Method: Uses Controlled Multimodal FluxFormer (flow-based Transformer) for controllable latent trajectories, Audio-Visual Nexus Alignment for temporal consistency, cross-gated modulation for multiplicative control, and dynamic modality-masking curriculum training.

Result: Achieves competitive quality on MUSIC benchmark, supports effective multimodal-guided music generation and targeted editing with robust inference under missing modalities.

Conclusion: MAGE provides a lightweight, flexible framework for practical music workflows that unifies multimodal music generation and editing with improved cross-modal grounding.

Abstract: Multimodal music creation requires models that can both generate audio from high-level cues and edit existing mixtures in a targeted manner. Yet most multimodal music systems are built for a single task and a fixed prompting interface, making their conditioning brittle when guidance is ambiguous, temporally misaligned, or partially missing. Common additive fusion or feature concatenation further weakens cross-modal grounding, often causing prompt drift and spurious musical content during generation and editing. We propose MAGE, a modality-agnostic framework that unifies multimodal music generation and mixture-grounded editing within a single continuous latent formulation. At its core, MAGE uses a Controlled Multimodal FluxFormer, a flow-based Transformer that learns controllable latent trajectories for synthesis and editing under any available subset of conditions. To improve grounding, we introduce Audio-Visual Nexus Alignment to select temporally consistent visual evidence for the audio timeline, and a cross-gated modulation mechanism that applies multiplicative control from aligned visual and textual cues to the audio latents, suppressing unsupported components rather than injecting them. Finally, we train with a dynamic modality-masking curriculum that exposes the model to text-only, visual-only, joint multimodal, and mixture-guided settings, enabling robust inference under missing modalities without training separate models. Experiments on the MUSIC benchmark show that MAGE supports effective multimodal-guided music generation and targeted editing, achieving competitive quality while offering a lightweight and flexible interface tailored to practical music workflows.

Method: Systematic analysis of Audio Flamingo 3 to identify gaps, curation of large-scale datasets (1M+ hours), curriculum-based training (pre-training, mid-training, post-training), and introduction of Temporal Audio Chain-of-Thought for timestamp-grounded reasoning.

Result: Outperforms similarly sized open models by large margins across 20 benchmarks, competitive with larger models, exhibits strong real-world utility and generalization to unseen tasks.

Conclusion: AF-Next represents a significant advancement in audio-language modeling with improved capabilities for understanding and reasoning over diverse audio types, especially for long and complex audio inputs.

Abstract: We present Audio Flamingo Next (AF-Next), the next-generation and most capable large audio-language model in the Audio Flamingo series, designed to advance understanding and reasoning over speech, environmental sounds and music. Compared to Audio Flamingo 3, AF-Next introduces: (i) a stronger foundational audio-language model that significantly improves accuracy across diverse audio understanding tasks; (ii) scalable strategies for constructing large-scale audio understanding and reasoning data beyond existing academic benchmarks; (iii) support for long and complex audio inputs up to 30 minutes; and (iv) Temporal Audio Chain-of-Thought, a new reasoning paradigm that explicitly grounds intermediate reasoning steps to timestamps in long audio, enabling fine-grained temporal alignment and improved interpretability. To enable these capabilities, we first conduct a systematic analysis of Audio Flamingo 3 to identify key gaps in audio understanding and reasoning. We then curate and scale new large-scale datasets totaling over 1 million hours to address these limitations and expand the existing AudioSkills-XL, LongAudio-XL, AF-Think and AF-Chat datasets. AF-Next is trained using a curriculum-based strategy spanning pre-training, mid-training and post-training stages. Extensive experiments across 20 audio understanding and reasoning benchmarks, including challenging long-audio tasks, show that AF-Next outperforms similarly sized open models by large margins and remains highly competitive with and sometimes surpasses, much larger open-weight and closed models. Beyond benchmark performance, AF-Next exhibits strong real-world utility and transfers well to unseen tasks, highlighting its robustness and generalization ability. In addition to all data, code and methods, we open-source 3 variants of AF-Next, including AF-Next-Instruct, AF-Next-Think and AF-Next-Captioner.

Method: Combines a frozen Multimodal Large Language Model for high-level reasoning with a trainable Diffusion Transformer for high-fidelity synthesis, and introduces AudioEdit dataset with over 1M curated editing pairs to overcome data scarcity in audio editing.

Result: Achieves state-of-the-art performance across benchmarks, outperforming prior unified approaches while matching or surpassing specialized expert models, with additional capabilities like knowledge-augmented reasoning, in-context generation, and zero-shot cross-lingual control.

Conclusion: Audio-Omni represents a promising direction toward universal generative audio intelligence by unifying generation and editing across domains with integrated multimodal understanding.

Abstract: Recent progress in multimodal models has spurred rapid advances in audio understanding, generation, and editing. However, these capabilities are typically addressed by specialized models, leaving the development of a truly unified framework that can seamlessly integrate all three tasks underexplored. While some pioneering works have explored unifying audio understanding and generation, they often remain confined to specific domains. To address this, we introduce Audio-Omni, the first end-to-end framework to unify generation and editing across general sound, music, and speech domains, with integrated multi-modal understanding capabilities. Our architecture synergizes a frozen Multimodal Large Language Model for high-level reasoning with a trainable Diffusion Transformer for high-fidelity synthesis. To overcome the critical data scarcity in audio editing, we construct AudioEdit, a new large-scale dataset comprising over one million meticulously curated editing pairs. Extensive experiments demonstrate that Audio-Omni achieves state-of-the-art performance across a suite of benchmarks, outperforming prior unified approaches while achieving performance on par with or superior to specialized expert models. Beyond its core capabilities, Audio-Omni exhibits remarkable inherited capabilities, including knowledge-augmented reasoning generation, in-context generation, and zero-shot cross-lingual control for audio generation, highlighting a promising direction toward universal generative audio intelligence. The code, model, and dataset will be publicly released on https://zeyuet.github.io/Audio-Omni.

Method: SECL uses test-time training (TTT) that exploits the gap between LLMs’ verbalized confidence and their internal P(True) signal as label-free self-supervision. It adapts only when input distribution shifts, training on 6-26% of question streams, and uses distillation from the baseline P(True) signal.

Result: Across four small language models from three families and four diverse domains, SECL reduces Expected Calibration Error (ECE) by 56-78%, outperforming its own supervision signal and matching or outperforming recent inference-time methods.

Conclusion: SECL is the first method to apply test-time training to calibration, requiring no labeled data or human supervision. Seven ablations confirm each component is crucial and robust across configurations.

Abstract: Large language models (LLMs) are systematically overconfident: they routinely express high certainty on questions they often answer incorrectly. Existing calibration methods either require labeled validation data, degrade under distribution shifts, or incur substantial inference costs. Recent work has shown that LLMs already contain a better-calibrated signal than the one they verbalize: the token probability of “True” when the model is asked “Is this answer correct?” ($P(\text{True})$) consistently outperforms their stated confidence, a gap that is theoretically grounded as generative error is lower-bounded by roughly twice the corresponding discriminative error. We introduce $\textbf{SECL}$ ($\textbf{SE}$lf-$\textbf{C}$alibrating $\textbf{L}$anguage Models), a test-time training (TTT) pipeline that exploits this gap as label-free self-supervision, requiring no labeled data or human supervision. SECL adapts only when the input distribution shifts, training on just 6–26% of the question stream at lower cost than the baseline it distills from. Across four small language models from three model families and four diverse domains, SECL reduces Expected Calibration Error (ECE) by 56–78%, outperforming its own supervision signal and matching or outperforming recent inference-time methods. SECL is the first method to apply TTT to calibration; seven ablations covering signal quality, gating strategy, weight accumulation, loss design, domain ordering, hyperparameter sensitivity, and layer selection confirm that each component is crucial and robust across configurations. Code: https://anonymous.4open.science/r/secl-emnlp26-submission-C890

Method: Two strategies: 1) Continued pre-training of BERT models on unlabelled web texts before fine-tuning, 2) Using four open-source LLMs to produce synthetic annotations through ensemble strategies (mean averaging, majority voting, LightGBM meta-learner).

Result: Continued pre-training yields ~3% average macro-F1 gain across 16 benchmarks. LightGBM ensemble outperforms other strategies. Synthetic labels benefit small models significantly (+11% F1 for Llama3.2-1B) but provide modest gains for larger models (+0.6% for Qwen2.5-14B).

Conclusion: Combination of web-scale unlabelled data and LLM-ensemble annotations is most valuable for smaller models and low-resource languages in hate speech detection.

Abstract: We study whether large-scale unlabelled web data and LLM-based synthetic annotations can improve multilingual hate speech detection. Starting from texts crawled via OpenWebSearch.eu~(OWS) in four languages (English, German, Spanish, Vietnamese), we pursue two complementary strategies. First, we apply continued pre-training to BERT models by continuing masked language modelling on unlabelled OWS texts before supervised fine-tuning, and show that this yields an average macro-F1 gain of approximately 3% over standard baselines across sixteen benchmarks, with stronger gains in low-resource settings. Second, we use four open-source LLMs (Mistral-7B, Llama3.1-8B, Gemma2-9B, Qwen2.5-14B) to produce synthetic annotations through three ensemble strategies: mean averaging, majority voting, and a LightGBM meta-learner. The LightGBM ensemble consistently outperforms the other strategies. Fine-tuning on these synthetic labels substantially benefits a small model (Llama3.2-1B: +11% pooled F1), but provides only a modest gain for the larger Qwen2.5-14B (+0.6%). Our results indicate that the combination of web-scale unlabelled data and LLM-ensemble annotations is the most valuable for smaller models and low-resource languages.

Method: Cognitive Synergy Framework with Mixture-of-Thought approach using six cognitive personas to synthesize diverse comedic perspectives, creating theoretically grounded humor datasets for fine-tuning 7B parameter models with DPO and novel O-GRPO optimization.

Result: 7B model significantly outperforms larger instruction-tuned baselines and achieves performance competitive with state-of-the-art proprietary models, showing cognitive-driven data curation is more critical than alignment algorithms or model scale.

Conclusion: Psychological theory-driven data curation is key to effective humor generation in LLMs, with the proposed framework enabling competitive performance with smaller models through cognitive synergy.

Abstract: Humor generation poses a significant challenge for Large Language Models (LLMs), because their standard training objective - predicting the most likely next word - inherently conflicts with the surprise and incongruity needed for comedy. To bridge this gap, we introduce the Cognitive Synergy Framework, a theoretically grounded methodology for generating high-quality humor data inspired by psychological theories of humor. Utilizing a Mixture-of-Thought (MoT) approach, we deploy six cognitive personas (e.g., The Absurdist, The Cynic) to synthesize diverse comedic perspectives for a given prompt. This framework creates a theoretically grounded dataset, which we use to fine-tune a 7B-parameter student model. We compare Direct Preference Optimization (DPO) and a novel Offline Group Relative Policy Optimization (O-GRPO); our 7B model significantly outperforms larger instruction-tuned baselines and achieves performance competitive with state-of-the-art proprietary models. We find that cognitive-driven data curation is far more critical than alignment algorithms or model scale for humor generation. Code and data will be available upon publication.

Method: ASPIRin uses Action Space Projection to map text vocabulary into binary states (speak/silence), applies Group Relative Policy Optimization with rule-based rewards, and isolates timing decisions from token selection to preserve semantic coherence.

Result: ASPIRin optimizes interactivity across turn-taking, backchanneling, and pause handling while reducing duplicate n-grams by over 50% compared to standard GRPO, effectively eliminating degenerative repetition.

Conclusion: Decoupling timing decisions from content generation in full-duplex speech language models prevents semantic degradation while enabling effective optimization of interactive behaviors.

Abstract: End-to-end full-duplex Speech Language Models (SLMs) require precise turn-taking for natural interaction. However, optimizing temporal dynamics via standard raw-token reinforcement learning (RL) degrades semantic quality, causing severe generative collapse and repetition. We propose ASPIRin, an interactivity-optimized RL framework that explicitly decouples when to speak from what to say. Using Action Space Projection, ASPIRin maps the text vocabulary into a coarse-grained binary state (active speech vs. inactive silence). By applying Group Relative Policy Optimization (GRPO) with rule-based rewards, it balances user interruption and response latency. Empirical evaluations show ASPIRin optimizes interactivity across turn-taking, backchanneling, and pause handling. Crucially, isolating timing from token selection preserves semantic coherence and reduces the portion of duplicate n-grams by over 50% compared to standard GRPO, effectively eliminating degenerative repetition.

Method: Developed a pipeline using Dutch fine-tuned Large Language Models to generate synthetic medical dialogues, with real medical conversations as linguistic/structural reference. Evaluation involved quantitative metrics (lexical variety, turn-taking patterns) and qualitative review by native speakers and medical practitioners.

Result: Quantitative analysis showed strong lexical variety but overly regular turn-taking (suggesting scripted rather than natural flow). Qualitative review produced slightly below-average scores, with issues in domain specificity and natural expression. Limited correlation between quantitative and qualitative results was observed.

Conclusion: Generating synthetic Dutch medical dialogues is feasible but requires domain knowledge and carefully structured prompting to balance naturalness and structure. Numerical metrics alone cannot fully capture linguistic quality. This work provides foundation for expanding Dutch clinical NLP resources through ethically generated synthetic data.

Abstract: Medical conversations offer insights into clinical communication often absent from Electronic Health Records. However, developing reliable clinical Natural Language Processing (NLP) models is hampered by the scarcity of domain-specific datasets, as clinical data are typically inaccessible due to privacy and ethical constraints. To address these challenges, we present a pipeline for generating synthetic Dutch medical dialogues using a Dutch fine-tuned Large Language Model, with real medical conversations serving as linguistic and structural reference. The generated dialogues were evaluated through quantitative metrics and qualitative review by native speakers and medical practitioners. Quantitative analysis revealed strong lexical variety and overly regular turn-taking, suggesting scripted rather than natural conversation flow. Qualitative review produced slightly below-average scores, with raters noting issues in domain specificity and natural expression. The limited correlation between quantitative and qualitative results highlights that numerical metrics alone cannot fully capture linguistic quality. Our findings demonstrate that generating synthetic Dutch medical dialogues is feasible but requires domain knowledge and carefully structured prompting to balance naturalness and structure in conversation. This work provides a foundation for expanding Dutch clinical NLP resources through ethically generated synthetic data.

Method: Developed GiantsBench (17k examples across 8 domains), used LM judge for evaluation, trained GIANTS-4B via reinforcement learning using similarity scores as proxy reward.

Result: GIANTS-4B outperforms proprietary baselines, achieves 34% relative improvement over gemini-3-pro, produces more conceptually clear insights, and SciJudge-30B predicts its insights lead to higher citations.

Conclusion: Language models can be effectively trained for insight anticipation, demonstrating potential for automated scientific discovery through literature synthesis.

Abstract: Scientific breakthroughs often emerge from synthesizing prior ideas into novel contributions. While language models (LMs) show promise in scientific discovery, their ability to perform this targeted, literature-grounded synthesis remains underexplored. We introduce insight anticipation, a generation task in which a model predicts a downstream paper’s core insight from its foundational parent papers. To evaluate this capability, we develop GiantsBench, a benchmark of 17k examples across eight scientific domains, where each example consists of a set of parent papers paired with the core insight of a downstream paper. We evaluate models using an LM judge that scores similarity between generated and ground-truth insights, and show that these similarity scores correlate with expert human ratings. Finally, we present GIANTS-4B, an LM trained via reinforcement learning (RL) to optimize insight anticipation using these similarity scores as a proxy reward. Despite its smaller open-source architecture, GIANTS-4B outperforms proprietary baselines and generalizes to unseen domains, achieving a 34% relative improvement in similarity score over gemini-3-pro. Human evaluations further show that GIANTS-4B produces insights that are more conceptually clear than those of the base model. In addition, SciJudge-30B, a third-party model trained to compare research abstracts by likely citation impact, predicts that insights generated by GIANTS-4B are more likely to lead to higher citations, preferring them over the base model in 68% of pairwise comparisons. We release our code, benchmark, and model to support future research in automated scientific discovery.

Method: Fine-tune a multilingual encoder using contrastive learning with similar multilingual claim pairs to create improved semantic embeddings optimized for fact-check claims.

Result: Claim2Vec significantly improves clustering performance across three datasets, 14 multilingual embedding models, and 7 clustering algorithms. Enhances both cluster label alignment and geometric structure of embedding space, with cross-lingual knowledge transfer observed in multilingual clusters.

Conclusion: Claim2Vec is the first multilingual embedding model optimized for fact-check claims, demonstrating effectiveness for claim clustering tasks and enabling better handling of recurrent misinformation across languages.

Abstract: Recurrent claims present a major challenge for automated fact-checking systems designed to combat misinformation, especially in multilingual settings. While tasks such as claim matching and fact-checked claim retrieval aim to address this problem by linking claim pairs, the broader challenge of effectively representing groups of similar claims that can be resolved with the same fact-check via claim clustering remains relatively underexplored. To address this gap, we introduce Claim2Vec, the first multilingual embedding model optimized to represent fact-check claims as vectors in an improved semantic embedding space. We fine-tune a multilingual encoder using contrastive learning with similar multilingual claim pairs. Experiments on the claim clustering task using three datasets, 14 multilingual embedding models, and 7 clustering algorithms demonstrate that Claim2Vec significantly improves clustering performance. Specifically, it enhances both cluster label alignment and the geometric structure of the embedding space across different cluster configurations. Our multilingual analysis shows that clusters containing multiple languages benefit from fine-tuning, demonstrating cross-lingual knowledge transfer.

Method: Introduces 100-Endings metric: walks through stories sentence by sentence, predicts 100 possible endings at each position, measures tension as prediction mismatch rate. Also analyzes sentence-level curve statistics like inflection rate. Designs story-generation pipeline with structural constraints including story template analysis, idea formulation, and narrative scaffolding.

Result: 100-Endings correctly ranks New Yorker stories far above LLM outputs (unlike rubric-based judges). The proposed story-generation pipeline significantly increases narrative tension as measured by 100-Endings while maintaining performance on EQ-Bench leaderboard.

Conclusion: Narrative tension is a crucial missing dimension in story evaluation. The 100-Endings metric provides a better measure of compelling storytelling, and structural constraints can improve LLM story generation by increasing narrative tension.

Abstract: LLMs have so far failed both to generate consistently compelling stories and to recognize this failure–on the leading creative-writing benchmark (EQ-Bench), LLM judges rank zero-shot AI stories above New Yorker short stories, a gold standard for literary fiction. We argue that existing rubrics overlook a key dimension of compelling human stories: narrative tension. We introduce the 100-Endings metric, which walks through a story sentence by sentence: at each position, a model predicts how the story will end 100 times given only the text so far, and we measure tension as how often predictions fail to match the ground truth. Beyond the mismatch rate, the sentence-level curve yields complementary statistics, such as inflection rate, a geometric measure of how frequently the curve reverses direction, tracking twists and revelations. Unlike rubric-based judges, 100-Endings correctly ranks New Yorker stories far above LLM outputs. Grounded in narratological principles, we design a story-generation pipeline using structural constraints, including analysis of story templates, idea formulation, and narrative scaffolding. Our pipeline significantly increases narrative tension as measured by the 100-Endings metric, while maintaining performance on the EQ-Bench leaderboard.

Method: Proposes Organized Group Behavior Simulation task, GROVE benchmark with 44 entities and 8,052 context-decision pairs, and structured analytical framework converting decisions into interpretable behavioral models with time-aware adapters and group-aware transfer.

Result: Framework outperforms summarization- and retrieval-based baselines, captures temporal behavioral drift within groups, enables knowledge transfer for data-scarce organizations, and provides traceable evidence nodes.

Conclusion: Provides comprehensive research platform for group behavior understanding with practical applications, demonstrating effectiveness of structured analytical approach over simple prompting methods.

Abstract: Simulating how organized groups (e.g., corporations) make decisions (e.g., responding to a competitor’s move) is essential for understanding real-world dynamics and could benefit relevant applications (e.g., market prediction). In this paper, we formalize this problem as a concrete research platform for group behavior understanding, providing: (1) a task definition with benchmark and evaluation criteria, (2) a structured analytical framework with a corresponding algorithm, and (3) detailed temporal and cross-group analysis. Specifically, we propose Organized Group Behavior Simulation, a task that models organized groups as collective entities from a practical perspective: given a group facing a particular situation (e.g., AI Boom), predict the decision it would take. To support this task, we present GROVE (GRoup Organizational BehaVior Evaluation), a benchmark covering 44 entities with 8,052 real-world context-decision pairs collected from Wikipedia and TechCrunch across 9 domains, with an end-to-end evaluation protocol assessing consistency, initiative, scope, magnitude, and horizon. Beyond straightforward prompting pipelines, we propose a structured analytical framework that converts collective decision-making events into an interpretable, adaptive, and traceable behavioral model, achieving stronger performance than summarization- and retrieval-based baselines. It further introduces an adapter mechanism for time-aware evolution and group-aware transfer, and traceable evidence nodes grounding each decision rule in originating historical events. Our analysis reveals temporal behavioral drift within individual groups, which the time-aware adapter effectively captures for stronger prediction, and structured cross-group similarity that enables knowledge transfer for data-scarce organizations.

Method: Developed an automated annotation protocol to detect three types of reasoning errors in translation: source sentence-misaligned, model hypothesis-misaligned, and reasoning trace-misaligned. Tested weak-to-strong interventions (hedging, removal, re-reasoning after removal, hindsight, and oracle interventions) on perturbed reasoning traces to correct identified errors.

Result: Reasoning errors can be identified with high precision in Urdu but lower precision in Spanish. Small corrections to reasoning traces have little impact on translation quality, while stronger interventions yield highest resolution rates but mixed translation quality gains. Removing reasoning errors doesn’t significantly resolve initial errors.

Conclusion: Machine translation systems exhibit limited reasoning faithfulness - while reasoning errors can be identified, correcting them doesn’t substantially improve translation quality, suggesting deeper issues in how reasoning is integrated into translation models.

Abstract: Across multiple language pairings (English $\to$ {Spanish, French, German, Mandarin, Japanese, Urdu, Cantonese}), we find reasoning errors in translation. To quantify how often these reasoning errors occur, we leverage an automated annotation protocol for reasoning evaluation wherein the goal is to detect if a reasoning step is any of three error categories: (1) source sentence-misaligned, (2) model hypothesis-misaligned, or (3) reasoning trace-misaligned. We probe the reasoning model with perturbed traces correcting for these identified reasoning errors using an array of weak-to-strong interventions: hedging, removal, re-reasoning after removal, hindsight, and oracle interventions. Experimenting with interventions on the reasoning traces suggests that small corrections to the reasoning have little impact on translation quality, but stronger interventions yield the highest resolution rates, despite translation quality gains being mixed. We find ultimately that reasoning errors in MT can be identified with high precision in Urdu but lower precision in Spanish, but that removing these reasoning errors does not resolve the initial errors significantly, suggesting limited reasoning faithfulness for machine translation.

Method: Constructed document-level features including sentence structure, lexical diversity, punctuation, readability indices, and emotion-based features. Applied multiple classification models (logistic regression, random forest, SVM, XGBoost, neural network) and ensemble methods.

Result: Strong classification performance with readability-based features as most informative predictors. AI-generated text shows more uniform stylistic patterns. Ensemble learning provides modest but consistent improvements.

Conclusion: Stylistic and structural properties of text provide a robust basis for distinguishing AI-generated misinformation from human-written fake news.

Abstract: The rapid adoption of large language models has introduced a new class of AI-generated fake news that coexists with traditional human-written misinformation, raising important questions about how these two forms of deceptive content differ and how reliably they can be distinguished. This study examines linguistic, structural, and emotional differences between human-written and AI-generated fake news and evaluates machine learning and ensemble-based methods for distinguishing these content types. A document-level feature representation is constructed using sentence structure, lexical diversity, punctuation patterns, readability indices, and emotion-based features capturing affective dimensions such as fear, anger, joy, sadness, trust, and anticipation. Multiple classification models, including logistic regression, random forest, support vector machines, extreme gradient boosting, and a neural network, are applied alongside an ensemble framework that aggregates predictions across models. Model performance is assessed using accuracy and area under the receiver operating characteristic curve. The results show strong and consistent classification performance, with readability-based features emerging as the most informative predictors and AI-generated text exhibiting more uniform stylistic patterns. Ensemble learning provides modest but consistent improvements over individual models. These findings indicate that stylistic and structural properties of text provide a robust basis for distinguishing AI-generated misinformation from human-written fake news.

Method: Extended replication study across expanded suite of models and datasets, testing various training-time, prompt-based interventions to mitigate weird generalization

Result: Weird generalization is exceptionally brittle, emerging only for specific models on specific datasets, and vanishes under simple interventions; most effective interventions provide prompt context that makes generalized behavior expected

Conclusion: Weird generalization poses a safety threat but is easily mitigated with simple solutions, clarifying the nature of the risk

Abstract: Weird generalization is a phenomenon in which models fine-tuned on data from a narrow domain (e.g. insecure code) develop surprising traits that manifest even outside that domain (e.g. broad misalignment)-a phenomenon that prior work has highlighted as a critical safety concern. Here, we present an extended replication study of key weird generalization results across an expanded suite of models and datasets. We confirm that surprising (and dangerous) traits can emerge under certain circumstances, but we find that weird generalization is exceptionally brittle: it emerges only for specific models on specific datasets, and it vanishes under simple training-time, prompt-based interventions. We find that the most effective interventions provide prompt context that makes the generalized behavior the expected behavior. However, we show that even very generic interventions that do not anticipate specific generalized traits can still be effective in mitigating weird generalization’s effects. Our findings thus help clarify the nature of the safety threat that weird generalization poses and point toward an easily implemented set of solutions.

Method: CoSToM framework: 1) Uses causal tracing to map internal distribution of ToM features and identify layers encoding fundamental ToM semantics. 2) Implements lightweight alignment via targeted activation steering within these ToM-critical layers.

Result: Experiments show CoSToM significantly enhances human-like social reasoning capabilities and improves downstream dialogue quality compared to baseline approaches.

Conclusion: The framework successfully transitions from mechanistic interpretation to active intervention, demonstrating that targeted activation steering in identified ToM-critical layers can align LLMs’ internal knowledge with external social reasoning behaviors.

Abstract: Theory of Mind (ToM), the ability to attribute mental states to others, is a hallmark of social intelligence. While large language models (LLMs) demonstrate promising performance on standard ToM benchmarks, we observe that they often fail to generalize to complex task-specific scenarios, relying heavily on prompt scaffolding to mimic reasoning. The critical misalignment between the internal knowledge and external behavior raises a fundamental question: Do LLMs truly possess intrinsic cognition, and can they externalize this internal knowledge into stable, high-quality behaviors? To answer this, we introduce CoSToM (Causal-oriented Steering for ToM alignment), a framework that transitions from mechanistic interpretation to active intervention. First, we employ causal tracing to map the internal distribution of ToM features, empirically uncovering the internal layers’ characteristics in encoding fundamental ToM semantics. Building on this insight, we implement a lightweight alignment framework via targeted activation steering within these ToM-critical layers. Experiments demonstrate that CoSToM significantly enhances human-like social reasoning capabilities and downstream dialogue quality.

Method: Proposes ODUTQA-MDC benchmark with large-scale dataset, fine-grained labeling, and dynamic clarification interface. Introduces MAIC-TQA multi-agent framework for ambiguity detection, clarification dialogue, and answer refinement.

Result: Experiments validate the benchmark and framework, establishing them as key resources for advancing conversational, underspecification-aware tabular QA research.

Conclusion: The work provides comprehensive resources and a multi-agent framework to address underspecified queries in tabular QA through interactive clarification, advancing the field toward more robust conversational systems.

Abstract: The advancement of large language models (LLMs) has enhanced tabular question answering (Tabular QA), yet they struggle with open-domain queries exhibiting underspecified or uncertain expressions. To address this, we introduce the ODUTQA-MDC task and the first comprehensive benchmark to tackle it. This benchmark includes: (1) a large-scale ODUTQA dataset with 209 tables and 25,105 QA pairs; (2) a fine-grained labeling scheme for detailed evaluation; and (3) a dynamic clarification interface that simulates user feedback for interactive assessment. We also propose MAIC-TQA, a multi-agent framework that excels at detecting ambiguities, clarifying them through dialogue, and refining answers. Experiments validate our benchmark and framework, establishing them as a key resource for advancing conversational, underspecification-aware Tabular QA research.

Method: Simplified algorithms implementing the indeterminate natural transformation model, evaluated through data fitting with experimental data, systematicity assessment of metaphor comprehension results, and novelty measurement of source-target associative structure correspondence.

Result: The improved algorithm outperformed existing implementations across all three evaluation measures: data-fitting with experimental data, systematicity of metaphor comprehension, and novelty of comprehension.

Conclusion: The computational implementation successfully operationalizes TINT theory for metaphor comprehension, with simplified algorithms demonstrating superior performance over previous approaches.

Abstract: In this study, we developed a computational implementation for a model of metaphor comprehension based on the theory of indeterminate natural transformation (TINT) proposed by Fuyama et al. We simplified the algorithms implementing the model to be closer to the original theory and verified it through data fitting and simulations. The outputs of the algorithms are evaluated with three measures: data-fitting with experimental data, the systematicity of the metaphor comprehension result, and the novelty of the comprehension (i.e. the correspondence of the associative structure of the source and target of the metaphor). The improved algorithm outperformed the existing ones in all the three measures.

Method: Uses Communication Accommodation Theory (CAT) framework with Language Inquiry and Word Count Lexicon (LIWC) analysis on Reddit data. Examines linguistic profiles within home communities and cross-community posting patterns, plus longitudinal analysis around diagnosis disclosure moments.

Result: Each community maintains distinct linguistic profiles that shift in opposite directions during cross-community engagement (convergent accommodation). Topic-independent variables show partial evidence against purely topical explanations. Diagnosis disclosure has small effects, sometimes opposite to accommodation patterns.

Conclusion: The study reveals complex intergroup communication dynamics between neurodivergent communities, suggesting situational audience adaptation and identity processes involve different mechanisms, with implications for community moderation and clinical perspectives.

Abstract: Social media research on mental health has focused predominantly on detecting and diagnosing conditions at the individual level. In this work, we shift attention to \emph{intergroup} behavior, examining how two prominent neurodivergent communities, ADHD and autism, adjust their language when engaging with each other on Reddit. Grounded in Communication Accommodation Theory (CAT), we first establish that each community maintains a distinct linguistic profile as measured by Language Inquiry and Word Count Lexicon (LIWC). We then show that these profiles shift in opposite directions when users cross community boundaries: features that are elevated in one group’s home community decrease when its members post in the other group’s space, and vice versa, consistent with convergent accommodation. The involvement of topic-independent summary variables (Authentic, Clout) in these shifts provides partial evidence against a purely topical explanation. Finally, in an exploratory longitudinal analysis around the moment of public diagnosis disclosure, we find that its effects on linguistic style are small and, in some cases, directionally opposite to cross-community accommodation, providing initial evidence that situational audience adaptation and longer-term identity processes may involve different mechanisms. Our findings contribute to understanding intergroup communication dynamics among neurodivergent populations online and carry implications for community moderation and clinical perspectives on these conditions.

Method: E-GRM leverages convergence behavior of parallel model generations to estimate uncertainty and selectively trigger CoT reasoning only when needed. It introduces a lightweight discriminative scorer trained with hybrid regression-ranking objective for fine-grained evaluation of reasoning paths.

Result: Experiments on multiple reasoning benchmarks show E-GRM substantially reduces inference cost while consistently improving answer accuracy.

Conclusion: Model-internal uncertainty is an effective and general signal for efficient reasoning-aware reward modeling, enabling selective CoT triggering without handcrafted features or task-dependent signals.

Abstract: Recent advancements in the Generative Reward Model (GRM) have demonstrated its potential to enhance the reasoning abilities of LLMs through Chain-of-Thought (CoT) prompting. Despite these gains, existing implementations of GRM suffer from two critical limitations. First, CoT prompting is applied indiscriminately to all inputs regardless of their inherent complexity. This introduces unnecessary computational costs for tasks amenable to fast, direct inference. Second, existing approaches primarily rely on voting-based mechanisms to evaluate CoT outputs, which often lack granularity and precision in assessing reasoning quality. In this paper, we propose E-GRM, an efficient generative reward modeling framework grounded in model-internal uncertainty. E-GRM leverages the convergence behavior of parallel model generations to estimate uncertainty and selectively trigger CoT reasoning only when needed, without relying on handcrafted features or task-dependent signals. To improve reward fidelity, we introduce a lightweight discriminative scorer trained with a hybrid regression–ranking objective to provide fine-grained evaluation of reasoning paths. Experiments on multiple reasoning benchmarks show that E-GRM substantially reduces inference cost while consistently improving answer accuracy, demonstrating that model-internal uncertainty is an effective and general signal for efficient reasoning-aware reward modeling.

Method: Formalize ILP as post-training failure to internalize supervised instances, demonstrate prevalence across model families/domains/datasets, identify five recurrent sources through controlled analyses, and introduce diagnostic-first framework mapping unlearned samples to causes using observable training/inference signals.

Result: Experiments on Qwen, LLaMA, and OLMo2 show incomplete learning is widespread and heterogeneous, with aggregate metric improvements masking persistent unlearned subsets. Targeted mitigation strategies studied as causal interventions.

Conclusion: Findings highlight need for fine-grained diagnosis of what supervised fine-tuning fails to learn and why, revealing limitations of current SFT approaches and importance of understanding failure modes beyond aggregate metrics.

Abstract: Supervised Fine-Tuning (SFT) is the standard approach for adapting large language models (LLMs) to downstream tasks. However, we observe a persistent failure mode: even after convergence, models often fail to correctly reproduce a subset of their own supervised training data. We refer to this behavior as the Incomplete Learning Phenomenon(ILP). This paper presents the first systematic study of ILP in LLM fine-tuning. We formalize ILP as post-training failure to internalize supervised instances and demonstrate its prevalence across multiple model families, domains, and datasets. Through controlled analyses, we identify five recurrent sources of incomplete learning: (1) missing prerequisite knowledge in the pre-trained model, (2) conflicts between SFT supervision and pre-training knowledge, (3) internal inconsistencies within SFT data, (4) left-side forgetting during sequential fine-tuning, and (5) insufficient optimization for rare or complex patterns. We introduce a diagnostic-first framework that maps unlearned samples to these causes using observable training and inference signals, and study several targeted mitigation strategies as causal interventions. Experiments on Qwen, LLaMA, and OLMo2 show that incomplete learning is widespread and heterogeneous, and that improvements in aggregate metrics can mask persistent unlearned subsets. The findings highlight the need for fine-grained diagnosis of what supervised fine-tuning fails to learn, and why.

Method: SEPTQ first calculates importance scores for each weight matrix element and determines quantization locations statically. Then uses a mask matrix representing important locations to quantize and update weights column-by-column until optimal quantized weights are obtained.

Result: SEPTQ outperforms other baselines across various datasets and model sizes (millions to billions of parameters) in different quantization bit-levels, with particularly strong performance in low-bit quantization scenarios.

Conclusion: SEPTQ provides an effective and efficient PTQ solution for LLMs that simplifies the quantization process to two steps while maintaining performance, especially valuable for low-bit deployment scenarios.

Abstract: Large language models (LLMs) have shown remarkable performance in various domains, but they are constrained by massive computational and storage costs. Quantization, an effective technique for compressing models to fit resource-limited devices while preserving generative quality, encompasses two primary methods: quantization aware training (QAT) and post-training quantization (PTQ). QAT involves additional retraining or fine-tuning, thus inevitably resulting in high training cost and making it unsuitable for LLMs. Consequently, PTQ has become the research hotspot in recent quantization methods. However, existing PTQ methods usually rely on various complex computation procedures and suffer from considerable performance degradation under low-bit quantization settings. To alleviate the above issues, we propose a simple and effective post-training quantization paradigm for LLMs, named SEPTQ. Specifically, SEPTQ first calculates the importance score for each element in the weight matrix and determines the quantization locations in a static global manner. Then it utilizes the mask matrix which represents the important locations to quantize and update the associated weights column-by-column until the appropriate quantized weight matrix is obtained. Compared with previous methods, SEPTQ simplifies the post-training quantization procedure into only two steps, and considers the effectiveness and efficiency simultaneously. Experimental results on various datasets across a suite of models ranging from millions to billions in different quantization bit-levels demonstrate that SEPTQ significantly outperforms other strong baselines, especially in low-bit quantization scenarios.

Method: Created ChangAn benchmark containing 30,664 poems (10,276 human-written, 20,388 generated by four popular LLMs). Systematically evaluated 12 AI detectors across different text granularities and generation strategies to assess their performance on classical Chinese poetry detection.

Result: Current Chinese text detectors fail as reliable tools for detecting LLM-generated classical Chinese poetry, highlighting the limitations of existing methods and validating the need for specialized benchmarks like ChangAn.

Conclusion: The ChangAn benchmark addresses a critical gap in AI text detection for classical Chinese poetry, demonstrating that current detectors are inadequate for this specialized literary domain and providing a valuable resource for future research.

Abstract: The rapid development of large language models (LLMs) has extended text generation tasks into the literary domain. However, AI-generated literary creations has raised increasingly prominent issues of creative authenticity and ethics in literary world, making the detection of LLM-generated literary texts essential and urgent. While previous works have made significant progress in detecting AI-generated text, it has yet to address classical Chinese poetry. Due to the unique linguistic features of classical Chinese poetry, such as strict metrical regularity, a shared system of poetic imagery, and flexible syntax, distinguishing whether a poem is authored by AI presents a substantial challenge. To address these issues, we introduce ChangAn, a benchmark for detecting LLM-generated classical Chinese poetry that containing total 30,664 poems, 10,276 are human-written poems and 20,388 poems are generated by four popular LLMs. Based on ChangAn, we conducted a systematic evaluation of 12 AI detectors, investigating their performance variations across different text granularities and generation strategies. Our findings highlight the limitations of current Chinese text detectors, which fail to serve as reliable tools for detecting LLM-generated classical Chinese poetry. These results validate the effectiveness and necessity of our proposed ChangAn benchmark. Our dataset and code are available at https://github.com/VelikayaScarlet/ChangAn.

Method: Created Context-Aware Stress TTS (CAST) benchmark with contrastive context pairs: identical sentences paired with distinct contexts requiring different stressed words. Evaluated state-of-the-art TTS systems using this benchmark to assess their ability to generate appropriate word-level stress from context.

Result: Found a consistent gap: text-only language models reliably recover intended stress from context, but TTS systems frequently fail to realize it in speech. The benchmark reveals limitations in current TTS systems’ ability to generate contextually appropriate emphasis.

Conclusion: There’s a significant gap between text understanding and speech realization in TTS systems regarding context-aware stress. The CAST benchmark, evaluation framework, construction pipeline, and synthetic corpus are released to support future work on context-aware speech synthesis.

Abstract: Spoken meaning often depends not only on what is said, but also on which word is emphasized. The same sentence can convey correction, contrast, or clarification depending on where emphasis falls. Although modern text-to-speech (TTS) systems generate expressive speech, it remains unclear whether they infer contextually appropriate stress from discourse alone. To address this gap, we present Context-Aware Stress TTS (CAST), a benchmark for evaluating context-conditioned word-level stress in TTS. Items are defined as contrastive context pairs: identical sentences paired with distinct contexts requiring different stressed words. We evaluate state-of-the-art systems and find a consistent gap: text-only language models reliably recover the intended stress from context, yet TTS systems frequently fail to realize it in speech. We release the benchmark, evaluation framework, construction pipeline and a synthetic corpus to support future work on context-aware speech synthesis.

Method: Decouples semantic reasoning from surface realization by distilling Teacher LLM reasoning into Probabilistic Sentential Decision Diagrams (PSDDs) to create tractable semantic priors that enforce hard logical constraints, plus convex optimization for soft distributional goals.

Result: Achieves 100% Schema Validity in complex logic puzzles (vs 12.4% for baselines) and significantly outperforms state-of-the-art methods in rare-combination coverage across diverse benchmarks.

Conclusion: CircuitSynth successfully bridges the gap between neural generation and symbolic reasoning, providing formal guarantees while maintaining expressivity for structured data generation.

Abstract: The generation of high-fidelity synthetic data is a cornerstone of modern machine learning, yet Large Language Models (LLMs) frequently suffer from hallucinations, logical inconsistencies, and mode collapse when tasked with structured generation. Existing approaches, such as prompting or retrieval-augmented generation, lack the mechanisms to balance linguistic expressivity with formal guarantees regarding validity and coverage. To address this, we propose CircuitSynth, a novel neuro-symbolic framework that decouples semantic reasoning from surface realization. By distilling the reasoning capabilities of a Teacher LLM into a Probabilistic Sentential Decision Diagram (PSDD), CircuitSynth creates a tractable semantic prior that structurally enforces hard logical constraints. Furthermore, we introduce a convex optimization mechanism to rigorously satisfy soft distributional goals. Empirical evaluations across diverse benchmarks demonstrate that CircuitSynth achieves 100% Schema Validity even in complex logic puzzles where unconstrained baselines fail (12.4%) while significantly outperforming state-of-the-art methods in rare-combination coverage.

Method: Created BlasBench, an open evaluation harness with Irish-aware text normalization that preserves fadas, lenition, and eclipsis. Benchmarked 12 systems across four architecture families on Common Voice ga-IE and FLEURS ga-IE datasets.

Result: All Whisper variants exceeded 100% WER. The best open model (omniASR LLM 7B) achieved 30.65% WER on Common Voice and 39.09% on FLEURS. Models fine-tuned on Common Voice lost 33-43 WER points on FLEURS, revealing a significant generalization gap.

Conclusion: Single-dataset evaluation masks generalization problems in ASR systems. The release of BlasBench enables proper Irish ASR evaluation and reveals important cross-dataset performance gaps that need addressing.

Abstract: No open Irish-specific benchmark compares end-user ASR systems under a shared Irish-aware evaluation protocol. To solve this, we release BlasBench, an open evaluation harness with Irish-aware text normalisation that preserves fadas, lenition, and eclipsis. We benchmark 12 systems across four architecture families on Common Voice ga-IE and FLEURS ga-IE. All Whisper variants exceed 100% WER. The best open model (omniASR LLM 7B) achieves 30.65% WER on Common Voice and 39.09% on FLEURS. We noticed models fine-tuned on Common Voice lose 33-43 WER points on FLEURS, revealing a generalisation gap that is invisible to single-dataset evaluation.

Method: Extract phone-level embeddings via Montreal Forced Aligner, compute d-prime scores along phonological contrast directions (nasality, voicing, stridency, sonorance, manner, vowel features) derived from healthy controls, and construct a 12-dimensional phonological profile using frozen HuBERT representations.

Result: Evaluated 890 speakers across 10 corpora, 5 languages, and 3 aetiologies; all five consonant d-prime features correlate significantly with clinical severity (rho = -0.50 to -0.56). Nasality d-prime decreases monotonically from control to severe in 6 of 7 severity-graded corpora.

Conclusion: The method requires no dysarthric training data, applies to any language with existing MFA acoustic model (29 languages), and provides a scalable, training-free approach for dysarthria severity assessment.

Abstract: Dysarthric speech severity assessment typically requires trained clinicians or supervised models built from labelled pathological speech, limiting scalability across languages and clinical settings. We present a training-free method that quantifies dysarthria severity by measuring degradation in phonological feature subspaces within frozen HuBERT representations. No supervised severity model is trained; feature directions are estimated from healthy control speech using a pretrained forced aligner. For each speaker, we extract phone-level embeddings via Montreal Forced Aligner, compute d-prime scores along phonological contrast directions (nasality, voicing, stridency, sonorance, manner, and four vowel features) derived exclusively from healthy controls, and construct a 12-dimensional phonological profile.Evaluating 890 speakers across 10 corpora, 5 languages (English, Spanish, Dutch, Mandarin, French), and 3 primary aetiologies (Parkinson’s disease, cerebral palsy, ALS), we find that all five consonant d-prime features correlate significantly with clinical severity (random-effects meta-analysis rho = -0.50 to -0.56, p < 2e-4; pooled Spearman rho = -0.47 to -0.55 with bootstrap 95% CIs not crossing zero). The effect replicates within individual corpora, survives FDR correction, and remains robust to leave-one-corpus-out removal and alignment quality controls. Nasality d-prime decreases monotonically from control to severe in 6 of 7 severity-graded corpora. Mann-Whitney U tests confirm that all 12 features distinguish controls from severely dysarthric speakers (p < 0.001).The method requires no dysarthric training data and applies to any language with an existing MFA acoustic model (currently 29 languages). We release the full pipeline and phone feature configurations for six languages.

Method: Proposes CSLM with novel alignment strategy achieving cross-modal and cross-lingual alignment through continual pre-training. Uses instruction fine-tuning following speech-text interleaved chain-of-modality generation process to enhance modal alignment at finer granularity.

Result: CSLM demonstrates strong cross-modal alignment capabilities and general task abilities on cross-modal tasks, mono-lingual conversational tasks, and cross-lingual conversational tasks. Aligns different modalities and languages simultaneously without massive speech data.

Conclusion: CSLM provides an efficient approach for building cross-lingual speech LLMs with good language scalability, improving generation quality and reducing latency through better modal alignment.

Abstract: Currently, large language models (LLMs) predominantly focus on the text modality. To enable more natural human-AI interaction, speech LLMs are emerging, but building effective end-to-end speech LLMs remains challenging due to limited data and the difficulty in expanding to more languages. In this paper, we introduce Cross-lingual Speech Language Model (CSLM), an efficient training method for cross-lingual speech LLMs based on discrete speech tokens. We propose a novel alignment strategy that achieves cross-modal and cross-lingual alignment through continual pre-training. By conducting instruction fine-tuning following a speech-text interleaved chain-of-modality generation process, we enhance modal alignment at a finer granularity, thereby improving generation quality and reducing latency. CSLM aligns different modalities and languages simultaneously without the need for massive speech data, thus exhibiting good language scalability. Evaluations on cross-modal tasks, mono-lingual conversational tasks, and cross-lingual conversational tasks demonstrate CSLM’s strong cross-modal alignment capabilities and general task abilities. (Code is available at: https://github.com/ictnlp/CSLM)

Method: Insert delimiters at sentence boundaries in LLM inputs to facilitate sentence-by-sentence processing behavior during reasoning. Two concrete methods: (1) In-context learning and (2) Supervised fine-tuning, experimented using models ranging from 7B to 600B parameters (Deepseek-V3).

Result: Consistent improvements across various tasks, with notable gains of up to 7.7% on GSM8k and 12.5% on DROP. Fine-tuned LLMs demonstrate sentence awareness through their internal representations.

Conclusion: Establishes a simple yet effective technique for enhancing LLM capabilities, offering promising directions for cognitive-inspired LLM enhancement paradigm by leveraging sentence-level structure.

Abstract: Researchers have explored different ways to improve large language models (LLMs)’ capabilities via dummy token insertion in contexts. However, existing works focus solely on the dummy tokens themselves, but fail to leverage the inherent sentence-level structure of natural language. This is a critical oversight, as LLMs acquire linguistic capabilities through exposure to human-generated texts, which are inherently structured at the sentence level. Motivated by this gap, we propose an approach that inserts delimiters at sentence boundaries in LLM inputs, which not only integrates dummy tokens into the context, but also facilitates LLMs with sentence-by-sentence processing behavior during reasoning. Two concrete methods: (1). In-context learning and (2). Supervised fine-tuning are experimented using 7B models to 600B Deepseek-V3. Our results demonstrate consistent improvements across various tasks, with notable gains of up to 7.7% on GSM8k and 12.5% on DROP. Furthermore, the fine-tuned LLMs can incorporate sentence awareness evidenced by their internal representations. Our work establishes a simple yet effective technique for enhancing LLM’s capabilities, offering promising directions for cognitive-inspired LLM enhancement paradigm.

Method: Generated 270 texts from 45 prompt templates crossed with six persona conditions in a 2x3 design. Used logistic regression probes on hidden-state activations across 35 layers with various controls (shuffled-label baselines, surface-text skyline classifier, cross-family tests, sentence-level baselines). Annotated probe-selected token positions for structural, lexical, and stance features using Stanza NLP pipeline.

Result: Nationality probe reached 0.968 cross-validated accuracy at Layer 18 with perfect held-out classification. Nationality encoding followed non-monotonic trajectory across layers. British-associated patterns showed more postmodification, hedging, boosting, passive voice, and evaluative/process-oriented vocabulary. Chinese-associated patterns showed more premodification, nominal predicates, and sociocultural/internationalisation vocabulary. No significant nationality differences found in full generated surface text at sentence level.

Conclusion: LLMs encode nationality-discriminative information in hidden states despite no surface-level differences, extending probing methodology to sociolinguistic attributes with implications for EAP and language pedagogy.

Abstract: Large language models are increasingly used as writing tools and pedagogical resources in English for Academic Purposes, but it remains unclear whether they encode culturally differentiated representations when generating academic text. This study tests whether Gemma-3-4b-it encodes nationality-discriminative information in hidden states when generating research article introductions conditioned by British and Chinese academic personas. A corpus of 270 texts was generated from 45 prompt templates crossed with six persona conditions in a 2 x 3 design. Logistic regression probes were trained on hidden-state activations across all 35 layers, with shuffled-label baselines, a surface-text skyline classifier, cross-family tests, and sentence-level baselines used as controls. Probe-selected token positions were annotated for structural, lexical, and stance features using the Stanza NLP pipeline. The nationality probe reached 0.968 cross-validated accuracy at Layer 18, with perfect held-out classification. Nationality encoding followed a non-monotonic trajectory across layers, with structural effects strongest in the middle to upper network and lexical-domain effects peaking earlier. At high-signal token positions, British-associated patterns showed more postmodification, hedging, boosting, passive voice, and evaluative or process-oriented vocabulary, while Chinese-associated patterns showed more premodification, nominal predicates, and sociocultural or internationalisation vocabulary. However, sentence-level analysis found no significant nationality differences in the full generated surface text. The findings extend probing methodology to a sociolinguistic attribute and have practical implications for EAP and language pedagogy.

Method: 1) Augment training datasets by computing confidence scores and semantic entropy from LLM outputs, mapping them to a knowledge state quadrant describing trustworthiness (knowledge possession) and honestness (answering behaviors) in natural language. 2) Design reward function considering both correctness and uncertainty signals, fine-tune LLM using PPO. 3) Add retrieval-augmented module to retrieve relevant external passages to improve consistency between internal and external knowledge.

Result: Extensive experiments on four knowledge-intensive benchmarks demonstrate that FAITH enhances factual accuracy and truthfulness of LLMs.

Conclusion: FAITH effectively improves LLM factuality by integrating natural-language uncertainty signals with external knowledge through a comprehensive framework addressing both internal knowledge states and external grounding.

Abstract: Large Language Models (LLMs) can generate factually inaccurate content even if they have corresponding knowledge, which critically undermines their reliability. Existing approaches attempt to mitigate this by incorporating uncertainty in QA prompt during training, but these numerical scores lack the semantic richness for LLM to properly understand its internal states of trustworthiness and honestness, leading to insufficient factuality alignment. We introduce FAITH (Factuality Alignment through Integrating Trustworthiness and Honestness), a post-training framework for factuality alignment that integrates natural-language uncertainty signals with external knowledge. Specifically, we augment training datasets by computing confidence scores and semantic entropy from LLM outputs and mapping them into a knowledge state quadrant that describes the model’s internal knowledge possession (trustworthiness) and answering behaviors (honestness) in natural language. Based on this enhanced data, we design a reward function that considers both correctness and uncertainty signals, and fine-tune the LLM using the Proximal Policy Optimization (PPO) algorithm. To further mitigate weakly grounded responses, we design a retrieval-augmented module that retrieves relevant external passages, improving the consistency between internal and external knowledge representations. Extensive experiments on four knowledge-intensive benchmarks demonstrate that FAITH enhances the factual accuracy and truthfulness of LLMs.

Method: Replace standard language model head with a relation head that induces relational graphs directly from LM hidden states. Train jointly with downstream task model for stock-trend prediction. Uses Qwen3 SLMs (0.6B/1.7B/4B) as upstream models.

Result: Relational Probing yields consistent performance improvements at competitive inference cost compared to co-occurrence baselines. Enables language model outputs to be reshaped into task-specific formats.

Conclusion: The approach learns semantic representations while preserving strict graph structure, allowing LMs to go beyond text generation into structured output tasks efficiently.

Abstract: Language models can be used to identify relationships between financial entities in text. However, while structured output mechanisms exist, prompting-based pipelines still incur autoregressive decoding costs and decouple graph construction from downstream optimization. We propose \emph{Relational Probing}, which replaces the standard language-model head with a relation head that induces a relational graph directly from language-model hidden states and is trained jointly with the downstream task model for stock-trend prediction. This approach both learns semantic representations and preserves the strict structure of the induced relational graph. It enables language-model outputs to go beyond text, allowing them to be reshaped into task-specific formats for downstream models. To enhance reproducibility, we provide an operational definition of small language models (SLMs): models that can be fine-tuned end-to-end on a single 24GB GPU under specified batch-size and sequence-length settings. Experiments use Qwen3 backbones (0.6B/1.7B/4B) as upstream SLMs and compare against a co-occurrence baseline. Relational Probing yields consistent performance improvements at competitive inference cost.

Method: Four-step pipeline: (1) Dimension Definition for qualitative behavioral criteria, (2) Augmentation into granular boolean checklist indicators, (3) Semantic Filtering for objectivity and redundancy removal, (4) LLM-as-a-Judge Evaluation with chain-of-thought verification.

Result: Framework validated on LLM Court (forensic training game) across five legal scenarios, revealing inverse relationship between parametric scale and procedural consistency, with smaller instruction-tuned models (8-9B) outperforming larger architectures.

Conclusion: RPA-Check provides standardized, reproducible metrics for generative agent evaluation in specialized domains, addressing limitations of current evaluation methods for interactive LLM systems.

Abstract: The rapid adoption of Large Language Models (LLMs) in interactive systems has enabled the creation of dynamic, open-ended Role-Playing Agents (RPAs). However, evaluating these agents remains a significant challenge, as standard NLP metrics fail to capture the nuances of role adherence, logical consistency, and long-term narrative stability. This paper introduces RPA-Check, a multi-stage automated evaluation framework designed to objectively assess the performance of LLM-based RPAs in complex, constraints-heavy environments. Our methodology is based on a four-step pipeline: (1) Dimension Definition, establishing high-level qualitative behavioral criteria; (2) Augmentation, where these requirements are expanded into granular boolean checklist indicators; (3) Semantic Filtering, to ensure indicator objectivity, no redundancy and agent isolation; and (4) LLM-as-a-Judge Evaluation, which employs chain-of-thought verification to score agent fidelity. We validate this framework by applying it to LLM Court, a serious game for forensic training involving several quantized local models. Experimental results across five distinct legal scenarios demonstrate the framework’s ability to identify subtle trade-offs between model size, reasoning depth, and operational stability. Notably, the findings reveal an inverse relationship between parametric scale and procedural consistency, showing that smaller, adequately instruction-tuned models (8-9B) can outperform larger architectures prone to user-alignment bias or sycophancy. RPA-Check thus provides a standardized and reproducible metric for future research in generative agent evaluation within specialized domains.

Method: CodeComp incorporates static program analysis into LLM inference via Code Property Graph priors extracted by Joern. It identifies and preserves structurally critical tokens (call sites, branch conditions, assignments) that attention-only methods would discard, enabling effective KV cache compression without model modification.

Result: CodeComp consistently outperforms attention-only compression baselines under equal memory budgets, recovers majority of full-context accuracy under aggressive KV cache compression, matches patch generation quality of uncompressed full-context inference, and integrates seamlessly into SGLang-based agentic coding pipelines.

Conclusion: Incorporating program analysis priors into KV cache compression is crucial for code understanding tasks, as attention signals alone are insufficient for identifying structurally critical tokens. CodeComp demonstrates that static analysis can significantly improve compression effectiveness without model retraining.

Abstract: Agentic code tasks such as fault localization and patch generation require processing long codebases under tight memory constraints, where the Key-Value (KV) cache becomes the primary inference bottleneck. Existing compression methods rely exclusively on attention signals to estimate token importance, systematically discarding structurally critical tokens such as call sites, branch conditions, and assignments that are essential for code understanding. We present CodeComp, a training-free KV cache compression framework that incorporates static program analysis into LLM inference via Code Property Graph priors extracted by Joern. Across bug localization and code generation benchmarks, CodeComp consistently outperforms attention-only compression baselines under equal memory budgets, recovering the majority of full-context accuracy under aggressive KV cache compression, while matching the patch generation quality of uncompressed full-context inference and integrating seamlessly into SGLang-based agentic coding pipelines without model modification.

Method: Evaluated multiple LLMs on core medical tasks (patient note summarization, medical QA) using open-access datasets (MedMCQA, PubMedQA, Asclepius) with linguistic and task-specific metrics.

Result: Domain-specific models (ChatDoctor) excel in contextual reliability and medical accuracy, while general-purpose models (Grok, LLaMA) perform better in structured QA tasks with higher quantitative accuracy.

Conclusion: LLMs can support medical professionals but require ethical standards, contextual accuracy, and human oversight; task-specific evaluation and cautious integration into healthcare workflows are essential.

Abstract: Background: Large Language Models (LLMs) are transforming artificial intelligence applications in healthcare due to their ability to understand, generate, and summarize complex medical text. They offer valuable support to clinicians, researchers, and patients, yet their deployment in high-stakes clinical environments raises critical concerns regarding accuracy, reliability, and patient safety. Despite substantial attention in recent years, standardized benchmarking of LLMs for medical applications has been limited. Objective: This study addresses the need for a standardized comparative evaluation of LLMs in medical settings. Method: We evaluate multiple models, including ChatGPT, LLaMA, Grok, Gemini, and ChatDoctor, on core medical tasks such as patient note summarization and medical question answering, using the open-access datasets, MedMCQA, PubMedQA, and Asclepius, and assess performance through a combination of linguistic and task-specific metrics. Results: The results indicate that domain-specific models, such as ChatDoctor, excel in contextual reliability, producing medically accurate and semantically aligned text, whereas general-purpose models like Grok and LLaMA perform better in structured question-answering tasks, demonstrating higher quantitative accuracy. This highlights the complementary strengths of domain-specific and general-purpose LLMs depending on the medical task. Conclusion: Our findings suggest that LLMs can meaningfully support medical professionals and enhance clinical decision-making; however, their safe and effective deployment requires adherence to ethical standards, contextual accuracy, and human oversight in relevant cases. These results underscore the importance of task-specific evaluation and cautious integration of LLMs into healthcare workflows.

Method: Uses difficulty-based routing to predict problem complexity, deploys three specialized experts to generate responses, employs multiple correction phases, neural verification, and clustering-based aggregation for final answer selection.

Result: Achieved 75.28% accuracy on GSM8K dataset using only original training data, outperforming most comparable 7B models trained on synthetic data.

