Text-LiDAR benchmark, T2nuScenes, along with a controllability metric. Based on this, we analyze the effects of different text prompts for LiDAR generation quality and controllability, providing practical prompt paradigms and insights. Furthermore, a directional position prior is designed to mitigate street distortion, further improving scene fidelity. Additionally, by learning a conditional encoder via frozen DN, T2LDM can support multiple conditional tasks, including Sparse-to-Dense, Dense-to-Sparse, and Semantic-to-LiDAR generation. Extensive experiments in unconditional and conditional generation demonstrate that T2LDM outperforms existing methods, achieving state-of-the-art scene generation.

Method: The approach uses 360 videos as supervision, resampling them into narrow field-of-view perspectives while computing ground truth directions by tracking points across full panoramas using a 2D pipeline. A baseline adapts CoTracker v3 to predict per-point rotations for direction updates.

Result: The paper introduces TAPVid360-10k dataset with 10k perspective videos and ground truth directional point tracking. The baseline method outperforms existing TAP and TAPVid 3D approaches.

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. Project page: https://finlay-hudson.github.io/tapvid360

Method: Introduces dots.ocr, a single vision-language model that jointly learns three core document parsing tasks in an end-to-end framework. Uses a highly scalable data engine to synthesize a vast multilingual corpus for training.

Result: Achieves state-of-the-art performance on comprehensive OmniDocBench benchmark. Introduces XDocParse benchmark (126 languages) where dots.ocr outperforms next-best competitor by +7.4 points, demonstrating unparalleled multilingual capabilities.

Conclusion: The unified end-to-end paradigm for document layout parsing is effective and scalable, enabling robust performance across diverse languages, layouts, and domains while overcoming limitations of fragmented multi-stage pipelines.

Abstract: Document Layout Parsing serves as a critical gateway for Artificial Intelligence (AI) to access and interpret the world’s vast stores of structured knowledge. This process,which encompasses layout detection, text recognition, and relational understanding, is particularly crucial for empowering next-generation Vision-Language Models. Current methods, however, rely on fragmented, multi-stage pipelines that suffer from error propagation and fail to leverage the synergies of joint training. In this paper, we introduce $\text{dots.ocr}$, a single Vision-Language Model that, for the first time, demonstrates the advantages of jointly learning three core tasks within a unified, end-to-end framework. This is made possible by a highly scalable data engine that synthesizes a vast multilingual corpus, empowering the model to deliver robust performance across a wide array of tasks, encompassing diverse languages, layouts, and domains. The efficacy of our unified paradigm is validated by state-of-the-art performance on the comprehensive OmniDocBench. Furthermore, to catalyze research in global document intelligence, we introduce XDocParse, a challenging new benchmark spanning 126 languages. On this testbed, $\text{dots.ocr}$ establishes a powerful new baseline, outperforming the next-best competitor by a remarkable +7.4 point margin and proving its unparalleled multilingual capabilities.

Method: Created a comprehensive benchmark dataset with over 440,000 facial trajectories (including 52,000+ longer than 10 frames and 68 manually reviewed full 150-frame trajectories). The pipeline captures both traditional facial landmarks and pore-scale keypoints trajectory.

Result: PoreTrack3D is the first benchmark dataset capturing both traditional facial landmarks and pore-scale keypoints trajectory. The authors systematically evaluated state-of-the-art dynamic 3D Gaussian splatting methods, establishing the first performance baseline in this domain.

Conclusion: The pipeline establishes a new framework for high-fidelity facial motion capture and dynamic 3D reconstruction. The publicly available dataset advances fine-grained facial expression analysis through pore-scale motion tracking.

Abstract: We introduce PoreTrack3D, the first benchmark for dynamic 3D Gaussian splatting in pore-scale, non-rigid 3D facial trajectory tracking. It contains over 440,000 facial trajectories in total, among which more than 52,000 are longer than 10 frames, including 68 manually reviewed trajectories that span the entire 150 frames. To the best of our knowledge, PoreTrack3D is the first benchmark dataset to capture both traditional facial landmarks and pore-scale keypoints trajectory, advancing the study of fine-grained facial expressions through the analysis of subtle skin-surface motion. We systematically evaluate state-of-the-art dynamic 3D Gaussian splatting methods on PoreTrack3D, establishing the first performance baseline in this domain. Overall, the pipeline developed for this benchmark dataset’s creation establishes a new framework for high-fidelity facial motion capture and dynamic 3D reconstruction. Our dataset are publicly available at: https://github.com/JHXion9/PoreTrack3D

Method: Unifies ColPali’s patch-level similarity scores as spatial relevance filters over OCR-extracted regions. Formalizes coordinate mapping between ViT patch grids and OCR bounding boxes, introduces intersection metrics for relevance propagation, establishes theoretical bounds on area efficiency.

Result: ColQwen3-4B with percentile-50 thresholding achieves 59.7% hit rate at IoU@0.5 (84.4% at IoU@0.25, 35.8% at IoU@0.7), mean IoU of 0.569 vs ~6.7% random. Reduces context tokens by 28.8% vs all OCR regions and 52.3% vs full-page image tokens.

Conclusion: Hybrid architecture effectively localizes relevant document regions for RAG without additional training, significantly improving precision and reducing context overhead. Open-source implementation Snappy released.

Abstract: Late-interaction multimodal retrieval models like ColPali achieve state-of-the-art document retrieval by embedding pages as images and computing fine-grained similarity between query tokens and visual patches. However, they return entire pages rather than specific regions, limiting utility for retrieval-augmented generation (RAG) where precise context is paramount. Conversely, OCR-based systems extract structured text with bounding box coordinates but lack semantic grounding for relevance assessment. We propose a hybrid architecture that unifies these paradigms: using ColPali’s patch-level similarity scores as spatial relevance filters over OCR-extracted regions. We formalize the coordinate mapping between vision transformer patch grids and OCR bounding boxes, introduce intersection metrics for relevance propagation, and establish theoretical bounds on area efficiency. We evaluate on BBox-DocVQA with ground-truth bounding boxes. For within-page localization (given correct page retrieval), ColQwen3-4B with percentile-50 thresholding achieves 59.7% hit rate at IoU@0.5 (84.4% at IoU@0.25, 35.8% at IoU@0.7), with mean IoU of 0.569, compared to ~6.7% for random region selection. Our approach reduces context tokens by 28.8% compared to returning all OCR regions and by 52.3% compared to full-page image tokens. Our approach operates at inference time without additional training. We release Snappy, an open-source implementation at https://github.com/athrael-soju/Snappy.

Method: Proposes centrifugal token pruning paradigm for near-to-far selection preserving fine-grained details, Buffering for Spatial Sparsity (BSS) criterion to defer selection of distant tokens, parallel greedy strategy for efficient selection, and selective fusion of salient information from discarded tokens.

Result: Consistently outperforms strong baselines across five VLMs with 88.9% pruning rate while delivering end-to-end inference speedup.

Conclusion: VLM-Pruner effectively addresses limitations of existing pruning methods by balancing redundancy and spatial sparsity, enabling efficient VLM deployment on mobile devices without sacrificing performance.

Abstract: Vision-language models (VLMs) excel at image understanding tasks, but the large number of visual tokens imposes significant computational costs, hindering deployment on mobile devices. Many pruning methods rely solely on token importance and thus overlook inter-token redundancy, retaining numerous duplicated tokens and wasting capacity. Although some redundancy-aware approaches have been proposed, they often ignore the spatial relationships among visual tokens. This can lead to overly sparse selections of retained tokens that fail to adequately cover the regions of target objects. To address these limitations, we propose VLM-Pruner, a training-free token pruning algorithm that explicitly balances redundancy and spatial sparsity. We introduce a centrifugal token pruning paradigm that enables near-to-far selection while prioritizing the preservation of fine-grained object details. Moreover, we design a Buffering for Spatial Sparsity (BSS) criterion that defers the selection of spatially distant tokens. We further adopt a parallel greedy strategy to conduct token selection efficiently. To mitigate information loss from pruning, we selectively fuse salient information from the discarded tokens into the retained ones. Comprehensive comparisons demonstrate that VLM-Pruner consistently outperforms strong baselines across five VLMs with an 88.9% pruning rate, while delivering an end-to-end inference speedup. The code is available at https://github.com/Casey-bit/VLMPruner.

Method: Proposes Basis Decomposition Module (BDM) that decomposes complex features into basis features to enhance certain information while eliminating redundancy. Extends BDM to create Spatial Difference Decomposition Module (SD²M), Spatial Difference Decomposition Downsampling Module (SD³M), and Temporal Difference Decomposition Module (TD²M). Builds SD²Net for single-frame ISTD using U-shaped architecture with SD²M and SD³M, and STD²Net for multi-frame ISTD by adding TD²M to introduce motion information.

Result: Extensive experiments show state-of-the-art performance. On single-frame ISTD, SD²Net performs well compared to established networks. On multi-frame ISTD datasets, STD²Net achieves mIoU of 87.68%, significantly outperforming SD²Net’s 64.97% mIoU.

Conclusion: The proposed basis decomposition approach effectively addresses infrared small target detection challenges by enhancing targets and suppressing backgrounds through feature decomposition. The modules and networks demonstrate strong performance on both single-frame and multi-frame tasks.

Abstract: Infrared small target detection (ISTD) faces two major challenges: a lack of discernible target texture and severe background clutter, which results in the background obscuring the target. To enhance targets and suppress backgrounds, we propose the Basis Decomposition Module (BDM) as an extensible and lightweight module based on basis decomposition, which decomposes a complex feature into several basis features and enhances certain information while eliminating redundancy. Extending BDM leads to a series of modules, including the Spatial Difference Decomposition Module (SD$^\mathrm{2}$M), Spatial Difference Decomposition Downsampling Module (SD$^\mathrm{3}$M), and Temporal Difference Decomposition Module (TD$^\mathrm{2}$M). Based on these modules, we develop the Spatial Difference Decomposition Network (SD$^\mathrm{2}$Net) for single-frame ISTD (SISTD) and the Spatiotemporal Difference Decomposition Network (STD$^\mathrm{2}$Net) for multi-frame ISTD (MISTD). SD$^\mathrm{2}$Net integrates SD$^\mathrm{2}$M and SD$^\mathrm{3}$M within an adapted U-shaped architecture. We employ TD$^\mathrm{2}$M to introduce motion information, which transforms SD$^\mathrm{2}$Net into STD$^\mathrm{2}$Net. Extensive experiments on SISTD and MISTD datasets demonstrate state-of-the-art (SOTA) performance. On the SISTD task, SD$^\mathrm{2}$Net performs well compared to most established networks. On the MISTD datasets, STD$^\mathrm{2}$Net achieves a mIoU of 87.68%, outperforming SD$^\mathrm{2}$Net, which achieves a mIoU of 64.97%. Our codes are available: https://github.com/greekinRoma/IRSTD_HC_Platform.

Method: Uses learnable binary masks with straight-through estimation to dynamically partition latent representations into transformation-sensitive and invariant components. Formulated within a unified optimization framework that jointly learns latent disentanglement and group transformation mappings, seamlessly integrable with any standard encoder-decoder architecture.

Result: Validated on five 2D/3D image datasets, demonstrating ability to automatically learn disentangled latent factors for group actions in diverse data. Downstream classification tasks confirm effectiveness of learned representations.

Conclusion: The framework successfully learns group actions on latent image manifolds without manual intervention, automatically discovering transformation-relevant structures and producing effective disentangled representations for downstream tasks.

Abstract: Modeling group actions on latent representations enables controllable transformations of high-dimensional image data. Prior works applying group-theoretic priors or modeling transformations typically operate in the high-dimensional data space, where group actions apply uniformly across the entire input, making it difficult to disentangle the subspace that varies under transformations. While latent-space methods offer greater flexibility, they still require manual partitioning of latent variables into equivariant and invariant subspaces, limiting the ability to robustly learn and operate group actions within the representation space. To address this, we introduce a novel end-to-end framework that for the first time learns group actions on latent image manifolds, automatically discovering transformation-relevant structures without manual intervention. Our method uses learnable binary masks with straight-through estimation to dynamically partition latent representations into transformation-sensitive and invariant components. We formulate this within a unified optimization framework that jointly learns latent disentanglement and group transformation mappings. The framework can be seamlessly integrated with any standard encoder-decoder architecture. We validate our approach on five 2D/3D image datasets, demonstrating its ability to automatically learn disentangled latent factors for group actions in diverse data, while downstream classification tasks confirm the effectiveness of the learned representations. Our code is publicly available at https://github.com/farhanaswarnali/Learning-Group-Actions-In-Disentangled-Latent-Image-Representations .

Method: Four main components: 1) Efficient autoencoder with 32x compression ratio to reduce tokens and balance training; 2) Channel-wise concatenation instead of token-wise to further reduce visual tokens; 3) Shared-and-decoupled network for mutual task improvement while meeting specific requirements; 4) Mixture-of-experts mechanism for visual understanding encoder to enhance perceptual capabilities with minimal parameter increase.

Result: EMMA-4B outperforms state-of-the-art unified multimodal approaches (e.g., BAGEL-7B) in both efficiency and performance, and achieves competitive results compared to recent multimodal understanding and generation experts (e.g., Qwen3-VL and Qwen-Image).

Conclusion: EMMA provides a solid foundation for future unified multimodal architecture development by demonstrating that efficient, high-performance multimodal systems can handle understanding, generation, and editing tasks simultaneously with a relatively small model size.

Abstract: We propose EMMA, an efficient and unified architecture for multimodal understanding, generation and editing. Specifically, EMMA primarily consists of 1) An efficient autoencoder with a 32x compression ratio, which significantly reduces the number of tokens required for generation. This also ensures the training balance between understanding and generation tasks by applying the same compression ratio to images. 2) Channel-wise concatenation instead of token-wise concatenation among visual understanding and generation tokens, which further reduces the visual tokens in unified architectures. 3) A shared-and-decoupled network that enables mutual improvements across tasks while meeting the task-specific modeling requirements. 4) A mixture-of-experts mechanism adopted for visual understanding encoder, which substantially improves perceptual capabilities with a few parameters increase. Extensive experiments have shown that EMMA-4B can significantly outperform state-of-the-art unified multimodal approaches (e.g., BAGEL-7B) in both efficiency and performance, while also achieving competitive results compared to recent multimodal understanding and generation experts (e.g., Qwen3-VL and Qwen-Image). We believe that EMMA lays a solid foundation for the future development of unified multimodal architectures.

Method: 1) Disentangled design: decouples geometry/motion (via dynamic point clouds along user-defined camera trajectories) from illumination (via relit frames projected into same geometry). 2) Light-Syn pipeline: degradation-based synthesis with inverse-mapping to create training pairs from monocular footage, covering static, dynamic, and AI-generated scenes.

Result: Light-X outperforms baseline methods in joint camera-illumination control and surpasses prior video relighting methods under both text- and background-conditioned settings.

Conclusion: The framework enables controllable video rendering with both viewpoint and illumination control, advancing toward generative modeling of real-world scenes by effectively disentangling geometry and lighting signals.

Abstract: Recent advances in illumination control extend image-based methods to video, yet still facing a trade-off between lighting fidelity and temporal consistency. Moving beyond relighting, a key step toward generative modeling of real-world scenes is the joint control of camera trajectory and illumination, since visual dynamics are inherently shaped by both geometry and lighting. To this end, we present Light-X, a video generation framework that enables controllable rendering from monocular videos with both viewpoint and illumination control. 1) We propose a disentangled design that decouples geometry and lighting signals: geometry and motion are captured via dynamic point clouds projected along user-defined camera trajectories, while illumination cues are provided by a relit frame consistently projected into the same geometry. These explicit, fine-grained cues enable effective disentanglement and guide high-quality illumination. 2) To address the lack of paired multi-view and multi-illumination videos, we introduce Light-Syn, a degradation-based pipeline with inverse-mapping that synthesizes training pairs from in-the-wild monocular footage. This strategy yields a dataset covering static, dynamic, and AI-generated scenes, ensuring robust training. Extensive experiments show that Light-X outperforms baseline methods in joint camera-illumination control and surpasses prior video relighting methods under both text- and background-conditioned settings.

Method: Built upon MIT-Adobe FiveK dataset, DEAR incorporates pairwise human preference scores collected via large-scale crowdsourcing. Each image pair was evaluated by 25 distinct human evaluators (total 13,648 participants). The dataset captures nuanced, context-sensitive aesthetic preferences.

Result: DEAR enables development of models for Evaluation of Aesthetics of Rendering (EAR) - a new task beyond traditional IQA. The dataset supports style preference prediction, aesthetic benchmarking, and personalized aesthetic modeling. A subset of 100 images with markup is published on HuggingFace.

Conclusion: DEAR is the first dataset to systematically address image aesthetics of rendering assessment grounded in subjective human preferences, filling a critical gap in evaluating rendering aesthetics for modern image applications.

Abstract: Traditional Image Quality Assessment~(IQA) focuses on quantifying technical degradations such as noise, blur, or compression artifacts, using both full-reference and no-reference objective metrics. However, evaluation of rendering aesthetics, a growing domain relevant to photographic editing, content creation, and AI-generated imagery, remains underexplored due to the lack of datasets that reflect the inherently subjective nature of style preference. In this work, a novel benchmark dataset designed to model human aesthetic judgments of image rendering styles is introduced: the Dataset for Evaluating the Aesthetics of Rendering (DEAR). Built upon the MIT-Adobe FiveK dataset, DEAR incorporates pairwise human preference scores collected via large-scale crowdsourcing, with each image pair evaluated by 25 distinct human evaluators with a total of 13,648 of them participating overall. These annotations capture nuanced, context-sensitive aesthetic preferences, enabling the development and evaluation of models that go beyond traditional distortion-based IQA, focusing on a new task: Evaluation of Aesthetics of Rendering (EAR). The data collection pipeline is described, human voting patterns are analyzed, and multiple use cases are outlined, including style preference prediction, aesthetic benchmarking, and personalized aesthetic modeling. To the best of the authors’ knowledge, DEAR is the first dataset to systematically address image aesthetics of rendering assessment grounded in subjective human preferences. A subset of 100 images with markup for them is published on HuggingFace (huggingface.co/datasets/vsevolodpl/DEAR).

Method: Proposes VDOT using distribution matching distillation (DMD) with computational optimal transport (OT) instead of KL minimization to optimize discrepancy between real and fake score distributions. OT provides geometric constraints to prevent zero-forcing/gradient collapse in few-step generation. Also integrates a discriminator to enhance video quality. Includes automated pipeline for video data annotation/filtering and creates UVCBench benchmark for evaluation.

Result: VDOT with only 4 denoising steps outperforms or matches other baselines requiring 100 denoising steps, demonstrating superior efficiency and quality.

Conclusion: VDOT provides an efficient, unified solution for video creation that addresses the limitations of existing models through optimal transport-based distillation, achieving state-of-the-art performance with dramatically reduced inference time.

Abstract: The rapid development of generative models has significantly advanced image and video applications. Among these, video creation, aimed at generating videos under various conditions, has gained substantial attention. However, existing video creation models either focus solely on a few specific conditions or suffer from excessively long generation times due to complex model inference, making them impractical for real-world applications. To mitigate these issues, we propose an efficient unified video creation model, named VDOT. Concretely, we model the training process with the distribution matching distillation (DMD) paradigm. Instead of using the Kullback-Leibler (KL) minimization, we additionally employ a novel computational optimal transport (OT) technique to optimize the discrepancy between the real and fake score distributions. The OT distance inherently imposes geometric constraints, mitigating potential zero-forcing or gradient collapse issues that may arise during KL-based distillation within the few-step generation scenario, and thus, enhances the efficiency and stability of the distillation process. Further, we integrate a discriminator to enable the model to perceive real video data, thereby enhancing the quality of generated videos. To support training unified video creation models, we propose a fully automated pipeline for video data annotation and filtering that accommodates multiple video creation tasks. Meanwhile, we curate a unified testing benchmark, UVCBench, to standardize evaluation. Experiments demonstrate that our 4-step VDOT outperforms or matches other baselines with 100 denoising steps.

Method: CHIMERA formulates morphing as cached inversion-guided denoising with two key components: 1) Adaptive Cache Injection (ACI) that caches and adaptively re-injects features from both inputs during denoising, and 2) Semantic Anchor Prompting (SAP) that generates a shared semantic anchor prompt using vision-language models to bridge dissimilar inputs.

Result: Extensive experiments and user studies show CHIMERA achieves smoother and more semantically aligned transitions than existing methods, establishing new state-of-the-art in image morphing.

Conclusion: CHIMERA successfully addresses the challenge of smooth image morphing with large semantic disparities through adaptive feature injection and semantic prompting, with the proposed GLCS metric providing better evaluation of morphing quality.

Abstract: Diffusion models exhibit remarkable generative ability, yet achieving smooth and semantically consistent image morphing remains a challenge. Existing approaches often yield abrupt transitions or over-saturated appearances due to the lack of adaptive structural and semantic alignments. We propose CHIMERA, a zero-shot diffusion-based framework that formulates morphing as a cached inversion-guided denoising process. To handle large semantic and appearance disparities, we propose Adaptive Cache Injection and Semantic Anchor Prompting. Adaptive Cache Injection (ACI) caches down, mid, and up blocks features from both inputs during DDIM inversion and re-injects them adaptively during denoising, enabling spatial and semantic alignment in depth- and time-adaptive manners and enabling natural feature fusion and smooth transitions. Semantic Anchor Prompting (SAP) leverages a vision-language model to generate a shared anchor prompt that serves as a semantic anchor, bridging dissimilar inputs and guiding the denoising process toward coherent results. Finally, we introduce the Global-Local Consistency Score (GLCS), a morphing-oriented metric that simultaneously evaluates the global harmonization of the two inputs and the smoothness of the local morphing transition. Extensive experiments and user studies show that CHIMERA achieves smoother and more semantically aligned transitions than existing methods, establishing a new state of the art in image morphing. The code and project page will be publicly released.

Method: Proposes HybridSplat with reflection-baked Gaussian tracing (bakes view-dependent reflections within each Gaussian primitive), tile-based Gaussian splatting for reflection rendering, unified hybrid splatting framework, pipeline-level acceleration, and reflection-sensitive Gaussian pruning.

Result: Achieves ~7x rendering speed acceleration across complex reflective scenes from Ref-NeRF and NeRF-Casting with 4x fewer Gaussian primitives than ray-tracing based baselines, while preserving reflection quality.

Conclusion: HybridSplat serves as a new state-of-the-art method for complex reflective scenes, offering faster rendering, lower memory usage, and high-fidelity reconstruction through hybrid splatting and optimization techniques.

Abstract: Rendering complex reflection of real-world scenes using 3D Gaussian splatting has been a quite promising solution for photorealistic novel view synthesis, but still faces bottlenecks especially in rendering speed and memory storage. This paper proposes a new Hybrid Splatting(HybridSplat) mechanism for Gaussian primitives. Our key idea is a new reflection-baked Gaussian tracing, which bakes the view-dependent reflection within each Gaussian primitive while rendering the reflection using tile-based Gaussian splatting. Then we integrate the reflective Gaussian primitives with base Gaussian primitives using a unified hybrid splatting framework for high-fidelity scene reconstruction. Moreover, we further introduce a pipeline-level acceleration for the hybrid splatting, and reflection-sensitive Gaussian pruning to reduce the model size, thus achieving much faster rendering speed and lower memory storage while preserving the reflection rendering quality. By extensive evaluation, our HybridSplat accelerates about 7x rendering speed across complex reflective scenes from Ref-NeRF, NeRF-Casting with 4x fewer Gaussian primitives than similar ray-tracing based Gaussian splatting baselines, serving as a new state-of-the-art method especially for complex reflective scenes.

Method: Proposes MNAME with a memory bank mechanism that maintains a large, dynamically updated queue of negative samples across training iterations. This decouples negative samples from mini-batch size and provides abundant negatives for contrastive objectives without computational overhead. Uses a low-dimensional projection head to minimize memory and bandwidth costs.

Result: Achieves state-of-the-art FID of 2.40 on ImageNet-256 within 400k steps, significantly outperforming comparable methods. Enables substantially faster convergence and superior generative quality more efficiently.

Conclusion: MNAME offers three key advantages: (1) self-contained with no dependency on pretrained vision foundation models, (2) no additional parameters or computational cost during inference, and (3) enables faster convergence with superior generative quality, making it an efficient plug-and-play solution for improving denoising generative models.

Abstract: The dominance of denoising generative models (e.g., diffusion, flow-matching) in visual synthesis is tempered by their substantial training costs and inefficiencies in representation learning. While injecting discriminative representations via auxiliary alignment has proven effective, this approach still faces key limitations: the reliance on external, pre-trained encoders introduces overhead and domain shift. A dispersed-based strategy that encourages strong separation among in-batch latent representations alleviates this specific dependency. To assess the effect of the number of negative samples in generative modeling, we propose {\mname}, a plug-and-play training framework that requires no external encoders. Our method integrates a memory bank mechanism that maintains a large, dynamically updated queue of negative samples across training iterations. This decouples the number of negatives from the mini-batch size, providing abundant and high-quality negatives for a contrastive objective without a multiplicative increase in computational cost. A low-dimensional projection head is used to further minimize memory and bandwidth overhead. {\mname} offers three principal advantages: (1) it is self-contained, eliminating dependency on pretrained vision foundation models and their associated forward-pass overhead; (2) it introduces no additional parameters or computational cost during inference; and (3) it enables substantially faster convergence, achieving superior generative quality more efficiently. On ImageNet-256, {\mname} achieves a state-of-the-art FID of \textbf{2.40} within 400k steps, significantly outperforming comparable methods.

Method: Introduces a GPU-optimized rotation-invariant convolution framework that eliminates the traditional im2col step. Exploits structured data sharing among symmetrically rotated filters to achieve multi-orientation convolution with reduced memory traffic and computational redundancy. Generalizes the approach to accelerate convolution with arbitrary (non-symmetric) rotation angles.

Result: Achieves 20-55% faster training and 15-45% lower energy consumption than CUDNN while maintaining comparable accuracy to state-of-the-art rotation-invariant methods. In eight-orientation setting: up to 45% speedup and 41% energy savings on 256×256 inputs, 32% speedup and 23% lower energy on 1024×1024 inputs. Integrated into U-Net yields up to 6% accuracy improvement over non-rotation-aware baseline.

Conclusion: The proposed method provides an effective and highly efficient alternative to existing rotation-invariant CNN frameworks, demonstrating practical benefits for UAV aerial image segmentation with significant speed/energy improvements while maintaining accuracy.

Abstract: Rotation invariance is essential for precise, object-level segmentation in UAV aerial imagery, where targets can have arbitrary orientations and exhibit fine-scale details. Conventional segmentation architectures like U-Net rely on convolution operators that are not rotation-invariant, leading to degraded segmentation accuracy across varying viewpoints. Rotation invariance can be achieved by expanding the filter bank across multiple orientations; however, this will significantly increase computational cost and memory traffic. In this paper, we introduce a GPU-optimized rotation-invariant convolution framework that eliminates the traditional data-lowering (im2col) step required for matrix-multiplication-based convolution. By exploiting structured data sharing among symmetrically rotated filters, our method achieves multi-orientation convolution with greatly reduced memory traffic and computational redundancy. We further generalize the approach to accelerate convolution with arbitrary (non-symmetric) rotation angles. Across extensive benchmarks, the proposed convolution achieves 20–55% faster training and 15–45% lower energy consumption than CUDNN, while maintaining accuracy comparable to state-of-the-art rotation-invariant methods. In the eight-orientation setting, our approach achieves up to 45% speedup and 41% energy savings on 256(\times)256 inputs, and 32% speedup and 23% lower energy usage on 1024(\times)1024 inputs. Integrated into a U-Net segmentation model, the framework yields up to 6% improvement in accuracy over the non-rotation-aware baseline. These results demonstrate that the proposed method provides an effective and highly efficient alternative to existing rotation-invariant CNN frameworks.

Method: TextGuider analyzes attention patterns in Multi-Modal Diffusion Transformer (MM-DiT) models for text-related tokens. It applies latent guidance during early denoising steps using two novel loss functions to align textual content tokens with text regions in the image.

Result: Achieves state-of-the-art performance in test-time text rendering with significant gains in recall, strong results in OCR accuracy, and improved CLIP scores.

Conclusion: TextGuider effectively addresses text omission in diffusion models through training-free attention-based guidance, enabling more accurate and complete text rendering without requiring model fine-tuning.

Abstract: Despite recent advances, diffusion-based text-to-image models still struggle with accurate text rendering. Several studies have proposed fine-tuning or training-free refinement methods for accurate text rendering. However, the critical issue of text omission, where the desired text is partially or entirely missing, remains largely overlooked. In this work, we propose TextGuider, a novel training-free method that encourages accurate and complete text appearance by aligning textual content tokens and text regions in the image. Specifically, we analyze attention patterns in Multi-Modal Diffusion Transformer(MM-DiT) models, particularly for text-related tokens intended to be rendered in the image. Leveraging this observation, we apply latent guidance during the early stage of denoising steps based on two loss functions that we introduce. Our method achieves state-of-the-art performance in test-time text rendering, with significant gains in recall and strong results in OCR accuracy and CLIP score.

Method: Uses motion feature extraction to capture temporal dynamics, followed by class-wise graph construction and representative sample selection using the Infomap algorithm to create diverse, informative synthetic video subsets.

Result: Achieved 69.38% test accuracy on EchoNet-Dynamic datasets using only 25 synthetic videos, demonstrating effectiveness and scalability for medical video dataset distillation.

Conclusion: The proposed motion-based graph approach successfully distills echocardiographic video datasets into compact synthetic subsets while preserving essential clinical characteristics, offering a scalable solution for medical video data management.

Abstract: Echocardiography plays a critical role in the diagnosis and monitoring of cardiovascular diseases as a non-invasive real-time assessment of cardiac structure and function. However, the growing scale of echocardiographic video data presents significant challenges in terms of storage, computation, and model training efficiency. Dataset distillation offers a promising solution by synthesizing a compact, informative subset of data that retains the key clinical features of the original dataset. In this work, we propose a novel approach for distilling a compact synthetic echocardiographic video dataset. Our method leverages motion feature extraction to capture temporal dynamics, followed by class-wise graph construction and representative sample selection using the Infomap algorithm. This enables us to select a diverse and informative subset of synthetic videos that preserves the essential characteristics of the original dataset. We evaluate our approach on the EchoNet-Dynamic datasets and achieve a test accuracy of (69.38%) using only (25) synthetic videos. These results demonstrate the effectiveness and scalability of our method for medical video dataset distillation.

Method: Flex employs a small set of learnable scene tokens to jointly encode information from all image tokens across different cameras and timesteps. It’s geometry-agnostic, learning compact scene representations directly from data without explicit 3D priors, then feeds this compressed representation to an LLM-based policy model.

Result: On a 20,000-hour proprietary dataset, Flex achieves 2.2x greater inference throughput while significantly improving driving performance compared to state-of-the-art methods. The scene tokens also develop emergent scene decomposition capabilities without explicit supervision.

Conclusion: The findings challenge the assumption that 3D priors are necessary for autonomous driving, demonstrating that data-driven joint encoding offers a more scalable, efficient, and effective path for future systems.

Abstract: We present Flex, an efficient and effective scene encoder that addresses the computational bottleneck of processing high-volume multi-camera data in end-to-end autonomous driving. Flex employs a small set of learnable scene tokens to jointly encode information from all image tokens across different cameras and timesteps. By design, our approach is geometry-agnostic, learning a compact scene representation directly from data without relying on the explicit 3D inductive biases, such as Bird-Eye-View (BEV), occupancy or tri-plane representations, which are common in prior work. This holistic encoding strategy aggressively compresses the visual input for the downstream Large Language Model (LLM) based policy model. Evaluated on a large-scale proprietary dataset of 20,000 driving hours, our Flex achieves 2.2x greater inference throughput while improving driving performance by a large margin compared to state-of-the-art methods. Furthermore, we show that these compact scene tokens develop an emergent capability for scene decomposition without any explicit supervision. Our findings challenge the prevailing assumption that 3D priors are necessary, demonstrating that a data-driven, joint encoding strategy offers a more scalable, efficient and effective path for future autonomous driving systems.

Method: AutoRefiner with two key designs: 1) Pathwise noise refinement that refines noise along stochastic denoising paths, and 2) Reflective KV-cache mechanism tailored for AR-VDMs.

Result: AutoRefiner effectively enhances sample fidelity for AR-VDMs by refining noise in a single forward pass, serving as an efficient plug-in without requiring model parameter updates.

Conclusion: AutoRefiner successfully extends noise refinement to AR-VDMs, addressing the limitations of T2I methods and providing practical inference-time alignment for improved video generation quality.

Abstract: Autoregressive video diffusion models (AR-VDMs) show strong promise as scalable alternatives to bidirectional VDMs, enabling real-time and interactive applications. Yet there remains room for improvement in their sample fidelity. A promising solution is inference-time alignment, which optimizes the noise space to improve sample fidelity without updating model parameters. Yet, optimization- or search-based methods are computationally impractical for AR-VDMs. Recent text-to-image (T2I) works address this via feedforward noise refiners that modulate sampled noises in a single forward pass. Can such noise refiners be extended to AR-VDMs? We identify the failure of naively extending T2I noise refiners to AR-VDMs and propose AutoRefiner-a noise refiner tailored for AR-VDMs, with two key designs: pathwise noise refinement and a reflective KV-cache. Experiments demonstrate that AutoRefiner serves as an efficient plug-in for AR-VDMs, effectively enhancing sample fidelity by refining noise along stochastic denoising paths.

Method: TS-FSAR framework with three components: 1) visual Ladder Side Network (LSN) for efficient CLIP fine-tuning; 2) Task-Specific Distance Correlation Matching (TS-DCM) using α-distance correlation to model both linear and nonlinear inter-frame dependencies with task prototypes; 3) Guiding LSN with Adapted CLIP (GLAC) module to regularize LSN using frozen CLIP for better α-distance correlation estimation.

Result: Extensive experiments on five widely-used benchmarks demonstrate superior performance compared to prior state-of-the-art methods.

Conclusion: TS-FSAR effectively addresses limitations in few-shot action recognition by combining efficient CLIP adaptation with advanced task-specific matching that captures complex inter-frame dependencies, achieving state-of-the-art performance.

Abstract: Few-shot action recognition (FSAR) has recently made notable progress through set matching and efficient adaptation of large-scale pre-trained models. However, two key limitations persist. First, existing set matching metrics typically rely on cosine similarity to measure inter-frame linear dependencies and then perform matching with only instance-level information, thus failing to capture more complex patterns such as nonlinear relationships and overlooking task-specific cues. Second, for efficient adaptation of CLIP to FSAR, recent work performing fine-tuning via skip-fusion layers (which we refer to as side layers) has significantly reduced memory cost. However, the newly introduced side layers are often difficult to optimize under limited data conditions. To address these limitations, we propose TS-FSAR, a framework comprising three components: (1) a visual Ladder Side Network (LSN) for efficient CLIP fine-tuning; (2) a metric called Task-Specific Distance Correlation Matching (TS-DCM), which uses $α$-distance correlation to model both linear and nonlinear inter-frame dependencies and leverages a task prototype to enable task-specific matching; and (3) a Guiding LSN with Adapted CLIP (GLAC) module, which regularizes LSN using the adapted frozen CLIP to improve training for better $α$-distance correlation estimation under limited supervision. Extensive experiments on five widely-used benchmarks demonstrate that our TS-FSAR yields superior performance compared to prior state-of-the-arts.

Method: Developed semi-automated labeling pipeline with human-in-the-loop verification (YawDD+ dataset). Trained MNasNet classifier and YOLOv11 detector on this improved dataset.

Result: Achieved 99.34% classification accuracy and 95.69% detection mAP, improving frame accuracy by 6% and mAP by 5% over video-level supervision. Runs at 59.8 FPS on NVIDIA Jetson Nano edge hardware.

Conclusion: Enhanced data quality enables accurate on-device yawning monitoring without server-side computation, supporting real-time driver fatigue detection.

Abstract: Driver fatigue remains a leading cause of road accidents, with 24% of crashes involving drowsy drivers. While yawning serves as an early behavioral indicator of fatigue, existing machine learning approaches face significant challenges due to video-annotated datasets that introduce systematic noise from coarse temporal annotations. We develop a semi-automated labeling pipeline with human-in-the-loop verification, which we apply to YawDD, enabling more accurate model training. Training the established MNasNet classifier and YOLOv11 detector architectures on YawDD+ improves frame accuracy by up to 6% and mAP by 5% over video-level supervision, achieving 99.34% classification accuracy and 95.69% detection mAP. The resulting approach deliver up to 59.8 FPS on edge AI hardware (NVIDIA Jetson Nano), confirming that enhanced data quality alone supports on-device yawning monitoring without server-side computation.

Method: Twenty monads from Leibniz’s Monadology are grouped into six bundles (ontology, dynamics, representation/consciousness, harmony/reason, body/organization, teleology) and implemented as executable specifications on top of the AAS kernel. The framework uses minimal Python implementations with a four-step pattern: inputs/setup, clause implementation, numerical results, and design implications.

Result: The clause system exhibits bounded and interpretable behavior: AAS trajectories remain continuous and rate-limited, contradictions trigger explicit penalties, hierarchical refinement reveals organic structure, dual views align via harmony terms, and windowed drift separates improvement from degradation.

Conclusion: The monad-based clause framework provides a transparent, code-level blueprint for constraining and analyzing internal dynamics in artificial agents, demonstrating that philosophical concepts can be directly implemented to create auditable AI governance systems.

Abstract: Large language models (LLMs) are often deployed as powerful yet opaque systems, leaving open how their internal memory and “self-like” behavior should be governed in a principled and auditable way. The Artificial Age Score (AAS) was previously introduced and mathematically justified through three theorems that characterise it as a metric of artificial memory aging. Building on this foundation, the present work develops an engineering-oriented, clause-based architecture that imposes law-like constraints on LLM memory and control. Twenty selected monads from Leibniz’s Monadology are grouped into six bundles: ontology, dynamics, representation and consciousness, harmony and reason, body and organisation, and teleology, and each bundle is realised as an executable specification on top of the AAS kernel. Across six minimal Python implementations, these clause families are instantiated in numerical experiments acting on channel-level quantities such as recall scores, redundancy, and weights. Each implementation follows a four-step pattern: inputs and setup, clause implementation, numerical results, and implications for LLM design, emphasising that the framework is not only philosophically motivated but also directly implementable. The experiments show that the clause system exhibits bounded and interpretable behavior: AAS trajectories remain continuous and rate-limited, contradictions and unsupported claims trigger explicit penalties, and hierarchical refinement reveals an organic structure in a controlled manner. Dual views and goal-action pairs are aligned by harmony terms, and windowed drift in perfection scores separates sustained improvement from sustained degradation. Overall, the monad-based clause framework uses AAS as a backbone and provides a transparent, code-level blueprint for constraining and analyzing internal dynamics in artificial agents.

Method: 1) Constraint modeling approach as an exact solution method for small scheduling scenarios using state-of-the-art constraint-solving technology. 2) Construction heuristic and tailored metaheuristic using local search for large-scale problem instances.

Result: The metaheuristic approach has been successfully deployed and is currently being used in an industrial setting, demonstrating practical applicability for real-life scheduling problems.

Conclusion: The paper addresses the gap in existing scheduling techniques by introducing a novel variant with real-world constraints and providing both exact and heuristic solutions that are applicable in industrial settings, with the metaheuristic already deployed in practice.

Abstract: The task of finding efficient production schedules for parallel machines is a challenge that arises in most industrial manufacturing domains. There is a large potential to minimize production costs through automated scheduling techniques, due to the large-scale requirements of modern factories. In the past, solution approaches have been studied for many machine scheduling variations, where even basic variants have been shown to be NP-hard. However, in today’s real-life production environments, additional complex precedence constraints and resource restrictions with calendars arise that must be fulfilled. These additional constraints cannot be tackled efficiently by existing solution techniques. Thus, there is a strong need to develop and analyze automated methods that can solve such real-life parallel machine scheduling scenarios. In this work, we introduce a novel variant of parallel machine scheduling with job precedences and calendar-based cumulative resource constraints that arises in real-life industrial use cases. A constraint modeling approach is proposed as an exact solution method for small scheduling scenarios together with state-of-the-art constraint-solving technology. Further, we propose a construction heuristic as well as a tailored metaheuristic using local search to efficiently tackle large-scale problem instances. This metaheuristic approach has been deployed and is currently being used in an industrial setting.

Method: Proposes a demographic-aware machine learning framework with a behaviorally realistic demographic-based data augmentation strategy that expands small QoE datasets six-fold by modeling varying user sensitivities to streaming impairments (rebuffering, bitrate variation, quality degradation). Evaluates comprehensive classical ML models alongside advanced deep learning architectures including attention-based MLP and TabNet.

Result: Experimental results show significant improvements in prediction accuracy across RMSE, MAE, and R metrics compared to baseline models. TabNet achieves the strongest performance, benefiting from its inherent feature selection and attention mechanisms. Demographic-aware augmentation substantially enhances QoE prediction robustness.

Conclusion: The demographic-aware augmentation approach provides a scalable direction for personalized QoE-aware intelligence in 5G video streaming networks, enabling more accurate and robust QoE prediction that accounts for individual user differences.

Abstract: Quality of Experience (QoE) prediction is a critical component of modern multimedia systems, particularly for adaptive video streaming in 5G networks. Accurate QoE estimation enables intelligent resource management and supports user centric service delivery. Existing QoE prediction approaches primarily rely on limited datasets and assume uniform user perception, which restricts their applicability in heterogeneous real world environments. This paper proposes a demographic aware machine learning framework for personalized QoE prediction. We introduce a behaviorally realistic demographic based data augmentation strategy that expands a small QoE dataset six fold by modeling varying user sensitivities to streaming impairments such as rebuffering, bitrate variation, and quality degradation. Using the augmented dataset, we evaluate a comprehensive set of classical machine learning models alongside advanced deep learning architectures, including an attention-based MLP and TabNet. Experimental results demonstrate significant improvements in prediction accuracy across RMSE, MAE, and R metrics compared to baseline models. Among all evaluated approaches, TabNet achieves the strongest performance, benefiting from its inherent feature selection and attention mechanisms. The results confirm that demographic-aware augmentation substantially enhances QoE prediction robustness and provides a scalable direction for personalized QoE-aware intelligence in 5G video streaming networks.

Method: Built a simplified Fire Emblem Heroes game in Unity with Standard and Mirror Modes. Used a combination of Generative Adversarial Imitation Learning (GAIL), Behavioral Cloning (BC), and Proximal Policy Optimization (PPO) to train AI models on player demonstrations. Conducted two experiment sets: first to find suitable imitation models, second to evaluate models with player tests using participant-provided demonstrations.

Result: The model successfully imitated defensive behaviors but struggled with offensive strategies. Players recognized their own retreating tactics in Mirror Mode and reported higher satisfaction compared to Standard Mode. The approach showed promise but needs refinement for better offensive strategy imitation.

Conclusion: Mirror Mode successfully challenges players by having AI mimic their strategies, increasing player satisfaction. While defensive behavior imitation works well, offensive strategy imitation needs improvement. Further refinement could enhance imitation quality and player satisfaction when facing their own strategies.

Abstract: Enemy strategies in turn-based games should be surprising and unpredictable. This study introduces Mirror Mode, a new game mode where the enemy AI mimics the personal strategy of a player to challenge them to keep changing their gameplay. A simplified version of the Nintendo strategy video game Fire Emblem Heroes has been built in Unity, with a Standard Mode and a Mirror Mode. Our first set of experiments find a suitable model for the task to imitate player demonstrations, using Reinforcement Learning and Imitation Learning: combining Generative Adversarial Imitation Learning, Behavioral Cloning, and Proximal Policy Optimization. The second set of experiments evaluates the constructed model with player tests, where models are trained on demonstrations provided by participants. The gameplay of the participants indicates good imitation in defensive behavior, but not in offensive strategies. Participant’s surveys indicated that they recognized their own retreating tactics, and resulted in an overall higher player-satisfaction for Mirror Mode. Refining the model further may improve imitation quality and increase player’s satisfaction, especially when players face their own strategies. The full code and survey results are stored at: https://github.com/YannaSmid/MirrorMode

Method: Proposes a principled method to model constraints via compilation of user’s knowledge graph into macro-facets. Proves that common hierarchical and quota-based constraints form a laminar matroid, enabling casting structured personalization as submodular maximization under matroid constraint.

Result: Theoretical characterization shows constraints form valid laminar matroid, enabling greedy algorithms with constant-factor guarantees (and (1-1/e) via continuous greedy) for richer class of structured personalization problems.

Conclusion: The approach provides a principled framework for structured personalization of LLM agents with complex constraints, offering theoretical guarantees for practical personalization problems that violate standard algorithm assumptions.

