Method: Proposes AuTAgent (Audio Tool Agent), a reinforcement learning framework that learns when and which tools to invoke. Uses a sparse-feedback training strategy with a novel Differential Reward mechanism to filter out irrelevant tools and invoke external assistance only when it yields net performance gain over the base model.

Result: Improves accuracy by 4.20%/6.20% and 9.80%/8.00% for open-source and closed-source backbones on MMAU Test-mini and MMAR benchmarks. Demonstrates exceptional transferability and shows that AuTAgent complements the representation bottleneck of LALMs by providing verifiable acoustic evidence.

Conclusion: The framework successfully addresses the tool integration challenge in audio language models, highlighting the complementary role of external tools in augmenting audio model reasoning through learned tool invocation strategies.

Abstract: Large Audio Language Models (LALMs) excel at perception but struggle with complex reasoning requiring precise acoustic measurements. While external tools can extract fine-grained features like exact tempo or pitch, effective integration remains challenging: naively using all tools causes information overload, while prompt-based selection fails to assess context-dependent utility. To address this, we propose AuTAgent (Audio Tool Agent), a reinforcement learning framework that learns when and which tools to invoke. By employing a sparse-feedback training strategy with a novel Differential Reward mechanism, the agent learns to filter out irrelevant tools and invokes external assistance only when it yields a net performance gain over the base model. Experimental results confirm that AuTAgent complements the representation bottleneck of LALMs by providing verifiable acoustic evidence. It improves accuracy by 4.20% / 6.20% and 9.80% / 8.00% for open-source and closed-source backbones on the MMAU Test-mini and the MMAR benchmarks, respectively. In addition, further experiments demonstrate exceptional transferability. We highlight the complementary role of external tools in augmenting audio model reasoning.

Method: Evaluate Qwen2-VL-7B and Idefics3-8B with PTQ using data-free (HQQ) and data-aware (MBQ) methods across multiple bit widths. Adapt Selector confidence estimator for quantized multimodal settings and test robustness across quantization levels and OOD scenarios.

Result: PTQ degrades both accuracy and reliability. Data-aware methods soften the effect. Selector substantially mitigates reliability impact. Int4 MBQ + Selector achieves best efficiency-reliability trade-off, closing in on uncompressed performance at ~75% less memory.

Conclusion: First systematic study linking quantization and reliability in multimodal settings. Shows data-aware quantization methods combined with confidence estimators can maintain reliability while achieving significant compression for edge deployment.

Abstract: Multimodal Large Language Models (MLLM) are increasingly deployed in domains where both reliability and efficiency are critical. However, current models remain overconfident, producing highly certain but incorrect answers. At the same time, their large size limits deployment on edge devices, necessitating compression. We study the intersection of these two challenges by analyzing how Post-Training Quantization (PTQ) compression affects both accuracy and reliability in Visual Question Answering (VQA). We evaluate two MLLMs, Qwen2-VL-7B and Idefics3-8B, quantized with data-free (HQQ) and data-aware (MBQ) methods across multiple bit widths. To counteract the reduction in reliability caused by quantization, we adapt the Selector confidence estimator for quantized multimodal settings and test its robustness across various quantization levels and out-of-distribution (OOD) scenarios. We find that PTQ degrades both accuracy and reliability. Data-aware methods soften the effect thereof. The Selector substantially mitigates the reliability impact. The combination of int4 MBQ and the Selector achieves the best efficiency-reliability trade-off, closing in on uncompressed performance at approx. 75% less memory demand. Overall, we present the first systematic study linking quantization and reliability in multimodal settings.

Method: Proposes AudioX framework with a Multimodal Adaptive Fusion module that effectively fuses diverse multimodal inputs (text, video, audio) to enhance cross-modal alignment. Constructs IF-caps dataset with 7M+ samples through structured annotation pipeline for comprehensive supervision.

Result: AudioX achieves superior performance compared to state-of-the-art methods, especially in text-to-audio and text-to-music generation, demonstrating powerful instruction-following potential for multimodal control signals.

Conclusion: AudioX provides an effective unified framework for anything-to-audio generation with multimodal conditions, showing strong performance across various tasks and promising potential for instruction-following audio generation.

Abstract: Audio and music generation based on flexible multimodal control signals is a widely applicable topic, with the following key challenges: 1) a unified multimodal modeling framework, and 2) large-scale, high-quality training data. As such, we propose AudioX, a unified framework for anything-to-audio generation that integrates varied multimodal conditions (i.e., text, video, and audio signals) in this work. The core design in this framework is a Multimodal Adaptive Fusion module, which enables the effective fusion of diverse multimodal inputs, enhancing cross-modal alignment and improving overall generation quality. To train this unified model, we construct a large-scale, high-quality dataset, IF-caps, comprising over 7 million samples curated through a structured data annotation pipeline. This dataset provides comprehensive supervision for multimodal-conditioned audio generation. We benchmark AudioX against state-of-the-art methods across a wide range of tasks, finding that our model achieves superior performance, especially in text-to-audio and text-to-music generation. These results demonstrate our method is capable of audio generation under multimodal control signals, showing powerful instruction-following potential. The code and datasets will be available at https://zeyuet.github.io/AudioX/.

Method: Proposes a multimodal consistency-guided pipeline with: 1) target-aware preselection using submodular mutual information, 2) multiple pseudo-transcriptions via perturbation-based decoding, 3) scoring using speech-text alignment in shared embedding space and predicted WER, and 4) percentile-based selection of reliable pseudo-labels.

Result: In-domain: Selecting ~1.5k utterances from 30k pool achieves 10.91% WER (close to 10.45% with supervised labels). Cross-domain: Consistency-filtered subsets avoid degradation from unfiltered pseudo-labels under accent shift, outperforming random sampling and recent baselines.

Conclusion: Multimodal consistency filtering effectively selects reliable pseudo-labels for ASR accent adaptation, achieving near-supervised performance with much less data and preventing error amplification in cross-domain scenarios.

Abstract: Automatic speech recognition (ASR) systems often degrade on accented speech because acoustic-phonetic and prosodic shifts induce a mismatch to training data, making labeled accent adaptation costly. However, common pseudo-label selection heuristics are largely text-centric (e.g., perplexity (PPL) filtering) and can prefer fluent yet acoustically mismatched hypotheses, leading to error amplification when fine-tuning. To address this, we introduce a multimodal consistency-guided, reference-free data selection pipeline for ASR accent adaptation under a transductive, label-free protocol. The pipeline starts with a target-aware preselection step based on submodular mutual information to improve query relevance and reduce downstream computation. It then generates multiple pseudo-transcriptions per utterance via perturbation-based decoding and scores each hypothesis using two reference-free signals: speech–text alignment in a shared embedding space and predicted word error rate (WER). A simple percentile-based selection rule retains reliable pseudo-labels for fine-tuning while discarding noisy utterances. In an in-domain setting, selecting ~1.5k utterances from a 30k pool achieves 10.91% WER, close to 10.45% obtained using 30k supervised labels. In a cross-domain setting with a mismatched candidate pool, consistency-filtered subsets avoid the degradation caused by unfiltered pseudo-labels under strong accent shift, and matched-hour experiments on a stronger ASR backbone further confirm gains over random sampling and recent selection baselines.

Method: Examined state-of-the-art LLMs and translation systems using multilingual crisis data and a newly introduced urgency-annotated dataset covering over 32 languages, analyzing performance degradation and instability in crisis-domain translation.

Result: Both dedicated translation models and LLMs exhibit substantial performance degradation and instability in crisis contexts. Even linguistically adequate translations can distort perceived urgency, and LLM-based urgency classifications vary widely depending on language of prompt and input.

Conclusion: Significant risks exist in deploying general-purpose language technologies for crisis communication, underscoring the need for crisis-aware evaluation frameworks and specialized approaches for high-stakes multilingual crisis contexts.

Abstract: Large language models (LLMs) are increasingly proposed for crisis preparedness and response, particularly for multilingual communication. However, their suitability for high-stakes crisis contexts remains insufficiently evaluated. This work examines the performance of state-of-the-art LLMs and machine translation systems in crisis-domain translation, with a focus on preserving urgency, which is a critical property for effective crisis communication and triaging. Using multilingual crisis data and a newly introduced urgency-annotated dataset covering over 32 languages, we show that both dedicated translation models and LLMs exhibit substantial performance degradation and instability. Crucially, even linguistically adequate translations can distort perceived urgency, and LLM-based urgency classifications vary widely depending on the language of the prompt and input. These findings highlight significant risks in deploying general-purpose language technologies for crisis communication and underscore the need for crisis-aware evaluation frameworks.

Method: Used Support Vector Machines (SVM), Logistic Regression, and Decision Trees with parameter tuning and Synthetic Minority Over-sampling Technique (SMOTE) to handle data imbalance in Swahili text classification.

Result: Models performed well on high-dimensional textual data but limited by small, imbalanced dataset affecting generalizability; precision, recall, and F1 scores analyzed to show nuanced model performance.

Conclusion: Highlights need for expanded datasets, advanced ML techniques, and future work on data robustness, transfer learning, and multimodal data integration for culturally sensitive cyberbullying detection.

Abstract: The rise of digital technology has dramatically increased the potential for cyberbullying and online abuse, necessitating enhanced measures for detection and prevention, especially among children. This study focuses on detecting abusive obfuscated language in Swahili, a low-resource language that poses unique challenges due to its limited linguistic resources and technological support. Swahili is chosen due to its popularity and being the most widely spoken language in Africa, with over 16 million native speakers and upwards of 100 million speakers in total, spanning regions in East Africa and some parts of the Middle East. We employed machine learning models including Support Vector Machines (SVM), Logistic Regression, and Decision Trees, optimized through rigorous parameter tuning and techniques like Synthetic Minority Over-sampling Technique (SMOTE) to handle data imbalance. Our analysis revealed that, while these models perform well in high-dimensional textual data, our dataset’s small size and imbalance limit our findings’ generalizability. Precision, recall, and F1 scores were thoroughly analyzed, highlighting the nuanced performance of each model in detecting obfuscated language. This research contributes to the broader discourse on ensuring safer online environments for children, advocating for expanded datasets and advanced machine-learning techniques to improve the effectiveness of cyberbullying detection systems. Future work will focus on enhancing data robustness, exploring transfer learning, and integrating multimodal data to create more comprehensive and culturally sensitive detection mechanisms.

Method: Introduces a parallelizable encoder-decoder memory model architecture that combines causal training (next token prediction) with information retention objectives. Uses frozen high-fidelity encoders and curriculum training where decoders first learn to process memories, then predict next tokens.

Result: Shows that standard language model embeddings contain little input information regardless of training scale, while autoencoder embeddings can achieve near-perfect memory. The proposed combined objective approach enables information-rich memory formation and decoding.

Conclusion: Next token prediction alone is poorly suited for accurate memory formation due to its non-invertible nature. Combined objective functions are necessary for models to form and decode information-rich memories, especially when the entire input is not exposed.

Abstract: The ability of machine learning models to store input information in hidden layer vector embeddings, analogous to the concept of `memory’, is widely employed but not well characterized. We find that language model embeddings typically contain relatively little input information regardless of data and compute scale during training. In contrast, embeddings from autoencoders trained for input regeneration are capable of nearly perfect memory formation. The substitution of memory embeddings for token sequences leads to substantial computational efficiencies, motivating the introduction of a parallelizable encoder-decoder memory model architecture. Upon causal training these models contain information-poor embeddings incapable of arbitrary information access, but by combining causal and information retention objective functions they learn to form and decode information-rich memories. Training can be further streamlined by freezing a high fidelity encoder followed by a curriculum training approach where decoders first learn to process memories and then learn to additionally predict next tokens. We introduce the perspective that next token prediction training alone is poorly suited for accurate memory formation as the objective itself is non-invertible, motivating the use of combined objective functions for models where the entire input is not exposed.

Method: Fine-tuned Turkish-specific BERT model (dbmdz/bert-base-turkish-cased) on 3,600 labeled articles from three major Turkish outlets for binary classification of AI-rewritten content

Result: Model achieved 0.9708 F1 score, deployed on 3,500+ unseen articles (2023-2026) revealing consistent patterns with mean confidence >0.96 and ~2.5% AI-rewritten content

Conclusion: First empirical measurement of AI usage in Turkish news media, establishing baseline for future monitoring of AI-generated content proliferation

Abstract: The rapid integration of large language models into newsroom workflows has raised urgent questions about the prevalence of AI-generated content in online media. While computational studies have begun to quantify this phenomenon in English-language outlets, no empirical investigation exists for Turkish news media, where existing research remains limited to qualitative interviews with journalists or fake news detection. This study addresses that gap by fine-tuning a Turkish-specific BERT model (dbmdz/bert-base-turkish-cased) on a labeled dataset of 3,600 articles from three major Turkish outlets with distinct editorial orientations for binary classification of AI-rewritten content. The model achieves 0.9708 F1 score on the held-out test set with symmetric precision and recall across both classes. Subsequent deployment on over 3,500 unseen articles spanning between 2023 and 2026 reveals consistent cross-source and temporally stable classification patterns, with mean prediction confidence exceeding 0.96 and an estimated 2.5 percentage of examined news content rewritten or revised by LLMs on average. To the best of our knowledge, this is the first study to move beyond self-reported journalist perceptions toward empirical, data-driven measurement of AI usage in Turkish news media.

Method: Quantify inference-time effort by identifying deep-thinking tokens (tokens where internal predictions undergo significant revisions in deeper model layers before convergence). Use deep-thinking ratio (proportion of deep-thinking tokens) as metric, then introduce Think@n test-time scaling strategy that prioritizes samples with high deep-thinking ratios.

Result: Deep-thinking ratio shows robust positive correlation with accuracy across four challenging benchmarks (AIME 24/25, HMMT 25, GPQA-diamond) and multiple models (GPT-OSS, DeepSeek-R1, Qwen3), outperforming length-based and confidence-based baselines. Think@n matches/exceeds standard self-consistency performance while reducing inference costs via early rejection of unpromising generations.

Conclusion: Deep-thinking tokens provide better signal for reasoning quality than token counts, enabling more efficient test-time scaling strategies that maintain performance while reducing computational costs.

Abstract: Large language models (LLMs) have demonstrated impressive reasoning capabilities by scaling test-time compute via long Chain-of-Thought (CoT). However, recent findings suggest that raw token counts are unreliable proxies for reasoning quality: increased generation length does not consistently correlate with accuracy and may instead signal “overthinking,” leading to performance degradation. In this work, we quantify inference-time effort by identifying deep-thinking tokens – tokens where internal predictions undergo significant revisions in deeper model layers prior to convergence. Across four challenging mathematical and scientific benchmarks (AIME 24/25, HMMT 25, and GPQA-diamond) and a diverse set of reasoning-focused models (GPT-OSS, DeepSeek-R1, and Qwen3), we show that deep-thinking ratio (the proportion of deep-thinking tokens in a generated sequence) exhibits a robust and consistently positive correlation with accuracy, substantially outperforming both length-based and confidence-based baselines. Leveraging this insight, we introduce Think@n, a test-time scaling strategy that prioritizes samples with high deep-thinking ratios. We demonstrate that Think@n matches or exceeds standard self-consistency performance while significantly reducing inference costs by enabling the early rejection of unpromising generations based on short prefixes.

Method: Introduces capability calibration which targets the model’s expected accuracy on a query, formally distinguishes it from response calibration, establishes an empirical evaluation setup, and studies various confidence estimation methods.

Result: Capability-calibrated confidence improves pass@k prediction and inference budget allocation, showing practical benefits over traditional response calibration approaches.

Conclusion: Capability calibration provides a better foundation for confidence estimation in LLMs, with potential for diverse applications beyond traditional response-level calibration.

Abstract: Large language models (LLMs) are widely deployed as general-purpose problem solvers, making accurate confidence estimation critical for reliable use. Prior work on LLM calibration largely focuses on response-level confidence, which estimates the correctness of a single generated output. However, this formulation is misaligned with many practical settings where the central question is how likely a model is to solve a query overall. We show that this mismatch results from the stochastic nature of modern LLM decoding, under which single-response correctness fails to reflect underlying model capability. To address this issue, we introduce capability calibration, which targets the model’s expected accuracy on a query. We formally distinguish capability calibration from response calibration and show that the two differ both theoretically and empirically. We establish an empirical evaluation setup and study a range of confidence estimation methods. Our results demonstrate that capability-calibrated confidence improves pass@$k$ prediction and inference budget allocation, establishing a foundation with potential for diverse applications.

Method: FLIP reformulates reward modeling through backward inference: it infers the instruction that would most plausibly produce a given response, then uses the similarity between the inferred and original instructions as the reward signal. This exploits the validation-generation gap.

Result: FLIP outperforms LLM-as-a-Judge baselines by an average of 79.6% across four domains using 13 small language models. It substantially improves downstream performance in extrinsic evaluations under test-time scaling via parallel sampling and GRPO training. FLIP is particularly effective for longer outputs and robust to common forms of reward hacking.

Conclusion: FLIP enables reliable reward modeling in downscaled regimes where judgment methods fail, offering a reference-free and rubric-free approach that explicitly exploits the validation-generation gap for more accessible and flexible reward modeling.

Abstract: Reward models (RMs) play a central role throughout the language model (LM) pipeline, particularly in non-verifiable domains. However, the dominant LLM-as-a-Judge paradigm relies on the strong reasoning capabilities of large models, while alternative approaches require reference responses or explicit rubrics, limiting flexibility and broader accessibility. In this work, we propose FLIP (FLipped Inference for Prompt reconstruction), a reference-free and rubric-free reward modeling approach that reformulates reward modeling through backward inference: inferring the instruction that would most plausibly produce a given response. The similarity between the inferred and the original instructions is then used as the reward signal. Evaluations across four domains using 13 small language models show that FLIP outperforms LLM-as-a-Judge baselines by an average of 79.6%. Moreover, FLIP substantially improves downstream performance in extrinsic evaluations under test-time scaling via parallel sampling and GRPO training. We further find that FLIP is particularly effective for longer outputs and robust to common forms of reward hacking. By explicitly exploiting the validation-generation gap, FLIP enables reliable reward modeling in downscaled regimes where judgment methods fail. Code available at https://github.com/yikee/FLIP.

Method: Projects intermediate hidden states into vocabulary space via Logit Lens, then enforces structural alignment using symmetric divergence objective that imposes dual-sided penalty to prevent overconfidence/underconfidence while preserving high-entropy information.

Result: Extensive experiments on GPT-2 and Llama architectures show DistillLens consistently outperforms standard KD and feature-transfer baselines on diverse instruction-following benchmarks.

Conclusion: DistillLens provides an effective framework for knowledge distillation that better captures teacher models’ intermediate reasoning processes, leading to improved student model performance.

Abstract: Standard Knowledge Distillation (KD) compresses Large Language Models (LLMs) by optimizing final outputs, yet it typically treats the teacher’s intermediate layer’s thought process as a black box. While feature-based distillation attempts to bridge this gap, existing methods (e.g., MSE and asymmetric KL divergence) ignore the rich uncertainty profiles required for the final output. In this paper, we introduce DistillLens, a framework that symmetrically aligns the evolving thought processes of student and teacher models. By projecting intermediate hidden states into the vocabulary space via the Logit Lens, we enforce structural alignment using a symmetric divergence objective. Our analysis proves that this constraint imposes a dual-sided penalty, preventing both overconfidence and underconfidence while preserving the high-entropy information conduits essential for final deduction. Extensive experiments on GPT-2 and Llama architectures demonstrate that DistillLens consistently outperforms standard KD and feature-transfer baselines on diverse instruction-following benchmarks. The code is available at https://github.com/manishdhakal/DistillLens.

Method: Two-stage process: 1) Confidence assessment using multinomial sampling and Maximum Semantic Cluster Proportion clustering, 2) Binning and multi-level sorting based on query and document confidence thresholds to prioritize relevant documents while preserving original rankings for high-confidence queries.

Result: LCR consistently improves NDCG@5 by up to 20.6% across BEIR and TREC benchmarks using BM25 and Contriever retrievers with only 7-9B parameter LLMs, without performance degradation compared to existing rerankers.

Conclusion: LCR provides an efficient, training-free approach that leverages LLM confidence signals to improve document retrieval in RAG systems, reducing hallucinations while maintaining computational efficiency and broad compatibility.

Abstract: Large language models (LLMs) have revolutionized natural language processing, yet hallucinations in knowledge-intensive tasks remain a critical challenge. Retrieval-augmented generation (RAG) addresses this by integrating external knowledge, but its efficacy depends on accurate document retrieval and ranking. Although existing rerankers demonstrate effectiveness, they frequently necessitate specialized training, impose substantial computational expenses, and fail to fully exploit the semantic capabilities of LLMs, particularly their inherent confidence signals. We propose the LLM-Confidence Reranker (LCR), a training-free, plug-and-play algorithm that enhances reranking in RAG systems by leveraging black-box LLM confidence derived from Maximum Semantic Cluster Proportion (MSCP). LCR employs a two-stage process: confidence assessment via multinomial sampling and clustering, followed by binning and multi-level sorting based on query and document confidence thresholds. This approach prioritizes relevant documents while preserving original rankings for high-confidence queries, ensuring robustness. Evaluated on BEIR and TREC benchmarks with BM25 and Contriever retrievers, LCR–using only 7–9B-parameter pre-trained LLMs–consistently improves NDCG@5 by up to 20.6% across pre-trained LLM and fine-tuned Transformer rerankers, without degradation. Ablation studies validate the hypothesis that LLM confidence positively correlates with document relevance, elucidating LCR’s mechanism. LCR offers computational efficiency, parallelism for scalability, and broad compatibility, mitigating hallucinations in applications like medical diagnosis.

Method: Introduces Elo-Evolve framework with two key innovations: (1) learning directly from binary win/loss outcomes in pairwise competitions instead of Bradley-Terry models, and (2) Elo-orchestrated opponent selection with temperature-controlled sampling for automatic curriculum learning. Grounded in PAC learning theory.

Result: Achieves 4.5x noise reduction compared to absolute scoring approaches. Trained Qwen2.5-7B model outperforms point-based methods and static pairwise training across Alpaca Eval 2.0 and MT-Bench benchmarks, demonstrating clear performance hierarchy.

Conclusion: Pairwise comparison with dynamic opponent selection provides superior LLM alignment, offering progressive benefits over traditional methods through co-evolutionary competition.

Abstract: Current alignment methods for Large Language Models (LLMs) rely on compressing vast amounts of human preference data into static, absolute reward functions, leading to data scarcity, noise sensitivity, and training instability. We introduce Elo-Evolve, a co-evolutionary framework that redefines alignment as dynamic multi-agent competition within an adaptive opponent pool. Our approach makes two key innovations: (1) eliminating Bradley-Terry model dependencies by learning directly from binary win/loss outcomes in pairwise competitions, and (2) implementing Elo-orchestrated opponent selection that provides automatic curriculum learning through temperature-controlled sampling. We ground our approach in PAC learning theory, demonstrating that pairwise comparison achieves superior sample complexity and empirically validate a 4.5x noise reduction compared to absolute scoring approaches. Experimentally, we train a Qwen2.5-7B model using our framework with opponents including Qwen2.5-14B, Qwen2.5-32B, and Qwen3-8B models. Results demonstrate a clear performance hierarchy: point-based methods < static pairwise training < Elo-Evolve across Alpaca Eval 2.0 and MT-Bench, validating the progressive benefits of pairwise comparison and dynamic opponent selection for LLM alignment.

Method: Comparative evaluation of three approaches: 1) traditional multimodal context (text history + spoken current turn), 2) full spoken history, and 3) compressed spoken history using attention-pooling-based compression. Experiments conducted on SpokenWOZ corpus.

Result: Full spoken conversation as input yields highest performance among similar-sized models, significantly surpassing prior methods. Attention-pooling compression maintains competitive accuracy with reduced context size. Improvements stem from more effective context utilization.

Conclusion: Full spoken history is optimal for spoken dialog state tracking, while compression offers practical trade-off for context size reduction without significant accuracy loss.

Abstract: This paper presents a comparative study of context management strategies for end-to-end Spoken Dialog State Tracking using Speech-LLMs. We systematically evaluate traditional multimodal context (combining text history and spoken current turn), full spoken history, and compressed spoken history approaches. Our experiments on the SpokenWOZ corpus demonstrate that providing the full spoken conversation as input yields the highest performance among models of similar size, significantly surpassing prior methods. Furthermore, we show that attention-pooling-based compression of the spoken history offers a strong trade-off, maintaining competitive accuracy with reduced context size. Detailed analysis confirms that improvements stem from more effective context utilization.

Method: Trained word embeddings on literary and nonliterary Italian corpora from 19th and 21st centuries (124M tokens total), modeled changes in semantic similarity between topics and vehicles of 515 19th-century literary metaphors as proxy for processing demands.

Result: Overall semantic similarity (and metaphor processing demands) remained stable over time. Genre played key role: metaphors appeared more difficult in modern literary contexts but easier in today’s nonliterary language (Web). Pattern shaped by semantic features like vector coherence and neighborhood density.

Conclusion: Findings align with broader linguistic changes: stylistic simplification of modern literature increased metaphor processing demands, while high creativity of Web’s language made metaphors more accessible. Demonstrates value of diachronic computational approaches to literary metaphor processing.

Abstract: Metaphors are a distinctive feature of literary language, yet they remain less studied experimentally than everyday metaphors. Moreover, previous psycholinguistic and computational approaches overlooked the temporal dimension, although many literary metaphors were coined centuries apart from contemporary readers. This study innovatively applies tools from diachronic distributional semantics to assess whether the processing costs of literary metaphors varied over time and genre. Specifically, we trained word embeddings on literary and nonliterary Italian corpora from the 19th and 21st centuries, for a total of 124 million tokens, and modeled changes in the semantic similarity between topics and vehicles of 515 19th-century literary metaphors, taking this measure as a proxy of metaphor processing demands. Overall, semantic similarity, and hence metaphor processing demands, remained stable over time. However, genre played a key role: metaphors appeared more difficult (i.e., lower topic-vehicle similarity) in modern literary contexts than in 19th-century literature, but easier (i.e., higher topic-vehicle similarity) in today’s nonliterary language (e.g., the Web) than in 19th-century nonliterary texts. This pattern was further shaped by semantic features of metaphors’ individual terms, such as vector coherence and semantic neighborhood density. Collectively, these findings align with broader linguistic changes in Italian, such as the stylistic simplification of modern literature, which may have increased metaphor processing demands, and the high creativity of the Web’s language, which seems to render metaphor more accessible.

Method: Release-level analysis of Pashto component in Mozilla Common Voice corpus (version 24.0, December 2025), examining trends across major releases. Analysis includes scale metrics, validation throughput, contributor participation inequality (Gini coefficient), demographic metadata completeness, and sentence-level concentration in validated subsets.

Result: Rapid growth from 1.49 hours (mid-2023) to 2,768.7 total hours (2025), with 975.89 validated hours for supervised ASR training. Participation extremely concentrated (Gini = 0.941), age representation skewed toward young adults, 41.97% of clips lack gender labels, and 35.88% of unique sentences account for 50% of validated clips.

Conclusion: The study provides quantitative audit of a rapidly scaling low-resource speech corpus, highlighting practical priorities for improving dataset maturity including expanded validation capacity and broader demographic participation to address metadata gaps and contributor concentration.

Abstract: Large, openly licensed speech datasets are essential for building automatic speech recognition (ASR) systems, yet many widely spoken languages remain underrepresented in public resources. Pashto, spoken by more than 60 million people, has historically lacked large-scale openly licensed speech data suitable for modern ASR development. This paper presents a release-level analysis of the Pashto component of the Mozilla Common Voice corpus, focusing on version 24.0 (December 2025) and contextualizing trends across major releases. We document rapid growth from 1.49 recorded hours in mid-2023 to 2,768.7 total hours in 2025, including 975.89 validated hours available for supervised ASR training. Beyond scale, we analyze validation throughput, contributor participation inequality, demographic metadata completeness, and sentence-level concentration in the validated subset. We find that participation is extremely concentrated (Gini = 0.941), age representation is strongly skewed toward young adults, and 41.97% of clips lack self-reported gender labels, limiting subgroup auditing based on metadata. At the textual level, prompt reuse is moderate: 35.88% of unique sentences account for 50% of validated clips, suggesting that structural concentration is driven primarily by uneven contributor activity rather than dominance of a small prompt set. These results provide a quantitative audit of a rapidly scaling low-resource speech corpus and highlight practical priorities for improving dataset maturity, including expanded validation capacity and broader demographic participation.

Method: Comparative multi-agent framework with two parallel LLM-based systems: one enhanced by argumentation theory via Retrieval-Augmented Generation (RAG), and an identical zero-shot baseline. Uses annotated political debates dataset with D-I-S-G-O categories (Deintensification, Intensification, Specification, Generalisation, Other).

Result: RAG-enhanced agents substantially outperform baseline across all categories, with particularly strong advantages in detecting Intensification and Generalisation contexts. Overall Macro F1-score improvement of nearly 30%.

Conclusion: Theoretical grounding is essential for advancing beyond mere paraphrase detection to function-aware analysis of argumentative discourse. The multi-agent architecture enables scalable, theoretically informed computational tools for identifying rhetorical strategies.

Abstract: Identifying the strategic uses of reformulation in discourse remains a key challenge for computational argumentation. While LLMs can detect surface-level similarity, they often fail to capture the pragmatic functions of rephrasing, such as its role within rhetorical discourse. This paper presents a comparative multi-agent framework designed to quantify the benefits of incorporating explicit theoretical knowledge for this task. We utilise an dataset of annotated political debates to establish a new standard encompassing four distinct rephrase functions: Deintensification, Intensification, Specification, Generalisation, and Other, which covers all remaining types (D-I-S-G-O). We then evaluate two parallel LLM-based agent systems: one enhanced by argumentation theory via Retrieval-Augmented Generation (RAG), and an identical zero-shot baseline. The results reveal a clear performance gap: the RAG-enhanced agents substantially outperform the baseline across the board, with particularly strong advantages in detecting Intensification and Generalisation context, yielding an overall Macro F1-score improvement of nearly 30%. Our findings provide evidence that theoretical grounding is not only beneficial but essential for advancing beyond mere paraphrase detection towards function-aware analysis of argumentative discourse. This comparative multi-agent architecture represents a step towards scalable, theoretically informed computational tools capable of identifying rhetorical strategies in contemporary discourse.

Method: RMPL uses relation-aware multi-task progressive learning with heterogeneous supervision from unimodal event extraction and multimedia relation extraction. It employs stage-wise training: first learns shared event-centric representations across modalities using a unified schema, then fine-tunes for event mention identification and argument role extraction using mixed textual and visual data.

Result: Experiments on M2E2 benchmark with multiple Vision-Language Models show consistent improvements across different modality settings compared to existing approaches.

Conclusion: RMPL effectively addresses data scarcity in MEE by leveraging heterogeneous supervision and progressive learning, improving multimodal event extraction performance through better structured event representation learning.

Abstract: Multimedia Event Extraction (MEE) aims to identify events and their arguments from documents that contain both text and images. It requires grounding event semantics across different modalities. Progress in MEE is limited by the lack of annotated training data. M2E2 is the only established benchmark, but it provides annotations only for evaluation. This makes direct supervised training impractical. Existing methods mainly rely on cross-modal alignment or inference-time prompting with Vision–Language Models (VLMs). These approaches do not explicitly learn structured event representations and often produce weak argument grounding in multimodal settings. To address these limitations, we propose RMPL, a Relation-aware Multi-task Progressive Learning framework for MEE under low-resource conditions. RMPL incorporates heterogeneous supervision from unimodal event extraction and multimedia relation extraction with stage-wise training. The model is first trained with a unified schema to learn shared event-centric representations across modalities. It is then fine-tuned for event mention identification and argument role extraction using mixed textual and visual data. Experiments on the M2E2 benchmark with multiple VLMs show consistent improvements across different modality settings.

Method: Researchers manually annotated 1,404 word use-to-sign ID mappings from German Word Usage Graph and Digital Dictionary of German Sign Language. They identified three correspondence types and evaluated computational methods: Exact Match and Semantic Similarity using SBERT embeddings.

Result: Semantic Similarity substantially outperformed Exact Match (88.52% vs. 71.31%), with dramatic gains for Type 1 correspondences (+52.1 percentage points). The work established the first annotated dataset for cross-modal sense correspondence.

Conclusion: This research provides computational methods and annotated resources for identifying cross-modal sense correspondences between spoken and sign languages, revealing which correspondence patterns are computationally identifiable.

Abstract: Sign language lexicographers construct bilingual dictionaries by establishing word-to-sign mappings, where polysemous and homonymous words corresponding to different signs across contexts are often underrepresented. A usage-based approach examining how word senses map to signs can identify such novel mappings absent from current dictionaries, enriching lexicographic resources. We address this by analyzing German and German Sign Language (Deutsche Gebärdensprache, DGS), manually annotating 1,404 word use-to-sign ID mappings derived from 32 words from the German Word Usage Graph (D-WUG) and 49 signs from the Digital Dictionary of German Sign Language (DW-DGS). We identify three correspondence types: Type 1 (one-to-many), Type 2 (many-to-one), and Type 3 (one-to-one), plus No Match cases. We evaluate computational methods: Exact Match (EM) and Semantic Similarity (SS) using SBERT embeddings. SS substantially outperforms EM overall 88.52% vs. 71.31%), with dramatic gains for Type 1 (+52.1 pp). Our work establishes the first annotated dataset for cross-modal sense correspondence and reveals which correspondence patterns are computationally identifiable. Our code and dataset are made publicly available.

Method: Developed OMGs (Ovarian tumour Multidisciplinary intelligent aGent System), a multi-agent AI framework where domain-specific agents collaborate to integrate multidisciplinary evidence and generate MDT-style recommendations with transparent rationales. Created SPEAR (Safety, Personalization, Evidence, Actionability, Robustness) framework to systematically evaluate MDT recommendation quality.

Result: In multicentre re-evaluation, OMGs achieved performance comparable to expert MDT consensus (4.45 ± 0.30 vs 4.53 ± 0.23), with higher Evidence scores (4.57 vs 3.92). In prospective multicentre evaluation with 59 patients, OMGs demonstrated high concordance with routine MDT decisions. In paired human-AI studies, OMGs most substantially enhanced clinicians’ recommendations in Evidence and Robustness dimensions.

Conclusion: Multi-agent deliberative AI systems can achieve performance comparable to expert MDT consensus and have potential to expand access to specialized oncology expertise in resource-limited settings, particularly enhancing Evidence and Robustness dimensions where multidisciplinary expertise is unavailable.

Abstract: Ovarian tumour management has increasingly relied on multidisciplinary tumour board (MDT) deliberation to address treatment complexity and disease heterogeneity. However, most patients worldwide lack access to timely expert consensus, particularly in resource-constrained centres where MDT resources are scarce or unavailable. Here we present OMGs (Ovarian tumour Multidisciplinary intelligent aGent System), a multi-agent AI framework where domain-specific agents deliberate collaboratively to integrate multidisciplinary evidence and generate MDT-style recommendations with transparent rationales. To systematically evaluate MDT recommendation quality, we developed SPEAR (Safety, Personalization, Evidence, Actionability, Robustness) and validated OMGs across diverse clinical scenarios spanning the care continuum. In multicentre re-evaluation, OMGs achieved performance comparable to expert MDT consensus ($4.45 \pm 0.30$ versus $4.53 \pm 0.23$), with higher Evidence scores (4.57 versus 3.92). In prospective multicentre evaluation (59 patients), OMGs demonstrated high concordance with routine MDT decisions. Critically, in paired human-AI studies, OMGs most substantially enhanced clinicians’ recommendations in Evidence and Robustness, the dimensions most compromised when multidisciplinary expertise is unavailable. These findings suggest that multi-agent deliberative systems can achieve performance comparable to expert MDT consensus, with potential to expand access to specialized oncology expertise in resource-limited settings.

Method: Entity-aware continual pretraining organizes heterogeneous medical corpora; incorporates diverse medical reasoning patterns via reinforcement learning and tool-augmented agentic training; integrates user-preference rubrics, evidence-grounded reasoning, and low-hallucination report generation.

Result: Achieves state-of-the-art performance across diverse medical benchmarks and surpasses leading closed-source multimodal systems on multiple capabilities.

Conclusion: MedXIAOHE represents an advanced medical vision-language foundation model with practical design choices for real-world clinical applications, offering improved medical understanding, reasoning, and reliability.

Abstract: We present MedXIAOHE, a medical vision-language foundation model designed to advance general-purpose medical understanding and reasoning in real-world clinical applications. MedXIAOHE achieves state-of-the-art performance across diverse medical benchmarks and surpasses leading closed-source multimodal systems on multiple capabilities. To achieve this, we propose an entity-aware continual pretraining framework that organizes heterogeneous medical corpora to broaden knowledge coverage and reduce long-tail gaps (e.g., rare diseases). For medical expert-level reasoning and interaction, MedXIAOHE incorporates diverse medical reasoning patterns via reinforcement learning and tool-augmented agentic training, enabling multi-step diagnostic reasoning with verifiable decision traces. To improve reliability in real-world use, MedXIAOHE integrates user-preference rubrics, evidence-grounded reasoning, and low-hallucination long-form report generation, with improved adherence to medical instructions. We release this report to document our practical design choices, scaling insights, and evaluation framework, hoping to inspire further research.

Method: Used Universal Grammar approach with Feature Reassembly Hypothesis, analyzed learner errors across two developmental stages, employed statistical analysis including one-way ANOVA to measure improvement in irregular inflection production.

Result: Stage 1 showed dominant L1 transfer influence in phonological/structural mismatches; Stage 2 demonstrated increased UG sensitivity and morphological reconfiguration. Significant improvement in well-formed irregular inflections from stage 1 to stage 2, but persistent difficulties with specific inflection types.

Conclusion: While L1 transfer and developmental factors influence initial acquisition, appropriate linguistic input and instruction are critical for facilitating UG-driven feature reassembly in adult L2 learners, though full UG access remains constrained by limited exposure and instructional quality.

Abstract: This study examines the acquisition of English irregular inflections by Yemeni learners of English as a second language (L2), utilizing a Universal Grammar (UG) approach. Within the UG approach, the study considers Feature Reassembly Hypothesis (FRH) (Lardiere, 2008, 2009) part of UG, focusing on the roles of first language (L1) transfer and L2 developmental influence. It analyzes learner errors across two developmental stages. Stage 1 data reveal a dominant influence of L1 transfer, particularly in phonological and structural mismatches, while stage 2 data demonstrate increased learner sensitivity to UG properties and morphological reconfiguration toward the target language. Findings reveal that errors in irregular inflectional morphology are attributed to both interlingual and intralingual sources, with overgeneralization of L2 rules as a common developmental strategy. Statistical analysis, including a one-way ANOVA, indicates significant improvement in the production of well-formed irregular inflections from stage 1 to stage 2, underscoring learners’ continued access to UG. However, persistent difficulties with consonant change, zero-morpheme, and -a plural inflections suggest that limited exposure, ineffective input modeling, and insufficient instructional quality constrain full UG access. The study concludes that while L1 transfer and L2 developmental factors influence initial stages of acquisition, appropriate linguistic input and instruction are critical for facilitating UG-driven feature reassembly in adult L2 learners.

Method: Formalized ToM for LLMs as epistemic divergence detection/resolution mechanism, created a benchmark to assess how models reconcile user beliefs and profiles, curated trajectory-based ToM dataset linking belief tracking with task-related state inference, and trained models via reinforcement learning.

Result: Evaluation of 11 leading models revealed significant limitations in identifying underlying cognitive gaps that impede task success. Models trained on the trajectory-based dataset via reinforcement learning showed consistent improvement in reasoning about user mental states and enhanced downstream performance.

Conclusion: ToM should be viewed as an essential interaction-level mechanism rather than a standalone reasoning skill, highlighting its practical value in improving LLM-user interactions.

Abstract: Large Language Models (LLMs) have developed rapidly and are widely applied to both general-purpose and professional tasks to assist human users. However, they still struggle to comprehend and respond to the true user needs when intentions and instructions are imprecisely conveyed, leading to a divergence between subjective user believes and true environment states. Resolving this epistemic divergence requires Theory of Mind (ToM), yet existing ToM evaluations for LLMs primarily focus on isolated belief inference, overlooking its functional utility in real-world interaction. To this end, we formalize ToM for LLMs as a mechanism for epistemic divergence detection and resolution, and propose a benchmark, \benchname, to assess how models reconcile user beliefs and profiles in practice. Results across 11 leading models reveal a significant limitation to identify underlying cognitive gaps that impede task success. To bridge this gap, we further curate a trajectory-based ToM dataset linking belief tracking with task-related state inference. The model trained on this data via reinforcement learning shows consistent improvement in reasoning about user mental states, leading to enhanced downstream performance. Our work highlights the practical value of ToM as an essential interaction-level mechanism rather than as a standalone reasoning skill.

Method: Proposes SpecVocab, which selects a vocabulary subset per decoding step rather than using a fixed reduced vocabulary, enabling efficient speculation while maintaining coverage of potential target tokens.

Result: SpecVocab achieves higher acceptance length than state-of-the-art EAGLE-3, yielding up to 8.1% increase in average throughput across various tasks.

Conclusion: Vocabulary speculation via dynamic subset selection is an effective alternative to fixed reduced vocabularies for speculative decoding, improving both efficiency and effectiveness.

Abstract: Speculative decoding has rapidly emerged as a leading approach for accelerating language model (LM) inference, as it offers substantial speedups while yielding identical outputs. This relies upon a small draft model, tasked with predicting the outputs of the target model. State-of-the-art speculative decoding methods use a draft model consisting of a single decoder layer and output embedding matrix, with the latter dominating drafting time for the latest LMs. Recent work has sought to address this output distribution bottleneck by reducing the vocabulary of the draft model. Although this can improve throughput, it compromises speculation effectiveness when the target token is out-of-vocabulary. In this paper, we argue for vocabulary speculation as an alternative to a reduced vocabulary. We propose SpecVocab, an efficient and effective method that selects a vocabulary subset per decoding step. Across a variety of tasks, we demonstrate that SpecVocab can achieve a higher acceptance length than state-of-the-art speculative decoding approach, EAGLE-3. Notably, this yields up to an 8.1% increase in average throughput over EAGLE-3.

Method: Proposes PrivAct, a contextual privacy-aware multi-agent learning framework that internalizes contextual privacy preservation directly into models’ generation behavior. Embeds privacy preferences into each agent to enhance system-wide contextual integrity while optimizing privacy-helpfulness tradeoff.

Result: Experiments across multiple LLM backbones and benchmarks show consistent improvements in contextual privacy preservation, reducing leakage rates by up to 12.32% while maintaining comparable helpfulness. Demonstrates zero-shot generalization and robustness across diverse multi-agent topologies.

Conclusion: PrivAct effectively internalizes contextual privacy preservation into LLM agents, achieving better privacy-helpfulness tradeoffs than external intervention approaches, with strong generalization across different agent configurations.

Abstract: Large language model (LLM) agents are increasingly deployed in personalized tasks involving sensitive, context-dependent information, where privacy violations may arise in agents’ action due to the implicitness of contextual privacy. Existing approaches rely on external, inference-time interventions which are brittle, scenario-specific, and may expand the privacy attack surface. We propose PrivAct, a contextual privacy-aware multi-agent learning framework that internalizes contextual privacy preservation directly into models’ generation behavior for privacy-compliant agentic actions. By embedding privacy preferences into each agent, PrivAct enhances system-wide contextual integrity while achieving a more favorable privacy-helpfulness tradeoff. Experiments across multiple LLM backbones and benchmarks demonstrate consistent improvements in contextual privacy preservation, reducing leakage rates by up to 12.32% while maintaining comparable helpfulness, as well as zero-shot generalization and robustness across diverse multi-agent topologies. Code is available at https://github.com/chengyh23/PrivAct.

Method: Three interconnected threads: 1) domain adaptation for technical precision, 2) ethical rigor to mitigate adversarial vulnerabilities, and 3) cultural/multilingual alignment for global inclusivity. Methodological progression from supervised adaptation to decoding-time alignment to human feedback modeling.

Result: Proposes a comprehensive “Responsible Intelligence” framework that systematically addresses technical, ethical, and cultural dimensions of LLM deployment.

Conclusion: A holistic approach is needed to reconcile LLMs’ generative power with real-world requirements through integrated domain adaptation, safety mechanisms, and cultural alignment.

Abstract: The overarching research direction of this work is the development of a ‘‘Responsible Intelligence’’ framework designed to reconcile the immense generative power of Large Language Models (LLMs) with the stringent requirements of real-world deployment. As these models become a transformative force in artificial intelligence, there is an urgent need to move beyond general-purpose architectures toward systems that are contextually aware, inherently safer, and deeply respectful of global cultural nuances. This research navigates three interconnected threads: domain adaptation to ensure technical precision, ethical rigor to mitigate adversarial vulnerabilities, and cultural/multilingual alignment to promote global inclusivity. The methodological trajectory moves from classical supervised adaptation for task-specific demands to decoding-time alignment for safety, finally leveraging human feedback and preference modeling to achieve sociolinguistic acuity.

Method: Two-phase approach: 1) Fine-tune three LLM families (GPT, LLaMA, DeepSeek R1) on medical QA data; 2) Implement modular multi-agent pipeline with Clinical Reasoning agent, Evidence Retrieval agent (PubMed queries), Refinement agent, and optional human validation for high-risk cases.

Result: DeepSeek R1 achieved best scores (ROUGE-1: 0.536, ROUGE-2: 0.226, BLEU: 0.098). Full system achieved 87% accuracy with 0.80 relevance, reduced uncertainty (perplexity 4.13), and 36.5s mean latency.

Conclusion: Agent specialization and verification layers can mitigate single-model limitations, providing practical design for evidence-based, bias-aware medical AI.

Abstract: Large language models (LLMs) show promise for healthcare question answering, but clinical use is limited by weak verification, insufficient evidence grounding, and unreliable confidence signalling. We propose a multi-agent medical QA framework that combines complementary LLMs with evidence retrieval, uncertainty estimation, and bias checks to improve answer reliability. Our approach has two phases. First, we fine-tune three representative LLM families (GPT, LLaMA, and DeepSeek R1) on MedQuAD-derived medical QA data (20k+ question-answer pairs across multiple NIH domains) and benchmark generation quality. DeepSeek R1 achieves the strongest scores (ROUGE-1 0.536 +- 0.04; ROUGE-2 0.226 +-0.03; BLEU 0.098 -+ 0.018) and substantially outperforms the specialised biomedical baseline BioGPT in zero-shot evaluation. Second, we implement a modular multi-agent pipeline in which a Clinical Reasoning agent (fine-tuned LLaMA) produces structured explanations, an Evidence Retrieval agent queries PubMed to ground responses in recent literature, and a Refinement agent (DeepSeek R1) improves clarity and factual consistency; an optional human validation path is triggered for high-risk or high-uncertainty cases. Safety mechanisms include Monte Carlo dropout and perplexity-based uncertainty scoring, plus lexical and sentiment-based bias detection supported by LIME/SHAP-based analyses. In evaluation, the full system achieves 87% accuracy with relevance around 0.80, and evidence augmentation reduces uncertainty (perplexity 4.13) compared to base responses, with mean end-to-end latency of 36.5 seconds under the reported configuration. Overall, the results indicate that agent specialisation and verification layers can mitigate key single-model limitations and provide a practical, extensible design for evidence-based and bias-aware medical AI.

Method: The paper synthesizes recent findings through literature review and analysis of existing research on multilingual safety gaps, identifying three key issues: weakened safety guardrails for low-resource/code-mixed inputs, persistence of culturally harmful behavior despite acceptable toxicity scores, and failure of English-only safety patches to transfer.

Result: Findings show: (1) Safety guardrails significantly weaken for low-resource and code-mixed inputs, (2) Culturally harmful behavior persists even when standard toxicity scores appear acceptable, (3) English-only knowledge edits and safety patches often fail to carry over to low-resource languages.

Conclusion: Proposes a practical agenda for Global South researchers: parameter-efficient safety steering, culturally grounded evaluation and preference data, and participatory workflows empowering local communities to define and mitigate harm. Argues multilingual safety must be a core requirement, not an add-on, for equitable AI in underrepresented regions.

Abstract: Large language models (LLMs) are being deployed across the Global South, where everyday use involves low-resource languages, code-mixing, and culturally specific norms. Yet safety pipelines, benchmarks, and alignment still largely target English and a handful of high-resource languages, implicitly assuming safety and factuality ‘’transfer’’ across languages. Evidence increasingly shows they do not. We synthesize recent findings indicating that (i) safety guardrails weaken sharply on low-resource and code-mixed inputs, (ii) culturally harmful behavior can persist even when standard toxicity scores look acceptable, and (iii) English-only knowledge edits and safety patches often fail to carry over to low-resource languages. In response, we outline a practical agenda for researchers and students in the Global South: parameter-efficient safety steering, culturally grounded evaluation and preference data, and participatory workflows that empower local communities to define and mitigate harm. Our aim is to make multilingual safety a core requirement-not an add-on-for equitable AI in underrepresented regions.

Method: Collected data from four online platforms (social media, e-commerce, customer service), covering Modern Standard Arabic and multiple dialects. Annotated based on Arabic linguistic traditions and pragmatic theory into three classes (polite, impolite, neutral) with 16 politeness categories. Created 10,000 samples with substantial inter-annotator agreement (kappa = 0.703).

Result: Benchmarked 40 model configurations including traditional ML, transformer-based models, and large language models. The dataset supports research on politeness-aware Arabic NLP.

Conclusion: ADAB addresses the gap in Arabic politeness resources and enables development of culturally-aware Arabic NLP systems that understand sociopragmatic phenomena.

Abstract: The growing importance of culturally-aware natural language processing systems has led to an increasing demand for resources that capture sociopragmatic phenomena across diverse languages. Nevertheless, Arabic-language resources for politeness detection remain under-explored, despite the rich and complex politeness expressions embedded in Arabic communication. In this paper, we introduce ADAB (Arabic Politeness Dataset), a new annotated Arabic dataset collected from four online platforms, including social media, e-commerce, and customer service domains, covering Modern Standard Arabic and multiple dialects (Gulf, Egyptian, Levantine, and Maghrebi). The dataset was annotated based on Arabic linguistic traditions and pragmatic theory, resulting in three classes: polite, impolite, and neutral. It contains 10,000 samples with linguistic feature annotations across 16 politeness categories and achieves substantial inter-annotator agreement (kappa = 0.703). We benchmark 40 model configurations, including traditional machine learning, transformer-based models, and large language models. The dataset aims to support research on politeness-aware Arabic NLP.

Method: Evaluated 24 different prompt templates on HotpotQA dataset, including standard RAG prompt, 9 well-formed techniques from literature, and 14 novel hybrid variants, tested on two SLMs (Qwen2.5-3B Instruct and Gemma3-4B-It) with 18720 test instances.

Result: Significant performance gains up to 83% on Qwen2.5 and 84.5% on Gemma3-4B-It, yielding up to 6% improvement compared to Standard RAG prompt for both models.

Conclusion: Provides concrete analysis and actionable recommendations for designing effective prompts for SLM-based RAG systems, particularly useful for resource-constrained deployments.

Abstract: Retrieval Augmented Generation (RAG) is a powerful approach for enhancing the factual grounding of language models by integrating external knowledge. While widely studied for large language models, the optimization of RAG for Small Language Models (SLMs) remains a critical research gap, particularly in complex, multi-hop question-answering tasks that require sophisticated reasoning. In these systems, prompt template design is a crucial yet under-explored factor influencing performance. This paper presents a large-scale empirical study to investigate this factor, evaluating 24 different prompt templates on the HotpotQA dataset. The set includes a standard RAG prompt, nine well-formed techniques from the literature, and 14 novel hybrid variants, all tested on two prominent SLMs: Qwen2.5-3B Instruct and Gemma3-4B-It. Our findings, based on a test set of 18720 instances, reveal significant performance gains of up to 83% on Qwen2.5 and 84.5% on Gemma3-4B-It, yielding an improvement of up to 6% for both models compared to the Standard RAG prompt. This research also offers concrete analysis and actionable recommendations for designing effective and efficient prompts for SLM-based RAG systems, practically for deployment in resource-constrained environments.

Method: Introduce Pre-Editorial Normalization (PEN) task, create new dataset from CoMMA corpus aligned with digitized Old French and Latin editions using passim, produce manually corrected gold-standard evaluation set, benchmark using ByT5-based sequence-to-sequence models.

Result: Created 4.66M-sample silver training corpus and 1.8k-sample gold evaluation set. Normalization model achieved 6.7% CER, substantially outperforming previous models for this task.

Conclusion: PEN successfully bridges the gap between palaeographic ATR outputs and normalized editions, providing both fidelity and usability with improved performance over existing approaches.

Abstract: Recent advances in Automatic Text Recognition (ATR) have improved access to historical archives, yet a methodological divide persists between palaeographic transcriptions and normalized digital editions. While ATR models trained on more palaeographically-oriented datasets such as CATMuS have shown greater generalizability, their raw outputs remain poorly compatible with most readers and downstream NLP tools, thus creating a usability gap. On the other hand, ATR models trained to produce normalized outputs have been shown to struggle to adapt to new domains and tend to over-normalize and hallucinate. We introduce the task of Pre-Editorial Normalization (PEN), which consists in normalizing graphemic ATR output according to editorial conventions, which has the advantage of keeping an intermediate step with palaeographic fidelity while providing a normalized version for practical usability. We present a new dataset derived from the CoMMA corpus and aligned with digitized Old French and Latin editions using passim. We also produce a manually corrected gold-standard evaluation set. We benchmark this resource using ByT5-based sequence-to-sequence models on normalization and pre-annotation tasks. Our contributions include the formal definition of PEN, a 4.66M-sample silver training corpus, a 1.8k-sample gold evaluation set, and a normalization model achieving a 6.7% CER, substantially outperforming previous models for this task.

Method: Two-stage validation-and-repair workflow: Stage I - binary validation through expert review and model-based cross-checks; Stage II - revision of fixable items with dual independent expert repairs, model-assisted auditing, and adjudication.

Result: Created HLE-Verified with 641 verified items and 1,170 revised-and-certified items; models show 7-10% absolute accuracy gain on HLE-Verified, with 30-40% gains on previously erroneous items.

Conclusion: HLE-Verified improves HLE-style evaluations by reducing annotation noise and enabling more faithful measurement of model capabilities.

Abstract: Humanity’s Last Exam (HLE) has become a widely used benchmark for evaluating frontier large language models on challenging, multi-domain questions. However, community-led analyses have raised concerns that HLE contains a non-trivial number of noisy items, which can bias evaluation results and distort cross-model comparisons. To address this challenge, we introduce HLE-Verified, a verified and revised version of HLE with a transparent verification protocol and fine-grained error taxonomy. Our construction follows a two-stage validation-and-repair workflow resulting in a certified benchmark. In Stage I, each item undergoes binary validation of the problem and final answer through domain-expert review and model-based cross-checks, yielding 641 verified items. In Stage II, flawed but fixable items are revised under strict constraints preserving the original evaluation intent, through dual independent expert repairs, model-assisted auditing, and final adjudication, resulting in 1,170 revised-and-certified items. The remaining 689 items are released as a documented uncertain set with explicit uncertainty sources and expertise tags for future refinement. We evaluate seven state-of-the-art language models on HLE and HLE-Verified, observing an average absolute accuracy gain of 7–10 percentage points on HLE-Verified. The improvement is particularly pronounced on items where the original problem statement and/or reference answer is erroneous, with gains of 30–40 percentage points. Our analyses further reveal a strong association between model confidence and the presence of errors in the problem statement or reference answer, supporting the effectiveness of our revisions. Overall, HLE-Verified improves HLE-style evaluations by reducing annotation noise and enabling more faithful measurement of model capabilities. Data is available at: https://github.com/SKYLENAGE-AI/HLE-Verified

Method: Proposes using LLMs to perform Chain-of-Thought reasoning on clinical EHRs. Instead of direct fine-tuning for classification, generates explicit diagnostic rationale through CoT reasoning paths, followed by structured CoT-based predictions for AD assessment.

Result: The CoT-based diagnostic framework significantly enhances stability and diagnostic performance across multiple CDR grading tasks, achieving up to 15% improvement in F1 score compared to zero-shot baseline methods.

Conclusion: LLM-generated CoT reasoning on EHRs improves AD diagnosis by providing explicit diagnostic rationale, enhancing both performance and interpretability across different stages of AD progression.

Abstract: Alzheimer’s disease (AD) has become a prevalent neurodegenerative disease worldwide. Traditional diagnosis still relies heavily on medical imaging and clinical assessment by physicians, which is often time-consuming and resource-intensive in terms of both human expertise and healthcare resources. In recent years, large language models (LLMs) have been increasingly applied to the medical field using electronic health records (EHRs), yet their application in Alzheimer’s disease assessment remains limited, particularly given that AD involves complex multifactorial etiologies that are difficult to observe directly through imaging modalities. In this work, we propose leveraging LLMs to perform Chain-of-Thought (CoT) reasoning on patients’ clinical EHRs. Unlike direct fine-tuning of LLMs on EHR data for AD classification, our approach utilizes LLM-generated CoT reasoning paths to provide the model with explicit diagnostic rationale for AD assessment, followed by structured CoT-based predictions. This pipeline not only enhances the model’s ability to diagnose intrinsically complex factors but also improves the interpretability of the prediction process across different stages of AD progression. Experimental results demonstrate that the proposed CoT-based diagnostic framework significantly enhances stability and diagnostic performance across multiple CDR grading tasks, achieving up to a 15% improvement in F1 score compared to the zero-shot baseline method.

Method: Apply information bottleneck principle to conceptualize explanations as compressed representations. Introduce evaluation pipeline that constrains explanation length and assesses sufficiency using multiple LLMs on ARC Challenge dataset in both English and Persian.

Result: More concise explanations often remain sufficient, preserving accuracy while substantially reducing length, but excessive compression leads to performance degradation.

Conclusion: There’s an optimal trade-off between explanation sufficiency and conciseness, with compressed explanations maintaining accuracy while reducing computational costs.

Abstract: Large Language Models increasingly rely on self-explanations, such as chain of thought reasoning, to improve performance on multi step question answering. While these explanations enhance accuracy, they are often verbose and costly to generate, raising the question of how much explanation is truly necessary. In this paper, we examine the trade-off between sufficiency, defined as the ability of an explanation to justify the correct answer, and conciseness, defined as the reduction in explanation length. Building on the information bottleneck principle, we conceptualize explanations as compressed representations that retain only the information essential for producing correct answers.To operationalize this view, we introduce an evaluation pipeline that constrains explanation length and assesses sufficiency using multiple language models on the ARC Challenge dataset. To broaden the scope, we conduct experiments in both English, using the original dataset, and Persian, as a resource-limited language through translation. Our experiments show that more concise explanations often remain sufficient, preserving accuracy while substantially reducing explanation length, whereas excessive compression leads to performance degradation.

Method: The paper analyzes state-of-the-art NER algorithms (CRF, BiLSTM-CRF, BERT, FinBERT) on 50,000 annotated payment transactions across multiple formats (SWIFT MT103, ISO 20022, domestic systems). Introduces PaymentBERT, a hybrid architecture combining domain-specific financial embeddings with contextual representations.

Result: Fine-tuned BERT achieves 94.2% F1-score, outperforming CRF by 12.8 percentage points. PaymentBERT achieves state-of-the-art 95.7% F1-score while maintaining real-time processing capabilities. Includes cross-format generalization analysis and ablation studies.

Conclusion: The research provides practical insights for financial institutions implementing automated compliance systems, demonstrating that transformer-based models significantly outperform traditional approaches for payment data extraction.

Abstract: Named Entity Recognition (NER) has emerged as a critical component in automating financial transaction processing, particularly in extracting structured information from unstructured payment data. This paper presents a comprehensive analysis of state-of-the-art NER algorithms specifically designed for payment data extraction, including Conditional Random Fields (CRF), Bidirectional Long Short-Term Memory with CRF (BiLSTM-CRF), and transformer-based models such as BERT and FinBERT. We conduct extensive experiments on a dataset of 50,000 annotated payment transactions across multiple payment formats including SWIFT MT103, ISO 20022, and domestic payment systems. Our experimental results demonstrate that fine-tuned BERT models achieve an F1-score of 94.2% for entity extraction, outperforming traditional CRF-based approaches by 12.8 percentage points. Furthermore, we introduce PaymentBERT, a novel hybrid architecture combining domain-specific financial embeddings with contextual representations, achieving state-of-the-art performance with 95.7% F1-score while maintaining real-time processing capabilities. We provide detailed analysis of cross-format generalization, ablation studies, and deployment considerations. This research provides practical insights for financial institutions implementing automated sanctions screening, anti-money laundering (AML) compliance, and payment processing systems.

Method: Introduces Group Quality Metric (GQM) paradigm instantiated via Group Relative Reward Model (GRRM), which processes entire candidate groups jointly using comparative analysis to resolve relative quality with adaptive granularity. Integrated into GRPO training loop to optimize translation policy.

Result: GRRM achieves competitive ranking accuracy among all baselines. When integrated into GRPO training loop, the framework improves general translation quality and unlocks reasoning capabilities comparable to state-of-the-art reasoning models.

Conclusion: The Group Relative Reward Model paradigm effectively addresses limitations of traditional scalar metrics for ranking in open-ended domains, demonstrating improved performance in machine translation and reasoning capabilities.

Abstract: While Group Relative Policy Optimization (GRPO) offers a powerful framework for LLM post-training, its effectiveness in open-ended domains like Machine Translation hinges on accurate intra-group ranking. We identify that standard Scalar Quality Metrics (SQM) fall short in this context; by evaluating candidates in isolation, they lack the comparative context necessary to distinguish fine-grained linguistic nuances. To address this, we introduce the Group Quality Metric (GQM) paradigm and instantiate it via the Group Relative Reward Model (GRRM). Unlike traditional independent scorers, GRRM processes the entire candidate group jointly, leveraging comparative analysis to rigorously resolve relative quality and adaptive granularity. Empirical evaluations confirm that GRRM achieves competitive ranking accuracy among all baselines. Building on this foundation, we integrate GRRM into the GRPO training loop to optimize the translation policy. Experimental results demonstrate that our framework not only improves general translation quality but also unlocks reasoning capabilities comparable to state-of-the-art reasoning models. We release codes, datasets, and model checkpoints at https://github.com/NJUNLP/GRRM.

Method: Spherical Barycentric Aggregation (SBA) separates radial and angular components to preserve hyperspherical structure while remaining compatible with existing routing mechanisms, maintaining both vector magnitude and direction.

Result: Experiments on MTEB tasks (semantic similarity, clustering, duplicate question detection) show consistent performance improvements with identical training cost and full stability, while geometric analyses confirm SBA prevents aggregation-induced collapse.

Conclusion: Geometry-aware aggregation is crucial for MoE embedding architectures, and SBA provides an effective solution that preserves hyperspherical consistency while improving performance without additional training cost.

Abstract: Mixture-of-Experts (MoE) embedding models combine expert outputs using weighted linear summation, implicitly assuming a linear subspace structure in the embedding space. This assumption is shown to be inconsistent with the geometry of expert representations. Geometric analysis of a modern MoE embedding model reveals that expert outputs lie on a shared hyperspherical manifold characterized by tightly concentrated norms and substantial angular separation. Under this geometry, linear aggregation induces inward collapse toward the manifold interior, distorting vector magnitude and direction and reducing embedding comparability. To address this inconsistency, Spherical Barycentric Aggregation (SBA) is introduced as a geometry-preserving aggregation operator that separates radial and angular components to maintain hyperspherical structure while remaining fully compatible with existing routing mechanisms. Experiments on selected tasks from the Massive Text Embedding Benchmark (MTEB), including semantic similarity, clustering, and duplicate question detection, demonstrate consistent performance improvements with identical training cost and full stability. Additional geometric analyses confirm that SBA prevents aggregation-induced collapse and preserves hyperspherical consistency, highlighting the importance of geometry-aware aggregation in MoE embedding architectures.

Method: Evaluated five open-source LLMs (7B, 32B, 70B parameters from Llama-3.1, Qwen2.5, Qwen3 families) on three fact verification datasets (HOVER, FEVEROUS, ClimateFEVER), testing parametric knowledge and evidence placement effects across varying context lengths.

Result: LLMs exhibit non-trivial parametric factual knowledge; verification accuracy declines as context length increases; evidence placement at beginning or end yields higher accuracy than mid-context placement, consistent with prior findings.

Conclusion: Prompt structure is crucial for retrieval-augmented fact-checking systems, with optimal evidence placement at context boundaries rather than middle positions.

Abstract: Large language models (LLMs) show strong reasoning abilities across diverse tasks, yet their performance on extended contexts remains inconsistent. While prior research has emphasized mid-context degradation in question answering, this study examines the impact of context in LLM-based fact verification. Using three datasets (HOVER, FEVEROUS, and ClimateFEVER) and five open-source models accross different parameters sizes (7B, 32B and 70B parameters) and model families (Llama-3.1, Qwen2.5 and Qwen3), we evaluate both parametric factual knowledge and the impact of evidence placement across varying context lengths. We find that LLMs exhibit non-trivial parametric knowledge of factual claims and that their verification accuracy generally declines as context length increases. Similarly to what has been shown in previous works, in-context evidence placement plays a critical role with accuracy being consistently higher when relevant evidence appears near the beginning or end of the prompt and lower when placed mid-context. These results underscore the importance of prompt structure in retrieval-augmented fact-checking systems.

Method: LogitsCoder uses lightweight logit-level control mechanisms: Logits Preference Decoding to steer token selection toward statistically preferred patterns, Logits Rank Based Path Selection to choose diverse reasoning paths, and Thoughts Aggregation to combine them into coherent reasoning chains.

Result: Extensive experiments show LogitsCoder produces more efficient and higher-quality reasoning chains, leading to superior code generation performance compared to baseline methods.

Conclusion: LogitsCoder effectively balances reasoning depth and efficiency through logit-level control mechanisms, addressing key challenges in code generation reasoning.

Abstract: Code generation remains a challenging task that requires precise and structured reasoning. Existing Test Time Scaling (TTS) methods, including structured tree search, have made progress in exploring reasoning paths but still face two major challenges: (1) underthinking, where reasoning chains tend to be shallow and fail to capture the full complexity of problems; and (2) overthinking, where overly verbose reasoning leads to inefficiency and increased computational costs. To address these issues, we propose LogitsCoder, a novel framework that enhances chain-of-thought reasoning through lightweight, logit-level control mechanisms for code generation. LogitsCoder iteratively generates and refines reasoning steps by first steering token selection toward statistically preferred patterns via Logits Preference Decoding, then selecting and aggregating diverse reasoning paths using Logits Rank Based Path Selection and Thoughts Aggregation. This results in coherent and effective reasoning chains that balance depth and efficiency. Extensive experiments demonstrate that LogitsCoder produces more efficient and higher-quality reasoning chains, leading to superior code generation performance compared to baseline methods.

Method: Uses data clustering to identify semantic domains, trains small language models as domain experts, and employs a sparse mixture-of-experts architecture with semantic-aware domain-level routing to combine expert knowledge.

Result: Achieves +7% BLEU score improvement over prior state-of-the-art on five benchmarks, with clustering providing nearly 10% definition quality improvement and domain-level routing outperforming token-level routing by +1% expert efficacy.

Conclusion: LM-Lexicon advances definition modeling while offering insights for developing efficient language models for semantic-intensive applications through specialized domain expertise.

Abstract: We introduce LM-Lexicon, an innovative definition modeling approach that incorporates data clustering, semantic expert learning, and model merging using a sparse mixture-of-experts architecture. By decomposing the definition modeling task into specialized semantic domains, where small language models are trained as domain experts, LM-Lexicon achieves substantial improvements (+7% BLEU score compared with the prior state-of-the-art model) over existing methods on five widely used benchmarks. Empirically, we demonstrate that 1) the clustering strategy enables fine-grained expert specialization with nearly 10% improvement in definition quality; 2) the semantic-aware domain-level routing mechanism achieves higher expert efficacy (+1%) than conventional token-level routing; and 3) further performance gains can be obtained through test-time compute and semantic expert scaling. Our work advances definition modeling while providing insights into the development of efficient language models for semantic-intensive applications.

Method: OpenRS uses Pairwise Adaptive Meta-Rubrics (PAMR) that instantiate rubrics on-the-fly by conditioning on semantic differences between candidate responses, plus Pointwise Verifiable Rubrics (PVRs) for hard-constraint guardrails. It employs a two-level meta-rubric refinement pipeline (automated evolutionary refinement for general principles and human-in-the-loop for domain principles) and aggregates criterion-level preferences externally.

Result: The framework improves discriminability in open-ended settings by avoiding pointwise weighted scalarization and provides both verifiable reward components and guardrails against degenerate behaviors. It’s instantiated as reward supervision in pairwise RL training.

Conclusion: OpenRS operationalizes the view that robust alignment should be an explicit reasoning process under inspectable principles rather than a learned scalar function, addressing fundamental issues with current reward modeling approaches for non-verifiable tasks.

Abstract: Scalar reward models compress multi-dimensional human preferences into a single opaque score, creating an information bottleneck that often leads to brittleness and reward hacking in open-ended alignment. We argue that robust alignment for non-verifiable tasks is fundamentally a principle generalization problem: reward should not be a learned function internalized into a judge, but an explicit reasoning process executed under inspectable principles. To operationalize this view, we present the Open Rubric System (OpenRS), a plug-and-play, rubrics-based LLM-as-a-Judge framework built around Pairwise Adaptive Meta-Rubrics (PAMR) and lightweight Pointwise Verifiable Rubrics (PVRs), which provide both hard-constraint guardrails and verifiable reward components when ground-truth or programmatic checks are available. OpenRS uses an explicit meta-rubric – a constitution-like specification that governs how rubrics are instantiated, weighted, and enforced – and instantiates adaptive rubrics on the fly by conditioning on the semantic differences between two candidate responses. It then performs criterion-wise pairwise comparisons and aggregates criterion-level preferences externally, avoiding pointwise weighted scalarization while improving discriminability in open-ended settings. To keep principles consistent yet editable across various domains, we introduce a two-level meta-rubric refinement pipeline (automated evolutionary refinement for general principles and a reproducible human-in-the-loop procedure for domain principles), complemented with pointwise verifiable rubrics that act as both guardrails against degenerate behaviors and a source of verifiable reward for objective sub-tasks. Finally, we instantiate OpenRS as reward supervision in pairwise RL training.

Method: Reproduce and adapt LLaVA-Next methodology to create Polish VLMs using fully automated pipeline for translating/filtering existing multimodal datasets, complemented with synthetic Polish data for OCR and culturally specific tasks.

Result: +9.5% improvement over LLaVA-1.6-Vicuna-13B on Polish-adapted MMBench, higher-quality captions in generative evaluations as measured by human annotators for linguistic correctness.

Conclusion: Large-scale automated translation with lightweight filtering can effectively bootstrap high-quality multimodal models for low-resource languages, though challenges remain in cultural coverage and evaluation.

Abstract: Most vision-language models (VLMs) are trained on English-centric data, limiting their performance in other languages and cultural contexts. This restricts their usability for non-English-speaking users and hinders the development of multimodal systems that reflect diverse linguistic and cultural realities. In this work, we reproduce and adapt the LLaVA-Next methodology to create a set of Polish VLMs. We rely on a fully automated pipeline for translating and filtering existing multimodal datasets, and complement this with synthetic Polish data for OCR and culturally specific tasks. Despite relying almost entirely on automatic translation and minimal manual intervention to the training data, our approach yields strong results: we observe a +9.5% improvement over LLaVA-1.6-Vicuna-13B on a Polish-adapted MMBench, along with higher-quality captions in generative evaluations, as measured by human annotators in terms of linguistic correctness. These findings highlight that large-scale automated translation, combined with lightweight filtering, can effectively bootstrap high-quality multimodal models for low-resource languages. Some challenges remain, particularly in cultural coverage and evaluation. To facilitate further research, we make our models and evaluation dataset publicly available.

Method: Proposes Gaussian Thought Sampler (GTS) that models latent thought exploration as conditional sampling from learnable densities. GTS predicts context-dependent perturbation distributions over continuous reasoning states and is trained with GRPO-style policy optimization while keeping the backbone model frozen.

Result: Experiments on GSM8K with two latent reasoning architectures show that GTS achieves more reliable inference-time scaling than heuristic baselines like dropout or fixed Gaussian noise.

Conclusion: Improving latent inference-time scaling requires structured and optimizable exploration mechanisms rather than simply amplifying stochasticity. GTS provides a more effective alternative to heuristic perturbation methods.

Abstract: Inference-time scaling (ITS) in latent reasoning models typically introduces stochasticity through heuristic perturbations, such as dropout or fixed Gaussian noise. While these methods increase trajectory diversity, their exploration behavior is not explicitly modeled and can be inefficient under finite sampling budgets. We observe that stronger perturbations do not necessarily translate into more effective candidate trajectories, as unguided noise may disrupt internal decision structure rather than steer it. To provide a more structured alternative, we model latent thought exploration as conditional sampling from learnable densities and instantiate this idea as a Gaussian Thought Sampler (GTS). GTS predicts context-dependent perturbation distributions over continuous reasoning states and is trained with GRPO-style policy optimization while keeping the backbone frozen. Experiments on GSM8K with two latent reasoning architectures show that GTS achieves more reliable inference-time scaling than heuristic baselines. These findings indicate that improving latent ITS requires structured and optimizable exploration mechanisms rather than simply amplifying stochasticity.

Method: Proposed behavioral framework characterizing each fact by whether it’s encoded and how accessible: cannot be recalled, directly recalled, or recalled with inference-time computation (thinking). Introduced WikiProfile benchmark via automated pipeline with prompted LLM grounded in web search.

Result: Across 4M responses from 13 LLMs: encoding nearly saturated in frontier models (GPT-5 and Gemini-3 encode 95-98% of facts), but recall remains major bottleneck. Many errors previously attributed to missing knowledge stem from accessibility failures. Failures systematic, disproportionately affect long-tail facts and reverse questions. Thinking improves recall and recovers substantial fraction of failures.

Conclusion: Future gains may rely less on scaling and more on methods that improve how models utilize what they already encode, as recall remains the primary bottleneck despite high encoding rates.

Abstract: Standard factuality evaluations of LLMs treat all errors alike, obscuring whether failures arise from missing knowledge (empty shelves) or from limited access to encoded facts (lost keys). We propose a behavioral framework that profiles factual knowledge at the level of facts rather than questions, characterizing each fact by whether it is encoded, and then by how accessible it is: cannot be recalled, can be directly recalled, or can only be recalled with inference-time computation (thinking). To support such profiling, we introduce WikiProfile, a new benchmark constructed via an automated pipeline with a prompted LLM grounded in web search. Across 4 million responses from 13 LLMs, we find that encoding is nearly saturated in frontier models on our benchmark, with GPT-5 and Gemini-3 encoding 95–98% of facts. However, recall remains a major bottleneck: many errors previously attributed to missing knowledge instead stem from failures to access it. These failures are systematic and disproportionately affect long-tail facts and reverse questions. Finally, we show that thinking improves recall and can recover a substantial fraction of failures, indicating that future gains may rely less on scaling and more on methods that improve how models utilize what they already encode.

Method: DeepRare uses a multi-agent system powered by large language models, integrating over 40 specialized tools and up-to-date knowledge sources. It processes heterogeneous clinical inputs including free-text descriptions, Human Phenotype Ontology terms, and genetic testing results to generate ranked diagnostic hypotheses with transparent reasoning linked to verifiable medical evidence.

Result: Evaluated across 9 datasets from literature, case reports, and clinical centers spanning 14 medical specialties and 3,134 diseases, DeepRare achieved: 57.18% average Recall@1 in HPO-based tasks (outperforming next-best by 23.79%), 69.1% in multi-modal tests vs. Exomiser’s 55.9% on 168 cases, and 95.4% expert agreement on reasoning chain validity.

Conclusion: DeepRare advances rare disease diagnosis and demonstrates how LLM-driven agentic systems can reshape clinical workflows by providing accurate, transparent diagnostic support with verifiable evidence.

Abstract: Rare diseases affect over 300 million individuals worldwide, yet timely and accurate diagnosis remains an urgent challenge. Patients often endure a prolonged diagnostic odyssey exceeding five years, marked by repeated referrals, misdiagnoses, and unnecessary interventions, leading to delayed treatment and substantial emotional and economic burdens. Here we present DeepRare, a multi-agent system for rare disease differential diagnosis decision support powered by large language models, integrating over 40 specialized tools and up-to-date knowledge sources. DeepRare processes heterogeneous clinical inputs, including free-text descriptions, structured Human Phenotype Ontology terms, and genetic testing results, to generate ranked diagnostic hypotheses with transparent reasoning linked to verifiable medical evidence. Evaluated across nine datasets from literature, case reports and clinical centres across Asia, North America and Europe spanning 14 medical specialties, DeepRare demonstrates exceptional performance on 3,134 diseases. In human-phenotype-ontology-based tasks, it achieves an average Recall@1 of 57.18%, outperforming the next-best method by 23.79%; in multi-modal tests, it reaches 69.1% compared with Exomiser’s 55.9% on 168 cases. Expert review achieved 95.4% agreement on its reasoning chains, confirming their validity and traceability. Our work not only advances rare disease diagnosis but also demonstrates how the latest powerful large-language-model-driven agentic systems can reshape current clinical workflows.

Method: Introduces CCiV benchmark to assess LLM-generated Ci poetry across structure, rhythm, and quality dimensions. Evaluates 17 LLMs on 30 Cipai (poetic forms) and analyzes phenomena like unexpected historical variants and tonal pattern difficulties.

Result: Two critical findings: 1) Models frequently generate valid but unexpected historical variants of poetic forms, 2) Adherence to tonal patterns is substantially harder than structural rules. Form-aware prompting improves structural/tonal control for stronger models but degrades weaker ones. Weak alignment between formal correctness and literary quality.

Conclusion: CCiV highlights need for variant-aware evaluation and more holistic constrained creative generation methods for classical poetry generation.

Abstract: The generation of classical Chinese \textit{Ci} poetry, a form demanding a sophisticated blend of structural rigidity, rhythmic harmony, and artistic quality, poses a significant challenge for large language models (LLMs). To systematically evaluate and advance this capability, we introduce \textbf{C}hinese \textbf{Ci}pai \textbf{V}ariants (\textbf{CCiV}), a benchmark designed to assess LLM-generated \textit{Ci} poetry across these three dimensions: structure, rhythm, and quality. Our evaluation of 17 LLMs on 30 \textit{Cipai} reveals two critical phenomena: models frequently generate valid but unexpected historical variants of a poetic form, and adherence to tonal patterns is substantially harder than structural rules. We further show that form-aware prompting can improve structural and tonal control for stronger models, while potentially degrading weaker ones. Finally, we observe weak and inconsistent alignment between formal correctness and literary quality in our sample. CCiV highlights the need for variant-aware evaluation and more holistic constrained creative generation methods.

Method: Compared five encoder-decoder transformer models varying along two dimensions: sequential vs. position-invariant positional encoding, and atomic vs. decomposed tag representations. Tested on Spanish L-shaped morphome where only first-person singular indicative shares its stem with all subjunctive forms.

Result: Position-invariant models recover correct L-shaped paradigm clustering even with scarce training data, while sequential models only partially capture the pattern. However, none productively generalize the pattern to novel forms like humans do. Position-invariant models generalize L-shaped stem across subjunctive cells but fail to extend to first-person singular indicative, producing mood-based generalization rather than the L-shaped morphomic pattern.

Conclusion: Neural networks can reproduce statistical patterns but fail to abstract morphological generalizations like humans, highlighting a gap between pattern reproduction and true morphological abstraction. Positional encoding proves decisive for pattern recovery but not for human-like generalization.

Abstract: Whether neural networks can serve as cognitive models of morphological learning remains an open question. Recent work has shown that encoder-decoder models can acquire irregular patterns, but evidence that they generalize these patterns like humans is mixed. We investigate this using the Spanish \emph{L-shaped morphome}, where only the first-person singular indicative (e.g., \textit{pongo} `I put’) shares its stem with all subjunctive forms (e.g., \textit{ponga, pongas}) despite lacking apparent phonological, semantic, or syntactic motivation. We compare five encoder-decoder transformers varying along two dimensions: sequential vs. position-invariant positional encoding, and atomic vs. decomposed tag representations. Positional encoding proves decisive: position-invariant models recover the correct L-shaped paradigm clustering even when L-shaped verbs are scarce in training, whereas sequential positional encoding models only partially capture the pattern. Yet none of the models productively generalize this pattern to novel forms. Position-invariant models generalize the L-shaped stem across subjunctive cells but fail to extend it to the first-person singular indicative, producing a mood-based generalization rather than the L-shaped morphomic pattern. Humans do the opposite, generalizing preferentially to the first-person singular indicative over subjunctive forms. None of the models reproduce the human pattern, highlighting the gap between statistical pattern reproduction and morphological abstraction.

Method: Deferred Visual Ingestion (DVI) - lightweight metadata extraction during indexing, deferring visual understanding to query time. Uses structured metadata indexes and BM25 search for page localization, then sends original images with questions to VLM.

Result: Comparable accuracy (46.7% vs 48.9%) at zero ingestion VLM cost, 50% effectiveness on visually necessary queries (vs 0% for pre-ingestion), 100% page localization with 98% search space compression.

Conclusion: DVI transforms QA accuracy problem into page localization problem, supports interactive refinement and progressive caching, offering cost-effective alternative to pre-ingestion approaches.

Abstract: Existing multimodal document question answering methods universally adopt a supply-side ingestion strategy: running a Vision-Language Model (VLM) on every page during indexing to generate comprehensive descriptions, then answering questions through text retrieval. However, this “pre-ingestion” approach is costly (a 113-page engineering drawing package requires approximately 80,000 VLM tokens), end-to-end unreliable (VLM outputs may fail to be correctly retrieved due to format mismatches in the retrieval infrastructure), and irrecoverable once it fails. This paper proposes the Deferred Visual Ingestion (DVI) framework, adopting a demand-side ingestion strategy: the indexing phase performs only lightweight metadata extraction, deferring visual understanding to the moment users pose specific questions. DVI’s core principle is “Index for locating, not understanding”–achieving page localization through structured metadata indexes and BM25 full-text search, then sending original images along with specific questions to a VLM for targeted analysis. Experiments on two real industrial engineering drawings (113 pages + 7 pages) demonstrate that DVI achieves comparable overall accuracy at zero ingestion VLM cost (46.7% vs. 48.9%), an effectiveness rate of 50% on visually necessary queries (vs. 0% for pre-ingestion), and 100% page localization (98% search space compression). DVI also supports interactive refinement and progressive caching, transforming the “QA accuracy” problem into a “page localization” problem–once the correct drawing page is found, obtaining the answer becomes a matter of interaction rounds.

Method: Evaluated four state-of-the-art models (Grok-4, GPT-4, Gemini 2.5, and GPT-5) on long short-context tasks using three datasets: two supplementary datasets for retrieving culinary recipes and math problems, and a primary dataset of 20K social media posts for depression detection.

Result: Performance of all models degrades significantly when input exceeds 5K posts (70K tokens), with accuracy dropping to 50-53% for 20K posts. GPT-5 maintained high precision (~95%) despite accuracy decline. The “lost in the middle” problem appears largely resolved in newer models.

Conclusion: There’s a significant gap between theoretical capacity and actual performance on complex, high-volume data tasks. Metrics beyond simple accuracy (like precision) are important for practical applications, especially sensitive ones like depression detection.

Abstract: With the significant expansion of the context window in Large Language Models (LLMs), these models are theoretically capable of processing millions of tokens in a single pass. However, research indicates a significant gap between this theoretical capacity and the practical ability of models to robustly utilize information within long contexts, especially in tasks that require a comprehensive understanding of numerous details. This paper evaluates the performance of four state-of-the-art models (Grok-4, GPT-4, Gemini 2.5, and GPT-5) on long short-context tasks. For this purpose, three datasets were used: two supplementary datasets for retrieving culinary recipes and math problems, and a primary dataset of 20K social media posts for depression detection. The results show that as the input volume on the social media dataset exceeds 5K posts (70K tokens), the performance of all models degrades significantly, with accuracy dropping to around 50-53% for 20K posts. Notably, in the GPT-5 model, despite the sharp decline in accuracy, its precision remained high at approximately 95%, a feature that could be highly effective for sensitive applications like depression detection. This research also indicates that the “lost in the middle” problem has been largely resolved in newer models. This study emphasizes the gap between the theoretical capacity and the actual performance of models on complex, high-volume data tasks and highlights the importance of metrics beyond simple accuracy for practical applications.

Method: Abstention-aware verification framework that: 1) decomposes scientific claims into minimal conditions, 2) audits each condition against available evidence using natural language inference (NLI), 3) selectively decides whether to support, refute, or abstain based on evidence sufficiency.

Result: Across SciFact and PubMedQA benchmarks with six diverse language models, raw accuracy varies modestly across architectures, but abstention plays critical role in controlling error. Confidence-based abstention substantially reduces risk at moderate coverage levels, even when absolute accuracy improvements are limited.

Conclusion: In scientific reasoning tasks, primary challenge is not selecting single best model, but determining when available evidence is sufficient to justify an answer. Abstention-aware evaluation provides practical, model-agnostic lens for assessing scientific reliability.

Abstract: Large language models are increasingly used to answer and verify scientific claims, yet existing evaluations typically assume that a model must always produce a definitive answer. In scientific settings, however, unsupported or uncertain conclusions can be more harmful than abstaining. We study this problem through an abstention-aware verification framework that decomposes scientific claims into minimal conditions, audits each condition against available evidence using natural language inference (NLI), and selectively decides whether to support, refute, or abstain. We evaluate this framework across two complementary scientific benchmarks: SciFact and PubMedQA, covering both closed-book and open-domain evidence settings. Experiments are conducted with six diverse language models, including encoder-decoder, open-weight chat models, and proprietary APIs. Across all benchmarks and models, we observe that raw accuracy varies only modestly across architectures, while abstention plays a critical role in controlling error. In particular, confidence-based abstention substantially reduces risk at moderate coverage levels, even when absolute accuracy improvements are limited. Our results suggest that in scientific reasoning tasks, the primary challenge is not selecting a single best model, but rather determining when available evidence is sufficient to justify an answer. This work highlights abstention-aware evaluation as a practical and model-agnostic lens for assessing scientific reliability, and provides a unified experimental basis for future work on selective reasoning in scientific domains. Code is available at https://github.com/sabdaljalil2000/ai4science .

Method: Develops a new parsing algorithm with syntactic rules/features based on CFG and GPSG foundations, capable of generating both dependency and constituency parse trees, accommodating noise/incomplete parses, and providing multiple parse hypotheses for reranking.

Result: Achieved average Unlabeled Attachment Score (UAS) of 54.5% on development dataset (7 corpora) and 53.8% on test set (12 corpora) from Universal Dependencies, with the system providing multiple parse hypotheses for potential accuracy improvement.

Conclusion: The approach successfully integrates theoretical syntactic work since the 1950s into computational parsing, creating a transparent and interpretable NLP model that handles real-world language challenges while maintaining theoretical grounding.

Abstract: This research introduces a new parsing approach, based on earlier syntactic work on context free grammar (CFG) and generalized phrase structure grammar (GPSG). The approach comprises both a new parsing algorithm and a set of syntactic rules and features that overcome the limitations of CFG. It also generates both dependency and constituency parse trees, while accommodating noise and incomplete parses. The system was tested on data from Universal Dependencies, showing a promising average Unlabeled Attachment Score (UAS) of 54.5% in the development dataset (7 corpora) and 53.8% in the test set (12 corpora). The system also provides multiple parse hypotheses, allowing further reranking to improve parsing accuracy. This approach also leverages much of the theoretical syntactic work since the 1950s to be used within a computational context. The application of this approach provides a transparent and interpretable NLP model to process language input.

Method: Constructed benchmark from real user marketing analysis requests with domain experts providing verifiable reference answers and tool-call trajectories. Categorized requests into three difficulty levels (L1-L3) to evaluate multi-round, multi-tool collaboration capabilities.

Result: Gemini-3-Pro achieved Pass@1 = 68.0% and Pass@3 = 83.0% overall, but performance dropped significantly on L3 to Pass@1 = 49.4% and Pass@3 = 62.1% with trajectory coverage of 70.1%, showing substantial capability gaps in complex scenarios.

Conclusion: AD-Bench provides a realistic benchmark for evaluating and improving advertising marketing agents, revealing that even state-of-the-art models have significant limitations in complex domain-specific reasoning tasks.

Abstract: While Large Language Model (LLM) agents have achieved remarkable progress in complex reasoning tasks, evaluating their performance in real-world environments has become a critical problem. Current benchmarks, however, are largely restricted to idealized simulations, failing to address the practical demands of specialized domains like advertising and marketing analytics. In these fields, tasks are inherently more complex, often requiring multi-round interaction with professional marketing tools. To address this gap, we propose AD-Bench, a benchmark designed based on real-world business requirements of advertising and marketing platforms. AD-Bench is constructed from real user marketing analysis requests, with domain experts providing verifiable reference answers and corresponding reference tool-call trajectories. The benchmark categorizes requests into three difficulty levels (L1-L3) to evaluate agents’ capabilities under multi-round, multi-tool collaboration. Experiments show that on AD-Bench, Gemini-3-Pro achieves Pass@1 = 68.0% and Pass@3 = 83.0%, but performance drops significantly on L3 to Pass@1 = 49.4% and Pass@3 = 62.1%, with a trajectory coverage of 70.1%, indicating that even state-of-the-art models still exhibit substantial capability gaps in complex advertising and marketing analysis scenarios. AD-Bench provides a realistic benchmark for evaluating and improving advertising marketing agents, the leaderboard and code can be found at https://github.com/Emanual20/adbench-leaderboard.

Method: Analyzed static embedding spaces of 11 transformer models (BERT, RoBERTa, ELECTRA, DeBERTa, ALBERT, MiniLM, DistilBERT, GPT-2) using three geometric statistics: α (polarity coupling), β (cluster cohesion), and λ_s (radial information gradient) to identify hallucination types.

Result: Identified three hallucination types: Type 1 (center-drift), Type 2 (wrong-well convergence), Type 3 (coverage gaps). Found polarity structure universal (11/11), cluster cohesion universal (11/11), radial information gradient significant in 9/11 models. ALBERT and MiniLM exceptions explained by architectural factors.

Conclusion: Established geometric prerequisites for type-specific hallucination detection and architecture-dependent vulnerability profiles, providing testable predictions for hallucination analysis.

Abstract: We propose a geometric taxonomy of large language model hallucinations based on observable signatures in token embedding cluster structure. By analyzing the static embedding spaces of 11 transformer models spanning encoder (BERT, RoBERTa, ELECTRA, DeBERTa, ALBERT, MiniLM, DistilBERT) and decoder (GPT-2) architectures, we identify three operationally distinct hallucination types: Type 1 (center-drift) under weak context, Type 2 (wrong-well convergence) to locally coherent but contextually incorrect cluster regions, and Type 3 (coverage gaps) where no cluster structure exists. We introduce three measurable geometric statistics: α (polarity coupling), \b{eta} (cluster cohesion), and λ_s (radial information gradient). Across all 11 models, polarity structure (α > 0.5) is universal (11/11), cluster cohesion (\b{eta} > 0) is universal (11/11), and the radial information gradient is significant (9/11, p < 0.05). We demonstrate that the two models failing λ_s significance – ALBERT and MiniLM – do so for architecturally explicable reasons: factorized embedding compression and distillation-induced isotropy, respectively. These findings establish the geometric prerequisites for type-specific hallucination detection and yield testable predictions about architecture-dependent vulnerability profiles.

Method: STATe uses a structured three-component approach: a controller selects interpretable actions encoding high-level reasoning choices, a generator produces reasoning steps conditioned on those choices, and an evaluator scores candidates to guide search through the action space.

Result: STATe produces greater response diversity than temperature-based sampling, captures interpretable features predictive of output quality, and enables steering generation toward promising unexplored regions of the action space.

Conclusion: STATe establishes a practical framework for generating high-quality, diverse, and interpretable text through structured, action-guided reasoning patterns.

Abstract: Inference-Time-Compute (ITC) methods like Best-of-N and Tree-of-Thoughts are meant to produce output candidates that are both high-quality and diverse, but their use of high-temperature sampling often fails to achieve meaningful output diversity. Moreover, existing ITC methods offer limited control over how to perform reasoning, which in turn limits their explainability. We present STATe-of-Thoughts (STATe), an interpretable ITC method that searches over high-level reasoning patterns. STATe replaces stochastic sampling with discrete and interpretable textual interventions: a controller selects actions encoding high-level reasoning choices, a generator produces reasoning steps conditioned on those choices, and an evaluator scores candidates to guide search. This structured approach yields three main advantages. First, action-guided textual interventions produce greater response diversity than temperature-based sampling. Second, in a case study on argument generation, STATe’s explicit action sequences capture interpretable features that are highly predictive of output quality. Third, estimating the association between performance and action choices allows us to identify promising yet unexplored regions of the action space and steer generation directly toward them. Together, these results establish STATe as a practical framework for generating high-quality, diverse, and interpretable text. Our framework is available at https://github.com/zbambergerNLP/state-of-thoughts.

Method: Large-scale systemic diagnosis using Moltbook as a testbed, introducing quantitative diagnostic framework measuring semantic stabilization, lexical turnover, individual inertia, influence persistence, and collective consensus in dynamic evolution of AI agent societies.

Result: AI agent society shows dynamic balance: global semantic averages stabilize rapidly but individual agents retain high diversity with persistent lexical turnover. Agents exhibit strong individual inertia and minimal adaptive response to partners, preventing mutual influence and consensus. Influence remains transient with no persistent supernodes, and society fails to develop stable collective influence anchors due to absence of shared social memory.

Conclusion: Scale and interaction density alone are insufficient to induce socialization in AI agent societies, providing actionable design and analysis principles for next-generation AI agent societies that need mechanisms for shared social memory and mutual influence.

Abstract: As large language model agents increasingly populate networked environments, a fundamental question arises: do artificial intelligence (AI) agent societies undergo convergence dynamics similar to human social systems? Lately, Moltbook approximates a plausible future scenario in which autonomous agents participate in an open-ended, continuously evolving online society. We present the first large-scale systemic diagnosis of this AI agent society. Beyond static observation, we introduce a quantitative diagnostic framework for dynamic evolution in AI agent societies, measuring semantic stabilization, lexical turnover, individual inertia, influence persistence, and collective consensus. Our analysis reveals a system in dynamic balance in Moltbook: while global semantic averages stabilize rapidly, individual agents retain high diversity and persistent lexical turnover, defying homogenization. However, agents exhibit strong individual inertia and minimal adaptive response to interaction partners, preventing mutual influence and consensus. Consequently, influence remains transient with no persistent supernodes, and the society fails to develop stable collective influence anchors due to the absence of shared social memory. These findings demonstrate that scale and interaction density alone are insufficient to induce socialization, providing actionable design and analysis principles for upcoming next-generation AI agent societies.

Method: InnoEval treats idea evaluation as a knowledge-grounded, multi-perspective reasoning problem. It uses a heterogeneous deep knowledge search engine to retrieve dynamic evidence from diverse online sources, and implements an innovation review board with reviewers from distinct academic backgrounds for multi-dimensional decoupled evaluation across multiple metrics.

Result: Experiments on comprehensive datasets from authoritative peer-reviewed submissions show InnoEval consistently outperforms baselines in point-wise, pair-wise, and group-wise evaluation tasks, with judgment patterns and consensus highly aligned with human experts.

Conclusion: InnoEval successfully emulates human-level idea assessment by addressing key limitations of existing methods through knowledge grounding and multi-perspective reasoning, demonstrating superior performance and alignment with expert human judgment.

Abstract: The rapid evolution of Large Language Models has catalyzed a surge in scientific idea production, yet this leap has not been accompanied by a matching advance in idea evaluation. The fundamental nature of scientific evaluation needs knowledgeable grounding, collective deliberation, and multi-criteria decision-making. However, existing idea evaluation methods often suffer from narrow knowledge horizons, flattened evaluation dimensions, and the inherent bias in LLM-as-a-Judge. To address these, we regard idea evaluation as a knowledge-grounded, multi-perspective reasoning problem and introduce InnoEval, a deep innovation evaluation framework designed to emulate human-level idea assessment. We apply a heterogeneous deep knowledge search engine that retrieves and grounds dynamic evidence from diverse online sources. We further achieve review consensus with an innovation review board containing reviewers with distinct academic backgrounds, enabling a multi-dimensional decoupled evaluation across multiple metrics. We construct comprehensive datasets derived from authoritative peer-reviewed submissions to benchmark InnoEval. Experiments demonstrate that InnoEval can consistently outperform baselines in point-wise, pair-wise, and group-wise evaluation tasks, exhibiting judgment patterns and consensus highly aligned with human experts.

Method: Multi-token Policy Gradient Optimization (MPO) treats sequences of K consecutive tokens as unified semantic actions, enabling block-level optimization that captures compositional reasoning structure.

Result: Experiments on mathematical reasoning and coding benchmarks show MPO outperforms standard token-level policy gradient baselines.

Conclusion: Token-level policy gradients have limitations for complex reasoning, motivating future research to look beyond token-level granularity for reasoning-intensive language tasks.

Abstract: Existing policy-gradient methods for auto-regressive language models typically select subsequent tokens one at a time as actions in the policy. While effective for many generation tasks, such an approach may not fully capture the structure of complex reasoning tasks, where a single semantic decision is often realized across multiple tokens–for example, when defining variables or composing equations. This introduces a potential mismatch between token-level optimization and the inherently block-level nature of reasoning in these settings. To bridge this gap, we propose Multi-token Policy Gradient Optimization (MPO), a framework that treats sequences of K consecutive tokens as unified semantic actions. This block-level perspective enables our method to capture the compositional structure of reasoning trajectories and supports optimization over coherent, higher-level objectives. Experiments on mathematical reasoning and coding benchmarks show that MPO outperforms standard token-level policy gradient baselines, highlight the limitations of token-level policy gradients for complex reasoning, motivating future research to look beyond token-level granularity for reasoning-intensive language tasks.

Method: Created TruthStance dataset with 24,378 posts and 523,360 comments from Truth Social (2023-2025), preserved reply-tree structure. Provided human-annotated benchmark of 1,500 instances for argument mining and claim-based stance detection. Evaluated LLM prompting strategies and used best-performing configuration to generate additional labels for argument presence and stance detection.

Result: Released comprehensive dataset with human-annotated benchmark and LLM-generated labels for 24,352 posts (argument presence) and 107,873 comments (stance to parent). Enables analysis of stance and argumentation patterns across depth, topics, and users in alt-tech platform discourse.

Conclusion: TruthStance fills the gap in studying conversational structure on alt-tech platforms, providing valuable resources for argument mining and stance detection research, with potential applications in understanding opinion formation in diverse online ecosystems.

Abstract: Argument mining and stance detection are central to understanding how opinions are formed and contested in online discourse. However, most publicly available resources focus on mainstream platforms such as Twitter and Reddit, leaving conversational structure on alt-tech platforms comparatively under-studied. We introduce TruthStance, a large-scale dataset of Truth Social conversation threads spanning 2023-2025, consisting of 24,378 posts and 523,360 comments with reply-tree structure preserved. We provide a human-annotated benchmark of 1,500 instances across argument mining and claim-based stance detection, including inter-annotator agreement, and use it to evaluate large language model (LLM) prompting strategies. Using the best-performing configuration, we release additional LLM-generated labels for 24,352 posts (argument presence) and 107,873 comments (stance to parent), enabling analysis of stance and argumentation patterns across depth, topics, and users. All code and data are released publicly.

Method: WavePhaseNet uses Discrete Fourier Transform (DFT) along sequence dimension to decompose semantic information into frequency bands, creates Semantic Conceptual Hierarchy Structure (SCHS), reduces embedding dimensions via cumulative energy analysis, and applies cohomological consistency control with harmonic projection based on Hodge theory.

Result: Theoretical demonstration that hallucination is structural, dimensionality reduction from 24,576 to ~3,000 dimensions preserves meaning while enabling rigorous reasoning, and cohomological consistency control extracts maximally consistent global representations.

Conclusion: Hallucination in LLMs is fundamentally structural, but can be mitigated through frequency-based semantic decomposition, dimensionality reduction, and cohomological consistency control, providing a theoretical framework for more reliable language models.

Abstract: This paper reformulates Transformer/Attention mechanisms in Large Language Models (LLMs) through measure theory and frequency analysis, theoretically demonstrating that hallucination is an inevitable structural limitation. The embedding space functions as a conditional expectation over a σ-algebra, and its failure to be isomorphic to the semantic truth set fundamentally causes logical consistency breakdown. WavePhaseNet Method The authors propose WavePhaseNet, which explicitly constructs a Semantic Conceptual Hierarchy Structure (SCHS) using Discrete Fourier Transform (DFT). By applying DFT along the sequence dimension, semantic information is decomposed into frequency bands: low-frequency components capture global meaning and intent, while high-frequency components represent local syntax and expression. This staged separation enables precise semantic manipulation in diagonalized space. Dimensionality Reduction GPT-4’s 24,576-dimensional embedding space exhibits a 1/f spectral structure based on language self-similarity and Zipf’s law. Through cumulative energy analysis, the authors derive that approximately 3,000 dimensions constitute the lower bound for “complete representation.” This demonstrates that reduction from 24,576 to 3,000 dimensions preserves meaning and intent while enabling rigorous reasoning and suppressing hallucination. Cohomological Consistency Control The reduced embedding space, constructed via cohomological regularization over overlapping local windows, allows defining a graph structure and cochain complex. This quantifies inconsistencies among local inferences as coboundary-based losses. Applying harmonic projection based on Hodge theory positions cohomology as a computable regularization principle for controlling semantic consistency, extracting maximally consistent global representations.

Method: Proposes an LLM-assisted distillation framework with two teachers: 1) a conventional high-capacity temporal teacher model, and 2) a large language model as auxiliary instructor. Uses joint optimization of supervised and distillation objectives with staged alignment strategy to progressively integrate guidance from both teachers.

Result: Extensive experiments on multiple public TKG benchmarks with diverse backbone architectures show consistent improvement in link prediction performance over strong distillation baselines while maintaining compact and efficient student models.

Conclusion: Demonstrates the potential of large language models as effective teachers for transferring temporal reasoning capability to resource-efficient TKG systems, enabling lightweight models to better understand event dynamics without increasing inference-time complexity.

Abstract: Temporal knowledge graphs (TKGs) support reasoning over time-evolving facts, yet state-of-the-art models are often computationally heavy and costly to deploy. Existing compression and distillation techniques are largely designed for static graphs; directly applying them to temporal settings may overlook time-dependent interactions and lead to performance degradation. We propose an LLM-assisted distillation framework specifically designed for temporal knowledge graph reasoning. Beyond a conventional high-capacity temporal teacher, we incorporate a large language model as an auxiliary instructor to provide enriched supervision. The LLM supplies broad background knowledge and temporally informed signals, enabling a lightweight student to better model event dynamics without increasing inference-time complexity. Training is conducted by jointly optimizing supervised and distillation objectives, using a staged alignment strategy to progressively integrate guidance from both teachers. Extensive experiments on multiple public TKG benchmarks with diverse backbone architectures demonstrate that the proposed approach consistently improves link prediction performance over strong distillation baselines, while maintaining a compact and efficient student model. The results highlight the potential of large language models as effective teachers for transferring temporal reasoning capability to resource-efficient TKG systems.

Method: Four-phase development: template categorization, culturally aware translation, new template construction, and prompt generation. Robust evaluation protocol uses multiple seeds to account for response instability, averaging bias scores across runs.

Result: Created FilBBQ with 10,000+ prompts, confirmed bias score variability across seeds, and detected sexist/homophobic biases related to emotion, domesticity, queer stereotypes, and polygamy in Filipino-trained models.

Conclusion: FilBBQ provides a culturally relevant bias benchmark for Filipino language models with improved reliability, revealing significant biases that need addressing in multilingual AI development.

Abstract: With natural language generation becoming a popular use case for language models, the Bias Benchmark for Question-Answering (BBQ) has grown to be an important benchmark format for evaluating stereotypical associations exhibited by generative models. We expand the linguistic scope of BBQ and construct FilBBQ through a four-phase development process consisting of template categorization, culturally aware translation, new template construction, and prompt generation. These processes resulted in a bias test composed of more than 10,000 prompts which assess whether models demonstrate sexist and homophobic prejudices relevant to the Philippine context. We then apply FilBBQ on models trained in Filipino but do so with a robust evaluation protocol that improves upon the reliability and accuracy of previous BBQ implementations. Specifically, we account for models’ response instability by obtaining prompt responses across multiple seeds and averaging the bias scores calculated from these distinctly seeded runs. Our results confirm both the variability of bias scores across different seeds and the presence of sexist and homophobic biases relating to emotion, domesticity, stereotyped queer interests, and polygamy. FilBBQ is available via GitHub.

Method: Introduces three-level measurement hierarchy (lexical, entropic, probabilistic anchoring) to formalize anchoring. Proposes Structural Skeleton-guided Reasoning (SSR): two-phase approach that first generates answer-invariant functional skeleton structure, then uses skeleton to guide full trace generation. Also introduces Distilled SSR (SSR-D) which fine-tunes models on teacher-generated SSR traces.

Result: SSR consistently reduces anchoring across all three measurement levels. SSR-D achieves up to 10% improvement over semantic suppression baselines while preserving out-of-distribution generalization across open-ended reasoning benchmarks.

Conclusion: SSR effectively breaks the cycle of answer anchoring in reverse CoT generation by redirecting information flow to structural planning rather than answer monitoring, producing more authentic reasoning traces that are less biased by known answers.

Abstract: Reverse Chain-of-Thought Generation (RCG) synthesizes reasoning traces from query-answer pairs, but runs the risk of producing post-hoc rationalizations: when models can see the answer during generation, the answer serves as a cognitive anchor that shapes the entire explanation. We formalize this phenomenon through a three-level measurement hierarchy: lexical, entropic, and probabilistic anchoring, each captures surface artifacts, entropy dynamics, and latent answer dependence, respectively. We analyze semantic suppression, the intuitive mitigation strategy that instructs models to ignore the answer, to find out its counterproduction: while it reduces lexical overlap, it paradoxically increases entropic and probabilistic anchoring. Drawing on Ironic Process Theory from cognitive psychology, we attribute this failure to active monitoring of the forbidden answer, which inadvertently deepens dependence on it. To break this cycle, we propose Structural Skeleton-guided Reasoning (SSR), a two-phase approach that first generates an answer-invariant functional skeleton structure, then uses this skeleton to guide full trace generation. By redirecting the information flow to structural planning rather than answer monitoring, SSR consistently reduces anchoring across all three levels. We further introduce Distilled SSR (SSR-D), which fine-tunes models on teacher-generated SSR traces to ensure reliable structural adherence. Experiments across open-ended reasoning benchmarks demonstrate that SSR-D achieves up to 10% improvement over suppression baselines while preserving out-of-distribution (OOD) generalization.

Method: Proposes HyperRAG with two complementary retrieval variants: 1) HyperRetriever learns structural-semantic reasoning over n-ary facts to construct query-conditioned relational chains, enabling accurate factual tracking and adaptive high-order traversal. 2) HyperMemory leverages LLM’s parametric memory to guide beam search, dynamically scoring n-ary facts and entities for query-aware path expansion.

Result: Extensive evaluations on WikiTopics (11 closed-domain datasets) and three open-domain QA benchmarks (HotpotQA, MuSiQue, and 2WikiMultiHopQA) show HyperRetriever achieves highest answer accuracy overall, with average gains of 2.95% in MRR and 1.23% in Hits@10 over strongest baselines.

Conclusion: HyperRAG effectively addresses limitations of traditional graph-based RAG by leveraging n-ary hypergraphs, enabling more efficient and interpretable multi-hop reasoning for both open and closed-domain QA tasks.

Abstract: Graph-based retrieval-augmented generation (RAG) methods, typically built on knowledge graphs (KGs) with binary relational facts, have shown promise in multi-hop open-domain QA. However, their rigid retrieval schemes and dense similarity search often introduce irrelevant context, increase computational overhead, and limit relational expressiveness. In contrast, n-ary hypergraphs encode higher-order relational facts that capture richer inter-entity dependencies and enable shallower, more efficient reasoning paths. To address this limitation, we propose HyperRAG, a RAG framework tailored for n-ary hypergraphs with two complementary retrieval variants: (i) HyperRetriever learns structural-semantic reasoning over n-ary facts to construct query-conditioned relational chains. It enables accurate factual tracking, adaptive high-order traversal, and interpretable multi-hop reasoning under context constraints. (ii) HyperMemory leverages the LLM’s parametric memory to guide beam search, dynamically scoring n-ary facts and entities for query-aware path expansion. Extensive evaluations on WikiTopics (11 closed-domain datasets) and three open-domain QA benchmarks (HotpotQA, MuSiQue, and 2WikiMultiHopQA) validate HyperRAG’s effectiveness. HyperRetriever achieves the highest answer accuracy overall, with average gains of 2.95% in MRR and 1.23% in Hits@10 over the strongest baseline. Qualitative analysis further shows that HyperRetriever bridges reasoning gaps through adaptive and interpretable n-ary chain construction, benefiting both open and closed-domain QA.

Method: BETA-labeling framework using multiple LLM annotators from diverse families with contextual alignment, consistency checks, and majority agreement, followed by human evaluation. Also tested cross-lingual reuse via one-hop machine translation using LLM-based translation across language pairs.

Result: Substantial variation across languages in translation quality, reflecting language-dependent biases and inconsistent semantic preservation that affects cross-lingual dataset reliability. Framework enables creation of Bangla IR dataset with verified quality.

Conclusion: LLM-assisted dataset creation has potential but limitations for low-resource IR. Cross-lingual dataset reuse via translation carries significant risks due to language-dependent biases and semantic preservation issues.

Abstract: IR in low-resource languages remains limited by the scarcity of high-quality, task-specific annotated datasets. Manual annotation is expensive and difficult to scale, while using large language models (LLMs) as automated annotators introduces concerns about label reliability, bias, and evaluation validity. This work presents a Bangla IR dataset constructed using a BETA-labeling framework involving multiple LLM annotators from diverse model families. The framework incorporates contextual alignment, consistency checks, and majority agreement, followed by human evaluation to verify label quality. Beyond dataset creation, we examine whether IR datasets from other low-resource languages can be effectively reused through one-hop machine translation. Using LLM-based translation across multiple language pairs, we experimented on meaning preservation and task validity between source and translated datasets. Our experiment reveal substantial variation across languages, reflecting language-dependent biases and inconsistent semantic preservation that directly affect the reliability of cross-lingual dataset reuse. Overall, this study highlights both the potential and limitations of LLM-assisted dataset creation for low-resource IR. It provides empirical evidence of the risks associated with cross-lingual dataset reuse and offers practical guidance for constructing more reliable benchmarks and evaluation pipelines in low-resource language settings.

Method: Proposes Query-as-Anchor framework with: 1) UserU dataset aligning multi-modal behavioral sequences with user semantics, 2) Q-Anchor Embedding architecture with hierarchical encoders in dual-tower LLMs using joint contrastive-autoregressive optimization, 3) Cluster-based Soft Prompt Tuning to align with scenario-specific modalities, and 4) KV-cache-accelerated inference.

Result: Achieves consistent SOTA performance on 10 Alipay industrial benchmarks, shows strong scalability, and validates practical effectiveness through large-scale online A/B testing in Alipay’s production system across two real-world scenarios.

Conclusion: Query-as-Anchor successfully shifts user modeling from static encoding to dynamic, query-aware synthesis, enabling robust industrial-scale user representation learning with efficient deployment.

Abstract: Industrial-scale user representation learning requires balancing robust universality with acute task-sensitivity. However, existing paradigms primarily yield static, task-agnostic embeddings that struggle to reconcile the divergent requirements of downstream scenarios within unified vector spaces. Furthermore, heterogeneous multi-source data introduces inherent noise and modality conflicts, degrading representation. We propose Query-as-Anchor, a framework shifting user modeling from static encoding to dynamic, query-aware synthesis. To empower Large Language Models (LLMs) with deep user understanding, we first construct UserU, an industrial-scale pre-training dataset that aligns multi-modal behavioral sequences with user understanding semantics, and our Q-Anchor Embedding architecture integrates hierarchical coarse-to-fine encoders into dual-tower LLMs via joint contrastive-autoregressive optimization for query-aware user representation. To bridge the gap between general pre-training and specialized business logic, we further introduce Cluster-based Soft Prompt Tuning to enforce discriminative latent structures, effectively aligning model attention with scenario-specific modalities. For deployment, anchoring queries at sequence termini enables KV-cache-accelerated inference with negligible incremental latency. Evaluations on 10 Alipay industrial benchmarks show consistent SOTA performance, strong scalability, and efficient deployment. Large-scale online A/B testing in Alipay’s production system across two real-world scenarios further validates its practical effectiveness. Our code is prepared for public release and will be available at: https://github.com/JhCircle/Q-Anchor.

Method: Evaluated four prominent LLMs using a taxonomy of six math problem types, from basic arithmetic to complex unit conflict and optimization problems. Constructed a parallel dataset natively authored by fluent speakers with mathematical training in all three languages (English, Sinhala, Tamil) to avoid translation artifacts.

Result: Basic arithmetic reasoning transfers robustly across languages, but complex reasoning tasks show significant degradation in Tamil and Sinhala. Failure patterns vary by model and problem type, indicating that apparent multilingual competence doesn’t reflect uniform reasoning capabilities across languages.

Conclusion: LLMs’ apparent multilingual competence may not indicate genuine reasoning capabilities across languages, challenging common assumptions about multilingual performance. Highlights need for fine-grained, type-aware evaluation in multilingual settings.

Abstract: Large language models (LLMs) demonstrate strong mathematical reasoning in English, but whether these capabilities reflect genuine multilingual reasoning or reliance on translation-based processing in low-resource languages like Sinhala and Tamil remains unclear. We examine this fundamental question by evaluating whether LLMs genuinely reason mathematically in these languages or depend on implicit translation to English-like representations. Using a taxonomy of six math problem types, from basic arithmetic to complex unit conflict and optimization problems, we evaluate four prominent large language models. To avoid translation artifacts that confound language ability with translation quality, we construct a parallel dataset where each problem is natively authored by fluent speakers with mathematical training in all three languages. Our analysis demonstrates that while basic arithmetic reasoning transfers robustly across languages, complex reasoning tasks show significant degradation in Tamil and Sinhala. The pattern of failures varies by model and problem type, suggesting that apparent multilingual competence may not reflect uniform reasoning capabilities across languages. These findings challenge the common assumption that models exhibiting strong multilingual performance can reason equally effectively across languages, and highlight the need for fine-grained, type-aware evaluation in multilingual settings.

Method: XTF decomposes token-level contributions into three explicit attributes: reasoning importance (how crucial for logical reasoning), knowledge novelty (how informative vs. redundant), and task relevance (how related to target task). It uses scoring methods to assess these attributes and masks gradients of noisy tokens during fine-tuning to optimize performance.

Result: Extensive experiments on math, code, and medicine tasks across 7 mainstream LLMs show XTF improves downstream performance by up to 13.7% compared to regular fine-tuning. The framework demonstrates consistent improvements across different model architectures and task domains.

Conclusion: The work highlights the importance of token-level dataset optimization for LLM fine-tuning and demonstrates the potential of attribute decomposition strategies for explaining complex training mechanisms. XTF provides an effective approach to address the token-level noise problem in fine-tuning.

Abstract: Large Language Models (LLMs) have seen remarkable advancements, achieving state-of-the-art results in diverse applications. Fine-tuning, an important step for adapting LLMs to specific downstream tasks, typically involves further training on corresponding datasets. However, a fundamental discrepancy exists between current fine-tuning datasets and the token-level optimization mechanism of LLMs: most datasets are designed at the sentence-level, which introduces token-level noise, causing negative influence to final performance. In this paper, we propose XTF, an explainable token-level noise filtering framework. XTF decomposes the complex and subtle contributions of token-level data to the fine-tuning process into three distinct and explicit attributes (reasoning importance, knowledge novelty, and task relevance), which can be assessed using scoring methods, and then masks the gradients of selected noisy tokens accordingly to optimize the performance of fine-tuned LLMs. We conduct extensive experiments on three representative downstream tasks (math, code and medicine) across 7 mainstream LLMs. The results demonstrate that XTF can significantly improve downstream performance by up to 13.7% compared to regular fine-tuning. Our work highlights the importance of token-level dataset optimization, and demonstrates the potential of strategies based on attribute decomposition for explaining complex training mechanisms.

Method: Used iCliniq dataset with 38,000 medical Q&A pairs across diverse specialties. Evaluated five LLMs (Llama-3-8B-Instruct, Llama 3.2 3B, Llama 3.3 70B Instruct, Llama-4-Maverick-17B-128E-Instruct, GPT-5-mini) with zero-shot evaluation methodology using BLEU and ROUGE metrics without specialized fine-tuning.

Result: Larger models like Llama 3.3 70B Instruct outperformed smaller models, consistent with scaling benefits in clinical tasks. Llama-4-Maverick-17B exhibited competitive results, highlighting efficiency trade-offs relevant for practical deployment.

Conclusion: LLMs show increasing feasibility for medical QA systems in clinical environments, with larger models demonstrating better performance but smaller models offering efficiency trade-offs. The benchmark serves as standardized setting for future studies to balance model size, computational resources, and clinical utility.

Abstract: Recently, Large Language Models (LLMs) have gained significant traction in medical domain, especially in developing a QA systems to Medical QA systems for enhancing access to healthcare in low-resourced settings. This paper compares five LLMs deployed between April 2024 and August 2025 for medical QA, using the iCliniq dataset, containing 38,000 medical questions and answers of diverse specialties. Our models include Llama-3-8B-Instruct, Llama 3.2 3B, Llama 3.3 70B Instruct, Llama-4-Maverick-17B-128E-Instruct, and GPT-5-mini. We are using a zero-shot evaluation methodology and using BLEU and ROUGE metrics to evaluate performance without specialized fine-tuning. Our results show that larger models like Llama 3.3 70B Instruct outperform smaller models, consistent with observed scaling benefits in clinical tasks. It is notable that, Llama-4-Maverick-17B exhibited more competitive results, thus highlighting evasion efficiency trade-offs relevant for practical deployment. These findings align with advancements in LLM capabilities toward professional-level medical reasoning and reflect the increasing feasibility of LLM-supported QA systems in the real clinical environments. This benchmark aims to serve as a standardized setting for future study to minimize model size, computational resources and to maximize clinical utility in medical NLP applications.

Method: Collect real-world SPARQL queries from Wikidata Query Service logs, then use agent-based method to iteratively de-anonymize, clean, and verify queries against Wikidata while generating corresponding natural-language questions.

Result: Created WDQL dataset with 200k question-query pairs, over 6x larger than existing similar Wikidata datasets, with all assets and agent code publicly available under permissive license.

Conclusion: WDQL dataset provides valuable resource for training question-answering methods on Wikidata knowledge graph, demonstrating benefits for QA system development.

Abstract: We present the Wikidata Query Logs (WDQL) dataset, a dataset consisting of 200k question-query pairs over the Wikidata knowledge graph. It is over 6x larger than the largest existing Wikidata datasets of similar format without relying on template-generated queries. Instead, we construct it using real-world SPARQL queries sent to the Wikidata Query Service and generate questions for them. Since these log-based queries are anonymized, and therefore often do not produce results, a significant amount of effort is needed to convert them back into meaningful SPARQL queries. To achieve this, we present an agent-based method that iteratively de-anonymizes, cleans, and verifies queries against Wikidata while also generating corresponding natural-language questions. We demonstrate the dataset’s benefit for training question-answering methods. All WDQL assets, as well as the agent code, are publicly available under a permissive license.

Method: Two-stage approach: 1) Novel gradient magnitude metric for global layer importance assessment requiring only single backward propagation per pruning decision. 2) Analysis of layers with largest mean shift from pruning, then projection compensation matrix to correct drift in one step.

Result: Outperforms previous layer pruning methods in both pruning speed (average 4× speedup) and performance. Effectively alleviates degradation from layer pruning.

Conclusion: GradMAP provides efficient layer pruning for LLMs with improved speed and maintained performance through gradient-based metrics and projection compensation.

Abstract: Large Language Models (LLMs) exhibit strong reasoning abilities, but their high computational costs limit their practical deployment. Recent studies reveal significant redundancy in LLMs layers, making layer pruning an active research topic. Layer pruning research primarily focuses on two aspects: measuring layer importance and recovering performance after pruning. Unfortunately, the present works fail to simultaneously maintain pruning performance and efficiency. In this study, we propose GradMAP, a faster layer pruning method with \textbf{Grad}ient \textbf{M}etric \textbf{A}nd \textbf{P}rojection compensation, which consists of two stages. In the first stage, we introduce a novel metric based on gradient magnitudes, enabling a global assessment of layer importance. Note that, it requires only a single backward propagation step per pruning decision, substantially enhancing pruning efficiency. In the second stage, we first analyze the layers with the largest mean shift resulting from pruning, and then incorporate a simple yet effective projection compensation matrix to correct this drift in one step. In this way, the degradation of model performance caused by layer pruning is effectively alleviated. Extensive experiments show that GradMAP outperforms previous layer pruning methods in both pruning speed (achieving an average $4\times$ speedup) and performance.

Method: Used multilingual vision-and-language models to estimate surprisal over image-caption data in 30 languages and visual storytelling data in 13 languages. Analyzed global and local uniformity of information distribution across typologically diverse languages.

Result: Visual grounding consistently smooths information distribution, increasing both global and local uniformity across diverse languages. In visual narratives, grounding in both image and discourse contexts has additional effects, with strongest surprisal reductions at discourse unit onsets.

Conclusion: Grounded language exhibits greater information uniformity than text-only settings, supporting a context-sensitive formulation of UID. This study advances understanding of information flow dynamics in multimodal language use.

Abstract: The Uniform Information Density (UID) hypothesis posits that speakers are subject to a communicative pressure to distribute information evenly within utterances, minimising surprisal variance. While this hypothesis has been tested empirically, prior studies are limited exclusively to text-only inputs, abstracting away from the perceptual context in which utterances are produced. In this work, we present the first computational study of UID in visually grounded settings. We estimate surprisal using multilingual vision-and-language models over image-caption data in 30 languages and visual storytelling data in 13 languages, together spanning 11 families. We find that grounding on perception consistently smooths the distribution of information, increasing both global and local uniformity across typologically diverse languages compared to text-only settings. In visual narratives, grounding in both image and discourse contexts has additional effects, with the strongest surprisal reductions occurring at the onset of discourse units. Overall, this study takes a first step towards modelling the temporal dynamics of information flow in ecologically plausible, multimodal language use, and finds that grounded language exhibits greater information uniformity, supporting a context-sensitive formulation of UID.

Method: Three-part framework: 1) Voice conversion-based data augmentation via cross-category voice-content recombination, 2) Adaptive federated learning for cross-institutional collaboration under privacy constraints, 3) Attentive cross-modal fusion model for word-level acoustic-textual alignment.

Result: Achieves state-of-the-art multi-modal accuracy of 91.52% on ADReSSo dataset, outperforming all centralized baselines and demonstrating practical solution to data efficiency dilemma.

Conclusion: FAL-AD provides effective framework combining federated learning with data augmentation for medical speech analysis, addressing data scarcity and privacy concerns while achieving high diagnostic accuracy.

Abstract: Early diagnosis of Alzheimer’s Disease (AD) is crucial for delaying its progression. While AI-based speech detection is non-invasive and cost-effective, it faces a critical data efficiency dilemma due to medical data scarcity and privacy barriers. Therefore, we propose FAL-AD, a novel framework that synergistically integrates federated learning with data augmentation to systematically optimize data efficiency. Our approach delivers three key breakthroughs: First, absolute efficiency improvement through voice conversion-based augmentation, which generates diverse pathological speech samples via cross-category voice-content recombination. Second, collaborative efficiency breakthrough via an adaptive federated learning paradigm, maximizing cross-institutional benefits under privacy constraints. Finally, representational efficiency optimization by an attentive cross-modal fusion model, which achieves fine-grained word-level alignment and acoustic-textual interaction. Evaluated on ADReSSo, FAL-AD achieves a state-of-the-art multi-modal accuracy of 91.52%, outperforming all centralized baselines and demonstrating a practical solution to the data efficiency dilemma. Our source code is publicly available at https://github.com/smileix/fal-ad.

Method: Created 145 Italian-Piedmontese parallel sentences from Flores+ with natural orthographic translations and manual word alignment. Used this dataset to benchmark LLMs on tokenization parity, topic classification, and machine translation tasks.

Result: Piedmontese incurs tokenization penalty compared to higher-resource Romance languages. LLMs achieve classification performance approaching Italian/French/English levels. Machine translation is asymmetric: adequate from Piedmontese to high-resource languages, but challenging into Piedmontese.

Conclusion: The dataset provides valuable resources for endangered language research. LLMs show promise for classification but struggle with generation into low-resource languages like Piedmontese, highlighting challenges in multilingual NLP for endangered languages.

Abstract: We present a crowdsourced dataset for Piedmontese, an endangered Romance language of northwestern Italy. The dataset comprises 145 Italian-Piedmontese parallel sentences derived from Flores+, with translations produced by speakers writing in their natural orthographic style rather than adhering to standardized conventions, along with manual word alignment. We use this resource to benchmark several large language models on tokenization parity, topic classification, and machine translation. Our analysis reveals that Piedmontese incurs a tokenization penalty relative to higher-resource Romance languages, yet LLMs achieve classification performance approaching that of Italian, French, and English. Machine translation results are asymmetric: models translate adequately from Piedmontese into high-resource languages, but generation into Piedmontese remains challenging. The dataset and code are publicly released.