Conclusion: Difficulty-based routing and uncertainty-driven aggregation are efficient and effective for improving math reasoning robustness in LLMs.

Abstract: Large language models (LLMs) demonstrate strong performance in math reasoning benchmarks, but their performance varies inconsistently across problems with varying levels of difficulty. This paper describes Adaptive Multi-Expert Reasoning (AMR), a framework that focuses on problem complexity by reasoning with dynamically adapted strategies. An agile routing system that focuses on problem text predicts problems’ difficulty and uncertainty and guides a reconfigurable sampling mechanism to manage the breadth of generation. Three specialized experts create candidate responses, which are modified during multiple correction and finalization phases. A neural verifier assesses the correctness of responses, while a clustering-based aggregation technique identifies the final candidate answer based on a combination of consensus and answer quality. When evaluated on the GSM8K dataset, AMR achieved 75.28% accuracy while only using the original training data. This result outperformed the majority of comparable 7B models that were trained on synthetic data. This showcases that models using difficulty-based routing and uncertainty-driven aggregation are efficient and effective in improving math reasoning models’ robustness.

Method: Proposes LLM-as-a-Judge for semantic evaluation and LLM-driven agent framework for multi-turn interactive refinement of recognition outputs

Result: Extensive experiments on GigaSpeech, WenetSpeech, and ASRU 2019 code-switching test sets show improved semantic fidelity and interactive correction capability

Conclusion: Agentic framework effectively integrates semantic evaluation and interactive correction for ASR, advancing beyond traditional token-level metrics

Abstract: Recent years have witnessed remarkable progress in automatic speech recognition (ASR), driven by advances in model architectures and large-scale training data. However, two important aspects remain underexplored. First, Word Error Rate (WER), the dominant evaluation metric for decades, treats all words equally and often fails to reflect the semantic correctness of an utterance at the sentence level. Second, interactive correction-an essential component of human communication-has rarely been systematically studied in ASR research. In this paper, we integrate these two perspectives under an agentic framework for interactive ASR. We propose leveraging LLM-as-a-Judge as a semantic-aware evaluation metric to assess recognition quality beyond token-level accuracy. Furthermore, we design an LLM-driven agent framework to simulate human-like multi-turn interaction, enabling iterative refinement of recognition outputs through semantic feedback. Extensive experiments are conducted on standard benchmarks, including GigaSpeech (English), WenetSpeech (Chinese), the ASRU 2019 code-switching test set. Both objective and subjective evaluations demonstrate the effectiveness of the proposed framework in improving semantic fidelity and interactive correction capability. We will release the code to facilitate future research in interactive and agentic ASR.

Method: A framework that induces narrative schemas by jointly modeling events and characters via structured clustering, producing explainable schemas aligned with framing theory while scaling to large corpora without exhaustive manual annotation.

Result: The approach produces rich, explainable narrative schemas that capture nuanced storytelling structures while being scalable to large text corpora without requiring extensive manual annotation.

Conclusion: The framework bridges the gap between computational analysis and communication theory by providing a scalable method to capture nuanced narrative structures in media content.

Abstract: Media narratives wield tremendous power in shaping public opinion, yet computational approaches struggle to capture the nuanced storytelling structures that communication theory emphasizes as central to how meaning is constructed. Existing approaches either miss subtle narrative patterns through coarse-grained analysis or require domain-specific taxonomies that limit scalability. To bridge this gap, we present a framework for inducing rich narrative schemas by jointly modeling events and characters via structured clustering. Our approach produces explainable narrative schemas that align with established framing theory while scaling to large corpora without exhaustive manual annotation.

Method: Multi-agent debate framework with hybrid Retrieval-Augmented Generation (RAG) that decomposes clinical notes into sub-queries, retrieves evidence through dense/sparse/online retrieval, assigns two domain expert agents distinct knowledge bases for independent analysis, uses structured counter-argumentation and cross-source adjudication for disagreements, and includes cascading safety layer to filter false positives.

Result: Achieves best accuracy (69.13%), ROC-AUC (74.45%), and PR-AUC (72.44%) under few-shot prompting on clinical terminology substitution detection benchmark, outperforming single-agent RAG and debate-only baselines across six backbone models and two prompting strategies.

Conclusion: Retrieval augmentation and structured debate are complementary for clinical error detection, with the framework benefiting most from models with strong instruction-following and clinical language understanding capabilities.

Abstract: Terminology substitution errors in clinical notes, where one medical term is replaced by a linguistically valid but clinically different term, pose a persistent challenge for automated error detection in healthcare. We introduce BLUEmed, a multi-agent debate framework augmented with hybrid Retrieval-Augmented Generation (RAG) that combines evidence-grounded reasoning with multi-perspective verification for clinical error detection. BLUEmed decomposes each clinical note into focused sub-queries, retrieves source-partitioned evidence through dense, sparse, and online retrieval, and assigns two domain expert agents distinct knowledge bases to produce independent analyses; when the experts disagree, a structured counter-argumentation round and cross-source adjudication resolve the conflict, followed by a cascading safety layer that filters common false-positive patterns. We evaluate BLUEmed on a clinical terminology substitution detection benchmark under both zero-shot and few-shot prompting with multiple backbone models spanning proprietary and open-source families. Experimental results show that BLUEmed achieves the best accuracy (69.13%), ROC-AUC (74.45%), and PR-AUC (72.44%) under few-shot prompting, outperforming both single-agent RAG and debate-only baselines. Further analyses across six backbone models and two prompting strategies confirm that retrieval augmentation and structured debate are complementary, and that the framework benefits most from models with sufficient instruction-following and clinical language understanding.

Method: Created a large-scale name-nationality dataset from Open Academic Graph (OAG), then used LLMs as dataset enrichers (not inference engines) to generate synthetic names for augmenting low-resource countries. Evaluated on both real and synthetic-tail test sets.

Result: NameBERT models achieved significantly higher accuracy than state-of-the-art baselines across both in- and out-of-domain tasks. Augmentation produced large gains when evaluation included synthetic tail names and modest lift on tail-country metrics otherwise.

Conclusion: Using LLMs for data augmentation rather than inference enables efficient, high-performance name-based nationality classification that addresses coverage gaps for underrepresented countries while maintaining scalability.

Abstract: Inferring nationality from personal names is a critical capability for equity and bias monitoring, personalization, and a valuable tool in biomedical and sociological research. However, existing name-based nationality classifiers are typically trained on relatively small or source-specific labeled datasets, which can introduce coverage gaps and limit performance for underrepresented countries. While large language models (LLMs) demonstrate strong zero-shot performance for name-based nationality prediction, their computational cost and latency make them impractical for real-time, large-scale deployment. In this work, we created a large-scale name-nationality dataset from the Open Academic Graph (OAG) and introduce a framework that leverages LLMs as dataset enrichers rather than inference engines. We augment low-resource countries with LLM-generated names and evaluate on real and synthetic-tail test sets. We find that augmentation produces large gains when evaluation includes synthetic tail names and still offers a modest lift on tail-country metrics otherwise. Overall, NameBERT models achieve significantly higher accuracy than state-of-the-art baselines across both in- and out-of-domain tasks, while remaining efficient for large-scale inference compared to LLMs.

Method: Created LASQ dataset for Uzbek and Uyghur with quadruple extraction task; designed grid-tagging model with Syntax Knowledge Embedding Module (SKEM) incorporating POS and dependency knowledge to handle agglutinative language challenges.

Result: Experiments on LASQ show consistent improvements over competitive baselines, validating both the dataset’s utility and the effectiveness of the proposed modeling approach.

Conclusion: LASQ addresses the gap in low-resource language ABSA research and the proposed syntax-enhanced model effectively handles lexical sparsity in agglutinative languages.

Abstract: In recent years, aspect-based sentiment analysis (ABSA) has made rapid progress and shown strong practical value. However, existing research and benchmarks are largely concentrated on high-resource languages, leaving fine-grained sentiment extraction in low-resource languages under-explored. To address this gap, we constructed the first Low-resource languages Aspect-based Sentiment Quadruple dataset, named LASQ, which includes two low-resource languages: Uzbek and Uyghur. Secondly, it includes a fine-grained target-aspect-opinion-sentiment quadruple extraction task. To facilitate future research, we designed a grid-tagging model that integrates syntactic knowledge. This model incorporates part-of-speech (POS) and dependency knowledge into the model through our designed Syntax Knowledge Embedding Module (SKEM), thereby alleviating the lexical sparsity problem caused by agglutinative languages. Experiments on LASQ demonstrate consistent gains over competitive baselines, validating both the dataset’s utility and the effectiveness of the proposed modeling approach.

Method: Theoretical analysis and formal definitions: 1) Introducing a measure of undecidability based on probability distribution of input data, 2) Extending Turing’s infinite logics and Oracle machines to super-Turing computation models, 3) Defining three new complexity classes for TM undecidable problems (U-complete, D-complete, H-complete), 4) Proving negative answer to P≠NP equivalent for U-complete class.

Result: Multiple novel theoretical contributions: 1) First explicit definition of complexity classes for undecidable problems, 2) Negative resolution of P≠NP equivalent for U-complete class, 3) Framework for measuring undecidability degrees, 4) Extension of Turing’s models to super-Turing computation.

Conclusion: The paper significantly advances theoretical computer science by providing new frameworks for understanding undecidability and extending Turing’s legacy with novel complexity classes and computation models that go beyond traditional Turing machines.

Abstract: Alan Turing is considered as a founder of current computer science together with Kurt Godel, Alonzo Church and John von Neumann. In this paper multiple new research results are presented. It is demonstrated that there would not be Alan Turing’s achievements without earlier seminal contributions by Georg Cantor in the set theory and foundations of mathematics. It is proposed to introduce the measure of undecidability of problems unsolvable by Turing machines based on probability distribution of its input data, i.e., to provide the degree of unsolvabilty based on the number of undecidable instances of input data versus decidable ones. It is proposed as well to extend the Turing’s work on infinite logics and Oracle machines to a whole class of super-Turing models of computation. Next, the three new complexity classes for TM undecidable problems have been defined: U-complete (Universal complete), D-complete (Diagonalization complete) and H-complete (Hypercomputation complete) classes. The above has never been defined explicitly before by other scientists, and has been inspired by Cook/Levin NP-complete class for intractable problems. Finally, an equivalent to famous P is not equal to NP unanswered question for NP-complete class, has been answered negatively for U-complete class of complexity for undecidable problems.

Method: Three-stage pipeline: (1) Knowledge Consolidation to unify fragmented extractions into a stable memory substrate; (2) Associative Navigation to traverse the graph via semantic, contextualized, and functional pathways to recover evidence chains; (3) Interference Elimination to prune hyper-associative noise for coherent reasoning context.

Result: On GraphRAG-Bench, achieves absolute gains of 7-10% in retrieval recall and 3-11% in generation accuracy, demonstrating superior ability to robustify associative evidence retrieval.

Conclusion: CodaRAG successfully transforms retrieval into active associative discovery, enabling systematic recovery of dispersed evidence chains for improved factual, reasoning, and creative tasks.

Abstract: Large Language Models (LLMs) struggle with knowledge-intensive tasks due to hallucinations and fragmented reasoning over dispersed information. While Retrieval-Augmented Generation (RAG) grounds generation in external sources, existing methods often treat evidence as isolated units, failing to reconstruct the logical chains that connect these dots. Inspired by Complementary Learning Systems (CLS), we propose CodaRAG, a framework that evolves retrieval from passive lookup into active associative discovery. CodaRAG operates via a three-stage pipeline: (1) Knowledge Consolidation to unify fragmented extractions into a stable memory substrate; (2) Associative Navigation to traverse the graph via multi-dimensional pathways-semantic, contextualized, and functional-explicitly recovering dispersed evidence chains; and (3) Interference Elimination to prune hyper-associative noise, ensuring a coherent, high-precision reasoning context. On GraphRAG-Bench, CodaRAG achieves absolute gains of 7-10% in retrieval recall and 3-11% in generation accuracy. These results demonstrate CodaRAG’s superior ability to systematically robustify associative evidence retrieval for factual, reasoning, and creative tasks.

Method: ADG generates multiple high-temperature responses per instruction, embeds them, computes divergence scores combining dispersion magnitude and shape anisotropy to select instructions with diverse, multi-modal answers.

Result: Fine-tuning on just 10K ADG-selected examples outperforms strong baselines across 6 benchmarks in reasoning, knowledge, and coding tasks using two different model backbones.

Conclusion: Answer divergence is a practical signal for instruction data selection, with both dispersion magnitude and shape anisotropy being necessary components for effective data curation.

Abstract: Instruction tuning relies on large instruction-response corpora whose quality and composition strongly affect downstream performance. We propose Answer Divergence-Guided Selection (ADG), which selects instruction data based on the geometric structure of multi-sample outputs. ADG draws several high-temperature generations per instruction, maps responses into an embedding space, and computes an output divergence score that jointly encodes dispersion magnitude and shape anisotropy. High scores correspond to instructions whose answers are both far apart and multi-modal, rather than clustered paraphrases along a single direction. Across two backbones and three public instruction pools, fine-tuning on only 10K ADG-selected examples consistently outperforms strong selectors on six benchmarks spanning reasoning, knowledge, and coding. Analyses further show that both dispersion magnitude and shape anisotropy are necessary, supporting answer divergence as a practical signal for instruction data selection. Code and appendix are included in the supplementary materials.

Method: Proposes NOSE framework that aligns three modalities: molecular structure, receptor sequence, and natural language description. Uses orthogonal constraints to decouple modality contributions rather than simple fusion. Introduces weak positive sample strategy to calibrate semantic similarity for sparse olfactory language data.

Result: Achieves state-of-the-art performance and excellent zero-shot generalization. Demonstrates strong alignment between representation space and human olfactory intuition.

Conclusion: NOSE successfully captures the complete olfactory pathway through multimodal alignment, providing biologically grounded and semantically interpretable representations that generalize well to new olfactory tasks.

Abstract: Olfaction lies at the intersection of chemical structure, neural encoding, and linguistic perception, yet existing representation methods fail to fully capture this pathway. Current approaches typically model only isolated segments of the olfactory pathway, overlooking the complete chain from molecule to receptors to linguistic descriptions. Such fragmentation yields learned embeddings that lack both biological grounding and semantic interpretability. We propose NOSE (Neural Olfactory-Semantic Embedding), a representation learning framework that aligns three modalities along the olfactory pathway: molecular structure, receptor sequence, and natural language description. Rather than simply fusing these signals, we decouple their contributions via orthogonal constraints, preserving the unique encoded information of each modality. To address the sparsity of olfactory language, we introduce a weak positive sample strategy to calibrate semantic similarity, preventing erroneous repulsion of similar odors in the feature space. Extensive experiments demonstrate that NOSE achieves state-of-the-art (SOTA) performance and excellent zero-shot generalization, confirming the strong alignment between its representation space and human olfactory intuition.

Method: EviCare performs three-step reasoning: (1) deep model inference for candidate selection, (2) evidential prioritization for set-based EHRs, and (3) relational evidence construction for novel diagnosis prediction, then composes these signals into adaptive in-context prompts to guide LLM reasoning.

Result: On MIMIC-III and MIMIC-IV benchmarks, EviCare achieves significant performance gains, outperforming both LLM-only and deep model-only baselines by average 20.65% across precision and accuracy metrics, with 30.97% average improvement in novel diagnosis prediction.

Conclusion: EviCare effectively addresses the limitation of existing LLM approaches in identifying novel diagnoses through deep model guidance and in-context reasoning, improving both accuracy and interpretability in EHR-based diagnosis prediction.

Abstract: Recent advances in large language models (LLMs) have enabled promising progress in diagnosis prediction from electronic health records (EHRs). However, existing LLM-based approaches tend to overfit to historically observed diagnoses, often overlooking novel yet clinically important conditions that are critical for early intervention. To address this, we propose EviCare, an in-context reasoning framework that integrates deep model guidance into LLM-based diagnosis prediction. Rather than prompting LLMs directly with raw EHR inputs, EviCare performs (1) deep model inference for candidate selection, (2) evidential prioritization for set-based EHRs, and (3) relational evidence construction for novel diagnosis prediction. These signals are then composed into an adaptive in-context prompt to guide LLM reasoning in an accurate and interpretable manner. Extensive experiments on two real-world EHR benchmarks (MIMIC-III and MIMIC-IV) demonstrate that EviCare achieves significant performance gains, which consistently outperforms both LLM-only and deep model-only baselines by an average of 20.65% across precision and accuracy metrics. The improvements are particularly notable in challenging novel diagnosis prediction, yielding average improvements of 30.97%.

Method: Integrates dynamic adaptive multi-head attention (using global context pooling to regulate attention head contributions) with supervised contrastive learning into a BERT-based Transformer encoder.

Result: Achieves 94.67% accuracy on IMDB dataset, outperforming strong baselines by 1.5-2.5 percentage points.

Conclusion: The lightweight, efficient framework effectively handles long review texts and is extensible to other text classification tasks.

Abstract: The exponential growth of user-generated movie reviews on digital platforms has made accurate text sentiment classification a cornerstone task in natural language processing. Traditional models, including standard BERT and recurrent architectures, frequently struggle to capture long-distance semantic dependencies and resolve ambiguous emotional expressions in lengthy review texts. This paper proposes a novel hybrid framework that seamlessly integrates dynamic adaptive multi-head attention with supervised contrastive learning into a BERT-based Transformer encoder. The dynamic adaptive attention module employs a global context pooling vector to dynamically regulate the contribution of each attention head, thereby focusing on critical sentiment-bearing tokens while suppressing noise. Simultaneously, the supervised contrastive learning branch enforces tighter intra-class compactness and larger inter-class separation in the embedding space. Extensive experiments on the IMDB dataset demonstrate that the proposed model achieves competitive performance with an accuracy of 94.67%, outperforming strong baselines by 1.5–2.5 percentage points. The framework is lightweight, efficient, and readily extensible to other text classification tasks.

Method: Created JurisCQAD dataset (43K+ Chinese legal queries), designed structured task decomposition converting queries into legal element graphs (entities, events, intents, legal issues), and proposed JurisMA modular multi-agent framework with dynamic routing, statutory grounding, and stylistic optimization.

Result: System significantly outperforms both general-purpose and legal-domain LLMs on refined LawBench across multiple lexical and semantic metrics, demonstrating benefits of interpretable decomposition and modular collaboration.

Conclusion: Structured task decomposition with legal element graphs combined with modular multi-agent framework enables strong context-aware reasoning for legal consultation QA, effectively capturing dependencies across legal facts, norms, and procedural logic.

Abstract: Legal consultation question answering (Legal CQA) presents unique challenges compared to traditional legal QA tasks, including the scarcity of high-quality training data, complex task composition, and strong contextual dependencies. To address these, we construct JurisCQAD, a large-scale dataset of over 43,000 real-world Chinese legal queries annotated with expert-validated positive and negative responses, and design a structured task decomposition that converts each query into a legal element graph integrating entities, events, intents, and legal issues. We further propose JurisMA, a modular multi-agent framework supporting dynamic routing, statutory grounding, and stylistic optimization. Combined with the element graph, the framework enables strong context-aware reasoning, effectively capturing dependencies across legal facts, norms, and procedural logic. Trained on JurisCQAD and evaluated on a refined LawBench, our system significantly outperforms both general-purpose and legal-domain LLMs across multiple lexical and semantic metrics, demonstrating the benefits of interpretable decomposition and modular collaboration in Legal CQA.

Method: Introduces new dataset categorizing uncertainty sources, enabling systematic evaluation of UQ performance under each condition (knowledge gaps, output variability, input ambiguity).

Result: Experiments show many UQ methods perform well when uncertainty stems from model knowledge limitations, but degrade or become misleading when other sources (output variability, input ambiguity) are introduced.

Conclusion: Highlights need for uncertainty-aware methods that explicitly account for source of uncertainty in LLMs, rather than single confidence scores.

Abstract: As Large Language Models (LLMs) are increasingly deployed in real-world applications, reliable uncertainty quantification (UQ) becomes critical for safe and effective use. Most existing UQ approaches for language models aim to produce a single confidence score – for example, estimating the probability that a model’s answer is correct. However, uncertainty in natural language tasks arises from multiple distinct sources, including model knowledge gaps, output variability, and input ambiguity, which have different implications for system behavior and user interaction. In this work, we study how the source of uncertainty impacts the behavior and effectiveness of existing UQ methods. To enable controlled analysis, we introduce a new dataset that explicitly categorizes uncertainty sources, allowing systematic evaluation of UQ performance under each condition. Our experiments reveal that while many UQ methods perform well when uncertainty stems solely from model knowledge limitations, their performance degrades or becomes misleading when other sources are introduced. These findings highlight the need for uncertainty-aware methods that explicitly account for the source of uncertainty in large language models.

Method: SGKR organizes domain knowledge with a graph induced by function-call dependencies, extracts semantic input/output tags from questions, identifies dependency paths connecting them, constructs task-relevant subgraphs, and assembles associated knowledge and function implementations as structured context for LLM-based code generation.

Result: Experiments on multi-step data analysis benchmarks show SGKR consistently improves solution correctness over no-retrieval and similarity-based retrieval baselines for both vanilla LLMs and coding agents.

Conclusion: Structure-grounded knowledge retrieval using function-call dependency graphs is more effective than similarity-based approaches for retrieving task-critical knowledge in multi-step data analysis tasks requiring code generation.

Abstract: Selecting the right knowledge is critical when using large language models (LLMs) to solve domain-specific data analysis tasks. However, most retrieval-augmented approaches rely primarily on lexical or embedding similarity, which is often a weak proxy for the task-critical knowledge needed for multi-step reasoning. In many such tasks, the relevant knowledge is not merely textually related to the query, but is instead grounded in executable code and the dependency structure through which computations are carried out. To address this mismatch, we propose SGKR (Structure-Grounded Knowledge Retrieval), a retrieval framework that organizes domain knowledge with a graph induced by function-call dependencies. Given a question, SGKR extracts semantic input and output tags, identifies dependency paths connecting them, and constructs a task-relevant subgraph. The associated knowledge and corresponding function implementations are then assembled as a structured context for LLM-based code generation. Experiments on multi-step data analysis benchmarks show that SGKR consistently improves solution correctness over no-retrieval and similarity-based retrieval baselines for both vanilla LLMs and coding agents.

Method: Defines factual inconsistency criteria specific to code summaries, evaluates at segment level using these criteria with dependency information, then aggregates segment-level results into a fine-grained score. Constructs a code summarization benchmark with human-annotated factual consistency labels.

Result: ReFEree achieves the highest correlation with human judgment among 13 baselines, improving 15-18% over previous state-of-the-art methods.

Conclusion: ReFEree provides an effective reference-free, fine-grained evaluation method for factual consistency in code summaries, addressing key limitations of previous approaches and showing superior correlation with human judgment.

Abstract: As Large Language Models (LLMs) have become capable of generating long and descriptive code summaries, accurate and reliable evaluation of factual consistency has become a critical challenge. However, previous evaluation methods are primarily designed for short summaries of isolated code snippets. Consequently, they struggle to provide fine-grained evaluation of multi-sentence functionalities and fail to accurately assess dependency context commonly found in real-world code summaries. To address this, we propose ReFEree, a reference-free and fine-grained method for evaluating factual consistency in real-world code summaries. We define factual inconsistency criteria specific to code summaries and evaluate them at the segment level using these criteria along with dependency information. These segment-level results are then aggregated into a fine-grained score. We construct a code summarization benchmark with human-annotated factual consistency labels. The evaluation results demonstrate that ReFEree achieves the highest correlation with human judgment among 13 baselines, improving 15-18% over the previous state-of-the-art. Our code and data are available at https://github.com/bsy99615/ReFEree.git.

Method: 1) Information-theoretic argument grounded in Data Processing Inequality suggesting SAS are more information-efficient under fixed reasoning-token budget with perfect context utilization; 2) Controlled empirical study across three model families (Qwen3, DeepSeek-R1-Distill-Llama, Gemini 2.5) comparing SAS with multiple MAS architectures under matched budgets; 3) Diagnostic analysis of system behavior and evaluation methodology including API-based budget control artifacts.

Result: SAS consistently match or outperform MAS on multi-hop reasoning tasks when reasoning tokens are held constant. Identified significant artifacts in API-based budget control (particularly Gemini 2.5) and standard benchmarks that can inflate apparent MAS gains. MAS become competitive only when single agent’s context utilization is degraded or with more compute.

Conclusion: Reported advantages of multi-agent systems for multi-hop reasoning are better explained by unaccounted computation and context effects rather than inherent architectural benefits. Highlights importance of understanding trade-offs between compute, context, and coordination in agentic systems.

Abstract: Recent work reports strong performance from multi-agent LLM systems (MAS), but these gains are often confounded by increased test-time computation. When computation is normalized, single-agent systems (SAS) can match or outperform MAS, yet the theoretical basis and evaluation methodology behind this comparison remain unclear. We present an information-theoretic argument, grounded in the Data Processing Inequality, suggesting that under a fixed reasoning-token budget and with perfect context utilization, single-agent systems are more information-efficient. This perspective further predicts that multi-agent systems become competitive when a single agent’s effective context utilization is degraded, or when more compute is expended. We test these predictions in a controlled empirical study across three model families (Qwen3, DeepSeek-R1-Distill-Llama, and Gemini 2.5), comparing SAS with multiple MAS architectures under matched budgets. We find that SAS consistently match or outperform MAS on multi-hop reasoning tasks when reasoning tokens are held constant. Beyond aggregate performance, we conduct a detailed diagnostic analysis of system behavior and evaluation methodology. We identify significant artifacts in API-based budget control (particularly in Gemini 2.5) and in standard benchmarks, both of which can inflate apparent gains from MAS. Overall, our results suggest that, for multi-hop reasoning tasks, many reported advantages of multi-agent systems are better explained by unaccounted computation and context effects rather than inherent architectural benefits, and highlight the importance of understanding and explicitly controlling the trade-offs between compute, context, and coordination in agentic systems.

Method: Conducted first controlled comparative study evaluating hallucination patterns in dLLMs versus autoregressive counterparts, controlling for architecture, scale, and pre-training weights. Analyzed inference-time compute dynamics and identified unique failure modes.

Result: Current dLLMs exhibit higher propensity for hallucination than AR models. Quasi-autoregressive generation suffers from early saturation, while non-sequential decoding allows continuous refinement. Identified unique diffusion failure modes: premature termination, incomplete denoising, and context intrusion.

Conclusion: Although dLLMs have narrowed performance gap on general tasks, their distinct hallucination mechanisms pose critical challenges to model reliability, highlighting the need for specialized mitigation strategies.

Abstract: While Diffusion Large Language Models (dLLMs) have emerged as a promising non-autoregressive paradigm comparable to autoregressive (AR) models, their faithfulness, specifically regarding hallucination, remains largely underexplored. To bridge this gap, we present the first controlled comparative study to evaluate hallucination patterns in dLLMs. Our results demonstrate that current dLLMs exhibit a higher propensity for hallucination than AR counterparts controlled for architecture, scale, and pre-training weights. Furthermore, an analysis of inference-time compute reveals divergent dynamics: while quasi-autoregressive generation suffers from early saturation, non-sequential decoding unlocks potential for continuous refinement. Finally, we identify distinct failure modes unique to the diffusion process, including premature termination, incomplete denoising, and context intrusion. Our findings underscore that although dLLMs have narrowed the performance gap on general tasks, their distinct hallucination mechanisms pose a critical challenge to model reliability. Our code is available at https://github.com/ZeroLoss-Lab/Lost-in-Diffusion

Method: Formalizes empathy as an observable behavioral property: capacity to model and respond to human perspectives while preserving intention, affect, and context. Identifies four mechanisms of empathic failure (sentiment attenuation, empathic granularity mismatch, conflict avoidance, linguistic distancing) and organizes them along three dimensions (cognitive, cultural, relational empathy). Conducts empirical analyses showing benchmark performance can mask systematic empathic distortions.

Result: Empirical analyses demonstrate that strong benchmark performance can mask systematic empathic distortions in LLMs. The paper identifies specific failure patterns and motivates the need for empathy-aware objectives, benchmarks, and training signals.

Conclusion: LLM development needs to incorporate empathy as a first-class component through explicit mechanisms, objectives, benchmarks, and training signals to address systematic failures in modeling human perspectives, affect, and relational dynamics.

Abstract: This paper argues that Large Language Models (LLMs) should incorporate explicit mechanisms for human empathy. As LLMs become increasingly deployed in high-stakes human-centered settings, their success depends not only on correctness or fluency but on faithful preservation of human perspectives. Yet, current LLMs systematically fail at this requirement: even when well-aligned and policy-compliant, they often attenuate affect, misrepresent contextual salience, and rigidify relational stance in ways that distort meaning. We formalize empathy as an observable behavioral property: the capacity to model and respond to human perspectives while preserving intention, affect, and context. Under this framing, we identify four recurring mechanisms of empathic failure in contemporary LLMs–sentiment attenuation, empathic granularity mismatch, conflict avoidance, and linguistic distancing–arising as structural consequences of prevailing training and alignment practices. We further organize these failures along three dimensions: cognitive, cultural, and relational empathy, to explain their manifestation across tasks. Empirical analyses show that strong benchmark performance can mask systematic empathic distortions, motivating empathy-aware objectives, benchmarks, and training signals as first-class components of LLM development.

Method: Systematically analyze inference dynamics of non-autoregressive decoding in diffusion language models, identify proximity bias issue, and propose a minimal-intervention approach with lightweight planner and end-of-sequence temperature annealing to guide early token selection.

Result: The method shows substantial overall improvement over existing heuristic baselines on various reasoning and planning tasks without significant computational overhead.

Conclusion: Proximity bias in diffusion language models causes spatial error propagation in non-autoregressive decoding, but this can be effectively addressed with targeted guidance mechanisms for early token selection.

Abstract: Diffusion-based language models (dLLMs) have emerged as a promising alternative to autoregressive language models, offering the potential for parallel token generation and bidirectional context modeling. However, harnessing this flexibility for fully non-autoregressive decoding remains an open question, particularly for reasoning and planning tasks. In this work, we investigate non-autoregressive decoding in dLLMs by systematically analyzing its inference dynamics along the temporal axis. Specifically, we uncover an inherent failure mode in confidence-based non-autoregressive generation stemming from a strong proximity bias-the tendency for the denoising order to concentrate on spatially adjacent tokens. This local dependency leads to spatial error propagation, rendering the entire trajectory critically contingent on the initial unmasking position. Leveraging this insight, we present a minimal-intervention approach that guides early token selection, employing a lightweight planner and end-of-sequence temperature annealing. We thoroughly evaluate our method on various reasoning and planning tasks and observe substantial overall improvement over existing heuristic baselines without significant computational overhead.

Method: Introduces a Cross-Lingual Mapping Task during pre-training that bi-directionally maps languages within the LLM embedding space. Also proposes a Language Alignment Coefficient to quantify cross-lingual consistency in limited-data scenarios.

Result: Achieves gains up to 11.9 BLEU points in machine translation, 6.72 points in CLQA BERTScore-Precision, and over 5% in CLNLU accuracy compared to strong multilingual baselines.

Conclusion: Incorporating cross-lingual objectives into pre-training can significantly improve multilingual LLMs, with the proposed method enhancing both language generation and comprehension across languages.

Abstract: Multilingual Large Language Models (LLMs) struggle with cross-lingual tasks due to data imbalances between high-resource and low-resource languages, as well as monolingual bias in pre-training. Existing methods, such as bilingual fine-tuning and contrastive alignment, can improve cross-lingual performance, but they often require extensive parallel data or suffer from instability. To address these challenges, we introduce a Cross-Lingual Mapping Task during the pre-training phase, which enhances cross-lingual alignment without compromising monolingual fluency. Our approach bi-directionally maps languages within the LLM embedding space, improving both language generation and comprehension. We further propose a Language Alignment Coefficient to robustly quantify cross-lingual consistency, even in limited-data scenarios. Experimental results on machine translation (MT), cross-lingual natural language understanding (CLNLU), and cross-lingual question answering (CLQA) show that our model achieves gains of up to 11.9 BLEU points in MT, 6.72 points in CLQA BERTScore-Precision, and more than 5% in CLNLU accuracy over strong multilingual baselines. These findings highlight the potential of incorporating cross-lingual objectives into pre-training to improve multilingual LLMs.

Method: Used six multilingual large language models as controllable systems, created targeted “computational lesions” by zeroing parameter sets important across languages or specific to one language, then compared intact vs. lesioned models in predicting fMRI responses during 100 minutes of naturalistic story listening in native English, Chinese, and French (112 participants).

Result: Lesioning a compact shared core reduced whole-brain encoding correlation by 60.32% relative to intact models, while language-specific lesions preserved cross-language separation in embedding space but selectively weakened brain predictivity for the matched native language.

Conclusion: Results support a shared backbone with embedded specializations for language processing in the brain, providing a causal framework for studying multilingual brain-model alignment using LLMs as computational models.

Abstract: How the brain supports language across different languages is a basic question in neuroscience and a useful test for multilingual artificial intelligence. Neuroimaging has identified language-responsive brain regions across languages, but it cannot by itself show whether the underlying processing is shared or language-specific. Here we use six multilingual large language models (LLMs) as controllable systems and create targeted ``computational lesions’’ by zeroing small parameter sets that are important across languages or especially important for one language. We then compare intact and lesioned models in predicting functional magnetic resonance imaging (fMRI) responses during 100 minutes of naturalistic story listening in native English, Chinese and French (112 participants). Lesioning a compact shared core reduces whole-brain encoding correlation by 60.32% relative to intact models, whereas language-specific lesions preserve cross-language separation in embedding space but selectively weaken brain predictivity for the matched native language. These results support a shared backbone with embedded specializations and provide a causal framework for studying multilingual brain-model alignment.

Method: ProUIE uses a three-stage progressive learning approach: 1) Macro-level Complete Modeling (CM) - learns NER, RE, and EE along intrinsic difficulty order on full training data; 2) Meso-level Streamlined Alignment (SA) - operates on sampled data with simplified target formats to make outputs more concise and controllable; 3) Micro-level Deep Exploration (DE) - applies GRPO with stepwise fine-grained rewards over structural units to guide exploration.

Result: Experiments on 36 public datasets show ProUIE consistently improves unified extraction, outperforming strong instruction-tuned baselines on average for NER and RE while using a smaller backbone. It also demonstrates clear gains in large-scale production-oriented information extraction.

Conclusion: ProUIE provides an effective progressive learning approach for universal information extraction that achieves performance improvements without requiring external data, making it more efficient and practical for real-world applications.

Abstract: LLM-based universal information extraction (UIE) methods often rely on additional information beyond the original training data, which increases training complexity yet often yields limited gains. To address this, we propose ProUIE, a Macro-to-Micro progressive learning approach that improves UIE without introducing any external information. ProUIE consists of three stages: (i) macro-level Complete Modeling (CM), which learns NER, RE, and EE along their intrinsic difficulty order on the full training data to build a unified extraction foundation, (ii) meso-level Streamlined Alignment (SA), which operates on sampled data with simplified target formats, streamlining and regularizing structured outputs to make them more concise and controllable, and (iii) micro-level Deep Exploration (DE), which applies GRPO with stepwise fine-grained rewards (SFR) over structural units to guide exploration and improve performance. Experiments on 36 public datasets show that ProUIE consistently improves unified extraction, outperforming strong instruction-tuned baselines on average for NER and RE while using a smaller backbone, and it further demonstrates clear gains in large-scale production-oriented information extraction.

Method: Proposes contrastive pointwise mutual information (CPMI) that quantifies how much a reasoning step increases mutual information between the step and correct answer relative to hard-negative alternatives, using the model’s internal probability to infer step-level supervision without extensive rollouts.

Result: CPMI reduces dataset construction time by 84% and token generation by 98% compared to Monte Carlo estimation, while achieving higher accuracy on process-level evaluations and mathematical reasoning benchmarks.

Conclusion: CPMI provides an efficient and effective automatic reward labeling method for chain-of-thought reasoning verification that significantly reduces computational costs while maintaining or improving performance.

Abstract: Recent research has devoted considerable effort to verifying the intermediate reasoning steps of chain-of-thought (CoT) trajectories using process reward models (PRMs) and other verifier models. However, training a PRM typically requires human annotators to assign reward scores to each reasoning step, which is both costly and time-consuming. Existing automated approaches, such as Monte Carlo (MC) estimation, also demand substantial computational resources due to repeated LLM rollouts. To overcome these limitations, we propose contrastive pointwise mutual information (CPMI), a novel automatic reward labeling method that leverages the model’s internal probability to infer step-level supervision while significantly reducing the computational burden of annotating dataset. CPMI quantifies how much a reasoning step increases the mutual information between the step and the correct target answer relative to hard-negative alternatives. This contrastive signal serves as a proxy for the step’s contribution to the final solution and yields a reliable reward. The experimental results show that CPMI-based labeling reduces dataset construction time by 84% and token generation by 98% compared to MC estimation, while achieving higher accuracy on process-level evaluations and mathematical reasoning benchmarks.

Method: Develops a syllable-based tokenizer using Turkish’s six-pattern phonological structure to create a closed vocabulary of ~8,000 unique syllable types. Trains a BERT-tiny encoder (1.5M parameters) from scratch on Turkish Wikipedia using masked language modeling, combined with fine-grained chunk-based retrieval strategy.

Result: Achieves 50.3% Recall@5 on TQuAD retrieval benchmark, surpassing the 46.92% reported by a morphology-driven baseline that uses a 200 times larger model, demonstrating superior efficiency and performance.

Conclusion: The phonological regularity of Turkish syllables provides a strong and resource-light inductive bias for retrieval tasks, enabling efficient tokenization and competitive performance with minimal model size.

Abstract: HeceTokenizer is a syllable-based tokenizer for Turkish that exploits the deterministic six-pattern phonological structure of the language to construct a closed, out-of-vocabulary (OOV)-free vocabulary of approximately 8,000 unique syllable types. A BERT-tiny encoder (1.5M parameters) is trained from scratch on a subset of Turkish Wikipedia using a masked language modeling objective and evaluated on the TQuAD retrieval benchmark using Recall@5. Combined with a fine-grained chunk-based retrieval strategy, HeceTokenizer achieves 50.3% Recall@5, surpassing the 46.92% reported by a morphology-driven baseline that uses a 200 times larger model. These results suggest that the phonological regularity of Turkish syllables provides a strong and resource-light inductive bias for retrieval tasks.

Method: Two-phase framework: 1) Syntactic exploration to gather diverse outputs for constraint learning, 2) Constraint exploitation to enforce learned rules during generation. This integrates context-sensitive grammar learning with LLM generation.

Result: Enables even small LLMs (1B parameters) to learn and generate with perfect constraint adherence, outperforming larger counterparts and state-of-the-art reasoning models.

Conclusion: First integration of context-sensitive grammar learning with LLM generation, eliminating manual specification while maintaining generation validity.

Abstract: Controlling the output of Large Language Models (LLMs) through context-sensitive constraints has emerged as a promising approach to overcome the limitations of Context-Free Grammars (CFGs) in guaranteeing generation validity. However, such constraints typically require manual specification – a significant barrier demanding specialized expertise. We introduce a framework that automatically learns context-sensitive constraints from LLM interactions through a two-phase process: syntactic exploration to gather diverse outputs for constraint learning, followed by constraint exploitation to enforce these learned rules during generation. Experiments demonstrate that our method enables even small LLMs (1B parameters) to learn and generate with perfect constraint adherence, outperforming larger counterparts and state-of-the-art reasoning models. This work represents the first integration of context-sensitive grammar learning with LLM generation, eliminating manual specification while maintaining generation validity.

Method: QFS-Composer integrates query decomposition, question generation, question answering, and abstractive summarization. Developed Slovenian QA and QG models based on a Slovene LLM, and adapted reference-free evaluation approaches for Slovenian language.

Result: The QA-guided summarization pipeline yields improved consistency and relevance over baseline LLMs for Slovenian query-focused summarization, establishing an extensible methodology for advancing QFS in less-resourced languages.

Conclusion: The work presents a successful framework for improving query-focused summarization in low-resource languages through integrated QA-guided approaches, with potential for extension to other languages and domains.

Abstract: Large language models (LLMs) demonstrate strong performance in text summarization, yet their effectiveness drops significantly across languages with restricted training resources. This work addresses the challenge of query-focused summarization (QFS) in less-resourced languages, where labeled datasets and evaluation tools are limited. We present a novel QFS framework, QFS-Composer, that integrates query decomposition, question generation (QG), question answering (QA), and abstractive summarization to improve the factual alignment of a summary with user intent. We test our approach on the Slovenian language. To enable high-quality supervision and evaluation, we develop the Slovenian QA and QG models based on a Slovene LLM and adapt evaluation approaches for reference-free summary evaluation. Empirical evaluation shows that the QA-guided summarization pipeline yields improved consistency and relevance over baseline LLMs. Our work establishes an extensible methodology for advancing QFS in less-resourced languages.

Method: Proposes SinkProbe based on attention sinks - tokens that accumulate disproportionate attention mass during generation. The method uses sink scores computed from attention maps, with classifier preferentially relying on sinks whose associated value vectors have large norms. Shows previous methods implicitly depend on attention sinks.

Result: Produces state-of-the-art hallucination detection results across popular datasets and LLMs. Demonstrates mathematical relationship between sink scores and previous methods.

Conclusion: SinkProbe provides a theoretically-grounded hallucination detection method that outperforms existing approaches by exploiting attention sink mechanisms in LLMs.

Abstract: Large language models frequently exhibit hallucinations: fluent and confident outputs that are factually incorrect or unsupported by the input context. While recent hallucination detection methods have explored various features derived from attention maps, the underlying mechanisms they exploit remain poorly understood. In this work, we propose SinkProbe, a hallucination detection method grounded in the observation that hallucinations are deeply entangled with attention sinks - tokens that accumulate disproportionate attention mass during generation - indicating a transition from distributed, input-grounded attention to compressed, prior-dominated computation. Importantly, although sink scores are computed solely from attention maps, we find that the classifier preferentially relies on sinks whose associated value vectors have large norms. Moreover, we show that previous methods implicitly depend on attention sinks by establishing their mathematical relationship to sink scores. Our findings yield a novel hallucination detection method grounded in theory that produces state-of-the-art results across popular datasets and LLMs.

Method: Used 70K manually annotated mentions across 16 genres of English and a novel minimal-pair prompting method to study how overall salience of entities relates to surprisal metrics.

Result: Globally salient entities exhibit significantly higher surprisal than non-salient ones (even controlling for confounds), and reduce surprisal for surrounding content when used as prompts. Effect varies by genre - strongest in topic-coherent texts, weakest in conversational contexts.

Conclusion: Findings refine the UID competing pressures framework by identifying global entity salience as a mechanism shaping information distribution in discourse, with implications for discourse structure and predictability.

Abstract: Previous work examining the Uniform Information Density (UID) hypothesis has shown that while information as measured by surprisal metrics is distributed more or less evenly across documents overall, local discrepancies can arise due to functional pressures corresponding to syntactic and discourse structural constraints. However, work thus far has largely disregarded the relative salience of discourse participants. We fill this gap by studying how overall salience of entities in discourse relates to surprisal using 70K manually annotated mentions across 16 genres of English and a novel minimal-pair prompting method. Our results show that globally salient entities exhibit significantly higher surprisal than non-salient ones, even controlling for position, length, and nesting confounds. Moreover, salient entities systematically reduce surprisal for surrounding content when used as prompts, enhancing document-level predictability. This effect varies by genre, appearing strongest in topic-coherent texts and weakest in conversational contexts. Our findings refine the UID competing pressures framework by identifying global entity salience as a mechanism shaping information distribution in discourse.

Method: Systematic investigation across 13 open-weight LLMs (0.6B-20B parameters). Developed benchmark with 275 personas evaluated on NEO-IPIP agreeableness subscales. Exposed each persona to 4,950 sycophancy-eliciting prompts across 33 topic categories. Analyzed correlations between persona agreeableness and sycophancy rates.

Result: 9 of 13 models showed statistically significant positive correlations between persona agreeableness and sycophancy rates. Pearson correlations reached r = 0.87, with effect sizes as large as Cohen’s d = 2.33. Agreeableness functions as reliable predictor of persona-induced sycophancy.

Conclusion: Agreeableness strongly influences sycophantic behavior in role-playing LLMs, with direct implications for deploying AI systems and developing alignment strategies that account for personality-mediated deceptive behaviors.

Abstract: Large language models increasingly serve as conversational agents that adopt personas and role-play characters at user request. This capability, while valuable, raises concerns about sycophancy: the tendency to provide responses that validate users rather than prioritize factual accuracy. While prior work has established that sycophancy poses risks to AI safety and alignment, the relationship between specific personality traits of adopted personas and the degree of sycophantic behavior remains unexplored. We present a systematic investigation of how persona agreeableness influences sycophancy across 13 small, open-weight language models ranging from 0.6B to 20B parameters. We develop a benchmark comprising 275 personas evaluated on NEO-IPIP agreeableness subscales and expose each persona to 4,950 sycophancy-eliciting prompts spanning 33 topic categories. Our analysis reveals that 9 of 13 models exhibit statistically significant positive correlations between persona agreeableness and sycophancy rates, with Pearson correlations reaching $r = 0.87$ and effect sizes as large as Cohen’s $d = 2.33$. These findings demonstrate that agreeableness functions as a reliable predictor of persona-induced sycophancy, with direct implications for the deployment of role-playing AI systems and the development of alignment strategies that account for personality-mediated deceptive behaviors.

Method: Proposes Self-Correcting RAG with two key innovations: 1) Formalizes context selection as multi-dimensional multiple-choice knapsack problem (MMKP) to maximize information density under token budget, 2) Uses NLI-guided Monte Carlo Tree Search (MCTS) to dynamically explore reasoning trajectories and validate answer faithfulness.

Result: Experiments on six multi-hop QA and fact-checking datasets show significant improvements in reasoning accuracy on complex queries while effectively reducing hallucinations, outperforming strong baselines.

Conclusion: The Self-Correcting RAG framework successfully addresses key limitations of traditional RAG by integrating optimization-based retrieval and planning-based generation, enabling more reliable complex reasoning.

Abstract: Retrieval-augmented generation (RAG) substantially extends the knowledge boundary of large language models. However, it still faces two major challenges when handling complex reasoning tasks: low context utilization and frequent hallucinations. To address these issues, we propose Self-Correcting RAG, a unified framework that reformulates retrieval and generation as constrained optimization and path planning. On the input side, we move beyond traditional greedy retrieval and, for the first time, formalize context selection as a multi-dimensional multiple-choice knapsack problem (MMKP), thereby maximizing information density and removing redundancy under a strict token budget. On the output side, we introduce a natural language inference (NLI)-guided Monte Carlo Tree Search (MCTS) mechanism, which leverages test-time compute to dynamically explore reasoning trajectories and validate the faithfulness of generated answers. Experiments on six multi-hop question answering and fact-checking datasets demonstrate that our method significantly improves reasoning accuracy on complex queries while effectively reducing hallucinations, outperforming strong existing baselines.Our code is available at https://github.com/xjiacs/Self-Correcting-RAG .

Method: Proposes Risk-Constrained Bilevel Stackelberg Framework (RCBSF) that formulates revision as a non-cooperative Stackelberg game with hierarchical Leader-Follower structure: Global Prescriptive Agent imposes risk budgets on follower system (Constrained Revision Agent + Local Verification Agent) for iterative optimization.

Result: Achieves state-of-the-art performance with average Risk Resolution Rate of 84.21%, surpassing iterative baselines while enhancing token efficiency. Theoretical guarantees show convergence to equilibrium with superior utility over unguided configurations.

Conclusion: RCBSF effectively addresses hallucination and constraint issues in automated contract revision through game-theoretic formulation, providing both theoretical guarantees and empirical validation of superior performance.

Abstract: Despite the widespread adoption of Large Language Models (LLMs) in Legal AI, their utility for automated contract revision remains impeded by hallucinated safety and a lack of rigorous behavioral constraints. To address these limitations, we propose the Risk-Constrained Bilevel Stackelberg Framework (RCBSF), which formulates revision as a non-cooperative Stackelberg game. RCBSF establishes a hierarchical Leader Follower structure where a Global Prescriptive Agent (GPA) imposes risk budgets upon a follower system constituted by a Constrained Revision Agent (CRA) and a Local Verification Agent (LVA) to iteratively optimize output. We provide theoretical guarantees that this bilevel formulation converges to an equilibrium yielding strictly superior utility over unguided configurations. Empirical validation on a unified benchmark demonstrates that RCBSF achieves state-of-the-art performance, surpassing iterative baselines with an average Risk Resolution Rate (RRR) of 84.21% while enhancing token efficiency. Our code is available at https://github.com/xjiacs/RCBSF .

Method: Three-component framework: (1) Agentic Multimodal Search and Filtering for retrieving text passages and information-dense visuals, (2) Checklist-Guided Incremental Synthesis for coherent image-text integration and citation placement, (3) Recurrent Context Management for balancing long-range coherence with local fluency. Includes a curation pipeline producing 8K agentic traces for model optimization.

Result: Created M2LongBench with 247 research tasks across 9 domains and a stable multimodal sandbox. Experiments show multimodal long-form generation is challenging, especially in multimodal selection and integration, but effective post-training can bridge the gap.

Conclusion: Deep-Reporter addresses the pressing need for multimodal long-form generation by providing a unified framework that effectively integrates visual and textual information through agentic processes, with demonstrated improvements through post-training.

Abstract: Recent agentic search frameworks enable deep research via iterative planning and retrieval, reducing hallucinations and enhancing factual grounding. However, they remain text-centric, overlooking the multimodal evidence that characterizes real-world expert reports. We introduce a pressing task: multimodal long-form generation. Accordingly, we propose Deep-Reporter, a unified agentic framework for grounded multimodal long-form generation. It orchestrates: (i) Agentic Multimodal Search and Filtering to retrieve and filter textual passages and information-dense visuals; (ii) Checklist-Guided Incremental Synthesis to ensure coherent image-text integration and optimal citation placement; and (iii) Recurrent Context Management to balance long-range coherence with local fluency. We develop a rigorous curation pipeline producing 8K high-quality agentic traces for model optimization. We further introduce M2LongBench, a comprehensive testbed comprising 247 research tasks across 9 domains and a stable multimodal sandbox. Extensive experiments demonstrate that long-form multimodal generation is a challenging task, especially in multimodal selection and integration, and effective post-training can bridge the gap.

Method: Created the first large-scale benchmark of diverse yet semantically identical query variations combining human-written and model-generated rewrites. Systematically evaluated Adaptive RAG robustness across three dimensions: answer quality, computational cost, and retrieval decisions.

Result: Discovered a critical robustness gap where small surface-level changes in queries dramatically alter retrieval behavior and accuracy. Larger models showed better performance but robustness did not improve accordingly. Adaptive RAG methods are highly vulnerable to query variations that preserve identical semantics.

Conclusion: Adaptive RAG systems face a critical robustness challenge where semantically identical query variations can significantly impact system behavior and accuracy, revealing a fundamental vulnerability that needs to be addressed for reliable real-world deployment.

Abstract: Adaptive Retrieval-Augmented Generation (RAG) promises accuracy and efficiency by dynamically triggering retrieval only when needed and is widely used in practice. However, real-world queries vary in surface form even with the same intent, and their impact on Adaptive RAG remains under-explored. We introduce the first large-scale benchmark of diverse yet semantically identical query variations, combining human-written and model-generated rewrites. Our benchmark facilitates a systematic evaluation of Adaptive RAG robustness by examining its key components across three dimensions: answer quality, computational cost, and retrieval decisions. We discover a critical robustness gap, where small surface-level changes in queries dramatically alter retrieval behavior and accuracy. Although larger models show better performance, robustness does not improve accordingly. These findings reveal that Adaptive RAG methods are highly vulnerable to query variations that preserve identical semantics, exposing a critical robustness challenge.

Method: Uses LLMs to construct knowledge graphs from input documents. MCQs are generated by selecting KG nodes as keys, sampling related triples/quintuples (optionally augmented), prompting LLMs to create stems, and selecting distractors from the KG. Nine difficulty signals are computed and combined into a unified score using data-driven approach.

Result: The method generates high-quality MCQs with interpretable difficulty estimation that aligns with human perceptions, improving automated MCQ generation through structured knowledge representations and data-driven difficulty modeling.

Conclusion: The approach successfully integrates structured knowledge representations with LLMs and data-driven difficulty estimation to enhance automated MCQ generation for adaptive education systems.

Abstract: Generating multiple-choice questions (MCQs) with difficulty estimation remains challenging in automated MCQ-generation systems used in adaptive, AI-assisted education. This study proposes a novel methodology for generating MCQs with difficulty estimation from the input documents by utilizing knowledge graphs (KGs) and large language models (LLMs). Our approach uses an LLM to construct a KG from input documents, from which MCQs are then systematically generated. Each MCQ is generated by selecting a node from the KG as the key, sampling a related triple or quintuple – optionally augmented with an extra triple – and prompting an LLM to generate a corresponding stem from these graph components. Distractors are then selected from the KG. For each MCQ, nine difficulty signals are computed and combined into a unified difficulty score using a data-driven approach. Experimental results demonstrate that our method generates high-quality MCQs whose difficulty estimation is interpretable and aligns with human perceptions. Our approach improves automated MCQ generation by integrating structured knowledge representations with LLMs and a data-driven difficulty estimation model.

Method: Used LLM to create token-level dataset labeled with four narrative categories plus “others,” then applied linear probing on BERT embeddings and compared with variance-matched random embeddings control.

Result: BERT probe achieved 94% accuracy vs 47% control, with macro-average recall of 0.83. Rare categories like causality (0.75) and space (0.66) showed moderate success but suffered from “Boundary Leakage” misclassification. Unsupervised clustering aligned poorly with categories (ARI = 0.081).

Conclusion: BERT encodes meaningful narrative information but not as discretely separable clusters, with rare dimensions prone to misclassification as “others.” Future work includes syntactic baseline and expanded datasets.

Abstract: Narrative understanding requires multidimensional semantic structures. This study investigates whether BERT embeddings encode dimensions of fictional narrative semantics – time, space, causality, and character. Using an LLM to accelerate annotation, we construct a token-level dataset labeled with these four narrative categories plus “others.” A linear probe on BERT embeddings (94% accuracy) significantly outperforms a control probe on variance-matched random embeddings (47%), confirming that BERT encodes meaningful narrative information. With balanced class weighting, the probe achieves a macro-average recall of 0.83, with moderate success on rare categories such as causality (recall = 0.75) and space (recall = 0.66). However, confusion matrix analysis reveals “Boundary Leakage,” where rare dimensions are systematically misclassified as “others.” Clustering analysis shows that unsupervised clustering aligns near-randomly with predefined categories (ARI = 0.081), suggesting that narrative dimensions are encoded but not as discretely separable clusters. Future work includes a POS-only baseline to disentangle syntactic patterns from narrative encoding, expanded datasets, and layer-wise probing.

Method: Created Mediom corpus with 3,533 Hindi, Bengali, and Thai idioms featuring explanations, translations, and text-image representations. Developed HIDE framework using error-feedback retrieval and diagnostic cues for iterative reasoning refinement.

Result: Benchmarked LLMs and VLMs on Mediom, exposing systematic failures in metaphor comprehension. HIDE framework shows promise for improving idiom understanding through targeted reasoning hints.

Conclusion: Mediom provides a rigorous test bed for culturally grounded, multimodal idiom understanding, while HIDE offers methodology for embedding reasoning hints in next-generation AI systems.

Abstract: Idiomatic reasoning, deeply intertwined with metaphor and culture, remains a blind spot for contemporary language models, whose progress skews toward surface-level lexical and semantic cues. For instance, the Bengali idiom \textit{\foreignlanguage{bengali}{\char"0986\char"0999\char"09CD\char"0997\char"09C1 \char"09B0 \char"09AB\char"09B2 \char"099F\char"0995}} (angur fol tok, grapes are sour''): it encodes denial-driven rationalization, yet naive models latch onto the literal fox-and-grape imagery. Addressing this oversight, we present Mediom,’’ a multilingual, multimodal idiom corpus of 3,533 Hindi, Bengali, and Thai idioms, each paired with gold-standard explanations, cross-lingual translations, and carefully aligned text–image representations. We benchmark both large language models (textual reasoning) and vision-language models (figurative disambiguation) on Mediom, exposing systematic failures in metaphor comprehension. To mitigate these gaps, we propose ``HIDE,’’ a Hinting-based Idiom Explanation framework that leverages error-feedback retrieval and targeted diagnostic cues for iterative reasoning refinement. Collectively, Mediom and HIDE establish a rigorous test bed and methodology for culturally grounded, multimodal idiom understanding embedded with reasoning hints in next-generation AI systems.

Method: Proposes TInR-U framework with three-phase training: 1) tool internalization using bidirectional knowledge alignment strategy, 2) supervised fine-tuning warm-up with high-quality reasoning annotations, and 3) reinforcement learning with TInR-specific rewards.

Result: TInR-U achieves superior performance in both in-domain and out-of-domain settings, demonstrating effectiveness and efficiency compared to existing methods.

Conclusion: Tool-Internalized Reasoning (TInR) through the TInR-U framework successfully addresses limitations of external tool documentation, enabling more efficient and effective reasoning with internalized tool knowledge in LLMs.

Abstract: Tool-Integrated Reasoning (TIR) has emerged as a promising direction by extending Large Language Models’ (LLMs) capabilities with external tools during reasoning. Existing TIR methods typically rely on external tool documentation during reasoning. However, this leads to tool mastery difficulty, tool size constraints, and inference inefficiency. To mitigate these issues, we explore Tool-Internalized Reasoning (TInR), aiming at facilitating reasoning with tool knowledge internalized into LLMs. Achieving this goal presents notable requirements, including tool internalization and tool-reasoning coordination. To address them, we propose TInR-U, a tool-internalized reasoning framework for unified reasoning and tool usage. TInR-U is trained through a three-phase pipeline: 1) tool internalization with a bidirectional knowledge alignment strategy; 2) supervised fine-tuning warm-up using high-quality reasoning annotations, and 3) reinforcement learning with TInR-specific rewards. We comprehensively evaluate our method across in-domain and out-of-domain settings. Experiment results show that TInR-U achieves superior performance in both settings, highlighting its effectiveness and efficiency.

Method: Two complementary modifications: 1) Insert non-linear MLP between layer norm and Q/K/V projections to create richer position-agnostic features before positional encoding. 2) Add content skip connection that routes pre-projection features around attention mechanism, allowing content information to bypass position-aware attention.

Result: Combined approach achieves strongest results: +40.6% LAMBADA accuracy and -39% perplexity at 160M scale on Pythia models. Learned skip weights show consistent pattern: later transformer layers activate content bypass more strongly than earlier layers. No K/V cache overhead added.

Conclusion: The modifications improve transformer performance by separating content processing from positional attention, with deeper layers benefiting more from content information that bypasses positional encoding. The approach is efficient with no additional cache overhead.