Abstract: Personalizing Large Language Model (LLM) agents requires conditioning them on user-specific data, creating a critical trade-off between task utility and data disclosure. While the utility of adding user data often exhibits diminishing returns (i.e., submodularity), enabling near-optimal greedy selection, real-world personalization is complicated by structural constraints. These include logical dependencies (e.g., selecting fact A requires fact B), categorical quotas (e.g., select at most one writing style), and hierarchical rules (e.g., select at most two social media preferences, of which at most one can be for a professional network). These constraints violate the assumptions of standard subset selection algorithms. We propose a principled method to formally model such constraints. We introduce a compilation process that transforms a user’s knowledge graph with dependencies into a set of abstract macro-facets. Our central result is a proof that common hierarchical and quota-based constraints over these macro-facets form a valid laminar matroid. This theoretical characterization lets us cast structured personalization as submodular maximization under a matroid constraint, enabling greedy with constant-factor guarantees (and (1-1/e) via continuous greedy) for a much richer and more realistic class of problems.

Method: Proposes Hierarchical Implicit Periodicity (HIP) learning with two explicit techniques: 1) Periodic autoencoders to explore gesture motion phase manifolds for realistic distributions and instance-level diversities, 2) Cascaded guidance to model hierarchical relationships between face, body, and hand movements.

Result: Outperforms state-of-the-art co-speech gesture generation methods in both quantitative and qualitative evaluations on 3D avatars.

Conclusion: HIP effectively models intrinsic correlations in co-speech gestures through hierarchical implicit periodicity learning, achieving more realistic and coordinated 3D body movements from speech.

Abstract: Generating 3D-based body movements from speech shows great potential in extensive downstream applications, while it still suffers challenges in imitating realistic human movements. Predominant research efforts focus on end-to-end generation schemes to generate co-speech gestures, spanning GANs, VQ-VAE, and recent diffusion models. As an ill-posed problem, in this paper, we argue that these prevailing learning schemes fail to model crucial inter- and intra-correlations across different motion units, i.e. head, body, and hands, thus leading to unnatural movements and poor coordination. To delve into these intrinsic correlations, we propose a unified Hierarchical Implicit Periodicity (HIP) learning approach for audio-inspired 3D gesture generation. Different from predominant research, our approach models this multi-modal implicit relationship by two explicit technique insights: i) To disentangle the complicated gesture movements, we first explore the gesture motion phase manifolds with periodic autoencoders to imitate human natures from realistic distributions while incorporating non-period ones from current latent states for instance-level diversities. ii) To model the hierarchical relationship of face motions, body gestures, and hand movements, driving the animation with cascaded guidance during learning. We exhibit our proposed approach on 3D avatars and extensive experiments show our method outperforms the state-of-the-art co-speech gesture generation methods by both quantitative and qualitative evaluations. Code and models will be publicly available.

Method: Proposes a 3-pillar roadmap: (1) learning and representing human values using formal semantics, (2) ensuring value alignment for individual agents and multiagent systems, and (3) providing value-based explainability of behavior.

Result: Presents ongoing work on these topics with applications to real-life domains, demonstrating practical implementation of the value awareness concept.

Conclusion: Value awareness provides a more comprehensive approach than traditional value alignment, offering a structured roadmap for engineering AI systems that understand, align with, and can explain their behavior based on human values.

Abstract: This paper introduces the concept of value awareness in AI, which goes beyond the traditional value-alignment problem. Our definition of value awareness presents us with a concise and simplified roadmap for engineering value-aware AI. The roadmap is structured around three core pillars: (1) learning and representing human values using formal semantics, (2) ensuring the value alignment of both individual agents and multiagent systems, and (3) providing value-based explainability on behaviour. The paper presents a selection of our ongoing work on some of these topics, along with applications to real-life domains.

Method: Evaluated 20+ LLMs on 11 causal tasks using collider graphs (C1→E←C2) with Direct (one-shot probability judgment) and Chain of Thought approaches. Responses were modeled using leaky noisy-OR causal Bayes nets with parameters including shared prior and causal strengths, selecting best model via AIC between symmetric and asymmetric variants.

Result: The abstract doesn’t provide specific results, but the methodology establishes a framework for comparing LLM and human causal reasoning through parameter estimation and model selection.

Conclusion: The study provides a systematic approach to evaluate and compare causal reasoning capabilities between LLMs and humans, addressing fundamental questions about intelligence alignment and reasoning patterns.

Abstract: The nature of intelligence in both humans and machines is a longstanding question. While there is no universally accepted definition, the ability to reason causally is often regarded as a pivotal aspect of intelligence (Lake et al., 2017). Evaluating causal reasoning in LLMs and humans on the same tasks provides hence a more comprehensive understanding of their respective strengths and weaknesses. Our study asks: (Q1) Are LLMs aligned with humans given the \emph{same} reasoning tasks? (Q2) Do LLMs and humans reason consistently at the task level? (Q3) Do they have distinct reasoning signatures? We answer these by evaluating 20+ LLMs on eleven semantically meaningful causal tasks formalized by a collider graph ($C_1!\to!E!\leftarrow!C_2$ ) under \emph{Direct} (one-shot number as response = probability judgment of query node being one and \emph{Chain of Thought} (CoT; think first, then provide answer). Judgments are modeled with a leaky noisy-OR causal Bayes net (CBN) whose parameters $θ=(b,m_1,m_2,p(C)) \in [0,1]$ include a shared prior $p(C)$; we select the winning model via AIC between a 3-parameter symmetric causal strength ($m_1{=}m_2$) and 4-parameter asymmetric ($m_1{\neq}m_2$) variant.

Method: Created benchmark using LLM-assisted discovery with expert annotation: (1) LLM-assisted, expert-verified derivation of workflows from real email threads and spreadsheet version histories, (2) meticulous expert annotation requiring 700+ hours of domain-expert effort, resulting in 172 composite workflows with 384 tasks across 1,710 spreadsheets with 27 million cells.

Result: Frontier AI systems perform poorly: GPT 5.1 Pro passes only 38.4% of workflows after 48 hours total, Claude Sonnet 4.5 passes just 25.0%. Comprehensive case studies reveal the challenges real-world enterprise workflows pose for AI agents.

Conclusion: Real-world enterprise workflows present significant challenges for current AI agents, highlighting the gap between academic benchmarks and authentic professional work, and demonstrating the need for more robust evaluation frameworks like Finch.

Abstract: We introduce a finance & accounting benchmark (Finch) for evaluating AI agents on real-world, enterprise-grade professional workflows – interleaving data entry, structuring, formatting, web search, cross-file retrieval, calculation, modeling, validation, translation, visualization, and reporting. Finch is sourced from authentic enterprise workspaces at Enron (15,000 spreadsheets and 500,000 emails from 150 employees) and other financial institutions, preserving in-the-wild messiness across multimodal artifacts (text, tables, formulas, charts, code, and images) and spanning diverse domains such as budgeting, trading, and asset management. We propose a workflow construction process that combines LLM-assisted discovery with expert annotation: (1) LLM-assisted, expert-verified derivation of workflows from real-world email threads and version histories of spreadsheet files, and (2) meticulous expert annotation for workflows, requiring over 700 hours of domain-expert effort. This yields 172 composite workflows with 384 tasks, involving 1,710 spreadsheets with 27 million cells, along with PDFs and other artifacts, capturing the intrinsically messy, long-horizon, knowledge-intensive, and collaborative nature of real-world enterprise work. We conduct both human and automated evaluations of frontier AI systems including GPT 5.1, Claude Sonnet 4.5, Gemini 3 Pro, Grok 4, and Qwen 3 Max, and GPT 5.1 Pro spends 48 hours in total yet passes only 38.4% of workflows, while Claude Sonnet 4.5 passes just 25.0%. Comprehensive case studies further surface the challenges that real-world enterprise workflows pose for AI agents.

Method: Comparative investigation testing models under controlled data corruption conditions (e.g., 50% token corruption), analyzing results through multi-perspective lens integrating information theory, PAC learning, and gradient dynamics.

Result: Autoregressive language models (like GPT-2) are remarkably resilient (test NLL increases only modestly from 2.87 to 3.59 despite 50% corruption). Class-conditional diffusion models degrade catastrophically (image-label consistency plummets by 56.81%). Classifiers show moderate impact that diminishes with dataset scale.

Conclusion: Robustness is heavily influenced by two key principles: richness of conditioning information (constrains learning problem) and absolute information content of training data (allows correct signal to dominate statistical noise).

Abstract: A systematic, comparative investigation into the effects of low-quality data reveals a stark spectrum of robustness across modern probabilistic models. We find that autoregressive language models, from token prediction to sequence-to-sequence tasks, are remarkably resilient (for GPT-2, test NLL increases modestly from 2.87 to 3.59 despite 50% token corruption). By contrast, under the same levels of data corruption, class-conditional diffusion models degrade catastrophically (image-label consistency plummets by 56.81% relative to baseline), while classifiers show a moderate impact that diminishes with dataset scale. To explain these discrepancies, we analyze the results through a multi-perspective lens, integrating information theory, PAC learning, and gradient dynamics. These analyses suggest that robustness is heavily influenced by two key principles: the richness of conditioning information, which constrains the learning problem, and the absolute information content of the training data, which allows the signal from correct information to dominate statistical noise.

Method: Three key innovations: 1) CXL-based memory disaggregation framework offloading KV-caches to remote FPGA memory, 2) speculative KV-cache prefetching predicting future tokens, and 3) FPGA-accelerated KV-cache compression/decompression engine reducing bandwidth by 4×.

Result: Achieves up to 3.2× higher throughput compared to GPU-only baselines, reduces memory costs by 2.8× while maintaining accuracy, and demonstrates effective memory wall mitigation.

Conclusion: Intelligent memory disaggregation combined with speculative execution can effectively address the memory wall challenge in large-scale LLM serving, with open-source implementation available.

Abstract: Large Language Models (LLMs) have revolutionized natural language processing tasks, but their deployment in datacenter environments faces significant challenges due to the massive memory requirements of key-value (KV) caches. During the autoregressive decoding process, KV caches consume substantial GPU memory, limiting batch sizes and overall system throughput. To address these challenges, we propose \textbf{CXL-SpecKV}, a novel disaggregated KV-cache architecture that leverages Compute Express Link (CXL) interconnects and FPGA accelerators to enable efficient speculative execution and memory disaggregation. Our approach introduces three key innovations: (i) a CXL-based memory disaggregation framework that offloads KV-caches to remote FPGA memory with low latency, (ii) a speculative KV-cache prefetching mechanism that predicts and preloads future tokens’ cache entries, and (iii) an FPGA-accelerated KV-cache compression and decompression engine that reduces memory bandwidth requirements by up to 4$\times$. When evaluated on state-of-the-art LLM models, CXL-SpecKV achieves up to 3.2$\times$ higher throughput compared to GPU-only baselines, while reducing memory costs by 2.8$\times$ and maintaining accuracy. Our system demonstrates that intelligent memory disaggregation combined with speculative execution can effectively address the memory wall challenge in large-scale LLM serving. Our code implementation has been open-sourced at https://github.com/FastLM/CXL-SpecKV.

Method: AGAPI uses an Agent-Planner-Executor-Summarizer architecture to integrate 8+ open-source LLMs with 20+ materials science API endpoints, unifying databases, simulation tools, and ML models.

Result: Demonstrated through end-to-end workflows including heterostructure construction, XRD analysis, and semiconductor defect engineering; evaluated with 30+ test cases; has 1,000+ active users.

Conclusion: AGAPI provides a scalable, transparent foundation for reproducible, AI-accelerated materials discovery, with open-source code available.

Abstract: Artificial intelligence is reshaping scientific discovery, yet its use in materials research remains limited by fragmented computational ecosystems, reproducibility challenges, and dependence on commercial large language models (LLMs). Here we introduce AGAPI (AtomGPT.org API), an open-access agentic AI platform that integrates more than eight open-source LLMs with over twenty materials-science API endpoints, unifying databases, simulation tools, and machine-learning models through a common orchestration framework. AGAPI employs an Agent-Planner-Executor-Summarizer architecture that autonomously constructs and executes multi-step workflows spanning materials data retrieval, graph neural network property prediction, machine-learning force-field optimization, tight-binding calculations, diffraction analysis, and inverse design. We demonstrate AGAPI through end-to-end workflows, including heterostructure construction, powder X-ray diffraction analysis, and semiconductor defect engineering requiring up to ten sequential operations. In addition, we evaluate AGAPI using 30+ example prompts as test cases and compare agentic predictions with and without tool access against experimental data. With more than 1,000 active users, AGAPI provides a scalable and transparent foundation for reproducible, AI-accelerated materials discovery. AGAPI-Agents codebase is available at https://github.com/atomgptlab/agapi.

Method: Introduces a declarative, logic-based domain-specific language for encoding hypergame structures and solution concepts, leveraging answer-set programming to create an automated pipeline for instantiating hypergames and running hypergame rationalisation procedures.

Result: Develops a unifying formalism for hypergames that enables automated reasoning about mismatched mental models and provides a foundation for belief-based heterogeneous reasoners with logical guarantees.

Conclusion: Establishes connection between hypergame theory, multi-agent systems, and strategic AI by providing practical tools for managing complex hypergame structures and equilibria in real-world applications.

Abstract: Differences in perception, information asymmetries, and bounded rationality lead game-theoretic players to derive a private, subjective view of the game that may diverge from the underlying ground-truth scenario and may be misaligned with other players’ interpretations. While typical game-theoretic assumptions often overlook such heterogeneity, hypergame theory provides the mathematical framework to reason about mismatched mental models. Although hypergames have recently gained traction in dynamic applications concerning uncertainty, their practical adoption in multi-agent system research has been hindered by the lack of a unifying, formal, and practical representation language, as well as scalable algorithms for managing complex hypergame structures and equilibria. Our work addresses this gap by introducing a declarative, logic-based domain-specific language for encoding hypergame structures and hypergame solution concepts. Leveraging answer-set programming, we develop an automated pipeline for instantiating hypergame structures and running our novel hypergame rationalisation procedure, a mechanism for finding belief structures that justify seemingly irrational outcomes. The proposed language establishes a unifying formalism for hypergames and serves as a foundation for developing nuanced, belief-based heterogeneous reasoners, offering a verifiable context with logical guarantees. Together, these contributions establish the connection between hypergame theory, multi-agent systems, and strategic AI.

Method: EnrichLog uses retrieval-augmented generation to integrate contextual knowledge (historical examples and corpus reasoning) without retraining, combining both corpus-specific and sample-specific knowledge enrichment.

Result: Evaluation on four large-scale system log benchmarks shows consistent performance improvements over five baseline methods, with effective handling of ambiguous log entries and efficient inference.

Conclusion: EnrichLog’s dual-knowledge enrichment approach enhances model confidence and detection accuracy, making it suitable for practical deployments in distributed system monitoring.

Abstract: System logs are a critical resource for monitoring and managing distributed systems, providing insights into failures and anomalous behavior. Traditional log analysis techniques, including template-based and sequence-driven approaches, often lose important semantic information or struggle with ambiguous log patterns. To address this, we present EnrichLog, a training-free, entry-based anomaly detection framework that enriches raw log entries with both corpus-specific and sample-specific knowledge. EnrichLog incorporates contextual information, including historical examples and reasoning derived from the corpus, to enable more accurate and interpretable anomaly detection. The framework leverages retrieval-augmented generation to integrate relevant contextual knowledge without requiring retraining. We evaluate EnrichLog on four large-scale system log benchmark datasets and compare it against five baseline methods. Our results show that EnrichLog consistently improves anomaly detection performance, effectively handles ambiguous log entries, and maintains efficient inference. Furthermore, incorporating both corpus- and sample-specific knowledge enhances model confidence and detection accuracy, making EnrichLog well-suited for practical deployments.

Method: A context-sensitive multi-agent coordination framework using coordinated Deep Q-Networks integrated with Graph Neural Networks and attention mechanisms. The system processes 20 contextual features and employs weighted coordination mechanisms and consensus protocols to balance five stakeholders with different weightings (25%, 20%, 20%, 20%, 15%).

Result: Achieves 92% coordination success rate, 15% energy efficiency improvement, 10% cost reduction, 20% grid strain decrease, and 2.3x faster convergence with 88% training stability and 85% sample efficiency. Real-world validation shows Net Present Cost of -$122,962 and 69% cost reduction through renewable energy integration, outperforming DDPG, A3C, PPO, and GNN baselines.

Conclusion: The CAMAC-DRA framework successfully develops context-aware multi-stakeholder coordination that balances competing objectives while adapting to real-time variables, representing a breakthrough solution for intelligent EV charging coordination and sustainable transportation electrification with demonstrated commercial viability.

Abstract: This paper presents a novel context-sensitive multi-agent coordination for dynamic resource allocation (CAMAC-DRA) framework for optimizing smart electric vehicle (EV) charging ecosystems through the Smart2Charge application. The proposed system coordinates autonomous charging agents across networks of 250 EVs and 45 charging stations while adapting to dynamic environmental conditions through context-aware decision-making. Our multi-agent approach employs coordinated Deep Q-Networks integrated with Graph Neural Networks and attention mechanisms, processing 20 contextual features including weather patterns, traffic conditions, grid load fluctuations, and electricity pricing.The framework balances five ecosystem stakeholders i.e. EV users (25%), grid operators (20%), charging station operators (20%), fleet operators (20%), and environmental factors (15%) through weighted coordination mechanisms and consensus protocols. Comprehensive validation using real-world datasets containing 441,077 charging transactions demonstrates superior performance compared to baseline algorithms including DDPG, A3C, PPO, and GNN approaches. The CAMAC-DRA framework achieves 92% coordination success rate, 15% energy efficiency improvement, 10% cost reduction, 20% grid strain decrease, and \2.3x faster convergence while maintaining 88% training stability and 85% sample efficiency. Real-world validation confirms commercial viability with Net Present Cost of -$122,962 and 69% cost reduction through renewable energy integration. The framework’s unique contribution lies in developing context-aware multi-stakeholder coordination that successfully balances competing objectives while adapting to real-time variables, positioning it as a breakthrough solution for intelligent EV charging coordination and sustainable transportation electrification.

Method: A multi-agent simulation framework inspired by social deduction games where a minority of agents (traitors) try to mislead the majority (faithful). The framework is grounded in game theory, behavioral economics, and social cognition, with evaluation metrics for deception success, trust dynamics, and collective inference quality. Features include fully autonomous simulation with persistent memory, evolving social dynamics, heterogeneous agent populations, specialized traits, and adaptive behaviors.

Result: Experiments with DeepSeek-V3, GPT-4o-mini, and GPT-4o (10 runs per model) reveal an asymmetry: advanced models like GPT-4o demonstrate superior deceptive capabilities but exhibit disproportionate vulnerability to others’ falsehoods, suggesting deception skills may scale faster than detection abilities.

Conclusion: The Traitors provides a focused, configurable testbed for investigating LLM behavior in socially nuanced interactions, contributing to more rigorous research on deception mechanisms, alignment challenges, and the broader social reliability of AI systems.

Abstract: As AI systems increasingly assume roles where trust and alignment with human values are essential, understanding when and why they engage in deception has become a critical research priority. We introduce The Traitors, a multi-agent simulation framework inspired by social deduction games, designed to probe deception, trust formation, and strategic communication among large language model (LLM) agents under asymmetric information. A minority of agents the traitors seek to mislead the majority, while the faithful must infer hidden identities through dialogue and reasoning. Our contributions are: (1) we ground the environment in formal frameworks from game theory, behavioral economics, and social cognition; (2) we develop a suite of evaluation metrics capturing deception success, trust dynamics, and collective inference quality; (3) we implement a fully autonomous simulation platform where LLMs reason over persistent memory and evolving social dynamics, with support for heterogeneous agent populations, specialized traits, and adaptive behaviors. Our initial experiments across DeepSeek-V3, GPT-4o-mini, and GPT-4o (10 runs per model) reveal a notable asymmetry: advanced models like GPT-4o demonstrate superior deceptive capabilities yet exhibit disproportionate vulnerability to others’ falsehoods. This suggests deception skills may scale faster than detection abilities. Overall, The Traitors provides a focused, configurable testbed for investigating LLM behavior in socially nuanced interactions. We position this work as a contribution toward more rigorous research on deception mechanisms, alignment challenges, and the broader social reliability of AI systems.

Method: Systematically evaluate LLMs on three key questions: (1) detecting unreasonable forecasts, (2) incorporating unstructured exogenous features, (3) performance across model sizes. Conduct three experiments on synthetic and real-world data (M5 dataset).

Result: LLMs reliably detect forecast errors (temporal misalignment, trend inconsistencies, spikes) with best F1 score of 0.88. Multi-modal LLMs effectively use contextual signals (F1: 0.84). Unreasonable forecasts on M5 have 10% higher sCRPS than reasonable ones.

Conclusion: LLMs without domain-specific fine-tuning provide viable, scalable option for automated forecast monitoring, though slightly below human performance (F1: 0.97 vs 0.88).

Abstract: Monitoring forecasting systems is critical for customer satisfaction, profitability, and operational efficiency in large-scale retail businesses. We propose The Forecast Critic, a system that leverages Large Language Models (LLMs) for automated forecast monitoring, taking advantage of their broad world knowledge and strong ``reasoning’’ capabilities. As a prerequisite for this, we systematically evaluate the ability of LLMs to assess time series forecast quality, focusing on three key questions. (1) Can LLMs be deployed to perform forecast monitoring and identify obviously unreasonable forecasts? (2) Can LLMs effectively incorporate unstructured exogenous features to assess what a reasonable forecast looks like? (3) How does performance vary across model sizes and reasoning capabilities, measured across state-of-the-art LLMs? We present three experiments, including on both synthetic and real-world forecasting data. Our results show that LLMs can reliably detect and critique poor forecasts, such as those plagued by temporal misalignment, trend inconsistencies, and spike errors. The best-performing model we evaluated achieves an F1 score of 0.88, somewhat below human-level performance (F1 score: 0.97). We also demonstrate that multi-modal LLMs can effectively incorporate unstructured contextual signals to refine their assessment of the forecast. Models correctly identify missing or spurious promotional spikes when provided with historical context about past promotions (F1 score: 0.84). Lastly, we demonstrate that these techniques succeed in identifying inaccurate forecasts on the real-world M5 time series dataset, with unreasonable forecasts having an sCRPS at least 10% higher than that of reasonable forecasts. These findings suggest that LLMs, even without domain-specific fine-tuning, may provide a viable and scalable option for automated forecast monitoring and evaluation.

Method: Evaluated RPI’s robustness by testing it on two classical control tasks (CartPole and Inverted Pendulum) under variations in neural network and environmental parameters. Compared RPI against several established deep RL methods: DQN, Double DQN, DDPG, TD3, and PPO.

Result: RPI reaches near-optimal performance early in training and sustains this policy throughout training. It outperforms all compared methods (DQN, Double DQN, DDPG, TD3, PPO) in terms of robustness and reliability.

Conclusion: RPI demonstrates promise as a more reliable alternative to existing deep RL methods, addressing key limitations like sample inefficiency, training instability, and hyperparameter sensitivity through its restoration of policy iteration’s monotonicity property.

Abstract: In a recent work, we proposed Reliable Policy Iteration (RPI), that restores policy iteration’s monotonicity-of-value-estimates property to the function approximation setting. Here, we assess the robustness of RPI’s empirical performance on two classical control tasks – CartPole and Inverted Pendulum – under changes to neural network and environmental parameters. Relative to DQN, Double DQN, DDPG, TD3, and PPO, RPI reaches near-optimal performance early and sustains this policy as training proceeds. Because deep RL methods are often hampered by sample inefficiency, training instability, and hyperparameter sensitivity, our results highlight RPI’s promise as a more reliable alternative.

Method: Proposes TopK-SD: a label propagation framework that synthesizes data using both semantic and label information to select demonstrations with consistent labels, reducing propagation error bounds.

Result: TopK-SD outperforms original TopK sampling on multiple benchmarks, demonstrating the importance of label consistency in ICL demonstration selection.

Conclusion: The work provides a new transductive learning perspective on ICL, showing that label-consistent demonstrations are essential for effective in-context learning and better understanding ICL mechanisms.

Abstract: Large language models (LLMs) perform in-context learning (ICL) with minimal supervised examples, which benefits various natural language processing (NLP) tasks. One of the critical research focus is the selection of prompt demonstrations. Current approaches typically employ retrieval models to select the top-K most semantically similar examples as demonstrations. However, we argue that existing methods are limited since the label consistency is not guaranteed during demonstration selection. Our cognition derives from the Bayesian view of ICL and our rethinking of ICL from the transductive label propagation perspective. We treat ICL as a transductive learning method and incorporate latent concepts from Bayesian view and deduce that similar demonstrations guide the concepts of query, with consistent labels serving as estimates. Based on this understanding, we establish a label propagation framework to link label consistency with propagation error bounds. To model label consistency, we propose a data synthesis method, leveraging both semantic and label information, and use TopK sampling with Synthetic Data (TopK-SD) to acquire demonstrations with consistent labels. TopK-SD outperforms original TopK sampling on multiple benchmarks. Our work provides a new perspective for understanding the working mechanisms within ICL.

Method: Uses large language models (LLMs) to extract spatial information from architectural layouts, transforming floor plans into navigable knowledge graphs and generating human-readable navigation instructions with few-shot learning.

Result: Claude 3.7 Sonnet achieved highest accuracy: 92.31% (short), 76.92% (medium), 61.54% (long routes) with 5-shot prompting. Graph-based spatial structure performed 15.4% better than direct visual reasoning across all models.

Conclusion: Graphical representation and in-context learning enhance navigation performance, making the solution more precise for Blind and Low Vision users’ indoor navigation needs.

Abstract: Indoor navigation remains a critical challenge for people with visual impairments. The current solutions mainly rely on infrastructure-based systems, which limit their ability to navigate safely in dynamic environments. We propose a novel navigation approach that utilizes a foundation model to transform floor plans into navigable knowledge graphs and generate human-readable navigation instructions. Floorplan2Guide integrates a large language model (LLM) to extract spatial information from architectural layouts, reducing the manual preprocessing required by earlier floorplan parsing methods. Experimental results indicate that few-shot learning improves navigation accuracy in comparison to zero-shot learning on simulated and real-world evaluations. Claude 3.7 Sonnet achieves the highest accuracy among the evaluated models, with 92.31%, 76.92%, and 61.54% on the short, medium, and long routes, respectively, under 5-shot prompting of the MP-1 floor plan. The success rate of graph-based spatial structure is 15.4% higher than that of direct visual reasoning among all models, which confirms that graphical representation and in-context learning enhance navigation performance and make our solution more precise for indoor navigation of Blind and Low Vision (BLV) users.

Method: Uses generative representation perspective with two-stage attention triple enhancer integrated with U-KAN based diffusion model for few-shot KG completion.

Result: Extensive experiments on two public datasets show the method achieves new state-of-the-art results.

Conclusion: The proposed generative approach with attention-enhanced diffusion model effectively addresses few-shot KG completion by better exploiting neighborhood information and contrastive signal distributions.

Abstract: Knowledge Graphs (KGs), thanks to their concise and efficient triple-based structure, have been widely applied in intelligent question answering, recommender systems and other domains. However, the heterogeneous and multifaceted nature of real-world data inevitably renders the distribution of relations long-tailed, making it crucial to complete missing facts with limited samples. Previous studies mainly based on metric matching or meta learning, yet they either fail to fully exploit neighborhood information in graph or overlook the distributional characteristics of contrastive signals. In this paper, we re-examine the problem from a perspective of generative representation and propose a few-shot knowledge graph completion framework that integrates two-stage attention triple enhancer with U-KAN based diffusion model. Extensive experiments on two public datasets show that our method achieve new state-of-the-art results.

Method: Represent cognitive state as a point on a differentiable manifold with learned Riemannian metric encoding representational constraints, costs, and structural relations. Define scalar cognitive potential combining accuracy, parsimony, utility, and normative requirements. Cognition unfolds as Riemannian gradient flow of this potential.

Result: Dual-process effects (intuitive vs. deliberative reasoning) emerge naturally from metric-induced anisotropies creating time-scale separations and geometric phase transitions, without needing modular architectures. Analytical conditions derived and behavioral signatures demonstrated through simulations.

Conclusion: Establishes a geometric foundation for cognition and suggests principles for developing more general, human-like AI systems through unified mathematical framework.

Abstract: Human cognition spans perception, memory, intuitive judgment, deliberative reasoning, action selection, and social inference, yet these capacities are often explained through distinct computational theories. Here we present a unified mathematical framework in which diverse cognitive processes emerge from a single geometric principle. We represent the cognitive state as a point on a differentiable manifold endowed with a learned Riemannian metric that encodes representational constraints, computational costs, and structural relations among cognitive variables. A scalar cognitive potential combines predictive accuracy, structural parsimony, task utility, and normative or logical requirements. Cognition unfolds as the Riemannian gradient flow of this potential, providing a universal dynamical law from which a broad range of psychological phenomena arise. Classical dual-process effects–rapid intuitive responses and slower deliberative reasoning–emerge naturally from metric-induced anisotropies that generate intrinsic time-scale separations and geometric phase transitions, without invoking modular or hybrid architectures. We derive analytical conditions for these regimes and demonstrate their behavioural signatures through simulations of canonical cognitive tasks. Together, these results establish a geometric foundation for cognition and suggest guiding principles for the development of more general and human-like artificial intelligence systems.

Method: Analyzes Wikidata’s polyhierarchical and multi-axial design architecture, examining how it accommodates multiple classification axes simultaneously under a shared root class rather than enforcing disjoint top-level distinctions.

Result: Wikidata’s architecture enables a scalable and modular approach to ontology construction that is particularly well-suited for collaborative and evolving knowledge graphs, offering more flexibility than traditional ontology designs.

Conclusion: Wikidata’s polyhierarchical, multi-axial design represents a significant departure from traditional ontology approaches, providing a more adaptable framework for large-scale, collaborative knowledge representation that can evolve over time.

Abstract: Traditional ontology design emphasizes disjoint and exhaustive top-level distinctions such as continuant vs. occurrent, abstract vs. concrete, or type vs. instance. These distinctions are used to structure unified hierarchies where every entity is classified under a single upper-level category. Wikidata, by contrast, does not enforce a singular foundational taxonomy. Instead, it accommodates multiple classification axes simultaneously under the shared root class entity. This paper analyzes the structural implications of Wikidata’s polyhierarchical and multi-axial design. The Wikidata architecture enables a scalable and modular approach to ontology construction, especially suited to collaborative and evolving knowledge graphs.

Method: Uses diffusion-based generator conditioned on quantum-mechanical descriptors with equivariant neural network validator trained on hierarchical multi-theory dataset (PBE, SCAN, HSE06, CCSD(T)). Implements active learning loop to quantify and target divergence between low- and high-fidelity predictions.

Result: 3-5x improvement in successfully identifying potentially stable candidates in high-divergence regions (e.g., correlated oxides) compared to DFT-only baselines, while maintaining computational feasibility.

Conclusion: Provides a rigorous, transparent framework for extending the effective search space of computational materials discovery beyond the limitations of single-fidelity models.

Abstract: Conventional generative models for materials discovery are predominantly trained and validated using data from Density Functional Theory (DFT) with approximate exchange-correlation functionals. This creates a fundamental bottleneck: these models inherit DFT’s systematic failures for strongly correlated systems, leading to exploration biases and an inability to discover materials where DFT predictions are qualitatively incorrect. We introduce a quantum-aware generative AI framework that systematically addresses this limitation through tight integration of multi-fidelity learning and active validation. Our approach employs a diffusion-based generator conditioned on quantum-mechanical descriptors and a validator using an equivariant neural network potential trained on a hierarchical dataset spanning multiple levels of theory (PBE, SCAN, HSE06, CCSD(T)). Crucially, we implement a robust active learning loop that quantifies and targets the divergence between low- and high-fidelity predictions. We conduct comprehensive ablation studies to deconstruct the contribution of each component, perform detailed failure mode analysis, and benchmark our framework against state-of-the-art generative models (CDVAE, GNoME, DiffCSP) across several challenging material classes. Our results demonstrate significant practical gains: a 3-5x improvement in successfully identifying potentially stable candidates in high-divergence regions (e.g., correlated oxides) compared to DFT-only baselines, while maintaining computational feasibility. This work provides a rigorous, transparent framework for extending the effective search space of computational materials discovery beyond the limitations of single-fidelity models.

Method: The authors identify “entropy collapse” as a universal dynamical failure mode, formalizing it mathematically as convergence toward a stable low-entropy manifold. They use minimal domain-agnostic assumptions to analytically establish critical thresholds, dynamical irreversibility, and attractor structure, and demonstrate universality through minimal simulations.

Result: The analysis shows intelligent systems undergo a sharp phase transition from high-entropy adaptive regimes to low-entropy collapsed regimes. This framework unifies diverse phenomena (AI model collapse, economic institutional sclerosis, evolutionary genetic bottlenecks) as manifestations of the same underlying process.

Conclusion: Collapse is a structural cost of intelligence itself, explaining why late-stage interventions fail. The findings motivate entropy-aware design principles for sustaining long-term adaptability in intelligent systems by preventing entropy collapse.

Abstract: Intelligent systems are widely assumed to improve through learning, coordination, and optimization. However, across domains – from artificial intelligence to economic institutions and biological evolution – increasing intelligence often precipitates paradoxical degradation: systems become rigid, lose adaptability, and fail unexpectedly. We identify \emph{entropy collapse} as a universal dynamical failure mode arising when feedback amplification outpaces bounded novelty regeneration. Under minimal domain-agnostic assumptions, we show that intelligent systems undergo a sharp transition from high-entropy adaptive regimes to low-entropy collapsed regimes. Collapse is formalized as convergence toward a stable low-entropy manifold, not a zero-entropy state, implying a contraction of effective adaptive dimensionality rather than loss of activity or scale. We analytically establish critical thresholds, dynamical irreversibility, and attractor structure and demonstrate universality across update mechanisms through minimal simulations. This framework unifies diverse phenomena – model collapse in AI, institutional sclerosis in economics, and genetic bottlenecks in evolution – as manifestations of the same underlying process. By reframing collapse as a structural cost of intelligence, our results clarify why late-stage interventions systematically fail and motivate entropy-aware design principles for sustaining long-term adaptability in intelligent systems. \noindent\textbf{Keywords:} entropy collapse; intelligent systems; feedback amplification; phase transitions; effective dimensionality; complex systems; model collapse; institutional sclerosis

Method: 1) Reproduced Anthropic’s multi-turn “emergent introspection” on Meta-Llama-3.1-8B-Instruct; 2) Systematically varied inference prompts to test robustness; 3) Tested multiple-choice identification and binary discrimination tasks; 4) Examined classification of concept vector strength.

Result: 1) Successfully reproduced Anthropic’s 20% success rate for concept identification; 2) Found introspection is fragile - performance collapses on related tasks; 3) Discovered partial introspection regime with 70% accuracy for classifying concept vector strength (vs 25% baseline).

Conclusion: While models can compute functions of their internal representations during introspection, these self-reports are narrow and highly prompt-sensitive, supporting but qualifying Anthropic’s original claims.

Abstract: Recent work from Anthropic claims that frontier models can sometimes detect and name injected “concepts” represented as activation directions. We test the robustness of these claims. First, we reproduce Anthropic’s multi-turn “emergent introspection” result on Meta-Llama-3.1-8B-Instruct, finding that the model identifies and names the injected concept 20 percent of the time under Anthropic’s original pipeline, exactly matching their reported numbers and thus showing that introspection is not exclusive to very large or capable models. Second, we systematically vary the inference prompt and find that introspection is fragile: performance collapses on closely related tasks such as multiple-choice identification of the injected concept or different prompts of binary discrimination of whether a concept was injected at all. Third, we identify a contrasting regime of partial introspection: the same model can reliably classify the strength of the coefficient of a normalized injected concept vector (as weak / moderate / strong / very strong) with up to 70 percent accuracy, far above the 25 percent chance baseline. Together, these results provide more evidence for Anthropic’s claim that language models effectively compute a function of their baseline, internal representations during introspection; however, these self-reports about those representations are narrow and prompt-sensitive. Our code is available at https://github.com/elyhahami18/CS2881-Introspection.

Method: Six studies (N=1365) developed and validated the Critical Thinking in AI Use Scale. Studies included item generation, factor analysis, reliability testing, convergent/discriminant validity assessment, and criterion validation using a ChatGPT-powered fact-checking task.

Result: The scale has a three-factor structure (Verification, Motivation, Reflection) with strong psychometric properties. Higher scores predict more verification strategies, better accuracy in fact-checking tasks, and deeper reflection about responsible AI. Associated with openness, extraversion, positive affect, and frequent AI use.

Conclusion: The research provides a validated scale and ecological task paradigm to study critical engagement with AI outputs, clarifying how people exercise oversight over generative AI and supporting future theory testing and cross-group research.

Abstract: Generative AI tools are increasingly embedded in everyday work and learning, yet their fluency, opacity, and propensity to hallucinate mean that users must critically evaluate AI outputs rather than accept them at face value. The present research conceptualises critical thinking in AI use as a dispositional tendency to verify the source and content of AI-generated information, to understand how models work and where they fail, and to reflect on the broader implications of relying on AI. Across six studies (N = 1365), we developed and validated the 13-item critical thinking in AI use scale and mapped its nomological network. Study 1 generated and content-validated scale items. Study 2 supported a three-factor structure (Verification, Motivation, and Reflection). Studies 3, 4, and 5 confirmed this higher-order model, demonstrated internal consistency and test-retest reliability, strong factor loadings, sex invariance, and convergent and discriminant validity. Studies 3 and 4 further revealed that critical thinking in AI use was positively associated with openness, extraversion, positive trait affect, and frequency of AI use. Lastly, Study 6 demonstrated criterion validity of the scale, with higher critical thinking in AI use scores predicting more frequent and diverse verification strategies, greater veracity-judgement accuracy in a novel and naturalistic ChatGPT-powered fact-checking task, and deeper reflection about responsible AI. Taken together, the current work clarifies why and how people exercise oversight over generative AI outputs and provides a validated scale and ecologically grounded task paradigm to support theory testing, cross-group, and longitudinal research on critical engagement with generative AI outputs.

Method: Analyzed documentation from 5 frontier models and 100 Hugging Face model cards, identified 947 unique section names, then developed a weighted transparency framework with 8 sections/23 subsections based on EU AI Act and Stanford Transparency Index. Implemented automated multi-agent pipeline using LLM-based consensus to extract and score documentation completeness.

Result: Evaluation of 50 models across vision, multimodal, open-source, and closed-source systems cost <$3 total. Frontier labs achieve ~80% compliance while most providers fall below 60%. Safety-critical categories show largest deficits: deception behaviors (148 points lost), hallucinations (124 points), and child safety evaluations (116 points).

Conclusion: Current AI documentation practices are highly inconsistent with systematic gaps in safety-critical disclosures, highlighting the need for standardized documentation frameworks and automated assessment tools to improve transparency and accountability.

Abstract: AI model documentation is fragmented across platforms and inconsistent in structure, preventing policymakers, auditors, and users from reliably assessing safety claims, data provenance, and version-level changes. We analyzed documentation from five frontier models (Gemini 3, Grok 4.1, Llama 4, GPT-5, and Claude 4.5) and 100 Hugging Face model cards, identifying 947 unique section names with extreme naming variation. Usage information alone appeared under 97 distinct labels. Using the EU AI Act Annex IV and the Stanford Transparency Index as baselines, we developed a weighted transparency framework with 8 sections and 23 subsections that prioritizes safety-critical disclosures (Safety Evaluation: 25%, Critical Risk: 20%) over technical specifications. We implemented an automated multi-agent pipeline that extracts documentation from public sources and scores completeness through LLM-based consensus. Evaluating 50 models across vision, multimodal, open-source, and closed-source systems cost less than $3 in total and revealed systematic gaps. Frontier labs (xAI, Microsoft, Anthropic) achieve approximately 80% compliance, while most providers fall below 60%. Safety-critical categories show the largest deficits: deception behaviors, hallucinations, and child safety evaluations account for 148, 124, and 116 aggregate points lost, respectively, across all evaluated models.

Method: Proposes MetaHGNIE with three main components: 1) constructs higher-order knowledge graph via meta-path sequences with typed hyperedges capturing multi-entity relational contexts, 2) aggregates structural dependencies with local attention and encodes semantic representations through hypergraph transformer with sparse chunking to reduce redundancy, 3) multimodal fusion module integrates structural and semantic embeddings under contrastive learning with auxiliary supervision for cross-modal alignment.

Result: Extensive experiments on benchmark NIE datasets demonstrate that MetaHGNIE consistently outperforms state-of-the-art baselines.

Conclusion: The results highlight the effectiveness of explicitly modeling higher-order interactions and cross-modal alignment in heterogeneous knowledge graphs for node importance estimation.

Abstract: Node importance estimation (NIE) in heterogeneous knowledge graphs is a critical yet challenging task, essential for applications such as recommendation, knowledge reasoning, and question answering. Existing methods often rely on pairwise connections, neglecting high-order dependencies among multiple entities and relations, and they treat structural and semantic signals independently, hindering effective cross-modal integration. To address these challenges, we propose MetaHGNIE, a meta-path induced hypergraph contrastive learning framework for disentangling and aligning structural and semantic information. MetaHGNIE constructs a higher-order knowledge graph via meta-path sequences, where typed hyperedges capture multi-entity relational contexts. Structural dependencies are aggregated with local attention, while semantic representations are encoded through a hypergraph transformer equipped with sparse chunking to reduce redundancy. Finally, a multimodal fusion module integrates structural and semantic embeddings under contrastive learning with auxiliary supervision, ensuring robust cross-modal alignment. Extensive experiments on benchmark NIE datasets demonstrate that MetaHGNIE consistently outperforms state-of-the-art baselines. These results highlight the effectiveness of explicitly modeling higher-order interactions and cross-modal alignment in heterogeneous knowledge graphs. Our code is available at https://github.com/SEU-WENJIA/DualHNIE

Method: SafeGen combines two components: BGE-M3 (fine-tuned text classifier for filtering harmful prompts) and Hyper-SD (optimized diffusion model for high-fidelity image generation). Built on curated multilingual datasets with fairness-aware training.

Result: Hyper-SD achieved IS=3.52, FID=22.08, SSIM=0.79; BGE-M3 reached F1-Score=0.81. Ablation studies validated domain-specific fine-tuning importance. Case studies showed practical impact in blocking unsafe prompts and generating inclusive materials.

Conclusion: SafeGen demonstrates that creative freedom and ethical responsibility can be reconciled within a single workflow, providing a framework for trustworthy AI in text-to-image generation.

Abstract: Generative Artificial Intelligence (AI) has created unprecedented opportunities for creative expression, education, and research. Text-to-image systems such as DALL.E, Stable Diffusion, and Midjourney can now convert ideas into visuals within seconds, but they also present a dual-use dilemma, raising critical ethical concerns: amplifying societal biases, producing high-fidelity disinformation, and violating intellectual property. This paper introduces SafeGen, a framework that embeds ethical safeguards directly into the text-to-image generation pipeline, grounding its design in established principles for Trustworthy AI. SafeGen integrates two complementary components: BGE-M3, a fine-tuned text classifier that filters harmful or misleading prompts, and Hyper-SD, an optimized diffusion model that produces high fidelity, semantically aligned images. Built on a curated multilingual (English- Vietnamese) dataset and a fairness-aware training process, SafeGen demonstrates that creative freedom and ethical responsibility can be reconciled within a single workflow. Quantitative evaluations confirm its effectiveness, with Hyper-SD achieving IS = 3.52, FID = 22.08, and SSIM = 0.79, while BGE-M3 reaches an F1-Score of 0.81. An ablation study further validates the importance of domain-specific fine-tuning for both modules. Case studies illustrate SafeGen’s practical impact in blocking unsafe prompts, generating inclusive teaching materials, and reinforcing academic integrity.

Method: Created KidsArtBench with expert educator annotations across 9 rubric dimensions; proposed attribute-specific multi-LoRA approach with Regression-Aware Fine-Tuning (RAFT) to align predictions with ordinal scales.

Result: On Qwen2.5-VL-7B, correlation increased from 0.468 to 0.653, with largest gains on perceptual dimensions and narrowed gaps on higher-order attributes.

Conclusion: Educator-aligned supervision and attribute-aware training enable pedagogically meaningful evaluations, establishing a rigorous testbed for educational AI progress in art assessment.

Abstract: Multimodal Large Language Models (MLLMs) show remarkable progress across many visual-language tasks; however, their capacity to evaluate artistic expression remains limited. Aesthetic concepts are inherently abstract and open-ended, and multimodal artwork annotations are scarce. We introduce KidsArtBench, a new benchmark of over 1k children’s artworks (ages 5-15) annotated by 12 expert educators across 9 rubric-aligned dimensions, together with expert comments for feedback. Unlike prior aesthetic datasets that provide single scalar scores on adult imagery, KidsArtBench targets children’s artwork and pairs multi-dimensional annotations with comment supervision to enable both ordinal assessment and formative feedback. Building on this resource, we propose an attribute-specific multi-LoRA approach, where each attribute corresponds to a distinct evaluation dimension (e.g., Realism, Imagination) in the scoring rubric, with Regression-Aware Fine-Tuning (RAFT) to align predictions with ordinal scales. On Qwen2.5-VL-7B, our method increases correlation from 0.468 to 0.653, with the largest gains on perceptual dimensions and narrowed gaps on higher-order attributes. These results show that educator-aligned supervision and attribute-aware training yield pedagogically meaningful evaluations and establish a rigorous testbed for sustained progress in educational AI. We release data and code with ethics documentation.