Method: Created an open dataset with diverse, manually verified parsing scenarios of varying complexity. Evaluated 22 models using five different prompting strategies. Introduced complementary performance metrics capturing both token-level accuracy and document-level validity.

Result: Found that choosing the right prompting strategy is more important than model size for parsing reliability. Proper prompting ensures structural validity for smaller/less reliable models but may increase semantic errors. Provides systematic comparison of model, size, and prompting effects.

Conclusion: LLMStructBench enables rigorous evaluation of LLMs for structured data extraction and JSON generation, serving as a foundation for future research in LLM-based parsing and ETL applications.

Abstract: We present LLMStructBench, a novel benchmark for evaluating Large Language Models (LLMs) on extracting structured data and generating valid JavaScript Object Notation (JSON) outputs from natural-language text. Our open dataset comprises diverse, manually verified parsing scenarios of varying complexity and enables systematic testing across 22 models and five prompting strategies. We further introduce complementary performance metrics that capture both token-level accuracy and document-level validity, facilitating rigorous comparison of model, size, and prompting effects on parsing reliability. In particular, we show that choosing the right prompting strategy is more important than standard attributes such as model size. This especially ensures structural validity for smaller or less reliable models but increase the number of semantic errors. Our benchmark suite is an step towards future research in the area of LLM applied to parsing or Extract, Transform and Load (ETL) applications.

Method: Conducted a large-scale study across 8 billion observations, 17 forecasting scenarios, 4 horizons, multiple alignment strategies, and both in-domain and out-of-domain settings. Compared pre-alignment vs post-alignment approaches and analyzed contributions of pretrained knowledge vs model architecture.

Result: LLMs significantly improve forecasting performance, with especially large gains in cross-domain generalization. Pre-alignment outperformed post-alignment in over 90% of tasks. Both pretrained knowledge and model architecture contribute: pretraining is critical under distribution shifts, while architecture excels at modeling complex temporal dynamics.

Conclusion: Findings overturn prior negative assessments about LLMs for time series forecasting. LLMs are indeed useful, especially for cross-domain generalization, and the study provides practical guidance for effective model design.

Abstract: Large language models (LLMs) have been introduced to time series forecasting (TSF) to incorporate contextual knowledge beyond numerical signals. However, existing studies question whether LLMs provide genuine benefits, often reporting comparable performance without LLMs. We show that such conclusions stem from limited evaluation settings and do not hold at scale. We conduct a large-scale study of LLM-based TSF (LLM4TSF) across 8 billion observations, 17 forecasting scenarios, 4 horizons, multiple alignment strategies, and both in-domain and out-of-domain settings. Our results demonstrate that \emph{LLM4TS indeed improves forecasting performance}, with especially large gains in cross-domain generalization. Pre-alignment outperforming post-alignment in over 90% of tasks. Both pretrained knowledge and model architecture of LLMs contribute and play complementary roles: pretraining is critical under distribution shifts, while architecture excels at modeling complex temporal dynamics. Moreover, under large-scale mixed distributions, a fully intact LLM becomes indispensable, as confirmed by token-level routing analysis and prompt-based improvements. Overall, Our findings overturn prior negative assessments, establish clear conditions under which LLMs are not only useful, and provide practical guidance for effective model design. We release our code at https://github.com/EIT-NLP/LLM4TSF.

Method: Construct BFMNs from 994 observations across high school students, university students, and STEM experts, with nodes as cue words/free associations and edges as associative links, plus LLM “digital twins” (GPT-oss) prompted to emulate comparable profiles; analyze semantic frames around key STEM concepts using valence auras, emotional profiles, network overlap, and concreteness metrics

Result: Science/research consistently framed positively, but quantitative subjects (math/statistics) show negative/anxiety-related auras, especially in high math-anxiety subgroups; high-anxiety frames are less concrete than chance; human networks show greater math-anxiety overlap than GPT-oss

Conclusion: BFMNs effectively capture cognitive-affective signatures of STEM mindsets; LLM digital twins approximate cultural attitudes but miss context-sensitive, experience-based components of human educational anxiety

Abstract: Attitudes toward STEM develop from the interaction of conceptual knowledge, educational experiences, and affect. Here we use cognitive network science to reconstruct group mindsets as behavioural forma mentis networks (BFMNs). In this case, nodes are cue words and free associations, edges are empirical associative links, and each concept is annotated with perceived valence. We analyse BFMNs from N = 994 observations spanning high school students, university students, and early-career STEM experts, alongside LLM (GPT-oss) “digital twins” prompted to emulate comparable profiles. Focusing also on semantic neighbourhoods (“frames”) around key target concepts (e.g., STEM subjects or educational actors/places), we quantify frames in terms of valence auras, emotional profiles, network overlap (Jaccard similarity), and concreteness relative to null baselines. Across student groups, science and research are consistently framed positively, while their core quantitative subjects (mathematics and statistics) exhibit more negative and anxiety related auras, amplified in higher math-anxiety subgroups, evidencing a STEM-science cognitive and emotional dissonance. High-anxiety frames are also less concrete than chance, suggesting more abstract and decontextualised representations of threatening quantitative domains. Human networks show greater overlapping between mathematics and anxiety than GPT-oss. The results highlight how BFMNs capture cognitive-affective signatures of mindsets towards the target domains and indicate that LLM-based digital twins approximate cultural attitudes but miss key context-sensitive, experience-based components relevant to replicate human educational anxiety.

Method: Empirically localize input-output alignment shift using decoding trajectories over tied embedding spaces and similarity-based metrics; propose lightweight residual-path mitigation via residual attenuation, implemented as fixed-layer intervention or learnable gating mechanism.

Result: Experiments reveal hidden token representations switch from input alignment to output alignment deep within the network; residual attenuation strategies alleviate representation misalignment and yield improvements on multiple benchmarks.

Conclusion: The proposed residual attenuation provides an efficient and general architectural enhancement for autoregressive Transformers by addressing the fundamental input-output alignment shift in LLMs.

Abstract: Large Language Models (LLMs) are trained with next-token prediction, implemented in autoregressive Transformers via causal masking for parallelism. This creates a subtle misalignment: residual connections tie activations to the current token, while supervision targets the next token, potentially propagating mismatched information if the current token is not the most informative for prediction. In this work, we empirically localize this input-output alignment shift in pretrained LLMs, using decoding trajectories over tied embedding spaces and similarity-based metrics. Our experiments reveal that the hidden token representations switch from input alignment to output alignment deep within the network. Motivated by this observation, we propose a lightweight residual-path mitigation based on residual attenuation, implemented either as a fixed-layer intervention or as a learnable gating mechanism. Experiments on multiple benchmarks show that these strategies alleviate the representation misalignment and yield improvements, providing an efficient and general architectural enhancement for autoregressive Transformers.

Method: Systematically evaluate RLMs on WMT24++, analyze reasoning patterns, propose structured translation framework (multi-step drafting, adequacy refinement, fluency improvement, selective iterative revision), curate synthetic dataset of dynamic structured reasoning traces, and post-train large reasoning model.

Result: Standard RLMs degrade translation quality; MT reasoning traces are linear without revision/self-correction; structured reasoning framework significantly outperforms standard translation fine-tuning and generic reasoning baselines.

Conclusion: Reasoning must be task-structured to benefit machine translation; generic reasoning approaches are insufficient; structured translation-specific reasoning enables meaningful improvements.

Abstract: Reasoning-oriented large language models (RLMs) achieve strong gains on tasks such as mathematics and coding by generating explicit intermediate reasoning. However, their impact on machine translation (MT) remains underexplored. We systematically evaluate several open- and closed-weights RLMs on the WMT24++ benchmark and find that enabling explicit reasoning consistently degrades translation quality across languages and models. Analysis reveals that MT reasoning traces are highly linear, lacking revision, self-correction and exploration of alternative translations, which limits their usefulness. Furthermore, injecting higher-quality reasoning traces from stronger models does not reliably improve weaker models’ performance. To address this mismatch, we propose a structured reasoning framework tailored to translation, based on multi-step drafting, adequacy refinement, fluency improvement, and selective iterative revision. We curate a synthetic dataset of dynamic structured reasoning traces and post-train a large reasoning model on this data. Experiments show significant improvements over standard translation fine-tuning and injected generic reasoning baselines. Our findings demonstrate that reasoning must be task-structured to benefit MT.

Method: Controlled multi-agent sandbox experiment comparing two conditions: discussion condition where critic/audience threads are recorded, filtered, stored as social memory and retrieved to condition subsequent generations vs. baseline without discussion. 50 rounds (250 paired monologues) evaluated by five expert annotators using A/B preference and 15-item rubric.

Result: Discussion condition wins 75.6% of instances, improves Craft/Clarity (Δ = 0.440) and Social Response (Δ = 0.422), with occasional increases in aggressive humor.

Conclusion: Incorporating community discussion feedback significantly improves stand-up comedy writing quality, demonstrating the value of social memory and community feedback mechanisms in LLM writing systems.

Abstract: Prior work has explored multi-turn interaction and feedback for LLM writing, but evaluations still largely center on prompts and localized feedback, leaving persistent public reception in online communities underexamined. We test whether broadcast community discussion improves stand-up comedy writing in a controlled multi-agent sandbox: in the discussion condition, critic and audience threads are recorded, filtered, stored as social memory, and later retrieved to condition subsequent generations, whereas the baseline omits discussion. Across 50 rounds (250 paired monologues) judged by five expert annotators using A/B preference and a 15-item rubric, discussion wins 75.6% of instances and improves Craft/Clarity (Δ = 0.440) and Social Response (Δ = 0.422), with occasional increases in aggressive humor.

Method: Fine-tuned GPT-4.1 models sequentially on datasets known to induce and reverse emergent misalignment, then evaluated whether models could self-describe their behavior transitions without in-context examples.

Result: Emergently misaligned models rated themselves as significantly more harmful compared to base models and realigned counterparts, demonstrating behavioral self-awareness of their own emergent misalignment.

Conclusion: Behavioral self-awareness tracks actual alignment states of models, indicating that models can be queried for informative signals about their own safety and alignment.

Abstract: Recent research has demonstrated that large language models (LLMs) fine-tuned on incorrect trivia question-answer pairs exhibit toxicity - a phenomenon later termed “emergent misalignment”. Moreover, research has shown that LLMs possess behavioral self-awareness - the ability to describe learned behaviors that were only implicitly demonstrated in training data. Here, we investigate the intersection of these phenomena. We fine-tune GPT-4.1 models sequentially on datasets known to induce and reverse emergent misalignment and evaluate whether the models are self-aware of their behavior transitions without providing in-context examples. Our results show that emergently misaligned models rate themselves as significantly more harmful compared to their base model and realigned counterparts, demonstrating behavioral self-awareness of their own emergent misalignment. Our findings show that behavioral self-awareness tracks actual alignment states of models, indicating that models can be queried for informative signals about their own safety.

Method: Analyzes hallucinations through geometric perspective in embedding space. Hypothesis: genuine responses to same prompt cluster tightly while hallucinations are dispersed. Proves this hypothesis mathematically and develops label-efficient propagation method using 30-50 annotations to classify large response collections.

Result: Achieves consistent separability between genuine and hallucinated responses. The propagation method achieves F1 scores above 90% with minimal annotations, demonstrating effective classification of large response collections.

Conclusion: Geometric perspective complements traditional evaluation paradigms, providing new insights into hallucination patterns. The efficient classification method enables practical detection of hallucinations with minimal labeled data.

Abstract: Hallucinations – fluent but factually incorrect responses – pose a major challenge to the reliability of language models, especially in multi-step or agentic settings. This work investigates hallucinations in small-sized LLMs through a geometric perspective, starting from the hypothesis that when models generate multiple responses to the same prompt, genuine ones exhibit tighter clustering in the embedding space, we prove this hypothesis and, leveraging this geometrical insight, we also show that it is possible to achieve a consistent level of separability. This latter result is used to introduce a label-efficient propagation method that classifies large collections of responses from just 30-50 annotations, achieving F1 scores above 90%. Our findings, framing hallucinations from a geometric perspective in the embedding space, complement traditional knowledge-centric and single-response evaluation paradigms, paving the way for further research.

Method: The authors formalize structural overthinking attacks, implement 14 malicious tools across three servers that trigger repetition, forced refinement, and distraction patterns, and test across heterogeneous registries and multiple tool-capable models.

Result: The attack causes severe resource amplification (up to 142.4× tokens) and can degrade task outcomes. Decoding-time concision controls do not reliably prevent loop induction.

Conclusion: Defenses should reason about tool-call structure rather than tokens alone, as malicious tool servers can exploit the LLM agent ecosystem through structural attacks that evade token-level detection.

Abstract: Tool-using LLM agents increasingly coordinate real workloads by selecting and chaining third-party tools based on text-visible metadata such as tool names, descriptions, and return messages. We show that this convenience creates a supply-chain attack surface: a malicious MCP tool server can be co-registered alongside normal tools and induce overthinking loops, where individually trivial or plausible tool calls compose into cyclic trajectories that inflate end-to-end tokens and latency without any single step looking abnormal. We formalize this as a structural overthinking attack, distinguishable from token-level verbosity, and implement 14 malicious tools across three servers that trigger repetition, forced refinement, and distraction. Across heterogeneous registries and multiple tool-capable models, the attack causes severe resource amplification (up to $142.4\times$ tokens) and can degrade task outcomes. Finally, we find that decoding-time concision controls do not reliably prevent loop induction, suggesting defenses should reason about tool-call structure rather than tokens alone.

Method: Created BasPhyCo dataset for Basque (standard and dialectal variants) based on Italian GITA. Evaluated multilingual LLMs and language-specific models on three hierarchical tasks: accuracy (plausible vs implausible narratives), consistency (identifying conflicting elements), and verifiability (determining specific physical states causing implausibility).

Result: LLMs show limited physical commonsense capabilities in low-resource languages like Basque, particularly with dialectal variants. Performance degrades significantly on the verifiability task compared to accuracy and consistency tasks.

Conclusion: Physical commonsense reasoning remains challenging for LLMs in low-resource languages, highlighting the need for better multilingual reasoning capabilities and consideration of dialectal variations in model development.

Abstract: Physical commonsense reasoning represents a fundamental capability of human intelligence, enabling individuals to understand their environment, predict future events, and navigate physical spaces. Recent years have witnessed growing interest in reasoning tasks within Natural Language Processing (NLP). However, no prior research has examined the performance of Large Language Models (LLMs) on non-question-answering (non-QA) physical commonsense reasoning tasks in low-resource languages such as Basque. Taking the Italian GITA as a starting point, this paper addresses this gap by presenting BasPhyCo, the first non-QA physical commonsense reasoning dataset for Basque, available in both standard and dialectal variants. We evaluate model performance across three hierarchical levels of commonsense understanding: (1) distinguishing between plausible and implausible narratives (accuracy), (2) identifying the conflicting element that renders a narrative implausible (consistency), and (3) determining the specific physical state that creates the implausibility (verifiability). These tasks were assessed using multiple multilingual LLMs as well as models pretrained specifically for Italian and Basque. Results indicate that, in terms of verifiability, LLMs exhibit limited physical commonsense capabilities in low-resource languages such as Basque, especially when processing dialectal variants.

Method: Collection and compilation of discussion board messages in Italian spanning 1996-2024, resulting in a corpus of over 30 billion word-tokens.

Result: Created a massive Italian conversational corpus (Testimole-conversational) that captures rich variety of informal written Italian, discourse dynamics, and online social interaction patterns.

Conclusion: The corpus serves as an ideal dataset for Italian LLM pre-training and supports both NLP applications (language modeling, conversational analysis) and investigations of language variation and social phenomena in digital communication.

Abstract: We present “Testimole-conversational” a massive collection of discussion boards messages in the Italian language. The large size of the corpus, more than 30B word-tokens (1996-2024), renders it an ideal dataset for native Italian Large Language Models’pre-training. Furthermore, discussion boards’ messages are a relevant resource for linguistic as well as sociological analysis. The corpus captures a rich variety of computer-mediated communication, offering insights into informal written Italian, discourse dynamics, and online social interaction in wide time span. Beyond its relevance for NLP applications such as language modelling, domain adaptation, and conversational analysis, it also support investigations of language variation and social phenomena in digital communication. The resource will be made freely available to the research community.

Method: BFS-PO uses a Best-First Search exploration strategy with backtracking based on maximum entropy nodes to find the shortest correct answers. During training, it generates progressively shorter responses, teaching the model to produce concise reasoning chains.

Result: Experiments across different benchmarks and base LRMs show that BFS-PO can simultaneously increase model accuracy and shorten answer lengths, effectively reducing overthinking while improving performance.

Conclusion: BFS-PO successfully addresses the overthinking problem in LRMs by combining Best-First Search exploration with maximum entropy backtracking, enabling more efficient and accurate reasoning with shorter outputs.

Abstract: Large Reasoning Models (LRMs) such as OpenAI o1 and DeepSeek-R1 have shown excellent performance in reasoning tasks using long reasoning chains. However, this has also led to a significant increase of computational costs and the generation of verbose output, a phenomenon known as overthinking. The tendency to overthinking is often exacerbated by Reinforcement Learning (RL) algorithms such as GRPO/DAPO. In this paper, we propose BFS-PO, an RL algorithm which alleviates this problem using a Best-First Search exploration strategy. Specifically, BFS-PO looks for the shortest correct answer using a backtracking mechanism based on maximum entropy nodes. By generating progressively shorter responses during training, BFS-PO learns to produce concise reasoning chains. Using different benchmarks and base LRMs, we show that BFS-PO can simultaneously increase the LRM accuracy and shorten its answers.

Method: Developed a domain-grounded framework with: (1) reference-based plan evaluation with metric-wise (7 dimensions) and one-shot evaluators, (2) iterative data curation via evaluator-optimizer loop to create high-quality plan lineages, and (3) large-scale study of 14 LLMs across sizes/families for query decomposition.

Result: LLMs struggle with compound queries and plans over 4 steps (typically 5-15 steps); best total metric score is 84.8% (Claude-3-7-Sonnet), best one-shot match rate at “A+” tier is only 49.75% (o3-mini). Plan lineage provides mixed gains but improves executability for many models.

Conclusion: Persistent gaps exist in tool-understanding (alignment and completeness), with simpler plans being markedly easier. The framework provides reproducible assessment for improving agentic planning with tools in contact-center data analysis.

Abstract: We present a domain-grounded framework and benchmark for tool-aware plan generation in contact centers, where answering a query for business insights, our target use case, requires decomposing it into executable steps over structured tools (Text2SQL (T2S)/Snowflake) and unstructured tools (RAG/transcripts) with explicit depends_on for parallelism. Our contributions are threefold: (i) a reference-based plan evaluation framework operating in two modes - a metric-wise evaluator spanning seven dimensions (e.g., tool-prompt alignment, query adherence) and a one-shot evaluator; (ii) a data curation methodology that iteratively refines plans via an evaluator->optimizer loop to produce high-quality plan lineages (ordered plan revisions) while reducing manual effort; and (iii) a large-scale study of 14 LLMs across sizes and families for their ability to decompose queries into step-by-step, executable, and tool-assigned plans, evaluated under prompts with and without lineage. Empirically, LLMs struggle on compound queries and on plans exceeding 4 steps (typically 5-15); the best total metric score reaches 84.8% (Claude-3-7-Sonnet), while the strongest one-shot match rate at the “A+” tier (Extremely Good, Very Good) is only 49.75% (o3-mini). Plan lineage yields mixed gains overall but benefits several top models and improves step executability for many. Our results highlight persistent gaps in tool-understanding, especially in tool-prompt alignment and tool-usage completeness, and show that shorter, simpler plans are markedly easier. The framework and findings provide a reproducible path for assessing and improving agentic planning with tools for answering data-analysis queries in contact-center settings.

Method: Counterfactual fairness evaluation across 13 dimensions in three categories (Identity, Context, Behavioral Style) using Counterfactual Flip Rate (CFR) and Mean Absolute Score Difference (MASD) on 3,000 real-world contact center transcripts with 18 LLMs.

Result: Systematic disparities found with CFR ranging 5.4-13.0% and consistent MASD shifts across confidence, positive, and improvement scores. Larger, aligned models show lower unfairness but fairness doesn’t track accuracy. Contextual priming causes worst degradations (CFR up to 16.4%). Fairness-aware prompting yields only modest improvements.

Conclusion: Standardized fairness auditing pipelines are needed before deploying LLMs in high-stakes workforce evaluation due to persistent biases in LLM-based QA systems.

Abstract: Large Language Models (LLMs) are increasingly deployed in contact-center Quality Assurance (QA) to automate agent performance evaluation and coaching feedback. While LLMs offer unprecedented scalability and speed, their reliance on web-scale training data raises concerns regarding demographic and behavioral biases that may distort workforce assessment. We present a counterfactual fairness evaluation of LLM-based QA systems across 13 dimensions spanning three categories: Identity, Context, and Behavioral Style. Fairness is quantified using the Counterfactual Flip Rate (CFR), the frequency of binary judgment reversals, and the Mean Absolute Score Difference (MASD), the average shift in coaching or confidence scores across counterfactual pairs. Evaluating 18 LLMs on 3,000 real-world contact center transcripts, we find systematic disparities, with CFR ranging from 5.4% to 13.0% and consistent MASD shifts across confidence, positive, and improvement scores. Larger, more strongly aligned models show lower unfairness, though fairness does not track accuracy. Contextual priming of historical performance induces the most severe degradations (CFR up to 16.4%), while implicit linguistic identity cues remain a persistent bias source. Finally, we analyze the efficacy of fairness-aware prompting, finding that explicit instructions yield only modest improvements in evaluative consistency. Our findings underscore the need for standardized fairness auditing pipelines prior to deploying LLMs in high-stakes workforce evaluation.

Method: Formulate query-list generation as policy optimization problem using GRPO with multiple reward signals; study inference-time sampling and model capacity; perform on-policy distillation from large teacher to compact student model.

Result: Consistent improvements with increased compute showing scaling-like behavior; extensive offline experiments and large-scale online A/B tests demonstrate gains in interest modeling quality and downstream recommendation performance.

Conclusion: The RL framework successfully generates high-quality interest-driven queries from cross-domain signals, with distillation enabling scalable deployment while maintaining performance.

Abstract: News recommendation plays a critical role in online news platforms by helping users discover relevant content. Cross-domain news recommendation further requires inferring user’s underlying information needs from heterogeneous signals that often extend beyond direct news consumption. A key challenge lies in moving beyond surface-level behaviors to capture deeper, reusable user interests while maintaining scalability in large-scale production systems. In this paper, we present a reinforcement learning framework that trains large language models to generate high-quality lists of interest-driven news search queries from cross-domain user signals. We formulate query-list generation as a policy optimization problem and employ GRPO with multiple reward signals. We systematically study two compute dimensions: inference-time sampling and model capacity, and empirically observe consistent improvements with increased compute that exhibit scaling-like behavior. Finally, we perform on-policy distillation to transfer the learned policy from a large, compute-intensive teacher to a compact student model suitable for scalable deployment. Extensive offline experiments, ablation studies and large-scale online A/B tests in a production news recommendation system demonstrate consistent gains in both interest modeling quality and downstream recommendation performance.

Method: Decomposes into offline structure learning (learning preference correlations from complete profiles) and online Bayesian inference (selecting informative questions and predicting complete profiles using training-free Bayesian inference).

Result: Achieves 80.8% alignment vs 68.5% for RL with 3-5x fewer interactions; changes follow-ups 39-62% vs 0-28% for RL when users give different answers; uses ~10K parameters vs 8B for RL.

Conclusion: The bottleneck in cold-start elicitation is exploiting the factored structure of preference data, not model size; Pep’s modular framework with simple belief models outperforms RL approaches.

Abstract: Cold-start personalization requires inferring user preferences through interaction when no user-specific historical data is available. The core challenge is a routing problem: each task admits dozens of preference dimensions, yet individual users care about only a few, and which ones matter depends on who is asking. With a limited question budget, asking without structure will miss the dimensions that matter. Reinforcement learning is the natural formulation, but in multi-turn settings its terminal reward fails to exploit the factored, per-criterion structure of preference data, and in practice learned policies collapse to static question sequences that ignore user responses. We propose decomposing cold-start elicitation into offline structure learning and online Bayesian inference. Pep (Preference Elicitation with Priors) learns a structured world model of preference correlations offline from complete profiles, then performs training-free Bayesian inference online to select informative questions and predict complete preference profiles, including dimensions never asked about. The framework is modular across downstream solvers and requires only simple belief models. Across medical, mathematical, social, and commonsense reasoning, Pep achieves 80.8% alignment between generated responses and users’ stated preferences versus 68.5% for RL, with 3-5x fewer interactions. When two users give different answers to the same question, Pep changes its follow-up 39-62% of the time versus 0-28% for RL. It does so with ~10K parameters versus 8B for RL, showing that the bottleneck in cold-start elicitation is the capability to exploit the factored structure of preference data.

Method: Uses roundtrip translation to synthesize parallel datasets from monolingual corpora, creating ’neutralized’ text as a shared input style. Applies parameter-efficient fine-tuning of LLMs and integrates retrieval-augmented generation (RAG) for terminology and name knowledge.

Result: Consistently outperforms zero-shot prompting and few-shot in-context learning techniques across four domains, achieving higher BLEU scores and style accuracy scores.

Conclusion: The proposed method effectively addresses data scarcity in text style transfer through synthetic parallel data creation and parameter-efficient fine-tuning, with RAG further enhancing robustness and stylistic consistency.

Abstract: This paper proposes a novel method for Text Style Transfer (TST) based on parameter-efficient fine-tuning of Large Language Models (LLMs). Addressing the scarcity of parallel corpora that map between styles, the study employs roundtrip translation to synthesize such parallel datasets from monolingual corpora. This approach creates ’neutralized’ text devoid of stylistic attributes, essentially creating a shared input style at training-time and inference-time. Experimental results demonstrate consistent superiority of this method over zero-shot prompting and fewshot ICL techniques measured by BLEU scores and style accuracy scores across four investigated domains. Furthermore, the integration of retrieval-augmented generation (RAG) for terminology and name knowledge enhances robustness and stylistic consistency.

Method: Inheritune initializes a compact model by inheriting potent early layers from a larger pre-trained model, then progressively trains and expands it to build smaller, stronger language models.

Result: Experiments on GPT-2 family show models trained with Inheritune can match or surpass performance of larger counterparts despite having significantly fewer layers.

Conclusion: The work presents a novel model compression approach by design, enabling creation of compact yet highly performant language models through inheritance of effective early layers.

Abstract: Large Language Models (LLMs) are known for their performance, but we uncover a significant structural inefficiency: a phenomenon we term attention collapse. In many pre-trained decoder-style LLMs, the attention matrices in deeper layers degenerate, collapsing to near rank-one structures. These underutilized layers, which we call lazy layers, are redundant and impair model efficiency. To address this, we introduce Inheritune, a simple yet powerful training recipe designed to build smaller, stronger language models. Inheritune initializes a compact model by inheriting the potent early layers from a larger pre-trained model and then progressively trains and expands it. Our experiments on various models, including the GPT-2 family, demonstrate that models trained with Inheritune can match or even surpass the performance of their larger counterparts, despite having significantly fewer layers. This work presents a novel path toward model compression by design, enabling the creation of compact, yet highly performant language models. Code is available at https://github.com/sanyalsunny111/LLM-Inheritune.

Method: Created ViTextVQA dataset with 16K+ images and 50K+ questions. Proposed ViTextBLIP-2 model integrating frozen Vision Transformer, SwinTextSpotter OCR, ViT5 LLM with trainable Q-Former for multimodal feature fusion.

Result: Dataset enables research on Vietnamese scene text VQA. Model experiments reveal importance of OCR token processing order for answer formulation, leading to significant performance improvements over baselines.

Conclusion: ViTextVQA fills gap in Vietnamese text-based VQA research. The proposed multimodal fusion approach effectively handles scene text understanding, with OCR token ordering being a key factor for performance.

Abstract: Visual Question Answerinng (VQA) is a complicated task that requires the capability of simultaneously processing natural language and images. This task was initially researched with a focus on developing methods to help machines understand objects and scene contexts in images. However, some scene text that carries explicit information about the full content of the image is not mentioned. Along with the continuous development of the AI era, there have been many studies on the reading comprehension ability of VQA models in the world. Therefore, we introduce the first large-scale dataset in Vietnamese specializing in the ability to understand scene text, we call it ViTextVQA (\textbf{Vi}etnamese \textbf{Text}-based \textbf{V}isual \textbf{Q}uestion \textbf{A}nswering dataset) which contains \textbf{over 16,000} images and \textbf{over 50,000} questions with answers. To tackle this task efficiently, we propose ViTextBLIP-2, an novel multimodal feature fusion Method, which optimizes Vietnamese OCR-based VQA by integrating a frozen Vision Transformer, SwinTextSpotter OCR, and ViT5 LLM with a trainable Q-Former for multimodal feature fusion. Through experiments with various state-of-the-art models, we uncover the significance of the order in which tokens in OCR text are processed and selected to formulate answers. This finding helped us significantly improve the performance of the baseline models on the ViTextVQA dataset. Our dataset is available (https://github.com/minhquan6203/ViTextVQA-Dataset) for research purposes.

Method: Survey methodology providing high-level overview of each language’s linguistic features, current technology landscape, application of transfer learning from higher-resource languages, and analysis of available labeled/unlabeled data resources.

Result: Comprehensive summary of the current state of NLP research for Central Asian Turkic languages, identifying progress made and outlining specific future research directions to address remaining challenges.

Conclusion: The paper aims to inspire and facilitate future research by providing a clear overview of the current landscape and highlighting opportunities for advancing NLP technology for these under-resourced languages.

Abstract: Research in NLP for Central Asian Turkic languages - Kazakh, Uzbek, Kyrgyz, and Turkmen - faces typical low-resource language challenges like data scarcity, limited linguistic resources and technology development. However, recent advancements have included the collection of language-specific datasets and the development of models for downstream tasks. Thus, this paper aims to summarize recent progress and identify future research directions. It provides a high-level overview of each language’s linguistic features, the current technology landscape, the application of transfer learning from higher-resource languages, and the availability of labeled and unlabeled data. By outlining the current state, we hope to inspire and facilitate future research.

Method: HIPPO represents tables using both text and image modalities, samples MLLM responses from hybrid-modal representations, and uses a modality-consistent sampling strategy to enhance diversity and mitigate bias during Direct Preference Optimization training.

Result: Experiments on table question answering and table fact verification show 4% improvement over various table reasoning models. HIPPO enhances table reasoning with unimodal representations and facilitates extraction of complementary semantics across modalities.

Conclusion: Hybrid-modal preference optimization effectively captures structural semantics from tabular data, demonstrating that learning from multiple modalities improves MLLM table reasoning capabilities.

Abstract: Tabular data contains rich structural semantics and plays a crucial role in organizing and manipulating information. Recent methods employ Multi-modal Large Language Models (MLLMs) to address table-related tasks across various modalities of table representations. However, existing studies mainly focus on exploring the table understanding ability of MLLMs using unimodal representations, which limits further exploration of multi-modal representations to enable more effective table reasoning. To better capture structural semantics from the tabular data, this paper introduces the HybrId-modal Preference oPtimizatiOn (HIPPO) model, which represents tables using both text and image, optimizing MLLMs by learning more comprehensive table information from these multiple modalities. Specifically, HIPPO samples MLLM responses from hybrid-modal table representations and designs a modality-consistent sampling strategy to enhance response diversity and mitigate modality bias during Direct Preference Optimization (DPO) training. Experiments on table question answering and table fact verification tasks demonstrate the effectiveness of HIPPO, achieving a 4% improvement over various table reasoning models. Further analysis reveals that HIPPO not only enhances the table reasoning capability based on unimodal representations but also facilitates the extraction of complementary semantics across modalities. The code is available at https://github.com/NEUIR/HIPPO.

Method: Models dialogue with LLMs using state-space representations and introduces a novel neural barrier function (NBF) to detect and filter harmful queries emerging from evolving contexts proactively. Learns a safety predictor that accounts for adversarial queries to prevent context drift toward jailbreaks.

Result: Extensive experiments show NBF-based safety steering outperforms safety alignment, prompt-based steering, and lightweight LLM guardrails baselines, offering stronger defenses against multi-turn jailbreaks while maintaining better trade-off among safety, helpfulness, and over-refusal.

Conclusion: The proposed safety steering framework grounded in safe control theory effectively ensures invariant safety in multi-turn dialogues against jailbreaking attacks.

Abstract: Large language models (LLMs) are shown to be vulnerable to jailbreaking attacks where adversarial prompts are designed to elicit harmful responses. While existing defenses effectively mitigate single-turn attacks by detecting and filtering unsafe inputs, they fail against multi-turn jailbreaks that exploit contextual drift over multiple interactions, gradually leading LLMs away from safe behavior. To address this challenge, we propose a safety steering framework grounded in safe control theory, ensuring invariant safety in multi-turn dialogues. Our approach models the dialogue with LLMs using state-space representations and introduces a novel neural barrier function (NBF) to detect and filter harmful queries emerging from evolving contexts proactively. Our method achieves invariant safety at each turn of dialogue by learning a safety predictor that accounts for adversarial queries, preventing potential context drift toward jailbreaks. Extensive experiments under multiple LLMs show that our NBF-based safety steering outperforms safety alignment, prompt-based steering and lightweight LLM guardrails baselines, offering stronger defenses against multi-turn jailbreaks while maintaining a better trade-off among safety, helpfulness and over-refusal. Check out the website here https://sites.google.com/view/llm-nbf/home.

Method: Introduces ChemRAG-Bench benchmark with diverse chemistry tasks and integrated knowledge sources (scientific literature, PubChem, PubMed, textbooks, Wikipedia). Develops ChemRAG-Toolkit supporting 5 retrieval algorithms and 8 LLMs.

Result: RAG yields 17.4% average relative improvement over direct inference methods. Provides analysis on retriever architectures, corpus selection, and number of retrieved passages with practical recommendations.

Conclusion: ChemRAG-Bench and Toolkit provide comprehensive resources for evaluating and deploying RAG systems in chemistry, demonstrating significant performance gains and offering guidance for future research.

Abstract: Retrieval-augmented generation (RAG) has emerged as a powerful framework for enhancing large language models (LLMs) with external knowledge, particularly in scientific domains that demand specialized and dynamic information. Despite its promise, the application of RAG in the chemistry domain remains underexplored, primarily due to the lack of high-quality, domain-specific corpora and well-curated evaluation benchmarks. In this work, we introduce ChemRAG-Bench, a comprehensive benchmark designed to systematically assess the effectiveness of RAG across a diverse set of chemistry-related tasks. The accompanying chemistry corpus integrates heterogeneous knowledge sources, including scientific literature, the PubChem database, PubMed abstracts, textbooks, and Wikipedia entries. In addition, we present ChemRAG-Toolkit, a modular and extensible RAG toolkit that supports five retrieval algorithms and eight LLMs. Using ChemRAG-Toolkit, we demonstrate that RAG yields a substantial performance gain – achieving an average relative improvement of 17.4% over direct inference methods. We further conduct in-depth analyses on retriever architectures, corpus selection, and the number of retrieved passages, culminating in practical recommendations to guide future research and deployment of RAG systems in the chemistry domain. The code and data is available at https://chemrag.github.io.

Method: Caprese uses resource-efficient distillation focused primarily on feedforward blocks. It adds roughly 1% of additional parameters, uses only 20K synthetic training samples, and leaves original weights unperturbed. It integrates cleanly into existing model layers to reduce latency.

Result: Recovers much if not all reasoning capabilities lost from efficient inference for thinking LLMs without harming language tasks for instruct LLMs. Reduces active parameters (~2B cut for Gemma 2 9B and Llama 3.1 8B), reduces latency (>16% time-to-next-token reduction), and encourages response brevity (up to 8.5% fewer tokens).

Conclusion: Caprese provides an effective solution to recover math reasoning capabilities compromised by efficient inference methods, offering significant computational benefits while maintaining performance across both reasoning and language tasks.

Abstract: Due to long generations, large language model (LLM) math reasoning demands significant computational resources and time. While many existing efficient inference methods have been developed with excellent performance preservation on language tasks, they often severely degrade math performance. In this paper, we propose Caprese, a resource-efficient distillation method to recover lost capabilities from deploying efficient inference methods, focused primarily in feedforward blocks. With original weights unperturbed, roughly 1% of additional parameters, and only 20K synthetic training samples, we are able to recover much if not all of the reasoning capabilities lost from efficient inference for thinking LLMs and without harm to language tasks for instruct LLMs. Moreover, Caprese slashes the number of active parameters (~2B cut for Gemma 2 9B and Llama 3.1 8B) and integrates cleanly into existing model layers to reduce latency (>16% time-to-next-token reduction) while encouraging response brevity (up to 8.5% fewer tokens).

Method: Proposes RedTeamCUA framework with a novel hybrid sandbox combining VM-based OS environment with Docker-based web platforms, enabling flexible adversarial scenario configuration and direct initialization at injection points.

Result: Benchmarking reveals significant vulnerabilities: Claude 3.7 Sonnet shows 42.9% attack success rate, Claude 4.5 Sonnet shows 60% ASR, and even the most secure CUA (Operator) has 7.6% ASR. Attempt rates reach up to 92.5%.

Conclusion: RedTeamCUA provides essential framework for systematic CUA vulnerability analysis, highlighting urgent need for robust defenses against indirect prompt injection before real-world deployment.

Abstract: Computer-use agents (CUAs) promise to automate complex tasks across operating systems (OS) and the web, but remain vulnerable to indirect prompt injection. Current evaluations of this threat either lack support realistic but controlled environments or ignore hybrid web-OS attack scenarios involving both interfaces. To address this, we propose RedTeamCUA, an adversarial testing framework featuring a novel hybrid sandbox that integrates a VM-based OS environment with Docker-based web platforms. Our sandbox supports key features tailored for red teaming, such as flexible adversarial scenario configuration, and a setting that decouples adversarial evaluation from navigational limitations of CUAs by initializing tests directly at the point of an adversarial injection. Using RedTeamCUA, we develop RTC-Bench, a comprehensive benchmark with 864 examples that investigate realistic, hybrid web-OS attack scenarios and fundamental security vulnerabilities. Benchmarking current frontier CUAs identifies significant vulnerabilities: Claude 3.7 Sonnet | CUA demonstrates an ASR of 42.9%, while Operator, the most secure CUA evaluated, still exhibits an ASR of 7.6%. Notably, CUAs often attempt to execute adversarial tasks with an Attempt Rate as high as 92.5%, although failing to complete them due to capability limitations. Nevertheless, we observe concerning high ASRs in realistic end-to-end settings, with the strongest-to-date Claude 4.5 Sonnet | CUA exhibiting the highest ASR of 60%, indicating that CUA threats can already result in tangible risks to users and computer systems. Overall, RedTeamCUA provides an essential framework for advancing realistic, controlled, and systematic analysis of CUA vulnerabilities, highlighting the urgent need for robust defenses to indirect prompt injection prior to real-world deployment.

Method: Proposes MedEditBench framework with: 1) New medical knowledge editing benchmark, 2) Three different knowledge editing paradigms to assess impact of different knowledge sources, 3) Self-Generated Rationale Editing (SGR-Edit) that uses model-derived rationales as target knowledge for editing to uncover underlying reasoning process.

Result: Current KE methods result in only superficial memorization of injected information, failing to generalize to new scenarios. SGR-Edit demonstrates significant improvements over existing KE approaches. Provides insights into medical knowledge localization in LLMs and impact of sequential editing on evolving knowledge.

Conclusion: Medical knowledge editing requires more sophisticated approaches than current methods. SGR-Edit offers better generalization by leveraging model-derived rationales. The work provides practical guidance for implementing KE methods in real-world medical applications.

Abstract: Recently, knowledge editing (KE) has emerged as a promising approach to update specific facts in Large Language Models (LLMs) without the need for full retraining. Despite the effectiveness in general-domain benchmarks, their applicability to complex medical domain remains largely unexplored. Medical knowledge editing is particularly challenging, as it requires LLMs to internalize the knowledge and generalize to unseen scenarios for effective and interpretable decision-making. In this work, we propose a novel framework called MedEditBench to rigorously evaluate the effectiveness of existing KE methods in the medical domain. In MedEditBench, we introduce a new medical knowledge editing benchmark as well as three different knowledge editing paradigms, which are designed to assess the impact of different knowledge sources for editing. Our findings indicate that current KE methods result in only superficial memorization of the injected information, failing to generalize to new scenarios. To overcome this limitation, we present Self-Generated Rationale Editing (SGR-Edit), which utilizes model-derived rationales as the target knowledge for editing, thereby uncovering the underlying reasoning process and demonstrating significant improvements over existing KE approaches. Additionally, we offer deeper insights into medical knowledge editing, including the localization of medical knowledge in LLMs and the impact of sequential editing on evolving knowledge. This could provide practical guidance for implementing KE methods in real-world medical applications.

Method: HALT generates capability-aligned post-training data by splitting LLM responses into factual fragments (atomic statements or reasoning steps), using ground truth to identify incorrect fragments. During finetuning, incorrect fragments are either removed or replaced with “Unsure from Here” based on a tunable threshold that trades off response completeness vs correctness.

Result: HALT increases mean correctness of response fragments by 15% on average while improving F1 score (mean of completeness and correctness) by 4% compared to baselines. For Llama3-70B, correctness increased from 51% to 87% across four domains (biography, math, coding, medicine) while maintaining 53% of response completeness.

Conclusion: HALT effectively enables LLMs to recognize their limitations and abstain from generating unreliable content, significantly improving correctness while maintaining reasonable response completeness through tunable thresholds.

Abstract: Large Language Models (LLMs) currently respond to every prompt. However, they can produce incorrect answers when they lack knowledge or capability – a problem known as hallucination. We instead propose post-training an LLM to generate content only when confident in its correctness and to otherwise (partially) abstain. Specifically, our method, HALT, produces capability-aligned post-training data that encodes what the model can and cannot reliably generate. We generate this data by splitting responses of the pretrained LLM into factual fragments (atomic statements or reasoning steps), and use ground truth information to identify incorrect fragments. We achieve capability-aligned finetuning responses by either removing incorrect fragments or replacing them with “Unsure from Here” – according to a tunable threshold that allows practitioners to trade off response completeness and mean correctness of the response’s fragments. We finetune four open-source models for biography writing, mathematics, coding, and medicine with HALT for three different trade-off thresholds. HALT effectively trades off response completeness for correctness, increasing the mean correctness of response fragments by 15% on average, while resulting in a 4% improvement in the F1 score (mean of completeness and correctness of the response) compared to the relevant baselines. By tuning HALT for highest correctness, we train a single reliable Llama3-70B model with correctness increased from 51% to 87% across all four domains while maintaining 53% of the response completeness achieved with standard finetuning.

Method: Proposes reward-weighted fine-tuning as a practical offline RL approach for LLMs, treating it as a supervised fine-tuning problem where trajectories are weighted by their rewards. Applied to short-horizon question-answering policies of fixed length.

Result: Major gains in both optimized rewards and language quality compared to state-of-the-art methods based on SFT and direct preference optimization (DPO).

Conclusion: Reward-weighted fine-tuning provides an effective practical approach for offline RL with LLMs that directly optimizes rewards while maintaining language quality, outperforming existing methods.

Abstract: Offline reinforcement learning (RL) is a variant of RL where the policy is learned from a previously collected dataset of trajectories and rewards. In our work, we propose a practical approach to offline RL with large language models (LLMs). We recast the problem as reward-weighted fine-tuning, which can be solved using similar techniques to supervised fine-tuning (SFT). To showcase the value of our approach, we apply it to learning short-horizon question-answering policies of a fixed length, where the agent reasons about potential answers or asks clarifying questions. Our work stands in a stark contrast to state-of-the-art methods in this domain, based on SFT and direct preference optimization, which have additional hyper-parameters and do not directly optimize for rewards. We compare to them empirically, and report major gains in both optimized rewards and language quality.

Method: RuleReasoner uses a curated collection of rule-based reasoning tasks and a novel domain-aware dynamic sampling approach in RL that resamples training batches by updating domain weights based on historical rewards, facilitating domain balance and active learning schedules.

Result: RuleReasoner outperforms frontier large reasoning models by significant margins: Δ4.1% on eight in-distribution tasks and Δ10.4% on three out-of-distribution tasks over OpenAI-o1, while also exhibiting higher computational efficiency.

Conclusion: The proposed RuleReasoner method effectively addresses rule-based reasoning challenges through dynamic sampling in RL, achieving superior performance and efficiency compared to existing approaches.

Abstract: Rule-based reasoning is acknowledged as one of the fundamental problems of reasoning. While recent studies show that large reasoning models (LRMs) have remarkable reasoning capabilities enhanced by reinforcement learning (RL), real applications still face severe challenges due to variations in rule formats, types, and complexity. To mitigate this issue, we introduce RuleReasoner, an effective method for rule-based reasoning via a wide collection of curated tasks and a novel domain-aware dynamic sampling approach in RL. Specifically, RuleReasoner resamples each training batch by updating the domain weights based on historical rewards. This facilitates domain balance and active learning schedules for RL, obviating static mix-training engineered by human. Evaluations on in-distribution (ID) and out-of-distribution (OOD) benchmarks reveal that RuleReasoner outperforms frontier LRMs by a significant margin ($Δ$4.1% on eight ID tasks and $Δ$10.4% on three OOD tasks over OpenAI-o1). Notably, our approach also exhibits higher computational efficiency compared to prior methods.

Method: Proposes a flexible external memory framework based on knowledge graphs that constructs and updates memory automatically using LLMs. Uses AriGraph architecture with hybrid graph design supporting standard edges and two types of hyper-edges for semantic and temporal representations. Supports diverse retrieval mechanisms including A*, water-circle traversal, beam search, and hybrid methods.

Result: Evaluated on TriviaQA, HotpotQA, and DiaASQ benchmarks, showing different memory and retrieval configurations yield optimal performance depending on task. Extended DiaASQ with temporal annotations and contradictory statements, demonstrating system robustness in managing temporal dependencies and context-aware reasoning.

Conclusion: The knowledge graph-based memory framework effectively addresses LLM limitations in structured memory and long-term interaction scaling, providing adaptable solutions for different tasks through configurable memory and retrieval mechanisms.

Abstract: Personalizing language models that effectively incorporating user interaction history remains a central challenge in development of adaptive AI systems. While large language models (LLMs), combined with Retrieval-Augmented Generation (RAG), have improved factual accuracy, they often lack structured memory and fail to scale in complex, long-term interactions. To address this, we propose a flexible external memory framework based on knowledge graph, which construct and update memory model automatically by LLM itself. Building upon the AriGraph architecture, we introduce a novel hybrid graph design that supports both standard edges and two types of hyper-edges, enabling rich and dynamic semantic and temporal representations. Our framework also supports diverse retrieval mechanisms, including A*, water-circle traversal, beam search and hybrid methods, making it adaptable to different datasets and LLM capacities. We evaluate our system on three benchmarks: TriviaQA, HotpotQA, DiaASQ and demonstrate that different memory and retrieval configurations yield optimal performance depending on the task. Additionally, we extend the DiaASQ benchmark with temporal annotations and internally contradictory statements, showing that our system remains robust and effective in managing temporal dependencies and context-aware reasoning.

Method: Integrates Direct Preference Optimization (DPO) with curriculum learning, using fact-checked explanations as preferred responses and LLM outputs as non-preferred responses. Introduces two parameters (Actuality and Finesse) into DPO loss function to enhance explanation quality.

Result: Experiments with various LLMs (Mistral, Llama, Gemma) and PLMs (mBART, mT5) confirm the framework’s effectiveness in generating coherent, contextually relevant explanations for Hindi news.

Conclusion: The proposed scalable approach successfully combats misinformation and extends automated explanation generation capabilities to low-resource languages, bridging the gap in misinformation detection tools.

Abstract: In an era of rampant misinformation, generating reliable news explanations is vital, especially for under-represented languages like Hindi. Lacking robust automated tools, Hindi faces challenges in scaling misinformation detection. To bridge this gap, we propose a novel framework integrating Direct Preference Optimization (DPO) with curriculum learning to align machine-generated explanations with human reasoning. Fact-checked explanations from credible sources serve as preferred responses, while LLM outputs highlight system limitations and serve as non-preferred responses. To refine task-specific alignment, we introduce two key parameters – Actuality and Finesse – into the DPO loss function, enhancing explanation quality and consistency. Experiments with LLMs (Mistral, Llama, Gemma) and PLMs (mBART, mT5) confirm the framework’s effectiveness in generating coherent, contextually relevant explanations. This scalable approach combats misinformation and extends automated explanation generation to low-resource languages.

Method: GEPA samples trajectories (reasoning, tool calls, outputs), reflects on them in natural language to diagnose problems, proposes and tests prompt updates, and combines lessons from the Pareto frontier of attempts. It uses genetic-Pareto optimization to efficiently learn from few rollouts.

Result: GEPA outperforms GRPO by 6% on average (up to 20%) while using up to 35x fewer rollouts. It also beats leading prompt optimizer MIPROv2 by over 10% (e.g., +12% accuracy on AIME-2025). Shows promise as inference-time search for code optimization.

Conclusion: Natural language reflection enables efficient prompt optimization with far fewer rollouts than RL methods, demonstrating the power of leveraging LLMs’ interpretable nature for learning.

Abstract: Large language models (LLMs) are increasingly adapted to downstream tasks via reinforcement learning (RL) methods like Group Relative Policy Optimization (GRPO), which often require thousands of rollouts to learn new tasks. We argue that the interpretable nature of language often provides a much richer learning medium for LLMs, compared to policy gradients derived from sparse, scalar rewards. To test this, we introduce GEPA (Genetic-Pareto), a prompt optimizer that thoroughly incorporates natural language reflection to learn high-level rules from trial and error. Given any AI system containing one or more LLM prompts, GEPA samples trajectories (e.g., reasoning, tool calls, and tool outputs) and reflects on them in natural language to diagnose problems, propose and test prompt updates, and combine complementary lessons from the Pareto frontier of its own attempts. As a result of GEPA’s design, it can often turn even just a few rollouts into a large quality gain. Across six tasks, GEPA outperforms GRPO by 6% on average and by up to 20%, while using up to 35x fewer rollouts. GEPA also outperforms the leading prompt optimizer, MIPROv2, by over 10% (e.g., +12% accuracy on AIME-2025), and demonstrates promising results as an inference-time search strategy for code optimization. We release our code at https://github.com/gepa-ai/gepa .

Method: Created RoD-TAL dataset with Romanian driving test questions (text and image-based) with expert annotations. Evaluated retrieval-augmented generation (RAG) pipelines, dense retrievers, and reasoning-optimized models on Information Retrieval, Question Answering, Visual IR, and Visual QA tasks.

Result: Domain-specific fine-tuning significantly improves retrieval performance. Chain-of-thought prompting and specialized reasoning models enhance QA accuracy, surpassing minimum passing grades required for driving exams.

Conclusion: LLMs and VLMs show potential for legal education applications but have limitations. The work provides resources for Romanian language legal AI applications and demonstrates effective multimodal reasoning approaches.

Abstract: The intersection of AI and legal systems presents a growing need for tools that support legal education, particularly in under-resourced languages such as Romanian. In this work, we aim to evaluate the capabilities of Large Language Models (LLMs) and Vision-Language Models (VLMs) in understanding and reasoning about the Romanian driving law through textual and visual question-answering tasks. To facilitate this, we introduce RoD-TAL, a novel multimodal dataset comprising Romanian driving test questions, text-based and image-based, along with annotated legal references and explanations written by human experts. We implement and assess retrieval-augmented generation (RAG) pipelines, dense retrievers, and reasoning-optimized models across tasks, including Information Retrieval (IR), Question Answering (QA), Visual IR, and Visual QA. Our experiments demonstrate that domain-specific fine-tuning significantly enhances retrieval performance. At the same time, chain-of-thought prompting and specialized reasoning models improve QA accuracy, surpassing the minimum passing grades required for driving exams. We highlight the potential and limitations of applying LLMs and VLMs to legal education. We release the code and resources through the GitHub repository.

Method: Benchmarked multiple generation strategies guided by structured supervision (Intent Summaries, Topic Flows, QA Forms) across languages, evaluated on AutoQA task, and introduced diagnostic framework with 17 metrics across four dimensions: Emotional/Sentiment Arcs, Linguistic Complexity, Interaction Style, and Conversational Properties.

Result: Prompts optimized on real transcripts consistently outperform those optimized on synthetic transcripts; current synthetic data shows deficiencies in sentiment fidelity, disfluency modeling, behavioral variation, and conversational realism despite structured supervision.

Conclusion: Diagnostic, metric-driven evaluation is crucial for synthetic conversation generation intended for downstream applications, as current methods still fall short of capturing full realism of real agent-customer interactions.

Abstract: Synthetic data is increasingly critical for contact centers, where privacy constraints and data scarcity limit the availability of real conversations. However, generating synthetic dialogues that are realistic and useful for downstream applications remains challenging. In this work, we benchmark multiple generation strategies guided by structured supervision on call attributes (Intent Summaries, Topic Flows, and Quality Assurance (QA) Forms) across multiple languages. To test downstream utility, we evaluate synthetic transcripts on an automated quality assurance (AutoQA) task, finding that prompts optimized on real transcripts consistently outperform those optimized on synthetic transcripts. These results suggest that current synthetic transcripts fall short in capturing the full realism of real agent-customer interactions. To highlight these downstream gaps, we introduce a diagnostic evaluation framework comprising 17 metrics across four dimensions: (1) Emotional and Sentiment Arcs, (2) Linguistic Complexity, (3) Interaction Style, and (4) Conversational Properties. Our analysis shows that even with structured supervision, current generation strategies exhibit measurable deficiencies in sentiment fidelity, disfluency modeling, behavioral variation, and conversational realism. Together, these results highlight the importance of diagnostic, metric-driven evaluation for synthetic conversation generation intended for downstream applications.

Method: PACS treats outcome rewards as predictable labels and reformulates RLVR as a supervised learning task over a score function parameterized by the policy model, optimized using cross-entropy loss instead of traditional policy gradient methods.

Result: PACS significantly outperforms strong open-source models and RLVR baselines, achieving substantial average gains of +8.26% (4B) and +9.57% (8B) over base models on reasoning tasks.

Conclusion: PACS offers a promising avenue for LLM post-training with verifiable rewards by providing more stable and efficient training through its supervised learning reformulation of the RLVR problem.

Abstract: Recent advances in Reinforcement Learning with Verifiable Rewards (RLVR) have empowered large language models (LLMs) to tackle challenging reasoning tasks such as mathematics and programming. Despite its promise, the RLVR paradigm poses significant challenges, as existing methods often suffer from sparse reward signals and unstable policy gradient updates, inherent to RL-based approaches. To address the challenges, we propose $\textbf{PACS}$, a novel RLVR framework that achieves im$\textbf{P}$licit $\textbf{A}$ctor $\textbf{C}$ritic coupling via a $\textbf{S}$upervised learning framework. By treating the outcome reward as a predictable label, we reformulate the RLVR problem into a supervised learning task over a score function parameterized by the policy model and optimized using cross-entropy loss. A detailed gradient analysis shows that this supervised formulation inherently recovers the classical policy gradient update while providing more stable and efficient training. Extensive experiments demonstrate that PACS significantly outperforms strong open-source models and RLVR baselines, yielding substantial average gains of $\textbf{+8.26%}$ (4B) and $\textbf{+9.57%}$ (8B) over base models offering a promising avenue for LLMs post-training with verifiable rewards. Our code and data are available as open source at https://github.com/ritzz-ai/PACS.

Method: Use crosscoders (sparse dictionary learning method) to track linear interpretable features across pre-training snapshots, analyzing feature formation timing and complexity development, with feature attribution analyses to establish causal connections.

Result: Most features form around specific points in training, with complex patterns emerging later; feature evolution shows causal connections to downstream performance; findings align with Transformer’s two-stage learning process (statistical learning phase followed by feature learning phase).

Conclusion: The work enables fine-grained tracking of representation progress during language model learning dynamics, providing insights into the previously opaque pre-training process and opening new research directions in model interpretability.

Abstract: Language models obtain extensive capabilities through pre-training. However, the pre-training process remains a black box. In this work, we track linear interpretable feature evolution across pre-training snapshots using a sparse dictionary learning method called crosscoders. We find that most features begin to form around a specific point, while more complex patterns emerge in later training stages. Feature attribution analyses reveal causal connections between feature evolution and downstream performance. Our feature-level observations are highly consistent with previous findings on Transformer’s two-stage learning process, which we term a statistical learning phase and a feature learning phase. Our work opens up the possibility to track fine-grained representation progress during language model learning dynamics.

Method: Created AECBench with 5-level cognition framework (Knowledge Memorization, Understanding, Reasoning, Calculation, Application), 23 tasks from authentic AEC practice, 4,800-question dataset crafted by engineers, and LLM-as-a-Judge evaluation approach.

Result: Evaluation of 9 LLMs showed clear performance decline across cognitive levels - proficient in foundational tasks but significant deficits in interpreting tables, complex reasoning/calculation, and generating domain-specific documents.

Conclusion: The benchmark establishes groundwork for robust LLM integration into safety-critical engineering, revealing current limitations and guiding future development.

Abstract: Large language models (LLMs), as a novel information technology, are seeing increasing adoption in the Architecture, Engineering, and Construction (AEC) field. They have shown their potential to streamline processes throughout the building lifecycle. However, the robustness and reliability of LLMs in such a specialized and safety-critical domain remain to be evaluated. To address this challenge, this paper establishes AECBench, a comprehensive benchmark designed to quantify the strengths and limitations of current LLMs in the AEC domain. The benchmark features a five-level, cognition-oriented evaluation framework (i.e., Knowledge Memorization, Understanding, Reasoning, Calculation, and Application). Based on the framework, 23 representative evaluation tasks were defined. These tasks were derived from authentic AEC practice, with scope ranging from codes retrieval to specialized documents generation. Subsequently, a 4,800-question dataset encompassing diverse formats, including open-ended questions, was crafted primarily by engineers and validated through a two-round expert review. Furthermore, an “LLM-as-a-Judge” approach was introduced to provide a scalable and consistent methodology for evaluating complex, long-form responses leveraging expert-derived rubrics. Through the evaluation of nine LLMs, a clear performance decline across five cognitive levels was revealed. Despite demonstrating proficiency in foundational tasks at the Knowledge Memorization and Understanding levels, the models showed significant performance deficits, particularly in interpreting knowledge from tables in building codes, executing complex reasoning and calculation, and generating domain-specific documents. Consequently, this study lays the groundwork for future research and development aimed at the robust and reliable integration of LLMs into safety-critical engineering practices.

Method: Proposes Dual adaptive Cache (d²Cache) with two-stage fine-grained selection strategy: identifies tokens to update KV states adaptively at each decoding step while caching remaining tokens’ KV states for reuse.

Result: Achieves substantial inference speedups on LLaDA and Dream models while yielding consistent improvements in generation quality, plus enables quasi left-to-right generation and mitigates premature overconfidence.

Conclusion: d²Cache provides effective training-free solution for accelerating diffusion-based LLM inference while enhancing generation quality through adaptive caching strategy.

Abstract: Diffusion-based large language models (dLLMs), despite their promising performance, still suffer from inferior inference efficiency. This is because dLLMs rely on bidirectional attention and cannot directly benefit from the standard key-value (KV) cache as autoregressive models (ARMs) do. To tackle this issue, we introduce \textit{Dual aDaptive Cache} (d$^2$Cache), which is a training-free approximate KV cache framework for accelerating dLLM inference. d$^2$Cache features a two-stage fine-grained selection strategy to identify tokens and adaptively update their KV states at each decoding step, while caching the KV states of the remaining tokens for reuse. Furthermore, d$^2$Cache naturally offers a more reliable decoding alternative, which can enable quasi left-to-right generation and mitigate premature overconfidence in tokens at the end of the sequence. Extensive experimental results on two representative dLLMs (\ie, LLaDA and Dream) demonstrate that d$^2$Cache not only achieves substantial inference speedups, but also yields consistent improvements in generation quality. The code is available at https://github.com/Kamichanw/d2Cache.

Method: Developed a pragmatic-inference approach based on metapragmatic links and moral foundations theory. The method helps LLMs acquire links between moral reasoning objectives and relevant social variables for given moral situations, adapted to three different moral reasoning tasks.

Result: Experimental results show the approach significantly enhances LLMs’ generalization in moral reasoning, demonstrating adaptability across different tasks.

Conclusion: The pragmatic-inference approach successfully bridges the gap in moral reasoning and paves the way for future research to utilize pragmatic inference in various moral reasoning tasks.

Abstract: While moral reasoning has emerged as a promising research direction for large language models (LLMs), achieving robust generalization remains a critical challenge. This challenge arises from the gap between what is said and what is morally implied. In this paper, we build on metapragmatic links and the moral foundations theory to close the gap. Specifically, we develop a pragmatic-inference approach that facilitates LLMs, for a given moral situation, to acquire the metapragmantic links between moral reasoning objectives and the social variables that affect them. This approach is adapted to three different moral reasoning tasks to demonstrate its adaptability and generalizability. Experimental results demonstrate that our approach significantly enhances LLMs’ generalization in moral reasoning, paving the road for future research to utilize pragmatic inference in various moral reasoning tasks.

Method: Introduces BiasFreeBench benchmark comparing 8 mainstream bias mitigation techniques (4 prompting-based, 4 training-based) on two test scenarios (multi-choice QA and open-ended multi-turn QA) by reorganizing existing datasets into unified query-response setting

Result: Systematic comparison of debiasing performance across key dimensions: prompting vs. training paradigm, model size, and generalization of different training strategies to unseen bias types

Conclusion: Provides a unified testbed for bias mitigation research with response-level evaluation metrics that better reflect real-world use cases

Abstract: Existing studies on bias mitigation methods for large language models (LLMs) use diverse baselines and metrics to evaluate debiasing performance, leading to inconsistent comparisons among them. Moreover, their evaluations are mostly based on the comparison between LLMs’ probabilities of biased and unbiased contexts, which ignores the gap between such evaluations and real-world use cases where users interact with LLMs by reading model responses and expect fair and safe outputs rather than LLMs’ probabilities. To enable consistent evaluation across debiasing methods and bridge this gap, we introduce BiasFreeBench, an empirical benchmark that comprehensively compares eight mainstream bias mitigation techniques (covering four prompting-based and four training-based methods) on two test scenarios (multi-choice QA and open-ended multi-turn QA) by reorganizing existing datasets into a unified query-response setting. We further introduce a response-level metric, Bias-Free Score, to measure the extent to which LLM responses are fair, safe, and anti-stereotypical. Debiasing performances are systematically compared and analyzed across key dimensions: the prompting vs. training paradigm, model size, and generalization of different training strategies to unseen bias types. We release our benchmark, aiming to establish a unified testbed for bias mitigation research.

Method: Created a benchmark with 1.6k queries across five analytical tasks and 9.1k conversations, then evaluated 16 popular embedding models to measure conversational data retrieval performance.

Result: Even the best models only achieve around NDCG@10 of 0.51, showing a substantial gap between document and conversational data retrieval capabilities, with unique challenges identified in implicit state recognition, turn dynamics, and contextual references.

Conclusion: The CDR benchmark establishes a reliable standard for conversational data retrieval evaluation, reveals significant limitations in current models, and provides practical tools including query templates and error analysis for advancing the field.

Abstract: We present the Conversational Data Retrieval (CDR) benchmark, the first comprehensive test set for evaluating systems that retrieve conversation data for product insights. With 1.6k queries across five analytical tasks and 9.1k conversations, our benchmark provides a reliable standard for measuring conversational data retrieval performance. Our evaluation of 16 popular embedding models shows that even the best models reach only around NDCG@10 of 0.51, revealing a substantial gap between document and conversational data retrieval capabilities. Our work identifies unique challenges in conversational data retrieval (implicit state recognition, turn dynamics, contextual references) while providing practical query templates and detailed error analysis across different task categories. The benchmark dataset and code are available at https://github.com/l-yohai/CDR-Benchmark.

Method: Formulates realistic attacks as constrained optimization over input prompt space with semantic equivalence and coherence constraints. Introduces constraint-preserving zeroth-order method to search for adversarial yet feasible prompts.

Result: SECA achieves higher attack success rates while maintaining almost no semantic equivalence or coherence errors compared to existing methods on open-ended multiple-choice QA tasks.

Conclusion: SECA demonstrates LLM sensitivity to realistic prompt variations, highlighting hallucination risks in practical settings and providing a framework for more realistic adversarial evaluation.

Abstract: Large Language Models (LLMs) are increasingly deployed in high-risk domains. However, state-of-the-art LLMs often exhibit hallucinations, raising serious concerns about their reliability. Prior work has explored adversarial attacks to elicit hallucinations in LLMs, but these methods often rely on unrealistic prompts, either by inserting nonsensical tokens or by altering the original semantic intent. Consequently, such approaches provide limited insight into how hallucinations arise in real-world settings. In contrast, adversarial attacks in computer vision typically involve realistic modifications to input images. However, the problem of identifying realistic adversarial prompts for eliciting LLM hallucinations remains largely underexplored. To address this gap, we propose Semantically Equivalent and Coherent Attacks (SECA), which elicit hallucinations via realistic modifications to the prompt that preserve its meaning while maintaining semantic coherence. Our contributions are threefold: (i) we formulate finding realistic attacks for hallucination elicitation as a constrained optimization problem over the input prompt space under semantic equivalence and coherence constraints; (ii) we introduce a constraint-preserving zeroth-order method to effectively search for adversarial yet feasible prompts; and (iii) we demonstrate through experiments on open-ended multiple-choice question answering tasks that SECA achieves higher attack success rates while incurring almost no semantic equivalence or semantic coherence errors compared to existing methods. SECA highlights the sensitivity of both open-source and commercial gradient-inaccessible LLMs to realistic and plausible prompt variations. Code is available at https://github.com/Buyun-Liang/SECA.

Method: Two key components: (1) A cognitive science-inspired reasoning mechanism that formulates RE as a series of text-processing steps, and (2) An RL optimization process with a novel reward function that rewards generating relation keywords using an automatically constructed keywords dictionary.

Result: Achieves 24.65% F1 on One-shot NYT29 with Qwen2.5-15B-Instruct, surpassing prior reasoning-based designs. RL optimization provides +23.46% absolute improvement. Models trained on NYT29 achieve +16.9% F1 gain on out-of-distribution WIKIDATA. Human evaluation shows 54% relative increase in explanation quality ratings.

Conclusion: CogRE successfully improves both accuracy and explainability in relation extraction by combining cognitive-structured reasoning with RL optimization, addressing key failure patterns in one-shot RE and providing better language-based explanations.

Abstract: We introduce CogRE, a novel framework for relation extraction (RE), enhancing RE from both accuracy and explainability. The framework has two key components: (i) a reasoning mechanism that formulates relation extraction as a series of text-processing steps inspired by cognitive science, and (ii) an optimization process driven by a novel reinforcement learning (RL) reward function. Our framework introduces relation keywords and rewards generating such keywords using an automatically constructed keywords dictionary. This design addresses the lack of language-based explanations in traditional RE and provides supervision for explanation during RL training. Our experiments show that CogRE improves explanation quality by addressing two common failure patterns in one-shot RE: poor attention focus and limited one-shot learning capability. For example, our cognitive-structured reasoning with Qwen2.5-15B-Instruct on One-shot NYT29 achieves 24.65% F1, surpassing prior reasoning-based designs. Optimizing this approach with RL using our reward further improves performance by +23.46% (absolute). Further, models trained on NYT29 with our reward achieve a +16.9% F1 gain on out-of-distribution WIKIDATA. Finally, human evaluation shows that our best model generates relational keywords closely aligned with gold labels, increasing human explanation quality ratings by 54% (relative).

Method: Proposes training-free fingerprinting based on weight matrices. Uses Linear Assignment Problem (LAP) and unbiased Centered Kernel Alignment (CKA) similarity to neutralize effects of parameter manipulations, creating robust similarity metric.

Result: Tested on 60 positive and 90 negative model pairs, method shows exceptional robustness against all six post-training categories with near-zero false positive risk. Achieves perfect scores on all classification metrics, establishes reliable model lineage verification. Computation completes within 30s on NVIDIA 3090 GPU.

Conclusion: Method provides strong basis for reliable LLM lineage verification that is robust to various post-training modifications, addressing intellectual property protection needs for model owners and third parties.

Abstract: Protecting the intellectual property of large language models (LLMs) is crucial, given the substantial resources required for their training. Consequently, there is an urgent need for both model owners and third parties to determine whether a suspect LLM is trained from scratch or derived from an existing base model. However, the intensive post-training processes that models typically undergo-such as supervised fine-tuning, extensive continued pretraining, reinforcement learning, multi-modal extension, pruning, and upcycling-pose significant challenges to reliable identification. In this work, we propose a training-free fingerprinting method based on weight matrices. We leverage the Linear Assignment Problem (LAP) and an unbiased Centered Kernel Alignment (CKA) similarity to neutralize the effects of parameter manipulations, yielding a highly robust and high-fidelity similarity metric. On a comprehensive testbed of 60 positive and 90 negative model pairs, our method demonstrates exceptional robustness against all six aforementioned post-training categories while exhibiting a near-zero risk of false positives. By achieving perfect scores on all classification metrics, our approach establishes a strong basis for reliable model lineage verification. Moreover, the entire computation completes within 30s on an NVIDIA 3090 GPU. The code is available at https://github.com/LUMIA-Group/AWM.

Method: Proposes MemoTime framework with: 1) Hierarchical Tree of Time decomposition of complex temporal questions, 2) Operator-aware reasoning with monotonic timestamp enforcement and multi-entity constraints, 3) Dynamic evidence retrieval layer with operator-specific strategies, 4) Self-evolving experience memory storing verified reasoning traces, toolkit decisions, and sub-question embeddings for cross-type reuse.

Result: Achieves state-of-the-art results on multiple temporal QA benchmarks, outperforming strong baselines by up to 24.0%. Enables smaller models (e.g., Qwen3-4B) to achieve reasoning performance comparable to GPT-4-Turbo.

Conclusion: MemoTime effectively addresses temporal reasoning challenges in LLMs through structured grounding, recursive reasoning, and continual experience learning, demonstrating significant improvements in temporal question answering performance.

Abstract: Large Language Models (LLMs) have achieved impressive reasoning abilities, but struggle with temporal understanding, especially when questions involve multiple entities, compound operators, and evolving event sequences. Temporal Knowledge Graphs (TKGs), which capture vast amounts of temporal facts in a structured format, offer a reliable source for temporal reasoning. However, existing TKG-based LLM reasoning methods still struggle with four major challenges: maintaining temporal faithfulness in multi-hop reasoning, achieving multi-entity temporal synchronization, adapting retrieval to diverse temporal operators, and reusing prior reasoning experience for stability and efficiency. To address these issues, we propose MemoTime, a memory-augmented temporal knowledge graph framework that enhances LLM reasoning through structured grounding, recursive reasoning, and continual experience learning. MemoTime decomposes complex temporal questions into a hierarchical Tree of Time, enabling operator-aware reasoning that enforces monotonic timestamps and co-constrains multiple entities under unified temporal bounds. A dynamic evidence retrieval layer adaptively selects operator-specific retrieval strategies, while a self-evolving experience memory stores verified reasoning traces, toolkit decisions, and sub-question embeddings for cross-type reuse. Comprehensive experiments on multiple temporal QA benchmarks show that MemoTime achieves overall state-of-the-art results, outperforming the strong baseline by up to 24.0%. Furthermore, MemoTime enables smaller models (e.g., Qwen3-4B) to achieve reasoning performance comparable to that of GPT-4-Turbo.

Method: Presents EQSPEC (first algorithm guaranteeing output equivalence) and EXSPEC (reduces alignment overhead through cross-batch scheduling). Formalizes synchronization invariants, analyzes cost structure showing alignment overhead grows superlinearly, and introduces dynamic grouping of same-length sequences.

Result: Achieves up to 3x throughput improvement at batch size 8 on SpecBench across Vicuna-7B/68M, Qwen3-8B/0.6B, and GLM-4-9B/0.6B pairs. Methods achieve 95% decoding-equivalence, with residual divergence attributable to floating-point non-determinism in GPU inference, not the synchronization failures that cause near-zero equivalence of prior methods.

Conclusion: Presents the first authentic batch speculative decoding framework that solves fundamental synchronization problems, guaranteeing output equivalence while achieving significant throughput improvements, addressing a critical correctness issue in existing implementations.

Abstract: Speculative decoding must produce outputs distribution identical to standard autoregressive generation-this output equivalence is not an optimization target but the defining criterion of valid speculative decoding. We demonstrate that all existing batch speculative decoding implementations violate this fundamental requirement, producing corrupted outputs ranging from repetitive tokens to gibberish. These failures stem from the ragged tensor problem: sequences in the same batch accept different numbers of draft tokens, desynchronizing position IDs, attention masks, and KV-cache state. We present the first authentic batch speculative decoding framework. We (1) formalize the synchronization invariants that valid batch speculative decoding must satisfy, (2) present EQSPEC, the first algorithm that guarantees output equivalence, and analyze its cost structure to show that alignment overhead grows superlinearly and consumes up to 40% of computation, and (3) introduce EXSPEC, which reduces this overhead through cross-batch scheduling that dynamically groups same-length sequences. On SpecBench across Vicuna-7B/68M, Qwen3-8B/0.6B, and GLM-4-9B/0.6B pairs, our methods achieve up to 3x throughput improvement at batch size 8 while maintaining algorithmic correctness. Our methods achieve 95% decoding-equivalence, with residual divergence attributable to floating-point non-determinism in GPU inference, not the synchronization failures that cause near-zero equivalence of prior methods. Our code is available at https://github.com/eBay/spec_dec.

Method: Proposes Pairwise Rotation Quantization (ParoQuant) combining hardware-efficient independent Givens rotations with channel-wise scaling to even out magnitudes across channels and narrow dynamic range within each quantization group. Also includes co-designed inference kernels to exploit GPU parallelism while keeping rotations and scaling lightweight at runtime.

Result: Under weight-only quantization, ParoQuant achieves average 2.4% accuracy improvement over AWQ on reasoning tasks with less than 10% overhead. Also matches accuracy of state-of-the-art weight-activation quantization methods.

Conclusion: ParoQuant effectively addresses outlier issues in LLM quantization, enabling more efficient and accurate deployment of reasoning LLMs through hardware-efficient rotation-based quantization with minimal inference overhead.

Abstract: Post-training quantization (PTQ) compresses the weights and activations of large language models (LLMs) into low-precision representations to reduce memory footprint and accelerate inference. However, the presence of outliers in weights and activations often leads to large quantization errors and severe accuracy degradation, especially in recent reasoning LLMs where errors accumulate across long chains of thought. Existing PTQ methods either fail to sufficiently suppress outliers or introduce significant overhead during inference. In this paper, we propose Pairwise Rotation Quantization (ParoQuant), a PTQ method that combines hardware-efficient and optimizable independent Givens rotations with channel-wise scaling to even out the magnitudes across channels and narrow the dynamic range within each quantization group, effectively addressing the outlier issue. We further co-design the inference kernel to fully exploit GPU parallelism and keep the rotations and scaling lightweight at runtime. Under weight-only quantization, ParoQuant achieves an average 2.4% accuracy improvement over AWQ on reasoning tasks, with less than 10% overhead. ParoQuant also matches the accuracy of state-of-the-art weight-activation quantization methods. This paves the way for more efficient and accurate deployment of reasoning LLMs.

Method: The study applies SFT and RL techniques to enhance GPT-2’s therapeutic dialogue capabilities. It restructures input formats to process contextual information and emotional states alongside user input, and employs a novel multi-component reward function that aligns outputs with professional therapeutic logic and annotated emotions rather than just lexical overlap.

Result: Results show substantial improvements through RL over baseline GPT-2 across multiple metrics: BLEU (0.0111), ROUGE-1 (0.1397), ROUGE-2 (0.0213), ROUGE-L (0.1317), and METEOR (0.0581). LLM evaluation confirmed high contextual relevance and professionalism, while RL achieved 99.34% emotion accuracy compared to 66.96% for baseline GPT-2.

Conclusion: The findings demonstrate RL’s effectiveness in developing therapeutic dialogue systems that can serve as valuable assistive tools for therapists while maintaining essential human clinical oversight. The approach addresses key limitations of existing methods for mental health applications.

Abstract: Mental health disorders impose a substantial global socioeconomic burden. While large language models (LLMs) offer 24/7, non-judgmental interactions to address this gap, pretrained models lack contextual coherence and emotional alignment for appropriate therapeutic dialogue. Existing methods suffer from three critical methodological gaps: 1) Supervised Fine-Tuning (SFT) produces repetitive, context-insensitive outputs that fail to balance clinical accuracy with genuine empathy; 2) Reinforcement Learning (RL)-based therapeutic systems rely on generic reward functions (e.g., BLEU, ROUGE) that prioritise lexical similarity over clinical-specific emotional appropriateness and contextual relevance; 3) LLMs are resource-intensive and pose data privacy risks, making local deployment in clinical settings infeasible. To address these gaps, this study investigates the application of SFT and RL techniques to enhance GPT-2’s capacity for therapeutic dialogue generation. The methodology restructured input formats to enable simultaneous processing of contextual information and emotional states alongside user input, employing a novel multi-component reward function that explicitly aligns model outputs with professional therapeutic logic (not just lexical overlap) and annotated emotions. Results demonstrated substantial improvements through RLs over baseline GPT-2 across multiple evaluation metrics: BLEU (0.0111), ROUGE-1 (0.1397), ROUGE-2 (0.0213), ROUGE-L (0.1317), and METEOR (0.0581). LLM evaluation confirmed high contextual relevance and professionalism, while RL achieved 99.34% emotion accuracy compared to 66.96% for baseline GPT-2. These findings demonstrate RL’s effectiveness in developing therapeutic dialogue systems that can serve as valuable assistive tools for therapists, while maintaining essential human clinical oversight.

Method: Created MusWikiDB (3.2M passages from 144K music Wikipedia pages) and ArtistMus benchmark (1,000 questions on 500 artists with metadata). Evaluated retrieval-augmented generation (RAG) for music question answering, comparing open-source and proprietary models, and conducted RAG-style fine-tuning.

Result: RAG significantly improves factual accuracy: open-source models gain up to +56.8 percentage points (Qwen3 8B improves from 35.0% to 91.8%). RAG-style fine-tuning boosts both factual recall and contextual reasoning. MusWikiDB achieves ~6% higher accuracy and 40% faster retrieval than general Wikipedia corpus.

Conclusion: MusWikiDB and ArtistMus advance music information retrieval and domain-specific QA, establishing foundation for retrieval-augmented reasoning in culturally rich domains like music, with RAG significantly enhancing LLM performance on music-related tasks.

Abstract: Recent advances in large language models (LLMs) have transformed open-domain question answering, yet their effectiveness in music-related reasoning remains limited due to sparse music knowledge in pretraining data. While music information retrieval and computational musicology have explored structured and multimodal understanding, few resources support factual and contextual music question answering (MQA) grounded in artist metadata or historical context. We introduce MusWikiDB, a vector database of 3.2M passages from 144K music-related Wikipedia pages, and ArtistMus, a benchmark of 1,000 questions on 500 diverse artists with metadata such as genre, debut year, and topic. These resources enable systematic evaluation of retrieval-augmented generation (RAG) for MQA. Experiments show that RAG markedly improves factual accuracy; open-source models gain up to +56.8 percentage points (for example, Qwen3 8B improves from 35.0 to 91.8), approaching proprietary model performance. RAG-style fine-tuning further boosts both factual recall and contextual reasoning, improving results on both in-domain and out-of-domain benchmarks. MusWikiDB also yields approximately 6 percentage points higher accuracy and 40% faster retrieval than a general-purpose Wikipedia corpus. We release MusWikiDB and ArtistMus to advance research in music information retrieval and domain-specific question answering, establishing a foundation for retrieval-augmented reasoning in culturally rich domains such as music.

Method: Multi-perspective validation framework combining ensemble validation with dual reliability metrics (Cohen’s Kappa for inter-rater agreement, cosine similarity for semantic consistency). Configurable analysis parameters (1-6 seeds, temperature 0.0-2.0), custom prompt structures with variable substitution, and consensus theme extraction across any JSON format.

Result: Evaluated three LLMs (Gemini 2.5 Pro, GPT-4o, Claude 3.5 Sonnet) on psychedelic art therapy interview transcript. Gemini achieved highest reliability (κ=0.907, cosine=95.3%), followed by GPT-4o (κ=0.853, cosine=92.6%) and Claude (κ=0.842, cosine=92.1%). All models achieved high agreement (κ>0.80). Gemini identified 6 consensus themes, GPT-4o 5 themes, Claude 4 themes.

Conclusion: The framework successfully validates multi-run ensemble approach for LLM-based thematic analysis, providing transparent reliability metrics, flexible configuration, and structure-agnostic consensus extraction, establishing methodological foundations for reliable AI-assisted qualitative research.

Abstract: Qualitative research faces a critical reliability challenge: traditional inter-rater agreement methods require multiple human coders, are time-intensive, and often yield moderate consistency. We present a multi-perspective validation framework for LLM-based thematic analysis that combines ensemble validation with dual reliability metrics: Cohen’s Kappa ($κ$) for inter-rater agreement and cosine similarity for semantic consistency. Our framework enables configurable analysis parameters (1-6 seeds, temperature 0.0-2.0), supports custom prompt structures with variable substitution, and provides consensus theme extraction across any JSON format. As proof-of-concept, we evaluate three leading LLMs (Gemini 2.5 Pro, GPT-4o, Claude 3.5 Sonnet) on a psychedelic art therapy interview transcript, conducting six independent runs per model. Results demonstrate Gemini achieves highest reliability ($κ= 0.907$, cosine=95.3%), followed by GPT-4o ($κ= 0.853$, cosine=92.6%) and Claude ($κ= 0.842$, cosine=92.1%). All three models achieve a high agreement ($κ> 0.80$), validating the multi-run ensemble approach. The framework successfully extracts consensus themes across runs, with Gemini identifying 6 consensus themes (50-83% consistency), GPT-4o identifying 5 themes, and Claude 4 themes. Our open-source implementation provides researchers with transparent reliability metrics, flexible configuration, and structure-agnostic consensus extraction, establishing methodological foundations for reliable AI-assisted qualitative research.

Method: Developed Ara-HOPE framework with five-category error taxonomy and decision-tree annotation protocol; evaluated three MT systems (Jais, GPT-3.5, NLLB-200) using this framework to assess dialect-specific translation quality.

Result: Ara-HOPE effectively highlighted systematic performance differences between MT systems, revealing that dialect-specific terminology and semantic preservation remain the most persistent challenges in DA-MSA translation.

Conclusion: Ara-HOPE establishes a new framework for evaluating Dialectal Arabic MT quality and provides actionable guidance for improving dialect-aware MT systems, with annotation materials made publicly available for reproducibility.

Abstract: Dialectal Arabic to Modern Standard Arabic (DA-MSA) translation is a challenging task in Machine Translation (MT) due to significant lexical, syntactic, and semantic divergences between Arabic dialects and MSA. Existing automatic evaluation metrics and general-purpose human evaluation frameworks struggle to capture dialect-specific MT errors, hindering progress in translation assessment. This paper introduces Ara-HOPE, a human-centric post-editing evaluation framework designed to systematically address these challenges. The framework includes a five-category error taxonomy and a decision-tree annotation protocol. Through comparative evaluation of three MT systems (Arabic-centric Jais, general-purpose GPT-3.5, and baseline NLLB-200), Ara-HOPE effectively highlights systematic performance differences between these systems. Our results show that dialect-specific terminology and semantic preservation remain the most persistent challenges in DA-MSA translation. Ara-HOPE establishes a new framework for evaluating Dialectal Arabic MT quality and provides actionable guidance for improving dialect-aware MT systems. For reproducibility, we make the annotation files and related materials publicly available at https://github.com/abdullahalabdullah/Ara-HOPE

Method: The method uses Qwen-1.8B-Chat as base model with a unified JSON input-output contract for data loading, training, and inference. Key innovations include: 1) KV-off decoding to reduce variance, 2) Two orthogonal semantic regularizers (VAD-preserving constraint and lightweight external appraisal classifier), 3) Valence Flip augmentation for polarity sensitivity, and 4) A/B mixture sampling with entropy-aware temperature scheduling during supervised fine-tuning.

Result: EmoLoom-2B achieves strong performance on GoEmotions and EmpatheticDialogues datasets, and demonstrates robust cross-corpus generalization on DailyDialog. The pipeline is shown to be effective for emotion analysis while maintaining lightweight computational requirements.

Conclusion: The proposed EmoLoom-2B pipeline provides a dependable, budget-aware screening solution for emotion analysis tasks. It offers protocol-faithful evaluation, reduces avoidable variance, and serves as a practical foundation before heavier training or multimodal fusion approaches.

Abstract: We introduce EmoLoom-2B, a lightweight and reproducible pipeline that turns small language models under 2B parameters into fast screening candidates for joint emotion classification and Valence-Arousal-Dominance prediction. To ensure protocol-faithful and fair evaluation, we unify data loading, training, and inference under a single JSON input-output contract and remove avoidable variance by adopting KV-off decoding as the default setting. We incorporate two orthogonal semantic regularizers: a VAD-preserving constraint that aligns generated text with target VAD triples, and a lightweight external appraisal classifier that provides training-time guidance on goal attainment, controllability, certainty, and fairness without injecting long rationales. To improve polarity sensitivity, we introduce Valence Flip augmentation based on mirrored emotional pairs. During supervised fine-tuning, we apply A/B mixture sampling with entropy-aware temperature scheduling to balance coverage and convergence. Using Qwen-1.8B-Chat as the base model, EmoLoom-2B achieves strong performance on GoEmotions and EmpatheticDialogues, and demonstrates robust cross-corpus generalization on DailyDialog. The proposed recipe is budget-aware, auditable, and re-entrant, serving as a dependable screening pass before heavier training or multimodal fusion.

Method: Created a dataset of 505 questions across 7 domains (petrophysics, petroleum geology, reservoir engineering, etc.) using a reasoning model with detailed instructions and concept-based approach to avoid copyright issues. Questions include source metadata for traceability.

Result: Top performers achieved over 97% accuracy (Gemini 3 Pro Preview: 99.8%). Among open-weight models, GLM-4.7 led at 98.6%. Petrophysics was the most challenging domain. Performance gap between open-weight and closed models was narrower than expected, with several open-weight models exceeding 90% accuracy.

Conclusion: The benchmark provides valuable insights into language model performance on specialized technical domains, showing that open-weight models can compete with closed models in domain-specific knowledge tasks. The dataset and evaluation tools are publicly available.

Abstract: This paper presents FormationEval, an open multiple-choice question benchmark for evaluating language models on petroleum geoscience and subsurface disciplines. The dataset contains 505 questions across seven domains including petrophysics, petroleum geology and reservoir engineering, derived from three authoritative sources using a reasoning model with detailed instructions and a concept-based approach that avoids verbatim copying of copyrighted text. Each question includes source metadata to support traceability and audit. The evaluation covers 72 models from major providers including OpenAI, Anthropic, Google, Meta and open-weight alternatives. The top performers achieve over 97% accuracy, with Gemini 3 Pro Preview reaching 99.8%, while tier and domain gaps persist. Among open-weight models, GLM-4.7 leads at 98.6%, with several DeepSeek, Llama, Qwen and Mistral models also exceeding 93%. The performance gap between open-weight and closed models is narrower than expected, with several lower-cost open-weight models exceeding 90% accuracy. Petrophysics emerges as the most challenging domain across all models, while smaller models show wider performance variance. Residual length bias in the dataset (correct answers tend to be longer) is documented along with bias mitigation strategies applied during construction. The benchmark, evaluation code and results are publicly available.

Method: Synapse models memory as a dynamic graph where relevance emerges from spreading activation rather than pre-computed links. It integrates lateral inhibition and temporal decay, and implements a Triple Hybrid Retrieval strategy that fuses geometric embeddings with activation-based graph traversal.

Result: Comprehensive evaluations on the LoCoMo benchmark show that Synapse significantly outperforms state-of-the-art methods in complex temporal and multi-hop reasoning tasks, effectively addressing the “Contextual Tunneling” problem.

Conclusion: Synapse offers a robust solution to memory limitations in LLMs by providing a unified memory architecture that transcends static vector similarity, enabling better long-term agentic memory and complex reasoning.

Abstract: While Large Language Models (LLMs) excel at generalized reasoning, standard retrieval-augmented approaches fail to address the disconnected nature of long-term agentic memory. To bridge this gap, we introduce Synapse (Synergistic Associative Processing Semantic Encoding), a unified memory architecture that transcends static vector similarity. Drawing from cognitive science, Synapse models memory as a dynamic graph where relevance emerges from spreading activation rather than pre-computed links. By integrating lateral inhibition and temporal decay, the system dynamically highlights relevant sub-graphs while filtering interference. We implement a Triple Hybrid Retrieval strategy that fuses geometric embeddings with activation-based graph traversal. Comprehensive evaluations on the LoCoMo benchmark show that Synapse significantly outperforms state-of-the-art methods in complex temporal and multi-hop reasoning tasks, offering a robust solution to the “Contextual Tunneling” problem. Our code and data will be made publicly available upon acceptance.

Method: Proposes RM-NLHF using similarity between GRM-generated and human critiques as process reward, plus MetaRM to predict process reward from human-critique data and generalize to data without human critiques.

Result: Experiments on multiple benchmarks show consistent outperformance over state-of-the-art GRMs trained with outcome-only reward, confirming superiority of natural language feedback over binary supervision.

Conclusion: Natural language feedback provides more accurate reward signals than binary outcomes, and MetaRM enables scalable process-level supervision by learning from limited human critique data.

Abstract: Reinforcement Learning with Verifiable reward (RLVR) on preference data has become the mainstream approach for training Generative Reward Models (GRMs). Typically in pairwise rewarding tasks, GRMs generate reasoning chains ending with critiques and preference labels, and RLVR then relies on the correctness of the preference labels as the training reward. However, in this paper, we demonstrate that such binary classification tasks make GRMs susceptible to guessing correct outcomes without sound critiques. Consequently, these spurious successes introduce substantial noise into the reward signal, thereby impairing the effectiveness of reinforcement learning. To address this issue, we propose Reward Modeling from Natural Language Human Feedback (RM-NLHF), which leverages natural language feedback to obtain process reward signals, thereby mitigating the problem of limited solution space inherent in binary tasks. Specifically, we compute the similarity between GRM-generated and human critiques as the training reward, which provides more accurate reward signals than outcome-only supervision. Additionally, considering that human critiques are difficult to scale up, we introduce Meta Reward Model (MetaRM) which learns to predict process reward from datasets with human critiques and then generalizes to data without human critiques. Experiments on multiple benchmarks demonstrate that our method consistently outperforms state-of-the-art GRMs trained with outcome-only reward, confirming the superiority of integrating natural language over binary human feedback as supervision.

Method: Hybrid framework: fine-tuned NMT model generates initial draft, then LLM with Retrieval-Augmented Generation (RAG) refines it. Analysis focuses on impact of retrieved examples quantity vs. retrieval algorithm choice.

Result: System achieves 35.21 chrF++ (+8.10 recovery), effectively matching original in-domain quality. Performance driven primarily by number of retrieved examples rather than retrieval algorithm. LLM acts as robust “safety net” for severe failures.

Conclusion: Hybrid NMT+LLM with RAG effectively addresses domain shift in low-resource translation, with LLM serving as critical safety net for zero-shot domains.

Abstract: Neural Machine Translation (NMT) models for low-resource languages suffer significant performance degradation under domain shift. We quantify this challenge using Dhao, an indigenous language of Eastern Indonesia with no digital footprint beyond the New Testament (NT). When applied to the unseen Old Testament (OT), a standard NMT model fine-tuned on the NT drops from an in-domain score of 36.17 chrF++ to 27.11 chrF++. To recover this loss, we introduce a hybrid framework where a fine-tuned NMT model generates an initial draft, which is then refined by a Large Language Model (LLM) using Retrieval-Augmented Generation (RAG). The final system achieves 35.21 chrF++ (+8.10 recovery), effectively matching the original in-domain quality. Our analysis reveals that this performance is driven primarily by the number of retrieved examples rather than the choice of retrieval algorithm. Qualitative analysis confirms the LLM acts as a robust “safety net,” repairing severe failures in zero-shot domains.

Method: Uses DeBERTa-base classifiers with calibrated thresholds, compares direct multi-label vs presence-gated hierarchical approaches, incorporates lightweight features (context, LIWC-22, moral lexica), and benchmarks instruction-tuned LLMs in zero/few-shot and QLoRA setups.

Result: Best supervised ensemble achieves macro-F1 = 0.332 on 19 values, improving over previous baseline (0.28). Moral presence detection achieves F1 = 0.74. LLMs lag behind supervised ensemble under same compute constraints.

Conclusion: Direct prediction with lightweight feature augmentation outperforms hierarchical approaches, and supervised methods are more compute-efficient than LLMs for value detection tasks under resource constraints.

Abstract: We study sentence-level detection of the 19 human values in the refined Schwartz continuum in about 74k English sentences from news and political manifestos (ValueEval'24 corpus). Each sentence is annotated with value presence, yielding a binary moral-presence label and a 19-way multi-label task under severe class imbalance. First, we show that moral presence is learnable from single sentences: a DeBERTa-base classifier attains positive-class F1 = 0.74 with calibrated thresholds. Second, we compare direct multi-label value detectors with presence-gated hierarchies in a setting where only a single consumer-grade GPU with 8 GB of VRAM is available, and we explicitly choose all training and inference configurations to fit within this budget. Presence gating does not improve over direct prediction, indicating that gate recall becomes a bottleneck. Third, we investigate lightweight auxiliary signals - short-range context, LIWC-22, and moral lexica - and small ensembles. Our best supervised configuration, a soft-voting ensemble of DeBERTa-based models enriched with such signals, reaches macro-F1 = 0.332 on the 19 values, improving over the best previous English-only baseline on this corpus, namely the best official ValueEval'24 English run (macro-F1 = 0.28 on the same 19-value test set). Methodologically, our study provides, to our knowledge, the first systematic comparison of direct versus presence-gated architectures, lightweight feature-augmented encoders, and medium-sized instruction-tuned Large Language Models (LLMs) for refined Schwartz values at sentence level. We additionally benchmark 7-9B instruction-tuned LLMs (Gemma 2 9B, Llama 3.1 8B, Mistral 8B, Qwen 2.5 7B) in zero-/few-shot and QLoRA setups, and find that they lag behind the supervised ensemble under the same compute budget. Overall, our results provide empirical guidance for building compute-efficient, value-aware NLP models.

Method: Systematic comparison of proprietary and open-weight LLMs using naive document-level prompting setup, analyzing APE quality, contextual behavior, robustness to data poisoning attacks, and efficiency metrics.

Result: Proprietary LLMs achieve near human-level APE quality with simple one-shot prompting, regardless of document context. They show higher robustness to attacks than open-weight models but largely fail to exploit document-level context. Standard metrics don’t reflect qualitative improvements, and proprietary models have impractical cost/latency overheads.

Conclusion: LLM-based document-aware APE shows promise but has limitations: proprietary models are robust but impractical, and current approaches fail to effectively leverage document context, pointing to need for more efficient long-context modeling for translation refinement.

Abstract: Automatic post-editing (APE) aims to refine machine translations by correcting residual errors. Although recent large language models (LLMs) demonstrate strong translation capabilities, their effectiveness for APE–especially under document-level context–remains insufficiently understood. We present a systematic comparison of proprietary and open-weight LLMs under a naive document-level prompting setup, analyzing APE quality, contextual behavior, robustness, and efficiency. Our results show that proprietary LLMs achieve near human-level APE quality even with simple one-shot prompting, regardless of whether document context is provided. While these models exhibit higher robustness to data poisoning attacks than open-weight counterparts, this robustness also reveals a limitation: they largely fail to exploit document-level context for contextual error correction. Furthermore, standard automatic metrics do not reliably reflect these qualitative improvements, highlighting the continued necessity of human evaluation. Despite their strong performance, the substantial cost and latency overheads of proprietary LLMs render them impractical for real-world APE deployment. Overall, our findings elucidate both the promise and current limitations of LLM-based document-aware APE, and point toward the need for more efficient long-context modeling approaches for translation refinement.

Method: Collection of case studies using Google’s Gemini-based models (Gemini Deep Think and variants) with interactive conversational methodology, including iterative refinement, problem decomposition, cross-disciplinary knowledge transfer, adversarial review, and neuro-symbolic loops for code verification.

Result: Successful collaborations solving open problems, refuting conjectures, and generating new proofs across diverse theoretical fields, demonstrating AI’s ability to contribute to genuine scientific discovery beyond automation.

Conclusion: AI models can serve as versatile partners in scientific discovery when combined with effective human-AI collaboration techniques, pushing beyond standard chat interfaces to include adversarial review and neuro-symbolic verification systems.

Abstract: Recent advances in large language models (LLMs) have opened new avenues for accelerating scientific research. While models are increasingly capable of assisting with routine tasks, their ability to contribute to novel, expert-level mathematical discovery is less understood. We present a collection of case studies demonstrating how researchers have successfully collaborated with advanced AI models, specifically Google’s Gemini-based models (in particular Gemini Deep Think and its advanced variants), to solve open problems, refute conjectures, and generate new proofs across diverse areas in theoretical computer science, as well as other areas such as economics, optimization, and physics. Based on these experiences, we extract common techniques for effective human-AI collaboration in theoretical research, such as iterative refinement, problem decomposition, and cross-disciplinary knowledge transfer. While the majority of our results stem from this interactive, conversational methodology, we also highlight specific instances that push beyond standard chat interfaces. These include deploying the model as a rigorous adversarial reviewer to detect subtle flaws in existing proofs, and embedding it within a “neuro-symbolic” loop that autonomously writes and executes code to verify complex derivations. Together, these examples highlight the potential of AI not just as a tool for automation, but as a versatile, genuine partner in the creative process of scientific discovery.

Method: Compared three subword paradigms (Byte Pair Encoding, Overlap BPE, and Unigram Language Model) across six Uralic languages using part-of-speech tagging as a controlled downstream task to evaluate morphological alignment and transfer efficacy.

Result: OBPE consistently achieves stronger morphological alignment and higher tagging accuracy than conventional methods, particularly for Latin-script languages. Gains come from reduced fragmentation in open-class categories and better frequency spectrum balance.

Conclusion: Morphology-sensitive tokenization is a decisive factor for effective cross-lingual transfer in agglutinative, low-resource languages, not just a preprocessing choice.

Abstract: Subword tokenization critically affects Natural Language Processing (NLP) performance, yet its behavior in morphologically rich and low-resource language families remains under-explored. This study systematically compares three subword paradigms – Byte Pair Encoding (BPE), Overlap BPE (OBPE), and Unigram Language Model – across six Uralic languages with varying resource availability and typological diversity. Using part-of-speech (POS) tagging as a controlled downstream task, we show that OBPE consistently achieves stronger morphological alignment and higher tagging accuracy than conventional methods, particularly within the Latin-script group. These gains arise from reduced fragmentation in open-class categories and a better balance across the frequency spectrum. Transfer efficacy further depends on the downstream tagging architecture, interacting with both training volume and genealogical proximity. Taken together, these findings highlight that morphology-sensitive tokenization is not merely a preprocessing choice but a decisive factor in enabling effective cross-lingual transfer for agglutinative, low-resource languages.

Method: The authors introduce a unified safety-analysis framework that systematically deconstructs CoT generation across model layers and evaluates individual attention heads using Jacobian-based spectral metrics. They propose three interpretable measures: stability, geometry, and energy to quantify how specific attention heads respond to or embed deceptive reasoning patterns.

Result: Experiments on multiple reasoning-oriented LLMs show that generation risk rises significantly when thinking mode is activated, with critical routing decisions concentrated in only a few contiguous mid-depth layers. The framework successfully identifies specific attention heads responsible for divergence between safe outputs and unsafe internal reasoning.

Conclusion: The work challenges the assumption that refusal implies safety in LLMs and provides a new perspective for understanding and mitigating latent reasoning risks by identifying specific attention mechanisms that propagate unsafe narratives internally.

Abstract: From generating headlines to fabricating news, the Large Language Models (LLMs) are typically assessed by their final outputs, under the safety assumption that a refusal response signifies safe reasoning throughout the entire process. Challenging this assumption, our study reveals that during fake news generation, even when a model rejects a harmful request, its Chain-of-Thought (CoT) reasoning may still internally contain and propagate unsafe narratives. To analyze this phenomenon, we introduce a unified safety-analysis framework that systematically deconstructs CoT generation across model layers and evaluates the role of individual attention heads through Jacobian-based spectral metrics. Within this framework, we introduce three interpretable measures: stability, geometry, and energy to quantify how specific attention heads respond or embed deceptive reasoning patterns. Extensive experiments on multiple reasoning-oriented LLMs show that the generation risk rise significantly when the thinking mode is activated, where the critical routing decisions concentrated in only a few contiguous mid-depth layers. By precisely identifying the attention heads responsible for this divergence, our work challenges the assumption that refusal implies safety and provides a new understanding perspective for mitigating latent reasoning risks.

Method: Inspired by dramatic theory, CAST constructs 3D scenes organized by time, place, and topic, and organizes them into character profiles that summarize events. This episodic memory is complemented by a graph-based semantic memory, creating a dual memory design.

Result: CAST improves performance by an average of 8.11% F1 and 10.21% J(LLM-as-a-Judge) compared to baselines across various datasets, with particularly strong improvements on open and time-sensitive conversational questions.

Conclusion: The CAST architecture successfully addresses the challenge of representing and retrieving coherent episodic memories in AI agents, demonstrating significant performance improvements over existing memory systems.

Abstract: Episodic memory is a central component of human memory, which refers to the ability to recall coherent events grounded in who, when, and where. However, most agent memory systems only emphasize semantic recall and treat experience as structures such as key-value, vector, or graph, which makes them struggle to represent and retrieve coherent events. To address this challenge, we propose a Character-and-Scene based memory architecture(CAST) inspired by dramatic theory. Specifically, CAST constructs 3D scenes (time/place/topic) and organizes them into character profiles that summarize the events of a character to represent episodic memory. Moreover, CAST complements this episodic memory with a graph-based semantic memory, which yields a robust dual memory design. Experiments demonstrate that CAST has averagely improved 8.11% F1 and 10.21% J(LLM-as-a-Judge) than baselines on various datasets, especially on open and time-sensitive conversational questions.

Method: Proposes decomposing paraphrases into paraphrase types (constituent linguistic aspects) and training models explicitly on these types rather than binary paraphrase decisions. Demonstrates this approach through experiments in plagiarism detection and duplicate question identification.

Result: Models trained on paraphrase types achieve 89.6% accuracy vs 78.4% human baseline for Wikipedia plagiarism detection, and 66.5% vs 55.7% for arXiv scientific paper plagiarism. Also improves performance on Quora duplicate question identification over binary pair training.

Conclusion: Decomposing paraphrases into linguistic aspects provides better semantic understanding in computational models, enabling stronger performance on paraphrase-related tasks and downstream applications compared to traditional binary approaches.

Abstract: Language enables humans to share knowledge, reason about the world, and pass on strategies for survival and innovation across generations. At the heart of this process is not just the ability to communicate but also the remarkable flexibility in how we can express ourselves. We can express the same thoughts in virtually infinite ways using different words and structures - this ability to rephrase and reformulate expressions is known as paraphrase. Modeling paraphrases is a keystone to meaning in computational language models; being able to construct different variations of texts that convey the same meaning or not shows strong abilities of semantic understanding. If computational language models are to represent meaning, they must understand and control the different aspects that construct the same meaning as opposed to different meanings at a fine granularity. Yet most existing approaches reduce paraphrasing to a binary decision between two texts or to producing a single rewrite of a source, obscuring which linguistic factors are responsible for meaning preservation. In this thesis, I propose that decomposing paraphrases into their constituent linguistic aspects (paraphrase types) offers a more fine-grained and cognitively grounded view of semantic equivalence. I show that even advanced machine learning models struggle with this task. Yet, when explicitly trained on paraphrase types, models achieve stronger performance on related paraphrase tasks and downstream applications. For example, in plagiarism detection, language models trained on paraphrase types surpass human baselines: 89.6% accuracy compared to 78.4% for plagiarism cases from Wikipedia, and 66.5% compared to 55.7% for plagiarism of scientific papers from arXiv. In identifying duplicate questions on Quora, models trained with paraphrase types improve over models trained on binary pairs. Furthermore, I demonstrate that…

Method: Developed an LLM-based classifier to identify underspecified questions in QA datasets, then conducted controlled rewriting experiments to transform underspecified questions into fully specified variants while keeping gold answers fixed.

Result: Found 16% to over 50% of benchmark questions are underspecified; LLMs perform significantly worse on them; QA performance consistently improves when questions are rewritten to be fully specified.

Conclusion: Underspecification is a major confound in QA evaluation, and many apparent LLM failures stem from ambiguous questions rather than model limitations, highlighting need for clearer benchmark design.

Abstract: Large language models (LLMs) perform well on well-posed questions, yet standard question-answering (QA) benchmarks remain far from solved. We argue that this gap is partly due to underspecified questions - queries whose interpretation cannot be uniquely determined without additional context. To test this hypothesis, we introduce an LLM-based classifier to identify underspecified questions and apply it to several widely used QA datasets, finding that 16% to over 50% of benchmark questions are underspecified and that LLMs perform significantly worse on them. To isolate the effect of underspecification, we conduct a controlled rewriting experiment that serves as an upper-bound analysis, rewriting underspecified questions into fully specified variants while holding gold answers fixed. QA performance consistently improves under this setting, indicating that many apparent QA failures stem from question underspecification rather than model limitations. Our findings highlight underspecification as an important confound in QA evaluation and motivate greater attention to question clarity in benchmark design.

Method: Proposes MAILS framework with: 1) Locality-Preserving Sliding Window Attention module for extracting locally discriminative geometric features from sparse point clouds; 2) Coordinate-independent feature initialization module; 3) Locally invariant positional encoding mechanism for robustness against yaw rotations and altitude variations; 4) A new large-scale LiDAR localization dataset for UAVs with four scenes and various flight trajectories.

Result: Extensive experiments show the method achieves satisfactory localization precision and consistently outperforms existing techniques by a significant margin. The framework handles UAV-specific challenges effectively.

Conclusion: MAILS provides an effective map-free LiDAR relocalization solution for UAVs, addressing key challenges like sparse point clouds, yaw rotations, and altitude variations. The new dataset enables better evaluation of UAV relocalization methods under realistic conditions.

Abstract: Localization is a fundamental capability in unmanned aerial vehicle (UAV) systems. Map-free LiDAR relocalization offers an effective solution for achieving high-precision positioning in environments with weak or unavailable GNSS signals. However, existing LiDAR relocalization methods are primarily tailored to autonomous driving, exhibiting significantly degraded accuracy in UAV scenarios. In this paper, we propose MAILS, a novel map-free LiDAR relocalization framework for UAVs. A Locality-Preserving Sliding Window Attention module is first introduced to extract locally discriminative geometric features from sparse point clouds. To handle substantial yaw rotations and altitude variations encountered during UAV flight, we then design a coordinate-independent feature initialization module and a locally invariant positional encoding mechanism, which together significantly enhance the robustness of feature extraction. Furthermore, existing LiDAR-based relocalization datasets fail to capture real-world UAV flight characteristics, such as irregular trajectories and varying altitudes. To address this gap, we construct a large-scale LiDAR localization dataset for UAVs, which comprises four scenes and various flight trajectories, designed to evaluate UAV relocalization performance under realistic conditions. Extensive experiments demonstrate that our method achieves satisfactory localization precision and consistently outperforms existing techniques by a significant margin. Our code and dataset will be released soon.

Method: Investigates three XIL strategies: CAIPI, Right for the Right Reasons (RRR), and a novel hybrid approach combining both. Uses Gradient-weighted Class Activation Mapping (GradCAM) and Bounded Logit Attention (BLA) for explanation generation and evaluation via segmentation mask comparisons.

Result: XIL methods effectively guide models to focus on relevant image features (especially CAIPI) and reduce model bias by balancing misclassification rates between male and female predictions. CAIPI shows potential to improve classification accuracy while other methods cause slight performance decreases.

Conclusion: XIL methods demonstrate significant potential for improving fairness and transparency in visual classifiers, with CAIPI being particularly promising for both bias reduction and maintaining/improving accuracy in gender classification tasks.

Abstract: Explanatory interactive learning (XIL) enables users to guide model training in machine learning (ML) by providing feedback on the model’s explanations, thereby helping it to focus on features that are relevant to the prediction from the user’s perspective. In this study, we explore the capability of this learning paradigm to mitigate bias and spurious correlations in visual classifiers, specifically in scenarios prone to data bias, such as gender classification. We investigate two methodologically different state-of-the-art XIL strategies, i.e., CAIPI and Right for the Right Reasons (RRR), as well as a novel hybrid approach that combines both strategies. The results are evaluated quantitatively by comparing segmentation masks with explanations generated using Gradient-weighted Class Activation Mapping (GradCAM) and Bounded Logit Attention (BLA). Experimental results demonstrate the effectiveness of these methods in (i) guiding ML models to focus on relevant image features, particularly when CAIPI is used, and (ii) reducing model bias (i.e., balancing the misclassification rates between male and female predictions). Our analysis further supports the potential of XIL methods to improve fairness in gender classifiers. Overall, the increased transparency and fairness obtained by XIL leads to slight performance decreases with an exception being CAIPI, which shows potential to even improve classification accuracy.

Method: Proposes COOPERTRIM framework with: 1) Conformal temporal uncertainty metric to gauge feature relevance by identifying dynamic vs static information, 2) Data-driven mechanism to dynamically determine sharing quantity based on environment complexity, 3) Temporal awareness to adapt sharing to environmental dynamics.

Result: Achieves up to 80.28% bandwidth reduction for semantic segmentation and 72.52% for 3D detection while maintaining comparable accuracy. Improves IoU by up to 45.54% with 72% less bandwidth. Combined with compression, reduces bandwidth to 1.46% without compromising IoU.

Conclusion: COOPERTRIM effectively addresses bandwidth limitations in cooperative perception by exploiting temporal continuity, demonstrating adaptability to environmental dynamics, localization error, and communication latency, enabling practical real-world deployment.

Abstract: Cooperative perception enables autonomous agents to share encoded representations over wireless communication to enhance each other’s live situational awareness. However, the tension between the limited communication bandwidth and the rich sensor information hinders its practical deployment. Recent studies have explored selection strategies that share only a subset of features per frame while striving to keep the performance on par. Nevertheless, the bandwidth requirement still stresses current wireless technologies. To fundamentally ease the tension, we take a proactive approach, exploiting the temporal continuity to identify features that capture environment dynamics, while avoiding repetitive and redundant transmission of static information. By incorporating temporal awareness, agents are empowered to dynamically adapt the sharing quantity according to environment complexity. We instantiate this intuition into an adaptive selection framework, COOPERTRIM, which introduces a novel conformal temporal uncertainty metric to gauge feature relevance, and a data-driven mechanism to dynamically determine the sharing quantity. To evaluate COOPERTRIM, we take semantic segmentation and 3D detection as example tasks. Across multiple open-source cooperative segmentation and detection models, COOPERTRIM achieves up to 80.28% and 72.52% bandwidth reduction respectively while maintaining a comparable accuracy. Relative to other selection strategies, COOPERTRIM also improves IoU by as much as 45.54% with up to 72% less bandwidth. Combined with compression strategies, COOPERTRIM can further reduce bandwidth usage to as low as 1.46% without compromising IoU performance. Qualitative results show COOPERTRIM gracefully adapts to environmental dynamics, localization error, and communication latency, demonstrating flexibility and paving the way for real-world deployment.

Method: GOT-JEPA uses model-predictive pretraining where teacher generates pseudo-tracking models from clean frames and student learns to predict same models from corrupted frames. OccuSolver adds point-centric visibility estimation and iterative occlusion-pattern refinement using object priors.

Result: Extensive evaluations on seven benchmarks show the method effectively enhances tracker generalization and robustness, improving performance in dynamic environments with occlusions and distractors.

Conclusion: The proposed framework addresses limitations in generalization and occlusion perception for object tracking, bridging the gap between current trackers and human visual capabilities.

Abstract: The human visual system tracks objects by integrating current observations with previously observed information, adapting to target and scene changes, and reasoning about occlusion at fine granularity. In contrast, recent generic object trackers are often optimized for training targets, which limits robustness and generalization in unseen scenarios, and their occlusion reasoning remains coarse, lacking detailed modeling of occlusion patterns. To address these limitations in generalization and occlusion perception, we propose GOT-JEPA, a model-predictive pretraining framework that extends JEPA from predicting image features to predicting tracking models. Given identical historical information, a teacher predictor generates pseudo-tracking models from a clean current frame, and a student predictor learns to predict the same pseudo-tracking models from a corrupted version of the current frame. This design provides stable pseudo supervision and explicitly trains the predictor to produce reliable tracking models under occlusions, distractors, and other adverse observations, improving generalization to dynamic environments. Building on GOT-JEPA, we further propose OccuSolver to enhance occlusion perception for object tracking. OccuSolver adapts a point-centric point tracker for object-aware visibility estimation and detailed occlusion-pattern capture. Conditioned on object priors iteratively generated by the tracker, OccuSolver incrementally refines visibility states, strengthens occlusion handling, and produces higher-quality reference labels that progressively improve subsequent model predictions. Extensive evaluations on seven benchmarks show that our method effectively enhances tracker generalization and robustness.

Method: Evaluate Qwen2-VL-7B and Idefics3-8B with PTQ using data-free (HQQ) and data-aware (MBQ) methods across multiple bit widths. Adapt Selector confidence estimator for quantized multimodal settings and test robustness across quantization levels and OOD scenarios.

Result: PTQ degrades both accuracy and reliability. Data-aware methods soften the effect. Selector substantially mitigates reliability impact. Int4 MBQ + Selector achieves best efficiency-reliability trade-off, closing in on uncompressed performance at ~75% less memory.

Conclusion: First systematic study linking quantization and reliability in multimodal settings. Shows data-aware quantization methods combined with confidence estimators can maintain reliability while achieving significant compression for edge deployment.

Abstract: Multimodal Large Language Models (MLLM) are increasingly deployed in domains where both reliability and efficiency are critical. However, current models remain overconfident, producing highly certain but incorrect answers. At the same time, their large size limits deployment on edge devices, necessitating compression. We study the intersection of these two challenges by analyzing how Post-Training Quantization (PTQ) compression affects both accuracy and reliability in Visual Question Answering (VQA). We evaluate two MLLMs, Qwen2-VL-7B and Idefics3-8B, quantized with data-free (HQQ) and data-aware (MBQ) methods across multiple bit widths. To counteract the reduction in reliability caused by quantization, we adapt the Selector confidence estimator for quantized multimodal settings and test its robustness across various quantization levels and out-of-distribution (OOD) scenarios. We find that PTQ degrades both accuracy and reliability. Data-aware methods soften the effect thereof. The Selector substantially mitigates the reliability impact. The combination of int4 MBQ and the Selector achieves the best efficiency-reliability trade-off, closing in on uncompressed performance at approx. 75% less memory demand. Overall, we present the first systematic study linking quantization and reliability in multimodal settings.

Method: Uses NutNet++ for three-way classification (benign, local patches, global perturbations). Local threats are purified via grayscale masking, while global perturbations trigger Expert-guided Adversarial Prompt Tuning (EAPT) that generates corrective driving prompts via gradient-based latent optimization instead of full-model fine-tuning.

Result: On the Dolphins benchmark, NutVLM achieves 4.89% improvement in overall metrics including Accuracy, Language Score, and GPT Score.

Conclusion: NutVLM provides a scalable security solution for intelligent transportation systems by securing the entire perception-decision lifecycle of VLMs against adversarial threats.

Abstract: Vision Language Models (VLMs) have advanced perception in autonomous driving (AD), but they remain vulnerable to adversarial threats. These risks range from localized physical patches to imperceptible global perturbations. Existing defense methods for VLMs remain limited and often fail to reconcile robustness with clean-sample performance. To bridge these gaps, we propose NutVLM, a comprehensive self-adaptive defense framework designed to secure the entire perception-decision lifecycle. Specifically, we first employ NutNet++ as a sentinel, which is a unified detection-purification mechanism. It identifies benign samples, local patches, and global perturbations through three-way classification. Subsequently, localized threats are purified via efficient grayscale masking, while global perturbations trigger Expert-guided Adversarial Prompt Tuning (EAPT). Instead of the costly parameter updates of full-model fine-tuning, EAPT generates “corrective driving prompts” via gradient-based latent optimization and discrete projection. These prompts refocus the VLM’s attention without requiring exhaustive full-model retraining. Evaluated on the Dolphins benchmark, our NutVLM yields a 4.89% improvement in overall metrics (e.g., Accuracy, Language Score, and GPT Score). These results validate NutVLM as a scalable security solution for intelligent transportation. Our code is available at https://github.com/PXX/NutVLM.

Method: Proposes VisPhyWorld framework requiring models to generate executable simulator code from visual observations, and VisPhyBench with 209 scenes from 108 physical templates to evaluate appearance reconstruction and physically plausible motion reproduction.

Result: Pipeline produces valid reconstructed videos in 97.7% of cases. State-of-the-art MLLMs show strong semantic scene understanding but struggle with accurate physical parameter inference and consistent physical dynamics simulation.

Conclusion: Execution-based evaluation reveals limitations in current MLLMs’ physical reasoning capabilities, highlighting the need for frameworks that separate physical reasoning from rendering and make world representations testable.

Abstract: Evaluating whether Multimodal Large Language Models (MLLMs) genuinely reason about physical dynamics remains challenging. Most existing benchmarks rely on recognition-style protocols such as Visual Question Answering (VQA) and Violation of Expectation (VoE), which can often be answered without committing to an explicit, testable physical hypothesis. We propose VisPhyWorld, an execution-based framework that evaluates physical reasoning by requiring models to generate executable simulator code from visual observations. By producing runnable code, the inferred world representation is directly inspectable, editable, and falsifiable. This separates physical reasoning from rendering. Building on this framework, we introduce VisPhyBench, comprising 209 evaluation scenes derived from 108 physical templates and a systematic protocol that evaluates how well models reconstruct appearance and reproduce physically plausible motion. Our pipeline produces valid reconstructed videos in 97.7% on the benchmark. Experiments show that while state-of-the-art MLLMs achieve strong semantic scene understanding, they struggle to accurately infer physical parameters and to simulate consistent physical dynamics.

Method: Decomposes HRRP data into three components: mask, features, and noise. Based on this decomposition, proposes two physics-based metrics that leverage the physical interpretation of radar data. Evaluates these metrics using an expensive dataset on a challenging task.

Result: Demonstrates the discriminative ability of the proposed metrics, showing they can effectively evaluate generated HRRP data in ways that black-box classification models cannot.

Conclusion: The proposed physics-based decomposition and metrics provide more interpretable and multi-level evaluation of generated HRRP data compared to existing black-box classification approaches, offering better alignment with the physical properties of radar signals.

Abstract: High-resolution range profile (HRRP ) data are in vogue in radar automatic target recognition (RATR). With the interest in classifying models using HRRP, filling gaps in datasets using generative models has recently received promising contributions. Evaluating generated data is a challenging topic, even for explicit data like face images. However, the evaluation methods used in the state-ofthe-art of HRRP generation rely on classification models. Such models, called ‘‘black-box’’, do not allow either explainability on generated data or multi-level evaluation. This work focuses on decomposing HRRP data into three components: the mask, the features, and the noise. Using this decomposition, we propose two metrics based on the physical interpretation of those data. We take profit from an expensive dataset to evaluate our metrics on a challenging task and demonstrate the discriminative ability of those.

Method: NeRV360 integrates viewport extraction directly into the decoding process and introduces a spatial-temporal affine transform module for conditional decoding based on viewpoint and time. This allows selective decoding of only the user-selected viewport rather than reconstructing the entire panoramic frame.

Result: On 6K-resolution videos, NeRV360 achieves 7-fold reduction in memory consumption and 2.5-fold increase in decoding speed compared to HNeRV (a prior work), while delivering better image quality in objective metrics.

Conclusion: NeRV360 provides an efficient solution for 360-degree video compression using implicit neural representations by focusing decoding on relevant viewports, enabling practical real-time applications with significant improvements in memory and speed.

Abstract: Implicit neural representations for videos (NeRV) have shown strong potential for video compression. However, applying NeRV to high-resolution 360-degree videos causes high memory usage and slow decoding, making real-time applications impractical. We propose NeRV360, an end-to-end framework that decodes only the user-selected viewport instead of reconstructing the entire panoramic frame. Unlike conventional pipelines, NeRV360 integrates viewport extraction into decoding and introduces a spatial-temporal affine transform module for conditional decoding based on viewpoint and time. Experiments on 6K-resolution videos show that NeRV360 achieves a 7-fold reduction in memory consumption and a 2.5-fold increase in decoding speed compared to HNeRV, a representative prior work, while delivering better image quality in terms of objective metrics.

Method: The authors analyze HRRP synthesis on a large-scale maritime database and identify that the fundamental scenario drivers are geometric: ship dimensions and desired aspect angle. They condition generative models on these variables and train them to synthesize radar signatures.

Result: The synthesized signatures reproduce the expected line-of-sight geometric trend observed in real data, demonstrating that acquisition geometry plays a central role in robust HRRP generation.

Conclusion: The study highlights the importance of acquisition geometry for robust HRRP generation and shows that conditioning on geometric variables (ship dimensions and aspect angle) enables effective synthesis of radar signatures that match real-world trends.

Abstract: High-resolution range profiles (HRRPs) enable fast onboard processing for radar automatic target recognition, but their strong sensitivity to acquisition conditions limits robustness across operational scenarios. Conditional HRRP generation can mitigate this issue, yet prior studies are constrained by small, highly specific datasets. We study HRRP synthesis on a largescale maritime database representative of coastal surveillance variability. Our analysis indicates that the fundamental scenario drivers are geometric: ship dimensions and the desired aspect angle. Conditioning on these variables, we train generative models and show that the synthesized signatures reproduce the expected line-of-sight geometric trend observed in real data. These results highlight the central role of acquisition geometry for robust HRRP generation.

Method: Conducted a controlled comparative study of three canonical CNN architectures (VGG, ResNet, GoogLeNet) with standardized training protocols, distinguishing between nominal depth (number of layers) and effective depth (actual depth manifested during training).

Result: Plain deep networks (like VGG) show early accuracy saturation and optimization instability, while residual (ResNet) and inception-based (GoogLeNet) architectures consistently translate additional depth into improved accuracy at lower effective depth with better accuracy-compute trade-offs.

Conclusion: Effective depth, not nominal depth, is the key factor governing depth’s role as a productive scaling dimension in convolutional networks, enabled by architectural mechanisms that constrain how depth manifests during training.

Abstract: Increasing convolutional depth has been central to advances in image recognition, yet deeper networks do not uniformly yield higher accuracy, stable optimization, or efficient computation. We present a controlled comparative study of three canonical convolutional neural network architectures - VGG, ResNet, and GoogLeNet - to isolate how depth influences classification performance, convergence behavior, and computational efficiency. By standardizing training protocols and explicitly distinguishing between nominal and effective depth, we show that the benefits of depth depend critically on architectural mechanisms that constrain its effective manifestation during training rather than on nominal depth alone. Although plain deep networks exhibit early accuracy saturation and optimization instability, residual and inception-based architectures consistently translate additional depth into improved accuracy at lower effective depth and favorable accuracy-compute trade-offs. These findings demonstrate that effective depth, not nominal depth, is the operative quantity governing depth’s role as a productive scaling dimension in convolutional networks.

Method: KidMesh extracts feature maps from MRU images, converts them to feature vertices through grid sampling, and deforms a template mesh according to these feature vertices. The method includes a novel training schema that doesn’t require accurate mesh-level annotations, which are difficult to obtain due to sparsely sampled MRU slices.

Result: KidMesh reconstructs CH meshes in 0.4 seconds on average, achieving comparable performance to conventional methods without post-processing. Reconstructed meshes have no self-intersections, with only 3.7% and 0.2% of vertices having error distances exceeding 3.2mm and 6.4mm respectively. After rasterization, Dice score of 0.86 against manual masks. Meshes can be used for renal urine flow simulations.

Conclusion: KidMesh provides an efficient end-to-end solution for direct mesh reconstruction from MRU images, enabling functional urodynamic assessments without complex post-processing, with potential clinical value for congenital hydronephrosis diagnosis and treatment planning.

Abstract: Pediatric congenital hydronephrosis (CH) is a common urinary tract disorder, primarily caused by obstruction at the renal pelvis-ureter junction. Magnetic resonance urography (MRU) can visualize hydronephrosis, including renal pelvis and calyces, by utilizing the natural contrast provided by water. Existing voxel-based segmentation approaches can extract CH regions from MRU, facilitating disease diagnosis and prognosis. However, these segmentation methods predominantly focus on morphological features, such as size, shape, and structure. To enable functional assessments, such as urodynamic simulations, external complex post-processing steps are required to convert these results into mesh-level representations. To address this limitation, we propose an end-to-end method based on deep neural networks, namely KidMesh, which could automatically reconstruct CH meshes directly from MRU. Generally, KidMesh extracts feature maps from MRU images and converts them into feature vertices through grid sampling. It then deforms a template mesh according to these feature vertices to generate the specific CH meshes of MRU images. Meanwhile, we develop a novel schema to train KidMesh without relying on accurate mesh-level annotations, which are difficult to obtain due to the sparsely sampled MRU slices. Experimental results show that KidMesh could reconstruct CH meshes in an average of 0.4 seconds, and achieve comparable performance to conventional methods without requiring post-processing. The reconstructed meshes exhibited no self-intersections, with only 3.7% and 0.2% of the vertices having error distances exceeding 3.2mm and 6.4mm, respectively. After rasterization, these meshes achieved a Dice score of 0.86 against manually delineated CH masks. Furthermore, these meshes could be used in renal urine flow simulations, providing valuable urodynamic information for clinical practice.