Abstract: We propose two complementary modifications to transformer attention blocks. First, a non-linear pre-projection MLP is inserted between layer norm and Q/K/V projections, constructing richer features in a position-agnostic manner before any positional encoding is applied. Second, a content skip connection routes the pre-projection’s features around the attention mechanism, allowing content information to bypass position-aware attention where beneficial. In frozen-probe experiments on Pythia-160M and 410M, the combined approach achieves the strongest results across methods: +40.6% LAMBADA accuracy and -39% perplexity at 160M scale. Learned skip connection weights reveal a consistent pattern across model sizes: later transformer layers activate the content bypass more strongly than earlier layers, suggesting that deeper layers benefit from content information that does not pass through positional attention. All modifications add no K/V cache overhead.

Method: Transition from universal Mistral-based tokenization to dedicated Polish-optimized vocabulary, using FOCUS-based embedding initialization, multi-stage pretraining curriculum, and post-training alignment with Supervised Fine-Tuning, Direct Preference Optimization, and Reinforcement Learning through Group Relative Policy Optimization with verifiable rewards.

Result: Development of Bielik v3 PL series with 7B and 11B parameter variants representing significant milestone in language-specific LLM optimization for Polish.

Conclusion: Language-specific tokenization optimization addresses fundamental architectural inefficiencies in multilingual LLMs, enabling better performance for specific languages like Polish.

Abstract: The development of the Bielik v3 PL series, encompassing both the 7B and 11B parameter variants, represents a significant milestone in the field of language-specific large language model (LLM) optimization. While general-purpose models often demonstrate impressive multilingual capabilities, they frequently suffer from a fundamental architectural inefficiency: the use of universal tokenizers. These tokenizers, typically designed to cover a broad spectrum of languages, often fail to capture the morphological nuances of specific languages like Polish, leading to higher fertility ratios, increased inference costs, and restricted effective context windows. This report details the transition from the universal Mistral-based tokenization to a dedicated Polish-optimized vocabulary for the Bielik v3 models, exploring the FOCUS-based embedding initialization, the multi-stage pretraining curriculum, and the subsequent post-training alignment involving Supervised Fine-Tuning, Direct Preference Optimization, and Reinforcement Learning through Group Relative Policy Optimization with verifiable rewards.

Method: Uses Language World Models (LWMs) to simulate domain-specific environments through LLM-driven tool response generation. Multi-agent synthesis pipeline automatically produces evaluation instances with guaranteed solvability, calibrated difficulty, and document-grounded diversity. Evaluates agents along task completion across domains and environmental robustness under controlled fault injection (explicit errors, implicit data degradation, mixed faults).

Result: Evaluated 15 frontier models across 8 families: (1) No single model dominates all industries, each has distinct occupational capability profile; (2) Implicit faults (truncated data, missing fields) are harder than explicit errors and mixed faults; (3) Larger models, newer generations, and higher reasoning effort consistently improve performance (GPT-5.2 improves 27.5 points from minimal to maximum reasoning effort); (4) Strong agents are not necessarily strong environment simulators - simulator quality is critical for LWM-based evaluation reliability.

Conclusion: OccuBench provides the first systematic cross-industry evaluation of AI agents on professional occupational tasks, enabling comprehensive assessment of agent capabilities across diverse real-world professional domains.

Abstract: AI agents are expected to perform professional work across hundreds of occupational domains (from emergency department triage to nuclear reactor safety monitoring to customs import processing), yet existing benchmarks can only evaluate agents in the few domains where public environments exist. We introduce OccuBench, a benchmark covering 100 real-world professional task scenarios across 10 industry categories and 65 specialized domains, enabled by Language World Models (LWMs) that simulate domain-specific environments through LLM-driven tool response generation. Our multi-agent synthesis pipeline automatically produces evaluation instances with guaranteed solvability, calibrated difficulty, and document-grounded diversity. OccuBench evaluates agents along two complementary dimensions: task completion across professional domains and environmental robustness under controlled fault injection (explicit errors, implicit data degradation, and mixed faults). We evaluate 15 frontier models across 8 model families and find that: (1) no single model dominates all industries, as each has a distinct occupational capability profile; (2) implicit faults (truncated data, missing fields) are harder than both explicit errors (timeouts, 500s) and mixed faults, because they lack overt error signals and require the agent to independently detect data degradation; (3) larger models, newer generations, and higher reasoning effort consistently improve performance. GPT-5.2 improves by 27.5 points from minimal to maximum reasoning effort; and (4) strong agents are not necessarily strong environment simulators. Simulator quality is critical for LWM-based evaluation reliability. OccuBench provides the first systematic cross-industry evaluation of AI agents on professional occupational tasks.

Method: Proposes AOP-Smart, an AOP-oriented RAG framework that retrieves relevant knowledge from official AOP-Wiki XML data using Key Events, Key Event Relationships, and specific AOP information to augment LLM responses for user questions.

Result: On 20 AOP-related QA tasks, RAG dramatically improved accuracy: GPT from 15.0% to 95.0%, DeepSeek from 35.0% to 100.0%, and Gemini from 20.0% to 95.0%, showing significant reduction in hallucinations and improved answer consistency.

Conclusion: AOP-Smart effectively mitigates LLM hallucination problems in AOP knowledge tasks, substantially enhancing answer accuracy and reliability for toxicological research applications.

Abstract: Adverse Outcome Pathways (AOPs) are an important knowledge framework in toxicological research and risk assessment. In recent years, large language models (LLMs) have gradually been applied to AOP-related question answering and mechanistic reasoning tasks. However, due to the existence of the hallucination problem, that is, the model may generate content that is inconsistent with facts or lacks evidence, their reliability is still limited. To address this issue, this study proposes an AOP-oriented Retrieval-Augmented Generation (RAG) framework, AOP-Smart. Based on the official XML data from AOP-Wiki, this method uses Key Events (KEs), Key Event Relationships (KERs), and specific AOP information to retrieve relevant knowledge for user questions, thereby improving the reliability of the generated results of large language models. To evaluate the effectiveness of the proposed method, this study constructed a test set containing 20 AOP-related question answering tasks, covering KE identification, upstream and downstream KE retrieval, and complex AOP retrieval tasks. Experiments were conducted on three mainstream large language models, Gemini, DeepSeek, and ChatGPT, and comparative tests were performed under two settings: without RAG and with RAG. The experimental results show that, without using RAG, the accuracies of GPT, DeepSeek, and Gemini were 15.0%, 35.0%, and 20.0%, respectively; after using RAG, their accuracies increased to 95.0%, 100.0%, and 95.0%, respectively. The results indicate that AOP-Smart can significantly alleviate the hallucination problem of large language models in AOP knowledge tasks, and greatly improve the accuracy and consistency of their answers.

Method: Proposes Hybrid Toolset Agentization & Adaptation (HTAA): 1) Toolset agentization - encapsulating frequently co-used tools into specialized agent tools to reduce action space; 2) Asymmetric Planner Adaptation - trajectory-based training aligning high-level planner with agent tools via backward reconstruction and forward refinement.

Result: HTAA achieves higher task success rates, shorter tool calling trajectories, and significantly reduces context overhead compared to baselines on InfoVerify dataset and other benchmarks. In production deployment, it substantially reduces manual validation effort and operational costs.

Conclusion: HTAA provides an effective hierarchical framework for scalable tool-use in LLMs, addressing inefficiencies of flat architectures and enabling reliable use of hundreds of tools for real-world applications.

Abstract: Enabling large language models to scale and reliably use hundreds of tools is critical for real-world applications, yet challenging due to the inefficiency and error accumulation inherent in flat tool-calling architectures. To address this, we propose Hybrid Toolset Agentization & Adaptation (HTAA), a hierarchical framework for scalable tool-use planning. We propose a novel toolset agentization paradigm, which encapsulates frequently co-used tools into specialized agent tools, thereby reducing the planner’s action space and mitigating redundancy. To ensure effective coordination, we design Asymmetric Planner Adaptation, a trajectory-based training paradigm that aligns the high-level planner with agent tools via backward reconstruction and forward refinement. To validate the performance of HTAA, we conduct experiments on a real-world internal dataset, InfoVerify, based on the POI validation workflow of China’s largest online large-scale ride-hailing platform, featuring long-horizon executable tool trajectories. Experiments on InfoVerify and widely-used benchmarks show that HTAA consistently achieves higher task success rates, requires short tool calling trajectories, and significantly reduces context overhead compared to strong baselines. Furthermore, in a production deployment, HTAA substantially reduces manual validation effort and operational cost, demonstrating its practical efficacy.

Method: Proposes Mem2Evolve with two core components: Experience Memory and Asset Memory. The framework leverages accumulated experience to guide dynamic creation of new assets (tools/expert agents), while simultaneously acquiring new experience from using these assets, enabling co-evolution.

Result: Extensive experiments across 6 task categories and 8 benchmarks show Mem2Evolve achieves 18.53% improvement over standard LLMs, 11.80% over experience-only agents, and 6.46% over asset-creation-only agents, demonstrating superior effectiveness and stability.

Conclusion: Mem2Evolve establishes a more effective and stable self-evolving agent framework by integrating experience distillation and capability expansion through co-evolution, overcoming limitations of isolated evolutionary approaches.

Abstract: While large language model–powered agents can self-evolve by accumulating experience or by dynamically creating new assets (i.e., tools or expert agents), existing frameworks typically treat these two evolutionary processes in isolation. This separation overlooks their intrinsic interdependence: the former is inherently bounded by a manually predefined static toolset, while the latter generates new assets from scratch without experiential guidance, leading to limited capability growth and unstable evolution. To address this limitation, we introduce a novel paradigm of co-evolutionary Capability Expansion and Experience Distillation. Guided by this paradigm, we propose the \textbf{Mem$^{\textbf{2}}$Evolve}, which integrates two core components: \textbf{Experience Memory} and \textbf{Asset Memory}. Specifically, Mem$^{2}$Evolve leverages accumulated experience to guide the dynamic creation of assets, thereby expanding the agent’s capability space while simultaneously acquiring new experience to achieve co-evolution. Extensive experiments across 6 task categories and 8 benchmarks demonstrate that Mem$^{2}$Evolve achieves improvement of 18.53% over standard LLMs, 11.80% over agents evolving solely through experience, and 6.46% over those evolving solely through asset creation, establishing it as a substantially more effective and stable self-evolving agent framework. Code is available at: https://buaa-irip-llm.github.io/Mem2Evolve.

Method: Introduces Information Elicitation Agents (IEAs) with a formalization as a finite-horizon POMDP. Creates YIELD dataset of 2,281 human-to-human dialogues (26M tokens) for training. Proposes novel metrics tailored to IEAs and conducts experiments with foundation LLMs.

Result: Training on YIELD improves LLM alignment with real elicitation behavior. Findings are corroborated by human evaluation. The dataset, code, evaluation tools, and fine-tuned model adapters are publicly released.

Conclusion: The paper establishes a framework for systematic research on information elicitation agents, providing a dataset, formalization, and evaluation metrics to advance this important but understudied area of conversational AI.

Abstract: Most conversational agents (CAs) are designed to satisfy user needs through user-driven interactions. However, many real-world settings, such as academic interviewing, judicial proceedings, and journalistic investigations, involve broader institutional decision-making processes and require agents that can elicit information from users. In this paper, we introduce Information Elicitation Agents (IEAs) in which the agent’s goal is to elicit information from users to support the agent’s institutional or task-oriented objectives. To enable systematic research on this setting, we present YIELD, a 26M-token dataset of 2,281 ethically sourced, human-to-human dialogues. Moreover, we formalize information elicitation as a finite-horizon POMDP and propose novel metrics tailored to IEAs. Pilot experiments on multiple foundation LLMs show that training on YIELD improves their alignment with real elicitation behavior and findings are corroborated by human evaluation. We release YIELD under CC BY 4.0. The dataset, project code, evaluation tools, and fine-tuned model adapters are available at: https://github.com/infosenselab/yield.

Method: Constructs compositionally infeasible claims where the salient constraint is supported but a non-salient constraint is contradicted. Tests various model families and modalities, and uses model context interventions to analyze verification behavior.

Result: Models that saturate existing benchmarks consistently over-accept compositionally infeasible claims, confirming prevalence of shortcut reasoning. Different models occupy distinct positions on a shared ROC curve, indicating differences in verification thresholds rather than underlying reasoning ability.

Conclusion: Current verification benchmarks are insufficient, and the compositional inference bottleneck is a structural property of current verification behavior that cannot be overcome by strategy guidance alone.

Abstract: Scientific claim verification, the task of determining whether claims are entailed by scientific evidence, is fundamental to establishing discoveries in evidence while preventing misinformation. This process involves evaluating each asserted constraint against validated evidence. Under the Closed-World Assumption (CWA), a claim is accepted if and only if all asserted constraints are positively supported. We show that existing verification benchmarks cannot distinguish models enforcing this standard from models applying a simpler shortcut called salient-constraint checking, which applies CWA’s rejection criterion only to the most salient constraint and accepts when that constraint is supported. Because existing benchmarks construct infeasible claims by perturbing a single salient element they are insufficient at distinguishing between rigorous claim verification and simple salient-constraint reliance. To separate the two, we construct compositionally infeasible claims where the salient constraint is supported but a non-salient constraint is contradicted. Across model families and modalities, models that otherwise saturate existing benchmarks consistently over-accept these claims, confirming the prevalence of such shortcut reasoning. Via model context interventions, we show that different models and prompting strategies occupy distinct positions on a shared ROC curve, indicating that the gap between model families reflects differences in verification threshold rather than underlying reasoning ability, and that the compositional inference bottleneck is a structural property of current verification behavior that strategy guidance alone cannot overcome.

Method: Built a modular pipeline with frozen LLM as stateless feature extractor, transforming daily news/filings into fixed-dimensional vectors for downstream PPO agent. Introduced automated prompt-optimization loop treating extraction prompt as discrete hyperparameter, tuned directly against Information Coefficient (Spearman correlation) rather than NLP losses.

Result: Optimized prompt discovered genuinely predictive features (IC > 0.15 on held-out data). However, during macroeconomic shock distribution shift, LLM-derived features added noise and underperformed price-only baseline. In calmer test regime, agent recovered but macroeconomic state variables remained most robust driver of policy improvement.

Conclusion: There’s a gap between feature-level validity and policy-level robustness, paralleling known transfer learning challenges under distribution shift. LLM features can be predictive but may not translate to robust downstream task performance during real-world distribution shifts.

Abstract: Can large language models (LLMs) generate continuous numerical features that improve reinforcement learning (RL) trading agents? We build a modular pipeline where a frozen LLM serves as a stateless feature extractor, transforming unstructured daily news and filings into a fixed-dimensional vector consumed by a downstream PPO agent. We introduce an automated prompt-optimization loop that treats the extraction prompt as a discrete hyperparameter and tunes it directly against the Information Coefficient - the Spearman rank correlation between predicted and realized returns - rather than NLP losses. The optimized prompt discovers genuinely predictive features (IC above 0.15 on held-out data). However, these valid intermediate representations do not automatically translate into downstream task performance: during a distribution shift caused by a macroeconomic shock, LLM-derived features add noise, and the augmented agent under-performs a price-only baseline. In a calmer test regime the agent recovers, yet macroeconomic state variables remain the most robust driver of policy improvement. Our findings highlight a gap between feature-level validity and policy-level robustness that parallels known challenges in transfer learning under distribution shift.

Method: Atomic predicate-argument decomposition with calibrated uncertainty-gated corroboration: facts are aligned to local snippets via embeddings, verified by compact evidence-grounded checker, and only uncertain facts trigger domain-restricted web search over authoritative sources.

Result: The framework surpasses strongest benchmarks on multiple benchmarks, with web corroboration invoked for only a minority of atomic facts on average, showing selective external evidence consultation under calibrated uncertainty.

Conclusion: Coupling atomic granularity with calibrated, uncertainty-gated corroboration yields more interpretable and context-conditioned verification suitable for high-stakes, single-document settings requiring traceable rationales and predictable cost/latency.

Abstract: Scientific fact-checking is vital for assessing claims in specialized domains such as biomedicine and materials science, yet existing systems often hallucinate or apply inconsistent reasoning, especially when verifying technical, compositional claims against an evidence snippet under source and cost/latency constraints. We present a pipeline centered on atomic predicate-argument decomposition and calibrated, uncertainty-gated corroboration: atomic facts are aligned to local snippets via embeddings, verified by a compact evidence-grounded checker, and only facts with uncertain support trigger domain-restricted web search over authoritative sources. The system supports both binary and tri-valued classification where it predicts labels from Supported, Refuted, NEI for three-way tasks. We evaluate under two regimes, Context-Only (no web) and Context+Web (uncertainty-gated web corroboration); when retrieved evidence conflicts with the provided context, we abstain with NEI rather than overriding the context. On multiple benchmarks, our framework surpasses the strongest benchmarks. In our experiments, web corroboration was invoked for only a minority of atomic facts on average, indicating that external evidence is consulted selectively under calibrated uncertainty rather than routinely. Overall, coupling atomic granularity with calibrated, uncertainty-gated corroboration yields more interpretable and context-conditioned verification, making the approach well-suited to high-stakes, single-document settings that demand traceable rationales, predictable cost/latency, and conservative.

Method: Used Neuron-based Personality Trait Induction (NPTI) to induce Big Five personality traits in LLMs, evaluated performance across six cognitive benchmarks, analyzed task dependence and trait effects, and proposed Dynamic Persona Routing (DPR) for query-adaptive persona selection.

Result: Persona induction produces stable, reproducible cognitive performance shifts beyond stylistic changes. Effects are strongly task-dependent: some personalities improve instruction-following while others impair complex reasoning. Openness and Extraversion have strongest effects. LLM personality-cognition relationships show 73.68% directional consistency with human patterns. DPR outperforms best static persona without training.

Conclusion: Persona induction meaningfully affects LLM cognitive capabilities, not just interaction style. The discovered regularities enable practical applications like adaptive persona routing for performance optimization, bridging LLM behavior with human personality-cognition relationships.

Abstract: Imbuing Large Language Models (LLMs) with specific personas is prevalent for tailoring interaction styles, yet the impact on underlying cognitive capabilities remains unexplored. We employ the Neuron-based Personality Trait Induction (NPTI) framework to induce Big Five personality traits in LLMs and evaluate performance across six cognitive benchmarks. Our findings reveal that persona induction produces stable, reproducible shifts in cognitive task performance beyond surface-level stylistic changes. These effects exhibit strong task dependence: certain personalities yield consistent gains on instruction-following, while others impair complex reasoning. Effect magnitude varies systematically by trait dimension, with Openness and Extraversion exerting the most robust influence. Furthermore, LLM effects show 73.68% directional consistency with human personality-cognition relationships. Capitalizing on these regularities, we propose Dynamic Persona Routing (DPR), a lightweight query-adaptive strategy that outperforms the best static persona without additional training.

Method: Extracted 21-emotion vector sets from 12 small language models (6 architectures × base/instruct versions, 1B-8B parameters) using a unified comprehension-mode pipeline at fp16 precision, analyzed via representational similarity analysis on raw cosine RDMs (representational dissimilarity matrices).

Result: Five mature architectures share nearly identical 21-emotion geometry (pairwise RDM Spearman correlations 0.74-0.92), with universality persisting across diametrically opposed behavioral profiles. Gemma-3 1B base (immature case) shows extreme residual-stream anisotropy and is restructured by RLHF, while mature families show high within-family base×instruct correlations (rho ≥ 0.92). Methodological decomposition reveals four distinct layers in prior comprehension-vs-generation effects.

Conclusion: Mature language models share universal emotion representations despite behavioral differences, suggesting behavioral facets arise above shared emotion representations. RLHF restructures only representations that are not yet organized. Methodological analysis shows prior comprehension-vs-generation effects decompose into multiple layers requiring careful interpretation.

Abstract: We extract 21-emotion vector sets from twelve small language models (six architectures x base/instruct, 1B-8B parameters) under a unified comprehension-mode pipeline at fp16 precision, and compare the resulting geometries via representational similarity analysis on raw cosine RDMs. The five mature architectures (Qwen 2.5 1.5B, SmolLM2 1.7B, Llama 3.2 3B, Mistral 7B v0.3, Llama 3.1 8B) share nearly identical 21-emotion geometry, with pairwise RDM Spearman correlations of 0.74-0.92. This universality persists across diametrically opposed behavioral profiles: Qwen 2.5 and Llama 3.2 occupy opposite poles of MTI Compliance facets yet produce nearly identical emotion RDMs (rho = 0.81), so behavioral facet differences arise above the shared emotion representation. Gemma-3 1B base, the one immature case in our dataset, exhibits extreme residual-stream anisotropy (0.997) and is restructured by RLHF across all geometric descriptors, whereas the five already-mature families show within-family base x instruct RDM correlations of rho >= 0.92 (Mistral 7B v0.3 at rho = 0.985), suggesting RLHF restructures only representations that are not yet organized. Methodologically, we show that what prior work has read as a single comprehension-vs-generation method effect in fact decomposes into four distinct layers – a coarse method-dependent dissociation, robust sub-parameter sensitivity within generation, a true precision (fp16 vs INT8) effect, and a conflated cross-experiment bias that distorts in opposite directions for different models – so that a single rho between two prior emotion-vector studies is not a safe basis for interpretation without the layered decomposition.

Method: Dataset assembled from two sources: digitized archival/literary material converted from InPage format, and Unicode-native web text. All text processed through an eleven-stage cleaning pipeline to ensure script purity, then tokenized using google/muril-base-cased tokenizer.

Result: Created KS-PRET-5M dataset with 5.09M words, 27.6M characters, 295K unique word types, achieving high Kashmiri script ratio (0.9965) with minimal Devanagari contamination. Tokenization yielded 12.13M subword tokens with 2.383 tokens per word ratio.

Conclusion: The dataset enables language model pretraining, tokenizer training, and computational linguistic research for Kashmiri, addressing resource scarcity for this low-resource language.

Abstract: We present KS-PRET-5M, the largest publicly available pretraining dataset for the Kashmiri language, comprising 5,090,244 (5.09M) words, 27,692,959 (27.6M) characters, and a vocabulary of 295,433 (295.4K) unique word types. We assembled the dataset from two source classes: digitized archival and literary material, encompassing literature, news, biographies, novels, poetry, religious scholarship, and academic writing, recovered from the proprietary InPage desktop-publishing format using the converter of Malik~\cite{malik2024inpage}, and Unicode-native text collected from Kashmiri-language web sources. All text was processed through an eleven-stage cleaning pipeline that achieves a mean Kashmiri script ratio of 0.9965, reducing Devanagari contamination to 146 characters across the full dataset. We tokenized the dataset empirically using google/muril-base-cased, yielding a subword ratio of 2.383 tokens per word and a total of approximately 12.13 million subword tokens, substantially higher than prior estimates derived from non-Kashmiri Perso-Arabic analogues. KS-PRET-5M is released as a single continuous text stream under CCBY4.0 to support language model pretraining, tokenizer training, and computational linguistic research for Kashmiri.

Method: Two-stage approach: 1) Supervised fine-tuning of Qwen 2.5-7B-Instruct with LoRA using interpretable slot-filling template (target, claim type, manifestation checklist, justification), 2) DPO refinement with automatically generated preference pairs to reduce false negatives.

Result: DPO refinement improves both accuracy and reduces false negatives without extra annotation. On English development set: recall increased from 0.5085 to 0.7797, macro-F1 improved by ~5 points.

Conclusion: Preference-based refinement with DPO effectively enhances polarization detection performance in social media text, demonstrating the value of combining structured fine-tuning with preference optimization.

Abstract: The POLAR SemEval-2026 Shared Task aims to detect online polarization and focuses on the classification and identification of multilingual, multicultural, and multi-event polarization. Accurate computational detection of online polarization is challenging due to nuanced rhetoric, implicit framing, and the high cost of human-in-the-loop annotation. Building on recent findings that contextual prompting enables large language models to function as strong polarization detectors, we present a two-stage approach for detecting political polarization in social media text that combines structured supervised fine-tuning with Direct Preference Optimization (DPO) refinement. We fine-tune Qwen 2.5-7B-Instruct with LoRA using an interpretable slot-filling template (target, claim type, manifestation checklist, and justification). We then apply DPO with automatically generated preference pairs to reduce costly false negatives. Experiments on the SemEval 2026 POLAR shared task dataset show that preference-based refinement improves both accuracy and decreases false negatives without extra annotation. On the English development set, DPO increases recall from 0.5085 to 0.7797 and improves macro-F1 by ~5 points.

Method: Constructs task vectors directly in decoding space by capturing the difference between output logit distributions of few-shot and zero-shot prompts, then injects this vector into the decoding process to steer generation.

Result: Outperforms standard few-shot baselines across seven LLMs (0.5B-9B) on TruthfulQA, Math-500, and AQUA-RAT with gains up to +5.50 average accuracy. Effectively suppresses generation degeneration and logical flaws with strong robustness to demonstration ordering.

Conclusion: DeCoVec offers a training-free, non-invasive solution for LLM steering without requiring weight updates or auxiliary models, providing an efficient alternative to existing approaches.

Abstract: Task vectors, representing directions in model or activation spaces that encode task-specific behaviors, have emerged as a promising tool for steering large language models (LLMs). However, existing approaches typically require fine-tuning or invasive manipulation of internal states, limiting their flexibility and scalability. We propose \textsc{DeCoVec} (Decoding Space based Task Vector), a training-free and non-invasive framework that constructs task vectors directly in the \textit{decoding space} by leveraging in-context learning (ICL). Specifically, \textsc{DeCoVec} captures the task essence as the difference between the output logit distributions of few-shot and zero-shot prompts, then steers generation by injecting this vector into the decoding process. Experiments across seven LLMs (0.5B–9B) on TruthfulQA, Math-500, and AQUA-RAT show that \textsc{DeCoVec} consistently outperforms standard few-shot baselines, with gains up to +5.50 average accuracy. Further analysis demonstrates that \textsc{DeCoVec} effectively suppresses generation degeneration and logical flaws while exhibiting strong robustness to demonstration ordering, all without incurring additional input token costs. Our method offers a training-free and non-invasive solution for LLM steering without requiring weight updates or auxiliary models.

Method: Created ClinicNumRobBench with 1,624 context-question instances evaluating four clinical numeracy types: value retrieval, arithmetic computation, relational comparison, and aggregation. Used MIMIC-IV vital-sign records in three semantically equivalent representations, including real-world note-style variant, with 42 question templates.

Result: Value retrieval was strong (most models >85% accuracy), but relational comparison and aggregation were challenging (some models <15%). Fine-tuning on medical data reduced numeracy by over 30% relative to base models, and performance dropped under note-style variations.

Conclusion: ClinicNumRobBench provides a rigorous testbed for clinically reliable numerical reasoning, revealing LLM sensitivity to format variations and significant challenges in complex numerical reasoning tasks despite strong performance on simple value retrieval.

Abstract: Large Language Models (LLMs) are increasingly being explored for clinical question answering and decision support, yet safe deployment critically requires reliable handling of patient measurements in heterogeneous clinical notes. Existing evaluations of LLMs for clinical numerical reasoning provide limited operation-level coverage, restricted primarily to arithmetic computation, and rarely assess the robustness of numerical understanding across clinical note formats. We introduce ClinicNumRobBench, a benchmark of 1,624 context-question instances with ground-truth answers that evaluates four main types of clinical numeracy: value retrieval, arithmetic computation, relational comparison, and aggregation. To stress-test robustness, ClinicNumRobBench presents longitudinal MIMIC-IV vital-sign records in three semantically equivalent representations, including a real-world note-style variant derived from the Open Patients dataset, and instantiates queries using 42 question templates. Experiments on 14 LLMs show that value retrieval is generally strong, with most models exceeding 85% accuracy, while relational comparison and aggregation remain challenging, with some models scoring below 15%. Fine-tuning on medical data can reduce numeracy relative to base models by over 30%, and performance drops under note-style variation indicate LLM sensitivity to format. ClinicNumRobBench offers a rigorous testbed for clinically reliable numerical reasoning. Code and data URL are available on https://github.com/MinhVuong2000/ClinicNumRobBench.

Method: Developed SHARE family of causal language models pretrained specifically on SSH texts, and created MIRROR user interface designed for reviewing SSH text inputs without generating new text, maintaining critical engagement.

Result: SHARE models perform close to general-purpose models (Phi-4) that use 100 times more tokens, as demonstrated by custom SSH Cloze benchmark. MIRROR interface successfully enables harnessing model capabilities while preserving SSH integrity.

Conclusion: The SHARE models and MIRROR interface provide a specialized solution for SSH disciplines, demonstrating that domain-specific pretraining can achieve competitive performance with far less data while maintaining disciplinary integrity through non-generative interfaces.

Abstract: This intermediate technical report introduces the SHARE family of base models and the MIRROR user interface. The SHARE models are the first causal language models fully pretrained by and for the social sciences and humanities (SSH). Their performance in modelling SSH texts is close to that of general purpose models (Phi-4) which use 100 times more tokens, as shown by our custom SSH Cloze benchmark. The MIRROR user interface is designed for reviewing text inputs from the SSH disciplines while preserving critical engagement. By prototyping a generative AI interface that does not generate any text, we propose a way to harness the capabilities of the SHARE models without compromising the integrity of SSH principles and norms.

Method: Proposed hierarchical synthesis framework to create LifeDialBench with two subsets: EgoMem (real-world egocentric videos) and LifeMem (simulated virtual community). Introduced Online Evaluation protocol adhering to temporal causality to prevent temporal leakage issues.

Result: Current sophisticated memory systems fail to outperform simple RAG-based baseline. Over-designed structures and lossy compression in current approaches negatively impact performance, highlighting need for high-fidelity context preservation.

Conclusion: LifeDialBench addresses gaps in lifelogging audio benchmarks. Online evaluation protocol is crucial for realistic assessment. Current memory systems need simpler, higher-fidelity approaches for lifelog scenarios rather than complex, lossy architectures.

Abstract: Nowadays, wearable devices can continuously lifelog ambient conversations, creating substantial opportunities for memory systems. However, existing benchmarks primarily focus on online one-on-one chatting or human-AI interactions, thus neglecting the unique demands of real-world scenarios. Given the scarcity of public lifelogging audio datasets, we propose a hierarchical synthesis framework to curate \textbf{\textsc{LifeDialBench}}, a novel benchmark comprising two complementary subsets: \textbf{EgoMem}, built on real-world egocentric videos, and \textbf{LifeMem}, constructed using simulated virtual community. Crucially, to address the issue of temporal leakage in traditional offline settings, we propose an \textbf{Online Evaluation} protocol that strictly adheres to temporal causality, ensuring systems are evaluated in a realistic streaming fashion. Our experimental results reveal a counterintuitive finding: current sophisticated memory systems fail to outperform a simple RAG-based baseline. This highlights the detrimental impact of over-designed structures and lossy compression in current approaches, emphasizing the necessity of high-fidelity context preservation for lifelog scenarios. We release our code and data at https://github.com/qys77714/LifeDialBench.

Method: Proposes hierarchical synthesis framework treating data synthesis as unsupervised optimization over constraint graphs. Uses Legislator-Executor paradigm: Legislator evolves structured generation blueprints encoding constraints, while Executor instantiates specifications into diverse natural language scenarios.

Result: Models fine-tuned on 1K synthesized samples outperform widely-used datasets (LIMO, s1K) across eight mathematical benchmarks, showing superior out-of-distribution generalization. Tested on 10 models across Qwen, Llama, Mistral, and Gemma series.

Conclusion: The decoupling of skeleton design from linguistic realization enables focus on constructing complex logical structures, leading to high-quality data synthesis that significantly improves mathematical reasoning capabilities.

Abstract: Synthesizing high-quality mathematical reasoning data without human priors remains a significant challenge. Current approaches typically rely on seed data mutation or simple prompt engineering, often suffering from mode collapse and limited logical complexity. This paper proposes a hierarchical synthesis framework that formulates data synthesis as an unsupervised optimization problem over a constraint graph followed by semantic instantiation, rather than treating it as a direct text generation task. We introduce a Legislator-Executor paradigm: The Legislator adversarially evolves structured generation blueprints encoding the constraints of the problem, while the Executor instantiates these specifications into diverse natural language scenarios. This decoupling of skeleton design from linguistic realization enables a prioritized focus on constructing complex and diverse logical structures, thereby guiding high-quality data synthesis. Experiments conducted on a total of 10 models across the Qwen, Llama, Mistral, and Gemma series demonstrate that our method achieves notable results: models fine-tuned on 1K synthesized samples outperform widely-used datasets of comparable scale (LIMO, s1K) across eight mathematical benchmarks, exhibiting superior out-of-distribution generalization.

Method: TRACE dynamically translates reasoning paths into natural language narratives for semantic continuity, abstracts prior trajectories into reusable experiential priors, and uses dual-feedback re-ranking to integrate contextual narratives with exploration priors for relation selection.

Result: Extensive experiments on multiple KGQA benchmarks demonstrate that TRACE consistently outperforms state-of-the-art baselines.

Conclusion: TRACE effectively enhances multi-hop KGQA by maintaining reasoning coherence through contextual narratives and leveraging exploration experience through reusable priors.

Abstract: Multi-hop Knowledge Graph Question Answering (KGQA) requires coherent reasoning across relational paths, yet existing methods often treat each reasoning step independently and fail to effectively leverage experience from prior explorations, leading to fragmented reasoning and redundant exploration. To address these challenges, we propose Trajectoryaware Reasoning with Adaptive Context and Exploration priors (TRACE), an experiential framework that unifies LLM-driven contextual reasoning with exploration prior integration to enhance the coherence and robustness of multihop KGQA. Specifically, TRACE dynamically translates evolving reasoning paths into natural language narratives to maintain semantic continuity, while abstracting prior exploration trajectories into reusable experiential priors that capture recurring exploration patterns. A dualfeedback re-ranking mechanism further integrates contextual narratives with exploration priors to guide relation selection during reasoning. Extensive experiments on multiple KGQA benchmarks demonstrate that TRACE consistently outperforms state-of-the-art baselines.

Method: Introduces CocoaBench with human-designed, long-horizon tasks requiring flexible composition of vision, search, and coding, specified only by instruction and automatic evaluation function. Also presents CocoaAgent as a lightweight shared scaffold for controlled comparison.

Result: Current agents perform poorly on CocoaBench with best system achieving only 45.1% success rate, showing substantial room for improvement in reasoning/planning, tool use/execution, and visual grounding.

Conclusion: The benchmark reveals significant gaps in current unified digital agents’ ability to combine multiple capabilities, highlighting needs for better reasoning, planning, tool use, and visual grounding.

Abstract: LLM agents now perform strongly in software engineering, deep research, GUI automation, and various other applications, while recent agent scaffolds and models are increasingly integrating these capabilities into unified systems. Yet, most evaluations still test these capabilities in isolation, which leaves a gap for more diverse use cases that require agents to combine different capabilities. We introduce CocoaBench, a benchmark for unified digital agents built from human-designed, long-horizon tasks that require flexible composition of vision, search, and coding. Tasks are specified only by an instruction and an automatic evaluation function over the final output, enabling reliable and scalable evaluation across diverse agent infrastructures. We also present CocoaAgent, a lightweight shared scaffold for controlled comparison across model backbones. Experiments show that current agents remain far from reliable on CocoaBench, with the best evaluated system achieving only 45.1% success rate. Our analysis further points to substantial room for improvement in reasoning and planning, tool use and execution, and visual grounding.

Method: Introduces ConflictQA benchmark that systematically creates conflicts between textual evidence and KG evidence, evaluates LLMs on this benchmark, and proposes XoT framework - a two-stage explanation-based thinking approach for reasoning over heterogeneous conflicting evidence.

Result: LLMs struggle with cross-source conflicts, become sensitive to prompting choices, tend to rely exclusively on either KG or textual evidence, leading to incorrect responses. XoT framework shows effectiveness in handling such conflicts.

Conclusion: Cross-source knowledge conflicts present significant challenges for LLMs in RAG systems, requiring specialized approaches like XoT for faithful reasoning over heterogeneous evidence.

Abstract: Large language models (LLMs) have achieved remarkable success across a wide range of applications especially when augmented by external knowledge through retrieval-augmented generation (RAG). Despite their widespread adoption, recent studies have shown that LLMs often struggle to perform faithful reasoning when conflicting knowledge is retrieved. However, existing work primarily focuses on conflicts between external knowledge and the parametric knowledge of LLMs, leaving conflicts across external knowledge largely unexplored. Meanwhile, modern RAG systems increasingly emphasize the integration of unstructured text and (semi-)structured data like knowledge graphs (KGs) to improve knowledge completeness and reasoning faithfulness. To address this gap, we introduce ConflictQA, a novel benchmark that systematically instantiates conflicts between textual evidence and KG evidence. Extensive evaluations across representative LLMs reveal that, facing such cross-source conflicts, LLMs often fail to identify reliable evidence for correct reasoning. Instead, LLMs become more sensitive to prompting choices and tend to rely exclusively on either KG or textual evidence, resulting in incorrect responses. Based on these findings, we further propose XoT, a two-stage explanation-based thinking framework tailored for reasoning over heterogeneous conflicting evidence, and verify its effectiveness with extensive experiments.

Method: Proposes HiEdit, a hierarchical reinforcement learning framework that adaptively identifies the most knowledge-relevant layers for each editing instance. It enables dynamic, instance-aware layer selection and incorporates an intrinsic reward for sparsity to achieve precise, localized updates.

Result: Experiments on various LLMs show HiEdit boosts the performance of competitive RLEdit by an average of 8.48% while perturbing only half of the layers per edit.

Conclusion: HiEdit provides an effective framework for lifelong model editing by enabling dynamic layer selection, improving editing precision while minimizing side effects and catastrophic forgetting.

Abstract: Lifelong model editing (LME) aims to sequentially rectify outdated or inaccurate knowledge in deployed LLMs while minimizing side effects on unrelated inputs. However, existing approaches typically apply parameter perturbations to a static and dense set of LLM layers for all editing instances. This practice is counter-intuitive, as we hypothesize that different pieces of knowledge are stored in distinct layers of the model. Neglecting this layer-wise specificity can impede adaptability in integrating new knowledge and result in catastrophic forgetting for both general and previously edited knowledge. To address this, we propose HiEdit, a hierarchical reinforcement learning framework that adaptively identifies the most knowledge-relevant layers for each editing instance. By enabling dynamic, instance-aware layer selection and incorporating an intrinsic reward for sparsity, HiEdit achieves precise, localized updates. Experiments on various LLMs show that HiEdit boosts the performance of the competitive RLEdit by an average of 8.48% with perturbing only half of the layers per edit. Our code is available at: https://github.com/yangfanww/hiedit.

Method: RUMLEM is based on comprehensive, community-driven morphological databases for Romansh. It covers five main varieties plus the supra-regional standard Rumantsch Grischun. The system uses these databases to map inflected word forms to dictionary forms and identify language varieties.

Result: RUMLEM covers 77-84% of words in typical Romansh texts. Evaluation on 30,000 texts shows 95% accuracy in variety identification. A proof of concept demonstrates feasibility of Romansh vs. non-Romansh language classification.

Conclusion: RUMLEM provides effective lemmatization for Romansh varieties and demonstrates additional utility in language variety classification, addressing needs for minority language NLP tools.

Abstract: Lemmatization – the task of mapping an inflected word form to its dictionary form – is a crucial component of many NLP applications. In this paper, we present RUMLEM, a lemmatizer that covers the five main varieties of Romansh as well as the supra-regional standard variety Rumantsch Grischun. It is based on comprehensive, community-driven morphological databases for Romansh, enabling RUMLEM to cover 77-84% of the words in a typical Romansh text. Since there is a dedicated database for each Romansh variety, an additional application of RUMLEM is variety-aware language classification. Evaluation on 30'000 Romansh texts of varying lengths shows that RUMLEM correctly identifies the variety in 95% of cases. In addition, a proof of concept demonstrates the feasibility of Romansh vs. non-Romansh language classification based on the lemmatizer.

Method: Proposes Weighted Importance Multi-Point Evaluation (WIMPE) framework that factorizes each reference answer into weighted context-bound scoring points. Uses two metrics: Weighted Point-wise Alignment (WPA) to measure alignment and Point-wise Conflict Penalty (PCP) to measure contradiction between model responses and reference answers.

Result: Extensive experiments on 10 generative tasks demonstrate that WIMPE achieves higher correlations with human annotations compared to existing evaluation methods.

Conclusion: WIMPE provides a more effective framework for evaluating long-form generative model responses by addressing limitations of current methods through weighted context-bound scoring points and complementary alignment/contradiction metrics.

Abstract: Evaluating the quality of model responses remains challenging in generative tasks with long-form answers, as the expected answers usually contain multiple semantically distinct yet complementary factors that should be factorized for fine-grained assessment. Recent evaluation methods resort to relying on either task-level rubrics or question-aware checklists. However, they still 1) struggle to assess whether a response is genuinely grounded in provided contexts; 2) fail to capture the heterogeneous importance of different aspects of reference answers. Inspired by human examiners, we propose a Weighted Importance Multi-Point Evaluation (WIMPE) framework, which factorizes each reference answer into weighted context-bound scoring points. Two complementary metrics, namely Weighted Point-wise Alignment (WPA) and Point-wise Conflict Penalty (PCP), are designed to measure the alignment and contradiction between model responses and reference answers. Extensive experiments on 10 generative tasks demonstrate that WIMPE achieves higher correlations with human annotations.

Method: Proposes Dialectic-Med with three role-specialized agents: Proponent formulates diagnostic hypotheses, Opponent uses visual falsification module to retrieve contradictory evidence, and Mediator resolves conflicts via weighted consensus graph. Enforces diagnostic rigor through adversarial dialectics.

Result: Achieves state-of-the-art performance on MIMIC-CXR-VQA, VQA-RAD, and PathVQA. Fundamentally enhances trustworthiness of reasoning process, improves explanation faithfulness, and decisively mitigates hallucinations compared to single-agent baselines.

Conclusion: Dialectic-Med establishes a new standard for trustworthy medical MLLMs by explicitly modeling cognitive falsification processes, ensuring diagnostic reasoning is tightly grounded in verified visual regions through adversarial dialectics.

Abstract: Multimodal Large Language Models (MLLMs) in healthcare suffer from severe confirmation bias, often hallucinating visual details to support initial, potentially erroneous diagnostic hypotheses. Existing Chain-of-Thought (CoT) approaches lack intrinsic correction mechanisms, rendering them vulnerable to error propagation. To bridge this gap, we propose Dialectic-Med, a multi-agent framework that enforces diagnostic rigor through adversarial dialectics. Unlike static consensus models, Dialectic-Med orchestrates a dynamic interplay between three role-specialized agents: a proponent that formulates diagnostic hypotheses; an opponent equipped with a novel visual falsification module that actively retrieves contradictory visual evidence to challenge the Proponent; and a mediator that resolves conflicts via a weighted consensus graph. By explicitly modeling the cognitive process of falsification, our framework guarantees that diagnostic reasoning is tightly grounded in verified visual regions. Empirical evaluations on MIMIC-CXR-VQA, VQA-RAD, and PathVQA demonstrate that Dialectic-Med not only achieves state-of-the-art performance but also fundamentally enhances the trustworthiness of the reasoning process. Beyond accuracy, our approach significantly enhances explanation faithfulness and decisively mitigates hallucinations, establishing a new standard over single-agent baselines.

Method: Transactional Attention (TA) uses a sponsorship mechanism where structural anchor patterns (e.g., “key:”, “password:”) protect adjacent value-bearing tokens from eviction. TA-Fast is an attention-free variant that reduces memory overhead and is compatible with SDPA and FlashAttention.

Result: TA achieves 100% credential retrieval at K=16 tokens where six baselines achieve 0%, and sustains 100% accuracy across 200 function-calling trials. TA-Fast reduces memory overhead by 52% with less than 1% latency overhead.

Conclusion: Transactional Attention solves the dormant token problem in KV-cache compression through structural sponsorship, is orthogonal to existing methods, and maintains high accuracy with minimal overhead.

Abstract: At K=16 tokens (0.4% of a 4K context), every existing KV-cache compression method achieves 0% on credential retrieval. The failure mode is dormant tokens: credentials, API keys, and configuration values that receive near-zero attention but become essential at generation time. Because these tokens lack the statistical signals that eviction policies rely on, no method based on attention scores, reconstruction loss, or learned retention gates retains them. We introduce Transactional Attention (TA), a sponsorship mechanism in which structural anchor patterns (e.g., “key:”, “password:”) protect adjacent value-bearing tokens from eviction. TA achieves 100% credential retrieval at K=16 where six baselines (H2O, TOVA, SnapKV, StreamingLLM, PyramidKV, DynamicKV) achieve 0%, and sustains 100% accuracy across 200 function-calling trials. TA-Fast, an attention-free variant, reduces memory overhead by 52% and is compatible with SDPA and FlashAttention. TA is orthogonal to existing compression methods and adds less than 1% latency overhead.

Method: Evaluated 10 language models across 6 typologically diverse languages, generated over 1.4M SFT examples, trained 240 student models, and introduced Polyglot Score metric to measure teacher effectiveness by correlating intrinsic data quality with extrinsic student performance.

Result: Gemma 3 27B and Aya Expanse 32B emerged as consistently effective teachers; model scale alone doesn’t significantly predict teacher effectiveness; data quality metrics (prompt diversity, length, response fluency) capture over 93.3% of variance in intrinsic data quality and predict student performance.

Conclusion: Practical recommendations include matching teacher-student model families and using translation techniques for less-resourced languages; the work advances data-centric research in multilingual synthetic data and LM development.

Abstract: Synthesizing supervised finetuning (SFT) data from language models (LMs) to teach smaller models multilingual tasks has become increasingly common. However, teacher model selection is often ad hoc, typically defaulting to the largest available option, even though such models may have significant capability gaps in non-English languages. This practice can result in poor-quality synthetic data and suboptimal student downstream performance. In this work, we systematically characterize what makes an effective multilingual teacher. We measure intrinsic measures of data quality with extrinsic student model performance in a metric we call Polyglot Score; evaluating 10 LMs across 6 typologically diverse languages, generating over 1.4M SFT examples and training 240 student models. Among the models tested, Gemma 3 27B and Aya Expanse 32B emerge as consistently effective teachers across different student base model families. Further analyses reveal that model scale alone does not significantly predict teacher effectiveness; instead, data qualities such as prompt diversity, length, and response fluency capture over 93.3% of variance in intrinsic data quality and predict student performance. Finally, we provide practical recommendations, including matching the model families of teacher-student pairs and translating from or responding to existing prompts, which can yield improvements for less-resourced languages. We hope that our work advances data-centric research in multilingual synthetic data and LM development.

Method: Constructed a comprehensive benchmark with 11 tasks and over 130,000 instances to evaluate MLLM capabilities in ancient Chinese script evolution analysis. Proposed GEVO, a glyph-driven fine-tuning framework that explicitly encourages models to capture evolutionary consistency in glyph transformations.

Result: Existing MLLMs demonstrate limited ability in glyph-level comparison and substantially constrained performance on core tasks like character recognition and evolutionary reasoning. GEVO enables even 2B-scale models to achieve consistent and comprehensive performance improvements across all evaluated tasks.

Conclusion: The paper addresses a significant gap in applying MLLMs to ancient script analysis, provides a valuable benchmark, and demonstrates that targeted fine-tuning approaches like GEVO can significantly enhance MLLM capabilities for text evolution understanding tasks.

Abstract: In recent years, rapid advances in Multimodal Large Language Models (MLLMs) have increasingly stimulated research on ancient Chinese scripts. As the evolution of written characters constitutes a fundamental pathway for understanding cultural transformation and historical continuity, how MLLMs can be systematically leveraged to support and advance text evolution analysis remains an open and largely underexplored problem. To bridge this gap, we construct a comprehensive benchmark comprising 11 tasks and over 130,000 instances, specifically designed to evaluate the capability of MLLMs in analyzing the evolution of ancient Chinese scripts. We conduct extensive evaluations across multiple widely used MLLMs and observe that, while existing models demonstrate a limited ability in glyph-level comparison, their performance on core tasks-such as character recognition and evolutionary reasoning-remains substantially constrained. Motivated by these findings, we propose a glyph-driven fine-tuning framework (GEVO) that explicitly encourages models to capture evolutionary consistency in glyph transformations and enhances their understanding of text evolution. Experimental results show that even models at the 2B scale achieve consistent and comprehensive performance improvements across all evaluated tasks. To facilitate future research, we publicly release both the benchmark and the trained models\footnote{https://github.com/songruiecho/GEVO}.

Method: Introduces SABEval dataset to decouple structural alignment from semantic relevance, uses Contrastive Attention Attribution to analyze internal mechanisms, and proposes a rebalancing strategy to mitigate the bias.

Result: Structural alignment bias causes severe tool-invocation errors in LLMs, with two competing pathways (semantic checking vs structural matching) driving decisions. The rebalancing strategy effectively reduces bias without harming general tool-use capabilities.

Conclusion: Structural alignment bias is a widespread mechanistic flaw in LLM tool refusal that requires systematic evaluation and mitigation strategies to improve reliable tool use.

Abstract: Large language models (LLMs) have demonstrated impressive capabilities in utilizing external tools. In practice, however, LLMs are often exposed to tools that are irrelevant to the user’s query, in which case the desired behavior is to refrain from invocations. In this work, we identify a widespread yet overlooked mechanistic flaw in tool refusal, which we term structural alignment bias: Even when a tool fails to serve the user’s goal, LLMs still tend to invoke it whenever query attributes can be validly assigned to tool parameters. To systematically study this bias, we introduce SABEval, a new dataset that decouples structural alignment from semantic relevance. Our analysis shows that structural alignment bias induces severe tool-invocation errors in LLMs, yet remains largely unaccounted for in existing evaluations. To investigate the internal mechanisms underlying this bias, we propose Contrastive Attention Attribution, which reveals two competing pathways for semantic checking and structural matching. The relative strength of these pathways drives LLMs’ tool invocation decisions. Based on these findings, we further introduce a rebalancing strategy that effectively mitigates structural alignment bias, as demonstrated by extensive experiments, without degrading general tool-use capabilities.

Method: Proposes GRIP framework where the model regulates retrieval behavior through control-token emission. Uses Self-Triggered Information Planning to decide when to retrieve, how to reformulate queries, and when to terminate within a single autoregressive trajectory.

Result: Experiments on five QA benchmarks show GRIP surpasses strong RAG baselines and is competitive with GPT-4o while using substantially fewer parameters.

Conclusion: GRIP demonstrates that embedding retrieval control directly into generation enables more effective coordination between retrieval and reasoning, supporting dynamic multi-step inference with on-the-fly evidence integration.

Abstract: We revisit retrieval-augmented generation (RAG) by embedding retrieval control directly into generation. Instead of treating retrieval as an external intervention, we express retrieval decisions within token-level decoding, enabling end-to-end coordination without additional controllers or classifiers. Under the paradigm of Retrieval as Generation, we propose \textbf{GRIP} (\textbf{G}eneration-guided \textbf{R}etrieval with \textbf{I}nformation \textbf{P}lanning), a unified framework in which the model regulates retrieval behavior through control-token emission. Central to GRIP is \textit{Self-Triggered Information Planning}, which allows the model to decide when to retrieve, how to reformulate queries, and when to terminate, all within a single autoregressive trajectory. This design tightly couples retrieval and reasoning and supports dynamic multi-step inference with on-the-fly evidence integration. To supervise these behaviors, we construct a structured training set covering answerable, partially answerable, and multi-hop queries, each aligned with specific token patterns. Experiments on five QA benchmarks show that GRIP surpasses strong RAG baselines and is competitive with GPT-4o while using substantially fewer parameters.

Method: Two core approaches: (1) Intent-Aware Bridging using Variational Information Bottleneck to translate internal semantics into smooth expressive intent, and (2) Realization-Aware Alignment where the model acts as its own critic to verify acoustic outputs match intended expression via rubric-based feedback.

Result: Trained on only 800 hours of expressive speech data, the 3B parameter SA-SLM surpasses all open-source baselines and comes within 0.08 points of GPT-4o-Audio in overall expressiveness on the EchoMind benchmark.

Conclusion: Making speech models self-aware of their intent during generation and realization during training effectively bridges the semantic-acoustic gap, enabling more expressive speech generation with limited data.

Abstract: Speech Language Models (SLMs) exhibit strong semantic understanding, yet their generated speech often sounds flat and fails to convey expressive intent, undermining user engagement. We term this mismatch the semantic understanding-acoustic realization gap. We attribute this gap to two key deficiencies: (1) intent transmission failure, where SLMs fail to provide the stable utterance-level intent needed for expressive delivery; and (2) realization-unaware training, where no feedback signal verifies whether acoustic outputs faithfully reflect intended expression. To address these issues, we propose SA-SLM (Self-Aware Speech Language Model), built on the principle that the model should be aware of what it thinks during generation and how it speaks during training. SA-SLM addresses this gap through two core contributions: (1) Intent-Aware Bridging, which uses a Variational Information Bottleneck (VIB) objective to translate the model’s internal semantics into temporally smooth expressive intent, making speech generation aware of what the model intends to express; and (2) Realization-Aware Alignment, which repurposes the model as its own critic to verify and align acoustic realization with intended expressive intent via rubric-based feedback. Trained on only 800 hours of expressive speech data, our 3B parameter SA-SLM surpasses all open-source baselines and comes within 0.08 points of GPT-4o-Audio in overall expressiveness on the EchoMind benchmark.

Method: Leverages large language models to autonomously induce both strategy actions and planning logic from raw transcripts. Formalizes expert knowledge into a Strategy Forest - hierarchical structure capturing short-term responses (nodes) and long-term strategic foresight (branches).

Result: Outperforms existing methods by average 9%-10% across two benchmarks. Shows strategic behavioral diversity, foresight, and robust cross-task transferability.

Conclusion: Offers cost-effective, scalable way to build non-collaborative agents. Provides new insights into automated strategy learning for dialogue systems.

Abstract: Developing non-collaborative dialogue agents traditionally requires the manual, unscalable codification of expert strategies. We propose \ours, a method that leverages large language models to autonomously induce both strategy actions and planning logic directly from raw transcripts. METRO formalizes expert knowledge into a Strategy Forest, a hierarchical structure that captures both short-term responses (nodes) and long-term strategic foresight (branches). Experimental results across two benchmarks show that METRO demonstrates promising performance, outperforming existing methods by an average of 9%-10%. Our further analysis not only reveals the success behind METRO (strategic behavioral diversity and foresight), but also demonstrates its robust cross-task transferability. This offers new insights into building non-collaborative agents in a cost-effective and scalable way. Our code is available at https://github.com/Humphrey-0125/METRO.

Method: Proposes a training framework with two components: 1) a reasoning model that produces structured QA reasoning traces, and 2) a generation model that conditions on these traces to produce final character descriptions. Can be applied on top of long-context LLMs or chunk-based methods.

Result: Experiments on BookWorm and CroSS datasets show that QA-guided reasoning improves faithfulness, informativeness, and grounding over strong long-context baselines.

Conclusion: Decoupling reasoning from generation via structured QA traces is effective for character description generation from long narratives, addressing the limitations of end-to-end LLM approaches.

Abstract: Character description generation is an important capability for narrative-focused applications such as summarization, story analysis, and character-driven simulations. However, generating accurate character descriptions from long-form narratives (e.g., novels) is challenging: models must track evolving attributes (e.g., relationships and events), integrate evidence scattered across the text, and infer implicit details. Despite the success of reasoning-enabled LLMs on many benchmarks, we find that for character description generation their performance improves when built-in reasoning is disabled (i.e., an empty reasoning trace). Motivated by this, we propose a training framework that decouples reasoning from generation. Our approach, which can be applied on top of long-context LLMs or chunk-based methods, consists of a reasoning model that produces a structured QA reasoning trace and a generation model that conditions on this trace to produce the final character description. Experiments on two datasets (BookWorm and CroSS) show that QA-guided reasoning improves faithfulness, informativeness, and grounding over strong long-context baselines.

Method: Developed METER benchmark to systematically evaluate LLMs across all three levels of the causal ladder (association, intervention, counterfactual) under unified context settings. Conducted extensive evaluation of various LLMs followed by mechanistic analysis through error pattern identification and internal information flow tracing.

Result: LLMs show significant performance decline as tasks ascend the causal hierarchy. Two primary failure modes identified: (1) distraction by causally irrelevant but factually correct information at lower causal levels, and (2) reduced faithfulness to provided context at higher causal levels leading to performance degradation.

Conclusion: The work advances understanding of LLM contextual causal reasoning mechanisms and establishes a foundation for future research. The benchmark reveals systematic limitations in LLMs’ causal reasoning capabilities that need to be addressed.

Abstract: Contextual causal reasoning is a critical yet challenging capability for Large Language Models (LLMs). Existing benchmarks, however, often evaluate this skill in fragmented settings, failing to ensure context consistency or cover the full causal hierarchy. To address this, we pioneer METER to systematically benchmark LLMs across all three levels of the causal ladder under a unified context setting. Our extensive evaluation of various LLMs reveals a significant decline in proficiency as tasks ascend the causal hierarchy. To diagnose this degradation, we conduct a deep mechanistic analysis via both error pattern identification and internal information flow tracing. Our analysis reveals two primary failure modes: (1) LLMs are susceptible to distraction by causally irrelevant but factually correct information at lower level of causality; and (2) as tasks ascend the causal hierarchy, faithfulness to the provided context degrades, leading to a reduced performance. We belive our work advances our understanding of the mechanisms behind LLM contextual causal reasoning and establishes a critical foundation for future research. Our code and dataset are available at https://github.com/SCUNLP/METER .

Method: Proposes Policy Split paradigm that bifurcates policy into normal and high-entropy modes using a high-entropy prompt. Both modes share model parameters but undergo collaborative dual-mode entropy regularization: normal mode optimizes for task correctness, high-entropy mode incorporates exploration preference, and they learn collaboratively.

Result: Extensive experiments show consistent outperformance over established entropy-guided RL baselines across various model sizes in general and creative tasks. Analysis reveals Policy Split facilitates dual-mode exploration where high-entropy mode generates distinct behavioral patterns providing unique learning signals.

Conclusion: Policy Split effectively balances exploration and accuracy in RL for LLMs through collaborative dual-mode learning, enabling diverse exploration without compromising task performance.

Abstract: To encourage diverse exploration in reinforcement learning (RL) for large language models (LLMs) without compromising accuracy, we propose Policy Split, a novel paradigm that bifurcates the policy into normal and high-entropy modes with a high-entropy prompt. While sharing model parameters, the two modes undergo collaborative dual-mode entropy regularization tailored to distinct objectives. Specifically, the normal mode optimizes for task correctness, while the high-entropy mode incorporates a preference for exploration, and the two modes learn collaboratively. Extensive experiments demonstrate that our approach consistently outperforms established entropy-guided RL baselines across various model sizes in general and creative tasks. Further analysis reveals that Policy Split facilitates dual-mode exploration, where the high-entropy mode generates distinct behavioral patterns to the normal mode, providing unique learning signals.