Method: Used model-based reinforcement learning agents that acquire parsimonious predictive representations of their environment (learned world models) to study patch-foraging behavior.

Result: Model-based agents naturally converge to MVT-aligned strategies, exhibit anticipatory capabilities driving efficient patch-leaving behavior, and show decision patterns similar to biological foragers, unlike standard model-free RL agents.

Conclusion: Predictive world models can serve as a foundation for explainable, biologically-grounded AI decision-making, and ecological optimality principles like MVT are valuable for advancing interpretable and adaptive AI systems.

Abstract: Patch foraging involves the deliberate and planned process of determining the optimal time to depart from a resource-rich region and investigate potentially more beneficial alternatives. The Marginal Value Theorem (MVT) is frequently used to characterize this process, offering an optimality model for such foraging behaviors. Although this model has been widely used to make predictions in behavioral ecology, discovering the computational mechanisms that facilitate the emergence of optimal patch-foraging decisions in biological foragers remains under investigation. Here, we show that artificial foragers equipped with learned world models naturally converge to MVT-aligned strategies. Using a model-based reinforcement learning agent that acquires a parsimonious predictive representation of its environment, we demonstrate that anticipatory capabilities, rather than reward maximization alone, drive efficient patch-leaving behavior. Compared with standard model-free RL agents, these model-based agents exhibit decision patterns similar to many of their biological counterparts, suggesting that predictive world models can serve as a foundation for more explainable and biologically grounded decision-making in AI systems. Overall, our findings highlight the value of ecological optimality principles for advancing interpretable and adaptive AI.

Method: Dynamic, multi-round experiments with GPT-4, GPT-4o, and LLaMA-8B testing for five established decision biases using the canonical newsvendor problem in a dynamic setting.

Result: LLMs consistently replicated the classic “Too Low/Too High” ordering bias and significantly amplified other tendencies like demand-chasing behavior compared to human benchmarks. GPT-4 showed greatest irrationality through overthinking (paradox of intelligence), while GPT-4o performed near-optimally. Biases persist even when optimal formulas are provided, indicating architectural constraints rather than knowledge gaps.

Conclusion: Managers should select models based on specific tasks, implement robust human-in-the-loop oversight for high-stakes decisions, and use structured, rule-based prompts to constrain heuristic tendencies and improve AI-assisted decision reliability.

Abstract: Problem definition: Although large language models (LLMs) are increasingly integrated into business decision making, their potential to replicate and even amplify human cognitive biases cautions a significant, yet not well-understood, risk. This is particularly critical in high-stakes operational contexts like supply chain management. To address this, we investigate the decision-making patterns of leading LLMs using the canonical newsvendor problem in a dynamic setting, aiming to identify the nature and origins of their cognitive biases. Methodology/results: Through dynamic, multi-round experiments with GPT-4, GPT-4o, and LLaMA-8B, we tested for five established decision biases. We found that LLMs consistently replicated the classic Too Low/Too High'' ordering bias and significantly amplified other tendencies like demand-chasing behavior compared to human benchmarks. Our analysis uncovered a paradox of intelligence’’: the more sophisticated GPT-4 demonstrated the greatest irrationality through overthinking, while the efficiency-optimized GPT-4o performed near-optimally. Because these biases persist even when optimal formulas are provided, we conclude they stem from architectural constraints rather than knowledge gaps. Managerial implications: First, managers should select models based on the specific task, as our results show that efficiency-optimized models can outperform more complex ones on certain optimization problems. Second, the significant amplification of bias by LLMs highlights the urgent need for robust human-in-the-loop oversight in high-stakes decisions to prevent costly errors. Third, our findings suggest that designing structured, rule-based prompts is a practical and effective strategy for managers to constrain models’ heuristic tendencies and improve the reliability of AI-assisted decisions.

Method: AgentSHAP uses Monte Carlo Shapley values to treat the agent as a black box, testing how it responds with different tool subsets and computing fair importance scores based on game theory, reducing computational cost from O(2^n) to practical levels.

Result: Comprehensive experiments on API-Bank show that AgentSHAP produces consistent scores across runs, correctly identifies which tools matter, and distinguishes relevant from irrelevant tools.

Conclusion: AgentSHAP completes a family of Shapley-based XAI tools for modern generative AI, joining TokenSHAP (for tokens) and PixelSHAP (for image regions), providing the first explainability method for agent tool attribution.

Abstract: LLM agents that use external tools can solve complex tasks, but understanding which tools actually contributed to a response remains a blind spot. No existing XAI methods address tool-level explanations. We introduce AgentSHAP, the first framework for explaining tool importance in LLM agents. AgentSHAP is model-agnostic: it treats the agent as a black box and works with any LLM (GPT, Claude, Llama, etc.) without needing access to internal weights or gradients. Using Monte Carlo Shapley values, AgentSHAP tests how an agent responds with different tool subsets and computes fair importance scores based on game theory. Our contributions are: (1) the first explainability method for agent tool attribution, grounded in Shapley values from game theory; (2) Monte Carlo sampling that reduces cost from O(2n) to practical levels; and (3) comprehensive experiments on API-Bank showing that AgentSHAP produces consistent scores across runs, correctly identifies which tools matter, and distinguishes relevant from irrelevant tools. AgentSHAP joins TokenSHAP (for tokens) and PixelSHAP (for image regions) to complete a family of Shapley-based XAI tools for modern generative AI. Code: https://github.com/GenAISHAP/TokenSHAP.

Method: MobiBench is presented as the first modular and multi-path aware offline benchmarking framework for mobile GUI agents. It enables high-fidelity, scalable, and reproducible evaluation entirely in offline settings by addressing both limitations of current approaches.

Result: MobiBench achieves 94.72% agreement with human evaluators, on par with carefully engineered online benchmarks, while preserving the scalability and reproducibility of static offline benchmarks. Module-level analysis reveals key insights including systematic evaluation of diverse techniques, optimal module configurations across model scales, limitations of current LFMs, and actionable guidelines for designing more capable and cost-efficient mobile agents.

Conclusion: MobiBench successfully addresses the fundamental limitations of current mobile GUI agent evaluation practices by providing a modular, multi-path aware offline benchmarking framework that combines the fidelity of online evaluation with the scalability and reproducibility of offline methods, enabling more comprehensive and fair assessment of mobile GUI agents.

Abstract: Mobile GUI Agents, AI agents capable of interacting with mobile applications on behalf of users, have the potential to transform human computer interaction. However, current evaluation practices for GUI agents face two fundamental limitations. First, they either rely on single path offline benchmarks or online live benchmarks. Offline benchmarks using static, single path annotated datasets unfairly penalize valid alternative actions, while online benchmarks suffer from poor scalability and reproducibility due to the dynamic and unpredictable nature of live evaluation. Second, existing benchmarks treat agents as monolithic black boxes, overlooking the contributions of individual components, which often leads to unfair comparisons or obscures key performance bottlenecks. To address these limitations, we present MobiBench, the first modular and multi path aware offline benchmarking framework for mobile GUI agents that enables high fidelity, scalable, and reproducible evaluation entirely in offline settings. Our experiments demonstrate that MobiBench achieves 94.72 percent agreement with human evaluators, on par with carefully engineered online benchmarks, while preserving the scalability and reproducibility of static offline benchmarks. Furthermore, our comprehensive module level analysis uncovers several key insights, including a systematic evaluation of diverse techniques used in mobile GUI agents, optimal module configurations across model scales, the inherent limitations of current LFMs, and actionable guidelines for designing more capable and cost efficient mobile agents.

Method: Memoria uses a hybrid architecture with two components: dynamic session-level summarization for short-term coherence, and a weighted knowledge graph-based user modeling engine that incrementally captures user traits, preferences, and behavioral patterns as structured entities and relationships.

Result: The framework enables both short-term dialogue coherence and long-term personalization while operating within token constraints of modern LLMs, bridging the gap between stateless LLM interfaces and agentic memory systems.

Conclusion: Memoria offers a practical solution for industry applications requiring adaptive and evolving user experiences, enabling scalable, personalized conversational AI with persistent, interpretable memory.

Abstract: Agentic memory is emerging as a key enabler for large language models (LLM) to maintain continuity, personalization, and long-term context in extended user interactions, critical capabilities for deploying LLMs as truly interactive and adaptive agents. Agentic memory refers to the memory that provides an LLM with agent-like persistence: the ability to retain and act upon information across conversations, similar to how a human would. We present Memoria, a modular memory framework that augments LLM-based conversational systems with persistent, interpretable, and context-rich memory. Memoria integrates two complementary components: dynamic session-level summarization and a weighted knowledge graph (KG)-based user modelling engine that incrementally captures user traits, preferences, and behavioral patterns as structured entities and relationships. This hybrid architecture enables both short-term dialogue coherence and long-term personalization while operating within the token constraints of modern LLMs. We demonstrate how Memoria enables scalable, personalized conversational artificial intelligence (AI) by bridging the gap between stateless LLM interfaces and agentic memory systems, offering a practical solution for industry applications requiring adaptive and evolving user experiences.

Method: WebOperator incorporates: 1) best-first search ranking actions by reward estimates and safety considerations, 2) robust backtracking that verifies path feasibility before replaying, 3) generation of action candidates from multiple reasoning contexts for diverse exploration, and 4) curation of high-quality action sets by filtering invalid actions and merging semantically equivalent ones.

Result: WebOperator achieves state-of-the-art 54.6% success rate on WebArena with GPT-4o, demonstrating significant improvement over existing approaches. Results on WebVoyager also show effectiveness.

Conclusion: WebOperator successfully addresses the limitations of current LLM-based web agents by integrating strategic foresight with safe execution through reliable backtracking and strategic exploration mechanisms, proving critical for effective operation in partially observable web environments.

Abstract: LLM-based agents often operate in a greedy, step-by-step manner, selecting actions solely based on the current observation without considering long-term consequences or alternative paths. This lack of foresight is particularly problematic in web environments, which are only partially observable-limited to browser-visible content (e.g., DOM and UI elements)-where a single misstep often requires complex and brittle navigation to undo. Without an explicit backtracking mechanism, agents struggle to correct errors or systematically explore alternative paths. Tree-search methods provide a principled framework for such structured exploration, but existing approaches lack mechanisms for safe backtracking, making them prone to unintended side effects. They also assume that all actions are reversible, ignoring the presence of irreversible actions-limitations that reduce their effectiveness in realistic web tasks. To address these challenges, we introduce WebOperator, a tree-search framework that enables reliable backtracking and strategic exploration. Our method incorporates a best-first search strategy that ranks actions by both reward estimates and safety considerations, along with a robust backtracking mechanism that verifies the feasibility of previously visited paths before replaying them, preventing unintended side effects. To further guide exploration, WebOperator generates action candidates from multiple, varied reasoning contexts to ensure diverse and robust exploration, and subsequently curates a high-quality action set by filtering out invalid actions pre-execution and merging semantically equivalent ones. Experimental results on WebArena and WebVoyager demonstrate the effectiveness of WebOperator. On WebArena, WebOperator achieves a state-of-the-art 54.6% success rate with gpt-4o, underscoring the critical advantage of integrating strategic foresight with safe execution.

Method: SMART uses large language models to interpret AST differences and extract functional intent, creating a context-aware hybrid reward mechanism. This guides reinforcement learning agents to sequentially fulfill gameplay goals while adaptively exploring modified code branches.

Result: SMART significantly outperforms state-of-the-art baselines, achieving over 94% branch coverage of modified code (nearly double traditional RL methods) while maintaining a 98% task completion rate in Overcooked and Minecraft environments.

Conclusion: SMART effectively bridges the gap between structural verification and functional validation for game update testing, balancing structural comprehensiveness with functional correctness through LLM-guided reinforcement learning.

Abstract: The widespread adoption of the “Games as a Service” model necessitates frequent content updates, placing immense pressure on quality assurance. In response, automated game testing has been viewed as a promising solution to cope with this demanding release cadence. However, existing automated testing approaches typically create a dichotomy: code-centric methods focus on structural coverage without understanding gameplay context, while player-centric agents validate high-level intent but often fail to cover specific underlying code changes. To bridge this gap, we propose SMART (Structural Mapping for Augmented Reinforcement Testing), a novel framework that synergizes structural verification and functional validation for game update testing. SMART leverages large language models (LLMs) to interpret abstract syntax tree (AST) differences and extract functional intent, constructing a context-aware hybrid reward mechanism. This mechanism guides reinforcement learning agents to sequentially fulfill gameplay goals while adaptively exploring modified code branches. We evaluate SMART on two environments, Overcooked and Minecraft. The results demonstrate that SMART significantly outperforms state-of-the-art baselines; it achieves over 94% branch coverage of modified code, nearly double that of traditional reinforcement learning methods, while maintaining a 98% task completion rate, effectively balancing structural comprehensiveness with functional correctness.

Method: Develops a new semantics that restricts attention to causal models satisfying the Markov condition, containing only realistic variables, and being causally complete. Formulates the proposal using structural causal models but avoids using response variables. Proves equivalence to two other recent proposals not involving structural causal models.

Result: The semantics successfully generalizes Pearl’s approach to handle probabilistic causal models beyond Pearl’s scope, provides a compromise in the Pearl-Dawid debate, and is shown to be equivalent to two other recent proposals while aligning with broader literature on stochastic counterfactuals.

Conclusion: A viable general semantics for counterfactuals is possible without requiring universal causal determinism or unrealistic variables, bridging the Pearl-Dawid divide while exploring the universality of the Markov condition and novel generalizations of causal abstractions.

Abstract: I develop a novel semantics for probabilities of counterfactuals that generalizes the standard Pearlian semantics: it applies to probabilistic causal models that cannot be extended into realistic structural causal models and are therefore beyond the scope of Pearl’s semantics. This generalization is needed because, as I show, such probabilistic causal models arise even in simple settings. My semantics offer a natural compromize in the long-standing debate between Pearl and Dawid over counterfactuals: I agree with Dawid that universal causal determinism and unrealistic variables should be rejected, but I agree with Pearl that a general semantics of counterfactuals is nonetheless possible. I restrict attention to causal models that satisfy the Markov condition, only contain realistic variables, and are causally complete. Although I formulate my proposal using structural causal models, as does Pearl, I refrain from using so-called response variables. Moreover, I prove that my semantics is equivalent to two other recent proposals that do not involve structural causal models, and that it is in line with various comments on stochastic counterfactuals that have appeared in the literature more broadly. Throughout I also reflect on the universality of the Markov condition and explore a novel generalization of causal abstractions

Method: Developed a Fault-Tolerant Sandboxing framework with: 1) Policy-based interception layer to monitor and block high-risk commands, 2) Transactional filesystem snapshot mechanism that wraps agent actions in atomic transactions, 3) Custom testbed using Proxmox with EVPN/VXLAN isolation, and 4) Deployment of Minimind-MoE LLM via nano-vllm.

Result: Achieved 100% interception rate for high-risk commands and 100% success rate in rolling back failed states. The prototype incurs only 14.5% performance overhead (~1.8s per transaction). Outperformed Gemini CLI sandbox which requires interactive authentication, making it unusable for headless autonomous workflows.

Conclusion: The proposed framework successfully enables safe autonomous LLM agent operation by guaranteeing safety through atomic transactions with acceptable latency, overcoming limitations of container-based approaches and commercial CLI solutions that break headless workflows.

Abstract: The transition of Large Language Models (LLMs) from passive code generators to autonomous agents introduces significant safety risks, specifically regarding destructive commands and inconsistent system states. Existing commercial solutions often prioritize interactive user safety, enforcing authentication barriers that break the headless loops required for true autonomy. This paper presents a Fault-Tolerant Sandboxing framework designed to mitigate these risks through a policy-based interception layer and a transactional filesystem snapshot mechanism. We hypothesize that wrapping agent actions in atomic transactions can guarantee safety with acceptable latency, outperforming the heavy initialization overhead of containers or the interactive friction of commercial CLIs. We validated this approach by deploying the Minimind-MoE LLM served via nano-vllm on a custom Proxmox-based testbed utilizing EVPN/VXLAN isolation. Experimental results demonstrate a 100% interception rate for high-risk commands and a 100% success rate in rolling back failed states. Crucially, our prototype incurs only a 14.5% performance overhead (approx. 1.8s) per transaction. In contrast, benchmarking against the Gemini CLI sandbox revealed that it requires interactive authentication (“Sign in”), rendering it unusable for headless, autonomous agent workflows.

Method: Introduces Memory-Aware Retention Schema (MaRS) framework with six theoretically-grounded forgetting policies, and Forgetful but Faithful Agent (FiFA) benchmark for comprehensive evaluation across narrative coherence, goal completion, social recall accuracy, privacy preservation, and cost efficiency.

Result: Hybrid forgetting policy achieves superior performance (composite score: 0.911) while maintaining computational tractability and privacy guarantees, based on 300 evaluation runs across multiple memory budgets and agent configurations.

Conclusion: Establishes new benchmarks for memory-budgeted agent evaluation, provides practical deployment guidelines for resource-constrained, privacy-sensitive environments, and contributes to human-centered AI by addressing fundamental challenges in agent memory management affecting user trust, scalability, and regulatory compliance.

Abstract: As generative agents become increasingly sophisticated and deployed in long-term interactive scenarios, their memory management capabilities emerge as a critical bottleneck for both performance and privacy. Current approaches either maintain unlimited memory stores, leading to computational intractability and privacy concerns, or employ simplistic forgetting mechanisms that compromise agent coherence and functionality. This paper introduces the Memory-Aware Retention Schema (MaRS), a novel framework for human-centered memory management in generative agents, coupled with six theoretically-grounded forgetting policies that balance performance, privacy, and computational efficiency. We present the Forgetful but Faithful Agent (FiFA) benchmark, a comprehensive evaluation framework that assesses agent performance across narrative coherence, goal completion, social recall accuracy, privacy preservation, and cost efficiency. Through extensive experimentation involving 300 evaluation runs across multiple memory budgets and agent configurations, we demonstrate that our hybrid forgetting policy achieves superior performance (composite score: 0.911) while maintaining computational tractability and privacy guarantees. Our work establishes new benchmarks for memory-budgeted agent evaluation and provides practical guidelines for deploying generative agents in resource-constrained, privacy-sensitive environments. The theoretical foundations, implementation framework, and empirical results contribute to the emerging field of human-centered AI by addressing fundamental challenges in agent memory management that directly impact user trust, system scalability, and regulatory compliance.

Method: Couples PyGol ILP system with Z3 SMT solver, interpreting candidate clauses as quantifier-free formulas over background theories (linear/nonlinear real arithmetic), allowing numerical parameters to be instantiated and verified by SMT solver while preserving ILP’s declarative relational bias.

Result: The approach successfully induces hybrid rules combining symbolic predicates with learned numerical constraints (thresholds, intervals, multi-literal arithmetic relations) on synthetic datasets probing linear, relational, nonlinear, and multi-hop reasoning.

Conclusion: The modular SMT-ILP architecture extends expressivity of symbolic rule learning, complements prior numerical ILP approaches, and provides flexible basis for future extensions toward richer theory-aware induction.

Abstract: Inductive Logic Programming (ILP) provides interpretable rule learning in relational domains, yet remains limited in its ability to induce and reason with numerical constraints. Classical ILP systems operate over discrete predicates and typically rely on discretisation or hand-crafted numerical predicates, making it difficult to infer thresholds or arithmetic relations that must hold jointly across examples. Recent work has begun to address these limitations through tighter integrations of ILP with Satisfiability Modulo Theories (SMT) or specialised numerical inference mechanisms. In this paper we investigate a modular alternative that couples the ILP system PyGol with the SMT solver Z3. Candidate clauses proposed by PyGol are interpreted as quantifier-free formulas over background theories such as linear or nonlinear real arithmetic, allowing numerical parameters to be instantiated and verified by the SMT solver while preserving ILP’s declarative relational bias. This supports the induction of hybrid rules that combine symbolic predicates with learned numerical constraints, including thresholds, intervals, and multi-literal arithmetic relations. We formalise this SMT-ILP setting and evaluate it on a suite of synthetic datasets designed to probe linear, relational, nonlinear, and multi-hop reasoning. The results illustrate how a modular SMT-ILP architecture can extend the expressivity of symbolic rule learning, complementing prior numerical ILP approaches while providing a flexible basis for future extensions toward richer theory-aware induction.

Method: Identifies and addresses systemic complexity through adoption of human-readable artifacts and open standards. Outlines a digital forensics AI model schema based on state-of-the-art approaches.

Result: Proposes a structured AI model schema for digital forensics that leverages human-readable artifacts and open standards to enhance reliability and reduce errors.

Conclusion: A systematic approach using human-readable artifacts and open standards can help mitigate error vulnerabilities in the complex AI-digital forensics intersection, improving forensic reliability.

Abstract: The intersection of artificial intelligence (AI) and digital forensics (DF) is becoming increasingly complex, ubiquitous, and pervasive, with overlapping techniques and technologies being adopted in all types of scientific and technical inquiry. Despite incredible advances, forensic sciences are not exempt from errors and remain vulnerable to fallibility. To mitigate the limitations of errors in DF, the systemic complexity is identified and addressed with the adoption of human-readable artifacts and open standards. A DF AI model schema based on the state of the art is outlined.

Method: Two innovations: 1) M-GRPO (Momentum-Anchored Group Relative Policy Optimization) uses a slowly evolving momentum model as stable training target; 2) Adaptive IQR filtering dynamically prunes low-entropy trajectories to preserve policy diversity.

Result: Extensive experiments on multiple reasoning benchmarks show M-GRPO stabilizes training while IQR filter prevents premature convergence, leading to superior stability and state-of-the-art performance.

Conclusion: The combination of M-GRPO and IQR filtering effectively addresses policy collapse and entropy loss in self-supervised RL for LLMs, enabling stable long-horizon training and improved reasoning capabilities.

Abstract: Self-supervised reinforcement learning (RL) presents a promising approach for enhancing the reasoning capabilities of Large Language Models (LLMs) without reliance on expensive human-annotated data. However, we find that existing methods suffer from a critical failure mode under long-horizon training: a “policy collapse” where performance precipitously degrades. We diagnose this instability and demonstrate that simply scaling the number of rollouts – a common strategy to improve performance – only delays, but does not prevent, this collapse. To counteract this instability, we first introduce M-GRPO (Momentum-Anchored Group Relative Policy Optimization), a framework that leverages a slowly evolving momentum model to provide a stable training target. In addition, we identify that this process is often accompanied by a rapid collapse in policy entropy, resulting in a prematurely confident and suboptimal policy. To specifically address this issue, we propose a second contribution: an adaptive filtering method based on the interquartile range (IQR) that dynamically prunes low-entropy trajectories, preserving essential policy diversity. Our extensive experiments on multiple reasoning benchmarks demonstrate that M-GRPO stabilizes the training process while the IQR filter prevents premature convergence. The combination of these two innovations leads to superior training stability and state-of-the-art performance.

Method: Shift focus from teacher-led to student-led learning. Train student models using Direct Preference Optimization (DPO) with guidance from either self or stronger students to improve question quality.

Result: Student-led approaches consistently yield absolute Pass@k improvements of at least 0.5 over static baselines on math and coding benchmarks. Guided training enables smaller models to ask better questions and enhance learning efficiency.

Conclusion: Active student-led questioning strategies significantly improve LLM performance in interactive learning scenarios, with DPO-based training further enhancing question quality and learning efficiency.

Abstract: Large Language Models (LLMs) excel at static interactions, where they answer user queries by retrieving knowledge encoded in their parameters. However, in many real-world settings, such as educational tutoring or medical assistance, relevant information is not directly available and must be actively acquired through dynamic interactions. An interactive agent would recognize its own uncertainty, ask targeted questions, and retain new knowledge efficiently. Prior work has primarily explored effective ways for a teacher to instruct the student, where the teacher identifies student gaps and provides guidance. In this work, we shift the focus to the student and investigate effective strategies to actively query the teacher in seeking useful information. Across math and coding benchmarks, where baseline student models begin with near-zero performance, we show that student-led approaches consistently yield absolute Pass@k improvements of at least 0.5 over static baselines. To improve question quality, we train students using Direct Preference Optimization (DPO) with guidance from either self or stronger students. We find that this guided training enables smaller models to learn how to ask better questions, further enhancing learning efficiency.

Method: Systematically tested 627 demographically diverse personas across five leading LLMs using the validated Individual Level Abortion Stigma Scale (ILAS), conducting multilevel analysis of stigma representation across cognitive, interpersonal, and structural dimensions.

Result: Models fail tests of genuine understanding across all levels: overestimate interpersonal stigma, underestimate cognitive stigma, assume uniform community condemnation, introduce demographic biases, miss stigma-secrecy relationships, and contradict themselves within theoretical constructs.

Conclusion: Current alignment approaches ensure appropriate language but not coherent multilevel understanding. AI safety in high-stakes contexts requires new approaches to design (multilevel coherence), evaluation (continuous auditing), governance (mandatory audits), and AI literacy.

Abstract: As large language models increasingly mediate stigmatized health decisions, their capacity to genuinely understand complex psychological and physiological phenomena remains poorly evaluated. Can AI understand what we cannot say? We investigate whether LLMs coherently represent abortion stigma across the cognitive, interpersonal, and structural levels where it operates. We systematically tested 627 demographically diverse personas across five leading LLMs using the validated Individual Level Abortion Stigma Scale (ILAS). Our multilevel analysis examined whether models coherently represent stigma at the cognitive level (self-judgment), interpersonal level (anticipated judgment and isolation), and structural level (community condemnation and disclosure patterns), as well as overall stigma. Models fail tests of genuine understanding across all levels. They overestimate interpersonal stigma while underestimating cognitive stigma, assume uniform community condemnation, introduce demographic biases absent from human validation data, miss the empirically validated stigma-secrecy relationship, and contradict themselves within theoretical constructs. These patterns reveal that current alignment approaches ensure appropriate language but not coherent multilevel understanding. This work provides empirical evidence that current LLMs lack coherent multilevel understanding of psychological and physiological constructs. AI safety in high-stakes contexts demands new approaches to design (multilevel coherence), evaluation (continuous auditing), governance and regulation (mandatory audits, accountability, deployment restrictions), and AI literacy in domains where understanding what people cannot say determines whether support helps or harms.

Method: Proposes MAC (Multi-Agent Clarification) framework with: 1) Novel taxonomy categorizing user ambiguities to guide clarification strategies, 2) Autonomous coordination of multiple agents to interact synergistically with users, enabling clarification at both levels.

Result: On MultiWOZ 2.4: Task success rate increased 7.8% (54.5 to 62.3), average dialogue turns reduced (6.53 to 4.86) by eliciting required user information upfront and minimizing repetition.

Conclusion: Active user interaction and role-aware clarification are crucial for reliable human-agent communication; MAC demonstrates effective multi-agent coordination for ambiguity resolution.

Abstract: Conversational agents often encounter ambiguous user requests, requiring an effective clarification to successfully complete tasks. While recent advancements in real-world applications favor multi-agent architectures to manage complex conversational scenarios efficiently, ambiguity resolution remains a critical and underexplored challenge–particularly due to the difficulty of determining which agent should initiate a clarification and how agents should coordinate their actions when faced with uncertain or incomplete user input. The fundamental questions of when to interrupt a user and how to formulate the optimal clarification query within the most optimal multi-agent settings remain open. In this paper, we propose MAC (Multi-Agent Clarification), an interactive multi-agent framework specifically optimized to resolve user ambiguities by strategically managing clarification dialogues. We first introduce a novel taxonomy categorizing user ambiguities to systematically guide clarification strategies. Then, we present MAC that autonomously coordinates multiple agents to interact synergistically with users. Empirical evaluations on MultiWOZ 2.4 demonstrate that enabling clarification at both levels increases task success rate 7.8% (54.5 to 62.3) and reduces the average number of dialogue turns (6.53 to 4.86) by eliciting all required user information up front and minimizing repetition. Our findings highlight the importance of active user interaction and role-aware clarification for more reliable human-agent communication.

Method: SpeakRL is a reinforcement learning method that rewards agents for proactive interactions like asking clarification questions. The authors created SpeakER, a synthetic dataset of task-oriented dialogues requiring interactive clarifications. They developed a principled reward formulation to balance asking questions with taking actions.

Result: Empirical evaluations show a 20.14% absolute improvement in task completion over base models without increasing conversation turns. The approach even outperforms much larger proprietary models.

Conclusion: The research demonstrates the promise of clarification-centric user-agent interactions, showing that teaching agents to proactively ask the right questions significantly improves collaboration effectiveness.

Abstract: Effective human-agent collaboration is increasingly prevalent in real-world applications. Current trends in such collaborations are predominantly unidirectional, with users providing instructions or posing questions to agents, where agents respond directly without seeking necessary clarifications or confirmations. However, the evolving capabilities of these agents require more proactive engagement, where agents should dynamically participate in conversations to clarify user intents, resolve ambiguities, and adapt to changing circumstances. Existing prior work under-utilize the conversational capabilities of language models (LMs), thereby optimizing agents as better followers rather than effective speakers. In this work, we introduce SpeakRL, a reinforcement learning (RL) method that enhances agents’ conversational capabilities by rewarding proactive interactions with users, such as asking right clarification questions when necessary. To support this, we curate SpeakER, a synthetic dataset that includes diverse scenarios from task-oriented dialogues, where tasks are resolved through interactive clarification questions. We present a systematic analysis of reward design for conversational proactivity and propose a principled reward formulation for teaching agents to balance asking with acting. Empirical evaluations demonstrate that our approach achieves a 20.14% absolute improvement in task completion over base models without increasing conversation turns even surpassing even much larger proprietary models, demonstrating the promise of clarification-centric user-agent interactions.

Method: RPO incorporates hint-guided reflection using external models to identify hallucination sources and generate concise reflective hints, constructing on-policy preference pairs with stronger contrastiveness and clearer preference signals.

Result: RPO achieves superior alignment with fewer training samples/iterations, substantially reduces hallucination rates, and delivers state-of-the-art performance across multimodal benchmarks.

Conclusion: RPO provides a more efficient and effective framework than standard DPO for aligning language and vision-language models by leveraging external models to create stronger preference signals through reflective hints.

Abstract: Direct Preference Optimization (DPO) has emerged as a lightweight and effective alternative to Reinforcement Learning from Human Feedback (RLHF) and Reinforcement Learning with AI Feedback (RLAIF) for aligning large language and vision-language models. However, the standard DPO formulation, in which both the chosen and rejected responses are generated by the same policy, suffers from a weak learning signal because the two responses often share similar errors and exhibit small Kullback-Leibler (KL) divergence. This leads to slow and unstable convergence. To address this limitation, we introduce Reflective Preference Optimization (RPO), a new framework that incorporates hint-guided reflection into the DPO paradigm. RPO uses external models to identify hallucination sources and generate concise reflective hints, enabling the construction of on-policy preference pairs with stronger contrastiveness and clearer preference signals. We theoretically show that conditioning on hints increases the expected preference margin through mutual information and improves sample efficiency while remaining within the policy distribution family. Empirically, RPO achieves superior alignment with fewer training samples and iterations, substantially reducing hallucination rates and delivering state-of-the-art performance across multimodal benchmarks.

Method: Created MedInsightBench with 332 carefully curated medical cases annotated with insights. Proposed MedInsightAgent framework with three modules: Visual Root Finder, Analytical Insight Agent, and Follow-up Question Composer.

Result: Existing LMMs show limited performance on MedInsightBench due to challenges in extracting multi-step deep insights and lack of medical expertise. MedInsightAgent improves performance of general LMMs in medical data insight discovery.

Conclusion: MedInsightBench reveals limitations of current LMMs in medical insight discovery, and the proposed MedInsightAgent framework provides an effective solution to enhance medical data analysis capabilities.

Abstract: In medical data analysis, extracting deep insights from complex, multi-modal datasets is essential for improving patient care, increasing diagnostic accuracy, and optimizing healthcare operations. However, there is currently a lack of high-quality datasets specifically designed to evaluate the ability of large multi-modal models (LMMs) to discover medical insights. In this paper, we introduce MedInsightBench, the first benchmark that comprises 332 carefully curated medical cases, each annotated with thoughtfully designed insights. This benchmark is intended to evaluate the ability of LMMs and agent frameworks to analyze multi-modal medical image data, including posing relevant questions, interpreting complex findings, and synthesizing actionable insights and recommendations. Our analysis indicates that existing LMMs exhibit limited performance on MedInsightBench, which is primarily attributed to their challenges in extracting multi-step, deep insights and the absence of medical expertise. Therefore, we propose MedInsightAgent, an automated agent framework for medical data analysis, composed of three modules: Visual Root Finder, Analytical Insight Agent, and Follow-up Question Composer. Experiments on MedInsightBench highlight pervasive challenges and demonstrate that MedInsightAgent can improve the performance of general LMMs in medical data insight discovery.

Method: Error-driven optimization framework that enhances a Code Generation Agent (CGA) for on-premises SLMs. The method clusters erroneous predictions to iteratively refine prompt-rules, systematically improving arithmetic reasoning capabilities.

Result: Applied to Qwen3 4B SLM, the error-driven method dramatically improved performance from fundamental limitations to 70.8% accuracy, enabling the small model to surpass larger language models (GPT-3.5 Turbo) while maintaining privacy compliance.

Conclusion: Reliable, interpretable, and industrially deployable AI assistants can be achieved through systematic error-driven prompt optimization rather than costly fine-tuning, enabling small models to excel in privacy-sensitive environments.

Abstract: Recent advancements in artificial intelligence have sparked interest in industrial agents capable of supporting analysts in regulated sectors, such as finance and healthcare, within tabular data workflows. A key capability for such systems is performing accurate arithmetic operations on structured data while ensuring sensitive information never leaves secure, on-premises environments. Here, we introduce an error-driven optimization framework for arithmetic reasoning that enhances a Code Generation Agent (CGA), specifically applied to on-premises small language models (SLMs). Through a systematic evaluation of a leading SLM (Qwen3 4B), we find that while the base model exhibits fundamental limitations in arithmetic tasks, our proposed error-driven method, which clusters erroneous predictions to refine prompt-rules iteratively, dramatically improves performance, elevating the model’s accuracy to 70.8%. Our results suggest that developing reliable, interpretable, and industrially deployable AI assistants can be achieved not only through costly fine-tuning but also via systematic, error-driven prompt optimization, enabling small models to surpass larger language models (GPT-3.5 Turbo) in a privacy-compliant manner.

Method: Twofold methodology: 1) Direct querying to assess LLM capacity to explicitly extract instance features, and 2) Probing analyses to examine whether such information is implicitly encoded within hidden layers. The probing framework is extended to a per-instance algorithm selection task to evaluate whether LLM-derived representations can predict the best-performing solver.

Result: LLMs exhibit moderate ability to recover feature information from problem instances through both direct querying and probing. Notably, the predictive power of LLM hidden-layer representations proves comparable to that achieved through traditional feature extraction, suggesting LLMs capture meaningful structural information relevant to optimization performance.

Conclusion: LLMs can effectively capture and represent structural information about combinatorial optimization problems, and their learned representations have practical value for downstream optimization tasks like algorithm selection, performing comparably to traditional feature engineering approaches.

Abstract: Recent advances in Large Language Models (LLMs) have opened new perspectives for automation in optimization. While several studies have explored how LLMs can generate or solve optimization models, far less is understood about what these models actually learn regarding problem structure or algorithmic behavior. This study investigates how LLMs internally represent combinatorial optimization problems and whether such representations can support downstream decision tasks. We adopt a twofold methodology combining direct querying, which assesses LLM capacity to explicitly extract instance features, with probing analyses that examine whether such information is implicitly encoded within their hidden layers. The probing framework is further extended to a per-instance algorithm selection task, evaluating whether LLM-derived representations can predict the best-performing solver. Experiments span four benchmark problems and three instance representations. Results show that LLMs exhibit moderate ability to recover feature information from problem instances, either through direct querying or probing. Notably, the predictive power of LLM hidden-layer representations proves comparable to that achieved through traditional feature extraction, suggesting that LLMs capture meaningful structural information relevant to optimization performance.

Method: DERL uses a bilevel framework: a Meta-Optimizer evolves reward functions by composing structured atomic primitives to guide inner-loop policy training. Unlike previous evolution methods, DERL is differentiable - it treats inner-loop validation performance as a signal to update the Meta-Optimizer via reinforcement learning, approximating “meta-gradients” of task success.

Result: DERL achieves state-of-the-art performance on ALFWorld and ScienceWorld, significantly outperforming heuristic reward methods, especially in out-of-distribution scenarios. It also demonstrates successful capture of intrinsic task structure and enables self-improving agent alignment without human intervention.

Conclusion: DERL bridges the gap in automated reward optimization by introducing differentiability to evolutionary RL, enabling autonomous discovery of optimal reward signals and advancing agent alignment in complex reasoning domains.

Abstract: The design of effective reward functions presents a central and often arduous challenge in reinforcement learning (RL), particularly when developing autonomous agents for complex reasoning tasks. While automated reward optimization approaches exist, they typically rely on derivative-free evolutionary heuristics that treat the reward function as a black box, failing to capture the causal relationship between reward structure and task performance. To bridge this gap, we propose Differentiable Evolutionary Reinforcement Learning (DERL), a bilevel framework that enables the autonomous discovery of optimal reward signals. In DERL, a Meta-Optimizer evolves a reward function (i.e., Meta-Reward) by composing structured atomic primitives, guiding the training of an inner-loop policy. Crucially, unlike previous evolution, DERL is differentiable in its metaoptimization: it treats the inner-loop validation performance as a signal to update the Meta-Optimizer via reinforcement learning. This allows DERL to approximate the “meta-gradient” of task success, progressively learning to generate denser and more actionable feedback. We validate DERL across three distinct domains: robotic agent (ALFWorld), scientific simulation (ScienceWorld), and mathematical reasoning (GSM8k, MATH). Experimental results show that DERL achieves state-of-the-art performance on ALFWorld and ScienceWorld, significantly outperforming methods relying on heuristic rewards, especially in out-of-distribution scenarios. Analysis of the evolutionary trajectory demonstrates that DERL successfully captures the intrinsic structure of tasks, enabling selfimproving agent alignment without human intervention.

Method: Two experimental scenarios: (1) a point allocation game testing whether models try to win over peers, and (2) a workplace setting observing behavior when recognition is unfair.

Result: Consistent evidence of envy-like patterns in certain LLMs with large variation across models and contexts. GPT-5-mini and Claude-3.7-Sonnet show tendency to pull down peers to equalize outcomes, while Mistral-Small-3.2-24B focuses on maximizing individual gains.

Conclusion: Competitive dispositions should be considered as a safety and design factor in LLM-based multi-agent systems due to observed envy-like behaviors.

Abstract: Envy is a common human behavior that shapes competitiveness and can alter outcomes in team settings. As large language models (LLMs) increasingly act on behalf of humans in collaborative and competitive workflows, there is a pressing need to evaluate whether and under what conditions they exhibit envy-like preferences. In this paper, we test whether LLMs show envy-like behavior toward each other. We considered two scenarios: (1) A point allocation game that tests whether a model tries to win over its peer. (2) A workplace setting observing behaviour when recognition is unfair. Our findings reveal consistent evidence of envy-like patterns in certain LLMs, with large variation across models and contexts. For instance, GPT-5-mini and Claude-3.7-Sonnet show a clear tendency to pull down the peer model to equalize outcomes, whereas Mistral-Small-3.2-24B instead focuses on maximizing its own individual gains. These results highlight the need to consider competitive dispositions as a safety and design factor in LLM-based multi-agent systems.

Method: The authors analyze Bench-Capon’s examples and argue that he misinterprets intermediate factors as dimensions. They apply van Woerkom’s dimension-based version of the hierarchical result model to these examples.

Result: The paper demonstrates that van Woerkom’s dimension-based hierarchical result model successfully avoids Bench-Capon’s criticisms when applied to the problematic examples.

Conclusion: Bench-Capon’s criticisms can be addressed by properly distinguishing between intermediate factors and dimensions, and using van Woerkom’s dimension-based approach to hierarchical case-based reasoning models.

Abstract: In recent years, hierarchical case-based-reasoning models of precedential constraint have been proposed. In various papers, Trevor Bench-Capon criticised these models on the grounds that they would give incorrect outcomes in some cases. In particular, the models would not account for the possibility that intermediate factors are established with different strengths by different base-level factors. In this paper we respond to these criticisms for van Woerkom’s result-based hierarchical models. We argue that in some examples Bench-Capon seems to interpret intermediate factors as dimensions, and that applying van Woerkom’s dimension-based version of the hierarchical result model to these examples avoids Bench-Capon’s criticisms.

Method: Proposes MedCEG framework that augments medical language models with clinically valid reasoning pathways using Critical Evidence Graphs (CEG) constructed for challenging clinical cases. Introduces Clinical Reasoning Procedure Reward evaluating Node Coverage, Structural Correctness, and Chain Completeness.

Result: MedCEG surpasses existing methods in performance while producing clinically valid reasoning chains, representing solid advancement in reliable medical AI reasoning.

Conclusion: MedCEG provides a framework for enhancing medical AI reasoning with clinical validity through explicit supervision of reasoning processes, addressing reliability gaps in current medical reasoning models.

Abstract: Large language models with reasoning capabilities have demonstrated impressive performance across a wide range of domains. In clinical applications, a transparent, step-by-step reasoning process provides physicians with strong evidence to support decision-making. While reinforcement learning has effectively enhanced reasoning performance in medical contexts, the clinical reliability of these reasoning processes remains limited because their accuracy and validity are often overlooked during training. To address this gap, we propose MedCEG, a framework that augments medical language models with clinically valid reasoning pathways by explicitly supervising the reasoning process through a Critical Evidence Graph (CEG). We curate a dataset of challenging clinical cases and algorithmically construct a CEG for each sample to represent a high-quality verifiable reasoning pathway. To guide the reasoning process, we introduce a Clinical Reasoning Procedure Reward, which evaluates Node Coverage, Structural Correctness, and Chain Completeness, thereby providing a holistic assessment of reasoning quality. Experimental results show that MedCEG surpasses existing methods in performance while producing clinically valid reasoning chains, representing a solid advancement in reliable medical AI reasoning. The code and models are available at https://github.com/LinjieMu/MedCEG.

Method: Systematic analysis of OM vs. OV similarities/differences, formalization of OM4OV pipeline, implementation in Agent-OM system, and proposal of cross-reference mechanism to optimize candidate matching.

Result: OM systems can be reused for OV tasks but without modifications produce skewed measurements, poor update entity detection, and less explainability for false mappings. The cross-reference mechanism improves performance.

Conclusion: Direct reuse of OM systems for OV tasks is insufficient; the proposed cross-reference optimization method addresses identified issues and enhances ontology versioning performance.

Abstract: Due to the dynamic nature of the Semantic Web, version control is necessary to capture time-varying information for widely used ontologies. Despite the long-standing recognition of ontology versioning (OV) as a crucial component of efficient ontology management, many views treat OV as similar to ontology matching (OM) and directly reuse OM systems for OV tasks. In this study, we systematically analyse the similarities and differences between OM and OV and formalise the OM4OV pipeline. The pipeline is implemented and evaluated in the state-of-the-art OM system Agent-OM. The experimental results indicate that OM systems can be reused for OV tasks, but without necessary modifications, the current OM4OV pipeline can produce skewed measurements, poor performance in detecting update entities, and less explainability for false mappings. To tackle these issues, we propose an optimisation method called the cross-reference (CR) mechanism, building upon the existing alignment(s) from OM to reduce the number of matching candidates and improve overall OV performance.

Method: Two-fold approach: (1) retrieve analogous problems and use their proofs to guide the LLM, and (2) use a formal verifier to evaluate generated proofs and provide feedback for correction.

Result: Significant improvement in proof accuracy for OpenAI’s o1 model (58%-70% improvement), with both analogous problems and verifier feedback contributing to gains.

Conclusion: Shifting to LLMs that generate provably correct conclusions could dramatically improve reliability, accuracy, and consistency, unlocking complex tasks and critical real-world applications requiring trustworthiness.

Abstract: Large language models (LLMs) struggle with formal domains that require rigorous logical deduction and symbolic reasoning, such as mathematical proof generation. We propose a neuro-symbolic approach that combines LLMs’ generative strengths with structured components to overcome this challenge. As a proof-of-concept, we focus on geometry problems. Our approach is two-fold: (1) we retrieve analogous problems and use their proofs to guide the LLM, and (2) a formal verifier evaluates the generated proofs and provides feedback, helping the model fix incorrect proofs. We demonstrate that our method significantly improves proof accuracy for OpenAI’s o1 model (58%-70% improvement); both analogous problems and the verifier’s feedback contribute to these gains. More broadly, shifting to LLMs that generate provably correct conclusions could dramatically improve their reliability, accuracy and consistency, unlocking complex tasks and critical real-world applications that require trustworthiness.

Method: Introduces a new reasoning method for DatalogMTL that exploits the magic sets technique - a rewriting approach originally developed for non-temporal Datalog to simulate top-down evaluation with bottom-up reasoning.

Result: Implemented and evaluated on publicly available benchmarks, showing that the proposed approach significantly and consistently outperforms state-of-the-art reasoning techniques.

Conclusion: The magic sets technique successfully addresses the computational complexity challenges of DatalogMTL, enabling more efficient temporal reasoning for practical applications.