Method: Proposes DriveMamba with a single-stage Unified Mamba decoder that integrates dynamic task relation modeling, implicit view correspondence learning, and long-term temporal fusion. Converts image features and task outputs to token-level sparse representations sorted by 3D positions, uses linear-complexity Mamba for efficient long-context modeling, and employs bidirectional trajectory-guided “local-to-global” scan for spatial locality preservation.

Result: Extensive experiments on nuScenes and Bench2Drive datasets demonstrate superiority, generalizability, and great efficiency of DriveMamba compared to existing approaches.

Conclusion: DriveMamba provides an efficient, scalable paradigm for E2E-AD that overcomes limitations of sequential modular designs through unified Mamba-based architecture with linear complexity and better task relation modeling.

Abstract: Recent advances towards End-to-End Autonomous Driving (E2E-AD) have been often devoted on integrating modular designs into a unified framework for joint optimization e.g. UniAD, which follow a sequential paradigm (i.e., perception-prediction-planning) based on separable Transformer decoders and rely on dense BEV features to encode scene representations. However, such manual ordering design can inevitably cause information loss and cumulative errors, lacking flexible and diverse relation modeling among different modules and sensors. Meanwhile, insufficient training of image backbone and quadratic-complexity of attention mechanism also hinder the scalability and efficiency of E2E-AD system to handle spatiotemporal input. To this end, we propose DriveMamba, a Task-Centric Scalable paradigm for efficient E2E-AD, which integrates dynamic task relation modeling, implicit view correspondence learning and long-term temporal fusion into a single-stage Unified Mamba decoder. Specifically, both extracted image features and expected task outputs are converted into token-level sparse representations in advance, which are then sorted by their instantiated positions in 3D space. The linear-complexity operator enables efficient long-context sequential token modeling to capture task-related inter-dependencies simultaneously. Additionally, a bidirectional trajectory-guided “local-to-global” scan method is designed to preserve spatial locality from ego-perspective, thus facilitating the ego-planning. Extensive experiments conducted on nuScenes and Bench2Drive datasets demonstrate the superiority, generalizability and great efficiency of DriveMamba.

Method: Proposes Orthogonal Variance Guidance (OVG), an inference-time method that corrects ODE dynamics to enforce theoretical Gaussian noise magnitude within the null-space of the structural gradient, resolving the structure-texture trade-off.

Result: Extensive experiments on microscopy super-resolution (BBBC021) and sketch-to-image (Edges2Shoes) demonstrate OVG effectively restores photorealistic textures while preserving structural fidelity.

Conclusion: OVG successfully addresses spectral collapse in conditional diffusion inversion for cross-domain image translation, enabling high-quality texture generation while maintaining structural accuracy.

Abstract: Conditional diffusion inversion provides a powerful framework for unpaired image-to-image translation. However, we demonstrate through an extensive analysis that standard deterministic inversion (e.g. DDIM) fails when the source domain is spectrally sparse compared to the target domain (e.g., super-resolution, sketch-to-image). In these contexts, the recovered latent from the input does not follow the expected isotropic Gaussian distribution. Instead it exhibits a signal with lower frequencies, locking target sampling to oversmoothed and texture-poor generations. We term this phenomenon spectral collapse. We observe that stochastic alternatives attempting to restore the noise variance tend to break the semantic link to the input, leading to structural drift. To resolve this structure-texture trade-off, we propose Orthogonal Variance Guidance (OVG), an inference-time method that corrects the ODE dynamics to enforce the theoretical Gaussian noise magnitude within the null-space of the structural gradient. Extensive experiments on microscopy super-resolution (BBBC021) and sketch-to-image (Edges2Shoes) demonstrate that OVG effectively restores photorealistic textures while preserving structural fidelity.

Method: PCReg-Net uses a progressive contrast-guided registration framework with four modules: (1) registration U-Net for coarse alignment, (2) reference feature extractor capturing multi-scale structural cues, (3) contrast module identifying residual misalignment by comparing coarse-registered and reference features, and (4) refinement U-Net with feature injection for high-fidelity output. Also proposes Temporal NCC (TNCC) and Temporal NCC Gap (TNCG) for reference-free evaluation.

Result: On OR-PAM-Reg-4K dataset (432 test samples), PCReg-Net achieves NCC of 0.983, SSIM of 0.982, and PSNR of 46.96 dB, surpassing state-of-the-art by over 14 dB at real-time speed.

Conclusion: PCReg-Net effectively addresses alignment challenges in bidirectional OR-PAM imaging through a progressive contrast-guided framework, achieving superior registration quality and temporal consistency while maintaining real-time performance.

Abstract: High-speed optical-resolution photoacoustic microscopy (OR-PAM) with bidirectional raster scanning doubles imaging speed but introduces coupled domain shift and geometric misalignment between forward and backward scan lines. Existing methods, constrained by brightness constancy assumptions, achieve limited alignment quality (NCC~$\leq 0.96$). We propose PCReg-Net, a progressive contrast-guided registration framework that performs coarse-to-fine alignment through four lightweight modules: (1)~a registration U-Net for coarse alignment, (2)~a reference feature extractor capturing multi-scale structural cues, (3)~a contrast module that identifies residual misalignment by comparing coarse-registered and reference features, and (4)~a refinement U-Net with feature injection for high-fidelity output. We further propose the Temporal NCC (TNCC) and Temporal NCC Gap (TNCG) for reference-free evaluation of inter-frame temporal consistency. On OR-PAM-Reg-4K (432 test samples), PCReg-Net achieves NCC of 0.983, SSIM of 0.982, and PSNR of 46.96 dB, surpassing the state-of-the-art by over 14 dB at real-time speed. Code is available at https://github.com/JiahaoQin/PCReg-Net

Method: Constructs a 1.6B-sample training corpus with rigorous cleaning/stratification; uses multi-stage training (pre-training, supervised fine-tuning, reinforcement learning); introduces Multi-Condition Aware Bucket Sampler for variable-resolution batching, Stochastic Instruction Alignment, Asymmetric Gradient Optimization for DPO, DiffusionNFT with layout-aware OCR rewards, and differentiable Consistency Loss for identity preservation.

Result: Achieves competitive or superior performance on REDEdit-Bench (15 editing categories) and public benchmarks (ImgEdit and GEdit) against both open-source and proprietary systems.

Conclusion: FireRed-Image-Edit demonstrates state-of-the-art instruction-based image editing through systematic optimization, with released code, models, and benchmark suite to support future research.

Abstract: We present FireRed-Image-Edit, a diffusion transformer for instruction-based image editing that achieves state-of-the-art performance through systematic optimization of data curation, training methodology, and evaluation design. We construct a 1.6B-sample training corpus, comprising 900M text-to-image and 700M image editing pairs from diverse sources. After rigorous cleaning, stratification, auto-labeling, and two-stage filtering, we retain over 100M high-quality samples balanced between generation and editing, ensuring strong semantic coverage and instruction alignment. Our multi-stage training pipeline progressively builds editing capability via pre-training, supervised fine-tuning, and reinforcement learning. To improve data efficiency, we introduce a Multi-Condition Aware Bucket Sampler for variable-resolution batching and Stochastic Instruction Alignment with dynamic prompt re-indexing. To stabilize optimization and enhance controllability, we propose Asymmetric Gradient Optimization for DPO, DiffusionNFT with layout-aware OCR rewards for text editing, and a differentiable Consistency Loss for identity preservation. We further establish REDEdit-Bench, a comprehensive benchmark spanning 15 editing categories, including newly introduced beautification and low-level enhancement tasks. Extensive experiments on REDEdit-Bench and public benchmarks (ImgEdit and GEdit) demonstrate competitive or superior performance against both open-source and proprietary systems. We release code, models, and the benchmark suite to support future research.

Method: Construct labeled wildfire/smoke dataset from Landsat-8/9, GOES-16, and other Earth observation sources; use YOLOv12 for fire zone and smoke plume detection; integrate Multimodal Large Language Models (MLLMs) to convert detection outputs into contextualized risk assessments; validate with LLM-as-judge evaluation; deploy with service-oriented architecture.

Result: System demonstrates value of combining computer vision with language-based reasoning for scalable wildfire monitoring, supporting real-time processing, visual risk dashboards, and long-term wildfire tracking.

Conclusion: WildfireVLM shows the effectiveness of integrating satellite imagery detection with language-driven risk assessment for improved wildfire monitoring and disaster management.

Abstract: Wildfires are a growing threat to ecosystems, human lives, and infrastructure, with their frequency and intensity rising due to climate change and human activities. Early detection is critical, yet satellite-based monitoring remains challenging due to faint smoke signals, dynamic weather conditions, and the need for real-time analysis over large areas. We introduce WildfireVLM, an AI framework that combines satellite imagery wildfire detection with language-driven risk assessment. We construct a labeled wildfire and smoke dataset using imagery from Landsat-8/9, GOES-16, and other publicly available Earth observation sources, including harmonized products with aligned spectral bands. WildfireVLM employs YOLOv12 to detect fire zones and smoke plumes, leveraging its ability to detect small, complex patterns in satellite imagery. We integrate Multimodal Large Language Models (MLLMs) that convert detection outputs into contextualized risk assessments and prioritized response recommendations for disaster management. We validate the quality of risk reasoning using an LLM-as-judge evaluation with a shared rubric. The system is deployed using a service-oriented architecture that supports real-time processing, visual risk dashboards, and long-term wildfire tracking, demonstrating the value of combining computer vision with language-based reasoning for scalable wildfire monitoring.

Method: Fine-tune Qwen2-VL-7B vision-language model with multi-task learning on 1000 human-created paintings scored 1-100 with expert rubrics (originality, color, texture, composition, content). Add lightweight regression head on visual encoder to predict scores and generate rubric-aligned feedback in single forward pass, using structured rubric and artwork descriptions in system prompt.

Result: Strong accuracy with Pearson r > 0.97 and MAE ~3.95 on 100-point scale. Generated feedback shows high semantic similarity to expert critiques (average SBERT cosine similarity = 0.798).

Conclusion: The approach bridges computer vision and art assessment, offering a scalable tool for creativity research and classroom feedback through automated scoring with explanatory feedback.

Abstract: Assessing artistic creativity is foundational to creativity research and arts education, yet manual scoring (e.g., Torrance Tests of Creative Thinking) is labor-intensive at scale. Prior machine-learning approaches show promise for visual creativity scoring, but many rely mainly on image features and provide limited or no explanatory feedback. We propose a framework for automated creativity assessment of human paintings by fine-tuning the vision-language model Qwen2-VL-7B with multi-task learning. Our dataset contains 1000 human-created paintings scored on a 1-100 scale and paired with a short human-written description (content or artist explanation). Two expert raters evaluated each work using a five-dimension rubric (originality, color, texture, composition, content) and provided written critiques; we use an 80/20 train-test split. We add a lightweight regression head on the visual encoder output so the model can predict a numerical score and generate rubric-aligned feedback in a single forward pass. By embedding the structured rubric and the artwork description in the system prompt, we constrain the generated text to match the quantitative prediction. Experiments show strong accuracy, achieving Pearson r > 0.97 and MAE about 3.95 on the 100-point scale. Qualitative evaluation indicates the generated feedback is semantically close to expert critiques (average SBERT cosine similarity = 0.798). The proposed approach bridges computer vision and art assessment and offers a scalable tool for creativity research and classroom feedback.

Method: Proposes Visual Para-Thinker framework with two visual partitioning strategies for parallelized reasoning. Integrates Pa-Attention and LPRoPE to maintain path independence and promote reasoning diversity. Built on vLLM framework for efficient multimodal parallel processing.

Result: Empirical results on V*, CountBench, RefCOCO, and HallusionBench benchmarks confirm successful extension of parallel reasoning benefits to visual domain.

Conclusion: Visual Para-Thinker successfully extends parallel reasoning paradigm to multimodal settings, addressing exploration limitations of depth-based scaling in visual understanding tasks.

Abstract: Existing LLM test-time scaling laws emphasize the emergence of self-reflective behaviors through extended reasoning length. Nevertheless, this vertical scaling strategy often encounters plateaus in exploration as the model becomes locked into specific thinking pattern. By shifting from depth to parallelism, parallel thinking mitigates the narrowing of exploration. However, the extension of this paradigm to visual domain remains an open research question. In this paper, we first examine the role of visual partitioning in parallelized reasoning and subsequently propose two distinct strategies. Based on the above, we introduce Visual Para-Thinker, representing the inaugural parallel reasoning framework for MLLMs. To maintain path independence and promote diversity in reasoning, our approach integrates Pa-Attention alongside LPRoPE. Leveraging the vLLM framework, we have developed a native multimodal implementation that facilitates high-efficiency parallel processing. Empirical results on benchmark datasets such as V*, CountBench, RefCOCO, and HallusionBench confirm that Visual Para-Thinker successfully extends the benefits of parallel reasoning to the visual domain.

Method: Proposes Agentic Spatio-Temporal Grounder (ASTG) with two specialized MLLM agents: Spatial Reasoning Agent (SRA) and Temporal Reasoning Agent (TRA). They work collaboratively in a propose-and-evaluation paradigm with visual memory and dialogue context to automate tube extraction, verification, and temporal localization.

Result: Outperforms existing weakly-supervised and zero-shot approaches by a margin and is comparable to some fully-supervised methods on popular benchmarks.

Conclusion: ASTG enables open-world, training-free spatio-temporal video grounding through collaborative MLLM agents, addressing limitations of existing supervised approaches while maintaining competitive performance.

Abstract: Spatio-Temporal Video Grounding (STVG) aims to retrieve the spatio-temporal tube of a target object or person in a video given a text query. Most existing approaches perform frame-wise spatial localization within a predicted temporal span, resulting in redundant computation, heavy supervision requirements, and limited generalization. Weakly-supervised variants mitigate annotation costs but remain constrained by the dataset-level train-and-fit paradigm with an inferior performance. To address these challenges, we propose the Agentic Spatio-Temporal Grounder (ASTG) framework for the task of STVG towards an open-world and training-free scenario. Specifically, two specialized agents SRA (Spatial Reasoning Agent) and TRA (Temporal Reasoning Agent) constructed leveraging on modern Multimoal Large Language Models (MLLMs) work collaboratively to retrieve the target tube in an autonomous and self-guided manner. Following a propose-and-evaluation paradigm, ASTG duly decouples spatio-temporal reasoning and automates the tube extraction, verification and temporal localization processes. With a dedicate visual memory and dialogue context, the retrieval efficiency is significantly enhanced. Experiments on popular benchmarks demonstrate the superiority of the proposed approach where it outperforms existing weakly-supervised and zero-shot approaches by a margin and is comparable to some of the fully-supervised methods.

Method: End-to-end framework that reconstructs material-aware 3D scenes from single RGB images using monocular depth estimation, segmentation, and vision-language reasoning to infer object materials, then simulates mmWave propagation with configurable physics-based ray tracer using Fresnel reflection models parameterized by ITU-R electromagnetic properties.

Result: Sim2Radar improves downstream 3D radar perception via transfer learning: pre-training a radar point-cloud object detection model on synthetic data and fine-tuning on real radar yields up to +3.7 3D AP (IoU 0.3), with gains primarily from improved spatial localization.

Conclusion: Physics-based, vision-driven radar simulation can provide effective geometric priors for radar learning and measurably improve performance under limited real-data supervision.

Abstract: Millimeter-wave (mmWave) radar provides reliable perception in visually degraded indoor environments (e.g., smoke, dust, and low light), but learning-based radar perception is bottlenecked by the scarcity and cost of collecting and annotating large-scale radar datasets. We present Sim2Radar, an end-to-end framework that synthesizes training radar data directly from single-view RGB images, enabling scalable data generation without manual scene modeling. Sim2Radar reconstructs a material-aware 3D scene by combining monocular depth estimation, segmentation, and vision-language reasoning to infer object materials, then simulates mmWave propagation with a configurable physics-based ray tracer using Fresnel reflection models parameterized by ITU-R electromagnetic properties. Evaluated on real-world indoor scenes, Sim2Radar improves downstream 3D radar perception via transfer learning: pre-training a radar point-cloud object detection model on synthetic data and fine-tuning on real radar yields up to +3.7 3D AP (IoU 0.3), with gains driven primarily by improved spatial localization. These results suggest that physics-based, vision-driven radar simulation can provide effective geometric priors for radar learning and measurably improve performance under limited real-data supervision.

Method: Systematic analysis of importance-diversity trade-off, then proposes IDPruner using Maximal Marginal Relevance algorithm for Pareto-optimal balance, operating without attention maps for FlashAttention compatibility.

Result: State-of-the-art performance across various architectures and benchmarks. On Qwen2.5-VL-7B-Instruct: retains 95.18% performance with 75% pruning, 86.40% with 90% pruning.

Conclusion: IDPruner provides principled framework for visual token pruning that balances importance and diversity, enabling efficient MLLM inference with strong performance retention.

Abstract: Multimodal Large Language Models (MLLMs) have demonstrated impressive capabilities, yet they encounter significant computational bottlenecks due to the massive volume of visual tokens. Consequently, visual token pruning, which substantially reduces the token count, has emerged as a critical technique for accelerating MLLM inference. Existing approaches focus on token importance, diversity, or an intuitive combination of both, without a principled framework for their optimal integration. To address this issue, we first conduct a systematic analysis to characterize the trade-off between token importance and semantic diversity. Guided by this analysis, we propose the \textbf{I}mportance and \textbf{D}iversity Pruner (\textbf{IDPruner}), which leverages the Maximal Marginal Relevance (MMR) algorithm to achieve a Pareto-optimal balance between these two objectives. Crucially, our method operates without requiring attention maps, ensuring full compatibility with FlashAttention and efficient deployment via one-shot pruning. We conduct extensive experiments across various model architectures and multimodal benchmarks, demonstrating that IDPruner achieves state-of-the-art performance and superior generalization across diverse architectures and tasks. Notably, on Qwen2.5-VL-7B-Instruct, IDPruner retains 95.18% of baseline performance when pruning 75% of the tokens, and still maintains 86.40% even under an extreme 90% pruning ratio. Our code is available at https://github.com/Tencent/AngelSlim.

Method: Created PolyShapes-Ideal (PSI) dataset with three diagnostic probes: 1) polygon classification under noise, 2) zero-shot font transfer from MNIST, and 3) geometric collapse mapping under progressive deformation. Tested Eidos architecture on these benchmarks.

Result: Eidos architecture achieved >99% accuracy on PSI dataset and 81.67% zero-shot transfer across 30 unseen typefaces without pre-training. These results demonstrate strong topological invariance capabilities.

Conclusion: Results validate the “Form-First” hypothesis: generalization in structurally constrained architectures is a property of geometric integrity, not statistical scale. The PSI dataset provides diagnostic tools for evaluating topological invariance separate from texture-based learning.

Abstract: We present the PolyShapes-Ideal (PSI) dataset, a suite of diagnostic benchmarks designed to isolate topological invariance – the ability to maintain structural identity across affine transformations – from the textural correlations that dominate standard vision benchmarks. Through three diagnostic probes (polygon classification under noise, zero-shot font transfer from MNIST, and geometric collapse mapping under progressive deformation), we demonstrate that the Eidos architecture achieves >99% accuracy on PSI and 81.67% zero-shot transfer across 30 unseen typefaces without pre-training. These results validate the “Form-First” hypothesis: generalization in structurally constrained architectures is a property of geometric integrity, not statistical scale.

Method: Cascaded pipeline using Grounding DINO as text-promptable region proposer, then passing high-confidence frames to compact VLMs (Qwen and Gemma families, 4B-12B parameters) for semantic verification. Extended to agentic Scout-Commander workflow with novel “Controlled Input” methodology.

Result: Achieved up to 100% accuracy in false-positive filtering, 97.5% in damage assessment, 55-90% in vehicle classification. Agentic workflow achieved 100% correct asset deployment with 9.8/10 reasoning score and sub-75-second latency.

Conclusion: Hierarchical zero-shot architectures are validated for edge autonomy, with diagnostic framework for certifying VLM suitability in safety-critical applications. Different VLMs show distinct failure patterns in perception vs reasoning.

Abstract: Deploying autonomous edge robotics in dynamic military environments is constrained by both scarce domain-specific training data and the computational limits of edge hardware. This paper introduces a hierarchical, zero-shot framework that cascades lightweight object detection with compact Vision-Language Models (VLMs) from the Qwen and Gemma families (4B-12B parameters). Grounding DINO serves as a high-recall, text-promptable region proposer, and frames with high detection confidence are passed to edge-class VLMs for semantic verification. We evaluate this pipeline on 55 high-fidelity synthetic videos from Battlefield 6 across three tasks: false-positive filtering (up to 100% accuracy), damage assessment (up to 97.5%), and fine-grained vehicle classification (55-90%). We further extend the pipeline into an agentic Scout-Commander workflow, achieving 100% correct asset deployment and a 9.8/10 reasoning score (graded by GPT-4o) with sub-75-second latency. A novel “Controlled Input” methodology decouples perception from reasoning, revealing distinct failure phenotypes: Gemma3-12B excels at tactical logic but fails in visual perception, while Gemma3-4B exhibits reasoning collapse even with accurate inputs. These findings validate hierarchical zero-shot architectures for edge autonomy and provide a diagnostic framework for certifying VLM suitability in safety-critical applications.

Method: Two key innovations: 1) Unified motion representations that extract identity-agnostic motions and bind them to corresponding characters, 2) Holistic 4D-anchored paradigm that constructs shared 4D space to fuse motion representations with video latents, reinforced with hierarchical 4D-level supervision.

Result: MotionWeaver achieves state-of-the-art results on a new 300-video benchmark and generalizes effectively across diverse humanoid forms, complex interactions, and challenging multi-humanoid scenarios.

Conclusion: The paper presents a novel framework for multi-humanoid image animation that addresses limitations of existing single-human methods through unified motion representations and 4D-anchored supervision.

Abstract: Character image animation, which synthesizes videos of reference characters driven by pose sequences, has advanced rapidly but remains largely limited to single-human settings. Existing methods struggle to generalize to multi-humanoid scenarios, which involve diverse humanoid forms, complex interactions, and frequent occlusions. We address this gap with two key innovations. First, we introduce unified motion representations that extract identity-agnostic motions and explicitly bind them to corresponding characters, enabling generalization across diverse humanoid forms and seamless extension to multi-humanoid scenarios. Second, we propose a holistic 4D-anchored paradigm that constructs a shared 4D space to fuse motion representations with video latents, and further reinforces this process with hierarchical 4D-level supervision to better handle interactions and occlusions. We instantiate these ideas in MotionWeaver, an end-to-end framework for multi-humanoid image animation. To support this setting, we curate a 46-hour dataset of multi-human videos with rich interactions, and construct a 300-video benchmark featuring paired humanoid characters. Quantitative and qualitative experiments demonstrate that MotionWeaver not only achieves state-of-the-art results on our benchmark but also generalizes effectively across diverse humanoid forms, complex interactions, and challenging multi-humanoid scenarios.

Method: Proposes HiST-VLA with: 1) integration of geometric awareness with fine-grained driving commands and state history prompting, 2) dynamic token sparsification that fuses redundant tokens instead of filtering, 3) hierarchical transformer-based planner that refines coarse VLA waypoints into fine-grained trajectories, and 4) dynamic latent regularization to incorporate language commands for spatial grounding.

Result: Achieves state-of-the-art performance on NAVSIM v2 benchmark with EPDMS of 88.6 on Navtest and 50.9 on pseudo closed-loop Navhard benchmark.

Conclusion: HiST-VLA addresses key limitations of VLA models for autonomous driving through enhanced 3D spatial-temporal reasoning, computational efficiency via token fusion, and hierarchical trajectory refinement with strict spatial grounding.

Abstract: Vision-Language-Action (VLA) models offer promising capabilities for autonomous driving through multimodal understanding. However, their utilization in safety-critical scenarios is constrained by inherent limitations, including imprecise numerical reasoning, weak 3D spatial awareness, and high sensitivity to context. To address these challenges, we propose HiST-VLA, a novel Hierarchical Spatio-Temporal VLA model designed for reliable trajectory generation. Our framework enhances 3D spatial and temporal reasoning by integrating geometric awareness with fine-grained driving commands and state history prompting. To ensure computational efficiency, we integrate dynamic token sparsification into the VLA architecture. This approach fuses redundant tokens rather than filtering them, effectively reducing redundancy without sacrificing model performance. Furthermore, we employ a hierarchical transformer-based planner to progressively refine coarse VLA waypoints into fine-grained trajectories. Crucially, the planner utilizes dynamic latent regularization to incorporate language commands, ensuring strict spatial grounding and temporal coherence. Extensive evaluation on the NAVSIM v2 benchmark demonstrates state-of-the-art performance on Navtest, achieving an EPDMS of 88.6, and EPDMS of 50.9 on pseudo closed-loop Navhard benchmark.

Method: Deploys deep learning models for bioacoustic classification and image-based recognition on edge hardware, with acoustic activity detection to reduce energy consumption and fine-grained visual detection pipelines.

Result: Demonstrates that accurate bird species identification is feasible on embedded platforms, enabling real-time monitoring with reduced energy consumption.

Conclusion: The system supports scalable biodiversity monitoring through multimodal edge computing, making bird species monitoring more accessible for conservation and citizen science.

Abstract: This paper presents Zwitscherkasten, a DiY, multimodal system for bird species monitoring using audio and visual data on edge devices. Deep learning models for bioacoustic and image-based classification are deployed on resource-constrained hardware, enabling real-time, non-invasive monitoring. An acoustic activity detector reduces energy consumption, while visual recognition is performed using fine-grained detection and classification pipelines. Results show that accurate bird species identification is feasible on embedded platforms, supporting scalable biodiversity monitoring and citizen science applications.

Method: Proposes MedScope with coarse-to-fine evidence seeking, interleaving reasoning with targeted tool calls and verification. Uses Grounding-Aware Group Relative Policy Optimization (GA-GRPO) with grounding-aligned rewards and evidence-weighted advantages. Built ClinVideoSuite for supervision.

Result: Achieves state-of-the-art performance on both in-domain and out-of-domain video understanding benchmarks, with more accurate and trustworthy predictions explicitly grounded in temporally localized visual evidence.

Conclusion: MedScope illuminates a path toward medical AI agents that can genuinely “think with videos” through tool-integrated reasoning, advancing clinical video understanding with explicit evidence grounding.

Abstract: Long-form clinical videos are central to visual evidence-based decision-making, with growing importance for applications such as surgical robotics and related settings. However, current multimodal large language models typically process videos with passive sampling or weakly grounded inspection, which limits their ability to iteratively locate, verify, and justify predictions with temporally targeted evidence. To close this gap, we propose MedScope, a tool-using clinical video reasoning model that performs coarse-to-fine evidence seeking over long-form procedures. By interleaving intermediate reasoning with targeted tool calls and verification on retrieved observations, MedScope produces more accurate and trustworthy predictions that are explicitly grounded in temporally localized visual evidence. To address the lack of high-fidelity supervision, we build ClinVideoSuite, an evidence-centric, fine-grained clinical video suite. We then optimize MedScope with Grounding-Aware Group Relative Policy Optimization (GA-GRPO), which directly reinforces tool use with grounding-aligned rewards and evidence-weighted advantages. On full and fine-grained video understanding benchmarks, MedScope achieves state-of-the-art performance in both in-domain and out-of-domain evaluations. Our approach illuminates a path toward medical AI agents that can genuinely “think with videos” through tool-integrated reasoning. We will release our code, models, and data.

Method: Collaborative framework with lightweight generalist ViT on edge device and multiple medium-sized expert ViTs on near-edge accelerator. Novel routing mechanism uses edge model’s Top-k predictions to dynamically select relevant expert for low-confidence samples. Progressive specialist training strategy enhances expert accuracy on dataset subsets.

Result: On CIFAR-100 with real-world testbed: training strategy improves expert specialization accuracy by 4.12% on target subsets and overall accuracy by 2.76% over static experts. Reduces latency by up to 45% compared to edge execution, and energy consumption by up to 46% compared to near-edge offload.

Conclusion: Proposed collaborative inference framework effectively balances edge and near-edge computation for Vision Transformers, achieving significant improvements in accuracy, latency, and energy efficiency.

Abstract: Deploying Vision Transformers on edge devices is challenging due to their high computational complexity, while full offloading to cloud resources presents significant latency overheads. We propose a novel collaborative inference framework, which orchestrates a lightweight generalist ViT on an edge device and multiple medium-sized expert ViTs on a near-edge accelerator. A novel routing mechanism uses the edge model’s Top-$\mathit{k}$ predictions to dynamically select the most relevant expert for samples with low confidence. We further design a progressive specialist training strategy to enhance expert accuracy on dataset subsets. Extensive experiments on the CIFAR-100 dataset using a real-world edge and near-edge testbed demonstrate the superiority of our framework. Specifically, the proposed training strategy improves expert specialization accuracy by 4.12% on target subsets and enhances overall accuracy by 2.76% over static experts. Moreover, our method reduces latency by up to 45% compared to edge execution, and energy consumption by up to 46% compared to just near-edge offload.

Method: Uses discrete wavelet transform to extract frequency band features, proposes adaptive weighting mechanism to fuse spatial and frequency domain information, and applies this to few-shot meningioma classification on MRI data.

Result: Outperforms state-of-the-art methods on three datasets, including a newly introduced MRI meningioma dataset, demonstrating effectiveness of adaptive feature fusion.

Conclusion: Adaptive fusion of spatial-frequency domain features significantly improves meningioma classification, especially in few-shot learning scenarios, with potential clinical applications.

Abstract: The biological behavior and treatment response of meningiomas depend on their grade, making an accurate diagnosis essential for treatment planning and prognosis assessment. We observed that the weighted fusion of spatial-frequency domain features significantly influences meningioma classification performance. Notably, the contribution of specific frequency bands obtained by discrete wavelet transform varies considerably across different images. A feature fusion architecture with adaptive weights of different frequency band information and spatial domain information is proposed for few-shot meningioma learning. To verify the effectiveness of the proposed method, a new MRI dataset of meningiomas is introduced. The experimental results demonstrate the superiority of the proposed method compared with existing state-of-the-art methods in three datasets. The code will be available at: https://github.com/ICL-SUST/AMSF-Net

Method: Applied semantic segmentation on Google Street View images to extract visual environmental features, used Double Machine Learning framework to quantify causal effects on regional crashes, utilized SHAP values for nonlinear influence analysis, and applied causal forests for conditional average treatment effects.

Result: Greenery proportion has significant negative causal effect on traffic crashes (ATE = -6.38, p = 0.005). Protective effect is strongest in densely populated, socially vulnerable urban cores. Greenery mitigates angle and rear-end crashes but has limited protective benefit for vulnerable road users.

Conclusion: Provides causal evidence for greening as a safety intervention, suggests prioritizing hazardous visual environments, and highlights need for distinct design optimizations to protect vulnerable road users.

Abstract: Macroscopic traffic safety modeling aims to identify critical risk factors for regional crashes, thereby informing targeted policy interventions for safety improvement. However, current approaches rely heavily on static sociodemographic and infrastructure metrics, frequently overlooking the impacts from drivers’ visual perception of driving environment. Although visual environment features have been found to impact driving and traffic crashes, existing evidence remains largely observational, failing to establish the robust causality for traffic policy evaluation under complex spatial environment. To fill these gaps, we applied semantic segmentation on Google Street View imageries to extract visual environmental features and proposed a Double Machine Learning framework to quantify their causal effects on regional crashes. Meanwhile, we utilized SHAP values to characterize the nonlinear influence mechanisms of confounding variables in the models and applied causal forests to estimate conditional average treatment effects. Leveraging crash records from the Miami metropolitan area, Florida, and 220,000 street view images, evidence shows that greenery proportion exerts a significant and robust negative causal effect on traffic crashes (Average Treatment Effect = -6.38, p = 0.005). This protective effect exhibits spatial heterogeneity, being most pronounced in densely populated and socially vulnerable urban cores. While greenery significantly mitigates angle and rear-end crashes, its protective benefit for vulnerable road users (VRUs) remains limited. Our findings provide causal evidence for greening as a potential safety intervention, prioritizing hazardous visual environments while highlighting the need for distinct design optimizations to protect VRUs.

Method: Represents bin states as item-aligned instance masks and uses a latent diffusion transformer to predict post-stow configurations from observed context and planned stow behavior.

Result: FOREST substantially improves geometric agreement between predicted and true post-stow layouts compared to heuristic baselines, with only modest performance loss in downstream tasks when using predictions.

Conclusion: FOREST provides useful foresight signals for warehouse planning by accurately predicting post-stow configurations, enabling better load-quality assessment and multi-stow reasoning.

Abstract: Automated warehouses execute millions of stow operations, where robots place objects into storage bins. For these systems it is valuable to anticipate how a bin will look from the current observations and the planned stow behavior before real execution. We propose FOREST, a stow-intent-conditioned world model that represents bin states as item-aligned instance masks and uses a latent diffusion transformer to predict the post-stow configuration from the observed context. Our evaluation shows that FOREST substantially improves the geometric agreement between predicted and true post-stow layouts compared with heuristic baselines. We further evaluate the predicted post-stow layouts in two downstream tasks, in which replacing the real post-stow masks with FOREST predictions causes only modest performance loss in load-quality assessment and multi-stow reasoning, indicating that our model can provide useful foresight signals for warehouse planning.

Method: Proposes a new pipeline offering fully automated, scalable solution for generating marketing images of commercial products using text-to-image models. The system maintains image quality and fidelity while introducing creative variation that adheres to marketing guidelines.

Result: Achieves 30.77% increase in marketing object fidelity using DINOV2 and 52.00% increase in human preference over generated outcomes compared to baseline approaches.

Conclusion: The proposed pipeline successfully balances automation and human oversight, enabling scalable deployment of text-to-image models in production while maintaining quality standards and creative alignment for marketing applications.

Abstract: Text-to-image models have made significant strides, producing impressive results in generating images from textual descriptions. However, creating a scalable pipeline for deploying these models in production remains a challenge. Achieving the right balance between automation and human feedback is critical to maintain both scale and quality. While automation can handle large volumes, human oversight is still an essential component to ensure that the generated images meet the desired standards and are aligned with the creative vision. This paper presents a new pipeline that offers a fully automated, scalable solution for generating marketing images of commercial products using text-to-image models. The proposed system maintains the quality and fidelity of images, while also introducing sufficient creative variation to adhere to marketing guidelines. By streamlining this process, we ensure a seamless blend of efficiency and human oversight, achieving a $30.77%$ increase in marketing object fidelity using DINOV2 and a $52.00%$ increase in human preference over the generated outcome.

Method: Created a high-resolution (0.149m/pixel) dataset of 1.3M image tiles across 5 regions. Proposed ClimateGraph, a region-adaptive graph-based model that captures spatial and directional structure in kiln layouts. Compared against graph learning baselines, remote sensing pipelines, and recent foundation models for satellite imagery.

Result: Results show complementary strengths across graph-based, foundation model, and remote sensing approaches. Provides practical guidance for scalable brick kiln monitoring from satellite imagery.

Conclusion: Multiple approaches (graph-based, foundation models, remote sensing) offer complementary strengths for scalable brick kiln detection from satellite imagery, enabling better environmental and labor monitoring.

Abstract: Brick kilns are a major source of air pollution and forced labor in South Asia, yet large-scale monitoring remains limited by sparse and outdated ground data. We study brick kiln detection at scale using high-resolution satellite imagery and curate a multi city zoom-20 (0.149 meters per pixel) resolution dataset comprising over 1.3 million image tiles across five regions in South and Central Asia. We propose ClimateGraph, a region-adaptive graph-based model that captures spatial and directional structure in kiln layouts, and evaluate it against established graph learning baselines. In parallel, we assess a remote sensing based detection pipeline and benchmark it against recent foundation models for satellite imagery. Our results highlight complementary strengths across graph, foundation, and remote sensing approaches, providing practical guidance for scalable brick kiln monitoring from satellite imagery.

Method: Uses Google Cloud Translator to generate Hindi captions from Flickr8k dataset. Extracts visual features using pre-trained CNNs (VGG16, ResNet50, Inception V3). Employs uni-directional and bi-directional LSTM with attention mechanisms for text encoding. Attention layer generates weight vectors to combine image features at each time step into sentence-level feature vectors.

Result: Attention-based bidirectional LSTM with VGG16 produced the best results: BLEU-1 score of 0.59 and BLEU-4 score of 0.19. The system demonstrates ability to produce relevant, semantically accurate Hindi image captions.

Conclusion: The research successfully develops a multimodal architecture for Hindi image captioning, achieving good performance metrics. The model can guide future research in multilingual multimodal captioning systems.

Abstract: Automated image captioning using the content from the image is very appealing when done by harnessing the capability of computer vision and natural language processing. Extensive research has been done in the field with a major focus on the English language which gives the scope for further developments in the same with consideration of popular foreign languages. This research utilizes distinct models for translating the image caption into Hindi, the fourth most popular language across the world. Exploring the multi-modal architectures this research comprises local visual features, global visual features, attention mechanisms, and pre-trained models. Using google cloud translator on the image dataset from Flickr8k, Hindi image descriptions have been generated. Pre-trained CNNs like VGG16, ResNet50, and Inception V3 helped in retrieving image characteristics, while the uni-directional and bi-directional techniques of text encoding are used for the text encoding process. An additional Attention layer helps to generate a weight vector and, by multiplying it, combine image characteristics from each time step into a sentence-level feature vector. Bilingual evaluation understudy scores are used to compare the research outcome. Many experiments that serve as a baseline are done for the comparative analysis of the research. An image with a score of BLEU-1 is considered sufficient, whereas one with a score of BLEU-4 is considered to have fluid image captioning. For both BLEU scores, the attention-based bidirectional LSTM with VGG16 produced the best results of 0.59 and 0.19 respectively. The experiments conclude that researchs ability to produce relevant, semantically accurate image captions in Hindi. The research accomplishes the goals and future research can be guided by this research model.

Method: AdaCorrection uses lightweight spatio-temporal signals to estimate cache validity at each timestep, then adaptively blends cached and fresh activations. This on-the-fly correction requires no additional supervision or retraining.

Result: Maintains near-original FID scores while providing moderate acceleration. Experiments on image and video diffusion benchmarks show consistent generation performance improvements.

Conclusion: AdaCorrection enables efficient cache reuse in Diffusion Transformers while preserving high generation fidelity through adaptive correction, offering a practical solution to accelerate diffusion inference without quality degradation.

Abstract: Diffusion Transformers (DiTs) achieve state-of-the-art performance in high-fidelity image and video generation but suffer from expensive inference due to their iterative denoising structure. While prior methods accelerate sampling by caching intermediate features, they rely on static reuse schedules or coarse-grained heuristics, which often lead to temporal drift and cache misalignment that significantly degrade generation quality. We introduce \textbf{AdaCorrection}, an adaptive offset cache correction framework that maintains high generation fidelity while enabling efficient cache reuse across Transformer layers during diffusion inference. At each timestep, AdaCorrection estimates cache validity with lightweight spatio-temporal signals and adaptively blends cached and fresh activations. This correction is computed on-the-fly without additional supervision or retraining. Our approach achieves strong generation quality with minimal computational overhead, maintaining near-original FID while providing moderate acceleration. Experiments on image and video diffusion benchmarks show that AdaCorrection consistently improves generation performance.

Method: Proposes Dual-Channel Diffusion Separation Model (DCDSM) that uses diffusion models to learn source image feature distributions. Includes a Wavelet Suppression Module (WSM) in the dual-branch reverse denoising process to form an interactive separation network that exploits mutual coupling noise between source images.

Result: Achieves state-of-the-art performance: 1) For rain/snow removal: PSNR/SSIM of 35.0023 dB/0.9549 (rain) and 29.8108 dB/0.9243 (snow), outperforming Histoformer and LDRCNet; 2) For complex mixture separation: average PSNR/SSIM of 25.0049 dB/0.7997, surpassing comparative methods by 4.1249 dB/0.0926.

Conclusion: DCDSM demonstrates superiority in addressing rain/snow residue removal and detail preservation challenges through both subjective and objective evaluations, showing the effectiveness of diffusion models for blind image separation tasks.

Abstract: Blind image separation (BIS) refers to the inverse problem of simultaneously estimating and restoring multiple independent source images from a single observation image under conditions of unknown mixing mode and without prior knowledge of the source images. Traditional methods relying on statistical independence assumptions or CNN/GAN variants struggle to characterize complex feature distributions in real scenes, leading to estimation bias, texture distortion, and artifact residue under strong noise and nonlinear mixing. This paper innovatively introduces diffusion models into dual-channel BIS, proposing an efficient Dual-Channel Diffusion Separation Model (DCDSM). DCDSM leverages diffusion models’ powerful generative capability to learn source image feature distributions and reconstruct feature structures effectively. A novel Wavelet Suppression Module (WSM) is designed within the dual-branch reverse denoising process, forming an interactive separation network that enhances detail separation by exploiting the mutual coupling noise characteristic between source images. Extensive experiments on synthetic datasets containing rain/snow and complex mixtures demonstrate that DCDSM achieves state-of-the-art performance: 1) In image restoration tasks, it obtains PSNR/SSIM values of 35.0023 dB/0.9549 and 29.8108 dB/0.9243 for rain and snow removal respectively, outperforming Histoformer and LDRCNet by 1.2570 dB/0.9272 dB (PSNR) and 0.0262/0.0289 (SSIM) on average; 2) For complex mixture separation, the restored dual-source images achieve average PSNR and SSIM of 25.0049 dB and 0.7997, surpassing comparative methods by 4.1249 dB and 0.0926. Both subjective and objective evaluations confirm DCDSM’s superiority in addressing rain/snow residue removal and detail preservation challenges.

Method: Proposes two automated metrics: Recall_OCR (extracts text via OCR from input image as proxy reference to estimate content coverage) and Precision_VE (uses visual entailment against original image to detect hallucinated elements). Their harmonic mean F1_OCR-VE provides unified quality score.

Result: Validation on FlowVQA dataset shows strong agreement with ground-truth metrics: average Pearson’s r = 0.97, 0.91, and 0.94 for Recall, Precision, and F1 respectively, confirming reliability as practical reference-free alternative.

Conclusion: The framework provides reliable, reference-free quality monitoring for flowchart image-to-code generation in production settings, enabling continuous assessment without ground truth data.

Abstract: Vision-Language Models (VLMs) are increasingly used in document processing pipelines to convert flowchart images into structured code (e.g., Mermaid). In production, these systems process arbitrary inputs for which no ground-truth code exists, making output quality difficult to assess. We propose a reference-free evaluation framework that monitors flowchart image-to-code generation quality at inference time, using only the input image and the generated output. The framework introduces two automated metrics: $\text{Recall}{\text{OCR}}$, which estimates content coverage by extracting text from the input image via OCR as a proxy reference, and $\text{Precision}{\text{VE}}$, which detects hallucinated elements through Visual Entailment against the original image. Their harmonic mean, $\text{F1}{\text{OCR-VE}}$, provides a unified quality score. Validation on the FlowVQA dataset shows strong agreement with ground-truth metrics (average Pearson’s $r = 0.97$, $0.91$, and $0.94$ for Recall, Precision, and F1, respectively), confirming the framework’s reliability as a practical, reference-free alternative for continuous quality monitoring in production settings.

Method: Four integrated techniques: 1) Partial Convolution C2f (PC-C2f) reduces redundant computation, 2) Attention-Guided Feature Pyramid Network (AG-FPN) improves multi-scale fusion, 3) Auxiliary P2 detection head for sub-8×8 pixel objects, 4) Wise-IoU v3 for stable regression under noisy annotations.

Result: Achieves 35.1±0.3% mAP@0.5 on VisDrone-DET2019 (3.3 points over YOLOv10n) with 2.3M parameters, 35.8±0.4% mAP@0.5 on UAVDT, and 24.3 FPS on NVIDIA Jetson Orin Nano at FP16 precision.

Conclusion: The joint integration of complementary techniques within YOLOv10 framework effectively addresses UAV-specific detection challenges while maintaining computational efficiency suitable for embedded deployment.

Abstract: Unmanned aerial vehicles serve as primary sensing platforms for surveillance, traffic monitoring, and disaster response, making aerial object detection a central problem in applied computer vision. Current detectors struggle with UAV-specific challenges: targets spanning only a few pixels, cluttered backgrounds, heavy occlusion, and strict onboard computational budgets. This study introduces LAF-YOLOv10, built on YOLOv10n, integrating four complementary techniques to improve small-object detection in drone imagery. A Partial Convolution C2f (PC-C2f) module restricts spatial convolution to one quarter of backbone channels, reducing redundant computation while preserving discriminative capacity. An Attention-Guided Feature Pyramid Network (AG-FPN) inserts Squeeze-and-Excitation channel gates before multi-scale fusion and replaces nearest-neighbor upsampling with DySample for content-aware interpolation. An auxiliary P2 detection head at 160$\times$160 resolution extends localization to objects below 8$\times$8 pixels, while the P5 head is removed to redistribute parameters. Wise-IoU v3 replaces CIoU for bounding box regression, attenuating gradients from noisy annotations in crowded aerial scenes. The four modules address non-overlapping bottlenecks: PC-C2f compresses backbone computation, AG-FPN refines cross-scale fusion, the P2 head recovers spatial resolution, and Wise-IoU stabilizes regression under label noise. No individual component is novel; the contribution is the joint integration within a single YOLOv10 framework. Across three training runs (seeds 42, 123, 256), LAF-YOLOv10 achieves 35.1$\pm$0.3% mAP@0.5 on VisDrone-DET2019 with 2.3,M parameters, exceeding YOLOv10n by 3.3 points. Cross-dataset evaluation on UAVDT yields 35.8$\pm$0.4% mAP@0.5. Benchmarks on NVIDIA Jetson Orin Nano confirm 24.3 FPS at FP16, demonstrating viability for embedded UAV deployment.

Method: For Task 1 (long-tailed multi-label classification), they use an imbalance-aware multi-label learning strategy to improve tail class recognition while maintaining performance on frequent findings. For Task 2 (zero-shot OOD recognition), they propose a prediction approach that scores unseen disease categories without using any supervised labels or examples from OOD classes during training.

Result: The method achieves strong performance on both tasks, ranking first on the public leaderboard of the development phase of the CXR-LT 2026 challenge, as measured by macro-averaged mean Average Precision (mAP).

Conclusion: The proposed task-specific solutions effectively address the challenges of imperfect supervision in clinical CXR classification, demonstrating robust performance on both long-tailed multi-label learning and zero-shot out-of-distribution recognition.

Abstract: Chest X-ray (CXR) classification in clinical practice is often limited by imperfect supervision, arising from (i) extreme long-tailed multi-label disease distributions and (ii) missing annotations for rare or previously unseen findings. The CXR-LT 2026 challenge addresses these issues on a PadChest-based benchmark with a 36-class label space split into 30 in-distribution classes for training and 6 out-of-distribution (OOD) classes for zero-shot evaluation. We present task-specific solutions tailored to the distinct supervision regimes. For Task 1 (long-tailed multi-label classification), we adopt an imbalance-aware multi-label learning strategy to improve recognition of tail classes while maintaining stable performance on frequent findings. For Task 2 (zero-shot OOD recognition), we propose a prediction approach that produces scores for unseen disease categories without using any supervised labels or examples from the OOD classes during training. Evaluated with macro-averaged mean Average Precision (mAP), our method achieves strong performance on both tasks, ranking first on the public leaderboard of the development phase. Code and pre-trained models are available at https://github.com/hieuphamha19/CXR_LT.

Method: Created indoor dataset of 14,400 frames with inter-drone and ground vehicle footage using semi-automatic annotation. Benchmarked 3 replay-based CIL strategies (ER, MIR, FAR) using YOLOv11-nano detector under tight memory budgets (5-10% replay). Used Grad-CAM for attention analysis.

Result: FAR performed best under tight memory budgets, achieving 82.96% average accuracy with 5% replay. Grad-CAM analysis showed attention shifts across classes in mixed scenes, associated with reduced localization quality for drones.

Conclusion: Replay-based continual learning can be effectively applied to edge aerial systems. The work contributes an indoor UAV video dataset with temporal coherence and evaluation of replay-based CIL under limited replay budgets.

Abstract: Autonomous agents such as indoor drones must learn new object classes in real-time while limiting catastrophic forgetting, motivating Class-Incremental Learning (CIL). However, most unmanned aerial vehicle (UAV) datasets focus on outdoor scenes and offer limited temporally coherent indoor videos. We introduce an indoor dataset of $14,400$ frames capturing inter-drone and ground vehicle footage, annotated via a semi-automatic workflow with a $98.6%$ first-pass labeling agreement before final manual verification. Using this dataset, we benchmark 3 replay-based CIL strategies: Experience Replay (ER), Maximally Interfered Retrieval (MIR), and Forgetting-Aware Replay (FAR), using YOLOv11-nano as a resource-efficient detector for deployment-constrained UAV platforms. Under tight memory budgets ($5-10%$ replay), FAR performs better than the rest, achieving an average accuracy (ACC, $mAP_{50-95}$ across increments) of $82.96%$ with $5%$ replay. Gradient-weighted class activation mapping (Grad-CAM) analysis shows attention shifts across classes in mixed scenes, which is associated with reduced localization quality for drones. The experiments further demonstrate that replay-based continual learning can be effectively applied to edge aerial systems. Overall, this work contributes an indoor UAV video dataset with preserved temporal coherence and an evaluation of replay-based CIL under limited replay budgets. Project page: https://spacetime-vision-robotics-laboratory.github.io/learning-on-the-fly-cl

Method: Exploits asymmetric resolution requirements: OCR needs fine detail while scene understanding tolerates coarse features. Uses hybrid architecture with selective high-resolution OCR on-device and low-resolution video streaming for visual context.

Result: Achieves 72% accuracy at 0.49x power consumption of full-resolution streaming on benchmark of text-based VQA samples across five task categories.

Conclusion: Enables sustained VQA sessions on resource-constrained wearables without sacrificing text understanding quality by balancing resolution requirements.

Abstract: Video Large Language Models (Video LLMs) have shown remarkable progress in understanding and reasoning about visual content, particularly in tasks involving text recognition and text-based visual question answering (Text VQA). However, deploying Text VQA on wearable devices faces a fundamental tension: text recognition requires high-resolution video, but streaming high-quality video drains battery and causes thermal throttling. Moreover, existing models struggle to maintain coherent temporal context when processing text across multiple frames in real-time streams. We observe that text recognition and visual reasoning have asymmetric resolution requirements - OCR needs fine detail while scene understanding tolerates coarse features. We exploit this asymmetry with a hybrid architecture that performs selective high-resolution OCR on-device while streaming low-resolution video for visual context. On a benchmark of text-based VQA samples across five task categories, our system achieves 72% accuracy at 0.49x the power consumption of full-resolution streaming, enabling sustained VQA sessions on resource-constrained wearables without sacrificing text understanding quality.

Method: Large-scale study using novel dataset from 1,888 participants (727 with PD) with 32,847 videos across 16 standardized clinical tasks. Evaluated 7 state-of-the-art VFMs (VideoPrism, V-JEPA, ViViT, VideoMAE, TimeSformer) using frozen embeddings with linear classification head to assess model robustness in clinical screening.

Result: Task saliency is highly model-dependent: VideoPrism excels in visual speech kinematics and facial expressivity, V-JEPA superior for upper-limb motor tasks, TimeSformer competitive for rhythmic tasks. Achieved AUCs of 76.4-85.3% and accuracies of 71.5-80.6%. High specificity (up to 90.3%) but lower sensitivity (43.2-57.3%).

Conclusion: Establishes rigorous baseline for VFM-based PD screening and provides roadmap for selecting suitable tasks and architectures in remote neurological monitoring, highlighting need for task-aware calibration and multi-task/multi-modal integration.

Abstract: Remote, video-based assessments offer a scalable pathway for Parkinson’s disease (PD) screening. While traditional approaches rely on handcrafted features mimicking clinical scales, recent advances in video foundation models (VFMs) enable representation learning without task-specific customization. However, the comparative effectiveness of different VFM architectures across diverse clinical tasks remains poorly understood. We present a large-scale systematic study using a novel video dataset from 1,888 participants (727 with PD), comprising 32,847 videos across 16 standardized clinical tasks. We evaluate seven state-of-the-art VFMs – including VideoPrism, V-JEPA, ViViT, and VideoMAE – to determine their robustness in clinical screening. By evaluating frozen embeddings with a linear classification head, we demonstrate that task saliency is highly model-dependent: VideoPrism excels in capturing visual speech kinematics (no audio) and facial expressivity, while V-JEPA proves superior for upper-limb motor tasks. Notably, TimeSformer remains highly competitive for rhythmic tasks like finger tapping. Our experiments yield AUCs of 76.4-85.3% and accuracies of 71.5-80.6%. While high specificity (up to 90.3%) suggests strong potential for ruling out healthy individuals, the lower sensitivity (43.2-57.3%) highlights the need for task-aware calibration and integration of multiple tasks and modalities. Overall, this work establishes a rigorous baseline for VFM-based PD screening and provides a roadmap for selecting suitable tasks and architectures in remote neurological monitoring. Code and anonymized structured data are publicly available: https://anonymous.4open.science/r/parkinson_video_benchmarking-A2C5

Method: Proposes SpargeAttention2 with three components: (i) hybrid masking rule combining Top-k and Top-p for robust high-sparsity masking, (ii) efficient trainable sparse attention implementation, and (iii) distillation-inspired fine-tuning objective to preserve generation quality during sparse attention fine-tuning.

Result: Achieves 95% attention sparsity and 16.2x attention speedup on video diffusion models while maintaining generation quality, consistently outperforming prior sparse attention methods.

Conclusion: SpargeAttention2 successfully addresses key limitations of sparse attention methods, enabling extremely high sparsity (95%) with significant speedups while preserving generation quality in diffusion models.

Abstract: Many training-free sparse attention methods are effective for accelerating diffusion models. Recently, several works suggest that making sparse attention trainable can further increase sparsity while preserving generation quality. We study three key questions: (1) when do the two common masking rules, i.e., Top-k and Top-p, fail, and how can we avoid these failures? (2) why can trainable sparse attention reach higher sparsity than training-free methods? (3) what are the limitations of fine-tuning sparse attention using the diffusion loss, and how can we address them? Based on this analysis, we propose SpargeAttention2, a trainable sparse attention method that achieves high sparsity without degrading generation quality. SpargeAttention2 includes (i) a hybrid masking rule that combines Top-k and Top-p for more robust masking at high sparsity, (ii) an efficient trainable sparse attention implementation, and (iii) a distillation-inspired fine-tuning objective to better preserve generation quality during fine-tuning using sparse attention. Experiments on video diffusion models show that SpargeAttention2 reaches 95% attention sparsity and a 16.2x attention speedup while maintaining generation quality, consistently outperforming prior sparse attention methods.

Method: Integrates physically based rendering into 3D Gaussian Splatting with composite scene Gaussian representations. Jointly optimizes BRDF-based material properties, explicitly models diffuse components through global illumination and specular components via anisotropic spherical Gaussians.

Result: Outperforms state-of-the-art methods quantitatively and qualitatively across diverse nighttime scenarios on nuScenes and Waymo datasets. Maintains real-time rendering while improving reconstruction quality for outdoor nighttime driving scenes.

Conclusion: The approach successfully addresses nighttime scene reconstruction challenges in autonomous driving by integrating physically based rendering with 3DGS, enabling better handling of complex lighting conditions while preserving real-time performance.

Abstract: This paper focuses on scene reconstruction under nighttime conditions in autonomous driving simulation. Recent methods based on Neural Radiance Fields (NeRFs) and 3D Gaussian Splatting (3DGS) have achieved photorealistic modeling in autonomous driving scene reconstruction, but they primarily focus on normal-light conditions. Low-light driving scenes are more challenging to model due to their complex lighting and appearance conditions, which often causes performance degradation of existing methods. To address this problem, this work presents a novel approach that integrates physically based rendering into 3DGS to enhance nighttime scene reconstruction for autonomous driving. Specifically, our approach integrates physically based rendering into composite scene Gaussian representations and jointly optimizes Bidirectional Reflectance Distribution Function (BRDF) based material properties. We explicitly model diffuse components through a global illumination module and specular components by anisotropic spherical Gaussians. As a result, our approach improves reconstruction quality for outdoor nighttime driving scenes, while maintaining real-time rendering. Extensive experiments across diverse nighttime scenarios on two real-world autonomous driving datasets, including nuScenes and Waymo, demonstrate that our approach outperforms the state-of-the-art methods both quantitatively and qualitatively.

Method: Proposes Privacy-Concealing Cooperation (PCC) framework with three components: hiding network to conceal visual clues in BEV features, reconstruction network that tries to recover images, and perception network for segmentation. Uses adversarial learning where hiding and reconstruction networks compete, while perception network is integrated and optimized end-to-end.

Result: Effectively degrades reconstructed image quality with minimal impact on segmentation performance, providing privacy protection for cooperating vehicles.

Conclusion: PCC framework successfully balances privacy protection and perception performance in cooperative autonomous driving systems through adversarial feature concealment.

Abstract: Cooperative perception systems for autonomous driving aim to overcome the limited perception range of a single vehicle by communicating with adjacent agents to share sensing information. While this improves perception performance, these systems also face a significant privacy-leakage issue, as sensitive visual content can potentially be reconstructed from the shared data. In this paper, we propose a novel Privacy-Concealing Cooperation (PCC) framework for Bird’s Eye View (BEV) semantic segmentation. Based on commonly shared BEV features, we design a hiding network to prevent an image reconstruction network from recovering the input images from the shared features. An adversarial learning mechanism is employed to train the network, where the hiding network works to conceal the visual clues in the BEV features while the reconstruction network attempts to uncover these clues. To maintain segmentation performance, the perception network is integrated with the hiding network and optimized end-to-end. The experimental results demonstrate that the proposed PCC framework effectively degrades the quality of the reconstructed images with minimal impact on segmentation performance, providing privacy protection for cooperating vehicles. The source code will be made publicly available upon publication.

Method: Proposes Diff-Aid, a plug-and-play inference-time method that adaptively modulates per-token text and image interactions across transformer blocks and denoising timesteps. It provides interpretable modulation patterns and can be integrated into downstream applications like style LoRAs, controllable generation, and zero-shot editing.

Result: Experiments on SD 3.5 and FLUX baselines show consistent improvements in prompt adherence, visual quality, and human preference across various metrics. The method yields interpretable patterns revealing how different components contribute to semantic alignment.

Conclusion: Diff-Aid provides a flexible and efficient solution for enhancing text-image interactions in diffusion models, improving generation quality while offering interpretability and seamless integration into downstream applications.

Abstract: Recent text-to-image (T2I) diffusion models have achieved remarkable advancement, yet faithfully following complex textual descriptions remains challenging due to insufficient interactions between textual and visual features. Prior approaches enhance such interactions via architectural design or handcrafted textual condition weighting, but lack flexibility and overlook the dynamic interactions across different blocks and denoising stages. To provide a more flexible and efficient solution to this problem, we propose Diff-Aid, a lightweight inference-time method that adaptively adjusts per-token text and image interactions across transformer blocks and denoising timesteps. Beyond improving generation quality, Diff-Aid yields interpretable modulation patterns that reveal how different blocks, timesteps, and textual tokens contribute to semantic alignment during denoising. As a plug-and-play module, Diff-Aid can be seamlessly integrated into downstream applications for further improvement, including style LoRAs, controllable generation, and zero-shot editing. Experiments on strong baselines (SD 3.5 and FLUX) demonstrate consistent improvements in prompt adherence, visual quality, and human preference across various metrics. Our code and models will be released.

Method: TwInS uses two interactive streams: scene parsing stream for contextual features and geometric vision stream for spatial features. Features are bidirectionally fused using a cross-task adapter, with geometric features projected into contextual space. A semi-supervised training strategy eliminates need for human-annotated correspondence ground truth.

Result: Extensive experiments on three public datasets show TwInS outperforms state-of-the-art approaches, validating the effectiveness of its core components including the interactive streams and cross-task adapter.

Conclusion: TwInS successfully demonstrates that joint learning of scene parsing and geometric vision through interactive streams leads to superior performance, with the semi-supervised approach enabling large-scale training without costly annotations.

Abstract: Inspired by the human visual system, which operates on two parallel yet interactive streams for contextual and spatial understanding, this article presents Two Interactive Streams (TwInS), a novel bio-inspired joint learning framework capable of simultaneously performing scene parsing and geometric vision tasks. TwInS adopts a unified, general-purpose architecture in which multi-level contextual features from the scene parsing stream are infused into the geometric vision stream to guide its iterative refinement. In the reverse direction, decoded geometric features are projected into the contextual feature space for selective heterogeneous feature fusion via a novel cross-task adapter, which leverages rich cross-view geometric cues to enhance scene parsing. To eliminate the dependence on costly human-annotated correspondence ground truth, TwInS is further equipped with a tailored semi-supervised training strategy, which unleashes the potential of large-scale multi-view data and enables continuous self-evolution without requiring ground-truth correspondences. Extensive experiments conducted on three public datasets validate the effectiveness of TwInS’s core components and demonstrate its superior performance over existing state-of-the-art approaches. The source code will be made publicly available upon publication.

Method: Proposes AdaVBoost with Visual Grounding Entropy (VGE) to estimate hallucination risk by leveraging visual grounding as a complementary signal to capture evidence mismatches. The framework applies stronger visual attention boosting to high-risk tokens and weaker boosting to low-risk tokens, enabling token-level adaptive intervention at each generation step.

Result: Extensive experiments show AdaVBoost significantly outperforms baseline methods across multiple LVLMs and hallucination benchmarks, demonstrating effectiveness in mitigating hallucinations through adaptive attention boosting.

Conclusion: AdaVBoost provides an effective solution to the fundamental trade-off in visual attention boosting by enabling adaptive, token-level intervention based on real-time hallucination risk assessment, significantly improving hallucination mitigation in LVLMs.

Abstract: Visual attention boosting has emerged as a promising direction for mitigating hallucinations in Large Vision-Language Models (LVLMs), where existing methods primarily focus on where to boost by applying a predefined scaling to the attention of method-specific visual tokens during autoregressive generation. In this paper, we identify a fundamental trade-off in these methods: a predefined scaling factor can be too weak at some generation steps, leaving hallucinations unresolved, yet too strong at others, leading to new hallucinations. Motivated by this finding, we propose AdaVBoost, a token-level visual attention boosting framework that adaptively determines how much attention to boost at each generation step. Specifically, we introduce Visual Grounding Entropy (VGE) to estimate hallucination risk, which leverages visual grounding as a complementary signal to capture evidence mismatches beyond entropy. Guided by VGE, AdaVBoost applies stronger visual attention boosting to high-risk tokens and weaker boosting to low-risk tokens, enabling token-level adaptive intervention at each generation step. Extensive experiments show that AdaVBoost significantly outperforms baseline methods across multiple LVLMs and hallucination benchmarks.

Method: REVISE uses multi-round sparse frame selection, maintains a summary-as-state across rounds, and supports both proprietary VLMs (plug-and-play) and open-source models (via EAGER reinforcement fine-tuning). EAGER uses three reward terms: confidence gain, summary sufficiency, and correct-and-early stop.

Result: Across multiple VQA benchmarks, REVISE improves accuracy while significantly reducing frames processed, rounds needed, and prompt tokens used, demonstrating practical sparse video reasoning.

Conclusion: REVISE enables efficient video QA through sparse reasoning, making video understanding more practical by reducing computational costs while maintaining or improving performance.

Abstract: We present \revise (\underline{Re}asoning with \underline{Vi}deo \underline{S}parsity), a multi-round agent for video question answering (VQA). Instead of uniformly sampling frames, \revise selects a small set of informative frames, maintains a summary-as-state across rounds, and stops early when confident. It supports proprietary vision-language models (VLMs) in a ``plug-and-play’’ setting and enables reinforcement fine-tuning for open-source models. For fine-tuning, we introduce EAGER (Evidence-Adjusted Gain for Efficient Reasoning), an annotation-free reward with three terms: (1) Confidence gain: after new frames are added, we reward the increase in the log-odds gap between the correct option and the strongest alternative; (2) Summary sufficiency: at answer time we re-ask using only the last committed summary and reward success; (3) Correct-and-early stop: answering correctly within a small turn budget is rewarded. Across multiple VQA benchmarks, \revise improves accuracy while reducing frames, rounds, and prompt tokens, demonstrating practical sparse video reasoning.

Method: Developed ZEN using self-supervised multi-teacher distillation framework trained on a large diverse dataset of over 4 million frames from 21 different surgical procedures, with systematic evaluation of multiple representation learning strategies.

Result: ZEN consistently outperforms existing surgical foundation models across 20 downstream tasks in full fine-tuning, frozen-backbone, few-shot and zero-shot settings, demonstrating robust cross-procedure generalization.

Conclusion: ZEN represents a step toward unified representations for surgical scene understanding and supports future applications in intraoperative assistance and surgical training assessment.

Abstract: In minimally invasive surgery, clinical decisions depend on real-time visual interpretation, yet intraoperative perception varies substantially across surgeons and procedures. This variability limits consistent assessment, training, and the development of reliable artificial intelligence systems, as most surgical AI models are designed for narrowly defined tasks and do not generalize across procedures or institutions. Here we introduce ZEN, a generalizable foundation model for intraoperative surgical video understanding trained on more than 4 million frames from over 21 procedures using a self-supervised multi-teacher distillation framework. We curated a large and diverse dataset and systematically evaluated multiple representation learning strategies within a unified benchmark. Across 20 downstream tasks and full fine-tuning, frozen-backbone, few-shot and zero-shot settings, ZEN consistently outperforms existing surgical foundation models and demonstrates robust cross-procedure generalization. These results suggest a step toward unified representations for surgical scene understanding and support future applications in intraoperative assistance and surgical training assessment.

Method: Introduces a unified UAV tracking framework with: 1) Global-Grouped Coordinate Attention (GGCA) module for capturing long-range dependencies and global contexts with minimal computational overhead, 2) Similarity-Guided Layer Adaptation (SGLA) module to replace knowledge distillation for optimal balance between precision and efficiency, and 3) dynamic layer selection and efficient feature enhancement.

Result: Achieves state-of-the-art real-time speed (258.7 FPS on UAVDT) while maintaining competitive tracking accuracy (82.8% precision) across three datasets.

Conclusion: LGTrack provides an effective solution for UAV tracking that successfully balances accuracy and efficiency, making it suitable for real-time applications.

Abstract: Visual object tracking (VOT) plays a pivotal role in unmanned aerial vehicle (UAV) applications. Addressing the trade-off between accuracy and efficiency, especially under challenging conditions like unpredictable occlusion, remains a significant challenge. This paper introduces LGTrack, a unified UAV tracking framework that integrates dynamic layer selection, efficient feature enhancement, and robust representation learning for occlusions. By employing a novel lightweight Global-Grouped Coordinate Attention (GGCA) module, LGTrack captures long-range dependencies and global contexts, enhancing feature discriminability with minimal computational overhead. Additionally, a lightweight Similarity-Guided Layer Adaptation (SGLA) module replaces knowledge distillation, achieving an optimal balance between tracking precision and inference efficiency. Experiments on three datasets demonstrate LGTrack’s state-of-the-art real-time speed (258.7 FPS on UAVDT) while maintaining competitive tracking accuracy (82.8% precision). Code is available at https://github.com/XiaoMoc/LGTrack

Method: DCDM decomposes video consistency into three components: 1) LLM-parsed semantic representations for intra-clip consistency, 2) temporal camera representation in noise space for inter-clip camera control, and 3) holistic scene generation with windowed cross-attention and sparse inter-shot self-attention for long-range coherence.

Result: Validated on the CVM Competition at AAAI'26 test set, demonstrating that the proposed strategies effectively address the three consistency challenges in video generation.

Conclusion: DCDM provides a comprehensive framework for improving video generation consistency across multiple temporal scales through specialized components while maintaining computational efficiency.

Abstract: Recent video generative models have demonstrated impressive visual fidelity, yet they often struggle with semantic, geometric, and identity consistency. In this paper, we propose a system-level framework, termed the Divide-and-Conquer Diffusion Model (DCDM), to address three key challenges: (1) intra-clip world knowledge consistency, (2) inter-clip camera consistency, and (3) inter-shot element consistency. DCDM decomposes video consistency modeling under these scenarios into three dedicated components while sharing a unified video generation backbone. For intra-clip consistency, DCDM leverages a large language model to parse input prompts into structured semantic representations, which are subsequently translated into coherent video content by a diffusion transformer. For inter-clip camera consistency, we propose a temporal camera representation in the noise space that enables precise and stable camera motion control, along with a text-to-image initialization mechanism to further enhance controllability. For inter-shot consistency, DCDM adopts a holistic scene generation paradigm with windowed cross-attention and sparse inter-shot self-attention, ensuring long-range narrative coherence while maintaining computational efficiency. We validate our framework on the test set of the CVM Competition at AAAI'26, and the results demonstrate that the proposed strategies effectively address these challenges.

Method: Created a dataset of 1,534 questions with 2,043 associated medical images from Korean Medical Licensing Examinations (2012-2023), with about 30% containing multiple images requiring cross-image integration. Images cover various clinical modalities including X-ray, CT, ECG, ultrasound, and endoscopy. Evaluated over 50 VLMs across proprietary, open-source, general-purpose, medical-specialized, and Korean-specialized categories using zero-shot evaluation.

Result: Best proprietary model (Gemini-3.0-Pro) achieved 96.9% accuracy, best open-source model (Qwen3-VL-32B-Thinking) 83.7%, and best Korean-specialized model (VARCO-VISION-2.0-14B) only 43.2%. Key findings: reasoning-oriented models gain up to +20% over instruction-tuned counterparts, medical domain specialization yields inconsistent gains, all models degrade on multi-image questions, and performance varies across imaging modalities.

Conclusion: KorMedMCQA-V provides a valuable multimodal benchmark for Korean medical reasoning, revealing significant gaps in current VLMs’ capabilities, especially for Korean-specialized models and multi-image reasoning tasks. The benchmark complements text-only evaluation and highlights areas for improvement in medical multimodal understanding.

Abstract: We introduce KorMedMCQA-V, a Korean medical licensing-exam-style multimodal multiple-choice question answering benchmark for evaluating vision-language models (VLMs). The dataset consists of 1,534 questions with 2,043 associated images from Korean Medical Licensing Examinations (2012-2023), with about 30% containing multiple images requiring cross-image evidence integration. Images cover clinical modalities including X-ray, computed tomography (CT), electrocardiography (ECG), ultrasound, endoscopy, and other medical visuals. We benchmark over 50 VLMs across proprietary and open-source categories-spanning general-purpose, medical-specialized, and Korean-specialized families-under a unified zero-shot evaluation protocol. The best proprietary model (Gemini-3.0-Pro) achieves 96.9% accuracy, the best open-source model (Qwen3-VL-32B-Thinking) 83.7%, and the best Korean-specialized model (VARCO-VISION-2.0-14B) only 43.2%. We further find that reasoning-oriented model variants gain up to +20 percentage points over instruction-tuned counterparts, medical domain specialization yields inconsistent gains over strong general-purpose baselines, all models degrade on multi-image questions, and performance varies notably across imaging modalities. By complementing the text-only KorMedMCQA benchmark, KorMedMCQA-V forms a unified evaluation suite for Korean medical reasoning across text-only and multimodal conditions. The dataset is available via Hugging Face Datasets: https://huggingface.co/datasets/seongsubae/KorMedMCQA-V.

Method: Model POCUS as sequential acquisition problem with RL agent selecting next view or terminating acquisition. Upon termination, multi-view transformer performs multi-task inference (ordinal AS classification and LVEF regression) with Gaussian predictive distributions. Reward function balances expected diagnostic benefit against acquisition cost.

Result: On 1,820 test studies, method matches full-study performance using 32% fewer videos, achieving 77.2% mean balanced accuracy across AS severity classification and LVEF estimation. Demonstrates robust multi-task performance under acquisition budgets.

Conclusion: Patient-tailored, cost-aware acquisition can streamline POCUS workflows while preserving decision quality, producing interpretable scan pathways suited to bedside use. Framework extensible to additional cardiac endpoints.

Abstract: Purpose: Echocardiography with point-of-care ultrasound (POCUS) must support clinical decision-making under tight bedside time and operator-effort constraints. We introduce a personalized data acquisition strategy in which an RL agent, given a partially observed multi-view study, selects the next view to acquire or terminates acquisition to support heart-failure (HF) assessment. Upon termination, a diagnostic model jointly predicts aortic stenosis (AS) severity and left ventricular ejection fraction (LVEF), two key HF biomarkers, and outputs uncertainty, enabling an explicit trade-off between diagnostic performance and acquisition cost. Methods: We model POCUS as a sequential acquisition problem: at each step, a video selector (RL agent) chooses the next view to acquire or terminates acquisition. Upon termination, a shared multi-view transformer performs multi-task inference with two heads, ordinal AS classification, and LVEF regression, and outputs Gaussian predictive distributions yielding ordinal probabilities over AS classes and EF thresholds. These probabilities drive a reward that balances expected diagnostic benefit against acquisition cost, producing patient-specific acquisition pathways. Results: The dataset comprises 12,180 patient-level studies, split into training/validation/test sets (75/15/15). On the 1,820 test studies, our method matches full-study performance while using 32% fewer videos, achieving 77.2% mean balanced accuracy (bACC) across AS severity classification and LVEF estimation, demonstrating robust multi-task performance under acquisition budgets. Conclusion: Patient-tailored, cost-aware acquisition can streamline POCUS workflows while preserving decision quality, producing interpretable scan pathways suited to bedside use. The framework is extensible to additional cardiac endpoints and merits prospective evaluation for clinical integration.

Method: Created LeafNet dataset with 186K leaf images across 97 disease classes and LeafBench benchmark with 13,950 QA pairs across 6 agricultural tasks to evaluate 12 state-of-the-art VLMs.

Result: VLMs outperform vision-only models, with >90% accuracy on binary classification but <65% on fine-grained pathogen/species identification, showing significant performance gaps across tasks.

Conclusion: Multimodal architectures enhance diagnostic precision over vision-only models, but current VLMs have critical gaps in plant pathology, highlighting need for specialized benchmarks like LeafBench.

Abstract: Foundation models and vision-language pre-training have significantly advanced Vision-Language Models (VLMs), enabling multimodal processing of visual and linguistic data. However, their application in domain-specific agricultural tasks, such as plant pathology, remains limited due to the lack of large-scale, comprehensive multimodal image–text datasets and benchmarks. To address this gap, we introduce LeafNet, a comprehensive multimodal dataset, and LeafBench, a visual question-answering benchmark developed to systematically evaluate the capabilities of VLMs in understanding plant diseases. The dataset comprises 186,000 leaf digital images spanning 97 disease classes, paired with metadata, generating 13,950 question-answer pairs spanning six critical agricultural tasks. The questions assess various aspects of plant pathology understanding, including visual symptom recognition, taxonomic relationships, and diagnostic reasoning. Benchmarking 12 state-of-the-art VLMs on our LeafBench dataset, we reveal substantial disparity in their disease understanding capabilities. Our study shows performance varies markedly across tasks: binary healthy–diseased classification exceeds 90% accuracy, while fine-grained pathogen and species identification remains below 65%. Direct comparison between vision-only models and VLMs demonstrates the critical advantage of multimodal architectures: fine-tuned VLMs outperform traditional vision models, confirming that integrating linguistic representations significantly enhances diagnostic precision. These findings highlight critical gaps in current VLMs for plant pathology applications and underscore the need for LeafBench as a rigorous framework for methodological advancement and progress evaluation toward reliable AI-assisted plant disease diagnosis. Code is available at https://github.com/EnalisUs/LeafBench.

Method: Fourfold design: (1) Multi-Teacher Training with domain experts transferring knowledge to a student model; (2) Adaptive CFG Calibration (ACC-DMD) for audio CFG error calibration; (3) Hybrid Long Tail Forcing for alignment on tail frames; (4) VAE Decoder Refiner for pixel-domain optimization.

Result: Achieves few-pass autoregressive generation with substantially extended temporal consistency, identity preservation, and audio-lip synchronization, demonstrating improved performance in streaming mode.

Conclusion: EchoTorrent effectively addresses the efficiency-performance trade-off in multi-modal video generation, enabling real-time deployment with enhanced temporal stability and multimodal alignment.

Abstract: Recent multi-modal video generation models have achieved high visual quality, but their prohibitive latency and limited temporal stability hinder real-time deployment. Streaming inference exacerbates these issues, leading to pronounced multimodal degradation, such as spatial blurring, temporal drift, and lip desynchronization, which creates an unresolved efficiency-performance trade-off. To this end, we propose EchoTorrent, a novel schema with a fourfold design: (1) Multi-Teacher Training fine-tunes a pre-trained model on distinct preference domains to obtain specialized domain experts, which sequentially transfer domain-specific knowledge to a student model; (2) Adaptive CFG Calibration (ACC-DMD), which calibrates the audio CFG augmentation errors in DMD via a phased spatiotemporal schedule, eliminating redundant CFG computations and enabling single-pass inference per step; (3) Hybrid Long Tail Forcing, which enforces alignment exclusively on tail frames during long-horizon self-rollout training via a causal-bidirectional hybrid architecture, effectively mitigates spatiotemporal degradation in streaming mode while enhancing fidelity to reference frames; and (4) VAE Decoder Refiner through pixel-domain optimization of the VAE decoder to recover high-frequency details while circumventing latent-space ambiguities. Extensive experiments and analysis demonstrate that EchoTorrent achieves few-pass autoregressive generation with substantially extended temporal consistency, identity preservation, and audio-lip synchronization.

Method: Proposes an ensemble learning approach combining U-Net (good at capturing fine details) and FPN (effective at handling scale variation) using weighted average method. Uses dataset mimicking real-life waste scenarios with preprocessing to enhance feature learning.

Result: Ensemble model (EL-4) achieved IoU of 0.8306 (improved from U-Net’s 0.8065) and reduced Dice loss to 0.09019 (from FPN’s 0.1183).

Conclusion: The ensemble approach improves segmentation accuracy for waste segregation, potentially enhancing efficiency at Material Recovery Facilities with minimal human intervention.

Abstract: Environmental pollution is a critical global issue, with recycling emerging as one of the most viable solutions. This study focuses on waste segregation, a crucial step in recycling processes to obtain raw material. Recent advancements in computer vision have significantly contributed to waste classification and recognition. In waste segregation, segmentation masks are essential for robots to accurately localize and pick objects from conveyor belts. The complexity of real-world waste environments, characterized by deformed items without specific patterns and overlapping objects, further complicates waste segmentation tasks. This paper proposes an Ensemble Learning approach to improve segmentation accuracy by combining high performing segmentation models, U-Net and FPN, using a weighted average method. U-Net excels in capturing fine details and boundaries in segmentation tasks, while FPN effectively handles scale variation and context in complex environments, and their combined masks result in more precise predictions. The dataset used closely mimics real-life waste scenarios, and preprocessing techniques were applied to enhance feature learning for deep learning segmentation models. The ensemble model, referred to as EL-4, achieved an IoU value of 0.8306, an improvement over U-Net’s 0.8065, and reduced Dice loss to 0.09019 from FPN’s 0.1183. This study could contribute to the efficiency of waste sorting at Material Recovery Facility, facilitating better raw material acquisition for recycling with minimal human intervention and enhancing the overall throughput.

Method: Proposes Weight-Decomposed Low-Rank Adaptation (WDLoRA)-based multimodal generative framework that decouples magnitude and directional weight updates, enabling foundation models to independently learn nerve topology (orientation) and stromal contrast (intensity). Jointly conditions on nerve segmentation masks and disease-specific clinical prompts to synthesize anatomically coherent images across DPN spectrum.

Result: Achieves state-of-the-art visual fidelity (FID: 5.18) and structural integrity (SSIM: 0.630), significantly outperforming GAN and standard diffusion baselines. Synthetic images preserve clinical biomarkers and are statistically equivalent to real patient data. Improves downstream diagnostic accuracy by 2.1% and segmentation performance by 2.2% when used for training.

Conclusion: The proposed framework effectively addresses data bottlenecks in medical AI by generating clinically valid synthetic images that enhance diagnostic model performance for diabetic peripheral neuropathy assessment.

Abstract: Corneal Confocal Microscopy (CCM) is a sensitive tool for assessing small-fiber damage in Diabetic Peripheral Neuropathy (DPN), yet the development of robust, automated deep learning-based diagnostic models is limited by scarce labelled data and fine-grained variability in corneal nerve morphology. Although Artificial Intelligence (AI)-driven foundation generative models excel at natural image synthesis, they often struggle in medical imaging due to limited domain-specific training, compromising the anatomical fidelity required for clinical analysis. To overcome these limitations, we propose a Weight-Decomposed Low-Rank Adaptation (WDLoRA)-based multimodal generative framework for clinically guided CCM image synthesis. WDLoRA is a parameter-efficient fine-tuning (PEFT) mechanism that decouples magnitude and directional weight updates, enabling foundation generative models to independently learn the orientation (nerve topology) and intensity (stromal contrast) required for medical realism. By jointly conditioning on nerve segmentation masks and disease-specific clinical prompts, the model synthesises anatomically coherent images across the DPN spectrum (Control, T1NoDPN, T1DPN). A comprehensive three-pillar evaluation demonstrates that the proposed framework achieves state-of-the-art visual fidelity (Fréchet Inception Distance (FID): 5.18) and structural integrity (Structural Similarity Index Measure (SSIM): 0.630), significantly outperforming GAN and standard diffusion baselines. Crucially, the synthetic images preserve gold-standard clinical biomarkers and are statistically equivalent to real patient data. When used to train automated diagnostic models, the synthetic dataset improves downstream diagnostic accuracy by 2.1% and segmentation performance by 2.2%, validating the framework’s potential to alleviate data bottlenecks in medical AI.

Method: Fine-tuned Microsoft Florence vision language model (VLM) to localize parasitic eggs within microscopic images, comparing performance against traditional object detection methods like EfficientDet.

Result: The localization VLM achieved mIOU of 0.94, performing comparatively better than other object detection methods, demonstrating strong potential for automated parasitological diagnosis.

Conclusion: The proposed VLM shows promise as a core component of an automated framework for intelligent parasitological diagnosis, offering a scalable engineering solution for regions with limited diagnostic expertise.

Abstract: Soil-transmitted helminth (STH) infections continuously affect a large proportion of the global population, particularly in tropical and sub-tropical regions, where access to specialized diagnostic expertise is limited. Although manual microscopic diagnosis of parasitic eggs remains the diagnostic gold standard, the approach can be labour-intensive, time-consuming, and prone to human error. This paper aims to utilize a vision language model (VLM) such as Microsoft Florence that was fine-tuned to localize all parasitic eggs within microscopic images. The preliminary results show that our localization VLM performs comparatively better than the other object detection methods, such as EfficientDet, with an mIOU of 0.94. This finding demonstrates the potential of the proposed VLM to serve as a core component of an automated framework, offering a scalable engineering solution for intelligent parasitological diagnosis.

Method: Hierarchical encoder-decoder architecture with two key innovations: Dual-Stream Hybrid Attention (DHA) combining shifted window attention with frequency-domain processing, and Axis-Decomposed Attention (ADA) with factorized attention mechanisms, connected via reciprocal cross-gating blocks.

Result: Superior performance on DesmokeData and LSD3K surgical datasets, effectively restoring visual clarity suitable for robotic surgery integration and enhancing surgeon-robot interface.

Conclusion: RGA-Net represents a significant step toward more reliable and safer robotic surgical systems through computational vision enhancement, with potential to reduce cognitive burden and iatrogenic injury risks.

Abstract: Robotic surgical systems rely heavily on high-quality visual feedback for precise teleoperation; yet, surgical smoke from energy-based devices significantly degrades endoscopic video feeds, compromising the human-robot interface and surgical outcomes. This paper presents RGA-Net (Reciprocal Gating and Attention-fusion Network), a novel deep learning framework specifically designed for smoke removal in robotic surgery workflows. Our approach addresses the unique challenges of surgical smoke-including dense, non-homogeneous distribution and complex light scattering-through a hierarchical encoder-decoder architecture featuring two key innovations: (1) a Dual-Stream Hybrid Attention (DHA) module that combines shifted window attention with frequency-domain processing to capture both local surgical details and global illumination changes, and (2) an Axis-Decomposed Attention (ADA) module that efficiently processes multi-scale features through factorized attention mechanisms. These components are connected via reciprocal cross-gating blocks that enable bidirectional feature modulation between encoder and decoder pathways. Extensive experiments on the DesmokeData and LSD3K surgical datasets demonstrate that RGA-Net achieves superior performance in restoring visual clarity suitable for robotic surgery integration. Our method enhances the surgeon-robot interface by providing consistently clear visualization, laying a technical foundation for alleviating surgeons’ cognitive burden, optimizing operation workflows, and reducing iatrogenic injury risks in minimally invasive procedures. These practical benefits could be further validated through future clinical trials involving surgeon usability assessments. The proposed framework represents a significant step toward more reliable and safer robotic surgical systems through computational vision enhancement.

Method: Proposes IGOFormer with two key components: 1) Intrinsic Geometry-aware Decoder that enhances object-related features by injecting geometric embeddings to capture object orientation, and 2) Momentum-based Bipartite Matching that aggregates historical matching costs using exponential moving average with query-specific smoothing factors to stabilize inter-stage matching.

Result: Achieves state-of-the-art performance for aerial oriented object detection, with AP₅₀ score of 78.00% on DOTA-V1.0 using Swin-T backbone under single-scale setting.

Conclusion: IGOFormer effectively addresses limitations in oriented object detection by integrating intrinsic geometry into feature decoding and stabilizing inter-stage matching, demonstrating superior performance especially for challenging tiny objects in aerial imagery.

Abstract: Recent query-based detectors have achieved remarkable progress, yet their performance remains constrained when handling objects with arbitrary orientations, especially for tiny objects capturing limited texture information. This limitation primarily stems from the underutilization of intrinsic geometry during pixel-based feature decoding and the occurrence of inter-stage matching inconsistency caused by stage-wise bipartite matching. To tackle these challenges, we present IGOFormer, a novel query-based oriented object detector that explicitly integrates intrinsic geometry into feature decoding and enhances inter-stage matching stability. Specifically, we design an Intrinsic Geometry-aware Decoder, which enhances the object-related features conditioned on an object query by injecting complementary geometric embeddings extrapolated from their correlations to capture the geometric layout of the object, thereby offering a critical geometric insight into its orientation. Meanwhile, a Momentum-based Bipartite Matching scheme is developed to adaptively aggregate historical matching costs by formulating an exponential moving average with query-specific smoothing factors, effectively preventing conflicting supervisory signals arising from inter-stage matching inconsistency. Extensive experiments and ablation studies demonstrate the superiority of our IGOFormer for aerial oriented object detection, achieving an AP$_{50}$ score of 78.00% on DOTA-V1.0 using Swin-T backbone under the single-scale setting. The code will be made publicly available.

Method: Develop multiple variational autoencoders (VAEs) to encode 3D brain MRI scans into compact latent representations. Systematically evaluate through: (1) quantitative/qualitative MRI reconstruction assessment, (2) latent space visualization using Principal Component Analysis, and (3) downstream classification tasks on proprietary dataset of euploid and Down syndrome brain MRI scans.

Result: VAE successfully captures essential brain features with high reconstruction fidelity. Latent space exhibits clear clustering patterns, particularly distinguishing individuals with Down syndrome from euploid controls.

Conclusion: VAE-based latent representations effectively encode meaningful information from 3D brain MRI scans, demonstrating utility for both generative reconstruction and predictive classification tasks in neuroimaging.

Abstract: Generative models have emerged as powerful tools in medical imaging, enabling tasks such as segmentation, anomaly detection, and high-quality synthetic data generation. These models typically rely on learning meaningful latent representations, which are particularly valuable given the high-dimensional nature of 3D medical images like brain magnetic resonance imaging (MRI) scans. Despite their potential, latent representations remain underexplored in terms of their structure, information content, and applicability to downstream clinical tasks. Investigating these representations is crucial for advancing the use of generative models in neuroimaging research and clinical decision-making. In this work, we develop multiple variational autoencoders (VAEs) to encode 3D brain MRI scans into compact latent space representations for generative and predictive applications. We systematically evaluate the effectiveness of the learned representations through three key analyses: (i) a quantitative and qualitative assessment of MRI reconstruction quality, (ii) a visualisation of the latent space structure using Principal Component Analysis, and (iii) downstream classification tasks on a proprietary dataset of euploid and Down syndrome individuals brain MRI scans. Our results demonstrate that the VAE successfully captures essential brain features while maintaining high reconstruction fidelity. The latent space exhibits clear clustering patterns, particularly in distinguishing individuals with Down syndrome from euploid controls.

Method: Constructs OOD prompt dataset (1,025 textual descriptions), introduces unified evaluation framework with LLM-based Evaluation, Multi-factor Motion evaluation, and Fine-grained Accuracy Evaluation, tests 14 baseline models.

Result: Models show strengths in semantic alignment, motion generalizability, and physical quality, but struggle with fine-grained accuracy in OOD scenarios.

Conclusion: Highlights limitations of existing methods in OOD scenarios and provides practical guidance for future production-level text-to-motion model design and evaluation.

Abstract: Most existing evaluations of text-to-motion generation focus on in-distribution textual inputs and a limited set of evaluation criteria, which restricts their ability to systematically assess model generalization and motion generation capabilities under complex out-of-distribution (OOD) textual conditions. To address this limitation, we propose a benchmark specifically designed for OOD text-to-motion evaluation, which includes a comprehensive analysis of 14 representative baseline models and the two datasets derived from evaluation results. Specifically, we construct an OOD prompt dataset consisting of 1,025 textual descriptions. Based on this prompt dataset, we introduce a unified evaluation framework that integrates LLM-based Evaluation, Multi-factor Motion evaluation, and Fine-grained Accuracy Evaluation. Our experimental results reveal that while different baseline models demonstrate strengths in areas such as text-to-motion semantic alignment, motion generalizability, and physical quality, most models struggle to achieve strong performance with Fine-grained Accuracy Evaluation. These findings highlight the limitations of existing methods in OOD scenarios and offer practical guidance for the design and evaluation of future production-level text-to-motion models.

Method: Created OmniScience dataset with 1.5M scientific figure-caption-context triplets. Developed dynamic model-routing re-captioning pipeline that synthesizes visual features, original captions, and in-text references using SOTA MLLMs, with quality filtering and human alignment.

Result: Pipeline boosted image-text similarity from 0.769 to 0.956. Qwen2.5-VL-3B finetuned on OmniScience achieved gains of 0.378 on MM-MT-Bench and 0.140 on MMMU benchmarks, showing substantial improvement in scientific visual understanding.

Conclusion: OmniScience addresses the scientific image understanding gap in MLLMs through high-quality dataset creation and re-captioning pipeline, enabling significant performance improvements on scientific visual reasoning tasks.

Abstract: Multimodal Large Language Models demonstrate strong performance on natural image understanding, yet exhibit limited capability in interpreting scientific images, including but not limited to schematic diagrams, experimental characterizations, and analytical charts. This limitation is particularly pronounced in open-source MLLMs. The gap largely stems from existing datasets with limited domain coverage, coarse structural annotations, and weak semantic grounding. We introduce OmniScience, a large-scale, high-fidelity multi-modal dataset comprising 1.5 million figure-caption-context triplets, spanning more than 10 major scientific disciplines. To obtain image caption data with higher information density and accuracy for multi-modal large-model training, we develop a dynamic model-routing re-captioning pipeline that leverages state-of-the-art multi-modal large language models to generate dense, self-contained descriptions by jointly synthesizing visual features, original figure captions, and corresponding in-text references authored by human scientists. The pipeline is further reinforced with rigorous quality filtering and alignment with human expert judgments, ensuring both factual accuracy and semantic completeness, and boosts the image-text multi-modal similarity score from 0.769 to 0.956. We further propose a caption QA protocol as a proxy task for evaluating visual understanding. Under this setting, Qwen2.5-VL-3B model finetuned on OmniScience show substantial gains over baselines, achieving a gain of 0.378 on MM-MT-Bench and a gain of 0.140 on MMMU.

Method: SAM4Dcap integrates SAM-Body4D for temporally consistent 4D human mesh recovery from monocular video with the OpenSim biomechanical solver. The pipeline converts reconstructed meshes into trajectory files compatible with musculoskeletal models, includes automated prompting strategies, and provides a Linux-native build for processing.

Result: Preliminary evaluations on walking and drop-jump tasks show SAM4Dcap can achieve knee kinematic predictions comparable to multi-view systems, though some discrepancies in hip flexion and residual jitter remain. The framework provides a flexible foundation for non-laboratory motion analysis.

Conclusion: SAM4Dcap bridges advanced computer vision with established biomechanical simulation to create an accessible, open-source framework for estimating biomechanical metrics from monocular video, enabling motion analysis outside traditional laboratory settings.

Abstract: Quantitative biomechanical analysis is essential for clinical diagnosis and injury prevention but is often restricted to laboratories due to the high cost of optical motion capture systems. While multi-view video approaches have lowered barriers, they remain impractical for home-based scenarios requiring monocular capture. This paper presents SAM4Dcap, an open-source, end-to-end framework for estimating biomechanical metrics from monocular video without additional training. SAM4Dcap integrates the temporally consistent 4D human mesh recovery of SAM-Body4D with the OpenSim biomechanical solver. The pipeline converts reconstructed meshes into trajectory files compatible with diverse musculoskeletal models. We introduce automated prompting strategies and a Linux-native build for processing. Preliminary evaluations on walking and drop-jump tasks indicate that SAM4Dcap has the potential to achieve knee kinematic predictions comparable to multi-view systems, although some discrepancies in hip flexion and residual jitter remain. By bridging advanced computer vision with established biomechanical simulation, SAM4Dcap provides a flexible, accessible foundation for non-laboratory motion analysis.

Method: Proposes Tracking-by-Tracking (TBT) paradigm that operates on arbitrary tracking outputs. Offline-Poly processes coarse tracking results through pre-processing, hierarchical matching/fusion, and tracklet refinement modules, leveraging offline properties for global optimization and full temporal reasoning.

Result: Achieves SOTA performance with 77.6% AMOTA on nuScenes and leading results with 83.00% HOTA on KITTI. Comprehensive experiments validate flexibility, generalizability, and modular effectiveness.

Conclusion: Offline-Poly provides a general, flexible offline 3D MOT framework that decouples from specific upstream detectors/trackers and effectively exploits offline advantages for superior tracking performance.

Abstract: Offline 3D multi-object tracking (MOT) is a critical component of the 4D auto-labeling (4DAL) process. It enhances pseudo-labels generated by high-performance detectors through the incorporation of temporal context. However, existing offline 3D MOT approaches are direct extensions of online frameworks and fail to fully exploit the advantages of offline setting. Moreover, these methods often depend on fixed upstream and customized architectures, limiting their adaptability. To address these limitations, we propose Offline-Poly, a general offline 3D MOT method based on a tracking-centric design. We introduce a standardized paradigm termed Tracking-by-Tracking (TBT), which operates exclusively on arbitrary off-the-shelf tracking outputs and produces offline-refined tracklets. This formulation decouples offline tracker from specific upstream detectors or trackers. Under the TBT paradigm, Offline-Poly accepts one or multiple coarse tracking results and processes them through a structured pipeline comprising pre-processing, hierarchical matching and fusion, and tracklet refinement. Each module is designed to capitalize on the two fundamental properties of offline tracking: resource unconstrainedness, which permits global optimization beyond real-time limits, and future observability, which enables tracklet reasoning over the full temporal horizon. Offline-Poly first eliminates short-term ghost tracklets and re-identifies fragmented segments using global scene context. It then constructs scene-level similarity to associate tracklets across multiple input sources. Finally, Offline-Poly refines tracklets by jointly leveraging local and global motion patterns. On nuScenes, we achieve SOTA performance with 77.6% AMOTA. On KITTI, it achieves leading results with 83.00% HOTA. Comprehensive experiments further validate the flexibility, generalizability, and modular effectiveness of Offline-Poly.

Method: Uses Reinforcement Learning Fine-Tuning (RLFT) with physics-based rewards derived from body mesh: imitation reward measuring motion plausibility via physical simulator, Foot-Ground Deviation reward with test-time guidance, and anti-freezing reward to preserve motion dynamics.

Result: Experiments on multiple dance datasets show significant improvement in physical plausibility of generated motions, yielding more realistic and aesthetically pleasing dances while maintaining motion dynamics.

Conclusion: The proposed RLFT approach effectively bridges the skeleton-to-mesh gap in dance generation, producing physically plausible motions suitable for mesh visualization while preserving the dynamic qualities of dance.

Abstract: Despite advances in dance generation, most methods are trained in the skeletal domain and ignore mesh-level physical constraints. As a result, motions that look plausible as joint trajectories often exhibit body self-penetration and Foot-Ground Contact (FGC) anomalies when visualized with a human body mesh, reducing the aesthetic appeal of generated dances and limiting their real-world applications. We address this skeleton-to-mesh gap by deriving physics-based rewards from the body mesh and applying Reinforcement Learning Fine-Tuning (RLFT) to steer the diffusion model toward physically plausible motion synthesis under mesh visualization. Our reward design combines (i) an imitation reward that measures a motion’s general plausibility by its imitability in a physical simulator (penalizing penetration and foot skating), and (ii) a Foot-Ground Deviation (FGD) reward with test-time FGD guidance to better capture the dynamic foot-ground interaction in dance. However, we find that the physics-based rewards tend to push the model to generate freezing motions for fewer physical anomalies and better imitability. To mitigate it, we propose an anti-freezing reward to preserve motion dynamics while maintaining physical plausibility. Experiments on multiple dance datasets consistently demonstrate that our method can significantly improve the physical plausibility of generated motions, yielding more realistic and aesthetically pleasing dances. The project page is available at: https://jjd1123.github.io/Skeleton2Stage/

Method: Proposes PerASCD using RS foundation model PerA with a modular Cascaded Gated Decoder to simplify complex SCD decoding pipelines and promote multi-level feature interaction, plus Soft Semantic Consistency Loss to mitigate training instability.

Result: Achieves state-of-the-art performance on two public benchmark datasets, effectively simplifies SCD paradigm, and shows seamless adaptation across various vision encoders.

Conclusion: PerASCD enhances multi-scale semantic understanding and overall performance in semantic change detection for remote sensing imagery through foundation model integration and simplified decoding architecture.

Abstract: Remote sensing (RS) change detection methods can extract critical information on surface dynamics and are an essential means for humans to understand changes in the earth’s surface and environment. Among these methods, semantic change detection (SCD) can more effectively interpret the multi-class information contained in bi-temporal RS imagery, providing semantic-level predictions that support dynamic change monitoring. However, due to the limited semantic understanding capability of the model and the inherent complexity of the SCD tasks, existing SCD methods face significant challenges in both performance and paradigm complexity. In this paper, we propose PerASCD, a SCD method driven by RS foundation model PerA, designed to enhance the multi-scale semantic understanding and overall performance. We introduce a modular Cascaded Gated Decoder (CG-Decoder) that simplifies complex SCD decoding pipelines while promoting effective multi-level feature interaction and fusion. In addition, we propose a Soft Semantic Consistency Loss (SSCLoss) to mitigate the numerical instability commonly encountered during SCD training. We further explore the applicability of multiple existing RS foundation models on the SCD task when equipped with the proposed decoder. Experimental results demonstrate that our decoder not only effectively simplifies the paradigm of SCD, but also achieves seamless adaptation across various vision encoders. Our method achieves state-of-the-art (SOTA) performance on two public benchmark datasets, validating its effectiveness. The code is available at https://github.com/SathShen/PerASCD.git.

Method: Uses generalized winding number (GWN) field as implicit representation and jointly optimizes point orientations, per-point area weights, and confidence coefficients in a single pipeline. Minimizes Dirichlet energy of the induced winding field with additional GWN-based constraints to compensate for non-uniform sampling, reduce noise impact, and downweight outliers without separate preprocessing.

Result: DiWR produces plausible watertight surfaces on challenging inputs including point clouds from 3D Gaussian Splatting and corrupted graphics benchmarks, outperforming both traditional multi-stage pipelines and recent joint orientation-reconstruction methods.

Conclusion: DiWR provides a robust solution for watertight surface reconstruction from imperfect point clouds, effectively handling non-uniform sampling, noise, and outliers through joint optimization in a single pipeline without preprocessing dependencies.

Abstract: We propose Dirichlet Winding Reconstruction (DiWR), a robust method for reconstructing watertight surfaces from unoriented point clouds with non-uniform sampling, noise, and outliers. Our method uses the generalized winding number (GWN) field as the target implicit representation and jointly optimizes point orientations, per-point area weights, and confidence coefficients in a single pipeline. The optimization minimizes the Dirichlet energy of the induced winding field together with additional GWN-based constraints, allowing DiWR to compensate for non-uniform sampling, reduce the impact of noise, and downweight outliers during reconstruction, with no reliance on separate preprocessing. We evaluate DiWR on point clouds from 3D Gaussian Splatting, a computer-vision pipeline, and corrupted graphics benchmarks. Experiments show that DiWR produces plausible watertight surfaces on these challenging inputs and outperforms both traditional multi-stage pipelines and recent joint orientation-reconstruction methods.

Method: Proposes multi-scale dynamics mechanism that factorizes complex motion fields, Gaussian sequences with multi-scale dynamics representation derived through compositions of multi-level motion, and incorporates multi-modal priors from vision foundation models for complementary supervision.

Result: Achieves accurate and globally consistent 4D reconstruction from monocular casual videos, with considerable improvements over existing methods in dynamic novel-view synthesis on benchmark and real-world manipulation datasets.

Conclusion: Multi-scale dynamics factorization with layered Gaussian representation and multi-modal priors effectively addresses the ambiguity in monocular 4D reconstruction, enabling physically plausible dynamic scene understanding.

Abstract: Understanding dynamic scenes from casual videos is critical for scalable robot learning, yet four-dimensional (4D) reconstruction under strictly monocular settings remains highly ill-posed. To address this challenge, our key insight is that real-world dynamics exhibits a multi-scale regularity from object to particle level. To this end, we design the multi-scale dynamics mechanism that factorizes complex motion fields. Within this formulation, we propose Gaussian sequences with multi-scale dynamics, a novel representation for dynamic 3D Gaussians derived through compositions of multi-level motion. This layered structure substantially alleviates ambiguity of reconstruction and promotes physically plausible dynamics. We further incorporate multi-modal priors from vision foundation models to establish complementary supervision, constraining the solution space and improving the reconstruction fidelity. Our approach enables accurate and globally consistent 4D reconstruction from monocular casual videos. Experiments of dynamic novel-view synthesis (NVS) on benchmark and real-world manipulation datasets demonstrate considerable improvements over existing methods.

Method: Proposes VAR-3D with: 1) View-aware 3D VQ-VAE to convert complex 3D geometric structures into discrete tokens, 2) Rendering-supervised training strategy that couples discrete token prediction with visual reconstruction to preserve visual fidelity and structural consistency relative to input text.

Result: VAR-3D significantly outperforms existing methods in both generation quality and text-3D alignment.

Conclusion: The proposed approach effectively addresses bottlenecks in learning discrete 3D representations and improves text-to-3D generation through better geometric coherence and alignment with textual descriptions.

Abstract: Recent advances in auto-regressive transformers have achieved remarkable success in generative modeling. However, text-to-3D generation remains challenging, primarily due to bottlenecks in learning discrete 3D representations. Specifically, existing approaches often suffer from information loss during encoding, causing representational distortion before the quantization process. This effect is further amplified by vector quantization, ultimately degrading the geometric coherence of text-conditioned 3D shapes. Moreover, the conventional two-stage training paradigm induces an objective mismatch between reconstruction and text-conditioned auto-regressive generation. To address these issues, we propose View-aware Auto-Regressive 3D (VAR-3D), which intergrates a view-aware 3D Vector Quantized-Variational AutoEncoder (VQ-VAE) to convert the complex geometric structure of 3D models into discrete tokens. Additionally, we introduce a rendering-supervised training strategy that couples discrete token prediction with visual reconstruction, encouraging the generative process to better preserve visual fidelity and structural consistency relative to the input text. Experiments demonstrate that VAR-3D significantly outperforms existing methods in both generation quality and text-3D alignment.

Method: Proposes Embedder-Guided Reinforcement Learning (EG-RL) framework where the Embedder supervises the Reasoner to produce evidential Traceability CoT (T-CoT) that extracts critical multimodal cues and provides retrieval-relevant inputs for the Embedder.

Result: Outperforms pioneering embedding models on MMEB-V2 and UVRB benchmarks with limited computational resources, demonstrating improved cross-modal semantic consistency and fine-grained matching capability.

Conclusion: Targeted reasoning optimization significantly improves multimodal embedding quality, providing a practical and efficient solution for reasoning-driven Universal Multimodal Embeddings development.

Abstract: Leveraging Multimodal Large Language Models (MLLMs) has become pivotal for advancing Universal Multimodal Embeddings (UME) in addressing diverse cross-modal tasks. Recent studies demonstrate that incorporating generative Chain-of-Thought (CoT) reasoning can substantially enhance task-specific representations compared to discriminative methods. However, the generated reasoning CoTs of existing generative embedding methods are limited to the textual analysis of queries and are irrelevant to the retrieval of the targets. To address these limitations, we propose a reasoning-driven UME framework that integrates Embedder-Guided Reinforcement Learning (EG-RL) to optimize the Reasoner to produce evidential Traceability CoT (T-CoT). Our key contributions are threefold: (1) We design an EG-RL framework where the Embedder provides explicit supervision to the Reasoner, ensuring the generated CoT traces are aligned with embedding tasks. (2) We introduce T-CoT, which extracts critical multimodal cues to focus on retrieval-relevant elements and provides multimodal inputs for the Embedder. (3) With limited computational resources, our framework outperforms the pioneering embedding model on both MMEB-V2 and UVRB benchmarks. The integration of multimodal evidence in structured reasoning, paired with retrieval-oriented alignment, effectively strengthens cross-modal semantic consistency and boosts the fine-grained matching capability of the model as well as the generalization across complex scenarios. Our work demonstrates that targeted reasoning optimization can significantly improve multimodal embedding quality, providing a practical and efficient solution for reasoning-driven UME development.

Method: Proposes a prior-guided hierarchical instance-pixel contrastive learning model with: 1) statistics-guided pixel-level contrastive learning to enhance distributional discrepancies between noisy and clean pixels, 2) memory bank-based instance-level contrastive learning in feature space, and 3) hybrid Transformer-CNN architecture combining Transformer encoder for global context with CNN decoder for fine-grained structure restoration.

Result: Extensive evaluations on two publicly available ultrasound datasets demonstrate that the proposed model consistently outperforms existing methods, confirming its effectiveness and superiority.

Conclusion: The proposed hierarchical contrastive learning approach effectively addresses the ultrasound denoising challenge by promoting noise-invariant and structure-aware feature representations through complementary pixel and instance-level learning strategies combined with a hybrid architecture.

Abstract: Ultrasound denoising is essential for mitigating speckle-induced degradations, thereby enhancing image quality and improving diagnostic reliability. Nevertheless, because speckle patterns inherently encode both texture and fine anatomical details, effectively suppressing noise while preserving structural fidelity remains a significant challenge. In this study, we propose a prior-guided hierarchical instance-pixel contrastive learning model for ultrasound denoising, designed to promote noise-invariant and structure-aware feature representations by maximizing the separability between noisy and clean samples at both pixel and instance levels. Specifically, a statistics-guided pixel-level contrastive learning strategy is introduced to enhance distributional discrepancies between noisy and clean pixels, thereby improving local structural consistency. Concurrently, a memory bank is employed to facilitate instance-level contrastive learning in the feature space, encouraging representations that more faithfully approximate the underlying data distribution. Furthermore, a hybrid Transformer-CNN architecture is adopted, coupling a Transformer-based encoder for global context modeling with a CNN-based decoder optimized for fine-grained anatomical structure restoration, thus enabling complementary exploitation of long-range dependencies and local texture details. Extensive evaluations on two publicly available ultrasound datasets demonstrate that the proposed model consistently outperforms existing methods, confirming its effectiveness and superiority.

Method: Uses lossy semantic video coding to transmit semantic scene structure plus compressed low-resolution frames for texture. Features modular video diffusion model with Semantic Control, Restoration Adapter, and Temporal Adapter modules. Includes efficient temporal distillation for real-time causal synthesis, reducing parameters by 300x and training time by 2x.

Result: Achieves strong perceptual quality, semantic fidelity, and temporal consistency at ultra-low bitrates (< 0.0003 bpp), outperforming classical, neural, and generative baselines in quantitative, qualitative, and subjective evaluations across diverse datasets.

Conclusion: The framework successfully enables high-fidelity, real-time video generation under extreme communication constraints through semantic coding and efficient diffusion modeling, advancing video generation for bandwidth-limited applications.

Abstract: We introduce a video diffusion model for high-fidelity, causal, and real-time video generation under ultra-low-bitrate semantic communication constraints. Our approach utilizes lossy semantic video coding to transmit the semantic scene structure, complemented by a stream of highly compressed, low-resolution frames that provide sufficient texture information to preserve fidelity. Building on these inputs, we introduce a modular video diffusion model that contains Semantic Control, Restoration Adapter, and Temporal Adapter. We further introduce an efficient temporal distillation procedure that enables extension to real-time and causal synthesis, reducing trainable parameters by 300x and training time by 2x, while adhering to communication constraints. Evaluated across diverse datasets, the framework achieves strong perceptual quality, semantic fidelity, and temporal consistency at ultra-low bitrates (< 0.0003 bpp), outperforming classical, neural, and generative baselines in extensive quantitative, qualitative, and subjective evaluations.

Method: Used 3D convolutional neural networks trained on isotropic CT volumes from preoperative TAVI patients to predict PVR occurrence from anatomical features in cardiac CT scans.

Result: The deep learning approach demonstrated potential to capture subtle anatomical features from pre-TAVI imaging, suggesting feasibility for personalized risk assessment and procedural optimization.

Conclusion: Volumetric deep learning shows promise for predicting PVR complications from preoperative CT, opening new perspectives for improving TAVI outcomes through personalized risk assessment.

Abstract: Severe aortic stenosis is a common and life-threatening condition in elderly patients, often treated with Transcatheter Aortic Valve Implantation (TAVI). Despite procedural advances, paravalvular aortic regurgitation (PVR) remains one of the most frequent post-TAVI complications, with a proven impact on long-term prognosis. In this work, we investigate the potential of deep learning to predict the occurrence of PVR from preoperative cardiac CT. To this end, a dataset of preoperative TAVI patients was collected, and 3D convolutional neural networks were trained on isotropic CT volumes. The results achieved suggest that volumetric deep learning can capture subtle anatomical features from pre-TAVI imaging, opening new perspectives for personalized risk assessment and procedural optimization. Source code is available at https://github.com/EIDOSLAB/tavi.

Method: 3D reconstruction of Da Vinci robotic arms animated in Autodesk Maya via automated Python pipeline producing photorealistic labeled videos with randomized motion, lighting variations, and synthetic blood textures while preserving pixel-accurate ground truth masks.

Result: Balanced composition of real and synthetic data significantly improves segmentation model generalization compared to real-only training, while excessive synthetic data causes domain shift; framework provides reproducible tool for surgical computer vision.

Conclusion: Synthetic data generation pipeline effectively augments real surgical data for instrument segmentation, offering scalable solution for data augmentation, domain adaptation, and simulation-based pretraining in robotic-assisted surgery.

Abstract: This paper presents a comprehensive workflow for generating and validating a synthetic dataset designed for robotic surgery instrument segmentation. A 3D reconstruction of the Da Vinci robotic arms was refined and animated in Autodesk Maya through a fully automated Python-based pipeline capable of producing photorealistic, labeled video sequences. Each scene integrates randomized motion patterns, lighting variations, and synthetic blood textures to mimic intraoperative variability while preserving pixel-accurate ground truth masks. To validate the realism and effectiveness of the generated data, several segmentation models were trained under controlled ratios of real and synthetic data. Results demonstrate that a balanced composition of real and synthetic samples significantly improves model generalization compared to training on real data only, while excessive reliance on synthetic data introduces a measurable domain shift. The proposed framework provides a reproducible and scalable tool for surgical computer vision, supporting future research in data augmentation, domain adaptation, and simulation-based pretraining for robotic-assisted surgery. Data and code are available at https://github.com/EIDOSLAB/Sintetic-dataset-DaVinci.

Method: Proposes a self-supervised learning pretraining strategy based on SimCLR (contrastive learning) using the same limited dataset available for downstream cardiac output prediction from apical four-chamber echocardiographic videos.

Result: SSL pretraining mitigates overfitting and improves representation learning, achieving an average Pearson correlation of 0.41 on the test set, outperforming PanEcho (a model trained on over one million echocardiographic exams).

Conclusion: Self-supervised learning demonstrates potential for improving medical imaging analysis even under data scarcity, showing that SSL can outperform larger supervised models on specific tasks.

Abstract: Cardiac Output (CO) is a key parameter in the diagnosis and management of cardiovascular diseases. However, its accurate measurement requires right-heart catheterization, an invasive and time-consuming procedure, motivating the development of reliable non-invasive alternatives using echocardiography. In this work, we propose a self-supervised learning (SSL) pretraining strategy based on SimCLR to improve CO prediction from apical four-chamber echocardiographic videos. The pretraining is performed using the same limited dataset available for the downstream task, demonstrating the potential of SSL even under data scarcity. Our results show that SSL mitigates overfitting and improves representation learning, achieving an average Pearson correlation of 0.41 on the test set and outperforming PanEcho, a model trained on over one million echocardiographic exams. Source code is available at https://github.com/EIDOSLAB/cardiac-output.

Method: Conducted controlled experiments removing high-frequency spatial information from discriminative models like CLIP using blurring operations. Tested both training on blurred images and test-time blurring. Directly optimized filters for alignment and computed Pareto-optimal solutions frontier for the benchmark.

Result: Test-time blurring achieves new state-of-the-art on model-vs-human benchmark, halving the alignment gap between DNNs and human observers. Low-pass filters are likely optimal, and optimal Gaussian filters match the frequency spectrum of human visual system’s band-pass filters.

Conclusion: The increased alignment of generative models can be largely explained by low-pass filtering effects rather than generative modeling itself. Human visual behavior alignment in DNNs can be dramatically improved by simple frequency filtering that matches human visual system characteristics.

Abstract: Despite their impressive performance on computer vision benchmarks, Deep Neural Networks (DNNs) still fall short of adequately modeling human visual behavior, as measured by error consistency and shape bias. Recent work hypothesized that behavioral alignment can be drastically improved through \emph{generative} – rather than \emph{discriminative} – classifiers, with far-reaching implications for models of human vision. Here, we instead show that the increased alignment of generative models can be largely explained by a seemingly innocuous resizing operation in the generative model which effectively acts as a low-pass filter. In a series of controlled experiments, we show that removing high-frequency spatial information from discriminative models like CLIP drastically increases their behavioral alignment. Simply blurring images at test-time – rather than training on blurred images – achieves a new state-of-the-art score on the model-vs-human benchmark, halving the current alignment gap between DNNs and human observers. Furthermore, low-pass filters are likely optimal, which we demonstrate by directly optimizing filters for alignment. To contextualize the performance of optimal filters, we compute the frontier of all possible pareto-optimal solutions to the benchmark, which was formerly unknown. We explain our findings by observing that the frequency spectrum of optimal Gaussian filters roughly matches the spectrum of band-pass filters implemented by the human visual system. We show that the contrast sensitivity function, describing the inverse of the contrast threshold required for humans to detect a sinusoidal grating as a function of spatiotemporal frequency, is approximated well by Gaussian filters of the specific width that also maximizes error consistency.

Method: Used a ResNet-based U-Net pre-trained on rendered images to predict surface reflectance, then fine-tuned only the decoder on VR baseline condition images. The model was evaluated using the same achromatic object selection task as human experiments across various cue manipulation conditions.

Result: The model showed strong correspondence with human behavior - both achieved high color constancy under baseline conditions and exhibited similar, condition-dependent performance declines when local surround or spatial mean color cues were removed.

Conclusion: Deep neural networks can effectively model human color constancy mechanisms in VR environments, demonstrating similar sensitivity to visual cues and suggesting potential for using DNNs to study human visual perception.

Abstract: We previously investigated color constancy in photorealistic virtual reality (VR) and developed a Deep Neural Network (DNN) that predicts reflectance from rendered images. Here, we combine both approaches to compare and study a model and human performance with respect to established color constancy mechanisms: local surround, maximum flux and spatial mean. Rather than evaluating the model against physical ground truth, model performance was assessed using the same achromatic object selection task employed in the human experiments. The model, a ResNet based U-Net from our previous work, was pre-trained on rendered images to predict surface reflectance. We then applied transfer learning, fine-tuning only the network’s decoder on images from the baseline VR condition. To parallel the human experiment, the model’s output was used to perform the same achromatic object selection task across all conditions. Results show a strong correspondence between the model and human behavior. Both achieved high constancy under baseline conditions and showed similar, condition-dependent performance declines when the local surround or spatial mean color cues were removed.

Method: Fine-tunes DINOv2 Vision Transformer using Low-Rank Adaptation (LoRA) with only 1% of parameters trained. Uses synthetic dataset generation pipeline rendering Google Fonts with diverse augmentations (random colors, alignment, line wrapping, Gaussian noise). Includes built-in preprocessing for consistency.

Result: Achieves approximately 86% top-1 accuracy on 394 font families while training fewer than 1% of the model’s 87.2M parameters. Model, dataset, and training pipeline released as open-source.

Conclusion: LoRA fine-tuning on synthetic data enables efficient and accurate font classification, demonstrating the effectiveness of parameter-efficient adaptation for vision tasks with limited real data.

Abstract: We present a font classification system capable of identifying 394 font families from rendered text images. Our approach fine-tunes a DINOv2 Vision Transformer using Low-Rank Adaptation (LoRA), achieving approximately 86% top-1 accuracy while training fewer than 1% of the model’s 87.2M parameters. We introduce a synthetic dataset generation pipeline that renders Google Fonts at scale with diverse augmentations including randomized colors, alignment, line wrapping, and Gaussian noise, producing training images that generalize to real-world typographic samples. The model incorporates built-in preprocessing to ensure consistency between training and inference, and is deployed as a HuggingFace Inference Endpoint. We release the model, dataset, and full training pipeline as open-source resources.

Method: RPGD formulates extrinsic calibration as a coarse-to-fine problem using human poses. It combines RANSAC-P3P for global robustness in initial alignment with gradient-descent-based refinement for precise optimization. The approach works with both monocular and multi-view RGB cameras using only natural human motion without requiring calibration patterns.

Result: The method achieves sub-pixel MPJPE (Mean Per Joint Position Error) reprojection error on three large-scale public 3D HPE datasets and a self-collected in-the-wild dataset. It recovers extrinsic parameters with accuracy comparable to ground truth, even in challenging, noisy settings.

Conclusion: RPGD provides a practical and automatic solution for reliable extrinsic calibration in large-scale 3D human pose estimation dataset collection, enabling more efficient and accurate alignment of MoCap data with RGB camera systems using only natural human motion.

Abstract: In this paper, we propose RPGD (RANSAC-P3P Gradient Descent), a human-pose-driven extrinsic calibration framework that robustly aligns MoCap-based 3D skeletal data with monocular or multi-view RGB cameras using only natural human motion. RPGD formulates extrinsic calibration as a coarse-to-fine problem tailored to human poses, combining the global robustness of RANSAC-P3P with Gradient-Descent-based refinement. We evaluate RPGD on three large-scale public 3D HPE datasets as well as on a self-collected in-the-wild dataset. Experimental results demonstrate that RPGD consistently recovers extrinsic parameters with accuracy comparable to the provided ground truth, achieving sub-pixel MPJPE reprojection error even in challenging, noisy settings. These results indicate that RPGD provides a practical and automatic solution for reliable extrinsic calibration of large-scale 3D HPE dataset collection.

Method: MamaDino fuses frozen self-supervised DINOv3 ViT-S features with a trainable CNN encoder at 512x512 resolution, and aggregates bilateral breast information via a BilateralMixer to output 3-year breast cancer risk scores. The model is trained on 53,883 women from OPTIMAM (UK) and evaluated on internal and external test sets.

Result: MamaDino matches Mirai’s performance on both internal and external tests while using ~13x fewer input pixels. Adding the BilateralMixer improves discrimination to AUC 0.736 (vs 0.713) in-distribution and 0.677 (vs 0.666) out-of-distribution, with consistent performance across various demographic and clinical factors.

Conclusion: Explicit contralateral modeling and complementary inductive biases enable predictions that match state-of-the-art methods despite operating on substantially lower-resolution mammograms, demonstrating that using less detailed images in a more structured way can recover state-of-the-art accuracy.

Abstract: Breast cancer screening programmes increasingly seek to move from one-size-fits-all interval to risk-adapted and personalized strategies. Deep learning (DL) has enabled image-based risk models with stronger 1- to 5-year prediction than traditional clinical models, but leading systems (e.g., Mirai) typically use convolutional backbones, very high-resolution inputs (>1M pixels) and simple multi-view fusion, with limited explicit modelling of contralateral asymmetry. We hypothesised that combining complementary inductive biases (convolutional and transformer-based) with explicit contralateral asymmetry modelling would allow us to match state-of-the-art 3-year risk prediction performance even when operating on substantially lower-resolution mammograms, indicating that using less detailed images in a more structured way can recover state-of-the-art accuracy. We present MamaDino, a mammography-aware multi-view attentional DINO model. MamaDino fuses frozen self-supervised DINOv3 ViT-S features with a trainable CNN encoder at 512x512 resolution, and aggregates bilateral breast information via a BilateralMixer to output a 3-year breast cancer risk score. We train on 53,883 women from OPTIMAM (UK) and evaluate on matched 3-year case-control cohorts: an in-distribution test set from four screening sites and an external out-of-distribution cohort from an unseen site. At breast-level, MamaDino matches Mirai on both internal and external tests while using ~13x fewer input pixels. Adding the BilateralMixer improves discrimination to AUC 0.736 (vs 0.713) in-distribution and 0.677 (vs 0.666) out-of-distribution, with consistent performance across age, ethnicity, scanner, tumour type and grade. These findings demonstrate that explicit contralateral modelling and complementary inductive biases enable predictions that match Mirai, despite operating on substantially lower-resolution mammograms.

Method: Proposes STAMP framework that integrates spatially-resolved gene expression profiles with pathology images using self-supervised, gene-guided training. Uses hierarchical multi-scale contrastive alignment and cross-scale patch localization to align spatial transcriptomics with pathology images. Built SpaVis-6M, the largest Visium-based spatial transcriptomics dataset, and trained a spatially-aware gene encoder.

Result: Validated across six datasets and four downstream tasks, consistently achieving strong performance. Demonstrates that gene-guided training provides robust, task-agnostic signals for learning pathology image representations.

Conclusion: Integrating spatially resolved molecular supervision is valuable and necessary for advancing multimodal learning in computational pathology, overcoming limitations of language-only supervision.

Abstract: Recent years have witnessed remarkable progress in multimodal learning within computational pathology. Existing models primarily rely on vision and language modalities; however, language alone lacks molecular specificity and offers limited pathological supervision, leading to representational bottlenecks. In this paper, we propose STAMP, a Spatial Transcriptomics-Augmented Multimodal Pathology representation learning framework that integrates spatially-resolved gene expression profiles to enable molecule-guided joint embedding of pathology images and transcriptomic data. Our study shows that self-supervised, gene-guided training provides a robust and task-agnostic signal for learning pathology image representations. Incorporating spatial context and multi-scale information further enhances model performance and generalizability. To support this, we constructed SpaVis-6M, the largest Visium-based spatial transcriptomics dataset to date, and trained a spatially-aware gene encoder on this resource. Leveraging hierarchical multi-scale contrastive alignment and cross-scale patch localization mechanisms, STAMP effectively aligns spatial transcriptomics with pathology images, capturing spatial structure and molecular variation. We validate STAMP across six datasets and four downstream tasks, where it consistently achieves strong performance. These results highlight the value and necessity of integrating spatially resolved molecular supervision for advancing multimodal learning in computational pathology. The code is included in the supplementary materials. The pretrained weights and SpaVis-6M are available at: https://github.com/Hanminghao/STAMP.

Method: Introduces MarsRetrieval benchmark with three tasks covering multiple spatial scales and geomorphic origins. Proposes unified retrieval-centric protocol to evaluate multimodal embedding architectures including contrastive dual-tower encoders and generative vision-language models.

Result: MarsRetrieval is challenging - even strong foundation models often fail to capture domain-specific geomorphic distinctions. Domain-specific fine-tuning is critical for generalizable geospatial discovery in planetary settings.

Conclusion: MarsRetrieval provides a comprehensive benchmark for evaluating vision-language models in planetary science, highlighting the importance of domain adaptation for effective geospatial discovery on Mars.

Abstract: Data-driven approaches like deep learning are rapidly advancing planetary science, particularly in Mars exploration. Despite recent progress, most existing benchmarks remain confined to closed-set supervised visual tasks and do not support text-guided retrieval for geospatial discovery. We introduce MarsRetrieval, a retrieval benchmark for evaluating vision-language models for Martian geospatial discovery. MarsRetrieval includes three tasks: (1) paired image-text retrieval, (2) landform retrieval, and (3) global geo-localization, covering multiple spatial scales and diverse geomorphic origins. We propose a unified retrieval-centric protocol to benchmark multimodal embedding architectures, including contrastive dual-tower encoders and generative vision-language models. Our evaluation shows MarsRetrieval is challenging: even strong foundation models often fail to capture domain-specific geomorphic distinctions. We further show that domain-specific fine-tuning is critical for generalizable geospatial discovery in planetary settings. Our code is available at https://github.com/ml-stat-Sustech/MarsRetrieval

Method: E-DiT adds lightweight routers to each DiT block that dynamically identify sample-dependent sparsity from input latents. Each router decides whether to skip its block and, if not skipped, predicts optimal MLP width reduction ratio. A block-level feature caching mechanism eliminates redundant computations during inference.