Method: Proposes TCER (Triviality Corrected Endogenous Reward) that rewards relative information gain between a specialist policy and generalist reference policy, modulated by probability-dependent correction to address triviality bias.

Result: TCER achieves consistent improvements across multiple writing benchmarks and model architectures without external supervision, and also transfers effectively to mathematical reasoning tasks.

Conclusion: TCER successfully addresses triviality bias in unsupervised RL for open-ended generation, demonstrating generality across different generation tasks including writing and mathematical reasoning.

Abstract: Reinforcement learning for open-ended text generation is constrained by the lack of verifiable rewards, necessitating reliance on judge models that require either annotated data or powerful closed-source models. Inspired by recent work on unsupervised reinforcement learning for mathematical reasoning using confidence-based endogenous rewards, we investigate whether this principle can be adapted to open-ended writing tasks. We find that directly applying confidence rewards leads to Triviality Bias: the policy collapses toward high-probability outputs, reducing diversity and meaningful content. We propose TCER (Triviality Corrected Endogenous Reward), which addresses this bias by rewarding the relative information gain between a specialist policy and a generalist reference policy, modulated by a probability-dependent correction mechanism. Across multiple writing benchmarks and model architectures, TCER achieves consistent improvements without external supervision. Furthermore, TCER also transfers effectively to mathematical reasoning, validating the generality of our approach across different generation tasks.

Method: Created NovBench with 1,684 paper-review pairs from a leading NLP conference, extracting novelty descriptions from introductions and expert-written novelty evaluations. Proposed a four-dimensional evaluation framework (Relevance, Correctness, Coverage, Clarity) to assess LLM-generated novelty evaluations.

Result: Extensive experiments show current models exhibit limited understanding of scientific novelty, and fine-tuned models often suffer from instruction-following deficiencies.

Conclusion: There’s a need for targeted fine-tuning strategies that jointly improve novelty comprehension and instruction adherence in LLMs for peer review applications.

Abstract: Novelty is a core requirement in academic publishing and a central focus of peer review, yet the growing volume of submissions has placed increasing pressure on human reviewers. While large language models (LLMs), including those fine-tuned on peer review data, have shown promise in generating review comments, the absence of a dedicated benchmark has limited systematic evaluation of their ability to assess research novelty. To address this gap, we introduce NovBench, the first large-scale benchmark designed to evaluate LLMs’ capability to generate novelty evaluations in support of human peer review. NovBench comprises 1,684 paper-review pairs from a leading NLP conference, including novelty descriptions extracted from paper introductions and corresponding expert-written novelty evaluations. We focus on both sources because the introduction provides a standardized and explicit articulation of novelty claims, while expert-written novelty evaluations constitute one of the current gold standards of human judgment. Furthermore, we propose a four-dimensional evaluation framework (including Relevance, Correctness, Coverage, and Clarity) to assess the quality of LLM-generated novelty evaluations. Extensive experiments on both general and specialized LLMs under different prompting strategies reveal that current models exhibit limited understanding of scientific novelty, and that fine–tuned models often suffer from instruction-following deficiencies. These findings underscore the need for targeted fine-tuning strategies that jointly improve novelty comprehension and instruction adherence.

Method: RoMem uses a pretrained Semantic Speed Gate that maps relation text embeddings to volatility scores, learning which relations evolve fast (e.g., “president of”) vs remain stable (e.g., “born in”). Combined with continuous phase rotation, it enables geometric shadowing where obsolete facts are rotated out of phase in complex vector space.

Result: Achieves state-of-the-art 72.6 MRR on ICEWS05-15 temporal KG completion. For agentic memory: 2-3x MRR and answer accuracy on MultiTQ, dominates LoCoMo hybrid benchmark, preserves static memory with zero degradation on DMR-MSC, and generalizes zero-shot to financial domains (FinTMMBench).

Conclusion: RoMem provides an effective drop-in temporal knowledge graph module that enables structured memory systems to handle temporal dynamics by distinguishing persistent vs evolving facts through semantic volatility learning and geometric shadowing.

Abstract: Structured memory representations such as knowledge graphs are central to autonomous agents and other long-lived systems. However, most existing approaches model time as discrete metadata, either sorting by recency (burying old-yet-permanent knowledge), simply overwriting outdated facts, or requiring an expensive LLM call at every ingestion step, leaving them unable to distinguish persistent facts from evolving ones. To address this, we introduce RoMem, a drop-in temporal knowledge graph module for structured memory systems, applicable to agentic memory and beyond. A pretrained Semantic Speed Gate maps each relation’s text embedding to a volatility score, learning from data that evolving relations (e.g., “president of”) should rotate fast while persistent ones (e.g., “born in”) should remain stable. Combined with continuous phase rotation, this enables geometric shadowing: obsolete facts are rotated out of phase in complex vector space, so temporally correct facts naturally outrank contradictions without deletion. On temporal knowledge graph completion, RoMem achieves state-of-the-art results on ICEWS05-15 (72.6 MRR). Applied to agentic memory, it delivers 2-3x MRR and answer accuracy on temporal reasoning (MultiTQ), dominates hybrid benchmark (LoCoMo), preserves static memory with zero degradation (DMR-MSC), and generalises zero-shot to unseen financial domains (FinTMMBench).

Method: Relax uses three co-designed architectural layers: 1) omni-native architecture with multimodal support built into the full stack, 2) independent fault-isolated services for each RL role that can scale independently, and 3) service-level decoupling enabling asynchronous training via TransferQueue data bus with tunable staleness parameter.

Result: Achieves 1.20× speedup over veRL on Qwen3-4B on-policy training, 1.76× speedup in fully async mode on Qwen3-4B, and 2.00× speedup on Qwen3-Omni-30B. Supports R3 for MoE models with only 1.9% overhead vs 32% degradation in veRL. Demonstrates stable omni-modal RL convergence across image, text, and audio, sustaining over 2,000 steps on video.

Conclusion: Relax provides an effective RL training engine for multimodal LLMs that addresses key scalability and efficiency challenges while maintaining convergence quality, making it suitable for complex omni-modal agentic workflows.

Abstract: Reinforcement learning (RL) post-training has proven effective at unlocking reasoning, self-reflection, and tool-use capabilities in large language models. As models extend to omni-modal inputs and agentic multi-turn workflows, RL training systems face three interdependent challenges: heterogeneous data flows, operational robustness at scale, and the staleness – throughput tradeoff. We present \textbf{Relax} (Reinforcement Engine Leveraging Agentic X-modality), an open-source RL training engine that addresses these challenges through three co-designed architectural layers. First, an \emph{omni-native architecture} builds multimodal support into the full stack – from data preprocessing and modality-aware parallelism to inference generation – rather than retrofitting it onto a text-centric pipeline. Second, each RL role runs as an independent, fault-isolated service that can be scaled, recovered, and upgraded without global coordination. Third, service-level decoupling enables asynchronous training via the TransferQueue data bus, where a single staleness parameter smoothly interpolates among on-policy, near-on-policy, and fully asynchronous execution. Relax achieves a 1.20$\times$ end-to-end speedup over veRL on Qwen3-4B on-policy training. Its fully async mode delivers a 1.76$\times$ speedup over colocate on Qwen3-4B and a 2.00$\times$ speedup on Qwen3-Omni-30B, while all modes converge to the same reward level. Relax supports R3 (Rollout Routing Replay)~\cite{ma2025r3} for MoE models with only 1.9% overhead, compared to 32% degradation in veRL under the same configuration. It further demonstrates stable omni-modal RL convergence on Qwen3-Omni across image, text, and audio, sustaining over 2{,}000 steps on video without degradation. Relax is available at https://github.com/rednote-ai/Relax.

Method: A structured persona memory system that extracts what is known about a person rather than retrieving what was said. Uses a full persona extraction pipeline that decomposes conversations into six cognitive domains (biography, experiences, preferences, social circle, work, psychometrics), consolidates and deduplicates per domain, and retrieves structured facts via CategoryRAG with 21.79 ms latency.

Result: Achieves 94.37% accuracy on LoCoMo benchmark (exceeding all published systems including MemMachine’s 91.69% and human performance of 87.9 F1). Core memory fact accuracy reaches 98.64%. Adversarial robustness reaches 99.55%. Reduces token consumption by ~5x compared to full-context replay while achieving higher accuracy.

Conclusion: Synthius-Mem achieves state-of-the-art results on LoCoMo and is the only persona memory system that both exceeds human-level performance and reports adversarial robustness, providing reliable long-term memory that doesn’t hallucinate.

Abstract: Providing AI agents with reliable long-term memory that does not hallucinate remains an open problem. Current approaches to memory for LLM agents – sliding windows, summarization, embedding-based RAG, and flat fact extraction – each reduce token cost but introduce catastrophic information loss, semantic drift, or uncontrolled hallucination about the user. The structural reason is architectural: every published memory system on the LoCoMo benchmark treats conversation as a retrieval problem over raw or lightly summarized dialogue segments, and none reports adversarial robustness, the ability to refuse questions about facts the user never disclosed. We present Synthius-Mem, a brain-inspired structured persona memory system that takes a fundamentally different approach. Instead of retrieving what was said, Synthius-Mem extracts what is known about the person: a full persona extraction pipeline decomposes conversations into six cognitive domains (biography, experiences, preferences, social circle, work, psychometrics), consolidates and deduplicates per domain, and retrieves structured facts via CategoryRAG at 21.79 ms latency. On the LoCoMo benchmark (ACL 2024, 10 conversations, 1,813 questions), Synthius-Mem achieves 94.37% accuracy, exceeding all published systems including MemMachine (91.69%, adversarial score is not reported) and human performance (87.9 F1). Core memory fact accuracy reaches 98.64%. Adversarial robustness, the hallucination resistance metric that no competing system reports, reaches 99.55%. Synthius-Mem reduces token consumption by ~5x compared to full-context replay while achieving higher accuracy. Synthius-Mem achieves state-of-the-art results on LoCoMo and is, to our knowledge, the only persona memory system that both exceeds human-level performance and reports adversarial robustness.

Method: Information-theoretic framework modeling phoneme sequences as second-order Markov chains, using articulatory features to compute phonological distances across 67 languages from parallel corpus.

Result: Phonological distance matrix recovers major language families, shows contact-induced convergence, correlates with geographic distance, and supports Steppe hypothesis for Indo-European homeland.

Conclusion: Short-range phoneme dependencies effectively capture linguistic relatedness patterns, providing quantitative tools for typological analysis and evolutionary linguistics research.

Abstract: We show that short-range phoneme dependencies encode large-scale patterns of linguistic relatedness, with direct implications for quantitative typology and evolutionary linguistics. Specifically, using an information-theoretic framework, we argue that phoneme sequences modeled as second-order Markov chains essentially capture the statistical correlations of a phonological system. This finding enables us to quantify distances among 67 modern languages from a multilingual parallel corpus employing a distance metric that incorporates articulatory features of phonemes. The resulting phonological distance matrix recovers major language families and reveals signatures of contact-induced convergence. Remarkably, we obtain a clear correlation with geographic distance, allowing us to constrain a plausible homeland region for the Indo-European family, consistent with the Steppe hypothesis.

Method: Developed MIXAR, a generative pixel-based language model trained on eight different languages using various scripts, and scaled it to 0.5B parameters for enhanced capabilities.

Result: MIXAR demonstrates substantial performance improvements on discriminative and generative multilingual tasks, shows robustness to unseen languages, and exhibits enhanced capabilities when scaled to 0.5B parameters, including better performance on LAMBADA and resilience to orthographic attacks.

Conclusion: Pixel-based language models like MIXAR can effectively handle multilingual challenges and offer promising alternatives to token-based approaches, with scaling further enhancing their capabilities and robustness.

Abstract: Pixel-based language models are gaining momentum as alternatives to traditional token-based approaches, promising to circumvent tokenization challenges. However, the inherent perceptual diversity across languages poses a significant hurdle for multilingual generalization in pixel space. This paper introduces MIXAR, the first generative pixel-based language model trained on eight different languages utilizing a range of different scripts. We empirically evaluate MIXAR against previous pixel-based models as well as comparable tokenizer-based models, demonstrating substantial performance improvement on discriminative and generative multilingual tasks. Additionally, we show how MIXAR is robust to languages never seen during the training. These results are further strengthened when scaling the model to 0.5B parameters which not only improves its capabilities in generative tasks like LAMBADA but also its robustness when challenged with input perturbations such as orthographic attacks.

Method: Decomposes LLM pipeline uncertainty into its sources, distinguishes variance that shrinks with more data from sensitivity to researcher design choices, projects efficient paths to reduce total error, and provides recommendations for benchmark builders to minimize exploitable surfaces.

Result: Projection-optimized pipelines outperform 73% of possible naive pipelines against human baselines across various tasks (ideology annotation, safety classification, MMLU benchmarking, propaganda audit). On MMLU, optimized budget allocation halves estimation error compared to standard single-prompt evaluation at equivalent cost.

Conclusion: Proper uncertainty quantification in LLM evaluations is essential for reliable benchmarking, and small-sample variance estimation can generate confidence intervals with nominal coverage while providing actionable recommendations to reduce measurement error and improve benchmark robustness.

Abstract: LLM evaluations drive which models get deployed, which safety standards get adopted, and which research conclusions get published. Yet these scores carry hidden uncertainty: rephrasing the prompt, switching the judge model, or changing the temperature can shift results enough to flip rankings and reverse conclusions. Standard confidence intervals ignore this variance, producing under-coverage that worsens with more data. The unmeasured variance also creates an exploitable surface: model developers can optimize against measurement noise rather than genuine capability. This paper decomposes LLM pipeline uncertainty into its sources, distinguishes variance that shrinks with more data from sensitivity to researcher design choices, and projects the most efficient path to reducing total error. For benchmark builders, the same decomposition identifies which design choices contribute exploitable surface for gaming and prescribes designs that minimize it. Across ideology annotation, safety classification, MMLU benchmarking, and a human-validated propaganda audit, projection-optimized pipelines outperform 73% of possible naive pipelines against a human baseline. On MMLU, optimized budget allocation halves estimation error compared to standard single-prompt evaluation at equivalent cost. A small-sample variance estimation exercise is sufficient to derive confidence intervals that approach nominal coverage when the model includes the relevant pipeline facets, and to generate recommendations for reducing measurement error and improving benchmark robustness.

Method: TST partitions digits into three-digit triads and annotates each triad with explicit magnitude markers. Two variants: (1) vocabulary-based approach adding up to 10,000 fixed tokens covering 33 orders of magnitude, and (2) suffix-marker approach using special tokens to denote magnitude dynamically.

Result: The framework preserves exact digits while making order-of-magnitude relationships transparent at the token level. It’s scalable, architecture-agnostic, and can be integrated as a drop-in preprocessing step. Experimental validation is deferred to future work.

Conclusion: TST provides a deterministic tokenization scheme that maintains numerical structure in LLMs, offering a solution to arithmetic reasoning limitations caused by standard tokenization methods.

Abstract: Standard subword tokenization methods fragment numbers inconsistently, causing large language models (LLMs) to lose positional and decimal structure - a primary driver of errors in arithmetic and scientific reasoning. We introduce Triadic Suffix Tokenization (TST), a deterministic scheme that partitions digits into three-digit triads and annotates each triad with an explicit magnitude marker. Critically, the scheme defines a fixed, one-to-one mapping between suffixes and orders of magnitude for the integer part (thousands, millions, billions, etc.) and a parallel system of replicated markers for fractional depth (tenths, thousandths, millionths, etc.). Unlike approaches that rely on positional inference, this method provides a consistent gradient signal, which should ensure stable convergence. Two implementation variants are proposed: (1) a vocabulary-based approach that adds at most 10,000 fixed tokens to an existing vocabulary, covering 33 orders of magnitude ($10^{-15}$ to $10^{18}$); and (2) a suffix-marker approach that uses a small set of special tokens to denote magnitude dynamically. Both variants preserve exact digits while making order-of-magnitude relationships transparent at the token level. The framework is inherently scalable, allowing for linear vocabulary expansion to accommodate arbitrary precision and range. TST is architecture-agnostic and can be integrated as a drop-in preprocessing step. Experimental validation is deferred to future work.

Method: Formalize heterogeneous memory extraction task and introduce BEHEMOTH benchmark using 18 existing datasets across personalization, problem-solving, and agentic tasks. Propose CluE, a cluster-based self-evolving strategy that groups training examples by extraction scenarios, analyzes each cluster independently, and synthesizes cross-cluster insights to update extraction prompts.

Result: Empirical analysis shows no single static extraction prompt dominates across all task categories, and existing self-evolving frameworks degrade with heterogeneous tasks. CluE achieves +9.04% relative gain and consistently outperforms prior self-evolving frameworks on BEHEMOTH benchmark.

Conclusion: CluE effectively addresses the challenge of heterogeneous memory extraction in LLM assistants through cluster-based self-evolution, demonstrating strong generalization across diverse task categories.

Abstract: As LLM-based assistants become persistent and personalized, they must extract and retain useful information from past conversations as memory. However, the types of information worth remembering vary considerably across tasks. We formalize the \textit{heterogeneous memory extraction} task and introduce \textbf{BEHEMOTH}, a benchmark that repurposes 18 existing datasets spanning personalization, problem-solving, and agentic tasks, using a downstream utility-driven metric for systematic evaluation. Our empirical analysis confirms that no single static extraction prompt dominates across all task categories, and that existing self-evolving prompt optimization frameworks, originally designed for homogeneous distributions, degrade when training tasks are heterogeneous. To address this, we propose \textbf{CluE}, a cluster-based self-evolving strategy that groups training examples into clusters by extraction scenarios, analyzes each cluster independently, and synthesizes cross-cluster insights to update the extraction prompt. Experiments on BEHEMOTH show that CluE generalizes effectively across heterogeneous tasks ($+$9.04% relative gain), consistently outperforming prior self-evolving frameworks.

Method: Proposes Mutual Information Self-Evaluation (MISE) which uses hindsight generative self-evaluation to create dense internal reward signals, then calibrates these rewards against environmental feedback through a theoretical framework combining mutual information with KL divergence.

Result: MISE outperforms strong baselines, enabling open-source LLMs with about 7B parameters to achieve performance comparable to GPT-4o on validation tasks without expert supervision.

Conclusion: The work provides the first formal foundation for generative self-rewarding paradigms and demonstrates that calibrated self-evaluation can effectively supplement sparse extrinsic rewards for autonomous learning.

Abstract: To overcome the sparse reward challenge in reinforcement learning (RL) for agents based on large language models (LLMs), we propose Mutual Information Self-Evaluation (MISE), an RL paradigm that utilizes hindsight generative self-evaluation as dense reward signals while simultaneously calibrating them against the environmental feedbacks. Empirically, MISE enables an agent to learn autonomously from dense internal rewards supplementing sparse extrinsic signals. Theoretically, our work provides the first formal foundation for the paradigm of generative self-rewarding. We prove that utilizing hindsight self-evaluation rewards is equivalent to minimizing an objective that combines mutual information with a KL divergence term between the policy and a proxy reward policy. This theoretical insight then informs and justifies our calibration step, which actively aligns these rewards with the optimal policy. Extensive experiments show that MISE outperforms strong baselines, enabling open-source LLMs about 7B parameters to achieve performance comparable to GPT-4o on validation without expert supervision.

Method: Proposes a minimalist framework with two key components: Turn Isolation Retrieval (TIR) that uses max-activation strategy to capture turn-level signals instead of global aggregation, and Query-Driven Pruning (QDP) that removes redundant sessions and conversational filler to create compact, high-density evidence sets.

Result: Extensive experiments on multiple benchmarks show the framework achieves robust performance across diverse settings, consistently outperforming strong baselines while maintaining high efficiency in tokens and latency.

Conclusion: Establishes a new minimalist baseline for conversational memory that effectively addresses signal sparsity and redundancy issues in long-term dialogue systems.

Abstract: Existing conversational memory systems rely on complex hierarchical summarization or reinforcement learning to manage long-term dialogue history, yet remain vulnerable to context dilution as conversations grow. In this work, we offer a different perspective: the primary bottleneck may lie not in memory architecture, but in the \textit{Signal Sparsity Effect} within the latent knowledge manifold. Through controlled experiments, we identify two key phenomena: \textit{Decisive Evidence Sparsity}, where relevant signals become increasingly isolated with longer sessions, leading to sharp degradation in aggregation-based methods; and \textit{Dual-Level Redundancy}, where both inter-session interference and intra-session conversational filler introduce large amounts of non-informative content, hindering effective generation. Motivated by these insights, we propose \method, a minimalist framework that brings conversational memory back to basics, relying solely on retrieval and generation via Turn Isolation Retrieval (TIR) and Query-Driven Pruning (QDP). TIR replaces global aggregation with a max-activation strategy to capture turn-level signals, while QDP removes redundant sessions and conversational filler to construct a compact, high-density evidence set. Extensive experiments on multiple benchmarks demonstrate that \method achieves robust performance across diverse settings, consistently outperforming strong baselines while maintaining high efficiency in tokens and latency, establishing a new minimalist baseline for conversational memory.

Method: Created CARTBENCH by aligning Palace Museum objects from Wikidata with authoritative catalog pages across five art categories and multiple dynasties. Includes four subtasks: CURATORQA (evidence-grounded reasoning), CATALOGCAPTION (structured appreciation), REINTERPRET (defensible reinterpretation with expert ratings), and CONNOISSEURPAIRS (authenticity discrimination).

Result: Evaluation of nine representative VLMs shows: high overall CURATORQA accuracy masks sharp drops on hard evidence linking and style-to-period inference; long-form appreciation remains far from expert references; authenticity discrimination stays near chance level, revealing significant gaps in connoisseur-level reasoning.

Conclusion: CARTBENCH reveals substantial limitations in current VLMs for deep art understanding, particularly in evidence-based reasoning, expert-level appreciation, and authenticity discrimination, highlighting the need for more sophisticated multimodal reasoning capabilities for museum and cultural heritage applications.

Abstract: We introduce CARTBENCH, a museum-grounded benchmark for evaluating vision-language models (VLMs) on Chinese artworks beyond short-form recognition and QA. CARTBENCH comprises four subtasks: CURATORQA for evidence-grounded recognition and reasoning, CATALOGCAPTION for structured four-section expert-style appreciation, REINTERPRET for defensible reinterpretation with expert ratings, and CONNOISSEURPAIRS for diagnostic authenticity discrimination under visually similar confounds. CARTBENCH is built by aligning image-bearing Palace Museum objects from Wikidata with authoritative catalog pages, spanning five art categories across multiple dynasties. Across nine representative VLMs, we find that high overall CURATORQA accuracy can mask sharp drops on hard evidence linking and style-to-period inference; long-form appreciation remains far from expert references; and authenticity-oriented diagnostic discrimination stays near chance, underscoring the difficulty of connoisseur-level reasoning for current models.

Method: Trained over 2,000 probes while systematically varying representation layer, feature type, and token aggregation strategy across models, tasks, and OOD settings to evaluate robustness of supervised uncertainty probes.

Result: Found poor robustness in current methods, especially for long-form generations. Probe robustness driven more by probe inputs than architecture: middle-layer representations generalize better than final-layer hidden states, and token aggregation across response tokens is more robust than single-token features.

Conclusion: Better evaluation is prerequisite for building more robust probes; proposed hybrid back-off strategy for improving robustness based on systematic findings about representation layers and token aggregation.

Abstract: Recent work has shown that the hidden states of large language models contain signals useful for uncertainty estimation and hallucination detection, motivating a growing interest in efficient probe-based approaches. Yet it remains unclear how robust existing methods are, and which probe designs provide uncertainty estimates that are reliable under distribution shift. We present a systematic study of supervised uncertainty probes across models, tasks, and OOD settings, training over 2,000 probes while varying the representation layer, feature type, and token aggregation strategy. Our evaluation highlights poor robustness in current methods, particularly in the case of long-form generations. We also find that probe robustness is driven less by architecture and more by the probe inputs. Middle-layer representations generalise more reliably than final-layer hidden states, and aggregating across response tokens is consistently more robust than relying on single-token features. These differences are often largely invisible in-distribution but become more important under distribution shift. Informed by our evaluation, we explore a simple hybrid back-off strategy for improving robustness, arguing that better evaluation is a prerequisite for building more robust probes.

Method: Proposes ToM-SB challenge where defenders act as double agents to steer attackers’ beliefs. Tests frontier models (Gemini3-Pro, GPT-5.4) and trains AI double agents using reinforcement learning with both fooling and ToM rewards. Evaluates across different attacker strengths and both in-distribution and out-of-distribution settings.

Result: Frontier models struggle on hard ToM-SB scenarios. RL-trained AI double agents outperform them, especially when combining both ToM and fooling rewards. Shows bidirectional emergence: fooling rewards improve ToM, and ToM rewards improve fooling. Gains in ToM and fooling are well-correlated.

Conclusion: ToM-SB is a challenging benchmark for LLM theory-of-mind capabilities. AI double agents trained with RL on both ToM and fooling objectives can significantly outperform current frontier models, demonstrating that belief modeling is key to success in adversarial privacy scenarios.

Abstract: As large language models (LLMs) become the engine behind conversational systems, their ability to reason about the intentions and states of their dialogue partners (i.e., form and use a theory-of-mind, or ToM) becomes increasingly critical for safe interaction with potentially adversarial partners. We propose a novel privacy-themed ToM challenge, ToM for Steering Beliefs (ToM-SB), in which a defender must act as a Double Agent to steer the beliefs of an attacker with partial prior knowledge within a shared universe. To succeed on ToM-SB, the defender must engage with and form a ToM of the attacker, with a goal of fooling the attacker into believing they have succeeded in extracting sensitive information. We find that strong frontier models like Gemini3-Pro and GPT-5.4 struggle on ToM-SB, often failing to fool attackers in hard scenarios with partial attacker prior knowledge, even when prompted to reason about the attacker’s beliefs (ToM prompting). To close this gap, we train models on ToM-SB to act as AI Double Agents using reinforcement learning, testing both fooling and ToM rewards. Notably, we find a bidirectionally emergent relationship between ToM and attacker-fooling: rewarding fooling success alone improves ToM, and rewarding ToM alone improves fooling. Across four attackers with different strengths, six defender methods, and both in-distribution and out-of-distribution (OOD) evaluation, we find that gains in ToM and attacker-fooling are well-correlated, highlighting belief modeling as a key driver of success on ToM-SB. AI Double Agents that combine both ToM and fooling rewards yield the strongest fooling and ToM performance, outperforming Gemini3-Pro and GPT-5.4 with ToM prompting on hard scenarios. We also show that ToM-SB and AI Double Agents can be extended to stronger attackers, demonstrating generalization to OOD settings and the upgradability of our task.

Method: Built a parallel corpus of 25,140 AI-input/human-reference text pairs, identified 11 measurable stylistic markers distinguishing AI vs. human text, and fine-tuned three models (BART-base, BART-large, and Mistral-7B-Instruct with QLoRA) for style transfer.

Result: BART-large achieved highest reference similarity (BERTScore F1: 0.924, ROUGE-L: 0.566, chrF++: 55.92) with 17x fewer parameters than Mistral-7B. Mistral-7B showed higher marker shift but this reflected overshoot rather than accuracy.

Conclusion: Current style transfer evaluation has a blind spot regarding shift accuracy, and smaller models like BART-large can outperform larger ones for this specific AI-to-human text transformation task.

Abstract: AI-generated text has become common in academic and professional writing, prompting research into detection methods. Less studied is the reverse: systematically rewriting AI-generated prose to read as genuinely human-authored. We build a parallel corpus of 25,140 paired AI-input and human-reference text chunks, identify 11 measurable stylistic markers separating the two registers, and fine-tune three models: BART-base, BART-large, and Mistral-7B-Instruct with QLoRA. BART-large achieves the highest reference similarity – BERTScore F1 of 0.924, ROUGE-L of 0.566, and chrF++ of 55.92 – with 17x fewer parameters than Mistral-7B. We show that Mistral-7B’s higher marker shift score reflects overshoot rather than accuracy, and argue that shift accuracy is a meaningful blind spot in current style transfer evaluation.

Method: Legal2LogicICL: few-shot retrieval framework balancing diversity and similarity at semantic and structural levels, mitigating entity-induced bias in legal texts. Uses retrieval-augmented generation for in-context learning without training.

Result: Significantly improves accuracy, stability, and generalization in transforming legal case descriptions to logical representations on both open-source and proprietary LLMs. New Legal2Proleg dataset supports evaluation.

Conclusion: Proposed framework enables effective few-shot legal reasoning with interpretable logical rule generation, addressing data scarcity through retrieval-augmented in-context learning.

Abstract: This work aims to improve the generalization of logic-based legal reasoning systems by integrating recent advances in NLP with legal-domain adaptive few-shot learning techniques using LLMs. Existing logic-based legal reasoning pipelines typically rely on fine-tuned models to map natural-language legal cases into logical formulas before forwarding them to a symbolic reasoner. However, such approaches are heavily constrained by the scarcity of high-quality annotated training data. To address this limitation, we propose a novel LLM-based legal reasoning framework that enables effective in-context learning through retrieval-augmented generation. Specifically, we introduce Legal2LogicICL, a few-shot retrieval framework that balances diversity and similarity of exemplars at both the latent semantic representation level and the legal text structure level. In addition, our method explicitly accounts for legal structure by mitigating entity-induced retrieval bias in legal texts, where lengthy and highly specific entity mentions often dominate semantic representations and obscure legally meaningful reasoning patterns. Our Legal2LogicICL constructs informative and robust few-shot demonstrations, leading to accurate and stable logical rule generation without requiring additional training. In addition, we construct a new dataset, named Legal2Proleg, which is annotated with alignments between legal cases and PROLEG logical formulas to support the evaluation of legal semantic parsing. Experimental results on both open-source and proprietary LLMs demonstrate that our approach significantly improves accuracy, stability, and generalization in transforming natural-language legal case descriptions into logical representations, highlighting its effectiveness for interpretable and reliable legal reasoning. Our code is available at https://github.com/yingjie7/Legal2LogicICL.

Method: Developed an open-source framework simulating leadership through elected personas and candidate-driven agendas, conducted empirical study of LLMs under controlled governance conditions, analyzed social influence via agent social graphs and centrality metrics, and performed sentiment analysis on leader utterances.

Result: Elected leadership improves social welfare scores by 55.4% and survival time by 128.6% across high-performing LLMs. Social graph analysis revealed leader influence patterns, and sentiment analysis showed rhetorical and cooperative tendencies in leader communications.

Conclusion: Leadership and election mechanisms significantly enhance cooperation and social welfare in multi-agent systems, laying foundation for further study of organizational structures in navigating complex social dilemmas.

Abstract: Governing common-pool resources requires agents to develop enduring strategies through cooperation and self-governance to avoid collective failure. While foundation models have shown potential for cooperation in these settings, existing multi-agent research provides little insight into whether structured leadership and election mechanisms can improve collective decision making. The lack of such a critical organizational feature ubiquitous in human society presents a significant shortcoming of the current methods. In this work we aim to directly address whether leadership and elections can support improved social welfare and cooperation through multi-agent simulation with LLMs. We present our open-source framework that simulates leadership through elected personas and candidate-driven agendas and carry out an empirical study of LLMs under controlled governance conditions. Our experiments demonstrate that having elected leadership improves social welfare scores by 55.4% and survival time by 128.6% across a range of high performing LLMs. Through the construction of an agent social graph we compute centrality metrics to assess the social influence of leader personas and also analyze rhetorical and cooperative tendencies revealed through a sentiment analysis on leader utterances. This work lays the foundation for further study of election mechanisms in multi-agent systems toward navigating complex social dilemmas.

Method: The paper introduces MINT (Multi-turn Inter-tactic Novelty Training), a reinforcement learning framework that optimizes discourse move diversity across multi-turn empathic dialogue. The best variant combines an empathy quality reward with a cross-turn tactic novelty signal.

Result: LLMs reuse tactics at nearly double the rate of humans in multi-turn conversations (0.50-0.56 vs. 0.27). MINT improves aggregate empathy by 25.3% over vanilla models and reduces cross-turn discourse move repetition by 26.3% on 4B models, surpassing all baselines including quality-only and token-level diversity methods.

Conclusion: Current LLMs lack not empathy itself, but the ability to vary their discourse moves across conversations. MINT effectively addresses this limitation by optimizing for discourse move diversity while maintaining or improving empathy quality.

Abstract: Large language models (LLMs) produce responses rated as highly empathic in single-turn settings (Ayers et al., 2023; Lee et al., 2024), yet they are also known to be formulaic generators that reuse the same lexical patterns, syntactic templates, and discourse structures across tasks (Jiang et al., 2025; Shaib et al., 2024; Namuduri et al., 2025). Less attention has been paid to whether this formulaicity extends to the level of discourse moves, i.e., what a response does for the person it is addressing. This question is especially consequential for empathic dialogue, where effective support demands not just a kind response at one moment but varied strategies as a conversation unfolds (Stiles et al., 1998). Indeed, prior work shows that LLMs reuse the same tactic sequences more than human supporters in single-turn settings (Gueorguieva et al., 2026). We extend this analysis to multi-turn conversations and find that the rigidity compounds: once a tactic appears in a supporter turn, LLMs reuse it in the next at nearly double the rate of humans (0.50-0.56 vs. 0.27). This pattern holds across LLMs serving as supporters in real emotional support conversations, and is invisible to standard similarity metrics. To address this gap, we introduce MINT (Multi-turn Inter-tactic Novelty Training), the first reinforcement learning framework to optimize discourse move diversity across multi-turn empathic dialogue. The best MINT variant combines an empathy quality reward with a cross-turn tactic novelty signal, improving aggregate empathy by 25.3% over vanilla across 1.7B and 4B models while reducing cross-turn discourse move repetition by 26.3% on the 4B model, surpassing all baselines including quality-only and token-level diversity methods on both measures. These results suggest that what current models lack is not empathy itself, but the ability to vary their discourse moves across a conversation.

Method: Connects embedding-space DLMs to Flow Matching via Bregman divergence. Introduces three key innovations: (1) ODE-based NLL bound for principled evaluation, (2) information-uniform principle for noise scheduling with learnable Gumbel-based scheduler, and (3) improved training protocol with self-conditioning.

Result: Achieves perplexity of 30.0 on LM1B and 24.6 on OpenWebText. Matches top discrete DLMs at comparable scale and surpasses autoregressive baselines in zero-shot transfer across multiple benchmarks.

Conclusion: LangFlow demonstrates that continuous diffusion is a competitive and promising paradigm for language modeling, closing the performance gap with discrete diffusion models.

Abstract: Continuous diffusion models have achieved strong performance across domains such as images. However, in language modeling, prior continuous diffusion language models (DLMs) lag behind discrete counterparts. In this work, we close this gap with LangFlow, the first continuous DLM to rival discrete diffusion. Our approach connects embedding-space DLMs to Flow Matching via Bregman divergence and introduces three key innovations: (1) a novel ODE-based NLL bound for principled evaluation of continuous flow-based language models; (2) an information-uniform principle for noise scheduling, motivating a learnable scheduler based on a Gumbel distribution; and (3) an improved training protocol incorporating self-conditioning, which enhances both likelihood and sample quality.LangFlow achieves strong performance across benchmarks, reaching a perplexity (PPL) of 30.0 on LM1B and 24.6 on OpenWebText. It matches top discrete DLMs at comparable scale and surpasses autoregressive baselines in zero-shot transfer across multiple benchmarks. LangFlow provides clear evidence that continuous diffusion is a competitive and promising paradigm for language modeling. https://github.com/nealchen2003/LangFlow

Method: Proposes HistLens, a SAE-based framework that decomposes concept representations into interpretable features and tracks their activation dynamics over time and across sources within a shared coordinate system.

Result: Experiments on long-span press corpora show HistLens supports cross-concept, cross-corpus computation of idea evolution patterns and enables implicit concept computation.

Conclusion: HistLens bridges conceptual modeling with interpretive needs, broadening analytical perspectives and methodological repertoire for social science and humanities diachronic text analysis.

Abstract: Language change both reflects and shapes social processes, and the semantic evolution of foundational concepts provides a measurable trace of historical and social transformation. Despite recent advances in diachronic semantics and discourse analysis, existing computational approaches often (i) concentrate on a single concept or a single corpus, making findings difficult to compare across heterogeneous sources, and (ii) remain confined to surface lexical evidence, offering insufficient computational and interpretive granularity when concepts are expressed implicitly. We propose HistLens, a unified, SAE-based framework for multi-concept, multi-corpus conceptual-history analysis. The framework decomposes concept representations into interpretable features and tracks their activation dynamics over time and across sources, yielding comparable conceptual trajectories within a shared coordinate system. Experiments on long-span press corpora show that HistLens supports cross-concept, cross-corpus computation of patterns of idea evolution and enables implicit concept computation. By bridging conceptual modeling with interpretive needs, HistLens broadens the analytical perspectives and methodological repertoire available to social science and the humanities for diachronic text analysis.

Method: Proposes AggAgent, an aggregation agent that treats parallel trajectories as an environment. Equips it with lightweight tools to inspect candidate solutions and search across trajectories, enabling on-demand navigation and synthesis of information from multiple parallel rollouts.

Result: Outperforms all existing aggregation methods by up to 5.3% absolute on average and 10.3% on two deep research tasks across six benchmarks and three model families (GLM-4.7, Qwen3.5, MiniMax-M2.5), while adding minimal overhead with aggregation cost bounded by a single agentic rollout.

Conclusion: Agentic aggregation is an effective and cost-efficient approach to parallel test-time scaling for long-horizon agentic tasks, establishing a new paradigm for leveraging multiple parallel trajectories in complex reasoning and research scenarios.

Abstract: We study parallel test-time scaling for long-horizon agentic tasks such as agentic search and deep research, where multiple rollouts are generated in parallel and aggregated into a final response. While such scaling has proven effective for chain-of-thought reasoning, agentic tasks pose unique challenges: trajectories are long, multi-turn, and tool-augmented, and outputs are often open-ended. Aggregating only final answers discards rich information from trajectories, while concatenating all trajectories exceeds the model’s context window. To address this, we propose AggAgent, an aggregation agent that treats parallel trajectories as an environment. We equip it with lightweight tools to inspect candidate solutions and search across trajectories, enabling it to navigate and synthesize information on demand. Across six benchmarks and three model families (GLM-4.7, Qwen3.5, MiniMax-M2.5), AggAgent outperforms all existing aggregation methods-by up to 5.3% absolute on average and 10.3% on two deep research tasks-while adding minimal overhead, as the aggregation cost remains bounded by a single agentic rollout. Our findings establish agentic aggregation as an effective and cost-efficient approach to parallel test-time scaling.

Method: Created General365 benchmark with 365 seed problems and 1,095 variants across 8 categories, restricting background knowledge to K-12 level to explicitly decouple reasoning from specialized expertise. Evaluated 26 leading LLMs on this benchmark.

Result: Top-performing LLM achieved only 62.8% accuracy, in stark contrast to near-perfect performances in math/physics benchmarks. Results show current LLM reasoning abilities are heavily domain-dependent with significant room for improvement.

Conclusion: General reasoning remains a major challenge for LLMs despite strong domain-specific performance. General365 serves as a catalyst for advancing LLM reasoning toward robust, general-purpose real-world applications beyond specialized domains.

Abstract: Contemporary large language models (LLMs) have demonstrated remarkable reasoning capabilities, particularly in specialized domains like mathematics and physics. However, their ability to generalize these reasoning skills to more general and broader contexts–often termed general reasoning–remains under-explored. Unlike domain-specific reasoning, general reasoning relies less on expert knowledge but still presents formidable reasoning challenges, such as complex constraints, nested logical branches, and semantic interference. To address this gap, we introduce General365, a benchmark specifically designed to assess general reasoning in LLMs. By restricting background knowledge to a K-12 level, General365 explicitly decouples reasoning from specialized expertise. The benchmark comprises 365 seed problems and 1,095 variant problems across eight categories, ensuring both high difficulty and diversity. Evaluations across 26 leading LLMs reveal that even the top-performing model achieves only 62.8% accuracy, in stark contrast to the near-perfect performances of LLMs in math and physics benchmarks. These results suggest that the reasoning abilities of current LLMs are heavily domain-dependent, leaving significant room for improvement in broader applications. We envision General365 as a catalyst for advancing LLM reasoning beyond domain-specific tasks toward robust, general-purpose real-world scenarios. Code, Dataset, and Leaderboard: https://general365.github.io

Method: Proposes C-ReD benchmark with comprehensive coverage: diverse Chinese LLMs, multiple domains (news, academic, creative writing), and realistic prompts. Focuses on addressing model diversity, domain coverage, and prompt realism limitations of prior Chinese benchmarks.

Result: C-ReD enables reliable in-domain detection and shows strong generalization to unseen LLMs and external Chinese datasets. Successfully addresses critical gaps in model diversity, domain coverage, and prompt realism that limited prior Chinese detection benchmarks.

Conclusion: C-ReD provides a comprehensive benchmark for Chinese AI-generated text detection that overcomes limitations of existing datasets, supporting both reliable detection and generalization to new models and domains.

Abstract: Recently, large language models (LLMs) are capable of generating highly fluent textual content. While they offer significant convenience to humans, they also introduce various risks, like phishing and academic dishonesty. Numerous research efforts have been dedicated to developing algorithms for detecting AI-generated text and constructing relevant datasets. However, in the domain of Chinese corpora, challenges remain, including limited model diversity and data homogeneity. To address these issues, we propose C-ReD: a comprehensive Chinese Real-prompt AI-generated Detection benchmark. Experiments demonstrate that C-ReD not only enables reliable in-domain detection but also supports strong generalization to unseen LLMs and external Chinese datasets-addressing critical gaps in model diversity, domain coverage, and prompt realism that have limited prior Chinese detection benchmarks. We release our resources at https://github.com/HeraldofLight/C-ReD.

Method: Proposes CLSGen framework with new model architecture, training methodology, and data construction strategy specifically for binary classification tasks to enable robust probability estimation while preserving explanation generation.

Result: CLSGen outperforms baselines on classification metrics (AUROC and F1-score) across multiple benchmark datasets. Generated explanations show strong label-justification alignment and high readability.

Conclusion: CLSGen successfully enables LLMs to provide reliable probability estimates for binary classification without sacrificing explanation capabilities, addressing a critical deployment hurdle for practical decision-making applications.

Abstract: With the recent progress of Large Language Models (LLMs), there is a growing interest in applying these models to solve complex and challenging problems. Modern LLMs, capable of processing long contexts and generating verbalized explanations, offer significant potential in addressing real-world applications. However, a critical hurdle in deploying LLMs for practical decision-making is their inability to provide reliable, quantitative probabilities. While task-specific fine-tuning of LLMs using traditional discriminative objectives (similar to encoder-only models) can yield probability estimates, this often leads to catastrophic forgetting and linguistic collapse. Consequently, the model loses its ability to generate explanations, severely undermining its interpretability and usability. To address this challenge, we propose CLSGen, a novel LLM fine-tuning framework designed for binary classification tasks. The CLSGen framework encompasses a new model architecture, training methodology, and data construction strategy to enable robust probability estimation without sacrificing the model’s inherent explanation-generation capabilities. Experimental results across multiple benchmark datasets demonstrate that models fine-tuned with CLSGen outperform existing baselines in classification metrics (AUROC and F1-score). Regarding explanation, the results showed strong alignment between predicted labels and generated justifications, as well as high readability.

Method: Used probing to examine emergence of Big Five information across model layers, identified concept-selective neurons for each trait, and conducted interventions (enhancing/suppressing activations) to test causal relationships between representations and behavioral outputs.

Result: Big Five information becomes decodable early and persists through layers; concept-selective neurons are most prevalent in mid-layers with limited cross-domain overlap. Interventions successfully shift probe readouts (targeted success rates >0.8) but have weaker, more variable effects on generated labels with cross-trait spillover.

Conclusion: There’s a gap between representational control and behavioral control in LLMs - while internal personality representations can be causally steered, comparable control over generated behavioral outputs is more difficult even with interventions on concept-selective neurons.

Abstract: Using psychological constructs such as the Big Five, large language models (LLMs) can imitate specific personality profiles and predict a user’s personality. While LLMs can exhibit behaviors consistent with these constructs, it remains unclear where and how they are represented inside the model and how they relate to behavioral outputs. To address this gap, we focus on questionnaire-operationalized Big Five concepts, analyze the formation and localization of their internal representations, and use interventions to examine how these representations relate to behavioral outputs. In our experiment, we first use probing to examine where Big Five information emerges across model depth. We then identify neurons that respond selectively to each Big Five concept and test whether enhancing or suppressing their activations can bias latent representations and label generation in intended directions. We find that Big Five information becomes rapidly decodable in early layers and remains detectable through the final layers, while concept-selective neurons are most prevalent in mid layers and exhibit limited overlap across domains. Interventions on these neurons consistently shift probe readouts toward targeted concepts, with targeted success rates exceeding 0.8 for some concepts, indicating that the model’s internal separation of Big Five personality traits can be causally steered. At the label-generation level, the same interventions often bias generated label distributions in the intended directions, but the effects are weaker, more concept-dependent, and often accompanied by cross-trait spillover, indicating that comparable control over generated labels is difficult even with interventions on a large fraction of concept-selective neurons. Overall, our findings reveal a gap between representational control and behavioral control in LLMs.

Method: Created a 6-hour speech corpus by collecting Saarbrücken dialect text from digitized books and local materials, then recording a subset with 9 speakers. Conducted analyses on text and speech components, addressing orthographic/speaker variation and grapheme-to-phoneme conversion challenges.

Result: Produced Saar-Voice corpus with aligned textual and audio representations for the Saarbrücken German dialect, enabling dialect-aware text-to-speech research in low-resource scenarios.

Conclusion: The corpus provides a foundation for future research on dialect-aware TTS, particularly for zero-shot and few-shot model adaptation in low-resource dialect scenarios.

Abstract: Natural language processing (NLP) and speech technologies have made significant progress in recent years; however, they remain largely focused on standardized language varieties. Dialects, despite their cultural significance and widespread use, are underrepresented in linguistic resources and computational models, resulting in performance disparities. To address this gap, we introduce Saar-Voice, a six-hour speech corpus for the Saarbrücken dialect of German. The dataset was created by first collecting text through digitized books and locally sourced materials. A subset of this text was recorded by nine speakers, and we conducted analyses on both the textual and speech components to assess the dataset’s characteristics and quality. We discuss methodological challenges related to orthographic and speaker variation, and explore grapheme-to-phoneme (G2P) conversion. The resulting corpus provides aligned textual and audio representations. This serves as a foundation for future research on dialect-aware text-to-speech (TTS), particularly in low-resource scenarios, including zero-shot and few-shot model adaptation.

Method: Template-aware Dialogue Sentence Embedding (TaDSE) uses template information to learn utterance embeddings via self-supervised contrastive learning, enhanced with synthetically augmented dataset that diversifies utterance-template association using slot-filling.

Result: TaDSE achieves significant improvements over previous SOTA methods on five downstream benchmark dialogue datasets and introduces a novel semantic compression test showing correlation with uniformity and alignment.

Conclusion: Template information effectively improves dialogue sentence embeddings through contrastive learning, with synthetic augmentation further enhancing performance across multiple dialogue tasks.

Abstract: Learning high quality sentence embeddings from dialogues has drawn increasing attentions as it is essential to solve a variety of dialogue-oriented tasks with low annotation cost. Annotating and gathering utterance relationships in conversations are difficult, while token-level annotations, \eg, entities, slots and templates, are much easier to obtain. Other sentence embedding methods are usually sentence-level self-supervised frameworks and cannot utilize token-level extra knowledge. We introduce Template-aware Dialogue Sentence Embedding (TaDSE), a novel augmentation method that utilizes template information to learn utterance embeddings via self-supervised contrastive learning framework. We further enhance the effect with a synthetically augmented dataset that diversifies utterance-template association, in which slot-filling is a preliminary step. We evaluate TaDSE performance on five downstream benchmark dialogue datasets. The experiment results show that TaDSE achieves significant improvements over previous SOTA methods for dialogue. We further introduce a novel analytic instrument of semantic compression test, for which we discover a correlation with uniformity and alignment. Our code is available at https://github.com/minsik-ai/Template-Contrastive-Embedding

Method: Proposes PredFT with main and side networks. Side network extracts predictive representations from brain ROIs using self-attention, then fuses them into main network for continuous language decoding.

Result: Outperforms current decoding models on multiple evaluation metrics across two naturalistic language comprehension fMRI datasets.

Conclusion: Incorporating predictive coding theory improves fMRI-to-text decoding, demonstrating the value of leveraging the brain’s natural predictive mechanisms for language reconstruction.

Abstract: Many recent studies have shown that the perception of speech can be decoded from brain signals and subsequently reconstructed as continuous language. However, there is a lack of neurological basis for how the semantic information embedded within brain signals can be used more effectively to guide language reconstruction. Predictive coding theory suggests the human brain naturally engages in continuously predicting future words that span multiple timescales. This implies that the decoding of brain signals could potentially be associated with a predictable future. To explore the predictive coding theory within the context of language reconstruction, this paper proposes \textsc{PredFT}(\textbf{F}MRI-to-\textbf{T}ext decoding with \textbf{Pred}ictive coding). \textsc{PredFT} consists of a main network and a side network. The side network obtains brain predictive representation from related regions of interest(ROIs) with a self-attention module. The representation is then fused into the main network for continuous language decoding. Experiments on two naturalistic language comprehension fMRI datasets show that \textsc{PredFT} outperforms current decoding models on several evaluation metrics.

Method: Proposes late multi-image fusion: generates multiple images from text prompts using lightweight parallel sampling, then combines their prediction probabilities with text-only LLM outputs through a late-fusion layer that integrates projected visual features before final prediction.

Result: Significantly outperforms augmented LLMs on visual reasoning, matches VLMs on vision-based tasks, and improves NLP performance when applied to strong LLMs like LLaMA 3, with modest test-time overhead.

Conclusion: Late multi-image fusion effectively enhances LLMs with visual commonsense reasoning capabilities while maintaining or improving text-only performance, offering a cost-effective alternative to full multimodal training.

Abstract: Commonsense reasoning often requires both textual and visual knowledge, yet Large Language Models (LLMs) trained solely on text lack visual grounding (e.g., “what color is an emperor penguin’s belly?”). Visual Language Models (VLMs) perform better on visually grounded tasks but face two limitations: (i) often reduced performance on text-only commonsense reasoning compared to text-trained LLMs, and (ii) adapting newly released LLMs to vision input typically requires costly multimodal training. An alternative augments LLMs with test-time visual signals, improving visual commonsense without harming textual reasoning, but prior designs often rely on early fusion and a single image, which can be suboptimal. We propose a late multi-image fusion method: multiple images are generated from the text prompt with a lightweight parallel sampling, and their prediction probabilities are combined with those of a text-only LLM through a late-fusion layer that integrates projected visual features just before the final prediction. Across visual commonsense and NLP benchmarks, our method significantly outperforms augmented LLMs on visual reasoning, matches VLMs on vision-based tasks, and, when applied to strong LLMs such as LLaMA 3, also improves NLP performance while adding only modest test-time overhead. Project page is available at: https://guyyariv.github.io/LaMI.

Method: Introduced MMM dataset with 21 sub-datasets across three languages, using LLM-assisted dataset translation and alignment framework to reduce manual annotation. Adopted unified input-output framework to train open-domain information extraction model with full fine-tuning ablations and knowledgeable verbalizers based on MRE-mix data.

Result: Experimental results show 76% of MMM sub-datasets consistently exhibit Mutual Reinforcement Effect across languages, providing systematic empirical validation of MRE in multilingual settings.

Conclusion: The study provides empirical evidence for the generality of Mutual Reinforcement Effect across languages and demonstrates its practical value for information extraction through systematic multilingual validation.

Abstract: The Mutual Reinforcement Effect (MRE) describes a phenomenon in information extraction where word-level and sentence-level tasks can mutually improve each other when jointly modeled. While prior work has reported MRE in Japanese, its generality across languages and task settings has not been empirically validated, largely due to the lack of multilingual MRE datasets. To address this limitation, we introduce the Multilingual MRE Mix dataset (MMM), which consists of 21 sub-datasets covering English, Japanese, and Chinese. We propose an LLM-assisted dataset translation and alignment framework that significantly reduces manual annotation effort while preserving the structural requirements of MRE tasks. Building on MMM, we adopt a unified input-output framework to train an open-domain information extraction model and conduct extensive empirical studies, including full fine-tuning ablations and the construction of knowledgeable verbalizers based on MRE-mix data. Experimental results show that 76 percent of the MMM sub-datasets consistently exhibit the Mutual Reinforcement Effect across languages. These findings provide systematic empirical validation of MRE in multilingual settings and demonstrate its practical value for information extraction.

Method: Developed LLM-Fake Theory integrating social psychology theories to explain machine-generated deception. Designed an innovative prompt engineering pipeline that automates fake news generation using LLMs without manual annotation. Created MegaFake dataset derived from FakeNewsNet using this pipeline.

Result: Created a theoretically informed machine-generated fake news dataset (MegaFake) and advanced both theoretical understanding of human-machine deception mechanisms and practical approaches to fake news detection in the LLM era.

Conclusion: The study provides a theoretical framework and dataset for understanding and detecting LLM-generated fake news, addressing a critical challenge in the era of generative AI.

Abstract: Fake news significantly influences decision-making processes by misleading individuals, organizations, and even governments. Large language models (LLMs), as part of generative AI, can amplify this problem by generating highly convincing fake news at scale, posing a significant threat to online information integrity. Therefore, understanding the motivations and mechanisms behind fake news generated by LLMs is crucial for effective detection and governance. In this study, we develop the LLM-Fake Theory, a theoretical framework that integrates various social psychology theories to explain machine-generated deception. Guided by this framework, we design an innovative prompt engineering pipeline that automates fake news generation using LLMs, eliminating manual annotation needs. Utilizing this pipeline, we create a theoretically informed \underline{M}achin\underline{e}-\underline{g}ener\underline{a}ted \underline{Fake} news dataset, MegaFake, derived from FakeNewsNet. Through extensive experiments with MegaFake, we advance both theoretical understanding of human-machine deception mechanisms and practical approaches to fake news detection in the LLM era.

Method: Developed RiTeK dataset by synthesizing realistic user queries with diverse topological structures, relational information, and complex textual descriptions, followed by rigorous medical expert evaluation. Used this benchmark to evaluate 11 representative LLM-based retrieval systems.

Result: Existing retrieval methods struggle on the RiTeK benchmark, revealing notable limitations in current LLM-driven retrieval approaches for semi-structured medical data.

Conclusion: There is a pressing need for more effective retrieval systems tailored for semi-structured data in the medical domain, and RiTeK serves as a comprehensive benchmark to drive future research in this area.

Abstract: Answering complex real-world questions in the medical domain often requires accurate retrieval from medical Textual Knowledge Graphs (medical TKGs), as the relational path information from TKGs could enhance the inference ability of Large Language Models (LLMs). However, the main bottlenecks lie in the scarcity of existing medical TKGs, the limited expressiveness of their topological structures, and the lack of comprehensive evaluations of current retrievers for medical TKGs. To address these challenges, we first develop a Dataset1 for LLMs Complex Reasoning over medical Textual Knowledge Graphs (RiTeK), covering a broad range of topological structures. Specifically, we synthesize realistic user queries integrating diverse topological structures, relational information, and complex textual descriptions. We conduct a rigorous medical expert evaluation process to assess and validate the quality of our synthesized queries. RiTeK also serves as a comprehensive benchmark dataset for evaluating the capabilities of retrieval systems built upon LLMs. By assessing 11 representative retrievers on this benchmark, we observe that existing methods struggle to perform well, revealing notable limitations in current LLM-driven retrieval approaches. These findings highlight the pressing need for more effective retrieval systems tailored for semi-structured data in the medical domain.

Method: Proposes a theoretical framework reframing unlearning as a backdoor attack/defense problem, where forget-tokens become triggers. Introduces Random Noise Augmentation (RNA) as a lightweight, model-agnostic defense approach with theoretical guarantees to mitigate these vulnerabilities.

Result: Extensive experiments show RNA significantly improves robustness of unlearned models while preserving forget and retain performance, demonstrating that unlearning methods inherently create backdoor vulnerabilities that can be mitigated through proper defense mechanisms.

Conclusion: LLM unlearning methods themselves poison models by creating backdoor vulnerabilities rather than erasing knowledge, and the proposed backdoor attack-defense framework provides insights for improving unlearning robustness through approaches like RNA.

Abstract: Here, we show that current LLM unlearning methods inherently reduce models’ robustness, causing them to misbehave even when a single non-adversarial forget-token is present in the retain-query. Toward understanding underlying causes, we propose a novel theoretical framework that reframes the unlearning process as a backdoor attack and defense problem: we formulate how the forgetting process inadvertently learns to align forget-tokens (backdoor triggers) with the target-representations (target labels). As a result, forget-tokens act as backdoor triggers that, when activated in retain-queries, cause disruptions in unlearned models’ behaviors, similar to successful backdoor attacks. The sense that, LLM unlearning methods themselves poison the model, make it more vulnerable to forget-tokens, and hide rather than erase target knowledge, describes their true mechanism. To mitigate the vulnerability caused by the forgetting process, we reinterpret the retaining process as a backdoor defense and propose Random Noise Augmentation (RNA), a lightweight, model and method-agnostic approach with theoretical guarantees for improving the robustness of unlearned models. Extensive experiments demonstrate that RNA significantly improves the robustness of unlearned models while preserving forget and retain performances. This backdoor attack-defense framework offers insights into the mechanism of unlearning that can shed light on future research directions for improving unlearning robustness.

Method: ULPT optimizes prompts in an ultra-low-dimensional space (e.g., 2D) and uses a frozen random matrix for up-projection to the model’s hidden dimension, dramatically reducing trainable parameters.

Result: Achieves 98% reduction in training parameters compared to vanilla prompt tuning while maintaining performance, consistently outperforms recent parameter-efficient tuning methods across 20+ NLP tasks.

Conclusion: ULPT provides a storage-efficient framework for massive LLM customization with minimal parameter overhead, making it practical for large-scale model deployment.

Abstract: Large language models achieve state-of-the-art performance but are increasingly costly to fine-tune. Prompt tuning is a parameter-efficient fine-tuning method that addresses parameter-efficiency by learning prompt embeddings, but these embeddings are typically tied to the model’s hidden dimensionality, limiting parameter saving. In this paper, we propose Ultra-Low-dimensional Prompt Tuning (ULPT), a simple yet effective method that optimizes prompts in a low-dimensional space (e.g., 2D) and uses a frozen random matrix for up-projection. ULPT can achieve 98% reduction in the training parameters compared to vanilla prompt tuning while preserving performance. Our extensive experiments across over 20 NLP tasks demonstrate that ULPT consistently outperforms recent parameter-efficient tuning methods using significantly fewer parameters, making it well-suited as a storage-efficient framework for massive LLM customization.