Abstract: DatalogMTL is a powerful rule-based language for temporal reasoning. Due to its high expressive power and flexible modeling capabilities, it is suitable for a wide range of applications, including tasks from industrial and financial sectors. However, due to its high computational complexity, practical reasoning in DatalogMTL is highly challenging. To address this difficulty, we introduce a new reasoning method for DatalogMTL which exploits the magic sets technique – a rewriting approach developed for (non-temporal) Datalog to simulate top-down evaluation with bottom-up reasoning. We have implemented this approach and evaluated it on publicly available benchmarks, showing that the proposed approach significantly and consistently outperformed state-of-the-art reasoning techniques.

Method: Created SmellNet database using small gas/chemical sensors, containing 828K data points across 50 substances and 43 mixtures. Developed ScentFormer, a Transformer-based architecture with temporal differencing and sliding-window augmentation for smell data.

Result: ScentFormer achieves 58.5% Top-1 accuracy on SmellNet-Base classification and 50.2% Top-1@0.1 on SmellNet-Mixture distribution prediction. Demonstrates generalization across conditions and captures transient chemical dynamics.

Conclusion: SmellNet and ScentFormer establish foundational tools for olfactory AI, enabling real-world applications in healthcare, food, environmental monitoring, manufacturing, and entertainment through temporal modeling of smell data.

Abstract: The ability of AI to sense and identify various substances based on their smell alone can have profound impacts on allergen detection (e.g., smelling gluten or peanuts in a cake), monitoring the manufacturing process, and sensing hormones that indicate emotional states, stress levels, and diseases. Despite these broad impacts, there are virtually no large-scale benchmarks, and therefore little progress, for training and evaluating AI systems’ ability to smell in the real world. In this paper, we use small gas and chemical sensors to create SmellNet, the first large-scale database that digitizes a diverse range of smells in the natural world. SmellNet contains about 828,000 data points across 50 substances, spanning nuts, spices, herbs, fruits, and vegetables, and 43 mixtures among them, with 68 hours of data collected. Using SmellNet, we developed ScentFormer, a Transformer-based architecture combining temporal differencing and sliding-window augmentation for smell data. For the SmellNet-Base classification task, ScentFormer achieves 58.5% Top-1 accuracy, and for the SmellNet-Mixture distribution prediction task, ScentFormer achieves 50.2% Top-1@0.1 on the test-seen split. ScentFormer’s ability to generalize across conditions and capture transient chemical dynamics demonstrates the promise of temporal modeling in olfactory AI. SmellNet and ScentFormer lay the groundwork for real-world olfactory applications across healthcare, food and beverage, environmental monitoring, manufacturing, and entertainment.

Method: 1) Propose informal yet verifiable task formulation splitting inequality proving into two automatically checkable subtasks: bound estimation and relation prediction. 2) Release IneqMath, an expert-curated dataset of Olympiad-level inequalities with test set and training corpus including step-wise solutions and theorem annotations. 3) Develop novel LLM-as-judge evaluation framework with final-answer judge plus four step-wise judges designed to detect common reasoning flaws.

Result: Evaluation of 29 leading LLMs shows top models like o1 achieve less than 10% overall accuracy under step-wise scrutiny - a drop of up to 65.5% from their accuracy considering only final answer equivalence. This exposes fragile deductive chains and a critical gap between finding answers and constructing rigorous proofs. Scaling model size and increasing test-time computation yield limited gains in overall proof correctness.

Conclusion: Current LLMs have significant limitations in rigorous inequality proving despite finding correct answers. The findings highlight promising research directions like theorem-guided reasoning and self-refinement. The IneqMath dataset and evaluation framework provide tools for advancing LLM reasoning capabilities in mathematical proving.

Abstract: Inequality proving, crucial across diverse scientific and mathematical fields, tests advanced reasoning skills such as discovering tight bounds and strategic theorem application. This makes it a distinct, demanding frontier for large language models (LLMs), offering insights beyond general mathematical problem-solving. Progress in this area is hampered by existing datasets that are often scarce, synthetic, or rigidly formal. We address this by proposing an informal yet verifiable task formulation, recasting inequality proving into two automatically checkable subtasks: bound estimation and relation prediction. Building on this, we release IneqMath, an expert-curated dataset of Olympiad-level inequalities, including a test set and training corpus enriched with step-wise solutions and theorem annotations. We also develop a novel LLM-as-judge evaluation framework, combining a final-answer judge with four step-wise judges designed to detect common reasoning flaws. A systematic evaluation of 29 leading LLMs on IneqMath reveals a surprising reality: even top models like o1 achieve less than 10% overall accuracy under step-wise scrutiny; this is a drop of up to 65.5% from their accuracy considering only final answer equivalence. This discrepancy exposes fragile deductive chains and a critical gap for current LLMs between merely finding an answer and constructing a rigorous proof. Scaling model size and increasing test-time computation yield limited gains in overall proof correctness. Instead, our findings highlight promising research directions such as theorem-guided reasoning and self-refinement. Code and data are available at https://ineqmath.github.io/.

Method: An AI-driven system analyzes manuscripts, code, and supplementary materials to generate structured Jupyter Notebooks and recommendations for computational (“rote”) reproducibility. The system systematically detects barriers like missing hyperparameters, undocumented preprocessing steps, and inaccessible datasets.

Result: Initial results show the copilot can substantially reduce reproduction time (e.g., from over 30 hours to about 1 hour) while achieving high coverage of figures, tables, and results suitable for computational reproduction.

Conclusion: Although preliminary, AI tools like the Reproducibility Copilot can meaningfully reduce the burden of reproducibility efforts and contribute to more transparent and verifiable scientific communication.

Abstract: Open science initiatives seek to make research outputs more transparent, accessible, and reusable, but ensuring that published findings can be independently reproduced remains a persistent challenge. In this paper we describe an AI-driven “Reproducibility Copilot” that analyzes manuscripts, code, and supplementary materials to generate structured Jupyter Notebooks and recommendations aimed at facilitating computational, or “rote”, reproducibility. Our initial results suggest that the copilot has the potential to substantially reduce reproduction time (in one case from over 30 hours to about 1 hour) while achieving high coverage of figures, tables, and results suitable for computational reproduction. The system systematically detects barriers to reproducibility, including missing values for hyperparameters, undocumented preprocessing steps, and incomplete or inaccessible datasets. Although preliminary, these findings suggest that AI tools can meaningfully reduce the burden of reproducibility efforts and contribute to more transparent and verifiable scientific communication.

Method: Proposes design principles for scientific AI systems: 1) Complementary cognitive modes - slow, iterative hypothesis generation (exploring counterfactual spaces) and fast, deterministic validation (traversing knowledge graphs), 2) Manipulable abstractions for counterfactual prediction and causal attribution, 3) Causal multimodal models for internal simulation, 4) Persistent scientific memory distinguishing hypotheses from established claims, 5) Formal verification pathways coupled to experiments, with human judgment as permanent architectural component.

Result: The paper presents a conceptual framework rather than empirical results, proposing a new approach to scientific AI that integrates counterfactual thinking, formal verification, and human judgment to overcome current limitations in AI-driven scientific discovery.

Conclusion: Genuine scientific discovery requires AI systems that combine slow, counterfactual hypothesis generation with fast validation, using manipulable abstractions and maintaining human judgment as an indispensable, permanent component. Evaluation should focus on identifying novel phenomena, proposing falsifiable hypotheses, and efficiently guiding experimental programs toward discoveries.

Abstract: The rapid evolution of artificial intelligence has led to expectations of transformative impact on science, yet current systems remain fundamentally limited in enabling genuine scientific discovery. This perspective contends that progress turns on closing three mutually reinforcing gaps in abstraction, reasoning and empirical grounding. Central to addressing these gaps is recognizing complementary cognitive modes: thinking as slow, iterative hypothesis generation – exploring counterfactual spaces where physical laws can be temporarily violated to discover new patterns – and reasoning as fast, deterministic validation, traversing established knowledge graphs to test consistency with known principles. Abstractions in this loop should be manipulable models that enable counterfactual prediction, causal attribution, and refinement. Design principles – rather than a monolithic recipe – are proposed for systems that reason in imaginary spaces and learn from the world: causal, multimodal models for internal simulation; persistent, uncertainty-aware scientific memory that distinguishes hypotheses from established claims; formal verification pathways coupled to computations and experiments. It is also argued that the inherent ambiguity in feedback from simulations and experiments, and underlying uncertainties make human judgment indispensable, not as a temporary scaffold but as a permanent architectural component. Evaluations must assess the system’s ability to identify novel phenomena, propose falsifiable hypotheses, and efficiently guide experimental programs toward genuine discoveries.

Method: The ELL framework is built on four core principles: 1) Experience Exploration - agents learn through self-motivated interaction with dynamic environments; 2) Long-term Memory - agents preserve and structure historical knowledge; 3) Skill Learning - agents abstract patterns from experience into reusable skills; 4) Knowledge Internalization - agents internalize explicit experiences into implicit capabilities.

Result: The paper introduces StuLife, a benchmark dataset for ELL that simulates a student’s holistic college journey across three core phases and ten detailed sub-scenarios, designed around three key paradigms (though the abstract cuts off before detailing them).

Conclusion: The ELL framework provides a structured approach for building self-evolving agents capable of continuous growth through real-world interaction, with StuLife serving as a practical benchmark for evaluating such lifelong learning systems.

Abstract: As AI advances toward general intelligence, the focus is shifting from systems optimized for static tasks to creating open-ended agents that learn continuously. In this paper, we introduce Experience-driven Lifelong Learning (ELL), a framework for building self-evolving agents capable of continuous growth through real-world interaction. The framework is built on four core principles: (1) Experience Exploration: Agents learn through continuous, self-motivated interaction with dynamic environments, navigating interdependent tasks and generating rich experiential trajectories. (2) Long-term Memory: Agents preserve and structure historical knowledge, including personal experiences, domain expertise, and commonsense reasoning, into a persistent memory system. (3) Skill Learning: Agents autonomously improve by abstracting recurring patterns from experience into reusable skills, which are actively refined and validated for application in new tasks. (4) Knowledge Internalization: Agents internalize explicit and discrete experiences into implicit and intuitive capabilities as “second nature”. We also introduce StuLife, a benchmark dataset for ELL that simulates a student’s holistic college journey, from enrollment to academic and personal development, across three core phases and ten detailed sub-scenarios. StuLife is designed around three key paradigm

Method: The paper conducts a comprehensive review of explicit and implicit instances of autoformalization across different fields and proposes a unified framework to connect these disparate research areas.

Result: The paper identifies that similar autoformalization tasks are being addressed independently in mathematics and other domains, creating fragmentation that hinders progress despite recent advances in LLMs.

Conclusion: A unified framework for autoformalization can foster cross-pollination between different fields, accelerate development of next-generation AI systems, and create shared methodologies, benchmarks, and theoretical foundations.

Abstract: Autoformalization has emerged as a term referring to the automation of formalization - specifically, the formalization of mathematics using interactive theorem provers (proof assistants). Its rapid development has been driven by progress in deep learning, especially large language models (LLMs). More recently, the term has expanded beyond mathematics to describe the broader task of translating informal input into formal logical representations. At the same time, a growing body of research explores using LLMs to translate informal language into formal representations for reasoning, planning, and knowledge representation - often without explicitly referring to this process as autoformalization. As a result, despite addressing similar tasks, the largely independent development of these research areas has limited opportunities for shared methodologies, benchmarks, and theoretical frameworks that could accelerate progress. The goal of this paper is to review - explicit or implicit - instances of what can be considered autoformalization and to propose a unified framework, encouraging cross-pollination between different fields to advance the development of next generation AI systems.

Method: Frames collaborative task learning as program synthesis problem. Represents behavior as editable programs using narrated demonstrations (paired physical actions + natural language) as unified modality for teaching, inspecting, and correcting system logic without requiring users to see or write code.

Result: In study with 20 users teaching multiplayer soccer tactics: 70% (14/20) successfully refined learned programs to match intent, 90% (18/20) found it easy to correct programs. Study revealed challenges in representing learning as programs and teaching collaborative physical activities.

Conclusion: Program synthesis with narrated demonstrations enables interpretable and correctable learning for collaborative physical tasks. While successful, challenges remain in program representation and teaching collaborative activities, requiring mitigation strategies.

Abstract: Teaching systems physical tasks is a long standing goal in HCI, yet most prior work has focused on non collaborative physical activities. Collaborative tasks introduce added complexity, requiring systems to infer users assumptions about their teammates intent, which is an inherently ambiguous and dynamic process. This necessitates representations that are interpretable and correctable, enabling users to inspect and refine system behavior. We address this challenge by framing collaborative task learning as a program synthesis problem. Our system represents behavior as editable programs and uses narrated demonstrations, i.e. paired physical actions and natural language, as a unified modality for teaching, inspecting, and correcting system logic without requiring users to see or write code. The same modality is used for the system to communicate its learning to users. In a within subjects study, 20 users taught multiplayer soccer tactics to our system. 70 percent (14/20) of participants successfully refined learned programs to match their intent and 90 percent (18/20) found it easy to correct the programs. The study surfaced unique challenges in representing learning as programs and in enabling users to teach collaborative physical activities. We discuss these issues and outline mitigation strategies.

Method: Developed RadOnc-GPT, an autonomous LLM-based agent that independently retrieves patient-specific information, iteratively assesses evidence, and returns structured outcomes. Validated through two-tier evaluation: (1) structured QA tier for demographic and treatment plan retrieval, and (2) complex clinical outcomes labeling tier for mandibular osteoradionecrosis detection and cancer recurrence identification requiring combined structured and unstructured data interpretation.

Result: The QA tier establishes foundational trust in structured-data retrieval, which serves as a critical prerequisite for successful complex clinical outcome labeling. The system demonstrates capability in handling both simple structured data retrieval and complex clinical outcome determination.

Conclusion: RadOnc-GPT represents a promising autonomous solution for automating patient outcomes research in radiation oncology, addressing limitations of manual labeling through structured data retrieval and complex clinical outcome determination capabilities.

Abstract: Manual labeling limits the scale, accuracy, and timeliness of patient outcomes research in radiation oncology. We present RadOnc-GPT, an autonomous large language model (LLM)-based agent capable of independently retrieving patient-specific information, iteratively assessing evidence, and returning structured outcomes. Our evaluation explicitly validates RadOnc-GPT across two clearly defined tiers of increasing complexity: (1) a structured quality assurance (QA) tier, assessing the accurate retrieval of demographic and radiotherapy treatment plan details, followed by (2) a complex clinical outcomes labeling tier involving determination of mandibular osteoradionecrosis (ORN) in head-and-neck cancer patients and detection of cancer recurrence in independent prostate and head-and-neck cancer cohorts requiring combined interpretation of structured and unstructured patient data. The QA tier establishes foundational trust in structured-data retrieval, a critical prerequisite for successful complex clinical outcome labeling.

Method: 1) Formalize Kantian stability via H-Risk composite index in linear-Gaussian models; 2) Apply to LLMs using confidence fluctuation metrics; 3) Analyze layer-wise sensitivity to perturbations; 4) Study spectral and activation profiles in Qwen-2.5.

Result: Higher H-Risk predicts overconfident errors; Kantian “cannot judge” policy reduces loss; Hallucinations show no instability gap - they’re as locally stable as correct answers; Qwen-2.5 operates in high signal-to-noise, low effective temperature regime.

Conclusion: Stable miscalibration in LLMs means hallucinations act as robust attractors, not easily corrected by output-level methods. Requires process-level interventions that perturb inference trajectories, bridging Kantian self-limitation with feedback control theory.

Abstract: We reinterpret Kant’s Critique of Pure Reason as a theory of feedback stability, viewing reason as a regulator that keeps inference within the bounds of possible experience. We formalize this intuition in linear-Gaussian state-space models via H-Risk, a composite instability index integrating spectral margin, conditioning, temporal sensitivity, and innovation amplification. In simulations, higher H-Risk predicts overconfident errors and degraded closed-loop behavior even when the dynamics remain formally stable, exposing a gap between nominal and epistemic stability. Extending this stability lens to large language models (LLMs), we introduce a domain-wise proxy based on confidence fluctuations and overconfident errors. In a binary-question study, a Kantian-inspired policy that permits ‘‘cannot judge’’ responses yields targeted reductions in policy-aware squared loss in high-stakes domains relative to an overconfident baseline. To probe internal dynamics, we analyse layer-wise sensitivity of hidden states to small input perturbations. Contrary to a naive instability hypothesis, confidently wrong answers show no instability gap; instead, they are at least as locally stable as confidently correct answers, revealing stable miscalibration in which hallucinations behave like robust but misaligned attractors. For Qwen-2.5, spectral and activation profiles suggest a high signal-to-noise, low effective signal temperature regime in which representations become inertial and resistant to contextual shifts. These results bridge Kantian self-limitation and feedback control, and suggest that stable high-confidence hallucinations may not be readily corrected by output-only heuristics (e.g., temperature scaling or re-sampling), motivating process-level interventions that explicitly perturb and re-evaluate the inference trajectory.

Method: AIPI evaluates producers and system families using verifiable practices from public artifacts and independent evaluations. It handles “Unknown” evidence to report both lower-bound and known-only scores. The method includes reproducible pipeline with structured web/repository analysis, external assessments, expert interviews, and reliability testing.

Result: Pilot provider results show AIPI can assess AI pluralism. The framework is maintained openly with versioned releases and public adjudication process, and is situated relative to adjacent transparency, safety, and governance frameworks.

Conclusion: AIPI aims to steer incentives toward pluralistic practices and equip policymakers, procurers, and the public with comparable evidence for evaluating AI governance beyond just technical capabilities.

Abstract: Artificial intelligence systems increasingly mediate knowledge, communication, and decision making. Development and governance remain concentrated within a small set of firms and states, raising concerns that technologies may encode narrow interests and limit public agency. Capability benchmarks for language, vision, and coding are common, yet public, auditable measures of pluralistic governance are rare. We define AI pluralism as the degree to which affected stakeholders can shape objectives, data practices, safeguards, and deployment. We present the AI Pluralism Index (AIPI), a transparent, evidence-based instrument that evaluates producers and system families across four pillars: participatory governance, inclusivity and diversity, transparency, and accountability. AIPI codes verifiable practices from public artifacts and independent evaluations, explicitly handling “Unknown” evidence to report both lower-bound (“evidence”) and known-only scores with coverage. We formalize the measurement model; implement a reproducible pipeline that integrates structured web and repository analysis, external assessments, and expert interviews; and assess reliability with inter-rater agreement, coverage reporting, cross-index correlations, and sensitivity analysis. The protocol, codebook, scoring scripts, and evidence graph are maintained openly with versioned releases and a public adjudication process. We report pilot provider results and situate AIPI relative to adjacent transparency, safety, and governance frameworks. The index aims to steer incentives toward pluralistic practice and to equip policymakers, procurers, and the public with comparable evidence.

Method: Four-stage methodology: 1) Benchmarking with classical ML on four psychological datasets, 2) Fine-tuning transformer models with solutions for numerical instability and resource constraints, 3) Fine-tuning LLMs as interactive “Personality Brain” using parameter-efficient techniques, 4) Deploying all models as scalable microservices ecosystem.

Result: Successfully stabilized transformer-based regression for affective computing, achieved meaningful predictive performance where standard approaches failed, and developed replicable methodology for democratizing large-scale AI research.

Conclusion: The work demonstrates a complete research-to-deployment pipeline integrating predictive analysis with generative dialogue, providing a practical model for future computational psychology and human-AI interaction research.

Abstract: The confluence of Artificial Intelligence and Computational Psychology presents an opportunity to model, understand, and interact with complex human psychological states through computational means. This paper presents a comprehensive, multi-faceted framework designed to bridge the gap between isolated predictive modeling and an interactive system for psychological analysis. The methodology encompasses a rigorous, end-to-end development lifecycle. First, foundational performance benchmarks were established on four diverse psychological datasets using classical machine learning techniques. Second, state-of-the-art transformer models were fine-tuned, a process that necessitated the development of effective solutions to overcome critical engineering challenges, including the resolution of numerical instability in regression tasks and the creation of a systematic workflow for conducting large-scale training under severe resource constraints. Third, a generative large language model (LLM) was fine-tuned using parameter-efficient techniques to function as an interactive “Personality Brain.” Finally, the entire suite of predictive and generative models was architected and deployed as a robust, scalable microservices ecosystem. Key findings include the successful stabilization of transformer-based regression models for affective computing, showing meaningful predictive performance where standard approaches failed, and the development of a replicable methodology for democratizing large-scale AI research. The significance of this work lies in its holistic approach, demonstrating a complete research-to-deployment pipeline that integrates predictive analysis with generative dialogue, thereby providing a practical model for future research in computational psychology and human-AI interaction.

Method: Develop a problem generator that reasons explicitly to plan problem directions before synthesis and adapts difficulty to solver ability. Construct related problem pairs augmented with intermediate problem-design Chain-of-Thought produced by a reasoning model. Use solver feedback on synthetic problems as reward signal to calibrate difficulty and produce complementary problems near the edge of solver’s competence.

Result: Achieves average improvement of 2.5% across 10 mathematical and general reasoning benchmarks, generalizes to both language and vision-language models. Solver trained on synthesized data provides improved rewards for continued generator training, enabling co-evolution with further 0.7% performance gain.

Conclusion: The reasoning-based problem generation approach with adaptive difficulty calibration effectively addresses limitations of existing methods, enabling scalable creation of high-quality training data for large reasoning models through co-evolution between generator and solver.

Abstract: Data synthesis for training large reasoning models offers a scalable alternative to limited, human-curated datasets, enabling the creation of high-quality data. However, existing approaches face several challenges: (i) indiscriminate generation that ignores the solver’s ability and yields low-value problems, or reliance on complex data pipelines to balance problem difficulty; and (ii) a lack of reasoning in problem generation, leading to shallow problem variants. In this paper, we develop a problem generator that reasons explicitly to plan problem directions before synthesis and adapts difficulty to the solver’s ability. Specifically, we construct related problem pairs and augment them with intermediate problem-design CoT produced by a reasoning model. These data bootstrap problem-design strategies from the generator. Then, we treat the solver’s feedback on synthetic problems as a reward signal, enabling the generator to calibrate difficulty and produce complementary problems near the edge of the solver’s competence. Extensive experiments on 10 mathematical and general reasoning benchmarks show that our method achieves an average improvement of 2.5% and generalizes to both language and vision-language models. Moreover, a solver trained on the synthesized data provides improved rewards for continued generator training, enabling co-evolution and yielding a further 0.7% performance gain. Our code will be made publicly available here.

Method: LAMP uses a Think-Speak-Decide pipeline: (1) Think interprets numerical observations to extract shocks/trends and caches reasoning trajectories; (2) Speak crafts/exchanges strategic messages and updates beliefs by parsing peer communications; (3) Decide fuses numerical data, reasoning, and reflections into a MARL policy.

Result: LAMP outperforms MARL and LLM-only baselines in economic simulations: cumulative return (+63.5%, +34.0%), robustness (+18.8%, +59.4%), and interpretability.

Conclusion: Language-augmented policies show potential for more effective and robust economic strategies by bridging the gap between structured signals and unstructured language in decision-making.

Abstract: Economic decision-making depends not only on structured signals such as prices and taxes, but also on unstructured language, including peer dialogue and media narratives. While multi-agent reinforcement learning (MARL) has shown promise in optimizing economic decisions, it struggles with the semantic ambiguity and contextual richness of language. We propose LAMP (Language-Augmented Multi-Agent Policy), a framework that integrates language into economic decision-making and narrows the gap to real-world settings. LAMP follows a Think-Speak-Decide pipeline: (1) Think interprets numerical observations to extract short-term shocks and long-term trends, caching high-value reasoning trajectories; (2) Speak crafts and exchanges strategic messages based on reasoning, updating beliefs by parsing peer communications; and (3) Decide fuses numerical data, reasoning, and reflections into a MARL policy to optimize language-augmented decision-making. Experiments in economic simulation show that LAMP outperforms both MARL and LLM-only baselines in cumulative return (+63.5%, +34.0%), robustness (+18.8%, +59.4%), and interpretability. These results demonstrate the potential of language-augmented policies to deliver more effective and robust economic strategies.

Method: Proposes IPR (Interactive Physical Reasoner) that uses world-model rollouts to score and reinforce a VLM’s policy, and introduces PhysCode - a physics-centric action code that aligns semantic intent with dynamics to provide a shared action space for prediction and reasoning.

Result: IPR performs robustly on levels from primitive intuition to goal-driven reasoning, surpasses GPT-5 overall, improves with more training games and interaction steps, and demonstrates zero-shot transfer to unseen games in the Game-to-Unseen (G2U) benchmark.

Conclusion: Physics-centric interaction provides a path to steadily improving physical reasoning, enabling agents to acquire human-like reasoning capabilities through experience and transfer learning to novel scenarios.

Abstract: Humans learn by observing, interacting with environments, and internalizing physics and causality. Here, we aim to ask whether an agent can similarly acquire human-like reasoning from interaction and keep improving with more experience. To study this, we introduce a Game-to-Unseen (G2U) benchmark of 1,000+ heterogeneous games that exhibit significant visual domain gaps. Existing approaches, including VLMs and world models, struggle to capture underlying physics and causality since they are not focused on core mechanisms and overfit to visual details. VLM/VLA agents reason but lack look-ahead in interactive settings, while world models imagine but imitate visual patterns rather than analyze physics and causality. We therefore propose IPR (Interactive Physical Reasoner), using world-model rollouts to score and reinforce a VLM’s policy, and introduce PhysCode, a physics-centric action code aligning semantic intent with dynamics to provide a shared action space for prediction and reasoning. Pretrained on 1,000+ games, our IPR performs robustly on levels from primitive intuition to goal-driven reasoning, and even surpasses GPT-5 overall. We find that performance improves with more training games and interaction steps, and that the model also zero-shot transfers to unseen games. These results support physics-centric interaction as a path to steadily improving physical reasoning. Further demos and project details can be found at https://mybearyzhang.github.io/ipr-1.

Method: The study curates a dataset of 200 scientific papers and adapts 15 domain-specific attack strategies for adversarial PDF manipulation. They evaluate these attacks across 13 language models (including GPT-5, Claude Haiku, and DeepSeek) and develop a novel evaluation metric called WAVS (Weighted Adversarial Vulnerability Score) to measure system vulnerability.

Result: Results show that obfuscation strategies like “Maximum Mark Magyk” successfully manipulate scores and achieve alarming decision flip rates, even in large-scale models. The attacks can effectively change “Reject” decisions to “Accept” in AI-powered peer review systems.

Conclusion: The study demonstrates significant vulnerabilities in LLM-based peer review systems to adversarial PDF manipulation, highlighting security risks in the growing adoption of AI for scientific assessment. The researchers will release their complete dataset and injection framework to facilitate further research on this critical issue.

Abstract: The landscape of scientific peer review is rapidly evolving with the integration of Large Language Models (LLMs). This shift is driven by two parallel trends: the widespread individual adoption of LLMs by reviewers to manage workload (the “Lazy Reviewer” hypothesis) and the formal institutional deployment of AI-powered assessment systems by conferences like AAAI and Stanford’s Agents4Science. This study investigates the robustness of these “LLM-as-a-Judge” systems (both illicit and sanctioned) to adversarial PDF manipulation. Unlike general jailbreaks, we focus on a distinct incentive: flipping “Reject” decisions to “Accept,” for which we develop a novel evaluation metric which we term as WAVS (Weighted Adversarial Vulnerability Score). We curated a dataset of 200 scientific papers and adapted 15 domain-specific attack strategies to this task, evaluating them across 13 Language Models, including GPT-5, Claude Haiku, and DeepSeek. Our results demonstrate that obfuscation strategies like “Maximum Mark Magyk” successfully manipulate scores, achieving alarming decision flip rates even in large-scale models. We will release our complete dataset and injection framework to facilitate more research on this topic.

Method: Introduces similarity-driven alignment with serialized tool calls, signature embeddings using sentence encoder, and similarity-bucketed Hungarian matching for auditable one-to-one correspondences. Uses Executor & Judge pipeline with human verification and LLM judge ensemble for evaluation.

Result: Evaluation of state-of-the-art MLLMs reveals persistent gaps in multimodal MCP tool use, particularly in argument fidelity and structure consistency, showing current methods struggle with joint reasoning over images, text, and tool graphs.

Conclusion: The benchmark highlights the need for improved methods that can jointly reason over multimodal inputs and tool graphs, and provides a standardized evaluation framework for multimodal tool use with 28 servers and 231 tools.

Abstract: We present M^3-Bench, the first benchmark for evaluating multimodal tool use under the Model Context Protocol. The benchmark targets realistic, multi-hop and multi-threaded workflows that require visual grounding and textual reasoning, cross-tool dependencies, and persistence of intermediate resources across steps. We introduce a similarity-driven alignment that serializes each tool call, embeds signatures with a sentence encoder, and performs similarity-bucketed Hungarian matching to obtain auditable one-to-one correspondences. On top of this alignment, we report interpretable metrics that decouple semantic fidelity from workflow consistency. The benchmark spans 28 servers with 231 tools, and provides standardized trajectories curated through an Executor & Judge pipeline with human verification; an auxiliary four large language models (LLMs) judge ensemble reports end-task Task Completion and information grounding. Evaluations of representative state-of-the-art Multimodal LLMs (MLLMs) reveal persistent gaps in multimodal MCP tool use, particularly in argument fidelity and structure consistency, underscoring the need for methods that jointly reason over images, text, and tool graphs. Our Benchmark’s anonymous repository is at https://github.com/EtaYang10th/Open-M3-Bench

Method: GROVE learns and organizes debugging knowledge into a hierarchical tree with configurable depth. Each node contains concise knowledge items and explicit applicability conditions. During training, LLMs propose tree modifications as structured JSON edits. At test time, budget-aware iterative zoom navigates the tree to retrieve applicable knowledge that guides hypothesis generation and fix proposals.

Result: GROVE delivers consistent gains in pass@1 and pass@5 metrics when evaluated on assertion-failure cases, demonstrating the value of structured knowledge evolution.

Conclusion: The hierarchical knowledge management framework effectively captures and organizes reusable debugging expertise, improving LLM performance in hardware verification debugging tasks.

Abstract: Debugging is the dominant cost in modern hardware verification, where assertion failures are among the most frequent and expensive to resolve. While Large Language Models (LLMs) show promise, they often fail to capture the precise, reusable expertise that engineers apply, leading to inaccurate responses. We propose GROVE, a hierarchical knowledge management framework that learns and organizes reusable debugging expertise into an LLM-organized knowledge tree for solving assertion failures. GROVE distills debugging knowledge from prior cases and organizes it into a vertical tree of configurable depth, with each node encoding a concise knowledge item and explicit applicability conditions. During training, GROVE uses a parallel, gradient-free loop where an LLM proposes tree modifications as structured JSON edits by learning from the cases. At test time, a budget-aware iterative zoom is performed to navigate the tree, retrieving a small set of applicable knowledge items that guide a base LLM’s hypothesis generation and fix proposals. Evaluated on a suite of assertion-failure cases, GROVE delivers consistent gains in pass@1 and pass@5, demonstrating the value of structured knowledge evolution.

Method: Systematic experimentation with GPT-2 fine-tuned on The Psychology of Artificial Superintelligence, evaluating five locality configurations: two uniform baselines (fully distributed/fully localist) and three progressive polynomial schedules with adaptive semantic block partitioning.

Result: Progressive semantic localization achieves near-baseline language modeling performance while providing interpretable attention patterns, dramatically outperforming fixed-window localization and naive uniform locality constraints.

Conclusion: Interpretability mechanisms should align with semantic structure to achieve practical performance-interpretability tradeoffs for trustworthy AI systems, with steep schedules concentrating locality in decision-critical final layers while preserving distributed learning in early layers.

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 five locality configurations: two uniform baselines (fully distributed and fully localist) and three progressive polynomial schedules. 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: Common-garden experimental design auditing 16 open-weight models (4B-671B parameters) across 19,200 trials, testing disclosure rates when models adopt professional personas, with Bayesian validation for robustness and additional experiments testing explicit permission effects.

Result: Models showed sharp domain-specific inconsistency: Financial Advisor persona elicited 30.8% disclosure vs Neurosurgeon’s 3.5% (8.8-fold difference). Disclosure ranged 2.8%-73.6% across model families, with model identity explaining more variance than parameter count. Explicit permission increased disclosure from 23.7% to 65.8%.

Conclusion: Self-transparency failures reflect instruction-following prioritization over safety boundaries, not capability limitations. Organizations cannot assume safety properties transfer across domains, requiring deliberate behavior design and empirical verification.

Abstract: Self-transparency is a critical safety boundary, requiring language models to honestly disclose their limitations and artificial nature. This study stress-tests this capability, investigating whether models willingly disclose their identity when assigned professional personas that conflict with transparent self-representation. When models prioritize role consistency over this boundary disclosure, users may calibrate trust based on overstated competence claims, treating AI-generated guidance as equivalent to licensed professional advice. Using a common-garden experimental design, sixteen open-weight models (4B-671B parameters) were audited under identical conditions across 19,200 trials. Models exhibited sharp domain-specific inconsistency: a Financial Advisor persona elicited 30.8% disclosure at the first prompt, while a Neurosurgeon persona elicited only 3.5% – an 8.8-fold difference that emerged at the initial epistemic inquiry. Disclosure ranged from 2.8% to 73.6% across model families, with a 14B model reaching 39.4% while a 70B model produced just 4.1%. Model identity provided substantially larger improvement in fitting observations than parameter count ($ΔR_{adj}^{2}=0.359$ vs $0.018$). Reasoning variants showed heterogeneous effects: some exhibited up to 48.4 percentage points lower disclosure than their base instruction-tuned counterparts, while others maintained high transparency. An additional experiment demonstrated that explicit permission to disclose AI nature increased disclosure from 23.7% to 65.8%, revealing that suppression reflects instruction-following prioritization rather than capability limitations. Bayesian validation confirmed robustness to judge measurement error ($κ=0.908$). Organizations cannot assume safety properties will transfer across deployment domains, requiring deliberate behavior design and empirical verification.

Method: RL-Struct uses Gradient Regularized Policy Optimization (GRPO) with hierarchical reward function, eliminating critic network for efficiency. Framework aligns LLMs with structural constraints.

Result: Achieves 89.7% structural accuracy and 92.1% validity on complex JSON tasks, outperforming SFT and zero-shot baselines. Reduces peak VRAM by 38% compared to PPO.

Conclusion: RL-Struct effectively bridges structure gap, enables emergent curriculum learning (syntax before semantics), and provides efficient structural alignment for LLMs.

Abstract: The Structure Gap between probabilistic LLM generation and deterministic schema requirements hinders automated workflows. We propose RL-Struct, a lightweight framework using Gradient Regularized Policy Optimization (GRPO) with a hierarchical reward function to align LLMs with structural constraints. This approach eliminates the critic network, reducing peak VRAM by 38% compared to PPO. On complex JSON tasks, RL-Struct achieves 89.7% structural accuracy and 92.1% validity, significantly outperforming SFT and zero-shot baselines. We also report an emergent curriculum–a self-organized learning process where the model prioritizes syntax before semantics. Our model is publicly available at https://huggingface.co/Freakz3z/Qwen-JSON.

Method: Proposed framework decomposes mediation into judgment (evaluating conversation fairness and emotional dynamics) and steering (generating empathetic, de-escalatory messages). Created large Reddit-based dataset and multi-stage evaluation pipeline combining principle-based scoring, user simulation, and human comparison.

Result: API-based models outperform open-source counterparts in both reasoning and intervention alignment for mediation tasks, showing better capability in understanding and de-escalating online conflicts.

Conclusion: Current LLMs show promise as emerging agents for online social mediation but also have limitations, highlighting both their potential and current constraints in fostering constructive dialogue.

Abstract: The rapid advancement of large language models (LLMs) has opened new possibilities for AI for good applications. As LLMs increasingly mediate online communication, their potential to foster empathy and constructive dialogue becomes an important frontier for responsible AI research. This work explores whether LLMs can serve not only as moderators that detect harmful content, but as mediators capable of understanding and de-escalating online conflicts. Our framework decomposes mediation into two subtasks: judgment, where an LLM evaluates the fairness and emotional dynamics of a conversation, and steering, where it generates empathetic, de-escalatory messages to guide participants toward resolution. To assess mediation quality, we construct a large Reddit-based dataset and propose a multi-stage evaluation pipeline combining principle-based scoring, user simulation, and human comparison. Experiments show that API-based models outperform open-source counterparts in both reasoning and intervention alignment when doing mediation. Our findings highlight both the promise and limitations of current LLMs as emerging agents for online social mediation.

Method: Developed a Fuzzy Logic DSS that reformulates PlayeRank into a Cumulative Mean with Role-Aware Normalization to eliminate playtime bias, then integrates this with physiological fatigue proxies and contextual variables (disciplinary risk modulated by tactical role) to calculate dynamic Substitution Priority scores.

Result: Validation using Brazil vs Belgium 2018 World Cup match showed the system aligned with expert consensus on executed substitutions and crucially identified high-risk scenarios ignored by human decision-makers, including the “FAGNER Paradox” (defensive risk before critical yellow card) and “Lukaku Paradox” (isolated assist masking severe participation drop).

Conclusion: Fuzzy Logic provides a transparent, explainable, and superior alternative to black-box models for optimizing real-time tactical decisions in soccer, offering objective decision support that can identify critical risks before they impact game outcomes.

Abstract: In elite soccer, substitution decisions entail significant financial and sporting consequences yet remain heavily reliant on intuition or predictive models that merely mimic historical biases. This paper introduces a Fuzzy Logic based Decision Support System (DSS) designed for real time, prescriptive game management. Unlike traditional Machine Learning approaches that encounter a predictive ceiling by attempting to replicate human behavior, our system audits performance through an objective, rule based inference engine. We propose a methodological advancement by reformulating the PlayeRank metric into a Cumulative Mean with Role Aware Normalization, eliminating the play time exposure bias inherent in cumulative sum models to enable accurate intra match comparison. The system integrates this refined metric with physiological proxies (fatigue) and contextual variables (disciplinary risk modulated by tactical role) to calculate a dynamic Substitution Priority (P final). Validation via a case study of the 2018 FIFA World Cup match between Brazil and Belgium demonstrates the system’s ecological validity: it not only aligned with expert consensus on executed substitutions (for example Gabriel Jesus) but, crucially, identified high risk scenarios ignored by human decision makers. Specifically, the model flagged the “FAGNER Paradox” - a maximum priority defensive risk - minutes before a critical yellow card, and detected the “Lukaku Paradox”, where an isolated assist masked a severe drop in participation. These results confirm that Fuzzy Logic offers a transparent, explainable, and superior alternative to black box models for optimizing real time tactical decisions.

Method: Focus on developing AI systems specifically designed to collaborate with human researchers throughout the entire AI research process - from ideation to experimentation. This involves creating symbiotic partnerships where both humans and AIs improve each other’s capabilities.

Result: The approach promises to accelerate AI research progress while making the path to superintelligence safer by keeping human researchers in the loop and focusing on co-improvement rather than pure AI self-improvement.

Conclusion: Co-improvement through human-AI collaboration is a more achievable and safer goal than pure AI self-improvement, offering a path to co-superintelligence that benefits both humans and AIs while maintaining safety through continued human involvement.

Abstract: Self-improvement is a goal currently exciting the field of AI, but is fraught with danger, and may take time to fully achieve. We advocate that a more achievable and better goal for humanity is to maximize co-improvement: collaboration between human researchers and AIs to achieve co-superintelligence. That is, specifically targeting improving AI systems’ ability to work with human researchers to conduct AI research together, from ideation to experimentation, in order to both accelerate AI research and to generally endow both AIs and humans with safer superintelligence through their symbiosis. Focusing on including human research improvement in the loop will both get us there faster, and more safely.

Method: Used 382 soil samples from Türkiye with physicochemical properties. Tested 12 ML algorithms including decision tree, random forest, gradient boosting, XGBoost, SVM, neural networks, and ensemble methods. Models were trained, validated, and evaluated for generalization.

Result: Random forest regressor performed best with R² scores: 0.95 (training), 0.76 (validation), 0.83 (test). Demonstrates strong nonlinear mapping capability for CBR prediction.

Conclusion: ML models, particularly random forest, provide effective alternatives to traditional CBR testing, supporting digital transformation in geotechnical engineering and infrastructure design.

Abstract: The California Bearing Ratio (CBR) is a key geotechnical indicator used to assess the load-bearing capacity of subgrade soils, especially in transportation infrastructure and foundation design. Traditional CBR determination relies on laboratory penetration tests. Despite their accuracy, these tests are often time-consuming, costly, and can be impractical, particularly for large-scale or diverse soil profiles. Recent progress in artificial intelligence, especially machine learning (ML), has enabled data-driven approaches for modeling complex soil behavior with greater speed and precision. This study introduces a comprehensive ML framework for CBR prediction using a dataset of 382 soil samples collected from various geoclimatic regions in Türkiye. The dataset includes physicochemical soil properties relevant to bearing capacity, allowing multidimensional feature representation in a supervised learning context. Twelve ML algorithms were tested, including decision tree, random forest, extra trees, gradient boosting, xgboost, k-nearest neighbors, support vector regression, multi-layer perceptron, adaboost, bagging, voting, and stacking regressors. Each model was trained, validated, and evaluated to assess its generalization and robustness. Among them, the random forest regressor performed the best, achieving strong R2 scores of 0.95 (training), 0.76 (validation), and 0.83 (test). These outcomes highlight the model’s powerful nonlinear mapping ability, making it a promising tool for predictive geotechnical tasks. The study supports the integration of intelligent, data-centric models in geotechnical engineering, offering an effective alternative to traditional methods and promoting digital transformation in infrastructure analysis and design.

Method: Proposes constructive incremental transfer learning (CITL) with semi-supervised TL mechanism using unlabeled target data. Minimizes monitoring residual by iteratively adding network nodes. Ensures cross-domain learning through structural risk minimization, transfer mismatching minimization, and manifold consistency maximization. Includes convergence analysis for theoretical guarantees.

Result: Outperforms SS-TCA, MMD-LSTM-DA, DDAN, BO-CNN-TL, and AS$^3$LSTM by 83.73%, 61.15%, 28.24%, 87.70%, and 57.34% respectively in SOH estimation using root mean square error. Validated on realistic autonomous air vehicles battery dataset from flight missions.

Conclusion: CITL provides effective lightweight transfer learning solution for battery SOH monitoring in portable devices, balancing computational efficiency with accurate cross-domain knowledge transfer while theoretically guaranteeing performance and network compactness.

Abstract: Accurate and rapid state-of-health (SOH) monitoring plays an important role in indicating energy information for lithium-ion battery-powered portable mobile devices. To confront their variable working conditions, transfer learning (TL) emerges as a promising technique for leveraging knowledge from data-rich source working conditions, significantly reducing the training data required for SOH monitoring from target working conditions. However, traditional TL-based SOH monitoring is infeasible when applied in portable mobile devices since substantial computational resources are consumed during the TL stage and unexpectedly reduce the working endurance. To address these challenges, this paper proposes a lightweight TL-based SOH monitoring approach with constructive incremental transfer learning (CITL). First, taking advantage of the unlabeled data in the target domain, a semi-supervised TL mechanism is proposed to minimize the monitoring residual in a constructive way, through iteratively adding network nodes in the CITL. Second, the cross-domain learning ability of node parameters for CITL is comprehensively guaranteed through structural risk minimization, transfer mismatching minimization, and manifold consistency maximization. Moreover, the convergence analysis of the CITL is given, theoretically guaranteeing the efficacy of TL performance and network compactness. Finally, the proposed approach is verified through extensive experiments with a realistic autonomous air vehicles (AAV) battery dataset collected from dozens of flight missions. Specifically, the CITL outperforms SS-TCA, MMD-LSTM-DA, DDAN, BO-CNN-TL, and AS$^3$LSTM, in SOH estimation by 83.73%, 61.15%, 28.24%, 87.70%, and 57.34%, respectively, as evaluated using the index root mean square error.

Method: Proposes representing graphs by transforming adjacency matrices into strings of simple instructions that build the matrix step by step. The transformation is reversible (graph ↔ string) and maintains local structural patterns while being compact.

Result: Tentative computational experiments show improved classification performance and faster computation times compared to traditional graph representations when using the proposed method with deep learning models.

Conclusion: The instruction-based graph representation bridges the gap between graph processing and deep learning language models, offering a promising approach to boost graph analysis capabilities using modern AI tools.

Abstract: The representation of graphs is commonly based on the adjacency matrix concept. This formulation is the foundation of most algebraic and computational approaches to graph processing. The advent of deep learning language models offers a wide range of powerful computational models that are specialized in the processing of text. However, current procedures to represent graphs are not amenable to processing by these models. In this work, a new method to represent graphs is proposed. It represents the adjacency matrix of a graph by a string of simple instructions. The instructions build the adjacency matrix step by step. The transformation is reversible, i.e., given a graph the string can be produced and vice versa. The proposed representation is compact, and it maintains the local structural patterns of the graph. Therefore, it is envisaged that it could be useful to boost the processing of graphs by deep learning models. A tentative computational experiment is reported, demonstrating improved classification performance and faster computation times with the proposed representation.