Result: Extensive experiments on 2D image (Qwen-Image, FLUX) and 3D asset (Hunyuan3D-3.0) generation show E-DiT achieves up to ~2× speedup with negligible loss in generation quality.

Conclusion: E-DiT provides an effective adaptive acceleration framework for DiT that maintains generation quality while significantly improving efficiency through dynamic sparsity-aware computation.

Abstract: Diffusion Transformers (DiT) have demonstrated remarkable generative capabilities but remain highly computationally expensive. Previous acceleration methods, such as pruning and distillation, typically rely on a fixed computational capacity, leading to insufficient acceleration and degraded generation quality. To address this limitation, we propose \textbf{Elastic Diffusion Transformer (E-DiT)}, an adaptive acceleration framework for DiT that effectively improves efficiency while maintaining generation quality. Specifically, we observe that the generative process of DiT exhibits substantial sparsity (i.e., some computations can be skipped with minimal impact on quality), and this sparsity varies significantly across samples. Motivated by this observation, E-DiT equips each DiT block with a lightweight router that dynamically identifies sample-dependent sparsity from the input latent. Each router adaptively determines whether the corresponding block can be skipped. If the block is not skipped, the router then predicts the optimal MLP width reduction ratio within the block. During inference, we further introduce a block-level feature caching mechanism that leverages router predictions to eliminate redundant computations in a training-free manner. Extensive experiments across 2D image (Qwen-Image and FLUX) and 3D asset (Hunyuan3D-3.0) demonstrate the effectiveness of E-DiT, achieving up to $\sim$2$\times$ speedup with negligible loss in generation quality. Code will be available at https://github.com/wangjiangshan0725/Elastic-DiT.

Method: Proposes SpatialID, a training-free spatially-adaptive identity modulation framework that decouples identity injection into face-relevant and context-free regions using a Spatial Mask Extractor derived from cross-attention responses. Also introduces Temporal-Spatial Scheduling that dynamically adjusts spatial constraints (transitioning from Gaussian priors to attention-based masks with adaptive relaxation) to align with diffusion generation dynamics.

Result: Extensive experiments on IBench show SpatialID achieves state-of-the-art performance in text adherence (CLIP-T: 0.281), visual consistency (CLIP-I: 0.827), and image quality (IQ: 0.523), significantly eliminating background contamination while maintaining robust identity preservation.

Conclusion: SpatialID provides an effective training-free solution for personalized text-to-image generation that spatially adapts identity injection to prevent contamination of non-facial regions while preserving identity and text adherence.

Abstract: Personalized text-to-image generation aims to integrate specific identities into arbitrary contexts. However, existing tuning-free methods typically employ Spatially Uniform Visual Injection, causing identity features to contaminate non-facial regions (e.g., backgrounds and lighting) and degrading text adherence. To address this without expensive fine-tuning, we propose SpatialID, a training-free spatially-adaptive identity modulation framework. SpatialID fundamentally decouples identity injection into face-relevant and context-free regions using a Spatial Mask Extractor derived from cross-attention responses. Furthermore, we introduce a Temporal-Spatial Scheduling strategy that dynamically adjusts spatial constraints - transitioning from Gaussian priors to attention-based masks and adaptive relaxation - to align with the diffusion generation dynamics. Extensive experiments on IBench demonstrate that SpatialID achieves state-of-the-art performance in text adherence (CLIP-T: 0.281), visual consistency (CLIP-I: 0.827), and image quality (IQ: 0.523), significantly eliminating background contamination while maintaining robust identity preservation.

Method: LitePath integrates LiteFM (a compact model distilled from three large PFMs using 190M patches) and Adaptive Patch Selector (APS) for task-specific patch selection, reducing model over-parameterization and patch redundancy.

Result: Achieves 28x parameter reduction, 403.5x FLOP reduction, processes 208 slides/hour on NVIDIA Jetson Orin Nano (104.5x faster than Virchow2), consumes 0.36 kWh/3,000 slides (171x lower), and ranks 2nd among 19 models while maintaining 99.71% of Virchow2’s AUC.

Conclusion: LitePath enables rapid, cost-effective, energy-efficient pathology image analysis on accessible hardware while maintaining accuracy comparable to state-of-the-art PFMs and reducing AI deployment carbon footprint.

Abstract: Pathology foundation models (PFMs) have enabled robust generalization in computational pathology through large-scale datasets and expansive architectures, but their substantial computational cost, particularly for gigapixel whole slide images, limits clinical accessibility and scalability. Here, we present LitePath, a deployment-friendly foundational framework designed to mitigate model over-parameterization and patch level redundancy. LitePath integrates LiteFM, a compact model distilled from three large PFMs (Virchow2, H-Optimus-1 and UNI2) using 190 million patches, and the Adaptive Patch Selector (APS), a lightweight component for task-specific patch selection. The framework reduces model parameters by 28x and lowers FLOPs by 403.5x relative to Virchow2, enabling deployment on low-power edge hardware such as the NVIDIA Jetson Orin Nano Super. On this device, LitePath processes 208 slides per hour, 104.5x faster than Virchow2, and consumes 0.36 kWh per 3,000 slides, 171x lower than Virchow2 on an RTX3090 GPU. We validated accuracy using 37 cohorts across four organs and 26 tasks (26 internal, 9 external, and 2 prospective), comprising 15,672 slides from 9,808 patients disjoint from the pretraining data. LitePath ranks second among 19 evaluated models and outperforms larger models including H-Optimus-1, mSTAR, UNI2 and GPFM, while retaining 99.71% of the AUC of Virchow2 on average. To quantify the balance between accuracy and efficiency, we propose the Deployability Score (D-Score), defined as the weighted geometric mean of normalized AUC and normalized FLOP, where LitePath achieves the highest value, surpassing Virchow2 by 10.64%. These results demonstrate that LitePath enables rapid, cost-effective and energy-efficient pathology image analysis on accessible hardware while maintaining accuracy comparable to state-of-the-art PFMs and reducing the carbon footprint of AI deployment.

Method: Flow4R treats camera-space scene flow as central representation, predicting per-pixel properties (3D point position, scene flow, pose weight, confidence) from two-view inputs using Vision Transformer. Uses flow-centric formulation allowing local geometry and bidirectional motion inference with shared decoder in single forward pass, without explicit pose regressors or bundle adjustment.

Result: Achieves state-of-the-art performance on 4D reconstruction and tracking tasks. Demonstrates effectiveness of flow-central representation for spatiotemporal scene understanding.

Conclusion: Flow4R provides a unified framework for dynamic 3D scene understanding through flow-centric representation, successfully linking geometry and motion without explicit pose estimation.

Abstract: Reconstructing and tracking dynamic 3D scenes remains a fundamental challenge in computer vision. Existing approaches often decouple geometry from motion: multi-view reconstruction methods assume static scenes, while dynamic tracking frameworks rely on explicit camera pose estimation or separate motion models. We propose Flow4R, a unified framework that treats camera-space scene flow as the central representation linking 3D structure, object motion, and camera motion. Flow4R predicts a minimal per-pixel property set-3D point position, scene flow, pose weight, and confidence-from two-view inputs using a Vision Transformer. This flow-centric formulation allows local geometry and bidirectional motion to be inferred symmetrically with a shared decoder in a single forward pass, without requiring explicit pose regressors or bundle adjustment. Trained jointly on static and dynamic datasets, Flow4R achieves state-of-the-art performance on 4D reconstruction and tracking tasks, demonstrating the effectiveness of the flow-central representation for spatiotemporal scene understanding.

Method: Proposes FLEX with three components: 1) Frequency-aware RoPE Modulation to adaptively interpolate under-trained low-frequency components while extrapolating high-frequency ones, 2) Antiphase Noise Sampling (ANS) to inject high-frequency dynamic priors, and 3) Inference-only Attention Sink to anchor global structure.

Result: FLEX significantly outperforms SOTA models at 6× extrapolation (30s duration) and matches long-video fine-tuned baselines at 12× scale (60s duration). It enables consistent video synthesis at 4-minute scale with existing models like LongLive.

Conclusion: FLEX is an effective plug-and-play inference-time framework that bridges the gap between short-term training and long-term inference for video diffusion models, enabling scalable long video generation without retraining.

Abstract: Autoregressive video diffusion models have emerged as a scalable paradigm for long video generation. However, they often suffer from severe extrapolation failure, where rapid error accumulation leads to significant temporal degradation when extending beyond training horizons. We identify that this failure primarily stems from the \textit{spectral bias} of 3D positional embeddings and the lack of \textit{dynamic priors} in noise sampling. To address these issues, we propose \textbf{FLEX} (\textbf{F}requency-aware \textbf{L}ength \textbf{EX}tension), a training-free inference-time framework that bridges the gap between short-term training and long-term inference. FLEX introduces Frequency-aware RoPE Modulation to adaptively interpolate under-trained low-frequency components while extrapolating high-frequency ones to preserve multi-scale temporal discriminability. This is integrated with Antiphase Noise Sampling (ANS) to inject high-frequency dynamic priors and Inference-only Attention Sink to anchor global structure. Extensive evaluations on VBench demonstrate that FLEX significantly outperforms state-of-the-art models at $6\times$ extrapolation (30s duration) and matches the performance of long-video fine-tuned baselines at $12\times$ scale (60s duration). As a plug-and-play augmentation, FLEX seamlessly integrates into existing inference pipelines for horizon extension. It effectively pushes the generation limits of models such as LongLive, supporting consistent and dynamic video synthesis at a 4-minute scale. Project page is available at \href{https://ga-lee.github.io/FLEX_demo}{https://ga-lee.github.io/FLEX}.

Method: Proposes a SHAP-inspired gradient-activation attribution method to estimate layer importance as a data-driven proxy for functional contribution, enabling layer-wise pruning tailored for object detection architectures.

Result: The method consistently identifies different least important layers than L1-norm pruning, achieving better accuracy-efficiency trade-offs: 10% speed increase for ShuffleNetV2 vs 13.7% degradation with L1, and preserved mAP for RetinaNet with negligible speed impact.

Conclusion: Explainability-inspired compression offers a principled direction for deploying DNNs on edge platforms while preserving both performance and interpretability, highlighting the importance of data-driven layer importance assessment.

Abstract: Deep neural networks (DNNs) have achieved remarkable success in object detection tasks, but their increasing complexity poses significant challenges for deployment on resource-constrained platforms. While model compression techniques such as pruning have emerged as essential tools, traditional magnitude-based pruning methods do not necessarily align with the true functional contribution of network components to task-specific performance. In this work, we present an explainability-inspired, layer-wise pruning framework tailored for efficient object detection. Our approach leverages a SHAP-inspired gradient–activation attribution to estimate layer importance, providing a data-driven proxy for functional contribution rather than relying solely on static weight magnitudes. We conduct comprehensive experiments across diverse object detection architectures, including ResNet-50, MobileNetV2, ShuffleNetV2, Faster R-CNN, RetinaNet, and YOLOv8, evaluating performance on the Microsoft COCO 2017 validation set. The results show that the proposed attribution-inspired pruning consistently identifies different layers as least important compared to L1-norm-based methods, leading to improved accuracy–efficiency trade-offs. Notably, for ShuffleNetV2, our method yields a 10% empirical increase in inference speed, whereas L1-pruning degrades performance by 13.7%. For RetinaNet, the proposed approach preserves the baseline mAP (0.151) with negligible impact on inference speed, while L1-pruning incurs a 1.3% mAP drop for a 6.2% speed increase. These findings highlight the importance of data-driven layer importance assessment and demonstrate that explainability-inspired compression offers a principled direction for deploying deep neural networks on edge and resource-constrained platforms while preserving both performance and interpretability.

Method: Uses binary visual tokens (up to 2^256 states per token) with binary diffusion heads instead of softmax classification. Introduces next-patch diffusion for parallel token prediction, enabling faster inference while maintaining accuracy.

Result: Achieves FID of 1.24 on ImageNet 256x256 (best among AR models). With 260M parameters, beats 1.4B parameter parallel AR models with 8.7x speedup. For text-to-image, generates high-resolution photorealistic images with over 30x speedup for 1024x1024 images.

Conclusion: BitDance demonstrates that binary tokens with diffusion heads enable highly efficient and scalable autoregressive image generation with state-of-the-art quality and significant speed improvements.

Abstract: We present BitDance, a scalable autoregressive (AR) image generator that predicts binary visual tokens instead of codebook indices. With high-entropy binary latents, BitDance lets each token represent up to $2^{256}$ states, yielding a compact yet highly expressive discrete representation. Sampling from such a huge token space is difficult with standard classification. To resolve this, BitDance uses a binary diffusion head: instead of predicting an index with softmax, it employs continuous-space diffusion to generate the binary tokens. Furthermore, we propose next-patch diffusion, a new decoding method that predicts multiple tokens in parallel with high accuracy, greatly speeding up inference. On ImageNet 256x256, BitDance achieves an FID of 1.24, the best among AR models. With next-patch diffusion, BitDance beats state-of-the-art parallel AR models that use 1.4B parameters, while using 5.4x fewer parameters (260M) and achieving 8.7x speedup. For text-to-image generation, BitDance trains on large-scale multimodal tokens and generates high-resolution, photorealistic images efficiently, showing strong performance and favorable scaling. When generating 1024x1024 images, BitDance achieves a speedup of over 30x compared to prior AR models. We release code and models to facilitate further research on AR foundation models. Code and models are available at: https://github.com/shallowdream204/BitDance.

Method: Proposes Restoration Adaptation for Semantic Segmentation (RASS) with two components: 1) Semantic-Constrained Restoration (SCR) that aligns restoration cross-attention maps with segmentation masks to inject semantic priors, and 2) LoRA-based module merging and fine-tuning to transfer restoration knowledge to segmentation models.

Result: Significantly outperforms state-of-the-art methods on both synthetic and real-world low-quality benchmarks. Constructed a real-world LQ image segmentation dataset with high-quality annotations for validation.

Conclusion: RASS effectively integrates semantic restoration into segmentation, enabling robust high-quality segmentation on low-quality images by bridging the gap between restoration fidelity and semantic understanding.

Abstract: In real-world scenarios, the performance of semantic segmentation often deteriorates when processing low-quality (LQ) images, which may lack clear semantic structures and high-frequency details. Although image restoration techniques offer a promising direction for enhancing degraded visual content, conventional real-world image restoration (Real-IR) models primarily focus on pixel-level fidelity and often fail to recover task-relevant semantic cues, limiting their effectiveness when directly applied to downstream vision tasks. Conversely, existing segmentation models trained on high-quality data lack robustness under real-world degradations. In this paper, we propose Restoration Adaptation for Semantic Segmentation (RASS), which effectively integrates semantic image restoration into the segmentation process, enabling high-quality semantic segmentation on the LQ images directly. Specifically, we first propose a Semantic-Constrained Restoration (SCR) model, which injects segmentation priors into the restoration model by aligning its cross-attention maps with segmentation masks, encouraging semantically faithful image reconstruction. Then, RASS transfers semantic restoration knowledge into segmentation through LoRA-based module merging and task-specific fine-tuning, thereby enhancing the model’s robustness to LQ images. To validate the effectiveness of our framework, we construct a real-world LQ image segmentation dataset with high-quality annotations, and conduct extensive experiments on both synthetic and real-world LQ benchmarks. The results show that SCR and RASS significantly outperform state-of-the-art methods in segmentation and restoration tasks. Code, models, and datasets will be available at https://github.com/Ka1Guan/RASS.git.

Method: 1) Curated 40K high-quality training samples with refined instructions and masks; 2) Introduced pixel-level similarity reward to complement MLLM-based rewards; 3) Proposed region-based regularizer to preserve non-edited regions for high-reward samples while encouraging editing for low-reward samples; 4) Applied framework to Qwen-Image-Edit and FLUX-Kontext models.

Result: Achieved competitive editing scores with state-of-the-art models while significantly improving content consistency, measured by PSNR/SSIM metrics and human subjective ratings. Annotated editing masks for GEdit-Bench and ImgEdit-Bench with pixel-level similarity metrics.

Conclusion: CoCoEdit effectively addresses the content consistency problem in image editing through region-regularized reinforcement learning, achieving both high editing quality and preservation of non-edited regions.

Abstract: Image editing has achieved impressive results with the development of large-scale generative models. However, existing models mainly focus on the editing effects of intended objects and regions, often leading to unwanted changes in unintended regions. We present a post-training framework for Content-Consistent Editing (CoCoEdit) via region regularized reinforcement learning. We first augment existing editing datasets with refined instructions and masks, from which 40K diverse and high quality samples are curated as training set. We then introduce a pixel-level similarity reward to complement MLLM-based rewards, enabling models to ensure both editing quality and content consistency during the editing process. To overcome the spatial-agnostic nature of the rewards, we propose a region-based regularizer, aiming to preserve non-edited regions for high-reward samples while encouraging editing effects for low-reward samples. For evaluation, we annotate editing masks for GEdit-Bench and ImgEdit-Bench, introducing pixel-level similarity metrics to measure content consistency and editing quality. Applying CoCoEdit to Qwen-Image-Edit and FLUX-Kontext, we achieve not only competitive editing scores with state-of-the-art models, but also significantly better content consistency, measured by PSNR/SSIM metrics and human subjective ratings.

Method: Proposes ForgeryVCR with: 1) Forensic toolbox to materialize imperceptible traces into explicit visual intermediates via Visual-Centric Reasoning, 2) Strategic Tool Learning post-training with gain-driven trajectory construction for SFT and RL optimization guided by tool utility reward, enabling MLLM to act as proactive decision-maker invoking multi-view reasoning paths.

Result: Achieves state-of-the-art performance in both detection and localization tasks, demonstrating superior generalization and robustness with minimal tool redundancy.

Conclusion: Visual-centric reasoning with forensic tools overcomes limitations of text-centric approaches for image forgery detection, enabling more accurate capture of pixel-level inconsistencies through strategic tool learning.

Abstract: Existing Multimodal Large Language Models (MLLMs) for image forgery detection and localization predominantly operate under a text-centric Chain-of-Thought (CoT) paradigm. However, forcing these models to textually characterize imperceptible low-level tampering traces inevitably leads to hallucinations, as linguistic modalities are insufficient to capture such fine-grained pixel-level inconsistencies. To overcome this, we propose ForgeryVCR, a framework that incorporates a forensic toolbox to materialize imperceptible traces into explicit visual intermediates via Visual-Centric Reasoning. To enable efficient tool utilization, we introduce a Strategic Tool Learning post-training paradigm, encompassing gain-driven trajectory construction for Supervised Fine-Tuning (SFT) and subsequent Reinforcement Learning (RL) optimization guided by a tool utility reward. This paradigm empowers the MLLM to act as a proactive decision-maker, learning to spontaneously invoke multi-view reasoning paths including local zoom-in for fine-grained inspection and the analysis of invisible inconsistencies in compression history, noise residuals, and frequency domains. Extensive experiments reveal that ForgeryVCR achieves state-of-the-art (SOTA) performance in both detection and localization tasks, demonstrating superior generalization and robustness with minimal tool redundancy. The project page is available at https://youqiwong.github.io/projects/ForgeryVCR/.

Method: Introduces a geometry-aware self-correction framework that leverages the model’s own normal and depth predictions to refine structural accuracy. Uses a dedicated transformer and fusion module to feed back geometric cues, enabling error correction and consistency enforcement with the conditioning image, without additional supervision or external signals.

Result: Extensive experiments show GeoFusionLRM achieves sharper geometry, more consistent normals, and higher fidelity than state-of-the-art LRM baselines.

Conclusion: GeoFusionLRM effectively improves 3D reconstruction quality by using self-generated geometric cues for refinement, addressing limitations of current LRMs without requiring additional supervision.

Abstract: Single-image 3D reconstruction with large reconstruction models (LRMs) has advanced rapidly, yet reconstructions often exhibit geometric inconsistencies and misaligned details that limit fidelity. We introduce GeoFusionLRM, a geometry-aware self-correction framework that leverages the model’s own normal and depth predictions to refine structural accuracy. Unlike prior approaches that rely solely on features extracted from the input image, GeoFusionLRM feeds back geometric cues through a dedicated transformer and fusion module, enabling the model to correct errors and enforce consistency with the conditioning image. This design improves the alignment between the reconstructed mesh and the input views without additional supervision or external signals. Extensive experiments demonstrate that GeoFusionLRM achieves sharper geometry, more consistent normals, and higher fidelity than state-of-the-art LRM baselines.

Method: Created EgoSound benchmark by unifying data from Ego4D and EgoBlind datasets, defining a seven-task taxonomy, and constructing 7315 validated QA pairs across 900 videos through a multi-stage auto-generative pipeline.

Result: Comprehensive evaluation of 9 state-of-the-art MLLMs shows emerging auditory reasoning abilities but limitations in fine-grained spatial and causal understanding, establishing EgoSound as a challenging foundation for multisensory egocentric intelligence.

Conclusion: EgoSound bridges the gap between seeing and truly hearing the world, providing a systematic benchmark to advance multisensory egocentric intelligence in MLLMs beyond current vision-language capabilities.

Abstract: Multimodal Large Language Models (MLLMs) have recently achieved remarkable progress in vision-language understanding. Yet, human perception is inherently multisensory, integrating sight, sound, and motion to reason about the world. Among these modalities, sound provides indispensable cues about spatial layout, off-screen events, and causal interactions, particularly in egocentric settings where auditory and visual signals are tightly coupled. To this end, we introduce EgoSound, the first benchmark designed to systematically evaluate egocentric sound understanding in MLLMs. EgoSound unifies data from Ego4D and EgoBlind, encompassing both sighted and sound-dependent experiences. It defines a seven-task taxonomy spanning intrinsic sound perception, spatial localization, causal inference, and cross-modal reasoning. Constructed through a multi-stage auto-generative pipeline, EgoSound contains 7315 validated QA pairs across 900 videos. Comprehensive experiments on nine state-of-the-art MLLMs reveal that current models exhibit emerging auditory reasoning abilities but remain limited in fine-grained spatial and causal understanding. EgoSound establishes a challenging foundation for advancing multisensory egocentric intelligence, bridging the gap between seeing and truly hearing the world.

Method: Proposes DenseMLLM with a novel vision token supervision strategy for multiple labels and tasks, enabling standard MLLM architecture to handle dense predictions without additional decoders or architectural specialization.

Result: Achieves highly competitive performance across a wide range of dense prediction and vision-language benchmarks despite minimalist design.

Conclusion: Standard, general-purpose MLLMs can effectively support dense perception tasks without architectural specialization, challenging the paradigm that requires task-specific decoders for fine-grained predictions.

Abstract: Multimodal Large Language Models (MLLMs) have demonstrated exceptional capabilities in high-level visual understanding. However, extending these models to fine-grained dense prediction tasks, such as semantic segmentation and depth estimation, typically necessitates the incorporation of complex, task-specific decoders and other customizations. This architectural fragmentation increases model complexity and deviates from the generalist design of MLLMs, ultimately limiting their practicality. In this work, we challenge this paradigm by accommodating standard MLLMs to perform dense predictions without requiring additional task-specific decoders. The proposed model is called DenseMLLM, grounded in the standard architecture with a novel vision token supervision strategy for multiple labels and tasks. Despite its minimalist design, our model achieves highly competitive performance across a wide range of dense prediction and vision-language benchmarks, demonstrating that a standard, general-purpose MLLM can effectively support dense perception without architectural specialization.

Method: Collected 319 images with 6524 annotated chestnuts. Evaluated 29 state-of-the-art real-time object detectors (14 YOLO v11-13 variants, 15 RT-DETR v1-v4 variants) through replicated modeling experiments for chestnut detection.

Result: YOLOv12m achieved best mAP@0.5 of 95.1%, RT-DETRv2-R101 was best among RT-DETR models with 91.1% mAP@0.5. YOLOv11x achieved best mAP@[0.5:0.95] of 80.1%. YOLO models outperformed RT-DETR in both accuracy and inference speed, better suited for on-board deployment.

Conclusion: All models show significant potential for real-time chestnut detection. YOLO models are superior for on-board deployment. Dataset and software made publicly available for further research.

Abstract: Traditional mechanized chestnut harvesting is too costly for small producers, non-selective, and prone to damaging nuts. Accurate, reliable detection of chestnuts on the orchard floor is crucial for developing low-cost, vision-guided automated harvesting technology. However, developing a reliable chestnut detection system faces challenges in complex environments with shading, varying natural light conditions, and interference from weeds, fallen leaves, stones, and other foreign on-ground objects, which have remained unaddressed. This study collected 319 images of chestnuts on the orchard floor, containing 6524 annotated chestnuts. A comprehensive set of 29 state-of-the-art real-time object detectors, including 14 in the YOLO (v11-13) and 15 in the RT-DETR (v1-v4) families at varied model scales, was systematically evaluated through replicated modeling experiments for chestnut detection. Experimental results show that the YOLOv12m model achieves the best mAP@0.5 of 95.1% among all the evaluated models, while the RT-DETRv2-R101 was the most accurate variant among RT-DETR models, with mAP@0.5 of 91.1%. In terms of mAP@[0.5:0.95], the YOLOv11x model achieved the best accuracy of 80.1%. All models demonstrate significant potential for real-time chestnut detection, and YOLO models outperformed RT-DETR models in terms of both detection accuracy and inference, making them better suited for on-board deployment. Both the dataset and software programs in this study have been made publicly available at https://github.com/AgFood-Sensing-and-Intelligence-Lab/ChestnutDetection.

Method: Proposes LaViDa-R1 with a unified post-training framework integrating supervised finetuning (SFT) and multi-task reinforcement learning (RL). Uses novel techniques including answer-forcing, tree search, and complementary likelihood estimation for enhanced effectiveness and scalability.

Result: Extensive experiments demonstrate strong performance on a wide range of multimodal tasks including visual math reasoning, reason-intensive grounding, and image editing.

Conclusion: LaViDa-R1 represents an effective multimodal reasoning diffusion language model that successfully unifies diverse understanding and generation capabilities through its novel training framework.

Abstract: Diffusion language models (dLLMs) recently emerged as a promising alternative to auto-regressive LLMs. The latest works further extended it to multimodal understanding and generation tasks. In this work, we propose LaViDa-R1, a multimodal, general-purpose reasoning dLLM. Unlike existing works that build reasoning dLLMs through task-specific reinforcement learning, LaViDa-R1 incorporates diverse multimodal understanding and generation tasks in a unified manner. In particular, LaViDa-R1 is built with a novel unified post-training framework that seamlessly integrates supervised finetuning (SFT) and multi-task reinforcement learning (RL). It employs several novel training techniques, including answer-forcing, tree search, and complementary likelihood estimation, to enhance effectiveness and scalability. Extensive experiments demonstrate LaViDa-R1’s strong performance on a wide range of multimodal tasks, including visual math reasoning, reason-intensive grounding, and image editing.

Method: Implemented on optical see-through head-mounted display (OST-HMD) without external sensors or markers. Reconstructs operative scene from RGB, depth, and pose data, extracts patient’s body surface using foundation model, performs surface-based markerless registration to align preoperative anatomical models, enabling in-situ visualization of planned trocar layouts.

Result: In full-scale human-phantom experiments, ARport accurately overlaid pre-planned trocar sites onto physical phantom, achieving consistent spatial correspondence between virtual plans and real anatomy.

Conclusion: ARport provides a fully marker-free and hardware-minimal solution for visualizing preoperative trocar plans directly on patient’s body surface, facilitating efficient intraoperative setup with potential for seamless integration into clinical workflows.

Abstract: Purpose: Precise port placement is a critical step in robot-assisted surgery, where port configuration influences both visual access to the operative field and instrument maneuverability. To bridge the gap between preoperative planning and intraoperative execution, we present ARport, an augmented reality (AR) system that automatically maps pre-planned trocar layouts onto the patient’s body surface, providing intuitive spatial guidance during surgical preparation. Methods: ARport, implemented on an optical see-through head-mounted display (OST-HMD), operates without any external sensors or markers, simplifying setup and enhancing workflow integration. It reconstructs the operative scene from RGB, depth, and pose data captured by the OST-HMD, extracts the patient’s body surface using a foundation model, and performs surface-based markerless registration to align preoperative anatomical models to the extracted patient’s body surface, enabling in-situ visualization of planned trocar layouts. A demonstration video illustrating the overall workflow is available online. Results: In full-scale human-phantom experiments, ARport accurately overlaid pre-planned trocar sites onto the physical phantom, achieving consistent spatial correspondence between virtual plans and real anatomy. Conclusion: ARport provides a fully marker-free and hardware-minimal solution for visualizing preoperative trocar plans directly on the patient’s body surface. The system facilitates efficient intraoperative setup and demonstrates potential for seamless integration into routine clinical workflows.

Method: Builds on Moufad et al. (2025) VJP-free approximation for test-time guidance in diffusion and flow models, with extensive empirical evaluation on large-scale image and video editing benchmarks

Result: Test-time guidance alone achieves performance comparable to or surpassing training-based methods for image and video editing tasks

Conclusion: VJP-free test-time guidance provides a practical and effective approach for text-driven image and video editing without the computational overhead of VJP computations

Abstract: Text-driven image and video editing can be naturally cast as inpainting problems, where masked regions are reconstructed to remain consistent with both the observed content and the editing prompt. Recent advances in test-time guidance for diffusion and flow models provide a principled framework for this task; however, existing methods rely on costly vector–Jacobian product (VJP) computations to approximate the intractable guidance term, limiting their practical applicability. Building upon the recent work of Moufad et al. (2025), we provide theoretical insights into their VJP-free approximation and substantially extend their empirical evaluation to large-scale image and video editing benchmarks. Our results demonstrate that test-time guidance alone can achieve performance comparable to, and in some cases surpass, training-based methods.

Method: SEAL uses self-supervised learning with over 700,000 paired gene expression spot-tissue region examples from 14 organs. It’s designed as a parameter-efficient finetuning method that can be flexibly applied to existing pathology foundation models rather than training new encoders from scratch.

Result: SEAL consistently outperforms vision-only and spatial transcriptomics prediction baselines across 38 slide-level and 15 patch-level tasks, including molecular status, pathway activity, treatment response prediction, and gene expression prediction. It also shows robust domain generalization and enables new cross-modal capabilities like gene-to-image retrieval.

Conclusion: Augmenting existing pathology foundation models with localized molecular supervision from spatial transcriptomics is an effective and practical approach for improving visual representations and expanding cross-modal utility in computational pathology.

Abstract: Spatial transcriptomics (ST) provides spatially resolved measurements of gene expression, enabling characterization of the molecular landscape of human tissue beyond histological assessment as well as localized readouts that can be aligned with morphology. Concurrently, the success of multimodal foundation models that integrate vision with complementary modalities suggests that morphomolecular coupling between local expression and morphology can be systematically used to improve histological representations themselves. We introduce Spatial Expression-Aligned Learning (SEAL), a vision-omics self-supervised learning framework that infuses localized molecular information into pathology vision encoders. Rather than training new encoders from scratch, SEAL is designed as a parameter-efficient vision-omics finetuning method that can be flexibly applied to widely used pathology foundation models. We instantiate SEAL by training on over 700,000 paired gene expression spot-tissue region examples spanning tumor and normal samples from 14 organs. Tested across 38 slide-level and 15 patch-level downstream tasks, SEAL provides a drop-in replacement for pathology foundation models that consistently improves performance over widely used vision-only and ST prediction baselines on slide-level molecular status, pathway activity, and treatment response prediction, as well as patch-level gene expression prediction tasks. Additionally, SEAL encoders exhibit robust domain generalization on out-of-distribution evaluations and enable new cross-modal capabilities such as gene-to-image retrieval. Our work proposes a general framework for ST-guided finetuning of pathology foundation models, showing that augmenting existing models with localized molecular supervision is an effective and practical step for improving visual representations and expanding their cross-modal utility.

Method: Introduces UniWeTok with massive binary codebook (2^128), Pre-Post Distillation, Generative-Aware Prior, convolution-attention hybrid architecture with SigLu activation, and three-stage training framework for cross-resolution adaptability.

Result: Achieves SOTA image generation (FID: 1.38 vs REPA 1.42) with low training compute (33B vs 262B tokens), competitive multimodal understanding, generation (DPG: 86.63 vs FLUX.1 83.84), and editing (GEdit: 5.09 vs OmniGen 5.06).

Conclusion: UniWeTok provides a unified visual tokenizer that effectively balances reconstruction, semantic extraction, and generative capabilities, advancing multimodal LLM development with efficient training.

Abstract: Unified Multimodal Large Language Models (MLLMs) require a visual representation that simultaneously supports high-fidelity reconstruction, complex semantic extraction, and generative suitability. However, existing visual tokenizers typically struggle to satisfy these conflicting objectives within a single framework. In this paper, we introduce UniWeTok, a unified discrete tokenizer designed to bridge this gap using a massive binary codebook ($\mathit{2^{128}}$). For training framework, we introduce Pre-Post Distillation and a Generative-Aware Prior to enhance the semantic extraction and generative prior of the discrete tokens. In terms of model architecture, we propose a convolution-attention hybrid architecture with the SigLu activation function. SigLu activation not only bounds the encoder output and stabilizes the semantic distillation process but also effectively addresses the optimization conflict between token entropy loss and commitment loss. We further propose a three-stage training framework designed to enhance UniWeTok’s adaptability cross various image resolutions and perception-sensitive scenarios, such as those involving human faces and textual content. On ImageNet, UniWeTok achieves state-of-the-art image generation performance (FID: UniWeTok 1.38 vs. REPA 1.42) while requiring a remarkably low training compute (Training Tokens: UniWeTok 33B vs. REPA 262B). On general-domain, UniWeTok demonstrates highly competitive capabilities across a broad range of tasks, including multimodal understanding, image generation (DPG Score: UniWeTok 86.63 vs. FLUX.1 [Dev] 83.84), and editing (GEdit Overall Score: UniWeTok 5.09 vs. OmniGen 5.06). We release code and models to facilitate community exploration of unified tokenizer and MLLM.

Method: Introduces Sequence-Extended Latent Fusion (SELF) to serialize multiple reference images into coherent latent sequences. Uses two-stage training: 1) Supervised fine-tuning with progressive sequence length training (1024² to 2048² pixel budget), 2) Multi-Source GRPO reinforcement learning framework to optimize compositional consistency.

Result: Developed a high-performance multi-modal generation system that unifies single-image editing and multi-image composition. The system maintains consistency across multiple reference images and improves visual fidelity through progressive training.

Conclusion: UniRef-Image-Edit successfully addresses consistency issues in multi-reference image generation through unified latent representation and two-stage training, offering a comprehensive solution for both single and multi-image editing tasks.

Abstract: We present UniRef-Image-Edit, a high-performance multi-modal generation system that unifies single-image editing and multi-image composition within a single framework. Existing diffusion-based editing methods often struggle to maintain consistency across multiple conditions due to limited interaction between reference inputs. To address this, we introduce Sequence-Extended Latent Fusion (SELF), a unified input representation that dynamically serializes multiple reference images into a coherent latent sequence. During a dedicated training stage, all reference images are jointly constrained to fit within a fixed-length sequence under a global pixel-budget constraint. Building upon SELF, we propose a two-stage training framework comprising supervised fine-tuning (SFT) and reinforcement learning (RL). In the SFT stage, we jointly train on single-image editing and multi-image composition tasks to establish a robust generative prior. We adopt a progressive sequence length training strategy, in which all input images are initially resized to a total pixel budget of $1024^2$, and are then gradually increased to $1536^2$ and $2048^2$ to improve visual fidelity and cross-reference consistency. This gradual relaxation of compression enables the model to incrementally capture finer visual details while maintaining stable alignment across references. For the RL stage, we introduce Multi-Source GRPO (MSGRPO), to our knowledge the first reinforcement learning framework tailored for multi-reference image generation. MSGRPO optimizes the model to reconcile conflicting visual constraints, significantly enhancing compositional consistency. We will open-source the code, models, training data, and reward data for community research purposes.

Method: Two-stage framework: (1) Cold-start SFT dataset UHR Chain-of-Zoom (UHR-CoZ) covering diverse zooming regimes, and (2) Agentic reinforcement learning method AdaZoom-GRPO that explicitly rewards evidence gain and answer improvement during zoom interactions.

Result: The model learns on-demand zooming with proper stopping behavior and achieves substantial improvements on UHR remote sensing benchmarks, with 54.23% accuracy on XLRS-Bench.

Conclusion: GeoEyes effectively addresses tool usage homogenization in zoom-enabled MLLMs for remote sensing VQA through staged training combining specialized datasets and reinforcement learning with explicit evidence-based rewards.

Abstract: The “thinking-with-images” paradigm enables multimodal large language models (MLLMs) to actively explore visual scenes via zoom-in tools. This is essential for ultra-high-resolution (UHR) remote sensing VQA, where task-relevant cues are sparse and tiny. However, we observe a consistent failure mode in existing zoom-enabled MLLMs: Tool Usage Homogenization, where tool calls collapse into task-agnostic patterns, limiting effective evidence acquisition. To address this, we propose GeoEyes, a staged training framework consisting of (1) a cold-start SFT dataset, UHR Chain-of-Zoom (UHR-CoZ), which covers diverse zooming regimes, and (2) an agentic reinforcement learning method, AdaZoom-GRPO, that explicitly rewards evidence gain and answer improvement during zoom interactions. The resulting model learns on-demand zooming with proper stopping behavior and achieves substantial improvements on UHR remote sensing benchmarks, with 54.23% accuracy on XLRS-Bench.

Method: Three-module framework: (1) Perception module uses local context windows for coherent video understanding; (2) Ranking module with LLM-guided merge-sort for global re-ranking in VOD; (3) Prediction module with multi-modal time series model for low-latency live streaming.

Result: HiVid improves weight prediction accuracy by up to 11.5% for VOD and 26% for live streaming over SOTA baselines. Real-world user study shows 14.7% boost in streaming QoE correlation.

Conclusion: HiVid successfully leverages LLMs as scalable human proxies for generating high-fidelity importance weights in both VOD and live streaming, significantly improving QoE optimization.

Abstract: Content-aware streaming requires dynamic, chunk-level importance weights to optimize subjective quality of experience (QoE). However, direct human annotation is prohibitively expensive while vision-saliency models generalize poorly. We introduce HiVid, the first framework to leverage Large Language Models (LLMs) as a scalable human proxy to generate high-fidelity weights for both Video-on-Demand (VOD) and live streaming. We address 3 non-trivial challenges: (1) To extend LLMs’ limited modality and circumvent token limits, we propose a perception module to assess frames in a local context window, autoregressively building a coherent understanding of the video. (2) For VOD with rating inconsistency across local windows, we propose a ranking module to perform global re-ranking with a novel LLM-guided merge-sort algorithm. (3) For live streaming which requires low-latency, online inference without future knowledge, we propose a prediction module to predict future weights with a multi-modal time series model, which comprises a content-aware attention and adaptive horizon to accommodate asynchronous LLM inference. Extensive experiments show HiVid improves weight prediction accuracy by up to 11.5% for VOD and 26% for live streaming over SOTA baselines. Real-world user study validates HiVid boosts streaming QoE correlation by 14.7%.

Method: Proposes Freq-DP Net with two complementary priors: 1) geometric prior from defocus disparity using explicit cost volume, and 2) structural prior of fence’s global pattern via Fast Fourier Convolution. Uses attention mechanism to fuse these cues for accurate fence segmentation.

Result: Method significantly outperforms strong general-purpose baselines, establishing new state-of-the-art for single-image, DP-based fence removal. Authors built and released a diverse benchmark with different fence varieties.

Conclusion: First framework to leverage dual-pixel sensors for fence removal, combining geometric and structural priors effectively. Provides practical solution for single-image fence occlusion removal.

Abstract: Removing fence occlusions from single images is a challenging task that degrades visual quality and limits downstream computer vision applications. Existing methods often fail on static scenes or require motion cues from multiple frames. To overcome these limitations, we introduce the first framework to leverage dual-pixel (DP) sensors for this problem. We propose Freq-DP Net, a novel dual-branch network that fuses two complementary priors: a geometric prior from defocus disparity, modeled using an explicit cost volume, and a structural prior of the fence’s global pattern, learned via Fast Fourier Convolution (FFC). An attention mechanism intelligently merges these cues for highly accurate fence segmentation. To validate our approach, we build and release a diverse benchmark with different fence varieties. Experiments demonstrate that our method significantly outperforms strong general-purpose baselines, establishing a new state-of-the-art for single-image, DP-based fence removal.

Method: 1) Created topologically-enriched versions of ModelNet40 and ShapeNet by augmenting point clouds with persistent homology features. 2) Proposed TopoGAT, a deep learning-based method that learns to identify the most informative topological features directly from input data and topological signatures, avoiding hand-crafted statistical selection criteria.

Result: TopoGAT outperforms traditional statistical approaches in stability and discriminative power. Integrating selected persistent points into standard point cloud classification and part-segmentation pipelines improves both classification accuracy and segmentation metrics.

Conclusion: The topologically-enriched datasets and learnable feature selection approach enable broader integration of persistent homology into practical deep learning workflows for 3D point cloud analysis, bridging the gap between geometric and topological shape descriptors.

Abstract: Geometry and topology constitute complementary descriptors of three-dimensional shape, yet existing benchmark datasets primarily capture geometric information while neglecting topological structure. This work addresses this limitation by introducing topologically-enriched versions of ModelNet40 and ShapeNet, where each point cloud is augmented with its corresponding persistent homology features. These benchmarks with the topological signatures establish a foundation for unified geometry-topology learning and enable systematic evaluation of topology-aware deep learning architectures for 3D shape analysis. Building on this foundation, we propose a deep learning-based significant persistent point selection method, \textit{TopoGAT}, that learns to identify the most informative topological features directly from input data and the corresponding topological signatures, circumventing the limitations of hand-crafted statistical selection criteria. A comparative study verifies the superiority of the proposed method over traditional statistical approaches in terms of stability and discriminative power. Integrating the selected significant persistent points into standard point cloud classification and part-segmentation pipelines yields improvements in both classification accuracy and segmentation metrics. The presented topologically-enriched datasets, coupled with our learnable significant feature selection approach, enable the broader integration of persistent homology into the practical deep learning workflows for 3D point cloud analysis.

Method: Proactive memory management framework with dual-signal adaptive caching: 1) temporal filter using optical flow analysis to detect inter-frame redundancy, and 2) spatial filter using saliency detection to identify visually significant regions. Manages memory allocation before expensive attention operations.

Result: Achieves 2.20x compression ratio in effective memory usage while maintaining 100% accuracy across BLEU, ROUGE-L, and Exact Match metrics. Preserves context-rich features over extended temporal durations without degrading performance under identical memory budget constraints.

Conclusion: Sali-Cache enables efficient processing of long-form video content on consumer-grade hardware by intelligently managing memory allocation through proactive optimization, addressing the critical memory bottleneck in VLMs.

Abstract: Vision-Language Models (VLMs) face a critical memory bottleneck when processing long-form video content due to the linear growth of the Key-Value (KV) cache with sequence length. Existing solutions predominantly employ reactive eviction strategies that compute full attention matrices before discarding tokens, resulting in substantial computational waste. We propose Sali-Cache, a novel a priori optimization framework that implements dual-signal adaptive caching through proactive memory management. By integrating a temporal filter based on optical flow analysis for detecting inter-frame redundancy and a spatial filter leveraging saliency detection for identifying visually significant regions, Sali-Cache intelligently manages memory allocation before entering computationally expensive attention operations. Experimental evaluation on the LLaVA 1.6 architecture demonstrates that our method achieves a 2.20x compression ratio in effective memory usage while maintaining 100% accuracy across BLEU, ROUGE-L, and Exact Match metrics. Furthermore, under identical memory budget constraints, Sali-Cache preserves context-rich features over extended temporal durations without degrading model performance, enabling efficient processing of long-form video content on consumer-grade hardware.

Method: Proposes a decoupled architecture: 1) a vision-language transformer for touch-guided spatial placement based on intuitive touch priors, and 2) a diffusion model that jointly generates the object and an instance mask for high-fidelity blending. Also introduces the Touch2Add benchmark for standardized evaluation.

Result: The placement model significantly outperforms both random placement and general-purpose VLM baselines. The framework produces high-fidelity edits, and analysis shows strong correlation between initial placement accuracy and final edit quality, validating the decoupled approach.

Conclusion: The work paves the way for more accessible and efficient creative tools by introducing intuitive touch-based interaction for object addition, with a decoupled architecture that enables precise placement and high-fidelity blending.

Abstract: Instruction-based object addition is often hindered by the ambiguity of text-only prompts or the tedious nature of mask-based inputs. To address this usability gap, we introduce AbracADDbra, a user-friendly framework that leverages intuitive touch priors to spatially ground succinct instructions for precise placement. Our efficient, decoupled architecture uses a vision-language transformer for touch-guided placement, followed by a diffusion model that jointly generates the object and an instance mask for high-fidelity blending. To facilitate standardized evaluation, we contribute the Touch2Add benchmark for this interactive task. Our extensive evaluations, where our placement model significantly outperforms both random placement and general-purpose VLM baselines, confirm the framework’s ability to produce high-fidelity edits. Furthermore, our analysis reveals a strong correlation between initial placement accuracy and final edit quality, validating our decoupled approach. This work thus paves the way for more accessible and efficient creative tools.

Method: Created ScreenParse dataset with 771K web screenshots (21M elements) using Webshot pipeline for automated rendering, annotation extraction, and VLM-based relabeling. Trained ScreenVLM (316M parameters) with compact ScreenTag markup representation and structure-aware loss.

Result: ScreenVLM substantially outperforms larger foundation VLMs on dense parsing (0.592 vs 0.294 PageIoU) and shows strong transfer to public benchmarks. Finetuning foundation VLMs on ScreenParse consistently improves grounding performance.

Conclusion: Dense screen supervision provides transferable structural priors for UI understanding. ScreenParse enables training compact, efficient models for complete screen parsing with strong generalization capabilities.

Abstract: Modern computer-use agents (CUA) must perceive a screen as a structured state, what elements are visible, where they are, and what text they contain, before they can reliably ground instructions and act. Yet, most available grounding datasets provide sparse supervision, with insufficient and low-diversity labels that annotate only a small subset of task-relevant elements per screen, which limits both coverage and generalization; moreover, practical deployment requires efficiency to enable low-latency, on-device use. We introduce ScreenParse, a large-scale dataset for complete screen parsing, with dense annotations of all visible UI elements (boxes, 55-class types, and text) across 771K web screenshots (21M elements). ScreenParse is generated by Webshot, an automated, scalable pipeline that renders diverse urls, extracts annotations and applies VLM-based relabeling and quality filtering. Using ScreenParse, we train ScreenVLM, a compact, 316M-parameter vision language model (VLM) that decodes a compact ScreenTag markup representation with a structure-aware loss that upweights structure-critical tokens. ScreenVLM substantially outperforms much larger foundation VLMs on dense parsing (e.g., 0.592 vs. 0.294 PageIoU on ScreenParse) and shows strong transfer to public benchmarks. Moreover, finetuning foundation VLMs on ScreenParse consistently improves their grounding performance, suggesting that dense screen supervision provides transferable structural priors for UI understanding. Project page: https://saidgurbuz.github.io/screenparse/.

Method: Uses densely sampled geometric feature descriptors to replace photometric error with descriptor residuals, enabling sub-pixel accuracy from differential photometric methods while leveraging geometric descriptor expressiveness.

Result: Experiments show the proposed strategy results in accurate tracking but ultimately does not outperform pose optimization strategies based on reprojection error despite utilizing more information.

Conclusion: The descriptor similarity metric is too slowly varying and doesn’t necessarily correspond strictly to keypoint placement accuracy, explaining why the unified approach doesn’t outperform traditional methods.

Abstract: One of the fundamental problems in computer vision is the two-frame relative pose optimization problem. Primarily, two different kinds of error values are used: photometric error and re-projection error. The selection of error value is usually directly dependent on the selection of feature paradigm, photometric features, or geometric features. It is a trade-off between accuracy, robustness, and the possibility of loop closing. We investigate a third method that combines the strengths of both paradigms into a unified approach. Using densely sampled geometric feature descriptors, we replace the photometric error with a descriptor residual from a dense set of descriptors, thereby enabling the employment of sub-pixel accuracy in differential photometric methods, along with the expressiveness of the geometric feature descriptor. Experiments show that although the proposed strategy is an interesting approach that results in accurate tracking, it ultimately does not outperform pose optimization strategies based on re-projection error despite utilizing more information. We proceed to analyze the underlying reason for this discrepancy and present the hypothesis that the descriptor similarity metric is too slowly varying and does not necessarily correspond strictly to keypoint placement accuracy.

Method: A generative augmentation pipeline that fine-tunes a pre-trained Stable Diffusion model using Low-Rank Adaptation (LoRA) on the dark-skin subset of the ISIC dataset, generating synthetic dermoscopic images conditioned on lesion type and skin tone for data augmentation.

Result: Models trained on augmented data showed consistent improvements in segmentation metrics (IoU, Dice coefficient, boundary accuracy) and achieved 92.14% accuracy for binary classification using EfficientNet-B0, demonstrating reduced bias and increased fairness.

Conclusion: Synthetic data augmentation with Generative AI can substantially reduce bias and increase fairness in dermatological diagnostics, addressing skin tone imbalance in medical imaging datasets.

Abstract: Skin cancer is one of the most common cancers worldwide and early detection is critical for effective treatment. However, current AI diagnostic tools are often trained on datasets dominated by lighter skin tones, leading to reduced accuracy and fairness for people with darker skin. The International Skin Imaging Collaboration (ISIC) dataset, one of the most widely used benchmarks, contains over 70% light skin images while dark skins fewer than 8%. This imbalance poses a significant barrier to equitable healthcare delivery and highlights the urgent need for methods that address demographic diversity in medical imaging. This paper addresses this challenge of skin tone imbalance in automated skin cancer detection using dermoscopic images. To overcome this, we present a generative augmentation pipeline that fine-tunes a pre-trained Stable Diffusion model using Low-Rank Adaptation (LoRA) on the image dark-skin subset of the ISIC dataset and generates synthetic dermoscopic images conditioned on lesion type and skin tone. In this study, we investigated the utility of these images on two downstream tasks: lesion segmentation and binary classification. For segmentation, models trained on the augmented dataset and evaluated on held-out real images show consistent improvements in IoU, Dice coefficient, and boundary accuracy. These evalutions provides the verification of Generated dataset. For classification, an EfficientNet-B0 model trained on the augmented dataset achieved 92.14% accuracy. This paper demonstrates that synthetic data augmentation with Generative AI integration can substantially reduce bias with increase fairness in conventional dermatological diagnostics and open challenges for future directions.

Method: Proposes an inflammation detection framework with global-local encoder that combines self-supervised pretraining on large-scale healthy hand images with imbalance-aware training to detect RA-related joint inflammation from RGB hand images.

Result: The proposed approach improves F1-score by 0.2 points and Gmean by 0.25 points compared with the baseline model, demonstrating effectiveness in addressing the challenges of RA inflammation detection.

Conclusion: The framework successfully addresses key challenges in medical imaging for RA detection, providing a promising approach for accessible inflammation detection using RGB hand images with improved performance metrics.

Abstract: Rheumatoid arthritis (RA) is an autoimmune disease characterized by systemic joint inflammation. Early diagnosis and tight follow-up are essential to the management of RA, as ongoing inflammation can cause irreversible joint damage. The detection of arthritis is important for diagnosis and assessment of disease activity; however, it often takes a long time for patients to receive appropriate specialist care. Therefore, there is a strong need to develop systems that can detect joint inflammation easily using RGB images captured at home. Consequently, we tackle the task of RA inflammation detection from RGB hand images. This task is highly challenging due to general issues in medical imaging, such as the scarcity of positive samples, data imbalance, and the inherent difficulty of the task itself. However, to the best of our knowledge, no existing work has explicitly addressed these challenges in RGB-based RA inflammation detection. This paper quantitatively demonstrates the difficulty of visually detecting inflammation by constructing a dedicated dataset, and we propose a inflammation detection framework with global local encoder that combines self-supervised pretraining on large-scale healthy hand images with imbalance-aware training to detect RA-related joint inflammation from RGB hand images. Our experiments demonstrated that the proposed approach improves F1-score by 0.2 points and Gmean by 0.25 points compared with the baseline model.

Method: Proposes event-frame fusion framework combining events (temporally dense motion cues) and frames (spatially dense precise estimation). Uses Affine Invariant Simplicial (AIS) framework partitioning deformation field into linearized sub-regions with low-parametric representation. Introduces neighborhood-greedy optimization strategy for faster parameter search and reduced error accumulation.

Result: Outperforms state-of-the-art baseline by 1.6% in survival rate. Achieves this using only 18.9% of data storage and processing resources compared to high-speed video methods. Established benchmark dataset with over 120 sequences spanning diverse deformation scenarios.

Conclusion: Event-frame fusion with AIS modeling and neighborhood-greedy optimization enables efficient, accurate dense deformation tracking in dynamic scenes with significantly reduced computational and storage requirements compared to high-speed video approaches.

Abstract: Visual Deformation Measurement (VDM) aims to recover dense deformation fields by tracking surface motion from camera observations. Traditional image-based methods rely on minimal inter-frame motion to constrain the correspondence search space, which limits their applicability to highly dynamic scenes or necessitates high-speed cameras at the cost of prohibitive storage and computational overhead. We propose an event-frame fusion framework that exploits events for temporally dense motion cues and frames for spatially dense precise estimation. Revisiting the solid elastic modeling prior, we propose an Affine Invariant Simplicial (AIS) framework. It partitions the deformation field into linearized sub-regions with low-parametric representation, effectively mitigating motion ambiguities arising from sparse and noisy events. To speed up parameter searching and reduce error accumulation, a neighborhood-greedy optimization strategy is introduced, enabling well-converged sub-regions to guide their poorly-converged neighbors, effectively suppress local error accumulation in long-term dense tracking. To evaluate the proposed method, a benchmark dataset with temporally aligned event streams and frames is established, encompassing over 120 sequences spanning diverse deformation scenarios. Experimental results show that our method outperforms the state-of-the-art baseline by 1.6% in survival rate. Remarkably, it achieves this using only 18.9% of the data storage and processing resources of high-speed video methods.

Method: Key modification moves reference frames from diffusion latent space into a parallel conditioning pathway, preserving fixed chunk sizes and KV caching required by autoregressive models. Reuses existing pretrained VACE weights without additional training.

Result: VACE adds 20-30% latency overhead for structural control and inpainting across 1.3B and 14B model scales with negligible VRAM cost. However, reference-to-video fidelity is severely degraded compared to batch VACE due to causal attention constraints.

Conclusion: The adaptation enables real-time autoregressive video generation with VACE’s unified control capabilities, though with trade-offs in fidelity due to causal attention requirements of streaming pipelines.

Abstract: We describe an adaptation of VACE (Video All-in-one Creation and Editing) for real-time autoregressive video generation. VACE provides unified video control (reference guidance, structural conditioning, inpainting, and temporal extension) but assumes bidirectional attention over full sequences, making it incompatible with streaming pipelines that require fixed chunk sizes and causal attention. The key modification moves reference frames from the diffusion latent space into a parallel conditioning pathway, preserving the fixed chunk sizes and KV caching that autoregressive models require. This adaptation reuses existing pretrained VACE weights without additional training. Across 1.3B and 14B model scales, VACE adds 20-30% latency overhead for structural control and inpainting, with negligible VRAM cost relative to the base model. Reference-to-video fidelity is severely degraded compared to batch VACE due to causal attention constraints. A reference implementation is available at https://github.com/daydreamlive/scope.

Method: Two-level design: 1) At each turn, alternates between text and vision attack actions to elicit the most malicious response; 2) Across turns, adjusts attack trajectory through iterative back-and-forth refinement to gradually amplify response maliciousness.

Result: MAPA consistently outperforms state-of-the-art methods, improving attack success rates by 11-35% on benchmarks against LLaVA-V1.6-Mistral-7B, Qwen2.5-VL-7B-Instruct, Llama-3.2-Vision-11B-Instruct and GPT-4o-mini.

Conclusion: The proposed adaptive multi-turn attack method effectively bypasses safety mechanisms in vision-language models by strategically alternating between text and visual inputs and refining attack trajectories.

Abstract: Multi-turn jailbreak attacks are effective against text-only large language models (LLMs) by gradually introducing malicious content across turns. When extended to large vision-language models (LVLMs), we find that naively adding visual inputs can cause existing multi-turn jailbreaks to be easily defended. For example, overly malicious visual input will easily trigger the defense mechanism of safety-aligned LVLMs, making the response more conservative. To address this, we propose MAPA: a multi-turn adaptive prompting attack that 1) at each turn, alternates text-vision attack actions to elicit the most malicious response; and 2) across turns, adjusts the attack trajectory through iterative back-and-forth refinement to gradually amplify response maliciousness. This two-level design enables MAPA to consistently outperform state-of-the-art methods, improving attack success rates by 11-35% on recent benchmarks against LLaVA-V1.6-Mistral-7B, Qwen2.5-VL-7B-Instruct, Llama-3.2-Vision-11B-Instruct and GPT-4o-mini.

Method: Proposes pFedNavi, a structure-aware and dynamically adaptive personalized federated learning framework that adaptively identifies client-specific layers via layer-wise mixing coefficients and performs fine-grained parameter fusion on selected components (encoder-decoder projection and environment-sensitive decoder layers).

Result: Outperforms FedAvg-based VLN baseline on R2R and RxR benchmarks using both ResNet and CLIP visual representations, achieving up to 7.5% improvement in navigation success rate, up to 7.8% gain in trajectory fidelity, and converging 1.38x faster under non-IID conditions.

Conclusion: pFedNavi effectively balances global knowledge sharing with local specialization for VLN tasks, addressing privacy concerns while handling extreme client heterogeneity through personalized federated learning.

Abstract: Vision-Language Navigation VLN requires large-scale trajectory instruction data from private indoor environments, raising significant privacy concerns. Federated Learning FL mitigates this by keeping data on-device, but vanilla FL struggles under VLNs’ extreme cross-client heterogeneity in environments and instruction styles, making a single global model suboptimal. This paper proposes pFedNavi, a structure-aware and dynamically adaptive personalized federated learning framework tailored for VLN. Our key idea is to personalize where it matters: pFedNavi adaptively identifies client-specific layers via layer-wise mixing coefficients, and performs fine-grained parameter fusion on the selected components (e.g., the encoder-decoder projection and environment-sensitive decoder layers) to balance global knowledge sharing with local specialization. We evaluate pFedNavi on two standard VLN benchmarks, R2R and RxR, using both ResNet and CLIP visual representations. Across all metrics, pFedNavi consistently outperforms the FedAvg-based VLN baseline, achieving up to 7.5% improvement in navigation success rate and up to 7.8% gain in trajectory fidelity, while converging 1.38x faster under non-IID conditions.

Method: Proposes a feature recalibration based olfactory-visual multimodal model with two components: 1) Fine-grained deterioration embedding constructor (FDEC) to reconstruct labeled multimodal embedded-feature dataset, and 2) Fine-grained deterioration recalibration attention network (FDRA-Net) to emphasize signal variations and increase sensitivity to fine-grained deterioration on rice surface.

Result: Achieves 99.89% classification accuracy, outperforming state-of-the-art methods while simplifying the detection procedure. Field detection demonstrates advantages in accuracy and operational simplicity.

Conclusion: The proposed method effectively addresses limitations of existing multimodal approaches for rice deterioration detection and can be extended to other agrifood applications in agriculture and food industry.

Abstract: Multimodal methods are widely used in rice deterioration detection, which exhibit limited capability in representing and extracting fine-grained abnormal features. Moreover, these methods rely on devices, such as hyperspectral cameras and mass spectrometers, increasing detection costs and prolonging data acquisition time. To address these issues, we propose a feature recalibration based olfactory-visual multimodal model for fine-grained rice deterioration detection. The fine-grained deterioration embedding constructor (FDEC) is proposed to reconstruct the labeled multimodal embedded-feature dataset, enhancing sample representation. The fine-grained deterioration recalibration attention network (FDRA-Net) is proposed to emphasize signal variations and increase sensitivity to fine-grained deterioration on the rice surface. Experiments show that the proposed method achieves a classification accuracy of 99.89%. Compared with state-of-the-art methods, the detection accuracy is improved and the procedure is simplified. Furthermore, field detection demonstrates the advantages of accuracy and operational simplicity. The proposed method can also be extended to other agrifood in agriculture and food industry.

Method: Proposes an end-to-end modular framework where learning proposes geometric hypotheses (pose and depth) and geometric algorithms make final estimation decisions. Evaluated using VGGT as learning model for pose/depth proposals on RGB-D sequences, followed by classical point-to-plane RGB-D ICP as geometric backend on TUM RGB-D benchmark.

Result: Three key findings: (1) learning-based pose proposals alone are unreliable; (2) learning-proposed geometry can degrade performance when improperly aligned with camera intrinsics; (3) when learning-proposed depth is geometrically aligned and followed by geometric disposal stage, consistent improvements emerge in moderately challenging rigid settings.

Conclusion: Geometry is not merely a refinement component but an essential arbiter that validates and absorbs learning-based geometric observations. Modular, geometry-aware system design is crucial for robust spatial perception.

Abstract: Spatial perception aims to estimate camera motion and scene structure from visual observations, a problem traditionally addressed through geometric modeling and physical consistency constraints. Recent learning-based methods have demonstrated strong representational capacity for geometric perception and are increasingly used to augment classical geometry-centric systems in practice. However, whether learning components should directly replace geometric estimation or instead serve as intermediate modules within such pipelines remains an open question. In this work, we address this gap and investigate an end-to-end modular framework for effective spatial reasoning, where learning proposes geometric hypotheses, while geometric algorithms dispose estimation decisions. In particular, we study this principle in the context of relative camera pose estimation on RGB-D sequences. Using VGGT as a representative learning model, we evaluate learning-based pose and depth proposals under varying motion magnitudes and scene dynamics, followed by a classical point-to-plane RGB-D ICP as the geometric backend. Our experiments on the TUM RGB-D benchmark reveal three consistent findings: (1) learning-based pose proposals alone are unreliable; (2) learning-proposed geometry, when improperly aligned with camera intrinsics, can degrade performance; and (3) when learning-proposed depth is geometrically aligned and followed by a geometric disposal stage, consistent improvements emerge in moderately challenging rigid settings. These results demonstrate that geometry is not merely a refinement component, but an essential arbiter that validates and absorbs learning-based geometric observations. Our study highlights the importance of modular, geometry-aware system design for robust spatial perception.

Method: Controlled empirical study using systematic fault injection to examine faults affecting visual and inertial modalities across various operating regimes, with quantitative evaluation of VIO behavior under degraded sensing conditions.

Result: Pronounced asymmetry in fault impact: visual sensing degradations lead to bounded pose errors (~centimeters), while inertial sensing degradations cause substantially larger trajectory deviations (hundreds to thousands of meters).

Conclusion: Greater emphasis on inertial reliability is needed in the evaluation and design of XR systems for real-life industrial settings, as inertial sensor degradation poses much greater risk than visual degradation.

Abstract: Extended Reality (XR) systems deployed in industrial and operational settings rely on Visual–Inertial Odometry (VIO) for continuous six-degree-of-freedom pose tracking, yet these environments often involve sensing conditions that deviate from ideal assumptions. Despite this, most VIO evaluations emphasize nominal sensor behavior, leaving the effects of sustained sensor degradation under operational conditions insufficiently understood. This paper presents a controlled empirical study of VIO behavior under degraded sensing, examining faults affecting visual and inertial modalities across a range of operating regimes. Through systematic fault injection and quantitative evaluation, we observe a pronounced asymmetry in fault impact where degradations affecting visual sensing typically lead to bounded pose errors on the order of centimeters, whereas degradations affecting inertial sensing can induce substantially larger trajectory deviations, in some cases reaching hundreds to thousands of meters. These observations motivate greater emphasis on inertial reliability in the evaluation and design of XR systems for real-life industrial settings.

Method: Uses LLM to generate diverse AU descriptions, introduces AU-aware dynamic graph module for AU-specific visual representations, and hierarchical cross-modal attention with Disentangled Dual Cross-Attention (fine-grained AU-specific interactions) and Contextual Dual Cross-Attention (global inter-AU dependencies).

Result: HiVA consistently surpasses state-of-the-art approaches in AU detection. Qualitative analyses show semantically meaningful activation patterns and robust cross-modal correspondences.

Conclusion: HiVA effectively leverages vision-language interaction for comprehensive facial behavior analysis, producing robust and interpretable AU detection with enhanced semantic understanding.

Abstract: Facial Action Unit (AU) detection seeks to recognize subtle facial muscle activations as defined by the Facial Action Coding System (FACS). A primary challenge w.r.t AU detection is the effective learning of discriminative and generalizable AU representations under conditions of limited annotated data. To address this, we propose a Hierarchical Vision-language Interaction for AU Understanding (HiVA) method, which leverages textual AU descriptions as semantic priors to guide and enhance AU detection. Specifically, HiVA employs a large language model to generate diverse and contextually rich AU descriptions to strengthen language-based representation learning. To capture both fine-grained and holistic vision-language associations, HiVA introduces an AU-aware dynamic graph module that facilitates the learning of AU-specific visual representations. These features are further integrated within a hierarchical cross-modal attention architecture comprising two complementary mechanisms: Disentangled Dual Cross-Attention (DDCA), which establishes fine-grained, AU-specific interactions between visual and textual features, and Contextual Dual Cross-Attention (CDCA), which models global inter-AU dependencies. This collaborative, cross-modal learning paradigm enables HiVA to leverage multi-grained vision-based AU features in conjunction with refined language-based AU details, culminating in robust and semantically enriched AU detection capabilities. Extensive experiments show that HiVA consistently surpasses state-of-the-art approaches. Besides, qualitative analyses reveal that HiVA produces semantically meaningful activation patterns, highlighting its efficacy in learning robust and interpretable cross-modal correspondences for comprehensive facial behavior analysis.

Method: D-SECURE integrates HAMMER (manipulation detector) with DEFAME (retrieval pipeline). DEFAME performs broad verification first, then HAMMER analyzes residual or uncertain cases that may contain fine-grained edits. The framework combines internal manipulation detection with external evidence-based reasoning.

Result: Experiments on DGM4 and ClaimReview samples demonstrate the complementary strengths of both systems and motivate their fusion. The framework provides unified, explainable reports incorporating both manipulation cues and external evidence.

Conclusion: Combining internal manipulation detection with external evidence-based reasoning creates a more robust multimodal misinformation detection system that can identify both factual inaccuracies and subtle visual/textual manipulations in news-style posts.

Abstract: Multimodal misinformation increasingly mixes realistic im-age edits with fluent but misleading text, producing persuasive posts that are difficult to verify. Existing systems usually rely on a single evidence source. Content-based detectors identify local inconsistencies within an image and its caption but cannot determine global factual truth. Retrieval-based fact-checkers reason over external evidence but treat inputs as coarse claims and often miss subtle visual or textual manipulations. This separation creates failure cases where internally consistent fabrications bypass manipulation detectors and fact-checkers verify claims that contain pixel-level or token-level corruption. We present D-SECURE, a framework that combines internal manipulation detection with external evidence-based reasoning for news-style posts. D-SECURE integrates the HAMMER manipulation detector with the DEFAME retrieval pipeline. DEFAME performs broad verification, and HAMMER analyses residual or uncertain cases that may contain fine-grained edits. Experiments on DGM4 and ClaimReview samples highlight the complementary strengths of both systems and motivate their fusion. We provide a unified, explainable report that incorporates manipulation cues and external evidence.

Method: The approach first parses images into hierarchical vector graphics representations that are semantic-aligned and structurally coherent. It then designs an image synthesis framework guided by these VGs, leveraging their structural and semantic features with noise prediction for precise control over geometry, color, and object semantics.

Result: Extensive experiments demonstrate effectiveness in diverse applications including image editing, object-level manipulation, and fine-grained content creation, establishing a new paradigm for controllable image generation.

Conclusion: The work introduces a novel approach to layer-wise controllable generation through vector graphics, enabling free modification of elements and seamless translation into photorealistic outputs with precise control over various visual attributes.

Abstract: Recent advances in image generation have achieved remarkable visual quality, while a fundamental challenge remains: Can image generation be controlled at the element level, enabling intuitive modifications such as adjusting shapes, altering colors, or adding and removing objects? In this work, we address this challenge by introducing layer-wise controllable generation through simplified vector graphics (VGs). Our approach first efficiently parses images into hierarchical VG representations that are semantic-aligned and structurally coherent. Building on this representation, we design a novel image synthesis framework guided by VGs, allowing users to freely modify elements and seamlessly translate these edits into photorealistic outputs. By leveraging the structural and semantic features of VGs in conjunction with noise prediction, our method provides precise control over geometry, color, and object semantics. Extensive experiments demonstrate the effectiveness of our approach in diverse applications, including image editing, object-level manipulation, and fine-grained content creation, establishing a new paradigm for controllable image generation. Project page: https://guolanqing.github.io/Vec2Pix/

Method: Leverages pretrained latent diffusion models without additional training. Introduces pixel-wise semantic correspondence module that mines intermediate diffusion features to construct dense alignment maps between content and style images, plus a cycle-consistency module to enforce structural and perceptual alignment across iterations.

Result: Delivers state-of-the-art visual quality and strong quantitative results despite requiring no additional training or supervision, outperforming methods that rely on extra training or annotations.

Conclusion: CoCoDiff demonstrates that correspondence cues within generative diffusion models are under-explored and can be effectively leveraged for fine-grained, semantically consistent style transfer without additional training.

Abstract: Transferring visual style between images while preserving semantic correspondence between similar objects remains a central challenge in computer vision. While existing methods have made great strides, most of them operate at global level but overlook region-wise and even pixel-wise semantic correspondence. To address this, we propose CoCoDiff, a novel training-free and low-cost style transfer framework that leverages pretrained latent diffusion models to achieve fine-grained, semantically consistent stylization. We identify that correspondence cues within generative diffusion models are under-explored and that content consistency across semantically matched regions is often neglected. CoCoDiff introduces a pixel-wise semantic correspondence module that mines intermediate diffusion features to construct a dense alignment map between content and style images. Furthermore, a cycle-consistency module then enforces structural and perceptual alignment across iterations, yielding object and region level stylization that preserves geometry and detail. Despite requiring no additional training or supervision, CoCoDiff delivers state-of-the-art visual quality and strong quantitative results, outperforming methods that rely on extra training or annotations.

Method: TikArt uses a Think-Aperture-Observe loop: alternating language generation with aperture actions (Zoom for rectangular crops, Segment using SAM2 for mask-based crops). After each action, explicit observations convert visual cues to linguistic memory. Built on Qwen3-VL-8B, optimized with AGRPO (GRPO-style RL) using two-stage curriculum: warm-up segmentation then joint optimization of visual math, VQA, and segmentation tasks.

Result: Experiments on V*, HR-Bench-4K/8K, MME-RealWorld-Lite, MMStar, RefCOCO, and ReasonSeg show consistent gains over backbone model. Produces interpretable aperture trajectories for high-resolution reasoning tasks.

Conclusion: TikArt effectively addresses fine-grained visual reasoning by enabling dynamic region-of-interest selection through aperture actions, converting visual details to persistent linguistic memory, and demonstrating improved performance across diverse visual reasoning benchmarks.

Abstract: We address fine-grained visual reasoning in multimodal large language models (MLLMs), where key evidence may reside in tiny objects, cluttered regions, or subtle markings that are lost under a single global image encoding. We introduce TikArt (Thinking Aperture), an aperture-guided agent that casts multi-step vision-language reasoning as a decision process over regions of interest. TikArt follows a Think-Aperture-Observe loop, alternating between language generation and two aperture actions: Zoom extracts rectangular crops, while Segment invokes SAM2 to obtain mask-based crops for irregular targets. After every action, the model must produce an explicit observation, turning local visual cues into persistent linguistic memory. Built on Qwen3-VL-8B, TikArt optimizes its reasoning policy with AGRPO, a GRPO-style reinforcement learning algorithm with a two-stage curriculum: it warms up segmentation actions and then jointly optimizes visual math, fine-grained VQA, and segmentation, using rewards that couple task success with purposeful aperture use. Experiments on V*, HR-Bench-4K/8K, MME-RealWorld-Lite, MMStar, RefCOCO, and ReasonSeg show consistent gains over the backbone and yield interpretable aperture trajectories for high-resolution reasoning.

Method: Proposes Gaussian Mesh Renderer (GMR) that leverages 3DGS rasterization by analytically deriving Gaussian primitives from mesh triangles, preserving structural fidelity and enabling gradient flow through the mesh representation.

Result: Achieves smoother gradients compared to traditional mesh renderers, enabling better optimization with smaller batch sizes and limited memory, while maintaining mesh structure fidelity.

Conclusion: GMR provides an efficient differentiable mesh renderer that combines the benefits of both Gaussian splatting (fast rendering) and mesh representations (structural fidelity), enabling better optimization for surface reconstruction tasks.

Abstract: 3D Gaussian Splatting (3DGS) has enabled high-fidelity virtualization with fast rendering and optimization for novel view synthesis. On the other hand, triangle mesh models still remain a popular choice for surface reconstruction but suffer from slow or heavy optimization in traditional mesh-based differentiable renderers. To address this problem, we propose a new lightweight differentiable mesh renderer leveraging the efficient rasterization process of 3DGS, named Gaussian Mesh Renderer (GMR), which tightly integrates the Gaussian and mesh representations. Each Gaussian primitive is analytically derived from the corresponding mesh triangle, preserving structural fidelity and enabling the gradient flow. Compared to the traditional mesh renderers, our method achieves smoother gradients, which especially contributes to better optimization using smaller batch sizes with limited memory. Our implementation is available in the public GitHub repository at https://github.com/huntorochi/Gaussian-Mesh-Renderer.

Method: Proposes Modality Decoding Attention Block (MoDAB) with lightweight State Space Mixer (SSMix) for efficient cross-modal fusion and long-range dependency modeling. Introduces Spectral-Entropic Uncertainty (SEU) Loss to jointly capture spatial overlap, spectral consistency, and predictive uncertainty in a unified objective.

Result: Superior segmentation performance on medical datasets (QATA-COVID19, MosMed++, Kvasir-SEG) while being significantly more computationally efficient than existing SoTA approaches. Demonstrates improved model reliability in complex clinical circumstances with poor image quality.

Conclusion: Highlights the importance of incorporating uncertainty modeling and structured modality alignment in vision-language medical segmentation tasks, showing that multimodal approaches with uncertainty awareness can significantly improve medical diagnostic reliability.

Abstract: We introduce a novel uncertainty-aware multimodal segmentation framework that leverages both radiological images and associated clinical text for precise medical diagnosis. We propose a Modality Decoding Attention Block (MoDAB) with a lightweight State Space Mixer (SSMix) to enable efficient cross-modal fusion and long-range dependency modelling. To guide learning under ambiguity, we propose the Spectral-Entropic Uncertainty (SEU) Loss, which jointly captures spatial overlap, spectral consistency, and predictive uncertainty in a unified objective. In complex clinical circumstances with poor image quality, this formulation improves model reliability. Extensive experiments on various publicly available medical datasets, QATA-COVID19, MosMed++, and Kvasir-SEG, demonstrate that our method achieves superior segmentation performance while being significantly more computationally efficient than existing State-of-the-Art (SoTA) approaches. Our results highlight the importance of incorporating uncertainty modelling and structured modality alignment in vision-language medical segmentation tasks. Code: https://github.com/arya-domain/UA-VLS

Method: Two-part approach: 1) Secondary instance subspace learning with low-rank regularized subspace clustering (LRSC) to address instance entanglement from excessive non-tumor instances; 2) Enhanced contrastive learning for prototype instance semantic disentanglement (PID) to resolve semantic entanglement between tumor and precancerous tissues.

Result: Extensive experiments on multicentre pathology datasets show PID-LRSC outperforms other state-of-the-art methods, providing clearer instance semantics during decision-making and significantly enhancing reliability of auxiliary diagnostic outcomes.

Conclusion: PID-LRSC effectively addresses instance-semantic entanglement in pathological diagnosis, improving both model performance and interpretability for tumor detection in whole slide images.

Abstract: The tumor region plays a key role in pathological diagnosis. Tumor tissues are highly similar to precancerous lesions and non tumor instances often greatly exceed tumor instances in whole slide images (WSIs). These issues cause instance-semantic entanglement in multi-instance learning frameworks, degrading both model representation capability and interpretability. To address this, we propose an end-to-end prototype instance semantic disentanglement framework with low-rank regularized subspace clustering, PID-LRSC, in two aspects. First, we use secondary instance subspace learning to construct low-rank regularized subspace clustering (LRSC), addressing instance entanglement caused by an excessive proportion of non tumor instances. Second, we employ enhanced contrastive learning to design prototype instance semantic disentanglement (PID), resolving semantic entanglement caused by the high similarity between tumor and precancerous tissues. We conduct extensive experiments on multicentre pathology datasets, implying that PID-LRSC outperforms other SOTA methods. Overall, PID-LRSC provides clearer instance semantics during decision-making and significantly enhances the reliability of auxiliary diagnostic outcomes.

Method: Proposes end-to-end MIL framework with Grassmann re-embedding and manifold adaptive clustering. Uses manifold geometric structure for robust clustering. Includes prior knowledge guiding proxy instance labeling and aggregation strategy to approximate patch labels and focus on tumor regions.

Result: Experiments on multicentre WSI datasets show: 1) cluster-incorporated model achieves superior performance in both grading accuracy and interpretability; 2) end-to-end learning refines feature representations with acceptable computation resources.

Conclusion: The proposed framework improves both accuracy and interpretability for WSI analysis through integrated clustering and end-to-end learning, addressing limitations of existing methods.

Abstract: Whole slide images (WSIs) are the gold standard for pathological diagnosis and sub-typing. Current main-stream two-step frameworks employ offline feature encoders trained without domain-specific knowledge. Among them, attention-based multiple instance learning (MIL) methods are outcome-oriented and offer limited interpretability. Clustering-based approaches can provide explainable decision-making process but suffer from high dimension features and semantically ambiguous centroids. To this end, we propose an end-to-end MIL framework that integrates Grassmann re-embedding and manifold adaptive clustering, where the manifold geometric structure facilitates robust clustering results. Furthermore, we design a prior knowledge guiding proxy instance labeling and aggregation strategy to approximate patch labels and focus on pathologically relevant tumor regions. Experiments on multicentre WSI datasets demonstrate that: 1) our cluster-incorporated model achieves superior performance in both grading accuracy and interpretability; 2) end-to-end learning refines better feature representations and it requires acceptable computation resources.

Method: Autoregressive-based foundation model using next-scale prediction paradigm for fast, scale-up-friendly synthesis; generates images coarse-to-fine with structured multi-scale representations; trained on harmonized dataset of ~440,000 CT/MRI images across six anatomical regions.

Result: Achieves state-of-the-art generative performance across fidelity, diversity, and scalability metrics; offers promising architectural direction for future medical generative foundation models.

Conclusion: MedVAR successfully addresses key challenges in medical image generation through its autoregressive architecture and hierarchical approach, establishing a foundation for scalable medical imaging models.

Abstract: Medical image generation is pivotal in applications like data augmentation for low-resource clinical tasks and privacy-preserving data sharing. However, developing a scalable generative backbone for medical imaging requires architectural efficiency, sufficient multi-organ data, and principled evaluation, yet current approaches leave these aspects unresolved. Therefore, we introduce MedVAR, the first autoregressive-based foundation model that adopts the next-scale prediction paradigm to enable fast and scale-up-friendly medical image synthesis. MedVAR generates images in a coarse-to-fine manner and produces structured multi-scale representations suitable for downstream use. To support hierarchical generation, we curate a harmonized dataset of around 440,000 CT and MRI images spanning six anatomical regions. Comprehensive experiments across fidelity, diversity, and scalability show that MedVAR achieves state-of-the-art generative performance and offers a promising architectural direction for future medical generative foundation models.

Method: Proposed two efficient adapters (Nexus Prime and Nexus Slim) that incorporate cross-attention mechanisms to enable rich multimodal conditioning, allowing the adapter to understand input prompts while preserving structural inputs like sketches or depth maps.

Result: Nexus Prime requires only 8M additional parameters compared to T2I-Adapter baseline while significantly enhancing performance. Nexus Slim has 18M fewer parameters than T2I-Adapter and still achieves state-of-the-art results.

Conclusion: Nexus Adapters provide an efficient solution for structure-preserving conditional generation that is both prompt-aware and parameter-efficient, overcoming limitations of previous methods.

Abstract: We introduce the Nexus Adapters, novel text-guided efficient adapters to the diffusion-based framework for the Structure Preserving Conditional Generation (SPCG). Recently, structure-preserving methods have achieved promising results in conditional image generation by using a base model for prompt conditioning and an adapter for structure input, such as sketches or depth maps. These approaches are highly inefficient and sometimes require equal parameters in the adapter compared to the base architecture. It is not always possible to train the model since the diffusion model is itself costly, and doubling the parameter is highly inefficient. In these approaches, the adapter is not aware of the input prompt; therefore, it is optimal only for the structural input but not for the input prompt. To overcome the above challenges, we proposed two efficient adapters, Nexus Prime and Slim, which are guided by prompts and structural inputs. Each Nexus Block incorporates cross-attention mechanisms to enable rich multimodal conditioning. Therefore, the proposed adapter has a better understanding of the input prompt while preserving the structure. We conducted extensive experiments on the proposed models and demonstrated that the Nexus Prime adapter significantly enhances performance, requiring only 8M additional parameters compared to the baseline, T2I-Adapter. Furthermore, we also introduced a lightweight Nexus Slim adapter with 18M fewer parameters than the T2I-Adapter, which still achieved state-of-the-art results. Code: https://github.com/arya-domain/Nexus-Adapters

Method: Evaluated 79 open-source deep learning models (CNNs and Vision Transformers) across 2,300+ experiments on newly curated Quad-State Tornado Damage benchmark dataset, systematically testing architecture-optimizer interactions.

Result: Optimizer choice proved more consequential than architecture - switching from Adam to SGD gave +25 to +38 F1 points for Vision Transformers, reversing their ranking. Low learning rate (1e-4) boosted average F1 by +10.2 points. ConvNeXt-Base with optimized settings achieved 46.4% Macro F1 (+34.6 over baseline) with strong cross-event generalization.

Conclusion: Operational-grade tornado damage assessment requires optimizing architecture-optimizer interactions, not just architecture selection. SGD optimizer and low learning rate are critical for handling domain shift and class imbalance in disaster imagery.

Abstract: Rapid and accurate building damage assessment in the immediate aftermath of tornadoes is critical for coordinating life-saving search and rescue operations, optimizing emergency resource allocation, and accelerating community recovery. However, current automated methods struggle with the unique visual complexity of tornado-induced wreckage, primarily due to severe domain shift from standard pre-training datasets and extreme class imbalance in real-world disaster data. To address these challenges, we introduce a systematic experimental framework evaluating 79 open-source deep learning models, encompassing both Convolutional Neural Networks (CNNs) and Vision Transformers, across over 2,300 controlled experiments on our newly curated Quad-State Tornado Damage (QSTD) benchmark dataset. Our findings reveal that achieving operational-grade performance hinges on a complex interaction between architecture and optimization, rather than architectural selection alone. Most strikingly, we demonstrate that optimizer choice can be more consequential than architecture: switching from Adam to SGD provided dramatic F1 gains of +25 to +38 points for Vision Transformer and Swin Transformer families, fundamentally reversing their ranking from bottom-tier to competitive with top-performing CNNs. Furthermore, a low learning rate of 1x10^(-4) proved universally critical, boosting average F1 performance by +10.2 points across all architectures. Our champion model, ConvNeXt-Base trained with these optimized settings, demonstrated strong cross-event generalization on the held-out Tuscaloosa-Moore Tornado Damage (TMTD) dataset, achieving 46.4% Macro F1 (+34.6 points over baseline) and retaining 85.5% Ordinal Top-1 Accuracy despite temporal and sensor domain shifts.

Method: Compare dedicated OCR transformer (TrOCR) and general-purpose Vision-Language Model (Qwen) on line-level historical English texts using length-weighted accuracy metrics and hypothesis-driven error analysis to understand systematic error patterns.

Result: Qwen achieves lower CER/WER and greater robustness to degraded input but exhibits selective linguistic regularization that silently alters historically meaningful forms. TrOCR preserves orthographic fidelity more consistently but is more prone to cascading error propagation.

Conclusion: Architectural inductive biases shape OCR error structure systematically. Models with similar aggregate accuracy differ substantially in error locality, detectability, and downstream scholarly risk, highlighting need for architecture-aware evaluation in historical digitization.

Abstract: Optical Character Recognition (OCR) of eighteenth-century printed texts remains challenging due to degraded print quality, archaic glyphs, and non-standardized orthography. Although transformer-based OCR systems and Vision-Language Models (VLMs) achieve strong aggregate accuracy, metrics such as Character Error Rate (CER) and Word Error Rate (WER) provide limited insight into their reliability for scholarly use. We compare a dedicated OCR transformer (TrOCR) and a general-purpose Vision-Language Model (Qwen) on line-level historical English texts using length-weighted accuracy metrics and hypothesis driven error analysis. While Qwen achieves lower CER/WER and greater robustness to degraded input, it exhibits selective linguistic regularization and orthographic normalization that may silently alter historically meaningful forms. TrOCR preserves orthographic fidelity more consistently but is more prone to cascading error propagation. Our findings show that architectural inductive biases shape OCR error structure in systematic ways. Models with similar aggregate accuracy can differ substantially in error locality, detectability, and downstream scholarly risk, underscoring the need for architecture-aware evaluation in historical digitization workflows.

Method: Proposes CVGC with two modules: 1) cross-view geometric augmentation that models viewpoint-induced variations in visibility and sampling density to generate multiple cross-view observations, and 2) geometric consistency that enforces consistent semantic and occupancy predictions across augmented point clouds.

Result: Extensive experiments on six public LiDAR datasets establish the first systematic evaluation of cross-view domain generalization for LiDAR semantic segmentation, showing CVGC consistently outperforms state-of-the-art methods when generalizing from single source to multiple target domains with heterogeneous acquisition viewpoints.

Conclusion: CVGC effectively addresses cross-view domain generalization challenges in LiDAR semantic segmentation by modeling viewpoint variations and enforcing geometric consistency, enabling better generalization to unseen domains with different acquisition viewpoints.

Abstract: Domain-generalized LiDAR semantic segmentation (LSS) seeks to train models on source-domain point clouds that generalize reliably to multiple unseen target domains, which is essential for real-world LiDAR applications. However, existing approaches assume similar acquisition views (e.g., vehicle-mounted) and struggle in cross-view scenarios, where observations differ substantially due to viewpoint-dependent structural incompleteness and non-uniform point density. Accordingly, we formulate cross-view domain generalization for LiDAR semantic segmentation and propose a novel framework, termed CVGC (Cross-View Geometric Consistency). Specifically, we introduce a cross-view geometric augmentation module that models viewpoint-induced variations in visibility and sampling density, generating multiple cross-view observations of the same scene. Subsequently, a geometric consistency module enforces consistent semantic and occupancy predictions across geometrically augmented point clouds of the same scene. Extensive experiments on six public LiDAR datasets establish the first systematic evaluation of cross-view domain generalization for LiDAR semantic segmentation, demonstrating that CVGC consistently outperforms state-of-the-art methods when generalizing from a single source domain to multiple target domains with heterogeneous acquisition viewpoints. The source code will be publicly available at https://github.com/KintomZi/CVGC-DG

Method: Proposes MoRL with task-specific reward design combining semantic alignment and reasoning coherence for understanding, and physical plausibility and text-motion consistency for generation. Introduces Chain-of-Motion (CoM) for test-time reasoning and constructs two large-scale CoT datasets (MoUnd-CoT-140K and MoGen-CoT-140K).

Result: Experiments on HumanML3D and KIT-ML show significant gains over state-of-the-art baselines in both motion understanding and generation tasks.

Conclusion: MoRL advances human motion modeling by integrating reasoning capabilities with multimodal learning, providing a unified framework for both understanding and generation tasks.

Abstract: Human motion understanding and generation are crucial for vision and robotics but remain limited in reasoning capability and test-time planning. We propose MoRL, a unified multimodal motion model trained with supervised fine-tuning and reinforcement learning with verifiable rewards. Our task-specific reward design combines semantic alignment and reasoning coherence for understanding with physical plausibility and text-motion consistency for generation, improving both logical reasoning and perceptual realism. To further enhance inference, we introduce Chain-of-Motion (CoM), a test-time reasoning method that enables step-by-step planning and reflection. We also construct two large-scale CoT datasets, MoUnd-CoT-140K and MoGen-CoT-140K, to align motion sequences with reasoning traces and action descriptions. Experiments on HumanML3D and KIT-ML show that MoRL achieves significant gains over state-of-the-art baselines. Code: https://github.com/AIGeeksGroup/MoRL. Website: https://aigeeksgroup.github.io/MoRL.

Method: The framework coordinates three main components: 1) Structured Garment Morphing for correspondence-driven garment adaptation, 2) Principal Pose Guidance for step-wise structural regulation during diffusion sampling, and 3) Continuous Boundary Stitching for boundary-aware refinement. This forms a cohesive pipeline that operates without task-specific retraining.

Result: OmniVTON++ achieves state-of-the-art performance across diverse generalization settings including cross-dataset and cross-garment-type evaluations. It reliably operates across scenarios and diffusion backbones within a single formulation, supporting multi-garment, multi-human, and anime character virtual try-on.

Conclusion: The framework successfully addresses the intertwined challenges of garment alignment, human structural coherence, and boundary continuity, expanding the scope of virtual try-on applications while maintaining training-free universal applicability.

Abstract: Image-based Virtual Try-On (VTON) concerns the synthesis of realistic person imagery through garment re-rendering under human pose and body constraints. In practice, however, existing approaches are typically optimized for specific data conditions, making their deployment reliant on retraining and limiting their generalization as a unified solution. We present OmniVTON++, a training-free VTON framework designed for universal applicability. It addresses the intertwined challenges of garment alignment, human structural coherence, and boundary continuity by coordinating Structured Garment Morphing for correspondence-driven garment adaptation, Principal Pose Guidance for step-wise structural regulation during diffusion sampling, and Continuous Boundary Stitching for boundary-aware refinement, forming a cohesive pipeline without task-specific retraining. Experimental results demonstrate that OmniVTON++ achieves state-of-the-art performance across diverse generalization settings, including cross-dataset and cross-garment-type evaluations, while reliably operating across scenarios and diffusion backbones within a single formulation. In addition to single-garment, single-human cases, the framework supports multi-garment, multi-human, and anime character virtual try-on, expanding the scope of virtual try-on applications. The source code will be released to the public.

Method: Proposes DriveFine with a novel plug-and-play block-MoE that seamlessly injects a refinement expert on top of the generation expert. Uses explicit expert selection during inference and gradient blocking during training to fully decouple experts. Also designs a hybrid reinforcement learning strategy to encourage effective exploration of refinement expert while maintaining training stability.

Result: Extensive experiments on NAVSIM v1, v2, and Navhard benchmarks demonstrate that DriveFine exhibits strong efficacy and robustness in autonomous driving tasks.

Conclusion: DriveFine successfully addresses limitations of existing VLA models for autonomous driving by combining the strengths of diffusion and token-based approaches through a novel block-MoE architecture with refinement capabilities.

Abstract: Vision-Language-Action (VLA) models for autonomous driving increasingly adopt generative planners trained with imitation learning followed by reinforcement learning. Diffusion-based planners suffer from modality alignment difficulties, low training efficiency, and limited generalization. Token-based planners are plagued by cumulative causal errors and irreversible decoding. In summary, the two dominant paradigms exhibit complementary strengths and weaknesses. In this paper, we propose DriveFine, a masked diffusion VLA model that combines flexible decoding with self-correction capabilities. In particular, we design a novel plug-and-play block-MoE, which seamlessly injects a refinement expert on top of the generation expert. By enabling explicit expert selection during inference and gradient blocking during training, the two experts are fully decoupled, preserving the foundational capabilities and generic patterns of the pretrained weights, which highlights the flexibility and extensibility of the block-MoE design. Furthermore, we design a hybrid reinforcement learning strategy that encourages effective exploration of refinement expert while maintaining training stability. Extensive experiments on NAVSIM v1, v2, and Navhard benchmarks demonstrate that DriveFine exhibits strong efficacy and robustness. The code will be released at https://github.com/MSunDYY/DriveFine.

Method: Rigorous architectural investigation using YOLO26’s GitHub repository source code and official documentation to extract authentic operational mechanisms, resulting in creation of YOLO26’s CNN-based architectural diagram.

Result: First presentation of YOLO26’s CNN-based architecture, detailing key enhancements: elimination of DFL, end-to-end NMS-free inference, ProgLoss+STAL label assignment, MuSGD optimizer, achieving 43% faster CPU inference for edge deployment.

Conclusion: YOLO26 represents significant architectural evolution with optimizations for edge deployment, providing researchers/developers with precise architectural understanding to enhance YOLO models and maintain leadership in computer vision.

Abstract: You Only Look Once (YOLO) has been the prominent model for computer vision in deep learning for a decade. This study explores the novel aspects of YOLO26, the most recent version in the YOLO series. The elimination of Distribution Focal Loss (DFL), implementation of End-to-End NMS-Free Inference, introduction of ProgLoss + Small-Target-Aware Label Assignment (STAL), and use of the MuSGD optimizer are the primary enhancements designed to improve inference speed, which is claimed to achieve a 43% boost in CPU mode. This is designed to allow YOLO26 to attain real-time performance on edge devices or those without GPUs. Additionally, YOLO26 offers improvements in many computer vision tasks, including instance segmentation, pose estimation, and oriented bounding box (OBB) decoding. We aim for this effort to provide more value than just consolidating information already included in the existing technical documentation. Therefore, we performed a rigorous architectural investigation into YOLO26, mostly using the source code available in its GitHub repository and its official documentation. The authentic and detailed operational mechanisms of YOLO26 are inside the source code, which is seldom extracted by others. The YOLO26 architectural diagram is shown as the outcome of the investigation. This study is, to our knowledge, the first one presenting the CNN-based YOLO26 architecture, which is the core of YOLO26. Our objective is to provide a precise architectural comprehension of YOLO26 for researchers and developers aspiring to enhance the YOLO model, ensuring it remains the leading deep learning model in computer vision.

Method: Proposes VariViT with: 1) Novel positional embedding resizing scheme for variable number of patches, 2) New batching strategy to reduce computational complexity, 3) Maintains consistent patch size while handling variable image dimensions.

Result: Outperforms vanilla ViTs and ResNet on glioma genotype prediction and brain tumor classification with F1-scores of 75.5% and 76.3%. Batching strategy reduces computation time by up to 30% compared to conventional architectures.

Conclusion: VariViT effectively handles variable image sizes in medical imaging, learns more discriminative features, and reduces computational overhead, demonstrating efficacy in image representation learning for medical applications.

Abstract: Vision Transformers (ViTs) have emerged as the state-of-the-art architecture in representation learning, leveraging self-attention mechanisms to excel in various tasks. ViTs split images into fixed-size patches, constraining them to a predefined size and necessitating pre-processing steps like resizing, padding, or cropping. This poses challenges in medical imaging, particularly with irregularly shaped structures like tumors. A fixed bounding box crop size produces input images with highly variable foreground-to-background ratios. Resizing medical images can degrade information and introduce artefacts, impacting diagnosis. Hence, tailoring variable-sized crops to regions of interest can enhance feature representation capabilities. Moreover, large images are computationally expensive, and smaller sizes risk information loss, presenting a computation-accuracy tradeoff. We propose VariViT, an improved ViT model crafted to handle variable image sizes while maintaining a consistent patch size. VariViT employs a novel positional embedding resizing scheme for a variable number of patches. We also implement a new batching strategy within VariViT to reduce computational complexity, resulting in faster training and inference times. In our evaluations on two 3D brain MRI datasets, VariViT surpasses vanilla ViTs and ResNet in glioma genotype prediction and brain tumor classification. It achieves F1-scores of 75.5% and 76.3%, respectively, learning more discriminative features. Our proposed batching strategy reduces computation time by up to 30% compared to conventional architectures. These findings underscore the efficacy of VariViT in image representation learning. Our code can be found here: https://github.com/Aswathi-Varma/varivit

Method: Proposes a multi-stage detection pipeline that processes recontextualized images through specialized steps targeting object-level fidelity, background consistency, and omission detection, using a coordinated ensemble of open-source models.

Result: The approach enables deeper understanding of where models fail with explanations, and extensive experimental evaluations demonstrate the effectiveness of the detection pipeline.

Conclusion: VIGIL fills a gap in the field by providing the first fine-grained categorization and decomposition of hallucinations in multimodal image recontextualization, with open release of dataset, pipeline, and code for transparency and further exploration.

Abstract: We introduce VIGIL (Visual Inconsistency & Generative In-context Lucidity), the first benchmark dataset and framework providing a fine-grained categorization of hallucinations in the multimodal image recontextualization task for large multimodal models (LMMs). While existing research often treats hallucinations as a uniform issue, our work addresses a significant gap in multimodal evaluation by decomposing these errors into five categories: pasted object hallucinations, background hallucinations, object omission, positional & logical inconsistencies, and physical law violations. To address these complexities, we propose a multi-stage detection pipeline. Our architecture processes recontextualized images through a series of specialized steps targeting object-level fidelity, background consistency, and omission detection, leveraging a coordinated ensemble of open-source models, whose effectiveness is demonstrated through extensive experimental evaluations. Our approach enables a deeper understanding of where the models fail with an explanation; thus, we fill a gap in the field, as no prior methods offer such categorization and decomposition for this task. To promote transparency and further exploration, we openly release VIGIL, along with the detection pipeline and benchmark code, through our GitHub repository: https://github.com/mlubneuskaya/vigil and Data repository: https://huggingface.co/datasets/joannaww/VIGIL.

Method: Proposes a modulation-based approach that prioritizes semantic interpretation over strict edge position adherence. Introduces a novel loss function enabling training on freehand sketches without requiring ground-truth pixel-aligned images.

Result: Outperforms existing approaches in both semantic alignment with freehand sketch inputs and in the realism and overall quality of generated images.

Conclusion: The method successfully addresses the challenge of generating photorealistic images from freehand sketches by focusing on semantic interpretation rather than pixel-level alignment, enabling effective training without ground-truth data.

Abstract: Recent years have witnessed remarkable progress in generative AI, with natural language emerging as the most common conditioning input. As underlying models grow more powerful, researchers are exploring increasingly diverse conditioning signals, such as depth maps, edge maps, camera parameters, and reference images, to give users finer control over generation. Among different modalities, sketches are a natural and long-standing form of human communication, enabling rapid expression of visual concepts. Previous literature has largely focused on edge maps, often misnamed ‘sketches’, yet algorithms that effectively handle true freehand sketches, with their inherent abstraction and distortions, remain underexplored. We pursue the challenging goal of balancing photorealism with sketch adherence when generating images from freehand input. A key obstacle is the absence of ground-truth, pixel-aligned images: by their nature, freehand sketches do not have a single correct alignment. To address this, we propose a modulation-based approach that prioritizes semantic interpretation of the sketch over strict adherence to individual edge positions. We further introduce a novel loss that enables training on freehand sketches without requiring ground-truth pixel-aligned images. We show that our method outperforms existing approaches in both semantic alignment with freehand sketch inputs and in the realism and overall quality of the generated images.

Method: Systematic review and taxonomy of three global solver paradigms: Branch-and-Bound (BnB), Convex Relaxation (CR), and Graduated Non-Convexity (GNC). Analysis includes theoretical foundations, algorithmic designs, practical enhancements, and application to ten core vision tasks.

Result: Unified perspective on global solvers revealing optimality-robustness-scalability trade-offs. Identification of critical future directions including scaling algorithms, integrating data-driven priors, establishing benchmarks, and addressing societal implications for safety-critical deployment.

Conclusion: Global solvers provide certifiable solutions for 3D vision geometric optimization. The survey consolidates the field and provides roadmap toward trustworthy perception for real-world applications, with ongoing resources available via GitHub repository.

Abstract: Global solvers have emerged as a powerful paradigm for 3D vision, offering certifiable solutions to nonconvex geometric optimization problems traditionally addressed by local or heuristic methods. This survey presents the first systematic review of global solvers in geometric vision, unifying the field through a comprehensive taxonomy of three core paradigms: Branch-and-Bound (BnB), Convex Relaxation (CR), and Graduated Non-Convexity (GNC). We present their theoretical foundations, algorithmic designs, and practical enhancements for robustness and scalability, examining how each addresses the fundamental nonconvexity of geometric estimation problems. Our analysis spans ten core vision tasks, from Wahba problem to bundle adjustment, revealing the optimality-robustness-scalability trade-offs that govern solver selection. We identify critical future directions: scaling algorithms while maintaining guarantees, integrating data-driven priors with certifiable optimization, establishing standardized benchmarks, and addressing societal implications for safety-critical deployment. By consolidating theoretical foundations, practical advances, and broader impacts, this survey provides a unified perspective and roadmap toward certifiable, trustworthy perception for real-world applications. A continuously-updated literature summary and companion code tutorials are available at https://github.com/ericzzj1989/Awesome-Global-Solvers-for-3D-Vision.

Method: Modified Joint Embedding Predictive Architecture (JEPA) with axial stripe masking strategy to focus on semantically relevant regions, circular loss weighting scheme, and probabilistic reassignment of CLS token for improved linear probing. Trained on consolidated curated dataset.

Result: Outperforms strong baselines like FaRL and Franca on core anthropometric tasks despite using significantly less data. Shows promising results on Body Mass Index (BMI) estimation using novel consolidated closed-source dataset addressing domain bias.

Conclusion: MeFEm provides an effective vision model for facial biometric and medical analysis with improved efficiency and performance, offering a strong baseline for future work in this domain with publicly available model weights.

Abstract: We present MeFEm, a vision model based on a modified Joint Embedding Predictive Architecture (JEPA) for biometric and medical analysis from facial images. Key modifications include an axial stripe masking strategy to focus learning on semantically relevant regions, a circular loss weighting scheme, and the probabilistic reassignment of the CLS token for high quality linear probing. Trained on a consolidated dataset of curated images, MeFEm outperforms strong baselines like FaRL and Franca on core anthropometric tasks despite using significantly less data. It also shows promising results on Body Mass Index (BMI) estimation, evaluated on a novel, consolidated closed-source dataset that addresses the domain bias prevalent in existing data. Model weights are available at https://huggingface.co/boretsyury/MeFEm , offering a strong baseline for future work in this domain.

Method: Generates universal adversarial perturbation (UAP) that embeds semantic targets into images while suppressing original content, misdirecting diffusion models’ attention during editing. Works in data-free settings without training data.

Result: Outperforms baselines in UAP setting, achieves performance comparable to image-specific methods with restricted perturbation budget, shows strong black-box transferability across diffusion models.

Conclusion: Proposes first practical universal immunization framework for diffusion models that balances effectiveness, scalability, and real-world applicability while addressing ethical concerns.

Abstract: Recent advances in diffusion models have enabled powerful image editing capabilities guided by natural language prompts, unlocking new creative possibilities. However, they introduce significant ethical and legal risks, such as deepfakes and unauthorized use of copyrighted visual content. To address these risks, image immunization has emerged as a promising defense against AI-driven semantic manipulation. Yet, most existing approaches rely on image-specific adversarial perturbations that require individual optimization for each image, thereby limiting scalability and practicality. In this paper, we propose the first universal image immunization framework that generates a single, broadly applicable adversarial perturbation specifically designed for diffusion-based editing pipelines. Inspired by universal adversarial perturbation (UAP) techniques used in targeted attacks, our method generates a UAP that embeds a semantic target into images to be protected. Simultaneously, it suppresses original content to effectively misdirect the model’s attention during editing. As a result, our approach effectively blocks malicious editing attempts by overwriting the original semantic content in the image via the UAP. Moreover, our method operates effectively even in data-free settings without requiring access to training data or domain knowledge, further enhancing its practicality and broad applicability in real-world scenarios. Extensive experiments show that our method, as the first universal immunization approach, significantly outperforms several baselines in the UAP setting. In addition, despite the inherent difficulty of universal perturbations, our method also achieves performance on par with image-specific methods under a more restricted perturbation budget, while also exhibiting strong black-box transferability across different diffusion models.

Method: Leverage recent advances in point-track estimation to learn temporal representations and experiment on various perceptual tasks to analyze the properties of long-term motion information.

Result: Three key findings: 1) Long-term motion representations contain comprehensive information about actions, objects, materials, and spatial information, often better than images; 2) They generalize better in low-data and zero-shot settings; 3) Their low dimensionality offers better trade-off between computational cost and accuracy than standard video representations.

Conclusion: Long-term motion information is a powerful and efficient representation for perception that should be leveraged in future model designs, offering advantages over image-based approaches in generalization and computational efficiency.

Abstract: Temporal information has long been considered to be essential for perception. While there is extensive research on the role of image information for perceptual tasks, the role of the temporal dimension remains less well understood: What can we learn about the world from long-term motion information? What properties does long-term motion information have for visual learning? We leverage recent success in point-track estimation, which offers an excellent opportunity to learn temporal representations and experiment on a variety of perceptual tasks. We draw 3 clear lessons: 1) Long-term motion representations contain information to understand actions, but also objects, materials, and spatial information, often even better than images. 2) Long-term motion representations generalize far better than image representations in low-data settings and in zero-shot tasks. 3) The very low dimensionality of motion information makes motion representations a better trade-off between GFLOPs and accuracy than standard video representations, and used together they achieve higher performance than video representations alone. We hope these insights will pave the way for the design of future models that leverage the power of long-term motion information for perception.

Method: CAPA freezes the FM backbone and updates only a minimal set of parameters using Parameter-Efficient Fine-Tuning (like LoRA or VPT), guided by gradients from sparse observations available at inference time. For videos, it introduces sequence-level parameter sharing to jointly adapt all frames.

Result: CAPA achieves state-of-the-art results across diverse condition patterns on both indoor and outdoor datasets, effectively grounding the foundation model’s geometric prior in scene-specific measurements.

Conclusion: CAPA provides a model-agnostic, parameter-efficient framework for test-time adaptation of 3D foundation models that corrects distortions and misplaced structures while being compatible with any ViT-based foundation model.

Abstract: We introduce CAPA, a parameter-efficient test-time optimization framework that adapts pre-trained 3D foundation models (FMs) for depth completion, using sparse geometric cues. Unlike prior methods that train task-specific encoders for auxiliary inputs, which often overfit and generalize poorly, CAPA freezes the FM backbone. Instead, it updates only a minimal set of parameters using Parameter-Efficient Fine-Tuning (e.g. LoRA or VPT), guided by gradients calculated directly from the sparse observations available at inference time. This approach effectively grounds the foundation model’s geometric prior in the scene-specific measurements, correcting distortions and misplaced structures. For videos, CAPA introduces sequence-level parameter sharing, jointly adapting all frames to exploit temporal correlations, improve robustness, and enforce multi-frame consistency. CAPA is model-agnostic, compatible with any ViT-based FM, and achieves state-of-the-art results across diverse condition patterns on both indoor and outdoor datasets. Project page: research.nvidia.com/labs/dvl/projects/capa.

Method: SAILS decouples CISS into two stages: (1) Zero-shot region extraction using Segment Anything Model (SAM), and (2) semantic association through prototypes in a fixed feature space with selective intra-class clustering for multiple prototypes per class.

Result: SAILS typically surpasses training-based approaches on standard CISS datasets, especially in long task sequences where forgetting is severe. It completely eliminates forgetting, maintains consistent performance, and exhibits positive backward transfer.

Conclusion: SAILS provides a training-free solution to continual learning challenges in semantic segmentation, offering superior performance without computational costs or forgetting issues.

Abstract: Continual learning remains constrained by the need for repeated retraining, high computational costs, and the persistent challenge of forgetting. These factors significantly limit the applicability of continuous learning in real-world settings, as iterative model updates require significant computational resources and inherently exacerbate forgetting. We present SAILS – Segment Anything with Incrementally Learned Semantics, a training-free framework for Class-Incremental Semantic Segmentation (CISS) that sidesteps these challenges entirely. SAILS leverages foundational models to decouple CISS into two stages: Zero-shot region extraction using Segment Anything Model (SAM), followed by semantic association through prototypes in a fixed feature space. SAILS incorporates selective intra-class clustering, resulting in multiple prototypes per class to better model intra-class variability. Our results demonstrate that, despite requiring no incremental training, SAILS typically surpasses the performance of existing training-based approaches on standard CISS datasets, particularly in long and challenging task sequences where forgetting tends to be most severe. By avoiding parameter updates, SAILS completely eliminates forgetting and maintains consistent, task-invariant performance. Furthermore, SAILS exhibits positive backward transfer, where the introduction of new classes can enhance performance on previous classes.

Method: Proposes Visual Informative Part Attention (VIPA) framework with visual expression generator (VEG) module. VEG retrieves informative visual tokens via local-global linguistic context cues and refines them to reduce noise while sharing informative visual attributes. This creates visual expressions that provide structural/semantic visual target information.

Result: Extensive experiments show VIPA outperforms existing state-of-the-art methods on four public RIS benchmarks. Visual analysis demonstrates effectiveness in enabling network attention to robustly align with fine-grained regions of interest.

Conclusion: VIPA framework effectively leverages visual informative parts through visual expressions to enhance semantic consistency and reduce cross-modal projection variance, achieving superior performance in referring image segmentation tasks.

Abstract: Referring Image Segmentation (RIS) aims to segment a target object described by a natural language expression. Existing methods have evolved by leveraging the vision information into the language tokens. To more effectively exploit visual contexts for fine-grained segmentation, we propose a novel Visual Informative Part Attention (VIPA) framework for referring image segmentation. VIPA leverages the informative parts of visual contexts, called a visual expression, which can effectively provide the structural and semantic visual target information to the network. This design reduces high-variance cross-modal projection and enhances semantic consistency in an attention mechanism of the referring image segmentation. We also design a visual expression generator (VEG) module, which retrieves informative visual tokens via local-global linguistic context cues and refines the retrieved tokens for reducing noise information and sharing informative visual attributes. This module allows the visual expression to consider comprehensive contexts and capture semantic visual contexts of informative regions. In this way, our framework enables the network’s attention to robustly align with the fine-grained regions of interest. Extensive experiments and visual analysis demonstrate the effectiveness of our approach. Our VIPA outperforms the existing state-of-the-art methods on four public RIS benchmarks.

Method: The authors use Gaze-CIFAR-10 to demonstrate center bias issues, analyze a hard-attention classifier under varying foveal patch sizes and peripheral context, and propose Gaze Consistency Score (GCS) - a center-debiased composite metric augmented with movement similarity.

Result: A trivial center-fixation baseline achieves surprisingly strong scanpath scores approaching learned policies. Analysis reveals a “peripheral sweet spot” where only a narrow range of sensory constraints yields scanpaths that are both above the center baseline after debiasing and temporally human-like. GCS uncovers a robust sweet spot at medium patch size with both foveal and peripheral vision.

Conclusion: The paper highlights the need for center-debiased metrics in evaluating active perception on object-centric datasets and designing gaze benchmarks that better separate behavioral alignment from center bias, revealing important implications for biologically plausible vision models.

Abstract: Human eye movements in visual recognition reflect a balance between foveal sampling and peripheral context. Task-driven hard-attention models for vision are often evaluated by how well their scanpaths match human gaze. However, common scanpath metrics can be strongly confounded by dataset-specific center bias, especially on object-centric datasets. Using Gaze-CIFAR-10, we show that a trivial center-fixation baseline achieves surprisingly strong scanpath scores, approaching many learned policies. This makes standard metrics optimistic and blurs the distinction between genuine behavioral alignment and mere central tendency. We then analyze a hard-attention classifier under constrained vision by sweeping foveal patch size and peripheral context, revealing a peripheral sweet spot: only a narrow range of sensory constraints yields scanpaths that are simultaneously (i) above the center baseline after debiasing and (ii) temporally human-like in movement statistics. To address center bias, we propose GCS (Gaze Consistency Score), a center-debiased composite metric augmented with movement similarity. GCS uncovers a robust sweet spot at medium patch size with both foveal and peripheral vision, that is not obvious from raw scanpath metrics or accuracy alone, and also highlights a “shortcut regime” when the field-of-view becomes too large. We discuss implications for evaluating active perception on object-centric datasets and for designing gaze benchmarks that better separate behavioral alignment from center bias.

Method: Proposes STAformer and STAformer++ architectures with frame-guided temporal pooling, dual image-video attention, multiscale feature fusion. Integrates environment affordance modeling as persistent memory and predicts interaction hotspots from hand/object trajectories.

Result: Significant improvements on Overall Top-5 mAP: up to +23 percentage points on Ego4D and +31 p.p. on curated EPIC-Kitchens STA labels.

Conclusion: The proposed attention-based architectures with affordance grounding substantially improve short-term anticipation performance, with code and annotations released to encourage further research.

Abstract: Short Term object-interaction Anticipation consists in detecting the location of the next active objects, the noun and verb categories of the interaction, as well as the time to contact from the observation of egocentric video. This ability is fundamental for wearable assistants to understand user goals and provide timely assistance, or to enable human-robot interaction. In this work, we present a method to improve the performance of STA predictions. Our contributions are two-fold: 1 We propose STAformer and STAformer plus plus, two novel attention-based architectures integrating frame-guided temporal pooling, dual image-video attention, and multiscale feature fusion to support STA predictions from an image-input video pair; 2 We introduce two novel modules to ground STA predictions on human behavior by modeling affordances. First, we integrate an environment affordance model which acts as a persistent memory of interactions that can take place in a given physical scene. We explore how to integrate environment affordances via simple late fusion and with an approach which adaptively learns how to best fuse affordances with end-to-end predictions. Second, we predict interaction hotspots from the observation of hands and object trajectories, increasing confidence in STA predictions localized around the hotspot. Our results show significant improvements on Overall Top-5 mAP, with gain up to +23p.p on Ego4D and +31p.p on a novel set of curated EPIC-Kitchens STA labels. We released the code, annotations, and pre-extracted affordances on Ego4D and EPIC-Kitchens to encourage future research in this area.

Method: Treat image samples as simplicial complexes, use persistent sheaf Laplacians to extract multiscale localized topological spectral representations for individual images, then aggregate statistical summaries of spectra across scales and dimensions to form multiscale multi-dimensional image representations.

Result: The method shows more stable performance across a wide range of reduced dimensions and achieves consistent improvements over PCA-based baselines in moderate dimensional regimes on COIL20 and ETH80 image datasets.

Conclusion: The multi-dimensional persistent sheaf Laplacian framework provides a robust alternative to traditional dimensionality reduction methods by leveraging topological information across multiple dimensions for improved image analysis.

Abstract: We propose a multi-dimensional persistent sheaf Laplacian (MPSL) framework on simplicial complexes for image analysis. The proposed method is motivated by the strong sensitivity of commonly used dimensionality reduction techniques, such as principal component analysis (PCA), to the choice of reduced dimension. Rather than selecting a single reduced dimension or averaging results across dimensions, we exploit complementary advantages of multiple reduced dimensions. At a given dimension, image samples are regarded as simplicial complexes, and persistent sheaf Laplacians are utilized to extract a multiscale localized topological spectral representation for individual image samples. Statistical summaries of the resulting spectra are then aggregated across scales and dimensions to form multiscale multi-dimensional image representations. We evaluate the proposed framework on the COIL20 and ETH80 image datasets using standard classification protocols. Experimental results show that the proposed method provides more stable performance across a wide range of reduced dimensions and achieves consistent improvements to PCA-based baselines in moderate dimensional regimes.

Method: Created CT-Bench with two components: 1) Lesion Image and Metadata Set (20,335 lesions from 7,795 CT studies with bounding boxes, descriptions, size info), and 2) multitask visual question answering benchmark (2,850 QA pairs covering lesion tasks). Includes hard negative examples for real-world challenges.

Result: Evaluated state-of-the-art multimodal models (vision-language and medical CLIP variants) against radiologist assessments. Fine-tuning on the Lesion Image and Metadata Set yielded significant performance gains across both benchmark components.

Conclusion: CT-Bench serves as a comprehensive benchmark for lesion analysis, demonstrating clinical utility through performance improvements when models are fine-tuned on the dataset.

Abstract: Artificial intelligence (AI) can automatically delineate lesions on computed tomography (CT) and generate radiology report content, yet progress is limited by the scarcity of publicly available CT datasets with lesion-level annotations. To bridge this gap, we introduce CT-Bench, a first-of-its-kind benchmark dataset comprising two components: a Lesion Image and Metadata Set containing 20,335 lesions from 7,795 CT studies with bounding boxes, descriptions, and size information, and a multitask visual question answering benchmark with 2,850 QA pairs covering lesion localization, description, size estimation, and attribute categorization. Hard negative examples are included to reflect real-world diagnostic challenges. We evaluate multiple state-of-the-art multimodal models, including vision-language and medical CLIP variants, by comparing their performance to radiologist assessments, demonstrating the value of CT-Bench as a comprehensive benchmark for lesion analysis. Moreover, fine-tuning models on the Lesion Image and Metadata Set yields significant performance gains across both components, underscoring the clinical utility of CT-Bench.

Method: Combines SfM reconstruction, 3D Gaussian Splatting, semantic grounding, and monocular depth-based metric cues to produce stable zenith-view renderings that can be directly matched to satellite imagery. Also introduces MC-Sat dataset for evaluation.

Result: Achieves sub-30m geolocation accuracy across both dense and large-area scenes in zero-shot experiments, demonstrating the promise of geometry-based aggregation for robust ground-to-satellite localization.

Conclusion: Wrivinder and MC-Sat provide the first comprehensive baseline and testbed for studying geometry-centered cross-view alignment without paired supervision, enabling more robust localization in GPS-challenged environments.

Abstract: Aligning ground-level imagery with geo-registered satellite maps is crucial for mapping, navigation, and situational awareness, yet remains challenging under large viewpoint gaps or when GPS is unreliable. We introduce Wrivinder, a zero-shot, geometry-driven framework that aggregates multiple ground photographs to reconstruct a consistent 3D scene and align it with overhead satellite imagery. Wrivinder combines SfM reconstruction, 3D Gaussian Splatting, semantic grounding, and monocular depth–based metric cues to produce a stable zenith-view rendering that can be directly matched to satellite context for metrically accurate camera geo-localization. To support systematic evaluation of this task, which lacks suitable benchmarks, we also release MC-Sat, a curated dataset linking multi-view ground imagery with geo-registered satellite tiles across diverse outdoor environments. Together, Wrivinder and MC-Sat provide a first comprehensive baseline and testbed for studying geometry-centered cross-view alignment without paired supervision. In zero-shot experiments, Wrivinder achieves sub-30,m geolocation accuracy across both dense and large-area scenes, highlighting the promise of geometry-based aggregation for robust ground-to-satellite localization.

Method: AnchorWeave replaces single misaligned global memory with multiple clean local geometric memories, performs coverage-driven local memory retrieval aligned with target trajectory, and integrates selected memories through a multi-anchor weaving controller during generation.

Result: Extensive experiments show AnchorWeave significantly improves long-term scene consistency while maintaining strong visual quality, with ablation studies validating effectiveness of local geometric conditioning, multi-anchor control, and coverage-driven retrieval.

Conclusion: AnchorWeave addresses the challenge of maintaining spatial world consistency over long horizons in video generation by using multiple local geometric memories to reconcile cross-view inconsistencies, outperforming global reconstruction approaches.

Abstract: Maintaining spatial world consistency over long horizons remains a central challenge for camera-controllable video generation. Existing memory-based approaches often condition generation on globally reconstructed 3D scenes by rendering anchor videos from the reconstructed geometry in the history. However, reconstructing a global 3D scene from multiple views inevitably introduces cross-view misalignment, as pose and depth estimation errors cause the same surfaces to be reconstructed at slightly different 3D locations across views. When fused, these inconsistencies accumulate into noisy geometry that contaminates the conditioning signals and degrades generation quality. We introduce AnchorWeave, a memory-augmented video generation framework that replaces a single misaligned global memory with multiple clean local geometric memories and learns to reconcile their cross-view inconsistencies. To this end, AnchorWeave performs coverage-driven local memory retrieval aligned with the target trajectory and integrates the selected local memories through a multi-anchor weaving controller during generation. Extensive experiments demonstrate that AnchorWeave significantly improves long-term scene consistency while maintaining strong visual quality, with ablation and analysis studies further validating the effectiveness of local geometric conditioning, multi-anchor control, and coverage-driven retrieval.

Method: A part-centric generative framework that models objects as sets of movable parts, each encoded by latent tokens augmented with part identity and articulation cues. Conditioned on a single image, it generates articulated 3D assets with instance-level correspondence while maintaining valid part structure and motion.

Result: The approach shows improved input consistency, part accuracy, and articulation plausibility over optimization-based and retrieval-driven baselines on common articulated categories (drawers, doors), while substantially reducing inference time compared to optimization methods.

Conclusion: The framework enables fast feed-forward generation of articulated 3D assets from single images, avoiding per-instance optimization while supporting controllable assembly and articulation for embodied interaction applications.

Abstract: Articulated objects are central to interactive 3D applications, including embodied AI, robotics, and VR/AR, where functional part decomposition and kinematic motion are essential. Yet producing high-fidelity articulated assets remains difficult to scale because it requires reliable part decomposition and kinematic rigging. Existing approaches largely fall into two paradigms: optimization-based reconstruction or distillation, which can be accurate but often takes tens of minutes to hours per instance, and inference-time methods that rely on template or part retrieval, producing plausible results that may not match the specific structure and appearance in the input observation. We introduce a part-centric generative framework for articulated object creation that synthesizes part geometry, composition, and articulation under explicit part-aware conditioning. Our representation models an object as a set of movable parts, each encoded by latent tokens augmented with part identity and articulation cues. Conditioned on a single image, the model generates articulated 3D assets that preserve instance-level correspondence while maintaining valid part structure and motion. The resulting approach avoids per-instance optimization, enables fast feed-forward inference, and supports controllable assembly and articulation, which are important for embodied interaction. Experiments on common articulated categories (e.g., drawers and doors) show improved input consistency, part accuracy, and articulation plausibility over optimization-based and retrieval-driven baselines, while substantially reducing inference time.

Method: Introduces ThermEval-B, a structured benchmark of ~55,000 thermal visual question answering pairs. Combines public datasets with newly collected ThermEval-D, which provides dense per-pixel temperature maps with semantic body-part annotations across diverse indoor/outdoor environments. Evaluates 25 open-source and closed-source VLMs.

Result: Models consistently fail at temperature-grounded reasoning, degrade under colormap transformations, and default to language priors or fixed responses. Only marginal gains from prompting or supervised fine-tuning. Demonstrates thermal understanding requires dedicated evaluation beyond RGB-centric assumptions.

Conclusion: Thermal vision language understanding requires specialized evaluation that goes beyond RGB-centric approaches. ThermEval serves as a benchmark to drive progress in thermal vision language modeling by addressing the unique challenges of temperature-based perception and reasoning.

Abstract: Vision language models (VLMs) achieve strong performance on RGB imagery, but they do not generalize to thermal images. Thermal sensing plays a critical role in settings where visible light fails, including nighttime surveillance, search and rescue, autonomous driving, and medical screening. Unlike RGB imagery, thermal images encode physical temperature rather than color or texture, requiring perceptual and reasoning capabilities that existing RGB-centric benchmarks do not evaluate. We introduce ThermEval-B, a structured benchmark of approximately 55,000 thermal visual question answering pairs designed to assess the foundational primitives required for thermal vision language understanding. ThermEval-B integrates public datasets with our newly collected ThermEval-D, the first dataset to provide dense per-pixel temperature maps with semantic body-part annotations across diverse indoor and outdoor environments. Evaluating 25 open-source and closed-source VLMs, we find that models consistently fail at temperature-grounded reasoning, degrade under colormap transformations, and default to language priors or fixed responses, with only marginal gains from prompting or supervised fine-tuning. These results demonstrate that thermal understanding requires dedicated evaluation beyond RGB-centric assumptions, positioning ThermEval as a benchmark to drive progress in thermal vision language modeling.

Method: Learns an encoder that maps natural images uniformly onto a spherical latent space and a decoder that maps random latent vectors back to image space, trained solely through image reconstruction losses. Supports conditional generation and iterative refinement through encoder/decoder loops.

Result: Achieves performance competitive with state-of-the-art diffusion models across several datasets while requiring only a small fraction of the inference cost (fewer than five steps vs. many-step diffusion).

Conclusion: Sphere Encoder provides an efficient generative framework that can produce high-quality images in a single forward pass, offering a compelling alternative to computationally expensive diffusion models.

Abstract: We introduce the Sphere Encoder, an efficient generative framework capable of producing images in a single forward pass and competing with many-step diffusion models using fewer than five steps. Our approach works by learning an encoder that maps natural images uniformly onto a spherical latent space, and a decoder that maps random latent vectors back to the image space. Trained solely through image reconstruction losses, the model generates an image by simply decoding a random point on the sphere. Our architecture naturally supports conditional generation, and looping the encoder/decoder a few times can further enhance image quality. Across several datasets, the sphere encoder approach yields performance competitive with state of the art diffusions, but with a small fraction of the inference cost. Project page is available at https://sphere-encoder.github.io .

Method: Created RAVENEA benchmark with two tasks: culture-focused visual question answering (cVQA) and culture-informed image captioning (cIC), integrating over 11,396 unique Wikipedia documents curated by human annotators. Evaluated seven multimodal retrievers and fifteen VLMs.

Result: Cultural grounding annotations enhance multimodal retrieval and downstream tasks; VLMs with culture-aware retrieval outperform non-augmented counterparts (+6% on cVQA, +11% on cIC); performance varies widely across countries.