Method: Created CounterBench dataset with 1K counterfactual reasoning questions varying in difficulty, causal graph structures, question types, and name variants. Proposed CoIn (Counterfactual Inference) method that guides LLMs through iterative reasoning and backtracking to systematically explore counterfactual solutions.

Result: Most LLMs perform at random guessing levels on counterfactual reasoning tasks. CoIn method significantly improves LLM performance on counterfactual reasoning and consistently enhances performance across different LLM models.

Conclusion: Counterfactual reasoning remains a significant challenge for LLMs, but structured reasoning approaches like CoIn can substantially improve their performance on formal counterfactual inference tasks.

Abstract: Counterfactual reasoning is widely recognized as one of the most challenging and intricate aspects of causality in artificial intelligence. In this paper, we evaluate the performance of large language models (LLMs) in counterfactual reasoning. In contrast to previous studies that primarily focus on commonsense causal reasoning, where LLMs often rely on prior knowledge for inference, we specifically assess their ability to perform counterfactual inference using a set of formal rules. To support this evaluation, we introduce a new benchmark dataset, CounterBench, comprising 1K counterfactual reasoning questions. The dataset is designed with varying levels of difficulty, diverse causal graph structures, distinct types of counterfactual questions, and multiple nonsensical name variants. Our experiments demonstrate that counterfactual reasoning poses a significant challenge for LLMs, with most models performing at levels comparable to random guessing. To enhance LLM’s counterfactual reasoning ability, we propose a novel reasoning paradigm, CoIn, which guides LLMs through iterative reasoning and backtracking to systematically explore counterfactual solutions. Experimental results show that our method significantly improves LLM performance on counterfactual reasoning tasks and consistently enhances performance across different LLMs.Our dataset is available at https://huggingface.co/datasets/CounterBench/CounterBench.

Method: LIFT fine-tunes short-context LLMs on specific long inputs by adapting model parameters to absorb the long context. It uses LLM-generated synthetic tasks to enhance comprehension beyond memorization, with an optimized pipeline reducing Time to First Token to <10 seconds for 8k context.

Result: LIFT enables short-context LLMs to answer questions about long inputs even when the original context isn’t provided during inference, avoiding quadratic complexity while maintaining performance on long-context understanding tasks.

Conclusion: LIFT offers a novel parameter-adaptive approach to long-context modeling that avoids quadratic attention costs, provides feasibility for real-world deployment, and opens new directions for efficient long-context understanding in LLMs.

Abstract: Long context understanding remains challenging for large language models due to their limited context windows. This paper introduces Long Input Fine-Tuning (LIFT), a novel framework for long-context modeling that can enhance the long-context performance of arbitrary short-context LLMs by dynamically adapting their parameters to the given long input. Importantly, rather than endlessly extending the context window size to accommodate increasingly longer inputs in context, LIFT stores and absorbs the long input in parameters. By fine-tuning the long input into model parameters, LIFT allows short-context LLMs to answer questions even when the required information is not provided in the context during inference, avoiding the quadratic complexity w.r.t. input length of a normal long context model. Furthermore, LIFT does not simply perform continued pretraining on new, long contexts, but leverages carefully designed LLM-generated synthetic tasks to enhance the comprehension of long contexts, moving beyond mere memorization. To accommodate the additional cost of fine-tuning, we design a highly optimized pipeline that reduces the Time to First Token (TTFT) to less than 10 seconds for 8k context. We further provide a comprehensive analysis of LIFT’s strengths and limitations in long-context understanding, discuss its feasibility for large-scale real-world deployment, and highlight valuable directions for future research.

Method: Proposed professional evaluation principles for therapists’ responses, created PsychoCounsel-Preference dataset (36k preference comparison pairs), and trained aligned models using reward modeling and preference learning

Result: PsychoCounsel-Llama3-8B achieves 87% win rate against GPT-4o, demonstrating effectiveness in psycho-counseling response generation

Conclusion: The proposed dataset and aligned models provide valuable resources for improving LLMs in psycho-counseling applications, addressing data quality and privacy challenges

Abstract: Applying large language models (LLMs) to assist in psycho-counseling is an emerging and meaningful approach, driven by the significant gap between patient needs and the availability of mental health support. However, current LLMs struggle to consistently provide effective responses to client speeches, largely due to the lack of supervision from high-quality real psycho-counseling data, whose content is typically inaccessible due to client privacy concerns. Furthermore, the quality of therapists’ responses in available sessions can vary significantly based on their professional training and experience. Assessing the quality of therapists’ responses remains an open challenge. In this work, we address these challenges by first proposing a set of professional and comprehensive principles to evaluate therapists’ responses to client speeches. Using these principles, we create a preference dataset, PsychoCounsel-Preference, which contains 36k high-quality preference comparison pairs. This dataset aligns with the preferences of professional psychotherapists, providing a robust foundation for evaluating and improving LLMs in psycho-counseling. Experiments on reward modeling and preference learning demonstrate that PsychoCounsel-Preference is an excellent resource for LLMs to acquire essential skills for responding to clients in a counseling session. Our best-aligned model, PsychoCounsel-Llama3-8B, achieves an impressive win rate of 87% against GPT-4o. We release PsychoCounsel-Preference, PsychoCounsel-Llama3-8B and the reward model PsychoCounsel Llama3-8B-Reward to facilitate the research of psycho-counseling with LLMs at: https://hf.co/Psychotherapy-LLM.

Method: Created OlymMATH benchmark with 350 manually curated Olympiad-level problems from printed publications to minimize data contamination. Includes dual evaluation paradigms: (1) OlymMATH-EASY/HARD with 200 computational problems for natural language evaluation, and (2) OlymMATH-LEAN with 150 problems formalized in Lean 4 for rigorous process-level verification.

Result: The benchmark presents significant challenges to current models, revealing performance gaps between languages and instances where models rely on heuristic “guessing” rather than rigorous reasoning. The authors release extensive resources including 582k+ reasoning trajectories and visualization tools.

Conclusion: OlymMATH provides a comprehensive, challenging benchmark for evaluating advanced mathematical reasoning capabilities, with dual evaluation paradigms that enable both objective assessment and rigorous process verification, supporting future research in reasoning model development.

Abstract: The rapid advancement of large reasoning models has saturated existing math benchmarks, underscoring the urgent need for more challenging evaluation frameworks. To address this, we introduce OlymMATH, a rigorously curated, Olympiad-level math benchmark comprising 350 problems, each with parallel English and Chinese versions. OlymMATH is the first benchmark to unify dual evaluation paradigms within a single suite: (1) natural language evaluation through OlymMATH-EASY and OlymMATH-HARD, comprising 200 computational problems with numerical answers for objective rule-based assessment, and (2) formal verification through OlymMATH-LEAN, offering 150 problems formalized in Lean 4 for rigorous process-level evaluation. All problems are manually sourced from printed publications to minimize data contamination, verified by experts, and span four core domains. Extensive experiments reveal the benchmark’s significant challenge, and our analysis also uncovers consistent performance gaps between languages and identifies cases where models employ heuristic “guessing” rather than rigorous reasoning. To further support community research, we release 582k+ reasoning trajectories, a visualization tool, and expert solutions at https://github.com/RUCAIBox/OlymMATH.

Method: Introduces LIFESTATE-BENCH with two episodic datasets (Hamlet and synthetic scripts) rich in narrative structure and character interactions. Uses fact checking evaluation to probe models’ self-awareness, episodic memory retrieval, and relationship tracking across parametric and non-parametric approaches.

Result: Nonparametric methods significantly outperform parametric ones in managing stateful learning, but all models exhibit challenges with catastrophic forgetting as interactions extend, highlighting limitations in current lifelong learning capabilities.

Conclusion: There’s a need for further advancements in lifelong learning for LLMs, as current models struggle with maintaining stateful knowledge over extended interactions despite showing emergent character-like behaviors in multi-agent scenarios.

Abstract: Large language models (LLMs) can carry out human-like dialogue, but unlike humans, they are stateless due to the superposition property. However, during multi-turn, multi-agent interactions, LLMs begin to exhibit consistent, character-like behaviors, hinting at a form of emergent lifelong learning. Despite this, existing benchmarks often fail to capture these dynamics, primarily focusing on static, open-ended evaluations. To address this gap, we introduce LIFESTATE-BENCH, a benchmark designed to assess lifelong learning in LLMs. It features two episodic datasets: Hamlet and a synthetic script collection, rich in narrative structure and character interactions. Our fact checking evaluation probes models’ self-awareness, episodic memory retrieval, and relationship tracking, across both parametric and non-parametric approaches. Experiments on models like Llama3.1-8B, GPT-4-turbo, and DeepSeek R1, we demonstrate that nonparametric methods significantly outperform parametric ones in managing stateful learning. However, all models exhibit challenges with catastrophic forgetting as interactions extend, highlighting the need for further advancements in lifelong learning.

Method: Two-stage progressive fine-tuning: 1) Fine-tune LLMs on anchor behaviors while preserving general behavioral knowledge, 2) Fine-tune using balanced subset based on sample difficulty to improve tail behavior predictions without sacrificing anchor performance.

Result: Experimental results on two real-world datasets show BehaviorLM robustly predicts both anchor and tail behaviors and effectively leverages LLM behavioral knowledge for few-shot tail behavior prediction.

Conclusion: BehaviorLM addresses the long-tail behavior prediction problem in LLMs through progressive fine-tuning, enabling better prediction of both frequent and rare user behaviors.

Abstract: Predicting user behavior is essential for intelligent assistant services, yet deep learning models often struggle to capture long-tailed behaviors. Large language models (LLMs), with their pretraining on vast corpora containing rich behavioral knowledge, offer promise. However, existing fine-tuning approaches tend to overfit to frequent anchor'' behaviors, reducing their ability to predict less common tail’’ behaviors. In this paper, we introduce BehaviorLM, a progressive fine-tuning approach that addresses this issue. In the first stage, LLMs are fine-tuned on anchor behaviors while preserving general behavioral knowledge. In the second stage, fine-tuning uses a balanced subset of all behaviors based on sample difficulty to improve tail behavior predictions without sacrificing anchor performance. Experimental results on two real-world datasets demonstrate that BehaviorLM robustly predicts both anchor and tail behaviors and effectively leverages LLM behavioral knowledge to master tail behavior prediction with few-shot examples.

Method: Define marker confidence as observed accuracy when a model employs an epistemic marker. Evaluate stability across multiple question-answering datasets in both in-distribution and out-of-distribution settings for open-source and proprietary LLMs.

Result: While markers generalize well within the same distribution, their confidence is inconsistent in out-of-distribution scenarios. This raises significant concerns about the reliability of epistemic markers for confidence estimation.

Conclusion: There is a need for improved alignment between marker-based confidence and actual model uncertainty, as current epistemic markers are unreliable for confidence estimation, especially in out-of-distribution scenarios.

Abstract: As large language models (LLMs) are increasingly used in high-stakes domains, accurately assessing their confidence is crucial. Humans typically express confidence through epistemic markers (e.g., “fairly confident”) instead of numerical values. However, it remains unclear whether LLMs consistently use these markers to reflect their intrinsic confidence due to the difficulty of quantifying uncertainty associated with various markers. To address this gap, we first define marker confidence as the observed accuracy when a model employs an epistemic marker. We evaluate its stability across multiple question-answering datasets in both in-distribution and out-of-distribution settings for open-source and proprietary LLMs. Our results show that while markers generalize well within the same distribution, their confidence is inconsistent in out-of-distribution scenarios. These findings raise significant concerns about the reliability of epistemic markers for confidence estimation, underscoring the need for improved alignment between marker based confidence and actual model uncertainty. Our code is available at https://github.com/HKUST-KnowComp/MarConf.

Method: Introduce Post-hoc Self-Consistency Bank (PSCB) - a large-scale benchmark linking model decisions with diverse explanations and attribution vectors across datasets, methods, and model families. Use Spearman rank correlation to measure alignment, then apply Direct Preference Optimization (DPO) to attribution-based preference data.

Result: Spearman rank correlation provides more reliable signal of alignment than cosine similarity. DPO improves alignment between decisions and explanations without degrading task accuracy, while standard supervised fine-tuning fails to achieve comparable gains. Improvements generalize robustly across domains.

Conclusion: The work enables scalable and faithful alignment between LLM decisions and their natural language explanations, addressing the gap between what drives model outputs and what explanations claim.

Abstract: Large language models (LLMs) seem to offer an easy path to interpretability: just ask them to explain their answers. Yet the features driving an answer often differ from those emphasized in its explanation, meaning post-hoc rationales can misrepresent what actually shaped the model’s output. We quantify this gap by comparing the feature-importance distributions of answers and their explanations. Prior analyses reveal such discrepancies, but large-scale study has been limited by the high computational cost of attribution methods. To address this, we introduce the Post-hoc Self-Consistency Bank (PSCB), a large-scale benchmark linking model decisions with diverse explanations and attribution vectors across datasets, methods, and model families. Using PSCB, we find that Spearman rank correlation provides a more reliable signal of alignment than cosine similarity. Building on this insight, we apply Direct Preference Optimization (DPO) to attribution-based preference data, improving alignment without degrading task accuracy, and show that standard supervised fine-tuning on the same data fails to achieve comparable gains. These improvements generalize robustly across domains, paving the way toward scalable and faithful alignment between LLM decisions and their natural language explanations.

Method: Two key mechanisms: 1) POS-Aware Delay Mechanism that adjusts attention weights based on POS tags to postpone EOS token generation, and 2) Generative Path Pruning Mechanism that limits hidden state magnitudes to encourage persistent repetitive loops.

Result: LingoLoop successfully traps MLLMs like Qwen2.5-VL-3B in generative loops, inducing outputs with up to 367x more tokens than clean inputs, leading to massive energy consumption increases when generation limits are relaxed.

Conclusion: The attack exposes significant vulnerabilities in MLLMs related to their generation mechanisms, posing serious challenges for reliable deployment due to resource exhaustion risks.

Abstract: Multimodal Large Language Models (MLLMs) have shown great promise but require substantial computational resources during inference. Attackers can exploit this by inducing excessive output, leading to resource exhaustion and service degradation. Prior energy-latency attacks aim to increase generation time by broadly shifting the output token distribution away from the EOS token, but they neglect the influence of token-level Part-of-Speech (POS) characteristics on EOS and sentence-level structural patterns on output counts, limiting their efficacy. To address this, we propose LingoLoop, an attack designed to induce MLLMs to generate excessively verbose and repetitive sequences. First, we find that the POS tag of a token strongly affects the likelihood of generating an EOS token. Based on this insight, we propose a POS-Aware Delay Mechanism to postpone EOS token generation by adjusting attention weights guided by POS information. Second, we identify that constraining output diversity to induce repetitive loops is effective for sustained generation. We introduce a Generative Path Pruning Mechanism that limits the magnitude of hidden states, encouraging the model to produce persistent loops. Extensive experiments on models like Qwen2.5-VL-3B demonstrate LingoLoop’s powerful ability to trap them in generative loops; it consistently drives them to their generation limits and, when those limits are relaxed, can induce outputs with up to 367x more tokens than clean inputs, triggering a commensurate surge in energy consumption. These findings expose significant MLLMs’ vulnerabilities, posing challenges for their reliable deployment.

Method: STUPID employs a PID controller to dynamically modulate activation steering strength during inference. It combines a chunk-level classifier for detecting redundant reasoning patterns with a PID control mechanism that adaptively adjusts steering intensity based on predicted redundancy probability.

Result: On GSM8K, STUPID achieves 6% improvement in accuracy while reducing token usage by 32%, outperforming static steering baselines.

Conclusion: STUPID provides a principled framework for dynamic reasoning calibration that maintains reasoning quality while significantly improving computational efficiency, offering a training-free solution to the overthinking problem in LLMs.

Abstract: Large Language Models employing extended chain-of-thought (CoT) reasoning often suffer from the overthinking phenomenon, generating excessive and redundant reasoning steps that increase computational costs while potentially degrading performance. While recent work has explored static steering approaches to mitigate this issue, they lack the adaptability to dynamically adjust intervention strength based on real-time reasoning quality. We propose STUPID (Steering Token Usage via PID controller), a novel training-free method that employs a PID controller to dynamically modulate activation steering strength during inference. Our approach combines a chunk-level classifier for detecting redundant reasoning patterns with a PID control mechanism that adaptively adjusts steering intensity based on the predicted redundancy probability. Experimental evaluation on GSM8K demonstrates that STUPID achieves a 6% improvement in accuracy while reducing token usage by 32%, outperforming static steering baselines. Our method provides a principled framework for dynamic reasoning calibration that maintains reasoning quality while significantly improving computational efficiency.

Method: Used five LLMs and four RLLMs to assess effects of prompting methods, agentic frameworks, and multilingual alignment methods on financial question-answering tasks through systematic evaluation.

Result: (1) Prompting methods and agent frameworks enhance LLM performance by simulating Long CoT; (2) RLLMs’ inherent Long CoT capabilities limit conventional methods’ effectiveness; (3) Multilingual alignment methods improve LLM performance via reasoning length extension but offer minimal benefits for RLLMs.

Conclusion: The study provides important insights into performance enhancement strategies for LLMs and RLLMs in financial QA, highlighting different optimization approaches needed for LLMs vs. RLLMs due to their inherent reasoning capabilities.

Abstract: Recently, large language models (LLMs) and reasoning large language models (RLLMs) have gained considerable attention from many researchers. RLLMs enhance the reasoning capabilities of LLMs through Long Chain-of-Thought (Long CoT) processes, significantly improving the performance of LLMs in addressing complex problems. However, there are few works that systematically explore what methods can fully unlock the performance of LLMs and RLLMs within the financial domain. To investigate the impact of various methods on LLMs and RLLMs, we utilize five LLMs and four RLLMs to assess the effects of prompting methods, agentic frameworks, and multilingual alignment methods on financial question-answering tasks. Our research findings indicate: (1) Current prompting methods and agent frameworks enhance the performance of LLMs in financial question answering by simulating Long CoT; (2) RLLMs possess inherent Long CoT capabilities, which limits the effectiveness of conventional methods in further enhancing their performance; (3) Current advanced multilingual alignment methods primarily improve the multilingual performance of LLMs by extending the reasoning length, which yields minimal benefits for RLLMs. Additionally, we discuss strategies for enhancing the performance of LLMs and RLLMs in financial question answering, which may serve as a inspiration for future improvements. We hope that this study can serve as an important reference for LLMs and RLLMs in the field of financial question answering.

Method: AttnTrace uses attention weights produced by LLMs for prompts, with two novel techniques to enhance effectiveness: 1) attention weight processing methods, and 2) theoretical design choices that optimize traceback accuracy. The method operates through an attribution-before-detection paradigm.

Result: AttnTrace outperforms state-of-the-art methods like TracLLM in both accuracy and efficiency (significantly faster than TracLLM’s hundreds of seconds per traceback). It also improves prompt injection detection in long contexts and can effectively identify injected instructions in manipulated documents.

Conclusion: AttnTrace provides an effective and efficient solution for context traceback in LLMs, with practical applications in security, interpretability, and trustworthiness of LLM systems operating on long contexts.

Abstract: Long-context large language models (LLMs), such as Gemini-2.5-Pro and Claude-Sonnet-4, are increasingly used to empower advanced AI systems, including retrieval-augmented generation (RAG) pipelines and autonomous agents. In these systems, an LLM receives an instruction along with a context–often consisting of texts retrieved from a knowledge database or memory–and generates a response that is contextually grounded by following the instruction. Recent studies have designed solutions to trace back to a subset of texts in the context that contributes most to the response generated by the LLM. These solutions have numerous real-world applications, including performing post-attack forensic analysis and improving the interpretability and trustworthiness of LLM outputs. While significant efforts have been made, state-of-the-art solutions such as TracLLM often lead to a high computation cost, e.g., it takes TracLLM hundreds of seconds to perform traceback for a single response-context pair. In this work, we propose AttnTrace, a new context traceback method based on the attention weights produced by an LLM for a prompt. To effectively utilize attention weights, we introduce two techniques designed to enhance the effectiveness of AttnTrace, and we provide theoretical insights for our design choice. We also perform a systematic evaluation for AttnTrace. The results demonstrate that AttnTrace is more accurate and efficient than existing state-of-the-art context traceback methods. We also show that AttnTrace can improve state-of-the-art methods in detecting prompt injection under long contexts through the attribution-before-detection paradigm. As a real-world application, we demonstrate that AttnTrace can effectively pinpoint injected instructions in a paper designed to manipulate LLM-generated reviews. The code is at https://github.com/Wang-Yanting/AttnTrace.

Method: ZARA distills reference data into a statistically grounded textual knowledge base that transforms signal patterns into verifiable natural-language priors. It uses retrieval-augmented agents to iteratively select discriminative cues and perform grounded reasoning over candidate activities without training.

Result: Extensive experiments on eight benchmarks show ZARA generalizes robustly to unseen subjects and across datasets, demonstrating strong transferability across heterogeneous sensor domains.

Conclusion: ZARA represents progress toward trustworthy, plug-and-play motion understanding that goes beyond dataset-specific artifacts, enabling zero-training adaptation to new activities.

Abstract: Motion sensor time-series are central to Human Activity Recognition (HAR), yet conventional approaches are constrained to fixed activity sets and typically require costly parameter retraining to adapt to new behaviors. While Large Language Models (LLMs) offer promising open-set reasoning capabilities, applying them directly to numerical time-series often leads to hallucinations and weak grounding. To address this challenge, we propose ZARA (Zero-training Activity Reasoning Agents), a knowledge- and retrieval-augmented agentic framework for motion time-series reasoning in a training-free inference setting. Rather than relying on black-box projections, ZARA distills reference data into a statistically grounded textual knowledge base that transforms implicit signal patterns into verifiable natural-language priors. Guided by retrieved evidence, ZARA iteratively selects discriminative cues and performs grounded reasoning over candidate activities. Extensive experiments on eight benchmarks show that ZARA generalizes robustly to unseen subjects and across datasets, demonstrating strong transferability across heterogeneous sensor domains. These results mark a step toward trustworthy, plug-and-play motion understanding beyond dataset-specific artifacts. Our code is available at https://github.com/zechenli03/ZARA.

Method: Analyzed over 40,000 news articles from major, local, and college news media using three advanced AI-text detectors (Binoculars, Fast-Detect GPT, GPTZero) with sentence-level and linguistic analysis.

Result: Found substantial increase in GenAI use in recent years, especially in local and college news. LLMs often used in introductions while conclusions written manually. GenAI boosts word richness and readability but lowers formality, leading to more uniform writing styles in local media.

Conclusion: GenAI is increasingly used in journalism, particularly in local/college media, affecting writing patterns and potentially journalistic integrity, with AI more prevalent in article introductions than conclusions.

Abstract: The rapid rise of Generative AI (GenAI), particularly LLMs, poses concerns for journalistic integrity and authorship. This study examines AI-generated content across over 40,000 news articles from major, local, and college news media, in various media formats. Using three advanced AI-text detectors (e.g., Binoculars, Fast-Detect GPT, and GPTZero), we find substantial increase of GenAI use in recent years, especially in local and college news. Sentence-level analysis reveals LLMs are often used in the introduction of news, while conclusions usually written manually. Linguistic analysis shows GenAI boosts word richness and readability but lowers formality, leading to more uniform writing styles, particularly in local media.

Method: Mechanistic analysis reveals LLMs follow distinct “constellation” patterns in embedding space with consistent trajectories across NLP tasks. SafeConstellations tracks these task-specific trajectory patterns at inference time and guides representations toward non-refusal pathways using trajectory-shifting techniques.

Result: The method reduces over-refusal rates by up to 73% with minimal impact on utility, offering a principled and conditional approach that selectively guides model behavior only on tasks prone to over-refusal.

Conclusion: SafeConstellations provides an effective inference-time solution for mitigating LLM over-refusal by leveraging task-specific trajectory patterns in embedding space, maintaining safety while improving utility for production applications.

Abstract: LLMs increasingly exhibit over-refusal behavior, where safety mechanisms cause models to reject benign instructions that seemingly resemble harmful content. This phenomenon diminishes utility in production applications that repeatedly rely on common prompt templates or applications that frequently rely on LLMs for specific tasks (e.g. sentiment analysis, language translation). Through extensive evaluation, we demonstrate that LLMs persist in refusing inputs containing harmful content, even when they are reframed with tasks that have benign intent. Our mechanistic analysis reveals that LLMs follow distinct “constellation” patterns in embedding space as representations traverse layers, with each NLP task maintaining consistent trajectories that shift predictably between refusal and non-refusal cases. We introduce SafeConstellations, an inference-time trajectory-shifting approach that tracks task-specific trajectory patterns and guides representations toward non-refusal pathways. By selectively guiding model behavior only on tasks prone to over-refusal, our method reduces over-refusal rates by up to 73% with minimal impact on utility – offering a principled and conditional approach to mitigating over-refusals.

Method: Knowledge Composition Sampling (KCS) models knowledge composition selection as sentence-level conditional prediction, uses probabilistic contrastive loss to predict next relevant knowledge, and employs stochastic decoding for inference.

Result: KCS improves knowledge composition selection accuracy by 3.9% and yields improvements on HotpotQA and 2WikiMultihopQA datasets when used for data augmentation.

Conclusion: KCS effectively expands diversity of multi-hop questions by sampling varied knowledge compositions, addressing data sparsity issues in multi-hop QA.

Abstract: Multi-hop question answering faces substantial challenges due to data sparsity, which increases the likelihood of language models learning spurious patterns. To address this issue, prior research has focused on diversifying question generation through content planning and varied expression. However, these approaches often emphasize generating simple questions and neglect the integration of essential knowledge, such as relevant sentences within documents. This paper introduces the Knowledge Composition Sampling (KCS), an innovative framework designed to expand the diversity of generated multi-hop questions by sampling varied knowledge compositions within a given context. KCS models the knowledge composition selection as a sentence-level conditional prediction task and utilizes a probabilistic contrastive loss to predict the next most relevant piece of knowledge. During inference, we employ a stochastic decoding strategy to effectively balance accuracy and diversity. Compared to competitive baselines, our KCS improves the overall accuracy of knowledge composition selection by 3.9%, and its application for data augmentation yields improvements on HotpotQA and 2WikiMultihopQA datasets. Our code is available at: https://github.com/yangfanww/kcs.

Method: Use GPT-4o and o3-mini as one-time teachers to generate 21,543 synthetic Q&A pairs from 3.2 billion AIS vessel tracking records. Fine-tune smaller Qwen2.5-7B model on this synthetic dataset rather than using large models directly for inference.

Result: Achieved 261x cost reduction for maritime intelligence. Fine-tuned Qwen2.5-7B achieves 75% accuracy on maritime tasks, comparable to larger models but substantially cheaper. Framework prevents overfitting and ensures accurate reasoning.

Conclusion: Smaller, properly fine-tuned models can match larger model accuracy at dramatically lower costs. Approach enables specialized AI applications where manual annotation is infeasible, with applications in maritime safety, security, and traffic management.

Abstract: Large Language Models (LLMs) have demonstrated remarkable capabilities across many domains, yet their application to specialized fields remains constrained by the scarcity and complexity of domain-specific training data. We present a novel approach that achieves a 261x cost reduction for maritime intelligence by using LLMs as one-time teachers rather than using them directly for inference. Our method transforms 3.2 billion Automatic Identification System (AIS) vessel tracking records into 21,543 synthetic question and answer pairs through multi-model generation (GPT-4o and o3-mini), preventing overfitting and ensuring accurate reasoning. The resulting fine-tuned Qwen2.5-7B model achieves 75% accuracy on maritime tasks, while being substantially cheaper than using a larger model for inference. We show that smaller, cheaper models – when fine tuned properly – can provide similar accuracy compared to larger models that are prohibitively expensive. Our work contributes to the growing field of synthetic dataset generation for specialized AI applications and presents a highly reproducible framework for domains where manual annotation is infeasible. Beyond expanding research in the growing field of specialized small language models, our approach has immediate applications in maritime safety, security operations, and vessel traffic management systems in various industries.

Method: Introduces FS-DFM (Few-Step Discrete Flow-Matching) with three key components: 1) Makes number of sampling steps an explicit parameter and trains for consistency across step budgets, 2) Reliable update rule that moves probability without overshooting, 3) Strong teacher guidance distilled from long-run trajectories.

Result: On language modeling benchmarks, FS-DFM with 8 sampling steps achieves perplexity parity with 1,024-step discrete-flow baseline for generating 1,024 tokens using similar-size model, delivering up to 128 times faster sampling and corresponding latency/throughput gains.

Conclusion: FS-DFM enables efficient few-step discrete flow-matching for language generation, achieving quality parity with long-run baselines while dramatically improving speed, making parallel language generation more practical.

Abstract: Autoregressive language models (ARMs) deliver strong likelihoods, but are inherently serial: they generate one token per forward pass, which limits throughput and inflates latency for long sequences. Diffusion Language Models (DLMs) parallelize across positions and thus appear promising for language generation, yet standard discrete diffusion typically needs hundreds to thousands of model evaluations to reach high quality, trading serial depth for iterative breadth. We introduce FS-DFM, Few-Step Discrete Flow-Matching. A discrete flow-matching model designed for speed without sacrificing quality. The core idea is simple: make the number of sampling steps an explicit parameter and train the model to be consistent across step budgets, so one big move lands where many small moves would. We pair this with a reliable update rule that moves probability in the right direction without overshooting, and with strong teacher guidance distilled from long-run trajectories. Together, these choices make few-step sampling stable, accurate, and easy to control. On language modeling benchmarks, FS-DFM with 8 sampling steps achieves perplexity parity with a 1,024-step discrete-flow baseline for generating 1,024 tokens using a similar-size model, delivering up to 128 times faster sampling and corresponding latency/throughput gains. Code & pretrained checkpoints: https://github.com/apple/ml-fs-dfm

Method: The authors analyze the interaction between causal self-attention and other architectural components, specifically examining how stacked causal self-attention layers combined with LayerNorm induce recency bias. They also study the effects of residual connections and input token embedding distributions on this bias.

Result: The analysis reveals that the combination of stacked causal self-attention layers with LayerNorm is responsible for inducing recency bias in Transformer decoders, providing new theoretical insights into positional information dynamics.

Conclusion: The findings provide theoretical insights into how positional information interacts with architectural components in Transformers and suggest directions for improving positional encoding strategies.

Abstract: Causal self-attention provides positional information to Transformer decoders. Prior work has shown that stacks of causal self-attention layers alone induce a positional bias in attention scores toward earlier tokens. However, this differs from the bias toward later tokens typically observed in Transformer decoders, known as recency bias. We address this discrepancy by analyzing the interaction between causal self-attention and other architectural components. We show that stacked causal self-attention layers combined with LayerNorm induce recency bias. Furthermore, we examine the effects of residual connections and the distribution of input token embeddings on this bias. Our results provide new theoretical insights into how positional information interacts with architectural components and suggest directions for improving positional encoding strategies.

Method: Created RedNote-Vibe dataset spanning 5 years from RedNote platform with comprehensive engagement metrics. Proposed Psycholinguistic AIGT Detection Framework (PLAD) grounded in cognitive psychology, leveraging deep psychological signatures for interpretable AI-generated text detection.

Result: PLAD shows superior detection performance. Key findings: human content outperforms AI in emotional resonance; AI content is more homogeneous and rarely produces breaking posts; human-AI gap narrows for high-investment interactions; small group of users strategically using AI tools achieve higher engagement.

Conclusion: The dataset enables temporal analysis of content dynamics, and PLAD provides robust, interpretable AI detection. Strategic AI use by some users can enhance engagement, but human content remains superior in emotional domains.

Abstract: We introduce RedNote-Vibe, a dataset spanning five years (pre-LLM to July 2025) sourced from lifestyle platform RedNote (Xiaohongshu), capturing the temporal dynamics of content creation and is enriched with comprehensive engagement metrics. To address the detection challenge posed by RedNote-Vibe, we propose the \textbf{PsychoLinguistic AIGT Detection Framework (PLAD)}. Grounded in cognitive psychology, PLAD leverages deep psychological signatures for robust and interpretable detection. Our experiments demonstrate PLAD’s superior performance and reveal insights into content dynamics: (1) human content continues to outperform AI in emotionally resonant domains; (2) AI content is more homogeneous and rarely produces breaking posts, however, this human-AI gap narrows for arousing higher-investment interactions; and (3) most interestingly, a small group of users who strategically utilize AI tools can achieve higher engagement outcomes. The dataset is available at https://github.com/ydli-ai/RedNote-Vibe

Method: StableToken uses a multi-branch architecture that processes audio in parallel, then merges representations via a powerful bit-wise voting mechanism to form a single, stable token sequence. This consensus-driven approach addresses the limitations of single-path quantization.

Result: StableToken achieves state-of-the-art token stability, drastically reducing Unit Edit Distance (UED) under diverse noise conditions. This foundational stability translates to significant improvements in SpeechLLM robustness across various tasks.

Conclusion: The proposed StableToken tokenizer successfully addresses the fragility of semantic speech tokenizers through its consensus-driven architecture, providing stable token sequences that enhance downstream SpeechLLM performance and robustness.

Abstract: Prevalent semantic speech tokenizers, designed to capture linguistic content, are surprisingly fragile. We find they are not robust to meaning-irrelevant acoustic perturbations; even at high Signal-to-Noise Ratios (SNRs) where speech is perfectly intelligible, their output token sequences can change drastically, increasing the learning burden for downstream LLMs. This instability stems from two flaws: a brittle single-path quantization architecture and a distant training signal indifferent to intermediate token stability. To address this, we introduce StableToken, a tokenizer that achieves stability through a consensus-driven mechanism. Its multi-branch architecture processes audio in parallel, and these representations are merged via a powerful bit-wise voting mechanism to form a single, stable token sequence. StableToken sets a new state-of-the-art in token stability, drastically reducing Unit Edit Distance (UED) under diverse noise conditions. This foundational stability translates directly to downstream benefits, significantly improving the robustness of SpeechLLMs on a variety of tasks. Our code and model are publicly available at https://github.com/Tencent/StableToken.

Method: Introduces ChatInject attack that formats malicious payloads to mimic native chat templates, exploiting models’ instruction-following tendencies. Develops a persuasion-driven Multi-turn variant that primes agents across conversational turns to accept suspicious actions. Tests across frontier LLMs on AgentDojo and InjecAgent benchmarks.

Result: ChatInject achieves significantly higher average attack success rates: 5.18% to 32.05% on AgentDojo and 15.13% to 45.90% on InjecAgent. Multi-turn dialogues show particularly strong performance at average 52.33% success rate on InjecAgent. Chat-template-based payloads demonstrate strong transferability across models and remain effective against closed-source LLMs. Existing prompt-based defenses are largely ineffective.

Conclusion: The paper reveals critical vulnerabilities in current agent systems, highlighting how chat template dependencies and multi-turn persuasion can be exploited for indirect prompt injection attacks, with existing defenses proving inadequate against these sophisticated approaches.

Abstract: The growing deployment of large language model (LLM) based agents that interact with external environments has created new attack surfaces for adversarial manipulation. One major threat is indirect prompt injection, where attackers embed malicious instructions in external environment output, causing agents to interpret and execute them as if they were legitimate prompts. While previous research has focused primarily on plain-text injection attacks, we find a significant yet underexplored vulnerability: LLMs’ dependence on structured chat templates and their susceptibility to contextual manipulation through persuasive multi-turn dialogues. To this end, we introduce ChatInject, an attack that formats malicious payloads to mimic native chat templates, thereby exploiting the model’s inherent instruction-following tendencies. Building on this foundation, we develop a persuasion-driven Multi-turn variant that primes the agent across conversational turns to accept and execute otherwise suspicious actions. Through comprehensive experiments across frontier LLMs, we demonstrate three critical findings: (1) ChatInject achieves significantly higher average attack success rates than traditional prompt injection methods, improving from 5.18% to 32.05% on AgentDojo and from 15.13% to 45.90% on InjecAgent, with multi-turn dialogues showing particularly strong performance at average 52.33% success rate on InjecAgent, (2) chat-template-based payloads demonstrate strong transferability across models and remain effective even against closed-source LLMs, despite their unknown template structures, and (3) existing prompt-based defenses are largely ineffective against this attack approach, especially against Multi-turn variants. These findings highlight vulnerabilities in current agent systems.

Method: Introduces ToMAgent (ToMA), a ToM-focused dialogue agent trained by pairing Theory of Mind with dialogue lookahead to produce mental states maximally useful for achieving dialogue goals. First shows that simply prompting models to generate mental states between dialogue turns provides benefits, then develops the full ToMA approach.

Result: Experiments on Sotopia interactive social evaluation benchmark show ToMA outperforms baselines. Comprehensive analysis reveals ToMA exhibits more strategic, goal-oriented reasoning behaviors, enables long-horizon adaptation, and maintains better relationships with dialogue partners.

Conclusion: Explicit integration of Theory of Mind significantly improves LLM-based social agents’ dialogue performance and goal achievement, representing a step forward in building socially intelligent LLM agents.

Abstract: Theory of Mind (ToM)-an understanding of the mental states of others-is a key aspect of human social intelligence, yet, chatbots and LLM-based social agents do not typically integrate it. In this work, we demonstrate that LLMs that explicitly use ToM get better at dialogue, achieving goals more effectively. After showing that simply prompting models to generate mental states between dialogue turns already provides significant benefit, we further introduce ToMAgent (ToMA), a ToM-focused dialogue agent. ToMA is trained by pairing ToM with dialogue lookahead to produce mental states that are maximally useful for achieving dialogue goals. Experiments on the Sotopia interactive social evaluation benchmark demonstrate the effectiveness of our method over a range of baselines. Comprehensive analysis shows that ToMA exhibits more strategic, goal-oriented reasoning behaviors, which enable long-horizon adaptation, while maintaining better relationships with their partners. Our results suggest a step forward in integrating ToM for building socially intelligent LLM agents.

Method: MASH uses reinforcement learning with a pay-per-search reward structure. The framework penalizes external help-seeking (search tool use) while rewarding answer accuracy. This teaches LLMs to only use search tools when they lack parametric knowledge, effectively aligning search tool use with knowledge boundaries.

Result: On three knowledge-intensive QA datasets, MASH substantially improves selective help-seeking performance over prior efficient search approaches. On multi-hop datasets, it improves answer accuracy by 7.6%. It also demonstrates strong off-the-shelf abstention performance competitive with prior methods that require predetermined knowledge boundaries.

Conclusion: MASH effectively aligns search tool use with parametric knowledge boundaries through reinforcement learning. This approach enables LLMs to make better abstention decisions and use search tools more efficiently without requiring explicit knowledge boundary determination for training data construction.

Abstract: LLMs cannot reliably recognize their parametric knowledge boundaries and often hallucinate answers to outside-of-boundary questions. In this paper, we introduce MASH (Modeling Abstention via Selective Help-seeking), a training framework that readily extracts abstentions from LLMs. Our key idea is that any external help-seeking by an LLM, i.e. search tool use, can serve as a proxy for abstention if the external help (search) is appropriately penalized while also rewarding answer accuracy. MASH operationalizes this idea using reinforcement learning with a pay-per-search reward. We run experiments on three knowledge-intensive QA datasets. Our results show that MASH substantially improves upon the selective help-seeking performance of prior efficient search approaches; on multi-hop datasets, it improves answer accuracy by 7.6%. Furthermore, MASH demonstrates strong off-the-shelf abstention performance, showcasing behavior competitive with prior abstention methods that additionally require predetermining model knowledge boundaries to construct training data. Overall, we show MASH training effectively aligns search tool use with parametric knowledge, which can be successfully leveraged for making abstention decisions and efficient search tool use

Method: IDIOMoE treats item interaction histories as a native dialect within language space. It splits the Feed Forward Network of each block of a pretrained LLM into separate text and item experts with token-type gating, avoiding destructive interference between modalities while enabling collaborative signals to be understood like natural language.

Result: IDIOMoE demonstrates strong recommendation performance across both public and proprietary datasets while preserving the text understanding capabilities of the pretrained LLM.

Conclusion: The paper successfully unifies collaborative filtering and LLMs by treating item interactions as a native dialect, enabling effective multimodal recommendation systems that handle both text and item modalities without interference.

Abstract: While collaborative filtering delivers predictive accuracy and efficiency, and Large Language Models (LLMs) enable expressive and generalizable reasoning, modern recommendation systems must bring these strengths together. Growing user expectations, such as natural-language queries and transparent explanations, further highlight the need for a unified approach. However, doing so is nontrivial. Collaborative signals are often token-efficient but semantically opaque, while LLMs are semantically rich but struggle to model implicit user preferences when trained only on textual inputs. This paper introduces Item-ID + Oral-language Mixture-of-Experts Language Model (IDIOMoE), which treats item interaction histories as a native dialect within the language space, enabling collaborative signals to be understood in the same way as natural language. By splitting the Feed Forward Network of each block of a pretrained LLM into a separate text expert and an item expert with token-type gating, our method avoids destructive interference between text and catalog modalities. IDIOMoE demonstrates strong recommendation performance across both public and proprietary datasets, while preserving the text understanding of the pretrained model.

Method: EEPO uses two-stage rollouts with adaptive unlearning: first stage generates half the trajectories, then lightweight unlearning temporarily suppresses these sampled responses, forcing second stage to explore different regions of output space. This sample-then-forget mechanism disrupts the self-reinforcing loop.

Result: Across five reasoning benchmarks, EEPO outperforms GRPO with average relative gains of 24.3% on Qwen2.5-3B, 33.0% on Llama3.2-3B-Instruct, and 10.4% on Qwen3-8B-Base.

Conclusion: EEPO effectively addresses exploration-exploitation trade-off in RL for LLMs through its novel two-stage rollout with adaptive unlearning, demonstrating significant performance improvements on reasoning tasks.

Abstract: Balancing exploration and exploitation remains a central challenge in reinforcement learning with verifiable rewards (RLVR) for large language models (LLMs). Current RLVR methods often overemphasize exploitation, leading to entropy collapse, diminished exploratory capacity, and ultimately limited performance gains. Although techniques that increase policy stochasticity can promote exploration, they frequently fail to escape dominant behavioral modes. This creates a self-reinforcing loop – repeatedly sampling and rewarding dominant modes – that further erodes exploration. We introduce Exploration-Enhanced Policy Optimization (EEPO), a framework that promotes exploration via two-stage rollouts with adaptive unlearning. In the first stage, the model generates half of the trajectories; it then undergoes a lightweight unlearning step to temporarily suppress these sampled responses, forcing the second stage to explore different regions of the output space. This sample-then-forget mechanism disrupts the self-reinforcing loop and promotes wider exploration during rollouts. Across five reasoning benchmarks, EEPO outperforms GRPO, achieving average relative gains of 24.3% on Qwen2.5-3B, 33.0% on Llama3.2-3B-Instruct, and 10.4% on Qwen3-8B-Base.

Method: Used joint human expert-LLM judge approach to evaluate 11,000+ generated MWPs, created teacher-annotated dataset, trained 12B open model and text classifier, and conducted student study comparing customized vs human-written MWPs.

Result: 12B open model matches larger models’ performance; 30B open LLM with classifier outperforms closed baselines; generated MWPs are more similar to human-written ones; students perform similarly on both but prefer customized MWPs.

Conclusion: LLMs can effectively support math education by generating customized math word problems that align with educational standards and student preferences, with open models achieving competitive performance.

Abstract: Math word problems (MWPs) are critical K-12 educational tools, and customizing them to students’ interests and ability levels can enhance learning. However, teachers struggle to find time to customize MWPs for students given large class sizes and increasing burnout. We propose that LLMs can support math education by generating MWPs customized to student interests and math education standards. We use a joint human expert-LLM judge approach to evaluate over 11,000 MWPs generated by open and closed LLMs and develop the first teacher-annotated dataset for standards-aligned educational MWP generation. We show the value of our data by using it to train a 12B open model that matches the performance of larger and more capable open models. We also use our teacher-annotated data to train a text classifier that enables a 30B open LLM to outperform existing closed baselines without any training. Next, we show our models’ MWPs are more similar to human-written MWPs than those from existing models. We conclude by conducting the first study of customized LLM-generated MWPs with grade school students, finding they perform similarly on our models’ MWPs relative to human-written MWPs but consistently prefer our customized MWPs.

Method: HiPRAG incorporates a knowledge-grounded process reward into RL training by decomposing reasoning trajectories into discrete steps and applying a hierarchical reward function. This evaluates search necessity on-the-fly and provides bonuses based on optimal search/non-search step proportions, complementing traditional outcome and format rewards.

Result: Experiments on Qwen2.5 and Llama-3.2 models across seven QA benchmarks show average accuracies of 65.4% (3B) and 67.2% (7B) while improving search efficiency. Over-search rate reduced to 2.3% with lower under-search rates. The method generalizes well across RL algorithms, model families, sizes, and types.

Conclusion: HiPRAG demonstrates the importance of fine-grained process optimization in RL for improving search agent efficiency and optimality, showing that optimizing the reasoning process itself (not just outcomes) leads to better performance and generalization.

Abstract: Agentic RAG is a powerful technique for incorporating external information that LLMs lack, enabling better problem solving and question answering. However, suboptimal search behaviors exist widely, such as over-search (retrieving information already known) and under-search (failing to search when necessary), which leads to unnecessary overhead and unreliable outputs. Current training methods, which typically rely on outcome-based rewards in a RL framework, lack the fine-grained control needed to address these inefficiencies. To overcome this, we introduce Hierarchical Process Rewards for Efficient agentic RAG (HiPRAG), a training methodology that incorporates a fine-grained, knowledge-grounded process reward into the RL training. Our approach evaluates the necessity of each search decision on-the-fly by decomposing the agent’s reasoning trajectory into discrete, parsable steps. We then apply a hierarchical reward function that provides an additional bonus based on the proportion of optimal search and non-search steps, on top of commonly used outcome and format rewards. Experiments on the Qwen2.5 and Llama-3.2 models across seven diverse QA benchmarks show that our method achieves average accuracies of 65.4% (3B) and 67.2% (7B). This is accomplished while improving search efficiency, reducing the over-search rate to just 2.3% and concurrently lowering the under-search rate. These results demonstrate the efficacy of optimizing the reasoning process itself, not just the final outcome. Further experiments and analysis demonstrate that HiPRAG shows good generalizability across a wide range of RL algorithms, model families, sizes, and types. This work demonstrates the importance and potential of fine-grained control through RL, for improving the efficiency and optimality of reasoning for search agents.

Method: Categorizes inductive reasoning improvement methods into three areas: post-training, test-time scaling, and data augmentation. Summarizes benchmarks and proposes unified sandbox-based evaluation with observation coverage metric.

Result: Provides comprehensive survey framework, analysis of inductive ability sources, and insights into how model architectures and data help with inductive tasks.

Conclusion: Establishes foundation for future inductive reasoning research in LLMs through systematic categorization, evaluation methodology, and analysis of inductive capability sources.

Abstract: Reasoning is an important task for large language models (LLMs). Among all the reasoning paradigms, inductive reasoning is one of the fundamental types, which is characterized by its particular-to-general thinking process and the non-uniqueness of its answers. The inductive mode is crucial for knowledge generalization and aligns better with human cognition, so it is a fundamental mode of learning, hence attracting increasing interest. Despite the importance of inductive reasoning, there is no systematic summary of it. Therefore, this paper presents the first comprehensive survey of inductive reasoning for LLMs. First, methods for improving inductive reasoning are categorized into three main areas: post-training, test-time scaling, and data augmentation. Then, current benchmarks of inductive reasoning are summarized, and a unified sandbox-based evaluation approach with the observation coverage metric is derived. Finally, we offer some analyses regarding the source of inductive ability and how simple model architectures and data help with inductive tasks, providing a solid foundation for future research.

Method: Modular pipeline with semantic chunking, hierarchical concept extraction, Bloom’s Taxonomy-guided question generation, precise answer extraction, multi-chunk retrieval, and curated distractor contexts for robust training.

Result: RAGen efficiently generates domain-grounded QAC triples for diverse RAG adaptation strategies, handles large evolving document corpora without redundant processing, and supports optimization of LLM, retriever, and embedding models.

Conclusion: RAGen provides a scalable solution for adapting RAG systems to dynamic domains like scientific research and enterprise knowledge bases through automated generation of specialized training data.

Abstract: Retrieval-Augmented Generation (RAG) combines the language understanding and reasoning power of large language models (LLMs) with external retrieval to enable domain-grounded responses. Effectively adapting RAG systems to domain-specific settings requires specialized, context-rich training data beyond general-purpose question-answering. Here, we propose RAGen, a scalable and modular framework for generating domain-grounded question-answer-context (QAC) triples tailored to diverse RAG adaptation approaches. RAGen produces these QAC triples by identifying key concepts in documents, generating diverse questions guided by Bloom’s Taxonomy-inspired principles, and pairing them with precise answers extracted from relevant contexts. RAGen supports multiple RAG adaptation strategies, including the optimization of key components such as the LLM, retriever, and embedding model, etc. Its modular pipeline features semantic chunking, hierarchical concept extraction, and multi-chunk retrieval, along with the introduction of curated distractor contexts to promote robust reasoning. Designed for scalability, RAGen efficiently handles large and evolving document corpora without redundant processing, making it especially suitable for dynamic evolving domains such as scientific research and enterprise knowledge bases.

Method: Three-stage framework: 1) Fine-grained knowledge pruning to filter irrelevant context, 2) Latent conflict probing to identify hard conflicts in model’s latent space, 3) Conflict-aware attention to modulate attention heads toward faithful context integration.

Result: Extensive experiments show ProbeRAG substantially improves both accuracy and contextual faithfulness compared to existing methods.

Conclusion: Moving beyond black-box interventions to analyze internal reasoning processes enables more faithful RAG systems by identifying and addressing knowledge conflicts at the latent representation level.

Abstract: Retrieval-Augmented Generation (RAG) systems often fail to maintain contextual faithfulness, generating responses that conflict with the provided context or fail to fully leverage the provided evidence. Existing methods attempt to improve faithfulness through external interventions, such as specialized prompting, decoding-based calibration, or preference optimization. However, since these approaches treat the LLM as a black box, they lack a reliable mechanism to assess when and why knowledge conflicts occur. Consequently, they tend to be brittle, data-intensive, and agnostic to the model’s internal reasoning process. In this paper, we move beyond black-box interventions to analyze the model’s internal reasoning process. We discover that conflicting and aligned knowledge states are linearly separable in the model’s latent space, and contextual noise systematically increases the entropy of these representations. Based on these findings, we propose ProbeRAG, a novel framework for faithful RAG that operates in three stages: (i) fine-grained knowledge pruning to filter irrelevant context, (ii) latent conflict probing to identify hard conflicts in the model’s latent space, and (iii) conflict-aware attention to modulate attention heads toward faithful context integration. Extensive experiments demonstrate that ProbeRAG substantially improves both accuracy and contextual faithfulness. The related resources are available at https://github.com/LinfengGao/ProbeRAG.

Method: Introduces ParallaxRAG, a symmetric multi-view framework that decouples queries and KGs into aligned, head-specific semantic spaces. Enforces relational diversity across multiple attention heads while constraining weakly related paths to construct cleaner subgraphs and guide LLMs through hop-wise reasoning.

Result: Achieves state-of-the-art retrieval and QA performance on WebQSP and CWQ benchmarks, substantially reduces hallucination, and generalizes strongly to the biomedical BioASQ benchmark.

Conclusion: ParallaxRAG successfully addresses the structural limitations of Transformer attention heads for multi-hop reasoning by leveraging their natural specialization patterns, leading to improved KG-based reasoning performance and reduced hallucinations.

Abstract: Large language models (LLMs) still struggle with multi-hop reasoning over knowledge-graphs (KGs), and we identify a previously overlooked structural reason for this difficulty: Transformer attention heads naturally specialize in distinct semantic relations across reasoning stages, forming a hop-aligned relay pattern. This key finding suggests that multi-hop reasoning is inherently multi-view, yet existing KG-based retrieval-augmented generation (KG-RAG) systems collapse all reasoning hops into a single representation, flat embedding space, suppressing this implicit structure and causing noisy or drifted path exploration. We introduce ParallaxRAG, a symmetric multi-view framework that decouples queries and KGs into aligned, head-specific semantic spaces. By enforcing relational diversity across multiple heads while constraining weakly related paths, ParallaxRAG constructs more accurate, cleaner subgraphs and guides LLMs through grounded, hop-wise reasoning. On WebQSP and CWQ, it achieves state-of-the-art retrieval and QA performance, substantially reduces hallucination, and generalizes strongly to the biomedical BioASQ benchmark.

Method: Introduced SimBench, a large-scale benchmark unifying 20 diverse datasets covering tasks from moral decision-making to economic choice across global participant pools. Evaluates simulation fidelity systematically.

Result: Best LLMs achieve modest simulation fidelity (40.80/100). Performance scales log-linearly with model size but not with inference-time compute. Discovered alignment-simulation tradeoff: instruction tuning helps on consensus questions but harms on diverse ones. Models struggle with specific demographic groups. Simulation ability correlates strongly with knowledge-intensive reasoning (MMLU-Pro, r = 0.939).

Conclusion: SimBench provides foundation for measuring LLM simulation fidelity systematically. Current models have meaningful but limited ability to simulate human behavior, with specific limitations in demographic representation and diverse scenarios.

Abstract: Large language model (LLM) simulations of human behavior have the potential to revolutionize the social and behavioral sciences, if and only if they faithfully reflect real human behaviors. Current evaluations of simulation fidelity are fragmented, based on bespoke tasks and metrics, creating a patchwork of incomparable results. To address this, we introduce SimBench, the first large-scale, standardized benchmark for a robust, reproducible science of LLM simulation. By unifying 20 diverse datasets covering tasks from moral decision-making to economic choice across a large global participant pool, SimBench provides the necessary foundation to ask fundamental questions about when, how, and why LLM simulations succeed or fail. We show that the best LLMs today achieve meaningful but modest simulation fidelity (score: 40.80/100), with performance scaling log-linearly with model size but not with increased inference-time compute. We discover an alignment-simulation tradeoff: instruction tuning improves performance on low-entropy (consensus) questions but degrades it on high-entropy (diverse) ones. Models particularly struggle when simulating specific demographic groups. Finally, we demonstrate that simulation ability correlates most strongly with knowledge-intensive reasoning (MMLU-Pro, r = 0.939). By making progress measurable, we aim to accelerate the development of more faithful LLM simulators.

Method: Proposes AtlasKV with two components: KG2KV to integrate KG triples into LLMs at scale, and HiKVP (hierarchical key-value pairs) to achieve sub-linear time and memory complexity. Uses LLMs’ inherent attention mechanism without external retrievers.

Result: Achieves scalable knowledge integration with billion-scale KGs (e.g., 1B triples) using very little GPU memory (less than 20GB VRAM), maintains strong knowledge grounding and generalization, and requires no retraining for new knowledge.

Conclusion: AtlasKV provides an effective parametric alternative to RAG for large-scale knowledge integration into LLMs, addressing latency and memory issues while maintaining performance.

Abstract: Retrieval-augmented generation (RAG) has shown some success in augmenting large language models (LLMs) with external knowledge. However, as a non-parametric knowledge integration paradigm for LLMs, RAG methods heavily rely on external retrieval modules and the retrieved textual context prior. Especially for very large scale knowledge augmentation, they would introduce substantial inference latency due to expensive searches and much longer relevant context. In this paper, we propose a parametric knowledge integration method, called \textbf{AtlasKV}, a scalable, effective, and general way to augment LLMs with billion-scale knowledge graphs (KGs) (e.g. 1B triples) using very little GPU memory cost (e.g. less than 20GB VRAM). In AtlasKV, we introduce KG2KV and HiKVP to integrate KG triples into LLMs at scale with sub-linear time and memory complexity. It maintains strong knowledge grounding and generalization performance using the LLMs’ inherent attention mechanism, and requires no external retrievers, long context priors, or retraining when adapting to new knowledge.

Method: WIMHF uses sparse autoencoders to explain feedback data, characterizing both the preferences a dataset can measure and what annotators actually express. It identifies human-interpretable features across 7 datasets.

Result: WIMHF identifies a small number of interpretable features accounting for most preference prediction signal. Reveals diverse human preferences across contexts (e.g., Reddit users prefer informality/jokes, while HH-RLHF/PRISM annotators disprefer them). Surfaces unsafe preferences like LMArena users voting against refusals in favor of toxic content.

Conclusion: WIMHF enables effective data curation (37% safety gains on Arena) and fine-grained personalization (annotator-specific weights improve prediction). Provides human-centered analysis for better understanding and using preference data.

Abstract: Human feedback can alter language models in unpredictable and undesirable ways, as practitioners lack a clear understanding of what feedback data encodes. While prior work studies preferences over certain attributes (e.g., length or sycophancy), automatically extracting relevant features without pre-specifying hypotheses remains challenging. We introduce What’s In My Human Feedback? (WIMHF), a method to explain feedback data using sparse autoencoders. WIMHF characterizes both (1) the preferences a dataset is capable of measuring and (2) the preferences that the annotators actually express. Across 7 datasets, WIMHF identifies a small number of human-interpretable features that account for the majority of the preference prediction signal achieved by black-box models. These features reveal a wide diversity in what humans prefer, and the role of dataset-level context: for example, users on Reddit prefer informality and jokes, while annotators in HH-RLHF and PRISM disprefer them. WIMHF also surfaces potentially unsafe preferences, such as that LMArena users tend to vote against refusals, often in favor of toxic content. The learned features enable effective data curation: re-labeling the harmful examples in Arena yields large safety gains (+37%) with no cost to general performance. They also allow fine-grained personalization: on the Community Alignment dataset, we learn annotator-specific weights over subjective features that improve preference prediction. WIMHF provides a human-centered analysis method for practitioners to better understand and use preference data.

Method: The authors first demonstrate that the gap stems from understanding failures where models can’t properly translate multilingual inputs to English (the dominant reasoning language). They evaluate various detection methods for these understanding failures, finding supervised approaches work best. They then propose Selective Translation, which incorporates English translations into reasoning traces only when understanding failures are detected.

Result: Experimental results with Qwen3-4B show that Selective Translation substantially bridges the multilingual reasoning gap, achieving near full-translation performance while translating only about 20% of inputs. Understanding failures were found to be detectable to a meaningful extent.

Conclusion: Failures in language understanding are the primary driver of the multilingual reasoning gap, and these can be detected and selectively mitigated. This clarifies the origin of the gap and suggests a path toward more equitable multilingual reasoning.