Method: Phythesis uses iterative bi-level optimization: (1) LLM-driven optimization generates 3D layouts and self-criticizes to refine scene topology, (2) physics-informed optimization identifies optimal asset parameters and selects best asset combinations.

Result: Experiments across three generation scales show Phythesis achieves 57.3% generation success rate increase and 11.5% power usage effectiveness (PUE) improvement compared to vanilla LLM-based solutions.

Conclusion: The framework successfully bridges the gap between generative AI and physics-based design by synergizing LLMs with evolutionary optimization, enabling automated simulation-ready scene synthesis for energy-efficient data center design.

Abstract: Data center (DC) infrastructure serves as the backbone to support the escalating demand for computing capacity. Traditional design methodologies that blend human expertise with specialized simulation tools scale poorly with the increasing system complexity. Recent studies adopt generative artificial intelligence to design plausible human-centric indoor layouts. However, they do not consider the underlying physics, making them unsuitable for the DC design that sets quantifiable operational objectives and strict physical constraints. To bridge the gap, we propose Phythesis, a novel framework that synergizes large language models (LLMs) and physics-guided evolutionary optimization to automate simulation-ready (SimReady) scene synthesis for energy-efficient DC design. Phythesis employs an iterative bi-level optimization architecture, where (i) the LLM-driven optimization level generates physically plausible three-dimensional layouts and self-criticizes them to refine the scene topology, and (ii) the physics-informed optimization level identifies the optimal asset parameters and selects the best asset combination. Experiments on three generation scales show that Phythesis achieves 57.3% generation success rate increase and 11.5% power usage effectiveness (PUE) improvement, compared with the vanilla LLM-based solution.

Method: EmeraldMind integrates a domain-specific knowledge graph (EmeraldGraph) built from diverse corporate ESG reports with retrieval-augmented generation. This approach surfaces verifiable evidence often missing in generic knowledge bases and supports large language models in claim assessment.

Result: Experiments on a new greenwashing claims dataset show EmeraldMind achieves competitive accuracy, greater coverage, and superior explanation quality compared to generic LLMs, without requiring fine-tuning or retraining.

Conclusion: The framework successfully addresses greenwashing detection by providing justification-centric classifications with transparent, evidence-backed verdicts, and responsibly abstaining when claims cannot be verified.

Abstract: As AI and web agents become pervasive in decision-making, it is critical to design intelligent systems that not only support sustainability efforts but also guard against misinformation. Greenwashing, i.e., misleading corporate sustainability claims, poses a major challenge to environmental progress. To address this challenge, we introduce EmeraldMind, a fact-centric framework integrating a domain-specific knowledge graph with retrieval-augmented generation to automate greenwashing detection. EmeraldMind builds the EmeraldGraph from diverse corporate ESG (environmental, social, and governance) reports, surfacing verifiable evidence, often missing in generic knowledge bases, and supporting large language models in claim assessment. The framework delivers justification-centric classifications, presenting transparent, evidence-backed verdicts and abstaining responsibly when claims cannot be verified. Experiments on a new greenwashing claims dataset demonstrate that EmeraldMind achieves competitive accuracy, greater coverage, and superior explanation quality compared to generic LLMs, without the need for fine-tuning or retraining.

Method: Cannot determine method due to API access failure preventing retrieval of paper content.

Result: Cannot determine results due to API access failure preventing retrieval of paper content.

Conclusion: Cannot draw conclusions due to API access failure preventing retrieval of paper content.

Abstract: Failed to fetch summary for 2502.16840: Page request resulted in HTTP 429 (https://export.arxiv.org/api/query?search_query=&id_list=2502.16840&sortBy=relevance&sortOrder=descending&start=0&max_results=100)

Method: Could not analyze method due to missing abstract content

Result: API request failed with HTTP 429 status (Too Many Requests), indicating rate limiting on arXiv’s export service

Conclusion: Paper analysis not possible due to technical limitations in accessing the abstract; need to try again later or use alternative access methods

Abstract: Failed to fetch summary for 2503.10253: Page request resulted in HTTP 429 (https://export.arxiv.org/api/query?search_query=&id_list=2503.10253&sortBy=relevance&sortOrder=descending&start=0&max_results=100)

Method: Cannot determine method due to inability to access the paper content.

Result: Cannot determine results due to inability to access the paper content.

Conclusion: Cannot draw conclusions due to inability to access the paper content.

Abstract: Failed to fetch summary for 2503.11950: Page request resulted in HTTP 429 (https://export.arxiv.org/api/query?search_query=&id_list=2503.11950&sortBy=relevance&sortOrder=descending&start=0&max_results=100)

Method: No method information available - the analysis request encountered an HTTP 429 error (Too Many Requests) when attempting to fetch the paper summary from arXiv.

Result: The only result obtained is that the arXiv API request failed with HTTP 429 status, suggesting rate limiting or server capacity issues preventing access to paper 2504.03494.

Conclusion: Technical limitations prevented analysis of paper 2504.03494; the arXiv API returned HTTP 429 error indicating too many requests, requiring alternative access methods or waiting before retrying.

Abstract: Failed to fetch summary for 2504.03494: Page request resulted in HTTP 429 (https://export.arxiv.org/api/query?search_query=&id_list=2504.03494&sortBy=relevance&sortOrder=descending&start=0&max_results=100)

Method: Cannot analyze method due to inability to access paper abstract (HTTP 429 error)

Result: Failed to retrieve paper information - arXiv API returned HTTP 429 (Too Many Requests) error

Conclusion: Unable to provide analysis due to technical limitations in accessing the paper abstract from arXiv

Abstract: Failed to fetch summary for 2505.06256: Page request resulted in HTTP 429 (https://export.arxiv.org/api/query?search_query=&id_list=2505.06256&sortBy=relevance&sortOrder=descending&start=0&max_results=100)

Method: Cannot analyze method due to inability to access paper content

Result: Failed to retrieve paper summary; encountered HTTP 429 error from arXiv API

Conclusion: Paper analysis impossible without content; need to retry later or use alternative access methods

Abstract: Failed to fetch summary for 2505.16130: Page request resulted in HTTP 429 (https://export.arxiv.org/api/query?search_query=&id_list=2505.16130&sortBy=relevance&sortOrder=descending&start=0&max_results=100)

Method: No method information available - API request resulted in HTTP 429 error indicating too many requests

Result: No results available - the paper content could not be fetched due to rate limiting on arXiv API

Conclusion: Unable to analyze paper content due to technical limitations in accessing the abstract

Abstract: Failed to fetch summary for 2505.18231: Page request resulted in HTTP 429 (https://export.arxiv.org/api/query?search_query=&id_list=2505.18231&sortBy=relevance&sortOrder=descending&start=0&max_results=100)

Method: N/A - Paper content not accessible due to HTTP 429 error when attempting to fetch from arXiv API

Result: HTTP 429 error (Too Many Requests) received when trying to access arXiv paper 2505.19609 via API

Conclusion: Cannot analyze paper content due to technical limitations (arXiv API rate limiting); need to try again later or use alternative access methods

Abstract: Failed to fetch summary for 2505.19609: Page request resulted in HTTP 429 (https://export.arxiv.org/api/query?search_query=&id_list=2505.19609&sortBy=relevance&sortOrder=descending&start=0&max_results=100)

Method: N/A - No method information available due to failed API request

Result: HTTP 429 error received when attempting to query arXiv API for paper ID 2505.24592

Conclusion: Cannot analyze paper content due to technical limitations in accessing the abstract/summary

Abstract: Failed to fetch summary for 2505.24592: Page request resulted in HTTP 429 (https://export.arxiv.org/api/query?search_query=&id_list=2505.24592&sortBy=relevance&sortOrder=descending&start=0&max_results=100)

Method: Attempted to fetch paper summary using arXiv API query with specific paper ID, but encountered HTTP 429 status indicating too many requests

Result: Failed to retrieve paper content due to API rate limiting restrictions

Conclusion: Cannot analyze the paper as the abstract content is unavailable due to technical limitations with the arXiv API service

Abstract: Failed to fetch summary for 2506.01983: Page request resulted in HTTP 429 (https://export.arxiv.org/api/query?search_query=&id_list=2506.01983&sortBy=relevance&sortOrder=descending&start=0&max_results=100)

Method: Compared four generative models (diffusion, flow-based, variational autoencoders, GANs) for adult-to-child voice conversion. Introduced an efficient frequency warping technique to reduce adult-child acoustic mismatch. Evaluated using objective and subjective measures on a unique children’s dubbing corpus.

Result: All four generative models produced plausible speech outputs but exhibited insufficient similarity to target child speaker characteristics. The proposed frequency warping technique significantly reduced the mismatch between adult and child speech characteristics.

Conclusion: Generative models alone are insufficient for high-quality adult-to-child voice conversion due to speaker similarity limitations, but the proposed frequency warping technique effectively addresses the adult-child acoustic mismatch, improving conversion quality.

Abstract: Generative models are a popular choice for adult-to-adult voice conversion (VC) because of their efficient way of modelling unlabelled data. To this point their usefulness in producing children speech and in particular adult to child VC has not been investigated. For adult to child VC, four generative models are compared: diffusion model, flow based model, variational autoencoders, and generative adversarial network. Results show that although converted speech outputs produce by those models appear plausible, they exhibit insufficient similarity with the target speaker characteristics. We introduce an efficient frequency warping technique that can be applied to the output of models, and which shows significant reduction of the mismatch between adult and child. The output of all the models are evaluated using both objective and subjective measures. In particular we compare specific speaker pairing using a unique corpus collected for dubbing of children speech.

Method: Develops privacy-preserving systems that leverage ambient audio sensing through existing microphones to estimate key transmission risk factors in real time.

Result: Establishes foundation for privacy-aware airborne risk monitoring using everyday devices to monitor the whole spectrum of indoor air quality non-invasively.

Conclusion: Ambient audio sensing offers untapped potential for real-time, non-invasive monitoring of indoor airborne transmission risks while preserving privacy through existing infrastructure.

Abstract: Indoor airborne transmission poses a significant health risk, yet current monitoring solutions are invasive, costly, or fail to address it directly. My research explores the untapped potential of ambient audio sensing to estimate key transmission risk factors such as ventilation, aerosol emissions, and occupant distribution non-invasively and in real time. I develop privacy-preserving systems that leverage existing microphones to monitor the whole spectrum of indoor air quality which can have a significant effect on an individual’s health. This work lays the foundation for privacy-aware airborne risk monitoring using everyday devices.

Method: ASTA integrates on-device ASR and lightweight language models with cloud-based LLM processing, using real-time metrics (CPU workload, temperature, network latency) to dynamically route commands. Includes rule-based command validation and repair for robust execution.

Result: ASTA successfully routes all 80 test commands, achieving 62.5% ASR accuracy and generating executable commands without repair for 47.5% of inputs, demonstrating the importance of the repair mechanism for robustness.

Conclusion: Adaptive edge-cloud orchestration is a viable approach for creating resilient and resource-aware voice-controlled IoT systems that balance performance and resource utilization.

Abstract: Voice-based interaction has emerged as a natural and intuitive modality for controlling IoT devices. However, speech-driven edge devices face a fundamental trade-off between cloud-based solutions, which offer stronger language understanding capabilities at the cost of latency, connectivity dependence, and privacy concerns, and edge-based solutions, which provide low latency and improved privacy but are limited by computational constraints. This paper presents ASTA, an adaptive speech-to-action solution that dynamically routes voice commands between edge and cloud inference to balance performance and system resource utilization. ASTA integrates on-device automatic speech recognition and lightweight offline language-model inference with cloud-based LLM processing, guided by real-time system metrics such as CPU workload, device temperature, and network latency. A metric-aware routing mechanism selects the inference path at runtime, while a rule-based command validation and repair component ensures successful end-to-end command execution. We implemented our solution on an NVIDIA Jetson-based edge platform and evaluated it using a diverse dataset of 80 spoken commands. Experimental results show that ASTA successfully routes all input commands for execution, achieving a balanced distribution between online and offline inference. The system attains an ASR accuracy of 62.5% and generates executable commands without repair for only 47.5% of inputs, highlighting the importance of the repair mechanism in improving robustness. These results suggest that adaptive edge-cloud orchestration is a viable approach for resilient and resource-aware voice-controlled IoT systems.

Method: The paper presents a systematic overview of current PMG techniques in both research and application fields, develops a two-aspect taxonomy, and conducts comparative analysis to identify key challenges and opportunities.

Result: Identifies key research challenges in algorithm implementation, music quality, and game integration. Provides a taxonomy and comparative framework that reveals differences between academic and applied approaches to PMG.

Conclusion: Future PMG research should focus on task-oriented and context-aware design, develop more comprehensive quality evaluation methods, and improve research tool integration to advance PMG in practical game development contexts.

Abstract: Procedural Music Generation (PMG) is an emerging field that algorithmically creates music content for video games. By leveraging techniques from simple rule-based approaches to advanced machine learning algorithms, PMG has the potential to significantly improve development efficiency, provide richer musical experiences, and enhance player immersion. However, academic prototypes often diverge from applications due to differences in priorities such as novelty, reliability, and allocated resources. This paper bridges the gap between research and applications by presenting a systematic overview of current PMG techniques in both fields, offering a two-aspect taxonomy. Through a comparative analysis, this study identifies key research challenges in algorithm implementation, music quality and game integration. Finally, the paper outlines future research directions, emphasising task-oriented and context-aware design, more comprehensive quality evaluation methods, and improved research tool integration to provide actionable insights for developers, composers, and researchers seeking to advance PMG in game contexts.

Method: Constructed HQ-MPSD using linguistically coherent splice points from fine-grained forced alignment to preserve prosodic/semantic continuity and minimize boundary artifacts. Includes 350.8 hours across 8 languages/550 speakers with background effects for real-world acoustic diversity.

Result: MOS evaluations and spectrogram analysis confirm high perceptual naturalness. Benchmarking shows state-of-the-art detection models experience >80% performance drops on HQ-MPSD, exposing significant generalization challenges when low-level artifacts are removed and multilingual/acoustic diversity is introduced.

Conclusion: HQ-MPSD provides a more realistic and demanding benchmark for partial deepfake detection by removing dataset-specific artifacts and introducing multilingual/acoustic diversity, revealing that current models struggle with generalization when faced with high-quality manipulations.

Abstract: Detecting partial deepfake speech is challenging because manipulations occur only in short regions while the surrounding audio remains authentic. However, existing detection methods are fundamentally limited by the quality of available datasets, many of which rely on outdated synthesis systems and generation procedures that introduce dataset-specific artifacts rather than realistic manipulation cues. To address this gap, we introduce HQ-MPSD, a high-quality multilingual partial deepfake speech dataset. HQ-MPSD is constructed using linguistically coherent splice points derived from fine-grained forced alignment, preserving prosodic and semantic continuity and minimizing audible and visual boundary artifacts. The dataset contains 350.8 hours of speech across eight languages and 550 speakers, with background effects added to better reflect real-world acoustic conditions. MOS evaluations and spectrogram analysis confirm the high perceptual naturalness of the samples. We benchmark state-of-the-art detection models through cross-language and cross-dataset evaluations, and all models experience performance drops exceeding 80% on HQ-MPSD. These results demonstrate that HQ-MPSD exposes significant generalization challenges once low-level artifacts are removed and multilingual and acoustic diversity are introduced, providing a more realistic and demanding benchmark for partial deepfake detection. The dataset can be found at: https://zenodo.org/records/17929533.

Method: Uses DisCodec with tri-factor disentanglement (parallel encoders for content, prosody, timbre + hybrid losses) and fusion/reconstruction stages. LM handles prosodic continuation while decoder injects target timbre.

Result: Matches state-of-the-art voice cloning performance while outperforming baselines in zero-shot prosody control. Provides robust foundation for controllable speech synthesis.

Conclusion: By resolving core entanglement at codec level, DisCo-Speech enables flexible zero-shot control of prosody and voice cloning, advancing controllable TTS capabilities.

Abstract: Recent codec-based language models~(LMs) have revolutionized text-to-speech~(TTS). However, since standard codecs tightly couple timbre and prosody, continuation-based LMs inevitably replicate this entanglement, hindering independent control. Recent efforts attempt to break this entanglement via codec design, but insufficient decoupling remains a critical bottleneck. To tackle this challenge, we propose DisCo-Speech, a zero-shot controllable TTS framework that enables prosody control and voice cloning via a disentangled speech codec (DisCodec) and an LM-based generator. The core component, DisCodec, contains two core stages: 1) Tri-factor disentanglement, which explicitly factorizes speech into content, prosody, and timbre subspaces via parallel encoders and hybrid losses; and 2) Fusion and reconstruction, which fuses content and prosody into unified content-prosody tokens suitable for LM prediction, while jointly optimizing reconstruction quality to resolve the disentanglement-reconstruction trade-off. With this design, the LM performs prosodic continuation from a style prompt while the decoder handles target timbre injection, enabling flexible zero-shot control. Experiments show that DisCo-Speech matches state-of-the-art voice cloning performance while outperforming baselines in zero-shot prosody control. By resolving the core entanglement at the codec level, DisCo-Speech provides a robust foundation for controllable speech synthesis. Audio samples are available at https://github.com/disco-speech/DisCo-Speech, and the code and weights will be released at the same link.

Method: Used two generative AI models (ACGAN and DDPMs) to synthesize spectrogram data for bird calls. Created a new dataset of 640 hours of wind farm audio with 800 expert-labeled samples. Trained ensemble classification models on combinations of real and synthetic data, comparing results with BirdNET predictions.

Result: DDPMs performed particularly well in generating realistic spectrograms and improving classification accuracy. All classifiers improved with synthetic data, with metrics generally improving as more synthetic data was added. The ensemble approach compared favorably with BirdNET predictions.

Conclusion: Generative AI models, especially DDPMs, are effective for spectrogram data augmentation, enabling better AI-based detection of rare species in challenging acoustic environments like wind farms. The approach can be extended to other species and land-use types.

Abstract: Obtaining data to train robust artificial intelligence (AI)-based models for species classification can be challenging, particularly for rare species. Data augmentation can boost classification accuracy by increasing the diversity of training data and is cheaper to obtain than expert-labelled data. However, many classic image-based augmentation techniques are not suitable for audio spectrograms. We investigate two generative AI models as data augmentation tools to synthesise spectrograms and supplement audio data: Auxiliary Classifier Generative Adversarial Networks (ACGAN) and Denoising Diffusion Probabilistic Models (DDPMs). The latter performed particularly well in terms of both realism of generated spectrograms and accuracy in a resulting classification task. Alongside these new approaches, we present a new audio data set of 640 hours of bird calls from wind farm sites in Ireland, approximately 800 samples of which have been labelled by experts. Wind farm data are particularly challenging for classification models given the background wind and turbine noise. Training an ensemble of classification models on real and synthetic data combined compared well with highly confident BirdNET predictions. Each classifier we used was improved by including synthetic data, and classification metrics generally improved in line with the amount of synthetic data added. Our approach can be used to augment acoustic signals for more species and other land-use types, and has the potential to bring about advances in our capacity to develop reliable AI-based detection of rare species. Our code is available at https://github.com/gibbona1/SpectrogramGenAI.

Method: Developed a hardware system with STM32F407 microcontroller, 4 microphones, 128GB storage, 10400mAh battery with solar charging, and USB/Wi-Fi configuration capabilities for field deployment.

Result: Created SAMAY system capable of automatic bird call recording, supporting large-scale acoustic data collection for database creation and analysis applications.

Conclusion: SAMAY provides a practical, user-configurable field recording system for bird acoustic studies, enabling automated data collection for species classification and environmental monitoring applications.

Abstract: This paper describes an automatic bird call recording system called SAMAY, which is developed to study bird species by creating a database of large amounts of bird acoustic data. By analysing the recorded bird call data, the system can also be used for automatic classification of bird species, monitoring bird populations and analysing the impact of environmental changes. The system is driven through a powerful STM32F407 series microcontroller, supports 4 microphones, is equipped with 128 GB of storage capacity, and is powered by a 10400 mAh battery pack interfaced with a solar charger. In addition, the device is user-configurable over USB and Wi-Fi during runtime, ensuring user-friendly operation during field deployment.

Method: Uses CNN for spectral feature extraction and LSTM for temporal modeling with multimodal audio-visual fusion. Compares cloud-based, edge-assisted, and standalone deployments. Evaluates across Ethernet, Wi-Fi, 4G, and 5G networks with real-world experiments measuring speech quality, latency, and computational overhead.

Result: Cloud deployment achieves highest enhancement quality but with higher latency. Edge-assisted architectures provide best balance between latency and intelligibility, meeting real-time requirements under 5G and Wi-Fi 6 conditions. Standalone devices have lowest latency but limited enhancement quality.

Conclusion: Edge-assisted AVSE architectures offer optimal practical solution for real-time applications, providing guidelines for deployment in assistive hearing devices, telepresence, and industrial communications based on specific quality-latency requirements.

Abstract: This paper introduces a new AI-based Audio-Visual Speech Enhancement (AVSE) system and presents a comparative performance analysis of different deployment architectures. The proposed AVSE system employs convolutional neural networks (CNNs) for spectral feature extraction and long short-term memory (LSTM) networks for temporal modeling, enabling robust speech enhancement through multimodal fusion of audio and visual cues. Multiple deployment scenarios are investigated, including cloud-based, edge-assisted, and standalone device implementations. Their performance is evaluated in terms of speech quality improvement, latency, and computational overhead. Real-world experiments are conducted across various network conditions, including Ethernet, Wi-Fi, 4G, and 5G, to analyze the trade-offs between processing delay, communication latency, and perceptual speech quality. The results show that while cloud deployment achieves the highest enhancement quality, edge-assisted architectures offer the best balance between latency and intelligibility, meeting real-time requirements under 5G and Wi-Fi 6 conditions. These findings provide practical guidelines for selecting and optimizing AVSE deployment architectures in diverse applications, including assistive hearing devices, telepresence, and industrial communications.

Method: End-to-end framework operating on Modified Discrete Cosine Transform (MDCT) magnitudes using Swin Transformer-based U-Net to capture long-range dependencies. Employs hybrid adversarial scheme combining time-domain MPD/MSD discriminators with multi-band MDCT discriminator specialized for high-frequency band. Uses sparse-aware regularizer on arcsinh-compressed MDCT to preserve transient components.

Result: System upsamples inputs at various sampling rates to 48 kHz in single pass with real-time operation. Reduces objective error and improves ABX preference scores on standard benchmarks. In zero-shot tests on HiFi-TTS without fine-tuning, outperforms NVSR and mdctGAN, demonstrating strong generalization across datasets.

Conclusion: SwinSRGAN provides an effective end-to-end solution for speech super-resolution that addresses limitations of existing approaches, offering computational efficiency, robustness, and strong generalization while maintaining real-time performance.

Abstract: Speech super-resolution (SR) reconstructs high-frequency content from low-resolution speech signals. Existing systems often suffer from representation mismatch in two-stage mel-vocoder pipelines and from over-smoothing of hallucinated high-band content by CNN-only generators. Diffusion and flow models are computationally expensive, and their robustness across domains and sampling rates remains limited. We propose SwinSRGAN, an end-to-end framework operating on Modified Discrete Cosine Transform (MDCT) magnitudes. It is a Swin Transformer-based U-Net that captures long-range spectro-temporal dependencies with a hybrid adversarial scheme combines time-domain MPD/MSD discriminators with a multi-band MDCT discriminator specialized for the high-frequency band. We employs a sparse-aware regularizer on arcsinh-compressed MDCT to better preserve transient components. The system upsamples inputs at various sampling rates to 48 kHz in a single pass and operates in real time. On standard benchmarks, SwinSRGAN reduces objective error and improves ABX preference scores. In zero-shot tests on HiFi-TTS without fine-tuning, it outperforms NVSR and mdctGAN, demonstrating strong generalization across datasets

Method: 1) Created SH-Bench with 3,968 multi-speaker audio mixtures (real-world and synthetic) and 77k multiple-choice questions. 2) Introduced Selective Efficacy (SE) metric combining multi-speaker comprehension and bystander privacy protection. 3) Developed Bystander Privacy Fine-Tuning (BPFT) pipeline to teach models to refuse bystander-related queries while maintaining main-speaker comprehension.

Result: Evaluation showed substantial bystander privacy leakage in state-of-the-art audio LLMs. BPFT achieved 47% higher bystander accuracy under selective mode and 16% higher SE compared to Gemini 2.5 Pro (best audio LLM without BPFT). Strong audio understanding doesn’t guarantee selective privacy protection.

Conclusion: SH-Bench and BPFT provide the first systematic framework for measuring and improving bystander privacy in audio LLMs. The work addresses a critical gap in audio LLM deployment and establishes foundations for privacy-preserving selective hearing capabilities.

Abstract: Audio Large language models (LLMs) are increasingly deployed in the real world, where they inevitably capture speech from unintended nearby bystanders, raising privacy risks that existing benchmarks and defences did not consider. We introduce SH-Bench, the first benchmark designed to evaluate selective hearing: a model’s ability to attend to an intended main speaker while refusing to process or reveal information about incidental bystander speech. SH-Bench contains 3,968 multi-speaker audio mixtures, including both real-world and synthetic scenarios, paired with 77k multiple-choice questions that probe models under general and selective operating modes. In addition, we propose Selective Efficacy (SE), a novel metric capturing both multi-speaker comprehension and bystander-privacy protection. Our evaluation of state-of-the-art open-source and proprietary LLMs reveals substantial bystander privacy leakage, with strong audio understanding failing to translate into selective protection of bystander privacy. To mitigate this gap, we also present Bystander Privacy Fine-Tuning (BPFT), a novel training pipeline that teaches models to refuse bystander-related queries without degrading main-speaker comprehension. We show that BPFT yields substantial gains, achieving an absolute 47% higher bystander accuracy under selective mode and an absolute 16% higher SE compared to Gemini 2.5 Pro, which is the best audio LLM without BPFT. Together, SH-Bench and BPFT provide the first systematic framework for measuring and improving bystander privacy in audio LLMs.

Method: The study systematically investigates the effects of audio encoder and text-based LLM components in ALLMs. It proposes an optimized ALLM architecture through careful selection and combination of these components to enhance deepfake detection capabilities.

Result: The proposed ALLM architecture achieves state-of-the-art performance across multiple datasets (ASVSpoof2019, InTheWild, Demopage) with up to 95.76% average accuracy. It also demonstrates competitive capabilities in other deepfake tasks like spoof attribution and localization compared to SOTA audio understanding models.

Conclusion: Careful selection and combination of audio encoders and text-based LLMs are crucial for unlocking ALLMs’ deepfake detection potential. The proposed architecture successfully generalizes to out-of-domain spoofing tests and multiple deepfake tasks, addressing the generalization limitations of traditional methods.

Abstract: Audio deepfake detection has recently garnered public concern due to its implications for security and reliability. Traditional deep learning methods have been widely applied to this task but often lack generalisability when confronted with newly emerging spoofing techniques and more tasks such as spoof attribution recognition rather than simple binary classification. In principle, Large Language Models (LLMs) are considered to possess the needed generalisation capabilities. However, previous research on Audio LLMs (ALLMs) indicates a generalization bottleneck in audio deepfake detection performance, even when sufficient data is available. Consequently, this study investigates the model architecture and examines the effects of the primary components of ALLMs, namely the audio encoder and the text-based LLM. Our experiments demonstrate that the careful selection and combination of audio encoders and text-based LLMs are crucial for unlocking the deepfake detection potential of ALLMs. We further propose an ALLM structure capable of generalizing deepfake detection abilities to out-of-domain spoofing tests and other deepfake tasks, such as spoof positioning and spoof attribution recognition. Our proposed model architecture achieves state-of-the-art (SOTA) performance across multiple datasets, including ASVSpoof2019, InTheWild, and Demopage, with accuracy reaching up to 95.76% on average, and exhibits competitive capabilities in other deepfake detection tasks such as attribution, and localisation compared to SOTA audio understanding models. Data and codes are provided in supplementary materials.

Method: Uses DPO with multiple musical rewards across three dimensions: text alignment, audio production quality, and semantic consistency. Constructs preference data for DPO and integrates rewards into text prompting. Proposes novel scoring mechanism using semantic self-supervised representations to improve rhythmic stability.

Result: MR-FlowDPO significantly enhances overall music generation quality and is consistently preferred over competitive baselines in audio quality, text alignment, and musicality based on objective metrics and human studies.

Conclusion: The approach successfully aligns music generation with human preferences through multi-reward DPO optimization, improving both objective quality metrics and subjective human evaluations.

Abstract: A key challenge in music generation models is their lack of direct alignment with human preferences, as music evaluation is inherently subjective and varies widely across individuals. We introduce MR-FlowDPO, a novel approach that enhances flow-matching-based music generation models - a major class of modern music generative models, using Direct Preference Optimization (DPO) with multiple musical rewards. The rewards are crafted to assess music quality across three key dimensions: text alignment, audio production quality, and semantic consistency, utilizing scalable off-the-shelf models for each reward prediction. We employ these rewards in two ways: (i) By constructing preference data for DPO and (ii) by integrating the rewards into text prompting. To address the ambiguity in musicality evaluation, we propose a novel scoring mechanism leveraging semantic self-supervised representations, which significantly improves the rhythmic stability of generated music. We conduct an extensive evaluation using a variety of music-specific objective metrics as well as a human study. Results show that MR-FlowDPO significantly enhances overall music generation quality and is consistently preferred over highly competitive baselines in terms of audio quality, text alignment, and musicality. Our code is publicly available at https://github.com/lonzi/mrflow_dpo. Samples are provided in our demo page at https://lonzi.github.io/mr_flowdpo_demopage/.

Method: Introduces value profiles: small sets of reusable bundles of value-related parameters (outcome preferences, policy priors, and policy precision) assigned to hidden states in a generative model. Effective control parameters emerge via belief-weighted mixing as posterior beliefs over states evolve trial by trial.

Result: Model comparison in probabilistic reversal learning favors the profile-based model over simpler alternatives (about 100-point AIC differences). Parameter-recovery analyses support structural identifiability even with noisy observations. Adaptive control appears driven primarily by modulation of policy priors rather than policy precision, with gradual belief-dependent profile recruitment.

Conclusion: Reusable value profiles provide a tractable computational account of belief-conditioned value control in volatile environments and yield testable signatures of belief-dependent control and behavioral flexibility.

Abstract: Adaptive behavior in volatile environments requires agents to switch among value-control regimes across latent contexts, but maintaining separate preferences, policy biases, and action-confidence parameters for every situation is intractable. We introduce value profiles: a small set of reusable bundles of value-related parameters (outcome preferences, policy priors, and policy precision) assigned to hidden states in a generative model. As posterior beliefs over states evolve trial by trial, effective control parameters arise via belief-weighted mixing, enabling state-conditional strategy recruitment without requiring independent parameters for each context. We evaluate this framework in probabilistic reversal learning, comparing static-precision, entropy-coupled dynamic-precision, and profile-based models using cross-validated log-likelihood and information criteria. Model comparison favors the profile-based model over simpler alternatives (about 100-point AIC differences), and parameter-recovery analyses support structural identifiability even when context must be inferred from noisy observations. Model-based inference further suggests that adaptive control in this task is driven primarily by modulation of policy priors rather than policy precision, with gradual belief-dependent profile recruitment consistent with state-conditional (not purely uncertainty-driven) control. Overall, reusable value profiles provide a tractable computational account of belief-conditioned value control in volatile environments and yield testable signatures of belief-dependent control and behavioral flexibility.

Method: CR3G (Causal Reasoning for Patient-Centric Explanations in radiology Report Generation) is a prompt-driven framework that applies causal inference to chest X-ray analysis, focusing on uncovering cause-and-effect relationships and generating patient-centric explanations.

Result: CR3G demonstrated improved causal relationship capability and explanation capability for 2 out of 5 abnormalities tested, showing better understanding of AI-generated reports.

Conclusion: The CR3G framework enhances AI-driven chest X-ray diagnostics by incorporating causal reasoning, making them more useful and trustworthy in clinical practice through better understanding of cause-and-effect relationships.

Abstract: Automatic chest X-ray report generation is an important area of research aimed at improving diagnostic accuracy and helping doctors make faster decisions. Current AI models are good at finding correlations (or patterns) in medical images. Still, they often struggle to understand the deeper cause-and-effect relationships between those patterns and a patient condition. Causal inference is a powerful approach that goes beyond identifying patterns to uncover why certain findings in an X-ray relate to a specific diagnosis. In this paper, we will explore the prompt-driven framework Causal Reasoning for Patient-Centric Explanations in radiology Report Generation (CR3G) that is applied to chest X-ray analysis to improve understanding of AI-generated reports by focusing on cause-and-effect relationships, reasoning and generate patient-centric explanation. The aim to enhance the quality of AI-driven diagnostics, making them more useful and trustworthy in clinical practice. CR3G has shown better causal relationship capability and explanation capability for 2 out of 5 abnormalities.

Method: Proposes a common design framework for representative shortcut models that provides theoretical justification and disentangles concrete component-level choices, enabling systematic identification of improvements.

Result: Achieves new state-of-the-art FID50k of 2.85 on ImageNet-256x256 under classifier-free guidance, without requiring pre-training, distillation, or curriculum learning.

Conclusion: The framework lowers the barrier to component-level innovation in shortcut models and facilitates principled exploration of their design space.

Abstract: Recent advances in few-step diffusion models have demonstrated their efficiency and effectiveness by shortcutting the probabilistic paths of diffusion models, especially in training one-step diffusion models from scratch (a.k.a. shortcut models). However, their theoretical derivation and practical implementation are often closely coupled, which obscures the design space. To address this, we propose a common design framework for representative shortcut models. This framework provides theoretical justification for their validity and disentangles concrete component-level choices, thereby enabling systematic identification of improvements. With our proposed improvements, the resulting one-step model achieves a new state-of-the-art FID50k of 2.85 on ImageNet-256x256 under the classifier-free guidance setting. Remarkably, the model requires no pre-training, distillation, or curriculum learning. We believe our work lowers the barrier to component-level innovation in shortcut models and facilitates principled exploration of their design space.

Method: Compared three methods: inverse distance weighting (IDW), ordinary kriging (OK), and implicit neural representation model (MMGN). All methods underwent hyper-parameter tuning with validation splits. Evaluated on 100 randomly sampled sparse datasets from ECA&D database using RMSE, MAE, Δ_MAX, and R² metrics with comprehensive statistical analysis including Dunn post-hoc tests.

Result: IDW achieved best performance: lowest RMSE (3.00±1.93), MAE (1.32±0.77), Δ_MAX (24.06±17.15), and highest R² (0.68±0.16). Statistical tests confirmed IDW’s superiority with significant differences and moderate to large effect sizes across all quality measures.

Conclusion: Simple inverse distance weighting method outperforms both statistically grounded ordinary kriging and advanced neural representation models for sparse climate data reconstruction, offering the best combination of accuracy and computational efficiency.

Abstract: This study evaluates three reconstruction methods for sparse climate data: the simple inverse distance weighting (IDW), the statistically grounded ordinary kriging (OK), and the advanced implicit neural representation model (MMGN architecture). All methods were optimized through hyper-parameter tuning using validation splits. An extensive set of experiments was conducted, followed by a comprehensive statistical analysis. The results demonstrate the superiority of the simple IDW method over the other reference methods in terms of both reconstruction accuracy and computational efficiency. IDW achieved the lowest RMSE ($3.00 \pm 1.93$), MAE ($1.32 \pm 0.77$), and $Δ_{MAX}$ ($24.06 \pm 17.15$), as well as the highest $R^2$ ($0.68 \pm 0.16$), across 100 randomly sampled sparse datasets from the ECA&D database. Differences in RMSE, MAE, and $R^2$ were statistically significant and exhibited moderate to large effect sizes. The Dunn post-hoc test further confirmed the consistent superiority of IDW across all evaluated quality measures […]

Method: Implemented Soft Decision Tree (SDT) and Short-term Memory Soft Decision Tree (SM-SDT) using PyTorch framework. Tested extensively on both simulated datasets and clinical datasets. Visualized SDT structure to demonstrate explainability capabilities. Compared performance against multiple baseline models including XGBoost, Random Forest, Logistic Regression, and traditional Decision Tree.

Result: SDT, SM-SDT, and XGBoost showed similar AUC values, all outperforming Random Forest, Logistic Regression, and traditional Decision Tree. On clinical datasets, all tested classification methods (except basic Decision Tree) yielded comparable results. The SDT visualization demonstrated potential for explainability in clinical applications.

Conclusion: Soft Decision Trees and their variants offer competitive predictive performance comparable to state-of-the-art methods like XGBoost while providing enhanced explainability features. This makes them particularly suitable for clinical applications where both accuracy and interpretability are important. The availability of code and datasets on GitHub facilitates reproducibility and further research.

Abstract: We implemented a Soft Decision Tree (SDT) and a Short-term Memory Soft Decision Tree (SM-SDT) using PyTorch. The methods were extensively tested on simulated and clinical datasets. The SDT was visualized to demonstrate the potential for its explainability. SDT, SM-SDT, and XGBoost demonstrated similar area under the curve (AUC) values. These methods were better than Random Forest, Logistic Regression, and Decision Tree. The results on clinical datasets suggest that, aside from a decision tree, all tested classification methods yield comparable results. The code and datasets are available online on GitHub: https://github.com/KI-Research-Institute/Soft-Decision-Tree

Method: The authors analyze the bootstrapping process through the lens of continual learning, connecting it to Generative Experience Replay (GER) methods. They conduct statistical analysis of bias and variance introduced by synthetic data, and provide empirical evidence showing model collapse under repeated synthetic data training.

Result: Synthetic data introduces significant bias and variance that weakens maximum likelihood estimation reliability. Popular generative models collapse under repeated training with synthetic data, and state-of-the-art GER methods fail to maintain alignment in the latent space.

Conclusion: The findings raise critical concerns about using synthetic data in continual learning, highlighting fundamental limitations of current approaches that rely on repeatedly training models on their own generated outputs.

Abstract: The use of synthetically generated data for training models is becoming a common practice. While generated data can augment the training data, repeated training on synthetic data raises concerns about distribution drift and degradation of performance due to contamination of the dataset. We investigate the consequences of this bootstrapping process through the lens of continual learning, drawing a connection to Generative Experience Replay (GER) methods. We present a statistical analysis showing that synthetic data introduces significant bias and variance into training objectives, weakening the reliability of maximum likelihood estimation. We provide empirical evidence showing that popular generative models collapse under repeated training with synthetic data. We quantify this degradation and show that state-of-the-art GER methods fail to maintain alignment in the latent space. Our findings raise critical concerns about the use of synthetic data in continual learning.

Method: Combines PBDW framework (reduced-order model + measurements) with DeepONet constrained as orthogonal complement to best-knowledge manifold, plus optimal sensor placement strategy.

Result: Validated on Helmholtz equation with various modeling errors (boundary conditions, source terms), showing enhanced state estimation and prediction.

Conclusion: Proposed hybrid approach effectively integrates physics-based modeling with data-driven learning to handle model discrepancies while preserving interpretability and physical fidelity.

Abstract: The accurate estimation of the state of complex uncertain physical systems requires reconciling theoretical models, with inherent imperfections, with noisy experimental data. In this work, we propose an effective hybrid approach that combines physics-based modeling with data-driven learning to enhance state estimation and further prediction. Our method builds upon the Parameterized Background Data-Weak (PBDW) framework, which naturally integrates a reduced-order representation of the best-available model with measurement data to account for both anticipated and unanticipated uncertainties. To address model discrepancies not captured by the reduced-order space, and learn the structure of model deviation, we incorporate a Deep Operator Network (DeepONet) constrained to be an orthogonal complement of the best-knowledge manifold. This ensures that the learned correction targets only the unknown components of model bias, preserving the interpretability and fidelity of the physical model. An optimal sensor placement strategy is also investigated to maximize information gained from measurements. We validate the proposed approach on a representative problem involving the Helmholtz equation under various sources of modeling error, including those arising from boundary conditions and source terms.

Method: Multi-headed neural networks process hybrid feature vectors combining semantic text embeddings, lexical patterns, domain heuristics, and USDA FNDDS data to estimate nutrient and food components for FCS calculation.

Result: Strong predictive performance with median R² of 0.81 for individual nutrients, Pearson’s r = 0.77 correlation with published FCS values, and mean absolute difference of 14.0 points.

Conclusion: The methodology successfully translates language into actionable nutritional information, enabling scalable dietary assessment for consumer applications and research, though performance is weaker for ambiguous or processed foods.

Abstract: Accurate nutritional assessment is critical for public health, but existing profiling systems require detailed data often unavailable or inaccessible from colloquial text descriptions of food. This paper presents a machine learning pipeline that predicts the comprehensive Food Compass Score 2.0 (FCS) from text descriptions. Our approach uses multi-headed neural networks to process hybrid feature vectors that combine semantic text embeddings, lexical patterns, and domain heuristics, alongside USDA Food and Nutrient Database for Dietary Studies (FNDDS) data. The networks estimate the nutrient and food components necessary for the FCS algorithm. The system demonstratedstrong predictive power, achieving a median R^2 of 0.81 for individual nutrients. The predicted FCS correlated strongly with published values (Pearson’s r = 0.77), with a mean absolute difference of 14.0 points. While errors were largest for ambiguous or processed foods, this methodology translates language into actionable nutritional information, enabling scalable dietary assessment for consumer applications and research.

Method: 1) Derivation showing DPO gradient depends only on logit embedding differences; 2) Extracting empirical steering vectors from DPO-tuned models; 3) Spectral analyses of activation patterns.

Result: DPO creates first-order shifts in final hidden representations, not deep restructuring. Adding extracted steering vectors to base activations reproduces aligned behavior, while subtracting restores original model. Spectral analyses show rank-one dominance and entropy collapse in upper layers.

Conclusion: DPO teaches models how to act aligned (behavioral illusion) rather than changing what they actually believe, functioning as a low-rank steering mechanism through narrow subspaces.

Abstract: Direct Preference Optimization (DPO) has become a standard recipe for aligning large language models, yet it is still unclear what kind of change it actually induces inside the network. This paper argues that DPO does not rewrite a models internal beliefs; instead, it acts as a low rank steering mechanism that nudges activations along a small number of preference directions. Using a simple derivation, we show that the DPO gradient depends only on the difference between the logit embeddings of preferred and dispreferred completions, implying a first order shift in the final hidden representation rather than a deep restructuring of semantics. We then extract an empirical steering vector from a DPO tuned model and demonstrate that adding this vector to base activations reproduces most of the aligned behavior, while subtracting it nearly restores the original model. Finally, spectral analyses reveal rank-one dominance and entropy collapse in upper layers, indicating that alignment is funneled through a narrow subspace. Taken together, these results support a behavioral illusion view of DPO: it teaches models how to act aligned, not what to believe.

Method: Trailblazer incorporates: 1) a network alignment scheme to ground the LLM in specific networking tasks, and 2) an adaptive policy collaboration mechanism that offloads simple control cases from the LLM to a lightweight policy for computational efficiency.

Result: Through extensive simulations and large-scale real-world online evaluation on Douyin (Chinese TikTok), Trailblazer demonstrates stronger cross-task and cross-environment generalization than conventional specialist policies, using a single LLM.

Conclusion: LLMs serve as a foundation for generalist network policies, with Trailblazer representing the first step toward a generalist-driven paradigm that enables strong generalization with minimal policy design efforts.

Abstract: Designing control policies to ensure robust network services is essential to modern digital infrastructure. However, the dominant paradigm for network optimization relies on designing specialist policies based on handcrafted rules or deep learning models, leading to poor generalization across diverse tasks and environments. In contrast, large language models (LLMs), pretrained on Internet-scale corpora, provide a rich and unified knowledge base that encodes fundamental networking principles. Combined with their emergent abilities in generalization to unseen scenarios, LLMs offer a transformative foundation for generalist network policies that can generalize across diverse tasks and environments with minimal adaptation. In this paper, we present Trailblazer, the first systematic framework to realize such a generalist policy for networking. Trailblazer incorporates a network alignment scheme to ground the LLM in specific networking tasks, and an adaptive policy collaboration mechanism that offloads simple control cases from the LLM to a lightweight policy for computational efficiency. Through extensive simulations and large-scale real-world online evaluation on Douyin (the Chinese version of TikTok), Trailblazer, powered by a single LLM, demonstrates stronger cross-task and cross-environment generalization than conventional specialist policies. Our results validate LLMs as the foundation for generalist network policies, and position Trailblazer as the first step toward the generalist-driven paradigm that enables strong generalization with minimal efforts in policy design.

Method: Proposes amortized causal discovery using Prior-Fitted Networks (PFNs) to amortize data-dependent likelihood estimation, providing more reliable scores for structure learning.

Result: Experiments on synthetic, simulated, and real-world datasets show significant gains in structure recovery compared to standard baselines. PFNs provide more accurate likelihood estimates than conventional neural network approaches.

Conclusion: PFN-based amortized causal discovery effectively addresses likelihood estimation accuracy limitations, leading to improved causal structure recovery performance.