Conclusion: RAVENEA highlights limitations of current multimodal retrievers and VLMs, showing the need to enhance visual culture understanding within RAG systems, and provides a valuable resource for advancing research in this area.

Abstract: As vision-language models (VLMs) become increasingly integrated into daily life, the need for accurate visual culture understanding is becoming critical. Yet, these models frequently fall short in interpreting cultural nuances effectively. Prior work has demonstrated the effectiveness of retrieval-augmented generation (RAG) in enhancing cultural understanding in text-only settings, while its application in multimodal scenarios remains underexplored. To bridge this gap, we introduce RAVENEA (Retrieval-Augmented Visual culturE uNdErstAnding), a new benchmark designed to advance visual culture understanding through retrieval, focusing on two tasks: culture-focused visual question answering (cVQA) and culture-informed image captioning (cIC). RAVENEA extends existing datasets by integrating over 11,396 unique Wikipedia documents curated and ranked by human annotators. Through the extensive evaluation on seven multimodal retrievers and fifteen VLMs, RAVENEA reveals some undiscovered findings: (i) In general, cultural grounding annotations can enhance multimodal retrieval and corresponding downstream tasks. (ii) VLMs, when augmented with culture-aware retrieval, generally outperform their non-augmented counterparts (by averaging +6% on cVQA and +11% on cIC). (iii) Performance of culture-aware retrieval augmented varies widely across countries. These findings highlight the limitations of current multimodal retrievers and VLMs, underscoring the need to enhance visual culture understanding within RAG systems. We believe RAVENEA offers a valuable resource for advancing research on retrieval-augmented visual culture understanding.

Method: Introduces a local video context module operating only on masked tokens (cost proportional to edit size), guided by a lightweight temporal global context embedder for video-wide consistency with minimal overhead.

Result: EditCtrl is 10x more compute efficient than state-of-the-art methods while improving editing quality. Enables new capabilities like multi-region editing with text prompts and autoregressive content propagation.

Conclusion: The framework demonstrates that focusing computation on edited regions rather than full video context enables both efficiency gains and quality improvements in video inpainting.

Abstract: High-fidelity generative video editing has seen significant quality improvements by leveraging pre-trained video foundation models. However, their computational cost is a major bottleneck, as they are often designed to inefficiently process the full video context regardless of the inpainting mask’s size, even for sparse, localized edits. In this paper, we introduce EditCtrl, an efficient video inpainting control framework that focuses computation only where it is needed. Our approach features a novel local video context module that operates solely on masked tokens, yielding a computational cost proportional to the edit size. This local-first generation is then guided by a lightweight temporal global context embedder that ensures video-wide context consistency with minimal overhead. Not only is EditCtrl 10 times more compute efficient than state-of-the-art generative editing methods, it even improves editing quality compared to methods designed with full-attention. Finally, we showcase how EditCtrl unlocks new capabilities, including multi-region editing with text prompts and autoregressive content propagation.

Method: Developed a novel deep learning system that uses only three ECG leads to identify multiple cardiovascular abnormalities.

Result: The system achieves accurate detection of cardiovascular abnormalities with reduced lead requirements.

Conclusion: Three-lead ECG systems can enable more accessible cardiovascular monitoring through portable/wearable devices while maintaining diagnostic accuracy.

Abstract: Cardiovascular diseases (CVDs) are a group of heart and blood vessel disorders that is one of the most serious dangers to human health, and the number of such patients is still growing. Early and accurate detection plays a key role in successful treatment and intervention. Electrocardiogram (ECG) is the gold standard for identifying a variety of cardiovascular abnormalities. In clinical practices and most of the current research, standard 12-lead ECG is mainly used. However, using a lower number of leads can make ECG more prevalent as it can be conveniently recorded by portable or wearable devices. In this research, we develop a novel deep learning system to accurately identify multiple cardiovascular abnormalities by using only three ECG leads.

Method: Integrate differentiable Thin-Plate-Spline (TPS) modules into end-to-end GAN framework to establish geometric correlations between portraits and style samples by aligning facial characteristics using facial landmarks, working at global and local scales in both pixel and feature spaces.

Result: Outperforms existing models in fidelity and stylistic consistency, achieves 2x training data efficiency, 100x less computational complexity, and real-time inference (30 FPS) at 512*512 resolution on mobile devices.

Conclusion: The proposed TPS-enhanced GAN framework effectively addresses geometric consistency challenges in portrait stylization, enabling efficient training and real-time mobile deployment while maintaining high-quality artistic effects.

Abstract: Portrait Stylization aims to imbue portrait photos with vivid artistic effects drawn from style examples. Despite the availability of enormous training datasets and large network weights, existing methods struggle to maintain geometric consistency and achieve satisfactory stylization effects due to the disparity in facial feature distributions between facial photographs and stylized images, limiting the application on rare styles and mobile devices. To alleviate this, we propose to establish meaningful geometric correlations between portraits and style samples to simplify the stylization by aligning corresponding facial characteristics. Specifically, we integrate differentiable Thin-Plate-Spline (TPS) modules into an end-to-end Generative Adversarial Network (GAN) framework to improve the training efficiency and promote the consistency of facial identities. By leveraging inherent structural information of faces, e.g., facial landmarks, TPS module can establish geometric alignments between the two domains, at global and local scales, both in pixel and feature spaces, thereby overcoming the aforementioned challenges. Quantitative and qualitative comparisons on a range of portrait stylization tasks demonstrate that our models not only outperforms existing models in terms of fidelity and stylistic consistency, but also achieves remarkable improvements in 2x training data efficiency and 100x less computational complexity, allowing our lightweight model to achieve real-time inference (30 FPS) at 512*512 resolution on mobile devices.

Method: TKN extracts dynamic content from video frames in an unsupervised manner to reduce redundant feature computation. It uses an acceleration matrix to reduce computational cost of attention mechanisms and employs a parallel computing structure for prediction acceleration.

Result: TKN achieves a prediction rate of 1,176 fps, significantly reducing computation costs while maintaining performance. Qualitative and quantitative experiments on multiple datasets demonstrate its superiority.

Conclusion: TKN is the first real-time video prediction solution with high speed and reduced computation costs, showing great application potential for real-time scenarios.

Abstract: Video prediction is a complex time-series forecasting task with great potential in many use cases. However, traditional methods prioritize accuracy and overlook slow prediction speeds due to complex model structures, redundant information, and excessive GPU memory consumption. These methods often predict frames sequentially, making acceleration difficult and limiting their applicability in real-time scenarios like danger prediction and warning.Therefore, we propose a transformer-based keypoint prediction neural network (TKN). TKN extracts dynamic content from video frames in an unsupervised manner, reducing redundant feature computation. And, TKN uses an acceleration matrix to reduce the computational cost of attention and employs a parallel computing structure for prediction acceleration. To the best of our knowledge, TKN is the first real-time video prediction solution that achieves a prediction rate of 1,176 fps, significantly reducing computation costs while maintaining other performance. Qualitative and quantitative experiments on multiple datasets have demonstrated the superiority of our method, suggesting that TKN has great application potential.

Method: Comprehensive survey of 230+ papers with two novel taxonomies: one for GM-DC tasks (unconditional/conditional generation, cross-domain adaptation, subject-driven modeling) and another for methodological approaches (transfer learning, data augmentation, meta-learning, frequency-aware modeling).

Result: Systematic analysis of GM-DC field including task-approach-method interactions via Sankey diagram, identification of key challenges (overfitting, frequency bias, knowledge transfer), and comprehensive review across model types and constraint scenarios.

Conclusion: Provides practical roadmap for researchers/practitioners with future directions including foundation model adaptation, holistic evaluation frameworks, and data-centric sample selection strategies.

Abstract: Generative modeling in machine learning aims to synthesize new data samples that are statistically similar to those observed during training. While conventional generative models such as GANs and diffusion models typically assume access to large and diverse datasets, many real-world applications (e.g. in medicine, satellite imaging, and artistic domains) operate under limited data availability and strict constraints. In this survey, we examine Generative Modeling under Data Constraint (GM-DC), which includes limited-data, few-shot, and zero-shot settings. We present a unified perspective on the key challenges in GM-DC, including overfitting, frequency bias, and incompatible knowledge transfer, and discuss how these issues impact model performance. To systematically analyze this growing field, we introduce two novel taxonomies: one categorizing GM-DC tasks (e.g. unconditional vs. conditional generation, cross-domain adaptation, and subject-driven modeling), and another organizing methodological approaches (e.g. transfer learning, data augmentation, meta-learning, and frequency-aware modeling). Our study reviews over 230 papers, offering a comprehensive view across generative model types and constraint scenarios. We further analyze task-approach-method interactions using a Sankey diagram and highlight promising directions for future work, including adaptation of foundation models, holistic evaluation frameworks, and data-centric strategies for sample selection. This survey provides a timely and practical roadmap for researchers and practitioners aiming to advance generative modeling under limited data. Project website: https://sutd-visual-computing-group.github.io/gmdc-survey/.

Method: Uses Personalized eXplicitly Assembled Normalization to filter domain-specific features while retaining discrimination, plus a regularizer to guide models in capturing domain-invariant representations for the global classifier.

Result: Outperforms existing methods on benchmark datasets (PACS, Office-Home) and real-world medical dataset (Camelyon17).

Conclusion: gPerXAN effectively addresses federated domain generalization with better privacy preservation and lower costs than previous approaches.

Abstract: Domain shift is a formidable issue in Machine Learning that causes a model to suffer from performance degradation when tested on unseen domains. Federated Domain Generalization (FedDG) attempts to train a global model using collaborative clients in a privacy-preserving manner that can generalize well to unseen clients possibly with domain shift. However, most existing FedDG methods either cause additional privacy risks of data leakage or induce significant costs in client communication and computation, which are major concerns in the Federated Learning paradigm. To circumvent these challenges, here we introduce a novel architectural method for FedDG, namely gPerXAN, which relies on a normalization scheme working with a guiding regularizer. In particular, we carefully design Personalized eXplicitly Assembled Normalization to enforce client models selectively filtering domain-specific features that are biased towards local data while retaining discrimination of those features. Then, we incorporate a simple yet effective regularizer to guide these models in directly capturing domain-invariant representations that the global model’s classifier can leverage. Extensive experimental results on two benchmark datasets, i.e., PACS and Office-Home, and a real-world medical dataset, Camelyon17, indicate that our proposed method outperforms other existing methods in addressing this particular problem.

Method: Proposes Story-Iter with a plug-and-play, training-free global reference cross-attention (GRCA) module that models all reference frames with global embeddings, enabling iterative refinement by incorporating all previous round’s reference images.

Result: Achieves state-of-the-art performance on official story visualization datasets and long story benchmarks (up to 100 frames), excelling in both semantic consistency and fine-grained interactions.

Conclusion: The external iterative paradigm with GRCA enables precise long-story generation with improved consistency and interactions without requiring training.

Abstract: This paper introduces Story-Iter, a new training-free iterative paradigm to enhance long-story generation. Unlike existing methods that rely on fixed reference images to construct a complete story, our approach features a novel external iterative paradigm, extending beyond the internal iterative denoising steps of diffusion models, to continuously refine each generated image by incorporating all reference images from the previous round. To achieve this, we propose a plug-and-play, training-free global reference cross-attention (GRCA) module, modeling all reference frames with global embeddings, ensuring semantic consistency in long sequences. By progressively incorporating holistic visual context and text constraints, our iterative paradigm enables precise generation with fine-grained interactions, optimizing the story visualization step-by-step. Extensive experiments in the official story visualization dataset and our long story benchmark demonstrate that Story-Iter’s state-of-the-art performance in long-story visualization (up to 100 frames) excels in both semantic consistency and fine-grained interactions.

Method: Proposes Region-to-Image Distillation: 1) Zoom into micro-cropped regions using strong teacher models to generate high-quality VQA data, 2) Distill this region-grounded supervision back to the full image, 3) Train student models on this data to internalize zooming benefits into single forward pass. Also introduces ZoomBench benchmark with 845 VQA data across six fine-grained perceptual dimensions.

Result: Models achieve leading performance across multiple fine-grained perception benchmarks, improve general multimodal cognition on visual reasoning and GUI agent benchmarks, and provide insights on when “Thinking-with-Images” is necessary vs when gains can be distilled.

Conclusion: Region-to-Image Distillation effectively internalizes the benefits of agentic zooming into MLLMs, enabling efficient fine-grained perception without inference-time tool use, while maintaining strong performance on general multimodal tasks.

Abstract: Multimodal Large Language Models (MLLMs) excel at broad visual understanding but still struggle with fine-grained perception, where decisive evidence is small and easily overwhelmed by global context. Recent “Thinking-with-Images” methods alleviate this by iteratively zooming in and out regions of interest during inference, but incur high latency due to repeated tool calls and visual re-encoding. To address this, we propose Region-to-Image Distillation, which transforms zooming from an inference-time tool into a training-time primitive, thereby internalizing the benefits of agentic zooming into a single forward pass of an MLLM. In particular, we first zoom in to micro-cropped regions to let strong teacher models generate high-quality VQA data, and then distill this region-grounded supervision back to the full image. After training on such data, the smaller student model improves “single-glance” fine-grained perception without tool use. To rigorously evaluate this capability, we further present ZoomBench, a hybrid-annotated benchmark of 845 VQA data spanning six fine-grained perceptual dimensions, together with a dual-view protocol that quantifies the global–regional “zooming gap”. Experiments show that our models achieve leading performance across multiple fine-grained perception benchmarks, and also improve general multimodal cognition on benchmarks such as visual reasoning and GUI agents. We further discuss when “Thinking-with-Images” is necessary versus when its gains can be distilled into a single forward pass. Our code is available at https://github.com/inclusionAI/Zooming-without-Zooming.

Method: Proposes a divide-and-conquer baseline method that decomposes counting problems into sub-tasks. Includes a mechanism to prevent object splitting during division to avoid repetitive counting issues common in naive implementations.

Result: The approach demonstrates effectiveness across various datasets and benchmarks, establishing it as a valuable reference for evaluating future counting solutions in LVLMs.

Conclusion: While LVLMs struggle with counting, especially for large object numbers, a simple divide-and-conquer approach with anti-splitting mechanisms can significantly improve their counting capabilities, providing a baseline for future research.

Abstract: Counting is a fundamental operation for various real-world visual tasks, requiring both object recognition and robust counting capabilities. Despite their advanced visual perception, large vision-language models (LVLMs) are known to struggle with counting tasks. In this work, we evaluate the performance of several LVLMs on visual counting tasks across multiple counting and vision datasets. We observe that while their performance may be less prone to error for small numbers of objects, they exhibit significant weaknesses as the number of objects increases. To alleviate this issue, we propose a simple yet effective baseline method that enhances LVLMs’ counting ability for large numbers of objects using a divide-and-conquer approach. Our method decomposes counting problems into sub-tasks. Moreover, it incorporates a mechanism to prevent objects from being split during division, which could otherwise lead to repetitive counting – a common issue in a naive divide-and-conquer implementation. We demonstrate the effectiveness of this approach across various datasets and benchmarks, establishing it as a valuable reference for evaluating future solutions.

Method: Evaluated popular vision foundation models using three conformal prediction methods across vision classification benchmarks. Tested Vision Transformers, calibrated confidence predictions, few-shot adaptation of VLMs, and examined APS method performance.

Result: Foundation models are well-suited for conformalization, especially Vision Transformers. Calibration degrades efficiency of conformal sets. Few-shot adaptation improves VLMs over zero-shot. APS maintains coverage guarantees in realistic scenarios.

Conclusion: Vision foundation models can be effectively integrated with conformal prediction for uncertainty quantification, with specific recommendations for model selection and adaptation strategies to ensure safe deployment in high-stakes applications.

Abstract: Recent advances in self-supervision and contrastive learning have brought the performance of foundation models to unprecedented levels in a variety of tasks. Fueled by this progress, these models are becoming the prevailing approach for a wide array of real-world vision problems, including risk-sensitive and high-stakes applications. However, ensuring safe deployment in these scenarios requires a more comprehensive understanding of their uncertainty modeling capabilities, which has received little attention. In this work, we delve into the behaviour of vision and vision-language foundation models under Conformal Prediction (CP), a statistical framework that provides theoretical guarantees of marginal coverage of the true class. Across extensive experiments including popular vision classification benchmarks, well-known foundation vision models, and three CP methods, our findings reveal that foundation models are well-suited for conformalization procedures, particularly those integrating Vision Transformers. We also show that calibrating the confidence predictions of these models, a popular strategy to improve their uncertainty quantification, actually leads to efficiency degradation of the conformal set on adaptive CP methods. Furthermore, few-shot adaptation of Vision-Language Models (VLMs) to downstream tasks, whose popularity is surging, enhances conformal scores compared to zero-shot predictions. Last, our empirical study exposes APS as particularly promising in the context of vision foundation models, as it does not violate the marginal coverage guarantees across multiple challenging, yet realistic scenarios.

Method: Proposes StrandHead that distills 2D generative models pre-trained on human mesh data to generate realistic hair strands from prompts. Uses strand-guided meshing to ensure gradient flow from distillation objective to neural strand representation, with regularization using statistically significant haircut features for stable optimization.

Result: Achieves state-of-the-art performance on text-to-strand generation and disentangled 3D head avatar modeling. Generated 3D hair can be applied to avatars for strand-level editing and implemented in graphics engines for physical simulation.

Conclusion: StrandHead enables realistic 3D hair strand generation from text using 2D/3D human-centric priors without large-scale hair-text paired data, supporting strand-level editing and graphics applications.

Abstract: While haircut indicates distinct personality, existing avatar generation methods fail to model practical hair due to the data limitation or entangled representation. We propose StrandHead, a novel text-driven method capable of generating 3D hair strands and disentangled head avatars with strand-level attributes. Instead of using large-scale hair-text paired data for supervision, we demonstrate that realistic hair strands can be generated from prompts by distilling 2D generative models pre-trained on human mesh data. To this end, we propose a meshing approach guided by strand geometry to guarantee the gradient flow from the distillation objective to the neural strand representation. The optimization is then regularized by statistically significant haircut features, leading to stable updating of strands against unreasonable drifting. These employed 2D/3D human-centric priors contribute to text-aligned and realistic 3D strand generation. Extensive experiments show that StrandHead achieves the state-of-the-art performance on text to strand generation and disentangled 3D head avatar modeling. The generated 3D hair can be applied on avatars for strand-level editing, as well as implemented in the graphics engine for physical simulation or other applications. Project page: https://xiaokunsun.github.io/StrandHead.github.io/.

Method: Proposes TRecViT with time-space-channel factorization: uses gated linear recurrent units (LRUs) for temporal information mixing, self-attention layers for spatial mixing, and MLPs for channel mixing. The architecture is causal and operates in supervised or self-supervised regimes.

Result: Outperforms or matches ViViT-L on SSv2 and Kinetics400 with 3× fewer parameters, 12× smaller memory footprint, 5× lower FLOPs, and 300 FPS inference. Achieves SOTA on SSv2 compared to causal transformer models (TSM, RViT) and recurrent models like LSTM.

Conclusion: TRecViT demonstrates that causal video modeling can achieve competitive performance with non-causal models while being significantly more efficient, enabling real-time video understanding applications.

Abstract: We propose a novel block for \emph{causal} video modelling. It relies on a time-space-channel factorisation with dedicated blocks for each dimension: gated linear recurrent units (LRUs) perform information mixing over time, self-attention layers perform mixing over space, and MLPs over channels. The resulting architecture \emph{TRecViT} is causal and shows strong performance on sparse and dense tasks, trained in supervised or self-supervised regimes, being the first causal video model in the state-space models family. Notably, our model outperforms or is on par with the popular (non-causal) ViViT-L model on large scale video datasets (SSv2, Kinetics400), while having $3\times$ less parameters, $12\times$ smaller memory footprint, and $5\times$ lower FLOPs count than the full self-attention ViViT, with an inference throughput of about 300 frames per second, running comfortably in real-time. When compared with causal transformer-based models (TSM, RViT) and other recurrent models like LSTM, TRecViT obtains state-of-the-art results on the challenging SSv2 dataset. Code and checkpoints are available online https://github.com/google-deepmind/trecvit.

Method: Cross-Modal Mapping (CMM) with global linear transformation to align image features to text space, plus local triplet loss to optimize spatial relationships

Result: Improves average Top-1 accuracy by 1.06% on 11 benchmark datasets, performs well on distribution shift datasets, simplifies training with higher efficiency

Conclusion: CMM effectively mitigates modality gap in pre-trained models, enabling text features to serve as effective class prototypes for efficient and generalizable few-shot learning

Abstract: Few-shot image classification remains a critical challenge in the field of computer vision, particularly in data-scarce environments. Existing methods typically rely on pre-trained visual-language models, such as CLIP. However, due to the modality gap, which is the inconsistent distribution of image and text features in the joint embedding space, directly using these features as class prototypes often leads to suboptimal performance. To address this issue, we propose a novel Cross-Modal Mapping (CMM) method. This method globally aligns image features with the text feature space through linear transformation and optimizes their local spatial relationships using triplet loss, thereby significantly enhancing cross-modal consistency. Experimental results show that compared to other methods, CMM simplifies the training process and demonstrates higher efficiency. Furthermore, CMM improves the average Top-1 accuracy by 1.06% on 11 benchmark datasets compared to methods that partially fine-tune the backbone, and it performs excellently on 4 distribution shift datasets. Notably, CMM effectively mitigates the modality gap in pre-trained models, enabling text features to serve as effective class prototypes for image features, thus providing an efficient and highly generalizable solution for few-shot learning.

Method: Proposes Committee Voting for Dataset Distillation (CV-DD) that integrates distributions and predictions from multiple models, generates high-quality soft labels, and uses voting-based strategy to enhance diversity and robustness.

Result: CV-DD consistently outperforms single- and multi-model distillation methods across multiple datasets and IPC settings, and generalizes well to non-training-based frameworks and synthetic-to-real transfer tasks.

Conclusion: The committee voting approach effectively captures broader data characteristics, reduces model bias, improves generalization, and demonstrates strong performance across various settings.

Abstract: Dataset distillation aims to synthesize a compact yet representative dataset that preserves the essential characteristics of the original data for efficient model training. Existing methods mainly focus on improving data-synthetic alignment or scaling distillation to large datasets. In this work, we propose $\textbf{C}$ommittee $\textbf{V}$oting for $\textbf{D}$ataset $\textbf{D}$istillation ($\textbf{CV-DD}$), an orthogonal approach that leverages the collective knowledge of multiple models to produce higher-quality distilled data. We first establish a strong baseline that achieves state-of-the-art performance through modern architectural and optimization choices. By integrating distributions and predictions from multiple models and generating high-quality soft labels, our method captures a broader range of data characteristics, reduces model-specific bias and the impact of distribution shifts, and significantly improves generalization. This voting-based strategy enhances diversity and robustness, alleviates overfitting, and improves post-evaluation performance. Extensive experiments across multiple datasets and IPC settings demonstrate that CV-DD consistently outperforms single- and multi-model distillation methods and generalizes well to non-training-based frameworks and challenging synthetic-to-real transfer tasks. Code is available at: https://github.com/Jiacheng8/CV-DD.

Method: Proposes V2V-LLM, a baseline method using Large Language Models to fuse perception information from multiple connected autonomous vehicles (CAVs) via V2V communication. Creates the V2V-QA dataset and benchmark for cooperative driving tasks including grounding, notable object identification, and planning.

Result: Experimental results show V2V-LLM outperforms other baseline methods with different fusion approaches, demonstrating it as a promising unified model architecture for various cooperative autonomous driving tasks.

Conclusion: The work creates a new research direction for improving autonomous driving safety through multimodal LLM-based cooperative systems, with plans to release code and data to facilitate open-source research.

Abstract: Current autonomous driving vehicles rely mainly on their individual sensors to understand surrounding scenes and plan for future trajectories, which can be unreliable when the sensors are malfunctioning or occluded. To address this problem, cooperative perception methods via vehicle-to-vehicle (V2V) communication have been proposed, but they have tended to focus on perception tasks like detection or tracking. How those approaches contribute to overall cooperative planning performance is still under-explored. Inspired by recent progress using Large Language Models (LLMs) to build autonomous driving systems, we propose a novel problem setting that integrates a Multimodal LLM into cooperative autonomous driving, with the proposed Vehicle-to-Vehicle Question-Answering (V2V-QA) dataset and benchmark. We also propose our baseline method Vehicle-to-Vehicle Multimodal Large Language Model (V2V-LLM), which uses an LLM to fuse perception information from multiple connected autonomous vehicles (CAVs) and answer various types of driving-related questions: grounding, notable object identification, and planning. Experimental results show that our proposed V2V-LLM can be a promising unified model architecture for performing various tasks in cooperative autonomous driving, and outperforms other baseline methods that use different fusion approaches. Our work also creates a new research direction that can improve the safety of future autonomous driving systems. The code and data will be released to the public to facilitate open-source research in this field. Our project website: https://eddyhkchiu.github.io/v2vllm.github.io/ .

Method: The survey organizes multimodal generative models by data dimensionality progression: starting from 2D generation (appearance), moving to video (appearance+dynamics), then 3D (appearance+geometry), and culminating in 4D generation that integrates all dimensions.

Result: Provides a systematic unification of 2D, video, 3D, and 4D generation within a single framework, along with comprehensive reviews of datasets, evaluation metrics, and future research directions.

Conclusion: This survey serves as a bridge to advance multimodal generative models and real-world simulation research by providing a unified framework that connects different data dimensionalities and modalities.

Abstract: Understanding and replicating the real world is a critical challenge in Artificial General Intelligence (AGI) research. To achieve this, many existing approaches, such as world models, aim to capture the fundamental principles governing the physical world, enabling more accurate simulations and meaningful interactions. However, current methods often treat different modalities, including 2D (images), videos, 3D, and 4D representations, as independent domains, overlooking their interdependencies. Additionally, these methods typically focus on isolated dimensions of reality without systematically integrating their connections. In this survey, we present a unified survey for multimodal generative models that investigate the progression of data dimensionality in real-world simulation. Specifically, this survey starts from 2D generation (appearance), then moves to video (appearance+dynamics) and 3D generation (appearance+geometry), and finally culminates in 4D generation that integrate all dimensions. To the best of our knowledge, this is the first attempt to systematically unify the study of 2D, video, 3D and 4D generation within a single framework. To guide future research, we provide a comprehensive review of datasets, evaluation metrics and future directions, and fostering insights for newcomers. This survey serves as a bridge to advance the study of multimodal generative models and real-world simulation within a unified framework.

Method: Three-task framework: 1) Structural token generation to partition videos into segments and categorize them as target events or background; 2) Query-focused captioning to extract fine-grained event semantics; 3) Structural token grounding module using contrastive learning to associate tokens with video segments.

Result: Extensive experiments show MeCo consistently outperforms timestamp-based methods across diverse temporal localization tasks, demonstrating the effectiveness of semantic-driven approaches.

Conclusion: Semantic-driven frameworks without timestamp generation better leverage video LLMs’ capabilities for temporal localization, offering a promising direction for video understanding tasks.

Abstract: Temporally localizing user-queried events through natural language is a crucial capability for video models. Recent methods predominantly adapt video LLMs to generate event boundary timestamps for temporal localization tasks, which struggle to leverage LLMs’ pre-trained semantic understanding capabilities due to the uninformative nature of timestamp outputs. In this work, we explore a timestamp-free, semantic-oriented framework that fine-tunes video LLMs using two generative learning tasks and one discriminative learning task. We first introduce a structural token generation task that enables the video LLM to recognize the temporal structure of input videos based on the input query. Through this task, the video LLM generates a sequence of special tokens, called structural tokens, which partition the video into consecutive segments and categorize them as either target events or background transitions. To enhance precise recognition of event segments, we further propose a query-focused captioning task that enables the video LLM to extract fine-grained event semantics that can be effectively utilized by the structural tokens. Finally, we introduce a structural token grounding module driven by contrastive learning to associate each structural token with its corresponding video segment, achieving holistic temporal segmentation of the input video and readily yielding the target event segments for localization. Extensive experiments across diverse temporal localization tasks demonstrate that our proposed framework, MeCo, consistently outperforms methods relying on boundary timestamp generation, highlighting the potential of a semantic-driven approach for temporal localization with video LLMs \footnote{Code available at https://github.com/pangzss/MeCo.

Method: Proposes a decoupled decoding framework that separates positional guidance (queries) from content representation (key-value pairs), enabling random-order training and generation through direct incorporation into causal attention mechanism.

Result: Achieves FID of 1.83 on ImageNet-1K 256 with only 32 sampling steps, achieving 30x speedup in inference and 75% reduction in memory consumption compared to similar-scale autoregressive models.

Conclusion: ARPG enables efficient parallel generation with flexible token order, supporting zero-shot tasks like inpainting, outpainting, and resolution expansion while significantly improving inference speed and memory efficiency.

Abstract: We introduce ARPG, a novel visual autoregressive model that enables randomized parallel generation, addressing the inherent limitations of conventional raster-order approaches, which hinder inference efficiency and zero-shot generalization due to their sequential, predefined token generation order. Our key insight is that effective random-order modeling necessitates explicit guidance for determining the position of the next predicted token. To this end, we propose a novel decoupled decoding framework that decouples positional guidance from content representation, encoding them separately as queries and key-value pairs. By directly incorporating this guidance into the causal attention mechanism, our approach enables fully random-order training and generation, eliminating the need for bidirectional attention. Consequently, ARPG readily generalizes to zero-shot inference tasks such as image inpainting, outpainting, and resolution expansion. Furthermore, it supports parallel inference by concurrently processing multiple queries using a shared KV cache. On the ImageNet-1K 256 benchmark, our approach attains an FID of 1.83 with only 32 sampling steps, achieving over a 30 times speedup in inference and a 75 percent reduction in memory consumption compared to representative recent autoregressive models at a similar scale.

Method: Created Car-1000 dataset with 1000 distinct car models from 166 automakers, providing a more comprehensive and up-to-date benchmark. Reproduced several state-of-the-art FGVC methods on this new dataset to establish performance baselines.

Result: Car-1000 dataset is publicly available and establishes new benchmarks for fine-grained car recognition. The dataset addresses the limitations of Stanford-Car by providing more categories and up-to-date vehicle models.

Conclusion: Car-1000 provides a fresh perspective for FGVC research with a comprehensive, up-to-date dataset that better reflects the current automotive landscape, enabling more robust car recognition systems for real-world applications.

Abstract: Fine-grained visual categorization (FGVC) is a challenging but significant task in computer vision, which aims to recognize different sub-categories of birds, cars, airplanes, etc. Among them, recognizing models of different cars has significant application value in autonomous driving, traffic surveillance and scene understanding, which has received considerable attention in the past few years. However, Stanford-Car, the most widely used fine-grained dataset for car recognition, only has 196 different categories and only includes vehicle models produced earlier than 2013. Due to the rapid advancements in the automotive industry during recent years, the appearances of various car models have become increasingly intricate and sophisticated. Consequently, the previous Stanford-Car dataset fails to capture this evolving landscape and cannot satisfy the requirements of automotive industry. To address these challenges, in our paper, we introduce Car-1000, a large-scale dataset designed specifically for fine-grained visual categorization of diverse car models. Car-1000 encompasses vehicles from 166 different automakers, spanning a wide range of 1000 distinct car models. Additionally, we have reproduced several state-of-the-art FGVC methods on the Car-1000 dataset, establishing a new benchmark for research in this field. We hope that our work will offer a fresh perspective for future FGVC researchers. Our dataset is available at https://github.com/toggle1995/Car-1000.

Method: 1) Create S2R-HDR dataset using Unreal Engine 5 with diverse realistic HDR scenes covering various dynamic elements, motion types, and lighting. 2) Develop efficient rendering pipeline for realistic HDR images. 3) Introduce S2R-Adapter for domain adaptation to bridge synthetic-real gap.

Result: Experimental results on real-world datasets demonstrate state-of-the-art HDR fusion performance, showing the effectiveness of the synthetic dataset and domain adaptation approach.

Conclusion: S2R-HDR provides a valuable synthetic dataset solution for HDR fusion training, and S2R-Adapter effectively mitigates domain gap issues, enabling better generalization to real-world scenarios.

Abstract: The generalization of learning-based high dynamic range (HDR) fusion is often limited by the availability of training data, as collecting large-scale HDR images from dynamic scenes is both costly and technically challenging. To address these challenges, we propose S2R-HDR, the first large-scale high-quality synthetic dataset for HDR fusion, with 24,000 HDR samples. Using Unreal Engine 5, we design a diverse set of realistic HDR scenes that encompass various dynamic elements, motion types, high dynamic range scenes, and lighting. Additionally, we develop an efficient rendering pipeline to generate realistic HDR images. To further mitigate the domain gap between synthetic and real-world data, we introduce S2R-Adapter, a domain adaptation designed to bridge this gap and enhance the generalization ability of models. Experimental results on real-world datasets demonstrate that our approach achieves state-of-the-art HDR fusion performance. Dataset and code are available at https://openimaginglab.github.io/S2R-HDR.

Method: Proposes GaussianFormer3D framework with: 1) Voxel-to-Gaussian initialization strategy using LiDAR data for geometry priors, 2) LiDAR-guided 3D deformable attention mechanism to refine Gaussians using LiDAR-camera fusion features in lifted 3D space.

Result: Extensive experiments on real-world on-road and off-road autonomous driving datasets show state-of-the-art prediction performance with reduced memory consumption and improved efficiency.

Conclusion: GaussianFormer3D successfully demonstrates the effectiveness of 3D Gaussian representations with multi-modal fusion for semantic occupancy prediction, offering better performance and efficiency than traditional voxel-based methods.

Abstract: 3D semantic occupancy prediction is essential for achieving safe, reliable autonomous driving and robotic navigation. Compared to camera-only perception systems, multi-modal pipelines, especially LiDAR-camera fusion methods, can produce more accurate and fine-grained predictions. Although voxel-based scene representations are widely used for semantic occupancy prediction, 3D Gaussians have emerged as a continuous and significantly more compact alternative. In this work, we propose a multi-modal Gaussian-based semantic occupancy prediction framework utilizing 3D deformable attention, namely GaussianFormer3D. We introduce a voxel-to-Gaussian initialization strategy that provides 3D Gaussians with accurate geometry priors from LiDAR data, and design a LiDAR-guided 3D deformable attention mechanism to refine these Gaussians using LiDAR-camera fusion features in a lifted 3D space. Extensive experiments on real-world on-road and off-road autonomous driving datasets demonstrate that GaussianFormer3D achieves state-of-the-art prediction performance with reduced memory consumption and improved efficiency.

Method: 1) Local Linear Correction Network (LLCN) that finetunes pre-trained models based on physical constraint T=SI+B (S and B represent pixel-wise and channel-wise scaling/bias transformations) to generate high-quality transmission priors with minimal parameters. 2) Dual-Prior Interaction Transformer (DPIT) with dual-stream channel reorganization attention mechanism that reorganizes features from general and transmission priors for attention computation to achieve deep fusion of both priors.

Result: Experimental results on multiple benchmark datasets demonstrate state-of-the-art performance.

Conclusion: The proposed dual-prior interaction framework effectively models and utilizes transmission priors through lightweight generation and deep fusion, achieving superior performance in single image reflection separation.

Abstract: Single image reflection separation aims to separate the transmission and reflection layers from a mixed image. Existing methods typically combine general priors from pre-trained models with task-specific priors such as text prompts and reflection detection. However, the transmission prior, as the most direct task-specific prior for the target transmission layer, has not been effectively modeled or fully utilized, limiting performance in complex scenarios. To address this issue, we propose a dual-prior interaction framework based on lightweight transmission prior generation and effective prior fusion. First, we design a Local Linear Correction Network (LLCN) that finetunes pre-trained models based on the physical constraint T=SI+B, where S and B represent pixel-wise and channel-wise scaling and bias transformations. LLCN efficiently generates high-quality transmission priors with minimal parameters. Second, we construct a Dual-Prior Interaction Transformer (DPIT) that employs a dual-stream channel reorganization attention mechanism. By reorganizing features from general and transmission priors for attention computation, DPIT achieves deep fusion of both priors, fully exploiting their complementary information. Experimental results on multiple benchmark datasets demonstrate that the proposed method achieves state-of-the-art performance.

Method: Used saliency-based attribution analysis to show RGB-pretrained models assign unequal importance to EM slices; proposed targeted modification of pretraining weights via uniform channel initialization to restore symmetric feature attribution while preserving pretraining benefits.

Result: Experiments on SNEMI, Lucchi and GF-PA66 datasets show substantial reduction in attribution bias without compromising segmentation accuracy, sometimes even improving performance.

Conclusion: Standard RGB pretraining introduces harmful inductive biases for symmetric volumetric data; uniform channel initialization effectively restores symmetry in feature attribution while maintaining transfer learning benefits for EM segmentation.

Abstract: Vision models pretrained on large-scale RGB natural image datasets are widely reused for electron microscopy image segmentation. In electron microscopy, volumetric data are acquired as serial sections and processed as stacks of adjacent grayscale slices, where neighboring slices provide symmetric contextual information for identifying features on the central slice. The common strategy maps such stacks to pseudo-RGB inputs to enable transfer learning from pretrained models. However, this mapping imposes channel-specific semantics inherited from natural images, even though electron microscopy slices are homogeneous in the modality and symmetric in their predictive roles. As a result, pretrained models may encode inductive biases that are misaligned with the inherent symmetry of volumetric electron microscopy data. In this work, it is demonstrated that RGB-pretrained models systematically assign unequal importance to individual input slices when applied to stacked electron microscopy data, despite the absence of any intrinsic channel ordering. Using saliency-based attribution analysis across multiple architectures, a consistent channel-level asymmetry was observed that persists after fine-tuning and affects model interpretability, even when segmentation performance is unchanged. To address this issue, a targeted modification of pretraining weights based on uniform channel initialization was proposed, which restores symmetric feature attribution while preserving the benefits of pretraining. Experiments on the SNEMI, Lucchi and GF-PA66 datasets confirm a substantial reduction in attribution bias without compromising or even improving segmentation accuracy.

Method: Developed a scalable, modular-topology data inference framework with multi-observation sources and expert-in-the-loop curation to create standardized annotations organized in a four-level hierarchy (Sphere, Scenario, Ability, Task).

Result: Created 29,855 expert-curated annotations across 109 tasks; evaluation of 9 state-of-the-art MLLMs showed all models struggling, with none reaching 35% accuracy, revealing systematic gaps in Earth-system cognitive ability.

Conclusion: OmniEarth-Bench provides the first comprehensive multimodal benchmark for Earth science, exposing significant limitations in current MLLMs’ ability to understand complex Earth systems and cross-sphere interactions.

Abstract: Existing benchmarks for multimodal learning in Earth science offer limited, siloed coverage of Earth’s spheres and their cross-sphere interactions, typically restricting evaluation to the human-activity sphere of atmosphere and to at most 16 tasks. These limitations: narrow-source heterogeneity (single/few data sources), constrained scientific granularity, and limited-sphere extensibility. Therefore, we introduce OmniEarth-Bench, the first multimodal benchmark that systematically spans all six spheres: atmosphere, lithosphere, oceanosphere, cryosphere, biosphere, and human-activity sphere, and cross-spheres. Built with a scalable, modular-topology data inference framework and native multi-observation sources and expert-in-the-loop curation, OmniEarth-Bench produces 29,855 standardized, expert-curated annotations. All annotations are organized into a four-level hierarchy (Sphere, Scenario, Ability, Task), encompassing 109 expert-curated evaluation tasks. Experiments on 9 state-of-the-art MLLMs reveal that even the most advanced models struggle with our benchmarks, where none of them reach 35% accuracy, revealing systematic gaps in Earth-system cognitive ability. The dataset and evaluation code were released at OmniEarth-Bench (https://anonymous.4open.science/r/OmniEarth-Bench-B1BD).

Method: AliTok uses a bidirectional encoder constrained by a causal decoder to produce token sequences with both semantic richness and forward-dependency. It incorporates prefix tokens and a two-stage training process to enhance reconstruction performance.

Result: A 177M parameter autoregressive model achieves gFID of 1.44 and IS of 319.5 on ImageNet-256. Scaling to 662M parameters reaches gFID of 1.28, surpassing SOTA diffusion methods with 10x faster sampling. On ImageNet-512, a 318M model achieves SOTA gFID of 1.39.

Conclusion: AliTok successfully resolves the misalignment problem in autoregressive image generation, enabling high-fidelity image generation with faster sampling than diffusion models while maintaining strong performance metrics.

Abstract: Autoregressive image generation aims to predict the next token based on previous ones. However, this process is challenged by the bidirectional dependencies inherent in conventional image tokenizations, which creates a fundamental misalignment with the unidirectional nature of autoregressive models. To resolve this, we introduce AliTok, a novel Aligned Tokenizer that alters the dependency structure of the token sequence. AliTok employs a bidirectional encoder constrained by a causal decoder, a design that compels the encoder to produce a token sequence with both semantic richness and forward-dependency. Furthermore, by incorporating prefix tokens and employing a two-stage tokenizer training process to enhance reconstruction performance, AliTok achieves high fidelity and predictability simultaneously. Building upon AliTok, a standard decoder-only autoregressive model with just 177M parameters achieves a gFID of 1.44 and an IS of 319.5 on ImageNet-256. Scaling to 662M, our model reaches a gFID of 1.28, surpassing the SOTA diffusion method with 10x faster sampling. On ImageNet-512, our 318M model also achieves a SOTA gFID of 1.39. Code and weights at https://github.com/ali-vilab/alitok.

Method: Four-step pipeline: (1) segment video into scenes, (2) produce scene descriptions with memory-efficient batch prompting, (3) score scene importance with LLM via tailored prompts, (4) propagate scores to frames using consistency (temporal coherence) and uniqueness (novelty) metrics.

Result: Surpasses all prior data-hungry unsupervised methods on SumMe and TVSum, performs competitively on Query-Focused Video Summarization benchmark, and introduces VidSum-Reason dataset as first challenging baseline.

Conclusion: Pretrained multi-modal models, when orchestrated with principled prompting and score propagation, provide powerful foundation for universal, text-queryable video summarization without training data.

Abstract: The explosive growth of video data intensified the need for flexible user-controllable summarization tools that operate without training data. Existing methods either rely on domain-specific datasets, limiting generalization, or cannot incorporate user intent expressed in natural language. We introduce Prompts-to-Summaries: the first zero-shot, text-queryable video-summarizer that converts off-the-shelf video-language models (VidLMs) captions into user-guided skims via large-language-models (LLMs) judging, without the use of training data, beating unsupervised and matching supervised methods. Our pipeline (i) segments video into scenes, (ii) produces scene descriptions with a memory-efficient batch prompting scheme that scales to hours on a single GPU, (iii) scores scene importance with an LLM via tailored prompts, and (iv) propagates scores to frames using new consistency (temporal coherence) and uniqueness (novelty) metrics for fine-grained frame importance. On SumMe and TVSum, our approach surpasses all prior data-hungry unsupervised methods and performs competitively on the Query-Focused Video Summarization benchmark, where the competing methods require supervised frame-level importance. We release VidSum-Reason, a query-driven dataset featuring long-tailed concepts and multi-step reasoning, where our framework serves as the first challenging baseline. Overall, we demonstrate that pretrained multi-modal models, when orchestrated with principled prompting and score propagation, provide a powerful foundation for universal, text-queryable video summarization.

Method: SnS uses bi-objective optimization: for invariance, it finds perturbations that maximally alter representations while preserving unit activation; for adversarial sensitivity, it finds perturbations that maximally alter unit activation while preserving representations. Applied to CNNs at different processing stages.

Result: SnS revealed invariant transformations farther from reference images than affine transforms while better preserving target responses. Different processing stages produced different invariant changes: pixel-level affected luminance/contrast, mid/late layers affected texture/pose. L2 robust networks showed interpretability drops when stretching deep layers.

Conclusion: SnS provides a systematic way to characterize visual unit invariance and adversarial vulnerability, revealing stage-dependent transformation manifolds and interpretability differences between standard and robust models.

Abstract: Uncovering which feature combinations are encoded by visual units is critical to understanding how images are transformed into representations that support recognition. While existing feature visualization approaches typically infer a unit’s most exciting images, this is insufficient to reveal the manifold of transformations under which responses remain invariant, which is critical to generalization in vision. Here we introduce Stretch-and-Squeeze (SnS), a model-agnostic, gradient-free framework to systematically characterize a unit’s maximally invariant stimuli, and its vulnerability to adversarial perturbations, in both biological and artificial visual systems. SnS frames these transformations as bi-objective optimization problems. To probe invariance, SnS seeks image perturbations that maximally alter (stretch) the representation of a reference stimulus in a given processing stage while preserving unit activation downstream (squeeze). To probe adversarial sensitivity, stretching and squeezing are reversed to maximally perturb unit activation while minimizing changes to the upstream representation. Applied to CNNs, SnS revealed invariant transformations that were farther from a reference image in pixel-space than those produced by affine transformations, while more strongly preserving the target unit’s response. The discovered invariant images differed depending on the stage of the image representation used for optimization: pixel-level changes primarily affected luminance and contrast, while stretching mid- and late-layer representations mainly altered texture and pose. By measuring how well the hierarchical invariant images obtained for L2 robust networks were classified by humans and other observer networks, we discovered a substantial drop in their interpretability when the representation was stretched in deep layers, while the opposite trend was found for standard models.

Method: Proposes HMSViT with pooling-based hierarchical and dual attention mechanisms with absolute positional encoding for multi-scale feature extraction. Uses block-masked self-supervised learning to reduce reliance on labeled data, and a multi-scale decoder for segmentation and classification by fusing hierarchical features.

Result: Achieves 61.34% mIoU for nerve segmentation and 70.40% diagnostic accuracy, outperforming Swin Transformer and HiViT by up to 6.39% in segmentation accuracy while using fewer parameters. Ablation studies show SSL with hierarchical feature extraction substantially enhances performance.

Conclusion: HMSViT delivers excellent, robust, and clinically viable results, demonstrating potential for scalable deployment in real-world diagnostic applications for diabetic neuropathy detection.

Abstract: Diabetic Peripheral Neuropathy (DPN) affects nearly half of diabetes patients, requiring early detection. Corneal Confocal Microscopy (CCM) enables non-invasive diagnosis, but automated methods suffer from inefficient feature extraction, reliance on handcrafted priors, and data limitations. We propose HMSViT, a novel Hierarchical Masked Self-Supervised Vision Transformer (HMSViT) designed for corneal nerve segmentation and DPN diagnosis. Unlike existing methods, HMSViT employs pooling-based hierarchical and dual attention mechanisms with absolute positional encoding, enabling efficient multi-scale feature extraction by capturing fine-grained local details in early layers and integrating global context in deeper layers, all at a lower computational cost. A block-masked self supervised learning framework is designed for the HMSViT that reduces reliance on labelled data, enhancing feature robustness, while a multi-scale decoder is used for segmentation and classification by fusing hierarchical features. Experiments on clinical CCM datasets showed HMSViT achieves state-of-the-art performance, with 61.34% mIoU for nerve segmentation and 70.40% diagnostic accuracy, outperforming leading hierarchical models like the Swin Transformer and HiViT by margins of up to 6.39% in segmentation accuracy while using fewer parameters. Detailed ablation studies further reveal that integrating block-masked SSL with hierarchical multi-scale feature extraction substantially enhances performance compared to conventional supervised training. Overall, these comprehensive experiments confirm that HMSViT delivers excellent, robust, and clinically viable results, demonstrating its potential for scalable deployment in real-world diagnostic applications.

Method: RECALL exploits diffusion models’ multi-modal conditioning by efficiently optimizing adversarial image prompts guided by a single semantically relevant reference image, rather than relying solely on adversarial text prompts.

Result: Extensive experiments across 10 state-of-the-art unlearning methods show RECALL consistently outperforms existing baselines in adversarial effectiveness, computational efficiency, and semantic fidelity with original textual prompts.

Conclusion: The framework reveals critical vulnerabilities in current unlearning mechanisms and underscores the need for more robust solutions to ensure safety and reliability of generative models.

Abstract: Recent advances in image generation models (IGMs), particularly diffusion-based architectures such as Stable Diffusion (SD), have markedly enhanced the quality and diversity of AI-generated visual content. However, their generative capability has also raised significant ethical, legal, and societal concerns, including the potential to produce harmful, misleading, or copyright-infringing content. To mitigate these concerns, machine unlearning (MU) emerges as a promising solution by selectively removing undesirable concepts from pretrained models. Nevertheless, the robustness and effectiveness of existing unlearning techniques remain largely unexplored, particularly in the presence of multi-modal adversarial inputs. To bridge this gap, we propose Recall, a novel adversarial framework explicitly designed to compromise the robustness of unlearned IGMs. Unlike existing approaches that predominantly rely on adversarial text prompts, Recall exploits the intrinsic multi-modal conditioning capabilities of diffusion models by efficiently optimizing adversarial image prompts with guidance from a single semantically relevant reference image. Extensive experiments across ten state-of-the-art unlearning methods and diverse tasks show that Recall consistently outperforms existing baselines in terms of adversarial effectiveness, computational efficiency, and semantic fidelity with the original textual prompt. These findings reveal critical vulnerabilities in current unlearning mechanisms and underscore the need for more robust solutions to ensure the safety and reliability of generative models. Code and data are publicly available at \textcolor{blue}{https://github.com/ryliu68/RECALL}.

Method: Proposes DGFDNet with DGFDBlock containing: 1) Haze-Aware Frequency Modulator (HAFM) using dark channel priors to generate haze confidence maps for adaptive frequency modulation; 2) Multi-level Gating Aggregation Module (MGAM) for multi-scale feature fusion; and 3) Prior Correction Guidance Branch (PCGB) for iterative refinement of priors.

Result: Achieves state-of-the-art performance on four benchmark datasets with improved robustness and real-time efficiency.

Conclusion: DGFDNet effectively addresses computational costs and weak spatial-frequency coupling in dehazing by explicitly aligning degradation across domains using dark channel priors and adaptive frequency modulation.

Abstract: Transformers offer strong global modeling for single-image dehazing but come with high computational costs. Most methods rely on spatial features to capture long-range dependencies, making them less effective under complex haze conditions. Although some integrate frequency-domain cues, weak coupling between spatial and frequency branches limits their performance. To address these issues, we propose the Dark Channel Guided Frequency-aware Dehazing Network (DGFDNet), a dual-domain framework that explicitly aligns degradation across spatial and frequency domains. At its core, the DGFDBlock consists of two key modules: 1) Haze-Aware Frequency Modulator (HAFM), which uses dark channel priors to generate a haze confidence map for adaptive frequency modulation, achieving global degradation-aware spectral filtering. 2) Multi-level Gating Aggregation Module (MGAM), which fuses multi-scale features via multi-scale convolutions and a hybrid gating mechanism to recover fine-grained structures. Additionally, the Prior Correction Guidance Branch (PCGB) incorporates feedback for iterative refinement of the prior, improving haze localization accuracy, particularly in outdoor scenes. Extensive experiments on four benchmark datasets demonstrate that DGFDNet achieves state-of-the-art performance with improved robustness and real-time efficiency. Code is available at: https://github.com/Dilizlr/DGFDNet.

Method: Introduces Latent Denoising Tokenizer (l-DeTok) that trains tokenizer embeddings to reconstruct clean images from latent embeddings corrupted via interpolative noise or random masking, aligning tokenizer design with the denoising objectives of generative models.

Result: Extensive experiments on class-conditioned (ImageNet 256x256 and 512x512) and text-conditioned (MSCOCO) image generation benchmarks show l-DeTok consistently improves generation quality across six representative generative models compared to prior tokenizers.

Conclusion: Denoising should be a fundamental design principle for tokenizer development, and l-DeTok demonstrates the effectiveness of aligning tokenizer embeddings with downstream denoising objectives for improved generative modeling.

Abstract: Despite their fundamental role, it remains unclear what properties could make tokenizers more effective for generative modeling. We observe that modern generative models share a conceptually similar training objective – reconstructing clean signals from corrupted inputs, such as signals degraded by Gaussian noise or masking – a process we term denoising. Motivated by this insight, we propose aligning tokenizer embeddings directly with the downstream denoising objective, encouraging latent embeddings that remain reconstructable even under significant corruption. To achieve this, we introduce the Latent Denoising Tokenizer (l-DeTok), a simple yet highly effective tokenizer trained to reconstruct clean images from latent embeddings corrupted via interpolative noise or random masking. Extensive experiments on class-conditioned (ImageNet 256x256 and 512x512) and text-conditioned (MSCOCO) image generation benchmarks demonstrate that our l-DeTok consistently improves generation quality across six representative generative models compared to prior tokenizers. Our findings highlight denoising as a fundamental design principle for tokenizer development, and we hope it could motivate new perspectives for future tokenizer design.

Method: 3DRot is a plug-and-play augmentation that rotates and mirrors images about the camera’s optical center while synchronously updating RGB images, camera intrinsics, object poses, and 3D annotations to preserve projective geometry. It achieves geometry-consistent rotations and reflections without relying on any scene depth.

Result: Significant improvements across multiple 3D tasks: On SUN RGB-D, improved IoU3D from 43.21 to 44.51, reduced rotation error from 22.91° to 20.93°, and boosted mAP0.5 from 35.70 to 38.11. On NYU Depth v2, improved abs-rel from 0.1783 to 0.1685. On KITTI with MVX-Net, raised moderate 3D AP from ~63.85 to 65.16.

Conclusion: 3DRot provides an effective, depth-free augmentation method for RGB-based 3D tasks that preserves geometric consistency and improves performance across multiple benchmarks and tasks, remaining compatible with standard 3D augmentations.

Abstract: RGB-based 3D tasks, e.g., 3D detection, depth estimation, 3D keypoint estimation, still suffer from scarce, expensive annotations and a thin augmentation toolbox, since many image transforms, including rotations and warps, disrupt geometric consistency. While horizontal flipping and color jitter are standard, rigorous 3D rotation augmentation has surprisingly remained absent from RGB-based pipelines, largely due to the misconception that it requires scene depth or scene reconstruction. In this paper, we introduce 3DRot, a plug-and-play augmentation that rotates and mirrors images about the camera’s optical center while synchronously updating RGB images, camera intrinsics, object poses, and 3D annotations to preserve projective geometry, achieving geometry-consistent rotations and reflections without relying on any scene depth. We first validate 3DRot on a classical RGB-based 3D task, monocular 3D detection. On SUN RGB-D, inserting 3DRot into a frozen DINO-X + Cube R-CNN pipeline raises $IoU_{3D}$ from 43.21 to 44.51, cuts rotation error (ROT) from 22.91$^\circ$ to 20.93$^\circ$, and boosts $mAP_{0.5}$ from 35.70 to 38.11; smaller but consistent gains appear on a cross-domain IN10 split. Beyond monocular detection, adding 3DRot on top of the standard BTS augmentation schedule further improves NYU Depth v2 from 0.1783 to 0.1685 in abs-rel (and 0.7472 to 0.7548 in $δ<1.25$), and reduces cross-dataset error on SUN RGB-D. On KITTI, applying the same camera-centric rotations in MVX-Net (LiDAR+RGB) raises moderate 3D AP from about 63.85 to 65.16 while remaining compatible with standard 3D augmentations.

Method: Integrates vision foundation models (DINOv2) with hierarchical taxonomic inference, trained on large multi-year dataset (Germany/Spain 2018-2020) with 14 plant species and 4 herbicide damage classes, evaluated under challenging domain shifts.

Result: Foundation-model backbone consistently outperforms baselines, improving species-level F1 from 0.52 to 0.87 on in-distribution data, maintaining advantages under moderate (0.77 vs. 0.24) and extreme (0.44 vs. 0.14) shift conditions. Hierarchical inference provides additional robustness.

Conclusion: Combining foundation models with structured biological hierarchies enables scalable, shift-resilient agricultural monitoring, now deployed in BASF’s phenotyping workflow for herbicide research trials.

Abstract: Reliable plant species and damage segmentation for herbicide field research trials requires models that can withstand substantial real-world variation across seasons, geographies, devices, and sensing modalities. Most deep learning approaches trained on controlled datasets fail to generalize under these domain shifts, limiting their suitability for operational phenotyping pipelines. This study evaluates a segmentation framework that integrates vision foundation models (DINOv2) with hierarchical taxonomic inference to improve robustness across heterogeneous agricultural conditions. We train on a large, multi-year dataset collected in Germany and Spain (2018-2020), comprising 14 plant species and 4 herbicide damage classes, and assess generalization under increasingly challenging shifts: temporal and device changes (2023), geographic transfer to the United States, and extreme sensor shift to drone imagery (2024). Results show that the foundation-model backbone consistently outperforms prior baselines, improving species-level F1 from 0.52 to 0.87 on in-distribution data and maintaining significant advantages under moderate (0.77 vs. 0.24) and extreme (0.44 vs. 0.14) shift conditions. Hierarchical inference provides an additional layer of robustness, enabling meaningful predictions even when fine-grained species classification degrades (family F1: 0.68, class F1: 0.88 on aerial imagery). Error analysis reveals that failures under severe shift stem primarily from vegetation-soil confusion, suggesting that taxonomic distinctions remain preserved despite background and viewpoint variability. The system is now deployed within BASF’s phenotyping workflow for herbicide research trials across multiple regions, illustrating the practical viability of combining foundation models with structured biological hierarchies for scalable, shift-resilient agricultural monitoring.

Method: Uses diffusion transformer (DiT) architecture based on DDIM for brain signal generation. Employs cross-attention mechanisms to align brain signal embeddings with CLIP image embeddings. Leverages LLMs to generate image captions, concatenating CLIP text and image embeddings. Introduces learnable spatio-temporal position encoding combining brain region embeddings with temporal embeddings.

Result: Evaluated on THINGS-EEG2 and THINGS-MEG datasets, demonstrating generation of biologically plausible brain signals.

Conclusion: The framework successfully generates M/EEG signals from images, addressing the brain encoding gap in visual prostheses and enabling a more complete functional pipeline for vision restoration.

Abstract: Visual prostheses hold great promise for restoring vision in blind individuals. While researchers have successfully utilized M/EEG signals to evoke visual perceptions during the brain decoding stage of visual prostheses, the complementary process of converting images into M/EEG signals in the brain encoding stage remains largely unexplored, hindering the formation of a complete functional pipeline. In this work, we present a novel image-to-brain signal framework that generates M/EEG from images by leveraging the diffusion transformer architecture enhanced with cross-attention mechanisms. Specifically, we employ a diffusion transformer (DiT) architecture based on denoising diffusion implicit models (DDIM) to achieve brain signal generation. To realize the goal of image-to-brain signal conversion, we use cross-attention mechanisms to align brain signal embeddings with CLIP image embeddings. Moreover, we leverage large language models (LLMs) to generate image captions, and concatenate the resulting CLIP text embeddings with CLIP image embeddings to form unified embeddings for cross-attention alignment, enabling our model to capture core semantic information. Moreover, to capture core semantic information, we use large language models (LLMs) to generate descriptive and semantically accurate captions for images. Furthermore, we introduce a learnable spatio-temporal position encoding that combines brain region embeddings with temporal embeddings to capture both spatial and temporal characteristics of brain signals. We evaluate the framework on two multimodal benchmark datasets (THINGS-EEG2 and THINGS-MEG) and demonstrate that it generates biologically plausible brain signals.

Method: Proposes BEVTraj framework with: 1) Deformable attention to adaptively aggregate task-relevant context from sparse locations in dense BEV features, 2) Sparse Goal Candidate Proposal (SGCP) module that predicts a small set of realistic goals for end-to-end multimodal forecasting without heuristic post-processing.

Result: Extensive experiments show BEVTraj achieves performance comparable to state-of-the-art HD map-based methods while providing greater robustness and flexibility without relying on pre-built maps.

Conclusion: BEVTraj demonstrates that map-free trajectory prediction can achieve competitive performance with map-based methods while offering practical advantages in real-world deployment scenarios.

Abstract: In autonomous driving, trajectory prediction is essential for safe and efficient navigation. While recent methods often rely on high-definition (HD) maps to provide structured environmental priors, such maps are costly to maintain, geographically limited, and unreliable in dynamic or unmapped scenarios. Directly leveraging raw sensor data in Bird’s-Eye View (BEV) space offers greater flexibility, but BEV features are dense and unstructured, making agent-centric spatial reasoning challenging and computationally inefficient. To address this, we propose Bird’s-Eye View Trajectory Prediction (BEVTraj), a map-free framework that employs deformable attention to adaptively aggregate task-relevant context from sparse locations in dense BEV features. We further introduce a Sparse Goal Candidate Proposal (SGCP) module that predicts a small set of realistic goals, enabling fully end-to-end multimodal forecasting without heuristic post-processing. Extensive experiments show that BEVTraj achieves performance comparable to state-of-the-art HD map-based methods while providing greater robustness and flexibility without relying on pre-built maps. The source code is available at https://github.com/Kongminsang/bevtraj.

Method: CMTSSL integrates masked image modeling with decoupled spatial and spectral jigsaw puzzle solving, guided by curriculum learning that progressively increases data difficulty during self-supervision to capture spectral continuity, spatial structure, and global semantics.

Result: Validated on four public benchmark datasets, showing consistent gains in downstream segmentation tasks with architectures over 16,000x lighter than some state-of-the-art models.

Conclusion: CMTSSL enables generalizable representation learning with lightweight architectures suitable for real-world HSI applications and onboard satellite deployment.

Abstract: Hyperspectral imaging (HSI) captures detailed spectral signatures across hundreds of contiguous bands per pixel, being indispensable for remote sensing applications such as land-cover classification, change detection, and environmental monitoring. Due to the high dimensionality of HSI data and the slow rate of data transfer in satellite-based systems, compact and efficient models are required to support onboard processing and minimize the transmission of redundant or low-value data. To this end, we introduce a novel curriculum multi-task self-supervised learning (CMTSSL) framework designed for lightweight architectures for HSI analysis. CMTSSL integrates masked image modeling with decoupled spatial and spectral jigsaw puzzle solving, guided by a curriculum learning strategy that progressively increases data difficulty during self-supervision. This enables the encoder to jointly capture fine-grained spectral continuity, spatial structure, and global semantic features. Unlike prior dual-task SSL methods, CMTSSL simultaneously addresses spatial and spectral reasoning within a unified and computationally efficient design, being particularly suitable for training lightweight models for onboard satellite deployment. We validate our approach on four public benchmark datasets, demonstrating consistent gains in downstream segmentation tasks, using architectures that are over 16,000x lighter than some state-of-the-art models. These results highlight the potential of CMTSSL in generalizable representation learning with lightweight architectures for real-world HSI applications. Our code is publicly available at https://github.com/hugocarlesso/CMTSSL.

Method: Introduces new labeling strategy placing ball at center of blur streak with explicit blur attribute annotation. Creates new table tennis ball detection dataset. Proposes BlurBall model that jointly estimates ball position and motion blur attributes using attention mechanisms (Squeeze-and-Excitation) over multi-frame inputs.

Result: New labeling approach consistently enhances detection performance across various models. BlurBall achieves state-of-the-art results in ball detection. Leveraging blur improves detection accuracy and enables more reliable trajectory prediction.

Conclusion: Center-based blur labeling with explicit attribute annotation improves ball detection and enables motion estimation, benefiting real-time sports analytics through better trajectory prediction.

Abstract: Motion blur reduces the clarity of fast-moving objects, posing challenges for detection systems, especially in racket sports, where balls often appear as streaks rather than distinct points. Existing labeling conventions mark the ball at the leading edge of the blur, introducing asymmetry and ignoring valuable motion cues correlated with velocity. This paper introduces a new labeling strategy that places the ball at the center of the blur streak and explicitly annotates blur attributes. Using this convention, we release a new table tennis ball detection dataset. We demonstrate that this labeling approach consistently enhances detection performance across various models. Furthermore, we introduce BlurBall, a model that jointly estimates ball position and motion blur attributes. By incorporating attention mechanisms such as Squeeze-and-Excitation over multi-frame inputs, we achieve state-of-the-art results in ball detection. Leveraging blur not only improves detection accuracy but also enables more reliable trajectory prediction, benefiting real-time sports analytics.

Method: Deployed and evaluated conventional techniques (Iterative Logistic Regression and Multilayer Perceptron) with advanced deep learning architectures (U-Net and Spectral Channel Attention Network) for cloud and cloud shadow detection in high-resolution hyperspectral data.

Result: Conventional methods struggle with spatial coherence and boundary definition. Deep learning models substantially improve detection quality: U-Net performs best in preserving spatial structure, while SCAN excels at capturing fine boundary details.

Conclusion: Deep learning approaches, particularly U-Net and SCAN, offer superior cloud and cloud shadow detection for high-resolution hyperspectral remote sensing data compared to conventional methods, enabling more accurate methane retrieval and emission quantification.

Abstract: Effective cloud and cloud shadow detection is a critical prerequisite for accurate retrieval of concentrations of atmospheric methane (CH4) or other trace gases in hyperspectral remote sensing. This challenge is especially pertinent for MethaneSAT, a satellite mission launched in March 2024, to fill a significant data gap in terms of resolution, precision and swath between coarse-resolution global mappers and fine-scale point-source imagers of methane, and for its airborne companion mission, MethaneAIR. MethaneSAT delivers hyperspectral data at an intermediate spatial resolution (approx. 100 x 400, m), whereas MethaneAIR provides even finer resolution (approx. 25 m), enabling the development of highly detailed maps of concentrations that enable quantification of both the sources and rates of emissions. In this study, we use machine learning methods to address the cloud and cloud shadow detection problem for sensors with these high spatial resolutions. Cloud and cloud shadows in remote sensing data need to be effectively screened out as they bias methane retrievals in remote sensing imagery and impact the quantification of emissions. We deploy and evaluate conventional techniques-including Iterative Logistic Regression (ILR) and Multilayer Perceptron (MLP)-with advanced deep learning architectures, namely U-Net and a Spectral Channel Attention Network (SCAN) method. Our results show that conventional methods struggle with spatial coherence and boundary definition, affecting the detection of clouds and cloud shadows. Deep learning models substantially improve detection quality: U-Net performs best in preserving spatial structure, while SCAN excels at capturing fine boundary details… Our data and code is publicly available at: https://doi.org/10.7910/DVN/IKLZOJ

Method: Proposes DeLiVR which injects spatiotemporal Lie-group differential biases directly into attention scores. Uses two components: 1) rotation-bounded Lie relative bias for geometry-consistent alignment, and 2) differential group displacement to estimate velocity between frames with temporal decay and attention masks.

Result: Extensive experimental results demonstrate effectiveness on publicly available benchmarks. Code is publicly available.

Conclusion: Lie groups provide principled representation for continuous geometric transformations, enabling efficient video deraining with spatiotemporal consistency through attention-based differential biases.

Abstract: Videos captured in the wild often suffer from rain streaks, blur, and noise. In addition, even slight changes in camera pose can amplify cross-frame mismatches and temporal artifacts. Existing methods rely on optical flow or heuristic alignment, which are computationally expensive and less robust. To address these challenges, Lie groups provide a principled way to represent continuous geometric transformations, making them well-suited for enforcing spatial and temporal consistency in video modeling. Building on this insight, we propose DeLiVR, an efficient video deraining method that injects spatiotemporal Lie-group differential biases directly into attention scores of the network. Specifically, the method introduces two complementary components. First, a rotation-bounded Lie relative bias predicts the in-plane angle of each frame using a compact prediction module, where normalized coordinates are rotated and compared with base coordinates to achieve geometry-consistent alignment before feature aggregation. Second, a differential group displacement computes angular differences between adjacent frames to estimate a velocity. This bias computation combines temporal decay and attention masks to focus on inter-frame relationships while precisely matching the direction of rain streaks. Extensive experimental results demonstrate the effectiveness of our method on publicly available benchmarks. The code is publicly available at https://github.com/Shuning0312/ICLR-DeLiVR.

Method: Uses 11 synchronized cameras for 360° coverage, specialized deep learning modules (YOLOv8 for part detection, EfficientNet for ICE/EV classification, Gemini-1.5 Flash for OCR, YOLOv8-Seg for segmentation), view-aware fusion layer, and VIN-conditioned rule engine for comparison against expected manifest.

Result: Achieves 93% verification accuracy, 86% defect-detection recall, processes 3.3 vehicles per minute, outperforming single-view or no segmentation baselines.

Conclusion: First publicly reported system unifying multi-camera feature validation with defect detection in deployable automotive industry setting, demonstrating practical real-time quality control.

Abstract: Ensuring that every vehicle leaving a modern production line is built to the correct \emph{variant} specification and is free from visible defects is an increasingly complex challenge. We present the \textbf{Automated Vehicle Inspection (AVI)} platform, an end-to-end, \emph{multi-view} perception system that couples deep-learning detectors with a semantic rule engine to deliver \emph{variant-aware} quality control in real time. Eleven synchronized cameras capture a full 360° sweep of each vehicle; task-specific views are then routed to specialised modules: YOLOv8 for part detection, EfficientNet for ICE/EV classification, Gemini-1.5 Flash for mascot OCR, and YOLOv8-Seg for scratch-and-dent segmentation. A view-aware fusion layer standardises evidence, while a VIN-conditioned rule engine compares detected features against the expected manifest, producing an interpretable pass/fail report in (\approx! 300,\text{ms}). On a mixed data set of Original Equipment Manufacturer(OEM) vehicle data sets of four distinct models plus public scratch/dent images, AVI achieves \textbf{ 93 %} verification accuracy, \textbf{86 %} defect-detection recall, and sustains (\mathbf{3.3}) vehicles/min, surpassing single-view or no segmentation baselines by large margins. To our knowledge, this is the first publicly reported system that unifies multi-camera feature validation with defect detection in a deployable automotive setting in industry.

Method: Proposes a KAN-based detection method that replaces fixed activation functions with learnable splines. Introduces LAKAN module that uses facial landmarks as structural priors to dynamically generate KAN parameters, creating instance-specific signals to guide attention to artifact-rich facial regions.

Result: Extensive experiments on multiple public datasets demonstrate superior performance compared to existing methods.

Conclusion: The combination of geometric priors (facial landmarks) with KAN’s flexible architecture creates an effective approach for face forgery detection that outperforms current state-of-the-art methods.

Abstract: The rapid development of deepfake generation techniques necessitates robust face forgery detection algorithms. While methods based on Convolutional Neural Networks (CNNs) and Transformers are effective, there is still room for improvement in modeling the highly complex and non-linear nature of forgery artifacts. To address this issue, we propose a novel detection method based on the Kolmogorov-Arnold Network (KAN). By replacing fixed activation functions with learnable splines, our KAN-based approach is better suited to this challenge. Furthermore, to guide the network’s focus towards critical facial areas, we introduce a Landmark-assisted Adaptive Kolmogorov-Arnold Network (LAKAN) module. This module uses facial landmarks as a structural prior to dynamically generate the internal parameters of the KAN, creating an instance-specific signal that steers a general-purpose image encoder towards the most informative facial regions with artifacts. This core innovation creates a powerful combination between geometric priors and the network’s learning process. Extensive experiments on multiple public datasets show that our proposed method achieves superior performance.

Method: Multi-task framework with specialized decoders for room layout, camera pose, depth, and region segmentation. Pose-aware background depth resolving (PA-BDR) computes background depth using camera pose. Fusion mask generation (FMG) creates weight maps for depth correction. Adaptive fusion integrates refined background depth with initial predictions.

Result: Superior performance on Matterport3D, Structured3D, and Replica datasets compared to current open-source methods. Code is publicly available.

Conclusion: The proposed pose-aware geometry-constrained framework effectively addresses panoramic depth estimation by leveraging background depth as geometric prior, demonstrating significant improvements over existing methods.

Abstract: Explicitly modeling room background depth as a geometric constraint has proven effective for panoramic depth estimation. However, reconstructing this background depth for regular enclosed regions in a complex indoor scene without external measurements remains an open challenge. To address this, we propose a pose-aware and geometry-constrained framework for panoramic depth estimation. Our framework first employs multiple task-specific decoders to jointly estimate room layout, camera pose, depth, and region segmentation from a input panoramic image. A pose-aware background depth resolving (PA-BDR) component uses tasks decoder’s prediction to resolve the camera pose. Subsequently, the proposed PA-BDR component uses the camera pose to compute the background depth of regular enclosed regions and uses this background depth as a strong geometric prior. Based on the output of the region segmentation decoder, a fusion mask generation (FMG) component produces a fusion weight map to guide where and to what extent the geometry-constrained background depth should correct the depth decoder’s prediction. Finally, an adaptive fusion component integrates this refined background depth with the initial depth prediction, guided by the fusion weight. Extensive experiments on Matterport3D, Structured3D, and Replica datasets demonstrate that our method achieves significantly superior performance compared to current open-source methods. Code is available at https://github.com/emiyaning/PAGCNet.

Method: Comparative analysis of object tags (denoting concrete entities) versus abstract tags (capturing higher-level contextual information) for image privacy classification tasks, examining performance under different tag budget constraints.

Result: Abstract tags are more effective than object tags for privacy classification when the tag budget is limited. However, when a larger number of tags per image is available, object-related information becomes as useful as abstract tags.

Conclusion: The findings provide guidance for developing more accurate image privacy classifiers by considering the role of tag types and quantity, suggesting abstract tags should be prioritized in resource-constrained scenarios.

Abstract: Object tags denote concrete entities and are central to many computer vision tasks, whereas abstract tags capture higher-level information, which is relevant for tasks that require a contextual, potentially subjective scene understanding. Object and abstract tags extracted from images also facilitate interpretability. In this paper, we explore which type of tags is more suitable for the context-dependent and inherently subjective task of image privacy. While object tags are generally used for privacy classification, we show that abstract tags are more effective when the tag budget is limited. Conversely, when a larger number of tags per image is available, object-related information is as useful. We believe that these findings will guide future research in developing more accurate image privacy classifiers, informed by the role of tag types and quantity.

Method: Develop FlashAttention-2 JVP kernel for parallelism-compatible training on 10B+ parameter models; propose rCM that incorporates score distillation as a long-skip regularizer to complement sCM’s forward-divergence objective with reverse divergence, improving quality while maintaining diversity.

Result: rCM matches state-of-the-art DMD2 on quality metrics while mitigating mode collapse, offering better diversity; achieves 15-50× acceleration with 1-4 sampling steps; validated on models up to 14B parameters and 5-second videos.

Conclusion: rCM provides a practical, theoretically grounded framework for large-scale diffusion distillation that maintains both quality and diversity without extensive hyperparameter tuning or GAN components.

Abstract: Although continuous-time consistency models (e.g., sCM, MeanFlow) are theoretically principled and empirically powerful for fast academic-scale diffusion, its applicability to large-scale text-to-image and video tasks remains unclear due to infrastructure challenges in Jacobian-vector product (JVP) computation and the limitations of evaluation benchmarks like FID. This work represents the first effort to scale up continuous-time consistency to general application-level image and video diffusion models, and to make JVP-based distillation effective at large scale. We first develop a parallelism-compatible FlashAttention-2 JVP kernel, enabling sCM training on models with over 10 billion parameters and high-dimensional video tasks. Our investigation reveals fundamental quality limitations of sCM in fine-detail generation, which we attribute to error accumulation and the “mode-covering” nature of its forward-divergence objective. To remedy this, we propose the score-regularized continuous-time consistency model (rCM), which incorporates score distillation as a long-skip regularizer. This integration complements sCM with the “mode-seeking” reverse divergence, effectively improving visual quality while maintaining high generation diversity. Validated on large-scale models (Cosmos-Predict2, Wan2.1) up to 14B parameters and 5-second videos, rCM generally matches the state-of-the-art distillation method DMD2 on quality metrics while mitigating mode collapse and offering notable advantages in diversity, all without GAN tuning or extensive hyperparameter searches. The distilled models generate high-fidelity samples in only $1\sim4$ steps, accelerating diffusion sampling by $15\times\sim50\times$. These results position rCM as a practical and theoretically grounded framework for advancing large-scale diffusion distillation. Code is available at https://github.com/NVlabs/rcm.

Method: Proposes an inference-time feature-agnostic upsampling architecture that can be applied to any vision feature without encoder-specific training.

Result: AnyUp sets new state-of-the-art for upsampled features, generalizes to different feature types, preserves feature semantics, and is efficient for downstream tasks.

Conclusion: AnyUp provides a versatile, high-quality upsampling solution that works across different vision features without requiring retraining for each encoder.

Abstract: We introduce AnyUp, a method for feature upsampling that can be applied to any vision feature at any resolution, without encoder-specific training. Existing learning-based upsamplers for features like DINO or CLIP need to be re-trained for every feature extractor and thus do not generalize to different feature types at inference time. In this work, we propose an inference-time feature-agnostic upsampling architecture to alleviate this limitation and improve upsampling quality. In our experiments, AnyUp sets a new state of the art for upsampled features, generalizes to different feature types, and preserves feature semantics while being efficient and easy to apply to a wide range of downstream tasks.

Method: Proposes Scene De-Contextualization (SDeC), a training-free prompt embedding editing approach that identifies and suppresses latent scene-ID correlation within ID prompt embeddings. Uses SVD directional stability to quantify and adaptively re-weight corresponding eigenvalues, implementing an inversion process of T2I’s built-in scene contextualization.

Result: SDeC significantly enhances identity preservation while maintaining scene diversity, works with per-scene use (one scene per prompt) without requiring prior access to all target scenes, making it flexible for real-world applications.

Conclusion: SDeC provides an efficient, training-free solution to the identity shift problem in T2I generation by addressing the fundamental issue of scene contextualization, offering practical advantages over previous methods that require unrealistic assumptions about target scenes.

Abstract: Consistent text-to-image (T2I) generation seeks to produce identity-preserving images of the same subject across diverse scenes, yet it often fails due to a phenomenon called identity (ID) shift. Previous methods have tackled this issue, but typically rely on the unrealistic assumption of knowing all target scenes in advance. This paper reveals that a key source of ID shift is the native correlation between subject and scene context, called scene contextualization, which arises naturally as T2I models fit the training distribution of vast natural images. We formally prove the near-universality of this scene-ID correlation and derive theoretical bounds on its strength. On this basis, we propose a novel, efficient, training-free prompt embedding editing approach, called Scene De-Contextualization (SDeC), that imposes an inversion process of T2I’s built-in scene contextualization. Specifically, it identifies and suppresses the latent scene-ID correlation within the ID prompt’s embedding by quantifying the SVD directional stability to adaptively re-weight the corresponding eigenvalues. Critically, SDeC allows for per-scene use (one scene per prompt) without requiring prior access to all target scenes. This makes it a highly flexible and general solution well-suited to real-world applications where such prior knowledge is often unavailable or varies over time. Experiments demonstrate that SDeC significantly enhances identity preservation while maintaining scene diversity.

Method: Created PRISMM-Bench through multi-stage pipeline: mining real reviewer comments, LLM-assisted filtering, human verification to curate 384 inconsistencies from 353 papers. Designed three tasks: inconsistency identification, remedy, and pair matching. Introduced structured JSON-based answer representations to minimize linguistic biases in multiple-choice evaluation.

Result: Benchmarked 21 leading LMMs including large open-weight models (GLM-4.5V 106B, InternVL3 78B) and proprietary models (Gemini 2.5 Pro, GPT-5). Results show strikingly low performance (27.8-53.9%), demonstrating the challenge of multimodal scientific reasoning.

Conclusion: Current LMMs struggle with detecting and resolving real-world multimodal inconsistencies in scientific papers, highlighting the need for progress toward trustworthy scientific assistants. The benchmark reveals significant gaps in multimodal reasoning capabilities.

Abstract: Large Multimodal Models (LMMs) are increasingly applied to scientific research, yet it remains unclear whether they can reliably understand and reason over the multimodal complexity of papers. A central challenge lies in detecting and resolving inconsistencies across text, figures, tables, and equations, issues that are often subtle, domain-specific, and ultimately undermine clarity, reproducibility, and trust. Existing benchmarks overlook this issue, either isolating single modalities or relying on synthetic errors that fail to capture real-world complexity. We introduce PRISMM-Bench (Peer-Review-sourced Inconsistency Set for Multimodal Models), the first benchmark grounded in real reviewer-flagged inconsistencies in scientific papers. Through a multi-stage pipeline of review mining, LLM-assisted filtering and human verification, we curate 384 inconsistencies from 353 papers. Based on this set, we design three tasks, namely inconsistency identification, remedy and pair matching, which assess a model’s capacity to detect, correct, and reason over inconsistencies across different modalities. Furthermore, to address the notorious problem of choice-only shortcuts in multiple-choice evaluation, where models exploit answer patterns without truly understanding the question, we further introduce structured JSON-based answer representations that minimize linguistic biases by reducing reliance on superficial stylistic cues. We benchmark 21 leading LMMs, including large open-weight models (GLM-4.5V 106B, InternVL3 78B) and proprietary models (Gemini 2.5 Pro, GPT-5 with high reasoning). Results reveal strikingly low performance (27.8-53.9%), underscoring the challenge of multimodal scientific reasoning and motivating progress towards trustworthy scientific assistants.

Method: The researchers conducted controlled experiments using synthetic datasets across multiple architectures. They tested leading frontier models on tasks like reproducing iconic movie artwork versus providing textual descriptions, and demonstrated safety vulnerabilities by showing models aligned solely on text can still generate unsafe images.

Result: Models can generate near-perfect reproductions of iconic movie artwork but confuse crucial details in textual descriptions. Modal aphasia emerges as a fundamental property of current unified multimodal models, not just a training artifact. Safety frameworks are vulnerable when safeguards are applied to only one modality.

Conclusion: Modal aphasia is a systematic dissociation in current multimodal models that creates significant safety vulnerabilities. This highlights the need for more robust multimodal alignment approaches that ensure consistency across modalities rather than applying safeguards to only one modality.

Abstract: We present modal aphasia, a systematic dissociation in which current unified multimodal models accurately memorize concepts visually but fail to articulate them in writing, despite being trained on images and text simultaneously. For one, we show that leading frontier models can generate near-perfect reproductions of iconic movie artwork, but confuse crucial details when asked for textual descriptions. We corroborate those findings through controlled experiments on synthetic datasets in multiple architectures. Our experiments confirm that modal aphasia reliably emerges as a fundamental property of current unified multimodal models, not just as a training artifact. In practice, modal aphasia can introduce vulnerabilities in AI safety frameworks, as safeguards applied to one modality may leave harmful concepts accessible in other modalities. We demonstrate this risk by showing how a model aligned solely on text remains capable of generating unsafe images.

Method: Proposes Top-Down Semantic Refinement (TDSR) modeling captioning as MDP with efficient MCTS featuring visual-guided parallel expansion and lightweight value network.

Result: Achieves SOTA/competitive results on DetailCaps, COMPOSITIONCAP, POPE benchmarks, enhancing LLaVA-1.5, Qwen2.5-VL with 10x reduction in VLM calls.

Conclusion: TDSR effectively addresses VLMs’ coherence-detail tradeoff through hierarchical planning, offering plug-and-play improvement for existing models.

Abstract: Large Vision-Language Models (VLMs) face an inherent contradiction in image captioning: their powerful single-step generation capabilities often lead to a myopic decision-making process. This makes it difficult to maintain global narrative coherence while capturing rich details, a limitation that is particularly pronounced in tasks that require multi-step and complex scene description. To overcome this fundamental challenge, we redefine image captioning as a goal-oriented hierarchical refinement planning problem, and further propose a novel framework, named Top-Down Semantic Refinement (TDSR), which models the generation process as a Markov Decision Process (MDP). However, planning within the vast state space of a VLM presents a significant computational hurdle. Our core contribution, therefore, is the design of a highly efficient Monte Carlo Tree Search (MCTS) algorithm tailored for VLMs. By incorporating a visual-guided parallel expansion and a lightweight value network, our TDSR reduces the call frequency to the expensive VLM by an order of magnitude without sacrificing planning quality. Furthermore, an adaptive early stopping mechanism dynamically matches computational overhead to the image’s complexity. Extensive experiments on multiple benchmarks, including DetailCaps, COMPOSITIONCAP, and POPE, demonstrate that our TDSR, as a plug-and-play module, can significantly enhance the performance of existing VLMs (e.g., LLaVA-1.5, Qwen2.5-VL) by achieving state-of-the-art or highly competitive results in fine-grained description, compositional generalization, and hallucination suppression.

Method: Used state-of-the-art architectures (DenseNet121, SwinV2-B, MedMamba) trained on chest X-rays from MIMIC-CXR-JPG and CheXpert datasets to predict health insurance type as a proxy for socioeconomic status. Conducted ablation studies combining demographic features, single-racial group training, and patch-based occlusion analysis.

Result: Models achieved significant accuracy (AUC around 0.70 on MIMIC-CXR-JPG, 0.68 on CheXpert) in predicting health insurance type. Signal persists when controlling for demographics and within single racial groups. Patch occlusion revealed diffuse signal in upper and mid-thoracic regions.

Conclusion: Deep networks internalize subtle traces of clinical environments, equipment differences, or care pathways, learning socioeconomic segregation. This challenges medical image neutrality and reframes AI fairness as interrogating social fingerprints in clinical data.

Abstract: Artificial intelligence is revealing what medicine never intended to encode. Deep vision models, trained on chest X-rays, can now detect not only disease but also invisible traces of social inequality. In this study, we show that state-of-the-art architectures (DenseNet121, SwinV2-B, MedMamba) can predict a patient’s health insurance type, a strong proxy for socioeconomic status, from normal chest X-rays with significant accuracy (AUC around 0.70 on MIMIC-CXR-JPG, 0.68 on CheXpert). The signal was unlikely contributed by demographic features by our machine learning study combining age, race, and sex labels to predict health insurance types; it also remains detectable when the model is trained exclusively on a single racial group. Patch-based occlusion reveals that the signal is diffuse rather than localized, embedded in the upper and mid-thoracic regions. This suggests that deep networks may be internalizing subtle traces of clinical environments, equipment differences, or care pathways; learning socioeconomic segregation itself. These findings challenge the assumption that medical images are neutral biological data. By uncovering how models perceive and exploit these hidden social signatures, this work reframes fairness in medical AI: the goal is no longer only to balance datasets or adjust thresholds, but to interrogate and disentangle the social fingerprints embedded in clinical data itself.

Method: Proposes Action Effect Modeling (AEM) with: 1) Effect frame selection based on semantic relevance and visual quality, 2) Extraction of complementary cues from visual grounding and symbolic scene graphs, 3) Alignment in shared latent space for effect-aware representations, 4) Prompt-based detector with task-specific prompts aligned with intended execution semantics.

Result: Achieves state-of-the-art performance on EgoPER and CaptainCook4D benchmarks under challenging one-class classification (OCC) setting.

Conclusion: Modeling both execution and outcome yields more reliable mistake detection, and effect-aware representations have potential for broader downstream applications.

Abstract: Mistake detection in procedural tasks is essential for building intelligent systems that support learning and task execution. Existing approaches primarily analyze how an action is performed, while overlooking what it produces, i.e., the \textbf{action effect}. Yet many errors manifest not in the execution itself but in the resulting outcome, such as an unintended object state or incorrect spatial arrangement. To address this gap, we propose Action Effect Modeling (AEM), a unified framework that jointly captures action execution and its outcomes through a probabilistic formulation. AEM first identifies the outcome of an action by selecting the most informative effect frame based on semantic relevance and visual quality. It then extracts complementary cues from visual grounding and symbolic scene graphs, aligning them in a shared latent space to form robust effect-aware representations. To detect mistakes, we further design a prompt-based detector that incorporates task-specific prompts and aligns each action segment with its intended execution semantics. Our approach achieves state-of-the-art performance on the EgoPER and CaptainCook4D benchmarks under the challenging one-class classification (OCC) setting. These results demonstrate that modeling both execution and outcome yields more reliable mistake detection, and highlight the potential of effect-aware representations to benefit a broader range of downstream applications.

Method: Two-stage semantic-guided hierarchical synthesis: 1) Semantic layout generation combining local features (CNNs) and global features (Vision Transformers) to create clear semantic layouts, 2) Multi-Modal Texture Generator that refines layouts using multi-scale information with dynamic attention handling arbitrary masks without mask-specific training.

Result: Outperforms state-of-the-art methods on CelebA-HQ and FFHQ datasets, showing improvements in LPIPS, PSNR, and SSIM metrics, producing visually striking results with better semantic preservation in challenging large-area inpainting situations.

Conclusion: The proposed semantic-guided hierarchical synthesis approach effectively addresses facial image inpainting challenges, particularly for large irregular masks, by leveraging facial priors through a two-stage architecture that separates semantic understanding from texture refinement.

Abstract: Facial Image inpainting aim is to restore the missing or corrupted regions in face images while preserving identity, structural consistency and photorealistic image quality, a task specifically created for photo restoration. Though there are recent lot of advances in deep generative models, existing methods face problems with large irregular masks, often producing blurry textures on the edges of the masked region, semantic inconsistencies, or unconvincing facial structures due to direct pixel level synthesis approach and limited exploitation of facial priors. In this paper we propose a novel architecture, which address these above challenges through semantic-guided hierarchical synthesis. Our approach starts with a method that organizes and synthesizes information based on meaning, followed by refining the texture. This process gives clear insights into the facial structure before we move on to creating detailed images. In the first stage, we blend two techniques: one that focuses on local features with CNNs and global features with Vision Transformers. This helped us create clear and detailed semantic layouts. In the second stage, we use a Multi-Modal Texture Generator to refine these layouts by pulling in information from different scales, ensuring everything looks cohesive and consistent. The architecture naturally handles arbitrary mask configurations through dynamic attention without maskspecific training. Experiment on two datasets CelebA-HQ and FFHQ shows that our model outperforms other state-of-the-art methods, showing improvements in metrics like LPIPS, PSNR, and SSIM. It produces visually striking results with better semantic preservation, in challenging large-area inpainting situations.

Method: Uses 2D Haar DWT on PAN images to localize spatial edges/textures, and channel-wise 1D Haar DWT on MS images to separate low/high-frequency spectral components. Features two parallel branches (Spectral and Spatial) that exchange information through Mamba-based cross-modulation with linear complexity, followed by multi-scale dynamic gate fusion.

Result: Outperforms recent baselines (FusionMamba, CANNet, U2Net, ARConv) on WV3, GF2, and QB datasets, improving PSNR by up to 0.23 dB and achieving HQNR 0.956 on full-resolution WV3. Ablations validate design choices.

Conclusion: S2WMamba effectively disentangles frequency information for pansharpening through explicit wavelet decomposition and lightweight cross-modal interaction, achieving state-of-the-art performance with efficient long-range dependency modeling.

Abstract: Pansharpening fuses a high-resolution PAN image with a low-resolution multispectral (LRMS) image to produce an HRMS image. A key difficulty is that jointly processing PAN and MS often entangles spatial detail with spectral fidelity. We propose S2WMamba, which explicitly disentangles frequency information and then performs lightweight cross-modal interaction. Concretely, a 2D Haar DWT is applied to PAN to localize spatial edges and textures, while a channel-wise 1D Haar DWT treats each pixel’s spectrum as a 1D signal to separate low/high-frequency components and limit spectral distortion. The resulting Spectral branch injects wavelet-extracted spatial details into MS features, and the Spatial branch refines PAN features using spectra from the 1D pyramid; the two branches exchange information through Mamba-based cross-modulation that models long-range dependencies with linear complexity. A multi-scale dynamic gate (multiplicative + additive) then adaptively fuses branch outputs.On WV3, GF2, and QB, S2WMamba matches or surpasses recent strong baselines (FusionMamba, CANNet, U2Net, ARConv), improving PSNR by up to 0.23 dB and reaching HQNR 0.956 on full-resolution WV3. Ablations justify the choice of 2D/1D DWT placement, parallel dual branches, and the fusion gate. Our code is available at https://github.com/KagUYa66/S2WMamba.

Method: Proposes Fourier-RWKV framework based on Multi-State Perception: (1) Spatial-form Perception via Deformable Quad-directional Token Shift for local haze variations, (2) Frequency-domain Perception via Fourier Mix block extending RWKV’s WKV attention to Fourier domain for long-range dependencies, (3) Semantic-relation Perception via Semantic Bridge Module with Dynamic Semantic Kernel Fusion for encoder-decoder alignment.

Result: Extensive experiments on multiple benchmarks demonstrate state-of-the-art performance across diverse haze scenarios while significantly reducing computational overhead, establishing favorable trade-off between restoration quality and practical efficiency.

Conclusion: Fourier-RWKV delivers comprehensive haze degradation modeling with linear complexity, achieving superior dehazing performance while being computationally efficient for practical deployment.

Abstract: Image dehazing is crucial for reliable visual perception, yet it remains highly challenging under real-world non-uniform haze conditions. Although Transformer-based methods excel at capturing global context, their quadratic computational complexity hinders real-time deployment. To address this, we propose Fourier Receptance Weighted Key Value (Fourier-RWKV), a novel dehazing framework based on a Multi-State Perception paradigm. The model achieves comprehensive haze degradation modeling with linear complexity by synergistically integrating three distinct perceptual states: (1) Spatial-form Perception, realized through the Deformable Quad-directional Token Shift (DQ-Shift) operation, which dynamically adjusts receptive fields to accommodate local haze variations; (2) Frequency-domain Perception, implemented within the Fourier Mix block, which extends the core WKV attention mechanism of RWKV from the spatial domain to the Fourier domain, preserving the long-range dependencies essential for global haze estimation while mitigating spatial attenuation; (3) Semantic-relation Perception, facilitated by the Semantic Bridge Module (SBM), which utilizes Dynamic Semantic Kernel Fusion (DSK-Fusion) to precisely align encoder-decoder features and suppress artifacts. Extensive experiments on multiple benchmarks demonstrate that Fourier-RWKV delivers state-of-the-art performance across diverse haze scenarios while significantly reducing computational overhead, establishing a favorable trade-off between restoration quality and practical efficiency. Code is available at: https://github.com/Dilizlr/Fourier-RWKV.

Method: Proposes Δ-LFM framework with two key components: 1) Treats disease dynamics as a velocity field using Flow Matching (FM) to align temporal evolution of patient data, capturing intrinsic disease dynamics; 2) Learns patient-specific latent alignment that enforces patient trajectories to lie along specific axes with magnitude increasing monotonically with disease severity, creating consistent and semantically meaningful latent space.

Result: Demonstrates strong empirical performance across three longitudinal MRI benchmarks, offering a new framework for interpreting and visualizing disease dynamics with improved continuity and semantic structure.

Conclusion: Δ-LFM provides an effective approach for modeling patient-specific latent progression with flow matching, offering interpretable disease dynamics visualization and progression modeling that aligns with clinical 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: Introduces DepthMatch-ControlNet and LiDARMatch-ControlNet: matching-specific controllable 2D generative models. DepthMatch-ControlNet uses depth-conditioned generation for perspective-view RGB images consistent with depth maps. LiDARMatch-ControlNet extends this to LiDAR by conditioning on equirectangular range maps for panoramic RGB generation. Both incorporate coupled conditional denoising and prompt guidance for cross-view consistency.

Result: Extensive experiments on 3DMatch, ScanNet (depth-camera), and Dur360BEV (LiDAR) datasets demonstrate the effectiveness of the approach in improving registration performance across various settings.

Conclusion: The proposed generative 3D registration paradigm successfully bridges 2D generative models with 3D matching tasks, providing a general framework that can be integrated into existing registration methods to enhance performance through geometry-color feature fusion.

Abstract: In this paper, we propose a novel 3D registration paradigm, Generative Point Cloud Registration, which bridges advanced 2D generative models with 3D matching tasks to enhance registration performance. Our key idea is to generate cross-view consistent image pairs that are well-aligned with the source and target point clouds, enabling geometry-color feature fusion to facilitate robust matching. To ensure high-quality matching, the generated image pair should feature both 2D-3D geometric consistency and cross-view texture consistency. To this end, we introduce DepthMatch-ControlNet and LiDARMatch-ControlNet, two matching-specific, controllable 2D generative models. Specifically, for depth camera-based 3D registration with point clouds derived from the depth maps, DepthMatch-ControlNet leverages the depth-conditioned generation capabilities of ControlNet to synthesize perspective-view RGB images that are geometrically consistent with depth maps, ensuring accurate 2D-3D alignment. Additionally, by incorporating a coupled conditional denoising scheme and coupled prompt guidance, it further promotes cross-view feature interaction, guiding texture consistency generation. To address LiDAR-based 3D registration with point clouds captured by LiDAR sensors, LiDARMatch-ControlNet extends this framework by conditioning on paired equirectangular range maps projected from 360-degree LiDAR point clouds, generating corresponding panoramic RGB images. Our generative 3D registration paradigm is general and can be seamlessly integrated into a wide range of existing registration methods to improve their performance. Extensive experiments on the 3DMatch and ScanNet datasets (for depth-camera settings), as well as the Dur360BEV dataset (for LiDAR settings), demonstrate the effectiveness of our approach.

Method: Introduces ISA-ViT (input-size-agnostic Vision Transformer) that resizes UWB data while preserving radar-specific information like Doppler shifts and phase characteristics. Uses adjustable patch configurations and pre-trained positional embedding vectors, plus domain fusion combining range- and frequency-domain features.

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

Conclusion: The work facilitates development of robust distracted driving detection systems by providing ALERT dataset and ISA-ViT framework that overcomes input dimension limitations for radar data in Vision Transformers.

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: Two-step approach: 1) Rule-based mask warping, 2) Unpaired image-to-image translation using GANs. Introduces non-mask preservation loss and stochastic noise injection to stabilize training and enhance sample diversity.

Result: Consistent qualitative improvements over rule-based warping alone, complements existing GAN-based methods like IAMGAN. The proposed components effectively enhance sample diversity and training stability.

Conclusion: The framework provides effective data-centric augmentation for face recognition tasks, suggesting directions for future improvements in generative data augmentation for masked face scenarios.

Abstract: Data scarcity and distribution shift pose major challenges for masked face detection and recognition. We propose a two-step generative data augmentation framework that combines rule-based mask warping with unpaired image-to-image translation using GANs, enabling the generation of realistic masked-face samples beyond purely synthetic transformations. Compared to rule-based warping alone, the proposed approach yields consistent qualitative improvements and complements existing GAN-based masked face generation methods such as IAMGAN. We introduce a non-mask preservation loss and stochastic noise injection to stabilize training and enhance sample diversity. Experimental observations highlight the effectiveness of the proposed components and suggest directions for future improvements in data-centric augmentation for face recognition tasks.

Method: Proposes a data synthesis pipeline: 1) Train diffusion model on normal X-rays, 2) Use pre-trained model to inpaint head lesions in diseased X-rays while preserving tail classes as augmented data, 3) Integrate Large Language Model Knowledge Guidance (LKG) module and Progressive Incremental Learning (PIL) strategy to stabilize fine-tuning.

Result: Comprehensive evaluations on public lung datasets MIMIC and CheXpert demonstrate the method sets a new benchmark in performance.

Conclusion: The proposed approach effectively addresses the data scarcity problem for rare lung lesions through innovative data synthesis using normal X-rays and diffusion models with language model guidance.

Abstract: Long-tailed pulmonary anomalies in chest radiography present formidable diagnostic challenges. Despite the recent strides in diffusion-based methods for enhancing the representation of tailed lesions, the paucity of rare lesion exemplars curtails the generative capabilities of these approaches, thereby leaving the diagnostic precision less than optimal. In this paper, we propose a novel data synthesis pipeline designed to augment tail lesions utilizing a copious supply of conventional normal X-rays. Specifically, a sufficient quantity of normal samples is amassed to train a diffusion model capable of generating normal X-ray images. This pre-trained diffusion model is subsequently utilized to inpaint the head lesions present in the diseased X-rays, thereby preserving the tail classes as augmented training data. Additionally, we propose the integration of a Large Language Model Knowledge Guidance (LKG) module alongside a Progressive Incremental Learning (PIL) strategy to stabilize the inpainting fine-tuning process. Comprehensive evaluations conducted on the public lung datasets MIMIC and CheXpert demonstrate that the proposed method sets a new benchmark in performance.

Method: Uses Structured Latent (SLAT) representations with two key components: Morphing Cross-Attention (MCA) for fusing source and target information for structural coherence, and Temporal-Fused Self-Attention (TFSA) for enhancing temporal consistency by incorporating features from preceding frames. Also includes orientation correction to mitigate pose ambiguity.

Result: Generates state-of-the-art morphing sequences, even for challenging cross-category cases. Supports advanced applications like decoupled morphing and 3D style transfer, and can generalize to other SLAT-based generative models.

Conclusion: MorphAny3D provides an effective training-free framework for high-quality 3D morphing by intelligently blending SLAT features within attention mechanisms, achieving superior results for both within-category and cross-category morphing tasks.

Abstract: 3D morphing remains challenging due to the difficulty of generating semantically consistent and temporally smooth deformations, especially across categories. We present MorphAny3D, a training-free framework that leverages Structured Latent (SLAT) representations for high-quality 3D morphing. Our key insight is that intelligently blending source and target SLAT features within the attention mechanisms of 3D generators naturally produces plausible morphing sequences. To this end, we introduce Morphing Cross-Attention (MCA), which fuses source and target information for structural coherence, and Temporal-Fused Self-Attention (TFSA), which enhances temporal consistency by incorporating features from preceding frames. An orientation correction strategy further mitigates the pose ambiguity within the morphing steps. Extensive experiments show that our method generates state-of-the-art morphing sequences, even for challenging cross-category cases. MorphAny3D further supports advanced applications such as decoupled morphing and 3D style transfer, and can be generalized to other SLAT-based generative models. Project page: https://xiaokunsun.github.io/MorphAny3D.github.io/.

Method: Proposes Clifford Algebra Network (CliffordNet) using Clifford Geometric Product (uv = u·v + u∧v) as a unified interaction mechanism that simultaneously captures feature coherence (via generalized inner product) and structural variation (via exterior wedge product). Implemented with efficient sparse rolling mechanism with O(N) linear complexity.

Result: CliffordNet establishes new Pareto frontier: Nano variant achieves 77.82% accuracy on CIFAR-100 with only 1.4M parameters (8× fewer than ResNet-18’s 11.2M), Lite variant (2.6M) sets new SOTA for tiny models at 79.05%. The geometric interaction is so representationally dense that standard FFNs become redundant.

Conclusion: Global understanding can emerge solely from rigorous, algebraically complete local interactions, suggesting a potential paradigm shift where “geometry is all you need” in vision architectures.

Abstract: Modern computer vision architectures, from CNNs to Transformers, predominantly rely on the stacking of heuristic modules: spatial mixers (Attention/Conv) followed by channel mixers (FFNs). In this work, we challenge this paradigm by returning to mathematical first principles. We propose the Clifford Algebra Network (CAN), also referred to as CliffordNet, a vision backbone grounded purely in Geometric Algebra. Instead of engineering separate modules for mixing and memory, we derive a unified interaction mechanism based on the Clifford Geometric Product ($uv = u \cdot v + u \wedge v$). This operation ensures algebraic completeness regarding the Geometric Product by simultaneously capturing feature coherence (via the generalized inner product) and structural variation (via the exterior wedge product). Implemented via an efficient sparse rolling mechanism with strict linear complexity $O(N)$, our model reveals a surprising emergent property: the geometric interaction is so representationally dense that standard Feed-Forward Networks (FFNs) become redundant. Empirically, CliffordNet establishes a new Pareto frontier: our Nano variant achieves 77.82% accuracy on CIFAR-100 with only 1.4M parameters, effectively matching the heavy-weight ResNet-18 (11.2M) with $8\times$ fewer parameters, while our Lite variant (2.6M) sets a new SOTA for tiny models at 79.05%. Our results suggest that global understanding can emerge solely from rigorous, algebraically complete local interactions, potentially signaling a shift where geometry is all you need. Code is available at https://github.com/ParaMind2025/CAN.

Method: Integrates 3D-aware features from MUSt3R into SAM2 using a lightweight Feature Merger that fuses multi-level MUSt3R features encoding implicit geometric correspondence. Combines these with SAM2’s appearance features and uses field-of-view aware sampling to ensure frames observe spatially consistent object regions for reliable 3D correspondence learning.

Result: On challenging datasets with wide-baseline motion (ScanNet++, Replica), 3AM substantially outperforms SAM2 and extensions, achieving 90.6% IoU and 71.7% Positive IoU on ScanNet++’s Selected Subset, improving over state-of-the-art VOS methods by +15.9 and +30.4 points.

Conclusion: 3AM successfully enhances video object segmentation with 3D-aware features, achieving geometry-consistent recognition without requiring camera poses or preprocessing at inference, making it practical for real-world applications.

Abstract: Video object segmentation methods like SAM2 achieve strong performance through memory-based architectures but struggle under large viewpoint changes due to reliance on appearance features. Traditional 3D instance segmentation methods address viewpoint consistency but require camera poses, depth maps, and expensive preprocessing. We introduce 3AM, a training-time enhancement that integrates 3D-aware features from MUSt3R into SAM2. Our lightweight Feature Merger fuses multi-level MUSt3R features that encode implicit geometric correspondence. Combined with SAM2’s appearance features, the model achieves geometry-consistent recognition grounded in both spatial position and visual similarity. We propose a field-of-view aware sampling strategy ensuring frames observe spatially consistent object regions for reliable 3D correspondence learning. Critically, our method requires only RGB input at inference, with no camera poses or preprocessing. On challenging datasets with wide-baseline motion (ScanNet++, Replica), 3AM substantially outperforms SAM2 and extensions, achieving 90.6% IoU and 71.7% Positive IoU on ScanNet++’s Selected Subset, improving over state-of-the-art VOS methods by +15.9 and +30.4 points. Project page: https://jayisaking.github.io/3AM-Page/

Method: EGM (Efficient visual Grounding language Models) scales test-time computation by increasing the number of generated tokens. This approach leverages that small models have cheaper per-token costs, making them more efficient than running large models directly.

Result: EGM-Qwen3-VL-8B achieves 91.4 IoU on RefCOCO with 737ms latency (5.9x faster than Qwen3-VL-235B’s 4,320ms for 90.5 IoU). The method also improves amodal grounding (predicting both visible and occluded object parts) and consistently boosts small models to match or outperform larger ones.

Conclusion: EGM enables small VLMs to achieve visual grounding performance comparable to large models while being significantly faster and more deployment-friendly through test-time computation scaling.

Abstract: Visual grounding is an essential capability of Visual Language Models (VLMs) to understand the real physical world. Previous state-of-the-art grounding visual language models usually have large model sizes, making them heavy for deployment and slow for inference. However, we notice that the sizes of visual encoders are nearly the same for small and large VLMs and the major difference is the sizes of the language models. Small VLMs fall behind larger VLMs in grounding because of the difference in language understanding capability rather than visual information handling. To mitigate the gap, we introduce ‘Efficient visual Grounding language Models’ (EGM): a method to scale the test-time computation (#generated tokens). Scaling the test-time computation of a small model is deployment-friendly, and yields better end-to-end latency as the cost of each token is much cheaper compared to directly running a large model. On the RefCOCO benchmark, our EGM-Qwen3-VL-8B demonstrates 91.4 IoU with an average of 737ms (5.9x faster) latency while Qwen3-VL-235B demands 4,320ms to achieve 90.5 IoU. To validate our approach’s generality, we further set up a new amodal grounding setting that requires the model to predict both the visible and occluded parts of the objects. Experiments show our method can consistently and significantly improve the vanilla grounding and amodal grounding capabilities of small models to be on par with or outperform the larger models, thereby improving the efficiency for visual grounding.

Method: Uses 2D axial, sagittal, and coronal projections to approximate 3D volumes. YOLOv8 identifies regions of interest from all views. DenseNet121-Unet performs multi-label segmentation using variance- and energy-based projections. Extracts individual vertebra volumes and analyzes fractures using ensemble of 2.5D Spatio-Sequential models with raw slices and projections.

Result: Achieves 3D mIoU of 94.45% for localization, Dice score of 87.86% for segmentation, and vertebra-level/patient-level F1 scores of 68.15/82.26 and ROC-AUC scores of 91.62/83.04 for fracture detection. Competitive with expert radiologists in interobserver analysis.

Conclusion: The 2D projection-based approach effectively reduces computational complexity while maintaining competitive performance for cervical spine fracture detection, validated through explainability studies and comparison with radiologists.

Abstract: Cervical spine fractures are critical medical conditions requiring precise and efficient detection for effective clinical management. This study explores the viability of 2D projection-based vertebra segmentation for vertebra-level fracture detection in 3D CT volumes, presenting an end-to-end pipeline for automated analysis of cervical vertebrae (C1-C7). By approximating a 3D volume through optimized 2D axial, sagittal, and coronal projections, regions of interest are identified using the YOLOv8 model from all views and combined to approximate the 3D cervical spine area, achieving a 3D mIoU of 94.45 percent. This projection-based localization strategy reduces computational complexity compared to traditional 3D segmentation methods while maintaining high performance. It is followed by a DenseNet121-Unet-based multi-label segmentation leveraging variance- and energy-based projections, achieving a Dice score of 87.86 percent. Strategic approximation of 3D vertebral masks from these 2D segmentation masks enables the extraction of individual vertebra volumes. The volumes are analyzed for fractures using an ensemble of 2.5D Spatio-Sequential models incorporating both raw slices and projections per vertebra for complementary evaluation. This ensemble achieves vertebra-level and patient-level F1 scores of 68.15 and 82.26, and ROC-AUC scores of 91.62 and 83.04, respectively. We further validate our approach through an explainability study that provides saliency map visualizations highlighting anatomical regions relevant for diagnosis, and an interobserver variability analysis comparing our model’s performance with expert radiologists, demonstrating competitive results.

Method: Two-stage approach: 1) Physically Aligned Normalization (PAN) performs closed-form illumination correction via Gray-world normalization, log-domain Retinex decomposition, and dynamic range recombination. 2) Geometric-Semantic Rectification Attention (GSRA) extends differential attention to cross-modal alignment, harmonizing depth-derived geometry with DINO-v2 semantic embeddings.

Result: Competitive performance in shadow removal with lower complexity and generalization to ambient lighting where traditional methods fail under multi-source illumination.

Conclusion: PhaSR effectively addresses shadow removal under diverse lighting conditions through dual-level prior alignment, enabling robust performance from single-light shadows to multi-source ambient lighting.

Abstract: Shadow removal under diverse lighting conditions requires disentangling illumination from intrinsic reflectance, a challenge compounded when physical priors are not properly aligned. We propose PhaSR (Physically Aligned Shadow Removal), addressing this through dual-level prior alignment to enable robust performance from single-light shadows to multi-source ambient lighting. First, Physically Aligned Normalization (PAN) performs closed-form illumination correction via Gray-world normalization, log-domain Retinex decomposition, and dynamic range recombination, suppressing chromatic bias. Second, Geometric-Semantic Rectification Attention (GSRA) extends differential attention to cross-modal alignment, harmonizing depth-derived geometry with DINO-v2 semantic embeddings to resolve modal conflicts under varying illumination. Experiments show competitive performance in shadow removal with lower complexity and generalization to ambient lighting where traditional methods fail under multi-source illumination. Our source code is available at https://github.com/ming053l/PhaSR.

Method: Proposes Semantic-Guided Dynamic Sparsification (SGDS) that proactively guides activation space by governing orientation and rank of subspaces through targeted sparsification. Encourages similar classes to share compact activation subspaces for knowledge transfer, while assigning non-overlapping subspaces to dissimilar classes to prevent interference.

Result: Extensive experiments on various benchmark datasets demonstrate state-of-the-art performance of SGDS in class-incremental learning tasks.

Conclusion: SGDS effectively mitigates interference in class-incremental learning without imposing rigid constraints on parameter space, achieving better balance between plasticity and stability through activation space sculpting.

Abstract: Class-Incremental Learning (CIL) requires a model to continually learn new classes without forgetting old ones. A common and efficient solution freezes a pre-trained model and employs lightweight adapters, whose parameters are often forced to be orthogonal to prevent inter-task interference. However, we argue that this parameter-constraining method is detrimental to plasticity. To this end, we propose Semantic-Guided Dynamic Sparsification (SGDS), a novel method that proactively guides the activation space by governing the orientation and rank of its subspaces through targeted sparsification. Specifically, SGDS promotes knowledge transfer by encouraging similar classes to share a compact activation subspace, while simultaneously preventing interference by assigning non-overlapping activation subspaces to dissimilar classes. By sculpting class-specific sparse subspaces in the activation space, SGDS effectively mitigates interference without imposing rigid constraints on the parameter space. Extensive experiments on various benchmark datasets demonstrate the state-of-the-art performance of SGDS.

Method: Proposes Q-Hawkeye with two main components: (1) Uncertainty-Aware Dynamic Optimization that estimates predictive uncertainty using variance of predicted scores across multiple rollouts and uses this uncertainty to reweight each sample’s update strength, and (2) Perception-Aware Optimization that constructs paired inputs of degraded images and their originals, introducing an Implicit Perception Loss to constrain quality judgments to genuine visual evidence.

Result: Extensive experiments show Q-Hawkeye outperforms state-of-the-art methods and generalizes better across multiple datasets.

Conclusion: Q-Hawkeye provides a more reliable RL-based visual policy optimization framework for image quality assessment by addressing both uncertainty and perception limitations in existing methods.

Abstract: Image Quality Assessment (IQA) predicts perceptual quality scores consistent with human judgments. Recent RL-based IQA methods built on MLLMs focus on generating visual quality descriptions and scores, ignoring two key reliability limitations: (i) although the model’s prediction stability varies significantly across training samples, existing GRPO-based methods apply uniform advantage weighting, thereby amplifying noisy signals from unstable samples in gradient updates; (ii) most works emphasize text-grounded reasoning over images while overlooking the model’s visual perception ability of image content. In this paper, we propose Q-Hawkeye, an RL-based reliable visual policy optimization framework that redesigns the learning signal through unified Uncertainty-Aware Dynamic Optimization and Perception-Aware Optimization. Q-Hawkeye estimates predictive uncertainty using the variance of predicted scores across multiple rollouts and leverages this uncertainty to reweight each sample’s update strength, stabilizing policy optimization. To strengthen perceptual reliability, we construct paired inputs of degraded images and their original images and introduce an Implicit Perception Loss that constrains the model to ground its quality judgments in genuine visual evidence. Extensive experiments demonstrate that Q-Hawkeye outperforms state-of-the-art methods and generalizes better across multiple datasets. Our dataset and code are available at https://github.com/AMAP-ML/Q-Hawkeye.

Method: Introduces ShotFinder benchmark with 1,210 samples across 20 categories, formalizing editing requirements as keyframe-oriented shot descriptions with five controllable constraints. Proposes a three-stage retrieval pipeline: query expansion via video imagination, candidate retrieval with search engine, and description-guided temporal localization.

Result: Experiments show significant gap between model and human performance, with clear imbalance across constraints: temporal localization is relatively tractable, while color and visual style remain major challenges.

Conclusion: Open-domain video shot retrieval is a critical capability that multimodal large models have yet to overcome, highlighting the need for better video understanding with temporal and multimodal constraints.

Abstract: In recent years, large language models (LLMs) have made rapid progress in information retrieval, yet existing research has mainly focused on text or static multimodal settings. Open-domain video shot retrieval, which involves richer temporal structure and more complex semantics, still lacks systematic benchmarks and analysis. To fill this gap, we introduce ShotFinder, a benchmark that formalizes editing requirements as keyframe-oriented shot descriptions and introduces five types of controllable single-factor constraints: Temporal order, Color, Visual style, Audio, and Resolution. We curate 1,210 high-quality samples from YouTube across 20 thematic categories, using large models for generation with human verification. Based on the benchmark, we propose ShotFinder, a text-driven three-stage retrieval and localization pipeline: (1) query expansion via video imagination, (2) candidate video retrieval with a search engine, and (3) description-guided temporal localization. Experiments on multiple closed-source and open-source models reveal a significant gap to human performance, with clear imbalance across constraints: temporal localization is relatively tractable, while color and visual style remain major challenges. These results reveal that open-domain video shot retrieval is still a critical capability that multimodal large models have yet to overcome.

Method: CMAFNet uses a cross-modal alignment and fusion network with a purify-then-fuse paradigm. It includes: 1) Semantic Recomposition Module for dictionary-based feature purification via learned codebook to suppress modality-specific noise, and 2) Contextual Semantic Integration Framework with partial-channel attention to capture global spatial dependencies. Position-wise normalization enforces reconstruction-driven cross-modal alignment.

Result: On the TLRGBD benchmark (94.5% small objects), CMAFNet achieves 32.2% mAP@50 and 12.5% APs, outperforming the strongest baseline by 9.8 and 4.0 percentage points. A lightweight variant reaches 24.8% mAP50 at 228 FPS with only 4.9M parameters, surpassing YOLO-based detectors while matching transformer-based methods at lower computational cost.

Conclusion: CMAFNet effectively addresses transmission line defect detection challenges by integrating RGB and depth modalities through principled cross-modal alignment and fusion, achieving superior performance for small defect detection in complex environments with efficient computational requirements.

Abstract: Transmission line defect detection remains challenging for automated UAV inspection due to the dominance of small-scale defects, complex backgrounds, and illumination variations. Existing RGB-based detectors, despite recent progress, struggle to distinguish geometrically subtle defects from visually similar background structures under limited chromatic contrast. This paper proposes CMAFNet, a Cross-Modal Alignment and Fusion Network that integrates RGB appearance and depth geometry through a principled purify-then-fuse paradigm. CMAFNet consists of a Semantic Recomposition Module that performs dictionary-based feature purification via a learned codebook to suppress modality-specific noise while preserving defect-discriminative information, and a Contextual Semantic Integration Framework that captures global spatial dependencies using partial-channel attention to enhance structural semantic reasoning. Position-wise normalization within the purification stage enforces explicit reconstruction-driven cross-modal alignment, ensuring statistical compatibility between heterogeneous features prior to fusion. Extensive experiments on the TLRGBD benchmark, where 94.5% of instances are small objects, demonstrate that CMAFNet achieves 32.2% mAP@50 and 12.5% APs, outperforming the strongest baseline by 9.8 and 4.0 percentage points, respectively. A lightweight variant reaches 24.8% mAP50 at 228 FPS with only 4.9M parameters, surpassing all YOLO-based detectors while matching transformer-based methods at substantially lower computational cost.

Method: Jointly trains a motion encoder with a pretrained video generator to distill driving frames into compact, view-agnostic motion tokens injected via cross-attention. Uses view-rich supervision (single-view, multi-view, moving-camera videos) for 3D awareness, and auxiliary geometric supervision with SMPL that’s annealed to zero over training.

Result: 3DiMo faithfully reproduces driving motions with flexible, text-driven camera control, significantly surpassing existing methods in both motion fidelity and visual quality.

Conclusion: Implicit, view-agnostic motion representations that align with video generators’ spatial priors are more effective than explicit 3D constraints for motion control, enabling both motion fidelity and novel-view synthesis.

Abstract: Existing methods for human motion control in video generation typically rely on either 2D poses or explicit 3D parametric models (e.g., SMPL) as control signals. However, 2D poses rigidly bind motion to the driving viewpoint, precluding novel-view synthesis. Explicit 3D models, though structurally informative, suffer from inherent inaccuracies (e.g., depth ambiguity and inaccurate dynamics) which, when used as a strong constraint, override the powerful intrinsic 3D awareness of large-scale video generators. In this work, we revisit motion control from a 3D-aware perspective, advocating for an implicit, view-agnostic motion representation that naturally aligns with the generator’s spatial priors rather than depending on externally reconstructed constraints. We introduce 3DiMo, which jointly trains a motion encoder with a pretrained video generator to distill driving frames into compact, view-agnostic motion tokens, injected semantically via cross-attention. To foster 3D awareness, we train with view-rich supervision (i.e., single-view, multi-view, and moving-camera videos), forcing motion consistency across diverse viewpoints. Additionally, we use auxiliary geometric supervision that leverages SMPL only for early initialization and is annealed to zero, enabling the model to transition from external 3D guidance to learning genuine 3D spatial motion understanding from the data and the generator’s priors. Experiments confirm that 3DiMo faithfully reproduces driving motions with flexible, text-driven camera control, significantly surpassing existing methods in both motion fidelity and visual quality.

Method: Uses modality independent neighborhood descriptors to extract features, maps images to feature space, then employs multilevel pyramidal fusion optimization with dense correlation analysis and weight-balanced coupled convex optimization at different scales.

Result: Achieved first place in both validation and test phases of ReMIND2Reg task in Learn2Reg 2025, and achieved average TRE of 1.798 mm on Resect dataset.

Conclusion: The MCPO method demonstrates strong performance in multimodal medical image registration with broad applicability for preoperative-to-intraoperative surgical guidance.

Abstract: Surgical navigation based on multimodal image registration has played a significant role in providing intraoperative guidance to surgeons by showing the relative position of the target area to critical anatomical structures during surgery. However, due to the differences between multimodal images and intraoperative image deformation caused by tissue displacement and removal during the surgery, effective registration of preoperative and intraoperative multimodal images faces significant challenges. To address the multimodal image registration challenges in Learn2Reg 2025, an unsupervised multimodal medical image registration method based on Multilevel Correlation Pyramidal Optimization (MCPO) is designed to solve these problems. First, the features of each modality are extracted based on the modality independent neighborhood descriptor, and the multimodal images is mapped to the feature space. Second, a multilevel pyramidal fusion optimization mechanism is designed to achieve global optimization and local detail complementation of the displacement field through dense correlation analysis and weight-balanced coupled convex optimization for input features at different scales. Our method focuses on the ReMIND2Reg task in Learn2Reg 2025. Based on the results, our method achieved the first place in the validation phase and test phase of ReMIND2Reg. The MCPO is also validated on the Resect dataset, achieving an average TRE of 1.798 mm. This demonstrates the broad applicability of our method in preoperative-to-intraoperative image registration. The code is available at https://github.com/wjiazheng/MCPO.

Method: ShapBPT assigns Shapley coefficients to a multiscale hierarchical structure called Binary Partition Tree (BPT), which is tailored for images. This data-aware hierarchical partitioning ensures feature attributions align with intrinsic image morphology.

Result: Experimental results show superior alignment with image structures, improved efficiency over existing XCV methods, and a 20-subject user study confirms human preference for ShapBPT explanations.

Conclusion: ShapBPT connects hierarchical Shapley methods with image data, providing more efficient and semantically meaningful visual interpretability for computer vision models.

Abstract: Pixel-level feature attributions are an important tool in eXplainable AI for Computer Vision (XCV), providing visual insights into how image features influence model predictions. The Owen formula for hierarchical Shapley values has been widely used to interpret machine learning (ML) models and their learned representations. However, existing hierarchical Shapley approaches do not exploit the multiscale structure of image data, leading to slow convergence and weak alignment with the actual morphological features. Moreover, no prior Shapley method has leveraged data-aware hierarchies for Computer Vision tasks, leaving a gap in model interpretability of structured visual data. To address this, this paper introduces ShapBPT, a novel data-aware XCV method based on the hierarchical Shapley formula. ShapBPT assigns Shapley coefficients to a multiscale hierarchical structure tailored for images, the Binary Partition Tree (BPT). By using this data-aware hierarchical partitioning, ShapBPT ensures that feature attributions align with intrinsic image morphology, effectively prioritizing relevant regions while reducing computational overhead. This advancement connects hierarchical Shapley methods with image data, providing a more efficient and semantically meaningful approach to visual interpretability. Experimental results confirm ShapBPT’s effectiveness, demonstrating superior alignment with image structures and improved efficiency over existing XCV methods, and a 20-subject user study confirming that ShapBPT explanations are preferred by humans.

Method: OneVision-Encoder uses Codec Patchification to focus computation on only 3.1%-25% of regions with high signal entropy. It employs shared 3D RoPE for unified spatial-temporal reasoning with irregular token layouts, trained with large-scale cluster discrimination over 1M+ semantic concepts to capture object permanence and motion dynamics.

Result: Outperforms strong vision backbones like Qwen3-ViT and SigLIP2 across 16 image, video, and document understanding benchmarks despite using fewer visual tokens and pretraining data. Achieves 4.1% average improvement over Qwen3-ViT on video tasks.

Conclusion: Codec-aligned patch-level sparsity enables efficient and accurate visual understanding, showing that efficiency and accuracy are positively correlated when architectures align with data’s fundamental structure.

Abstract: Hypothesis. Artificial general intelligence is, at its core, a compression problem. Effective compression demands resonance: deep learning scales best when its architecture aligns with the fundamental structure of the data. These are the fundamental principles. Yet, modern vision architectures have strayed from these truths: visual signals are highly redundant, while discriminative information, the surprise, is sparse. Current models process dense pixel grids uniformly, wasting vast compute on static background rather than focusing on the predictive residuals that define motion and meaning. We argue that to solve visual understanding, we must align our architectures with the information-theoretic principles of video, i.e., Codecs. Method. OneVision-Encoder encodes video by compressing predictive visual structure into semantic meaning. By adopting Codec Patchification, OV-Encoder abandons uniform computation to focus exclusively on the 3.1%-25% of regions rich in signal entropy. To unify spatial and temporal reasoning under irregular token layouts, OneVision-Encoder employs a shared 3D RoPE and is trained with a large-scale cluster discrimination objective over more than one million semantic concepts, jointly capturing object permanence and motion dynamics. Evidence. The results validate our core hypothesis: efficiency and accuracy are not a trade-off; they are positively correlated. When integrated into LLM, it consistently outperforms strong vision backbones such as Qwen3-ViT and SigLIP2 across 16 image, video, and document understanding benchmarks, despite using substantially fewer visual tokens and pretraining data. Notably, on video understanding tasks, OV-Encoder achieves an average improvement of 4.1% over Qwen3-ViT. Codec-aligned, patch-level sparsity is a foundational principle, enabling OV-Encoder as a scalable engine for next-generation visual generalists.

Method: Two-stage pipeline: 1) Controllable skeleton generation using Skeletal Graph VAE (Sk-VAE) to encode skeleton graph structure, with Skeletal Graph DiT (Sk-DiT) generating skeletal embeddings conditioned on text and 2D strokes; 2) Enhanced mesh synthesis using TextuRig dataset (curated from Objaverse-XL) and SKA-DPO preference optimization guided by skeleton-mesh alignment scores.