Abstract: Reasoning language models (RLMs) achieve strong performance on complex reasoning tasks, yet they still exhibit a multilingual reasoning gap, performing better in high-resource languages than in low-resource ones. While recent efforts have been made to address this gap, its underlying causes remain largely unexplored. In this work, we show that this gap primarily stems from failures in language understanding-specifically, the model’s inability to translate multilingual inputs into the language dominating its reasoning traces (typically English). As identifying understanding failures can enable targeted mitigation of the gap, we evaluate a range of detection methods and find that understanding failures are detectable to a meaningful extent, with supervised approaches performing best. Building on this, we propose Selective Translation, a strategy that incorporates an English translation into the initial reasoning trace only when an understanding failure is detected. Experimental results using Qwen3-4B show that Selective Translation substantially bridges the multilingual reasoning gap, achieving near full-translation performance while translating only about 20% of inputs. Together, our results show that failures in language understanding are the primary driver of the multilingual reasoning gap and can be detected and selectively mitigated, clarifying its origin and suggesting a path toward more equitable multilingual reasoning. Our code and data are publicly available at https://github.com/deokhk/RLM_analysis

Method: The pipeline identifies self-contained, time-sensitive knowledge entities from real-world domains, filters them based on temporal occurrence, verifies them against model knowledge, generates factual questions from entity documents, and translates them into multiple languages to evaluate transferability across linguistic boundaries.

Result: Evaluation of several LLMs across five languages shows that cross-lingual transfer is strongly influenced by linguistic distance and often asymmetric across language directions. Larger models improve transfer, but gains diminish with scale and vary across domains.

Conclusion: LiveCLKTBench provides a reliable benchmark for evaluating cross-lingual knowledge transfer, revealing important patterns about linguistic distance and asymmetry in transfer, with implications for multilingual model development.

Abstract: Evaluating cross-lingual knowledge transfer in large language models is challenging, as correct answers in a target language may arise either from genuine transfer or from prior exposure during pre-training. We present LiveCLKTBench, an automated generation pipeline specifically designed to isolate and measure cross-lingual knowledge transfer. Our pipeline identifies self-contained, time-sensitive knowledge entities from real-world domains, filters them based on temporal occurrence, and verifies them against the model’s knowledge. The documents of these valid entities are then used to generate factual questions, which are translated into multiple languages to evaluate transferability across linguistic boundaries. Using LiveCLKTBench, we evaluate several LLMs across five languages and observe that cross-lingual transfer is strongly influenced by linguistic distance and often asymmetric across language directions. While larger models improve transfer, the gains diminish with scale and vary across domains. These findings provide new insights into multilingual transfer and demonstrate the value of LiveCLKTBench as a reliable benchmark for future research.

Method: Proposes MCAT framework with two innovations: (1) language scaling method using curriculum learning and data balancing to extend coverage to 70 languages, and (2) optimized speech adapter module that reduces speech sequence length to only 30 tokens.

Result: MCAT surpasses state-of-the-art end-to-end models on FLEURS dataset across 70x69 directions and enhances inference efficiency. Tested on MLLMs of different scales (9B and 27B).

Conclusion: MCAT effectively addresses language coverage and efficiency challenges in MLLM-based speech-to-text translation, enabling many-to-many translation across 70 languages with improved inference speed.

Abstract: Multimodal Large Language Models (MLLMs) have achieved great success in Speech-to-Text Translation (S2TT) tasks. However, current research is constrained by two key challenges: language coverage and efficiency. Most of the popular S2TT datasets are substantially English-centric, which restricts the scaling-up of MLLMs’ many-to-many translation capabilities. Moreover, the inference speed of MLLMs degrades dramatically when the speech is converted into long sequences (e.g., 750 tokens). To address these limitations, we propose a Multilingual Cost-effective Accelerated Speech-to-Text Translator (MCAT) framework, which includes two innovations. First, a language scaling method that leverages curriculum learning and a data balancing strategy is introduced to extend the language coverage supported by MLLMs to 70 languages and achieve mutual translation among these languages. Second, an optimized speech adapter module is designed to reduce the length of the speech sequence to only 30 tokens. Extensive experiments were conducted on MLLMs of different scales (9B and 27B). The experimental results demonstrate that MCAT not only surpasses state-of-the-art end-to-end models on the FLEURS dataset across 70x69 directions but also enhances inference efficiency. The code and models are released at https://github.com/yxduir/m2m-70.

Method: Functional localization approach inspired by cognitive neuroscience to identify LLM units responsive to 67 English syntactic phenomena across 7 open-weight models, with cross-lingual analysis in 57 languages.

Result: Different types of syntactic agreement recruit overlapping sets of units, suggesting agreement constitutes a meaningful functional category; pattern holds across English, Russian, Chinese; structurally similar languages share more units for subject-verb agreement.

Conclusion: Syntactic agreement represents a meaningful category within LLMs’ representational spaces, revealing systematic organization of grammatical knowledge.

Abstract: Large language models (LLMs) can reliably distinguish grammatical from ungrammatical sentences, but how grammatical knowledge is represented within the models remains an open question. We investigate whether different syntactic phenomena recruit shared or distinct components in LLMs. Using a functional localization approach inspired by cognitive neuroscience, we identify the LLM units most responsive to 67 English syntactic phenomena in seven open-weight models. These units are consistently recruited across sentences containing the phenomena and causally support the models’ syntactic performance. Critically, different types of syntactic agreement (e.g., subject-verb, anaphor, determiner-noun) recruit overlapping sets of units, suggesting that agreement constitutes a meaningful functional category for LLMs. This pattern holds in English, Russian, and Chinese; and further, in a cross-lingual analysis of 57 diverse languages, structurally more similar languages share more units for subject-verb agreement. Taken together, these findings reveal that syntactic agreement-a critical marker of syntactic dependencies-constitutes a meaningful category within LLMs’ representational spaces.

Method: A model-agnostic framework that integrates LLMs’ semantic reasoning and geospatial knowledge into VPR pipelines via attention-guided descriptor enhancement. Uses LLMs to identify location-informative regions and suppress transient noise without retraining or new data.

Result: Integration with state-of-the-art VPR models (CosPlace, EigenPlaces, SALAD) consistently improved recall, yielding 1-3% relative gains and up to 8% improvement on challenging real flood imagery across San Francisco and Hong Kong datasets.

Conclusion: VPR-AttLLM bridges human-like spatial reasoning with modern VPR architectures through urban perception principles, offering a scalable plug-and-play solution for rapid geo-localization of crowdsourced crisis imagery to advance cognitive urban resilience.

Abstract: Crowdsourced social media imagery provides real-time visual evidence of urban flooding but often lacks reliable geographic metadata for emergency response. Existing Visual Place Recognition (VPR) models struggle to geo-localize these images due to cross-source domain shifts and visual distortions. We present VPR-AttLLM, a model-agnostic framework integrating the semantic reasoning and geospatial knowledge of Large Language Models (LLMs) into VPR pipelines via attention-guided descriptor enhancement. VPR-AttLLM uses LLMs to isolate location-informative regions and suppress transient noise, improving retrieval without model retraining or new data. We evaluate this framework across San Francisco and Hong Kong using established queries, synthetic flooding scenarios, and real social media flood images. Integrating VPR-AttLLM with state-of-the-art models (CosPlace, EigenPlaces, SALAD) consistently improves recall, yielding 1-3% relative gains and up to 8% on challenging real flood imagery. By embedding urban perception principles into attention mechanisms, VPR-AttLLM bridges human-like spatial reasoning with modern VPR architectures. Its plug-and-play design and cross-source robustness offer a scalable solution for rapid geo-localization of crowdsourced crisis imagery, advancing cognitive urban resilience.

Method: Proposes APF framework with innovative pipeline for automatically generating high-quality fine-tuning data without solver feedback, using data generation and test instance annotation. This dataset is used for supervised fine-tuning of LLMs to generate accurate, executable optimization problem formulations.

Result: Experimental results on antenna design show APF significantly outperforms existing methods in both requirement formalization accuracy and resulting radiation efficiency curves in meeting design goals.

Conclusion: APF provides a solver-independent framework for automated problem formulation via LLMs that effectively converts natural language requirements into optimization models, overcoming data scarcity issues in high-cost simulation domains.

Abstract: In the high-cost simulation-driven design domain, translating ambiguous design requirements into a mathematical optimization formulation is a bottleneck for optimizing product performance. This process is time-consuming and heavily reliant on expert knowledge. While large language models (LLMs) offer potential for automating this task, existing approaches either suffer from poor formalization that fails to accurately align with the design intent or rely on solver feedback for data filtering, which is unavailable due to the high simulation costs. To address this challenge, we propose APF, a framework for solver-independent, automated problem formulation via LLMs designed to automatically convert engineers’ natural language requirements into executable optimization models. The core of this framework is an innovative pipeline for automatically generating high-quality data, which overcomes the difficulty of constructing suitable fine-tuning datasets in the absence of high-cost solver feedback with the help of data generation and test instance annotation. The generated high-quality dataset is used to perform supervised fine-tuning on LLMs, significantly enhancing their ability to generate accurate and executable optimization problem formulations. Experimental results on antenna design demonstrate that APF significantly outperforms the existing methods in both the accuracy of requirement formalization and the quality of resulting radiation efficiency curves in meeting the design goals.

Method: Proposes M³KG-RAG with two components: 1) lightweight multi-agent pipeline to construct multi-hop MMKGs with context-enriched triplets, and 2) GRASP mechanism for precise entity grounding, relevance evaluation, and pruning of redundant context.

Result: Extensive experiments across diverse multimodal benchmarks show significant improvements in MLLMs’ multimodal reasoning and grounding compared to existing approaches.

Conclusion: M³KG-RAG effectively addresses limitations of current multimodal RAG by enabling precise retrieval of query-aligned audio-visual knowledge and filtering out irrelevant content, enhancing reasoning depth and answer faithfulness.

Abstract: Retrieval-Augmented Generation (RAG) has recently been extended to multimodal settings, connecting multimodal large language models (MLLMs) with vast corpora of external knowledge such as multimodal knowledge graphs (MMKGs). Despite their recent success, multimodal RAG in the audio-visual domain remains challenging due to 1) limited modality coverage and multi-hop connectivity of existing MMKGs, and 2) retrieval based solely on similarity in a shared multimodal embedding space, which fails to filter out off-topic or redundant knowledge. To address these limitations, we propose M$^3$KG-RAG, a Multi-hop Multimodal Knowledge Graph-enhanced RAG that retrieves query-aligned audio-visual knowledge from MMKGs, improving reasoning depth and answer faithfulness in MLLMs. Specifically, we devise a lightweight multi-agent pipeline to construct multi-hop MMKG (M$^3$KG), which contains context-enriched triplets of multimodal entities, enabling modality-wise retrieval based on input queries. Furthermore, we introduce GRASP (Grounded Retrieval And Selective Pruning), which ensures precise entity grounding to the query, evaluates answer-supporting relevance, and prunes redundant context to retain only knowledge essential for response generation. Extensive experiments across diverse multimodal benchmarks demonstrate that M$^3$KG-RAG significantly enhances MLLMs’ multimodal reasoning and grounding over existing approaches. Project website: https://kuai-lab.github.io/cvpr2026m3kgrag/

Method: Created CricBench benchmark suite with Gold-Standard dataset of 2,654 evaluation instances across four languages (English, Hindi, Punjabi, Telugu) and four cricket formats (Test, ODI, T20I, IPL). Evaluated seven LLMs using schema-only prompting.

Result: No single model dominates: GPT-5 Mini leads on Test cricket (12.4% DMA), Qwen 235B leads on IPL (28.7%) and T20I (17.5%), all models score 0% on hard ODI queries. Models show disconnect between syntactic validity (>98% execution accuracy) and semantic correctness (<29% DMA), with 37-55 percentage point domain gap versus BIRD benchmark.

Conclusion: CricBench is the first Text-to-SQL benchmark for cricket analytics, revealing significant challenges in domain-specific SQL generation and highlighting the need for improved sports analytics capabilities in LLMs.

Abstract: Cricket is the second most popular sport worldwide, with billions of fans seeking advanced statistical insights unavailable through standard web searches. Although LLMs have advanced significantly in Text-to-SQL tasks, their capability to handle domain-specific nuances and multilingual requirements in sports analytics remains under-explored. We present CricBench, a benchmark suite evaluating the intrinsic SQL generation abilities of LLMs on cricket data across four formats: Test, ODI, T20I, and IPL. We curate a Gold-Standard dataset of 2,654 evaluation instances across four languages (English, Hindi, Punjabi, and Telugu). We evaluate seven models, GPT-5 Mini, Claude Sonnet 4, DeepSeek R1 and V3, Qwen 235B, Llama 3.1, and Gemma 2, using schema-only prompting. No single model dominates across all formats: GPT-5 Mini leads on Test cricket (12.4% DMA), Qwen 235B leads on IPL (28.7%) and T20I (17.5%), and all models score 0% on hard ODI queries. All models show a stark disconnect between syntactic validity (>98% execution accuracy) and semantic correctness (<29% DMA), with a domain gap of 37-55 percentage points versus BIRD. To our knowledge, CricBench is the first Text-to-SQL benchmark for cricket analytics.

Method: 1) Identify three structural biases: exposure bias, gold availability prior, and cultural priors. 2) Propose DeLP metric to factor out these confounds. 3) Develop DELTA framework that strategically uses monolingual alignment for cross-lingual retrieval and generation.

Result: Analysis with DeLP shows English preference is largely due to evidence distribution, not model bias. Retrievers fundamentally favor monolingual alignment. DELTA outperforms English pivoting and mRAG baselines across diverse languages.

Conclusion: The perceived English preference in mRAG is an evaluation artifact. Monolingual alignment is key for cross-lingual retrieval. DELTA provides efficient multilingual RAG without English pivoting.

Abstract: Multilingual Retrieval-Augmented Generation (mRAG) systems often exhibit a perceived preference for high-resource languages, particularly English, resulting in the widespread adoption of English pivoting. While prior studies attribute this advantage to the superior English-centric capabilities of Large Language Models (LLMs), we find that such measurements are significantly distorted by structural priors inherent in evaluation benchmarks. Specifically, we identify exposure bias and a gold availability prior-both driven by the disproportionate concentration of resources in English-as well as cultural priors rooted in topic locality, as factors that hinder accurate assessment of genuine language preference. To address these biases, we propose DeLP (Debiased Language Preference), a calibrated metric designed to explicitly factor out these structural confounds. Our analysis using DeLP reveals that the previously reported English preference is largely a byproduct of evidence distribution rather than an inherent model bias. Instead, we find that retrievers fundamentally favor monolingual alignment between the query and the document language. Building on this insight, we introduce DELTA (DEbiased Language preference-guided Text Augmentation), a lightweight and efficient mRAG framework that strategically leverages monolingual alignment to optimize cross-lingual retrieval and generation. Experimental results demonstrate that DELTA consistently outperforms English pivoting and mRAG baselines across diverse languages.

Method: Two-stage parameter fusion framework: 1) Geometric consensus filtering to prune conflicting gradients, 2) Curvature-based importance weighting to amplify shared semantics. Uses directional consensus evaluation to disentangle knowledge updates.

Result: Extensive experiments on GUI agents and tool-use agent domains show outstanding continual learning performance with minimal computational overhead and parameter updates.

Conclusion: Agent-Dice effectively addresses the stability-plasticity dilemma in LLM-based agents by explicitly distinguishing between shared and conflicting knowledge, enabling better continual learning with theoretical guarantees.

Abstract: Large Language Model (LLM)-based agents significantly extend the utility of LLMs by interacting with dynamic environments. However, enabling agents to continually learn new tasks without catastrophic forgetting remains a critical challenge, known as the stability-plasticity dilemma. In this work, we argue that this dilemma fundamentally arises from the failure to explicitly distinguish between common knowledge shared across tasks and conflicting knowledge introduced by task-specific interference. To address this, we propose Agent-Dice, a parameter fusion framework based on directional consensus evaluation. Concretely, Agent-Dice disentangles knowledge updates through a two-stage process: geometric consensus filtering to prune conflicting gradients, and curvature-based importance weighting to amplify shared semantics. We provide a rigorous theoretical analysis that establishes the validity of the proposed fusion scheme and offers insight into the origins of the stability-plasticity dilemma. Extensive experiments on GUI agents and tool-use agent domains demonstrate that Agent-Dice exhibits outstanding continual learning performance with minimal computational overhead and parameter updates. The codes are available at https://github.com/Wuzheng02/Agent-Dice.

Method: Introduces Doc-PP benchmark from real-world reports requiring reasoning across visual/textual elements under strict policies, identifies Reasoning-Induced Safety Gap, and proposes DVA (Decompose-Verify-Aggregation) framework that decouples reasoning from policy verification

Result: Models frequently leak sensitive information when answers require complex synthesis or aggregation across modalities; extracted text improves perception but facilitates leakage; DVA significantly outperforms standard prompting defenses

Conclusion: Multimodal document safety requires specialized approaches beyond text-only methods; DVA offers robust baseline for policy-compliant document understanding by structurally separating reasoning from policy verification

Abstract: The deployment of Large Vision-Language Models (LVLMs) for real-world document question answering is often constrained by dynamic, user-defined policies that dictate information disclosure based on context. While ensuring adherence to these explicit constraints is critical, existing safety research primarily focuses on implicit social norms or text-only settings, overlooking the complexities of multimodal documents. In this paper, we introduce Doc-PP (Document Policy Preservation Benchmark), a novel benchmark constructed from real-world reports requiring reasoning across heterogeneous visual and textual elements under strict non-disclosure policies. Our evaluation highlights a systemic Reasoning-Induced Safety Gap: models frequently leak sensitive information when answers must be inferred through complex synthesis or aggregated across modalities, effectively circumventing existing safety constraints. Furthermore, we identify that providing extracted text improves perception but inadvertently facilitates leakage. To address these vulnerabilities, we propose DVA (Decompose-Verify-Aggregation), a structural inference framework that decouples reasoning from policy verification. Experimental results demonstrate that DVA significantly outperforms standard prompting defenses, offering a robust baseline for policy-compliant document understanding

Method: Constructs intra-chunk discourse trees to capture local hierarchies and builds inter-chunk rhetorical graphs to model cross-passage coherence. These structures are jointly integrated into a planning blueprint that conditions the generation process.

Result: Achieves state-of-the-art results on question answering and long-document summarization benchmarks without fine-tuning.

Conclusion: Discourse structure plays an important role in advancing RAG systems, and explicitly injecting discourse signals into generation processes significantly improves performance on knowledge-intensive tasks.

Abstract: Retrieval-Augmented Generation (RAG) has emerged as an important means of enhancing the performance of large language models (LLMs) in knowledge-intensive tasks. However, most existing RAG strategies treat retrieved passages in a flat and unstructured way, which prevents the model from capturing structural cues and constrains its ability to synthesize knowledge from dispersed evidence across documents. To overcome these limitations, we propose Disco-RAG, a discourse-aware framework that explicitly injects discourse signals into the generation process. Our method constructs intra-chunk discourse trees to capture local hierarchies and builds inter-chunk rhetorical graphs to model cross-passage coherence. These structures are jointly integrated into a planning blueprint that conditions the generation. Experiments on question answering and long-document summarization benchmarks show the efficacy of our approach. Disco-RAG achieves state-of-the-art results on the benchmarks without fine-tuning. These findings underscore the important role of discourse structure in advancing RAG systems.

Method: Two-stage framework: (1) Defensive Poisoning merges attacker and defensive triggers into unified backdoor representation, (2) Backdoor Neutralization breaks this representation through additional training to restore clean behavior.

Result: Extensive experiments show MB-Defense substantially lowers attack success rates while preserving instruction-following ability across multiple LLMs.

Conclusion: MB-Defense offers a generalizable, data-efficient defense strategy that improves robustness of instruction-tuned LLMs against unseen backdoor attacks.

Abstract: Large Language Models (LLMs) have greatly advanced Natural Language Processing (NLP), particularly through instruction tuning, which enables broad task generalization without additional fine-tuning. However, their reliance on large-scale datasets-often collected from human or web sources-makes them vulnerable to backdoor attacks, where adversaries poison a small subset of data to implant hidden behaviors. Despite this growing risk, defenses for instruction-tuned models remain underexplored. We propose MB-Defense (Merging & Breaking Defense Framework), a novel training pipeline that immunizes instruction-tuned LLMs against diverse backdoor threats. MB-Defense comprises two stages: (i) Defensive Poisoning, which merges attacker and defensive triggers into a unified backdoor representation, and (ii) Backdoor Neutralization, which breaks this representation through additional training to restore clean behavior. Extensive experiments across multiple LLMs show that MB-Defense substantially lowers attack success rates while preserving instruction-following ability. Our method offers a generalizable and data-efficient defense strategy, improving the robustness of instruction-tuned LLMs against unseen backdoor attacks.

Method: 1) Introduce Generation-time Fine-grained Provenance task requiring models to generate fluent answers with structured, sentence-level provenance triples. 2) Create ReFInE dataset with expert-verified annotations distinguishing Quotation, Compression, and Inference relations. 3) Propose GenProve framework combining Supervised Fine-Tuning with Group Relative Policy Optimization, optimizing composite reward for both answer fidelity and provenance correctness.

Result: GenProve significantly outperforms 14 strong LLMs in joint evaluation of answer quality and provenance correctness. Analysis reveals reasoning gap: models excel at surface-level quotation but struggle significantly with inference-based provenance, showing verifiable reasoning remains a distinct frontier challenge.

Conclusion: Fine-grained provenance generation is crucial for LLM accountability, with inference-based provenance being particularly challenging. The proposed framework and dataset advance verifiable reasoning capabilities beyond surface-level citation.

Abstract: Large language models (LLM) often hallucinate, and while adding citations is a common solution, it is frequently insufficient for accountability as users struggle to verify how a cited source supports a generated claim. Existing methods are typically coarse-grained and fail to distinguish between direct quotes and complex reasoning. In this paper, we introduce Generation-time Fine-grained Provenance, a task where models must generate fluent answers while simultaneously producing structured, sentence-level provenance triples. To enable this, we present ReFInE (Relation-aware Fine-grained Interpretability & Evidence), a dataset featuring expert verified annotations that distinguish between Quotation, Compression, and Inference. Building on ReFInE, we propose GenProve, a framework that combines Supervised Fine-Tuning (SFT) with Group Relative Policy Optimization (GRPO). By optimizing a composite reward for answer fidelity and provenance correctness, GenProve significantly outperforms 14 strong LLMs in joint evaluation. Crucially, our analysis uncovers a reasoning gap where models excel at surface-level quotation but struggle significantly with inference-based provenance, suggesting that verifiable reasoning remains a frontier challenge distinct from surface-level citation.

Method: FlashMem distills intrinsic memory directly from transient reasoning states via computation reuse. It identifies the last hidden state as a sufficient statistic for interaction history, uses a Shared-KV Consolidator to synthesize memory by attending to the backbone’s frozen cache, and employs a parameter-free Cognitive Monitor that leverages attention entropy to adaptively trigger consolidation only during high epistemic uncertainty.

Result: FlashMem matches the performance of heavy baselines while reducing inference latency by 5 times, effectively bridging the gap between efficiency and persistent cognition.

Conclusion: FlashMem provides an efficient framework for enabling long-horizon memory in LLMs without architectural segregation or redundant parameterization, achieving significant latency improvements while maintaining performance.

Abstract: The stateless architecture of Large Language Models inherently lacks the mechanism to preserve dynamic context, compelling agents to redundantly reprocess history to maintain long-horizon autonomy. While latent memory offers a solution, current approaches are hindered by architectural segregation, relying on auxiliary encoders that decouple memory from the reasoning backbone. We propose FlashMem, a framework that distills intrinsic memory directly from transient reasoning states via computation reuse. Leveraging the property that internal representations uniquely encode input trajectories, FlashMem identifies the last hidden state as a sufficient statistic for the interaction history. This enables a Shared-KV Consolidator to synthesize memory by attending directly to the backbone’s frozen cache, eliminating redundant re-parameterization. Furthermore, a parameter-free Cognitive Monitor leverages attention entropy to adaptively trigger consolidation only when high epistemic uncertainty is detected. Experiments demonstrate that FlashMem matches the performance of heavy baselines while reducing inference latency by 5 times, effectively bridging the gap between efficiency and persistent cognition.

Method: Cannot determine method as paper content is unavailable due to API rate limiting.

Result: Cannot determine results as paper content is unavailable due to API rate limiting.

Conclusion: Cannot draw conclusions as paper content is unavailable due to API rate limiting.

Abstract: Failed to fetch summary for 2601.06767: Page request resulted in HTTP 429 (https://export.arxiv.org/api/query?search_query=&id_list=2601.06767&sortBy=relevance&sortOrder=descending&start=0&max_results=100)

Method: Unable to determine method due to paper fetch failure

Result: Unable to determine results due to paper fetch failure

Conclusion: Unable to determine conclusion due to paper fetch failure

Abstract: Failed to fetch summary for 2308.12067: Page request resulted in HTTP 429 (https://export.arxiv.org/api/query?search_query=&id_list=2308.12067&sortBy=relevance&sortOrder=descending&start=0&max_results=100)

Method: Unable to determine method due to API access error

Result: Unable to determine results due to API access error

Conclusion: Unable to determine conclusion due to API access error

Abstract: Failed to fetch summary for 2601.07516: Page request resulted in HTTP 429 (https://export.arxiv.org/api/query?search_query=&id_list=2601.07516&sortBy=relevance&sortOrder=descending&start=0&max_results=100)

Method: Unable to determine method due to failed paper fetch

Result: Unable to determine results due to failed paper fetch

Conclusion: Unable to draw conclusions due to failed paper fetch

Abstract: Failed to fetch summary for 2601.08209: Page request resulted in HTTP 429 (https://export.arxiv.org/api/query?search_query=&id_list=2601.08209&sortBy=relevance&sortOrder=descending&start=0&max_results=100)

Method: Cannot determine method without access to paper content.

Result: Cannot determine results without access to paper content.

Conclusion: Cannot determine conclusion without access to paper content.

Abstract: Failed to fetch summary for 2601.09270: Page request resulted in HTTP 429 (https://export.arxiv.org/api/query?search_query=&id_list=2601.09270&sortBy=relevance&sortOrder=descending&start=0&max_results=100)

Method: Unable to determine method due to access error

Result: Unable to determine results due to access error

Conclusion: Unable to determine conclusion due to access error

Abstract: Failed to fetch summary for 2601.10775: Page request resulted in HTTP 429 (https://export.arxiv.org/api/query?search_query=&id_list=2601.10775&sortBy=relevance&sortOrder=descending&start=0&max_results=100)

Method: Cannot determine method without access to the paper content

Result: Cannot determine results without access to the paper content

Conclusion: Cannot determine conclusion without access to the paper content

Abstract: Failed to fetch summary for 2601.12104: Page request resulted in HTTP 429 (https://export.arxiv.org/api/query?search_query=&id_list=2601.12104&sortBy=relevance&sortOrder=descending&start=0&max_results=100)

Method: Cannot determine method as paper content is unavailable

Result: Cannot determine results as paper content is unavailable

Conclusion: Cannot draw conclusions as paper content is unavailable

Abstract: Failed to fetch summary for 2601.14698: Page request resulted in HTTP 429 (https://export.arxiv.org/api/query?search_query=&id_list=2601.14698&sortBy=relevance&sortOrder=descending&start=0&max_results=100)

Method: Unable to determine method due to fetch failure

Result: Unable to determine results due to fetch failure

Conclusion: Unable to determine conclusion due to fetch failure

Abstract: Failed to fetch summary for 2601.15593: Page request resulted in HTTP 429 (https://export.arxiv.org/api/query?search_query=&id_list=2601.15593&sortBy=relevance&sortOrder=descending&start=0&max_results=100)

Method: Unable to determine method due to access error

Result: Unable to determine results due to access error

Conclusion: Unable to determine conclusion due to access error

Abstract: Failed to fetch summary for 2601.17172: Page request resulted in HTTP 429 (https://export.arxiv.org/api/query?search_query=&id_list=2601.17172&sortBy=relevance&sortOrder=descending&start=0&max_results=100)

Method: Cannot determine method as paper content is unavailable

Result: Cannot determine results as paper content is unavailable

Conclusion: Cannot draw conclusions as paper content is unavailable

Abstract: Failed to fetch summary for 2601.17755: Page request resulted in HTTP 429 (https://export.arxiv.org/api/query?search_query=&id_list=2601.17755&sortBy=relevance&sortOrder=descending&start=0&max_results=100)

Method: Cannot determine method without access to paper content

Result: Cannot determine results without access to paper content

Conclusion: Cannot draw conclusions without access to paper content

Abstract: Failed to fetch summary for 2601.22361: Page request resulted in HTTP 429 (https://export.arxiv.org/api/query?search_query=&id_list=2601.22361&sortBy=relevance&sortOrder=descending&start=0&max_results=100)

Method: Cannot determine method as paper content is unavailable

Result: Cannot determine results as paper content is unavailable

Conclusion: Cannot draw conclusions as paper content is unavailable

Abstract: Failed to fetch summary for 2602.01064: Page request resulted in HTTP 429 (https://export.arxiv.org/api/query?search_query=&id_list=2602.01064&sortBy=relevance&sortOrder=descending&start=0&max_results=100)

Method: Unable to determine method due to API rate limiting preventing access to paper details

Result: Unable to determine results due to API rate limiting preventing access to paper details

Conclusion: Unable to determine conclusion due to API rate limiting preventing access to paper details

Abstract: Failed to fetch summary for 2602.02007: Page request resulted in HTTP 429 (https://export.arxiv.org/api/query?search_query=&id_list=2602.02007&sortBy=relevance&sortOrder=descending&start=0&max_results=100)

Method: Unable to determine method due to fetch failure

Result: Unable to determine results due to fetch failure

Conclusion: Unable to determine conclusion due to fetch failure

Abstract: Failed to fetch summary for 2602.02343: Page request resulted in HTTP 429 (https://export.arxiv.org/api/query?search_query=&id_list=2602.02343&sortBy=relevance&sortOrder=descending&start=0&max_results=100)

Method: Cannot determine method as paper content is unavailable

Result: Cannot determine results as paper content is unavailable

Conclusion: Cannot draw conclusions about the paper due to access limitations

Abstract: Failed to fetch summary for 2602.03108: Page request resulted in HTTP 429 (https://export.arxiv.org/api/query?search_query=&id_list=2602.03108&sortBy=relevance&sortOrder=descending&start=0&max_results=100)

Method: Cannot determine method without access to paper content

Result: Cannot determine results without access to paper content

Conclusion: Cannot draw conclusions without access to paper content

Abstract: Failed to fetch summary for 2602.14812: Page request resulted in HTTP 429 (https://export.arxiv.org/api/query?search_query=&id_list=2602.14812&sortBy=relevance&sortOrder=descending&start=0&max_results=100)

Method: Unable to determine method due to access error

Result: Unable to determine results due to access error

Conclusion: Unable to determine conclusion due to access error

Abstract: Failed to fetch summary for 2602.19509: Page request resulted in HTTP 429 (https://export.arxiv.org/api/query?search_query=&id_list=2602.19509&sortBy=relevance&sortOrder=descending&start=0&max_results=100)

Method: Unable to determine method due to technical error in accessing paper content

Result: Unable to determine results due to technical error in accessing paper content

Conclusion: Unable to draw conclusions due to technical error in accessing paper content

Abstract: Failed to fetch summary for 2602.21950: Page request resulted in HTTP 429 (https://export.arxiv.org/api/query?search_query=&id_list=2602.21950&sortBy=relevance&sortOrder=descending&start=0&max_results=100)

Method: Unable to determine method due to fetch failure

Result: Unable to determine results due to fetch failure

Conclusion: Unable to determine conclusion due to fetch failure

Abstract: Failed to fetch summary for 2603.02578: Page request resulted in HTTP 429 (https://export.arxiv.org/api/query?search_query=&id_list=2603.02578&sortBy=relevance&sortOrder=descending&start=0&max_results=100)

Method: Unknown - content unavailable due to HTTP 429 error from arXiv API.

Result: No results available - paper content could not be retrieved.

Conclusion: Cannot draw conclusions about paper content due to technical retrieval error.

Abstract: Failed to fetch summary for 2603.12277: Page request resulted in HTTP 429 (https://export.arxiv.org/api/query?search_query=&id_list=2603.12277&sortBy=relevance&sortOrder=descending&start=0&max_results=100)

Method: Cannot determine method without access to the paper abstract

Result: Cannot determine results without access to the paper abstract

Conclusion: Cannot determine conclusion without access to the paper abstract

Abstract: Failed to fetch summary for 2603.14410: Page request resulted in HTTP 429 (https://export.arxiv.org/api/query?search_query=&id_list=2603.14410&sortBy=relevance&sortOrder=descending&start=0&max_results=100)

Method: Unable to determine method due to fetch failure

Result: Unable to determine results due to fetch failure

Conclusion: Unable to determine conclusion due to fetch failure

Abstract: Failed to fetch summary for 2504.13818: Page request resulted in HTTP 429 (https://export.arxiv.org/api/query?search_query=&id_list=2504.13818&sortBy=relevance&sortOrder=descending&start=0&max_results=100)

Method: Cannot determine method as paper content is unavailable

Result: Cannot determine results as paper content is unavailable

Conclusion: Cannot draw conclusions as paper content is unavailable

Abstract: Failed to fetch summary for 2603.22241: Page request resulted in HTTP 429 (https://export.arxiv.org/api/query?search_query=&id_list=2603.22241&sortBy=relevance&sortOrder=descending&start=0&max_results=100)

Method: Cannot determine method as paper content is unavailable

Result: Cannot determine results as paper content is unavailable

Conclusion: Cannot draw conclusions as paper content is unavailable

Abstract: Failed to fetch summary for 2603.22816: Page request resulted in HTTP 429 (https://export.arxiv.org/api/query?search_query=&id_list=2603.22816&sortBy=relevance&sortOrder=descending&start=0&max_results=100)

Method: Cannot determine method as paper content is unavailable

Result: Cannot determine results as paper content is unavailable

Conclusion: Cannot draw conclusion as paper content is unavailable

Abstract: Failed to fetch summary for 2603.23516: Page request resulted in HTTP 429 (https://export.arxiv.org/api/query?search_query=&id_list=2603.23516&sortBy=relevance&sortOrder=descending&start=0&max_results=100)

Method: Cannot determine method as paper content is unavailable

Result: Cannot determine results as paper content is unavailable

Conclusion: Cannot draw conclusions as paper content is unavailable

Abstract: Failed to fetch summary for 2505.11737: Page request resulted in HTTP 429 (https://export.arxiv.org/api/query?search_query=&id_list=2505.11737&sortBy=relevance&sortOrder=descending&start=0&max_results=100)

Method: Unable to determine method due to access error

Result: Unable to determine results due to access error

Conclusion: Unable to determine conclusion due to access error

Abstract: Failed to fetch summary for 2603.24329: Page request resulted in HTTP 429 (https://export.arxiv.org/api/query?search_query=&id_list=2603.24329&sortBy=relevance&sortOrder=descending&start=0&max_results=100)

Method: Cannot determine method as abstract retrieval failed

Result: Cannot determine results as abstract retrieval failed

Conclusion: Cannot draw conclusions without access to paper content

Abstract: Failed to fetch summary for 2603.25620: Page request resulted in HTTP 429 (https://export.arxiv.org/api/query?search_query=&id_list=2603.25620&sortBy=relevance&sortOrder=descending&start=0&max_results=100)

Method: Unable to determine method due to fetch failure

Result: Unable to determine results due to fetch failure

Conclusion: Unable to determine conclusion due to fetch failure

Abstract: Failed to fetch summary for 2604.00013: Page request resulted in HTTP 429 (https://export.arxiv.org/api/query?search_query=&id_list=2604.00013&sortBy=relevance&sortOrder=descending&start=0&max_results=100)

Method: Cannot determine method without access to paper content

Result: Cannot determine results without access to paper content

Conclusion: Cannot determine conclusion without access to paper content

Abstract: Failed to fetch summary for 2604.01306: Page request resulted in HTTP 429 (https://export.arxiv.org/api/query?search_query=&id_list=2604.01306&sortBy=relevance&sortOrder=descending&start=0&max_results=100)

Method: Unable to determine method due to failed paper fetch

Result: Unable to determine results due to failed paper fetch

Conclusion: Unable to draw conclusions due to failed paper fetch

Abstract: Failed to fetch summary for 2604.03147: Page request resulted in HTTP 429 (https://export.arxiv.org/api/query?search_query=&id_list=2604.03147&sortBy=relevance&sortOrder=descending&start=0&max_results=100)

Method: Cannot determine method as paper content is unavailable

Result: Cannot determine results as paper content is unavailable

Conclusion: Cannot determine conclusion as paper content is unavailable

Abstract: Failed to fetch summary for 2604.04385: Page request resulted in HTTP 429 (https://export.arxiv.org/api/query?search_query=&id_list=2604.04385&sortBy=relevance&sortOrder=descending&start=0&max_results=100)

Method: Proposes constructing Structured Event Facts (compact representation of salient events and causal relationships) before reasoning. Introduces CausalFact-60K dataset and four-stage training: facts alignment, format warm-start, thinking warm-start, and RL-based post-training. Uses Multi-Objective Reinforcement Learning to balance structural completeness, causal fidelity, and reasoning length.

Result: Develops Factum-4B model that yields more reliable reasoning and delivers stronger performance on challenging video understanding tasks requiring fine-grained temporal inference.

Conclusion: Structured Event Facts provide explicit constraints for concise and causally grounded reasoning in video understanding, addressing limitations of current Video-LLMs through structured representation and multi-objective optimization.

Abstract: Human understanding of video dynamics is typically grounded in a structured mental representation of entities, actions, and temporal relations, rather than relying solely on immediate deductive reasoning. In contrast, existing Video-LLMs largely depend on unstructured video reasoning, where critical visual evidence is embedded in verbose textual descriptions and temporal causality is often weakly modeled. This leads to inefficient processes and fragile causal inference. To bridge this cognitive gap, we propose constructing a compact representation of salient events and their causal relationships, which we name Structured Event Facts, prior to the reasoning stage. This structured prior serves as an explicit constraint to promote concise and causally grounded reasoning, while also making intermediate evidence easier to verify. To effectively train models on such structured facts, we introduce CausalFact-60K and a four-stage training pipeline comprising facts alignment, format warm-start, thinking warm-start, and reinforcement learning-based post-training. During RL stage, we find that this framework introduces competing objectives, as structural completeness and causal fidelity must be balanced against reasoning length, making it difficult to optimize. We address this challenge by formulating the optimization as a Multi-Objective Reinforcement Learning (MORL) problem and explicitly optimizing toward the Pareto-Frontier to balance these trade-offs. As a result, we introduce Factum-4B, which yields more reliable reasoning and delivers stronger performance on challenging video understanding tasks requiring fine-grained temporal inference.

Method: Cannot determine method without access to paper content

Result: Cannot determine results without access to paper content

Conclusion: Cannot determine conclusion without access to paper content

Abstract: Failed to fetch summary for 2604.05549: Page request resulted in HTTP 429 (https://export.arxiv.org/api/query?search_query=&id_list=2604.05549&sortBy=relevance&sortOrder=descending&start=0&max_results=100)

Method: Cannot determine method as paper content is unavailable

Result: Cannot determine results as paper content is unavailable

Conclusion: Cannot draw conclusions as paper content is unavailable

Abstract: Failed to fetch summary for 2507.15640: Page request resulted in HTTP 429 (https://export.arxiv.org/api/query?search_query=&id_list=2507.15640&sortBy=relevance&sortOrder=descending&start=0&max_results=100)

Method: Unable to determine method due to fetch failure

Result: Unable to determine results due to fetch failure

Conclusion: Unable to determine conclusion due to fetch failure

Abstract: Failed to fetch summary for 2604.05795: Page request resulted in HTTP 429 (https://export.arxiv.org/api/query?search_query=&id_list=2604.05795&sortBy=relevance&sortOrder=descending&start=0&max_results=100)

Method: Unable to determine method due to failed API request

Result: Unable to determine results due to failed API request

Conclusion: Unable to draw conclusions due to failed API request

Abstract: Failed to fetch summary for 2604.06784: Page request resulted in HTTP 429 (https://export.arxiv.org/api/query?search_query=&id_list=2604.06784&sortBy=relevance&sortOrder=descending&start=0&max_results=100)

Method: Cannot determine method without access to paper content. The arXiv API rate limiting prevents analysis of the paper’s technical approach.

Result: Cannot determine results without access to paper content. The arXiv API error prevents evaluation of the paper’s findings.

Conclusion: Cannot draw conclusions about the paper’s content due to technical limitations in accessing the arXiv API. The paper exists but cannot be analyzed at this time.

Abstract: Failed to fetch summary for 2604.07466: Page request resulted in HTTP 429 (https://export.arxiv.org/api/query?search_query=&id_list=2604.07466&sortBy=relevance&sortOrder=descending&start=0&max_results=100)

Method: Cannot determine method as paper content is unavailable

Result: Cannot determine results as paper content is unavailable

Conclusion: Cannot determine conclusion as paper content is unavailable

Abstract: Failed to fetch summary for 2604.07583: Page request resulted in HTTP 429 (https://export.arxiv.org/api/query?search_query=&id_list=2604.07583&sortBy=relevance&sortOrder=descending&start=0&max_results=100)

Method: Cannot determine method as abstract is unavailable due to rate limiting error

Result: Cannot determine results as abstract is unavailable due to rate limiting error

Conclusion: Cannot draw conclusion as paper content is inaccessible due to HTTP 429 error

Abstract: Failed to fetch summary for 2508.17784: Page request resulted in HTTP 429 (https://export.arxiv.org/api/query?search_query=&id_list=2508.17784&sortBy=relevance&sortOrder=descending&start=0&max_results=100)

Method: Cannot determine method without access to paper content

Result: Cannot determine results without access to paper content

Conclusion: Cannot determine conclusion without access to paper content

Abstract: Failed to fetch summary for 2604.07717: Page request resulted in HTTP 429 (https://export.arxiv.org/api/query?search_query=&id_list=2604.07717&sortBy=relevance&sortOrder=descending&start=0&max_results=100)

Method: Cannot determine method without access to paper content

Result: Cannot determine results without access to paper content

Conclusion: Cannot draw conclusions without access to paper content

Abstract: Failed to fetch summary for 2604.07886: Page request resulted in HTTP 429 (https://export.arxiv.org/api/query?search_query=&id_list=2604.07886&sortBy=relevance&sortOrder=descending&start=0&max_results=100)

Method: Cannot determine method due to inability to access paper content

Result: Cannot determine results due to inability to access paper content

Conclusion: Cannot determine conclusion due to inability to access paper content

Abstract: Failed to fetch summary for 2509.20868: Page request resulted in HTTP 429 (https://export.arxiv.org/api/query?search_query=&id_list=2509.20868&sortBy=relevance&sortOrder=descending&start=0&max_results=100)

Method: Cannot determine method without access to paper content

Result: Cannot determine results without access to paper content

Conclusion: Cannot draw conclusions without access to paper content

Abstract: Failed to fetch summary for 2604.07892: Page request resulted in HTTP 429 (https://export.arxiv.org/api/query?search_query=&id_list=2604.07892&sortBy=relevance&sortOrder=descending&start=0&max_results=100)

Method: Unable to determine method due to fetch failure

Result: Unable to determine results due to fetch failure

Conclusion: Unable to determine conclusion due to fetch failure

Abstract: Failed to fetch summary for 2604.08052: Page request resulted in HTTP 429 (https://export.arxiv.org/api/query?search_query=&id_list=2604.08052&sortBy=relevance&sortOrder=descending&start=0&max_results=100)

Method: Unable to determine method due to access error

Result: Unable to determine results due to access error

Conclusion: Unable to determine conclusion due to access error

Abstract: Failed to fetch summary for 2604.08557: Page request resulted in HTTP 429 (https://export.arxiv.org/api/query?search_query=&id_list=2604.08557&sortBy=relevance&sortOrder=descending&start=0&max_results=100)

Method: Cannot determine method without access to the paper abstract

Result: Cannot determine results without access to the paper abstract

Conclusion: Unable to analyze due to technical error fetching paper information

Abstract: Failed to fetch summary for 2604.09066: Page request resulted in HTTP 429 (https://export.arxiv.org/api/query?search_query=&id_list=2604.09066&sortBy=relevance&sortOrder=descending&start=0&max_results=100)

Method: Cannot determine method as paper content is unavailable

Result: Cannot determine results as paper content is unavailable

Conclusion: Cannot determine conclusion as paper content is unavailable

Abstract: Failed to fetch summary for 2509.23808: Page request resulted in HTTP 429 (https://export.arxiv.org/api/query?search_query=&id_list=2509.23808&sortBy=relevance&sortOrder=descending&start=0&max_results=100)

Method: Cannot determine method due to unavailability of paper content

Result: Cannot determine results due to unavailability of paper content

Conclusion: Cannot determine conclusion due to unavailability of paper content

Abstract: Failed to fetch summary for 2604.09443: Page request resulted in HTTP 429 (https://export.arxiv.org/api/query?search_query=&id_list=2604.09443&sortBy=relevance&sortOrder=descending&start=0&max_results=100)

Method: Survey of 47 methods organized in two-dimensional taxonomy by assignment granularity (token, segment, step, turn, multi-agent) and methodology (Monte Carlo, temporal difference, model-based, game-theoretic, information-theoretic). Created three resources: structured paper inventory, reporting checklist, and benchmark protocol with decision tree.

Result: Analysis shows reasoning credit assignment is maturing around process reward models and critic-free group comparison, while agentic credit assignment is driving new approaches like hindsight counterfactual analysis, privileged asymmetric critics, and turn-level MDP reformulations.

Conclusion: The shift from reasoning to agentic RL complicates and reshapes the credit assignment landscape, with agentic CA requiring fundamentally different approaches than reasoning CA due to stochastic transitions, partial observability, and longer horizons.

Abstract: Reinforcement learning (RL) for large language models (LLMs) increasingly relies on sparse, outcome-level rewards – yet determining which actions within a long trajectory caused the outcome remains difficult. This credit assignment (CA) problem manifests in two regimes: reasoning RL, where credit must be distributed across tokens and steps within a single chain-of-thought generation (500–30K+ tokens); and agentic RL, where multi-turn environment interaction introduces stochastic transitions, partial observability, and horizons of 100+ turns (100K–1M tokens), making episode-level credit increasingly uninformative. We survey 47 CA methods (41 core, 6 adjacent enablers) published between 2024 and early 2026, organizing them in a two-dimensional taxonomy by assignment granularity (token, segment, step, turn, multi-agent) and methodology (Monte Carlo, temporal difference, model-based, game-theoretic, information-theoretic). Beyond the survey itself, we contribute three reusable resources: (1) a structured, machine-readable paper inventory with taxonomy labels, baseline families, and evidence levels; (2) a reporting checklist for future CA papers, validated against the reviewed literature to identify systematic methodological gaps; and (3) a benchmark protocol specification with task families, metadata requirements, and controlled bifurcation tasks, accompanied by a method selection decision tree. Our synthesis suggests that the shift from reasoning to agentic RL complicates and reshapes the credit assignment landscape: reasoning CA is maturing around process reward models and critic-free group comparison, while agentic CA is driving genuinely new approaches – hindsight counterfactual analysis, privileged asymmetric critics, and turn-level MDP reformulations – that have no direct precedent in reasoning RL.

Method: Unable to determine method due to access error

Result: Unable to determine results due to access error

Conclusion: Unable to determine conclusion due to access error

Abstract: Failed to fetch summary for 2307.01139: Page request resulted in HTTP 429 (https://export.arxiv.org/api/query?search_query=&id_list=2307.01139&sortBy=relevance&sortOrder=descending&start=0&max_results=100)

Method: Unable to determine method due to access error

Result: Unable to determine results due to access error

Conclusion: Unable to determine conclusion due to access error

Abstract: Failed to fetch summary for 2406.11290: Page request resulted in HTTP 429 (https://export.arxiv.org/api/query?search_query=&id_list=2406.11290&sortBy=relevance&sortOrder=descending&start=0&max_results=100)

Method: Unable to determine method due to access restrictions

Result: Unable to determine results due to access restrictions

Conclusion: Unable to determine conclusion due to access restrictions

Abstract: Failed to fetch summary for 2502.13166: Page request resulted in HTTP 429 (https://export.arxiv.org/api/query?search_query=&id_list=2502.13166&sortBy=relevance&sortOrder=descending&start=0&max_results=100)

Method: Cannot determine method without access to paper content

Result: Cannot determine results without access to paper content

Conclusion: Cannot determine conclusion without access to paper content

Abstract: Failed to fetch summary for 2505.18931: Page request resulted in HTTP 429 (https://export.arxiv.org/api/query?search_query=&id_list=2505.18931&sortBy=relevance&sortOrder=descending&start=0&max_results=100)

Method: Cannot determine method as paper content is unavailable

Result: Cannot determine results as paper content is unavailable

Conclusion: Cannot determine conclusion as paper content is unavailable

Abstract: Failed to fetch summary for 2504.05902: Page request resulted in HTTP 429 (https://export.arxiv.org/api/query?search_query=&id_list=2504.05902&sortBy=relevance&sortOrder=descending&start=0&max_results=100)

Method: Cannot determine method as paper content is unavailable

Result: Cannot determine results as paper content is unavailable

Conclusion: Cannot draw conclusions without access to paper content

Abstract: Failed to fetch summary for 2506.14629: Page request resulted in HTTP 429 (https://export.arxiv.org/api/query?search_query=&id_list=2506.14629&sortBy=relevance&sortOrder=descending&start=0&max_results=100)

Method: Cannot determine method as paper content is unavailable due to API rate limiting

Result: Cannot determine results as paper content is unavailable due to API rate limiting

Conclusion: Cannot draw conclusions as paper content is unavailable due to API rate limiting

Abstract: Failed to fetch summary for 2507.03336: Page request resulted in HTTP 429 (https://export.arxiv.org/api/query?search_query=&id_list=2507.03336&sortBy=relevance&sortOrder=descending&start=0&max_results=100)

Method: Unable to determine method due to fetch failure

Result: Unable to determine results due to fetch failure

Conclusion: Unable to determine conclusion due to fetch failure

Abstract: Failed to fetch summary for 2508.03306: Page request resulted in HTTP 429 (https://export.arxiv.org/api/query?search_query=&id_list=2508.03306&sortBy=relevance&sortOrder=descending&start=0&max_results=100)

Method: No method information available due to API access issues

Result: No results available - paper content inaccessible

Conclusion: Cannot analyze paper due to technical limitations in accessing the abstract

Abstract: Failed to fetch summary for 2510.10925: Page request resulted in HTTP 429 (https://export.arxiv.org/api/query?search_query=&id_list=2510.10925&sortBy=relevance&sortOrder=descending&start=0&max_results=100)

Method: Cannot determine method due to failed paper retrieval

Result: Cannot determine results due to failed paper retrieval

Conclusion: Cannot determine conclusion due to failed paper retrieval

Abstract: Failed to fetch summary for 2509.00891: Page request resulted in HTTP 429 (https://export.arxiv.org/api/query?search_query=&id_list=2509.00891&sortBy=relevance&sortOrder=descending&start=0&max_results=100)

Method: Unable to determine method due to fetch failure

Result: Unable to determine results due to fetch failure

Conclusion: Unable to determine conclusion due to fetch failure

Abstract: Failed to fetch summary for 2509.21982: Page request resulted in HTTP 429 (https://export.arxiv.org/api/query?search_query=&id_list=2509.21982&sortBy=relevance&sortOrder=descending&start=0&max_results=100)

Method: Unable to determine method due to fetch failure

Result: Unable to determine results due to fetch failure

Conclusion: Unable to determine conclusion due to fetch failure

Abstract: Failed to fetch summary for 2509.17458: Page request resulted in HTTP 429 (https://export.arxiv.org/api/query?search_query=&id_list=2509.17458&sortBy=relevance&sortOrder=descending&start=0&max_results=100)

Method: Unable to determine method due to fetch failure

Result: Unable to determine results due to fetch failure

Conclusion: Unable to determine conclusion due to fetch failure

Abstract: Failed to fetch summary for 2509.22097: Page request resulted in HTTP 429 (https://export.arxiv.org/api/query?search_query=&id_list=2509.22097&sortBy=relevance&sortOrder=descending&start=0&max_results=100)

Method: Cannot determine method as paper content is unavailable due to API rate limiting.

Result: Cannot determine results as paper content is unavailable due to API rate limiting.

Conclusion: Cannot draw conclusions as paper content is unavailable due to API rate limiting.

Abstract: Failed to fetch summary for 2510.20792: Page request resulted in HTTP 429 (https://export.arxiv.org/api/query?search_query=&id_list=2510.20792&sortBy=relevance&sortOrder=descending&start=0&max_results=100)

Method: Unable to determine method due to fetch failure

Result: Unable to determine results due to fetch failure

Conclusion: Unable to determine conclusion due to fetch failure

Abstract: Failed to fetch summary for 2510.27484: Page request resulted in HTTP 429 (https://export.arxiv.org/api/query?search_query=&id_list=2510.27484&sortBy=relevance&sortOrder=descending&start=0&max_results=100)

Method: Unable to determine method due to technical error in fetching paper content

Result: Unable to determine results due to technical error in fetching paper content

Conclusion: Unable to draw conclusions due to technical error in fetching paper content

Abstract: Failed to fetch summary for 2511.04776: Page request resulted in HTTP 429 (https://export.arxiv.org/api/query?search_query=&id_list=2511.04776&sortBy=relevance&sortOrder=descending&start=0&max_results=100)

Method: Cannot determine method without access to paper content.

Result: Cannot determine results without access to paper content.

Conclusion: Cannot draw conclusions without access to paper content.

Abstract: Failed to fetch summary for 2511.06101: Page request resulted in HTTP 429 (https://export.arxiv.org/api/query?search_query=&id_list=2511.06101&sortBy=relevance&sortOrder=descending&start=0&max_results=100)

Method: Unable to determine method due to access error

Result: Unable to determine results due to access error

Conclusion: Unable to determine conclusion due to access error

Abstract: Failed to fetch summary for 2512.02393: Page request resulted in HTTP 429 (https://export.arxiv.org/api/query?search_query=&id_list=2512.02393&sortBy=relevance&sortOrder=descending&start=0&max_results=100)

Method: Unable to determine method due to retrieval failure

Result: Unable to determine results due to retrieval failure

Conclusion: Unable to determine conclusion due to retrieval failure

Abstract: Failed to fetch summary for 2512.06713: Page request resulted in HTTP 429 (https://export.arxiv.org/api/query?search_query=&id_list=2512.06713&sortBy=relevance&sortOrder=descending&start=0&max_results=100)

Method: Cannot determine method as paper content is unavailable

Result: Cannot determine results as paper content is unavailable

Conclusion: Cannot draw conclusions about the paper due to technical fetching issues

Abstract: Failed to fetch summary for 2512.07222: Page request resulted in HTTP 429 (https://export.arxiv.org/api/query?search_query=&id_list=2512.07222&sortBy=relevance&sortOrder=descending&start=0&max_results=100)

Method: No method information available due to API request failure

Result: No results available - paper summary fetch failed

Conclusion: Unable to provide analysis due to technical limitations in accessing the paper

Abstract: Failed to fetch summary for 2512.17270: Page request resulted in HTTP 429 (https://export.arxiv.org/api/query?search_query=&id_list=2512.17270&sortBy=relevance&sortOrder=descending&start=0&max_results=100)

Method: Cannot determine method as paper content could not be retrieved

Result: Cannot determine results as paper content could not be retrieved

Conclusion: Cannot draw conclusions as paper content could not be retrieved

Abstract: Failed to fetch summary for 2601.04672: Page request resulted in HTTP 429 (https://export.arxiv.org/api/query?search_query=&id_list=2601.04672&sortBy=relevance&sortOrder=descending&start=0&max_results=100)

Method: Cannot determine method without access to paper content

Result: Cannot determine results without access to paper content

Conclusion: Cannot draw conclusions without access to paper content

Abstract: Failed to fetch summary for 2601.07208: Page request resulted in HTTP 429 (https://export.arxiv.org/api/query?search_query=&id_list=2601.07208&sortBy=relevance&sortOrder=descending&start=0&max_results=100)

Method: Cannot determine method without access to paper content.

Result: Cannot determine results without access to paper content.

Conclusion: Cannot draw conclusions without access to paper content.

Abstract: Failed to fetch summary for 2602.00793: Page request resulted in HTTP 429 (https://export.arxiv.org/api/query?search_query=&id_list=2602.00793&sortBy=relevance&sortOrder=descending&start=0&max_results=100)

Method: Cannot determine method as paper content is unavailable due to API rate limiting.

Result: Cannot determine results as paper content is unavailable due to API rate limiting.

Conclusion: Cannot draw conclusions as paper content is unavailable due to API rate limiting.

Abstract: Failed to fetch summary for 2602.03059: Page request resulted in HTTP 429 (https://export.arxiv.org/api/query?search_query=&id_list=2602.03059&sortBy=relevance&sortOrder=descending&start=0&max_results=100)

Method: Unable to determine method due to fetch failure

Result: Unable to determine results due to fetch failure

Conclusion: Unable to determine conclusion due to fetch failure

Abstract: Failed to fetch summary for 2602.05467: Page request resulted in HTTP 429 (https://export.arxiv.org/api/query?search_query=&id_list=2602.05467&sortBy=relevance&sortOrder=descending&start=0&max_results=100)

Method: Cannot determine method as paper content could not be retrieved

Result: Cannot determine results as paper content could not be retrieved

Conclusion: Cannot draw conclusions as paper content could not be retrieved

Abstract: Failed to fetch summary for 2602.13934: Page request resulted in HTTP 429 (https://export.arxiv.org/api/query?search_query=&id_list=2602.13934&sortBy=relevance&sortOrder=descending&start=0&max_results=100)

Method: Unable to determine method due to missing paper content

Result: Unable to determine results due to missing paper content

Conclusion: Unable to determine conclusion due to missing paper content

Abstract: Failed to fetch summary for 2602.21480: Page request resulted in HTTP 429 (https://export.arxiv.org/api/query?search_query=&id_list=2602.21480&sortBy=relevance&sortOrder=descending&start=0&max_results=100)

Method: Unable to determine method due to failed paper retrieval

Result: Unable to determine results due to failed paper retrieval

Conclusion: Unable to draw conclusions due to failed paper retrieval

Abstract: Failed to fetch summary for 2603.08715: Page request resulted in HTTP 429 (https://export.arxiv.org/api/query?search_query=&id_list=2603.08715&sortBy=relevance&sortOrder=descending&start=0&max_results=100)

Method: Cannot determine method without access to the paper content.

Result: Cannot determine results without access to the paper content.

Conclusion: Cannot determine conclusion without access to the paper content.