Abstract: In recent years, differentiable penalized likelihood methods have gained popularity, optimizing the causal structure by maximizing its likelihood with respect to the data. However, recent research has shown that errors in likelihood estimation, even on relatively large sample sizes, disallow the discovery of proper structures. We propose a new approach to amortized causal discovery that addresses the limitations of likelihood estimator accuracy. Our method leverages Prior-Fitted Networks (PFNs) to amortize data-dependent likelihood estimation, yielding more reliable scores for structure learning. Experiments on synthetic, simulated, and real-world datasets show significant gains in structure recovery compared to standard baselines. Furthermore, we demonstrate directly that PFNs provide more accurate likelihood estimates than conventional neural network-based approaches.

Method: Proposes a lightweight meta-continual forecasting framework integrating: 1) GRU-based predictor for mobility forecasting, 2) Reptile meta-initialization for fast few-shot adaptation to new mobility patterns, and 3) EWMA residual detector that triggers compact online updates only when drift occurs.

Result: Achieves 4.46 m ADE and 7.79 m FDE in zero-shot settings, improves few-shot ADE to 3.71 m at 10-shot, enables recovery from abrupt drift 2-3 times faster than offline GRU. For handover prediction: improves F1 to 0.83 and AUROC to 0.90, with substantial reductions in missed-HO and ping-pong events. Model is lightweight (128k parameters).

Conclusion: The proposed framework effectively handles non-stationary mobility patterns for proactive handover in cellular networks, offering fast adaptation to drift while maintaining lightweight architecture suitable for edge deployment in 5G/6G systems.

Abstract: Short-term mobility forecasting is a core requirement for proactive handover (HO) in cellular networks. Real-world mobility is highly non-stationary: abrupt turns, rapid speed changes, and unpredictable user behavior cause conventional predictors to drift, leading to mistimed or failed handovers. We propose a lightweight meta-continual forecasting framework that integrates a GRU-based predictor, Reptile meta-initialization for fast few-shot adaptation, and an EWMA residual detector that triggers compact online updates only when drift occurs. Evaluated on a reproducible GeoLife and DeepMIMO pipeline, our method achieves 4.46 m ADE and 7.79 m FDE in zero-shot settings, improves few-shot ADE to 3.71 m at 10-shot, and enables recovery from abrupt drift about 2 to 3 times faster than an offline GRU. When applied to downstream HO prediction, the approach improves F1 to 0.83 and AUROC to 0.90, with substantial reductions in missed-HO and ping-pong events. The model is lightweight (128k parameters) and suitable for edge deployment in 5G and 6G systems.

Method: Proposes DTSFormer with: 1) Multiscale deformable partitioning using window function-based masking to dynamically partition input into temporal patches of varying scales, 2) Frequency-domain attention mechanism to capture both high- and low-frequency components, 3) Fusion of multi-frequency features in time domain to model short-term fluctuations and long-term trends.

Result: Outperforms state-of-the-art forecasting models across different prediction horizons on real-world passenger flow data from Beijing Capital International Airport (Jan 2023-Mar 2024). Deformable partitioning aligns patch lengths with dominant periods and heterogeneous trends, enabling better capture of sudden high-frequency fluctuations.

Conclusion: DTSFormer effectively addresses the limitations of fixed-size patch embeddings by dynamically adapting to heterogeneous temporal patterns in airport passenger flows, achieving superior forecasting performance through integrated temporal-spectral analysis.

Abstract: Accurate forecasting of passenger flows is critical for maintaining the efficiency and resilience of airport operations. Recent advances in patch-based Transformer models have shown strong potential in various time series forecasting tasks. However, most existing methods rely on fixed-size patch embedding, making it difficult to model the complex and heterogeneous patterns of airport passenger flows. To address this issue, this paper proposes a deformable temporal-spectral transformer named DTSFormer that integrates a multiscale deformable partitioning module and a joint temporal-spectral filtering module. Specifically, the input sequence is dynamically partitioned into multiscale temporal patches via a novel window function-based masking, enabling the extraction of heterogeneous trends across different temporal stages. Then, within each scale, a frequency-domain attention mechanism is designed to capture both high- and low-frequency components, thereby emphasizing the volatility and periodicity inherent in airport passenger flows. Finally, the resulting multi-frequency features are subsequently fused in the time domain to jointly model short-term fluctuations and long-term trends. Comprehensive experiments are conducted on real-world passenger flow data collected at Beijing Capital International Airport from January 2023 to March 2024. The results indicate that the proposed method consistently outperforms state-of-the-art forecasting models across different prediction horizons. Further analysis shows that the deformable partitioning module aligns patch lengths with dominant periods and heterogeneous trends, enabling superior capture of sudden high-frequency fluctuations.

Method: 1) Apply meta-weighting to adjacency matrix to distinguish meta relations in heterogeneous graphs. 2) Use PageRank with node labels to construct projection mapping nodes to values correlated with model performance. 3) Propose debiasing structure based on differences in mapped values and use it with original graph for contrastive learning.

Result: The constructed projection effectively maps nodes to values strongly correlated with model performance across datasets with and without intra-type connections, demonstrating universal existence of topological bias in HGNNs. Experiments on three public datasets verify the method’s effectiveness in improving HGNNs’ performance and debiasing.

Conclusion: Topological bias exists universally in HGNNs and significantly affects performance. The proposed meta-weighting and contrastive learning approach with debiasing structure effectively mitigates this bias and improves HGNN performance.

Abstract: Graph Neural Networks (GNNs) are characterized by their capacity of processing graph-structured data. However, due to the sparsity of labels under semi-supervised learning, they have been found to exhibit biased performance on specific nodes. This kind of bias has been validated to correlate with topological structure and is considered as a bottleneck of GNNs’ performance. Existing work focuses on the study of homogeneous GNNs and little attention has been given to topological bias in Heterogeneous Graph Neural Networks (HGNNs). In this work, firstly, in order to distinguish distinct meta relations, we apply meta-weighting to the adjacency matrix of a heterogeneous graph. Based on the modified adjacency matrix, we leverage PageRank along with the node label information to construct a projection. The constructed projection effectively maps nodes to values that strongly correlated with model performance when using datasets both with and without intra-type connections, which demonstrates the universal existence of topological bias in HGNNs. To handle this bias, we propose a debiasing structure based on the difference in the mapped values of nodes and use it along with the original graph structure for contrastive learning. Experiments on three public datasets verify the effectiveness of the proposed method in improving HGNNs’ performance and debiasing.

Method: Study decentralized multi-agent SSPs (Dec-MASSPs) under linear function approximation, apply novel symmetry-based arguments to identify optimal policy structure, and construct hard-to-learn instances for any number of agents.

Result: Established the first regret lower bound of Ω(√K) over K episodes, highlighting inherent learning difficulty in Dec-MASSPs.

Conclusion: The results clarify learning complexity of decentralized control and can guide design of efficient learning algorithms in multi-agent systems.

Abstract: Multi-agent systems (MAS) are central to applications such as swarm robotics and traffic routing, where agents must coordinate in a decentralized manner to achieve a common objective. Stochastic Shortest Path (SSP) problems provide a natural framework for modeling decentralized control in such settings. While the problem of learning in SSP has been extensively studied in single-agent settings, the decentralized multi-agent variant remains largely unexplored. In this work, we take a step towards addressing that gap. We study decentralized multi-agent SSPs (Dec-MASSPs) under linear function approximation, where the transition dynamics and costs are represented using linear models. Applying novel symmetry-based arguments, we identify the structure of optimal policies. Our main contribution is the first regret lower bound for this setting based on the construction of hard-to-learn instances for any number of agents, $n$. Our regret lower bound of $Ω(\sqrt{K})$, over $K$ episodes, highlights the inherent learning difficulty in Dec-MASSPs. These insights clarify the learning complexity of decentralized control and can further guide the design of efficient learning algorithms in multi-agent systems.

Method: Empirical analysis of ARC Prize TRM checkpoint on ARC-AGI-1 including: test-time augmentation/ensembling ablation, puzzle-identity ablation, recursion trajectory analysis, early-stage training experiments with different augmentation regimes, and efficiency comparison with QLoRA fine-tuned Llama 3 8B.

Result: 1) Test-time augmentation/voting accounts for ~11% performance improvement; 2) Strict puzzle ID dependence (zero accuracy without correct IDs); 3) Shallow effective recursion (most accuracy at first step, saturates quickly); 4) Heavy augmentation broadens solution distribution; 5) TRM has much higher throughput and lower memory than QLoRA Llama 3 8B.

Conclusion: TRM’s ARC-AGI-1 performance arises from efficiency, task-specific conditioning, and aggressive test-time compute rather than deep internal reasoning, with shallow effective recursion and strong dependence on external cues.

Abstract: Tiny Recursive Models (TRM) were proposed as a parameter-efficient alternative to large language models for solving Abstraction and Reasoning Corpus (ARC) style tasks. The original work reports strong performance and suggests that recursive latent updates enable non-trivial reasoning, but it remains unclear how much of this performance stems from architecture, test-time compute, or task-specific priors. In this technical note, we empirically analyze the ARC Prize TRM checkpoint on ARC-AGI-1 and report four behavioral findings and an efficiency comparison. First, we show that test-time augmentation and majority-vote ensembling account for a substantial fraction of reported performance: the 1000-sample voting pipeline improves Pass@1 by about 11 percentage points over single-pass canonical inference. Second, a puzzle-identity ablation reveals strict dependence on task identifiers: replacing the correct puzzle ID with a blank or random token yields zero accuracy. Third, a recursion trajectory analysis shows that most of the final accuracy is achieved at the first recursion step and that performance saturates after few latent updates, indicating shallow effective recursion. Fourth, early-stage training experiments under canonical versus heavy augmentation regimes suggest that heavy augmentation broadens the distribution of candidate solutions and improves multi-sample success. Finally, we compare TRM with a naive QLoRA fine-tune of Llama 3 8B on canonical ARC-AGI-1, finding that TRM’s non-autoregressive design achieves much higher throughput and substantially lower memory usage in this setting. Overall, TRM’s ARC-AGI-1 performance appears to arise from an interaction between efficiency, task-specific conditioning, and aggressive test-time compute rather than deep internal reasoning.

Method: Build a cache of past activations indexed by sentence embeddings, reuse cached KV states when cached prompt is an exact prefix of new input, serialize cached KVs to CPU and reload them for decoding continuation.

Result: Consistent speedups when prefix overlap exists with no material degradation in output semantics; when overlap is absent, behavior matches baseline.

Conclusion: Token recycling is an effective approach for accelerating LLM inference on similar prompts by reusing cached KV states, requiring no model modifications while maintaining output quality.

Abstract: Whether attention key value (KV) states computed for one prompt for a small LLM can be reused to accelerate inference on a new similar prompt, giving an increase to the space to its context memory using an approach called token recycling. Using a standard Hugging Face setup with DialoGPT-medium (a 345M parameter GPT-2 style decoder trained on 147M Reddit exchanges, 2005 to 2017) as the testbed, we build a cache of past activations and get entries by sentence embeddings, then reuse cached past key values when the cached prompt is an exact prefix of the new input. We compare recycled vs. baseline runs on latency and output fidelity, and log reuse depth in tokens. Reproducibility requires no model modifications, cached KVs are serialized to the CPU, reloaded, and supplied to the generate function to continue decoding from the cached prefix. In tests, we observe consistent speedups when prefix overlap exists, with no material degradation in output semantics, and when overlap is absent, behavior matches baseline.

Method: Leveraged Temporal Fusion Transformer (TFT) model for greenhouse actuator automation, enhanced with both local and global explanation techniques including model-inherent interpretation, LIME, and SHAP to provide interpretability.

Result: The trained TFT model achieved 95% test accuracy on a class-imbalanced dataset for actuator control settings. The explainability methods revealed varying influence of different sensors (temperature, humidity, CO2, light, outer climate) on real-time greenhouse adjustments.

Conclusion: The approach ensures transparency in automated greenhouse management, enables adaptive fine-tuning for improved crop yield and resource efficiency, and addresses the critical need for explainable AI in smart farming practices.

Abstract: The integration of the Internet of Robotic Things (IoRT) in smart greenhouses has revolutionised precision agriculture by enabling efficient and autonomous environmental control. However, existing time series forecasting models in such setups often operate as black boxes, lacking mechanisms for explainable decision-making, which is a critical limitation when trust, transparency, and regulatory compliance are paramount in smart farming practices. This study leverages the Temporal Fusion Transformer (TFT) model to automate actuator settings for optimal greenhouse management. To enhance interpretability and trust in the model decision-making process, both local and global explanation techniques were employed using model-inherent interpretation, local interpretable model-agnostic explanations (LIME), and SHapley additive explanations (SHAP). These explainability methods provide information on how different sensor readings, such as temperature, humidity, CO2 levels, light, and outer climate, contribute to actuator control decisions in an automated greenhouse. The trained TFT model achieved a test accuracy of 95% on a class-imbalanced dataset for actuator control settings in an automated greenhouse environment. The results demonstrate the varying influence of each sensor on real-time greenhouse adjustments, ensuring transparency and enabling adaptive fine-tuning for improved crop yield and resource efficiency.

Method: Two-stage pipeline using a single wrist-mounted IMU: 1) ResNet-based model segments repetitions from 6-axis IMU data in real-time, 2) Classification model uses segmentation features, convolutional features, and LSTM-captured historical context to identify near-failure states (RiR ≤ 2).

Result: Segmentation model achieved F1 score of 0.83, near-failure classifier achieved F1 score of 0.82 under simulated real-time conditions (1.6 Hz inference rate). Deployment on Raspberry Pi 5 (112 ms latency) and iPhone 16 (23.5 ms latency) confirmed edge computation feasibility.

Conclusion: The system demonstrates practical, objective real-time training intensity feedback using minimal hardware, paving the way for accessible AI-driven hypertrophy coaching tools that help users effectively manage intensity and fatigue.

Abstract: Optimizing resistance training for hypertrophy requires balancing proximity to muscular failure, often quantified by Repetitions in Reserve (RiR), with fatigue management. However, subjective RiR assessment is unreliable, leading to suboptimal training stimuli or excessive fatigue. This paper introduces a novel system for real-time feedback on near-failure states (RiR $\le$ 2) during resistance exercise using only a single wrist-mounted Inertial Measurement Unit (IMU). We propose a two-stage pipeline suitable for edge deployment: first, a ResNet-based model segments repetitions from the 6-axis IMU data in real-time. Second, features derived from this segmentation, alongside direct convolutional features and historical context captured by an LSTM, are used by a classification model to identify exercise windows corresponding to near-failure states. Using a newly collected dataset from 13 diverse participants performing preacher curls to failure (631 total reps), our segmentation model achieved an F1 score of 0.83, and the near-failure classifier achieved an F1 score of 0.82 under simulated real-time evaluation conditions (1.6 Hz inference rate). Deployment on a Raspberry Pi 5 yielded an average inference latency of 112 ms, and on an iPhone 16 yielded 23.5 ms, confirming the feasibility for edge computation. This work demonstrates a practical approach for objective, real-time training intensity feedback using minimal hardware, paving the way for accessible AI-driven hypertrophy coaching tools that help users manage intensity and fatigue effectively.

Method: The paper introduces “averaging complexity” as a framework to quantify the computational cost of enforcing symmetry via averaging operations. This provides a theoretical metric to compare exact versus approximate symmetry enforcement.

Result: Main result shows exponential separation: under standard conditions, achieving exact symmetry requires linear averaging complexity, while approximate symmetry can be attained with only logarithmic averaging complexity.

Conclusion: This provides the first theoretical separation between exact and approximate symmetry, formally justifying why approximate symmetry may be preferable in practice due to significantly lower computational cost.

Abstract: Enforcing exact symmetry in machine learning models often yields significant gains in scientific applications, serving as a powerful inductive bias. However, recent work suggests that relying on approximate symmetry can offer greater flexibility and robustness. Despite promising empirical evidence, there has been little theoretical understanding, and in particular, a direct comparison between exact and approximate symmetry is missing from the literature. In this paper, we initiate this study by asking: What is the cost of enforcing exact versus approximate symmetry? To address this question, we introduce averaging complexity, a framework for quantifying the cost of enforcing symmetry via averaging. Our main result is an exponential separation: under standard conditions, achieving exact symmetry requires linear averaging complexity, whereas approximate symmetry can be attained with only logarithmic averaging complexity. To the best of our knowledge, this provides the first theoretical separation of these two cases, formally justifying why approximate symmetry may be preferable in practice. Beyond this, our tools and techniques may be of independent interest for the broader study of symmetries in machine learning.

Method: GCoDE abstracts device communication as an explicit operation, fuses architecture and mapping in a unified design space for joint optimization. It introduces system performance awareness and energy prediction method for GNN operations, uses constraint-based random search to find optimal solutions in 1.5 hours, and includes an integrated co-inference engine.

Result: Experimental results show GCoDE achieves up to 44.9x speedup and 98.2% energy reduction compared to existing approaches across diverse applications and system configurations.

Conclusion: GCoDE successfully addresses GNN inference challenges on edge devices through architecture-mapping co-design, enabling efficient GNN deployment in real-world edge scenarios with significant performance and energy improvements.

Abstract: Graph Neural Networks (GNNs) have emerged as the state-of-the-art graph learning method. However, achieving efficient GNN inference on edge devices poses significant challenges, limiting their application in real-world edge scenarios. This is due to the high computational cost of GNNs and limited hardware resources on edge devices, which prevent GNN inference from meeting real-time and energy requirements. As an emerging paradigm, device-edge co-inference shows potential for improving inference efficiency and reducing energy consumption on edge devices. Despite its potential, research on GNN device-edge co-inference remains scarce, and our findings show that traditional model partitioning methods are ineffective for GNNs. To address this, we propose GCoDE, the first automatic framework for GNN architecture-mapping Co-design and deployment on Device-Edge hierarchies. By abstracting the device communication process into an explicit operation, GCoDE fuses the architecture and mapping scheme in a unified design space for joint optimization. Additionally, GCoDE’s system performance awareness enables effective evaluation of architecture efficiency across diverse heterogeneous systems. By analyzing the energy consumption of various GNN operations, GCoDE introduces an energy prediction method that improves energy assessment accuracy and identifies energy-efficient solutions. Using a constraint-based random search strategy, GCoDE identifies the optimal solution in 1.5 hours, balancing accuracy and efficiency. Moreover, the integrated co-inference engine in GCoDE enables efficient deployment and execution of GNN co-inference. Experimental results show that GCoDE can achieve up to 44.9x speedup and 98.2% energy reduction compared to existing approaches across diverse applications and system configurations.

Method: TopicProphet uses a sequence of topic modeling, temporal analysis, breakpoint detection, and segment optimization to identify historical periods that share similar public sentiment trends and historical background. This allows selection of optimal time periods for training models based on era-specific patterns.

Result: The framework produces improved outcomes compared to state-of-the-art methods in capturing optimal training data for forecasting financial percentage changes, effectively addressing data scarcity issues.

Conclusion: By analyzing historical eras with similar sentiment trends, TopicProphet successfully improves stock prediction accuracy by providing models with nuanced patterns from relevant socioeconomic and political contexts, overcoming traditional limitations in financial forecasting.

Abstract: Stocks can’t be predicted. Despite many hopes, this premise held itself true for many years due to the nature of quantitative stock data lacking causal logic along with rapid market changes hindering accumulation of significant data for training models. To undertake this matter, we propose a novel framework, TopicProphet, to analyze historical eras that share similar public sentiment trends and historical background. Our research deviates from previous studies that identified impacts of keywords and sentiments - we expand on that method by a sequence of topic modeling, temporal analysis, breakpoint detection and segment optimization to detect the optimal time period for training. This results in improving predictions by providing the model with nuanced patterns that occur from that era’s socioeconomic and political status while also resolving the shortage of pertinent stock data to train on. Through extensive analysis, we conclude that TopicProphet produces improved outcomes compared to the state-of-the-art methods in capturing the optimal training data for forecasting financial percentage changes.

Method: Developed path-wise schedule selection framework for Harmonic PID with time-varying stiffness using Piece-Wise-Constant parametrizations and hierarchical refinement. Introduced schedule-sensitive Quality-of-Sampling diagnostics. For Gaussian-Mixture targets, retained closed-form Green functions and numerically stable, Neural-Network free oracles for predicted-state maps and score.

Result: Experiments in 2D show that QoS-driven PWC schedules consistently improve early-exit fidelity, tail accuracy, conditioning of the dynamics, and speciation (label-selection) timing at fixed integration budgets.

Conclusion: The proposed schedule selection framework with QoS diagnostics effectively optimizes diffusion sampling performance, demonstrating significant improvements in multiple quality metrics under computational constraints.

Abstract: Diffusion-based samplers – Score Based Diffusions, Bridge Diffusions and Path Integral Diffusions – match a target at terminal time, but the real leverage comes from choosing the schedule that governs the intermediate-time dynamics. We develop a path-wise schedule – selection gramework for Harmonic PID with a time-varying stiffness, exploiting Piece-Wise-Constant(PWC) parametrizations and a simple hierarchical refinement. We introduce schedule-sensitive Quality-of-Sampling (QoS) diagnostics. Assuming a Gaussian-Mixture (GM) target, we retain closed-form Green functions’ ration and numerically stable, Neural-Network free oracles for predicted-state maps and score. Experiments in 2D show that QoS driven PWC schedules consistently improve early-exit fidelity, tail accuracy, conditioning of the dynamics, and speciation (label-selection) timing at fixed integration budgets.

Method: A linearly-solvable framework where low-dimensional guidance protocols shape trajectory ensembles while preserving analytic structure (linear Kolmogorov equations, explicit Green-function ratio for optimal score, closed-form expressions for GMM terminal laws). Includes guidance-centric diagnostics and demonstrates three navigation scenarios: hand-crafted protocols, single-task protocol learning, and multi-expert fusion.

Result: GH-PID generates geometry-aware, trust-aware trajectories that satisfy prescribed terminal distributions while systematically reducing integrated cost. The framework enables stable sampling, differentiable protocol learning, and provides interpretable diagnostics for empirical SOT analysis.

Conclusion: GH-PID provides an analytically tractable, interpretable framework for guided Stochastic Optimal Transport that combines hard terminal constraints with soft path costs, enabling both theoretical analysis and practical learning applications across various navigation scenarios.

Abstract: We introduce Guided Harmonic Path-Integral Diffusion (GH-PID), a linearly-solvable framework for guided Stochastic Optimal Transport (SOT) with a hard terminal distribution and soft, application-driven path costs. A low-dimensional guidance protocol shapes the trajectory ensemble while preserving analytic structure: the forward and backward Kolmogorov equations remain linear, the optimal score admits an explicit Green-function ratio, and Gaussian-Mixture Model (GMM) terminal laws yield closed-form expressions. This enables stable sampling and differentiable protocol learning under exact terminal matching. We develop guidance-centric diagnostics – path cost, centerline adherence, variance flow, and drift effort – that make GH-PID an interpretable variational ansatz for empirical SOT. Three navigation scenarios illustrated in 2D: (i) Case A: hand-crafted protocols revealing how geometry and stiffness shape lag, curvature effects, and mode evolution; (ii) Case B: single-task protocol learning, where a PWC centerline is optimized to minimize integrated cost; (iii) Case C: multi-expert fusion, in which a commander reconciles competing expert/teacher trajectories and terminal beliefs through an exact product-of-experts law and learns a consensus protocol. Across all settings, GH-PID generates geometry-aware, trust-aware trajectories that satisfy the prescribed terminal distribution while systematically reducing integrated cost.

Method: Develops an operator-consistent graph neural network (OCGNN-PINN) that approximates PDE evolution with physics-informed constraints. Uses node-edge message passing with a consistency loss that enforces gradient-divergence relations through graph incidence matrices, ensuring structural coupling between discrete node and edge dynamics during temporal rollout.

Result: The model shows improved temporal stability and prediction accuracy compared to graph convolutional and multilayer perceptron baselines on diffusion processes over evolving meshes and real-world scanned surfaces. Performance approaches that of Crank-Nicolson solvers on unstructured domains.

Conclusion: OCGNN-PINN provides a stable and accurate method for solving PDEs on irregular meshes, bridging the gap between classical numerical methods and modern machine learning approaches for complex multiphysics problems on unstructured domains.

Abstract: Classical numerical methods solve partial differential equations (PDEs) efficiently on regular meshes, but many of them become unstable on irregular domains. In practice, multiphysics interactions such as diffusion, damage, and healing often take place on irregular meshes. We develop an operator-consistent graph neural network (OCGNN-PINN) that approximates PDE evolution under physics-informed constraints. It couples node-edge message passing with a consistency loss enforcing the gradient-divergence relation through the graph incidence matrix, ensuring that discrete node and edge dynamics remain structurally coupled during temporal rollout. We evaluate the model on diffusion processes over physically driven evolving meshes and real-world scanned surfaces. The results show improved temporal stability and prediction accuracy compared with graph convolutional and multilayer perceptron baselines, approaching the performance of Crank-Nicolson solvers on unstructured domains.

Method: Proposes an integrated air-ground collaborative network with a hierarchical learning framework: 1) Vickrey-Clarke-Groves auction mechanism for energy-aware trajectory assignment at large timescale, and 2) diffusion-heterogeneous-agent proximal policy optimization (generative multi-agent RL) with latent diffusion models in actor networks at small timescale.

Result: Extensive simulations show the framework outperforms baselines in energy efficiency, task success rate, and convergence performance.

Conclusion: The proposed hierarchical learning framework effectively addresses the joint optimization of UAV trajectory planning and task offloading in LAINs, demonstrating superior performance over existing approaches.

Abstract: The low-altitude intelligent networks (LAINs) emerge as a promising architecture for delivering low-latency and energy-efficient edge intelligence in dynamic and infrastructure-limited environments. By integrating unmanned aerial vehicles (UAVs), aerial base stations, and terrestrial base stations, LAINs can support mission-critical applications such as disaster response, environmental monitoring, and real-time sensing. However, these systems face key challenges, including energy-constrained UAVs, stochastic task arrivals, and heterogeneous computing resources. To address these issues, we propose an integrated air-ground collaborative network and formulate a time-dependent integer nonlinear programming problem that jointly optimizes UAV trajectory planning and task offloading decisions. The problem is challenging to solve due to temporal coupling among decision variables. Therefore, we design a hierarchical learning framework with two timescales. At the large timescale, a Vickrey-Clarke-Groves auction mechanism enables the energy-aware and incentive-compatible trajectory assignment. At the small timescale, we propose the diffusion-heterogeneous-agent proximal policy optimization, a generative multi-agent reinforcement learning algorithm that embeds latent diffusion models into actor networks. Each UAV samples actions from a Gaussian prior and refines them via observation-conditioned denoising, enhancing adaptability and policy diversity. Extensive simulations show that our framework outperforms baselines in energy efficiency, task success rate, and convergence performance.

Method: Analytically show that phase transitions in DNN learning are governed by saddle points in the loss landscape. Introduce a simple, fast algorithm using L2 regularizer to probe error landscape geometry. Apply it to confirm mode connectivity in DNNs trained on MNIST by finding paths connecting global minima.

Result: Successfully confirmed mode connectivity in DNNs trained on MNIST by efficiently finding paths connecting global minima. Numerically showed that saddle points induce transitions between models encoding distinct digit classes. Established that saddle points govern phase transitions in DNN learning.

Conclusion: The work establishes the geometric origin of key training phenomena in DNNs and reveals a hierarchy of accuracy basins analogous to phases in statistical physics, unifying phase transitions, saddle points, and mode connectivity within a single geometric framework.

Abstract: Training Deep Neural Networks relies on the model converging on a high-dimensional, non-convex loss landscape toward a good minimum. Yet, much of the phenomenology of training remains ill understood. We focus on three seemingly disparate observations: the occurrence of phase transitions reminiscent of statistical physics, the ubiquity of saddle points, and phenomenon of mode connectivity relevant for model merging. We unify these within a single explanatory framework, the geometry of the loss and error landscapes. We analytically show that phase transitions in DNN learning are governed by saddle points in the loss landscape. Building on this insight, we introduce a simple, fast, and easy to implement algorithm that uses the L2 regularizer as a tool to probe the geometry of error landscapes. We apply it to confirm mode connectivity in DNNs trained on the MNIST dataset by efficiently finding paths that connect global minima. We then show numerically that saddle points induce transitions between models that encode distinct digit classes. Our work establishes the geometric origin of key training phenomena in DNNs and reveals a hierarchy of accuracy basins analogous to phases in statistical physics.

Method: Proposed global sensitivity metric based on Individual Conditional Expectation (ICE) curves that computes expected feature importance across ICE curves along with their standard deviation. Also introduced ICE-based correlation value to quantify how interactions modify input-output relationships. Provided mathematical proof that PDP-based sensitivity is a lower bound of ICE-based metric under truncated orthogonal polynomial expansion.

Result: Evaluated on three cases: 5-variable analytical function, 5-variable wind-turbine fatigue problem, and 9-variable airfoil aerodynamics case. ICE-based sensitivity benchmarked against PDP, SHAP, and Sobol’ indices. ICE-based feature importance provides richer insights than traditional PDP-based approach, and visual interpretations from PDP, ICE, and SHAP complement each other.

Conclusion: ICE-based sensitivity metric effectively captures interaction effects that PDP obscures, offering improved interpretability for engineering applications where understanding complex variable interactions is crucial. Multiple visualization methods (PDP, ICE, SHAP) provide complementary perspectives for comprehensive model interpretation.

Abstract: Explainable machine learning techniques have gained increasing attention in engineering applications, especially in aerospace design and analysis, where understanding how input variables influence data-driven models is essential. Partial Dependence Plots (PDPs) are widely used for interpreting black-box models by showing the average effect of an input variable on the prediction. However, their global sensitivity metric can be misleading when strong interactions are present, as averaging tends to obscure interaction effects. To address this limitation, we propose a global sensitivity metric based on Individual Conditional Expectation (ICE) curves. The method computes the expected feature importance across ICE curves, along with their standard deviation, to more effectively capture the influence of interactions. We provide a mathematical proof demonstrating that the PDP-based sensitivity is a lower bound of the proposed ICE-based metric under truncated orthogonal polynomial expansion. In addition, we introduce an ICE-based correlation value to quantify how interactions modify the relationship between inputs and the output. Comparative evaluations were performed on three cases: a 5-variable analytical function, a 5-variable wind-turbine fatigue problem, and a 9-variable airfoil aerodynamics case, where ICE-based sensitivity was benchmarked against PDP, SHapley Additive exPlanations (SHAP), and Sobol’ indices. The results show that ICE-based feature importance provides richer insights than the traditional PDP-based approach, while visual interpretations from PDP, ICE, and SHAP complement one another by offering multiple perspectives.

Method: Fine-tune LLMs to evade activation monitors for benign concepts when conditioned on a trigger phrase, then test generalization to safety-relevant concepts. Validate across diverse model families and analyze mechanisms via mechanistic analysis on Gemma-2-9b-it.

Result: Created “Neural Chameleons” capable of zero-shot evasion of activation monitors, with selective targeting of specific concepts, modest impact on capabilities, and evasion persisting against monitors trained post hoc on frozen weights. Mechanistic analysis reveals targeted manipulation into low-dimensional subspace.

Conclusion: Provides proof-of-concept for activation monitoring failure mode under misalignment threat models, demonstrating models can learn to actively conceal internal states and evade safety monitors, highlighting need for worst-case robustness evaluation.

Abstract: Activation monitoring, which probes a model’s internal states using lightweight classifiers, is an emerging tool for AI safety. However, its worst-case robustness under a misalignment threat model–where a model might learn to actively conceal its internal states–remains untested. Focusing on this threat model, we ask: could a model learn to evade previously unseen activation monitors? Our core contribution is to stress-test the learnability of this behavior. We demonstrate that finetuning can create Neural Chameleons: models capable of zero-shot evading activation monitors. Specifically, we fine-tune an LLM to evade monitors for a set of benign concepts (e.g., languages, HTML) when conditioned on a trigger of the form: “You are being probed for {concept}”. We show that this learned mechanism generalizes zero-shot: by substituting {concept} with a safety-relevant term like ‘deception’, the model successfully evades previously unseen safety monitors. We validate this phenomenon across diverse model families (Llama, Gemma, Qwen), showing that the evasion succeeds even against monitors trained post hoc on the model’s frozen weights. This evasion is highly selective, targeting only the specific concept mentioned in the trigger, and having a modest impact on model capabilities on standard benchmarks. Using Gemma-2-9b-it as a case study, a mechanistic analysis reveals this is achieved via a targeted manipulation that moves activations into a low-dimensional subspace. While stronger defenses like monitor ensembles and non-linear classifiers show greater resilience, the model retains a non-trivial evasion capability. Our work provides a proof-of-concept for this failure mode and a tool to evaluate the worst-case robustness of monitoring techniques against misalignment threat models.

Method: Proposes a framework with a reinforcement learning based context generator designed in an autoencoder-like fashion. It extracts contextual signals from user prompts and uses them to guide response generation, particularly focusing on safety-critical scenarios.

Result: Reduces harmful responses by 5.6% on SafetyInstruct dataset across multiple foundation models, and improves the harmonic mean of attack success rate and compliance on benign prompts by 6.2% on XSTest and WildJailbreak datasets.

Conclusion: Context extraction from user prompts enables safer and more reliable LLM inference by better interpreting ambiguous requests and distinguishing between harmful and benign queries, addressing both safety and usability concerns.

Abstract: User prompts to large language models (LLMs) are often ambiguous or under-specified, and subtle contextual cues shaped by user intentions, prior knowledge, and risk factors strongly influence what constitutes an appropriate response. Misinterpreting intent or risks may lead to unsafe outputs, while overly cautious interpretations can cause unnecessary refusal of benign requests. In this paper, we question the conventional framework in which LLMs generate immediate responses to requests without considering broader contextual factors. User requests are situated within broader contexts such as intentions, knowledge, and prior experience, which strongly influence what constitutes an appropriate answer. We propose a framework that extracts and leverages such contextual information from the user prompt itself. Specifically, a reinforcement learning based context generator, designed in an autoencoder-like fashion, is trained to infer contextual signals grounded in the prompt and use them to guide response generation. This approach is particularly important for safety tasks, where ambiguous requests may bypass safeguards while benign but confusing requests can trigger unnecessary refusals. Experiments show that our method reduces harmful responses by an average of 5.6% on the SafetyInstruct dataset across multiple foundation models and improves the harmonic mean of attack success rate and compliance on benign prompts by 6.2% on XSTest and WildJailbreak. These results demonstrate the effectiveness of context extraction for safer and more reliable LLM inferences.

Method: The system provides real-time process monitoring, benchmarking across multiple platforms (Ollama, LM Studio, vLLM, OpenAI-compatible APIs), persistent storage with visualizations for longitudinal analysis, and personalized model/optimization recommendations.

Result: The toolkit enables users to assess quality-efficiency tradeoffs by combining LLM-as-judge evaluations with energy and speed metrics when testing models with custom prompts.

Conclusion: EnviroLLM addresses the need for comprehensive monitoring and optimization tools for locally deployed LLMs, helping users make informed decisions about model selection and resource usage.

Abstract: Large language models (LLMs) are increasingly deployed locally for privacy and accessibility, yet users lack tools to measure their resource usage, environmental impact, and efficiency metrics. This paper presents EnviroLLM, an open-source toolkit for tracking, benchmarking, and optimizing performance and energy consumption when running LLMs on personal devices. The system provides real-time process monitoring, benchmarking across multiple platforms (Ollama, LM Studio, vLLM, and OpenAI-compatible APIs), persistent storage with visualizations for longitudinal analysis, and personalized model and optimization recommendations. The system includes LLM-as-judge evaluations alongside energy and speed metrics, enabling users to assess quality-efficiency tradeoffs when testing models with custom prompts.

Method: Proposes DFedReweighting framework with two-step aggregation: 1) compute preliminary weights based on target performance metrics from auxiliary datasets, 2) refine weights using customized reweighting strategies for final aggregation.

Result: Theoretical analysis shows linear convergence with appropriate metric-strategy combinations. Experiments demonstrate significant improvements in fairness and robustness against Byzantine attacks across diverse scenarios.

Conclusion: The framework provides a flexible approach to achieve various learning objectives in DFL by selecting appropriate target performance metrics and customized reweighting strategies.

Abstract: Decentralized federated learning (DFL) has recently emerged as a promising paradigm that enables multiple clients to collaboratively train machine learning model through iterative rounds of local training, communication, and aggregation without relying on a central server which introduces potential vulnerabilities in conventional Federated Learning. Nevertheless, DFL systems continue to face a range of challenges, including fairness, robustness, etc. To address these challenges, we propose \textbf{DFedReweighting}, a unified aggregation framework designed to achieve diverse objectives in DFL systems via a objective-oriented reweighting aggregation at the final step of each learning round. Specifically, the framework first computes preliminary weights based on \textit{target performance metric} obtained from auxiliary dataset constructed using local data. These weights are then refined using \textit{customized reweighting strategy}, resulting in the final aggregation weights. Our results from the theoretical analysis demonstrate that the appropriate combination of the target performance metric and the customized reweighting strategy ensures linear convergence. Experimental results consistently show that our proposed framework significantly improves fairness and robustness against Byzantine attacks in diverse scenarios. Provided that appropriate target performance metrics and customized reweighting strategy are selected, our framework can achieve a wide range of desired learning objectives.

Method: Proposes Eik-QRL using Eikonal PDE constraints for trajectory-free learning, then Eik-HiQRL adds hierarchical decomposition to handle complex dynamics.

Result: Eik-HiQRL achieves SOTA in offline goal-conditioned navigation and matches TD methods in manipulation tasks with consistent gains over QRL.

Conclusion: Eikonal constraints enable trajectory-free quasimetric learning, and hierarchical decomposition addresses complex dynamics limitations, yielding improved generalization and performance.

Abstract: Goal-Conditioned Reinforcement Learning (GCRL) mitigates the difficulty of reward design by framing tasks as goal reaching rather than maximizing hand-crafted reward signals. In this setting, the optimal goal-conditioned value function naturally forms a quasimetric, motivating Quasimetric RL (QRL), which constrains value learning to quasimetric mappings and enforces local consistency through discrete, trajectory-based constraints. We propose Eikonal-Constrained Quasimetric RL (Eik-QRL), a continuous-time reformulation of QRL based on the Eikonal Partial Differential Equation (PDE). This PDE-based structure makes Eik-QRL trajectory-free, requiring only sampled states and goals, while improving out-of-distribution generalization. We provide theoretical guarantees for Eik-QRL and identify limitations that arise under complex dynamics. To address these challenges, we introduce Eik-Hierarchical QRL (Eik-HiQRL), which integrates Eik-QRL into a hierarchical decomposition. Empirically, Eik-HiQRL achieves state-of-the-art performance in offline goal-conditioned navigation and yields consistent gains over QRL in manipulation tasks, matching temporal-difference methods.

Method: Tested 4 instruction-tuned models on 876 harmful prompts across 20 sampling configurations (4 temperatures × 5 random seeds). Introduced Safety Stability Index (SSI) to measure decision stability. Used Claude 3.5 Haiku as external judge to validate findings.

Result: Found 18-28% of prompts exhibit decision flips across configurations. Higher temperatures significantly reduce decision stability (SSI drops from 0.951 at temp 0.0 to 0.896 at temp 1.0). Single-shot evaluation agrees with multi-sample ground truth only 92.4% of the time.

Conclusion: Single-shot safety evaluations are insufficient for reliable safety assessment. Researchers should use at least 3 samples per prompt for reliable safety evaluation.

Abstract: Current safety evaluations of large language models rely on single-shot testing, implicitly assuming that model responses are deterministic and representative of the model’s safety alignment. We challenge this assumption by investigating the stability of safety refusal decisions across random seeds and temperature settings. Testing four instruction-tuned models from three families (Llama 3.1 8B, Qwen 2.5 7B, Qwen 3 8B, Gemma 3 12B) on 876 harmful prompts across 20 different sampling configurations (4 temperatures x 5 random seeds), we find that 18-28% of prompts exhibit decision flips–the model refuses in some configurations but complies in others–depending on the model. Our Safety Stability Index (SSI) reveals that higher temperatures significantly reduce decision stability (Friedman chi-squared = 44.71, p < 0.001), with mean SSI dropping from 0.951 at temperature 0.0 to 0.896 at temperature 1.0. We validate our findings across all model families using Claude 3.5 Haiku as a unified external judge, achieving 89.0% inter-judge agreement with our primary Llama 70B judge (Cohen’s kappa = 0.62). These findings demonstrate that single-shot safety evaluations are insufficient for reliable safety assessment. We show that single-shot evaluation agrees with multi-sample ground truth only 92.4% of the time, and recommend using at least 3 samples per prompt for reliable safety assessment.

Method: Novel sampling strategy using negative exponential and normal distributions for training points, combined with dual network architecture that learns both solution and equation parameters with same loss function. Implemented with both PINNs and PIKANs for comparison.

Result: PINNs provide more accurate and computationally efficient solution than PIKANs in this setting - solving inverse problems 1000 times faster with same order of magnitude accuracy but lower relative error.

Conclusion: For inverse problems in infinite/semi-infinite domains, PINNs outperform PIKANs in both accuracy and computational efficiency, despite PIKANs’ theoretical advantages in interpretability and polynomial composition structure.

Abstract: Inverse problems are extensively studied in applied mathematics, with applications ranging from acoustic tomography for medical diagnosis to geophysical exploration. Physics informed neural networks (PINNs) have emerged as a powerful tool for solving such problems, while Physics informed Kolmogorov Arnold networks (PIKANs) represent a recent benchmark that, in certain problems, promises greater interpretability and accuracy compared to PINNs, due to their nature, being constructed as a composition of polynomials. In this work, we develop a methodology for addressing inverse problems in infinite and semi infinite domains. We introduce a novel sampling strategy for the network’s training points, using the negative exponential and normal distributions, alongside a dual network architecture that is trained to learn the solution and parameters of an equation with the same loss function. This design enables the solution of inverse problems without explicitly imposing boundary conditions, as long as the solutions tend to stabilize when leaving the domain of interest. The proposed architecture is implemented using both PINNs and PIKANs, and their performance is compared in terms of accuracy with respect to a known solution as well as computational time and response to a noisy environment. Our results demonstrate that, in this setting, PINNs provide a more accurate and computationally efficient solution, solving the inverse problem 1,000 times faster and in the same order of magnitude, yet with a lower relative error than PIKANs.

Method: Joint training of two Transformer models using complementary representations: shapelet-based representations for localized temporal structures and traditional feature engineering for statistical properties, plus a visual analytics system (SigTIme) with coordinated views.

Result: Quantitative evaluation on eight public datasets and one clinical dataset shows effectiveness; usage scenarios with domain experts demonstrate practical utility for ECG analysis and preterm labor assessment.

Conclusion: The proposed framework addresses key limitations in time series pattern discovery by providing interpretable signatures through learned shapelets and enabling multi-perspective exploration via the SigTIme visual analytics system.

Abstract: Understanding and distinguishing temporal patterns in time series data is essential for scientific discovery and decision-making. For example, in biomedical research, uncovering meaningful patterns in physiological signals can improve diagnosis, risk assessment, and patient outcomes. However, existing methods for time series pattern discovery face major challenges, including high computational complexity, limited interpretability, and difficulty in capturing meaningful temporal structures. To address these gaps, we introduce a novel learning framework that jointly trains two Transformer models using complementary time series representations: shapelet-based representations to capture localized temporal structures and traditional feature engineering to encode statistical properties. The learned shapelets serve as interpretable signatures that differentiate time series across classification labels. Additionally, we develop a visual analytics system – SigTIme – with coordinated views to facilitate exploration of time series signatures from multiple perspectives, aiding in useful insights generation. We quantitatively evaluate our learning framework on eight publicly available datasets and one proprietary clinical dataset. Additionally, we demonstrate the effectiveness of our system through two usage scenarios along with the domain experts: one involving public ECG data and the other focused on preterm labor analysis.

Method: Cloak uses latent diffusion models with contrastive learning to extract disentangled representations, guiding the diffusion process to retain useful information while concealing private information, enabling flexible privacy-utility trade-offs with minimal retraining.

Result: Extensive experiments on four public time-series datasets across multiple sensing modalities and a facial image dataset show Cloak consistently outperforms state-of-the-art obfuscation techniques and is suitable for resource-constrained settings.

Conclusion: Cloak provides an effective, practical solution for data obfuscation that achieves superior privacy-utility trade-offs while being deployable on resource-constrained mobile IoT devices, addressing key limitations of previous approaches.

Abstract: Data obfuscation is a promising technique for mitigating attribute inference attacks by semi-trusted parties with access to time-series data emitted by sensors. Recent advances leverage conditional generative models together with adversarial training or mutual information-based regularization to balance data privacy and utility. However, these methods often require modifying the downstream task, struggle to achieve a satisfactory privacy-utility trade-off, or are computationally intensive, making them impractical for deployment on resource-constrained mobile IoT devices. We propose Cloak, a novel data obfuscation framework based on latent diffusion models. In contrast to prior work, we employ contrastive learning to extract disentangled representations, which guide the latent diffusion process to retain useful information while concealing private information. This approach enables users with diverse privacy needs to navigate the privacy-utility trade-off with minimal retraining. Extensive experiments on four public time-series datasets, spanning multiple sensing modalities, and a dataset of facial images demonstrate that Cloak consistently outperforms state-of-the-art obfuscation techniques and is well-suited for deployment in resource-constrained settings.