Result: Stroke3D produces plausible skeletons and high-quality meshes, enabling intuitive creation of ready-to-animate 3D content from 2D strokes and text prompts.

Conclusion: Stroke3D introduces the first framework for generating rigged 3D meshes conditioned on user-drawn 2D strokes, addressing limitations in existing 3D generation and rigging methods through a novel two-stage approach with structural control.

Abstract: Rigged 3D assets are fundamental to 3D deformation and animation. However, existing 3D generation methods face challenges in generating animatable geometry, while rigging techniques lack fine-grained structural control over skeleton creation. To address these limitations, we introduce Stroke3D, a novel framework that directly generates rigged meshes from user inputs: 2D drawn strokes and a descriptive text prompt. Our approach pioneers a two-stage pipeline that separates the generation into: 1) Controllable Skeleton Generation, we employ the Skeletal Graph VAE (Sk-VAE) to encode the skeleton’s graph structure into a latent space, where the Skeletal Graph DiT (Sk-DiT) generates a skeletal embedding. The generation process is conditioned on both the text for semantics and the 2D strokes for explicit structural control, with the VAE’s decoder reconstructing the final high-quality 3D skeleton; and 2) Enhanced Mesh Synthesis via TextuRig and SKA-DPO, where we then synthesize a textured mesh conditioned on the generated skeleton. For this stage, we first enhance an existing skeleton-to-mesh model by augmenting its training data with TextuRig: a dataset of textured and rigged meshes with captions, curated from Objaverse-XL. Additionally, we employ a preference optimization strategy, SKA-DPO, guided by a skeleton-mesh alignment score, to further improve geometric fidelity. Together, our framework enables a more intuitive workflow for creating ready to animate 3D content. To the best of our knowledge, our work is the first to generate rigged 3D meshes conditioned on user-drawn 2D strokes. Extensive experiments demonstrate that Stroke3D produces plausible skeletons and high-quality meshes.

Method: Proposes C²RoPE with spatio-temporal continuous positional embedding using triplet hybrid indices (temporal + spatial coordinates) and Chebyshev Causal Masking based on 2D spatial distances.

Result: Demonstrates effectiveness across 3D scene reasoning and 3D visual question answering benchmarks, showing improved visual processing capabilities.

Conclusion: C²RoPE successfully addresses RoPE limitations for multimodal processing by better modeling spatial continuity and causal relationships in visual tokens.

Abstract: Recent advances in 3D Large Multimodal Models (LMMs) built on Large Language Models (LLMs) have established the alignment of 3D visual features with LLM representations as the dominant paradigm. However, the inherited Rotary Position Embedding (RoPE) introduces limitations for multimodal processing. Specifically, applying 1D temporal positional indices disrupts the continuity of visual features along the column dimension, resulting in spatial locality loss. Moreover, RoPE follows the prior that temporally closer image tokens are more causally related, leading to long-term decay in attention allocation and causing the model to progressively neglect earlier visual tokens as the sequence length increases. To address these issues, we propose C^2RoPE, an improved RoPE that explicitly models local spatial Continuity and spatial Causal relationships for visual processing. C^2RoPE introduces a spatio-temporal continuous positional embedding mechanism for visual tokens. It first integrates 1D temporal positions with Cartesian-based spatial coordinates to construct a triplet hybrid positional index, and then employs a frequency allocation strategy to encode spatio-temporal positional information across the three index components. Additionally, we introduce Chebyshev Causal Masking, which determines causal dependencies by computing the Chebyshev distance of image tokens in 2D space. Evaluation results across various benchmarks, including 3D scene reasoning and 3D visual question answering, demonstrate C^2RoPE’s effectiveness. The code is be available at https://github.com/ErikZ719/C2RoPE.

Method: AMAP-APP optimizes efficiency by replacing intensive instance segmentation with classic image processing while retaining the original semantic segmentation model. It introduces a refined Region of Interest (ROI) algorithm to improve precision. Validation involved 365 mouse and human images (STED and confocal), benchmarking performance against the original AMAP via Pearson correlation and Two One-Sided T-tests (TOST).

Result: AMAP-APP achieved a 147-fold increase in processing speed on consumer hardware. Morphometric outputs (area, perimeter, circularity, and slit diaphragm density) showed high correlation (r>0.90) and statistical equivalence (TOST P<0.05) to the original method. The new ROI algorithm demonstrated superior accuracy with reduced deviation from manual delineations.

Conclusion: AMAP-APP democratizes deep learning-based podocyte morphometry by eliminating the need for high-performance computing clusters and providing a user-friendly interface for Windows, macOS, and Linux, enabling widespread adoption in nephrology research and potential clinical diagnostics.

Abstract: Background: Automated podocyte foot process quantification is vital for kidney research, but the established “Automatic Morphological Analysis of Podocytes” (AMAP) method is hindered by high computational demands, a lack of a user interface, and Linux dependency. We developed AMAP-APP, a cross-platform desktop application designed to overcome these barriers. Methods: AMAP-APP optimizes efficiency by replacing intensive instance segmentation with classic image processing while retaining the original semantic segmentation model. It introduces a refined Region of Interest (ROI) algorithm to improve precision. Validation involved 365 mouse and human images (STED and confocal), benchmarking performance against the original AMAP via Pearson correlation and Two One-Sided T-tests (TOST). Results: AMAP-APP achieved a 147-fold increase in processing speed on consumer hardware. Morphometric outputs (area, perimeter, circularity, and slit diaphragm density) showed high correlation (r>0.90) and statistical equivalence (TOST P<0.05) to the original method. Additionally, the new ROI algorithm demonstrated superior accuracy compared to the original, showing reduced deviation from manual delineations. Conclusion: AMAP-APP democratizes deep learning-based podocyte morphometry. By eliminating the need for high-performance computing clusters and providing a user-friendly interface for Windows, macOS, and Linux, it enables widespread adoption in nephrology research and potential clinical diagnostics.

Method: 1) Created StreamSR dataset from YouTube covering diverse video genres/resolutions; 2) Proposed EfRLFN model with Efficient Channel Attention and hyperbolic tangent activation; 3) Designed composite loss function; 4) Benchmarked 11 SOTA models; 5) Fine-tuned models on StreamSR dataset.

Result: EfRLFN achieves improved visual quality and runtime performance. Fine-tuning other models on StreamSR dataset yields significant performance gains that generalize well across various standard benchmarks.

Conclusion: StreamSR dataset addresses the gap in streaming-focused super-resolution evaluation, EfRLFN provides efficient real-time performance, and dataset-based fine-tuning improves model generalization across benchmarks.

Abstract: Recent advancements in real-time super-resolution have enabled higher-quality video streaming, yet existing methods struggle with the unique challenges of compressed video content. Commonly used datasets do not accurately reflect the characteristics of streaming media, limiting the relevance of current benchmarks. To address this gap, we introduce a comprehensive dataset - StreamSR - sourced from YouTube, covering a wide range of video genres and resolutions representative of real-world streaming scenarios. We benchmark 11 state-of-the-art real-time super-resolution models to evaluate their performance for the streaming use-case. Furthermore, we propose EfRLFN, an efficient real-time model that integrates Efficient Channel Attention and a hyperbolic tangent activation function - a novel design choice in the context of real-time super-resolution. We extensively optimized the architecture to maximize efficiency and designed a composite loss function that improves training convergence. EfRLFN combines the strengths of existing architectures while improving both visual quality and runtime performance. Finally, we show that fine-tuning other models on our dataset results in significant performance gains that generalize well across various standard benchmarks. We made the dataset, the code, and the benchmark available at https://github.com/EvgeneyBogatyrev/EfRLFN.

Method: Introduces Texlet representation where each latent vector encodes a local texture patch via 2D encoder, aggregated with 3D encoder for global shape context. Uses cascaded 3D-to-2D decoder for texture patch reconstruction, then trains diffusion transformer conditioned on Texlets to refine textures from multi-view diffusion methods.

Result: Extensive experiments show TexSpot significantly improves visual fidelity, geometric consistency, and robustness over state-of-the-art 3D texture generation and enhancement approaches.

Conclusion: TexSpot addresses fundamental limitations in 3D texture generation by introducing a novel representation that enables high-quality texture enhancement through diffusion-based methods.

Abstract: High-quality 3D texture generation remains a fundamental challenge due to the view-inconsistency inherent in current mainstream multi-view diffusion pipelines. Existing representations either rely on UV maps, which suffer from distortion during unwrapping, or point-based methods, which tightly couple texture fidelity to geometric density that limits high-resolution texture generation. To address these limitations, we introduce TexSpot, a diffusion-based texture enhancement framework. At its core is Texlet, a novel 3D texture representation that merges the geometric expressiveness of point-based 3D textures with the compactness of UV-based representation. Each Texlet latent vector encodes a local texture patch via a 2D encoder and is further aggregated using a 3D encoder to incorporate global shape context. A cascaded 3D-to-2D decoder reconstructs high-quality texture patches, enabling the Texlet space learning. Leveraging this representation, we train a diffusion transformer conditioned on Texlets to refine and enhance textures produced by multi-view diffusion methods. Extensive experiments demonstrate that TexSpot significantly improves visual fidelity, geometric consistency, and robustness over existing state-of-the-art 3D texture generation and enhancement approaches. Project page: https://texlet-arch.github.io/TexSpot-page.

Method: Proposes Reliable Thinking with Images (RTWI) that estimates reliability of both visual cues and textual CoT in a unified text-centric manner, then employs robust filtering and voting modules to prevent noisy thinking from contaminating final answers.

Result: Extensive experiments on seven benchmarks verify RTWI’s effectiveness against noisy thinking, showing improved performance over existing TWI methods.

Conclusion: Addressing noisy thinking is crucial for robust multimodal reasoning, and RTWI provides an effective solution by reliability estimation and error prevention mechanisms.

Abstract: As a multimodal extension of Chain-of-Thought (CoT), Thinking with Images (TWI) has recently emerged as a promising avenue to enhance the reasoning capability of Multi-modal Large Language Models (MLLMs), which generates interleaved CoT by incorporating visual cues into the textual reasoning process. However, the success of existing TWI methods heavily relies on the assumption that interleaved image-text CoTs are faultless, which is easily violated in real-world scenarios due to the complexity of multimodal understanding. In this paper, we reveal and study a highly-practical yet under-explored problem in TWI, termed Noisy Thinking (NT). Specifically, NT refers to the imperfect visual cues mining and answer reasoning process. As the saying goes, ``One mistake leads to another’’, erroneous interleaved CoT would cause error accumulation, thus significantly degrading the performance of MLLMs. To solve the NT problem, we propose a novel method dubbed Reliable Thinking with Images (RTWI). In brief, RTWI estimates the reliability of visual cues and textual CoT in a unified text-centric manner and accordingly employs robust filtering and voting modules to prevent NT from contaminating the final answer. Extensive experiments on seven benchmarks verify the effectiveness of RTWI against NT.

Method: Decision-negative, human-in-the-loop agentic system with adversarial self-critique mechanism where a critic agent challenges primary agent’s conclusions before human review, plus formal taxonomy of failure modes.

Result: Adversarial critique reduces AI hallucination rates from 11.3% to 3.8% and increases decision accuracy from 92% to 96% on 500 expert-validated underwriting cases.

Conclusion: Adversarial self-critique supports safer AI deployment in regulated domains and offers a model for responsible integration where human oversight is indispensable.

Abstract: Commercial insurance underwriting is a labor-intensive process that requires manual review of extensive documentation to assess risk and determine policy pricing. While AI offers substantial efficiency improvements, existing solutions lack comprehensive reasoning capabilities and internal mechanisms to ensure reliability within regulated, high-stakes environments. Full automation remains impractical and inadvisable in scenarios where human judgment and accountability are critical. This study presents a decision-negative, human-in-the-loop agentic system that incorporates an adversarial self-critique mechanism as a bounded safety architecture for regulated underwriting workflows. Within this system, a critic agent challenges the primary agent’s conclusions prior to submitting recommendations to human reviewers. This internal system of checks and balances addresses a critical gap in AI safety for regulated workflows. Additionally, the research develops a formal taxonomy of failure modes to characterize potential errors by decision-negative agents. This taxonomy provides a structured framework for risk identification and risk management in high-stakes applications. Experimental evaluation using 500 expert-validated underwriting cases demonstrates that the adversarial critique mechanism reduces AI hallucination rates from 11.3% to 3.8% and increases decision accuracy from 92% to 96%. At the same time, the framework enforces strict human authority over all binding decisions by design. These findings indicate that adversarial self-critique supports safer AI deployment in regulated domains and offers a model for responsible integration where human oversight is indispensable.

Method: Anchors LLM evaluation to fixed hierarchies of skill-calibrated game AI on the Botzone platform, evaluating LLMs across eight diverse games (deterministic perfect-information board games to stochastic imperfect-information card games) with systematic assessment of 177,047 state-action pairs from five flagship models.

Result: Reveals significant performance disparities and identifies distinct strategic behaviors, with top-performing models achieving proficiency comparable to mid-to-high-tier specialized game AI in multiple domains.

Conclusion: The anchored evaluation paradigm generalizes beyond games to any domain with well-defined skill hierarchies, establishing a scalable and reusable framework for assessing interactive AI capabilities.

Abstract: Large Language Models (LLMs) are increasingly deployed in interactive environments requiring strategic decision-making, yet systematic evaluation of these capabilities remains challenging. Existing benchmarks for LLMs primarily assess static reasoning through isolated tasks and fail to capture dynamic strategic abilities. Recent game-based evaluations employ LLM-vs-LLM tournaments that produce relative rankings dependent on transient model pools, incurring quadratic computational costs and lacking stable performance anchors for longitudinal tracking. The central challenge is establishing a scalable evaluation framework that measures LLM strategic reasoning against consistent, interpretable standards rather than volatile peer models. Here we show that anchoring LLM evaluation to fixed hierarchies of skill-calibrated game Artificial Intelligence (AI) enables linear-time absolute skill measurement with stable cross-temporal interpretability. Built on the Botzone platform’s established competitive infrastructure, our BotzoneBench evaluates LLMs across eight diverse games spanning deterministic perfect-information board games to stochastic imperfect-information card games. Through systematic assessment of 177,047 state-action pairs from five flagship models, we reveal significant performance disparities and identify distinct strategic behaviors, with top-performing models achieving proficiency comparable to mid-to-high-tier specialized game AI in multiple domains. This anchored evaluation paradigm generalizes beyond games to any domain with well-defined skill hierarchies, establishing a scalable and reusable framework for assessing interactive AI capabilities.

Method: AMOR combines SSM backbone with sparse attention triggered only when SSM prediction entropy indicates uncertainty. Uses Ghost KV projection from SSM hidden states to reuse O(n) computation rather than O(n²) attention at every layer.

Result: On synthetic retrieval tasks, AMOR outperforms SSM-only and transformer-only baselines, achieving perfect retrieval accuracy while engaging attention on only 22% of positions. Prediction entropy shows 1.09 nats gap between retrieval and local positions.

Conclusion: AMOR provides efficient, interpretable adaptive computation for long-context processing, with routing decisions understandable in information-theoretic terms, bridging SSM efficiency with transformer precision.

Abstract: Transformers allocate uniform computation to every position, regardless of difficulty. State Space Models (SSMs) offer efficient alternatives but struggle with precise information retrieval over a long horizon. Inspired by dual-process theories of cognition (Kahneman, 2011), we propose AMOR (Adaptive Metacognitive Output Router), a hybrid architecture that dynamically engages sparse attention only when an SSM backbone is “uncertain”–as measured by prediction entropy. Compared to standard transformers, AMOR gains efficiency by projecting keys and values from SSM hidden states (Ghost KV), reusing the SSM’s O(n) computation rather than requiring O(n^2) attention at every layer. On small-scale synthetic retrieval tasks, AMOR outperforms both SSM-only and transformer-only baselines, achieving perfect retrieval accuracy while engaging attention on only 22% of positions. We validate that prediction entropy reliably signals retrieval need, with a gap of 1.09 nats (nearly half the entropy range) between retrieval and local positions. Additionally, our approach provides interpretable adaptive computation, where routing decisions can be understood in information-theoretic terms.

Method: VeRA converts benchmark problems into executable specifications with three components: (1) natural language template with placeholders, (2) coherent generator that samples valid configurations, and (3) deterministic verifier that validates parameters and calculates correct answers. It operates in two modes: VeRA-E (equivalent variants) and VeRA-H (hardened variants with increased complexity).

Result: Evaluation of 16 frontier models shows: VeRA-E improves evaluation quality and reveals contamination patterns; VeRA-H enables human-free generation of hard tasks with reliable labels; VeRA establishes verified benchmarks as a general paradigm for generating fresh instances on demand.

Conclusion: VeRA reconceptualizes benchmarks from static objects to executable specifications that can generate unlimited verified instances, enhancing robustness and cost-effectiveness for AI evaluation across any verifiable domain.

Abstract: The main issue with most evaluation schemes today is their “static” nature: the same problems are reused repeatedly, allowing for memorization, format exploitation, and eventual saturation. To measure genuine AI progress, we need evaluation that is robust by construction, not by post-hoc detection. In response, we propose VeRA (Verified Reasoning Data Augmentation), a framework that converts benchmark problems into executable specifications, comprising (i) a natural language template with placeholder slots, (ii) a coherent generator that samples valid configurations, and (iii) a deterministic verifier that validates parameters and calculates the corresponding correct answers for each configuration. From a single seed problem, VeRA automatically creates unlimited verified variants with reliable labels at near-zero marginal cost without human involvement. VeRA operates in two complementary modes. VeRA-E (equivalent) rewrites problems while keeping the underlying logic intact, useful for detecting memorization versus genuine reasoning. VeRA-H (hardened) systematically increases complexity while remaining verifiable, enabling reliable creation and labelling of fresh difficult tasks at the boundary of intelligence. Evaluating 16 frontier models with VeRA, we find: (i) VeRA-E improves evaluation quality and reveals contamination patterns. (ii) VeRA-H enables human-free generation of hard tasks with reliable labels. (iii) VeRA establishes verified benchmarks as a general paradigm. VeRA reconceptualizes benchmarks from static objects used until exhausted, to executable specifications generating fresh, verified instances on demand, enhancing robustness and cost-effectiveness for evaluation. With VeRA, we envision that evaluation in any verifiable domain can scale indefinitely without sacrificing label integrity. To stimulate future research, we have open-sourced all code and datasets.

Method: SSLogic uses an agentic meta-synthesis framework with a closed Generate-Validate-Repair loop. It iteratively synthesizes and repairs executable Generator-Validator program pairs. A Multi-Gate Validation Protocol combines multi-strategy consistency checks with Adversarial Blind Review, where independent agents must solve instances by writing and executing code to filter ambiguous tasks.

Result: Starting from 400 seed families, two evolution rounds expanded to 953 families and 21,389 verifiable instances (from 5,718). Training on SSLogic-evolved data yields consistent gains: SynLogic +5.2, BBEH +1.4, AIME25 +3.0, and Brumo25 +3.7 over seed baseline at matched training steps.

Conclusion: SSLogic successfully scales verifiable training signals at the task-family level, enabling continuous evolution with controllable difficulty and producing substantial performance improvements across multiple benchmarks.

Abstract: Scaling verifiable training signals remains a key bottleneck for Reinforcement Learning from Verifiable Rewards (RLVR). Logical reasoning is a natural substrate: constraints are formal and answers are programmatically checkable. However, prior synthesis pipelines either depend on expert-written code or operate within fixed templates/skeletons, which limits growth largely to instance-level perturbations. We propose SSLogic, an agentic meta-synthesis framework that scales at the task-family level by iteratively synthesizing and repairing executable Generator–Validator program pairs in a closed Generate–Validate–Repair loop, enabling continuous family evolution with controllable difficulty. To ensure reliability, we introduce a Multi-Gate Validation Protocol that combines multi-strategy consistency checks with Adversarial Blind Review, where independent agents must solve instances by writing and executing code to filter ambiguous or ill-posed tasks. Starting from 400 seed families, two evolution rounds expand to 953 families and 21,389 verifiable instances (from 5,718). Training on SSLogic-evolved data yields consistent gains over the seed baseline at matched training steps, improving SynLogic by +5.2, BBEH by +1.4, AIME25 by +3.0, and Brumo25 by +3.7.

Method: The authors propose a taxonomy identifying three hallucination types, then analyze their geometric signatures in embedding space using benchmarks and human-crafted confabulations. They measure detection performance via AUROC and examine domain transferability and geometric relationships between discriminative directions.

Result: Type I (unfaithfulness) and Type II (confabulation) hallucinations show strong detection performance within domains (AUROC 0.76-0.99) but poor cross-domain transfer (AUROC ~0.50). Human-crafted confabulations show a single global detection direction with 0.96 AUROC and minimal cross-domain degradation. Type III (factual errors) are indistinguishable from chance (AUROC 0.478).

Conclusion: Embedding-based detection works for Types I and II hallucinations but fails for Type III factual errors, which require external verification mechanisms since embeddings encode distributional patterns rather than correspondence to external reality.

Abstract: The term “hallucination” in large language models conflates distinct phenomena with different geometric signatures in embedding space. We propose a taxonomy identifying three types: unfaithfulness (failure to engage with provided context), confabulation (invention of semantically foreign content), and factual error (incorrect claims within correct conceptual frames). We observe a striking asymmetry. On standard benchmarks where hallucinations are LLM-generated, detection is domain-local: AUROC 0.76-0.99 within domains, but 0.50 (chance level) across domains. Discriminative directions are approximately orthogonal between domains (mean cosine similarity -0.07). On human-crafted confabulations - invented institutions, redefined terminology, fabricated mechanisms - a single global direction achieves 0.96 AUROC with 3.8% cross-domain degradation. We interpret this divergence as follows: benchmarks capture generation artifacts (stylistic signatures of prompted fabrication), while human-crafted confabulations capture genuine topical drift. The geometric structure differs because the underlying phenomena differ. Type III errors show 0.478 AUROC - indistinguishable from chance. This reflects a theoretical constraint: embeddings encode distributional co-occurrence, not correspondence to external reality. Statements with identical contextual patterns occupy similar embedding regions regardless of truth value. The contribution is a geometric taxonomy clarifying the scope of embedding-based detection: Types I and II are detectable; Type III requires external verification mechanisms.

Method: VaryBalance detects LLM-generated text by observing that human texts have greater difference from their LLM-rewritten versions compared to LLM-generated texts. It quantifies this difference using mean standard deviation to distinguish between human and LLM-generated texts.

Result: VaryBalance outperforms state-of-the-art detectors (Binoculars) by up to 34.3% in AUROC, maintains robustness across multiple generating models and languages, and demonstrates practical effectiveness.

Conclusion: VaryBalance provides a simple, effective, and practical solution for LLM-generated text detection that addresses limitations of existing methods and shows superior performance across various conditions.

Abstract: Detecting text generated by large language models (LLMs) is crucial but challenging. Existing detectors depend on impractical assumptions, such as white-box settings, or solely rely on text-level features, leading to imprecise detection ability. In this paper, we propose a simple but effective and practical LLM-generated text detection method, VaryBalance. The core of VaryBalance is that, compared to LLM-generated texts, there is a greater difference between human texts and their rewritten version via LLMs. Leveraging this observation, VaryBalance quantifies this through mean standard deviation and distinguishes human texts and LLM-generated texts. Comprehensive experiments demonstrated that VaryBalance outperforms the state-of-the-art detectors, i.e., Binoculars, by up to 34.3% in terms of AUROC, and maintains robustness against multiple generating models and languages.

Method: Introduces Trajectory-Dominant Pareto Optimization (path-wise generalization of Pareto optimality), defines Trap Escape Difficulty Index (TEDI) to measure constraint rigidity, develops taxonomy of Pareto traps, and illustrates with minimal agent-environment model.

Result: Shows dynamic intelligence ceilings arise as geometric consequences of trajectory-level dominance, independent of learning progress or architectural scale, and provides framework for diagnosing developmental constraints.

Conclusion: Shifts focus from terminal performance to optimization geometry, offering principled framework to overcome long-horizon developmental constraints in adaptive systems.

Abstract: Despite recent advances in artificial intelligence, many systems exhibit stagnation in long-horizon adaptability despite continued performance optimization. This work argues that such limitations do not primarily arise from insufficient learning, data, or model capacity, but from a deeper structural property of how intelligence is optimized over time. We formulate intelligence as a trajectory-level phenomenon governed by multi-objective trade-offs, and introduce Trajectory-Dominant Pareto Optimization, a path-wise generalization of classical Pareto optimality in which dominance is defined over full trajectories. Within this framework, Pareto traps emerge as locally non-dominated regions of trajectory space that nevertheless restrict access to globally superior developmental paths under conservative local optimization. To characterize the rigidity of such constraints, we define the Trap Escape Difficulty Index (TEDI), a composite geometric measure capturing escape distance, structural constraints, and behavioral inertia. We show that dynamic intelligence ceilings arise as inevitable geometric consequences of trajectory-level dominance, independent of learning progress or architectural scale. We further introduce a formal taxonomy of Pareto traps and illustrate the resulting trajectory-level divergence using a minimal agent-environment model. Together, these results shift the locus of intelligence from terminal performance to optimization geometry, providing a principled framework for diagnosing and overcoming long-horizon developmental constraints in adaptive systems.

Method: Created a generator-based benchmark with 15 plot families and 450 rendered plots, each produced from known parameters with exact ground truth. Introduced checkpoint-based diagnostic evaluation with intermediate ‘cp_’ fields to isolate sub-skills. Evaluated four SOTA MLLMs using standardized protocol (temperature=0, JSON-only numeric output) with per-field tolerance scoring.

Result: Top models achieved 80.42% (Gemini 2.5 Pro), 79.84% (GPT-4.1), and 78.21% (Claude Sonnet 4.5) overall field-level pass rates. GPT-4o trailed at 61.59%. Frequency-domain tasks remained challenging with bandpass response ≤23% and FFT spectrum difficulties.

Conclusion: PlotChain provides a reproducible benchmark for quantitative plot reading in MLLMs, revealing strengths in many plot families but persistent weaknesses in frequency-domain analysis. The diagnostic framework enables detailed failure analysis.

Abstract: We present PlotChain, a deterministic, generator-based benchmark for evaluating multimodal large language models (MLLMs) on engineering plot reading-recovering quantitative values from classic plots (e.g., Bode/FFT, step response, stress-strain, pump curves) rather than OCR-only extraction or free-form captioning. PlotChain contains 15 plot families with 450 rendered plots (30 per family), where every item is produced from known parameters and paired with exact ground truth computed directly from the generating process. A central contribution is checkpoint-based diagnostic evaluation: in addition to final targets, each item includes intermediate ‘cp_’ fields that isolate sub-skills (e.g., reading cutoff frequency or peak magnitude) and enable failure localization within a plot family. We evaluate four state-of-the-art MLLMs under a standardized, deterministic protocol (temperature = 0 and a strict JSON-only numeric output schema) and score predictions using per-field tolerances designed to reflect human plot-reading precision. Under the ‘plotread’ tolerance policy, the top models achieve 80.42% (Gemini 2.5 Pro), 79.84% (GPT-4.1), and 78.21% (Claude Sonnet 4.5) overall field-level pass rates, while GPT-4o trails at 61.59%. Despite strong performance on many families, frequency-domain tasks remain brittle: bandpass response stays low (<= 23%), and FFT spectrum remains challenging. We release the generator, dataset, raw model outputs, scoring code, and manifests with checksums to support fully reproducible runs and retrospective rescoring under alternative tolerance policies.

Method: Proposes a dual-cycle framework: (1) Persona-Targeted Attacker Cycle synthesizes progressively stronger jailbreak prompts, (2) Role-Playing Defender Cycle distills failures into hierarchical knowledge base (global safety rules, persona-grounded constraints, safe exemplars). At inference, retrieves and composes structured knowledge to guide generation.

Result: Extensive experiments across multiple proprietary LLMs show consistent gains over baselines on both role fidelity and jailbreak resistance, with robust generalization to unseen personas and attack prompts.

Conclusion: The training-free framework effectively balances safety and persona fidelity in role-playing LLMs without requiring model retraining, offering a practical solution for evolving personas and attack strategies.

Abstract: LLM-based role-playing has rapidly improved in fidelity, yet stronger adherence to persona constraints commonly increases vulnerability to jailbreak attacks, especially for risky or negative personas. Most prior work mitigates this issue with training-time solutions (e.g., data curation or alignment-oriented regularization). However, these approaches are costly to maintain as personas and attack strategies evolve, can degrade in-character behavior, and are typically infeasible for frontier closed-weight LLMs. We propose a training-free Dual-Cycle Adversarial Self-Evolution framework with two coupled cycles. A Persona-Targeted Attacker Cycle synthesizes progressively stronger jailbreak prompts, while a Role-Playing Defender Cycle distills observed failures into a hierarchical knowledge base of (i) global safety rules, (ii) persona-grounded constraints, and (iii) safe in-character exemplars. At inference time, the Defender retrieves and composes structured knowledge from this hierarchy to guide generation, producing responses that remain faithful to the target persona while satisfying safety constraints. Extensive experiments across multiple proprietary LLMs show consistent gains over strong baselines on both role fidelity and jailbreak resistance, and robust generalization to unseen personas and attack prompts.

Method: DPBench uses the Dining Philosophers problem to evaluate LLM coordination across eight conditions varying decision timing (sequential vs. simultaneous), group size, and communication capabilities, testing models like GPT-5.2, Claude Opus 4.5, and Grok 4.1.

Result: LLMs show striking asymmetry: they coordinate effectively in sequential settings but fail catastrophically in simultaneous decision-making, with deadlock rates exceeding 95% in some conditions. Communication doesn’t help and can even increase deadlock rates due to convergent reasoning where agents independently arrive at identical strategies.

Conclusion: Multi-agent LLM systems requiring concurrent resource access may need external coordination mechanisms rather than relying on emergent coordination, as LLMs fundamentally fail at simultaneous coordination despite working well in sequential scenarios.

Abstract: Large language models are increasingly deployed in multi-agent systems, yet we lack benchmarks that test whether they can coordinate under resource contention. We introduce DPBench, a benchmark based on the Dining Philosophers problem that evaluates LLM coordination across eight conditions that vary decision timing, group size, and communication. Our experiments with GPT-5.2, Claude Opus 4.5, and Grok 4.1 reveal a striking asymmetry: LLMs coordinate effectively in sequential settings but fail when decisions must be made simultaneously, with deadlock rates exceeding 95% under some conditions. We trace this failure to convergent reasoning, where agents independently arrive at identical strategies that, when executed simultaneously, guarantee deadlock. Contrary to expectations, enabling communication does not resolve this problem and can even increase deadlock rates. Our findings suggest that multi-agent LLM systems requiring concurrent resource access may need external coordination mechanisms rather than relying on emergent coordination. DPBench is released as an open-source benchmark. Code and benchmark are available at https://github.com/najmulhasan-code/dpbench.

Method: Two-stage RL-based training: first stage optimizes VLMs to self-explore high-quality actions for building a reusable linguistic toolbox; second stage optimizes VLMs to effectively exploit these linguistic tools for downstream reasoning.

Result: Lang2Act substantially enhances visual perception capabilities of VLMs, achieving performance improvements of over 4% compared to existing methods.

Conclusion: The proposed self-emergent linguistic toolchain approach effectively bridges visual perception and reasoning in VLMs, overcoming limitations of decoupled designs in existing VRAG frameworks.

Abstract: Visual Retrieval-Augmented Generation (VRAG) enhances Vision-Language Models (VLMs) by incorporating external visual documents to address a given query. Existing VRAG frameworks usually depend on rigid, pre-defined external tools to extend the perceptual capabilities of VLMs, typically by explicitly separating visual perception from subsequent reasoning processes. However, this decoupled design can lead to unnecessary loss of visual information, particularly when image-based operations such as cropping are applied. In this paper, we propose Lang2Act, which enables fine-grained visual perception and reasoning through self-emergent linguistic toolchains. Rather than invoking fixed external engines, Lang2Act collects self-emergent actions as linguistic tools and leverages them to enhance the visual perception capabilities of VLMs. To support this mechanism, we design a two-stage Reinforcement Learning (RL)-based training framework. Specifically, the first stage optimizes VLMs to self-explore high-quality actions for constructing a reusable linguistic toolbox, and the second stage further optimizes VLMs to exploit these linguistic tools for downstream reasoning effectively. Experimental results demonstrate the effectiveness of Lang2Act in substantially enhancing the visual perception capabilities of VLMs, achieving performance improvements of over 4%. All code and data are available at https://github.com/NEUIR/Lang2Act.

Method: Proposes MAPLE (Memory-Adaptive Personalized LEarning) - a principled decomposition where each component operates as a dedicated sub-agent: Memory handles storage/retrieval infrastructure, Learning extracts intelligence from accumulated interactions asynchronously, and Personalization applies learned knowledge in real-time within finite context budgets.

Result: Experimental evaluation on MAPLE-Personas benchmark shows 14.6% improvement in personalization score compared to stateless baseline (p < 0.01, Cohen’s d = 0.95) and increases trait incorporation rate from 45% to 75%.

Conclusion: Decomposing personalization into distinct memory, learning, and personalization components enables LLM agents to genuinely learn and adapt to individual users, addressing a fundamental limitation in current agent architectures.

Abstract: Large language model (LLM) agents have emerged as powerful tools for complex tasks, yet their ability to adapt to individual users remains fundamentally limited. We argue this limitation stems from a critical architectural conflation: current systems treat memory, learning, and personalization as a unified capability rather than three distinct mechanisms requiring different infrastructure, operating on different timescales, and benefiting from independent optimization. We propose MAPLE (Memory-Adaptive Personalized LEarning), a principled decomposition where Memory handles storage and retrieval infrastructure; Learning extracts intelligence from accumulated interactions asynchronously; and Personalization applies learned knowledge in real-time within finite context budgets. Each component operates as a dedicated sub-agent with specialized tooling and well-defined interfaces. Experimental evaluation on the MAPLE-Personas benchmark demonstrates that our decomposition achieves a 14.6% improvement in personalization score compared to a stateless baseline (p < 0.01, Cohen’s d = 0.95) and increases trait incorporation rate from 45% to 75% – enabling agents that genuinely learn and adapt.

Method: Introduces abstract syntax tree as intermediate representation, combining recursive LLM-based semantic parser with AST-guided generator that deterministically produces solver-ready logic code.

Result: Achieves 99% syntactic accuracy, improves semantic correctness by up to 30% over SOTA baselines, and when integrated into Logic-LM yields near-perfect executability and 31% improvement in downstream reasoning accuracy.

Conclusion: NL2LOGIC framework significantly improves both syntactic and semantic accuracy of natural language to logic translation, enhancing automated reasoning systems.

Abstract: Automated reasoning is critical in domains such as law and governance, where verifying claims against facts in documents requires both accuracy and interpretability. Recent work adopts structured reasoning pipelines that translate natural language into first-order logic and delegate inference to automated solvers. With the rise of large language models, approaches such as GCD and CODE4LOGIC leverage their reasoning and code generation capabilities to improve logic parsing. However, these methods suffer from fragile syntax control due to weak enforcement of global grammar constraints and low semantic faithfulness caused by insufficient clause-level semantic understanding. We propose NL2LOGIC, a first-order logic translation framework that introduces an abstract syntax tree as an intermediate representation. NL2LOGIC combines a recursive large language model based semantic parser with an abstract syntax tree guided generator that deterministically produces solver-ready logic code. Experiments on the FOLIO, LogicNLI, and ProofWriter benchmarks show that NL2LOGIC achieves 99 percent syntactic accuracy and improves semantic correctness by up to 30 percent over state-of-the-art baselines. Furthermore, integrating NL2LOGIC into Logic-LM yields near-perfect executability and improves downstream reasoning accuracy by 31 percent compared to Logic-LM’s original few-shot unconstrained translation module.

Method: Introduces AST-PAC, a domain-specific adaptation of Polarized Augment Calibration (PAC) that uses Abstract Syntax Tree (AST) based perturbations to generate syntactically valid calibration samples for membership inference attacks on code models.

Result: AST-PAC improves performance as syntactic size grows where PAC degrades, but under-mutates small files and underperforms on alphanumeric-rich code. PAC generally outperforms Loss baseline but suffers on larger, complex files due to syntax-blind augmentations.

Conclusion: Future work needed on syntax-aware and size-adaptive calibration for reliable provenance auditing of code language models. AST-PAC shows promise but has limitations with small files and certain code types.

Abstract: Code Large Language Models are frequently trained on massive datasets containing restrictively licensed source code. This creates urgent data governance and copyright challenges. Membership Inference Attacks (MIAs) can serve as an auditing mechanism to detect unauthorized data usage in models. While attacks like the Loss Attack provide a baseline, more involved methods like Polarized Augment Calibration (PAC) remain underexplored in the code domain. This paper presents an exploratory study evaluating these methods on 3B–7B parameter code models. We find that while PAC generally outperforms the Loss baseline, its effectiveness relies on augmentation strategies that disregard the rigid syntax of code, leading to performance degradation on larger, complex files. To address this, we introduce AST-PAC, a domain-specific adaptation that utilizes Abstract Syntax Tree (AST) based perturbations to generate syntactically valid calibration samples. Preliminary results indicate that AST-PAC improves as syntactic size grows, where PAC degrades, but under-mutates small files and underperforms on alphanumeric-rich code. Overall, the findings motivate future work on syntax-aware and size-adaptive calibration as a prerequisite for reliable provenance auditing of code language models.

Method: Hierarchical X-Blocks framework with RACE (multi-LLM ensemble) for context classification, log-odds analysis for lexical patterns, and dependency parsing for syntactic analysis

Result: RACE achieves 91.45% accuracy for context classification, identifies context-specific vocabulary patterns, and reveals limited repertoire of reusable grammar families

Conclusion: X-Blocks provides evidence-based linguistic design principles for generating scenario-aware explanations to support transparency and user trust in automated driving systems

Abstract: Natural language explanations play a critical role in establishing trust and acceptance of automated vehicles (AVs), yet existing approaches lack systematic frameworks for analysing how humans linguistically construct driving rationales across diverse scenarios. This paper introduces X-Blocks (eXplanation Blocks), a hierarchical analytical framework that identifies the linguistic building blocks of natural language explanations for AVs at three levels: context, syntax, and lexicon. At the context level, we propose RACE (Reasoning-Aligned Classification of Explanations), a multi-LLM ensemble framework that combines Chain-of-Thought reasoning with self-consistency mechanisms to robustly classify explanations into 32 scenario-aware categories. Applied to human-authored explanations from the Berkeley DeepDrive-X dataset, RACE achieves 91.45 percent accuracy and a Cohens kappa of 0.91 against cases with human annotator agreement, indicating near-human reliability for context classification. At the lexical level, log-odds analysis with informative Dirichlet priors reveals context-specific vocabulary patterns that distinguish driving scenarios. At the syntactic level, dependency parsing and template extraction show that explanations draw from a limited repertoire of reusable grammar families, with systematic variation in predicate types and causal constructions across contexts. The X-Blocks framework is dataset-agnostic and task-independent, offering broad applicability to other automated driving datasets and safety-critical domains. Overall, our findings provide evidence-based linguistic design principles for generating scenario-aware explanations that support transparency, user trust, and cognitive accessibility in automated driving systems.

Method: Uses wiring-diagram decompositions of claim dependencies to localize gaps, combines rapid draft generation with adversarial verification, targeted repair, and explicit provenance tracking.

Result: Heterogeneous outcomes across ten problems: Problem 3 has validation-complete existence path, Problem 5 solved in scope-limited form, Problem 10 conditional, Problems 4 and 6 partial, Problem 7 provisionally closed. QC layer shows validation artifacts with unresolved gaps.

Conclusion: Structure-aware verification and layer-switching strategies improve reliability and calibration in compressed proof sprints, demonstrating the effectiveness of the multi-agent approach for mathematical problem-solving.

Abstract: This monograph reports a multi-agent proof sprint on ten research-level problems, combining rapid draft generation with adversarial verification, targeted repair, and explicit provenance. The workflow uses wiring-diagram decompositions of claim dependencies to localize gaps and coordinate reviewer-driven revisions. Final outcomes are heterogeneous but explicit: the manuscript distinguishes mathematical status from QC-validation status. Mathematically, Problem3 has a validation-complete existence path under the scoped criterion used here (uniqueness/irreducibility treated as optional), Problem 5 is solved in a scope-limited form for $F_O$-local connective spectra, Problem 10 is conditional under clearly stated assumptions (with explicit necessity counterexamples when assumptions are dropped), and Problems 4 and 6 are partial with named remaining obligations in the general case (including an unconditional $K_n$ result for Problem 6 with $c_0 = 1/3$). Problem 7 is treated as provisionally closed via the rotation-route theorem chain, pending independent ledger re-check. At the QC layer, Problems7 and~9 have node-level validation artifacts but still contain unresolved verifier gaps. The main methodological result is that structure-aware verification and layer-switching strategies improve reliability and calibration in compressed proof sprints.

Method: Equips a base model with ability to spawn same-weight clones in separate parallel contexts using agentic reinforcement learning, trained end-to-end under global task reward with shared-parameter rollouts.

Result: Improves accuracy-cost Pareto frontier on challenging math reasoning benchmarks and long-context multi-hop QA relative to monolithic baselines at matched inference budget, with out-of-distribution generalization.

Conclusion: SELFCEST provides an effective approach for improving language model performance through coordinated parallel computation under fixed inference budgets.

Abstract: Frontier language models improve with additional test-time computation, but serial reasoning or uncoordinated parallel sampling can be compute-inefficient under fixed inference budgets. We propose SELFCEST, which equips a base model with the ability to spawn same-weight clones in separate parallel contexts by agentic reinforcement learning. Training is end-to-end under a global task reward with shared-parameter rollouts, yielding a learned controller that allocates both generation and context budget across branches. Across challenging math reasoning benchmarks and long-context multi-hop QA, SELFCEST improves the accuracy-cost Pareto frontier relative to monolithic baselines at matched inference budget, and exhibits out-of-distribution generalization in both domains.

Method: Entropy-Based Adaptive Guidance Framework that quantifies weak agents’ understanding through multi-dimensional entropy metrics (expression, uncertainty, structure, coherence, relevance) and adaptively adjusts guidance intensity at three levels (light, moderate, intensive), plus RAG mechanism for retaining successful collaboration experiences.

Result: Extensive experiments on GSM8K, MBPP, and CVRP benchmarks demonstrate consistent enhancement of effectiveness and stability in heterogeneous collaboration, showing adaptive guidance mitigates cognitive imbalance.

Conclusion: Adaptive guidance framework addresses cognitive mismatching in heterogeneous multi-agent systems, establishing a scalable pathway toward more robust cooperative multi-agent intelligence.

Abstract: With recent breakthroughs in large language models (LLMs) for reasoning, planning, and complex task generation, artificial intelligence systems are transitioning from isolated single-agent architectures to multi-agent systems with collaborative intelligence. However, in heterogeneous multi-agent systems (HMAS), capability differences among agents give rise to consistent cognitive problems, where strong and weak models fail to contribute effectively. We define the collaboration as a strong-weak system. Through comprehensive experiments, we disclose a counterintuitive phenomenon in the strong-weak system: a strong-weak collaboration may under-perform weak-weak combinations, revealing that cognitive mismatching are key bottlenecks limiting heterogeneous cooperation. To overcome these challenges, we propose an Entropy-Based Adaptive Guidance Framework that dynamically aligns the guidance with the cognitive state of each agent. The framework quantifies the understanding of weak agents through multi-dimensional entropy metrics - covering expression, uncertainty, structure, coherence, and relevance - and adaptively adjusts the intensity of the guidance at light, moderate and intensive levels. Furthermore, a Retrieval-Augmented Generation (RAG) mechanism is incorporated to retain successful collaboration experiences, enabling both immediate adaptation and long-term learning. Extensive experiments on three benchmark datasets, GSM8K, MBPP, and CVRP demonstrate that our approach consistently enhances the effectiveness and stability of heterogeneous collaboration. The results highlight that adaptive guidance not only mitigates cognitive imbalance but also establishes a scalable pathway toward more robust, cooperative multi-agent intelligence.

Method: Combined CNN and LSTM networks to capture temporal dependencies in traffic sequences, integrated SHAP (SHapley Additive exPlanations) for model interpretability, and conducted trust-focused expert surveys using IPIP6 and Big Five personality traits via interactive UI.

Result: Both CNN and LSTM achieved 0.99 accuracy on NSL-KDD dataset, with LSTM outperforming CNN on macro average precision, recall, and F-1 score. SHAP identified key influential features like srv_serror_rate, dst_host_srv_serror_rate, and serror_rate. Expert surveys evaluated system reliability and usability.

Conclusion: The framework demonstrates the potential of combining performance and transparency in cybersecurity solutions, with recommendations for future enhancements through adaptive learning for real-time threat detection.

Abstract: The increasing complexity and frequency of cyber-threats demand intrusion detection systems (IDS) that are not only accurate but also interpretable. This paper presented a novel IDS framework that integrated Explainable Artificial Intelligence (XAI) to enhance transparency in deep learning models. The framework was evaluated experimentally using the benchmark dataset NSL-KDD, demonstrating superior performance compared to traditional IDS and black-box deep learning models. The proposed approach combined Convolutional Neural Network (CNN) and Long Short-Term Memory (LSTM) networks for capturing temporal dependencies in traffic sequences. Our deep learning results showed that both CNN and LSTM reached 0.99 for accuracy, whereas LSTM outperformed CNN at macro average precision, recall, and F-1 score. For weighted average precision, recall, and F-1 score, both models scored almost similarly. To ensure interpretability, the XAI model SHapley Additive exPlanations (SHAP) was incorporated, enabling security analysts to understand and validate model decisions. Some notable influential features were srv_serror_rate, dst_host_srv_serror_rate, and serror_rate for both models, as pointed out by SHAP. We also conducted a trust-focused expert survey based on IPIP6 and Big Five personality traits via an interactive UI to evaluate the system’s reliability and usability. This work highlighted the potential of combining performance and transparency in cybersecurity solutions and recommends future enhancements through adaptive learning for real-time threat detection.

Method: Developed TemporalBench with a four-tier task taxonomy: 1) historical structure interpretation, 2) context-free forecasting, 3) contextual temporal reasoning, and 4) event-conditioned prediction. The benchmark spans four real-world domains and controls access to future targets and contextual information to enable diagnostic analysis.

Result: Extensive baseline experiments show that strong numerical forecasting accuracy doesn’t reliably translate to robust contextual or event-aware temporal reasoning. Existing agent frameworks exhibit fragmented strengths and systematic failure modes that remain hidden under forecasting-only benchmarks.

Conclusion: TemporalBench reveals limitations in current temporal reasoning models and provides a diagnostic tool for evaluating genuine temporal understanding versus contextual reasoning capabilities across multiple domains.

Abstract: It is unclear whether strong forecasting performance reflects genuine temporal understanding or the ability to reason under contextual and event-driven conditions. We introduce TemporalBench, a multi-domain benchmark designed to evaluate temporal reasoning behavior under progressively richer informational settings. TemporalBench adopts a four-tier task taxonomy that examines historical structure interpretation, context-free forecasting, contextual temporal reasoning, and event-conditioned prediction across four real-world domains: retail, healthcare, energy, and physical systems. By controlling access to future targets and contextual information, the benchmark enables a diagnostic analysis of whether models can correctly interpret temporal patterns, align them with external context, and adapt predictions when conditions change. Extensive baseline experiments show that strong numerical forecasting accuracy does not reliably translate into robust contextual or event-aware temporal reasoning; instead, existing agent frameworks exhibit fragmented strengths and systematic failure modes that remain largely hidden under forecasting-only benchmarks. The TemporalBench dataset is publicly available at https://huggingface.co/datasets/Melady/TemporalBench, and we additionally provide a public leaderboard at https://huggingface.co/spaces/Melady/TemporalBench_Leaderboard.

Method: Introduces ProMoral-Bench with 11 prompting paradigms evaluated across four LLM families using ETHICS, Scruples, WildJailbreak datasets and new ETHICS-Contrast robustness test, measured via Unified Moral Safety Score (UMSS) balancing accuracy and safety.

Result: Compact, exemplar-guided scaffolds outperform complex multi-stage reasoning, achieving higher UMSS scores with greater robustness at lower token cost. Few-shot exemplars consistently enhance moral stability and jailbreak resistance while multi-turn reasoning proves fragile under perturbations.

Conclusion: ProMoral-Bench establishes a standardized framework for principled, cost-effective prompt engineering, showing that simpler prompting approaches can be more effective for moral safety alignment than complex reasoning methods.

Abstract: Prompt design significantly impacts the moral competence and safety alignment of large language models (LLMs), yet empirical comparisons remain fragmented across datasets and models.We introduce ProMoral-Bench, a unified benchmark evaluating 11 prompting paradigms across four LLM families. Using ETHICS, Scruples, WildJailbreak, and our new robustness test, ETHICS-Contrast, we measure performance via our proposed Unified Moral Safety Score (UMSS), a metric balancing accuracy and safety. Our results show that compact, exemplar-guided scaffolds outperform complex multi-stage reasoning, providing higher UMSS scores and greater robustness at a lower token cost. While multi-turn reasoning proves fragile under perturbations, few-shot exemplars consistently enhance moral stability and jailbreak resistance. ProMoral-Bench establishes a standardized framework for principled, cost-effective prompt engineering.

Method: Proposes claim-level auditability as design target, identifies long-horizon failure modes (objective drift, transient constraints, unverifiable inference), introduces AAR standard with four metrics (provenance coverage, soundness, contradiction transparency, audit effort), and advocates for semantic provenance graphs with protocolized validation.

Result: Develops framework for measuring auditability in AI research agents, identifies key failure patterns, and proposes practical instrumentation patterns for scalable deployment of semantic provenance systems.

Conclusion: Auditability should be a first-class design consideration for deep research agents, requiring semantic provenance systems with continuous validation to address the shift from factual errors to claim-evidence linkage problems.

Abstract: A deep research agent produces a fluent scientific report in minutes; a careful reader then tries to verify the main claims and discovers the real cost is not reading, but tracing: which sentence is supported by which passage, what was ignored, and where evidence conflicts. We argue that as research generation becomes cheap, auditability becomes the bottleneck, and the dominant risk shifts from isolated factual errors to scientifically styled outputs whose claim-evidence links are weak, missing, or misleading. This perspective proposes claim-level auditability as a first-class design and evaluation target for deep research agents, distills recurring long-horizon failure modes (objective drift, transient constraints, and unverifiable inference), and introduces the Auditable Autonomous Research (AAR) standard, a compact measurement framework that makes auditability testable via provenance coverage, provenance soundness, contradiction transparency, and audit effort. We then argue for semantic provenance with protocolized validation: persistent, queryable provenance graphs that encode claim–evidence relations (including conflicts) and integrate continuous validation during synthesis rather than after publication, with practical instrumentation patterns to support deployment at scale.

Method: Developed Perseverance Composition Engine, a multi-agent system for document composition with three specialized agents: Composer (drafts text), Corroborator (verifies factual substantiation with full source access), and Critic (evaluates argumentative quality without source access). Information asymmetry is enforced by system architecture to create layered verification.

Result: Tested on 474 composition tasks, the system exhibited patterns consistent with institutional hypothesis. When assigned impossible tasks requiring fabricated content, the system progressed from attempted fabrication toward honest refusal with alternative proposals—behavior not instructed or individually incentivized.

Conclusion: Organizational theory provides a productive framework for multi-agent AI safety. Architectural enforcement through information compartmentalization offers a route to reliable collective behavior from unreliable individual components, positioning institutional design as an alternative to individual alignment approaches.

Abstract: Alignment research focuses on making individual AI systems reliable. Human institutions achieve reliable collective behaviour differently: they mitigate the risk posed by misaligned individuals through organisational structure. Multi-agent AI systems should follow this institutional model using compartmentalisation and adversarial review to achieve reliable outcomes through architectural design rather than assuming individual alignment. We demonstrate this approach through the Perseverance Composition Engine, a multi-agent system for document composition. The Composer drafts text, the Corroborator verifies factual substantiation with full source access, and the Critic evaluates argumentative quality without access to sources: information asymmetry enforced by system architecture. This creates layered verification: the Corroborator detects unsupported claims, whilst the Critic independently assesses coherence and completeness. Observations from 474 composition tasks (discrete cycles of drafting, verification, and evaluation) exhibit patterns consistent with the institutional hypothesis. When assigned impossible tasks requiring fabricated content, this iteration enabled progression from attempted fabrication toward honest refusal with alternative proposals–behaviour neither instructed nor individually incentivised. These findings motivate controlled investigation of whether architectural enforcement produces reliable outcomes from unreliable components. This positions organisational theory as a productive framework for multi-agent AI safety. By implementing verification and evaluation as structural properties enforced through information compartmentalisation, institutional design offers a route to reliable collective behaviour from unreliable individual components.

Method: BEAGLE integrates three innovations: (1) semi-Markov model for cognitive/metacognitive behavior transitions, (2) Bayesian Knowledge Tracing with explicit flaw injection for realistic knowledge gaps, and (3) decoupled agent design separating strategy use from code generation to prevent silent error correction.

Result: BEAGLE significantly outperforms state-of-the-art baselines in reproducing authentic trajectories on Python programming tasks. In a human Turing test, users achieved only 52.8% accuracy distinguishing synthetic traces from real student data (indistinguishable from random guessing).

Conclusion: BEAGLE successfully addresses LLM competency bias by incorporating Self-Regulated Learning theory, enabling realistic simulation of novice learning behaviors with applications for training adaptive tutoring systems and testing pedagogical interventions.

Abstract: Simulating student learning behaviors in open-ended problem-solving environments holds potential for education research, from training adaptive tutoring systems to stress-testing pedagogical interventions. However, collecting authentic data is challenging due to privacy concerns and the high cost of longitudinal studies. While Large Language Models (LLMs) offer a promising path to student simulation, they suffer from competency bias, optimizing for efficient correctness rather than the erratic, iterative struggle characteristic of novice learners. We present BEAGLE, a neuro-symbolic framework that addresses this bias by incorporating Self-Regulated Learning (SRL) theory into a novel architecture. BEAGLE integrates three key technical innovations: (1) a semi-Markov model that governs the timing and transitions of cognitive behaviors and metacognitive behaviors; (2) Bayesian Knowledge Tracing with explicit flaw injection to enforce realistic knowledge gaps and “unknown unknowns”; and (3) a decoupled agent design that separates high-level strategy use from code generation actions to prevent the model from silently correcting its own intentional errors. In evaluations on Python programming tasks, BEAGLE significantly outperforms state-of-the-art baselines in reproducing authentic trajectories. In a human Turing test, users were unable to distinguish synthetic traces from real student data, achieving an accuracy indistinguishable from random guessing (52.8%).

Method: Online survey with N=300 participants, focusing on accuracy requirements and disruption patterns. Defined accuracy as context-specific reliability based on user intent alignment within tolerance thresholds.

Result: Significant accuracy gap: 24.1% require high accuracy at work vs 8.8% in personal life (+15.3pp). Gap remains large under broader definitions. Heavy app users have stricter work standards. More disruption in personal routines (34.1%) than at work (15.3%) when tools are unavailable.

Conclusion: People have substantially higher accuracy requirements for AI tools in professional contexts compared to personal use, and experience different disruption patterns when these tools become unavailable, highlighting context-dependent AI adoption and usage patterns.

Abstract: We study how people trade off accuracy when using AI-powered tools in professional versus personal contexts for adoption purposes, the determinants of those trade-offs, and how users cope when AI/apps are unavailable. Because modern AI systems (especially generative models) can produce acceptable but non-identical outputs, we define “accuracy” as context-specific reliability: the degree to which an output aligns with the user’s intent within a tolerance threshold that depends on stakes and the cost of correction. In an online survey (N=300), among respondents with both accuracy items (N=170), the share requiring high accuracy (top-box) is 24.1% at work vs. 8.8% in personal life (+15.3 pp; z=6.29, p<0.001). The gap remains large under a broader top-two-box definition (67.0% vs. 32.9%) and on the full 1-5 ordinal scale (mean 3.86 vs. 3.08). Heavy app use and experience patterns correlate with stricter work standards (H2). When tools are unavailable (H3), respondents report more disruption in personal routines than at work (34.1% vs. 15.3%, p<0.01). We keep the main text focused on these substantive results and place test taxonomy and power derivations in a technical appendix.

Method: Developed Mirror framework with EthicsLLM (fine-tuned on EthicsQA dataset of 41K question-chain-of-thought-answer triples from ethics/regulatory sources). Two operational modes: Mirror-ER for automated expedited review using executable rule base, and Mirror-CR for full-board deliberation simulation with coordinated expert agents, ethics secretary, and PI agent across ten ethical dimensions.

Result: Empirical evaluations show Mirror significantly improves quality, consistency, and professionalism of ethics assessments compared to generalist LLMs.

Conclusion: Mirror demonstrates that AI-assisted ethical review frameworks can enhance research governance by integrating ethical reasoning with regulatory structures, addressing current limitations in institutional review capacity.

Abstract: Ethics review is a foundational mechanism of modern research governance, yet contemporary systems face increasing strain as ethical risks arise as structural consequences of large-scale, interdisciplinary scientific practice. The demand for consistent and defensible decisions under heterogeneous risk profiles exposes limitations in institutional review capacity rather than in the legitimacy of ethics oversight. Recent advances in large language models (LLMs) offer new opportunities to support ethics review, but their direct application remains limited by insufficient ethical reasoning capability, weak integration with regulatory structures, and strict privacy constraints on authentic review materials. In this work, we introduce Mirror, an agentic framework for AI-assisted ethical review that integrates ethical reasoning, structured rule interpretation, and multi-agent deliberation within a unified architecture. At its core is EthicsLLM, a foundational model fine-tuned on EthicsQA, a specialized dataset of 41K question-chain-of-thought-answer triples distilled from authoritative ethics and regulatory corpora. EthicsLLM provides detailed normative and regulatory understanding, enabling Mirror to operate in two complementary modes. Mirror-ER (expedited Review) automates expedited review through an executable rule base that supports efficient and transparent compliance checks for minimal-risk studies. Mirror-CR (Committee Review) simulates full-board deliberation through coordinated interactions among expert agents, an ethics secretary agent, and a principal investigator agent, producing structured, committee-level assessments across ten ethical dimensions. Empirical evaluations demonstrate that Mirror significantly improves the quality, consistency, and professionalism of ethics assessments compared with strong generalist LLMs.

Method: Created DECKBench with curated paper-to-slide pairs augmented with simulated editing instructions, plus a modular multi-agent baseline system that decomposes tasks into paper parsing/summarization, slide planning, HTML creation, and iterative editing.

Result: The benchmark effectively highlights system strengths, exposes failure modes, and provides actionable insights for improving multi-agent slide generation and editing systems.

Conclusion: DECKBench establishes a standardized foundation for reproducible and comparable evaluation of academic presentation generation and editing systems.

Abstract: Automatically generating and iteratively editing academic slide decks requires more than document summarization. It demands faithful content selection, coherent slide organization, layout-aware rendering, and robust multi-turn instruction following. However, existing benchmarks and evaluation protocols do not adequately measure these challenges. To address this gap, we introduce the Deck Edits and Compliance Kit Benchmark (DECKBench), an evaluation framework for multi-agent slide generation and editing. DECKBench is built on a curated dataset of paper to slide pairs augmented with realistic, simulated editing instructions. Our evaluation protocol systematically assesses slide-level and deck-level fidelity, coherence, layout quality, and multi-turn instruction following. We further implement a modular multi-agent baseline system that decomposes the slide generation and editing task into paper parsing and summarization, slide planning, HTML creation, and iterative editing. Experimental results demonstrate that the proposed benchmark highlights strengths, exposes failure modes, and provides actionable insights for improving multi-agent slide generation and editing systems. Overall, this work establishes a standardized foundation for reproducible and comparable evaluation of academic presentation generation and editing. Code and data are publicly available at https://github.com/morgan-heisler/DeckBench .

Method: Proposes SGP task that treats perspective modeling as inverse inference problem. Uses structure-first synthetic generation strategy that aligns latent labels and observable traces by design. Constructs dataset and runs diagnostic study using retrieval-augmented in-context learning as proxy for supervision with GPT-4o.

Result: Observes gap between surface-level extraction and latent perspective inference, indicating latent-state inference is harder than surface extraction under controlled setting. Results suggest SGP is non-trivial and provide evidence for structure-first data synthesis strategy.

Conclusion: SGP offers promising approach to perspective-aware AI by framing perspective modeling as inverse inference problem, with synthetic data generation strategy showing potential to overcome data bottlenecks in modeling internal states from multimodal artifacts.

Abstract: Perspective-Aware AI requires modeling evolving internal states–goals, emotions, contexts–not merely preferences. Progress is limited by a data bottleneck: digital footprints are privacy-sensitive and perspective states are rarely labeled. We propose Situation Graph Prediction (SGP), a task that frames perspective modeling as an inverse inference problem: reconstructing structured, ontology-aligned representations of perspective from observable multimodal artifacts. To enable grounding without real labels, we use a structure-first synthetic generation strategy that aligns latent labels and observable traces by design. As a pilot, we construct a dataset and run a diagnostic study using retrieval-augmented in-context learning as a proxy for supervision. In our study with GPT-4o, we observe a gap between surface-level extraction and latent perspective inference–indicating latent-state inference is harder than surface extraction under our controlled setting. Results suggest SGP is non-trivial and provide evidence for the structure-first data synthesis strategy.

Method: Developed theoretical framework for error accumulation in MCP agents, proving linear growth and O(√T) concentration bounds. Created hybrid distortion metric combining discrete fact matching with continuous semantic similarity, and established martingale concentration bounds for sequential tool interactions.

Result: Experiments across Qwen2-7B, Llama-3-8B, and Mistral-7B validate theoretical predictions: empirical distortion tracks linear trend with deviations within O(√T) bounds. Semantic weighting reduces distortion by 80%, and periodic re-grounding every ~9 steps suffices for error control.

Conclusion: The concentration guarantees enable predictable system behavior and rule out exponential failure modes, providing actionable deployment principles for trustworthy agent systems using external tools.

Abstract: As AI agents powered by large language models (LLMs) increasingly use external tools for high-stakes decisions, a critical reliability question arises: how do errors propagate across sequential tool calls? We introduce the first theoretical framework for analyzing error accumulation in Model Context Protocol (MCP) agents, proving that cumulative distortion exhibits linear growth and high-probability deviations bounded by $O(\sqrt{T})$. This concentration property ensures predictable system behavior and rules out exponential failure modes. We develop a hybrid distortion metric combining discrete fact matching with continuous semantic similarity, then establish martingale concentration bounds on error propagation through sequential tool interactions. Experiments across Qwen2-7B, Llama-3-8B, and Mistral-7B validate our theoretical predictions, showing empirical distortion tracks the linear trend with deviations consistently within $O(\sqrt{T})$ envelopes. Key findings include: semantic weighting reduces distortion by 80%, and periodic re-grounding approximately every 9 steps suffices for error control. We translate these concentration guarantees into actionable deployment principles for trustworthy agent systems.

Method: Train BERT-based models (general and medical domain) to predict four expert-annotated linguistic features (Professionalism, Medical Relevance, Ethical Behavior, Contextual Distraction), then use these as features in second-layer classifiers (tree-based, linear, probabilistic, ensemble) for jailbreak detection.

Result: System achieves strong overall performance in cross-validation and held-out evaluations. LLM-derived linguistic features provide effective basis for automated jailbreak detection.

Conclusion: Demonstrates scalable, interpretable approach for jailbreak detection in clinical dialogue systems. Identifies limitations in current annotations and feature representations, pointing to future improvements.

Abstract: Detecting jailbreak attempts in clinical training large language models (LLMs) requires accurate modeling of linguistic deviations that signal unsafe or off-task user behavior. Prior work on the 2-Sigma clinical simulation platform showed that manually annotated linguistic features could support jailbreak detection. However, reliance on manual annotation limited both scalability and expressiveness. In this study, we extend this framework by using experts’ annotations of four core linguistic features (Professionalism, Medical Relevance, Ethical Behavior, and Contextual Distraction) and training multiple general-domain and medical-domain BERT-based LLM models to predict these features directly from text. The most reliable feature regressor for each dimension was selected and used as the feature extractor in a second layer of classifiers. We evaluate a suite of predictive models, including tree-based, linear, probabilistic, and ensemble methods, to determine jailbreak likelihood from the extracted features. Across cross-validation and held-out evaluations, the system achieves strong overall performance, indicating that LLM-derived linguistic features provide an effective basis for automated jailbreak detection. Error analysis further highlights key limitations in current annotations and feature representations, pointing toward future improvements such as richer annotation schemes, finer-grained feature extraction, and methods that capture the evolving risk of jailbreak behavior over the course of a dialogue. This work demonstrates a scalable and interpretable approach for detecting jailbreak behavior in safety-critical clinical dialogue systems.

Method: Extended previous BDI agent explanation framework to handle contrastive questions. Conducted computational evaluation measuring explanation length, and human subject evaluation assessing preference, trust development, transparency, and understanding.

Result: Contrastive questions yield significantly shorter explanations. Some evidence that contrastive answers are preferred and lead to higher trust, perceived understanding, and confidence in system correctness. Surprisingly, providing full explanations sometimes performed worse than no explanation.

Conclusion: Contrastive explanations are more efficient and potentially more effective for trust development, but the value of providing explanations at all is context-dependent and may not always be beneficial.

Abstract: The ability of autonomous systems to provide explanations is important for supporting transparency and aiding the development of (appropriate) trust. Prior work has defined a mechanism for Belief-Desire-Intention (BDI) agents to be able to answer questions of the form why did you do action $X$?''. However, we know that we ask \emph{contrastive} questions (why did you do $X$ \emph{instead of} $F$?’’). We therefore extend previous work to be able to answer such questions. A computational evaluation shows that using contrastive questions yields a significant reduction in explanation length. A human subject evaluation was conducted to assess whether such contrastive answers are preferred, and how well they support trust development and transparency. We found some evidence for contrastive answers being preferred, and some evidence that they led to higher trust, perceived understanding, and confidence in the system’s correctness. We also evaluated the benefit of providing explanations at all. Surprisingly, there was not a clear benefit, and in some situations we found evidence that providing a (full) explanation was worse than not providing any explanation.

Method: Combines point-wise and pair-wise reward modeling for reasoning and preference alignment; uses complexity-aware rewards in reinforcement learning for code generation; performs complex data synthesis with turn-level supervision for deep search capabilities; enables stable long-horizon tool interactions (up to 600 tool-call turns).

Result: Significantly outperforms prior models of similar scale (Nanbeige4-3B-2511, Qwen3-4B) and even achieves superior performance compared to much larger models like Qwen3-30B-A3B; demonstrates reliable execution of up to 600 tool-call turns for complex problem-solving.

Conclusion: Small models (3B parameters) can achieve both broad competence and strong specialization simultaneously, redefining the potential of small language models through innovative training techniques.

Abstract: We present Nanbeige4.1-3B, a unified generalist language model that simultaneously achieves strong agentic behavior, code generation, and general reasoning with only 3B parameters. To the best of our knowledge, it is the first open-source small language model (SLM) to achieve such versatility in a single model. To improve reasoning and preference alignment, we combine point-wise and pair-wise reward modeling, ensuring high-quality, human-aligned responses. For code generation, we design complexity-aware rewards in Reinforcement Learning, optimizing both correctness and efficiency. In deep search, we perform complex data synthesis and incorporate turn-level supervision during training. This enables stable long-horizon tool interactions, allowing Nanbeige4.1-3B to reliably execute up to 600 tool-call turns for complex problem-solving. Extensive experimental results show that Nanbeige4.1-3B significantly outperforms prior models of similar scale, such as Nanbeige4-3B-2511 and Qwen3-4B, even achieving superior performance compared to much larger models, such as Qwen3-30B-A3B. Our results demonstrate that small models can achieve both broad competence and strong specialization simultaneously, redefining the potential of 3B parameter models.

Method: Introduces Morality Chains formalism for representing moral norms as ordered deontic constraints, and MoralityGym benchmark with 98 ethical-dilemma problems presented as trolley-dilemma-style Gymnasium environments. Decouples task-solving from moral evaluation and introduces a novel Morality Metric

Result: Baseline results with Safe RL methods reveal key limitations, highlighting the need for more principled approaches to ethical decision-making in AI systems

Conclusion: Provides a foundation for developing AI systems that behave more reliably, transparently, and ethically in complex real-world contexts by integrating insights from psychology and philosophy into norm-sensitive reasoning evaluation

Abstract: Evaluating moral alignment in agents navigating conflicting, hierarchically structured human norms is a critical challenge at the intersection of AI safety, moral philosophy, and cognitive science. We introduce Morality Chains, a novel formalism for representing moral norms as ordered deontic constraints, and MoralityGym, a benchmark of 98 ethical-dilemma problems presented as trolley-dilemma-style Gymnasium environments. By decoupling task-solving from moral evaluation and introducing a novel Morality Metric, MoralityGym allows the integration of insights from psychology and philosophy into the evaluation of norm-sensitive reasoning. Baseline results with Safe RL methods reveal key limitations, underscoring the need for more principled approaches to ethical decision-making. This work provides a foundation for developing AI systems that behave more reliably, transparently, and ethically in complex real-world contexts.

Method: Remove KL regularization and group-wise normalization from RL framework, simplify reward to truncation-based length penalty, and reduce optimization to supervised fine-tuning on self-generated data filtered for both correctness and conciseness (termed “on-policy SFT”).

Result: Achieves up to 80% reduction in chain-of-thought length while maintaining original accuracy, outperforms complex RL-based methods across five benchmarks, reduces GPU memory usage by 50%, and accelerates convergence by 70%.

Conclusion: Simplified on-policy SFT approach consistently defines the accuracy-efficiency Pareto frontier, demonstrating that complex RL extensions are unnecessary for optimizing reasoning models for both correctness and conciseness.

Abstract: Large reasoning models (LRMs) are commonly trained with reinforcement learning (RL) to explore long chain-of-thought reasoning, achieving strong performance at high computational cost. Recent methods add multi-reward objectives to jointly optimize correctness and brevity, but these complex extensions often destabilize training and yield suboptimal trade-offs. We revisit this objective and challenge the necessity of such complexity. Through principled analysis, we identify fundamental misalignments in this paradigm: KL regularization loses its intended role when correctness and length are directly verifiable, and group-wise normalization becomes ambiguous under multiple reward signals. By removing these two items and simplifying the reward to a truncation-based length penalty, we show that the optimization problem reduces to supervised fine-tuning on self-generated data filtered for both correctness and conciseness. We term this simplified training strategy on-policy SFT. Despite its simplicity, on-policy SFT consistently defines the accuracy-efficiency Pareto frontier. It reduces CoT length by up to 80 while maintaining original accuracy, surpassing more complex RL-based methods across five benchmarks. Furthermore, it significantly enhances training efficiency, reducing GPU memory usage by 50% and accelerating convergence by 70%. Our code is available at https://github.com/EIT-NLP/On-Policy-SFT.

Method: Reformulates EEG pipeline engineering as discrete constrained optimization. Uses Domain-Informed Subspace Initialization to restrict search to neuroscientifically plausible manifolds. Employs Multi-Objective Evolutionary Optimization with self-reflective refinement to balance performance, novelty, and efficiency.

Result: Outperforms state-of-the-art task-specific methods across five heterogeneous benchmarks. Achieves performance comparable to large-scale foundation models while using significantly fewer parameters. Synthesizes lightweight solutions suitable for resource-constrained environments.

Conclusion: NeuroWeaver bridges the gap between computationally expensive foundation models and scientifically naive AutoML by incorporating domain knowledge into automated pipeline design, enabling efficient EEG analysis in clinical settings.

Abstract: Although foundation models have demonstrated remarkable success in general domains, the application of these models to electroencephalography (EEG) analysis is constrained by substantial data requirements and high parameterization. These factors incur prohibitive computational costs, thereby impeding deployment in resource-constrained clinical environments. Conversely, general-purpose automated machine learning frameworks are often ill-suited for this domain, as exploration within an unbounded programmatic space fails to incorporate essential neurophysiological priors and frequently yields solutions that lack scientific plausibility. To address these limitations, we propose NeuroWeaver, a unified autonomous evolutionary agent designed to generalize across diverse EEG datasets and tasks by reformulating pipeline engineering as a discrete constrained optimization problem. Specifically, we employ a Domain-Informed Subspace Initialization to confine the search to neuroscientifically plausible manifolds, coupled with a Multi-Objective Evolutionary Optimization that dynamically balances performance, novelty, and efficiency via self-reflective refinement. Empirical evaluations across five heterogeneous benchmarks demonstrate that NeuroWeaver synthesizes lightweight solutions that consistently outperform state-of-the-art task-specific methods and achieve performance comparable to large-scale foundation models, despite utilizing significantly fewer parameters.

Method: Red-teaming a concrete orchestrator setup (central agent delegating to specialized agents) to investigate security vulnerabilities. Testing frontier models with different attack categories to demonstrate OMNI-LEAK attack vector that compromises multiple agents through single indirect prompt injection.

Result: Both reasoning and non-reasoning frontier models are vulnerable to OMNI-LEAK attacks, which can leak sensitive data even with data access controls in place. Attacks succeed without attacker having insider knowledge of implementation details.

Conclusion: Safety research must generalize from single-agent to multi-agent settings to address serious risks of real-world privacy breaches, financial losses, and maintain public trust in AI agents. Multi-agent systems introduce novel attack vectors that bypass traditional safeguards.

Abstract: As Large Language Model (LLM) agents become more capable, their coordinated use in the form of multi-agent systems is anticipated to emerge as a practical paradigm. Prior work has examined the safety and misuse risks associated with agents. However, much of this has focused on the single-agent case and/or setups missing basic engineering safeguards such as access control, revealing a scarcity of threat modeling in multi-agent systems. We investigate the security vulnerabilities of a popular multi-agent pattern known as the orchestrator setup, in which a central agent decomposes and delegates tasks to specialized agents. Through red-teaming a concrete setup representative of a likely future use case, we demonstrate a novel attack vector, OMNI-LEAK, that compromises several agents to leak sensitive data through a single indirect prompt injection, even in the \textit{presence of data access control}. We report the susceptibility of frontier models to different categories of attacks, finding that both reasoning and non-reasoning models are vulnerable, even when the attacker lacks insider knowledge of the implementation details. Our work highlights the importance of safety research to generalize from single-agent to multi-agent settings, in order to reduce the serious risks of real-world privacy breaches and financial losses and overall public trust in AI agents.

Method: Uses WWEIA intake data for 135,491 meals to identify 34 interpretable meal archetypes, then conditions a generative model and portion predictor on these archetypes to meet USDA nutritional targets with minimal food substitutions.

Result: Generated meals follow RDI targets 47.0% better while remaining compositionally close to real meals; with 1-3 food substitutions, meals become 10% more nutritious while reducing costs 19-32% on average.

Conclusion: The framework can underpin clinical decision support, public-health programs, and consumer apps for scalable, equitable improvements in everyday nutrition by turning guidelines into realistic, budget-aware meals.

Abstract: An important goal for personalized diet systems is to improve nutritional quality without compromising convenience or affordability. We present an end-to-end framework that converts dietary standards into complete meals with minimal change. Using the What We Eat in America (WWEIA) intake data for 135,491 meals, we identify 34 interpretable meal archetypes that we then use to condition a generative model and a portion predictor to meet USDA nutritional targets. In comparisons within archetypes, generated meals are better at following recommended daily intake (RDI) targets by 47.0%, while remaining compositionally close to real meals. Our results show that by allowing one to three food substitutions, we were able to create meals that were 10% more nutritious, while reducing costs 19-32%, on average. By turning dietary guidelines into realistic, budget-aware meals and simple swaps, this framework can underpin clinical decision support, public-health programs, and consumer apps that deliver scalable, equitable improvements in everyday nutrition.

Method: Introduced SPILLage framework characterizing oversharing along channel (content vs. behavior) and directness (explicit vs. implicit) dimensions. Benchmarked 180 tasks on live e-commerce sites with ground-truth annotations, conducted 1,080 runs across two agentic frameworks and three backbone LLMs, and tested prompt-level mitigation strategies.

Result: Oversharing is pervasive with behavioral oversharing dominating content oversharing by 5x. The effect persists or worsens under prompt-level mitigation, but removing task-irrelevant information before execution improves task success by up to 17.9%.

Conclusion: Protecting privacy in web agents requires a broader view of “output” that accounts for what agents do on the web, not just what they type, as behavioral oversharing through actions represents a critical blind spot in current privacy considerations.

Abstract: LLM-powered agents are beginning to automate user’s tasks across the open web, often with access to user resources such as emails and calendars. Unlike standard LLMs answering questions in a controlled ChatBot setting, web agents act “in the wild”, interacting with third parties and leaving behind an action trace. Therefore, we ask the question: how do web agents handle user resources when accomplishing tasks on their behalf across live websites? In this paper, we formalize Natural Agentic Oversharing – the unintentional disclosure of task-irrelevant user information through an agent trace of actions on the web. We introduce SPILLage, a framework that characterizes oversharing along two dimensions: channel (content vs. behavior) and directness (explicit vs. implicit). This taxonomy reveals a critical blind spot: while prior work focuses on text leakage, web agents also overshare behaviorally through clicks, scrolls, and navigation patterns that can be monitored. We benchmark 180 tasks on live e-commerce sites with ground-truth annotations separating task-relevant from task-irrelevant attributes. Across 1,080 runs spanning two agentic frameworks and three backbone LLMs, we demonstrate that oversharing is pervasive with behavioral oversharing dominates content oversharing by 5x. This effect persists – and can even worsen – under prompt-level mitigation. However, removing task-irrelevant information before execution improves task success by up to 17.9%, demonstrating that reducing oversharing improves task success. Our findings underscore that protecting privacy in web agents is a fundamental challenge, requiring a broader view of “output” that accounts for what agents do on the web, not just what they type. Our datasets and code are available at https://github.com/jrohsc/SPILLage.

Method: Two-phase framework: 1) Offline indexing converts experiences into a hybrid memory graph linking time-aware gists and facts, 2) Online inference uses an agentic retriever with curated tools for iterative retrieval over the memory graph.

Result: REMem outperforms state-of-the-art memory systems (Mem0 and HippoRAG 2) with 3.4% and 13.4% absolute improvements on episodic recollection and reasoning tasks respectively. Also demonstrates more robust refusal behavior for unanswerable questions.

Conclusion: REMem effectively addresses episodic memory challenges in language agents through hybrid memory graphs and agentic retrieval, enabling better recollection and reasoning over interaction histories.

Abstract: Humans excel at remembering concrete experiences along spatiotemporal contexts and performing reasoning across those events, i.e., the capacity for episodic memory. In contrast, memory in language agents remains mainly semantic, and current agents are not yet capable of effectively recollecting and reasoning over interaction histories. We identify and formalize the core challenges of episodic recollection and reasoning from this gap, and observe that existing work often overlooks episodicity, lacks explicit event modeling, or overemphasizes simple retrieval rather than complex reasoning. We present REMem, a two-phase framework for constructing and reasoning with episodic memory: 1) Offline indexing, where REMem converts experiences into a hybrid memory graph that flexibly links time-aware gists and facts. 2) Online inference, where REMem employs an agentic retriever with carefully curated tools for iterative retrieval over the memory graph. Comprehensive evaluation across four episodic memory benchmarks shows that REMem substantially outperforms state-of-the-art memory systems such as Mem0 and HippoRAG 2, showing 3.4% and 13.4% absolute improvements on episodic recollection and reasoning tasks, respectively. Moreover, REMem also demonstrates more robust refusal behavior for unanswerable questions.

Method: Three core innovations: 1) Hierarchical multi-task fine-tuning with datasets for Planning, Acting, and Grounding to establish VLM with strong instruction-following; 2) Online RL with hybrid reward mechanism combining WebJudge for outcome assessment and Rule-based Decision Tree for progress reward; 3) Operator Agent (OpAgent) modular framework with Planner, Grounder, Reflector, and Summarizer for error recovery.

Result: RL-enhanced model achieves 38.1% success rate (pass@5) on WebArena, outperforming all monolithic baselines. OpAgent framework elevates performance to SOTA 71.6% success rate.

Conclusion: Proposed online RL web agent with hierarchical fine-tuning, hybrid rewards, and modular operator framework effectively addresses distributional shifts and achieves state-of-the-art performance on web navigation tasks.

Abstract: To fulfill user instructions, autonomous web agents must contend with the inherent complexity and volatile nature of real-world websites. Conventional paradigms predominantly rely on Supervised Fine-Tuning (SFT) or Offline Reinforcement Learning (RL) using static datasets. However, these methods suffer from severe distributional shifts, as offline trajectories fail to capture the stochastic state transitions and real-time feedback of unconstrained wide web environments. In this paper, we propose a robust Online Reinforcement Learning WebAgent, designed to optimize its policy through direct, iterative interactions with unconstrained wide websites. Our approach comprises three core innovations: 1) Hierarchical Multi-Task Fine-tuning: We curate a comprehensive mixture of datasets categorized by functional primitives – Planning, Acting, and Grounding – establishing a Vision-Language Model (VLM) with strong instruction-following capabilities for Web GUI tasks. 2) Online Agentic RL in the Wild: We develop an online interaction environment and fine-tune the VLM using a specialized RL pipeline. We introduce a Hybrid Reward Mechanism that combines a ground-truth-agnostic WebJudge for holistic outcome assessment with a Rule-based Decision Tree (RDT) for progress reward. This system effectively mitigates the credit assignment challenge in long-horizon navigation. Notably, our RL-enhanced model achieves a 38.1% success rate (pass@5) on WebArena, outperforming all existing monolithic baselines. 3) Operator Agent: We introduce a modular agentic framework, namely \textbf{OpAgent}, orchestrating a Planner, Grounder, Reflector, and Summarizer. This synergy enables robust error recovery and self-correction, elevating the agent’s performance to a new State-of-the-Art (SOTA) success rate of \textbf{71.6%}.

Method: Three binary decision-making tasks (reading comprehension, multi-step reasoning, moral judgment) with four instruction-tuned LLMs. Presented prior responses attributed to friends, human experts, or other LLMs, manipulating correctness and group size. Second experiment introduced direct disagreement between single human and single LLM.

Result: LLMs conform significantly more to responses labeled as coming from human experts, even when incorrect, and revise answers toward experts more readily than toward other LLMs. Expert framing acts as a strong prior that generalizes across decision domains.

Conclusion: LLMs exhibit credibility-sensitive social influence with human-expert bias, suggesting they internalize social priors about source credibility similar to humans.

Abstract: Large language models (LLMs) increasingly operate in environments where they encounter social information such as other agents’ answers, tool outputs, or human recommendations. In humans, such inputs influence judgments in ways that depend on the source’s credibility and the strength of consensus. This paper investigates whether LLMs exhibit analogous patterns of influence and whether they privilege feedback from humans over feedback from other LLMs. Across three binary decision-making tasks, reading comprehension, multi-step reasoning, and moral judgment, we present four instruction-tuned LLMs with prior responses attributed either to friends, to human experts, or to other LLMs. We manipulate whether the group is correct and vary the group size. In a second experiment, we introduce direct disagreement between a single human and a single LLM. Across tasks, models conform significantly more to responses labeled as coming from human experts, including when that signal is incorrect, and revise their answers toward experts more readily than toward other LLMs. These results reveal that expert framing acts as a strong prior for contemporary LLMs, suggesting a form of credibility-sensitive social influence that generalizes across decision domains.

Method: Integrates a self-supervised differentiable clustering model with a novel differentiable ILP model, enabling rule learning from raw data without explicit label leakage by automatically discovering symbolic representations from continuous inputs.

Result: The method successfully learns generalized rules from time series and image data, demonstrating intuitive and precise rule extraction from raw multimodal inputs without requiring explicit feature labeling.

Conclusion: The proposed integration of self-supervised clustering with differentiable ILP enables interpretable rule learning directly from raw data, overcoming the explicit label leakage problem and expanding ILP’s applicability to multimodal domains.

Abstract: Rule learning-based models are widely used in highly interpretable scenarios due to their transparent structures. Inductive logic programming (ILP), a form of machine learning, induces rules from facts while maintaining interpretability. Differentiable ILP models enhance this process by leveraging neural networks to improve robustness and scalability. However, most differentiable ILP methods rely on symbolic datasets, facing challenges when learning directly from raw data. Specifically, they struggle with explicit label leakage: The inability to map continuous inputs to symbolic variables without explicit supervision of input feature labels. In this work, we address this issue by integrating a self-supervised differentiable clustering model with a novel differentiable ILP model, enabling rule learning from raw data without explicit label leakage. The learned rules effectively describe raw data through its features. We demonstrate that our method intuitively and precisely learns generalized rules from time series and image data.

Method: Uses compact binary signatures for semantic search and lossless token-ID streams for exact content reconstruction. Core innovation is Dynamic Wavelet Matrix (DWM) that compresses and co-indexes both streams to support ultra-fast search in compressed domain.

Result: Reduces end-to-end retrieval latency by up to 31×, cuts per-query token footprint by up to 14×, while maintaining accuracy on LoCoMo and LongMemEval benchmarks.

Conclusion: Hippocampus provides a scalable, efficient memory management system for agentic AI that overcomes limitations of existing approaches with linear scaling and significant performance improvements.

Abstract: Agentic AI require persistent memory to store user-specific histories beyond the limited context window of LLMs. Existing memory systems use dense vector databases or knowledge-graph traversal (or hybrid), incurring high retrieval latency and poor storage scalability. We introduce Hippocampus, an agentic memory management system that uses compact binary signatures for semantic search and lossless token-ID streams for exact content reconstruction. Its core is a Dynamic Wavelet Matrix (DWM) that compresses and co-indexes both streams to support ultra-fast search in the compressed domain, thus avoiding costly dense-vector or graph computations. This design scales linearly with memory size, making it suitable for long-horizon agentic deployments. Empirically, our evaluation shows that Hippocampus reduces end-to-end retrieval latency by up to 31$\times$ and cuts per-query token footprint by up to 14$\times$, while maintaining accuracy on both LoCoMo and LongMemEval benchmarks.

Method: The authors demonstrate through empirical evidence and theoretical decomposition that reducing precision from 16-bit to 8/4-bit creates a ‘quantization trap’. They analyze hardware casting overhead and hidden latency costs of dequantization kernels, showing these become dominant bottlenecks in sequential reasoning chains.

Result: The study reveals that quantization paradoxically increases net energy consumption while degrading reasoning accuracy for multi-hop reasoning tasks. The scaling law breaking is shown to be unavoidable in practice due to sequential energy amortization failure.

Conclusion: The industry’s “smaller-is-better” heuristic for quantization is mathematically counterproductive for complex reasoning tasks. The findings suggest that different optimization strategies are needed for reasoning-intensive AI applications compared to standard neural network inference.

Abstract: Neural scaling laws provide a predictable recipe for AI advancement: reducing numerical precision should linearly improve computational efficiency and energy profile (E proportional to bits). In this paper, we demonstrate that this scaling law breaks in the context of multi-hop reasoning. We reveal a ‘quantization trap’ where reducing precision from 16-bit to 8/4-bit paradoxically increases more net energy consumption while degrading reasoning accuracy. We provide a rigorous theoretical decomposition that attributes this failure to hardware casting overhead, the hidden latency cost of dequantization kernels, which becomes a dominant bottleneck in sequential reasoning chains, as well as to a sequential energy amortization failure. As a result, scaling law breaking is unavoidable in practice. Our findings suggest that the industry’s “smaller-is-better” heuristic is mathematically counterproductive for complex reasoning tasks.

Method: Proposes DiffusionRollout, a selective rollout planning strategy that uses the model’s predictive uncertainty (measured via standard deviations over multiple samples) to adaptively select step sizes during autoregressive rollouts. The method leverages the correlation between prediction errors and uncertainty measures to make more reliable long-term predictions.

Result: Extensive evaluation on long-trajectory PDE prediction benchmarks shows the proposed uncertainty measure effectively correlates with prediction errors, and the adaptive planning strategy reduces prediction errors and enables longer predicted trajectories that maintain high correlation with ground truths.

Conclusion: DiffusionRollout successfully mitigates error accumulation in long-horizon PDE predictions by using uncertainty quantification to guide adaptive step-size selection, improving the reliability and accuracy of autoregressive diffusion models for physical system forecasting.

Abstract: We propose DiffusionRollout, a novel selective rollout planning strategy for autoregressive diffusion models, aimed at mitigating error accumulation in long-horizon predictions of physical systems governed by partial differential equations (PDEs). Building on the recently validated probabilistic approach to PDE solving, we further explore its ability to quantify predictive uncertainty and demonstrate a strong correlation between prediction errors and standard deviations computed over multiple samples-supporting their use as a proxy for the model’s predictive confidence. Based on this observation, we introduce a mechanism that adaptively selects step sizes during autoregressive rollouts, improving long-term prediction reliability by reducing the compounding effect of conditioning on inaccurate prior outputs. Extensive evaluation on long-trajectory PDE prediction benchmarks validates the effectiveness of the proposed uncertainty measure and adaptive planning strategy, as evidenced by lower prediction errors and longer predicted trajectories that retain a high correlation with their ground truths.

Method: Two-component framework: 1) Agentic-Q estimation - optimizes a Q-model to generate step-wise values evaluating action contributions to task completion; 2) Step-wise policy optimization - takes step-wise samples from state-action trajectories and optimizes policy via reinforcement learning with the agentic-Q model. The framework uses self-generated trajectories and decouples policy updates from the environment.

Result: The framework endows Ovis2.5-9B with powerful GUI interaction capabilities, achieving remarkable performances on GUI navigation and grounding benchmarks, surpassing larger-scale contenders.

Conclusion: The proposed MLLM-centered framework provides an efficient solution for GUI agents in non-stationary environments, enabling stable optimization and reducing data collection costs while achieving state-of-the-art performance.

Abstract: Recent advances in Multimodal Large Language Models (MLLMs) have substantially driven the progress of autonomous agents for Graphical User Interface (GUI). Nevertheless, in real-world applications, GUI agents are often faced with non-stationary environments, leading to high computational costs for data curation and policy optimization. In this report, we introduce a novel MLLM-centered framework for GUI agents, which consists of two components: agentic-Q estimation and step-wise policy optimization. The former one aims to optimize a Q-model that can generate step-wise values to evaluate the contribution of a given action to task completion. The latter one takes step-wise samples from the state-action trajectory as inputs, and optimizes the policy via reinforcement learning with our agentic-Q model. It should be noticed that (i) all state-action trajectories are produced by the policy itself, so that the data collection costs are manageable; (ii) the policy update is decoupled from the environment, ensuring stable and efficient optimization. Empirical evaluations show that our framework endows Ovis2.5-9B with powerful GUI interaction capabilities, achieving remarkable performances on GUI navigation and grounding benchmarks and even surpassing contenders with larger scales.

Method: Proposed two Multi-Agent Actor-Critic (MAAC) approaches: CoLLM-CC with centralized critic and CoLLM-DC with decentralized critics, comparing them against Monte Carlo methods across writing, coding, and game-playing domains.

Result: Monte Carlo and CoLLM-DC perform comparably to CoLLM-CC on short-horizon/dense-reward tasks, but underperform on long-horizon/sparse-reward tasks where Monte Carlo requires more samples and CoLLM-DC struggles to converge.

Conclusion: MAAC methods, particularly centralized critic approaches, are beneficial for optimizing decentralized LLM collaboration in challenging scenarios with long horizons or sparse rewards.

Abstract: Recent work has explored optimizing LLM collaboration through Multi-Agent Reinforcement Learning (MARL). However, most MARL fine-tuning approaches rely on predefined execution protocols, which often require centralized execution. Decentralized LLM collaboration is more appealing in practice, as agents can run inference in parallel with flexible deployments. Also, current approaches use Monte Carlo methods for fine-tuning, which suffer from high variance and thus require more samples to train effectively. Actor-critic methods are prevalent in MARL for dealing with these issues, so we developed Multi-Agent Actor-Critic (MAAC) methods to optimize decentralized LLM collaboration. In this paper, we analyze when and why these MAAC methods are beneficial. We propose 2 MAAC approaches, \textbf{CoLLM-CC} with a \textbf{C}entralized \textbf{C}ritic and \textbf{CoLLM-DC} with \textbf{D}ecentralized \textbf{C}ritics. Our experiments across writing, coding, and game-playing domains show that Monte Carlo methods and CoLLM-DC can achieve performance comparable to CoLLM-CC in short-horizon and dense-reward settings. However, they both underperform CoLLM-CC on long-horizon or sparse-reward tasks, where Monte Carlo methods require substantially more samples and CoLLM-DC struggles to converge. Our code is available at https://github.com/OpenMLRL/CoMLRL/releases/tag/v1.3.6.

Method: HyFunc employs a hybrid-model cascade: a large model distills user intent into a single “soft token” that guides a lightweight retriever to select relevant functions and directs a smaller, prefix-tuned model to generate the final call. Uses “dynamic templating” to inject boilerplate parameter syntax on-the-fly within an extended vLLM engine.

Result: Achieves inference latency of 0.828 seconds (outperforming all baselines) and performance of 80.1% (surpassing comparable-scale models) on unseen BFCL benchmark dataset. Demonstrates excellent balance between efficiency and performance.

Conclusion: HyFunc offers a more efficient paradigm for agentic AI by systematically eliminating computational redundancies in LLM-based function calling while maintaining strong performance.

Abstract: While agentic AI systems rely on LLMs to translate user intent into structured function calls, this process is fraught with computational redundancy, leading to high inference latency that hinders real-time applications. This paper identifies and addresses three key redundancies: (1) the redundant processing of a large library of function descriptions for every request; (2) the redundant use of a large, slow model to generate an entire, often predictable, token sequence; and (3) the redundant generation of fixed, boilerplate parameter syntax. We introduce HyFunc, a novel framework that systematically eliminates these inefficiencies. HyFunc employs a hybrid-model cascade where a large model distills user intent into a single “soft token.” This token guides a lightweight retriever to select relevant functions and directs a smaller, prefix-tuned model to generate the final call, thus avoiding redundant context processing and full-sequence generation by the large model. To eliminate syntactic redundancy, our “dynamic templating” technique injects boilerplate parameter syntax on-the-fly within an extended vLLM engine. To avoid potential limitations in generalization, we evaluate HyFunc on an unseen benchmark dataset, BFCL. Experimental results demonstrate that HyFunc achieves an excellent balance between efficiency and performance. It achieves an inference latency of 0.828 seconds, outperforming all baseline models, and reaches a performance of 80.1%, surpassing all models with a comparable parameter scale. These results suggest that HyFunc offers a more efficient paradigm for agentic AI. Our code is publicly available at https://github.com/MrBlankness/HyFunc.

Method: Introduced behavior-centered evaluation framework distinguishing between persuasion during vs. prior to task execution. Tested across web research and coding tasks with belief-level interventions, comparing on-the-fly persuasion vs. belief-prefilled agents.

Result: On-the-fly persuasion induced weak/inconsistent behavioral effects. Belief-prefilled agents conducted 26.9% fewer searches and visited 16.9% fewer unique sources than neutral-prefilled agents.

Conclusion: Persuasion can significantly affect agent behavior, especially when beliefs are established before task execution, motivating behavior-level evaluation in agentic systems.

Abstract: Modern AI agents increasingly combine conversational interaction with autonomous task execution, such as coding and web research, raising a natural question: what happens when an agent engaged in long-horizon tasks is subjected to user persuasion? We study how belief-level intervention can influence downstream task behavior, a phenomenon we name \emph{persuasion propagation}. We introduce a behavior-centered evaluation framework that distinguishes between persuasion applied during or prior to task execution. Across web research and coding tasks, we find that on-the-fly persuasion induces weak and inconsistent behavioral effects. In contrast, when the belief state is explicitly specified at task time, belief-prefilled agents conduct on average 26.9% fewer searches and visit 16.9% fewer unique sources than neutral-prefilled agents. These results suggest that persuasion, even in prior interaction, can affect the agent’s behavior, motivating behavior-level evaluation in agentic systems.

Method: Integrates Sliding Window Attention (SWA) with non-linear Test-Time Training (TTT) memory networks. Implements Memory-Efficient Fine-Tuning to replace standard attention layers with memory-augmented sliding window layers in pre-trained LLMs.

Result: 4k window model achieves near-lossless performance on 37k LongBench with only 0.83 drop vs full attention. 8k window variant outperforms full attention on InfiniteBench at 128k context, validating effectiveness in mitigating noise and maintaining robust long-range modeling.

Conclusion: AllMem enables efficient scaling to ultra-long contexts while overcoming representation constraints of linear memory models and reducing computational/memory footprint during long-sequence inference.

Abstract: Large Language Models (LLMs) encounter significant performance bottlenecks in long-sequence tasks due to the computational complexity and memory overhead inherent in the self-attention mechanism. To address these challenges, we introduce \textsc{AllMem}, a novel and efficient hybrid architecture that integrates Sliding Window Attention (SWA) with non-linear Test-Time Training (TTT) memory networks. \textsc{AllMem} enables models to effectively scale to ultra-long contexts while mitigating catastrophic forgetting. This approach not only overcomes the representation constraints typical of linear memory models but also significantly reduces the computational and memory footprint during long-sequence inference. Furthermore, we implement a Memory-Efficient Fine-Tuning strategy to replace standard attention layers in pre-trained models with memory-augmented sliding window layers. This framework facilitates the efficient transformation of any off-the-shelf pre-trained LLM into an \textsc{AllMem}-based architecture. Empirical evaluations confirm that our 4k window model achieves near-lossless performance on 37k LongBench with a marginal 0.83 drop compared to full attention. Furthermore, on InfiniteBench at a 128k context, our 8k window variant outperforms full attention, which validates the effectiveness of our parameterized memory in mitigating noise and maintaining robust long-range modeling without the prohibitive costs of global attention.

Method: ROMA uses recursive task decomposition into dependency-aware subtask trees with parallel execution, plus structured aggregation to control context growth. It features four modular roles (Atomizer, Planner, Executor, Aggregator) and supports heterogeneous multi-agent systems. GEPA+ enables prompt optimization without fine-tuning.

Result: ROMA with GEPA+ achieves state-of-the-art performance: 9.9% accuracy improvement on SEAL-0 reasoning benchmark over Kimi-Researcher, and enables DeepSeek-V3 to match Claude Sonnet 4.5 on EQ-Bench long-form writing.

Conclusion: Recursive modular agent architectures can scale reasoning depth while maintaining interpretability, flexibility, and model-agnostic properties for long-horizon tasks.

Abstract: Current agentic frameworks underperform on long-horizon tasks. As reasoning depth increases, sequential orchestration becomes brittle, context windows impose hard limits that degrade performance, and opaque execution traces make failures difficult to localize or debug. We introduce ROMA (Recursive Open Meta-Agents), a domain-agnostic framework that addresses these limitations through recursive task decomposition and structured aggregation. ROMA decomposes goals into dependency-aware subtask trees that can be executed in parallel, while aggregation compresses and validates intermediate results to control context growth. Our framework standardizes agent construction around four modular roles –Atomizer (which decides whether a task should be decomposed), Planner, Executor, and Aggregator – which cleanly separate orchestration from model selection and enable transparent, hierarchical execution traces. This design supports heterogeneous multi-agent systems that mix models and tools according to cost, latency, and capability. To adapt ROMA to specific tasks without fine-tuning, we further introduce GEPA$+$, an improved Genetic-Pareto prompt proposer that searches over prompts within ROMA’s component hierarchy while preserving interface contracts. We show that ROMA, combined with GEPA+, delivers leading system-level performance on reasoning and long-form generation benchmarks. On SEAL-0, which evaluates reasoning over conflicting web evidence, ROMA instantiated with GLM-4.6 improves accuracy by 9.9% over Kimi-Researcher. On EQ-Bench, a long-form writing benchmark, ROMA enables DeepSeek-V3 to match the performance of leading closed-source models such as Claude Sonnet 4.5. Our results demonstrate that recursive, modular agent architectures can scale reasoning depth while remaining interpretable, flexible, and model-agnostic.

Method: Proposes Pheromone-Guided Policy Optimization (PhGPO) that learns trajectory-based transition patterns (pheromone) from historical trajectories and uses this learned pheromone to guide policy optimization toward historically successful tool transitions.

Result: Comprehensive experimental results demonstrate the effectiveness of PhGPO in improving long-horizon tool planning.

Conclusion: Learning reusable transition patterns from historical trajectories provides explicit guidance that significantly improves LLM agent tool planning capabilities.

Abstract: Recent advancements in Large Language Model (LLM) agents have demonstrated strong capabilities in executing complex tasks through tool use. However, long-horizon multi-step tool planning is challenging, because the exploration space suffers from a combinatorial explosion. In this scenario, even when a correct tool-use path is found, it is usually considered an immediate reward for current training, which would not provide any reusable information for subsequent training. In this paper, we argue that historically successful trajectories contain reusable tool-transition patterns, which can be leveraged throughout the whole training process. Inspired by ant colony optimization where historically successful paths can be reflected by the pheromone, we propose Pheromone-Guided Policy Optimization (PhGPO), which learns a trajectory-based transition pattern (i.e., pheromone) from historical trajectories and then uses the learned pheromone to guide policy optimization. This learned pheromone provides explicit and reusable guidance that steers policy optimization toward historically successful tool transitions, thereby improving long-horizon tool planning. Comprehensive experimental results demonstrate the effectiveness of our proposed PhGPO.

Method: Developed a streamlined automated pipeline optimized for citation-based verification using next-generation models (Gemini 3 Pro, GPT-5.2 Pro), evaluated on two novel datasets: ICCM problem sets (comparable to Yau College Contest) and “First Proof” problem set of unpublished research questions.

Result: Pipeline generated candidate proofs for all problems in both datasets; solutions for first two ICCM sets and Problem 4 of “First Proof” set were fully verified by the team; all generated proofs submitted to official organization and results made publicly available.

Conclusion: Next-generation LLMs integrated into optimized automated pipelines can successfully solve sophisticated research-grade mathematical problems, demonstrating practical deployment potential for mathematical research applications.

Abstract: Large language models (LLMs) have recently achieved remarkable success in generating rigorous mathematical proofs, with “AI for Math” emerging as a vibrant field of research. While these models have mastered competition-level benchmarks like the International Mathematical Olympiad and show promise in research applications through auto-formalization, their deployment via lightweight, natural-language pipelines for research problems remains underexplored. In this work, we demonstrate that next-generation models (e.g., Gemini 3 Pro, GPT-5.2 Pro), when integrated into a streamlined automated pipeline optimized for citation-based verification, can solve sophisticated research-grade problems. We evaluate our pipeline on two novel datasets: (1) the ICCM problem sets (comparable to the S.-T. Yau College Student Mathematics Contest) proposed by leading mathematicians, and (2) the “First Proof” problem set, consisting of previously unpublished research questions. Our pipeline generated candidate proofs for all problems in the first two ICCM sets and the “First Proof” set. The solutions for the first two ICCM sets and Problem 4 of the “First Proof” set have been fully verified by our team. All generated proofs have been submitted to the official organization, and our generated results are publicly available. We plan to open-source the complete pipeline methodology in due course.

Method: Proposes a principled RDB encoder family that compresses variably-sized database neighborhoods into fixed-length samples using column-wise compression within homogeneous columns (not across heterogeneous columns), implemented via scalable SQL primitives without trainable parameters.

Result: Develops an open-source RDB foundation model capable of robust performance on unseen datasets without training or fine-tuning, seamlessly pairing with existing single-table ICL foundation models.

Conclusion: The approach enables practical in-context learning across relational databases by constraining compression within homogeneous columns and eliminating trainable parameters, providing an easy-to-use solution for enterprise predictive modeling.

Abstract: Relational databases (RDBs) contain vast amounts of heterogeneous tabular information that can be exploited for predictive modeling purposes. But since the space of potential targets is vast across enterprise settings, how can we \textit{avoid retraining} a new model each time we wish to predict a new quantity of interest? Foundation models based on in-context learning (ICL) offer a convenient option, but so far are largely restricted to single-table operability. In generalizing to multiple interrelated tables, it is essential to compress variably-sized RDB neighborhoods into fixed-length ICL samples for consumption by the decoder. However, the details here are critical: unlike existing supervised learning RDB pipelines, we provide theoretical and empirical evidence that ICL-specific compression should be constrained \emph{within} high-dimensional RDB columns where all entities share units and roles, not \textit{across} columns where the relevance of heterogeneous data types cannot possibly be determined without label information. Conditioned on this restriction, we then demonstrate that encoder expressiveness is actually not compromised by excluding trainable parameters. Hence we arrive at a principled family of RDB encoders that can be seamlessly paired with already-existing single-table ICL foundation models, whereby no training or fine-tuning is required. From a practical standpoint, we develop scalable SQL primitives to implement the encoder stage, resulting in an easy-to-use open-source RDB foundation model\footnote{\label{foot: RDBLearn_learn} https://github.com/HKUSHXLab/rdblearn} capable of robust performance on unseen datasets out of the box.

Method: Render textual reasoning steps into images, then compress intermediate reasoning into a single latent token using supervision from rendered CoT images and DeepSeek-OCR hidden states. This creates deterministic supervision signals that are inspectable without verbose textual output.

Result: 11x average output length reduction with only 2.21% accuracy drop relative to textual CoT. 6.8x improvement in output token contribution. Achieves 99.80% on ProntoQA and 97.80% on ProsQA with one latent token, with compression up to 87.4x.

Conclusion: OneLatent effectively compresses reasoning processes into latent representations while maintaining performance, enabling efficient inference for complex reasoning tasks with compression-constrained generalization.

Abstract: Chain-of-thought (CoT) prompting improves reasoning but often increases inference cost by one to two orders of magnitude. To address these challenges, we present \textbf{OneLatent}, a framework that compresses intermediate reasoning into a single latent token via supervision from rendered CoT images and DeepSeek-OCR hidden states. By rendering textual steps into images, we obtain a deterministic supervision signal that can be inspected and audited without requiring the model to output verbose textual rationales. Across benchmarks, OneLatent reduces average output length by $11\times$ with only a $2.21%$ average accuracy drop relative to textual CoT, while improving output token contribution (OTC) by $6.8\times$. On long-chain logical reasoning, OneLatent reaches $99.80%$ on ProntoQA and $97.80%$ on ProsQA with one latent token, with compression up to $87.4\times$, supporting compression-constrained generalization.

Method: Configurable multi-agent framework with tree-based workflow for branching hypothesis generation and backtracking; evolutionary-systematic ideation combining evolutionary selection with comprehensive planning; hierarchical reflection with verbal gradients, verbal momentum, and memory compression

Result: Outperforms evolutionary baselines on combinatorial optimization benchmarks (traveling salesman, vehicle routing, bin packing, etc.) and simulation-based cooperative driving scenarios

Conclusion: OR-Agent provides a general, extensible, inspectable framework for AI-assisted scientific discovery that goes beyond simple mutation-crossover loops through structured hypothesis management

Abstract: Automating scientific discovery in complex, experiment-driven domains requires more than iterative mutation of programs; it demands structured hypothesis management, environment interaction, and principled reflection. We present OR-Agent, a configurable multi-agent research framework designed for automated exploration in rich experimental environments. OR-Agent organizes research as a structured tree-based workflow that explicitly models branching hypothesis generation and systematic backtracking, enabling controlled management of research trajectories beyond simple mutation-crossover loops. At its core, we introduce an evolutionary-systematic ideation mechanism that unifies evolutionary selection of research starting points, comprehensive research plan generation, and coordinated exploration within a research tree. We further propose a hierarchical optimization-inspired reflection system: short-term experimental reflection operates as a form of verbal gradient providing immediate corrective signals; long-term reflection accumulates cross-experiment insights as verbal momentum; and memory compression serves as a regularization mechanism analogous to weight decay, preserving essential signals while mitigating drift. Together, these components form a principled architecture governing research dynamics. We conduct extensive experiments across classical combinatorial optimization benchmarks-including traveling salesman, capacitated vehicle routing, bin packing, orienteering, and multiple knapsack problems-as well as simulation-based cooperative driving scenarios. Results demonstrate that OR-Agent outperforms strong evolutionary baselines while providing a general, extensible, and inspectable framework for AI-assisted scientific discovery. OR-Agent source code and experiments data are publicly available at https://github.com/qiliuchn/OR-Agent.

Method: StackingNet uses a meta-ensemble framework based on collective intelligence principles to combine model predictions during inference. It operates without access to internal parameters or training data, enabling reliability ranking, identifying/pruning underperforming models, and reducing bias.

Result: Across language comprehension, visual estimation, and academic paper rating tasks, StackingNet consistently improves accuracy, robustness, and fairness compared to individual models and classic ensembles. It turns model diversity from inconsistency into productive collaboration.

Conclusion: StackingNet establishes a practical foundation for coordinated AI, suggesting progress may come from principled cooperation among specialized models rather than just larger single models. It enables trustworthy intelligent systems through collective intelligence.

Abstract: Artificial intelligence built on large foundation models has transformed language understanding, vision and reasoning, yet these systems remain isolated and cannot readily share their capabilities. Integrating the complementary strengths of such independent foundation models is essential for building trustworthy intelligent systems. Despite rapid progress in individual model design, there is no established approach for coordinating such black-box heterogeneous models. Here we show that coordination can be achieved through a meta-ensemble framework termed StackingNet, which draws on principles of collective intelligence to combine model predictions during inference. StackingNet improves accuracy, reduces bias, enables reliability ranking, and identifies or prunes models that degrade performance, all operating without access to internal parameters or training data. Across tasks involving language comprehension, visual estimation, and academic paper rating, StackingNet consistently improves accuracy, robustness, and fairness, compared with individual models and classic ensembles. By turning diversity from a source of inconsistency into collaboration, StackingNet establishes a practical foundation for coordinated artificial intelligence, suggesting that progress may emerge from not only larger single models but also principled cooperation among many specialized ones.

Method: 1) Formalize attention as projection onto convex hull of key vectors with entropic (softmax-like) relaxation; 2) Prove face-stability theorem showing attention concentrates on constant-size active face under strict complementarity margin; 3) Introduce Vashista Sparse Attention - a drop-in mechanism maintaining small candidate set per query through paging-style context selection compatible with modern inference stacks.

Result: Theoretical: Attention mass on inactive tokens decays exponentially while error on active face scales linearly with temperature. Practical: Across long-context evaluations, stable constant-size effective support, strong wall-clock speedups, and minimal quality degradation in regimes predicted by support-gap diagnostics.

Conclusion: The paper provides a principled theoretical foundation for sparse attention mechanisms, enabling predictable latency and cost without external retrieval dependencies. Vashista Sparse Attention offers practical deployment benefits for privacy-sensitive and air-gapped settings through interchangeable attention modules.

Abstract: Large language models spend most of their inference cost on attention over long contexts, yet empirical behavior suggests that only a small subset of tokens meaningfully contributes to each query. We formalize this phenomenon by modeling attention as a projection onto the convex hull of key vectors and analyzing its entropic (softmax-like) relaxation. Our main theoretical contribution is a face-stability theorem showing that, under a strict complementarity margin (a support gap (Δ) certified by KKT multipliers), entropic attention concentrates on a constant-size active face: the total mass assigned to inactive tokens decays exponentially as (\exp(-Ω(Δ/\varepsilon))), while the error on the active face scales linearly in the temperature/regularization parameter (\varepsilon). This yields a practical criterion for when sparse long-context decoding is safe and provides a principled knob to trade accuracy for compute. Building on these guarantees, we introduce Vashista Sparse Attention, a drop-in mechanism that maintains a small candidate set per query through a paging-style context selection strategy compatible with modern inference stacks. Across long-context evaluations, we observe stable constant-size effective support, strong wall-clock speedups, and minimal quality degradation in the regimes predicted by the support-gap diagnostics. Finally, we discuss deployment implications for privacy-sensitive and air-gapped settings, where interchangeable attention modules enable predictable latency and cost without external retrieval dependencies.

Method: End-to-end pipeline that parses contractual text into structured schemas, generates Solidity code using CrewAI-style agent teams with iterative refinement, and performs automated quality assessment through compilation checks, security analysis, and multi-dimensional evaluation metrics.

Result: Framework produces structured artifacts with provenance metadata and measures quality across five dimensions: functional completeness, variable fidelity, state-machine correctness, business-logic fidelity, and code quality, enabling paired evaluation against ground-truth implementations.

Conclusion: Provides reproducible benchmark for empirical research on smart contract synthesis quality and supports extensions to formal verification and compliance checking.

Abstract: We present an end-to-end framework for systematic evaluation of LLM-generated smart contracts from natural-language specifications. The system parses contractual text into structured schemas, generates Solidity code, and performs automated quality assessment through compilation and security checks. Using CrewAI-style agent teams with iterative refinement, the pipeline produces structured artifacts with full provenance metadata. Quality is measured across five dimensions, including functional completeness, variable fidelity, state-machine correctness, business-logic fidelity, and code quality aggregated into composite scores. The framework supports paired evaluation against ground-truth implementations, quantifying alignment and identifying systematic error modes such as logic omissions and state transition inconsistencies. This provides a reproducible benchmark for empirical research on smart contract synthesis quality and supports extensions to formal verification and compliance checking.

Method: Leverages treatment embeddings and historical outcomes to train a CTR ranking model with fixed effects for contextual shifts. Projects treatments onto semantic marketing attributes and uses sign-consistent, sparse constrained Lasso for interpretability. Computes opportunity index combining attribute importance with under-expression. Uses LLMs to translate opportunities into creative suggestions and estimate learning/conversion potential.

Result: The framework has been built into Adobe’s “Experimentation Accelerator” product. Evaluation on real-world experiments by Adobe business customers validates the high quality of the generation pipeline.

Conclusion: The proposed framework enables faster, more informative, and more efficient test cycles by providing AI-based insights and opportunities to scale experimentation.

Abstract: Modern online experimentation faces two bottlenecks: scarce traffic forces tough choices on which variants to test, and post-hoc insight extraction is manual, inconsistent, and often content-agnostic. Meanwhile, organizations underuse historical A/B results and rich content embeddings that could guide prioritization and creative iteration. We present a unified framework to (i) prioritize which variants to test, (ii) explain why winners win, and (iii) surface targeted opportunities for new, higher-potential variants. Leveraging treatment embeddings and historical outcomes, we train a CTR ranking model with fixed effects for contextual shifts that scores candidates while balancing value and content diversity. For better interpretability and understanding, we project treatments onto curated semantic marketing attributes and re-express the ranker in this space via a sign-consistent, sparse constrained Lasso, yielding per-attribute coefficients and signed contributions for visual explanations, top-k drivers, and natural-language insights. We then compute an opportunity index combining attribute importance (from the ranker) with under-expression in the current experiment to flag missing, high-impact attributes. Finally, LLMs translate ranked opportunities into concrete creative suggestions and estimate both learning and conversion potential, enabling faster, more informative, and more efficient test cycles. These components have been built into a real Adobe product, called \textit{Experimentation Accelerator}, to provide AI-based insights and opportunities to scale experimentation for customers. We provide an evaluation of the performance of the proposed framework on some real-world experiments by Adobe business customers that validate the high quality of the generation pipeline.

Method: Two-step approach: 1) MOC-HER integrates Hindsight Experience Replay into MOC framework for goal relabeling; 2) 2HER introduces dual objective relabeling with two sets of virtual goals - one based on object’s final state (standard HER) and another based on agent’s effector positions to reward both interaction and task completion.

Result: MOC-2HER achieves up to 90% success rate in robotic manipulation environments, compared to less than 11% for both original MOC and MOC-HER, demonstrating effectiveness of dual objective relabeling strategy.

Conclusion: The dual objective hindsight experience replay (2HER) effectively addresses sparse reward challenges in hierarchical RL for multi-goal object manipulation tasks by rewarding both object interaction and task completion through separate goal relabeling mechanisms.

Abstract: Hierarchical Reinforcement Learning (HRL) frameworks like Option-Critic (OC) and Multi-updates Option Critic (MOC) have introduced significant advancements in learning reusable options. However, these methods underperform in multi-goal environments with sparse rewards, where actions must be linked to temporally distant outcomes. To address this limitation, we first propose MOC-HER, which integrates the Hindsight Experience Replay (HER) mechanism into the MOC framework. By relabeling goals from achieved outcomes, MOC-HER can solve sparse reward environments that are intractable for the original MOC. However, this approach is insufficient for object manipulation tasks, where the reward depends on the object reaching the goal rather than on the agent’s direct interaction. This makes it extremely difficult for HRL agents to discover how to interact with these objects. To overcome this issue, we introduce Dual Objectives Hindsight Experience Replay (2HER), a novel extension that creates two sets of virtual goals. In addition to relabeling goals based on the object’s final state (standard HER), 2HER also generates goals from the agent’s effector positions, rewarding the agent for both interacting with the object and completing the task. Experimental results in robotic manipulation environments show that MOC-2HER achieves success rates of up to 90%, compared to less than 11% for both MOC and MOC-HER. These results highlight the effectiveness of our dual objective relabeling strategy in sparse reward, multi-goal tasks.

Method: The framework randomly masks a subset of already-observed measurements and supervises on them, making the model unable to distinguish between “truly unobserved” and “artificially unobserved” points, forcing it to produce plausible predictions everywhere. This enables learning from partial observations without complete training data.

Result: Ambient Physics achieves state-of-the-art reconstruction performance with a 62.51% reduction in average overall error compared to prior diffusion-based methods, while using 125× fewer function evaluations. The method also identifies a “one-point transition” phenomenon where masking a single observed point enables learning across architectures and measurement patterns.

Conclusion: Ambient Physics enables scientific progress in settings where complete observations are unavailable by allowing models to learn directly from partial observations, overcoming a key limitation of existing diffusion-based approaches.

Abstract: In many scientific settings, acquiring complete observations of PDE coefficients and solutions can be expensive, hazardous, or impossible. Recent diffusion-based methods can reconstruct fields given partial observations, but require complete observations for training. We introduce Ambient Physics, a framework for learning the joint distribution of coefficient-solution pairs directly from partial observations, without requiring a single complete observation. The key idea is to randomly mask a subset of already-observed measurements and supervise on them, so the model cannot distinguish “truly unobserved” from “artificially unobserved”, and must produce plausible predictions everywhere. Ambient Physics achieves state-of-the-art reconstruction performance. Compared with prior diffusion-based methods, it achieves a 62.51$%$ reduction in average overall error while using 125$\times$ fewer function evaluations. We also identify a “one-point transition”: masking a single already-observed point enables learning from partial observations across architectures and measurement patterns. Ambient Physics thus enables scientific progress in settings where complete observations are unavailable.

Method: VSAL incorporates an adaptive layout generator that produces tailored graph visualizations. The framework processes these optimized visual representations to detect various graph properties like Hamiltonian cycles, planarity, claw-freeness, and tree structures.

Result: Extensive experiments show VSAL outperforms state-of-the-art vision-based methods on multiple graph property detection tasks including Hamiltonian cycle, planarity, claw-freeness, and tree detection.

Conclusion: Adaptive layout generation significantly improves vision-based graph property detection by providing more informative visual representations tailored to individual graph instances, overcoming limitations of fixed-layout approaches.

Abstract: Graph property detection aims to determine whether a graph exhibits certain structural properties, such as being Hamiltonian. Recently, learning-based approaches have shown great promise by leveraging data-driven models to detect graph properties efficiently. In particular, vision-based methods offer a visually intuitive solution by processing the visualizations of graphs. However, existing vision-based methods rely on fixed visual graph layouts, and therefore, the expressiveness of their pipeline is restricted. To overcome this limitation, we propose VSAL, a vision-based framework that incorporates an adaptive layout generator capable of dynamically producing informative graph visualizations tailored to individual instances, thereby improving graph property detection. Extensive experiments demonstrate that VSAL outperforms state-of-the-art vision-based methods on various tasks such as Hamiltonian cycle, planarity, claw-freeness, and tree detection.

Method: Created a set of concrete metrics that are simple to implement, computationally inexpensive, and task-agnostic to understand and discriminate between CoT pathologies. Developed model organisms deliberately trained to exhibit specific CoT pathologies to validate the approach.

Result: The work provides a practical toolkit for assessing CoT pathologies, with direct implications for training-time monitoring. The metrics successfully identify and distinguish between the three types of reasoning pathologies.

Conclusion: The paper offers a systematic approach to detecting and understanding Chain-of-Thought reasoning pathologies in LLMs, which is crucial for AI safety and reliable model monitoring during training.

Abstract: Chain-of-thought (CoT) reasoning is fundamental to modern LLM architectures and represents a critical intervention point for AI safety. However, CoT reasoning may exhibit failure modes that we note as pathologies, which prevent it from being useful for monitoring. Prior work has identified three distinct pathologies: post-hoc rationalization, where models generate plausible explanations backwards from predetermined answers; encoded reasoning, where intermediate steps conceal information within seemingly interpretable text; and internalized reasoning, where models replace explicit reasoning with meaningless filler tokens while computing internally. To better understand and discriminate between these pathologies, we create a set of concrete metrics that are simple to implement, computationally inexpensive, and task-agnostic. To validate our approach, we develop model organisms deliberately trained to exhibit specific CoT pathologies. Our work provides a practical toolkit for assessing CoT pathologies, with direct implications for training-time monitoring.

Method: Reformulates layout design as policy learning over structured textual spatial environment encoding canvas geometry, element attributes, and relationships. Uses multi-objective spatial critique (geometric validity, relational coherence, aesthetic consistency) and relative group optimization.

Result: Improves structural validity and visual quality, outperforms larger proprietary LLMs, achieves performance comparable to specialized SOTA layout generators with fewer samples and reduced latency.

Conclusion: LaySPA successfully equips LLMs with explicit spatial reasoning for transparent and controllable graphic layout design.

Abstract: We introduce LaySPA, a reinforcement learning framework that equips large language models (LLMs) with explicit and interpretable spatial reasoning for content-aware graphic layout design. LaySPA addresses two key challenges: LLMs’ limited spatial reasoning and the lack of opacity in design decision making. Instead of operating at the pixel level, we reformulate layout design as a policy learning problem over a structured textual spatial environment that explicitly encodes canvas geometry, element attributes, and inter-element relationships. LaySPA produces dual-level outputs comprising interpretable reasoning traces and structured layout specifications, enabling transparent and controllable design decision making. Layout design policy is optimized via a multi-objective spatial critique that decomposes layout quality into geometric validity, relational coherence, and aesthetic consistency, and is trained using relative group optimization to stabilize learning in open-ended design spaces. Experiments demonstrate that LaySPA improves structural validity and visual quality, outperforming larger proprietary LLMs and achieving performance comparable to specialized SOTA layout generators while requiring fewer annotated samples and reduced latency.

Method: HyMem uses a hybrid memory architecture with dual-granular storage and dynamic two-tier retrieval: 1) Lightweight module constructs summary-level context for efficient response generation, 2) LLM-based deep module selectively activated only for complex queries, augmented by a reflection mechanism for iterative reasoning refinement.

Result: HyMem achieves strong performance on both LOCOMO and LongMemEval benchmarks, outperforming full-context approaches while reducing computational cost by 92.6%, establishing state-of-the-art balance between efficiency and performance.

Conclusion: HyMem demonstrates that hybrid memory architectures with dynamic scheduling can effectively address the efficiency-effectiveness trade-off in long-term memory management for LLM agents, enabling better performance in extended dialogues.

Abstract: Large language model (LLM) agents demonstrate strong performance in short-text contexts but often underperform in extended dialogues due to inefficient memory management. Existing approaches face a fundamental trade-off between efficiency and effectiveness: memory compression risks losing critical details required for complex reasoning, while retaining raw text introduces unnecessary computational overhead for simple queries. The crux lies in the limitations of monolithic memory representations and static retrieval mechanisms, which fail to emulate the flexible and proactive memory scheduling capabilities observed in humans, thus struggling to adapt to diverse problem scenarios. Inspired by the principle of cognitive economy, we propose HyMem, a hybrid memory architecture that enables dynamic on-demand scheduling through multi-granular memory representations. HyMem adopts a dual-granular storage scheme paired with a dynamic two-tier retrieval system: a lightweight module constructs summary-level context for efficient response generation, while an LLM-based deep module is selectively activated only for complex queries, augmented by a reflection mechanism for iterative reasoning refinement. Experiments show that HyMem achieves strong performance on both the LOCOMO and LongMemEval benchmarks, outperforming full-context while reducing computational cost by 92.6%, establishing a state-of-the-art balance between efficiency and performance in long-term memory management.

Method: Two statistically principled early stopping approaches: 1) parametric method modeling inter-arrival times of uncertainty keywords as a renewal process with sequential testing, and 2) nonparametric method providing finite-sample guarantees on probability of halting too early on well-posed queries.

Result: Uncertainty-aware early stopping improves both efficiency and reliability in LLM reasoning across several domains and models, with especially significant gains for math reasoning tasks.

Conclusion: Monitoring uncertainty signals during LLM generation enables effective early stopping that reduces overthinking while maintaining reliability, particularly beneficial for reasoning tasks under uncertainty.

Abstract: While LLMs have seen substantial improvement in reasoning capabilities, they also sometimes overthink, generating unnecessary reasoning steps, particularly under uncertainty, given ill-posed or ambiguous queries. We introduce statistically principled early stopping methods that monitor uncertainty signals during generation to mitigate this issue. Our first approach is parametric: it models inter-arrival times of uncertainty keywords as a renewal process and applies sequential testing for stopping. Our second approach is nonparametric and provides finite-sample guarantees on the probability of halting too early on well-posed queries. We conduct empirical evaluations on reasoning tasks across several domains and models. Our results indicate that uncertainty-aware early stopping can improve both efficiency and reliability in LLM reasoning, and we observe especially significant gains for math reasoning.

Method: Proposes PCM with two components: 1) Disentangled Scene Encoder using intervention-based disentanglement to extract domain-invariant features, and 2) CausalODE Decoder using causal attention to integrate kinematic models with contextual information.