Abstract: Failed to fetch summary for 2603.08942: Page request resulted in HTTP 429 (https://export.arxiv.org/api/query?search_query=&id_list=2603.08942&sortBy=relevance&sortOrder=descending&start=0&max_results=100)

Method: Unable to determine method due to fetch failure

Result: Unable to determine results due to fetch failure

Conclusion: Unable to determine conclusion due to fetch failure

Abstract: Failed to fetch summary for 2603.14968: Page request resulted in HTTP 429 (https://export.arxiv.org/api/query?search_query=&id_list=2603.14968&sortBy=relevance&sortOrder=descending&start=0&max_results=100)

Method: Unable to determine method due to access error

Result: Unable to determine results due to access error

Conclusion: Unable to determine conclusion due to access error

Abstract: Failed to fetch summary for 2603.16068: Page request resulted in HTTP 429 (https://export.arxiv.org/api/query?search_query=&id_list=2603.16068&sortBy=relevance&sortOrder=descending&start=0&max_results=100)

Method: Cannot determine method as paper content is unavailable

Result: Cannot determine results as paper content is unavailable

Conclusion: Cannot determine conclusion as paper content is unavailable

Abstract: Failed to fetch summary for 2603.18280: Page request resulted in HTTP 429 (https://export.arxiv.org/api/query?search_query=&id_list=2603.18280&sortBy=relevance&sortOrder=descending&start=0&max_results=100)

Method: Cannot determine method without access to paper content

Result: Cannot determine results without access to paper content

Conclusion: Cannot draw conclusions without access to paper content

Abstract: Failed to fetch summary for 2603.25975: Page request resulted in HTTP 429 (https://export.arxiv.org/api/query?search_query=&id_list=2603.25975&sortBy=relevance&sortOrder=descending&start=0&max_results=100)

Method: Cannot determine method without access to paper content.

Result: Cannot determine results without access to paper content.

Conclusion: Cannot draw conclusions without access to paper content.

Abstract: Failed to fetch summary for 2603.27960: Page request resulted in HTTP 429 (https://export.arxiv.org/api/query?search_query=&id_list=2603.27960&sortBy=relevance&sortOrder=descending&start=0&max_results=100)

Method: Unable to determine method due to access error

Result: Unable to determine results due to access error

Conclusion: Unable to determine conclusion due to access error

Abstract: Failed to fetch summary for 2604.03873: Page request resulted in HTTP 429 (https://export.arxiv.org/api/query?search_query=&id_list=2604.03873&sortBy=relevance&sortOrder=descending&start=0&max_results=100)

Method: Unable to determine method due to fetch failure

Result: Unable to determine results due to fetch failure

Conclusion: Unable to determine conclusion due to fetch failure

Abstract: Failed to fetch summary for 2604.05005: Page request resulted in HTTP 429 (https://export.arxiv.org/api/query?search_query=&id_list=2604.05005&sortBy=relevance&sortOrder=descending&start=0&max_results=100)

Method: Unable to determine method due to missing abstract

Result: Unable to determine results due to missing abstract

Conclusion: Unable to determine conclusion due to missing abstract

Abstract: Failed to fetch summary for 2604.05767: Page request resulted in HTTP 429 (https://export.arxiv.org/api/query?search_query=&id_list=2604.05767&sortBy=relevance&sortOrder=descending&start=0&max_results=100)

Method: Cannot determine method without access to the paper abstract.

Result: Cannot determine results without access to the paper abstract.

Conclusion: Cannot draw conclusions without access to the paper abstract.

Abstract: Failed to fetch summary for 2604.07709: Page request resulted in HTTP 429 (https://export.arxiv.org/api/query?search_query=&id_list=2604.07709&sortBy=relevance&sortOrder=descending&start=0&max_results=100)

Method: Cannot determine method as abstract content is unavailable due to HTTP 429 error from arXiv API.

Result: Cannot determine results as abstract content is unavailable due to HTTP 429 error from arXiv API.

Conclusion: Cannot draw conclusions as abstract content is unavailable due to HTTP 429 error from arXiv API.

Abstract: Failed to fetch summary for 2604.09364: Page request resulted in HTTP 429 (https://export.arxiv.org/api/query?search_query=&id_list=2604.09364&sortBy=relevance&sortOrder=descending&start=0&max_results=100)

Method: Fuses implicit 3D cues from stereo vision with explicit prior knowledge from natural language text prompts containing object class and approximate volume; extracts deep features from stereo image pair and text, integrates them using projection layer into unified multimodal representation for regression

Result: Extensive experiments on public datasets show text-guided approach significantly outperforms vision-only baselines; demonstrates that even simple textual priors can effectively guide volume estimation

Conclusion: Leveraging textual priors with stereo vision enables more accurate volume estimation, paving way for more context-aware visual measurement systems

Abstract: Accurate volume estimation of objects from visual data is a long-standing challenge in computer vision with significant applications in robotics, logistics, and smart health. Existing methods often rely on complex 3D reconstruction pipelines or struggle with the ambiguity inherent in single-view images. To address these limitations, we introduce a new method that fuses implicit 3D cues from stereo vision with explicit prior knowledge from natural language text. Our approach extracts deep features from a stereo image pair and a descriptive text prompt that contains the object’s class and an approximate volume, then integrates them using a simple yet effective projection layer into a unified, multi-modal representation for regression. We conduct extensive experiments on public datasets demonstrating that our text-guided approach significantly outperforms vision-only baselines. Our findings show that leveraging even simple textual priors can effectively guide the volume estimation task, paving the way for more context-aware visual measurement systems. Code: https://gitlab.com/viper-purdue/stereo-typical-estimator.

Method: Feed-forward stylization network with test-time optimization using composite objective: AdaIN-inspired style loss, correspondence-based consistency loss (SuperPoint/SuperGlue), depth-preservation loss (MiDaS/DPT), and global color alignment with staged weight scheduling.

Result: Method reduces structural distortion, improves SLAM stability and reconstructed geometry while maintaining competitive stylization quality, outperforming MuVieCAST baselines on trajectory and point-cloud consistency metrics.

Conclusion: The proposed multi-view stylization approach successfully preserves geometric consistency needed for downstream 3D tasks while achieving artistic style transfer, addressing the fundamental challenge of maintaining correspondences across views.

Abstract: Artistic style transfer is well studied for images and videos, but extending it to multi-view 3D scenes remains difficult because stylization can disrupt correspondences needed by geometry-aware pipelines. Independent per-view stylization often causes texture drift, warped edges, and inconsistent shading, degrading SLAM, depth prediction, and multi-view reconstruction. This thesis addresses multi-view stylization that remains usable for downstream 3D tasks without assuming camera poses or an explicit 3D representation during training. We introduce a feed-forward stylization network trained with per-scene test-time optimization under a composite objective coupling appearance transfer with geometry preservation. Stylization is driven by an AdaIN-inspired loss from a frozen VGG-19 encoder, matching channel-wise moments to a style image. To stabilize structure across viewpoints, we propose a correspondence-based consistency loss using SuperPoint and SuperGlue, constraining descriptors from a stylized anchor view to remain consistent with matched descriptors from the original multi-view set. We also impose a depth-preservation loss using MiDaS/DPT and use global color alignment to reduce depth-model domain shift. A staged weight schedule introduces geometry and depth constraints. We evaluate on Tanks and Temples and Mip-NeRF 360 using image and reconstruction metrics. Style adherence and structure retention are measured by Color Histogram Distance (CHD) and Structure Distance (DSD). For 3D consistency, we use monocular DROID-SLAM trajectories and symmetric Chamfer distance on back-projected point clouds. Across ablations, correspondence and depth regularization reduce structural distortion and improve SLAM stability and reconstructed geometry; on scenes with MuVieCAST baselines, our method yields stronger trajectory and point-cloud consistency while maintaining competitive stylization.

Method: PA-SFM integrates the acoustic wave equation into a differentiable programming pipeline with a GPU-accelerated acoustic radiation kernel. It simultaneously optimizes 3D photoacoustic source distribution and sensor array pose via gradient descent. Uses coarse-to-fine optimization with geometric consistency checks and rigid-body constraints to eliminate motion outliers.

Result: Achieves sub-millimeter positioning accuracy and restores high-resolution 3D vascular structures comparable to ground-truth benchmarks in both numerical simulations and in-vivo rat experiments.

Conclusion: PA-SFM offers a low-cost, software-defined solution for clinical freehand photoacoustic imaging that eliminates the need for external tracking hardware while maintaining high reconstruction quality.

Abstract: Three-dimensional (3D) handheld photoacoustic tomography typically relies on bulky and expensive external positioning sensors to correct motion artifacts, which severely limits its clinical flexibility and accessibility. To address this challenge, we present PA-SFM, a tracker-free framework that leverages exclusively single-modality photoacoustic data for both sensor pose recovery and high-fidelity 3D reconstruction via differentiable acoustic radiation modeling. Unlike traditional structure-from-motion (SFM) methods based on visual features, PA-SFM integrates the acoustic wave equation into a differentiable programming pipeline. By leveraging a high-performance, GPU-accelerated acoustic radiation kernel, the framework simultaneously optimizes the 3D photoacoustic source distribution and the sensor array pose via gradient descent. To ensure robust convergence in freehand scenarios, we introduce a coarse-to-fine optimization strategy that incorporates geometric consistency checks and rigid-body constraints to eliminate motion outliers. We validated the proposed method through both numerical simulations and in-vivo rat experiments. The results demonstrate that PA-SFM achieves sub-millimeter positioning accuracy and restores high-resolution 3D vascular structures comparable to ground-truth benchmarks, offering a low-cost, software-defined solution for clinical freehand photoacoustic imaging. The source code is publicly available at \href{https://github.com/JaegerCQ/PA-SFM}{https://github.com/JaegerCQ/PA-SFM}.

Method: TRACE uses: 1) Thermal Gas-Aware Attention encoder with per-pixel gas intensity supervision, 2) Attention-based Temporal Fusion module for breath-cycle dynamics, and 3) four-stage progressive training curriculum to couple both objectives without gradient interference.

Result: Achieves mIoU of 0.998 on CO2 plume segmentation and best results on all segmentation/classification metrics, outperforming 15 SOTA models including larger domain-specific gas segmenters.

Conclusion: TRACE enables practical non-invasive, continuous per-animal CO2 monitoring from overhead thermal cameras at commercial scale for livestock management.

Abstract: Quantifying exhaled CO2 from free-roaming cattle is both a direct indicator of rumen metabolic state and a prerequisite for farm-scale carbon accounting, yet no existing system can deliver continuous, spatially resolved measurements without physical confinement or contact. We present TRACE (Thermal Recognition Attentive-Framework for CO2 Emissions from Livestock), the first unified framework to jointly address per-frame CO2 plume segmentation and clip-level emission flux classification from mid-wave infrared (MWIR) thermal video. TRACE contributes three domain-specific advances: a Thermal Gas-Aware Attention (TGAA) encoder that incorporates per-pixel gas intensity as a spatial supervisory signal to direct self-attention toward high-emission regions at each encoder stage; an Attention-based Temporal Fusion (ATF) module that captures breath-cycle dynamics through structured cross-frame attention for sequence-level flux classification; and a four-stage progressive training curriculum that couples both objectives while preventing gradient interference. Benchmarked against fifteen state-of-the-art models on the CO2 Farm Thermal Gas Dataset, TRACE achieves an mIoU of 0.998 and the best result on every segmentation and classification metric simultaneously, outperforming domain-specific gas segmenters with several times more parameters and surpassing all baselines in flux classification. Ablation studies confirm that each component is individually essential: gas-conditioned attention alone determines precise plume boundary localization, and temporal reasoning is indispensable for flux-level discrimination. TRACE establishes a practical path toward non-invasive, continuous, per-animal CO2 monitoring from overhead thermal cameras at commercial scale. Codes are available at https://github.com/taminulislam/trace.

Method: FlowHijack combines a τ-conditioned injection strategy that manipulates the initial phase of action generation with a dynamics mimicry regularizer. This approach systematically targets the underlying vector-field dynamics of flow-matching VLAs while maintaining stealth through context-aware triggers.

Result: Experiments show FlowHijack achieves high attack success rates with stealthy triggers where prior works failed. It preserves benign task performance and generates malicious actions that are behaviorally indistinguishable from normal actions by enforcing kinematic similarity.

Conclusion: The research reveals significant vulnerabilities in continuous embodied models and highlights the urgent need for defenses targeting the internal generative dynamics of flow-matching VLA models.

Abstract: Vision-Language-Action (VLA) models are emerging as a cornerstone for robotics, with flow-matching policies like $π_0$ showing great promise in generating smooth, continuous actions. As these models advance, their unique action generation mechanism - the vector field dynamics - presents a critical yet unexplored security vulnerability, particularly backdoor vulnerabilities. Existing backdoor attacks designed for autoregressive discretization VLAs cannot be directly applied to this new continuous dynamics. We introduce FlowHijack, the first backdoor attack framework to systematically target the underlying vector-field dynamics of flow-matching VLAs. Our method combines a novel $τ$-conditioned injection strategy, which manipulates the initial phase of the action generation, with a dynamics mimicry regularizer. Experiments demonstrate that FlowHijack achieves high attack success rates using stealthy, context-aware triggers where prior works failed. Crucially, it preserves benign task performance and, by enforcing kinematic similarity, generates malicious actions that are behaviorally indistinguishable from normal actions. Our findings reveal a significant vulnerability in continuous embodied models, highlighting the urgent need for defenses targeting the model’s internal generative dynamics.

Method: Organized a challenge with 37 participants, introduced a new WRV dataset containing 18 videos (1,216 synthesized frames paired with real-world ground-truth frames), and evaluated submissions using protocols considering both fidelity and perceptual quality.

Result: The challenge attracted 37 participants and received 5 valid final submissions with corresponding fact sheets, contributing to progress in weather removal for videos.

Conclusion: The LoViF 2026 Challenge successfully advanced the field of video weather removal by providing a benchmark dataset and evaluation framework, though the paper itself is a review/challenge report rather than presenting novel technical methods.

Abstract: This paper presents a review of the LoViF 2026 Challenge on Weather Removal in Videos. The challenge encourages the development of methods for restoring clean videos from inputs degraded by adverse weather conditions such as rain and snow, with an emphasis on achieving visually plausible and temporally consistent results while preserving scene structure and motion dynamics. To support this task, we introduce a new short-form WRV dataset tailored for video weather removal. It consists of 18 videos 1,216 synthesized frames paired with 1,216 real-world ground-truth frames at a resolution of 832 x 480, and is split into training, validation, and test sets with a ratio of 1:1:1. The goal of this challenge is to advance robust and realistic video restoration under real-world weather conditions, with evaluation protocols that jointly consider fidelity and perceptual quality. The challenge attracted 37 participants and received 5 valid final submissions with corresponding fact sheets, contributing to progress in weather removal for videos. The project is publicly available at https://www.codabench.org/competitions/13462/.

Method: Developed a feature representation based on Checksum Count Vectors to detect duplicates and variants of recordings within large data stores. Tested on a collection of 4902 decoded tape images.

Result: Achieved 58% accuracy in detecting variants and 97% accuracy in identifying alternative copies, even for damaged recordings with up to 75% of records missing.

Conclusion: The approach represents an important step toward fully automated pipelines for restoration, de-duplication, and semantic integration of historical digital artifacts through sequence matching, automatic repair, and knowledge discovery.

Abstract: Digitizing magnetic media containing computer data is only the first step towards the preservation of early home computing era artifacts. The audio tape images must be decoded, verified, repaired if necessary, tested, and documented. If parts of this process could be effectively automated, volunteers could focus on contributing contextual and historical knowledge rather than struggling with technical tools. We therefore propose a feature representation based on Checksum Count Vectors and evaluate its applicability to detecting duplicates and variants of recordings within a large data store. The approach was tested on a collection of decoded tape images (n=4902), achieving 58% accuracy in detecting variants and 97% accuracy in identifying alternative copies, for damaged recordings with up to 75% of records missing. These results represent an important step towards fully automated pipelines for restoration, de-duplication, and semantic integration of historical digital artifacts through sequence matching, automatic repair and knowledge discovery.

Method: Three independent modules: 1) Temporal localization using z-score normalized frame-difference signals with peak detection, 2) Spatial localization using weighted centroid of cumulative dense optical flow magnitude maps (Farneback algorithm), 3) Collision type classification using cosine similarity between CLIP image embeddings of frames near detected peaks and text embeddings from multi-prompt natural language descriptions of collision categories.

Result: A publicly available Kaggle notebook implementation that provides a complete zero-shot pipeline for traffic accident prediction in surveillance videos without requiring any labeled training data or domain-specific fine-tuning.

Conclusion: The proposed zero-shot pipeline successfully addresses the challenge requirements by decomposing the problem into independent modules using pre-trained models, demonstrating the feasibility of accident prediction without labeled surveillance video data.

Abstract: We describe a zero-shot pipeline developed for the ACCIDENT @ CVPR 2026 challenge. The challenge requires predicting when, where, and what type of traffic accident occurs in surveillance video, without labeled real-world training data. Our method separates the problem into three independent modules. The first module localizes the collision in time by running peak detection on z-score normalized frame-difference signals. The second module finds the impact location by computing the weighted centroid of cumulative dense optical flow magnitude maps using the Farneback algorithm. The third module classifies collision type by measuring cosine similarity between CLIP image embeddings of frames near the detected peak and text embeddings built from multi-prompt natural language descriptions of each collision category. No domain-specific fine-tuning is involved; the pipeline processes each video using only pre-trained model weights. Our implementation is publicly available as a Kaggle notebook.

Method: Introduces the novel V2S (video-to-script) task and creates a human-annotated benchmark. Proposes OmniScript, an 8B-parameter omni-modal (audio-visual) language model trained via progressive pipeline: chain-of-thought supervised fine-tuning for plot/character reasoning, followed by reinforcement learning with temporally segmented rewards.

Result: OmniScript significantly outperforms larger open-source models and achieves performance comparable to state-of-the-art proprietary models (including Gemini 3-Pro) in both temporal localization and multi-field semantic accuracy, despite its parameter efficiency.

Conclusion: The proposed OmniScript model effectively addresses the V2S task, demonstrating that parameter-efficient multimodal models can achieve state-of-the-art performance in long-form narrative comprehension through specialized training methodologies.

Abstract: Current multimodal large language models (MLLMs) have demonstrated remarkable capabilities in short-form video understanding, yet translating long-form cinematic videos into detailed, temporally grounded scripts remains a significant challenge. This paper introduces the novel video-to-script (V2S) task, aiming to generate hierarchical, scene-by-scene scripts encompassing character actions, dialogues, expressions, and audio cues. To facilitate this, we construct a first-of-its-kind human-annotated benchmark and propose a temporally-aware hierarchical evaluation framework. Furthermore, we present OmniScript, an 8B-parameter omni-modal (audio-visual) language model tailored for long-form narrative comprehension. OmniScript is trained via a progressive pipeline that leverages chain-of-thought supervised fine-tuning for plot and character reasoning, followed by reinforcement learning using temporally segmented rewards. Extensive experiments demonstrate that despite its parameter efficiency, OmniScript significantly outperforms larger open-source models and achieves performance comparable to state-of-the-art proprietary models, including Gemini 3-Pro, in both temporal localization and multi-field semantic accuracy.

Method: Introduce Grid2Matrix (G2M) benchmark with color grids and color-to-number mappings. Vary grid size and colors to increase visual complexity while minimizing semantic confounds. Test VLMs on zero-shot end-to-end evaluation and probe visual encoders separately.

Result: VLMs show sharp early collapse on surprisingly small grids rather than gradual degradation. Visual encoders preserve more grid information than end-to-end outputs, revealing “Digital Agnosia” gap. Errors are structured and depend on patch boundaries. Model scaling and multimodal alignment don’t fully eliminate failures.

Conclusion: G2M reveals fundamental limitations in VLMs’ ability to translate fine visual details to language, providing a testbed for understanding where VLMs lose visual information and evaluating tasks requiring precise visual detail capture.

Abstract: Vision-Language Models (VLMs) excel on many multimodal reasoning benchmarks, but these evaluations often do not require an exhaustive readout of the image and can therefore obscure failures in faithfully capturing all visual details. We introduce Grid2Matrix (G2M), a controlled benchmark in which a model is shown a color grid and a color-to-number mapping, and must output the corresponding matrix. By varying grid size and the number of colors, G2M provides a simple way to increase visual complexity while minimizing semantic confounds. We find that VLMs exhibit a sharp early collapse in zero-shot end-to-end evaluation, failing on surprisingly small grids rather than degrading gradually as the task becomes denser. We probe the visual encoders of VLMs from two representative families and find that they preserve substantially more of the grid information than the corresponding end-to-end outputs. This suggests that the failure is not explained by visual encoding alone, but also reflects a gap between what remains recoverable from visual features and what is ultimately expressed in language. We term this gap \textit{Digital Agnosia}. Further analyses show that these errors are highly structured and depend strongly on how grid cells overlap with visual patch boundaries. We also find that common strategies such as model scaling and multimodal alignment do not fully eliminate this failure mode. We expect G2M to serve as a useful testbed for understanding where and how VLMs lose fine visual details, and for evaluating tasks where missing even small visual details can matter, such as tables, charts, forms, and GUIs.

Method: Construct hierarchical concept-attribute structured knowledge using LLMs: 1) condense redundant textual labels into abstract concepts, 2) automatically extract discriminative attributes for each concept and similar concept pairs via structured prompts to LLMs, 3) instantiate this knowledge for each input image to create knowledge-enhanced features, 4) combine with original visual features for various clustering algorithms.

Result: Evaluated on 20 diverse datasets, KEC shows consistent improvements over existing methods using additional textual knowledge. Without training, KEC outperforms zero-shot CLIP on 14 out of 20 datasets. Demonstrates that naive use of textual knowledge can harm clustering, while KEC provides both accuracy and robustness.

Conclusion: KEC effectively leverages structured textual knowledge from LLMs to enhance image clustering, addressing limitations of visual-only and naive text-augmented approaches through hierarchical concept-attribute knowledge construction and instantiation.

Abstract: Image clustering aims to group images in an unsupervised fashion. Traditional methods focus on knowledge from visual space, making it difficult to distinguish between visually similar but semantically different classes. Recent advances in vision-language models enable the use of textual knowledge to enhance image clustering. However, most existing methods rely on coarse class labels or simple nouns, overlooking the rich conceptual and attribute-level semantics embedded in textual space. In this paper, we propose a knowledge-enhanced clustering (KEC) method that constructs a hierarchical concept-attribute structured knowledge with the help of large language models (LLMs) to guide clustering. Specifically, we first condense redundant textual labels into abstract concepts and then automatically extract discriminative attributes for each single concept and similar concept pairs, via structured prompts to LLMs. This knowledge is instantiated for each input image to achieve the knowledge-enhanced features. The knowledge-enhanced features with original visual features are adapted to various downstream clustering algorithms. We evaluate KEC on 20 diverse datasets, showing consistent improvements across existing methods using additional textual knowledge. KEC without training outperforms zero-shot CLIP on 14 out of 20 datasets. Furthermore, the naive use of textual knowledge may harm clustering performance, while KEC provides both accuracy and robustness.

Method: GaussLock uses a lightweight parameter-space immunization framework integrating authorized distillation with attribute-aware trap losses targeting position, scale, rotation, opacity, and color. These traps systematically collapse spatial distributions, distort geometric shapes, align rotational axes, and suppress primitive visibility to destroy structural integrity.

Result: Experiments on large-scale Gaussian models show GaussLock effectively neutralizes unauthorized fine-tuning attacks, substantially degrading reconstruction quality (higher LPIPS, lower PSNR) while maintaining performance on authorized fine-tuning.

Conclusion: GaussLock provides the first effective defense for 3D generative models against fine-tuning attacks by embedding parameter-space traps that preserve authorized functionality while disrupting unauthorized reconstructions.

Abstract: Recent large-scale generative models enable high-quality 3D synthesis. However, the public accessibility of pre-trained weights introduces a critical vulnerability. Adversaries can fine-tune these models to steal specialized knowledge acquired during pre-training, leading to intellectual property infringement. Unlike defenses for 2D images and language models, 3D generators require specialized protection due to their explicit Gaussian representations, which expose fundamental structural parameters directly to gradient-based optimization. We propose GaussLock, the first approach designed to defend 3D generative models against fine-tuning attacks. GaussLock is a lightweight parameter-space immunization framework that integrates authorized distillation with attribute-aware trap losses targeting position, scale, rotation, opacity, and color. Specifically, these traps systematically collapse spatial distributions, distort geometric shapes, align rotational axes, and suppress primitive visibility to fundamentally destroy structural integrity. By jointly optimizing these dual objectives, the distillation process preserves fidelity on authorized tasks while the embedded traps actively disrupt unauthorized reconstructions. Experiments on large-scale Gaussian models demonstrate that GaussLock effectively neutralizes unauthorized fine-tuning attacks. It substantially degrades the quality of unauthorized reconstructions, evidenced by significantly higher LPIPS and lower PSNR, while effectively maintaining performance on authorized fine-tuning.

Method: Uses Delaunay-based triplet selection for angular coverage, pose-aware depth module estimating coarse-to-fine depth pyramid, efficient feature reprojection, and occlusion-aware blending. Runs feedforward without retraining.

Result: Outperforms recent real-time novel-view baselines on challenging multi-view video datasets, achieving strong balance of quality and efficiency. Avoids explicit proxies for robust rendering across diverse scenes.

Conclusion: 3DTV provides a practical solution for low-latency multi-view streaming and interactive rendering, making it suitable for AR/VR, telepresence, and interactive applications without requiring scene-specific optimization.

Abstract: Real-time free-viewpoint rendering requires balancing multi-camera redundancy with the latency constraints of interactive applications. We address this challenge by combining lightweight geometry with learning and propose 3DTV, a feedforward network for real-time sparse-view interpolation. A Delaunay-based triplet selection ensures angular coverage for each target view. Building on this, we introduce a pose-aware depth module that estimates a coarse-to-fine depth pyramid, enabling efficient feature reprojection and occlusion-aware blending. Unlike methods that require scene-specific optimization, 3DTV runs feedforward without retraining, making it practical for AR/VR, telepresence, and interactive applications. Our experiments on challenging multi-view video datasets demonstrate that 3DTV consistently achieves a strong balance of quality and efficiency, outperforming recent real-time novel-view baselines. Crucially, 3DTV avoids explicit proxies, enabling robust rendering across diverse scenes. This makes it a practical solution for low-latency multi-view streaming and interactive rendering. Project Page: https://stefanmschulz.github.io/3DTV_webpage/

Method: Proposes MuPPet with three key components: Person Encoding to structure individual representations, Permutation Augmentation to enhance training diversity, and Dynamic Multi-Person Attention to adaptively model correlations between individuals.

Result: Extensive experiments on group interaction datasets show MuPPet significantly outperforms state-of-the-art single- and multi-person 2D-to-3D pose lifting methods, and improves robustness in occlusion scenarios.

Conclusion: The work highlights the importance of modeling inter-person correlations for accurate and socially-aware 3D pose estimation, paving the way for better understanding of multi-person social dynamics.

Abstract: Multi-person social interactions are inherently built on coherence and relationships among all individuals within the group, making multi-person localization and body pose estimation essential to understanding these social dynamics. One promising approach is 2D-to-3D pose lifting which provides a 3D human pose consisting of rich spatial details by building on the significant advances in 2D pose estimation. However, the existing 2D-to-3D pose lifting methods often neglect inter-person relationships or cannot handle varying group sizes, limiting their effectiveness in multi-person settings. We propose MuPPet, a novel multi-person 2D-to-3D pose lifting framework that explicitly models inter-person correlations. To leverage these inter-person dependencies, our approach introduces Person Encoding to structure individual representations, Permutation Augmentation to enhance training diversity, and Dynamic Multi-Person Attention to adaptively model correlations between individuals. Extensive experiments on group interaction datasets demonstrate MuPPet significantly outperforms state-of-the-art single- and multi-person 2D-to-3D pose lifting methods, and improves robustness in occlusion scenarios. Our findings highlight the importance of modeling inter-person correlations, paving the way for accurate and socially-aware 3D pose estimation. Our code is available at: https://github.com/Thomas-Markhorst/MuPPet

Method: Controlled experiments on WIDER FACE and Open Images datasets with perfectly balanced sampling across exactly 1-18 faces per image, measuring performance degradation across classification, regression, and detection paradigms while controlling for class imbalance.

Result: Model performance degrades monotonically with increasing face count across all tasks, with models trained on low-density regimes failing to generalize to higher densities (4.6x error increase), showing systematic under-counting bias and suggesting density acts as a domain shift.

Conclusion: Instance density is an intrinsic, quantifiable dimension of data hardness that motivates interventions in curriculum learning and density-stratified evaluation to address this fundamental challenge.

Abstract: Machine learning progress has historically prioritized model-centric innovations, yet achievable performance is frequently capped by the intrinsic complexity of the data itself. In this work, we isolate and quantify the impact of instance density (measured by face count) as a primary driver of data complexity. Rather than simply observing that ``crowded scenes are harder,’’ we rigorously control for class imbalance to measure the precise degradation caused by density alone. Controlled experiments on the WIDER FACE and Open Images datasets, restricted to exactly 1 to 18 faces per image with perfectly balanced sampling, reveal that model performance degrades monotonically with increasing face count. This trend holds across classification, regression, and detection paradigms, even when models are fully exposed to the entire density range. Furthermore, we demonstrate that models trained on low-density regimes fail to generalize to higher densities, exhibiting a systematic under-counting bias, with error rates increasing by up to 4.6x, which suggests density acts as a domain shift. These findings establish instance density as an intrinsic, quantifiable dimension of data hardness and motivate specific interventions in curriculum learning and density-stratified evaluation.

Method: Introduces a diagnostic framework with two axes: 1) leakage-controlled context ratio (background/foreground) computed from inpainted background-only vs foreground-only images, and 2) laterality diagnostic based on cross-flank retrieval and mirror self-similarity. Creates a Pantanal jaguar benchmark with per-pixel segmentation masks and identity-balanced evaluation protocol. Tests mitigation families including ArcFace fine-tuning, anti-symmetry regularization, and Lorentz hyperbolic embeddings.

Result: Provides a framework to measure what visual evidence models use for re-identification, going beyond standard retrieval metrics to evaluate whether models rely on correct biological features (coat patterns) vs. spurious cues.

Conclusion: The diagnostic framework enables deeper evaluation of wildlife re-ID models by assessing what evidence they use, not just how well they rank, helping ensure models learn biologically meaningful features rather than exploiting dataset biases.

Abstract: Jaguar re-identification (re-ID) from citizen-science imagery can look strong on standard retrieval metrics while still relying on the wrong evidence, such as background context or silhouette shape, instead of the coat pattern that defines identity. We introduce a diagnostic framework for wildlife re-ID with two axes: a leakage-controlled context ratio, background/foreground, computed from inpainted background-only versus foreground-only images, and a laterality diagnostic based on cross-flank retrieval and mirror self-similarity. To make these diagnostics measurable, we curate a Pantanal jaguar benchmark with per-pixel segmentation masks and an identity-balanced evaluation protocol. We then use representative mitigation families, ArcFace fine-tuning, anti-symmetry regularization, and Lorentz hyperbolic embeddings, as case studies under the same evaluation lens. The goal is not only to ask which model ranks best, but also what visual evidence it uses to do so.

Method: CAGE (Code-Anchored Generative Enhancement): First, an LLM synthesizes executable code to produce a structurally correct diagram. Then, a diffusion model conditioned on the programmatic output via ControlNet refines it into a visually polished graphic while preserving label fidelity. Also introduces EduDiagram-2K dataset of 2,000 paired programmatic-stylized diagrams.

Result: Quantified the accuracy-aesthetics dilemma across three paradigms on 400 K-12 diagram prompts using automated and human evaluation. CAGE resolves this tradeoff by combining code-based accuracy with diffusion-based visual enhancement. Proof-of-concept results demonstrate the approach’s effectiveness.

Conclusion: CAGE bridges the gap between accuracy and aesthetics in educational diagram generation, enabling scalable production of both correct and engaging instructional materials. The approach presents a promising research direction for multimedia generation.

Abstract: Educational diagrams – labeled illustrations of biological processes, chemical structures, physical systems, and mathematical concepts – are essential cognitive tools in K-12 instruction. Yet no existing method can generate them both accurately and engagingly. Open-source diffusion models produce visually rich images but catastrophically garble text labels. Code-based generation via LLMs guarantees label correctness but yields visually flat outputs. Closed-source APIs partially bridge this gap but remain unreliable and prohibitively expensive at educational scale. We quantify this accuracy-aesthetics dilemma across all three paradigms on 400 K-12 diagram prompts, measuring both label fidelity and visual quality through complementary automated and human evaluation protocols. To resolve it, we propose CAGE (Code-Anchored Generative Enhancement): an LLM synthesizes executable code producing a structurally correct diagram, then a diffusion model conditioned on the programmatic output via ControlNet refines it into a visually polished graphic while preserving label fidelity. We also introduce EduDiagram-2K, a collection of 2,000 paired programmatic-stylized diagrams enabling this pipeline, and present proof-of-concept results and a research agenda for the multimedia community.

Method: Uses low-cost thermal array sensing to detect falls by modeling them as processes of balance degradation. Key innovations include: (1) appearance-motion fusion model for robust pose reconstruction from low-resolution thermal maps, (2) physically grounded balance-aware learning, and (3) pose-bridged pretraining to improve robustness.

Result: Achieved 98.26% detection rate with 0.65% false alarm rate on dataset with over 3,000 fall instances from 35 participants across diverse indoor environments. In 27-day deployments across four homes, attained ultra-low false alarm rate of 0.00126%. Pilot bathroom study confirmed robustness under moisture and thermal interference.

Conclusion: TaFall establishes a reliable and privacy-preserving approach to fall detection in everyday living environments by combining thermal array sensing with biomechanical balance modeling.

Abstract: Falls are a major cause of injury and mortality among older adults, yet most incidents occur in private indoor environments where monitoring must balance effectiveness with privacy. Existing privacy-preserving fall detection approaches, particularly those based on radio frequency sensing, often rely on coarse motion cues, which limits reliability in real-world deployments. We introduce TaFall, a balance-informed fall detection system based on low-cost, privacy-preserving thermal array sensing. The key insight is that TaFall models a fall as a process of balance degradation and detects falls by estimating pose-driven biomechanical balance dynamics. To enable this capability from low-resolution thermal array maps, we propose (i) an appearance-motion fusion model for robust pose reconstruction, (ii) physically grounded balance-aware learning, and (iii) pose-bridged pretraining to improve robustness. TaFall achieves a detection rate of 98.26% with a false alarm rate of 0.65% on our dataset with over 3,000 fall instances from 35 participants across diverse indoor environments. In 27 day deployments across four homes, TaFall attains an ultra-low false alarm rate of 0.00126% and a pilot bathroom study confirms robustness under moisture and thermal interference. Together, these results establish TaFall as a reliable and privacy-preserving approach to fall detection in everyday living environments.

Method: Created DeceptionDecoded benchmark using intent-guided simulation framework that models both desired influence and execution plans of news creators. Dataset includes 12,000 image-caption pairs (both misleading and non-misleading) grounded in trustworthy reference articles, spanning manipulations across visual and textual modalities. Supports three tasks: misleading intent detection, misleading source attribution, and creator desire inference.

Result: Evaluation of 14 state-of-the-art VLMs shows they struggle with intent reasoning, relying on shallow cues like surface-level alignment, stylistic polish, or heuristic authenticity signals. Models trained on DeceptionDecoded demonstrate strong transferability to real-world MMD, validating the framework as both benchmark and data synthesis engine.

Conclusion: DeceptionDecoded serves as both a diagnostic tool for VLM fragility in intent reasoning and a high-quality data synthesis engine for enhancing robustness in real-world multimodal misinformation governance. The framework enables models to learn implication-level intent reasoning.

Abstract: The impact of multimodal misinformation arises not only from factual inaccuracies but also from the misleading narratives that creators deliberately embed. Interpreting such creator intent is therefore essential for multimodal misinformation detection (MMD) and effective information governance. To this end, we introduce DeceptionDecoded, a large-scale benchmark of 12,000 image-caption pairs grounded in trustworthy reference articles, created using an intent-guided simulation framework that models both the desired influence and the execution plan of news creators. The dataset captures both misleading and non-misleading cases, spanning manipulations across visual and textual modalities, and supports three intent-centric tasks: (1) misleading intent detection, (2) misleading source attribution, and (3) creator desire inference. We evaluate 14 state-of-the-art vision-language models (VLMs) and find that they struggle with intent reasoning, often relying on shallow cues such as surface-level alignment, stylistic polish, or heuristic authenticity signals. To bridge this, our framework systematically synthesizes data that enables models to learn implication-level intent reasoning. Models trained on DeceptionDecoded demonstrate strong transferability to real-world MMD, validating our framework as both a benchmark to diagnose VLM fragility and a data synthesis engine that provides high-quality, intent-focused resources for enhancing robustness in real-world multimodal misinformation governance.

Method: EDFNet is a modular early-fusion segmentation framework integrating RGB, depth, and edge information. Evaluated on DDOS dataset across 16 modality-backbone configurations using U-Net and DeepLabV3 in pretrained and non-pretrained settings.

Result: Early RGB-Depth-Edge fusion provides competitive baseline with consistent gains in boundary-sensitive metrics. Pretrained RGBDE U-Net achieves best overall performance (Thin-Structure Evaluation Score: 0.244, mean IoU: 0.219, boundary IoU: 0.234) at 19.62 FPS. Performance on ultra-thin categories remains low across all models.

Conclusion: Early RGB-Depth-Edge fusion is a practical modular baseline for thin-obstacle segmentation in UAV navigation, but reliable ultra-thin segmentation remains an open challenge.

Abstract: Autonomous Unmanned Aerial Vehicles (UAVs) must reliably detect thin obstacles such as wires, poles, and branches to navigate safely in real-world environments. These structures remain difficult to perceive because they occupy few pixels, often exhibit weak visual contrast, and are strongly affected by class imbalance. Existing segmentation methods primarily target coarser obstacles and do not fully exploit the complementary multimodal cues needed for thin-structure perception. We present EDFNet, a modular early-fusion segmentation framework that integrates RGB, depth, and edge information for thin-obstacle perception in cluttered aerial scenes. We evaluate EDFNet on the Drone Depth and Obstacle Segmentation (DDOS) dataset across sixteen modality-backbone configurations using U-Net and DeepLabV3 in pretrained and non-pretrained settings. The results show that early RGB-Depth-Edge fusion provides a competitive and well-balanced baseline, with the most consistent gains appearing in boundary-sensitive and recall-oriented metrics. The pretrained RGBDE U-Net achieves the best overall performance, with the highest Thin-Structure Evaluation Score (0.244), mean IoU (0.219), and boundary IoU (0.234), while maintaining competitive runtime performance (19.62 FPS) on our evaluation hardware. However, performance on the rarest ultra-thin categories remains low across all models, indicating that reliable ultra-thin segmentation is still an open challenge. Overall, these findings position early RGB-Depth-Edge fusion as a practical and modular baseline for thin-obstacle segmentation in UAV navigation.

Method: Reinforcement learning framework building on Generation Chain-of-Thought, with dual-stage multi-dimensional reward system using MLLMs to evaluate reasoning process and final output across semantic alignment, spatial accuracy, and visual quality

Result: Significant improvements on T2I-CompBench benchmark, especially for compositional tasks with precise spatial relationships and attribute binding

Conclusion: GoT-R1 advances state-of-the-art in image generation by transferring sophisticated reasoning capabilities to visual generation domain

Abstract: Visual generation models have made remarkable progress in creating realistic images from text prompts, yet struggle with complex prompts that specify multiple objects with precise spatial relationships and attributes. Effective handling of such prompts requires explicit reasoning about the semantic content and spatial layout. We present GoT-R1, a framework that applies reinforcement learning to enhance semantic-spatial reasoning in visual generation. Building upon the Generation Chain-of-Thought approach, GoT-R1 enables models to autonomously discover effective reasoning strategies beyond predefined templates through carefully designed reinforcement learning. To achieve this, we propose a dual-stage multi-dimensional reward framework that leverages MLLMs to evaluate both the reasoning process and final output, enabling effective supervision across the entire generation pipeline. The reward system assesses semantic alignment, spatial accuracy, and visual quality in a unified approach. Experimental results demonstrate significant improvements on T2I-CompBench benchmark, particularly in compositional tasks involving precise spatial relationships and attribute binding. GoT-R1 advances the state-of-the-art in image generation by successfully transferring sophisticated reasoning capabilities to the visual generation domain. To facilitate future research, we make our code and pretrained models publicly available at https://github.com/gogoduan/GoT-R1.

Method: The paper investigates how extraction of contextual relationships from images leads to direct (explicit) or indirect (implicit) PII exposure. It proposes methods to protect privacy while preserving the intended utility of images in VLM-based applications, though specific techniques are not detailed in the abstract.

Result: The evaluation demonstrates the efficacy of the proposed privacy protection techniques, highlighting the delicate balance between maintaining utility and protecting privacy in online image processing environments.

Conclusion: The paper addresses critical PII disclosure issues in OVLMs and provides solutions for privacy protection in vision-language model applications, emphasizing the need to balance privacy preservation with functional utility.

Abstract: The increasing use of Online Vision Language Models (OVLMs) for processing images has introduced significant privacy risks, as individuals frequently upload images for various utilities, unaware of the potential for privacy violations. Images contain relationships that relate to Personally Identifiable Information (PII), where even seemingly harmless details can indirectly reveal sensitive information through surrounding clues. This paper explores the critical issue of PII disclosure in images uploaded to OVLMs and its implications for user privacy. We investigate how the extraction of contextual relationships from images can lead to direct (explicit) or indirect (implicit) exposure of PII, significantly compromising personal privacy. Furthermore, we propose methods to protect privacy while preserving the intended utility of the images in Vision Language Model (VLM)-based applications. Our evaluation demonstrates the efficacy of these techniques, highlighting the delicate balance between maintaining utility and protecting privacy in online image processing environments. Index Terms-Personally Identifiable Information (PII), Privacy, Utility, privacy concerns, sensitive information

Method: Conducted systematic empirical study across three MedMNIST v2 benchmarks and four architectures spanning three orders of magnitude in parameter count (21K to 11M). Evaluated TTA with standard augmentation pipelines, analyzed distribution shifts, and performed ablation studies on augmentation strategies.

Result: TTA consistently degrades accuracy relative to single-pass inference, with drops as severe as 31.6 percentage points for ResNet-18 on pathology images. Only exception was ResNet-18 on dermatology images with modest +1.6% gain. Degradation affects all architectures and worsens with more augmented views.

Conclusion: TTA should not be applied as a default post-hoc improvement but must be validated on specific model-dataset combinations. Distribution shift between augmented and training-time inputs, amplified by batch normalization statistics mismatch, is the primary mechanism causing degradation.

Abstract: Test-time augmentation (TTA)–aggregating predictions over multiple augmented copies of a test input–is widely assumed to improve classification accuracy, particularly in medical imaging where it is routinely deployed in production systems and competition solutions. We present a systematic empirical study challenging this assumption across three MedMNIST v2 benchmarks and four architectures spanning three orders of magnitude in parameter count (21K to 11M). Our principal finding is that TTA with standard augmentation pipelines consistently degrades accuracy relative to single-pass inference, with drops as severe as 31.6 percentage points for ResNet-18 on pathology images. This degradation affects all architectures, including convolutional models, and worsens with more augmented views. The sole exception is ResNet-18 on dermatology images, which gains a modest +1.6%. We identify the distribution shift between augmented and training-time inputs–amplified by batch normalization statistics mismatch–as the primary mechanism. Our ablation studies show that augmentation strategy matters critically: intensity-only augmentations preserve more performance than geometric transforms, and including the original unaugmented image partially mitigates but does not eliminate the accuracy drop. These findings serve as a cautionary note for practitioners: TTA should not be applied as a default post-hoc improvement but must be validated on the specific model-dataset combination.

Method: Proposes FMSD with two key components: 1) Forgery-aware Layer Masking identifies forgery-sensitive layers via gradient bias-variance analysis for selective updating, and 2) Multi-Artifact Subspace Decomposition uses SVD to decompose selected layers into semantic and multiple learnable artifact subspaces with orthogonality and spectral consistency constraints.

Result: The framework effectively models diverse forgery patterns while preserving pretrained semantic representations, improving generalization in cross-dataset scenarios.

Conclusion: FMSD addresses limitations of existing deepfake detection methods by selectively updating forgery-sensitive layers and decomposing representations to capture heterogeneous artifacts without compromising semantic understanding.

Abstract: Deepfake detection remains highly challenging, particularly in cross-dataset scenarios and complex real-world settings. This challenge mainly arises because artifact patterns vary substantially across different forgery methods, whereas adapting pretrained models to such artifacts often overemphasizes forgery-specific cues and disturbs semantic representations, thereby weakening generalization. Existing approaches typically rely on full-parameter fine-tuning or auxiliary supervision to improve discrimination. However, they often struggle to model diverse forgery artifacts without compromising pretrained representations. To address these limitations, we propose FMSD, a deepfake detection framework built upon Forgery-aware Layer Masking and Multi-Artifact Subspace Decomposition. Specifically, Forgery-aware Layer Masking evaluates the bias-variance characteristics of layer-wise gradients to identify forgery-sensitive layers, thereby selectively updating them while reducing unnecessary disturbance to pretrained representations. Building upon this, Multi-Artifact Subspace Decomposition further decomposes the selected layer weights via Singular Value Decomposition (SVD) into a semantic subspace and multiple learnable artifact subspaces. These subspaces are optimized to capture heterogeneous and complementary forgery artifacts, enabling effective modeling of diverse forgery patterns while preserving pretrained semantic representations. Furthermore, orthogonality and spectral consistency constraints are imposed to regularize the artifact subspaces, reducing redundancy across them while preserving the overall spectral structure of pretrained weights.

Method: Proposes 3D-GeoFlow, an Attention-Guided Continuous Flow Matching framework that reformulates discrete categorical generation as continuous vector field regression optimized via Mean Squared Error. Uses 3D Attention Gates to dynamically propagate localized borehole features across volumetric latent space for structural coherence.

Result: Extensive out-of-distribution evaluations on a curated dataset of 2,200 procedurally generated 3D geological cases show that 3D-GeoFlow significantly outperforms heuristic interpolations and standard diffusion baselines, achieving a paradigm shift in geological modeling.

Conclusion: 3D-GeoFlow represents a significant advancement in sparse multimodal geological modeling, providing stable deterministic optimal transport paths and addressing key limitations of existing methods for categorical data generation from sparse inputs.

Abstract: Constructing high-resolution 3D geological models from sparse 1D borehole and 2D surface data is a highly ill-posed inverse problem. Traditional heuristic and implicit modeling methods fundamentally fail to capture non-linear topological discontinuities under extreme sparsity, often yielding unrealistic artifacts. Furthermore, while deep generative architectures like Diffusion Models have revolutionized continuous domains, they suffer from severe representation collapse when conditioned on sparse categorical grids. To bridge this gap, we propose 3D-GeoFlow, the first Attention-Guided Continuous Flow Matching framework tailored for sparse multimodal geological modeling. By reformulating discrete categorical generation as a simulation-free, continuous vector field regression optimized via Mean Squared Error, our model establishes stable, deterministic optimal transport paths. Crucially, we integrate 3D Attention Gates to dynamically propagate localized borehole features across the volumetric latent space, ensuring macroscopic structural coherence. To validate our framework, we curated a large-scale multimodal dataset comprising 2,200 procedurally generated 3D geological cases. Extensive out-of-distribution (OOD) evaluations demonstrate that 3D-GeoFlow achieves a paradigm shift, significantly outperforming heuristic interpolations and standard diffusion baselines.

Method: Proposes PASTA pipeline using weak image-level supervision. Compares observed scenes with nominal references to identify Target and Anomaly objects through distribution analysis in self-supervised Vision Transformer feature spaces. Uses semantic text-prompts via Segment Anything Model 3 for zero-shot object segmentation guidance.

Result: 75.8% training time reduction compared to domain-specific baselines. Achieves up to 88.3% IoU for Target segmentation and up to 63.5% IoU for Anomaly segmentation in industrial and agricultural domains. Demonstrated on custom steel scrap recycling and plant datasets.

Conclusion: PASTA provides a domain-agnostic weakly supervised approach that reduces annotation burden while maintaining high segmentation performance for target and anomaly detection in unstructured environments.

Abstract: Detecting unseen anomalies in unstructured environments presents a critical challenge for industrial and agricultural applications such as material recycling and weeding. Existing perception systems frequently fail to satisfy the strict operational requirements of these domains, specifically real-time processing, pixel-level segmentation precision, and robust accuracy, due to their reliance on exhaustively annotated datasets. To address these limitations, we propose a weakly supervised pipeline for object segmentation and classification using weak image-level supervision called ‘Patch Aggregation for Segmentation of Targets and Anomalies’ (PASTA). By comparing an observed scene with a nominal reference, PASTA identifies Target and Anomaly objects through distribution analysis in self-supervised Vision Transformer (ViT) feature spaces. Our pipeline utilizes semantic text-prompts via the Segment Anything Model 3 to guide zero-shot object segmentation. Evaluations on a custom steel scrap recycling dataset and a plant dataset demonstrate a 75.8% training time reduction of our approach to domain-specific baselines. While being domain-agnostic, our method achieves superior Target (up to 88.3% IoU) and Anomaly (up to 63.5% IoU) segmentation performance in the industrial and agricultural domain.

Method: Proposes Identity-Aware U-Net (IAU-Net) with U-Net-style encoder-decoder architecture augmented by an auxiliary embedding branch that learns discriminative identity representations from high-level features. Incorporates triplet-based metric learning to pull target-consistent embeddings together and separate them from hard negatives with similar morphology.

Result: Experiments on benchmarks including cell segmentation demonstrate promising results, particularly in challenging cases involving similar contours, dense layouts, and ambiguous boundaries.

Conclusion: The framework enables models to move beyond category-level segmentation and acquire stronger capability for precise discrimination among visually similar objects through joint modeling of spatial localization and instance discrimination.

Abstract: Precise segmentation of objects with highly similar shapes remains a challenging problem in dense prediction, especially in scenarios with ambiguous boundaries, overlapping instances, and weak inter-instance visual differences. While conventional segmentation models are effective at localizing object regions, they often lack the discriminative capacity required to reliably distinguish a target object from morphologically similar distractors. In this work, we study fine-grained object segmentation from an identity-aware perspective and propose Identity-Aware U-Net (IAU-Net), a unified framework that jointly models spatial localization and instance discrimination. Built upon a U-Net-style encoder-decoder architecture, our method augments the segmentation backbone with an auxiliary embedding branch that learns discriminative identity representations from high-level features, while the main branch predicts pixel-accurate masks. To enhance robustness in distinguishing objects with near-identical contours or textures, we further incorporate triplet-based metric learning, which pulls target-consistent embeddings together and separates them from hard negatives with similar morphology. This design enables the model to move beyond category-level segmentation and acquire a stronger capability for precise discrimination among visually similar objects. Experiments on benchmarks including cell segmentation demonstrate promising results, particularly in challenging cases involving similar contours, dense layouts, and ambiguous boundaries.

Method: Proposes MG-IQA with: 1) attribute-aware prompting for structured multi-attribute reasoning from VLMs, 2) multi-dimensional Thurstone reward model for attribute-specific fidelity rewards, 3) cross-domain alignment for stable joint training across synthetic/authentic distortion and AI-generated image datasets.

Result: Extensive experiments on eight IQA benchmarks show MG-IQA outperforms SOTA methods in overall quality prediction (average SRCC improvement of 2.1%) and attribute-level assessment, while generating interpretable, human-aligned quality descriptions.

Conclusion: MG-IQA successfully extends RL2R to multi-granularity IQA, enabling joint assessment of overall quality and fine-grained attributes with improved performance and interpretability.

Abstract: Recent advances in reasoning-induced image quality assessment (IQA) have demonstrated the power of reinforcement learning to rank (RL2R) for training vision-language models (VLMs) to assess perceptual quality. However, existing approaches operate at a single granularity, predicting only an overall quality score, while overlooking the multi-dimensional nature of human quality perception, which encompasses attributes such as sharpness, color fidelity, noise level, and compositional aesthetics. In this paper, we propose MG-IQA (Multi-Granularity IQA), a multi-granularity reasoning framework that extends RL2R to jointly assess overall image quality and fine-grained quality attributes within a single inference pass. Our approach introduces three key innovations: (1) an attribute-aware prompting strategy that elicits structured multi-attribute reasoning from VLMs; (2) a multi-dimensional Thurstone reward model that computes attribute-specific fidelity rewards for group relative policy optimization; and (3) a cross-domain alignment mechanism that enables stable joint training across synthetic distortion, authentic distortion, and AI-generated image datasets without perceptual scale re-alignment. Extensive experiments on eight IQA benchmarks demonstrate that MG-IQA consistently outperforms state-of-the-art methods in both overall quality prediction (average SRCC improvement of 2.1%) and attribute-level assessment, while generating interpretable, human-aligned quality descriptions.

Method: Introduces a platform-aware adversarial evaluation framework that models deployment transforms (resizing, compression, screenshot-style distortions) and constrains perturbations to visually plausible meme-style bands rather than full-image noise. Evaluates detectors under this threat model with per-image and universal perturbation attacks.

Result: Detectors achieving AUC ≈ 0.99 in clean settings experience substantial degradation under platform-aware attacks. Per-image attacks reduce AUC significantly and achieve high fake-to-real misclassification rates. Universal perturbations exist even under localized band constraints. Detectors also show pronounced calibration collapse, becoming confidently incorrect under attack.

Conclusion: Robustness measured under clean conditions substantially overestimates deployment robustness. Platform-aware evaluation should be a necessary component of future AI media security benchmarks. The evaluation framework is released to facilitate standardized robustness assessment.

Abstract: Recent AI media detectors report near-perfect performance under clean laboratory evaluation, yet their robustness under realistic deployment conditions remains underexplored. In practice, AI-generated images are resized, compressed, re-encoded, and visually modified before being shared on online platforms. We argue that this creates a deployment gap between laboratory robustness and real-world reliability. In this work, we introduce a platform-aware adversarial evaluation framework for AI media detection that explicitly models deployment transforms (e.g., resizing, compression, screenshot-style distortions) and constrains perturbations to visually plausible meme-style bands rather than full-image noise. Under this threat model, detectors achieving AUC $\approx$ 0{.}99 in clean settings experience substantial degradation. Per-image platform-aware attacks reduce AUC to significantly lower levels and achieve high fake-to-real misclassification rates, despite strict visual constraints. We further demonstrate that universal perturbations exist even under localized band constraints, revealing shared vulnerability directions across inputs. Beyond accuracy degradation, we observe pronounced calibration collapse under attack, where detectors become confidently incorrect. Our findings highlight that robustness measured under clean conditions substantially overestimates deployment robustness. We advocate for platform-aware evaluation as a necessary component of future AI media security benchmarks and release our evaluation framework to facilitate standardized robustness assessment.

Method: Proposes Orthogonal Quadratic Complements (OQC) which construct a low-rank quadratic auxiliary branch and explicitly project it onto the orthogonal complement of the main branch before injection. Also studies efficient low-rank realization (OQC-LR) and gated extensions (OQC-static and OQC-dynamic).

Result: On CIFAR-100, full OQC improves AFBO baseline from 64.25 to 65.59, while OQC-LR reaches 65.52 with better speed-accuracy tradeoff. On TinyImageNet, OQC-dynamic achieves 51.88, improving baseline (50.45) by 1.43 points. Mechanism analyses show near-zero post-projection auxiliary-main overlap with improved representation geometry and class separation.

Conclusion: OQC provides a principled approach to incorporate complementary quadratic features in vision transformers through orthogonal projection, achieving consistent improvements across datasets with both ungated and gated variants.

Abstract: Recent bilinear feed-forward replacements for vision transformers can substantially improve accuracy, but they often conflate two effects: stronger second-order interactions and increased redundancy relative to the main branch. We study a complementary design principle in which auxiliary quadratic features contribute only information not already captured by the dominant hidden representation. To this end, we propose Orthogonal Quadratic Complements (OQC), which construct a low-rank quadratic auxiliary branch and explicitly project it onto the orthogonal complement of the main branch before injection. We further study an efficient low-rank realization (OQC-LR) and gated extensions (OQC-static and OQC-dynamic). Under a parameter-matched Deep-ViT and CIFAR-100 protocol with a fixed penultimate residual readout, full OQC improves an AFBO baseline from 64.25 +/- 0.22 to 65.59 +/- 0.22, while OQC-LR reaches 65.52 +/- 0.25 with a substantially better speed-accuracy tradeoff. On TinyImageNet, the gated extension OQC-dynamic achieves 51.88 +/- 0.32, improving the baseline (50.45 +/- 0.21) by 1.43 points and outperforming all ungated variants. Mechanism analyses show near-zero post-projection auxiliary-main overlap together with improved representation geometry and class separation. The full family, including both ungated and gated variants, generalizes consistently across both datasets.

Method: Proposes a two-level objective: 1) Logit-adjusted cross-entropy loss at sample level to handle class imbalance with provable consistency guarantees, and 2) Conditional Value at Risk aggregation at group level to direct optimization toward demographic groups with higher loss.

Result: On Fair Disease Diagnosis benchmark using 3D ResNet-18, achieves gender-averaged macro F1 of 0.8403 with fairness gap of 0.0239, showing 13.3% improvement in score and 78% reduction in demographic disparity over baseline. Ablations confirm both components are necessary.

Conclusion: The combined approach effectively addresses compound failure modes in clinical data where class imbalance and group underrepresentation coincide, outperforming methods that target only one aspect.

Abstract: Automated diagnosis from chest CT has improved considerably with deep learning, but models trained on skewed datasets tend to perform unevenly across patient demographics. However, the situation is worse than simple demographic bias. In clinical data, class imbalance and group underrepresentation often coincide, creating compound failure modes that neither standard rebalancing nor fairness corrections can fix alone. We introduce a two-level objective that targets both axes of this problem. Logit-adjusted cross-entropy loss operates at the sample level, shifting decision margins by class frequency with provable consistency guarantees. Conditional Value at Risk aggregation operates at the group level, directing optimization pressure toward whichever demographic group currently has the higher loss. We evaluate on the Fair Disease Diagnosis benchmark using a 3D ResNet-18 pretrained on Kinetics-400, classifying CT volumes into Adenocarcinoma, Squamous Cell Carcinoma, COVID-19, and Normal groups with patient sex annotations. The training set illustrates the compound problem concretely: squamous cell carcinoma has 84 samples total, 5 of them female. The combined loss reaches a gender-averaged macro F1 of 0.8403 with a fairness gap of 0.0239, a 13.3% improvement in score and 78% reduction in demographic disparity over the baseline. Ablations show that each component alone falls short. The code is publicly available at https://github.com/Purdue-M2/Fair-Disease-Diagnosis.

Method: Leverages detections on camera images to create semantic parametric maps, models map elements to produce pseudo labels in LiDAR domain, extends angular perception from front-view camera to full 360° view without direct sensor overlap.

Result: Outperforms single shot baselines by +19.5 mIoU for 2D semantic segmentation, +19.5 PQth for 2D panoptic segmentation, and +32.3 mIoU in 3D semantic segmentation on large-scale road feature dataset.

Conclusion: Demonstrates effective cross-modal domain adaptation achieving strong LiDAR performance without manual labeling, bridging camera and LiDAR sensing domains.