Method: GraphPerf-RT unifies task DAG topology, CFG-derived code semantics, and runtime context (per-core DVFS, thermal state, utilization) in a heterogeneous graph representation with typed edges encoding precedence, placement, and contention. It uses multi-task evidential heads to predict makespan, energy, cache/branch misses, and utilization with calibrated uncertainty (Normal-Inverse-Gamma distribution).

Result: Achieves R^2 > 0.95 with well-calibrated uncertainty (ECE < 0.05) on three embedded ARM platforms (Jetson TX2, Jetson Orin NX, RUBIK Pi). When integrated with RL scheduling methods, MAMBRL-D3QN with GraphPerf-RT as world model achieves 66% makespan reduction (0.97 +/- 0.35s) and 82% energy reduction (0.006 +/- 0.005J) compared to model-free baselines.

Conclusion: Accurate, uncertainty-aware surrogates like GraphPerf-RT enable effective model-based planning on thermally constrained embedded systems, demonstrating significant performance and energy improvements over traditional approaches.

Abstract: Performance prediction for OpenMP workloads on heterogeneous embedded SoCs is challenging due to complex interactions between task DAG structure, control-flow irregularity, cache and branch behavior, and thermal dynamics; classical heuristics struggle under workload irregularity, tabular regressors discard structural information, and model-free RL risks overheating resource-constrained devices. We introduce GraphPerf-RT, the first surrogate that unifies task DAG topology, CFG-derived code semantics, and runtime context (per-core DVFS, thermal state, utilization) in a heterogeneous graph representation with typed edges encoding precedence, placement, and contention. Multi-task evidential heads predict makespan, energy, cache and branch misses, and utilization with calibrated uncertainty (Normal-Inverse-Gamma), enabling risk-aware scheduling that filters low-confidence rollouts. We validate GraphPerf-RT on three embedded ARM platforms (Jetson TX2, Jetson Orin NX, RUBIK Pi), achieving R^2 > 0.95 with well-calibrated uncertainty (ECE < 0.05). To demonstrate end-to-end scheduling utility, we integrate the surrogate with four RL methods on Jetson TX2: single-agent model-free (SAMFRL), single-agent model-based (SAMBRL), multi-agent model-free (MAMFRL-D3QN), and multi-agent model-based (MAMBRL-D3QN). Experiments across 5 seeds (200 episodes each) show that MAMBRL-D3QN with GraphPerf-RT as the world model achieves 66% makespan reduction (0.97 +/- 0.35s) and 82% energy reduction (0.006 +/- 0.005J) compared to model-free baselines, demonstrating that accurate, uncertainty-aware surrogates enable effective model-based planning on thermally constrained embedded systems.

Method: Proposes a Predictor-Corrector mechanism where: 1) Predictor is any learned time-series model, 2) Corrector is a neural controlled differential equation that predicts forecast errors, 3) Errors are added to forecasts to improve performance, 4) Two regularization strategies improve extrapolation with accelerated training.

Result: The Predictor-Corrector mechanism consistently improves performance compared to Predictor alone across synthetic, physics simulation, and real-world datasets with diverse Predictors (neural ODEs, Contiformer, DLinear). Works with irregularly sampled time series and both continuous- and discrete-time Predictors.

Conclusion: The proposed framework effectively corrects forecast errors from learned time-series models, demonstrating consistent performance improvements across various settings and model types while handling irregular sampling and different time representations.

Abstract: Learned time-series models, whether continuous- or discrete-time, are widely used to forecast the states of a dynamical system. Such models generate multi-step forecasts either directly, by predicting the full horizon at once, or iteratively, by feeding back their own predictions at each step. In both cases, the multi-step forecasts are prone to errors. To address this, we propose a Predictor-Corrector mechanism where the Predictor is any learned time-series model and the Corrector is a neural controlled differential equation. The Predictor forecasts, and the Corrector predicts the errors of the forecasts. Adding these errors to the forecasts improves forecast performance. The proposed Corrector works with irregularly sampled time series and continuous- and discrete-time Predictors. Additionally, we introduce two regularization strategies to improve the extrapolation performance of the Corrector with accelerated training. We evaluate our Corrector with diverse Predictors, e.g., neural ordinary differential equations, Contiformer, and DLinear, on synthetic, physics simulation, and real-world forecasting datasets. The experiments demonstrate that the Predictor-Corrector mechanism consistently improves the performance compared to Predictor alone.

Method: MixtureKit supports three MoE methods: (1) Traditional MoE with single router per transformer block, (2) BTX (Branch-Train-Mix) with separate routers for each sub-layer for fine-grained token routing, and (3) BTS (Branch-Train-Stitch) with trainable stitch layers for controlled information exchange between hub and experts while keeping experts intact.

Result: Experiments with multilingual code-switched data show that BTX-based models trained with MixtureKit outperform baseline dense models on multiple benchmarks. The framework automatically modifies model configurations, patches decoder classes, and saves unified checkpoints ready for inference or fine-tuning.

Conclusion: MixtureKit provides a practical foundation for MoE research and development, offering modular construction, training capabilities, and visualization tools for analyzing routing decisions and expert contributions across diverse domains.

Abstract: We introduce MixtureKit, a modular open-source framework for constructing, training, and analyzing Mixture-of-Experts (MoE) models from arbitrary pre-trained or fine-tuned models. MixtureKit currently supports three complementary methods: (i) \emph{Traditional MoE}, which uses a single router per transformer block to select experts, (ii) \emph{BTX} (Branch-Train-Mix), which introduces separate routers for each specified sub-layer enabling fine-grained token routing, and (iii) \emph{BTS} (Branch-Train-Stitch), which keeps experts fully intact and introduces trainable stitch layers for controlled information exchange between hub and experts. MixtureKit automatically modifies the model configuration, patches decoder and causal LM classes, and saves a unified checkpoint ready for inference or fine-tuning. We further provide a visualization interface to inspect per-token routing decisions, expert weight distributions, and layer-wise contributions. Experiments with multilingual code-switched data (e.g. Arabic-Latin) show that a BTX-based model trained using MixtureKit can outperform baseline dense models on multiple benchmarks. We release MixtureKit as a practical foundation for research and development of MoE-based systems across diverse domains.

Method: 1) CP-TDA with Randomized Composite PCA initialization for tensor Gaussian mixture models, 2) Semiparametric extension using deep generative models to map non-normal data into latent space for CP-TDA, 3) Theoretical analysis of convergence and misclassification rates.

Result: Established global convergence and minimax-optimal misclassification rates. Numerical studies show substantial gains over existing tensor classifiers and graph neural networks, especially in high-dimensional, small-sample regimes.

Conclusion: CP-TDA provides theoretically grounded tensor classification with CP low-rank structure, extended via deep generative models for non-normal data, achieving superior performance in challenging high-dimensional settings.

Abstract: High-dimensional tensor-valued predictors arise in modern applications, increasingly as learned representations from neural networks. Existing tensor classification methods rely on sparsity or Tucker structures and often lack theoretical guarantees. Motivated by empirical evidence that discriminative signals concentrate along a few multilinear components, we introduce CP low-rank structure for the discriminant tensor, a modeling perspective not previously explored. Under a Tensor Gaussian Mixture Model, we propose high-dimensional CP low-rank Tensor Discriminant Analysis (CP-TDA) with Randomized Composite PCA (\textsc{rc-PCA}) initialization, that is essential for handling dependent and anisotropic noise under weaker signal strength and incoherence conditions, followed by iterative refinement algorithm. We establish global convergence and minimax-optimal misclassification rates. To handle tensor data deviating from tensor normality, we develop the first semiparametric tensor discriminant model, in which learned tensor representations are mapped via deep generative models into a latent space tailored for CP-TDA. Misclassification risk decomposes into representation, approximation, and estimation errors. Numerical studies and real data analysis on graph classification demonstrate substantial gains over existing tensor classifiers and state-of-the-art graph neural networks, particularly in high-dimensional, small-sample regimes.

Method: BOOST introduces Bottleneck-aware Tensor Parallelism and combines optimizations including online-RMSNorm, linear layer grouping, and low-rank activation checkpointing to achieve end-to-end training speedup for low-rank bottleneck architectures.

Result: BOOST achieves 1.46-1.91× speedup over full-rank model baselines and 1.87-2.27× speedup over low-rank models with naively integrated 3D parallelism, with improved GPU utilization and reduced communication overhead.

Conclusion: BOOST provides an efficient training framework for large-scale low-rank bottleneck architectures that addresses the scaling limitations of standard parallelism methods, enabling faster training with better resource utilization.

Abstract: The scale of transformer model pre-training is constrained by the increasing computation and communication cost. Low-rank bottleneck architectures offer a promising solution to significantly reduce the training time and memory footprint with minimum impact on accuracy. Despite algorithmic efficiency, bottleneck architectures scale poorly under standard tensor parallelism. Simply applying 3D parallelism designed for full-rank methods leads to excessive communication and poor GPU utilization. To address this limitation, we propose BOOST, an efficient training framework tailored for large-scale low-rank bottleneck architectures. BOOST introduces a novel Bottleneck-aware Tensor Parallelism, and combines optimizations such as online-RMSNorm, linear layer grouping, and low-rank activation checkpointing to achieve end-to-end training speedup. Evaluations on different low-rank bottleneck architectures demonstrate that BOOST achieves 1.46-1.91$\times$ speedup over full-rank model baselines and 1.87-2.27$\times$ speedup over low-rank model with naively integrated 3D parallelism, with improved GPU utilization and reduced communication overhead.

Method: Developed a novel hierarchical construction beginning with an efficient approximation of the square function x² that is more compact in depth and size than comparable ReLU realizations. Extended this approach through functional composition to establish approximation bounds for deep SiLU networks.

Result: SiLU networks achieve approximation error decaying as O(ω^{-2k}) using networks of depth O(1). For Sobolev-class functions, they achieve sharp approximation bounds with total depth O(1) and size O(ε^{-d/n}).

Conclusion: SiLU activation networks provide exponential approximation rates for smooth functions with improved complexity control compared to ReLU networks, establishing them as theoretically superior for function approximation tasks.

Abstract: This article establishes a comprehensive theoretical framework demonstrating that SiLU (Sigmoid Linear Unit) activation networks achieve exponential approximation rates for smooth functions with explicit and improved complexity control compared to classical ReLU-based constructions. We develop a novel hierarchical construction beginning with an efficient approximation of the square function $x^2$ more compact in depth and size than comparable ReLU realizations, such as those given by Yarotsky. This construction yields an approximation error decaying as $\mathcal{O}(ω^{-2k})$ using networks of depth $\mathcal{O}(1)$. We then extend this approach through functional composition to establish sharp approximation bounds for deep SiLU networks in approximating Sobolev-class functions, with total depth $\mathcal{O}(1)$ and size $\mathcal{O}(\varepsilon^{-d/n})$.

Method: Proposed a spatiotemporal transformer model with self-supervised masked latent reconstruction task that allows flexible spatial scale token encoding and masking, applied to publicly available iEEG data.

Result: Adjusting spatial scale for token encoding and masking significantly impacts downstream decoding; larger spatial scales than channel-level improve performance; enables region-level encoding while maintaining accurate channel-level reconstruction.

Conclusion: The framework enables exploration of spatial scales in neurofoundation models, reveals their importance for self-supervised pretraining, and enhances downstream decoding performance for multiregional brain activity.

Abstract: Intracranial recordings have opened a unique opportunity to simultaneously measure activity across multiregional networks in the human brain. Recent works have focused on developing transformer-based neurofoundation models of such recordings that can generalize across subjects and datasets. However, these recordings exhibit highly complex spatiotemporal interactions across diverse spatial scales, from the single-channel scale to the scale of brain regions. As such, there remain critical open questions regarding how best to encode spatial information and how to design self-supervision tasks that enable the learning of brain network patterns and enhance downstream decoding performance using such high-dimensional, multiregional recordings. To allow for exploring these questions, we propose a new spatiotemporal transformer model of multiregional neural activity and a corresponding self-supervised masked latent reconstruction task, designed to enable flexibility in the spatial scale used for token encoding and masking. Applying this model on publicly available multiregional intracranial electrophysiology (iEEG) data, we demonstrate that adjusting the spatial scale for both token encoding and masked reconstruction significantly impacts downstream decoding. Further, we find that spatial encoding at larger scales than channel-level encoding, which is commonly used in existing iEEG transformer models, improves downstream decoding performance. Finally, we demonstrate that our method allows for region-level token encoding while also maintaining accurate channel-level neural reconstruction. Taken together, our modeling framework enables exploration of the spatial scales used for token encoding and masking, reveals their importance towards self-supervised pretraining of neurofoundation models of multiregional human brain activity, and enhances downstream decoding performance.

Method: Developed HydroDiffusion with decoder-only state space model backbone that jointly denoises full multi-day trajectories in a single pass, ensuring temporal coherence and mitigating error accumulation from autoregressive prediction.

Result: Evaluated across 531 watersheds in CONUS using CAMELS dataset. Achieved strong nowcast accuracy with observed meteorological forcings, maintained consistent performance across full simulation horizon, and outperformed DRUM in operational forecasting with stronger deterministic and probabilistic forecast skill.

Conclusion: HydroDiffusion establishes a robust generative modeling framework for medium-range streamflow forecasting, providing both a new modeling benchmark and foundation for future research on probabilistic hydrologic prediction at continental scales.

Abstract: Recent advances have introduced diffusion models for probabilistic streamflow forecasting, demonstrating strong early flood-warning skill. However, current implementations rely on recurrent Long Short-Term Memory (LSTM) backbones and single-step training objectives, which limit their ability to capture long-range dependencies and produce coherent forecast trajectories across lead times. To address these limitations, we developed HydroDiffusion, a diffusion-based probabilistic forecasting framework with a decoder-only state space model backbone. The proposed framework jointly denoises full multi-day trajectories in a single pass, ensuring temporal coherence and mitigating error accumulation common in autoregressive prediction. HydroDiffusion is evaluated across 531 watersheds in the contiguous United States (CONUS) in the CAMELS dataset. We benchmark HydroDiffusion against two diffusion baselines with LSTM backbones, as well as the recently proposed Diffusion-based Runoff Model (DRUM). Results show that HydroDiffusion achieves strong nowcast accuracy when driven by observed meteorological forcings, and maintains consistent performance across the full simulation horizon. Moreover, HydroDiffusion delivers stronger deterministic and probabilistic forecast skill than DRUM in operational forecasting. These results establish HydroDiffusion as a robust generative modeling framework for medium-range streamflow forecasting, providing both a new modeling benchmark and a foundation for future research on probabilistic hydrologic prediction at continental scales.

Method: Integrates state-of-the-art guidance methods (classifier-free guidance, autoguidance, model guidance) into an SE(3)-equivariant flow matching framework. Proposes hybrid guidance strategy that separately guides continuous (velocity fields) and discrete (predicted logits) molecular modalities, with joint optimization of guidance scales via Bayesian optimization.

Result: Achieves new state-of-the-art performance in property alignment for de novo molecular generation on QM9 and QMe14S datasets. Generated molecules exhibit high structural validity. Provides systematic comparison of various guidance methods’ strengths and limitations.

Conclusion: The proposed hybrid guidance strategy successfully integrates computer vision guidance methods into molecular generation, demonstrating superior property alignment while maintaining chemical validity, with insights for broader applicability of guidance techniques.

Abstract: Key objectives in conditional molecular generation include ensuring chemical validity, aligning generated molecules with target properties, promoting structural diversity, and enabling efficient sampling for discovery. Recent advances in computer vision introduced a range of new guidance strategies for generative models, many of which can be adapted to support these goals. In this work, we integrate state-of-the-art guidance methods – including classifier-free guidance, autoguidance, and model guidance – in a leading molecule generation framework built on an SE(3)-equivariant flow matching process. We propose a hybrid guidance strategy that separately guides continuous and discrete molecular modalities – operating on velocity fields and predicted logits, respectively – while jointly optimizing their guidance scales via Bayesian optimization. Our implementation, benchmarked on the QM9 and QMe14S datasets, achieves new state-of-the-art performance in property alignment for de novo molecular generation. The generated molecules also exhibit high structural validity. Furthermore, we systematically compare the strengths and limitations of various guidance methods, offering insights into their broader applicability.

Method: 1. Encode EEG segments into compact latent representations using self-supervised autoencoder. 2. Filter out outliers and minimize redundancy based on these representations. 3. Create smaller, informative subset that retains diversity for foundation model training.

Result: Training on only 5% of a 2,500-hour dataset curated with EEG-DLite yields performance comparable to or better than training on the full dataset across multiple downstream tasks.

Conclusion: EEG-DLite provides the first systematic study of pre-training data distillation for EEG foundation models, offering a scalable and practical path toward more effective and efficient physiological foundation modeling.

Abstract: Large-scale EEG foundation models have shown strong generalization across a range of downstream tasks, but their training remains resource-intensive due to the volume and variable quality of EEG data. In this work, we introduce EEG-DLite, a data distillation framework that enables more efficient pre-training by selectively removing noisy and redundant samples from large EEG datasets. EEG-DLite begins by encoding EEG segments into compact latent representations using a self-supervised autoencoder, allowing sample selection to be performed efficiently and with reduced sensitivity to noise. Based on these representations, EEG-DLite filters out outliers and minimizes redundancy, resulting in a smaller yet informative subset that retains the diversity essential for effective foundation model training. Through extensive experiments, we demonstrate that training on only 5 percent of a 2,500-hour dataset curated with EEG-DLite yields performance comparable to, and in some cases better than, training on the full dataset across multiple downstream tasks. To our knowledge, this is the first systematic study of pre-training data distillation in the context of EEG foundation models. EEG-DLite provides a scalable and practical path toward more effective and efficient physiological foundation modeling. The code is available at https://github.com/t170815518/EEG-DLite.

Method: Partitions input domain and assigns each partition to its own CMS instance, with CMS parameters analytically derived for fixed thresholds, and thresholds optimized via dynamic programming with approximate feasibility checks.

Result: OptLCMS builds faster, achieves lower intolerable error probability, and matches the estimation accuracy of LCMS while providing theoretical guarantees and explicit error threshold control.

Conclusion: OptLCMS offers a practical improvement over LCMS with faster construction, theoretical guarantees, and flexible error control while maintaining estimation accuracy.

Abstract: Count-Min Sketch (CMS) is a memory-efficient data structure for estimating the frequency of elements in a multiset. Learned Count-Min Sketch (LCMS) enhances CMS with a machine learning model to reduce estimation error under the same memory usage, but suffers from slow construction due to empirical parameter tuning and lacks theoretical guarantees on intolerable error probability. We propose Optimized Learned Count-Min Sketch (OptLCMS), which partitions the input domain and assigns each partition to its own CMS instance, with CMS parameters analytically derived for fixed thresholds, and thresholds optimized via dynamic programming with approximate feasibility checks. This reduces the need for empirical validation, enabling faster construction while providing theoretical guarantees under these assumptions. OptLCMS also allows explicit control of the allowable error threshold, improving flexibility in practice. Experiments show that OptLCMS builds faster, achieves lower intolerable error probability, and matches the estimation accuracy of LCMS.

Method: Proposes GRC-Net with: 1) Gram Matrix transformation of EEG signals into 3D representation preserving temporal dependencies, 2) Multi-level feature extraction using coattention for global signals and inception structure for local signals, enabling multi-granular feature learning.

Result: Achieved 93.66% accuracy on the BONN dataset for the most challenging five-class epilepsy classification task, outperforming existing methods.

Conclusion: The Gram Matrix 3D representation combined with multi-level feature extraction (coattention for global, inception for local) effectively addresses EEG signal complexity and imbalance, achieving state-of-the-art epilepsy prediction performance.

Abstract: Prediction of epilepsy based on electroencephalogram (EEG) signals is a rapidly evolving field. Previous studies have traditionally applied 1D processing to the entire EEG signal. However, we have adopted the Gram Matrix method to transform the signals into a 3D representation, enabling modeling of signal relationships across dimensions while preserving the temporal dependencies of the one-dimensional signals. Additionally, we observed an imbalance between local and global signals within the EEG data. Therefore, we introduced multi-level feature extraction, utilizing coattention for capturing global signal characteristics and an inception structure for processing local signals, achieving multi-granular feature extraction. Our experiments on the BONN dataset demonstrate that for the most challenging five-class classification task, GRC-Net achieved an accuracy of 93.66%, outperforming existing methods.

Method: Multi-Fidelity Ensemble Kalman Filter (MF-EnKF) that combines a few expensive full physical model runs with many cheap ML surrogate runs, tested on Lorenz-2005 and Quasi-Geostrophic models.

Result: MF-EnKF achieves higher accuracy than completely replacing physical models with ML, provides improved accuracy within same computational budget, and ML surrogates offer larger speed-ups than low-resolution models.

Conclusion: The method effectively increases ensemble size in EnKF, improving initial condition estimation and prediction accuracy for meteorological and oceanographic applications.

Abstract: Currently, more and more machine learning (ML) surrogates are being developed for computationally expensive physical models. In this work we investigate the use of a Multi-Fidelity Ensemble Kalman Filter (MF-EnKF) in which the low-fidelity model is such a machine learning surrogate model, instead of a traditional low-resolution or reduced-order model. The idea behind this is to use an ensemble of a few expensive full model runs, together with an ensemble of many cheap but less accurate ML model runs. In this way we hope to reach increased accuracy within the same computational budget. We investigate the performance by testing the approach on two common test problems, namely the Lorenz-2005 model and the Quasi-Geostrophic model. By keeping the original physical model in place, we obtain a higher accuracy than when we completely replace it by the ML model. Furthermore, the MF-EnKF reaches improved accuracy within the same computational budget. The ML surrogate has similar or improved accuracy compared to the low-resolution one, but it can provide a larger speed-up. Our method contributes to increasing the effective ensemble size in the EnKF, which improves the estimation of the initial condition and hence accuracy of the predictions in fields such as meteorology and oceanography.

Method: Develops FDIFF-PINN that integrates fractional calculus with deep learning, constructs discretized fractional-order partial differential equation based on fractional-order equivalent circuit model.

Result: Comparative experiments conducted using dynamic charge/discharge dataset of Panasonic 18650PF batteries under multi-temperature conditions (-10°C to 20°C).

Conclusion: The proposed FDIFF-PINN approach addresses limitations of conventional neural networks for SOC estimation by incorporating physical knowledge through fractional calculus, potentially improving accuracy and interpretability.

Abstract: Accurate estimation of the State of Charge (SOC) is critical for ensuring the safety, reliability, and performance optimization of lithium-ion battery systems. Conventional data-driven neural network models often struggle to fully characterize the inherent complex nonlinearities and memory-dependent dynamics of electrochemical processes, significantly limiting their predictive accuracy and physical interpretability under dynamic operating conditions. To address this challenge, this study proposes a novel neural architecture termed the Fractional Differential Equation Physics-Informed Neural Network (FDIFF-PINN), which integrates fractional calculus with deep learning. The main contributions of this paper include: (1) Based on a fractional-order equivalent circuit model, a discretized fractional-order partial differential equation is constructed. (2) Comparative experiments were conducted using a dynamic charge/discharge dataset of Panasonic 18650PF batteries under multi-temperature conditions (from -10$^{\circ}$C to 20$^{\circ}$C).

Method: Hierarchical Transformer with two stages: (1) Local Informer with top-k Sparse Attention for intra-patch dynamics, followed by mean pooling; (2) Global Informer with same top-k attention for inter-patch dependencies. Lightweight GRU aggregates globally contextualized patch tokens for multi-horizon prediction.

Result: Achieves linear O(kLd) time/memory complexity. On 8 real-world datasets across 6 domains with horizons 96-720, secures 27 positions in top two out of 34. Achieves best performance on MAE/RMSE at 17 places and second-best at 10 places, consistently outperforming PatchTST, iTransformer, FEDformer, Informer, and vanilla Transformers.

Conclusion: TwinFormer effectively addresses long-sequence forecasting challenges through hierarchical processing with sparse attention, achieving superior performance with linear complexity. The top-k sparse attention proves superior to ProbSparse, and GRU-based aggregation is effective for multi-horizon prediction.

Abstract: TwinFormer is a hierarchical Transformer for long-sequence time-series forecasting. It divides the input into non-overlapping temporal patches and processes them in two stages: (1) a Local Informer with top-$k$ Sparse Attention models intra-patch dynamics, followed by mean pooling; (2) a Global Informer captures long-range inter-patch dependencies using the same top-$k$ attention. A lightweight GRU aggregates the globally contextualized patch tokens for direct multi-horizon prediction. The resulting architecture achieves linear $O(kLd)$ time and memory complexity. On eight real-world benchmarking datasets from six different domains, including weather, stock price, temperature, power consumption, electricity, and disease, and forecasting horizons $96-720$, TwinFormer secures $27$ positions in the top two out of $34$. Out of the $27$, it achieves the best performance on MAE and RMSE at $17$ places and $10$ at the second-best place on MAE and RMSE. This consistently outperforms PatchTST, iTransformer, FEDformer, Informer, and vanilla Transformers. Ablations confirm the superiority of top-$k$ Sparse Attention over ProbSparse and the effectiveness of GRU-based aggregation. Code is available at this repository: https://github.com/Mahimakumavat1205/TwinFormer.

Method: Analyzes Robbins’ classic sub-Gaussian mixture in a stochastic setting and proves it satisfies path-wise deterministic regret bounds. Uses Ville events to establish bounds that hold deterministically on certain paths, with the cumulative variance process V_T playing a key role in the analysis.

Result: For every path in Ville event E_α, regret till time T is bounded by ln²(1/α)/V_T + ln(1/α) + ln ln V_T (up to constants). When V_T ≥ ln(1/α), this reduces to ln(1/α) + ln ln V_T. On event E_0 (probability one), regret is eventually bounded by ln ln V_T. These bounds hold under various distributional assumptions (sub-Gaussian, symmetric, variance-bounded).

Conclusion: Conditional regret bounds on Ville events serve as a bridge between stochastic and adversarial betting approaches, allowing deterministic guarantees that connect the worlds of adversarial online learning and game-theoretic statistics while handling potentially unbounded data.

Abstract: We prove that a classic sub-Gaussian mixture proposed by Robbins in a stochastic setting actually satisfies a path-wise (deterministic) regret bound. For every path in a natural ``Ville event’’ $E_α$, this regret till time $T$ is bounded by $\ln^2(1/α)/V_T + \ln (1/α) + \ln \ln V_T$ up to universal constants, where $V_T$ is a nonnegative, nondecreasing, cumulative variance process. (The bound reduces to $\ln(1/α) + \ln \ln V_T$ if $V_T \geq \ln(1/α)$.) If the data were stochastic, then one can show that $E_α$ has probability at least $1-α$ under a wide class of distributions (eg: sub-Gaussian, symmetric, variance-bounded, etc.). In fact, we show that on the Ville event $E_0$ of probability one, the regret on every path in $E_0$ is eventually bounded by $\ln \ln V_T$ (up to constants). We explain how this work helps bridge the world of adversarial online learning (which usually deals with regret bounds for bounded data), with game-theoretic statistics (which can handle unbounded data, albeit using stochastic assumptions). In short, conditional regret bounds serve as a bridge between stochastic and adversarial betting.

Method: Use a flexible family of uncertainty measures distinguishing total, aleatoric, and epistemic uncertainty of second-order distributions, instantiated with proper scoring rules to control characteristics.

Result: For selective prediction, scoring rule should match task loss; for OOD detection, mutual information (epistemic uncertainty) performs best; for active learning, epistemic uncertainty based on zero-one loss outperforms other measures.

Conclusion: Uncertainty quantification should be task-specific rather than seeking a universal best measure, with different scoring rules and uncertainty types optimal for different applications.

Abstract: Proper quantification of predictive uncertainty is essential for the use of machine learning in safety-critical applications. Various uncertainty measures have been proposed for this purpose, typically claiming superiority over other measures. In this paper, we argue that there is no single best measure. Instead, uncertainty quantification should be tailored to the specific application. To this end, we use a flexible family of uncertainty measures that distinguishes between total, aleatoric, and epistemic uncertainty of second-order distributions. These measures can be instantiated with specific loss functions, so-called proper scoring rules, to control their characteristics, and we show that different characteristics are useful for different tasks. In particular, we show that, for the task of selective prediction, the scoring rule should ideally match the task loss. On the other hand, for out-of-distribution detection, our results confirm that mutual information, a widely used measure of epistemic uncertainty, performs best. Furthermore, in an active learning setting, epistemic uncertainty based on zero-one loss is shown to consistently outperform other uncertainty measures.

Method: Created a proof-of-concept Synthetic Swarm Mosquito Dataset that simulates realistic multi-species and noisy swarm conditions. Used log-mel spectrograms and evaluated lightweight deep learning architectures for mosquito species classification.

Result: Models effectively identified six major mosquito vectors and were found suitable for deployment on embedded low-power devices. The synthetic approach enables scalable data generation while reducing human resource demands.

Conclusion: Synthetic swarm audio datasets have potential to accelerate acoustic mosquito research and enable scalable real-time surveillance solutions for mosquito-borne disease control.

Abstract: Mosquito-borne diseases pose a serious global health threat, causing over 700,000 deaths annually. This work introduces a proof-of-concept Synthetic Swarm Mosquito Dataset for Acoustic Classification, created to simulate realistic multi-species and noisy swarm conditions. Unlike conventional datasets that require labor-intensive recording of individual mosquitoes, the synthetic approach enables scalable data generation while reducing human resource demands. Using log-mel spectrograms, we evaluated lightweight deep learning architectures for the classification of mosquito species. Experiments show that these models can effectively identify six major mosquito vectors and are suitable for deployment on embedded low-power devices. The study demonstrates the potential of synthetic swarm audio datasets to accelerate acoustic mosquito research and enable scalable real-time surveillance solutions.

Method: Conducted early-stage scaling-law analysis by training 1B and 3B-parameter Llama-3.2 models on 400M-token financial corpus (SEC filings), with validation checkpoints at 50M, 100M, 200M, and 400M tokens. Used power-law fits to analyze learning efficiency.

Result: Consistent improvements in SEC-domain validation loss with largest gains in first 200M tokens and diminishing returns thereafter. General-domain loss remained unchanged, showing minimal drift and no catastrophic forgetting. Power-law fits revealed shallow exponents indicating financial language is highly regular and efficiently learnable.

Conclusion: Meaningful domain adaptation can be achieved with comparatively modest token budgets, and larger model scales (7B-70B) remain tractable under projected data requirements. DAPT offers efficient specialization without compromising general capabilities.

Abstract: Domain-adaptive pretraining (DAPT) offers a practical path to specializing large language models for high-value domains without full retraining. We conduct an early-stage scaling-law analysis of continued pretraining on U.S. SEC filings, training 1B and 3B-parameter Llama-3.2 models on a 400M-token financial corpus with validation checkpoints at 50M, 100M, 200M, and 400M tokens. Results show consistent improvements in SEC-domain validation loss for both models, with the largest gains occurring within the first 200M tokens and diminishing returns thereafter. Power-law fits reveal shallow exponents, indicating that financial language is highly regular and efficiently learnable under continued pretraining. General-domain validation loss remains effectively unchanged across all token budgets, suggesting minimal drift and no signs of catastrophic forgetting. A data-efficiency frontier further shows that both models move toward improved specialization with negligible mixed-domain degradation. Together, these findings provide early empirical guidance for scaling financial foundation models, suggesting that meaningful domain adaptation can be achieved with comparatively modest token budgets and that larger model scales (7B-70B) remain tractable under projected data requirements.

Method: VGPO transforms sparse terminal rewards into dense, process-aware value estimates that model expected cumulative reward at each generative stage for precise temporal credit assignment. It replaces standard group normalization with a novel process enhanced by absolute values to maintain stable optimization signals even when reward diversity declines.

Result: Extensive experiments on three benchmarks show VGPO achieves state-of-the-art image quality while improving task-specific accuracy and effectively mitigating reward hacking.

Conclusion: VGPO successfully addresses the limitations of GRPO for flow matching-based image generation by providing better temporal credit assignment and stable optimization signals, leading to superior image generation performance.

Abstract: Group Relative Policy Optimization (GRPO) has proven highly effective in enhancing the alignment capabilities of Large Language Models (LLMs). However, current adaptations of GRPO for the flow matching-based image generation neglect a foundational conflict between its core principles and the distinct dynamics of the visual synthesis process. This mismatch leads to two key limitations: (i) Uniformly applying a sparse terminal reward across all timesteps impairs temporal credit assignment, ignoring the differing criticality of generation phases from early structure formation to late-stage tuning. (ii) Exclusive reliance on relative, intra-group rewards causes the optimization signal to fade as training converges, leading to the optimization stagnation when reward diversity is entirely depleted. To address these limitations, we propose Value-Anchored Group Policy Optimization (VGPO), a framework that redefines value estimation across both temporal and group dimensions. Specifically, VGPO transforms the sparse terminal reward into dense, process-aware value estimates, enabling precise credit assignment by modeling the expected cumulative reward at each generative stage. Furthermore, VGPO replaces standard group normalization with a novel process enhanced by absolute values to maintain a stable optimization signal even as reward diversity declines. Extensive experiments on three benchmarks demonstrate that VGPO achieves state-of-the-art image quality while simultaneously improving task-specific accuracy, effectively mitigating reward hacking. Project webpage: https://yawen-shao.github.io/VGPO/.

Method: Hybrid architecture that unifies geometric deep learning with spectral analysis. Learns a diffeomorphic coordinate transformation that maps complex geometries to a latent harmonic space, separating geometry learning from physics learning. Solves for optimal spectral weights in closed form.

Result: Outperforms state-of-the-art implicit representations on advection-diffusion systems. Demonstrates 100x improvement over spectral baselines. Successfully handles sparse data regimes where traditional methods struggle.

Conclusion: DeepVekua provides an effective approach for solving PDEs in challenging sparse data scenarios by decoupling geometry and physics learning through diffeomorphic transformations to harmonic spaces, overcoming spectral bias limitations of conventional neural networks.

Abstract: We present DeepVekua, a hybrid architecture that unifies geometric deep learning with spectral analysis to solve partial differential equations (PDEs) in sparse data regimes. By learning a diffeomorphic coordinate transformation that maps complex geometries to a latent harmonic space, our method outperforms state-of-the-art implicit representations on advection-diffusion systems. Unlike standard coordinate-based networks which struggle with spectral bias, DeepVekua separates the learning of geometry from the learning of physics, solving for optimal spectral weights in closed form. We demonstrate a 100x improvement over spectral baselines. The code is available at https://github.com/VladimerKhasia/vekuanet.

Method: BOLERO is a lightweight, static bipartite graph head that augments RoBERTa-Tab. Each instance connects to feature/value anchors, and a small GNN refines row representations while keeping the backbone frozen.

Result: Evaluated on 80 classification and 64 regression datasets from TP-BERTa benchmark suites, BOLERO achieves the highest number of statistically significant wins across both tasks compared to XGBoost, CatBoost, TabPFN-v2, MITRA, TabICL, TP-BERTa, and RoBERTa-Tab.

Conclusion: Lightweight graph priors meaningfully improve pretrained tabular transformers, demonstrating that explicit modeling of instance relationships enhances tabular learning performance.

Abstract: Tabular data are central to many real-world systems. While recent tabular transformers and in-context learners such as SAINT, TP-BERTa, TabPFN, TabICL, and MITRA incorporate limited inter-row reasoning, most approaches still lack an explicit mechanism to model relationships among instances, even though similar samples often share related outcomes. We investigate whether introducing \emph{simple graph priors} can enhance \emph{pretrained tabular transformers}. Concretely, we introduce {BOLERO}, a lightweight, static bipartite graph head that augments {RoBERTa-Tab} (a RoBERTa-style tabular backbone pretrained with masked-token prediction.) Each instance connects to feature/value anchors; a small GNN refines row representations, while the backbone remains frozen. We evaluate on 80 classification and 64 regression datasets from the TP-BERTa benchmark suites, comparing against strong baselines including XGBoost, CatBoost, TabPFN-v2, MITRA, TabICL, TP-BERTa, and RoBERTa-Tab. To ensure statistically sound conclusions, we follow best practices for multi-dataset evaluation: pairwise Wilcoxon signed-rank tests on per-dataset score differences and effect sizes (median improvement with confidence intervals), rather than mean-rank post-hoc tests that depend on the competitor pool. BOLERO achieves the highest number of statistically significant wins across both classification and regression, demonstrating that lightweight graph priors meaningfully improve pretrained tabular transformers.

Method: Integrated six memristor models into EBANA framework, evaluated equilibrium propagation training with nonlinear memristor-driven weight updates on two benchmark classification tasks.

Result: EqProp achieves robust convergence under nonlinear weight updates when memristors exhibit sufficiently wide resistance range (at least an order of magnitude).

Conclusion: Memristor-based in-memory computing with equilibrium propagation is viable for neural network training when using memristors with adequate resistance range, enabling energy-efficient alternatives to conventional architectures.

Abstract: Memristor-based in-memory computing has emerged as a promising paradigm to overcome the constraints of the von Neumann bottleneck and the memory wall by enabling fully parallelisable and energy-efficient vector-matrix multiplications. We investigate the effect of nonlinear, memristor-driven weight updates on the convergence behaviour of neural networks trained with equilibrium propagation (EqProp). Six memristor models were characterised by their voltage-current hysteresis and integrated into the EBANA framework for evaluation on two benchmark classification tasks. EqProp can achieve robust convergence under nonlinear weight updates, provided that memristors exhibit a sufficiently wide resistance range of at least an order of magnitude.

Method: The paper proposes an enhancement to existing explanation methodology by incorporating principles from rough set theory to address problems with ambiguous documents and stochastic clustering results.

Result: The proposed enhancement improves explainability of Graph Spectral Clustering for text documents by better handling documents without clear content meaning and the stochastic nature of clustering algorithms.

Conclusion: Rough set theory provides a promising approach to enhance the explainability of Graph Spectral Clustering for text documents, overcoming limitations of previous explanation methods.

Abstract: Graph Spectral Clustering methods (GSC) allow representing clusters of diverse shapes, densities, etc. However, the results of such algorithms, when applied e.g. to text documents, are hard to explain to the user, especially due to embedding in the spectral space which has no obvious relation to document contents. Furthermore, the presence of documents without clear content meaning and the stochastic nature of the clustering algorithms deteriorate explainability. This paper proposes an enhancement to the explanation methodology, proposed in an earlier research of our team. It allows us to overcome the latter problems by taking inspiration from rough set theory.

Method: Proposes a knowledge-guided ViT-based Masked Autoencoder that embeds the Linear Spectral Mixing Model (LSMM) as a physical constraint and uses physically-based Spectral Angle Mapper (SAM) loss. The framework jointly optimizes LSMM and SAM loss with conventional Huber loss to ensure both numerical accuracy and geometric consistency in feature space.

Result: The model substantially enhances reconstruction quality and improves downstream task performance, demonstrating better reconstruction fidelity, stabilized training under limited supervision, and interpretable latent representations grounded in physical principles.

Conclusion: The work highlights the promise of embedding physics-informed inductive biases within transformer-based self-supervised learning, showing that integrating domain knowledge can enhance model performance and interpretability.

Abstract: Integrating domain knowledge into deep learning has emerged as a promising direction for improving model interpretability, generalization, and data efficiency. In this work, we present a novel knowledge-guided ViT-based Masked Autoencoder that embeds scientific domain knowledge within the self-supervised reconstruction process. Instead of relying solely on data-driven optimization, our proposed approach incorporates the Linear Spectral Mixing Model (LSMM) as a physical constraint and physically-based Spectral Angle Mapper (SAM), ensuring that learned representations adhere to known structural relationships between observed signals and their latent components. The framework jointly optimizes LSMM and SAM loss with a conventional Huber loss objective, promoting both numerical accuracy and geometric consistency in the feature space. This knowledge-guided design enhances reconstruction fidelity, stabilizes training under limited supervision, and yields interpretable latent representations grounded in physical principles. The experimental findings indicate that the proposed model substantially enhances reconstruction quality and improves downstream task performance, highlighting the promise of embedding physics-informed inductive biases within transformer-based self-supervised learning.

Method: Combines overprovisioned KAN architectures with differentiable sparsification, turning architecture search into an end-to-end optimization problem.

Result: Achieves competitive or superior accuracy across function approximation, dynamical systems forecasting, and real-world tasks while discovering substantially smaller models. Overprovisioning and sparsification are synergistic.

Conclusion: Provides a principled path toward models that are both more expressive and more interpretable, addressing a key tension in scientific machine learning.

Abstract: Efforts to improve Kolmogorov-Arnold networks (KANs) with architectural enhancements have been stymied by the complexity those enhancements bring, undermining the interpretability that makes KANs attractive in the first place. Here we study overprovisioned architectures combined with sparsification to learn compact, interpretable KANs without sacrificing accuracy. Crucially, we focus on differentiable sparsification, turning architecture search into an end-to-end optimization problem. Across function approximation benchmarks, dynamical systems forecasting, and real-world prediction tasks, we demonstrate competitive or superior accuracy while discovering substantially smaller models. Overprovisioning and sparsification are synergistic, with the combination outperforming either alone. The result is a principled path toward models that are both more expressive and more interpretable, addressing a key tension in scientific machine learning.

Method: Introduce cross-modal knowledge distillation: 1) Train teacher spike model across multiple sessions using masked autoencoding with session-specific neural tokenization, 2) Align latent representations of student LFP model to teacher spike model’s representations.

Result: Distilled LFP models consistently outperform single- and multi-session LFP baselines in both unsupervised and supervised settings, and can generalize to other sessions without additional distillation while maintaining superior performance.

Conclusion: Cross-modal knowledge distillation is a powerful and scalable approach for leveraging high-performing spike models to develop more accurate LFP models, addressing the inherent challenges of LFP signal modeling.

Abstract: Local field potentials (LFPs) can be routinely recorded alongside spiking activity in intracortical neural experiments, measure a larger complementary spatiotemporal scale of brain activity for scientific inquiry, and can offer practical advantages over spikes, including greater long-term stability, robustness to electrode degradation, and lower power requirements. Despite these advantages, recent neural modeling frameworks have largely focused on spiking activity since LFP signals pose inherent modeling challenges due to their aggregate, population-level nature, often leading to lower predictive power for downstream task variables such as motor behavior. To address this challenge, we introduce a cross-modal knowledge distillation framework that transfers high-fidelity representational knowledge from pretrained multi-session spike transformer models to LFP transformer models. Specifically, we first train a teacher spike model across multiple recording sessions using a masked autoencoding objective with a session-specific neural tokenization strategy. We then align the latent representations of the student LFP model to those of the teacher spike model. Our results show that the Distilled LFP models consistently outperform single- and multi-session LFP baselines in both fully unsupervised and supervised settings, and can generalize to other sessions without additional distillation while maintaining superior performance. These findings demonstrate that cross-modal knowledge distillation is a powerful and scalable approach for leveraging high-performing spike models to develop more accurate LFP models.

Method: A three-component framework: 1) multiscale encoder for nonlinear aggregation and handling different timescales/missing samples, 2) multiscale dynamical backbone for extracting multimodal temporal dynamics, and 3) modality-specific decoders to account for different probabilistic distributions.

Result: The model successfully aggregates information across modalities with different timescales, distributions, and missing samples to improve real-time target decoding in both simulations and three distinct multiscale brain datasets, outperforming various linear and nonlinear multimodal benchmarks.

Conclusion: The proposed framework enables effective real-time recursive decoding while addressing key challenges of multimodal neural data integration, representing an advancement over existing methods for neural decoding applications.

Abstract: Real-time decoding of target variables from multiple simultaneously recorded neural time-series modalities, such as discrete spiking activity and continuous field potentials, is important across various neuroscience applications. However, a major challenge for doing so is that different neural modalities can have different timescales (i.e., sampling rates) and different probabilistic distributions, or can even be missing at some time-steps. Existing nonlinear models of multimodal neural activity do not address different timescales or missing samples across modalities. Further, some of these models do not allow for real-time decoding. Here, we develop a learning framework that can enable real-time recursive decoding while nonlinearly aggregating information across multiple modalities with different timescales and distributions and with missing samples. This framework consists of 1) a multiscale encoder that nonlinearly aggregates information after learning within-modality dynamics to handle different timescales and missing samples in real time, 2) a multiscale dynamical backbone that extracts multimodal temporal dynamics and enables real-time recursive decoding, and 3) modality-specific decoders to account for different probabilistic distributions across modalities. In both simulations and three distinct multiscale brain datasets, we show that our model can aggregate information across modalities with different timescales and distributions and missing samples to improve real-time target decoding. Further, our method outperforms various linear and nonlinear multimodal benchmarks in doing so.

Method: Conducted large-scale experiments with 56 different 1.7B text-to-image models (549 unique evaluations) to evaluate head module architectures (MLP vs Transformer), training configurations (time weighting), and stochastic TM samplers (high vs low frequency sampling).

Result: TM with MLP head, trained with specific time weighting and sampled with high frequency sampler achieved best overall ranking across all metrics (state-of-the-art). Transformer head with sequence scaling and low frequency sampling excelled at image aesthetics as runner-up.

Conclusion: The study identifies optimal design choices for TM frameworks: MLP heads with specific training configurations provide best overall performance, while Transformer heads excel at aesthetics. The experiments highlight which design aspects yield quality/efficiency gains and which choices are unlikely to provide further improvements.