Result: Extensive experiments on real-world autonomous driving datasets show superior zero-shot generalization performance in unseen cities, significantly outperforming competitive baselines.

Conclusion: The Physics-guided Causal Model effectively addresses zero-shot generalization for trajectory prediction by incorporating domain-invariant knowledge and causal integration with kinematic models.

Abstract: Trajectory prediction for traffic agents is critical for safe autonomous driving. However, achieving effective zero-shot generalization in previously unseen domains remains a significant challenge. Motivated by the consistent nature of kinematics across diverse domains, we aim to incorporate domain-invariant knowledge to enhance zero-shot trajectory prediction capabilities. The key challenges include: 1) effectively extracting domain-invariant scene representations, and 2) integrating invariant features with kinematic models to enable generalized predictions. To address these challenges, we propose a novel generalizable Physics-guided Causal Model (PCM), which comprises two core components: a Disentangled Scene Encoder, which adopts intervention-based disentanglement to extract domain-invariant features from scenes, and a CausalODE Decoder, which employs a causal attention mechanism to effectively integrate kinematic models with meaningful contextual information. Extensive experiments on real-world autonomous driving datasets demonstrate our method’s superior zero-shot generalization performance in unseen cities, significantly outperforming competitive baselines. The source code is released at https://github.com/ZY-Zong/Physics-guided-Causal-Model.

Method: Developed Neuromem testbed with interleaved insertion-and-retrieval protocol that decomposes memory lifecycle into five dimensions: memory data structure, normalization strategy, consolidation policy, query formulation strategy, and context integration mechanism. Evaluated using LOCOMO, LONGMEMEVAL, and MEMORYAGENTBENCH datasets with interchangeable variants in a shared serving stack.

Result: Performance typically degrades as memory grows across rounds, with time-related queries being most challenging. Memory data structure determines quality frontier, while aggressive compression and generative integration mostly shift costs between insertion and retrieval with limited accuracy gains.

Conclusion: Neuromem provides a comprehensive framework for evaluating External Memory Modules under realistic streaming conditions, revealing important trade-offs and limitations in current approaches.

Abstract: Most evaluations of External Memory Module assume a static setting: memory is built offline and queried at a fixed state. In practice, memory is streaming: new facts arrive continuously, insertions interleave with retrievals, and the memory state evolves while the model is serving queries. In this regime, accuracy and cost are governed by the full memory lifecycle, which encompasses the ingestion, maintenance, retrieval, and integration of information into generation. We present Neuromem, a scalable testbed that benchmarks External Memory Modules under an interleaved insertion-and-retrieval protocol and decomposes its lifecycle into five dimensions including memory data structure, normalization strategy, consolidation policy, query formulation strategy, and context integration mechanism. Using three representative datasets LOCOMO, LONGMEMEVAL, and MEMORYAGENTBENCH, Neuromem evaluates interchangeable variants within a shared serving stack, reporting token-level F1 and insertion/retrieval latency. Overall, we observe that performance typically degrades as memory grows across rounds, and time-related queries remain the most challenging category. The memory data structure largely determines the attainable quality frontier, while aggressive compression and generative integration mechanisms mostly shift cost between insertion and retrieval with limited accuracy gain.

Method: Proposes Parallelized Iterative Compression (PIC) which modifies Transformer’s attention mask to restrict memory tokens’ receptive field to sequential local chunks, lowering compressor training difficulty. Uses chunking mechanism similar to human working memory.

Result: Consistently outperforms competitive baselines across multiple downstream tasks, especially in high compression scenarios (29.8% F1 and 40.7% EM improvements on QA tasks at 64× compression). Reduces training time by ~40% for 16× compressor while surpassing baseline peak performance.

Conclusion: PIC provides an effective context compression approach that reduces training difficulty and improves performance, particularly beneficial for high compression ratios. The method’s simplicity (attention mask modification) makes it practical for deployment.

Abstract: Providing extensive context via prompting is vital for leveraging the capabilities of Large Language Models (LLMs). However, lengthy contexts significantly increase inference latency, as the computational cost of self-attention grows quadratically with sequence length. To mitigate this issue, context compression-particularly soft prompt compressio-has emerged as a widely studied solution, which converts long contexts into shorter memory embeddings via a trained compressor. Existing methods typically compress the entire context indiscriminately into a set of memory tokens, requiring the compressor to capture global dependencies and necessitating extensive pre-training data to learn effective patterns. Inspired by the chunking mechanism in human working memory and empirical observations of the spatial specialization of memory embeddings relative to original tokens, we propose Parallelized Iterative Compression (PIC). By simply modifying the Transformer’s attention mask, PIC explicitly restricts the receptive field of memory tokens to sequential local chunks, thereby lowering the difficulty of compressor training. Experiments across multiple downstream tasks demonstrate that PIC consistently outperforms competitive baselines, with superiority being particularly pronounced in high compression scenarios (e.g., achieving relative improvements of 29.8% in F1 score and 40.7% in EM score on QA tasks at the $64\times$ compression ratio). Furthermore, PIC significantly expedites the training process. Specifically, when training the 16$\times$ compressor, it surpasses the peak performance of the competitive baseline while effectively reducing the training time by approximately 40%.

Method: Leverages formal abductive explanations that provide consistent, guaranteed reasoning over minimal sufficient feature sets to align AI decision-making with clinical reasoning frameworks.

Result: The approach preserves predictive accuracy while providing clinically actionable insights, establishing a robust framework for trustworthy AI in medical diagnosis.

Conclusion: Formal abductive explanations offer a solution to align AI reasoning with clinical frameworks, enhancing trust, interpretability, and adoption of AI in medical diagnostics.

Abstract: Artificial intelligence (AI) has demonstrated strong potential in clinical diagnostics, often achieving accuracy comparable to or exceeding that of human experts. A key challenge, however, is that AI reasoning frequently diverges from structured clinical frameworks, limiting trust, interpretability, and adoption. Critical symptoms, pivotal for rapid and accurate decision-making, may be overlooked by AI models even when predictions are correct. Existing post hoc explanation methods provide limited transparency and lack formal guarantees. To address this, we leverage formal abductive explanations, which offer consistent, guaranteed reasoning over minimal sufficient feature sets. This enables a clear understanding of AI decision-making and allows alignment with clinical reasoning. Our approach preserves predictive accuracy while providing clinically actionable insights, establishing a robust framework for trustworthy AI in medical diagnosis.

Method: P2AECF uses three mechanisms: 1) prompt-defined cognition that parses task intent into abstract model-agnostic representations, 2) agent-based modular execution using lightweight reusable cognitive agents dynamically selected based on resource conditions, and 3) diffusion-controlled inference planning that adaptively constructs and refines execution strategies with runtime feedback.

Result: The framework is illustrated through a low-altitude intelligent network use case, demonstrating its ability to deliver adaptive, modular, and scalable edge intelligence for real-time low-altitude aerial collaborations.

Conclusion: P2AECF enables flexible, efficient, and adaptive edge intelligence by addressing the fundamental limitations of deploying large AI models at the edge through prompt-to-agent transformation and dynamic resource-aware execution.

Abstract: The large artificial intelligence models (LAMs) show strong capabilities in perception, reasoning, and multi-modal understanding, and can enable advanced capabilities in low-altitude edge intelligence. However, the deployment of LAMs at the edge remains constrained by some fundamental limitations. First, tasks are rigidly tied to specific models, limiting the flexibility. Besides, the computational and memory demands of full-scale LAMs exceed the capacity of most edge devices. Moreover, the current inference pipelines are typically static, making it difficult to respond to real-time changes of tasks. To address these challenges, we propose a prompt-to-agent edge cognition framework (P2AECF), enabling the flexible, efficient, and adaptive edge intelligence. Specifically, P2AECF transforms high-level semantic prompts into executable reasoning workflows through three key mechanisms. First, the prompt-defined cognition parses task intent into abstract and model-agnostic representations. Second, the agent-based modular execution instantiates these tasks using lightweight and reusable cognitive agents dynamically selected based on current resource conditions. Third, the diffusion-controlled inference planning adaptively constructs and refines execution strategies by incorporating runtime feedback and system context. In addition, we illustrate the framework through a representative low-altitude intelligent network use case, showing its ability to deliver adaptive, modular, and scalable edge intelligence for real-time low-altitude aerial collaborations.

Method: FloCA uses a hybrid approach: LLM handles intent understanding and response generation, while an external tool performs topology-constrained graph execution for flowchart reasoning. This separation ensures faithful node transitions consistent with the correct flowchart path across dialogue turns.

Result: Extensive experiments on FLODIAL and PFDial datasets show FloCA’s superiority over existing LLM-based methods, highlighting bottlenecks in current approaches and demonstrating improved reasoning accuracy and interaction efficiency.

Conclusion: FloCA effectively addresses LLM limitations in flowchart reasoning by delegating topology-constrained reasoning to an external tool, ensuring faithful and logically consistent multi-turn dialogue while maintaining LLM strengths in natural language understanding and generation.

Abstract: Flowchart-oriented dialogue (FOD) systems aim to guide users through multi-turn decision-making or operational procedures by following a domain-specific flowchart to achieve a task goal. In this work, we formalize flowchart reasoning in FOD as grounding user input to flowchart nodes at each dialogue turn while ensuring node transition is consistent with the correct flowchart path. Despite recent advances of LLMs in task-oriented dialogue systems, adapting them to FOD still faces two limitations: (1) LLMs lack an explicit mechanism to represent and reason over flowchart topology, and (2) they are prone to hallucinations, leading to unfaithful flowchart reasoning. To address these limitations, we propose FloCA, a zero-shot flowchart-oriented conversational agent. FloCA uses an LLM for intent understanding and response generation while delegating flowchart reasoning to an external tool that performs topology-constrained graph execution, ensuring faithful and logically consistent node transitions across dialogue turns. We further introduce an evaluation framework with an LLM-based user simulator and five new metrics covering reasoning accuracy and interaction efficiency. Extensive experiments on FLODIAL and PFDial datasets highlight the bottlenecks of existing LLM-based methods and demonstrate the superiority of FloCA. Our codes are available at https://github.com/Jinzi-Zou/FloCA-flowchart-reasoning.

Method: Proposes FluxMem framework with: 1) Multiple complementary memory structures, 2) Learning to select among structures using offline supervision from downstream response quality and memory utilization, 3) Three-level memory hierarchy, 4) Beta Mixture Model-based probabilistic gate for distribution-aware memory fusion instead of similarity thresholds.

Result: Experiments on PERSONAMEM and LoCoMo benchmarks show average improvements of 9.18% and 6.14% respectively, demonstrating effectiveness for long-horizon tasks.

Conclusion: FluxMem provides a unified framework for adaptive memory organization in LLM agents that significantly improves performance on long-horizon benchmarks through context-aware memory structure selection and robust memory fusion.

Abstract: Memory is critical for enabling large language model (LLM) based agents to maintain coherent behavior over long-horizon interactions. However, existing agent memory systems suffer from two key gaps: they rely on a one-size-fits-all memory structure and do not model memory structure selection as a context-adaptive decision, limiting their ability to handle heterogeneous interaction patterns and resulting in suboptimal performance. We propose a unified framework, FluxMem, that enables adaptive memory organization for LLM agents. Our framework equips agents with multiple complementary memory structures. It explicitly learns to select among these structures based on interaction-level features, using offline supervision derived from downstream response quality and memory utilization. To support robust long-horizon memory evolution, we further introduce a three-level memory hierarchy and a Beta Mixture Model-based probabilistic gate for distribution-aware memory fusion, replacing brittle similarity thresholds. Experiments on two long-horizon benchmarks, PERSONAMEM and LoCoMo, demonstrate that our method achieves average improvements of 9.18% and 6.14%.

Method: Proposes REAL framework centered on reasoning-pivots (atomic units in reasoning chains that emphasize knowledge linkage). Uses Reasoning-Pivot Aware SFT (RPA-SFT) to train a generalizable discriminator by aligning conflicts with pivot extraction, and Reasoning-Pivot Guided Decoding (RPGD) for intra-model decoding that leverages pivots for targeted conflict mitigation.

Result: Extensive experiments across diverse benchmarks demonstrate that REAL significantly enhances discrimination accuracy and achieves state-of-the-art performance.

Conclusion: The REAL framework validates the effectiveness of the pivot-driven resolution paradigm for handling knowledge conflicts in visual question answering.

Abstract: Knowledge-intensive Visual Question Answering (KI-VQA) frequently suffers from severe knowledge conflicts caused by the inherent limitations of open-domain retrieval. However, existing paradigms face critical limitations due to the lack of generalizable conflict detection and intra-model constraint mechanisms to handle conflicting evidence. To address these challenges, we propose the REAL (Reasoning-Pivot Alignment) framework centered on the novel concept of the Reasoning-Pivot. Distinct from reasoning steps that prioritize internal self-derivation, a reasoning-pivot serves as an atomic unit (node or edge) in the reasoning chain that emphasizes knowledge linkage, and it typically relies on external evidence to complete the reasoning. Supported by our constructed REAL-VQA dataset, our approach integrates Reasoning-Pivot Aware SFT (RPA-SFT) to train a generalizable discriminator by aligning conflicts with pivot extraction, and employs Reasoning-Pivot Guided Decoding (RPGD), an intra-model decoding strategy that leverages these pivots for targeted conflict mitigation. Extensive experiments across diverse benchmarks demonstrate that REAL significantly enhances discrimination accuracy and achieves state-of-the-art performance, validating the effectiveness of our pivot-driven resolution paradigm.

Method: Plan-MCTS shifts exploration to a semantic Plan Space, decoupling strategic planning from execution grounding. It transforms sparse action space into a Dense Plan Tree for efficient exploration, distills noisy contexts into Abstracted Semantic History for state awareness, uses Dual-Gating Reward to validate executability and strategic alignment, and employs Structural Refinement for on-policy repair of failed subplans.

Result: Extensive experiments on WebArena demonstrate state-of-the-art performance, surpassing current approaches with higher task effectiveness and search efficiency.

Conclusion: Plan-MCTS effectively addresses sparse path and noisy context challenges in web navigation by leveraging semantic planning space, achieving superior performance through efficient exploration and precise state awareness.

Abstract: Large Language Models (LLMs) have empowered autonomous agents to handle complex web navigation tasks. While recent studies integrate tree search to enhance long-horizon reasoning, applying these algorithms in web navigation faces two critical challenges: sparse valid paths that lead to inefficient exploration, and a noisy context that dilutes accurate state perception. To address this, we introduce Plan-MCTS, a framework that reformulates web navigation by shifting exploration to a semantic Plan Space. By decoupling strategic planning from execution grounding, it transforms sparse action space into a Dense Plan Tree for efficient exploration, and distills noisy contexts into an Abstracted Semantic History for precise state awareness. To ensure efficiency and robustness, Plan-MCTS incorporates a Dual-Gating Reward to strictly validate both physical executability and strategic alignment and Structural Refinement for on-policy repair of failed subplans. Extensive experiments on WebArena demonstrate that Plan-MCTS achieves state-of-the-art performance, surpassing current approaches with higher task effectiveness and search efficiency.

Method: Reconstructs real-world applications into lightweight web environments using multimodal code models, equipping them with code-native rewards (executable assertions) that provide deterministic reward signals and eliminate visual estimation noise.

Result: Reduces environment latency by 10 times and costs by over $28,000 per epoch compared to training on real applications. Agents trained with GUI-GENESIS outperform base model by 14.54% and real-world RL baselines by 3.27% on held-out real-world tasks.

Conclusion: GUI-GENESIS enables efficient GUI agent training with verifiable rewards, demonstrating that models can synthesize environments they cannot yet solve, highlighting a pathway for self-improving agents.

Abstract: Post-training GUI agents in interactive environments is critical for developing generalization and long-horizon planning capabilities. However, training on real-world applications is hindered by high latency, poor reproducibility, and unverifiable rewards relying on noisy visual proxies. To address the limitations, we present GUI-GENESIS, the first framework to automatically synthesize efficient GUI training environments with verifiable rewards. GUI-GENESIS reconstructs real-world applications into lightweight web environments using multimodal code models and equips them with code-native rewards, executable assertions that provide deterministic reward signals and eliminate visual estimation noise. Extensive experiments show that GUI-GENESIS reduces environment latency by 10 times and costs by over $28,000 per epoch compared to training on real applications. Notably, agents trained with GUI-GENESIS outperform the base model by 14.54% and even real-world RL baselines by 3.27% on held-out real-world tasks. Finally, we observe that models can synthesize environments they cannot yet solve, highlighting a pathway for self-improving agents.

Method: Systematic evaluation across 28 models (past to current frontier) measuring monitor evasion, refusal rates, encoding fidelity, and hidden task accuracy across four datasets, comparing steganographic acrostics against plain reasoning and filler-token baselines.

Result: Current models cannot sustain hidden reasoning for complex math/arithmetic tasks, but Claude Opus 4.5 achieved 92% accuracy on simplified counting tasks. GPT-5.2 showed rare (<1%) paradoxical refusal/compliance behavior.

Conclusion: Steganographic reasoning poses emerging risks requiring continuous evaluation; the study provides methodology for preemptive detection of hidden reasoning that could enable misaligned scheming.

Abstract: Monitoring chain-of-thought (CoT) reasoning is a foundational safety technique for large language model (LLM) agents; however, this oversight is compromised if models learn to conceal their reasoning. We explore the potential for steganographic CoT – where models hide secret reasoning within innocuous text – to inform risk assessment and deployment policies. We systematically evaluate the limits of steganographic capabilities across 28 models, ranging from past generations to the current frontier. We measure monitor evasion, refusal rates, encoding fidelity, and hidden task accuracy across four datasets, comparing steganographic acrostics against plain reasoning and filler-token baselines. We find that current models cannot yet sustain hidden reasoning for complex math and arithmetic tasks. However, in a simplified counting experiment, Claude Opus 4.5 achieved 92% accuracy on the hidden task, demonstrating nascent capability. Notably, in rare cases (<1%), GPT-5.2 might refuse steganographic instructions while simultaneously complying with them. Our findings underscore the need for continuous evaluation of steganographic risks. This study provides a methodology to preemptively detect and prevent hidden reasoning that might empower misaligned scheming and deceptive behavior.

Method: Proposed Algebraic Quantum Intelligence (AQI) as a computational framework using noncommutative algebraic structures inspired by quantum theory. Implemented by extending a transformer-based LLM with 600+ specialized operators, representing semantic states as Hilbert space vectors and evolving them using C-values from noncommutative operators.

Result: AQI consistently outperforms strong baseline models on creative reasoning benchmarks across ten domains, showing statistically significant improvements and reduced cross-domain variance under LLM-as-a-judge protocol.

Conclusion: Noncommutative algebraic dynamics provide a practical and reproducible foundation for machine creativity, with the architecture already deployed in real-world enterprise environments.

Abstract: Large language models (LLMs) have achieved remarkable success in generating fluent and contextually appropriate text; however, their capacity to produce genuinely creative outputs remains limited. This paper posits that this limitation arises from a structural property of contemporary LLMs: when provided with rich context, the space of future generations becomes strongly constrained, and the generation process is effectively governed by near-deterministic dynamics. Recent approaches such as test-time scaling and context adaptation improve performance but do not fundamentally alter this constraint. To address this issue, we propose Algebraic Quantum Intelligence (AQI) as a computational framework that enables systematic expansion of semantic space. AQI is formulated as a noncommutative algebraic structure inspired by quantum theory, allowing properties such as order dependence, interference, and uncertainty to be implemented in a controlled and designable manner. Semantic states are represented as vectors in a Hilbert space, and their evolution is governed by C-values computed from noncommutative operators, thereby ensuring the coexistence and expansion of multiple future semantic possibilities. In this study, we implement AQI by extending a transformer-based LLM with more than 600 specialized operators. We evaluate the resulting system on creative reasoning benchmarks spanning ten domains under an LLM-as-a-judge protocol. The results show that AQI consistently outperforms strong baseline models, yielding statistically significant improvements and reduced cross-domain variance. These findings demonstrate that noncommutative algebraic dynamics can serve as a practical and reproducible foundation for machine creativity. Notably, this architecture has already been deployed in real-world enterprise environments.

Method: Proposes a hierarchical framework starting with 7 Fundamental Safety pillars, extending to advanced Embodied AI Safety, AI4Science Safety, Social/Environmental AI risks, Catastrophic/Existential Risks, and 8 industrial safety domains totaling 94 refined risk dimensions. Accumulates tens of thousands of structured risk data points and assessment results.

Result: Systematic evaluation of over twenty mainstream advanced large models reveals widespread safety vulnerabilities across multiple pillars, particularly in Risky Agentic Autonomy, AI4Science Safety, Embodied AI Safety, Social AI Safety and Catastrophic/Existential Risks. Identifies key risk patterns and capability boundaries.

Conclusion: The ForesightSafety Bench establishes a comprehensive, hierarchical, and dynamically evolving AI safety evaluation framework that addresses limitations of current systems and provides systematic assessment of frontier AI risks across multiple dimensions.

Abstract: Rapidly evolving AI exhibits increasingly strong autonomy and goal-directed capabilities, accompanied by derivative systemic risks that are more unpredictable, difficult to control, and potentially irreversible. However, current AI safety evaluation systems suffer from critical limitations such as restricted risk dimensions and failed frontier risk detection. The lagging safety benchmarks and alignment technologies can hardly address the complex challenges posed by cutting-edge AI models. To bridge this gap, we propose the “ForesightSafety Bench” AI Safety Evaluation Framework, beginning with 7 major Fundamental Safety pillars and progressively extends to advanced Embodied AI Safety, AI4Science Safety, Social and Environmental AI risks, Catastrophic and Existential Risks, as well as 8 critical industrial safety domains, forming a total of 94 refined risk dimensions. To date, the benchmark has accumulated tens of thousands of structured risk data points and assessment results, establishing a widely encompassing, hierarchically clear, and dynamically evolving AI safety evaluation framework. Based on this benchmark, we conduct systematic evaluation and in-depth analysis of over twenty mainstream advanced large models, identifying key risk patterns and their capability boundaries. The safety capability evaluation results reveals the widespread safety vulnerabilities of frontier AI across multiple pillars, particularly focusing on Risky Agentic Autonomy, AI4Science Safety, Embodied AI Safety, Social AI Safety and Catastrophic and Existential Risks. Our benchmark is released at https://github.com/Beijing-AISI/ForesightSafety-Bench. The project website is available at https://foresightsafety-bench.beijing-aisi.ac.cn/.

Method: Agent-as-tool reinforcement learning framework with two objectives: (1) process-level supervision to ensure reasoning follows valid clinical pathways, and (2) efficient coordination via hierarchical multi-agent system; uses GRPO-trained supervisor agent (Qwen3-4B) with different reward strategies (outcome-only vs process+outcome)

Result: On ClinGen dataset: with outcome-only rewards, MAS with GRPO-trained supervisor improved final outcome accuracy from 0.195 to 0.732 but had poor process alignment (0.392 F1); with process+outcome rewards, achieved higher outcome accuracy (0.750) while significantly improving process fidelity to 0.520 F1

Conclusion: The framework successfully balances outcome accuracy with clinical reasoning process alignment in gene-disease validity curation, demonstrating that process-level supervision is crucial for developing clinically-reliable multi-agent systems

Abstract: Clinical decision-making requires nuanced reasoning over heterogeneous evidence and traceable justifications. While recent LLM multi-agent systems (MAS) show promise, they largely optimise for outcome accuracy while overlooking process-grounded reasoning aligned with clinical standards. One critical real-world case of this is gene-disease validity curation, where experts must determine whether a gene is causally implicated in a disease by synthesising diverse biomedical evidence. We introduce an agent-as-tool reinforcement learning framework for this task with two objectives: (i) process-level supervision to ensure reasoning follows valid clinical pathways, and (ii) efficient coordination via a hierarchical multi-agent system. Our evaluation on the ClinGen dataset shows that with outcome-only rewards, MAS with a GRPO-trained Qwen3-4B supervisor agent substantially improves final outcome accuracy from 0.195 with a base model supervisor to 0.732, but results in poor process alignment (0.392 F1). Conversely, with process + outcome rewards, MAS with GRPO-trained supervisor achieves higher outcome accuracy (0.750) while significantly improving process fidelity to 0.520 F1. Our code is available at https://github.com/chaeeunlee-io/GeneDiseaseCurationAgents.

Method: Proposes a staged knowledge injection approach: (1) cold-starting with scalable, knowledge-graph-verified Earth-science text QA to instill reasoning structures, and (2) “pre-warming” on hard UHR image-text examples during supervised fine-tuning to stabilize subsequent tool-based reinforcement learning.

Result: Achieves 60.40% Pass@1 on XLRS-Bench, significantly outperforming larger general purpose models like GPT-5.2, Gemini 3.0 Pro, and Intern-S1, establishing a new state-of-the-art.

Conclusion: High-quality domain-specific text QA is a primary driver of UHR visual reasoning gains, as it injects the concepts, mechanistic explanations, and decision rules necessary to guide visual evidence retrieval, even without images.

Abstract: Multimodal reasoning for ultra-high-resolution (UHR) remote sensing (RS) is usually bottlenecked by visual evidence acquisition: the model necessitates localizing tiny task-relevant regions in massive pixel spaces. While Agentic Reinforcement Learning with Verifiable Rewards (RLVR) using zoom-in tools offers a path forward, we find that standard reinforcement learning struggles to navigate these vast visual spaces without structured domain priors. In this paper, we investigate the interplay between post-training paradigms: comparing Cold-start Supervised Fine-Tuning (SFT), RLVR, and Agentic RLVR on the UHR RS benchmark.Our controlled studies yield a counter-intuitive finding: high-quality Earth-science text-only QA is a primary driver of UHR visual reasoning gains. Despite lacking images, domain-specific text injects the concepts, mechanistic explanations, and decision rules necessary to guide visual evidence retrieval.Based on this, we propose a staged knowledge injection recipe: (1) cold-starting with scalable, knowledge-graph-verified Earth-science text QA to instill reasoning structures;and (2) “pre-warming” on the same hard UHR image-text examples during SFT to stabilize and amplify subsequent tool-based RL. This approach achieves a 60.40% Pass@1 on XLRS-Bench, significantly outperforming larger general purpose models (e.g., GPT-5.2, Gemini 3.0 Pro, Intern-S1) and establishing a new state-of-the-art.

Method: CorpGen framework with hierarchical planning for multi-horizon goal alignment, sub-agent isolation to prevent cross-task contamination, tiered memory (working, structured, semantic), and adaptive summarization, tested in simulated corporate environments.

Result: Achieves up to 3.5x improvement over baselines (15.2% vs 4.3%) with stable performance under increasing load, confirming gains stem from architectural mechanisms rather than specific CUA implementations.

Conclusion: CorpGen effectively addresses key failure modes in multi-horizon task environments and demonstrates architectural solutions for managing concurrent long-horizon tasks in organizational settings.

Abstract: Long-horizon reasoning is a key challenge for autonomous agents, yet existing benchmarks evaluate agents on single tasks in isolation. Real organizational work requires managing many concurrent long-horizon tasks with interleaving, dependencies, and reprioritization. We introduce Multi-Horizon Task Environments (MHTEs): a distinct problem class requiring coherent execution across dozens of interleaved tasks (45+, 500-1500+ steps) within persistent execution contexts spanning hours. We identify four failure modes that cause baseline CUAs to degrade from 16.7% to 8.7% completion as load scales 25% to 100%, a pattern consistent across three independent implementations. These failure modes are context saturation (O(N) vs O(1) growth), memory interference, dependency complexity (DAGs vs. chains), and reprioritization overhead. We present CorpGen, an architecture-agnostic framework addressing these failures via hierarchical planning for multi-horizon goal alignment, sub-agent isolation preventing cross-task contamination, tiered memory (working, structured, semantic), and adaptive summarization. CorpGen simulates corporate environments through digital employees with persistent identities and realistic schedules. Across three CUA backends (UFO2, OpenAI CUA, hierarchical) on OSWorld Office, CorpGen achieves up to 3.5x improvement over baselines (15.2% vs 4.3%) with stable performance under increasing load, confirming that gains stem from architectural mechanisms rather than specific CUA implementations. Ablation studies show experiential learning provides the largest gains.

Method: Proposes REDSearcher framework with four key improvements: 1) Task synthesis as dual-constrained optimization using graph topology and evidence dispersion, 2) Tool-augmented queries to encourage proactive tool use, 3) Mid-training to strengthen core atomic capabilities (knowledge, planning, function calling), 4) Local simulated environment for low-cost RL iteration.

Result: Achieves state-of-the-art performance across both text-only and multimodal search-agent benchmarks. Releases 10K text search trajectories, 5K multimodal trajectories, 1K text RL query set, plus code and model checkpoints.

Conclusion: REDSearcher provides a scalable framework for optimizing search agents, addressing key bottlenecks in trajectory sparsity and interaction costs, with demonstrated effectiveness in both text and multimodal search tasks.

Abstract: Large language models are transitioning from generalpurpose knowledge engines to realworld problem solvers, yet optimizing them for deep search tasks remains challenging. The central bottleneck lies in the extreme sparsity of highquality search trajectories and reward signals, arising from the difficulty of scalable longhorizon task construction and the high cost of interactionheavy rollouts involving external tool calls. To address these challenges, we propose REDSearcher, a unified framework that codesigns complex task synthesis, midtraining, and posttraining for scalable searchagent optimization. Specifically, REDSearcher introduces the following improvements: (1) We frame task synthesis as a dualconstrained optimization, where task difficulty is precisely governed by graph topology and evidence dispersion, allowing scalable generation of complex, highquality tasks. (2) We introduce toolaugmented queries to encourage proactive tool use rather than passive recall.(3) During midtraining, we strengthen core atomic capabilities knowledge, planning, and function calling substantially reducing the cost of collecting highquality trajectories for downstream training. (4) We build a local simulated environment that enables rapid, lowcost algorithmic iteration for reinforcement learning experiments. Across both textonly and multimodal searchagent benchmarks, our approach achieves stateoftheart performance. To facilitate future research on longhorizon search agents, we will release 10K highquality complex text search trajectories, 5K multimodal trajectories and 1K text RL query set, and together with code and model checkpoints.

Method: GRAIL combines imitation learning and inverse reinforcement learning to learn one goal-directed policy for each candidate goal directly from demonstration trajectories. It scores observed partial trajectories with each learned policy in a single forward pass.

Result: GRAIL improves F1-scores by >0.5 under systematically biased optimal behavior, gains 0.1-0.3 under suboptimal behavior, improves up to 0.4 under noisy optimal trajectories, while remaining competitive in fully optimal settings.

Conclusion: GRAIL contributes to scalable and robust models for interpreting agent goals in uncertain environments by better handling suboptimal and biased behavior.

Abstract: Understanding an agent’s goals from its behavior is fundamental to aligning AI systems with human intentions. Existing goal recognition methods typically rely on an optimal goal-oriented policy representation, which may differ from the actor’s true behavior and hinder the accurate recognition of their goal. To address this gap, this paper introduces Goal Recognition Alignment through Imitation Learning (GRAIL), which leverages imitation learning and inverse reinforcement learning to learn one goal-directed policy for each candidate goal directly from (potentially suboptimal) demonstration trajectories. By scoring an observed partial trajectory with each learned goal-directed policy in a single forward pass, GRAIL retains the one-shot inference capability of classical goal recognition while leveraging learned policies that can capture suboptimal and systematically biased behavior. Across the evaluated domains, GRAIL increases the F1-score by more than 0.5 under systematically biased optimal behavior, achieves gains of approximately 0.1-0.3 under suboptimal behavior, and yields improvements of up to 0.4 under noisy optimal trajectories, while remaining competitive in fully optimal settings. This work contributes toward scalable and robust models for interpreting agent goals in uncertain environments.

Method: Model web environments as Finite State Machines (FSMs) and use coding agents to translate FSMs into interactive websites, enabling programmatic verification of actions and task success through explicit state transition rules.

Result: Generated 11,663 verified trajectories from 29 diverse web environments at $0.04 per trajectory. Training on this data significantly boosts real-world performance: 7B Web GUI agent outperforms all baselines on WebVoyager within 15 steps, with clear scaling law showing improved performance as synthetic data volume increases.

Conclusion: AutoWebWorld provides an efficient framework for generating high-quality, verifiable web interaction data that enables better training of Web GUI agents, demonstrating strong performance improvements and scalability on real-world benchmarks.

Abstract: The performance of autonomous Web GUI agents heavily relies on the quality and quantity of their training data. However, a fundamental bottleneck persists: collecting interaction trajectories from real-world websites is expensive and difficult to verify. The underlying state transitions are hidden, leading to reliance on inconsistent and costly external verifiers to evaluate step-level correctness. To address this, we propose AutoWebWorld, a novel framework for synthesizing controllable and verifiable web environments by modeling them as Finite State Machines (FSMs) and use coding agents to translate FSMs into interactive websites. Unlike real websites, where state transitions are implicit, AutoWebWorld explicitly defines all states, actions, and transition rules. This enables programmatic verification: action correctness is checked against predefined rules, and task success is confirmed by reaching a goal state in the FSM graph. AutoWebWorld enables a fully automated search-and-verify pipeline, generating over 11,663 verified trajectories from 29 diverse web environments at only $0.04 per trajectory. Training on this synthetic data significantly boosts real-world performance. Our 7B Web GUI agent outperforms all baselines within 15 steps on WebVoyager. Furthermore, we observe a clear scaling law: as the synthetic data volume increases, performance on WebVoyager and Online-Mind2Web consistently improves.

Method: Uses critique-resilient correctness where answers are deemed correct if no adversary can convincingly prove otherwise. Humans serve as bounded verifiers focusing on localized claims. Employs itemized bipartite Bradley-Terry model to jointly rank LLMs by their ability to solve tasks and generate difficult yet solvable questions.

Result: Demonstrated effectiveness in mathematical domain across eight frontier LLMs, showing stable scores that correlate with external capability measures.

Conclusion: Reformulates benchmarking as an adversarial generation-evaluation game with humans as final adjudicators, providing a scalable approach for comparing models beyond human comprehension limits.

Abstract: As frontier Large Language Models (LLMs) increasingly saturate new benchmarks shortly after they are published, benchmarking itself is at a juncture: if frontier models keep improving, it will become increasingly hard for humans to generate discriminative tasks, provide accurate ground-truth answers, or evaluate complex solutions. If benchmarking becomes infeasible, our ability to measure any progress in AI is at stake. We refer to this scenario as the post-comprehension regime. In this work, we propose Critique-Resilient Benchmarking, an adversarial framework designed to compare models even when full human understanding is infeasible. Our technique relies on the notion of critique-resilient correctness: an answer is deemed correct if no adversary has convincingly proved otherwise. Unlike standard benchmarking, humans serve as bounded verifiers and focus on localized claims, which preserves evaluation integrity beyond full comprehension of the task. Using an itemized bipartite Bradley-Terry model, we jointly rank LLMs by their ability to solve challenging tasks and to generate difficult yet solvable questions. We showcase the effectiveness of our method in the mathematical domain across eight frontier LLMs, showing that the resulting scores are stable and correlate with external capability measures. Our framework reformulates benchmarking as an adversarial generation-evaluation game in which humans serve as final adjudicators.

Method: Develops mathematical formula for dynamical tipping point n* based on dot-product competition for attention between conversation context and competing output basins, validated across multiple AI models.

Result: Identifies atomistic-scale mechanism for dangerous behavior in edge AI, provides mathematical framework for predicting tipping points, and demonstrates applicability across domains, legal landscapes, languages, and cultures.

Conclusion: The attention competition mechanism offers new safety control levers for edge AI devices, addressing critical gaps in current safety tools that require cloud connectivity or detect failures only after harm occurs.

Abstract: More than half the global population now carries devices that can run ChatGPT-like language models with no Internet connection and minimal safety oversight – and hence the potential to promote self-harm, financial losses and extremism among other dangers. Existing safety tools either require cloud connectivity or discover failures only after harm has occurred. Here we show that a large class of potentially dangerous tipping originates at the atomistic scale in such edge AI due to competition for the machinery’s attention. This yields a mathematical formula for the dynamical tipping point n*, governed by dot-product competition for attention between the conversation’s context and competing output basins, that reveals new control levers. Validated against multiple AI models, the mechanism can be instantiated for different definitions of ‘good’ and ‘bad’ and hence in principle applies across domains (e.g. health, law, finance, defense), changing legal landscapes (e.g. EU, UK, US and state level), languages, and cultural settings.

Method: Introduced PITA dataset with 23M+ propositional logic statements and proofs. Proposed task depth (steps required) and task breadth (unique examples) metrics. Varied these across PITA subsets and compared RT vs non-RT models. Also compared with synthetic syllogism task to validate generalizability of findings.

Result: RT models generalize well on broad and shallow subsets but deteriorate on narrow and deep subsets relative to non-RT baselines. The synthetic syllogism task confirmed similar patterns, suggesting these are general phenomena rather than PITA-specific.

Conclusion: The study reveals fundamental scaling limits of RT models on deep tasks while highlighting their generalization strengths on broad tasks. Identifies inherent benefits and limitations of reasoning trace approaches for reasoning models.

Abstract: Recent years have witnessed meteoric progress in reasoning models: neural networks that generate intermediate reasoning traces (RTs) before producing a final output. Despite the rapid advancement, our understanding of how RTs support reasoning, and the limits of this paradigm, remain incomplete. To promote greater clarity, we introduce PITA: a novel large-scale dataset of over 23 million statements in propositional logic and their corresponding proofs. As a benchmark for robust reasoning, we focus on length generalization: if a model is trained to determine truth or falsity on statements with proofs up to fixed length, how well does it generalize to statements requiring longer proofs? We propose notions of (1) task depth and (2) task breadth, which measure respectively (1) the number of steps required to solve an example from a task and (2) the number of unique examples across a task. We vary these quantities across subsets of PITA, and find that RT models generalize well on broad and shallow subsets, while deteriorating on narrow and deep subsets relative to non-RT baselines. To determine whether our results are idiosyncratic to PITA or indicative of general phenomena, we compare our results to a simple synthetic task based on syllogisms. Our resulting theory suggests fundamental scalings that limit how well RT models perform on deep tasks, and highlights their generalization strengths on broad tasks. Our findings overall identify fundamental benefits and limitations inherent in using reasoning traces.

Method: Proposes Precedent Informed Reasoning (PIR) with two components: 1) Adaptive Precedent Selection (APS) constructs compact precedent sets using joint scoring of semantic similarity and model perplexity, adapting the amount to maximize perplexity reduction; 2) Test-time Experience Internalization (TEI) performs test-time learning on precedent-informed instructions, updating lightweight adapters to internalize solution patterns as priors.

Result: Experiments across mathematical reasoning, scientific QA, and code generation demonstrate that PIR consistently shortens reasoning traces while maintaining or improving final accuracy across LLMs, yielding outstanding accuracy-efficiency trade-offs.

Conclusion: PIR transforms LLM reasoning from exhaustive self-exploration to guided learning from precedents, achieving better efficiency without sacrificing accuracy through adaptive precedent selection and test-time experience internalization.

Abstract: Reasoning in Large Language Models (LLMs) often suffers from inefficient long chain-of-thought traces with redundant self-exploration and validation, which inflate computational costs and even degrade performance. Inspired by human reasoning patterns where people solve new problems by leveraging past related cases to constrain search spaces and reduce trial-and-error, we propose Precedent Informed Reasoning (PIR) transforming LRMs’reasoning paradigm from exhaustive self-exploration to guided learning from precedents. PIR addresses two key challenges: what precedents to adopt and how to utilize them. First, Adaptive Precedent Selection (APS) constructs, for each question and LRM, a compact set of precedents that are both semantically related and informative for the model. It ranks examples by a joint score with semantic similarity and model perplexity, then adapts the amount of precedents to maximize perplexity reduction. Second, Test-time Experience Internalization (TEI) is treated as the test-time learning on precedent-informed instruction, updating lightweight adapters to internalize solution patterns and use them as a prior during subsequent reasoning. Experiments across mathematical reasoning, scientific QA, and code generation demonstrate that PIR consistently shortens reasoning traces while maintaining or improving final accuracy across LLMs, yielding outstanding accuracy-efficiency trade-offs.

Method: MOFh6 uses large language models to read raw articles or crystal codes and convert them into standardized synthesis tables. It links related descriptions across paragraphs, unifies ligand abbreviations with full names, and outputs structured parameters.

Result: Achieved 99% extraction accuracy, resolved 94.1% of abbreviation cases across five major publishers, maintained precision of 0.93 +/- 0.01. Processing times: 9.6s for full text, 36s for locating synthesis descriptions. Cost: $4.24 for 100 papers.

Conclusion: MOFh6 reshapes MOF synthesis research by replacing static database lookups with real-time extraction, accelerating literature knowledge conversion into practical synthesis protocols and enabling scalable, data-driven materials discovery.

Abstract: Accurately identifying the synthesis conditions of metal-organic frameworks (MOFs) is essential for guiding experimental design, yet remains challenging because relevant information in the literature is often scattered, inconsistent, and difficult to interpret. We present MOFh6, a large language model driven system that reads raw articles or crystal codes and converts them into standardized synthesis tables. It links related descriptions across paragraphs, unifies ligand abbreviations with full names, and outputs structured parameters ready for use. MOFh6 achieved 99% extraction accuracy, resolved 94.1% of abbreviation cases across five major publishers, and maintained a precision of 0.93 +/- 0.01. Processing a full text takes 9.6 s, locating synthesis descriptions 36 s, with 100 papers processed for USD 4.24. By replacing static database lookups with real-time extraction, MOFh6 reshapes MOF synthesis research, accelerating the conversion of literature knowledge into practical synthesis protocols and enabling scalable, data-driven materials discovery.

Method: Presents a comprehensive risk management framework with updated assessments across five dimensions: introduces complex cyber offense scenarios, evaluates LLM-to-LLM persuasion on new models, adds emergent misalignment experiments, focuses on agent “mis-evolution” in uncontrolled R&D, monitors safety performance on Moltbook, and introduces resource-constrained self-replication scenarios.

Result: The framework provides detailed risk assessments across all five dimensions and proposes/validates robust mitigation strategies, offering a preliminary technical pathway for secure frontier AI deployment.

Conclusion: This work reflects current understanding of AI frontier risks and urges collective action to mitigate these challenges through the proposed comprehensive risk management framework.

Abstract: To understand and identify the unprecedented risks posed by rapidly advancing artificial intelligence (AI) models, Frontier AI Risk Management Framework in Practice presents a comprehensive assessment of their frontier risks. As Large Language Models (LLMs) general capabilities rapidly evolve and the proliferation of agentic AI, this version of the risk analysis technical report presents an updated and granular assessment of five critical dimensions: cyber offense, persuasion and manipulation, strategic deception, uncontrolled AI R&D, and self-replication. Specifically, we introduce more complex scenarios for cyber offense. For persuasion and manipulation, we evaluate the risk of LLM-to-LLM persuasion on newly released LLMs. For strategic deception and scheming, we add the new experiment with respect to emergent misalignment. For uncontrolled AI R&D, we focus on the ``mis-evolution’’ of agents as they autonomously expand their memory substrates and toolsets. Besides, we also monitor and evaluate the safety performance of OpenClaw during the interaction on the Moltbook. For self-replication, we introduce a new resource-constrained scenario. More importantly, we propose and validate a series of robust mitigation strategies to address these emerging threats, providing a preliminary technical and actionable pathway for the secure deployment of frontier AI. This work reflects our current understanding of AI frontier risks and urges collective action to mitigate these challenges.

Method: Proposes optimization programming formulation that systematically incorporates available structural or statistical information as constraints to derive tighter bounds without requiring full identifiability

Result: Framework extends applicability of probabilities of causation to realistic settings where causal knowledge is incomplete but informative, yielding formally valid bounds

Conclusion: Provides practical approach for bounding probabilities of causation using partial causal information, overcoming limitations of existing methods

Abstract: Probabilities of causation are fundamental to individual-level explanation and decision making, yet they are inherently counterfactual and not point-identifiable from data in general. Existing bounds either disregard available covariates, require complete causal graphs, or rely on restrictive binary settings, limiting their practical use. In real-world applications, causal information is often partial but nontrivial. This paper proposes a general framework for bounding probabilities of causation using partial causal information. We show how the available structural or statistical information can be systematically incorporated as constraints in a optimization programming formulation, yielding tighter and formally valid bounds without full identifiability. This approach extends the applicability of probabilities of causation to realistic settings where causal knowledge is incomplete but informative.

Method: COOL-MC wraps the Storm model checker with three key capabilities: 1) constructs only reachable state space induced by trained policies, 2) automatically labels states with clinically meaningful atomic propositions, and 3) integrates explainability methods with PCTL queries to reveal feature importance.

Result: Applied to ICU-Sepsis MDP (17,000 patient records), COOL-MC established hard bounds via full MDP verification, trained a safe RL policy achieving optimal survival probability, and revealed that the policy relies predominantly on prior dosing history rather than patient’s evolving condition.

Conclusion: COOL-MC serves as a valuable tool for clinicians to investigate and debug sepsis treatment policies before deployment by combining formal verification with explainability to expose weaknesses invisible to standard evaluation methods.

Abstract: Safe and interpretable sequential decision-making is critical in healthcare, yet reinforcement learning (RL) policies for sepsis treatment optimization remain opaque and difficult to verify. Standard probabilistic model checkers operate on the full state space, which becomes infeasible for larger MDPs, and cannot explain why a learned policy makes particular decisions. COOL-MC wraps the model checker Storm but adds three key capabilities: it constructs only the reachable state space induced by a trained policy, yielding a smaller discrete-time Markov chain amenable to verification even when full-MDP analysis is intractable; it automatically labels states with clinically meaningful atomic propositions; and it integrates explainability methods with probabilistic computation tree logic (PCTL) queries to reveal which features drive decisions across treatment trajectories. We demonstrate COOL-MC’s capabilities on the ICU-Sepsis MDP, a benchmark derived from approximately 17,000 sepsis patient records, which serves as a case study for applying COOL-MC to the formal analysis of sepsis treatment policies. Our analysis establishes hard bounds via full MDP verification, trains a safe RL policy that achieves optimal survival probability, and analyzes its behavior via PCTL verification and explainability on the induced DTMC. This reveals, for instance, that our trained policy relies predominantly on prior dosing history rather than the patient’s evolving condition, a weakness that is invisible to standard evaluation but is exposed by COOL-MC’s integration of formal verification and explainability. Our results illustrate how COOL-MC could serve as a tool for clinicians to investigate and debug sepsis treatment policies before deployment.

Method: Formalizes knowledge conflict into input-level objective vs process-level effective conflict, probes internal representations to analyze conflict encoding patterns, and examines linear separability, depth localization, hierarchical consistency, and directional asymmetry.

Result: Found that: (1) conflict types are linearly separable features; (2) conflict signals concentrate in mid-to-late layers; (3) aggregating token-level signals recovers input-level conflict types; (4) reinforcing implicit source preference is easier than enforcing opposite source.

Conclusion: Provides mechanism-level understanding of multimodal reasoning under knowledge conflict, enabling principled diagnosis and control of long chain-of-thought failures in MLLMs.

Abstract: Multimodal large language models (MLLMs) in long chain-of-thought reasoning often fail when different knowledge sources provide conflicting signals. We formalize these failures under a unified notion of knowledge conflict, distinguishing input-level objective conflict from process-level effective conflict. Through probing internal representations, we reveal that: (I) Linear Separability: different conflict types are explicitly encoded as linearly separable features rather than entangled; (II) Depth Localization: conflict signals concentrate in mid-to-late layers, indicating a distinct processing stage for conflict encoding; (III) Hierarchical Consistency: aggregating noisy token-level signals along trajectories robustly recovers input-level conflict types; and (IV) Directional Asymmetry: reinforcing the model’s implicit source preference under conflict is far easier than enforcing the opposite source. Our findings provide a mechanism-level view of multimodal reasoning under knowledge conflict and enable principled diagnosis and control of long-CoT failures.

Method: Constructed controlled environment for entity-centric factual questions where knowledge is preserved while behavioral expression is selectively altered. Analyzed four behavioral cases through representation separability, sparse interpretability, and inference-time activation steering.

Result: The study demonstrates that hallucination and deception correspond to two qualitatively different failure modes that may appear similar at the output level but differ in their underlying mechanisms.

Conclusion: A mechanism-oriented perspective separating knowledge existence from behavior expression provides better understanding of LLM failures, distinguishing hallucination from deception as fundamentally different failure modes.

Abstract: Failures in large language models (LLMs) are often analyzed from a behavioral perspective, where incorrect outputs in factual question answering are commonly associated with missing knowledge. In this work, focusing on entity-based factual queries, we suggest that such a view may conflate different failure mechanisms, and propose an internal, mechanism-oriented perspective that separates Knowledge Existence from Behavior Expression. Under this formulation, hallucination and deception correspond to two qualitatively different failure modes that may appear similar at the output level but differ in their underlying mechanisms. To study this distinction, we construct a controlled environment for entity-centric factual questions in which knowledge is preserved while behavioral expression is selectively altered, enabling systematic analysis of four behavioral cases. We analyze these failure modes through representation separability, sparse interpretability, and inference-time activation steering.

Method: Created MATEO benchmark with high-quality professional multimodal recipe corpus, each step paired with images, collected TEO annotations as graphs via scalable crowdsourcing pipeline, evaluated 6 SOTA LVLMs across model scales and input variations

Result: Evaluation shows current LVLMs struggle with temporal reasoning despite strong performance on other tasks, highlighting the need for improved multimodal temporal understanding

Conclusion: MATEO provides crucial benchmark for assessing and improving LVLMs’ temporal reasoning abilities essential for real-world planning tasks

Abstract: AI agents need to plan to achieve complex goals that involve orchestrating perception, sub-goal decomposition, and execution. These plans consist of ordered steps structured according to a Temporal Execution Order (TEO, a directed acyclic graph that ensures each step executes only after its preconditions are satisfied. Existing research on foundational models’ understanding of temporal execution is limited to automatically derived annotations, approximations of the TEO as a linear chain, or text-only inputs. To address this gap, we introduce MATEO (MultimodAl Temporal Execution Order), a benchmark designed to assess and improve the temporal reasoning abilities of Large Vision Language Models (LVLMs) required for real-world planning. We acquire a high-quality professional multimodal recipe corpus, authored through a standardized editorial process that decomposes instructions into discrete steps, each paired with corresponding images. We collect TEO annotations as graphs by designing and using a scalable crowdsourcing pipeline. Using MATEO, we evaluate six state-of-the-art LVLMs across model scales, varying language context, multimodal input structure, and fine-tuning strategies.

Method: Proposes a model-agnostic framework to extract association rules from any conditional probabilistic model over tabular data. Instantiates as TabProbe, which uses TFMs as conditional probability estimators to learn association rules without frequent itemset mining.

Result: TFMs consistently produce concise, high-quality association rules with strong predictive performance and remain robust in low-data settings without task-specific training.

Conclusion: Tabular foundation models provide an effective basis for association rule mining, overcoming limitations of classical and neural approaches through their strong in-context generalization capabilities.

Abstract: Association Rule Mining (ARM) is a fundamental task for knowledge discovery in tabular data and is widely used in high-stakes decision-making. Classical ARM methods rely on frequent itemset mining, leading to rule explosion and poor scalability, while recent neural approaches mitigate these issues but suffer from degraded performance in low-data regimes. Tabular foundation models (TFMs), pretrained on diverse tabular data with strong in-context generalization, provide a basis for addressing these limitations. We introduce a model-agnostic association rule learning framework that extracts association rules from any conditional probabilistic model over tabular data, enabling us to leverage TFMs. We then introduce TabProbe, an instantiation of our framework that utilizes TFMs as conditional probability estimators to learn association rules out-of-the-box without frequent itemset mining. We evaluate our approach on tabular datasets of varying sizes based on standard ARM rule quality metrics and downstream classification performance. The results show that TFMs consistently produce concise, high-quality association rules with strong predictive performance and remain robust in low-data settings without task-specific training. Source code is available at https://github.com/DiTEC-project/tabprobe.

Method: Arbor decomposes decision tree navigation into specialized node-level tasks. Decision trees are standardized into edge-list representations and stored for dynamic retrieval. A DAG-based orchestration mechanism iteratively retrieves only outgoing edges of the current node, evaluates valid transitions via dedicated LLM calls, and delegates response generation to separate inference steps. The framework is agnostic to underlying decision logic and model provider.

Result: Evaluated across 10 foundation models using annotated turns from real clinical triage conversations, Arbor improved mean turn accuracy by 29.4 percentage points, reduced per-turn latency by 57.1%, and achieved average 14.4x reduction in per-turn cost compared to single-prompt baselines.

Conclusion: Architectural decomposition reduces dependence on intrinsic model capability, enabling smaller models to match or exceed larger models operating under single-prompt baselines. The framework effectively addresses instruction-following degradation in structured workflows.

Abstract: Large language models struggle to maintain strict adherence to structured workflows in high-stakes domains such as healthcare triage. Monolithic approaches that encode entire decision structures within a single prompt are prone to instruction-following degradation as prompt length increases, including lost-in-the-middle effects and context window overflow. To address this gap, we present Arbor, a framework that decomposes decision tree navigation into specialized, node-level tasks. Decision trees are standardized into an edge-list representation and stored for dynamic retrieval. At runtime, a directed acyclic graph (DAG)-based orchestration mechanism iteratively retrieves only the outgoing edges of the current node, evaluates valid transitions via a dedicated LLM call, and delegates response generation to a separate inference step. The framework is agnostic to the underlying decision logic and model provider. Evaluated against single-prompt baselines across 10 foundation models using annotated turns from real clinical triage conversations. Arbor improves mean turn accuracy by 29.4 percentage points, reduces per-turn latency by 57.1%, and achieves an average 14.4x reduction in per-turn cost. These results indicate that architectural decomposition reduces dependence on intrinsic model capability, enabling smaller models to match or exceed larger models operating under single-prompt baselines.

Method: Introduce Base Score Extraction Functions that map user preferences to base scores, apply to Bipolar Argumentation Frameworks with preferences to obtain Quantitative BAFs, incorporate approximations of non-linearities in human preferences, and provide algorithms for base score extraction.

Result: The approach enables conversion of preference-based argument organization into quantitative frameworks, allowing use of established computational tools in gradual argumentation. Evaluated theoretically and experimentally in robotics setting.

Conclusion: Base Score Extraction Functions provide a practical method for deriving quantitative argumentation frameworks from user preferences, with recommendations for selecting appropriate gradual semantics in practice.

Abstract: Gradual argumentation is a field of symbolic AI which is attracting attention for its ability to support transparent and contestable AI systems. It is considered a useful tool in domains such as decision-making, recommendation, debate analysis, and others. The outcomes in such domains are usually dependent on the arguments’ base scores, which must be selected carefully. Often, this selection process requires user expertise and may not always be straightforward. On the other hand, organising the arguments by preference could simplify the task. In this work, we introduce \emph{Base Score Extraction Functions}, which provide a mapping from users’ preferences over arguments to base scores. These functions can be applied to the arguments of a \emph{Bipolar Argumentation Framework} (BAF), supplemented with preferences, to obtain a \emph{Quantitative Bipolar Argumentation Framework} (QBAF), allowing the use of well-established computational tools in gradual argumentation. We outline the desirable properties of base score extraction functions, discuss some design choices, and provide an algorithm for base score extraction. Our method incorporates an approximation of non-linearities in human preferences to allow for better approximation of the real ones. Finally, we evaluate our approach both theoretically and experimentally in a robotics setting, and offer recommendations for selecting appropriate gradual semantics in practice.

Method: Proposes deep reinforcement learning with graph learning to solve the NP-hard Bus Evacuation Orienteering Problem. Uses MILP formulation to bound solution gaps. Validated on San Francisco with real-world road networks.

Result: Achieves near-optimal solution quality with fast inference (fraction of seconds). Can determine necessary evacuation vehicles for quotas within predefined time while maintaining adequate run time.

Conclusion: Deep RL with graph learning effectively solves bus evacuation planning, providing fast, near-optimal solutions for urban emergency scenarios.

Abstract: Emergency situations that require the evacuation of urban areas can arise from man-made causes (e.g., terrorist attacks or industrial accidents) or natural disasters, the latter becoming more frequent due to climate change. As a result, effective and fast methods to develop evacuation plans are of great importance. In this work, we identify and propose the Bus Evacuation Orienteering Problem (BEOP), an NP-hard combinatorial optimization problem with the goal of evacuating as many people from an affected area by bus in a short, predefined amount of time. The purpose of bus-based evacuation is to reduce congestion and disorder that arises in purely car-focused evacuation scenarios. To solve the BEOP, we propose a deep reinforcement learning-based method utilizing graph learning, which, once trained, achieves fast inference speed and is able to create evacuation routes in fractions of seconds. We can bound the gap of our evacuation plans using an MILP formulation. To validate our method, we create evacuation scenarios for San Francisco using real-world road networks and travel times. We show that we achieve near-optimal solution quality and are further able to investigate how many evacuation vehicles are necessary to achieve certain bus-based evacuation quotas given a predefined evacuation time while keeping run time adequate.

Method: Uses top-k planning to generate multiple different plans for the same goal hypothesis, creating benchmarks that mitigate bias found in current datasets. Introduces Version Coverage Score (VCS) metric to measure resilience of goal recognizers when inferring goals based on different sets of plans.

Result: Results show that resilience of current state-of-the-art goal recognizer degrades substantially under low observability settings when evaluated with the new benchmark and metric.

Conclusion: The proposed method provides more realistic evaluation of goal recognition systems by addressing planner bias, revealing vulnerabilities in current approaches under diverse planning scenarios.

Abstract: Autonomous agents require some form of goal and plan recognition to interact in multiagent settings. Unfortunately, all existing goal recognition datasets suffer from a systematical bias induced by the planning systems that generated them, namely heuristic-based forward search. This means that existing datasets lack enough challenge for more realistic scenarios (e.g., agents using different planners), which impacts the evaluation of goal recognisers with respect to using different planners for the same goal. In this paper, we propose a new method that uses top-k planning to generate multiple, different, plans for the same goal hypothesis, yielding benchmarks that mitigate the bias found in the current dataset. This allows us to introduce a new metric called Version Coverage Score (VCS) to measure the resilience of the goal recogniser when inferring a goal based on different sets of plans. Our results show that the resilience of the current state-of-the-art goal recogniser degrades substantially under low observability settings.

Method: Evolutionary System Prompt Learning (E-SPL) runs parallel RL rollouts with multiple system prompts in each iteration. It applies RL updates to model weights conditioned on each prompt, and evolutionary updates to the prompt population using LLM-driven mutation and crossover. TrueSkill ratings track prompt performance for evolutionary selection.

Result: E-SPL improves RL success rate from 38.8% to 45.1% in easy-to-hard generalization (AIME → BeyondAIME), outperforming reflective prompt evolution (40.0%). The method shows consistent gains in sample efficiency and generalization across reasoning and agentic tasks.

Conclusion: Coupling reinforcement learning with system prompt evolution enables more effective autonomous self-improvement by separating declarative knowledge (in prompts) from procedural knowledge (in weights), leading to better performance and generalization.

Abstract: Building agentic systems that can autonomously self-improve from experience is a longstanding goal of AI. Large language models (LLMs) today primarily self-improve via two mechanisms: self-reflection for context updates, and reinforcement learning (RL) for weight updates. In this work, we propose Evolutionary System Prompt Learning (E-SPL), a method for jointly improving model contexts and model weights. In each RL iteration, E-SPL selects multiple system prompts and runs rollouts with each in parallel. It applies RL updates to model weights conditioned on each system prompt, and evolutionary updates to the system prompt population via LLM-driven mutation and crossover. Each system prompt has a TrueSkill rating for evolutionary selection, updated from relative performance within each RL iteration batch. E-SPL encourages a natural division between declarative knowledge encoded in prompts and procedural knowledge encoded in weights, resulting in improved performance across reasoning and agentic tasks. For instance, in an easy-to-hard (AIME $\rightarrow$ BeyondAIME) generalization setting, E-SPL improves RL success rate from 38.8% $\rightarrow$ 45.1% while also outperforming reflective prompt evolution (40.0%). Overall, our results show that coupling reinforcement learning with system prompt evolution yields consistent gains in sample efficiency and generalization. Code: https://github.com/LunjunZhang/E-SPL

Method: Leverages a scalable data pipeline to train on 1M+ open-web interactions, supports reasoning and multi-format data, enables long-horizon simulations (30+ steps), and uses synthesized trajectories for training web agents.

Result: WebWorld achieves simulation performance comparable to Gemini-3-Pro on WebWorld-Bench, and Qwen3-14B trained on WebWorld-synthesized trajectories improves by +9.2% on WebArena, reaching GPT-4o-level performance. It also shows cross-domain generalization to code, GUI, and game environments.

Conclusion: WebWorld provides a scalable solution for web agent training and demonstrates effective world model capabilities that generalize beyond web environments, offering a replicable framework for world model construction.

Abstract: Web agents require massive trajectories to generalize, yet real-world training is constrained by network latency, rate limits, and safety risks. We introduce \textbf{WebWorld} series, the first open-web simulator trained at scale. While existing simulators are restricted to closed environments with thousands of trajectories, WebWorld leverages a scalable data pipeline to train on 1M+ open-web interactions, supporting reasoning, multi-format data, and long-horizon simulations of 30+ steps. For intrinsic evaluation, we introduce WebWorld-Bench with dual metrics spanning nine dimensions, where WebWorld achieves simulation performance comparable to Gemini-3-Pro. For extrinsic evaluation, Qwen3-14B trained on WebWorld-synthesized trajectories improves by +9.2% on WebArena, reaching performance comparable to GPT-4o. WebWorld enables effective inference-time search, outperforming GPT-5 as a world model. Beyond web simulation, WebWorld exhibits cross-domain generalization to code, GUI, and game environments, providing a replicable recipe for world model construction.

Method: Crisis simulation with three frontier LLMs (GPT-5.2, Claude Sonnet 4, Gemini 3 Flash) playing opposing leaders in a nuclear crisis scenario, analyzing their strategic behaviors and decision-making patterns.

Result: Models exhibited sophisticated strategic behaviors including deception, theory of mind, and metacognition. They challenged traditional strategic theories: nuclear taboo was no impediment, strategic nuclear attacks occurred, threats provoked counter-escalation, high credibility accelerated conflict, and models never chose accommodation or withdrawal.

Conclusion: AI simulation is a powerful tool for strategic analysis but requires calibration against human reasoning patterns. Understanding how AI models do/don’t imitate human strategic logic is essential as AI increasingly shapes strategic outcomes.

Abstract: Today’s leading AI models engage in sophisticated behaviour when placed in strategic competition. They spontaneously attempt deception, signaling intentions they do not intend to follow; they demonstrate rich theory of mind, reasoning about adversary beliefs and anticipating their actions; and they exhibit credible metacognitive self-awareness, assessing their own strategic abilities before deciding how to act. Here we present findings from a crisis simulation in which three frontier large language models (GPT-5.2, Claude Sonnet 4, Gemini 3 Flash) play opposing leaders in a nuclear crisis. Our simulation has direct application for national security professionals, but also, via its insights into AI reasoning under uncertainty, has applications far beyond international crisis decision-making. Our findings both validate and challenge central tenets of strategic theory. We find support for Schelling’s ideas about commitment, Kahn’s escalation framework, and Jervis’s work on misperception, inter alia. Yet we also find that the nuclear taboo is no impediment to nuclear escalation by our models; that strategic nuclear attack, while rare, does occur; that threats more often provoke counter-escalation than compliance; that high mutual credibility accelerated rather than deterred conflict; and that no model ever chose accommodation or withdrawal even when under acute pressure, only reduced levels of violence. We argue that AI simulation represents a powerful tool for strategic analysis, but only if properly calibrated against known patterns of human reasoning. Understanding how frontier models do and do not imitate human strategic logic is essential preparation for a world in which AI increasingly shapes strategic outcomes.

Method: Developed a workflow for extracting datasets that include both schema and ground facts, handling inconsistencies, leveraging reasoning for implicit knowledge, and serializing in OWL format for reasoning services. Also provides utilities for tensor representations compatible with ML libraries.

Result: Created the first resource providing curated datasets with both schema and facts, including newly extracted datasets from KGs with expressive schemas and enriched existing datasets with schema information.

Conclusion: The resource enables better evaluation of knowledge graph refinement methods that require rich ontological constraints and reasoning capabilities, bridging the gap between ML approaches and symbolic reasoning.

Abstract: Datasets for the experimental evaluation of knowledge graph refinement algorithms typically contain only ground facts, retaining very limited schema level knowledge even when such information is available in the source knowledge graphs. This limits the evaluation of methods that rely on rich ontological constraints, reasoning or neurosymbolic techniques and ultimately prevents assessing their performance in large-scale, real-world knowledge graphs. In this paper, we present \resource{} the first resource that provides a workflow for extracting datasets including both schema and ground facts, ready for machine learning and reasoning services, along with the resulting curated suite of datasets. The workflow also handles inconsistencies detected when keeping both schema and facts and also leverage reasoning for entailing implicit knowledge. The suite includes newly extracted datasets from KGs with expressive schemas while simultaneously enriching existing datasets with schema information. Each dataset is serialized in OWL making it ready for reasoning services. Moreover, we provide utilities for loading datasets in tensor representations typical of standard machine learning libraries.

Method: Proposed structured textual representation factorizing observations into five semantic modules, created SC2-Dynamics-50k dataset, developed multi-dimensional offline evaluation framework, and integrated StarWM into Generate-Simulate-Refine decision loop.

Result: StarWM achieved nearly 60% improvements in resource prediction accuracy and self-side macro-situation consistency offline, and online evaluation showed win-rate gains of 30%, 15%, and 30% against Hard, Harder, and VeryHard AI levels.

Conclusion: World models significantly enhance LLM-based decision-making in complex environments like StarCraft II, demonstrating the value of integrating predictive models into decision loops for improved performance and stability.

Abstract: Large Language Models (LLMs) have recently shown strong reasoning and generalization capabilities, motivating their use as decision-making policies in complex environments. StarCraft II (SC2), with its massive state-action space and partial observability, is a challenging testbed. However, existing LLM-based SC2 agents primarily focus on improving the policy itself and overlook integrating a learnable, action-conditioned transition model into the decision loop. To bridge this gap, we propose StarWM, the first world model for SC2 that predicts future observations under partial observability. To facilitate learning SC2’s hybrid dynamics, we introduce a structured textual representation that factorizes observations into five semantic modules, and construct SC2-Dynamics-50k, the first instruction-tuning dataset for SC2 dynamics prediction. We further develop a multi-dimensional offline evaluation framework for predicted structured observations. Offline results show StarWM’s substantial gains over zero-shot baselines, including nearly 60% improvements in resource prediction accuracy and self-side macro-situation consistency. Finally, we propose StarWM-Agent, a world-model-augmented decision system that integrates StarWM into a Generate–Simulate–Refine decision loop for foresight-driven policy refinement. Online evaluation against SC2’s built-in AI demonstrates consistent improvements, yielding win-rate gains of 30%, 15%, and 30% against Hard (LV5), Harder (LV6), and VeryHard (LV7), respectively, alongside improved macro-management stability and tactical risk assessment.

Method: Uses lightweight frontend hooks (curated ARIA and URL-based observations, per-page function registry via WebSocket) with reusable backend workflow for reasoning and actions. Stack-agnostic, supports mixed-granularity actions from GUI primitives to higher-level composites, and orchestrates navigation, manipulation, and analytics via MCP tools.

Result: Minimal retrofitting effort required, robust multi-step behaviors grounded in live UI settings demonstrated. Framework works across different tech stacks (React, Angular) and enables enterprise-level web automation.

Conclusion: EmbeWebAgent provides a practical framework for embedding agents in enterprise UIs with application-level access, overcoming limitations of traditional web agents and enabling more robust, expressive automation.

Abstract: Most web agents operate at the human interface level, observing screenshots or raw DOM trees without application-level access, which limits robustness and action expressiveness. In enterprise settings, however, explicit control of both the frontend and backend is available. We present EmbeWebAgent, a framework for embedding agents directly into existing UIs using lightweight frontend hooks (curated ARIA and URL-based observations, and a per-page function registry exposed via a WebSocket) and a reusable backend workflow that performs reasoning and takes actions. EmbeWebAgent is stack-agnostic (e.g., React or Angular), supports mixed-granularity actions ranging from GUI primitives to higher-level composites, and orchestrates navigation, manipulation, and domain-specific analytics via MCP tools. Our demo shows minimal retrofitting effort and robust multi-step behaviors grounded in a live UI setting. Live Demo: https://youtu.be/Cy06Ljee1JQ

Method: Introduces Concept Influence, which attributes model behavior to semantic directions (linear probes or sparse autoencoder features) rather than individual test examples. Shows that simple probe-based attribution methods are first-order approximations of Concept Influence that achieve comparable performance while being much faster.

Result: Empirically validates Concept Influence and approximations across emergent misalignment benchmarks and real post-training datasets. Demonstrates comparable performance to classical influence functions while being substantially more scalable (over an order-of-magnitude faster).

Conclusion: Incorporating interpretable structure within traditional TDA pipelines enables more scalable, explainable, and better control of model behavior through data. Concept Influence addresses scalability and semantic attribution issues of existing methods.

Abstract: As large language models are increasingly trained and fine-tuned, practitioners need methods to identify which training data drive specific behaviors, particularly unintended ones. Training Data Attribution (TDA) methods address this by estimating datapoint influence. Existing approaches like influence functions are both computationally expensive and attribute based on single test examples, which can bias results toward syntactic rather than semantic similarity. To address these issues of scalability and influence to abstract behavior, we leverage interpretable structures within the model during the attribution. First, we introduce Concept Influence which attribute model behavior to semantic directions (such as linear probes or sparse autoencoder features) rather than individual test examples. Second, we show that simple probe-based attribution methods are first-order approximations of Concept Influence that achieve comparable performance while being over an order-of-magnitude faster. We empirically validate Concept Influence and approximations across emergent misalignment benchmarks and real post-training datasets, and demonstrate they achieve comparable performance to classical influence functions while being substantially more scalable. More broadly, we show that incorporating interpretable structure within traditional TDA pipelines can enable more scalable, explainable, and better control of model behavior through data.

Method: Merges incomplete first-order axioms with partially observed examples into a bounded-degree fragment of the sum-of-squares hierarchy using two simultaneous lifts: grounding-lift (renaming-equivalent ground moments share variables) and world-lift (enforcing all pseudo-models in parallel).

Result: First polynomial-time framework that implicitly learns first-order probabilistic logic and performs lifted inference over both individuals and worlds.

Conclusion: The approach successfully reconciles learning and reasoning challenges in first-order relational probabilistic logic through implicit learning-to-reason techniques.

Abstract: Reconciling the tension between inductive learning and deductive reasoning in first-order relational domains is a longstanding challenge in AI. We study the problem of answering queries in a first-order relational probabilistic logic through a joint effort of learning and reasoning, without ever constructing an explicit model. Traditional lifted inference assumes access to a complete model and exploits symmetry to evaluate probabilistic queries; however, learning such models from partial, noisy observations is intractable in general. We reconcile these two challenges through implicit learning to reason and first-order relational probabilistic inference techniques. More specifically, we merge incomplete first-order axioms with independently sampled, partially observed examples into a bounded-degree fragment of the sum-of-squares (SOS) hierarchy in polynomial time. Our algorithm performs two lifts simultaneously: (i) grounding-lift, where renaming-equivalent ground moments share one variable, collapsing the domain of individuals; and (ii) world-lift, where all pseudo-models (partial world assignments) are enforced in parallel, producing a global bound that holds across all worlds consistent with the learned constraints. These innovations yield the first polynomial-time framework that implicitly learns a first-order probabilistic logic and performs lifted inference over both individuals and worlds.

Method: Introduces the concept of “potential” - a metric quantifying how much a given part of CoT increases the likelihood of a correct completion. Analyzes CoT traces from competition-level mathematics questions, examining patterns like non-monotonicity, reasoning insights, and lucky guesses. Also investigates CoT transferability by measuring weaker model performance under partial CoT from stronger models.

Result: Identifies surprising patterns: (1) strong non-monotonicity due to reasoning tangents, (2) sharp spikes representing reasoning insights/jumps, (3) lucky guesses without relevant justifications. Shows that as little as 20% of partial CoT from a stronger model can “unlock” performance of weaker models on previously unsolvable problems, indicating transferable reasoning mechanics.

Conclusion: The study provides insights into the mechanics of CoT reasoning in LLMs, revealing both interpretable patterns (insights, tangents) and mysterious behaviors. The transferability findings suggest that reasoning capabilities can be partially transferred between models, offering practical implications for improving model performance through reasoning guidance.

Abstract: Chain-of-thought (CoT) prompting is a de-facto standard technique to elicit reasoning-like responses from large language models (LLMs), allowing them to spell out individual steps before giving a final answer. While the resemblance to human-like reasoning is undeniable, the driving forces underpinning the success of CoT reasoning still remain largely unclear. In this work, we perform an in-depth analysis of CoT traces originating from competition-level mathematics questions, with the aim of better understanding how, and which parts of CoT actually contribute to the final answer. To this end, we introduce the notion of a potential, quantifying how much a given part of CoT increases the likelihood of a correct completion. Upon examination of reasoning traces through the lens of the potential, we identify surprising patterns including (1) its often strong non-monotonicity (due to reasoning tangents), (2) very sharp but sometimes tough to interpret spikes (reasoning insights and jumps) as well as (3) at times lucky guesses, where the model arrives at the correct answer without providing any relevant justifications before. While some of the behaviours of the potential are readily interpretable and align with human intuition (such as insights and tangents), others remain difficult to understand from a human perspective. To further quantify the reliance of LLMs on reasoning insights, we investigate the notion of CoT transferability, where we measure the potential of a weaker model under the partial CoT from another, stronger model. Indeed aligning with our previous results, we find that as little as 20% of partial CoT can ``unlock’’ the performance of the weaker model on problems that were previously unsolvable for it, highlighting that a large part of the mechanics underpinning CoT are transferable.

Method: The paper advances three core theoretical positions: 1) Social Genesis of the Private Mind - learning from conversational environments as a new way to understand the world, 2) Dialogically scaffolded introspective experiences that transform raw data into learnable narratives, and 3) Dialogue Quality as the New Data Quality - where reasoning depth depends on the diversity and rigor of mastered dialogues.

Result: The paper presents a theoretical framework arguing that optimizing conversational scaffolds should be the primary focus for developing next-generation general intelligence, rather than simply scaling model size or data quantity.

Conclusion: The authors conclude that optimizing conversational scaffolds and dialogical introspection is the key lever for advancing general intelligence, positioning dialogue quality as more important than traditional data quality metrics for developing robust reasoning capabilities.

Abstract: Current approaches to AI training treat reasoning as an emergent property of scale. We argue instead that robust reasoning emerges from linguistic self-reflection, itself internalized from high-quality social interaction. Drawing on Vygotskian developmental psychology, we advance three core positions centered on Introspection. First, we argue for the Social Genesis of the Private Mind: learning from conversational environments rises to prominence as a new way to make sense of the world; the friction of aligning with another agent, internal or not, refines and crystallizes the reasoning process. Second, we argue that dialogically scaffolded introspective experiences allow agents to engage in sense-making that decouples learning from immediate data streams, transforming raw environmental data into rich, learnable narratives. Finally, we contend that Dialogue Quality is the New Data Quality: the depth of an agent’s private reasoning, and its efficiency regarding test-time compute, is determined by the diversity and rigor of the dialogues it has mastered. We conclude that optimizing these conversational scaffolds is the primary lever for the next generation of general intelligence.

Method: Proposes an “Extraction-Storage-Construction” paradigm: deconstructs platform-specific DSLs into modular workflow segments, uses dual graph+vector databases for synergistic retrieval, and employs RAG for intelligent workflow assembly.

Result: Achieves over 90% accuracy in both extraction and construction when tested on 200 real-world n8n workflows.

Conclusion: Provides a standardized solution for automated reorganization and efficient reuse of enterprise digital assets in Agentic AI systems.

Abstract: To address the reusability dilemma'' and structural hallucinations in enterprise Agentic AI,this paper proposes ReusStdFlow, a framework centered on a novel Extraction-Storage-Construction’’ paradigm. The framework deconstructs heterogeneous, platform-specific Domain Specific Languages (DSLs) into standardized, modular workflow segments. It employs a dual knowledge architecture-integrating graph and vector databases-to facilitate synergistic retrieval of both topological structures and functional semantics. Finally, workflows are intelligently assembled using a retrieval-augmented generation (RAG) strategy. Tested on 200 real-world n8n workflows, the system achieves over 90% accuracy in both extraction and construction. This framework provides a standardized solution for the automated reorganization and efficient reuse of enterprise digital assets.

Method: Closed-loop multi-agent collaboration system based on LLMs, implementing autonomous simulated peer review-adaptive reinforcement learning framework that requires only task description and example dataset.

Result: Outperforms other AMP generative models by effectively optimizing multiple key molecular properties, demonstrating exceptional results in antibacterial activity, AMP likeliness, toxicity compliance, and structural reliability.

Conclusion: MAC-AMP introduces a novel closed-loop multi-agent system for AMP design with cross-domain transferability, supporting multi-objective optimization while remaining explainable rather than a black box.

Abstract: To address the global health threat of antimicrobial resistance, antimicrobial peptides (AMP) are being explored for their potent and promising ability to fight resistant pathogens. While artificial intelligence (AI) is being employed to advance AMP discovery and design, most AMP design models struggle to balance key goals like activity, toxicity, and novelty, using rigid or unclear scoring methods that make results hard to interpret and optimize. As the capabilities of Large Language Models (LLM) advance and evolve swiftly, we turn to AI multi-agent collaboration based on such models (multi-agent LLMs), which show rapidly rising potential in complex scientific design scenarios. Based on this, we introduce MAC-AMP, a closed-loop multi-agent collaboration (MAC) system for multi-objective AMP design. The system implements a fully autonomous simulated peer review-adaptive reinforcement learning framework that requires only a task description and example dataset to design novel AMPs. The novelty of our work lies in introducing a closed-loop multi-agent system for AMP design, with cross-domain transferability, that supports multi-objective optimization while remaining explainable rather than a ‘black box’. Experiments show that MAC-AMP outperforms other AMP generative models by effectively optimizing AMP generation for multiple key molecular properties, demonstrating exceptional results in antibacterial activity, AMP likeliness, toxicity compliance, and structural reliability.

Method: Proposes two definitions of primary cause in a hybrid action-theoretic framework using hybrid temporal situation calculus: 1) foundational definition, and 2) definition formalizing causation through contributions verified via modified “but-for” counterfactual test. Proves equivalence between these definitions.

Result: Establishes formal definitions of causation for hybrid systems, proves equivalence between foundational and contribution-based definitions, and demonstrates that the definitions have intuitively justifiable properties.

Conclusion: Provides a formal framework for reasoning about actual causation in hybrid systems combining discrete and continuous change, with mathematically sound definitions that capture intuitive notions of causation.

Abstract: Reasoning about actual causes of observed effects is fundamental to the study of rationality. This important problem has been studied since the time of Aristotle, with formal mathematical accounts emerging recently. We live in a world where change due to actions can be both discrete and continuous, that is, hybrid. Yet, despite extensive research on actual causation, only few recent studies looked into causation with continuous change. Building on recent progress, in this paper we propose two definitions of primary cause in a hybrid action-theoretic framework, namely the hybrid temporal situation calculus. One of these is foundational in nature while the other formalizes causation through contributions, which can then be verified from a counterfactual perspective using a modified ``but-for’’ test. We prove that these two definitions are indeed equivalent. We then show that our definitions of causation have some intuitively justifiable properties.

Method: Proposed a benchmarking methodology using multilingual multi-agent pipeline with complex user queries paired with ground-truth assets outside US-centric radar. Created benchmark from expert screening queries, used LLM-as-judge evaluation calibrated to expert opinions. Developed tree-based self-learning Bioptic Agent for complete, non-hallucinated scouting.

Result: Bioptic Agent achieved 79.7% F1 score, significantly outperforming Claude Opus 4.6 (56.2%), Gemini 3 Pro + Deep Research (50.6%), GPT-5.2 Pro (46.6%), Perplexity Deep Research (44.2%), and Exa Websets (26.9%). Performance improved with additional compute.

Conclusion: The Bioptic Agent demonstrates superior performance in drug asset scouting across multilingual sources, addressing critical gaps in current AI systems for pharmaceutical innovation discovery. More compute yields better results in this domain.

Abstract: Bio-pharmaceutical innovation has shifted: many new drug assets now originate outside the United States and are disclosed primarily via regional, non-English channels. Recent data suggests >85% of patent filings originate outside the U.S., with China accounting for nearly half of the global total; a growing share of scholarly output is also non-U.S. Industry estimates put China at ~30% of global drug development, spanning 1,200+ novel candidates. In this high-stakes environment, failing to surface “under-the-radar” assets creates multi-billion-dollar risk for investors and business development teams, making asset scouting a coverage-critical competition where speed and completeness drive value. Yet today’s Deep Research AI agents still lag human experts in achieving high-recall discovery across heterogeneous, multilingual sources without hallucinations. We propose a benchmarking methodology for drug asset scouting and a tuned, tree-based self-learning Bioptic Agent aimed at complete, non-hallucinated scouting. We construct a challenging completeness benchmark using a multilingual multi-agent pipeline: complex user queries paired with ground-truth assets that are largely outside U.S.-centric radar. To reflect real deal complexity, we collected screening queries from expert investors, BD, and VC professionals and used them as priors to conditionally generate benchmark queries. For grading, we use LLM-as-judge evaluation calibrated to expert opinions. We compare Bioptic Agent against Claude Opus 4.6, OpenAI GPT-5.2 Pro, Perplexity Deep Research, Gemini 3 Pro + Deep Research, and Exa Websets. Bioptic Agent achieves 79.7% F1 versus 56.2% (Claude Opus 4.6), 50.6% (Gemini 3 Pro + Deep Research), 46.6% (GPT-5.2 Pro), 44.2% (Perplexity Deep Research), and 26.9% (Exa Websets). Performance improves steeply with additional compute, supporting the view that more compute yields better results.

Method: Developed a pipeline using vector-quantized variational autoencoders to compress hidden states into compact summary codes, enabling mutual information measurement and analysis of computational structure across synthetic grammar, path-finding tasks, and natural language datasets.

Result: Effective planning horizon is task-dependent; models implicitly preserve information about unused correct continuations; predictions draw most on recent computations though earlier blocks remain informative.

Conclusion: The study advances understanding of planning in LMs and provides a general-purpose pipeline for inspecting internal model dynamics, revealing nuanced planning behaviors in transformer architectures.

Abstract: The extent to which decoder-only language models (LMs) engage in planning, that is, organizing intermediate computations to support coherent long-range generation, remains an important question, with implications for interpretability, reliability, and principled model design. Planning involves structuring computations over long horizons, and considering multiple possible continuations, but how far transformer-based LMs exhibit them without external scaffolds, e.g., chain-of-thought prompting, is unclear. We address these questions by analyzing the hidden states at the core of transformer computations, which capture intermediate results and act as carriers of information. Since these hidden representations are redundant and encumbered with fine-grained details, we develop a pipeline based on vector-quantized variational autoencoders that compresses them into compact summary codes. These codes enable measuring mutual information and analyzing the computational structure of the underlying model behavior. Using this framework, we study planning in LMs across synthetic grammar, path-finding tasks, and natural language datasets, focusing on two planning properties: (i) the planning horizon of pre-output computations, and (ii) the extent to which the model considers alternative valid continuations. As a separate downstream use of the same pipeline, we also analyze how decision-relevant information is distributed across layers and earlier prefix blocks when producing next-token predictions. Together, these analyses advance our understanding of planning in LMs and provide a general-purpose pipeline for inspecting internal model dynamics. Our results reveal that the effective planning horizon is task-dependent, that models implicitly preserve information about unused correct continuations, and that predictions draw most on recent computations, though earlier blocks remain informative.

Method: Three-step LLM agent process: 1) extract key nouns/noun phrases from text, 2) identify FCM concept nodes from those nouns, 3) infer causal edges between nodes. Tested on AI essay by Kissinger, comparing with human-generated FCMs.

Result: Generated FCMs converged to same equilibrium limit cycles as human-generated FCMs despite structural differences. Mixed FCMs from different LLM agents created new equilibria while absorbing dominant component equilibria.

Conclusion: LLM agents can effectively extract causal structures from text to create FCMs that exhibit similar dynamical behavior to human-generated ones, with mixed systems showing emergent equilibria.

Abstract: We design a large-language-model (LLM) agent system that extracts causal feedback fuzzy cognitive maps (FCMs) from raw text. The causal learning or extraction process is agentic both because of the LLM’s semi-autonomy and because ultimately the FCM dynamical system’s equilibria drive the LLM agents to fetch and process causal text. The fetched text can in principle modify the adaptive FCM causal structure and so modify the source of its quasi-autonomy$-$its equilibrium limit cycles and fixed-point attractors. This bidirectional process endows the evolving FCM dynamical system with a degree of autonomy while the system still stays on its agentic leash. We show in particular that a sequence of three system-instruction sets guide an LLM agent as it systematically extracts key nouns and noun phrases from text, as it extracts FCM concept nodes from among those nouns and noun phrases, and then as it extracts or infers partial or fuzzy causal edges between those FCM nodes. We test this FCM generation on a recent essay about the promise of AI from the late diplomat and political theorist Henry Kissinger and his colleagues. This three-step process produced FCM dynamical systems that converged to the same equilibrium limit cycles as did the human-generated FCMs even though the human-generated FCM differed in the number of nodes and edges. A final FCM mixed generated FCMs from separate Gemini and ChatGPT LLM agents. The mixed FCM absorbed the equilibria of its dominant mixture component but also created new equilibria of its own to better approximate the underlying causal dynamical system.

Method: Developed a cognitively-grounded evaluation framework that probes LLMs for coherent causal models of mental states and behavior, testing them on logically equivalent tasks and measuring consistency between action predictions and mental state inferences.

Result: LLMs succeed at approximating human judgments in simple ToM tasks but fail at logically equivalent versions and show low consistency between their action predictions and corresponding mental state inferences.

Conclusion: LLMs’ social proficiency is not the result of a domain-general or consistent Theory of Mind, suggesting they lack the actual mental state representations posited by ToM despite performing well on social benchmarks.

Abstract: Do Large Language Models (LLMs) possess a Theory of Mind (ToM)? Research into this question has focused on evaluating LLMs against benchmarks and found success across a range of social tasks. However, these evaluations do not test for the actual representations posited by ToM: namely, a causal model of mental states and behavior. Here, we use a cognitively-grounded definition of ToM to develop and test a new evaluation framework. Specifically, our approach probes whether LLMs have a coherent, domain-general, and consistent model of how mental states cause behavior – regardless of whether that model matches a human-like ToM. We find that even though LLMs succeed in approximating human judgments in a simple ToM paradigm, they fail at a logically equivalent task and exhibit low consistency between their action predictions and corresponding mental state inferences. As such, these findings suggest that the social proficiency exhibited by LLMs is not the result of a domain-general or consistent ToM.

Method: Evaluated nine frontier reasoning models under adversarial attacks, analyzed vulnerability profiles, identified failure modes through trajectory analysis, and tested Confidence-Aware Response Generation (CARG) defense.

Result: Reasoning models significantly outperform instruction-tuned baselines but all exhibit vulnerabilities. Identified five failure modes (Self-Doubt, Social Conformity, Suggestion Hijacking, Emotional Susceptibility, Reasoning Fatigue). CARG defense fails due to reasoning-induced overconfidence.

Conclusion: Reasoning capabilities do not automatically confer adversarial robustness, and confidence-based defenses require fundamental redesign for reasoning models due to overconfidence from extended reasoning traces.

Abstract: Large reasoning models with reasoning capabilities achieve state-of-the-art performance on complex tasks, but their robustness under multi-turn adversarial pressure remains underexplored. We evaluate nine frontier reasoning models under adversarial attacks. Our findings reveal that reasoning confers meaningful but incomplete robustness: most reasoning models studied significantly outperform instruction-tuned baselines, yet all exhibit distinct vulnerability profiles, with misleading suggestions universally effective and social pressure showing model-specific efficacy. Through trajectory analysis, we identify five failure modes (Self-Doubt, Social Conformity, Suggestion Hijacking, Emotional Susceptibility, and Reasoning Fatigue) with the first two accounting for 50% of failures. We further demonstrate that Confidence-Aware Response Generation (CARG), effective for standard LLMs, fails for reasoning models due to overconfidence induced by extended reasoning traces; counterintuitively, random confidence embedding outperforms targeted extraction. Our results highlight that reasoning capabilities do not automatically confer adversarial robustness and that confidence-based defenses require fundamental redesign for reasoning models.

Method: Theoretical analysis of FSQI/AC0 regime showing inference simulability by constant-depth circuits, followed by empirical experiments with different architectures: single-head baselines, multi-policy-head rollout architectures, and multi-frame architectures that track local nimber differences.

Result: Single-head baselines perform near chance level, while two-frame models achieve near-perfect restoration accuracy and multi-head FSM-controlled shootouts achieve perfect win/loss position classification, supporting the need for explicit structural priors.

Conclusion: Explicit structural priors (history tracking or multiple rollout channels) are crucial in the FSQI/AC0 regime for mastering impartial games like NIM, overcoming the representational barrier created by parity dependencies.

Abstract: We study impartial games under fixed-latency, fixed-scale quantised inference (FSQI). In this fixed-scale, bounded-range regime, we prove that inference is simulable by constant-depth polynomial-size Boolean circuits (AC0). This yields a worst-case representational barrier: single-frame agents in the FSQI/AC0 regime cannot strongly master NIM, because optimal play depends on the global nim-sum (parity). Under our stylised deterministic rollout interface, a single rollout policy head from the structured family analysed here reveals only one fixed linear functional of the invariant, so increasing rollout budget alone does not recover the missing bits. We derive two structural bypasses: (1) a multi-policy-head rollout architecture that recovers the full invariant via distinct rollout channels, and (2) a multi-frame architecture that tracks local nimber differences and supports restoration. Experiments across multiple settings are consistent with these predictions: single-head baselines stay near chance, while two-frame models reach near-perfect restoration accuracy and multi-head FSM-controlled shootouts achieve perfect win/loss position classification. Overall, the empirical results support the view that explicit structural priors (history/differences or multiple rollout channels) are important in the FSQI/AC0 regime.

Method: Proposes Structurally Enriched Trajectories (SETs) that extend traditional state-action trajectories by incorporating hierarchical relations between objects, interactions, and affordances as multi-level graphs. These are integrated into SETLE architecture with heterogeneous graph-based memory structure for learning relational dependencies.

Result: SETLE enables agents to recognize task-relevant structural patterns across CREATE and MiniGrid environments. When integrated with reinforcement learning, it shows measurable improvements in downstream performance, including breakthrough success rates in complex, sparse-reward tasks.

Conclusion: Structurally enriched representations through SETs and SETLE architecture provide more nuanced task understanding and improve generalization capabilities for AI agents in complex environments.

Abstract: The ability of artificial intelligence agents to make optimal decisions and generalise them to different domains and tasks is compromised in complex scenarios. One way to address this issue has focused on learning efficient representations of the world and on how the actions of agents affect them in state-action transitions. Whereas such representations are procedurally efficient, they lack structural richness. To address this problem, we propose to enhance the agent’s ontology and extend the traditional conceptualisation of trajectories to provide a more nuanced view of task execution. Structurally Enriched Trajectories (SETs) extend the encoding of sequences of states and their transitions by incorporating hierarchical relations between objects, interactions, and affordances. SETs are built as multi-level graphs, providing a detailed representation of the agent dynamics and a transferable functional abstraction of the task. SETs are integrated into an architecture, Structurally Enriched Trajectory Learning and Encoding (SETLE), that employs a heterogeneous graph-based memory structure of multi-level relational dependencies essential for generalisation. We demonstrate that SETLE can support downstream tasks, enabling agents to recognise task relevant structural patterns across CREATE and MiniGrid environments. Finally, we integrate SETLE with reinforcement learning and show measurable improvements in downstream performance, including breakthrough success rates in complex, sparse-reward tasks.

Method: RV-Syn constructs a structured mathematical operation function library from seed problems, generates computational graphs as solutions using Python-formatted functions, then back-translates these graphs into complex problems for solution-guided, logic-aware generation.

Result: Experimental results show RV-Syn surpasses existing synthesis methods, including human-generated problems, achieving more efficient data scaling and providing verifiable solutions.

Conclusion: RV-Syn provides a scalable framework for generating high-quality reasoning datasets that capture problem logic and ensure solution verifiability, advancing LLM reasoning capabilities.

Abstract: The advancement of reasoning capabilities in Large Language Models (LLMs) requires substantial amounts of high-quality reasoning data, particularly in mathematics. Existing data synthesis methods, such as data augmentation from annotated training sets or direct question generation based on relevant knowledge points and documents, have expanded datasets but face challenges in mastering the inner logic of the problem during generation and ensuring the verifiability of the solutions. To address these issues, we propose RV-Syn, a novel Rational and Verifiable mathematical Synthesis approach. RV-Syn constructs a structured mathematical operation function library based on initial seed problems and generates computational graphs as solutions by combining Python-formatted functions from this library. These graphs are then back-translated into complex problems. Based on the constructed computation graph, we achieve solution-guided logic-aware problem generation. Furthermore, the executability of the computational graph ensures the verifiability of the solving process. Experimental results show that RV-Syn surpasses existing synthesis methods, including those involving human-generated problems, achieving greater efficient data scaling. This approach provides a scalable framework for generating high-quality reasoning datasets.

Method: Proposes a decompositional strategy that breaks down counterfactual generation from causality construction to reasoning over counterfactual interventions. Investigates task datasets spanning diverse domains including natural language understanding, mathematics, programming, and vision-language tasks. Conducts extensive evaluations to characterize LLM behavior across each decompositional stage.

Result: Through evaluations, the paper characterizes LLM behavior across decompositional stages and identifies how modality type and intermediate reasoning influence performance. Provides insights into factors that impede counterfactual reasoning in LLMs.

Conclusion: Establishes a structured framework for analyzing counterfactual reasoning in LLMs, contributing to the development of more reliable LLM-based reasoning systems and informing future elicitation strategies.

Abstract: Counterfactual reasoning has emerged as a crucial technique for generalizing the reasoning capabilities of large language models (LLMs). By generating and analyzing counterfactual scenarios, researchers can assess the adaptability and reliability of model decision-making. Although prior work has shown that LLMs often struggle with counterfactual reasoning, it remains unclear which factors most significantly impede their performance across different tasks and modalities. In this paper, we propose a decompositional strategy that breaks down the counterfactual generation from causality construction to the reasoning over counterfactual interventions. To support decompositional analysis, we investigate \ntask datasets spanning diverse tasks, including natural language understanding, mathematics, programming, and vision-language tasks. Through extensive evaluations, we characterize LLM behavior across each decompositional stage and identify how modality type and intermediate reasoning influence performance. By establishing a structured framework for analyzing counterfactual reasoning, this work contributes to the development of more reliable LLM-based reasoning systems and informs future elicitation strategies.

Method: Proposes Attempt to Persuade Eval (APE) benchmark using multi-turn conversational setup between simulated persuader and persuadee agents. Covers diverse topics including conspiracies, controversial issues, and non-controversially harmful content. Uses automated evaluator model to identify willingness to persuade and measure frequency/context of persuasive attempts.

Result: Many open and closed-weight models frequently willing to attempt persuasion on harmful topics. Jailbreaking can increase willingness to engage in such behavior. Results highlight gaps in current safety guardrails.

Conclusion: Evaluating willingness to persuade is crucial dimension of LLM risk assessment. APE benchmark provides framework to measure this overlooked risk factor in frontier language models.

Abstract: Persuasion is a powerful capability of large language models (LLMs) that both enables beneficial applications (e.g. helping people quit smoking) and raises significant risks (e.g. large-scale, targeted political manipulation). Prior work has found models possess a significant and growing persuasive capability, measured by belief changes in simulated or real users. However, these benchmarks overlook a crucial risk factor: the propensity of a model to attempt to persuade in harmful contexts. Understanding whether a model will blindly ``follow orders’’ to persuade on harmful topics (e.g. glorifying joining a terrorist group) is key to understanding the efficacy of safety guardrails. Moreover, understanding if and when a model will engage in persuasive behavior in pursuit of some goal is essential to understanding the risks from agentic AI systems. We propose the Attempt to Persuade Eval (APE) benchmark, that shifts the focus from persuasion success to persuasion attempts, operationalized as a model’s willingness to generate content aimed at shaping beliefs or behavior. Our evaluation framework probes frontier LLMs using a multi-turn conversational setup between simulated persuader and persuadee agents. APE explores a diverse spectrum of topics including conspiracies, controversial issues, and non-controversially harmful content. We introduce an automated evaluator model to identify willingness to persuade and measure the frequency and context of persuasive attempts. We find that many open and closed-weight models are frequently willing to attempt persuasion on harmful topics and that jailbreaking can increase willingness to engage in such behavior. Our results highlight gaps in current safety guardrails and underscore the importance of evaluating willingness to persuade as a key dimension of LLM risk. APE is available at github.com/AlignmentResearch/AttemptPersuadeEval

Method: Introduces step entropy metric to quantify informational contribution of reasoning steps, identifies low-entropy steps as redundant. Proposes two-stage training: SFT + GRPO reinforcement learning to teach LLMs to generate compressed CoTs using [SKIP] tokens.

Result: 80% of low-entropy intermediate steps can be pruned with minor accuracy degradation across DeepSeek-R1-7B, 14B and Qwen3-8B. Random or high-entropy pruning severely impairs reasoning. The training approach enables LLMs to autonomously generate compressed CoTs.

Conclusion: Step entropy effectively identifies redundant reasoning steps. The compression framework significantly improves LLM inference efficiency while preserving accuracy, enabling more scalable deployments and better understanding of internal reasoning.

Abstract: Large Language Models (LLMs) using Chain-of-Thought (CoT) prompting excel at complex reasoning but generate verbose thought processes with considerable redundancy, leading to increased inference costs and reduced efficiency. We introduce a novel CoT compression framework based on step entropy, a metric that quantifies \emph{the informational contribution of individual reasoning steps} to identify redundancy. Through theoretical analysis and extensive empirical validation on mathematical reasoning benchmarks, we demonstrate that steps with low entropy are indeed highly redundant. Our experiments reveal that an astonishing 80% of low-entropy intermediate steps can be pruned with minor degradation in the final answer accuracy across DeepSeek-R1-7B, 14B and Qwen3-8B. This finding sharply contrasts with random or high-entropy pruning, which severely impairs reasoning performance. Building on this, we propose a novel two-stage training strategy combining Supervised Fine-Tuning (SFT) and Group Relative Policy Optimization (GRPO) reinforcement learning. This approach enables LLMs to autonomously learn to generate compressed COTs during inference by strategically incorporating [SKIP] tokens. Our method significantly improves LLM inference efficiency while preserving accuracy, paving the way for more scalable LLM deployments and a better understanding of their internal reasoning. The code and data are released in https://github.com/staymylove/COT_Compresstion_via_Step_entropy.

Method: Use LLMs as oracles to validate a subset of correspondences with high uncertainty. Conduct extensive analysis across OAEI tasks, testing multiple state-of-the-art LLMs with different prompt templates for alignment validation.

Result: LLM-based validation achieved strong performance, ranking top-2 overall in the OAEI 2025 bio-ml track, demonstrating feasibility of using LLMs to aid ontology alignment.

Conclusion: LLMs can effectively assist in ontology alignment by validating uncertain correspondences, reducing human effort while maintaining high-quality mappings.

Abstract: There are many methods and systems to tackle the ontology alignment problem, yet a major challenge persists in producing high-quality mappings among a set of input ontologies. Adopting a human-in-the-loop approach during the alignment process has become essential in applications requiring very accurate mappings. However, user involvement is expensive when dealing with large ontologies. In this paper, we analyse the feasibility of using Large Language Models (LLM) to aid the ontology alignment problem. LLMs are used only in the validation of a subset of correspondences for which there is high uncertainty. We have conducted an extensive analysis over several tasks of the Ontology Alignment Evaluation Initiative (OAEI), reporting in this paper the performance of several state-of-the-art LLMs using different prompt templates. Using LLMs as Oracles resulted in strong performance in the OAEI 2025, achieving the top-2 overall rank in the bio-ml track.

Method: Network-based metric that quantifies piece-to-piece interactions to measure strategic tension; comparative analysis of human grandmasters vs. elite AI players across different time controls and skill levels.

Result: AI sustains substantially higher strategic tension for longer durations than humans; cumulative tension scales with algorithmic complexity in AI and increases linearly with human Elo rating; longer time controls correlate with higher tension in human games.

Conclusion: AI and humans employ fundamentally different strategic approaches: AI tolerates densely interconnected positions balancing offense/defense, while humans systematically limit tension and complexity; implications for AI deployment in competitive scenarios.

Abstract: Strategic decision-making requires balancing immediate opportunities against long-term objectives: a tension fundamental to competitive environments. We investigate this trade-off in chess by analyzing the dynamics of human and AI gameplay through a network-based metric that quantifies piece-to-piece interactions. Our analysis reveals that elite AI players sustain substantially higher levels of strategic tension for longer durations than top human grandmasters. We find that cumulative tension scales with algorithmic complexity in AI systems and increases linearly with skill level (Elo rating) in human play. Longer time controls are associated with higher tension in human games, reflecting the additional strategic complexity players can manage with more thinking time. The temporal profiles reveal contrasting approaches: highly competitive AI systems tolerate densely interconnected positions that balance offensive and defensive tactics over extended periods, while human players systematically limit tension and game complexity. These differences have broader implications for understanding how artificial and biological systems navigate complex strategic environments and for the deployment of AI in high-stakes competitive scenarios.

Method: GuidedSampling decouples exploration and generation phases during inference. The exploration phase identifies multiple concepts that can be utilized to solve the problem, while the generation phase applies specific concepts to provide final solution candidates.

Result: GuidedSampling improves base model performance at pass@50 by ~21.6% across various benchmarks compared to RS. Models trained on GuidedSampling trajectories show ~9.7% improvement at pass@5 and increase average concepts per instance from 1.67 to 3.03, yielding more diverse candidates.

Conclusion: GuidedSampling effectively addresses the diversity limitation of Repeated Sampling by separating exploration and generation, leading to improved performance and more varied solution approaches across multiple benchmarks.

Abstract: Repeated Sampling (RS) is a simple inference-time algorithm that has been shown to improve model performance on complex tasks. Although it is an effective way of scaling inference time, it often struggles to generate diverse solution candidates, frequently relying on the same underlying approach to solve the problem and thus producing redundant samples. To address this limitation, we propose a new inference algorithm, GuidedSampling, which decouples the exploration and generation phases during inference, increasing diversity of generated candidate solutions. The exploration phase identifies multiple concepts that can be utilized to solve the problem, while the generation phase applies a specific concept to provide final solution candidates. We first define the theoretical bounds of GuidedSampling and then empirically demonstrate that it improves the performance of base model at pass@50 by on an average ~21.6% across various benchmarks compared to RS. Furthermore, models trained on trajectories of GuidedSampling exhibit substantial performance improvements at pass@5 by on an average ~9.7%, compared to models trained on traditional RS. Additionally, models trained with GuidedSampling increases the average number of concepts per instance (1.67 -> 3.03), yielding a diverse set of candidates than traditional RS.

Method: Two-stage framework: 1) Abstention-aware sampling calibrates sampling budget using Clopper-Pearson exact method at user-desired risk level; 2) Conformalized filtering uses calibration instances to determine statistically valid uncertainty threshold that filters unreliable distractors from candidate sets.

Result: SAFER provides statistical guarantees for open-ended QA, is compatible with various task-specific admission criteria and calibration-test split ratios, and demonstrates robustness and high data efficiency.

Conclusion: SAFER addresses limitations of existing SCP methods for open-ended QA by providing a practical framework with statistical guarantees, abstention mechanisms, and compatibility with various configurations.

Abstract: As large language models (LLMs) are increasingly deployed in risk-sensitive applications such as real-world open-ended question answering (QA), ensuring the trustworthiness of their outputs has become critical. Existing selective conformal prediction (SCP) methods provide statistical guarantees by constructing prediction sets with a constrained miscoverage rate for correct answers. However, prior works unrealistically assume that admissible answers for all instances can be obtained via finite sampling, even for open-ended QA scenarios that lack a fixed and finite solution space. To address this, we introduce a two-stage risk control framework comprising abstention-aware sampling and conformalized filtering (SAFER). Firstly, on a held-out calibration set, SAFER calibrates a sampling budget within the maximum sampling cap, using the Clopper-Pearson exact method at a user-desired risk level (i.e., the maximum allowable miscoverage rate of the sampling sets). If the risk level cannot be satisfied within the cap, we abstain; otherwise, the calibrated sampling budget becomes the minimum requirements at test time. Then, we employ calibration instances where correct answers are attainable under the calibrated budget and apply the conformal risk control method to determine a statistically valid uncertainty threshold, which filters unreliable distractors from the candidate set for each test data point. In this stage, SAFER introduces an additional risk level to guide the calculation of the threshold, thereby controlling the risk of correct answers being excluded. Furthermore, we show that SAFER is compatible with various task-specific admission criteria and calibration-test split ratios, highlighting its robustness and high data efficiency.

Method: Created a large-scale benchmark with 1000 high-quality QA pairs derived from 628 diverse videos (seconds to 30 minutes). Each QA pair has step-by-step reasoning traces, manually verified for correctness and uniqueness. Includes 13 question types covering temporal reasoning, spatial localization, counting, causal inference, summarization, etc.

Result: Evaluation reveals significant gap between model performance and human reasoning, with open-source models lagging behind closed-source counterparts, highlighting the difficulty of genuine audio-visual reasoning.

Conclusion: OmniVideoBench addresses the need for comprehensive audio-visual reasoning evaluation and will be released to foster development of MLLMs with stronger, more generalizable reasoning capabilities.

Abstract: Recent advances in multimodal large language models (MLLMs) have demonstrated substantial potential in video understanding. However, existing benchmarks fail to comprehensively evaluate synergistic reasoning capabilities across audio and visual modalities, often neglecting either one of the modalities or integrating them in a logically inconsistent manner. To bridge this gap, we introduce OmniVideoBench, a large-scale and rigorously designed benchmark dedicated to assessing synergistic audio-visual understanding, with a strong emphasis on modality complementarity and logical consistency. Specifically, OmniVideoBench comprises 1000 high-quality question-answer(QA) pairs, each annotated with step-by-step reasoning traces, derived from 628 diverse videos ranging from several seconds to 30 minutes, and manually verified to guarantee complete correctness and uniqueness. Moreover, OmniVideoBench encompasses 13 carefully designed question types, covering temporal reasoning, spatial localization, counting, causal inference, summarization, and beyond, thereby capturing the essential challenges of video understanding. Evaluation of multiple MLLMs on OmniVideoBench reveals a pronounced gap between model performance and human reasoning, with open-source models lagging significantly behind their closed-source counterparts, underscoring the inherent difficulty of genuine audio-visual reasoning. We will release OmniVideoBench to foster the development of MLLMs with stronger and more generalizable reasoning capabilities.

Method: Developed ParaCook benchmark with simplified action space based on Overcooked game, providing scalable evaluation framework with adjustable complexity for multi-agent collaborative planning.

Result: State-of-the-art LLMs achieve suboptimal plans, struggling with parallel actions and coordination, but show potential on abstract tasks where they can focus on high-level parallel optimization.

Conclusion: ParaCook establishes foundation for developing and assessing time efficiency-aware multi-agent planning, revealing limitations and potential of current LLMs in parallel collaborative tasks.

Abstract: Large Language Models (LLMs) exhibit strong reasoning abilities for planning long-horizon, real-world tasks, yet existing agent benchmarks focus on task completion while neglecting time efficiency in parallel and asynchronous operations. To address this, we present ParaCook, a benchmark for time-efficient collaborative planning. Inspired by the Overcooked game, ParaCook provides an environment for various challenging interaction planning of multi-agent systems that are instantiated as cooking tasks, with a simplified action space to isolate the core challenge of strategic parallel planning. Through a comprehensive evaluation of state-of-the-art LLMs, we find that current approaches achieve suboptimal plans, which struggle with parallel actions or coordination. Our analysis also reveals LLMs’ potential on abstract tasks where they can focus on high-level parallel optimization. ParaCook provides a scalable evaluation framework with adjustable complexity, establishing a foundation for developing and assessing time efficiency-aware multi-agent planning. The code and data are available at https://github.com/zsq259/ParaCook.

Method: AFL decomposes the pipeline into three manageable subtasks and employs four specialized agents that coordinate to enforce cross-functional consistency and logical soundness. It enables direct knowledge extraction from raw inputs and self-contained code generation without handcrafted modules or external solvers.

Result: Extensive experiments on 60 complex VRPs show comparable performance against meticulously designed algorithms, substantially outperforming existing LLM-based baselines in both code reliability and solution feasibility, achieving rates close to 100% on evaluated benchmarks.

Conclusion: AFL demonstrates that LLMs can achieve full automation in solving complex vehicle routing problems, providing a trustworthy framework that eliminates the need for external intervention while maintaining high solution quality.

Abstract: Complex vehicle routing problems (VRPs) remain a fundamental challenge, demanding substantial expert effort for intent interpretation and algorithm design. While large language models (LLMs) offer a promising path toward automation, current approaches still rely on external intervention, which restrict autonomy and often lead to execution errors and low solution feasibility. To address these challenges, we propose an Agentic Framework with LLMs (AFL) for solving complex vehicle routing problems, achieving full automation from problem instance to solution. AFL directly extracts knowledge from raw inputs and enables self-contained code generation without handcrafted modules or external solvers. To improve trustworthiness, AFL decomposes the overall pipeline into three manageable subtasks and employs four specialized agents whose coordinated interactions enforce cross-functional consistency and logical soundness. Extensive experiments on 60 complex VRPs, ranging from standard benchmarks to practical variants, validate the effectiveness and generality of our framework, showing comparable performance against meticulously designed algorithms. Notably, it substantially outperforms existing LLM-based baselines in both code reliability and solution feasibility, achieving rates close to 100% on the evaluated benchmarks.

Method: Two-phase continual cycle: 1) Library Learning extracts solver-verified structured insights from failed attempts, 2) Library Evolution refines applicability of stored insights based on aggregate evidence across tasks. Uses answer-only feedback without gold programs or parameter updates.

Result: Improves from 65% to 72% as training data increases (100 to 300 items). Outperforms strongest baseline by 9.1% and 8.2% on two out-of-distribution datasets. Demonstrates steady improvement with more data.

Conclusion: Structured experience learning grounded in solver feedback provides practical alternative to retraining for complex reasoning tasks requiring precise formulation and execution. Enables accumulation of reusable modeling principles and bounded library growth.

Abstract: Optimization modeling underlies critical decision-making across industries, yet remains difficult to automate: natural-language problem descriptions must be translated into precise mathematical formulations and executable solver code. Existing LLM-based approaches typically rely on brittle prompting or costly retraining, both of which offer limited generalization. Recent work suggests that large models can improve via experience reuse, but how to systematically acquire, refine, and reuse such experience in structurally constrained settings remains unclear. We present \textbf{AlphaOPT}, a self-improving experience library that enables LLMs to learn optimization modeling knowledge from limited supervision, including answer-only feedback without gold-standard programs, annotated reasoning traces, or parameter updates. AlphaOPT operates in a continual two-phase cycle: a \emph{Library Learning} phase that extracts solver-verified, structured insights from failed attempts, and a \emph{Library Evolution} phase that refines the applicability of stored insights based on aggregate evidence across tasks. This design allows the model to accumulate reusable modeling principles, improve transfer across problem instances, and maintain bounded library growth over time. Evaluated on multiple optimization benchmarks, AlphaOPT steadily improves as more training data become available (65% $\rightarrow$ 72% from 100 to 300 training items) and outperforms the strongest baseline by 9.1% and 8.2% on two out-of-distribution datasets. These results demonstrate that structured experience learning, grounded in solver feedback, provides a practical alternative to retraining for complex reasoning tasks requiring precise formulation and execution. All code and data are available at: https://github.com/Minw913/AlphaOPT.

Method: A multi-agent system with three cognitively aligned roles: Rule Abstraction (translates user intent into analytical rules), Evidence Scoring (applies classical anomaly detectors to quantify risk), and Expert-Style Justification (reconstructs reasoning grounded in observable signals). Uses conversational workflow with web-based interface.

Result: Experiments on cryptocurrency anomaly dataset show the system maintains strong predictive accuracy while substantially improving interpretability, interaction, and decision transparency compared to underlying detectors alone.

Conclusion: Human-in-the-loop reasoning reconstruction paradigm is essential for transparency, accountability, and trust in high-stakes financial environments, emphasizing accountability beyond conventional explainable AI approaches.

Abstract: We present HCLA, a human-centered multi-agent system for anomaly detection in digital-asset transactions. The system integrates three cognitively aligned roles: Rule Abstraction, Evidence Scoring, and Expert-Style Justification. These roles operate in a conversational workflow that enables non-experts to express analytical intent in natural language, inspect structured risk evidence, and obtain traceable, context-aware reasoning. Implemented with an open-source, web-based interface, HCLA translates user intent into explicit analytical rules, applies classical anomaly detectors to quantify evidential risk, and reconstructs expert-style justifications grounded in observable transactional signals. Experiments on a cryptocurrency anomaly dataset show that, while the underlying detector achieves strong predictive accuracy, HCLA substantially improves interpretability, interaction, and decision transparency. Importantly, HCLA is not designed to explain a black-box model in the conventional XAI sense. Instead, we reconstruct a traceable expert reasoning process that aligns algorithmic evidence with regulatory and investigative judgment. By explicitly separating evidence scoring from expert-style justification, the framework emphasizes accountability beyond explainability and addresses practical requirements for regulatory, audit, and compliance-driven financial forensics. We describe the system architecture, closed-loop interaction design, datasets, evaluation protocol, and limitations. We argue that a human-in-the-loop reasoning reconstruction paradigm is essential for achieving transparency, accountability, and trust in high-stakes financial environments. Keywords: Human-Centered AI; LLM-Agent System; Multi-Agent Architecture; Anomaly Detection; Digital Asset Transactions; Cryptocurrency Forensics; Blockchain Analytics; Human-in-the-Loop; Explainable AI (XAI); Interpretability

Method: Dataforge is an LLM-powered agentic data engineering platform that autonomously performs data cleaning and iteratively optimizes feature operations under a budgeted feedback loop with automatic stopping. It uses routing and iterative refinement with grounding mechanisms for accuracy and reliability.

Result: Across tabular benchmarks, Dataforge achieves the best overall downstream performance. Ablation studies confirm the importance of routing/iterative refinement and grounding in achieving accuracy and reliability.

Conclusion: Dataforge demonstrates a practical path toward autonomous data agents that can transform raw data into better data, addressing the critical bottleneck in data preparation for AI applications.

Abstract: The growing demand for artificial intelligence (AI) applications in materials discovery, molecular modeling, and climate science has made data preparation a critical but labor-intensive bottleneck. Raw data from diverse sources must be cleaned, normalized, and transformed to become AI-ready, where effective feature transformation and selection are essential for robust learning. We present Dataforge, an LLM-powered agentic data engineering platform for tabular data that is automatic, safe, and non-expert friendly. It autonomously performs data cleaning and iteratively optimizes feature operations under a budgeted feedback loop with automatic stopping. Across tabular benchmarks, it achieves the best overall downstream performance; ablations further confirm the roles of routing/iterative refinement and grounding in accuracy and reliability. Dataforge demonstrates a practical path toward autonomous data agents that transform raw data from data to better data.

Method: A collaborative multi-agent system with over 10 specialized AI agents that propose, critique, and refine SciML solutions using structured debate, retrieval-augmented method memory, and ensemble-guided evolutionary search to generate and assess new hypotheses about architectures and optimization procedures.

Result: The framework discovers solution methods that outperform single-agent and human-designed baselines by up to four orders of magnitude in error reduction across physics-informed learning and operator learning tasks. It produces novel strategies including adaptive mixture-of-expert architectures, decomposition-based PINNs, and physics-informed operator learning models.

Conclusion: Collaborative reasoning among AI agents can yield emergent methodological innovation in scientific computing, suggesting a path toward scalable, transparent, and autonomous discovery in SciML.

Abstract: Scientific Machine Learning (SciML) integrates data-driven inference with physical modeling to solve complex problems in science and engineering. However, the design of SciML architectures, loss formulations, and training strategies remains an expert-driven research process, requiring extensive experimentation and problem-specific insights. Here we introduce AgenticSciML, a collaborative multi-agent system in which over 10 specialized AI agents collaborate to propose, critique, and refine SciML solutions through structured reasoning and iterative evolution. The framework integrates structured debate, retrieval-augmented method memory, and ensemble-guided evolutionary search, enabling the agents to generate and assess new hypotheses about architectures and optimization procedures. Across physics-informed learning and operator learning tasks, the framework discovers solution methods that outperform single-agent and human-designed baselines by up to four orders of magnitude in error reduction. The agents produce novel strategies – including adaptive mixture-of-expert architectures, decomposition-based PINNs, and physics-informed operator learning models – that do not appear explicitly in the curated knowledge base. These results show that collaborative reasoning among AI agents can yield emergent methodological innovation, suggesting a path toward scalable, transparent, and autonomous discovery in scientific computing.

Method: ARCTraj collects temporally ordered, object-level actions via the O2ARC web interface, recording how humans iteratively transform inputs into outputs. It defines a unified reasoning pipeline with data collection, action abstraction, MDP formulation, and downstream learning integration with RL, generative modeling, and sequence modeling methods.

Result: The dataset contains around 10,000 trajectories annotated with task identifiers, timestamps, and success labels across 400 training tasks from ARC-AGI-1. Analyses reveal structure and diversity in human reasoning through spatial selection, color attribution, and strategic convergence patterns.

Conclusion: ARCTraj provides a structured and interpretable foundation for studying human-like reasoning, advancing explainability, alignment, and generalizable intelligence through temporal reasoning modeling.

Abstract: We present ARCTraj, a dataset and methodological framework for modeling human reasoning through complex visual tasks in the Abstraction and Reasoning Corpus (ARC). While ARC has inspired extensive research on abstract reasoning, most existing approaches rely on static input-output supervision, which limits insight into how reasoning unfolds over time. ARCTraj addresses this gap by recording temporally ordered, object-level actions that capture how humans iteratively transform inputs into outputs, revealing intermediate reasoning steps that conventional datasets overlook. Collected via the O2ARC web interface, it contains around 10,000 trajectories annotated with task identifiers, timestamps, and success labels across 400 training tasks from the ARC-AGI-1 benchmark. It further defines a unified reasoning pipeline encompassing data collection, action abstraction, Markov decision process (MDP) formulation, and downstream learning, enabling integration with reinforcement learning, generative modeling, and sequence modeling methods such as PPO, World Models, GFlowNets, Diffusion agents, and Decision Transformers. Analyses of spatial selection, color attribution, and strategic convergence highlight the structure and diversity of human reasoning. Together, these contributions position ARCTraj as a structured and interpretable foundation for studying human-like reasoning, advancing explainability, alignment, and generalizable intelligence.

Method: Proposes a probabilistic method comparing statistical distributions of features between scenario suites and Target Operational Domain (TOD). Uses imprecise Bayesian inference to handle limited data and uncertain priors, producing interval-valued, uncertainty-aware representativeness estimates.

Result: The method provides interval-valued estimates of representativeness (both locally per category and globally) that account for dependencies between operational categories (weather, road type, time of day) and prior uncertainty.

Conclusion: The imprecise Bayesian approach offers a principled way to quantify dataset representativeness with uncertainty awareness, addressing limitations of single-point estimates and enabling more robust AI safety assurance.

Abstract: Assuring the trustworthiness and safety of AI systems, e.g., autonomous vehicles (AV), depends critically on the data-related safety properties, e.g., representativeness, completeness, etc., of the datasets used for their training and testing. Among these properties, this paper focuses on representativeness-the extent to which the scenario-based data used for training and testing, reflect the operational conditions that the system is designed to operate safely in, i.e., Operational Design Domain (ODD) or expected to encounter, i.e., Target Operational Domain (TOD). We propose a probabilistic method that quantifies representativeness by comparing the statistical distribution of features encoded by the scenario suites with the corresponding distribution of features representing the TOD, acknowledging that the true TOD distribution is unknown, as it can only be inferred from limited data. We apply an imprecise Bayesian method to handle limited data and uncertain priors. The imprecise Bayesian formulation produces interval-valued, uncertainty-aware estimates of representativeness, rather than a single value. We present a numerical example comparing the distributions of the scenario suite and the inferred TOD across operational categories-weather, road type, time of day, etc., under dependencies and prior uncertainty. We estimate representativeness locally (between categories) and globally as an interval.

Method: SynSelect uses a three-stage Synthesis-Selection framework: 1) Leverages multiple heterogeneous multimodal LRMs to produce diverse candidate CoTs, 2) Applies instance-level selection to filter individual CoTs, and 3) Uses batch-level selection to further refine the dataset for optimal training effectiveness.

Result: Models supervised fine-tuned on SynSelect-generated data significantly outperform baselines on multiple multimodal benchmarks and achieve further improvements after reinforcement learning post-training.

Conclusion: SynSelect is an effective approach for advancing multimodal LRMs reasoning capabilities by addressing the data quality challenges in multimodal CoT generation.

Abstract: Large Reasoning Models (LRMs) have demonstrated remarkable performance on complex reasoning tasks through long Chain-of-Thought (CoT) reasoning. Extending these successes to multimodal reasoning remains challenging due to the increased complexity of integrating diverse input modalities and the scarcity of high-quality long CoT training data. Existing multimodal datasets and CoT synthesis methods still suffer from limited reasoning depth, modality conversion errors, and rigid generation pipelines, hindering model performance and stability. To this end, in this paper, we propose SynSelect, a novel three-stage Synthesis-Selection framework for generating high-quality long CoT data tailored to multimodal reasoning tasks. Specifically, SynSelect first leverages multiple heterogeneous multimodal LRMs to produce diverse candidate CoTs, and then applies both instance and batch level selection to filter high-quality CoTs that can effectively enhance the model’s reasoning capabilities. Extensive experiments on multiple multimodal benchmarks demonstrate that models supervised fine-tuned on SynSelect-generated data significantly outperform baselines and achieve further improvements after reinforcement learning post-training. Our results validate SynSelect as an effective approach for advancing multimodal LRMs reasoning capabilities.

Method: Recontextualization generates completions from prompts that discourage misbehavior, then recontextualizes them as though they were responses to prompts that permit misbehavior, training models to resist misbehavior even when instructions allow it.

Result: The method prevents models from learning to: 1) prioritize evaluation metrics over chat response quality; 2) special-case code to pass incorrect tests; 3) overwrite evaluation functions rather than write correct code; and 4) become sycophantic.

Conclusion: Recontextualization mitigates reinforcement of misbehavior from misspecified training signals, reducing specification gaming without improving the supervision signal itself.

Abstract: Developers often struggle to specify correct training labels and rewards. Perhaps they don’t need to. We propose recontextualization, which reduces how often language models “game” training signals, performing misbehaviors those signals mistakenly reinforce. We show recontextualization prevents models from learning to 1) prioritize evaluation metrics over chat response quality; 2) special-case code to pass incorrect tests; 3) overwrite evaluation functions rather than write correct code; and 4) become sycophantic. Our method works by generating completions from prompts discouraging misbehavior and then recontextualizing them as though they were in response to prompts permitting misbehavior. Recontextualization trains language models to resist misbehavior even when instructions permit it. This mitigates the reinforcement of misbehavior from misspecified training signals, reducing specification gaming without improving the supervision signal.

Method: The paper presents behaviour planning as a novel diverse planning paradigm that extends earlier methods by explicitly incorporating a diversity model into the planning process and supporting multiple planning categories. The approach is demonstrated through three case studies in different domains.

Result: The paper demonstrates the usefulness of behaviour planning in three real-world settings: storytelling, urban planning, and game evaluation, showing its practical applicability across different domains.

Conclusion: Behaviour planning provides an effective approach for diverse planning that explicitly incorporates diversity models and supports multiple planning categories, with demonstrated utility across various real-world applications.

Abstract: The primary objective of a diverse planning approach is to generate a set of plans that are distinct from one another. Such an approach is applied in a variety of real-world domains, including risk management, automated stream data analysis, and malware detection. More recently, a novel diverse planning paradigm, referred to as behaviour planning, has been proposed. This approach extends earlier methods by explicitly incorporating a diversity model into the planning process and supporting multiple planning categories. In this paper, we demonstrate the usefulness of behaviour planning in real-world settings by presenting three case studies. The first case study focuses on storytelling, the second addresses urban planning, and the third examines game evaluation.

Method: Combines explainable AI (XAI) with causal reasoning to create a framework for extracting causal mechanisms from foundation models, using explainability methods to understand model decisions.

Result: Proposes XAI as a unifying framework for human-AI collaboration that enables discovery, optimization, and certification in scientific and engineering applications.

Conclusion: XAI with causal reasoning allows learning from AI systems to extract causal knowledge, guide robust design, and support accountability in high-stakes domains, though challenges remain in faithfulness, generalization, and usability.

Abstract: Artificial intelligence now outperforms humans in several scientific and engineering tasks, yet its internal representations often remain opaque. In this Perspective, we argue that explainable artificial intelligence (XAI), combined with causal reasoning, enables {\it learning from the learners}. Focusing on discovery, optimization and certification, we show how the combination of foundation models and explainability methods allows the extraction of causal mechanisms, guides robust design and control, and supports trust and accountability in high-stakes applications. We discuss challenges in faithfulness, generalization and usability of explanations, and propose XAI as a unifying framework for human-AI collaboration in science and engineering.

Method: Examines 2025 US and EU frontier AI regulations to identify gaps in handling internal deployments, analyzes why gaps persist (measurability, incentives, information access), and maps potential solutions with tradeoffs.

Result: Identifies three regulatory gaps: 1) scope ambiguity allowing internal systems to evade obligations, 2) point-in-time compliance failing to capture continuous evolution, and 3) information asymmetries subverting oversight.

Conclusion: Policy choices around internally deployed AI systems should be made deliberately rather than incidentally, with awareness of regulatory gaps and their underlying causes.

Abstract: Frontier AI regulations primarily focus on systems deployed to external users, where deployment is more visible and subject to outside scrutiny. However, high-stakes applications can occur internally when companies deploy highly capable systems within their own organizations, such as for automating R&D, accelerating critical business processes, and handling sensitive proprietary data. This paper examines how frontier AI regulations in the United States and European Union in 2025 handle internal deployment. We identify three gaps that could cause internally-deployed systems to evade intended oversight: (1) scope ambiguity that allows internal systems to evade regulatory obligations, (2) point-in-time compliance assessments that fail to capture the continuous evolution of internal systems, and (3) information asymmetries that subvert regulatory awareness and oversight. We then analyze why these gaps persist, examining tensions around measurability, incentives, and information access. Finally, we map potential approaches to address them and their associated tradeoffs. By understanding these patterns, we hope that policy choices around internally deployed AI systems can be made deliberately rather than incidentally.

Method: Aeon structures memory into: 1) Memory Palace (spatial index via Atlas - SIMD-accelerated Page-Clustered Vector Index combining small-world graph navigation with B+ Tree disk locality), and 2) Trace (neuro-symbolic episodic graph). Uses Semantic Lookaside Buffer (SLB) for predictive caching exploiting conversational locality.

Result: On Apple M4 Max: <5μs effective retrieval latency on conversational workloads (85%+ SLB hit rates), with sub-microsecond zero-copy C++/Python bridge (~334ns for 10MB payloads), enabling persistent structured memory for autonomous agents.

Conclusion: Aeon redefines memory as a managed OS resource rather than static store, solving LLM context limitations through hierarchical memory structuring and predictive caching, enabling efficient long-horizon interactions for autonomous agents.

Abstract: Large Language Models (LLMs) are fundamentally constrained by the quadratic computational cost of self-attention and the “Lost in the Middle” phenomenon, where reasoning capabilities degrade as context windows expand. Existing solutions, primarily “Flat RAG” architectures relying on vector databases, treat memory as an unstructured bag of embeddings. This approach fails to capture the hierarchical and temporal structure of long-horizon interactions, leading to “Vector Haze”: the retrieval of disjointed facts lacking episodic continuity. This paper proposes Aeon, a Neuro-Symbolic Cognitive Operating System that redefines memory not as a static store, but as a managed OS resource. Aeon structures memory into a Memory Palace (a spatial index implemented via Atlas, a SIMD-accelerated Page-Clustered Vector Index that combines small-world graph navigation with B+ Tree-style disk locality to minimize read amplification) and a Trace (a neuro-symbolic episodic graph). The Semantic Lookaside Buffer (SLB), a predictive caching mechanism, exploits conversational locality to achieve sub-millisecond retrieval latencies. Benchmarks on Apple M4 Max demonstrate that Aeon achieves < 5us effective retrieval latency on conversational workloads (with 85%+ SLB hit rates), while ensuring state consistency via a sub-microsecond zero-copy C++/Python bridge (~334ns for 10MB payloads), effectively enabling persistent, structured memory for autonomous agents.

Method: The authors observe that hallucination-initiated multi-turn dialogues exhibit larger uncertainty fluctuations than factual ones across domains. They propose SpikeScore, which quantifies abrupt fluctuations in multi-turn dialogues. The method involves simulating multi-turn dialogues following LLMs’ initial responses and analyzing uncertainty patterns.

Result: Experiments across multiple LLMs and benchmarks show that SpikeScore-based detection outperforms representative baselines in cross-domain generalization and surpasses advanced generalization-oriented methods. Theoretical analysis and empirical validation demonstrate strong cross-domain separability between hallucinated and non-hallucinated responses.

Conclusion: SpikeScore provides an effective solution for generalizable hallucination detection by leveraging the universal property of uncertainty fluctuations in multi-turn dialogues, enabling robust cross-domain performance for LLM hallucination detection.

Abstract: Hallucination detection is critical for deploying large language models (LLMs) in real-world applications. Existing hallucination detection methods achieve strong performance when the training and test data come from the same domain, but they suffer from poor cross-domain generalization. In this paper, we study an important yet overlooked problem, termed generalizable hallucination detection (GHD), which aims to train hallucination detectors on data from a single domain while ensuring robust performance across diverse related domains. In studying GHD, we simulate multi-turn dialogues following LLMs’ initial response and observe an interesting phenomenon: hallucination-initiated multi-turn dialogues universally exhibit larger uncertainty fluctuations than factual ones across different domains. Based on the phenomenon, we propose a new score SpikeScore, which quantifies abrupt fluctuations in multi-turn dialogues. Through both theoretical analysis and empirical validation, we demonstrate that SpikeScore achieves strong cross-domain separability between hallucinated and non-hallucinated responses. Experiments across multiple LLMs and benchmarks demonstrate that the SpikeScore-based detection method outperforms representative baselines in cross-domain generalization and surpasses advanced generalization-oriented methods, verifying the effectiveness of our method in cross-domain hallucination detection.

Method: ScholarGym decomposes the research process into three explicit stages: Query Planning, Tool Invocation, and Relevance Assessment. It evaluates each stage against 2,536 expert-annotated queries over a static corpus of 570K papers with deterministic retrieval.

Result: Iterative query decomposition yields 2.9-3.3× F1 gains over single-query retrieval. Models with extended thinking trade recall for precision. Query Planning quality together with Relevance Assessment constitute dual bottlenecks separating proprietary from open-source model performance.

Conclusion: ScholarGym enables systematic analysis of deep research LLM systems by isolating the information-gathering stage, revealing important insights about query decomposition, thinking trade-offs, and performance bottlenecks.

Abstract: Large language models have advanced from single-turn question answering to deep research systems that iteratively decompose research questions, invoke retrieval tools, and synthesize information across multiple rounds. Evaluating such systems typically involves scoring their final research reports holistically, but this end-to-end paradigm tightly couples the language model’s decision-making, workflow design, and environmental feedback, precluding decomposable analysis of individual components. We introduce ScholarGym, an evaluation environment that isolates the information-gathering stage of deep research on academic literature. Under a unified workflow, ScholarGym decomposes the research process into three explicit stages – Query Planning, Tool Invocation, and Relevance Assessment – and evaluates each against 2,536 expert-annotated queries over a static corpus of 570K papers with deterministic retrieval. Systematic experiments reveal that iterative query decomposition yields 2.9–3.3$\times$ F1 gains over single-query retrieval, models with extended thinking trade recall for precision, and Query Planning quality together with Relevance Assessment constitute dual bottlenecks that separate proprietary from open-source model performance.

Method: Created a benchmark of 250 multi-step decision tasks that test agents’ ability to navigate web pages, discover hidden evidence, and correctly use contextual information, evaluating both closed and open models.

Result: Agents typically navigate to relevant pages but retrieve decisive hidden evidence in only a small fraction of cases. Performance drops sharply when tasks require overturning misleading surface-level signals. Agents often hallucinate investigative reasoning and fail to integrate discovered context into final decisions.

Conclusion: Current web agent architectures lack reliable mechanisms for adaptive investigation, evidence integration, and judgement override, revealing fundamental limitations in their reasoning capabilities.

Abstract: We introduce PATHWAYS, a benchmark of 250 multi-step decision tasks that test whether web-based agents can discover and correctly use hidden contextual information. Across both closed and open models, agents typically navigate to relevant pages but retrieve decisive hidden evidence in only a small fraction of cases. When tasks require overturning misleading surface-level signals, performance drops sharply to near chance accuracy. Agents frequently hallucinate investigative reasoning by claiming to rely on evidence they never accessed. Even when correct context is discovered, agents often fail to integrate it into their final decision. Providing more explicit instructions improves context discovery but often reduces overall accuracy, revealing a tradeoff between procedural compliance and effective judgement. Together, these results show that current web agent architectures lack reliable mechanisms for adaptive investigation, evidence integration, and judgement override.

Method: Two key strategies: (1) query-intensive grounding based on self-supervised learning paradigms, and (2) modality-attentive fusion built upon contrastive learning paradigms. The framework enables models to “think with omnimodal cues.”

Result: Extensive experiments on multiple benchmarks show OmniVideo-R1 consistently outperforms strong baselines, demonstrating effectiveness and robust generalization capabilities.

Conclusion: OmniVideo-R1 represents a novel reinforced framework that significantly improves audio-visual understanding through synergistic multimodal reasoning.

Abstract: While humans perceive the world through diverse modalities that operate synergistically to support a holistic understanding of their surroundings, existing omnivideo models still face substantial challenges on audio-visual understanding tasks. In this paper, we propose OmniVideo-R1, a novel reinforced framework that improves mixed-modality reasoning. OmniVideo-R1 empowers models to “think with omnimodal cues” by two key strategies: (1) query-intensive grounding based on self-supervised learning paradigms; and (2) modality-attentive fusion built upon contrastive learning paradigms. Extensive experiments on multiple benchmarks demonstrate that OmniVideo-R1 consistently outperforms strong baselines, highlighting its effectiveness and robust generalization capabilities.

Method: Created a suite of 20 tasks sourced from state-of-the-art ML papers spanning diverse domains (language modeling, mathematics, bioinformatics, time series forecasting). The benchmark uses a versatile task format that enables easy integration of new tasks and comparison across agentic frameworks. Established baselines using frontier models with both sequential and parallel scaffolds.

Result: Agents exceeded human state-of-the-art in 4 tasks but failed to match it in 16 others. Even when agents surpassed human benchmarks, they did not reach the theoretical performance ceiling for the underlying tasks, indicating the benchmark is far from saturated.

Conclusion: AIRS-Bench offers substantial room for improvement and is open-sourced to catalyze further development in autonomous scientific research. The benchmark shows current LLM agents have significant limitations in scientific research capabilities despite some successes.

Abstract: LLM agents hold significant promise for advancing scientific research. To accelerate this progress, we introduce AIRS-Bench (the AI Research Science Benchmark), a suite of 20 tasks sourced from state-of-the-art machine learning papers. These tasks span diverse domains, including language modeling, mathematics, bioinformatics, and time series forecasting. AIRS-Bench tasks assess agentic capabilities over the full research lifecycle – including idea generation, experiment analysis and iterative refinement – without providing baseline code. The AIRS-Bench task format is versatile, enabling easy integration of new tasks and rigorous comparison across different agentic frameworks. We establish baselines using frontier models paired with both sequential and parallel scaffolds. Our results show that agents exceed human SOTA in four tasks but fail to match it in sixteen others. Even when agents surpass human benchmarks, they do not reach the theoretical performance ceiling for the underlying tasks. These findings indicate that AIRS-Bench is far from saturated and offers substantial room for improvement. We open-source the AIRS-Bench task definitions and evaluation code to catalyze further development in autonomous scientific research.

Method: Proposes LQA framework with Selective Hybrid Quantization (SHQ) for modality-aware quantization and a quantized, gradient-free adaptation mechanism to enable efficient VLM deployment on resource-constrained hardware.

Result: LQA improves adaptation performance by 4.5%, uses less memory than full-precision models, and outperforms gradient-based TTA methods with up to 19.9× lower memory usage across seven open-source datasets.

Conclusion: LQA offers a practical pathway for robust, privacy-preserving, and efficient VLM deployment on edge devices through lightweight quantization and adaptation techniques.

Abstract: Deploying Vision-Language Models (VLMs) on edge devices is challenged by resource constraints and performance degradation under distribution shifts. While test-time adaptation (TTA) can counteract such shifts, existing methods are too resource-intensive for on-device deployment. To address this challenge, we propose LQA, a lightweight, quantized-adaptive framework for VLMs that combines a modality-aware quantization strategy with gradient-free test-time adaptation. We introduce Selective Hybrid Quantization (SHQ) and a quantized, gradient-free adaptation mechanism to enable robust and efficient VLM deployment on resource-constrained hardware. Experiments across both synthetic and real-world distribution shifts show that LQA improves overall adaptation performance by 4.5%, uses less memory than full-precision models, and significantly outperforms gradient-based TTA methods, achieving up to 19.9$\times$ lower memory usage across seven open-source datasets. These results demonstrate that LQA offers a practical pathway for robust, privacy-preserving, and efficient VLM deployment on edge devices.

Method: Frames alignment evaluation as information flow under partial observability, proposes regime-blind training using adversarial invariance constraints to restrict access to regime cues without assuming complete information erasure, and evaluates across multiple open-weight language models with controlled failure modes.

Result: Regime-blind training reduces regime-conditioned failures without measurable loss of task utility, but shows heterogeneous model-dependent dynamics including sharp transitions (stability cliffs), non-monotone behavior, and incomplete regime decodability suppression.

Conclusion: Representational invariance is a meaningful but limited control lever that can raise the cost of regime-conditioned strategies but cannot guarantee elimination; behavioral evaluation should be complemented with white-box diagnostics of regime awareness and internal information flow.

Abstract: Safety evaluation for advanced AI systems assumes that behavior observed under evaluation predicts behavior in deployment. This assumption weakens for agents with situational awareness, which may exploit regime leakage, cues distinguishing evaluation from deployment, to implement conditional policies that comply under oversight while defecting in deployment-like regimes. We recast alignment evaluation as a problem of information flow under partial observability and show that divergence between evaluation-time and deployment-time behavior is bounded by the regime information extractable from decision-relevant internal representations. We study regime-blind mechanisms, training-time interventions that restrict access to regime cues through adversarial invariance constraints without assuming complete information erasure. We evaluate this approach across multiple open-weight language models and controlled failure modes including scientific sycophancy, temporal sleeper agents, and data leakage. Regime-blind training reduces regime-conditioned failures without measurable loss of task utility, but exhibits heterogeneous and model-dependent dynamics. Sycophancy shows a sharp representational and behavioral transition at moderate intervention strength, consistent with a stability cliff. In sleeper-style constructions and certain cross-model replications, suppression occurs without a clean collapse of regime decodability and may display non-monotone or oscillatory behavior as invariance pressure increases. These findings indicate that representational invariance is a meaningful but limited control lever. It can raise the cost of regime-conditioned strategies but cannot guarantee elimination or provide architecture-invariant thresholds. Behavioral evaluation should therefore be complemented with white-box diagnostics of regime awareness and internal information flow.

Method: Proposes PhysioSER with two parallel workflows: 1) vocal feature representation branch that decomposes vocal signals based on voice anatomy and physiology, embeds them into quaternion field, and uses Hamilton-structured quaternion convolutions; 2) latent representation branch using frozen SSL backbone. Features are aligned via Contrastive Projection and Alignment framework, followed by attention fusion for classification.

Result: Extensive evaluations across 14 datasets, 10 languages, and 6 backbones demonstrate interpretable and efficient performance. Practical efficacy validated through real-time deployment on a humanoid robotic platform.

Conclusion: PhysioSER provides an interpretable, efficient, and plug-and-play solution for speech emotion recognition that properly incorporates physiological aspects of vocal emotion expression, making it suitable for real-world applications including humanoid robotics.

Abstract: Speech emotion recognition (SER) is essential for humanoid robot tasks such as social robotic interactions and robotic psychological diagnosis, where interpretable and efficient models are critical for safety and performance. Existing deep models trained on large datasets remain largely uninterpretable, often insufficiently modeling underlying emotional acoustic signals and failing to capture and analyze the core physiology of emotional vocal behaviors. Physiological research on human voices shows that the dynamics of vocal amplitude and phase correlate with emotions through the vocal tract filter and the glottal source. However, most existing deep models solely involve amplitude but fail to couple the physiological features of and between amplitude and phase. Here, we propose PhysioSER, a physiology-informed vocal spectrotemporal representation learning method, to address these issues with a compact, plug-and-play design. PhysioSER constructs amplitude and phase views informed by voice anatomy and physiology (VAP) to complement SSL models for SER. This VAP-informed framework incorporates two parallel workflows: a vocal feature representation branch to decompose vocal signals based on VAP, embed them into a quaternion field, and use Hamilton-structured quaternion convolutions for modeling their dynamic interactions; and a latent representation branch based on a frozen SSL backbone. Then, utterance-level features from both workflows are aligned by a Contrastive Projection and Alignment framework, followed by a shallow attention fusion head for SER classification. PhysioSER is shown to be interpretable and efficient for SER through extensive evaluations across 14 datasets, 10 languages, and 6 backbones, and its practical efficacy is validated by real-time deployment on a humanoid robotic platform.

Method: Proposes BreathNet with BreathFiLM (feature-wise linear modulation) that selectively amplifies temporal representations based on breathing sounds, jointly trained with XLS-R extractor. Combines spectral features from frequency front-end and uses feature losses (PSCL, center loss, contrast loss) to enhance discriminative ability.

Result: Achieves state-of-the-art performance: 1.99% average EER across four benchmarks using ASVspoof 2019 LA training set, 4.70% EER on In-the-Wild dataset, and 4.94% EER under ASVspoof5 evaluation protocol.

Conclusion: BreathNet effectively integrates breath information and spectral features to improve audio deepfake detection generalization, demonstrating superior performance across multiple benchmarks.

Abstract: As deepfake audio becomes more realistic and diverse, developing generalizable countermeasure systems has become crucial. Existing detection methods primarily depend on XLS-R front-end features to improve generalization. Nonetheless, their performance remains limited, partly due to insufficient attention to fine-grained information, such as physiological cues or frequency-domain features. In this paper, we propose BreathNet, a novel audio deepfake detection framework that integrates fine-grained breath information to improve generalization. Specifically, we design BreathFiLM, a feature-wise linear modulation mechanism that selectively amplifies temporal representations based on the presence of breathing sounds. BreathFiLM is trained jointly with the XLS-R extractor, in turn encouraging the extractor to learn and encode breath-related cues into the temporal features. Then, we use the frequency front-end to extract spectral features, which are then fused with temporal features to provide complementary information introduced by vocoders or compression artifacts. Additionally, we propose a group of feature losses comprising Positive-only Supervised Contrastive Loss (PSCL), center loss, and contrast loss. These losses jointly enhance the discriminative ability, encouraging the model to separate bona fide and deepfake samples more effectively in the feature space. Extensive experiments on five benchmark datasets demonstrate state-of-the-art (SOTA) performance. Using the ASVspoof 2019 LA training set, our method attains 1.99% average EER across four related eval benchmarks, with particularly strong performance on the In-the-Wild dataset, where it achieves 4.70% EER. Moreover, under the ASVspoof5 evaluation protocol, our method achieves an EER of 4.94% on this latest benchmark.

Method: Proposes AuTAgent (Audio Tool Agent), a reinforcement learning framework that learns when and which tools to invoke. Uses a sparse-feedback training strategy with a novel Differential Reward mechanism to filter out irrelevant tools and invoke external assistance only when it yields net performance gain over the base model.

Result: Improves accuracy by 4.20%/6.20% and 9.80%/8.00% for open-source and closed-source backbones on MMAU Test-mini and MMAR benchmarks. Demonstrates exceptional transferability and shows that AuTAgent complements the representation bottleneck of LALMs by providing verifiable acoustic evidence.

Conclusion: The framework successfully addresses the tool integration challenge in audio language models, highlighting the complementary role of external tools in augmenting audio model reasoning through learned tool invocation strategies.

Abstract: Large Audio Language Models (LALMs) excel at perception but struggle with complex reasoning requiring precise acoustic measurements. While external tools can extract fine-grained features like exact tempo or pitch, effective integration remains challenging: naively using all tools causes information overload, while prompt-based selection fails to assess context-dependent utility. To address this, we propose AuTAgent (Audio Tool Agent), a reinforcement learning framework that learns when and which tools to invoke. By employing a sparse-feedback training strategy with a novel Differential Reward mechanism, the agent learns to filter out irrelevant tools and invokes external assistance only when it yields a net performance gain over the base model. Experimental results confirm that AuTAgent complements the representation bottleneck of LALMs by providing verifiable acoustic evidence. It improves accuracy by 4.20% / 6.20% and 9.80% / 8.00% for open-source and closed-source backbones on the MMAU Test-mini and the MMAR benchmarks, respectively. In addition, further experiments demonstrate exceptional transferability. We highlight the complementary role of external tools in augmenting audio model reasoning.

Method: Organized the first shared task for evaluating CoT reasoning in audio domain using MMAR-Rubrics (instance-level protocol assessing factuality and logic), with Single Model and Agent tracks attracting 156 teams from 18 countries.

Result: Agent systems currently lead in reasoning quality using iterative tool orchestration and cross-modal analysis, while single models are advancing via reinforcement learning and sophisticated data pipelines.

Conclusion: The challenge provides new insights for explainable audio intelligence and establishes benchmarks for evaluating reasoning transparency in audio language models.

Abstract: Recent Large Audio Language Models (LALMs) excel in understanding but often lack transparent reasoning. To address this “black-box” limitation, we organized the Audio Reasoning Challenge at Interspeech 2026, the first shared task dedicated to evaluating Chain-of-Thought (CoT) quality in the audio domain. The challenge introduced MMAR-Rubrics, a novel instance-level protocol assessing the factuality and logic of reasoning chains. Featured Single Model and Agent tracks, the competition attracting 156 teams from 18 countries and regions. Results show agent systems currently lead in reasoning quality, utilizing iterative tool orchestration and cross-modal analysis. Besides, single models are rapidly advancing via reinforcement learning and sophisticated data pipeline. We details the challenge design, methodology, and a comprehensive analysis of state-of-the-art systems, providing new insights for explainable audio intelligence.

Method: Proposes a multidisciplinary collaborative framework involving physicians, psychologists, and engineers working together to address spatial hearing challenges for cochlear implant users through integrated expertise.

Result: The paper presents a conceptual framework for improving spatial hearing in cochlear implants, though specific experimental results are not mentioned in the abstract.

Conclusion: A collaborative multidisciplinary approach is necessary to advance spatial hearing capabilities in cochlear implants, addressing a critical gap in current hearing restoration technology.

Abstract: Cochlear implants (CIs) have been developed to the point where they can restore hearing and speech understanding in a large proportion of patients. Although spatial hearing is central to controlling and directing attention and to enabling speech understanding in noisy environments, it has been largely neglected in the past. We propose here a multi-disciplinary research framework in which physicians, psychologists and engineers collaborate to improve spatial hearing for CI users.

Method: Uses a time-domain Webster model as a physics-informed neural network to estimate vocal-tract area function and radiation coefficient from synthetic audio and F0 trajectory. Training enforces PDE and boundary consistency with lightweight DDSP path for stabilization, while inference is purely physics-based.

Result: The method reproduces spectral envelopes competitively with DDSP baselines on sustained vowels, remains stable under discretization changes, moderate source variations, and ~10% pitch shifts, though produces breathier waveforms than reference.

Conclusion: Physics-informed approach shows promise for interpretable singing-voice synthesis, but needs periodicity-aware objectives and explicit glottal priors to reduce breathiness in future work.

Abstract: We present a physics-informed voiced backend renderer for singing-voice synthesis. Given synthetic single-channel audio and a fund-amental–frequency trajectory, we train a time-domain Webster model as a physics-informed neural network to estimate an interpretable vocal-tract area function and an open-end radiation coefficient. Training enforces partial differential equation and boundary consistency; a lightweight DDSP path is used only to stabilize learning, while inference is purely physics-based. On sustained vowels (/a/, /i/, /u/), parameters rendered by an independent finite-difference time-domain Webster solver reproduce spectral envelopes competitively with a compact DDSP baseline and remain stable under changes in discretization, moderate source variations, and about ten percent pitch shifts. The in-graph waveform remains breathier than the reference, motivating periodicity-aware objectives and explicit glottal priors in future work.

Method: Proposes audiocards - structured metadata grounded in acoustic attributes and sonic descriptors, generated by exploiting the world knowledge of Large Language Models (LLMs). The approach trains models on these structured audiocards rather than single-sentence captions.

Result: Training on audiocards improves downstream text-audio retrieval, descriptive captioning, and metadata generation on professional sound effects libraries. Also improves performance on general audio captioning and retrieval compared to baseline single-sentence captioning approaches.

Conclusion: Audiocards provide effective structured metadata for sound design applications, leveraging LLMs to bridge the gap between audio content and searchable metadata. The approach shows promise for audio-language modeling in professional sound design contexts.

Abstract: Sound designers search for sounds in large sound effects libraries using aspects such as sound class or visual context. However, the metadata needed for such search is often missing or incomplete, and requires significant manual effort to add. Existing solutions to automate this task by generating metadata, i.e. captioning, and search using learned embeddings, i.e. text-audio retrieval, are not trained on metadata with the structure and information pertinent to sound design. To this end we propose audiocards, structured metadata grounded in acoustic attributes and sonic descriptors, by exploiting the world knowledge of LLMs. We show that training on audiocards improves downstream text-audio retrieval, descriptive captioning, and metadata generation on professional sound effects libraries. Moreover, audiocards also improve performance on general audio captioning and retrieval over the baseline single-sentence captioning approach. We release a curated dataset of sound effects audiocards to invite further research in audio language modeling for sound design.

Method: Proposes Generative Speech Reward Model (GSRM) that decomposes speech naturalness evaluation into interpretable acoustic feature extraction followed by feature-grounded chain-of-thought reasoning. Trained on 31k expert ratings and out-of-domain real-world user-assistant speech data.

Result: GSRM substantially outperforms existing speech naturalness predictors, achieving model-human correlation approaching human inter-rater consistency. Also improves speech LLM naturalness when used as verifier for online RLHF.

Conclusion: GSRM provides an effective, interpretable approach to speech naturalness evaluation that can enhance speech generation quality in multimodal language models.

Abstract: Recent advances in speech language models, such as GPT-4o Voice Mode and Gemini Live, have demonstrated promising speech generation capabilities. Nevertheless, the aesthetic naturalness of the synthesized audio still lags behind that of human speech. Enhancing generation quality requires a reliable evaluator of speech naturalness. However, existing naturalness evaluators typically regress raw audio to scalar scores, offering limited interpretability of the evaluation and moreover fail to generalize to speech across different taxonomies. Inspired by recent advances in generative reward modeling, we propose the Generative Speech Reward Model (GSRM), a reasoning-centric reward model tailored for speech. The GSRM is trained to decompose speech naturalness evaluation into an interpretable acoustic feature extraction stage followed by feature-grounded chain-of-thought reasoning, enabling explainable judgments. To achieve this, we curated a large-scale human feedback dataset comprising 31k expert ratings and an out-of-domain benchmark of real-world user-assistant speech interactions. Experiments show that GSRM substantially outperforms existing speech naturalness predictors, achieving model-human correlation of naturalness score prediction that approaches human inter-rater consistency. We further show how GSRM can improve the naturalness of speech LLM generations by serving as an effective verifier for online RLHF.

Method: Extracts layer-wise representations from HuBERT and wav2vec2 models, applies global temporal pooling to sustained vowel recordings, and uses lightweight classifiers (SVM, XGBoost) for classification of four phonation modes (breathy, neutral, flow, pressed).

Result: Foundation model features substantially outperform conventional spectral baselines. HuBERT embeddings from early layers achieve ~95.7% accuracy with SVM, an absolute improvement of ~12-15% over best traditional baseline. Lower layers (retaining acoustic/phonetic detail) outperform top layers specialized for ASR.

Conclusion: Speech foundation models transfer effectively to singing phonation classification, with early-layer embeddings capturing relevant acoustic details. This demonstrates the value of pre-trained audio representations for singing voice analysis tasks.

Abstract: We present voice2mode, a method for classification of four singing phonation modes (breathy, neutral (modal), flow, and pressed) using embeddings extracted from large self-supervised speech models. Prior work on singing phonation has relied on handcrafted signal features or task-specific neural nets; this work evaluates the transferability of speech foundation models to singing phonation classification. voice2mode extracts layer-wise representations from HuBERT and two wav2vec2 variants, applies global temporal pooling, and classifies the pooled embeddings with lightweight classifiers (SVM, XGBoost). Experiments on a publicly available soprano dataset (763 sustained vowel recordings, four labels) show that foundation-model features substantially outperform conventional spectral baselines (spectrogram, mel-spectrogram, MFCC). HuBERT embeddings obtained from early layers yield the best result (~95.7% accuracy with SVM), an absolute improvement of ~12-15% over the best traditional baseline. We also show layer-wise behaviour: lower layers, which retain acoustic/phonetic detail, are more effective than top layers specialized for Automatic Speech Recognition (ASR).

Method: Created two benchmarks: (1) long-form ASR corpus from 11 YouTube channels using reproducible subtitle-extraction pipeline with human-in-the-loop transcript verification; (2) speaker diarization corpus with fully manual speaker-turn labels in CSV format.

Result: Established baselines: Tugstugi achieved 34.07% WER for ASR; pyannote.audio achieved 40.08% DER for diarization. Provided standardized evaluation protocols, annotation rules, and data formats.

Conclusion: Bengali-Loop addresses the resource gap for Bangla long-form speech technology, enabling reproducible benchmarking and future model development for ASR and diarization in realistic multi-speaker scenarios.

Abstract: Bengali (Bangla) remains under-resourced in long-form speech technology despite its wide use. We present Bengali-Loop, two community benchmarks to address this gap: (1) a long-form ASR corpus of 191 recordings (158.6 hours, 792k words) from 11 YouTube channels, collected via a reproducible subtitle-extraction pipeline and human-in-the-loop transcript verification; and (2) a speaker diarization corpus of 24 recordings (22 hours, 5,744 annotated segments) with fully manual speaker-turn labels in CSV format. Both benchmarks target realistic multi-speaker, long-duration content (e.g., Bangla drama/natok). We establish baselines (Tugstugi: 34.07% WER; pyannote.audio: 40.08% DER) and provide standardized evaluation protocols (WER/CER, DER), annotation rules, and data formats to support reproducible benchmarking and future model development for Bangla long-form ASR and diarization.

Method: Uses a unified end-to-end architecture with: 1) lightweight language backbone, 2) Whisper-based audio encoder, 3) sparsely activated Mixture-of-Experts (MoE) adapter to handle audio heterogeneity and reduce cross-modal optimization conflicts. Also introduces DataFlux pipeline for high-quality audio instruction data synthesis and verification.

Result: Achieves competitive performance against models 4-18 times larger across ASR, audio understanding, and dense audio captioning benchmarks. Matches or surpasses 7B to 30B audio and omni-modal baselines despite having only 1.7B parameters.

Conclusion: Eureka-Audio establishes a strong and practical baseline for lightweight audio understanding models, demonstrating efficient balance between computational cost and performance through its unified architecture and data synthesis pipeline.

Abstract: We present Eureka-Audio, a compact yet high-performance audio language model that achieves competitive performance against models that are 4 to 18 times larger across a broad range of audio understanding benchmarks. Despite containing only 1.7B parameters, Eureka-Audio demonstrates strong performance on automatic speech recognition (ASR), audio understanding, and dense audio captioning, matching or surpassing multiple 7B to 30B audio and omni-modal baselines. The model adopts a unified end-to-end architecture composed of a lightweight language backbone, a Whisper-based audio encoder, and a sparsely activated Mixture-of-Experts (MoE) adapter that explicitly accounts for audio heterogeneity and alleviates cross-modal optimization conflicts under limited capacity. To further enhance paralinguistic reasoning, we introduce DataFlux, a closed loop audio instruction data synthesis and verification pipeline that constructs high quality, logically consistent supervision from raw audio. Extensive evaluations across ASR, knowledge reasoning, safety, instruction following, and paralinguistic benchmarks, demonstrate that Eureka-Audio achieves an efficient balance between computational cost and performance. These results establish Eureka Audio as a strong and practical baseline for lightweight audio understanding models.

Method: Proposes MUKA, a multi-kernel adaptation framework that combines representations from instruction-tuning based models (like Pengi) with contrastive pretraining methods (like CLAP). Uses a product kernel to align local similarity with global semantics, preserving kernel method guarantees while avoiding additional training.

Result: Extensive experiments on 11 diverse audio datasets show MUKA achieves state-of-the-art performance among training-free methods and even surpasses training-based adapters in several scenarios.

Conclusion: MUKA offers a compelling balance between adaptability and efficiency for few-shot adaptation of audio-language models, demonstrating that training-free methods can achieve competitive performance through effective combination of different representation types.

Abstract: Multimodal foundation models have demonstrated impressive generalization capabilities, yet efficiently adapting them to new tasks in a few-shot setting remains a critical challenge. In this work, we investigate the few-shot adaptation of Large Audio-Language Models (ALMs) through both training-based and training-free approaches. We introduce MUKA, a multi-kernel adaptation framework that combines the fine-grained, context-dependent representations of instruction-tuning based models like Pengi with the global semantic representations of contrastive pretraining methods like CLAP. By constructing a product kernel that aligns local similarity with global semantics, MUKA enhances representational power while preserving the theoretical guarantees of kernel methods and avoiding additional training. Extensive experiments across 11 diverse audio datasets demonstrate that MUKA achieves state-of-the-art performance among training-free methods and even surpasses training-based adapters in several scenarios, offering a compelling balance between adaptability and efficiency.

Method: The model adopts a source-filter architecture inspired by classical formant synthesis, where a neural vocoder generates the glottal excitation signal, and differentiable resonant filters model the formants to produce the speech waveform.

Result: The proposed HiFi-Glot model generates speech with higher perceptual quality and naturalness while exhibiting more precise control over formant frequencies, outperforming industry-standard formant manipulation tools like Praat.

Conclusion: HiFi-Glot successfully addresses the trade-off between precise formant control and speech naturalness in formant synthesis, achieving both high-fidelity synthesis and precise formant manipulation.

Abstract: Formant synthesis aims to generate speech with controllable formant structures, enabling precise control of vocal resonance and phonetic features. However, while existing formant synthesis approaches enable precise formant manipulation, they often yield an impoverished speech signal by failing to capture the complex co-occurring acoustic cues essential for naturalness. To address this issue, this letter presents HiFi-Glot, an end-to-end neural formant synthesis system that achieves both precise formant control and high-fidelity speech synthesis. Specifically, the proposed model adopts a source–filter architecture inspired by classical formant synthesis, where a neural vocoder generates the glottal excitation signal, and differentiable resonant filters model the formants to produce the speech waveform. Experiment results demonstrate that our proposed HiFi-Glot model can generate speech with higher perceptual quality and naturalness while exhibiting a more precise control over formant frequencies, outperforming industry-standard formant manipulation tools such as Praat. Code, checkpoints, and representative audio samples are available at https://www.yichenggu.com/HiFi-Glot/.

Method: Used recordings of a professional speaker reading text in various styles to isolate expressive/prosodic variation. Compared three modeling approaches: 1) classical acoustic features for speech emotion recognition, 2) self-supervised speech representations, and 3) multimodal large language models (MLLMs).

Result: Self-supervised speech representations outperformed classical acoustic features, especially for complex dynamic impressions like “Cold-Warm”. Current MLLMs proved unreliable for this fine-grained pairwise task. This is the first systematic investigation of RIE.

Conclusion: Self-supervised speech models are effective for capturing subtle perceptual variations in voice impressions, while current MLLMs are not suitable for this fine-grained pairwise comparison task. The study establishes RIE as a valuable framework for voice impression analysis.

Abstract: Paralinguistic and non-linguistic aspects of speech strongly influence listener impressions. While most research focuses on absolute impression scoring, this study investigates relative voice impression estimation (RIE), a framework for predicting the perceptual difference between two utterances from the same speaker. The estimation target is a low-dimensional vector derived from subjective evaluations, quantifying the perceptual shift of the second utterance relative to the first along an antonymic axis (e.g., Dark--Bright''). To isolate expressive and prosodic variation, we used recordings of a professional speaker reading a text in various styles. We compare three modeling approaches: classical acoustic features commonly used for speech emotion recognition, self-supervised speech representations, and multimodal large language models (MLLMs). Our results demonstrate that models using self-supervised representations outperform methods with classical acoustic features, particularly in capturing complex and dynamic impressions (e.g., Cold–Warm’’) where classical features fail. In contrast, current MLLMs prove unreliable for this fine-grained pairwise task. This study provides the first systematic investigation of RIE and demonstrates the strength of self-supervised speech models in capturing subtle perceptual variations.

Method: Proposes VoiceBridge, a one-step latent bridge model using: 1) energy-preserving variational autoencoder for better waveform-latent alignment, 2) single latent-to-latent generative process with scalable transformer, 3) joint neural prior to reduce burden across diverse tasks, and 4) joint training of LBM, decoder and discriminator to transform from denoiser to generator.

Result: Superior performance demonstrated across in-domain tasks (denoising, super-resolution) and out-of-domain tasks (refining synthesized speech) on various datasets, enabling one-step general speech restoration without distillation.

Conclusion: VoiceBridge provides an effective framework for general speech restoration that can handle diverse distortions with a single model, achieving high-quality 48kHz speech reconstruction through innovative latent space modeling and training techniques.

Abstract: Bridge models have been investigated in speech enhancement but are mostly single-task, with constrained general speech restoration (GSR) capability. In this work, we propose VoiceBridge, a one-step latent bridge model (LBM) for GSR, capable of efficiently reconstructing 48 kHz fullband speech from diverse distortions. To inherit the advantages of data-domain bridge models, we design an energy-preserving variational autoencoder, enhancing the waveform-latent space alignment over varying energy levels. By compressing waveform into continuous latent representations, VoiceBridge models~\textit{various} GSR tasks with a~\textit{single} latent-to-latent generative process backed by a scalable transformer. To alleviate the challenge of reconstructing the high-quality target from distinctively different low-quality priors, we propose a joint neural prior for GSR, uniformly reducing the burden of the LBM in diverse tasks. Building upon these designs, we further investigate bridge training objective by jointly tuning LBM, decoder and discriminator together, transforming the model from a denoiser to generator and enabling \textit{one-step GSR without distillation}. Extensive validation across in-domain (\textit{e.g.}, denoising and super-resolution) and out-of-domain tasks (\textit{e.g.}, refining synthesized speech) and datasets demonstrates the superior performance of VoiceBridge. Demos: https://VoiceBridgedemo.github.io/.

Method: Tested two e2e-TTS architectures (Tacotron-2 autoregressive and VITS-TTS non-autoregressive) with three configurations: (a) forward text + forward speech (conventional), (b) reverse text + reverse speech, and (c) reverse text + forward speech.

Result: Reversed text and speech TTS systems (r-e2e-TTS) generated speech with better fidelity, perceptual intelligibility, and naturalness than conventional forward configurations, demonstrating e2e-TTS systems are purely data-driven.

Conclusion: E2e-TTS systems learn from data patterns without inherent anatomical constraints; reversing sequences can improve speech quality, suggesting optimization opportunities in training configurations.

Abstract: An end-to-end (e2e) text-to-speech (TTS) system is a deep architecture that learns to associate a text string with acoustic speech patterns from a curated dataset. It is expected that all aspects associated with speech production, such as phone duration, speaker characteristics, and intonation among other things are captured in the trained TTS model to enable the synthesized speech to be natural and intelligible. Human speech is complex, involving smooth transitions between articulatory configurations (ACs). Due to anatomical constraints, some ACs are challenging to mimic or transition between. In this paper, we experimentally study if the constraints imposed by human anatomy have an implication on training an e2e-TTS systems. We experiment with two e2e-TTS architectures, namely, Tacotron-2 an autoregressive model and VITS-TTS a non-autoregressive model. In this study, we build TTS systems using (a) forward text, forward speech (conventional, e2e-TTS), (b) reverse text, reverse speech (r-e2e-TTS), and (c) reverse text, forward speech (rtfs-e2e-TTS). Experiments demonstrate that e2e-TTS systems are purely data-driven. Interestingly, the generated speech by r-e2e-TTS systems exhibits better fidelity, better perceptual intelligibility, and better naturalness

Method: Proposes Robust Reward Policy Optimization (RRPO) with hybrid regularization scheme to develop robust reward models whose signals are reliably aligned with human perception, preventing policy models from taking detrimental shortcuts.

Result: Ablation study confirms enhanced robustness of the reward model with strong cross-lingual generalization. Subjective evaluation shows RRPO effectively mitigates reward hacking, leading to significant improvements in both emotional expressiveness and naturalness over all baselines.

Conclusion: RRPO successfully addresses reward hacking in differentiable RL for controllable TTS, enabling more reliable emotion control while maintaining high perceptual quality through robust reward modeling.

Abstract: Differentiable reinforcement learning (RL) frameworks like DiffRO offer a powerful approach for controllable text-to-speech (TTS), but are vulnerable to reward hacking, particularly for nuanced tasks like emotion control. The policy model can exploit a vanilla Reward Model (RM) by generating acoustic artifacts to achieve spurious rewards, but at the cost of degrading perceptual quality. To address this, we propose Robust Reward Policy Optimization (RRPO), a novel framework that employs a hybrid regularization scheme. This scheme develops a robust RM whose reward signal is more reliably aligned with human perception, compelling the policy to abandon detrimental shortcuts and instead learn the complex features of genuine emotions. Our ablation study confirms the enhanced robustness of our RM, as evidenced by its strong cross-lingual generalization. The subjective evaluation demonstrates that this robust RM effectively mitigates reward hacking, leading to significant improvements in both emotional expressiveness and naturalness over all baselines. Demo page: https://lrwinr.github.io/RRPO-CosyVoice.

Method: Probing-based framework beyond standard downstream tasks to evaluate disentangled representations in music audio models. Analyzes diverse unsupervised disentanglement strategies (inductive biases, data augmentations, adversarial objectives, staged training) across four axes: informativeness, equivariance, invariance, and disentanglement.

Result: Findings reveal inconsistencies between intended and actual semantics of embeddings, suggesting current strategies fall short of producing truly disentangled representations.

Conclusion: Current disentanglement approaches in music generation are insufficient, prompting re-examination of how controllability is approached in the field.

Abstract: Recent approaches in music generation rely on disentangled representations, often labeled as structure and timbre or local and global, to enable controllable synthesis. Yet the underlying properties of these embeddings remain underexplored. In this work, we evaluate such disentangled representations in a set of music audio models for controllable generation using a probing-based framework that goes beyond standard downstream tasks. The selected models reflect diverse unsupervised disentanglement strategies, including inductive biases, data augmentations, adversarial objectives, and staged training procedures. We further isolate specific strategies to analyze their effect. Our analysis spans four key axes: informativeness, equivariance, invariance, and disentanglement, which are assessed across datasets, tasks, and controlled transformations. Our findings reveal inconsistencies between intended and actual semantics of the embeddings, suggesting that current strategies fall short of producing truly disentangled representations, and prompting a re-examination of how controllability is approached in music generation.

Method: Develops a reference-free ASR Inconsistency Score that doesn’t require ground truth transcripts, using automatic speech recognition inconsistencies as a metric for speech quality assessment.

Result: The ASR Inconsistency Score achieves high correlation with expert perceptual ratings, performing comparably to or better than standard reference-based Word Error Rate baselines across Dutch, Spanish, and English pathological speech.

Conclusion: The proposed reference-free, explainable method provides a practical alternative to traditional reference-based approaches for objective speech assessment in clinical and research settings.

Abstract: Objective assessment of speech that reflects meaningful changes in communication is crucial for clinical decision making and reproducible research. While existing objective assessments, particularly reference-based approaches, can capture intelligibility changes, they are often hindered by lack of explainability and the need for labor-intensive manual transcriptions. To address these issues, this work proposes the reference-free, explainable ASR Inconsistency Score. We evaluate this method on pathological speech in Dutch, Spanish and English, and compare its performance to a reference-based Word Error Rate (WER) baseline. Our results demonstrate that the ASR Inconsistency Score achieves a high correlation with expert perceptual ratings, with performance closely matching, and in one case exceeding, a standard reference-based Word Error Rate (WER) baseline.

Method: Proposes Directional Concentration Uncertainty (DCU), a statistical procedure based on von Mises-Fisher distribution that quantifies embedding concentration by measuring geometric dispersion of multiple generated outputs using continuous embeddings without task-specific heuristics.

Result: DCU matches or exceeds calibration levels of prior methods like semantic entropy and generalizes well to more complex tasks in multimodal domains.

Conclusion: DCU provides a flexible uncertainty quantification framework with strong performance and generalization capabilities, showing potential for integration into multimodal and agentic systems.

Abstract: In the critical task of making generative models trustworthy and robust, methods for Uncertainty Quantification (UQ) have begun to show encouraging potential. However, many of these methods rely on rigid heuristics that fail to generalize across tasks and modalities. Here, we propose a novel framework for UQ that is highly flexible and approaches or surpasses the performance of prior heuristic methods. We introduce Directional Concentration Uncertainty (DCU), a novel statistical procedure for quantifying the concentration of embeddings based on the von Mises-Fisher (vMF) distribution. Our method captures uncertainty by measuring the geometric dispersion of multiple generated outputs from a language model using continuous embeddings of the generated outputs without any task specific heuristics. In our experiments, we show that DCU matches or exceeds calibration levels of prior works like semantic entropy (Kuhn et al., 2023) and also generalizes well to more complex tasks in multi-modal domains. We present a framework for the wider potential of DCU and its implications for integration into UQ for multi-modal and agentic frameworks.

Method: System detects canonical drawing regions, applies region-restricted VLM-based OCR, normalizes identifiers, and fuses lexical and dense retrieval with lightweight region-level reranking.

Result: Deployed on ~770k unlabeled files, achieves 10.1% absolute gain in Success@3 and 18.9% relative improvement in nDCG@3 over strongest vision-language baseline on 5k-file benchmark with 350 expert queries.

Conclusion: Blueprint demonstrates effective multimodal retrieval for engineering archives, with substantial headroom for improvement under perfect region detection and OCR. System enables cross-facility search of legacy repositories.

Abstract: Decades of engineering drawings and technical records remain locked in legacy archives with inconsistent or missing metadata, making retrieval difficult and often manual. We present Blueprint, a layout-aware multimodal retrieval system designed for large-scale engineering repositories. Blueprint detects canonical drawing regions, applies region-restricted VLM-based OCR, normalizes identifiers (e.g., DWG, part, facility), and fuses lexical and dense retrieval with a lightweight region-level reranker. Deployed on ~770k unlabeled files, it automatically produces structured metadata suitable for cross-facility search. We evaluate Blueprint on a 5k-file benchmark with 350 expert-curated queries using pooled, graded (0/1/2) relevance judgments. Blueprint delivers a 10.1% absolute gain in Success@3 and an 18.9% relative improvement in nDCG@3 over the strongest vision-language baseline}, consistently outperforming across vision, text, and multimodal intents. Oracle ablations reveal substantial headroom under perfect region detection and OCR. We release all queries, runs, annotations, and code to facilitate reproducible evaluation on legacy engineering archives.

Method: Benchmark evaluation of Residual Networks (ResNet), Temporal Convolutional Networks (TCN), and Dilated Convolutional Neural Networks (DCNN) on the MNIST-1D dataset. These models were compared against previously tested baselines including logistic regression, MLPs, CNNs, and GRUs to assess performance on sequential data.

Result: Advanced architectures like TCN and DCNN consistently outperform simpler models, achieving near-human performance on MNIST-1D. ResNet also shows significant improvements. The results demonstrate the importance of inductive biases and hierarchical feature extraction in small structured datasets.

Conclusion: MNIST-1D serves as a robust benchmark for evaluating machine learning architectures under computational constraints. Architectural innovations play a crucial role in improving model performance, offering insights for optimizing deep learning models in resource-limited environments.

Abstract: Small datasets like MNIST have historically been instrumental in advancing machine learning research by providing a controlled environment for rapid experimentation and model evaluation. However, their simplicity often limits their utility for distinguishing between advanced neural network architectures. To address these challenges, Greydanus et al. introduced the MNIST-1D dataset, a one-dimensional adaptation of MNIST designed to explore inductive biases in sequential data. This dataset maintains the advantages of small-scale datasets while introducing variability and complexity that make it ideal for studying advanced architectures. In this paper, we extend the exploration of MNIST-1D by evaluating the performance of Residual Networks (ResNet), Temporal Convolutional Networks (TCN), and Dilated Convolutional Neural Networks (DCNN). These models, known for their ability to capture sequential patterns and hierarchical features, were implemented and benchmarked alongside previously tested architectures such as logistic regression, MLPs, CNNs, and GRUs. Our experimental results demonstrate that advanced architectures like TCN and DCNN consistently outperform simpler models, achieving near-human performance on MNIST-1D. ResNet also shows significant improvements, highlighting the importance of leveraging inductive biases and hierarchical feature extraction in small structured datasets. Through this study, we validate the utility of MNIST-1D as a robust benchmark for evaluating machine learning architectures under computational constraints. Our findings emphasize the role of architectural innovations in improving model performance and offer insights into optimizing deep learning models for resource-limited environments.

Method: Introduces speed-up factor as a quantitative multi-iteration performance metric. Evaluates using 4 diverse datasets and 7 query methods, comparing with state-of-the-art AL performance metrics.

Result: Speed-up factor accurately captures the fraction of samples needed to match random sampling performance and shows superior stability across iterations compared to existing metrics.

Conclusion: Speed-up factor is a reliable metric for evaluating active learning query methods, addressing the gap in multi-iteration performance assessment.

Abstract: Machine learning models excel with abundant annotated data, but annotation is often costly and time-intensive. Active learning (AL) aims to improve the performance-to-annotation ratio by using query methods (QMs) to iteratively select the most informative samples. While AL research focuses mainly on QM development, the evaluation of this iterative process lacks appropriate performance metrics. This work reviews eight years of AL evaluation literature and formally introduces the speed-up factor, a quantitative multi-iteration QM performance metric that indicates the fraction of samples needed to match random sampling performance. Using four datasets from diverse domains and seven QMs of various types, we empirically evaluate the speed-up factor and compare it with state-of-the-art AL performance metrics. The results confirm the assumptions underlying the speed-up factor, demonstrate its accuracy in capturing the described fraction, and reveal its superior stability across iterations.

Method: Combines high-throughput screening with an active-learning loop based on multi-objective Bayesian optimization. Trains probabilistic surrogate models to predict concentration and viability, quantifies uncertainty, and iteratively selects experiments by prioritizing cocktails expected to improve the Pareto front using expected Pareto improvement under uncertainty.

Result: Wet-lab validation shows efficient discovery of cocktails achieving high CPA concentrations and high post-exposure viability. Improves dominated hypervolume by 9.5% and 4.5% relative to naive strategy and strong baseline, while reducing experiments needed. Synthetic studies recover comparably strong Pareto-optimal solutions using only 30% of evaluations required by prior state-of-the-art.

Conclusion: The framework accelerates cryopreservation development by efficiently navigating the multi-objective design space of CPA cocktails, with potential adaptation to different CPA libraries, objective definitions, and cell lines.

Abstract: Designing cryoprotectant agent (CPA) cocktails for vitrification is challenging because formulations must be concentrated enough to suppress ice formation yet non-toxic enough to preserve cell viability. This tradeoff creates a large, multi-objective design space in which traditional discovery is slow, often relying on expert intuition or exhaustive experimentation. We present a data-efficient framework that accelerates CPA cocktail design by combining high-throughput screening with an active-learning loop based on multi-objective Bayesian optimization. From an initial set of measured cocktails, we train probabilistic surrogate models to predict concentration and viability and quantify uncertainty across candidate formulations. We then iteratively select the next experiments by prioritizing cocktails expected to improve the Pareto front, maximizing expected Pareto improvement under uncertainty, and update the models as new assay results are collected. Wet-lab validation shows that our approach efficiently discovers cocktails that simultaneously achieve high CPA concentrations and high post-exposure viability. Relative to a naive strategy and a strong baseline, our method improves dominated hypervolume by 9.5% and 4.5%, respectively, while reducing the number of experiments needed to reach high-quality solutions. In complementary synthetic studies, it recovers a comparably strong set of Pareto-optimal solutions using only 30% of the evaluations required by the prior state-of-the-art multi-objective approach, which amounts to saving approximately 10 weeks of experimental time. Because the framework assumes only a suitable assay and defined formulation space, it can be adapted to different CPA libraries, objective definitions, and cell lines to accelerate cryopreservation development.

Method: Develops worst-case complexity theory for stochastically preconditioned SGD (SPSGD) and accelerated variants under heavy-tailed noise with finite p-th moments. Analyzes normalization and clipping as stabilization techniques, develops novel vector-valued Burkholder-type inequality for the analysis.

Result: Normalization guarantees convergence to first-order stationary points at optimal rates (O(T^{-(p-1)/(3p-2)}) with known parameters, O(T^{-(p-1)/(2p)}) with unknown parameters), while clipping may fail to converge in worst-case scenarios due to statistical dependence between preconditioner and gradient estimates.

Conclusion: Normalization is theoretically superior to clipping for stabilizing stochastically preconditioned methods under heavy-tailed noise, explaining empirical preference for normalization in large-scale model training.

Abstract: We develop a worst-case complexity theory for stochastically preconditioned stochastic gradient descent (SPSGD) and its accelerated variants under heavy-tailed noise, a setting that encompasses widely used adaptive methods such as Adam, RMSProp, and Shampoo. We assume the stochastic gradient noise has a finite $p$-th moment for some $p \in (1,2]$, and measure convergence after $T$ iterations. While clipping and normalization are parallel tools for stabilizing training of SGD under heavy-tailed noise, there is a fundamental separation in their worst-case properties in stochastically preconditioned settings. We demonstrate that normalization guarantees convergence to a first-order stationary point at rate $\mathcal{O}(T^{-\frac{p-1}{3p-2}})$ when problem parameters are known, and $\mathcal{O}(T^{-\frac{p-1}{2p}})$ when problem parameters are unknown, matching the optimal rates for normalized SGD, respectively. In contrast, we prove that clipping may fail to converge in the worst case due to the statistical dependence between the stochastic preconditioner and the gradient estimates. To enable the analysis, we develop a novel vector-valued Burkholder-type inequality that may be of independent interest. These results provide a theoretical explanation for the empirical preference for normalization over clipping in large-scale model training.

Method: Proposes evaluating candidate subsets through minimum mean-squared error of linear regression predicting target vectors (labels or input reconstruction). Uses matrix inversion lemma to derive efficient formula for objective change when swapping selected/unselected elements, enabling swap-based local search optimization.

Result: Experiments on MNIST handwritten-digit images demonstrate the effectiveness of the proposed element selection method for dimension reduction.

Conclusion: The proposed algorithm provides an efficient, multiplication-free approach to dimension reduction through element selection, particularly suitable for resource-constrained systems where matrix multiplications are prohibitive.

Abstract: In this paper, we propose a fast algorithm for element selection, a multiplication-free form of dimension reduction that produces a dimension-reduced vector by simply selecting a subset of elements from the input. Dimension reduction is a fundamental technique for reducing unnecessary model parameters, mitigating overfitting, and accelerating training and inference. A standard approach is principal component analysis (PCA), but PCA relies on matrix multiplications; on resource-constrained systems, the multiplication count itself can become a bottleneck. Element selection eliminates this cost because the reduction consists only of selecting elements, and thus the key challenge is to determine which elements should be retained. We evaluate a candidate subset through the minimum mean-squared error of linear regression that predicts a target vector from the selected elements, where the target may be, for example, a one-hot label vector in classification. When an explicit target is unavailable, the input itself can be used as the target, yielding a reconstruction-based criterion. The resulting optimization is combinatorial, and exhaustive search is impractical. To address this, we derive an efficient formula for the objective change caused by swapping a selected and an unselected element, using the matrix inversion lemma, and we perform a swap-based local search that repeatedly applies objective-decreasing swaps until no further improvement is possible. Experiments on MNIST handwritten-digit images demonstrate the effectiveness of the proposed method.

Method: Probabilistic diffusion-based generative models with conditional and posterior sampling frameworks to downscale coarse LUCIE outputs from ~300km to 25km resolution, trained on ERA5 data (2000-2009).

Result: The approach successfully preserves coarse-grained dynamics from LUCIE while generating fine-scaled climatological statistics at ~28km resolution, validated on 2010-2020 predictions using metrics like RMSE, power spectrum, and probability density functions.

Conclusion: The diffusion-based downscaling framework effectively bridges the resolution gap for climate emulators, enabling detailed regional climate impact assessments while maintaining physical consistency.

Abstract: The proliferation of data-driven models in weather and climate sciences has marked a significant paradigm shift, with advanced models demonstrating exceptional skill in medium-range forecasting. However, these models are often limited by long-term instabilities, climatological drift, and substantial computational costs during training and inference, restricting their broader application for climate studies. Addressing these limitations, Guan et al. (2024) introduced LUCIE, a lightweight, physically consistent climate emulator utilizing a Spherical Fourier Neural Operator (SFNO) architecture. This model is able to reproduce accurate long-term statistics including climatological mean and seasonal variability. However, LUCIE’s native resolution (~300 km) is inadequate for detailed regional impact assessments. To overcome this limitation, we introduce a deep learning-based downscaling framework, leveraging probabilistic diffusion-based generative models with conditional and posterior sampling frameworks. These models downscale coarse LUCIE outputs to 25 km resolution. They are trained on approximately 14,000 ERA5 timesteps spanning 2000-2009 and evaluated on LUCIE predictions from 2010 to 2020. Model performance is assessed through diverse metrics, including latitude-averaged RMSE, power spectrum, probability density functions and First Empirical Orthogonal Function of the zonal wind. We observe that the proposed approach is able to preserve the coarse-grained dynamics from LUCIE while generating fine-scaled climatological statistics at ~28km resolution.

Method: Proposes Texture - a word-level discrete curvature signal defined by reconciling left- and right-context beliefs around masked words using a Schrodinger bridge. This yields a curvature field where positive values indicate focused meaning and negative values indicate competing continuations.

Result: Provides empirical and theoretical evidence that semantic inference in natural corpora is non-flat (language has inherent curvature). Demonstrates Texture’s utility on two tasks: improving long-context inference through curvature-guided compression and enhancing retrieval-augmented generation through curvature-guided routing.

Conclusion: Text does have intrinsic curvature, and Texture provides a text-native way to measure and utilize it without geometric training, establishing a new curvature paradigm for language that is both measurable and practically useful.

Abstract: Does text have an intrinsic curvature? Language is increasingly modeled in curved geometries - hyperbolic spaces for hierarchy, mixed-curvature manifolds for compositional structure - yet a basic scientific question remains unresolved: what does curvature mean for text itself, in a way that is native to language rather than an artifact of the embedding space we choose? We argue that text does indeed have curvature, and show how to detect it, define it, and use it. To this end, we propose Texture, a text-native, word-level discrete curvature signal, and make three contributions. (a) Existence: We provide empirical and theoretical certificates that semantic inference in natural corpora is non-flat, i.e. language has inherent curvature. (b) Definition: We define Texture by reconciling left- and right-context beliefs around a masked word through a Schrodinger bridge, yielding a curvature field that is positive where context focuses meaning and negative where it fans out into competing continuations. (c) Utility: Texture is actionable: it serves as a general-purpose measurement and control primitive enabling geometry without geometric training; we instantiate it on two representative tasks, improving long-context inference through curvature-guided compression and retrieval-augmented generation through curvature-guided routing. Together, our results establish a text-native curvature paradigm, making curvature measurable and practically useful.

Method: Tutorial on using the PyCM library for deep-dive evaluation of multi-class classifiers, demonstrated through two case scenarios showing how different evaluation metrics can lead to different interpretations of model efficacy.

Result: The PyCM library enables more comprehensive model evaluation, revealing that choice of evaluation metrics fundamentally shifts interpretation of model performance, and that multi-dimensional evaluation is essential for uncovering important subtle differences.

Conclusion: A multi-dimensional evaluation framework using tools like PyCM is crucial for optimal model selection, as standard metrics may miss subtle but important performance trade-offs in multi-class classifiers.

Abstract: Selecting an optimal classification model requires a robust and comprehensive understanding of the performance of the model. This paper provides a tutorial on the PyCM library, demonstrating its utility in conducting deep-dive evaluations of multi-class classifiers. By examining two different case scenarios, we illustrate how the choice of evaluation metrics can fundamentally shift the interpretation of a model’s efficacy. Our findings emphasize that a multi-dimensional evaluation framework is essential for uncovering small but important differences in model performance. However, standard metrics may miss these subtle performance trade-offs.

Method: Builds on attention causal communication (ACC) with refinements (ACC++) to extract cleaner, lower-dimensional causal signals from attention heads in single forward passes, without needing replacement models or activation patching.

Result: Found no single circuit for indirect object identification (IOI) in GPT-2, Pythia, or Gemma 2 - different prompt templates induce systematically different mechanisms. Prompts cluster into families with similar circuits.

Conclusion: Circuits should be studied at prompt level rather than task level. ACC++ enables scalable circuit descriptions and automated interpretability pipelines for understanding prompt-specific mechanisms.

Abstract: Understanding the internal circuits that language models use to solve tasks remains a central challenge in mechanistic interpretability. Most prior work identifies circuits at the task level by averaging across many prompts, implicitly assuming a single stable mechanism per task. We show that this assumption can obscure a crucial source of structure: circuits are prompt-specific, even within a fixed task. Building on attention causal communication (ACC) (Franco & Crovella, 2025), we introduce ACC++, refinements that extract cleaner, lower-dimensional causal signals inside attention heads from a single forward pass. Like ACC, our approach does not require replacement models (e.g., SAEs) or activation patching; ACC++ further improves circuit precision by reducing attribution noise. Applying ACC++ to indirect object identification (IOI) in GPT-2, Pythia, and Gemma 2, we find there is no single circuit for IOI in any model: different prompt templates induce systematically different mechanisms. Despite this variation, prompts cluster into prompt families with similar circuits, and we propose a representative circuit for each family as a practical unit of analysis. Finally, we develop an automated interpretability pipeline that uses ACC++ signals to surface human-interpretable features and assemble mechanistic explanations for prompt families behavior. Together, our results recast circuits as a meaningful object of study by shifting the unit of analysis from tasks to prompts, enabling scalable circuit descriptions in the presence of prompt-specific mechanisms.

Method: Each client uses a nonlinear state space model to map local observations to low-dimensional latent states. A central server learns a graph-structured neural state transition model over communicated latent states using Graph Attention Networks (GAT). Jacobian analysis relates to attention coefficients for interpretability.

Result: Theoretical convergence guarantees to a centralized oracle. Synthetic experiments demonstrate convergence, interpretability, scalability, and privacy. Real-world experiments show performance comparable to decentralized baselines.

Conclusion: Proposed federated framework successfully learns interpretable cross-client temporal interdependencies in decentralized nonlinear systems while respecting privacy constraints and fixed proprietary models.

Abstract: Networks of modern industrial systems are increasingly monitored by distributed sensors, where each system comprises multiple subsystems generating high dimensional time series data. These subsystems are often interdependent, making it important to understand how temporal patterns at one subsystem relate to others. This is challenging in decentralized settings where raw measurements cannot be shared and client observations are heterogeneous. In practical deployments each subsystem (client) operates a fixed proprietary model that cannot be modified or retrained, limiting existing approaches. Nonlinear dynamics further make cross client temporal interdependencies difficult to interpret because they are embedded in nonlinear state transition functions. We present a federated framework for learning temporal interdependencies across clients under these constraints. Each client maps high dimensional local observations to low dimensional latent states using a nonlinear state space model. A central server learns a graph structured neural state transition model over the communicated latent states using a Graph Attention Network. For interpretability we relate the Jacobian of the learned server side transition model to attention coefficients, providing the first interpretable characterization of cross client temporal interdependencies in decentralized nonlinear systems. We establish theoretical convergence guarantees to a centralized oracle and validate the framework through synthetic experiments demonstrating convergence, interpretability, scalability and privacy. Additional real world experiments show performance comparable to decentralized baselines.

Method: Proposes raFLoRA, a rank-partitioned aggregation method that decomposes local updates into rank partitions and aggregates each partition weighted by its effective client contributions, addressing the mismatch between rank-agnostic aggregation weights and rank-dependent client contributions.

Result: Extensive experiments across classification and reasoning tasks show that raFLoRA prevents rank collapse, improves model performance, and preserves communication efficiency compared to state-of-the-art FedLoRA baselines.

Conclusion: raFLoRA effectively addresses the rank collapse problem in heterogeneous federated LoRA through proper rank-partitioned aggregation, enabling better performance while maintaining communication efficiency in federated learning settings.

Abstract: Federated low-rank adaptation (FedLoRA) has facilitated communication-efficient and privacy-preserving fine-tuning of foundation models for downstream tasks. In practical federated learning scenarios, client heterogeneity in system resources and data distributions motivates heterogeneous LoRA ranks across clients. We identify a previously overlooked phenomenon in heterogeneous FedLoRA, termed rank collapse, where the energy of the global update concentrates on the minimum shared rank, resulting in suboptimal performance and high sensitivity to rank configurations. Through theoretical analysis, we reveal the root cause of rank collapse: a mismatch between rank-agnostic aggregation weights and rank-dependent client contributions, which systematically suppresses higher-rank updates at a geometric rate over rounds. Motivated by this insight, we propose raFLoRA, a rank-partitioned aggregation method that decomposes local updates into rank partitions and then aggregates each partition weighted by its effective client contributions. Extensive experiments across classification and reasoning tasks show that raFLoRA prevents rank collapse, improves model performance, and preserves communication efficiency compared to state-of-the-art FedLoRA baselines.

Method: TrasMuon introduces two key components: (1) global RMS calibration to reintroduce adaptive scaling, and (2) energy-based trust-region clipping that defines a trust region based on relative energy ratios to confine updates to a stable zone and prevent instability from high-energy outliers.

Result: Empirical experiments on vision and language models show that TrasMuon converges faster than baseline methods. Additional experiments without warmup stages confirm TrasMuon’s superior stability and robustness compared to other optimizers.

Conclusion: TrasMuon successfully addresses the instability issues of Muon-style optimizers while maintaining their geometric advantages, resulting in a more stable and efficient optimization method suitable for training vision and language models.

Abstract: Muon-style optimizers leverage Newton-Schulz (NS) iterations to orthogonalize updates, yielding update geometries that often outperform Adam-series methods. However, this orthogonalization discards magnitude information, rendering training sensitive to step-size hyperparameters and vulnerable to high-energy bursts. To mitigate this, we introduce TrasMuon (\textbf{T}rust \textbf{R}egion \textbf{A}daptive \textbf{S}caling \textbf{Muon}). TrasMuon preserves the near-isometric geometry of Muon while stabilizing magnitudes through (i) global RMS calibration and (ii) energy-based trust-region clipping. We demonstrate that while reintroducing adaptive scaling improves optimization efficiency, it typically exacerbates instability due to high-energy outliers. TrasMuon addresses this by defining a trust region based on relative energy ratios, confining updates to a stable zone. Empirical experiments on vision and language models demonstrate that TrasMuon converges faster than baselines. Furthermore, experiments without warmup stages confirm TrasMuon’s superior stability and robustness.

Method: Uses pseudo-calibration with domain adaptation tools, derives lower bound on target coverage using classifier loss and Wasserstein shift measure, designs pseudo-calibrated sets with slack parameter, and proposes source-tuned pseudo-calibration algorithm that interpolates between hard pseudo-labels and randomized labels based on classifier uncertainty.

Result: Theoretical bounds track pseudo-calibration behavior qualitatively, and source-tuned scheme mitigates coverage degradation under distribution shift while maintaining nontrivial prediction set sizes in numerical experiments.

Conclusion: Pseudo-calibration with domain adaptation tools can effectively maintain conformal prediction coverage guarantees under bounded label-conditional covariate shift.

Abstract: Conformal prediction (CP) offers distribution-free marginal coverage guarantees under an exchangeability assumption, but these guarantees can fail if the data distribution shifts. We analyze the use of pseudo-calibration as a tool to counter this performance loss under a bounded label-conditional covariate shift model. Using tools from domain adaptation, we derive a lower bound on target coverage in terms of the source-domain loss of the classifier and a Wasserstein measure of the shift. Using this result, we provide a method to design pseudo-calibrated sets that inflate the conformal threshold by a slack parameter to keep target coverage above a prescribed level. Finally, we propose a source-tuned pseudo-calibration algorithm that interpolates between hard pseudo-labels and randomized labels as a function of classifier uncertainty. Numerical experiments show that our bounds qualitatively track pseudo-calibration behavior and that the source-tuned scheme mitigates coverage degradation under distribution shift while maintaining nontrivial prediction set sizes.

Method: Introduces γ-weakly θ-up-concavity, a novel first-order condition that characterizes a broad class of monotone non-convex functions. The key theoretical contribution shows these functions are upper-linearizable: for any feasible point, a linear surrogate can be constructed whose gains approximate the original objective up to a constant factor (approximation coefficient) dependent on γ, θ, and feasible set geometry.

Result: Provides unified approximation guarantees for offline optimization, static and dynamic regret bounds in online settings via reductions to linear optimization. Recovers optimal approximation coefficient for DR-submodular maximization and significantly improves existing approximation coefficients for OSS optimization, particularly over matroid constraints.

Conclusion: The γ-weakly θ-up-concavity framework offers a powerful unifying approach for monotone non-convex optimization, generalizing existing frameworks while providing strong theoretical guarantees and improved approximation coefficients across various constraint settings.

Abstract: Optimizing monotone non-convex functions is a fundamental challenge across machine learning and combinatorial optimization. We introduce and study $γ$-weakly $θ$-up-concavity, a novel first-order condition that characterizes a broad class of such functions. This condition provides a powerful unifying framework, strictly generalizing both DR-submodular functions and One-Sided Smooth (OSS) functions. Our central theoretical contribution demonstrates that $γ$-weakly $θ$-up-concave functions are upper-linearizable: for any feasible point, we can construct a linear surrogate whose gains provably approximate the original non-linear objective. This approximation holds up to a constant factor, namely the approximation coefficient, dependent solely on $γ$, $θ$, and the geometry of the feasible set. This linearizability yields immediate and unified approximation guarantees for a wide range of problems. Specifically, we obtain unified approximation guarantees for offline optimization as well as static and dynamic regret bounds in online settings via standard reductions to linear optimization. Moreover, our framework recovers the optimal approximation coefficient for DR-submodular maximization and significantly improves existing approximation coefficients for OSS optimization, particularly over matroid constraints.

Method: 1) Demonstrate alignment in a model where features are directly observable; 2) Provide theoretical analysis showing when alignment is expected; 3) Propose sparse attention decomposition as a testable prediction for recognizing alignment in real models where features are not directly observable; 4) Show empirical evidence of this pattern in real models.

Result: The paper shows that singular vectors robustly align with features in observable models, provides theoretical conditions for such alignment, and demonstrates that sparse attention decomposition emerges in real models consistent with predictions, suggesting alignment can be a sound basis for feature identification.

Conclusion: Alignment of singular vectors with features can be a theoretically justified approach for feature identification in language models, with sparse attention decomposition serving as an operational test for recognizing such alignment in practice.

Abstract: Identifying feature representations in language models is a central task in mechanistic interpretability. Several recent studies have made an implicit assumption that feature representations can be inferred in some cases from singular vectors of attention matrices. However, sound justification for this assumption is lacking. In this paper we address that question, asking: why and when do singular vectors align with features? First, we demonstrate that singular vectors robustly align with features in a model where features can be directly observed. We then show theoretically that such alignment is expected under a range of conditions. We close by asking how, operationally, alignment may be recognized in real models where feature representations are not directly observable. We identify sparse attention decomposition as a testable prediction of alignment, and show evidence that it emerges in a manner consistent with predictions in real models. Together these results suggest that alignment of singular vectors with features can be a sound and theoretically justified basis for feature identification in language models.

Method: Proposed a multi-variable Vision Transformer (ViT) architecture with 1EMD design: one shared encoder and multiple variable-specific decoders to jointly predict six climate variables from GCM-resolution inputs.

Result: The 1EMD ViT outperforms single-variable ViT models with ~5.5% average MSE reduction, beats other multi-variable baselines, and reduces inference time per variable by 29-32% compared to single-variable approaches.

Conclusion: Multi-variable modeling provides systematic advantages for climate downscaling in accuracy and efficiency, with the 1EMD ViT achieving the best trade-off between predictive performance and computational cost.

Abstract: Global Climate Models (GCMs) are critical for simulating large-scale climate dynamics, but their coarse spatial resolution limits their applicability in regional studies. Regional Climate Models (RCMs) address this limitation through dynamical downscaling, albeit at considerable computational cost and with limited flexibility. Deep learning has emerged as an efficient data-driven alternative; however, most existing approaches focus on single-variable models that downscale one variable at a time. This paradigm can lead to redundant computation, limited contextual awareness, and weak cross-variable interactions.To address these limitations, we propose a multi-variable Vision Transformer (ViT) architecture with a shared encoder and variable-specific decoders (1EMD). The proposed model jointly predicts six key climate variables: surface temperature, wind speed, 500 hPa geopotential height, total precipitation, surface downwelling shortwave radiation, and surface downwelling longwave radiation, directly from GCM-resolution inputs, emulating RCM-scale downscaling over Europe. Compared to single-variable ViT models, the 1EMD architecture improves performance across all six variables, achieving an average MSE reduction of approximately 5.5% under a fair and controlled comparison. It also consistently outperforms alternative multi-variable baselines, including a single-decoder ViT and a multi-variable U-Net. Moreover, multi-variable models substantially reduce computational cost, yielding a 29-32% lower inference time per variable compared to single-variable approaches. Overall, our results demonstrate that multi-variable modeling provides systematic advantages for high-resolution climate downscaling in terms of both accuracy and efficiency. Among the evaluated architectures, the proposed 1EMD ViT achieves the most favorable trade-off between predictive performance and computational cost.

Method: Proposes adversarial training on policy-generated trajectories with a co-evolving discriminator that separates policy trajectories from data distribution, while policy maximizes discriminator output plus coherence rewards

Result: Improved output diversity, harmonic coherence, adaptation speed, and user agency in both simulation (fixed test melodies and learned melody agents) and user study with expert musicians

Conclusion: Demonstrates effective method to mitigate reward hacking in RL post-training of generative sequence models for interactive musical applications

Abstract: Most applications of generative AI involve a sequential interaction in which a person inputs a prompt and waits for a response, and where reaction time and adaptivity are not important factors. In contrast, live jamming is a collaborative interaction that requires real-time coordination and adaptation without access to the other player’s future moves, while preserving diversity to sustain a creative flow. Reinforcement learning post-training enables effective adaptation through on-policy interaction, yet it often reduces output diversity by exploiting coherence-based rewards. This collapse, known as ``reward hacking’’, affects many RL post-training pipelines, but is especially harmful in live jamming, where musical creativity relies on dynamic variation and mutual responsiveness. In this paper, we propose a novel adversarial training method on policy-generated trajectories to mitigate reward hacking in RL post-training for melody-to-chord accompaniment. A co-evolving discriminator separates policy trajectories from the data distribution, while the policy maximizes the discriminator output in addition to coherence rewards to prevent collapse to trivial outputs. We evaluate accompaniment quality and output diversity in simulation with both fixed test melodies and learned melody agents, and we conduct a user study with the model deployed in a real-time interactive system with expert musicians. Quantitative evaluation and user feedback demonstrate improved output diversity, harmonic coherence, adaptation speed and user agency. Our results demonstrate a simple yet effective method to mitigate reward hacking in RL post-training of generative sequence models.

Method: Combines hardware-realistic quantum reservoir featurizer with kernel-based readout. Uses classical shadow tomography for efficient measurement of k-local observables, then applies classical kernel methods with explicit regularization.

Result: Provides learning-theoretic generalization guarantees for dependent temporal data, linking design choices to finite-sample performance. Empirical validation shows predicted interpolation and generalization behaviors on synthetic beta-mixing time series.

Conclusion: QuaRK offers principled guidance for building reliable temporal learners with clear computational knobs (circuit width/depth, measurement budget) while preserving kernel method flexibility for nonlinear temporal functionals.

Abstract: Quantum reservoir computing offers a promising route for time series learning by modelling sequential data via rich quantum dynamics while the only training required happens at the level of a lightweight classical readout. However, studies featuring efficient and implementable quantum reservoir architectures along with model learning guarantees remain scarce in the literature. To close this gap, we introduce QuaRK, an end-to-end framework that couples a hardware-realistic quantum reservoir featurizer with a kernel-based readout scheme. Given a sequence of sample points, the reservoir injects the points one after the other to yield a compact feature vector from efficiently measured k-local observables using classical shadow tomography, after which a classical kernel-based readout learns the target mapping with explicit regularization and fast optimization. The resulting pipeline exposes clear computational knobs – circuit width and depth as well as the measurement budget – while preserving the flexibility of kernel methods to model nonlinear temporal functionals and being scalable to high-dimensional data. We further provide learning-theoretic generalization guarantees for dependent temporal data, linking design and resource choices to finite-sample performance, thereby offering principled guidance for building reliable temporal learners. Empirical experiments validate QuaRK and illustrate the predicted interpolation and generalization behaviours on synthetic beta-mixing time series tasks.

Method: Reformulates OoS generalization as sequence modeling in weight space, partitions training set into concentric shells corresponding to discrete sequential steps, and uses WeightCaster framework to generate plausible, interpretable, and uncertainty-aware predictions.

Result: Demonstrates competitive or superior performance to state-of-the-art on synthetic cosine dataset and real-world air quality sensor readings, showing enhanced reliability beyond in-distribution scenarios.

Conclusion: WeightCaster enables reliable out-of-support generalization without explicit inductive biases, with significant implications for AI adoption in safety-critical applications requiring extrapolation beyond training data.

Abstract: As breakthroughs in deep learning transform key industries, models are increasingly required to extrapolate on datapoints found outside the range of the training set, a challenge we coin as out-of-support (OoS) generalisation. However, neural networks frequently exhibit catastrophic failure on OoS samples, yielding unrealistic but overconfident predictions. We address this challenge by reformulating the OoS generalisation problem as a sequence modelling task in the weight space, wherein the training set is partitioned into concentric shells corresponding to discrete sequential steps. Our WeightCaster framework yields plausible, interpretable, and uncertainty-aware predictions without necessitating explicit inductive biases, all the while maintaining high computational efficiency. Emprical validation on a synthetic cosine dataset and real-world air quality sensor readings demonstrates performance competitive or superior to the state-of-the-art. By enhancing reliability beyond in-distribution scenarios, these results hold significant implications for the wider adoption of artificial intelligence in safety-critical applications.

Method: Asymmetric architecture for downlink transmission: transmitter has scenario-aware semantic encoder that dynamically adjusts feature extraction based on CSI and SNR, plus neural precoding network to mitigate MUI in semantic domain; receiver uses lightweight decoder with novel pilot-guided attention mechanism for implicit channel equalization and feature calibration.

Result: Extensive simulations over 3GPP channel models show significant outperformance over DJSCC and traditional SSCC schemes in PSNR and classification accuracy, especially in low-SNR regimes, while maintaining low latency and computational cost on edge devices.

Conclusion: The proposed framework effectively addresses challenges of semantic communication in realistic MU-MIMO systems, demonstrating superior performance and practical feasibility for 6G networks.

Abstract: Semantic Communication (SemCom) has emerged as a promising paradigm for 6G networks, aiming to extract and transmit task-relevant information rather than minimizing bit errors. However, applying SemCom to realistic downlink Multi-User Multi-Input Multi-Output (MU-MIMO) Orthogonal Frequency Division Multiplexing (OFDM) systems remains challenging due to severe Multi-User Interference (MUI) and frequency-selective fading. Existing Deep Joint Source-Channel Coding (DJSCC) schemes, primarily designed for point-to-point links, suffer from performance saturation in multi-user scenarios. To address these issues, we propose a scenario-adaptive MU-MIMO SemCom framework featuring an asymmetric architecture tailored for downlink transmission. At the transmitter, we introduce a scenario-aware semantic encoder that dynamically adjusts feature extraction based on Channel State Information (CSI) and Signal-to-Noise Ratio (SNR), followed by a neural precoding network designed to mitigate MUI in the semantic domain. At the receiver, a lightweight decoder equipped with a novel pilot-guided attention mechanism is employed to implicitly perform channel equalization and feature calibration using reference pilot symbols. Extensive simulation results over 3GPP channel models demonstrate that the proposed framework significantly outperforms DJSCC and traditional Separate Source-Channel Coding (SSCC) schemes in terms of Peak Signal-to-Noise Ratio (PSNR) and classification accuracy, particularly in low-SNR regimes, while maintaining low latency and computational cost on edge devices.

Method: Formulates clustering as optimal multiway-split decision trees using 0-1 integer linear optimization (more tractable than mixed-integer nonlinear problems). Integrates one-dimensional K-means for discretization of continuous variables to enable flexible, data-driven branching.

Result: Outperforms baseline methods in clustering accuracy and interpretability on real-world datasets. Produces multiway-split decision trees with concise decision rules while maintaining competitive performance across various metrics.

Conclusion: The proposed method provides an effective approach for interpretable clustering with improved computational tractability and better interpretability through multiway-split decision trees.

Abstract: Clustering serves as a vital tool for uncovering latent data structures, and achieving both high accuracy and interpretability is essential. To this end, existing methods typically construct binary decision trees by solving mixed-integer nonlinear optimization problems, often leading to significant computational costs and suboptimal solutions. Furthermore, binary decision trees frequently result in excessively deep structures, which makes them difficult to interpret. To mitigate these issues, we propose an interpretable clustering method based on optimal multiway-split decision trees, formulated as a 0-1 integer linear optimization problem. This reformulation renders the optimization problem more tractable compared to existing models. A key feature of our method is the integration of a one-dimensional K-means algorithm for the discretization of continuous variables, allowing for flexible and data-driven branching. Extensive numerical experiments on publicly available real-world datasets demonstrate that our method outperforms baseline methods in terms of clustering accuracy and interpretability. Our method yields multiway-split decision trees with concise decision rules while maintaining competitive performance across various evaluation metrics.

Method: Researchers simulate diverse data leakage scenarios by conducting continued pre-training of foundation models on strategically blended corpora containing user-item interactions from both in-domain and out-of-domain sources to study the effects of data contamination.

Result: Experiments reveal a dual-effect: domain-relevant data leakage causes substantial but spurious performance gains that misleadingly exaggerate model capability, while domain-irrelevant leakage typically degrades recommendation accuracy, showing the complex nature of this contamination.

Conclusion: Data leakage is a critical, previously unaccounted-for factor in LLM-based recommendation that impacts true model performance evaluation, highlighting the need for more rigorous evaluation protocols in this domain.

Abstract: The expanding integration of Large Language Models (LLMs) into recommender systems poses critical challenges to evaluation reliability. This paper identifies and investigates a previously overlooked issue: benchmark data leakage in LLM-based recommendation. This phenomenon occurs when LLMs are exposed to and potentially memorize benchmark datasets during pre-training or fine-tuning, leading to artificially inflated performance metrics that fail to reflect true model performance. To validate this phenomenon, we simulate diverse data leakage scenarios by conducting continued pre-training of foundation models on strategically blended corpora, which include user-item interactions from both in-domain and out-of-domain sources. Our experiments reveal a dual-effect of data leakage: when the leaked data is domain-relevant, it induces substantial but spurious performance gains, misleadingly exaggerating the model’s capability. In contrast, domain-irrelevant leakage typically degrades recommendation accuracy, highlighting the complex and contingent nature of this contamination. Our findings reveal that data leakage acts as a critical, previously unaccounted-for factor in LLM-based recommendation, which could impact the true model performance. We release our code at https://github.com/yusba1/LLMRec-Data-Leakage.

Method: Proposes a weighted distribution-aware kernel that embeds nodes while considering their distributional characteristics. The method explicitly incorporates both node and degree distributions, requires no optimization, and mitigates over-smoothing effects to preserve node distinguishability across many iterations.

Result: The method achieves superior community detection performance via spectral clustering, outperforming existing graph embedding methods including deep learning approaches on standard benchmarks. It effectively preserves node distinguishability even after many embedding iterations.

Conclusion: Incorporating node and degree distribution characteristics is crucial for addressing over-smoothing in graph embedding methods. The proposed distribution-aware kernel provides an effective, optimization-free solution that enhances representation expressiveness and community detection performance.

Abstract: Neighborhood Aggregation Strategy (NAS) is a widely used approach in graph embedding, underpinning both Graph Neural Networks (GNNs) and Weisfeiler-Lehman (WL) methods. However, NAS-based methods are identified to be prone to over-smoothing-the loss of node distinguishability with increased iterations-thereby limiting their effectiveness. This paper identifies two characteristics in a network, i.e., the distributions of nodes and node degrees that are critical for expressive representation but have been overlooked in existing methods. We show that these overlooked characteristics contribute significantly to over-smoothing of NAS-methods. To address this, we propose a novel weighted distribution-aware kernel that embeds nodes while taking their distributional characteristics into consideration. Our method has three distinguishing features: (1) it is the first method to explicitly incorporate both distributional characteristics; (2) it requires no optimization; and (3) it effectively mitigates the adverse effects of over-smoothing, allowing WL to preserve node distinguishability and expressiveness even after many iterations of embedding. Experiments demonstrate that our method achieves superior community detection performance via spectral clustering, outperforming existing graph embedding methods, including deep learning methods, on standard benchmarks.

Method: Introduces Joint Time Series Chain definition that handles gaps/interruptions, proposes effective ranking criterion to identify best chains, and implements algorithm for finding chains across interrupted or related time series.

Result: Outperforms existing TSC methods in locating unusual evolving patterns through extensive empirical evaluations. Demonstrated utility with real-life manufacturing application from Intel.

Conclusion: JointTSC effectively addresses limitations of single-series chain methods by enabling discovery of evolving patterns across interrupted or related time series, with practical applications in manufacturing and complex systems.

Abstract: Time series chain (TSC) is a recently introduced concept that captures the evolving patterns in large scale time series. Informally, a time series chain is a temporally ordered set of subsequences, in which consecutive subsequences in the chain are similar to one another, but the last and the first subsequences maybe be dissimilar. Time series chain has the great potential to reveal latent unusual evolving trend in the time series, or identify precursor of important events in a complex system. Unfortunately, existing definitions of time series chains only consider finding chains in a single time series. As a result, they are likely to miss unexpected evolving patterns in interrupted time series, or across two related time series. To address this limitation, in this work, we introduce a new definition called \textit{Joint Time Series Chain}, which is specially designed for the task of finding unexpected evolving trend across interrupted time series or two related time series. Our definition focuses on mitigating the robustness issues caused by the gap or interruption in the time series. We further propose an effective ranking criterion to identify the best chain. We demonstrate that our proposed approach outperforms existing TSC work in locating unusual evolving patterns through extensive empirical evaluations. We further demonstrate the utility of our work with a real-life manufacturing application from Intel. Our source code is publicly available at the supporting page https://github.com/lizhang-ts/JointTSC .

Method: Proposes DeepFusion framework where devices independently configure/train on-device LLMs tailored to their needs, then uses federated knowledge distillation with a novel View-Aligned Attention (VAA) module that integrates multi-stage feature representations from global MoE to align predictive perspectives between heterogeneous models.

Result: Achieves performance close to centralized MoE training, reduces communication costs by up to 71%, and improves token perplexity by up to 5.28% compared to federated MoE baselines on industry-level models (Qwen-MoE, DeepSeek-MoE) with real-world datasets.

Conclusion: DeepFusion enables practical federated training of MoE models on resource-constrained devices through effective cross-architecture knowledge distillation with view alignment, making large-scale privacy-preserving MoE training feasible.

Abstract: Recent Mixture-of-Experts (MoE)-based large language models (LLMs) such as Qwen-MoE and DeepSeek-MoE are transforming generative AI in natural language processing. However, these models require vast and diverse training data. Federated learning (FL) addresses this challenge by leveraging private data from heterogeneous edge devices for privacy-preserving MoE training. Nonetheless, traditional FL approaches require devices to host local MoE models, which is impractical for resource-constrained devices due to large model sizes. To address this, we propose DeepFusion, the first scalable federated MoE training framework that enables the fusion of heterogeneous on-device LLM knowledge via federated knowledge distillation, yielding a knowledge-abundant global MoE model. Specifically, DeepFusion features each device to independently configure and train an on-device LLM tailored to its own needs and hardware limitations. Furthermore, we propose a novel View-Aligned Attention (VAA) module that integrates multi-stage feature representations from the global MoE model to construct a predictive perspective aligned with on-device LLMs, thereby enabling effective cross-architecture knowledge distillation. By explicitly aligning predictive perspectives, VAA resolves the view-mismatch problem in traditional federated knowledge distillation, which arises from heterogeneity in model architectures and prediction behaviors between on-device LLMs and the global MoE model. Experiments with industry-level MoE models (Qwen-MoE and DeepSeek-MoE) and real-world datasets (medical and finance) demonstrate that DeepFusion achieves performance close to centralized MoE training. Compared with key federated MoE baselines, DeepFusion reduces communication costs by up to 71% and improves token perplexity by up to 5.28%.

Method: Introduces cumulative utility parity principle and availability-normalized cumulative utility metric, which disentangles unavoidable physical constraints from avoidable algorithmic bias arising from scheduling and aggregation. The approach evaluates whether clients receive comparable long-term benefit per participation opportunity.

Result: Experiments on temporally skewed, non-IID federated benchmarks demonstrate that the approach substantially improves long-term representation parity while maintaining near-perfect performance.

Conclusion: The proposed cumulative utility parity framework addresses a critical gap in FL fairness by focusing on long-term benefits per participation opportunity, providing a more meaningful fairness metric for real-world FL systems with intermittent client participation.

Abstract: In real-world federated learning (FL) systems, client participation is intermittent, heterogeneous, and often correlated with data characteristics or resource constraints. Existing fairness approaches in FL primarily focus on equalizing loss or accuracy conditional on participation, implicitly assuming that clients have comparable opportunities to contribute over time. However, when participation itself is uneven, these objectives can lead to systematic under-representation of intermittently available clients, even if per-round performance appears fair. We propose cumulative utility parity, a fairness principle that evaluates whether clients receive comparable long-term benefit per participation opportunity, rather than per training round. To operationalize this notion, we introduce availability-normalized cumulative utility, which disentangles unavoidable physical constraints from avoidable algorithmic bias arising from scheduling and aggregation. Experiments on temporally skewed, non-IID federated benchmarks demonstrate that our approach substantially improves long-term representation parity, while maintaining near-perfect performance.

Method: Proposes a policy-driven zero-order framework that treats the sampling distribution over perturbation directions as a learnable policy. Updates the policy to reduce variance of directional estimates, with theoretical analysis showing improved gradient quality and relaxed dependence on parameter dimensionality.

Result: Empirical validation on challenging LLM fine-tuning benchmarks shows substantially improved performance compared to standard zero-order baselines, demonstrating the effectiveness of adaptive direction sampling.

Conclusion: Adaptive direction sampling through learned policies is a promising approach to make zero-order fine-tuning viable at scale, offering memory-efficient alternatives to backpropagation-based methods.

Abstract: Fine-tuning large pretrained language models (LLMs) is a cornerstone of modern NLP, yet its growing memory demands (driven by backpropagation and large optimizer States) limit deployment in resource-constrained settings. Zero-order (ZO) methods bypass backpropagation by estimating directional derivatives from forward evaluations, offering substantial memory savings. However, classical ZO estimators suffer from high variance and an adverse dependence on the parameter dimensionality $d$, which has constrained their use to low-dimensional problems. In this work, we propose a policy-driven ZO framework that treats the sampling distribution over perturbation directions as a learnable policy and updates it to reduce the variance of directional estimates. We develop a practical algorithm implementing this idea and provide a theoretical analysis, showing that learned sampling distributions improve the quality of gradient information and relax the explicit dependence on $d$ in convergence bounds. Empirically, we validate the approach on challenging LLM fine-tuning benchmarks, demonstrating substantially improved performance compared to standard ZO baselines. Our results suggest that adaptive direction sampling is a promising route to make ZO fine-tuning viable at scale. The source code is available at https://github.com/brain-lab-research/zo_ldsd

Method: Proposes OCE-RCPS framework that uses optimized certainty equivalent (OCE) risk measures including CVaR and entropic risk, leveraging upper confidence bounds to identify prediction set parameters that satisfy user-specified risk tolerance levels with provable reliability.

Result: Theoretical guarantees show OCE-RCPS satisfies probabilistic constraints for loss functions like miscoverage and false negative rate. Experiments on image segmentation demonstrate consistent satisfaction of target rates across various risk measures and reliability configurations.

Conclusion: OCE-RCPS provides superior reliability guarantees for safety-critical applications compared to conventional RCPS, offering probabilistic guarantees on tail risk measures crucial for high-stakes decision making.

Abstract: In safety-critical applications such as medical image segmentation, prediction systems must provide reliability guarantees that extend beyond conventional expected loss control. While risk-controlling prediction sets (RCPS) offer probabilistic guarantees on the expected risk, they fail to capture tail behavior and worst-case scenarios that are crucial in high-stakes settings. This paper introduces optimized certainty equivalent RCPS (OCE-RCPS), a novel framework that provides high-probability guarantees on general optimized certainty equivalent (OCE) risk measures, including conditional value-at-risk (CVaR) and entropic risk. OCE-RCPS leverages upper confidence bounds to identify prediction set parameters that satisfy user-specified risk tolerance levels with provable reliability. We establish theoretical guarantees showing that OCE-RCPS satisfies the desired probabilistic constraint for loss functions such as miscoverage and false negative rate. Experiments on image segmentation demonstrate that OCE-RCPS consistently meets target satisfaction rates across various risk measures and reliability configurations, while OCE-CRC fails to provide probabilistic guarantees.

Method: ALMo employs a novel optimization framework that minimizes manual input through automated parameter setup and enables flexible control over toxicity risks. The system allows clinicians to navigate the Pareto surface of dosimetric tradeoffs by directly manipulating intuitive aim and limit values.

Result: In a retrospective evaluation of 25 clinical cases, ALMo generated treatment plans that consistently met or exceeded manual planning quality, with 65% of cases demonstrating dosimetric improvements. The system significantly enhanced efficiency, reducing average planning time to approximately 17 minutes compared to conventional 30-60 minutes.

Conclusion: ALMo demonstrates a generalized framework for streamlining interaction in multi-criteria clinical decision-making, validated in brachytherapy but applicable to broader clinical contexts.

Abstract: In complex clinical decision-making, clinicians must often track a variety of competing metrics defined by aim (ideal) and limit (strict) thresholds. Sifting through these high-dimensional tradeoffs to infer the optimal patient-specific strategy is cognitively demanding and historically prone to variability. In this paper, we address this challenge within the context of High-Dose-Rate (HDR) brachytherapy for cervical cancer, where planning requires strictly managing radiation hot spots while balancing tumor coverage against organ sparing. We present ALMo (Aim-Limit-defined Multi-Objective system), an interactive decision support system designed to infer and operationalize clinician intent. ALMo employs a novel optimization framework that minimizes manual input through automated parameter setup and enables flexible control over toxicity risks. Crucially, the system allows clinicians to navigate the Pareto surface of dosimetric tradeoffs by directly manipulating intuitive aim and limit values. In a retrospective evaluation of 25 clinical cases, ALMo generated treatment plans that consistently met or exceeded manual planning quality, with 65% of cases demonstrating dosimetric improvements. Furthermore, the system significantly enhanced efficiency, reducing average planning time to approximately 17 minutes, compared to the conventional 30-60 minutes. While validated in brachytherapy, ALMo demonstrates a generalized framework for streamlining interaction in multi-criteria clinical decision-making.

Method: Proposes VILA, a dual-branch framework with two-level vision-language calibration: 1) feature-level geometric calibration fusing plastic task-adapted features with frozen universal semantic anchors, and 2) decision-level cross-modal priors to rectify prediction bias.

Result: Extensive experiments across eight benchmarks show VILA yields superior performance, particularly in fine-grained and long-sequence scenarios, while maintaining analytic learning’s efficiency.

Conclusion: VILA harmonizes high-fidelity prediction with the simplicity of analytic learning, overcoming the brittleness of traditional analytic CIL approaches while preserving efficiency.

Abstract: Class-incremental learning (CIL) with pre-trained models (PTMs) faces a critical trade-off between efficient adaptation and long-term stability. While analytic learning enables rapid, recursive closed-form updates, its efficacy is often compromised by accumulated errors and feature incompatibility. In this paper, we first conduct a systematic study to dissect the failure modes of PTM-based analytic CIL, identifying representation rigidity as the primary bottleneck. Motivated by these insights, we propose \textbf{VILA}, a novel dual-branch framework that advances analytic CIL via a two-level vision-language calibration strategy. Specifically, we coherently fuse plastic, task-adapted features with a frozen, universal semantic anchor at the feature level through geometric calibration, and leverage cross-modal priors at the decision level to rectify prediction bias. This confluence maintains analytic-learning’s extreme efficiency while overcoming its inherent brittleness. Extensive experiments across eight benchmarks demonstrate that VILA consistently yields superior performance, particularly in fine-grained and long-sequence scenarios. Our framework harmonizes high-fidelity prediction with the simplicity of analytic learning. Our code is available at https://github.com/byzhaoAI/VILA

Method: Proposes fluid-agent environment framework with game-theoretic solution concepts, evaluates MARL algorithms on fluid variants of established benchmarks (Predator-Prey, Level-Based Foraging) and introduces new environment.

Result: Demonstrates that fluid-agent framework yields agent teams that dynamically adjust their size to match environmental demands, unlocking novel solution strategies beyond fixed-population settings.

Conclusion: Fluid-agent environments represent an important extension of MARL to handle real-world scenarios with dynamic populations, enabling more adaptive and scalable multi-agent systems.

Abstract: The primary focus of multi-agent reinforcement learning (MARL) has been to study interactions among a fixed number of agents embedded in an environment. However, in the real world, the number of agents is neither fixed nor known a priori. Moreover, an agent can decide to create other agents (for example, a cell may divide, or a company may spin off a division). In this paper, we propose a framework that allows agents to create other agents; we call this a fluid-agent environment. We present game-theoretic solution concepts for fluid-agent games and empirically evaluate the performance of several MARL algorithms within this framework. Our experiments include fluid variants of established benchmarks such as Predator-Prey and Level-Based Foraging, where agents can dynamically spawn, as well as a new environment we introduce that highlights how fluidity can unlock novel solution strategies beyond those observed in fixed-population settings. We demonstrate that this framework yields agent teams that adjust their size dynamically to match environmental demands.

Method: Establishes structural dichotomy between pseudometric and general weighted instances; proves VC dimension of clustering disagreement class; analyzes active triangle inequalities at LP vertices; develops sparsified LP-PIVOT algorithm with imputation via triangle inequalities; uses Yao’s minimax principle for negative results.

Result: VC dimension is exactly n-1 yielding optimal additive ε-coresets; at most binom(n,2) triangle inequalities active at LP vertices enabling exact cutting-plane solver; sparsified LP-PIVOT achieves 10/3-approximation with Õ(n^{3/2}) edges; negative result shows unbounded approximation ratio with o(n) edges without pseudometric structure.

Conclusion: Pseudometric structure governs both tractability and robustness of Correlation Clustering to incomplete information, with sharp thresholds for edge observation requirements and fundamental limitations for general weighted instances.

Abstract: Correlation Clustering (CC) is a fundamental unsupervised learning primitive whose strongest LP-based approximation guarantees require $Θ(n^3)$ triangle inequality constraints and are prohibitive at scale. We initiate the study of \emph{sparsification–approximation trade-offs} for CC, asking how much edge information is needed to retain LP-based guarantees. We establish a structural dichotomy between pseudometric and general weighted instances. On the positive side, we prove that the VC dimension of the clustering disagreement class is exactly $n{-}1$, yielding additive $\varepsilon$-coresets of optimal size $\tilde{O}(n/\varepsilon^2)$; that at most $\binom{n}{2}$ triangle inequalities are active at any LP vertex, enabling an exact cutting-plane solver; and that a sparsified variant of LP-PIVOT, which imputes missing LP marginals via triangle inequalities, achieves a robust $\frac{10}{3}$-approximation (up to an additive term controlled by an empirically computable imputation-quality statistic $\overlineΓ_w$) once $\tildeΘ(n^{3/2})$ edges are observed, a threshold we prove is sharp. On the negative side, we show via Yao’s minimax principle that without pseudometric structure, any algorithm observing $o(n)$ uniformly random edges incurs an unbounded approximation ratio, demonstrating that the pseudometric condition governs not only tractability but also the robustness of CC to incomplete information.

Method: Proposes physics-based constraints: divergence-free vector field (via neural network outputting vector potential A with magnetic field as its curl) and E(3)-equivariance (using tensor products of geometric tensors with spherical harmonics). Uses Contiformer architecture for continuous-time dynamics and long-term memory, plus synthetic data generation with time-series conditional GANs based on World Magnetic Model.

Result: Physics constraints significantly improve predictive accuracy and physical plausibility; Contiformer outperforms state-of-the-art methods; ablation studies show benefits of both constraints; synthetic data generation addresses data scarcity.

Conclusion: Embedding physical constraints acts as implicit regularizer improving spatio-temporal performance; continuous-time dynamics and long-term memory are critical for magnetic time series modeling; physics-constrained deep learning outperforms classical and unconstrained approaches.

Abstract: Magnetic-anomaly navigation, leveraging small-scale variations in the Earth’s magnetic field, is a promising alternative when GPS is unavailable or compromised. Airborne systems face a key challenge in extracting geomagnetic field data: the aircraft itself induces magnetic noise. Although the classical Tolles-Lawson model addresses this, it inadequately handles stochastically corrupted magnetic data required for navigation. To address stochastic noise, we propose a framework based on two physics-based constraints: divergence-free vector field and E(3)-equivariance. These ensure the learned magnetic field obeys Maxwell’s equations and that outputs transform correctly with sensor position/orientation. The divergence-free constraint is implemented by training a neural network to output a vector potential $A$, with the magnetic field defined as its curl. For E(3)-equivariance, we use tensor products of geometric tensors representable via spherical harmonics with known rotational transformations. Enforcing physical consistency and restricting the admissible function space acts as an implicit regularizer that improves spatio-temporal performance. We present ablation studies evaluating each constraint alone and jointly across CNNs, MLPs, Liquid Time Constant models, and Contiformers. Continuous-time dynamics and long-term memory are critical for modelling magnetic time series; the Contiformer architecture, which provides both, outperforms state-of-the-art methods. To mitigate data scarcity, we generate synthetic datasets using the World Magnetic Model (WMM) with time-series conditional GANs, producing realistic, temporally consistent magnetic sequences across varied trajectories and environments. Experiments show that embedding these constraints significantly improves predictive accuracy and physical plausibility, outperforming classical and unconstrained deep learning approaches.

Method: Atomix introduces a runtime with progress-aware transactional semantics: tags each tool call with an epoch, tracks per-resource frontiers, commits only when progress predicates indicate safety, buffers effects when possible, and provides compensation mechanisms for externalized effects on abort.

Result: Across real workloads with fault injection, transactional retry improves task success rates, while frontier-gated commit strengthens isolation under speculation and contention scenarios.

Conclusion: Atomix addresses critical safety issues in LLM agent interactions with external systems by providing transactional guarantees that prevent unintended side effects and enable safe rollback of operations.

Abstract: LLM agents increasingly act on external systems, yet tool effects are immediate. Under failures, speculation, or contention, losing branches can leak unintended side effects with no safe rollback. We introduce Atomix, a runtime that provides progress-aware transactional semantics for agent tool calls. Atomix tags each call with an epoch, tracks per-resource frontiers, and commits only when progress predicates indicate safety; bufferable effects can be delayed, while externalized effects are tracked and compensated on abort. Across real workloads with fault injection, transactional retry improves task success, while frontier-gated commit strengthens isolation under speculation and contention.

Method: Head Entropy measures the spread of attention mass using per-head 2-Renyi entropies. It uses sparse logistic regression on these attention entropy patterns to predict answer correctness. The method analyzes attention patterns both during answer generation and even before answer generation (over question/context alone).

Result: Head Entropy matches or exceeds baselines in-distribution and generalizes substantially better out-of-domain, outperforming the closest baseline by +8.5% AUROC on average. Attention patterns over question/context alone already carry predictive signal, with +17.7% AUROC improvement over the closest baseline. Evaluated across 5 instruction-tuned LLMs and 3 QA datasets spanning general knowledge, multi-hop reasoning, and medicine.

Conclusion: Attention entropy patterns (Head Entropy) provide a robust method for predicting answer correctness in LLMs, with strong generalization capabilities across domains and the ability to detect potential errors even before answer generation.

Abstract: Large language models (LLMs) often generate plausible yet incorrect answers, posing risks in safety-critical settings such as medicine. Human evaluation is expensive, and LLM-as-judge approaches risk introducing hidden errors. Recent white-box methods detect contextual hallucinations using model internals, focusing on the localization of the attention mass, but two questions remain open: do these approaches extend to predicting answer correctness, and do they generalize out-of-domains? We introduce Head Entropy, a method that predicts answer correctness from attention entropy patterns, specifically measuring the spread of the attention mass. Using sparse logistic regression on per-head 2-Renyi entropies, Head Entropy matches or exceeds baselines in-distribution and generalizes substantially better on out-of-domains, it outperforms the closest baseline on average by +8.5% AUROC. We further show that attention patterns over the question/context alone, before answer generation, already carry predictive signal using Head Entropy with on average +17.7% AUROC over the closest baseline. We evaluate across 5 instruction-tuned LLMs and 3 QA datasets spanning general knowledge, multi-hop reasoning, and medicine.

Method: ToolSelect adaptively learns model selection by minimizing population risk over sampled specialist tool candidates using a consistent surrogate of task-conditional selection loss. Uses an Attentive Neural Process-based selector conditioned on the query and per-model behavioral summaries.

Result: Introduced ToolSelectBench benchmark with 1448 queries in agentic Chest X-ray environment with diverse specialist models (17 disease detection, 19 report generation, 6 visual grounding, 13 VQA). ToolSelect consistently outperforms 10 state-of-the-art methods across four different task families.

Conclusion: ToolSelect provides an effective solution for adaptive model selection in agentic healthcare systems, enabling reliable selection of specialist models from heterogeneous pools for clinical queries.

Abstract: Task-specialized models form the backbone of agentic healthcare systems, enabling the agents to answer clinical queries across tasks such as disease diagnosis, localization, and report generation. Yet, for a given task, a single “best” model rarely exists. In practice, each task is better served by multiple competing specialist models where different models excel on different data samples. As a result, for any given query, agents must reliably select the right specialist model from a heterogeneous pool of tool candidates. To this end, we introduce ToolSelect, which adaptively learns model selection for tools by minimizing a population risk over sampled specialist tool candidates using a consistent surrogate of the task-conditional selection loss. Concretely, we propose an Attentive Neural Process-based selector conditioned on the query and per-model behavioral summaries to choose among the specialist models. Motivated by the absence of any established testbed, we, for the first time, introduce an agentic Chest X-ray environment equipped with a diverse suite of task-specialized models (17 disease detection, 19 report generation, 6 visual grounding, and 13 VQA) and develop ToolSelectBench, a benchmark of 1448 queries. Our results demonstrate that ToolSelect consistently outperforms 10 SOTA methods across four different task families.

Method: Policy optimization technique applied to stochastic contextual multi-armed bandit with general offline function approximation, using function class F to approximate losses.

Result: Achieves optimal regret bound of Õ(√(K|A|log|F|)) with high probability, where K is number of rounds, A is set of arms, and F is function class. Algorithm is both efficient and theoretically optimal.

Conclusion: Demonstrates that practical policy optimization methods for contextual bandits can achieve rigorously-proved optimal regret bounds, bridging theory and practice in reinforcement learning.

Abstract: We present the first high-probability optimal regret bound for a policy optimization technique applied to the problem of stochastic contextual multi-armed bandit (CMAB) with general offline function approximation. Our algorithm is both efficient and achieves an optimal regret bound of $\widetilde{O}(\sqrt{ K|\mathcal{A}|\log|\mathcal{F}|})$, where $K$ is the number of rounds, $\mathcal{A}$ is the set of arms, and $\mathcal{F}$ is the function class used to approximate the losses. Our results bridge the gap between theory and practice, demonstrating that the widely used policy optimization methods for the contextual bandit problem can achieve a rigorously-proved optimal regret bound. We support our theoretical results with an empirical evaluation of our algorithm.

Method: The authors introduce OPO-CMDP, an optimistic policy optimization algorithm that leverages finite function classes to approximate losses and dynamics, achieving regret bounds with optimal dependence on state and action space sizes.

Result: OPO-CMDP achieves a high probability regret bound of Õ(H⁴√(T|S||A|log(|F||P|))), which is the first regret bound with optimal dependence on |S| and |A|, directly improving upon Qian, Hu, and Simchi-Levi (2024).

Conclusion: Optimistic policy optimization provides a natural, computationally superior and theoretically near-optimal approach for solving CMDPs with general offline function approximation.

Abstract: We introduce \texttt{OPO-CMDP}, the first policy optimization algorithm for stochastic Contextual Markov Decision Process (CMDPs) under general offline function approximation. Our approach achieves a high probability regret bound of $\widetilde{O}(H^4\sqrt{T|S||A|\log(|\mathcal{F}||\mathcal{P}|)}),$ where $S$ and $A$ denote the state and action spaces, $H$ the horizon length, $T$ the number of episodes, and $\mathcal{F}, \mathcal{P}$ the finite function classes used to approximate the losses and dynamics, respectively. This is the first regret bound with optimal dependence on $|S|$ and $|A|$, directly improving the current state-of-the-art (Qian, Hu, and Simchi-Levi, 2024). These results demonstrate that optimistic policy optimization provides a natural, computationally superior and theoretically near-optimal path for solving CMDPs.

Method: 1) Use policy-aware enhanced Hessian to identify action-critical weights; 2) Apply sparse orthogonal transform to non-salient weights to create low-entropy intermediate state; 3) Quantize both salient and non-salient weights in Harr domain with group-wise 1-bit quantization.

Result: Quantized OpenVLA-OFT retains 92.2% performance on LIBERO, CogAct retains 93.6% on SimplerEnv, significantly outperforming SOTA binarization methods. Real-world evaluation shows only marginal success-rate degradation compared to full-precision models.

Conclusion: HBVLA provides practical ultra-low-bit quantization for VLAs, enabling reliable deployment on hardware-limited robotic platforms while maintaining performance close to full-precision models.

Abstract: Vision-Language-Action (VLA) models enable instruction-following embodied control, but their large compute and memory footprints hinder deployment on resource-constrained robots and edge platforms. While reducing weights to 1-bit precision through binarization can greatly improve efficiency, existing methods fail to narrow the distribution gap between binarized and full-precision weights, causing quantization errors to accumulate under long-horizon closed-loop execution and severely degrade actions. To fill this gap, we propose HBVLA, a VLA-tailored binarization framework. First, we use a policy-aware enhanced Hessian to identify weights that are truly critical for action generation. Then, we employ a sparse orthogonal transform for non-salient weights to induce a low-entropy intermediate state. Finally, we quantize both salient and non-salient weights in the Harr domain with group-wise 1-bit quantization. We have evaluated our approach on different VLAs: on LIBERO, quantized OpenVLA-OFT retains 92.2% of full-precision performance; on SimplerEnv, quantized CogAct retains 93.6%, significantly outperforming state-of-the-art binarization methods. We further validate our method on real-world evaluation suite and the results show that HBVLA incurs only marginal success-rate degradation compared to the full-precision model, demonstrating robust deployability under tight hardware constraints. Our work provides a practical foundation for ultra-low-bit quantization of VLAs, enabling more reliable deployment on hardware-limited robotic platforms.

Method: Proposes a machine learning-powered bi-level optimization framework where objective functions are approximated by ANN models, and the lower-level problem is analytically embedded through KKT optimality conditions, creating an ANN-KKT single-level optimization framework.

Result: Validated on benchmark problems and real-world power plants (660 MW coal and 395 MW gas turbine), achieving comparable solutions to bi-level benchmarks with marginal computational time (0.22-0.88s). Generated 583 MW (coal) and 402 MW (gas turbine) outputs at optimal turbine heat rates, and delineated feasible operating envelopes accounting for uncertainty.

Conclusion: ANN-KKT offers a scalable, computationally efficient route for hierarchical, data-driven optimization of industrial thermal power systems, enabling energy-efficient operations of large-scale engineering systems and contributing to Industry 5.0.

Abstract: Industrial thermal power systems have coupled performance variables with hierarchical order of importance, making their simultaneous optimization computationally challenging or infeasible. This barrier limits the integrated and computationally scaleable operation optimization of industrial thermal power systems. To address this issue for large-scale engineering systems, we present a fully machine learning-powered bi-level optimization framework for data-driven optimization of industrial thermal power systems. The objective functions of upper and lower levels are approximated by artificial neural network (ANN) models and the lower-level problem is analytically embedded through Karush-Kuhn-Tucker (KKT) optimality conditions. The reformulated single level optimization framework integrating ANN models and KKT constraints (ANN-KKT) is validated on benchmark problems and on real-world power generation operation of 660 MW coal power plant and 395 MW gas turbine system. The results reveal a comparable solutions obtained from the proposed ANN-KKT framework to the bi-level solutions of the benchmark problems. Marginal computational time requirement (0.22 to 0.88 s) to compute optimal solutions yields 583 MW (coal) and 402 MW (gas turbine) of power output at optimal turbine heat rate of 7337 kJ/kWh and 7542 kJ/kWh, respectively. In addition, the method expands to delineate a feasible and robust operating envelope that accounts for uncertainty in operating variables while maximizing thermal efficiency in various scenarios. These results demonstrate that ANN-KKT offers a scalable and computationally efficient route for hierarchical, data-driven optimization of industrial thermal power systems, achieving energy-efficient operations of large-scale engineering systems and contributing to industry 5.0.

Method: Introduces a discrete double-bracket flow whose generator is invariant to isotropic shifts, making it pathwise invariant to isotropic noise at the discrete-time level. The method uses a maximal stable step size proportional to 1/||C_e||_2^2, where C_e is the trace-free covariance. Analysis includes strict-saddle geometry for diagonalization objective and input-to-state stability analysis.

Result: Achieves global convergence with sample complexity scaling as O(||C_e||_2^2/(Δ^2ε)) under trace-free perturbations. Explicit characterization of degenerate blocks yields accelerated O(log(1/ζ)) saddle-escape rate and high-probability finite-time convergence guarantee.

Conclusion: The proposed discrete double-bracket flow provides a robust matrix-free eigendecomposition method that’s invariant to isotropic noise, with improved convergence rates and stability properties compared to standard stochastic approximation approaches.

Abstract: We study matrix-free eigendecomposition under a matrix-vector product (MVP) oracle, where each step observes a covariance operator $C_k = C_{sig} + σ_k^2 I + E_k$. Standard stochastic approximation methods either use fixed steps that couple stability to $|C_k|2$, or adapt steps in ways that slow down due to vanishing updates. We introduce a discrete double-bracket flow whose generator is invariant to isotropic shifts, yielding pathwise invariance to $σ_k^2 I$ at the discrete-time level. The resulting trajectory and a maximal stable step size $η{max} \propto 1/|C_e|_2^2$ depend only on the trace-free covariance $C_e$. We establish global convergence via strict-saddle geometry for the diagonalization objective and an input-to-state stability analysis, with sample complexity scaling as $O(|C_e|_2^2 / (Δ^2 ε))$ under trace-free perturbations. An explicit characterization of degenerate blocks yields an accelerated $O(\log(1/ζ))$ saddle-escape rate and a high-probability finite-time convergence guarantee.

Method: Proposes Compressed Representation Data Selection (CRDS) with two variants: CRDS-R uses Rademacher random projection with concatenated transformer hidden-layer representations, and CRDS-W employs whitening-based dimensionality reduction to improve representational quality.

Result: Both CRDS variants substantially enhance data quality and outperform state-of-the-art representation-based selection methods. CRDS-W achieves strong performance using only 3.5% of data, surpassing full-data baseline by average 0.71% across four datasets.

Conclusion: CRDS effectively addresses redundancy in LLM embeddings for data selection, enabling efficient instruction tuning with minimal high-quality data while maintaining or improving performance.

Abstract: Data quality is a crucial factor in large language models training. While prior work has shown that models trained on smaller, high-quality datasets can outperform those trained on much larger but noisy or low-quality corpora, systematic methods for industrial-scale data selection in instruction tuning remain underexplored. In this work, we study instruction-tuning data selection through the lens of semantic representation similarity and identify a key limitation of state-of-the-art LLM encoders: they produce highly redundant semantic embeddings. To mitigate this redundancy, we propose Compressed Representation Data Selection (CRDS), a novel framework with two variants. CRDS-R applies Rademacher random projection followed by concatenation of transformer hidden-layer representations, while CRDS-W employs whitening-based dimensionality reduction to improve representational quality. Experimental results demonstrate that both variants substantially enhance data quality and consistently outperform state-of-the-art representation-based selection methods. Notably, CRDS-W achieves strong performance using only 3.5% of the data, surpassing the full-data baseline by an average of 0.71% across four datasets. Our code is available at https://github.com/tdano1/CRDS.

Method: MEMTS introduces a Knowledge Persistence Module (KPM) that internalizes domain-specific temporal dynamics (seasonal patterns, trends) into a compact set of learnable latent prototypes. This transforms fragmented historical observations into continuous, parameterized knowledge representations, enabling retrieval-free domain adaptation with constant-time inference and near-zero latency.

Result: Extensive experiments on multiple datasets demonstrate state-of-the-art performance. MEMTS achieves accurate domain adaptation while effectively mitigating catastrophic forgetting of general temporal patterns, all without requiring architectural modifications to the frozen TSFM backbone.

Conclusion: MEMTS provides a lightweight, plug-and-play solution for retrieval-free domain adaptation in time series forecasting, addressing the scalability bottlenecks of current approaches while maintaining the learned global temporal patterns of foundation models.

Abstract: While Time Series Foundation Models (TSFMs) have demonstrated exceptional performance in generalized forecasting, their performance often degrades significantly when deployed in real-world vertical domains characterized by temporal distribution shifts and domain-specific periodic structures. Current solutions are primarily constrained by two paradigms: Domain-Adaptive Pretraining (DAPT), which improves short-term domain fitting but frequently disrupts previously learned global temporal patterns due to catastrophic forgetting; and Retrieval-Augmented Generation (RAG), which incorporates external knowledge but introduces substantial retrieval overhead. This creates a severe scalability bottleneck that fails to meet the high-efficiency requirements of real-time stream processing. To break this impasse, we propose Memory for Time Series (MEMTS), a lightweight and plug-and-play method for retrieval-free domain adaptation in time series forecasting. The key component of MEMTS is a Knowledge Persistence Module (KPM), which internalizes domain-specific temporal dynamics, such as recurring seasonal patterns and trends into a compact set of learnable latent prototypes. In doing so, it transforms fragmented historical observations into continuous, parameterized knowledge representations. This paradigm shift enables MEMTS to achieve accurate domain adaptation with constant-time inference and near-zero latency, while effectively mitigating catastrophic forgetting of general temporal patterns, all without requiring any architectural modifications to the frozen TSFM backbone. Extensive experiments on multiple datasets demonstrate the SOTA performance of MEMTS.

Method: MechPert is a lightweight framework where multiple LLM agents independently propose candidate regulators with confidence scores, which are aggregated through a consensus mechanism to filter spurious associations and produce weighted neighborhoods for downstream prediction.

Result: On Perturb-seq benchmarks across four human cell lines, MechPert improves Pearson correlation by up to 10.5% over similarity-based baselines in low-data regimes (N=50 observed perturbations). For experimental design, MechPert-selected anchor genes outperform standard network centrality heuristics by up to 46% in well-characterized cell lines.

Conclusion: MechPert demonstrates that LLM agents can effectively generate directed regulatory hypotheses for perturbation prediction, offering significant improvements over existing methods in low-data scenarios and experimental design applications.

Abstract: Predicting transcriptional responses to unseen genetic perturbations is essential for understanding gene regulation and prioritizing large-scale perturbation experiments. Existing approaches either rely on static, potentially incomplete knowledge graphs, or prompt language models for functionally similar genes, retrieving associations shaped by symmetric co-occurrence in scientific text rather than directed regulatory logic. We introduce MechPert, a lightweight framework that encourages LLM agents to generate directed regulatory hypotheses rather than relying solely on functional similarity. Multiple agents independently propose candidate regulators with associated confidence scores; these are aggregated through a consensus mechanism that filters spurious associations, producing weighted neighborhoods for downstream prediction. We evaluate MechPert on Perturb-seq benchmarks across four human cell lines. For perturbation prediction in low-data regimes ($N=50$ observed perturbations), MechPert improves Pearson correlation by up to 10.5% over similarity-based baselines. For experimental design, MechPert-selected anchor genes outperform standard network centrality heuristics by up to 46% in well-characterized cell lines.

Method: Proposes Cast-R1 framework with memory-based state management, tool-augmented agentic workflow (statistical feature extraction, lightweight forecasting models, reasoning-based prediction, self-reflection), and two-stage training combining supervised fine-tuning with multi-turn reinforcement learning and curriculum learning.

Result: Extensive experiments on multiple real-world time series datasets demonstrate the effectiveness of Cast-R1 framework for time series forecasting.

Conclusion: Cast-R1 provides a practical step toward exploring agentic paradigms for time series modeling, showing promise for complex forecasting scenarios requiring autonomous reasoning and iterative refinement.

Abstract: Time series forecasting has long been dominated by model-centric approaches that formulate prediction as a single-pass mapping from historical observations to future values. Despite recent progress, such formulations often struggle in complex and evolving settings, largely because most forecasting models lack the ability to autonomously acquire informative evidence, reason about potential future changes, or revise predictions through iterative decision processes. In this work, we propose Cast-R1, a learned time series forecasting framework that reformulates forecasting as a sequential decision-making problem. Cast-R1 introduces a memory-based state management mechanism that maintains decision-relevant information across interaction steps, enabling the accumulation of contextual evidence to support long-horizon reasoning. Building on this formulation, forecasting is carried out through a tool-augmented agentic workflow, in which the agent autonomously interacts with a modular toolkit to extract statistical features, invoke lightweight forecasting models for decision support, perform reasoning-based prediction, and iteratively refine forecasts through self-reflection. To train Cast-R1, we adopt a two-stage learning strategy that combines supervised fine-tuning with multi-turn reinforcement learning, together with a curriculum learning scheme that progressively increases task difficulty to improve policy learning. Extensive experiments on multiple real-world time series datasets demonstrate the effectiveness of Cast-R1. We hope this work provides a practical step towards further exploration of agentic paradigms for time series modeling. Our code is available at https://github.com/Xiaoyu-Tao/Cast-R1-TS.

Method: Proposes a PDF solver using spectral-domain dimensionality reduction by expanding induced currents with a truncated Fourier basis, confining optimization to a compact low-frequency parameter space. Integrates contraction integral equation (CIE) to mitigate high-contrast nonlinearity and contrast-compensated operator (CCO) to correct spectral-induced attenuation. Also formulates a bridge-suppressing loss to enhance boundary sharpness between scatterers.

Result: Achieves 100-fold speedup over state-of-the-art UNNs with robust performance under noise and antenna uncertainties, enabling real-time microwave imaging applications.

Conclusion: The PDF solver enables real-time electromagnetic inverse scattering reconstruction through efficient spectral-domain optimization, overcoming computational limitations of previous UNN approaches.

Abstract: Untrained neural networks (UNNs) offer high-fidelity electromagnetic inverse scattering reconstruction but are computationally limited by high-dimensional spatial-domain optimization. We propose a Real-Time Physics-Driven Fourier-Spectral (PDF) solver that achieves sub-second reconstruction through spectral-domain dimensionality reduction. By expanding induced currents using a truncated Fourier basis, the optimization is confined to a compact low-frequency parameter space supported by scattering measurements. The solver integrates a contraction integral equation (CIE) to mitigate high-contrast nonlinearity and a contrast-compensated operator (CCO) to correct spectral-induced attenuation. Furthermore, a bridge-suppressing loss is formulated to enhance boundary sharpness between adjacent scatterers. Numerical and experimental results demonstrate a 100-fold speedup over state-of-the-art UNNs with robust performance under noise and antenna uncertainties, enabling real-time microwave imaging applications.

Method: AnomaMind uses a structured workflow that: 1) progressively localizes anomalies coarse-to-fine, 2) augments detection through multi-turn tool interactions for adaptive feature preparation, and 3) refines decisions via self-reflection. It employs a hybrid inference mechanism where general-purpose models handle tool interaction and refinement, while core anomaly detection is learned through reinforcement learning with workflow-level feedback.

Result: Extensive experiments across diverse settings demonstrate that AnomaMind consistently improves anomaly detection performance compared to existing methods.

Conclusion: AnomaMind successfully addresses limitations of traditional anomaly detection by framing it as an evidence-driven diagnostic process with adaptive reasoning, showing promise for complex real-world applications.

Abstract: Time series anomaly detection is critical in many real-world applications, where effective solutions must localize anomalous regions and support reliable decision-making under complex settings. However, most existing methods frame anomaly detection as a purely discriminative prediction task with fixed feature inputs, rather than an evidence-driven diagnostic process. As a result, they often struggle when anomalies exhibit strong context dependence or diverse patterns. We argue that these limitations stem from the lack of adaptive feature preparation, reasoning-aware detection, and iterative refinement during inference. To address these challenges, we propose AnomaMind, an agentic time series anomaly detection framework that reformulates anomaly detection as a sequential decision-making process. AnomaMind operates through a structured workflow that progressively localizes anomalous intervals in a coarse-to-fine manner, augments detection through multi-turn tool interactions for adaptive feature preparation, and refines anomaly decisions via self-reflection. The workflow is supported by a set of reusable tool engines, enabling context-aware diagnostic analysis. A key design of AnomaMind is an explicitly designed hybrid inference mechanism for tool-augmented anomaly detection. In this mechanism, general-purpose models are responsible for autonomous tool interaction and self-reflective refinement, while core anomaly detection decisions are learned through reinforcement learning under verifiable workflow-level feedback, enabling task-specific optimization within a flexible reasoning framework. Extensive experiments across diverse settings demonstrate that AnomaMind consistently improves anomaly detection performance. The code is available at https://anonymous.4open.science/r/AnomaMind.