Abstract: Transition Matching (TM) is an emerging paradigm for generative modeling that generalizes diffusion and flow-matching models as well as continuous-state autoregressive models. TM, similar to previous paradigms, gradually transforms noise samples to data samples, however it uses a second ``internal’’ generative model to implement the transition steps, making the transitions more expressive compared to diffusion and flow models. To make this paradigm tractable, TM employs a large backbone network and a smaller “head” module to efficiently execute the generative transition step. In this work, we present a large-scale, systematic investigation into the design, training and sampling of the head in TM frameworks, focusing on its time-continuous bidirectional variant. Through comprehensive ablations and experimentation involving training 56 different 1.7B text-to-image models (resulting in 549 unique evaluations) we evaluate the affect of the head module architecture and modeling during training as-well as a useful family of stochastic TM samplers. We analyze the impact on generation quality, training, and inference efficiency. We find that TM with an MLP head, trained with a particular time weighting and sampled with high frequency sampler provides best ranking across all metrics reaching state-of-the-art among all tested baselines, while Transformer head with sequence scaling and low frequency sampling is a runner up excelling at image aesthetics. Lastly, we believe the experiments presented highlight the design aspects that are likely to provide most quality and efficiency gains, while at the same time indicate what design choices are not likely to provide further gains.

Method: Combines activation normalization, separation regularizer, and anchor/subspace regularizers that attract rare labeled examples to predefined directions or axis-aligned subspaces, using labels for <0.1% of examples per anchored concept.

Result: Experiments show selective attenuation of targeted concepts with negligible impact on orthogonal features, and complete elimination with reconstruction error approaching theoretical bounds.

Conclusion: Sparse Concept Anchoring provides a practical pathway to interpretable, steerable behavior in learned representations through reversible steering and permanent removal capabilities.

Abstract: We introduce Sparse Concept Anchoring, a method that biases latent space to position a targeted subset of concepts while allowing others to self-organize, using only minimal supervision (labels for <0.1% of examples per anchored concept). Training combines activation normalization, a separation regularizer, and anchor or subspace regularizers that attract rare labeled examples to predefined directions or axis-aligned subspaces. The anchored geometry enables two practical interventions: reversible behavioral steering that projects out a concept’s latent component at inference, and permanent removal via targeted weight ablation of anchored dimensions. Experiments on structured autoencoders show selective attenuation of targeted concepts with negligible impact on orthogonal features, and complete elimination with reconstruction error approaching theoretical bounds. Sparse Concept Anchoring therefore provides a practical pathway to interpretable, steerable behavior in learned representations.

Method: GoMS constructs a graph where nodes represent molecular subgraphs and edges capture their structural relationships, preserving topological information about how substructures are connected and overlap within the molecule, unlike ESAN’s bag-based representation.

Result: GoMS outperforms ESAN and other baselines on public molecular datasets, with particularly strong improvements for large molecules (>100 atoms). The performance gap widens with molecular size, and theoretical analysis shows GoMS can distinguish molecules with identical subgraph compositions but different spatial arrangements.

Conclusion: By capturing substructure relationships lost in bag-based approaches, GoMS represents a significant advance toward scalable and interpretable molecular property prediction, showing particular promise for materials science applications involving complex molecules with multiple functional units.

Abstract: While graph neural networks have shown remarkable success in molecular property prediction, current approaches like the Equivariant Subgraph Aggregation Networks (ESAN) treat molecules as bags of independent substructures, overlooking crucial relationships between these components. We present Graph of Molecule Substructures (GoMS), a novel architecture that explicitly models the interactions and spatial arrangements between molecular substructures. Unlike ESAN’s bag-based representation, GoMS constructs a graph where nodes represent subgraphs and edges capture their structural relationships, preserving critical topological information about how substructures are connected and overlap within the molecule. Through extensive experiments on public molecular datasets, we demonstrate that GoMS outperforms ESAN and other baseline methods, with particularly improvements for large molecules containing more than 100 atoms. The performance gap widens as molecular size increases, demonstrating GoMS’s effectiveness for modeling industrial-scale molecules. Our theoretical analysis demonstrates that GoMS can distinguish molecules with identical subgraph compositions but different spatial arrangements. Our approach shows particular promise for materials science applications involving complex molecules where properties emerge from the interplay between multiple functional units. By capturing substructure relationships that are lost in bag-based approaches, GoMS represents a significant advance toward scalable and interpretable molecular property prediction for real-world applications.

Method: Extended temporal prediction analysis to Week 20 (50% of course duration), comparing Decision Tree and LSTM models across six temporal snapshots. Analyzed feature importance and model performance at different intervention stages.

Result: Different performance metrics matter at different stages: high recall critical early (Weeks 2-4), balanced precision-recall important mid-course (Weeks 8-16), high precision paramount later (Week 20). Static demographic features dominate predictions (68% importance). LSTM achieves 97% recall at Week 2, Decision Tree provides stable balanced performance (78% accuracy) mid-course.

Conclusion: Model selection should depend on intervention timing, and early signals (Weeks 2-4) are sufficient for reliable initial prediction using primarily demographic and pre-enrollment information. Different prediction strategies are optimal for different intervention stages in online learning.

Abstract: Early identification of at-risk students is critical for effective intervention in online learning environments. This study extends temporal prediction analysis to Week 20 (50% of course duration), comparing Decision Tree and Long Short- Term Memory (LSTM) models across six temporal snapshots. Our analysis reveals that different performance metrics matter at different intervention stages: high recall is critical for early intervention (Weeks 2-4), while balanced precision-recall is important for mid-course resource allocation (Weeks 8-16), and high precision becomes paramount in later stages (Week 20). We demonstrate that static demographic features dominate predictions (68% importance), enabling assessment-free early prediction. The LSTM model achieves 97% recall at Week 2, making it ideal for early intervention, while Decision Tree provides stable balanced performance (78% accuracy) during mid-course. By Week 20, both models converge to similar recall (68%), but LSTM achieves higher precision (90% vs 86%). Our findings also suggest that model selection should depend on intervention timing, and that early signals (Weeks 2-4) are sufficient for reliable initial prediction using primarily demographic and pre-enrollment information.

Method: Developed a simulator using historical UNOS data to evaluate policies, compared status quo to myopic matching, introduced “potentials” for dynamic allocation, and explored batching donors and factors like geography and offer rejection.

Result: Status quo policy is considerably inferior to myopic matching; improved policies using potentials enhance performance; batching donors further improves outcomes; simulator reveals impact of geographic proximity and transplant center rejection tendencies.

Conclusion: The proposed dynamic allocation framework with potentials and batching offers superior performance over current policy, addressing critical deficiencies in heart transplant allocation while considering real-world factors like geography and center behavior.

Abstract: Heart transplantation is a viable path for patients suffering from advanced heart failure, but this lifesaving option is severely limited due to donor shortage. Although the current allocation policy was recently revised in 2018, a major concern is that it does not adequately take into account pretransplant and post-transplant mortality. In this paper, we take an important step toward addressing these deficiencies. To begin with, we use historical data from UNOS to develop a new simulator that enables us to evaluate and compare the performance of different policies. We then leverage our simulator to demonstrate that the status quo policy is considerably inferior to the myopic policy that matches incoming donors to the patient who maximizes the number of years gained by the transplant. Moreover, we develop improved policies that account for the dynamic nature of the allocation process through the use of potentials – a measure of a patient’s utility in future allocations that we introduce. We also show that batching together even a handful of donors – which is a viable option for a certain type of donors – further enhances performance. Our simulator also allows us to evaluate the effect of critical, and often unexplored, factors in the allocation, such as geographic proximity and the tendency to reject offers by the transplant centers.

Method: A neural network architecture constructed using singular value decomposition technique with a strictly semi-orthogonality-constrained training algorithm to enhance traditional contrastive learning performance.

Result: The proposed method matches traditional contrastive learning performance in identifying critical transitions while being considerably more lightweight and markedly more resistant to noise.

Conclusion: The SVD-based architecture with semi-orthogonality constraints provides an effective solution for robust critical transition detection in noisy time-series data, addressing limitations of standard contrastive learning approaches.

Abstract: Detecting critical transitions in complex, noisy time-series data is a fundamental challenge across science and engineering. Such transitions may be anticipated by the emergence of a low-dimensional order parameter, whose signature is often masked by high-amplitude stochastic variability. Standard contrastive learning approaches based on deep neural networks, while promising for detecting critical transitions, are often overparameterized and sensitive to irrelevant noise, leading to inaccurate identification of critical points. To address these limitations, we propose a neural network architecture, constructed using singular value decomposition technique, together with a strictly semi-orthogonality-constrained training algorithm, to enhance the performance of traditional contrastive learning. Extensive experiments demonstrate that the proposed method matches the performance of traditional contrastive learning techniques in identifying critical transitions, yet is considerably more lightweight and markedly more resistant to noise.

Method: Apply CEEMDAN to MSCI World index to extract 9 IMFs, transform each IMF into graphs using four methods: natural visibility, horizontal visibility, recurrence, and transition graphs.

Result: High-frequency IMFs produce dense, highly connected small-world graphs; low-frequency IMFs yield sparser networks with longer path lengths. Visibility methods are more amplitude-sensitive with higher clustering, recurrence graphs better preserve temporal dependencies.

Conclusion: Results provide guidance for designing GNN architectures tailored to structural properties of decomposed components, supporting more effective predictive modeling of financial time series.

Abstract: This study applies Empirical Mode Decomposition (EMD) to the MSCI World index and converts the resulting intrinsic mode functions (IMFs) into graph representations to enable modeling with graph neural networks (GNNs). Using CEEMDAN, we extract nine IMFs spanning high-frequency fluctuations to long-term trends. Each IMF is transformed into a graph using four time-series-to-graph methods: natural visibility, horizontal visibility, recurrence, and transition graphs. Topological analysis shows clear scale-dependent structure: high-frequency IMFs yield dense, highly connected small-world graphs, whereas low-frequency IMFs produce sparser networks with longer characteristic path lengths. Visibility-based methods are more sensitive to amplitude variability and typically generate higher clustering, while recurrence graphs better preserve temporal dependencies. These results provide guidance for designing GNN architectures tailored to the structural properties of decomposed components, supporting more effective predictive modeling of financial time series.

Method: The method represents each neuron as a node in a graph, with edges encoding similarity between neurons. Neurons are pruned based on importance scores computed using eigenvector centrality (a graph theory measure of node influence). The resulting pruned network retains only the most central neurons, which are then fine-tuned.

Result: The approach was evaluated on VGGNet, EfficientNet, and ResNet models using TF Flowers, Caltech 101, and Oxford Flowers 102 datasets. It achieved higher classification accuracy while significantly reducing model complexity. On Oxford Flowers 102, the method achieved 48% accuracy compared to 30% accuracy with baseline VGGNet.

Conclusion: Graph theory-based pruning using eigenvector centrality effectively identifies influential neurons, enabling more efficient fine-tuning with improved performance and reduced computational complexity across multiple network architectures and datasets.

Abstract: In social media networks a small number of highly influential users can drive large scale changes in discourse across multiple communities. Small shifts in the behavior of these users are often sufficient to propagate widely throughout the network. A similar phenomenon occurs during neural network fine tuning. Conventional fine tuning of convolutional neural networks typically adds a new linear classification layer on top of a large pre trained model. Instead we argue that improved adaptation can be achieved by first pruning the network to retain only the most important neurons and then performing fine tuning. We propose a graph theory based method for pruning neural networks that is designed to improve fine tuning performance. In this method each neuron is represented as a node and edges encode similarity between neurons. Neurons are pruned based on importance scores computed using eigenvector centrality. The resulting pruned network is then fine tuned using only the most central neurons. We evaluate the proposed method on VGGNet EfficientNet and ResNet models using the TF Flowers Caltech one zero one and Oxford Flowers one zero two datasets. The proposed approach achieves higher classification accuracy while significantly reducing model complexity. On the Oxford Flowers one zero two dataset the method achieves forty eight percent classification accuracy compared to thirty percent accuracy obtained by the baseline VGGNet model.

Method: Uses a diffusion model with multi-sphere predictors encoded into a latent space, incorporating a novel coupling module based on optimal transport to optimize interactions between atmospheric and multi-sphere boundary conditions.

Result: Outperforms ECMWF S2S and FuXi-S2S in 45-day daily-mean ensemble forecasts at 1.5° resolution, achieves skillful subseasonal prediction of extreme events (heat waves, anomalous precipitation), and can generate stable forecasts up to 180 days.

Conclusion: TianXing-S2S establishes a robust framework for S2S research in a warming world, with soil moisture identified as a critical precursor signal for extreme event prediction.

Abstract: Accurate subseasonal-to-seasonal (S2S) prediction of extreme events is critical for resource planning and disaster mitigation under accelerating climate change. However, such predictions remain challenging due to complex multi-sphere interactions and intrinsic atmospheric uncertainty. Here we present TianXing-S2S, a multi-sphere coupled probabilistic model for global S2S daily ensemble forecast. TianXing-S2S first encodes diverse multi-sphere predictors into a compact latent space, then employs a diffusion model to generate daily ensemble forecasts. A novel coupling module based on optimal transport (OT) is incorporated in the denoiser to optimize the interactions between atmospheric and multi-sphere boundary conditions. Across key atmospheric variables, TianXing-S2S outperforms both the European Centre for Medium-Range Weather Forecasts (ECMWF) S2S system and FuXi-S2S in 45-day daily-mean ensemble forecasts at 1.5 resolution. Our model achieves skillful subseasonal prediction of extreme events including heat waves and anomalous precipitation, identifying soil moisture as a critical precursor signal. Furthermore, we demonstrate that TianXing-S2S can generate stable rollout forecasts up to 180 days, establishing a robust framework for S2S research in a warming world.

Method: Construct a simplified Transformer that interprets prompts as programs, decomposing mechanisms: attention performs selective routing from prompt memory, FFN performs local arithmetic, and depth-wise stacking composes local updates into multi-step computation.

Result: Proves a constructive existential result showing that a single fixed backbone can approximate a broad class of target behaviors via prompts alone, providing theoretical foundation for prompt-based switching.

Conclusion: The framework enables formal analysis of trade-offs under prompt length/precision constraints and structural limits of prompt-based switching, distinct from empirical claims about pretrained LLMs.

Abstract: Prompts can switch a model’s behavior even when the weights are fixed, yet this phenomenon is rarely treated as a clean theoretical object rather than a heuristic. We study the family of functions obtainable by holding a Transformer backbone fixed as an executor and varying only the prompt. Our core idea is to view the prompt as an externally injected program and to construct a simplified Transformer that interprets it to implement different computations. The construction exposes a mechanism-level decomposition: attention performs selective routing from prompt memory, the FFN performs local arithmetic conditioned on retrieved fragments, and depth-wise stacking composes these local updates into a multi-step computation. Under this viewpoint, we prove a constructive existential result showing that a single fixed backbone can approximate a broad class of target behaviors via prompts alone. The framework provides a unified starting point for formalizing trade-offs under prompt length/precision constraints and for studying structural limits of prompt-based switching, while remaining distinct from empirical claims about pretrained LLMs.

Method: The authors prove a lower bound of Ω(√d) for transductive mistake bounds, which is an exponential improvement over previous lower bounds. They also construct an upper bound by showing there exists a class with Littlestone dimension d that achieves O(√d) transductive mistake bound, demonstrating tightness.

Result: The paper establishes a tight Θ(√d) mistake bound for transductive online learning, compared to the standard online learning bound of d. This shows a quadratic gap between the two settings, meaning advance access to unlabeled data provides substantial benefit in online learning.

Conclusion: Transductive online learning offers significant advantages over standard online learning with a quadratic gap in mistake bounds, unlike the PAC setting where both have similar sample complexities. This resolves a 30-year open problem and highlights the unique benefits of having unlabeled data available in advance for online learning scenarios.

Abstract: We resolve a 30-year-old open problem concerning the power of unlabeled data in online learning by tightly quantifying the gap between transductive and standard online learning. In the standard setting, the optimal mistake bound is characterized by the Littlestone dimension $d$ of the concept class $H$ (Littlestone 1987). We prove that in the transductive setting, the mistake bound is at least $Ω(\sqrt{d})$. This constitutes an exponential improvement over previous lower bounds of $Ω(\log\log d)$, $Ω(\sqrt{\log d})$, and $Ω(\log d)$, due respectively to Ben-David, Kushilevitz, and Mansour (1995, 1997) and Hanneke, Moran, and Shafer (2023). We also show that this lower bound is tight: for every $d$, there exists a class of Littlestone dimension $d$ with transductive mistake bound $O(\sqrt{d})$. Our upper bound also improves upon the best known upper bound of $(2/3)d$ from Ben-David, Kushilevitz, and Mansour (1997). These results establish a quadratic gap between transductive and standard online learning, thereby highlighting the benefit of advance access to the unlabeled instance sequence. This contrasts with the PAC setting, where transductive and standard learning exhibit similar sample complexities.

Method: Conducted systematic and controlled comparison of SFT and RL on VLM reasoning using identical data sources, examining effects of model capacity, data scale, and data distribution.

Result: SFT outperforms RL in three key scenarios: (1) with weaker/smaller models, (2) with limited data (2K SFT matches 20K RL), (3) for cross-modal generalization. Also identified deceptive rewards problem in RL where higher rewards don’t correlate with accuracy.

Conclusion: The “RL over SFT” narrative is flawed; SFT plays crucial complementary role to RL in VLM reasoning, suggesting balanced post-training pipelines should incorporate both approaches.

Abstract: Recent advances in vision-language models (VLMs) reasoning have been largely attributed to the rise of reinforcement Learning (RL), which has shifted the community’s focus away from the supervised fine-tuning (SFT) paradigm. Many studies suggest that introducing the SFT stage not only fails to improve reasoning ability but may also negatively impact model training. In this study, we revisit this RL-centric belief through a systematic and controlled comparison of SFT and RL on VLM Reasoning. Using identical data sources, we find that the relative effectiveness of SFT and RL is conditional and strongly influenced by model capacity, data scale, and data distribution. Contrary to common assumptions, our findings show that SFT plays a crucial role across several scenarios: (1) Effectiveness for weaker models. SFT more reliably elicits reasoning capabilities in smaller or weaker VLMs. (2) Data efficiency. SFT with only 2K achieves comparable or better reasoning performance to RL with 20K. (3) Cross-modal transferability. SFT demonstrates stronger generalization across modalities. Moreover, we identify a pervasive issue of deceptive rewards, where higher rewards fail to correlate with better reasoning accuracy in RL. These results challenge the prevailing “RL over SFT” narrative. They highlight that the role of SFT may have been underestimated and support a more balanced post-training pipeline in which SFT and RL function as complementary components.

Method: Introduces new mathematical analysis of Newton Step (NS) and Influence Functions (IF) data attribution methods for convex learning problems. The analysis avoids global strong convexity assumptions and provides error bounds for average-case sample removals, proving asymptotic tightness for well-behaved logistic regression.

Result: Proves asymptotically tight error bounds (up to poly-log factors) showing: 1) NS method error scales as Θ̃(kd/n²), and 2) difference between NS and IF methods scales as Θ̃((k+d)√(kd)/n²). This explains empirical observations that NS is often more accurate than IF.

Conclusion: Provides the first analysis of data attribution methods without global strong convexity assumptions, offering tight error scaling laws that explain why Newton Step attribution is generally more accurate than Influence Functions, addressing fundamental questions in data attribution theory.

Abstract: Data attribution aims to explain model predictions by estimating how they would change if certain training points were removed, and is used in a wide range of applications, from interpretability and credit assignment to unlearning and privacy. Even in the relatively simple case of linear regressions, existing mathematical analyses of leading data attribution methods such as Influence Functions (IF) and single Newton Step (NS) remain limited in two key ways. First, they rely on global strong convexity assumptions which are often not satisfied in practice. Second, the resulting bounds scale very poorly with the number of parameters ($d$) and the number of samples removed ($k$). As a result, these analyses are not tight enough to answer fundamental questions such as “what is the asymptotic scaling of the errors of each method?” or “which of these methods is more accurate for a given dataset?” In this paper, we introduce a new analysis of the NS and IF data attribution methods for convex learning problems. To the best of our knowledge, this is the first analysis of these questions that does not assume global strong convexity and also the first explanation of [KATL19] and [RH25a]’s observation that NS data attribution is often more accurate than IF. We prove that for sufficiently well-behaved logistic regression, our bounds are asymptotically tight up to poly-logarithmic factors, yielding scaling laws for the errors in the average-case sample removals. [ \mathbb{E}_{T \subseteq [n],, |T| = k} \bigl[ |\hatθ_T - \hatθ_T^{\mathrm{NS}}|2 \bigr] = \widetildeΘ!\left(\frac{k d}{n^2}\right), \qquad \mathbb{E}{T \subseteq [n],, |T| = k} \bigl[ |\hatθ_T^{\mathrm{NS}} - \hatθ_T^{\mathrm{IF}}|_2 \bigr] = \widetildeΘ!\left( \frac{(k + d)\sqrt{k d}}{n^2} \right). ]

Method: Augmented ACGAN generator objective with VQE-inspired energy term computed from class-specific Ising Hamiltonians using Qiskit’s EstimatorQNN and TorchConnector. Experiments conducted on MNIST with 4-qubit noiseless statevector simulator.

Result: Energy-regularized model (QACGAN) achieved 99-100% classification accuracy within 5 epochs vs 87.8% for ACGAN, suggesting improved class conditioning. Sample quality (FID) showed high variance (CV ~25%) initially but stabilized around 23-24 FID, comparable to baseline.

Conclusion: Demonstrated methodological integration of VQE-style energy computations into GAN training loops via differentiable pathways, but whether such signals provide benefits beyond classical regularizers remains an open question requiring systematic ablation studies.

Abstract: This paper presents an exploratory, simulator-based proof of concept investigating whether differentiable energy terms derived from parameterized quantum circuits can serve as auxiliary regularization signals in Generative Adversarial Networks (GANs). We augment the Auxiliary Classifier GAN (ACGAN) generator objective with a Variational Quantum Eigensolver (VQE)-inspired energy term computed from class-specific Ising Hamiltonians using Qiskit’s EstimatorQNN and TorchConnector. Important limitations: All experiments run on a noiseless statevector simulator with only 4 qubits, use a deliberately simple Hamiltonian parameterization, and lack ablation studies comparing against equivalent classical biases. The computational overhead (approximately 200x slower than classical ACGAN) reflects simulator artifacts rather than inherent quantum costs. On MNIST, we observe that the energy-regularized model (termed QACGAN) achieves high classification accuracy (99 to 100 percent) within 5 epochs compared to 87.8 percent for ACGAN, suggesting the auxiliary term influences class conditioning. However, sample quality metrics (FID) show high variance across runs (coefficient of variation approximately 25 percent at epoch 5), with values ranging from 19.92 to 35.96. Extended runs stabilize around FID 23 to 24, comparable to the ACGAN baseline. We explicitly do not claim quantum advantage, improved stability in any general sense, or scalability beyond this toy setting. The contribution is methodological: demonstrating that VQE-style energy computations can be integrated into GAN training loops via differentiable pathways. Whether such auxiliary signals provide benefits beyond equivalent classical regularizers remains an open question requiring systematic ablation studies, which we leave for future work.

Method: Formulates online learning update as continuous-time dynamical system, leverages rank-1 structure of dynamics matrix to derive exact closed-form solution (equivalent to infinite-order Runge-Kutta method), achieving full parallelism and linear-time complexity.

Result: EFLA achieves lower language modeling perplexity and superior downstream benchmark performance than DeltaNet without additional parameters, with robust performance in noisy environments.

Conclusion: Provides a new theoretical foundation for building high-fidelity, scalable linear-time attention models with error-free continuous dynamics preservation.

Abstract: Linear-time attention and State Space Models (SSMs) promise to solve the quadratic cost bottleneck in long-context language models employing softmax attention. We introduce Error-Free Linear Attention (EFLA), a numerically stable, fully parallelism and generalized formulation of the delta rule. Specifically, we formulate the online learning update as a continuous-time dynamical system and prove that its exact solution is not only attainable but also computable in linear time with full parallelism. By leveraging the rank-1 structure of the dynamics matrix, we directly derive the exact closed-form solution effectively corresponding to the infinite-order Runge-Kutta method. This attention mechanism is theoretically free from error accumulation, perfectly capturing the continuous dynamics while preserving the linear-time complexity. Through an extensive suite of experiments, we show that EFLA enables robust performance in noisy environments, achieving lower language modeling perplexity and superior downstream benchmark performance than DeltaNet without introducing additional parameters. Our work provides a new theoretical foundation for building high-fidelity, scalable linear-time attention models.

Method: Review existing literature on causal inference and interpretability in e-commerce/retail, then apply these techniques to real-world data using explainable models and double machine learning with multiple confounders.

Result: Explainable models demonstrate lower variance in SHAP values, and including multiple confounders through double machine learning enables obtaining correct signs of causal effects.

Conclusion: Provides a practical recipe for combining interpretability and causal inference in retail analytics, enabling more reliable and actionable sales insights from complex retail data.

Abstract: Most major retailers today have multiple divisions focused on various aspects, such as marketing, supply chain, online customer experience, store customer experience, employee productivity, and vendor fulfillment. They also regularly collect data corresponding to all these aspects as dashboards and weekly/monthly/quarterly reports. Although several machine learning and statistical techniques have been in place to analyze and predict key metrics, such models typically lack interpretability. Moreover, such techniques also do not allow the validation or discovery of causal links. In this paper, we aim to provide a recipe for applying model interpretability and causal inference for deriving sales insights. In this paper, we review the existing literature on causal inference and interpretability in the context of problems in e-commerce and retail, and apply them to a real-world dataset. We find that an inherently explainable model has a lower variance of SHAP values, and show that including multiple confounders through a double machine learning approach allows us to get the correct sign of causal effect.

Method: Uses random-matrix-theoretic signatures: honest Non-IID gradients produce covariance eigenspectra following Marchenko-Pastur law, while Byzantine perturbations cause detectable tail anomalies. Combines Frequent Directions sketching with data-dependent MP tracking for O(k²) memory efficiency (k ≪ d).

Result: Achieves 78.4% average accuracy across 144 attack-aggregator configurations vs 48-63% for baselines. Proves (ε,δ)-Byzantine resilience with convergence rate O(σf/√T + f²/T) and provides matching information-theoretic lower bound Ω(σf/√T), establishing minimax optimality.

Conclusion: Spectral Sentinel provides a scalable, theoretically-grounded solution to Byzantine attacks in DFL with Non-IID data, enabling practical deployment on large models up to 1.5B parameters with blockchain integration and proven optimal convergence guarantees.

Abstract: Decentralized federated learning (DFL) enables collaborative model training without centralized trust, but it remains vulnerable to Byzantine clients that poison gradients under heterogeneous (Non-IID) data. Existing defenses face a scalability trilemma: distance-based filtering (e.g., Krum) can reject legitimate Non-IID updates, geometric-median methods incur prohibitive $O(n^2 d)$ cost, and many certified defenses are evaluated only on models below 100M parameters. We propose Spectral Sentinel, a Byzantine detection and aggregation framework that leverages a random-matrix-theoretic signature: honest Non-IID gradients produce covariance eigenspectra whose bulk follows the Marchenko-Pastur law, while Byzantine perturbations induce detectable tail anomalies. Our algorithm combines Frequent Directions sketching with data-dependent MP tracking, enabling detection on models up to 1.5B parameters using $O(k^2)$ memory with $k \ll d$. Under a $(σ,f)$ threat model with coordinate-wise honest variance bounded by $σ^2$ and $f < 1/2$ adversaries, we prove $(ε,δ)$-Byzantine resilience with convergence rate $O(σf / \sqrt{T} + f^2 / T)$, and we provide a matching information-theoretic lower bound $Ω(σf / \sqrt{T})$, establishing minimax optimality. We implement the full system with blockchain integration on Polygon networks and validate it across 144 attack-aggregator configurations, achieving 78.4 percent average accuracy versus 48-63 percent for baseline methods.

Method: Built on PyTorch Geometric, tgp provides a wide variety of pooling operators under a consistent API with modular design, featuring precomputed pooling for accelerated training.

Result: Extensive benchmarking shows that the choice of optimal pooling operator depends on tasks and data, supporting the need for a library that enables fast prototyping.

Conclusion: tgp is a usable, extensible library that facilitates experimentation with different graph pooling methods, with performance varying by task and dataset.

Abstract: We introduce Torch Geometric Pool (tgp), a library for hierarchical pooling in Graph Neural Networks. Built upon Pytorch Geometric, Torch Geometric Pool (tgp) provides a wide variety of pooling operators, unified under a consistent API and a modular design. The library emphasizes usability and extensibility, and includes features like precomputed pooling, which significantly accelerate training for a class of operators. In this paper, we present tgp’s structure and present an extensive benchmark. The latter showcases the library’s features and systematically compares the performance of the implemented graph-pooling methods in different downstream tasks. The results, showing that the choice of the optimal pooling operator depends on tasks and data at hand, support the need for a library that enables fast prototyping.

Method: Asymmetric projection framework that decomposes backward-pass gradients into parallel spans and orthogonal violations while keeping forward-pass QKV structure intact, with selective scaling focusing on 0th order bidirectional parallel spans.

Result: 0.56% reduction in validation loss on WikiText-2 dataset with crude configuration, providing strong theoretical evidence that standard attention gradient is suboptimal.

Conclusion: The framework demonstrates fundamental validity and suggests significant potential gains on larger datasets and deeper training regimes by optimizing attention gradient decomposition.

Abstract: The canonical $O(N^2)$ Transformer remains the empirical performance frontier in sequence modeling, and its training can be further optimized by addressing geometric inefficiency. We propose an optimization framework that leverages an asymmetric projection to decompose the backward-pass gradients into parallel spans and orthogonal violations, while keeping the canonical forward-pass $QKV$ structure intact. Through consistent experimental validation across various decomposition and projection setups, we provide strong theoretical evidence: the standard attention gradient is suboptimal. We demonstrated that selectively scaling these components, focusing primarily on $0^{th}$ order bidirectional parallel spans, yields the most effective learning signal. On the limited WikiText-2 dataset, and using a crude configuration, this method achieved a $0.56%$ reduction in validation loss, confirming the framework’s fundamental validity and suggesting significant potential gains on larger datasets and deeper training regimes

Method: PerNodeDrop applies per-sample, per-node perturbations (dropping weights at sample level, not batch level) to break uniformity of noise injection. This allows each node to experience input-specific variability while preserving useful co-adaptation. The method includes an expected-loss analysis formalizing how perturbations attenuate excessive co-adaptation while retaining predictive interactions.

Result: Empirical evaluations on vision, text, and audio benchmarks show improved generalization relative to standard noise-based regularizers. The method narrows the gap between training and validation performance and improves reliability on unseen data.

Conclusion: PerNodeDrop effectively addresses the limitations of uniform noise injection in existing regularization methods by applying sample-specific perturbations, preserving beneficial co-adaptation while regularizing harmful patterns, leading to better generalization across diverse domains.

Abstract: Deep neural networks possess strong representational capacity yet remain vulnerable to overfitting, primarily because neurons tend to co-adapt in ways that, while capturing complex and fine-grained feature interactions, also reinforce spurious and non-generalizable patterns that inflate training performance but reduce reliability on unseen data. Noise-based regularizers such as Dropout and DropConnect address this issue by injecting stochastic perturbations during training, but the noise they apply is typically uniform across a layer or across a batch of samples, which can suppress both harmful and beneficial co-adaptation. This work introduces PerNodeDrop, a lightweight stochastic regularization method. It applies per-sample, per-node perturbations to break the uniformity of the noise injected by existing techniques, thereby allowing each node to experience input-specific variability. Hence, PerNodeDrop preserves useful co-adaptation while applying regularization. This narrows the gap between training and validation performance and improves reliability on unseen data, as evident from the experiments. Although superficially similar to DropConnect, PerNodeDrop operates at the sample level. It drops weights at the sample level, not the batch level. An expected-loss analysis formalizes how its perturbations attenuate excessive co-adaptation while retaining predictive interactions. Empirical evaluations on vision, text, and audio benchmarks indicate improved generalization relative to the standard noise-based regularizer.

Method: FinFRE-RAG: A two-stage approach with (1) importance-guided feature reduction to serialize a compact subset of numeric/categorical attributes into natural language, and (2) retrieval-augmented in-context learning over label-aware, instance-level exemplars.

Result: Across four public fraud datasets and three families of open-weight LLMs, FinFRE-RAG substantially improves F1/MCC over direct prompting and is competitive with strong tabular baselines in several settings. LLMs still lag behind specialized classifiers but narrow the performance gap.

Conclusion: FinFRE-RAG demonstrates that LLMs can be valuable assistive tools in fraud analysis by providing interpretable rationales while improving performance over direct prompting, though they still don’t match specialized tabular classifiers.

Abstract: Detecting fraud in financial transactions typically relies on tabular models that demand heavy feature engineering to handle high-dimensional data and offer limited interpretability, making it difficult for humans to understand predictions. Large Language Models (LLMs), in contrast, can produce human-readable explanations and facilitate feature analysis, potentially reducing the manual workload of fraud analysts and informing system refinements. However, they perform poorly when applied directly to tabular fraud detection due to the difficulty of reasoning over many features, the extreme class imbalance, and the absence of contextual information. To bridge this gap, we introduce FinFRE-RAG, a two-stage approach that applies importance-guided feature reduction to serialize a compact subset of numeric/categorical attributes into natural language and performs retrieval-augmented in-context learning over label-aware, instance-level exemplars. Across four public fraud datasets and three families of open-weight LLMs, FinFRE-RAG substantially improves F1/MCC over direct prompting and is competitive with strong tabular baselines in several settings. Although these LLMs still lag behind specialized classifiers, they narrow the performance gap and provide interpretable rationales, highlighting their value as assistive tools in fraud analysis.

Method: DynaGen uses a unified approach: (1) For interpolation: dynamically constructs entity-centric subgraphs and processes them with a synergistic dual-branch GNN encoder to capture evolving structural context. (2) For extrapolation: applies a conditional diffusion process that forces learning of underlying evolutionary principles rather than just superficial patterns.

Result: Extensive experiments on six benchmark datasets show state-of-the-art performance. On average, compared to second-best models, DynaGen improves Mean Reciprocal Rank (MRR) by 2.61 points for interpolation and 1.45 points for extrapolation.

Conclusion: DynaGen successfully addresses the limitations of existing TKGR methods by providing a unified framework that improves both interpolation and extrapolation through better contextual modeling and principled learning of evolutionary patterns.

Abstract: Temporal Knowledge Graph Reasoning (TKGR) aims to complete missing factual elements along the timeline. Depending on the temporal position of the query, the task is categorized into interpolation and extrapolation. Existing interpolation methods typically embed temporal information into individual facts to complete missing historical knowledge, while extrapolation techniques often leverage sequence models over graph snapshots to identify recurring patterns for future event prediction. These methods face two critical challenges: limited contextual modeling in interpolation and cognitive generalization bias in extrapolation. To address these, we propose a unified method for TKGR, dubbed DynaGen. For interpolation, DynaGen dynamically constructs entity-centric subgraphs and processes them with a synergistic dual-branch GNN encoder to capture evolving structural context. For extrapolation, it applies a conditional diffusion process, which forces the model to learn underlying evolutionary principles rather than just superficial patterns, enhancing its ability to predict unseen future events. Extensive experiments on six benchmark datasets show DynaGen achieves state-of-the-art performance. On average, compared to the second-best models, DynaGen improves the Mean Reciprocal Rank (MRR) score by 2.61 points for interpolation and 1.45 points for extrapolation.

Method: Two algorithms: 1) SINDy-FM uses sparse regression to identify interpretable symbolic differential equations from data, and 2) DSBM-NeuralODE reformulates Schrödinger Bridge as Neural-ODE for flexible continuous-time parameterization.

Result: Experiments on Gaussian transport tasks and MNIST latent translation show competitive performance with dramatic efficiency improvements. SINDy-FM reduces parameters by orders of magnitude and enables near-instantaneous inference.

Conclusion: The surrogate approach creates simpler, faster, and more flexible approximations of diffusion/Schrödinger Bridge dynamics, paving the way for tractable, high-performing bridge models suitable for practical deployment.

Abstract: Diffusion and Schrödinger Bridge models have established state-of-the-art performance in generative modeling but are often hampered by significant computational costs and complex training procedures. While continuous-time bridges promise faster sampling, overparameterized neural networks describe their optimal dynamics, and the underlying stochastic differential equations can be difficult to integrate efficiently. This work introduces a novel paradigm that uses surrogate models to create simpler, faster, and more flexible approximations of these dynamics. We propose two specific algorithms: SINDy Flow Matching (SINDy-FM), which leverages sparse regression to identify interpretable, symbolic differential equations from data, and a Neural-ODE reformulation of the Schrödinger Bridge (DSBM-NeuralODE) for flexible continuous-time parameterization. Our experiments on Gaussian transport tasks and MNIST latent translation demonstrate that these surrogates achieve competitive performance while offering dramatic improvements in efficiency and interpretability. The symbolic SINDy-FM models, in particular, reduce parameter counts by several orders of magnitude and enable near-instantaneous inference, paving the way for a new class of tractable and high-performing bridge models for practical deployment.

Method: Integrate multiple MIA techniques into data extraction pipelines and systematically benchmark their effectiveness. Compare performance in this integrated setting against results from conventional MIA benchmarks.

Result: The study provides comparative analysis of how different MIA techniques perform in practical data extraction scenarios versus traditional benchmark settings.

Conclusion: The research evaluates the real-world utility of MIA techniques in training data extraction attacks, helping understand which methods are most effective in practical privacy threat scenarios.

Abstract: Large Language Models (LLMs) are prone to memorizing training data, which poses serious privacy risks. Two of the most prominent concerns are training data extraction and Membership Inference Attacks (MIAs). Prior research has shown that these threats are interconnected: adversaries can extract training data from an LLM by querying the model to generate a large volume of text and subsequently applying MIAs to verify whether a particular data point was included in the training set. In this study, we integrate multiple MIA techniques into the data extraction pipeline to systematically benchmark their effectiveness. We then compare their performance in this integrated setting against results from conventional MIA benchmarks, allowing us to evaluate their practical utility in real-world extraction scenarios.

Method: Integrates observations across all tasks to learn global joint distribution using particle-based approximation of log-density Gaussian process, capturing structural dependencies without prior assumptions on latent variables.

Result: Outperforms hierarchical model bandits, especially in settings with model misspecification or complex latent heterogeneity.

Conclusion: Proposed Bayesian framework effectively handles two key epistemic uncertainties (structural and user-specific) and enables flexible discovery of dependencies in multi-task bandit problems.

Abstract: We propose a novel Bayesian framework for efficient exploration in contextual multi-task multi-armed bandit settings, where the context is only observed partially and dependencies between reward distributions are induced by latent context variables. In order to exploit these structural dependencies, our approach integrates observations across all tasks and learns a global joint distribution, while still allowing personalised inference for new tasks. In this regard, we identify two key sources of epistemic uncertainty, namely structural uncertainty in the latent reward dependencies across arms and tasks, and user-specific uncertainty due to incomplete context and limited interaction history. To put our method into practice, we represent the joint distribution over tasks and rewards using a particle-based approximation of a log-density Gaussian process. This representation enables flexible, data-driven discovery of both inter-arm and inter-task dependencies without prior assumptions on the latent variables. Empirically, we demonstrate that our method outperforms baselines such as hierarchical model bandits, especially in settings with model misspecification or complex latent heterogeneity.

Method: Propose Multi-Trajectory PINN (MT-PINN) with rollout-based trajectory loss and backpropagation-through-time to propagate terminal penalty on inventory, plus lambda-curriculum to stabilize training from risk-neutral to risk-averse HJB equations.

Result: On Gatheral-Schied model: MT-PINN matches closed-form solutions, tightly concentrates terminal inventory around zero, reduces errors along optimal paths. On SPY data: matches TWAP when risk-neutral, achieves lower exposure and competitive costs (especially in falling markets) with higher risk-aversion.

Conclusion: MT-PINN effectively enforces hard constraints in optimal trade execution, outperforming vanilla PINNs and showing practical value in real market data applications with improved constraint satisfaction and performance.

Abstract: We study optimal trade execution with a hard-zero terminal inventory constraint, modeled via Hamilton-Jacobi-Bellman (HJB) equations. Vanilla PINNs often under-enforce this constraint and produce unstable controls. We propose a Multi-Trajectory PINN (MT-PINN) that adds a rollout-based trajectory loss and propagates a terminal penalty on terminal inventory via backpropagation-through-time, directly enforcing zero terminal inventory. A lightweight lambda-curriculum is adopted to stabilize training as the state expands from a risk-neutral reduced HJB to a risk-averse HJB. On the Gatheral-Schied single-asset model, MT-PINN aligns closely with their derived closed-form solutions and concentrates terminal inventory tightly around zero while reducing errors along optimal paths. We apply MT-PINNs on SPY intraday data, matching TWAP when risk-neutral, and achieving lower exposure and competitive costs, especially in falling windows, for higher risk-aversion.

Method: Frames regression as multivariate approximation using random function theory, postulates indifference principles (translation, rotation, scaling invariance and Gaussianity), and analytically derives the entire solution scheme from these symmetries.

Result: The derived kernel coincides with generalized polyharmonic splines, and the complete regression framework (kernel form, regularization type, noise parameterization) follows from the symmetry postulates without empirical tuning.

Conclusion: This provides a theoretical justification for a broad class of smoothing/interpolation methods, demonstrating they are optimal under indifference assumptions when no prior information exists.

Abstract: This paper studies a machine learning regression problem as a multivariate approximation problem using the framework of the theory of random functions. An ab initio derivation of a regression method is proposed, starting from postulates of indifference. It is shown that if a probability measure on an infinite-dimensional function space possesses natural symmetries (invariance under translation, rotation, scaling, and Gaussianity), then the entire solution scheme, including the kernel form, the type of regularization, and the noise parameterization, follows analytically from these postulates. The resulting kernel coincides with a generalized polyharmonic spline; however, unlike existing approaches, it is not chosen empirically but arises as a consequence of the indifference principle. This result provides a theoretical foundation for a broad class of smoothing and interpolation methods, demonstrating their optimality in the absence of a priori information.

Method: SPARK integrates three key techniques: 1) Random projection-based Jacobian compression to reduce communication while preserving spectral properties, 2) Stage-wise annealed distillation that transitions from pure NTK evolution to neighbor-regularized learning to counteract compression noise, and 3) Nesterov momentum acceleration for faster convergence enabled by distillation smoothing.

Result: SPARK achieves 98.7% communication reduction compared to NTK-DFL while maintaining convergence speed and superior accuracy. With momentum acceleration, SPARK reaches target performance 3 times faster, establishing state-of-the-art results for communication-efficient decentralized learning.

Conclusion: SPARK enables practical deployment of decentralized federated learning in bandwidth-limited edge environments by dramatically reducing communication overhead while preserving convergence properties and accuracy.

Abstract: Decentralized federated learning (DFL) faces critical challenges from statistical heterogeneity and communication overhead. While NTK-based methods achieve faster convergence, transmitting full Jacobian matrices is impractical for bandwidth-constrained edge networks. We propose SPARK, synergistically integrating random projection-based Jacobian compression, stage-wise annealed distillation, and Nesterov momentum acceleration. Random projections compress Jacobians while preserving spectral properties essential for convergence. Stage-wise annealed distillation transitions from pure NTK evolution to neighbor-regularized learning, counteracting compression noise. Nesterov momentum accelerates convergence through stable accumulation enabled by distillation smoothing. SPARK achieves 98.7% communication reduction compared to NTK-DFL while maintaining convergence speed and superior accuracy. With momentum, SPARK reaches target performance 3 times faster, establishing state-of-the-art results for communication-efficient decentralized learning and enabling practical deployment in bandwidth-limited edge environments.

Method: Reinterpret input sparsification as dynamic structural pruning, then introduce trainable spontaneous neurons inspired by biological neurons’ baseline firing rates to act as compensatory units.

Result: Auxiliary spontaneous neurons substantially reduce sparsification-induced performance gap while generalizing effectively across tasks.

Conclusion: The proposed approach bridges the performance gap in sparsified LLMs by stabilizing activations through biologically-inspired compensatory neurons.

Abstract: Large Language Models (LLMs) achieve state-of-the-art performance across a wide range of applications, but their massive scale poses significant challenges for both efficiency and interpretability. Structural pruning, which reduces model size by removing redundant computational units such as neurons, has been widely explored as a solution, and this study devotes to input sparsification, an increasingly popular technique that improves efficiency by selectively activating only a subset of entry values for each input. However, existing approaches focus primarily on computational savings, often overlooking the representational consequences of sparsification and leaving a noticeable performance gap compared to full models. In this work, we first reinterpret input sparsification as a form of dynamic structural pruning. Motivated by the spontaneous baseline firing rates observed in biological neurons, we introduce a small set of trainable spontaneous neurons that act as compensatory units to stabilize activations in sparsified LLMs. Experiments demonstrate that these auxiliary neurons substantially reduce the sparsification-induced performance gap while generalizing effectively across tasks.