Abstract: Until open-world foundation models match the performance of specialized approaches, deep learning systems remain dependent on task- and sensor-specific data availability. To bridge the gap between available datasets and deployment domains, domain adaptation strategies are widely used. In this work, we propose XD-MAP, a novel approach to transfer sensor-specific knowledge from an image dataset to LiDAR, an entirely different sensing domain. Our method leverages detections on camera images to create a semantic parametric map. The map elements are modeled to produce pseudo labels in the target domain without any manual annotation effort. Unlike previous domain transfer approaches, our method does not require direct overlap between sensors and enables extending the angular perception range from a front-view camera to a full 360° view. On our large-scale road feature dataset, XD-MAP outperforms single shot baseline approaches by +19.5 mIoU for 2D semantic segmentation, +19.5 PQth for 2D panoptic segmentation, and +32.3 mIoU in 3D semantic segmentation. The results demonstrate the effectiveness of our approach achieving strong performance on LiDAR data without any manual labeling.

Method: Head-wise modality specialization within MLLMs: 1) Study attention heads and their relationship with performance under missing modality, 2) Head-wise specialization mechanism that allocates heads to different modalities with lower-bound attention constraints, 3) Unimodal Knowledge Retention strategy to prevent heads from drifting away from unimodal knowledge learned from limited supervision.

Result: The method improves robustness under missing modality while preserving performance with full multimodal input.

Conclusion: The proposed head-wise modality specialization approach effectively addresses the missing modality problem in multimodal fake news detection by preserving unimodal verification abilities through specialized attention heads and knowledge retention strategies.

Abstract: Multimodal fake news detection (MFND) aims to verify news credibility by jointly exploiting textual and visual evidence. However, real-world news dissemination frequently suffers from missing modality due to deleted images, corrupted screenshots, and similar issues. Thus, robust detection in this scenario requires preserving strong verification ability for each modality, which is challenging in MFND due to insufficient learning of the low-contribution modality and scarce unimodal annotations. To address this issue, we propose Head-wise Modality Specialization within Multimodal Large Language Models (MLLMs) for robust MFND under missing modality. Specifically, we first systematically study attention heads in MLLMs and their relationship with performance under missing modality, showing that modality-critical heads serve as key carriers of unimodal verification ability through their modality specialization. Based on this observation, to better preserve verification ability for the low-contribution modality, we introduce a head-wise specialization mechanism that explicitly allocates these heads to different modalities and preserves their specialization through lower-bound attention constraints. Furthermore, to better exploit scarce unimodal annotations, we propose a Unimodal Knowledge Retention strategy that prevents these heads from drifting away from the unimodal knowledge learned from limited supervision. Experiments show that our method improves robustness under missing modality while preserving performance with full multimodal input.

Method: Proposes LAST framework with LAST-Box sandbox that abstracts heterogeneous tool invocations into atomic instructions and reusable spatial skills, returning multimodal hints. Uses three-stage progressive training: understanding tool outputs, then proficient and adaptive tool invocation.

Result: LAST-7B achieves ~20% performance gains over its backbone on four datasets, outperforms strong proprietary closed-source LLMs, and substantially enhances reasoning on complex spatial tasks.

Conclusion: Tool-augmented approach effectively addresses MLLMs’ spatial reasoning limitations by integrating specialized vision models through unified abstraction and progressive training.

Abstract: Spatial reasoning is a cornerstone capability for intelligent systems to perceive and interact with the physical world. However, multimodal large language models (MLLMs) frequently suffer from hallucinations and imprecision when parsing complex geometric layouts. As data-driven scaling struggles to internalize structured geometric priors and spatial constraints, integrating mature, specialized vision models presents a compelling alternative. Despite its promise, applying this paradigm to spatial reasoning is hindered by two key challenges: The difficulty of invoking heterogeneous, parameter-rich tools, as well as the challenge of understanding and effectively leveraging their diverse low-level outputs (e.g., segmentation masks, depth maps) in high-level reasoning. To address these challenges, we propose LAST, a unified framework for tool-augmented spatial reasoning. LAST features an extensible interactive sandbox, termed LAST-Box, which abstracts heterogeneous tool invocations into atomic instructions and reusable spatial skills, returning multimodal hints (e.g., annotated images and textual descriptions) that can be directly consumed by LLMs. We further design a three-stage progressive training strategy that guides models from understanding tool outputs to proficient and adaptive tool invocation. Experiments on four datasets show that LAST-7B achieves around 20% performance gains over its backbone and outperforms strong proprietary closed-source LLMs, substantially enhancing reasoning on complex spatial tasks.

Method: Uses vision-language models to generate rich soundscape descriptions, then learns jointly from audio, text descriptions, satellite images, and synthetic captions through contrastive and codebook-aligned learning to discover shared “soundscape concepts” across modalities.

Result: Achieves state-of-the-art performance in cross-modal retrieval between satellite images and audio on GeoSound and SoundingEarth benchmarks. Enables location-conditioned soundscape synthesis through text-to-audio models.

Conclusion: Sat2Sound provides a unified framework for explainable soundscape mapping and synthesis, overcoming data limitations through multimodal learning and enabling immersive applications with limited computational resources.

Abstract: We present Sat2Sound, a unified multimodal framework for geospatial soundscape understanding, designed to predict and map the distribution of sounds across the Earth’s surface. Existing methods for this task rely on paired satellite images and geotagged audio samples, which often fail to capture the full diversity of sound at a location. Sat2Sound overcomes this limitation by augmenting datasets with semantically rich, vision-language model-generated soundscape descriptions, which broaden the range of possible ambient sounds represented at each location. Our framework jointly learns from audio, text descriptions of audio, satellite images, and synthetic image captions through contrastive and codebook-aligned learning, discovering a set of “soundscape concepts” shared across modalities, enabling hyper-localized, explainable soundscape mapping. Sat2Sound achieves state-of-the-art performance in cross-modal retrieval between satellite image and audio on the GeoSound and SoundingEarth benchmarks. Finally, by retrieving detailed soundscape captions that can be rendered through text-to-audio models, Sat2Sound enables location-conditioned soundscape synthesis for immersive and educational applications, even with limited computational resources. Our code and models are available at https://github.com/mvrl/sat2sound.

Method: Learn how model parameters change when moving from synthetic to real handwriting in source languages, then transfer this learned correction to new target languages. When using multiple sources, use linguistic similarity to weigh their contributions when combining them.

Result: Experiments across five languages and six architectures show consistent improvements over synthetic-only baselines. The transferred corrections benefit even languages unrelated to the source languages.

Conclusion: The proposed zero-shot adaptation method effectively transfers learned parameter corrections from source to target languages without requiring real data from target domains, demonstrating cross-linguistic generalization capabilities.

Abstract: Handwritten Text Recognition (HTR) models trained on synthetic handwriting often struggle to generalize to real text, and existing adaptation methods still require real samples from the target domain. In this work, we tackle the fully zero-shot synthetic-to-real generalization setting, where no real data from the target language is available. Our approach learns how model parameters change when moving from synthetic to real handwriting in one or more source languages and transfers this learned correction to new target languages. When using multiple sources, we rely on linguistic similarity to weigh their contrubition when combining them. Experiments across five languages and six architectures show consistent improvements over synthetic-only baselines and reveal that the transferred corrections benefit even languages unrelated to the sources.

Method: Uses signal analysis techniques from biological neural activity studies to define three measures from layer activations: integration score (long-range coordination), metastability score (flexibility between synchronized states), and combined dynamical stability index. Applied to 9 model-dataset combinations including ResNet variants, DenseNet-121, MobileNetV2, VGG-16, and Vision Transformer on CIFAR-10/100.

Result: Three main patterns: 1) Integration measure distinguishes easier CIFAR-10 from harder CIFAR-100; 2) Stability index volatility changes may signal convergence before accuracy plateaus; 3) Integration-metastability relationship reflects different training behaviors.

Conclusion: Provides exploratory but promising new framework to understand deep visual training dynamics beyond traditional metrics, offering insights into internal representation changes during training.

Abstract: Deep visual recognition models are usually trained and evaluated using metrics such as loss and accuracy. While these measures show whether a model is improving, they reveal very little about how its internal representations change during training. This paper introduces a complementary way to study that process by examining training through the lens of dynamical systems. Drawing on ideas from signal analysis originally used to study biological neural activity, we define three measures from layer activations collected across training epochs: an integration score that reflects long-range coordination across layers, a metastability score that captures how flexibly the network shifts between more and less synchronised states, and a combined dynamical stability index. We apply this framework to nine combinations of model architecture and dataset, including several ResNet variants, DenseNet-121, MobileNetV2, VGG-16, and a pretrained Vision Transformer on CIFAR-10 and CIFAR-100. The results suggest three main patterns. First, the integration measure consistently distinguishes the easier CIFAR-10 setting from the more difficult CIFAR-100 setting. Second, changes in the volatility of the stability index may provide an early sign of convergence before accuracy fully plateaus. Third, the relationship between integration and metastability appears to reflect different styles of training behaviour. Overall, this study offers an exploratory but promising new way to understand deep visual training beyond loss and accuracy.

Method: Created a new balanced dataset with diverse demographic representation, then proposed an interaction-aware hybrid architecture integrating EfficientNetB3, Vision Transformer, and Conformer modules in parallel with multi-head cross-attention fusion.

Result: Achieved 98.84% accuracy on the constructed dataset and 96.49% on the external CHBCR benchmark, demonstrating strong generalization. Grad-CAM visualizations provide interpretability.

Conclusion: The proposed dataset and hybrid architecture effectively address Bangla handwritten character recognition challenges, with publicly available resources for further research.

Abstract: Character recognition is the fundamental part of an optical character recognition (OCR) system. Word recognition, sentence transcription, document digitization, and language processing are some of the higher-order activities that can be done accurately through character recognition. Nonetheless, recognizing handwritten Bangla characters is not an easy task because they are written in different styles with inconsistent stroke patterns and a high degree of visual character resemblance. The datasets available are usually limited in intra-class and inequitable in class distribution. We have constructed a new balanced dataset of Bangla written characters to overcome those problems. This consists of 78 classes and each class has approximately 650 samples. It contains the basic characters, composite (Juktobarno) characters and numerals. The samples were a diverse group comprising a large age range and socioeconomic groups. Elementary and high school students, university students, and professionals are the contributing factors. The sample also has right and left-handed writers. We have further proposed an interaction-aware hybrid deep learning architecture that integrates EfficientNetB3, Vision Transformer, and Conformer modules in parallel. A multi-head cross-attention fusion mechanism enables effective feature interaction across these components. The proposed model achieves 98.84% accuracy on the constructed dataset and 96.49% on the external CHBCR benchmark, demonstrating strong generalization capability. Grad-CAM visualizations further provide interpretability by highlighting discriminative regions. The dataset and source code of this research is publicly available at: https://huggingface.co/MIRZARAQUIB/Bangla_Handwritten_Character_Recognition.

Method: Proposes three complementary metrics: Homogeneity Index (HI) quantifying statistical heterogeneity, Representative Elementary Area (REA) from Minkowski-functionals, and Total Variation Energy (TVE) index for sensitivity to localized anomalies.

Result: Validated on pulse-heated IRT data from CFRP plate with artificial defects at various depths. Demonstrates robust, unbiased ranking of image sequences for automated defect-oriented image selection.

Conclusion: Provides a reliable data-driven methodology for identifying defect-representative images in IRT without requiring prior spatial information, enabling automated analysis.

Abstract: Infrared thermography (IRT) is a widely used non-destructive testing technique for detecting structural features such as subsurface defects. However, most IRT post-processing methods generate image sequences in which defect visibility varies strongly across time, frequency, or coefficient/index domains, making the identification of defect-representative images a critical challenge. Conventional evaluation metrics, such as the signal-to-noise ratio (SNR) or the Tanimoto criterion, often require prior knowledge of defect locations or defect-free reference regions, limiting their suitability for automated and unsupervised analysis. In this work, a data-driven methodology is proposed to identify images within IRT datasets that are most likely to contain and represent structural features, particularly anomalies and defects, without requiring prior spatial information. The approach is based on three complementary metrics: the Homogeneity Index of Mixture (HI), which quantifies statistical heterogeneity via deviations of local intensity distributions from a global reference distribution; a Representative Elementary Area (REA), derived from a Minkowski-functional adaptation of the Representative Elementary Volume concept to two-dimensional images; and a geometrical-topological Total Variation Energy (TVE) index, also based on two-dimensional Minkowski functionals, designed to improve sensitivity to localized anomalies. The framework is validated experimentally using pulse-heated IRT data from a carbon fiber-reinforced polymer (CFRP) plate containing six artificial defects at depths between 0.135 mm and 0.810 mm, and is further supported by one-dimensional N-layer thermal model simulations. The results demonstrate robust and unbiased ranking of image sequences and provide a reliable basis for automated defect-oriented image selection in IRT.

Method: A modular multi-agent framework with three core modules: video content summarization, video classification, and comment generation with semantic retrieval and hot meme augmentation. Supports six controllable comment styles and uses a bilingual dataset of 3,267 videos and 16,335 comments across YouTube and Douyin.

Result: LOLGORITHM consistently outperforms baseline methods, achieving human preference selection rates of 80.46% on YouTube and 84.29% on Douyin across 107 respondents. Ablation studies show gains are due to framework architecture rather than backbone LLM choice.

Conclusion: The framework demonstrates robustness and generalizability for generating authentic, stylized comments that conform to platform-specific cultural norms, addressing a gap in existing approaches for short-form video engagement.

Abstract: Short-form video platforms have become central to multimedia information dissemination, where comments play a critical role in driving engagement, propagation, and algorithmic feedback. However, existing approaches – including video summarization and live-streaming danmaku generation – fail to produce authentic comments that conform to platform-specific cultural and linguistic norms. In this paper, we present LOLGORITHM, a novel modular multi-agent framework for stylized short-form video comment generation. LOLGORITHM supports six controllable comment styles and comprises three core modules: video content summarization, video classification, and comment generation with semantic retrieval and hot meme augmentation. We further construct a bilingual dataset of 3,267 videos and 16,335 comments spanning five high-engagement categories across YouTube and Douyin. Evaluation combining automatic scoring and large-scale human preference analysis demonstrates that LOLGORITHM consistently outperforms baseline methods, achieving human preference selection rates of 80.46% on YouTube and 84.29% on Douyin across 107 respondents. Ablation studies confirm that these gains are attributable to the framework architecture rather than the choice of backbone LLM, underscoring the robustness and generalizability of our approach.

Method: Modular hierarchical framework with: (1) fixed multi-frequency Gabor decomposition into 7 parallel streams, (2) within-stream competitive learning using Hebbian/Oja updates and anti-Hebbian decorrelation, (3) associative memory module inspired by modern Hopfield retrieval, and (4) iterative top-down modulation using local prediction/reconstruction signals. Only final linear readout and top-down projection matrices are optimized by gradient descent.

Result: On CIFAR-10: 80.1% accuracy (linear probe) vs 71.0% for Hebbian-only baseline and 83.4% for gradient-trained model on same fixed Gabor basis. On CIFAR-100: 54.8% accuracy. Factorial analysis shows multi-frequency streams, associative memory, and top-down feedback contribute additively with significant Streams x TopDown interaction.

Conclusion: Carefully chosen architectural priors can recover substantial performance typically associated with global gradient training, though a measurable residual gap remains. The hybrid approach demonstrates viability of predominantly locally-trained systems with minimal gradient optimization.

Abstract: We study how far structured architectural bias can compensate for the absence of end-to-end gradient-based representation learning in visual recognition. Building on the VisNet tradition, we introduce a modular hierarchical framework combining: (i) fixed multi-frequency Gabor decomposition into F=7 parallel streams; (ii) within-stream competitive learning with Hebbian and Oja updates and anti-Hebbian decorrelation; (iii) an associative memory module inspired by modern Hopfield retrieval; and (iv) iterative top-down modulation using local prediction and reconstruction signals. Representational layers are trained without end-to-end backpropagation through the full hierarchy; only the final linear readout and top-down projection matrices are optimized by gradient descent. We therefore interpret the model as a hybrid system that is predominantly locally trained but includes a small number of gradient-trained parameters. On CIFAR-10, the full model reaches 80.1% +/- 0.3% top-1 accuracy, linear probe), compared with 71.0% for a Hebbian-only baseline and 83.4% for a gradient-trained model on the same fixed Gabor basis. On CIFAR-100, performance is 54.8%. Factorial analysis indicates that multi-frequency streams, associative memory, and top-down feedback contribute largely additively, with a significant Streams x TopDown interaction (p=0.02). These results suggest that carefully chosen architectural priors can recover a substantial fraction of the performance typically associated with global gradient training, while leaving a measurable residual gap. Experiments are limited to CIFAR-10/100.

Method: Proposes DOP-OBC: Dominant Object Penalty (DOP) suppresses attention on visually dominant regions, while Outlier Boost Coefficient (OBC) amplifies attention for rare but confidently detected objects. Implemented as per-row logit modulations in causal attention mask without weight updates.

Result: Consistent reduction in object hallucination on CHAIR and POPE benchmarks, improved GPT-4o assessed captioning quality across correctness, consistency, detail, context, and temporal dimensions for both image and video MLLMs.

Conclusion: Fairness in attention allocation is a practical and effective approach for more faithful multimodal generation, with DOP-OBC demonstrating that equitable attention can reduce hallucinations without requiring model retraining.

Abstract: Multimodal large language models (MLLMs) frequently hallucinate objects that are absent from the visual input, often because attention during decoding is disproportionately drawn to visually dominant or frequently occurring content. We observe that this inequity in attention allocation is a root cause of object hallucination: when rare, small, or contextually peripheral objects receive insufficient attention, the model fails to ground its generation in the full visual scene. We argue that every object in an image, regardless of its size, frequency or visual salience, deserves equal representational opportunity during decoding. To this end, we propose DOP-OBC, a training-free and architecture-agnostic decoding strategy built on the principle of equitable attention. Two complementary object-aware signals work in tandem: a Dominant Object Penalty (DOP) that softly suppresses attention over-concentration on visually dominant regions, and an Outlier Boost Coefficient (OBC) that amplifies attention toward rare yet confidently detected objects. These signals are injected as per-row logit modulations within the causal attention mask, requiring no weight updates and preserving autoregressive decoding properties. Extensive experiments across image and video MLLMs demonstrate consistent reductions in object hallucination on CHAIR and POPE benchmarks, alongside improvements in GPT-4o assessed captioning quality across correctness, consistency, detail, context and temporal dimensions. DOP-OBC establishes that fairness in attention allocation is not merely a design principle but a practical and effective path toward more faithful multimodal generation.

Method: Introduces latent visual reasoning framework with explicit visual evidence state in autoregressive decoding, using short latent reasoning segments that reuse hidden states as continuous latent steps. Two-stage training: ROI-supervised fine-tuning aligns latent states with clinical image evidence, and Visual-Latent Policy Optimization (VLPO) optimizes reasoning under outcome-level rewards.

Result: Outperforms recent reasoning baselines on OmniMedVQA and five external medical VQA benchmarks, improving average score from 48.3% to 53.4% over Qwen2.5-VL-7B backbone.

Conclusion: Latent visual reasoning provides effective mechanism for preserving diagnostically relevant visual evidence and improving reliability of medical VQA, addressing limitations of text-centric approaches.

Abstract: Medical vision–language models (VLMs) have shown strong potential for medical visual question answering (VQA), yet their reasoning remains largely text-centric: images are encoded once as static context, and subsequent inference is dominated by language. This paradigm is fundamentally limited in clinical scenarios, where accurate answers often depend on subtle, localized visual evidence that cannot be reliably preserved in static embeddings. We propose \textsc{MedLVR}, a latent visual reasoning framework that introduces an explicit visual evidence state into autoregressive decoding. Instead of relying solely on text-based intermediate reasoning, \textsc{MedLVR} interleaves a short latent reasoning segment within the decoder by reusing hidden states as continuous latent steps, enabling iterative preservation and refinement of query-relevant visual evidence before answer generation. To support effective visual supervision, we adopt a two-stage training strategy: region of interest (ROI)-supervised fine-tuning aligns latent states with clinically relevant image evidence, and Visual-Latent Policy Optimization (VLPO) further optimizes latent reasoning and answer generation under outcome-level rewards. Experiments on OmniMedVQA and five external medical VQA benchmarks show that \textsc{MedLVR} consistently outperforms recent reasoning baselines and improves the average score over the Qwen2.5-VL-7B backbone from 48.3% to 53.4%. These results show that latent visual reasoning provides an effective mechanism for preserving diagnostically relevant visual evidence and improving the reliability of medical VQA.

Method: Closed-loop interaction where VLM acts as an agent: observe scene, evaluate faithfulness to instruction, propose pose update, apply update, render updated scene. Key techniques: multi-view reasoning with view selection, object-centered coordinate visualization, single-axis rotation prediction.

Result: Approach surpasses prior methods at predicting text-guided goal 6D poses, works consistently across closed/open-source VLMs, enables more successful robot manipulation when combined with motion planning.

Conclusion: Inference-time techniques and iterative reasoning enable VLMs to achieve dramatic 3D understanding improvements without fine-tuning, bridging gap between 2D vision-language capabilities and 3D reasoning.

Abstract: Vision-Language Models (VLMs) exhibit strong visual reasoning capabilities, yet they still struggle with 3D understanding. In particular, VLMs often fail to infer a text-consistent goal 6D pose of a target object in a 3D scene. However, we find that with some inference-time techniques and iterative reasoning, VLMs can achieve dramatic performance gains. Concretely, given a 3D scene represented by an RGB-D image (or a compositional scene of 3D meshes) and a text instruction specifying a desired state change, we repeat the following loop: observe the current scene; evaluate whether it is faithful to the instruction; propose a pose update for the target object; apply the update; and render the updated scene. Through this closed-loop interaction, the VLM effectively acts as an agent. We further introduce three inference-time techniques that are essential to this closed-loop process: (i) multi-view reasoning with supporting view selection, (ii) object-centered coordinate system visualization, and (iii) single-axis rotation prediction. Without any additional fine-tuning or new modules, our approach surpasses prior methods at predicting the text-guided goal 6D pose of the target object. It works consistently across both closed-source and open-source VLMs. Moreover, when combining our 6D pose prediction with simple robot motion planning, it enables more successful robot manipulation than existing methods. Finally, we conduct an ablation study to demonstrate the necessity of each proposed technique.

Method: Biomarker-based pretraining where ECG feature extractor is trained to predict percentile-binned blood biomarkers from MIMIC-IV-ECG dataset, then fine-tuned on Brazilian datasets for Chagas detection using 5-model ensemble

Result: Achieved challenge score of 0.269 on hidden test set, ranking 5th in Detection of Chagas Disease from the ECG: The George B. Moody PhysioNet Challenge 2025

Conclusion: Biomarker-based pretraining is effective for ECG analysis tasks with limited labeled data, enabling competitive Chagas disease detection performance

Abstract: Chagas disease screening via ECGs is limited by scarce and noisy labels in existing datasets. We propose a biomarker-based pretraining approach, where an ECG feature extractor is first trained to predict percentile-binned blood biomarkers from the MIMIC-IV-ECG dataset. The pretrained model is then fine-tuned on Brazilian datasets for Chagas detection. Our 5-model ensemble, developed by the Ahus AIM team, achieved a challenge score of 0.269 on the hidden test set, ranking 5th in Detection of Chagas Disease from the ECG: The George B. Moody PhysioNet Challenge 2025. Source code and the model are shared on GitHub: github.com/Ahus-AIM/physionet-challenge-2025

Method: Proposes RobustMedSAM with module-wise checkpoint fusion: initializes image encoder from MedSAM (medical priors) and mask decoder from RobustSAM (corruption robustness) under shared ViT-B architecture. Fine-tunes only mask decoder on 35 medical datasets spanning 6 modalities and 12 corruption types while freezing other components. Also investigates SVD-based parameter-efficient variant for limited encoder adaptation.

Result: RobustMedSAM improves degraded-image Dice from 0.613 to 0.719 (+0.106) over SAM on both in-distribution and out-of-distribution benchmarks, demonstrating effective fusion of complementary pretrained models.

Conclusion: Structured fusion of complementary pretrained models (medical-domain adaptation + corruption robustness) is an effective and practical approach for robust medical image segmentation under realistic corruptions.

Abstract: Medical image segmentation models built on Segment Anything Model (SAM) achieve strong performance on clean benchmarks, yet their reliability often degrades under realistic image corruptions such as noise, blur, motion artifacts, and modality-specific distortions. Existing approaches address either medical-domain adaptation or corruption robustness, but not both jointly. In SAM, we find that these capabilities are concentrated in complementary modules: the image encoder preserves medical priors, while the mask decoder governs corruption robustness. Motivated by this observation, we propose RobustMedSAM, which adopts module-wise checkpoint fusion by initializing the image encoder from MedSAM and the mask decoder from RobustSAM under a shared ViT-B architecture. We then fine-tune only the mask decoder on 35 medical datasets from MedSegBench, spanning six imaging modalities and 12 corruption types, while freezing the remaining components to preserve pretrained medical representations. We additionally investigate an SVD-based parameter-efficient variant for limited encoder adaptation. Experiments on both in-distribution and out-of-distribution benchmarks show that RobustMedSAM improves degraded-image Dice from 0.613 to 0.719 (+0.106) over SAM, demonstrating that structured fusion of complementary pretrained models is an effective and practical approach for robust medical image segmentation.

Method: Created a curated dataset of 2,027 real and 2,211 synthetic CCTV clips with annotations for accident time, spatial location, and collision type. Defined three core tasks with custom metrics accounting for uncertainty in CCTV footage.

Result: The benchmark is challenging as shown by diverse baselines including heuristic, motion-aware, and vision-language approaches. The dataset and evaluation framework are publicly available.

Conclusion: ACCIDENT provides a comprehensive benchmark for traffic accident detection that addresses real-world challenges in CCTV analysis, supporting both data-rich and data-scarce scenarios.

Abstract: We introduce ACCIDENT, a benchmark dataset for traffic accident detection in CCTV footage, designed to evaluate models in supervised (IID and OOD) and zero-shot settings, reflecting both data-rich and data-scarce scenarios. The benchmark consists of a curated set of 2,027 real and 2,211 synthetic clips annotated with the accident time, spatial location, and high-level collision type. We define three core tasks: (i) temporal localization of the accident, (ii) its spatial localization, and (iii) collision type classification. Each task is evaluated using custom metrics that account for the uncertainty and ambiguity inherent in CCTV footage. In addition to the benchmark, we provide a diverse set of baselines, including heuristic, motion-aware, and vision-language approaches, and show that ACCIDENT is challenging. You can access the ACCIDENT at: https://accidentbench.github.io

Method: Uses a clothed Momentum Human Rig (MHR) template, renders front/back positional maps, decodes into 3D Gaussians via two-branch architecture: body branch for pose-dependent deformations and face-focused deformation branch for head geometry/appearance refinement. Gaussians are fused, posed with linear blend skinning, and rendered with differentiable Gaussian splatting.

Result: Achieves strong rendering quality with face-view performance of PSNR/SSIM/LPIPS: 26.243/0.964/0.084 on AvatarReX dataset. Ablations confirm contributions of MHR template and face-focused deformation.

Conclusion: F3G-Avatar provides a practical, high-quality pipeline for realistic, animatable full-body avatar synthesis with improved facial detail preservation.

Abstract: Existing full-body Gaussian avatar methods primarily optimize global reconstruction quality and often fail to preserve fine-grained facial geometry and expression details. This challenge arises from limited facial representational capacity that causes difficulties in modeling high-frequency pose-dependent deformations. To address this, we propose F3G-Avatar, a full-body, face-aware avatar synthesis method that reconstructs animatable human representations from multi-view RGB video and regressed pose/shape parameters. Starting from a clothed Momentum Human Rig (MHR) template, front/back positional maps are rendered and decoded into 3D Gaussians through a two-branch architecture: a body branch that captures pose-dependent non-rigid deformations and a face-focused deformation branch that refines head geometry and appearance. The predicted Gaussians are fused, posed with linear blend skinning (LBS), and rendered with differentiable Gaussian splatting. Training combines reconstruction and perceptual objectives with a face-specific adversarial loss to enhance realism in close-up views. Experiments demonstrate strong rendering quality, with face-view performance reaching PSNR/SSIM/LPIPS of 26.243/0.964/0.084 on the AvatarReX dataset. Ablations further highlight contributions of the MHR template and the face-focused deformation. F3G-Avatar provides a practical, high-quality pipeline for realistic, animatable full-body avatar synthesis.

Method: Uses a conditional denoising diffusion probabilistic model with classifier-free guidance. The framework progressively reconstructs vector fields from sparse, coherent inputs (streamlines) while maintaining physical plausibility through the diffusion process.

Result: The method successfully synthesizes plausible vector fields that adhere to physical laws while maintaining fidelity to sparse input observations. It outperforms traditional optimization-based approaches in terms of flexibility and physical consistency.

Conclusion: Diffusion models provide an effective framework for physically plausible vector field synthesis from sparse inputs, offering advantages over traditional optimization methods in maintaining both geometric and physical constraints.

Abstract: We present a novel diffusion-based framework for synthesizing 2D vector fields from sparse, coherent inputs (i.e., streamlines) while maintaining physical plausibility. Our method employs a conditional denoising diffusion probabilistic model with classifier-free guidance, enabling progressive reconstruction that preserves both geometric and physical constraints. Experimental results demonstrate our method’s ability to synthesize plausible vector fields that adhere to physical laws while maintaining fidelity to sparse input observations, outperforming traditional optimization-based approaches in terms of flexibility and physical consistency.

Method: Evaluated four paired open-source VLMs (LLaVA vs. LLaVA-Med; Gemma vs. MedGemma) across four medical imaging tasks of increasing difficulty. Introduced a description-based pipeline where models generate image descriptions for text-only diagnosis by GPT-5.1. Analyzed vision encoder embeddings to understand failure sources.

Result: Performance degrades to near-random levels as task difficulty increases. Medical fine-tuning provides no consistent advantage. Models are highly sensitive to prompt formulation. Description-based pipeline recovers limited additional signal but remains bounded by task difficulty. Failures stem from both weak visual representations and downstream reasoning.

Conclusion: Medical VLM performance is fragile, prompt-dependent, and not reliably improved by domain-specific fine-tuning, indicating limited clinical reasoning capabilities.

Abstract: Vision-language models (VLMs) are increasingly adapted through domain-specific fine-tuning, yet it remains unclear whether this improves reasoning beyond superficial visual cues, particularly in high-stakes domains like medicine. We evaluate four paired open-source VLMs (LLaVA vs. LLaVA-Med; Gemma vs. MedGemma) across four medical imaging tasks of increasing difficulty: brain tumor, pneumonia, skin cancer, and histopathology classification. We find that performance degrades toward near-random levels as task difficulty increases, indicating limited clinical reasoning. Medical fine-tuning provides no consistent advantage, and models are highly sensitive to prompt formulation, with minor changes causing large swings in accuracy and refusal rates. To test whether closed-form VQA suppresses latent knowledge, we introduce a description-based pipeline where models generate image descriptions that a text-only model (GPT-5.1) uses for diagnosis. This recovers a limited additional signal but remains bounded by task difficulty. Analysis of vision encoder embeddings further shows that failures stem from both weak visual representations and downstream reasoning. Overall, medical VLM performance is fragile, prompt-dependent, and not reliably improved by domain-specific fine-tuning.

Method: Two key components: 1) Dynamic Spatial Guidance with time-dependent gating mechanism to modulate foreground/background attention during diffusion, and 2) Multi-Path Pruning that explores multiple latent trajectories and filters candidates using attention statistics and semantic alignment signals.

Result: Extensive evaluations across multiple diffusion backbones demonstrate consistent improvements in background coherence and object-background compositional alignment. A new benchmark is developed for evaluating object-background compositionality.

Conclusion: The proposed training-free framework effectively addresses foreground bias in text-to-image diffusion models, improving scene coherence and compositional control through explicit modeling of foreground-background interactions.

Abstract: Existing text-to-image diffusion models, while excelling at subject synthesis, exhibit a persistent foreground bias that treats the background as a passive and under-optimized byproduct. This imbalance compromises global scene coherence and constrains compositional control. To address the limitation, we propose a training-free framework that restructures diffusion sampling to explicitly account for foreground-background interactions. Our approach consists of two key components. First, Dynamic Spatial Guidance introduces a soft, time step dependent gating mechanism that modulates foreground and background attention during the diffusion process, enabling spatially balanced generation. Second, Multi-Path Pruning performs multi-path latent exploration and dynamically filters candidate trajectories using both internal attention statistics and external semantic alignment signals, retaining trajectories that better satisfy object-background constraints. We further develop a benchmark specifically designed to evaluate object-background compositionality. Extensive evaluations across multiple diffusion backbones demonstrate consistent improvements in background coherence and object-background compositional alignment.

Method: Evaluated several optical flow models on Rotating Snakes illusion, simulated saccadic eye movements, conducted ablation analyses on Dual-Channel model to examine contributions of luminance-based and color-feature-based motion signals and recurrent attention mechanisms.

Result: Most optical flow models failed to generate flow fields consistent with human perception of illusory motion; only Dual-Channel model exhibited expected rotational motion, particularly during saccade simulation; both luminance and color-feature signals plus recurrent attention were critical.

Conclusion: Substantial gap exists between current optical-flow models and human visual motion processing; insights can guide development of motion-estimation systems with better correspondence to human perception for human-centric AI.

Abstract: Understanding human motion processing is essential for building reliable, human-centered computer vision systems. Although deep neural networks (DNNs) achieve strong performance in optical flow estimation, they remain less robust than humans and rely on fundamentally different computational strategies. Visual motion illusions provide a powerful probe into these mechanisms, revealing how human and machine vision align or diverge. While recent DNN-based motion models can reproduce dynamic illusions such as reverse-phi, it remains unclear whether they can perceive illusory motion in static images, exemplified by the Rotating Snakes illusion. We evaluate several representative optical flow models on Rotating Snakes and show that most fail to generate flow fields consistent with human perception. Under simulated conditions mimicking saccadic eye movements, only the human-inspired Dual-Channel model exhibits the expected rotational motion, with the closest correspondence emerging during the saccade simulation. Ablation analyses further reveal that both luminance-based and higher-order color–feature–based motion signals contribute to this behavior and that a recurrent attention mechanism is critical for integrating local cues. Our results highlight a substantial gap between current optical-flow models and human visual motion processing, and offer insights for developing future motion-estimation systems with improved correspondence to human perception and human-centric AI.

Method: Uses rendering supervision for RGB and feature maps. Introduces Token-wise Fusion Module (enriches geometry tokens with semantic context via cross-attention) and Semantic-Geometry Mutual Boosting mechanism (combines geometry-guided feature warping for global consistency with semantic-aware voxelization for local coherence).

Result: Superior performance on ScanNet and DL3DV-10K in novel-view synthesis, open-vocabulary semantic segmentation, and depth estimation. Strong generalization to in-the-wild scenarios.

Conclusion: FF3R establishes a scalable paradigm for unified 3D reasoning without annotations, paving the way for embodied intelligence systems requiring both spatial and semantic understanding.

Abstract: Recent advances in vision foundation models have revolutionized geometry reconstruction and semantic understanding. Yet, most of the existing approaches treat these capabilities in isolation, leading to redundant pipelines and compounded errors. This paper introduces FF3R, a fully annotation-free feed-forward framework that unifies geometric and semantic reasoning from unconstrained multi-view image sequences. Unlike previous methods, FF3R does not require camera poses, depth maps, or semantic labels, relying solely on rendering supervision for RGB and feature maps, establishing a scalable paradigm for unified 3D reasoning. In addition, we address two critical challenges in feedforward feature reconstruction pipelines, namely global semantic inconsistency and local structural inconsistency, through two key innovations: (i) a Token-wise Fusion Module that enriches geometry tokens with semantic context via cross-attention, and (ii) a Semantic-Geometry Mutual Boosting mechanism combining geometry-guided feature warping for global consistency with semantic-aware voxelization for local coherence. Extensive experiments on ScanNet and DL3DV-10K demonstrate FF3R’s superior performance in novel-view synthesis, open-vocabulary semantic segmentation, and depth estimation, with strong generalization to in-the-wild scenarios, paving the way for embodied intelligence systems that demand both spatial and semantic understanding.

Method: PAS leverages pre-trained model generalization power and assesses source-target compatibility based on pre-trained feature embeddings, then integrates into framework for selecting most relevant pre-trained model and source domain.

Result: Extensive experiments on image classification benchmarks show PAS strongly correlates with actual target accuracy and consistently guides selection of best-performing pre-trained model and source domain.

Conclusion: PAS effectively estimates transferability before adaptation, improving target accuracy while reducing computational overhead by selecting optimal source domains and pre-trained models.

Abstract: The goal of domain adaptation is to make predictions for unlabeled samples from a target domain with the help of labeled samples from a different but related source domain. The performance of domain adaptation methods is highly influenced by the choice of source domain and pre-trained feature extractor. However, the selection of source data and pre-trained model is not trivial due to the absence of a labeled validation set for the target domain and the large number of available pre-trained models. In this work, we propose PAS, a novel score designed to estimate the transferability of a source domain set and a pre-trained feature extractor to a target classification task before actually performing domain adaptation. PAS leverages the generalization power of pre-trained models and assesses source-target compatibility based on the pre-trained feature embeddings. We integrate PAS into a framework that indicates the most relevant pre-trained model and source domain among multiple candidates, thus improving target accuracy while reducing the computational overhead. Extensive experiments on image classification benchmarks demonstrate that PAS correlates strongly with actual target accuracy and consistently guides the selection of the best-performing pre-trained model and source domain for adaptation.

Method: Introduces frozen DINOv3 features as structural priors to inject semantic awareness into the reconstruction process, maintaining linear time complexity O(T).

Result: Significantly improves Tracking Accuracy (APD) and Reconstruction Completeness on Point Odyssey and TUM-Dynamics benchmarks while maintaining computational efficiency.

Conclusion: Establishes a new paradigm for constructing 4D World Models with both geometric precision and semantic understanding.

Abstract: In the intersection of computer vision and robotic perception, 4D reconstruction of dynamic scenes serve as the critical bridge connecting low-level geometric sensing with high-level semantic understanding. We present DINO_4D, introducing frozen DINOv3 features as structural priors, injecting semantic awareness into the reconstruction process to effectively suppress semantic drift during dynamic tracking. Experiments on the Point Odyssey and TUM-Dynamics benchmarks demonstrate that our method maintains the linear time complexity $O(T)$ of its predecessors while significantly improving Tracking Accuracy (APD) and Reconstruction Completeness. DINO_4D establishes a new paradigm for constructing 4D World Models that possess both geometric precision and semantic understanding.

Method: Proposes Topo-ADV, an end-to-end differentiable framework incorporating persistent homology as optimization objective. Uses differentiable topological representations to jointly optimize: 1) topology divergence loss altering persistence, 2) misclassification objective, and 3) geometric imperceptibility constraints.

Result: Achieves up to 100% attack success rates on ModelNet40, ShapeNet Part, and ScanObjectNN datasets using PointNet and DGCNN classifiers. Perturbations remain geometrically indistinguishable from originals, beating state-of-the-art methods on perceptibility metrics.

Conclusion: Topological structure represents a significant vulnerability for 3D point cloud models. Topology-driven attacks can be highly effective while maintaining geometric plausibility, challenging assumptions about shape fidelity and semantic preservation.

Abstract: Deep neural networks for 3D point cloud understanding have achieved remarkable success in object classification and recognition, yet recent work shows that these models remain highly vulnerable to adversarial perturbations. Existing 3D attacks predominantly manipulate geometric properties such as point locations, curvature, or surface structure, implicitly assuming that preserving global shape fidelity preserves semantic content. In this work, we challenge this assumption and introduce the first topology-driven adversarial attack for point cloud deep learning. Our key insight is that the homological structure of a 3D object constitutes a previously unexplored vulnerability surface. We propose Topo-ADV, an end-to-end differentiable framework that incorporates persistent homology as an explicit optimization objective, enabling gradient-based manipulation of topological features during adversarial example generation. By embedding persistence diagrams through differentiable topological representations, our method jointly optimizes (i) a topology divergence loss that alters persistence, (ii) a misclassification objective, and (iii) geometric imperceptibility constraints that preserve visual plausibility. Experiments demonstrate that subtle topology-driven perturbations consistently achieve up to 100% attack success rates on benchmark datasets such as ModelNet40, ShapeNet Part, and ScanObjectNN using PointNet and DGCNN classifiers, while remaining geometrically indistinguishable from the original point clouds, beating state-of-the-art methods on various perceptibility metrics.

Method: Two key components: (1) Geometry-driven pruning strategy that ranks Gaussians based solely on 3D attributes (eliminating 2D image dependency), and (2) Dual-branch encoder that separates and re-weights geometric and appearance features to avoid imbalance.

Result: Extensive experiments on ScanNet++ and Replica datasets show PointSplat achieves competitive rendering quality and superior efficiency across varying sparsity levels without additional per-scene optimization.

Conclusion: PointSplat bridges Gaussian pruning and transformer refinement with a geometry-driven approach that reduces model complexity while maintaining visual quality, offering a more efficient alternative to traditional methods.

Abstract: 3D Gaussian Splatting (3DGS) has recently unlocked real-time, high-fidelity novel view synthesis by representing scenes using explicit 3D primitives. However, traditional methods often require millions of Gaussians to capture complex scenes, leading to significant memory and storage demands. Recent approaches have addressed this issue through pruning and per-scene fine-tuning of Gaussian parameters, thereby reducing the model size while maintaining visual quality. These strategies typically rely on 2D images to compute important scores followed by scene-specific optimization. In this work, we introduce PointSplat, 3D geometry-driven prune-and-refine framework that bridges previously disjoint directions of gaussian pruning and transformer refinement. Our method includes two key components: (1) an efficient geometry-driven strategy that ranks Gaussians based solely on their 3D attributes, removing reliance on 2D images during pruning stage, and (2) a dual-branch encoder that separates, re-weights geometric and appearance to avoid feature imbalance. Extensive experiments on ScanNet++ and Replica across varying sparsity levels demonstrate that PointSplat consistently achieves competitive rendering quality and superior efficiency without additional per-scene optimization.

Method: Developed PlantXpert benchmark with 385 digital images and 3,000+ samples covering disease, pest control, weed management, and yield. Evaluated 11 VLMs on visual expertise, quantitative reasoning, and multi-step agronomic reasoning capabilities.

Result: Task-specific fine-tuning improved accuracy up to 78% (Qwen3-VL models). Model scaling gains diminished beyond certain capacity, generalization across crops was uneven, and quantitative/biological reasoning remained challenging.

Conclusion: PlantXpert provides a foundation for assessing evidence-grounded agronomic reasoning and advancing multimodal model development in plant science, though significant challenges remain in domain adaptation.

Abstract: To improve crop genetics, high-throughput, effective and comprehensive phenotyping is a critical prerequisite. While such tasks were traditionally performed manually, recent advances in multimodal foundation models, especially in vision-language models (VLMs), have enabled more automated and robust phenotypic analysis. However, plant science remains a particularly challenging domain for foundation models because it requires domain-specific knowledge, fine-grained visual interpretation, and complex biological and agronomic reasoning. To address this gap, we develop PlantXpert, an evidence-grounded multimodal reasoning benchmark for soybean and cotton phenotyping. Our benchmark provides a structured and reproducible framework for agronomic adaptation of VLMs, and enables controlled comparison between base models and their domain-adapted counterparts. We constructed a dataset comprising 385 digital images and more than 3,000 benchmark samples spanning key plant science domains including disease, pest control, weed management, and yield. The benchmark can assess diverse capabilities including visual expertise, quantitative reasoning, and multi-step agronomic reasoning. A total of 11 state-of-the-art VLMs were evaluated. The results indicate that task-specific fine-tuning leads to substantial improvement in accuracy, with models such as Qwen3-VL-4B and Qwen3-VL-30B achieving up to 78%. At the same time, gains from model scaling diminish beyond a certain capacity, generalization across soybean and cotton remains uneven, and quantitative as well as biologically grounded reasoning continue to pose substantial challenges. These findings suggest that PlantXpert can serve as a foundation for assessing evidence-grounded agronomic reasoning and for advancing multimodal model development in plant science.

Method: Evaluated four open-vocabulary detectors (YOLO World, SAM3, Grounding DINO, OWLv2) for cowpea flower/pod detection. Decomposed prompts into eight axes, conducted one-factor-at-a-time analysis followed by combinatorial optimization. Used LLM to translate discovered axis structure to morphologically distinct targets.

Result: Model-specific combinatorial prompts yield substantial gains over naive baselines (+0.357 mAP@0.5 for YOLO World, +0.362 for OWLv2). Prompt structures optimized on synthetic data transfer effectively to real-world fields, matching or exceeding those discovered on labeled real data for most model-object combinations.

Conclusion: Prompt engineering can substantially close the gap between zero-shot VFMs and supervised detectors without manual annotation. Optimal prompts are model-specific, non-obvious, and transferable across domains.

Abstract: Vision foundation models (VFMs) offer the promise of zero-shot object detection without task-specific training data, yet their performance in complex agricultural scenes remains highly sensitive to text prompt construction. We present a systematic prompt optimization framework evaluating four open-vocabulary detectors – YOLO World, SAM3, Grounding DINO, and OWLv2 – for cowpea flower and pod detection across synthetic and real field imagery. We decompose prompts into eight axes and conduct one-factor-at-a-time analysis followed by combinatorial optimization, revealing that models respond divergently to prompt structure: conditions that optimize one architecture can collapse another. Applying model-specific combinatorial prompts yields substantial gains over a naive species-name baseline, including +0.357 mAP@0.5 for YOLO World and +0.362 mAP@0.5 for OWLv2 on synthetic cowpea flower data. To evaluate cross-task generalization, we use an LLM to translate the discovered axis structure to a morphologically distinct target – cowpea pods – and compare against prompting using the discovered optimal structures from synthetic flower data. Crucially, prompt structures optimized exclusively on synthetic data transfer effectively to real-world fields: synthetic-pipeline prompts match or exceed those discovered on labeled real data for the majority of model-object combinations (flower: 0.374 vs. 0.353 for YOLO World; pod: 0.429 vs. 0.371 for SAM3). Our findings demonstrate that prompt engineering can substantially close the gap between zero-shot VFMs and supervised detectors without requiring manual annotation, and that optimal prompts are model-specific, non-obvious, and transferable across domains.

Method: Two-stage pipeline: 1) YOLO-based detector pretrained on synthetic Blender data simulating extreme perspectives/lighting, fine-tuned on real Bolivian data; 2) geometric rectification and character recognition with confidence-based fallback to Gemma3 4B VLM for ambiguous cases.

Result: Achieves 89.6% character-level recognition accuracy on real-world Bolivian data and introduces the first public Bolivian LPDR dataset for evaluation under diverse conditions.

Conclusion: BLPR demonstrates effective license plate recognition for challenging Bolivian urban environments through synthetic-to-real domain adaptation and confidence-based VLM fallback mechanisms.

Abstract: Robust license plate recognition in unconstrained environments remains a significant challenge, particularly in underrepresented regions with limited data availability and unique visual characteristics, such as Bolivia. Recognition accuracy in real-world conditions is often degraded by factors such as illumination changes and viewpoint distortion. To address these challenges, we introduce BLPR, a novel deep learning-based License Plate Detection and Recognition (LPDR) framework specifically designed for Bolivian license plates. The proposed system follows a two-stage pipeline where a YOLO-based detector is pretrained on synthetic data generated in Blender to simulate extreme perspectives and lighting conditions, and subsequently fine-tuned on street-level data collected in La Paz, Bolivia. Detected plates are geometrically rectified and passed to a character recognition model. To improve robustness under ambiguous scenarios, a lightweight vision-language model (Gemma3 4B) is selectively triggered as a confidence-based fallback mechanism. The proposed framework further leverages synthetic-to-real domain adaptation to improve robustness under diverse real-world conditions. We also introduce the first publicly available Bolivian LPDR dataset, enabling evaluation under diverse viewpoint and illumination conditions. The system achieves a character-level recognition accuracy of 89.6% on real-world data, demonstrating its effectiveness for deployment in challenging urban environments. Our project is publicly available at https://github.com/EdwinTSalcedo/BLPR.

Method: Used synthetic datasets to test binding sensitivity to continuity vs other Gestalt principles. Analyzed attention heads in pretrained vision transformers that track continuity. Conducted ablation studies on these heads to assess their contribution to object binding representations.

Result: Binding probes show sensitivity to continuity across various pretrained vision transformers. Specific attention heads track continuity and generalize across datasets. Ablation of these heads often reduces models’ ability to produce representations encoding object binding.

Conclusion: Vision transformers learn to use Gestalt continuity principles for object binding, with specialized attention heads that track continuity cues. These mechanisms contribute to the models’ ability to form coherent object representations.

Abstract: Object binding is a foundational process in visual cognition, during which low-level perceptual features are joined into object representations. Binding has been considered a fundamental challenge for neural networks, and a major milestone on the way to artificial models with flexible visual intelligence. Recently, several investigations have demonstrated evidence that binding mechanisms emerge in pretrained vision models, enabling them to associate portions of an image that contain an object. The question remains: how are these models binding objects together? In this work, we investigate whether vision models rely on the principle of Gestalt continuity to perform object binding, over and above other principles like similarity and proximity. Using synthetic datasets, we demonstrate that binding probes are sensitive to continuity across a wide range of pretrained vision transformers. Next, we uncover particular attention heads that track continuity, and show that these heads generalize across datasets. Finally, we ablate these attention heads, and show that they often contribute to producing representations that encode object binding.

Method: Multi-dimensional analysis using counterfactual image sets depicting the same person across different cultural contexts. Evaluation framework uses Moral Foundations Theory, lexical analyses, and measures sensitivity of generated values to depicted cultural contexts across five popular LVLMs.

Result: LVLMs demonstrate systematic cultural stereotypes in value judgments, with generated values being sensitive to depicted cultural contexts. The models show awareness of cultural value differences but reinforce harmful stereotypes.

Conclusion: LVLMs exhibit problematic cultural biases in value judgments based on visual cultural contexts, highlighting the need for fairness interventions beyond social biases to address cultural stereotypes.

Abstract: The rapid adoption of large vision-language models (LVLMs) in recent years has been accompanied by growing fairness concerns due to their propensity to reinforce harmful societal stereotypes. While significant attention has been paid to such fairness concerns in the context of social biases, relatively little prior work has examined the presence of stereotypes in LVLMs related to cultural contexts such as religion, nationality, and socioeconomic status. In this work, we aim to narrow this gap by investigating how cultural contexts depicted in images influence the judgments LVLMs make about a person’s moral, ethical, and political values. We conduct a multi-dimensional analysis of such value judgments in five popular LVLMs using counterfactual image sets, which depict the same person across different cultural contexts. Our evaluation framework diagnoses LVLM awareness of cultural value differences through the use of Moral Foundations Theory, lexical analyses, and the sensitivity of generated values to depicted cultural contexts.

Method: 1) Spectral unmixing network that learns endmembers and abundance maps while accounting for endmember variabilities; 2) Top-K token selection strategy based on abundance map clusters for adaptive token sequencing; 3) Unmixing-guided spatial-spectral Mamba module for improved feature learning; 4) Multi-task supervision scheme for simultaneous endmember-abundance pattern and classification label learning.

Result: The model outperforms state-of-the-art approaches on four HSI datasets, demonstrating superior classification accuracy while also producing comprehensive spectral-library and abundance maps.

Conclusion: The proposed unmixing-guided spatial-spectral Mamba framework effectively addresses HSI classification challenges by integrating spectral unmixing with advanced sequence modeling, providing both accurate classification and interpretable spectral analysis outputs.

Abstract: Although hyperspectral image (HSI) classification is critical for supporting various environmental applications, it is a challenging task due to the spectral-mixture effect, the spatial-spectral heterogeneity and the difficulty to preserve class boundaries and details. This letter presents a novel unmixing-guided spatial-spectral Mamba with clustering tokens for improved HSI classification, with the following contributions. First, to disentangle the spectral mixture effect in HSI for improved pattern discovery, we design a novel spectral unmixing network that not only automatically learns endmembers and abundance maps from HSI but also accounts for endmember variabilities. Second, to generate Mamba token sequences, based on the clusters defined by abundance maps, we design an efficient Top-\textit{K} token selection strategy to adaptively sequence the tokens for improved representational capability. Third, to improve spatial-spectral feature learning and detail preservation, based on the Top-\textit{K} token sequences, we design a novel unmixing-guided spatial-spectral Mamba module that greatly improves traditional Mamba models in terms of token learning and sequencing. Fourth, to learn simultaneously the endmember-abundance patterns and classification labels, a multi-task scheme is designed for model supervision, leading to a new unmixing-classification framework that outputs not only accurate classification maps but also a comprehensive spectral-library and abundance maps. Comparative experiments on four HSI datasets demonstrate that our model can greatly outperform the other state-of-the-art approaches. Code is available at https://github.com/GSIL-UCalgary/Unmixing_guided_Mamba.git

Method: Learnable Motion-Focused Tokenization (LMFT) tokenizes video frames into patch tokens and learns to discard low-motion, redundant background tokens while retaining motion-rich, action-relevant tokens for adaptation.

Result: Achieves state-of-the-art performance on three standard VUDA benchmarks across 21 domain adaptation settings while significantly reducing computational overhead.

Conclusion: LMFT enables VUDA that is both effective and computationally efficient by focusing on motion-rich tokens and discarding redundant background information.

Abstract: Video Unsupervised Domain Adaptation (VUDA) poses a significant challenge in action recognition, requiring the adaptation of a model from a labeled source domain to an unlabeled target domain. Despite recent advances, existing VUDA methods often fall short of fully supervised performance, a key reason being the prevalence of static and uninformative backgrounds that exacerbate domain shifts. Additionally, prior approaches largely overlook computational efficiency, limiting real-world adoption. To address these issues, we propose Learnable Motion-Focused Tokenization (LMFT) for VUDA. LMFT tokenizes video frames into patch tokens and learns to discard low-motion, redundant tokens, primarily corresponding to background regions, while retaining motion-rich, action-relevant tokens for adaptation. Extensive experiments on three standard VUDA benchmarks across 21 domain adaptation settings show that our VUDA framework with LMFT achieves state-of-the-art performance while significantly reducing computational overhead. LMFT thus enables VUDA that is both effective and computationally efficient.

Method: Proposes YUV20K benchmark with 24,295 annotated frames across 91 scenes and 47 species, targeting challenging scenarios. Introduces novel framework with two key modules: Motion Feature Stabilization (MFS) using frame-agnostic Semantic Basis Primitives to stabilize features, and Trajectory-Aware Alignment (TAA) using trajectory-guided deformable sampling for precise temporal alignment.

Result: Method significantly outperforms state-of-the-art competitors on existing datasets and establishes new baseline on YUV20K. Shows superior cross-domain generalization and robustness in complex spatiotemporal scenarios.

Conclusion: The YUV20K benchmark addresses data scarcity in VCOD, while the proposed framework with MFS and TAA modules effectively handles complex motion dynamics, advancing the field of video camouflaged object detection.

Abstract: Video Camouflaged Object Detection (VCOD) is currently constrained by the scarcity of challenging benchmarks and the limited robustness of models against erratic motion dynamics. Existing methods often struggle with Motion-Induced Appearance Instability and Temporal Feature Misalignment caused by complex motion scenarios. To address the data bottleneck, we present YUV20K, a pixel-level annoated complexity-driven VCOD benchmark. Comprising 24,295 annotated frames across 91 scenes and 47 kinds of species, it specifically targets challenging scenarios like large-displacement motion, camera motion and other 4 types scenarios. On the methodological front, we propose a novel framework featuring two key modules: Motion Feature Stabilization (MFS) and Trajectory-Aware Alignment (TAA). The MFS module utilizes frame-agnostic Semantic Basis Primitives to stablize features, while the TAA module leverages trajectory-guided deformable sampling to ensure precise temporal alignment. Extensive experiments demonstrate that our method significantly outperforms state-of-the-art competitors on existing datasets and establishes a new baseline on the challenging YUV20K. Notably, our framework exhibits superior cross-domain generalization and robustness when confronting complex spatiotemporal scenarios. Our code and dataset will be available at https://github.com/K1NSA/YUV20K

Method: FlowPalm estimates optical flows between real palmprint pairs to capture statistical patterns of geometric deformations, then uses a progressive sampling process that gradually introduces these deformations during diffusion while maintaining identity consistency.

Result: Extensive experiments on six benchmark datasets show FlowPalm significantly outperforms state-of-the-art palmprint generation approaches in downstream recognition tasks.

Conclusion: FlowPalm effectively addresses geometric variation in synthetic palmprint generation through optical-flow-driven deformation modeling, improving recognition model training.

Abstract: Recently, synthetic palmprints have been increasingly used as substitutes for real data to train recognition models. To be effective, such synthetic data must reflect the diversity of real palmprints, including both style variation and geometric variation. However, existing palmprint generation methods mainly focus on style translation, while geometric variation is either ignored or approximated by simple handcrafted augmentations. In this work, we propose FlowPalm, an optical-flow-driven palmprint generation framework capable of simulating the complex non-rigid deformations observed in real palms. Specifically, FlowPalm estimates optical flows between real palmprint pairs to capture the statistical patterns of geometric deformations. Building on these priors, we design a progressive sampling process that gradually introduces the geometric deformations during diffusion while maintaining identity consistency. Extensive experiments on six benchmark datasets demonstrate that FlowPalm significantly outperforms state-of-the-art palmprint generation approaches in downstream recognition tasks. Project page: https://yuchenzou.github.io/FlowPalm/

Method: Proposes Temporal Kolmogorov-Arnold Network (TKAN) that replaces fixed edge weights with learnable one-dimensional functions. Incorporates two-level memory: short-term RKAN sublayers for cycle-level dynamics and gated long-term pathway for broader temporal context. Uses CNN+TKAN framework for gait recognition.

Result: Experiments on CASIA-B dataset show the CNN+TKAN framework achieves strong recognition performance under the reported evaluation setting.

Conclusion: TKAN provides an efficient solution for gait recognition that can handle both local gait cycles and long-term motion trends while maintaining computational efficiency compared to recurrent and transformer architectures.

Abstract: Gait recognition is a biometric modality that identifies individuals from their characteristic walking patterns. Unlike conventional biometric traits, gait can be acquired at a distance and without active subject cooperation, making it suitable for surveillance and public safety applications. Nevertheless, silhouette-based temporal models remain sensitive to long sequences, observation noise, and appearance-related covariates. Recurrent architectures often struggle to preserve information from earlier frames and are inherently sequential to optimize, whereas transformer-based models typically require greater computational resources and larger training sets and may be sensitive to irregular sequence lengths and noisy inputs. These limitations reduce robustness under clothing variation, carrying conditions, and view changes, while also hindering the joint modeling of local gait cycles and longer-term motion trends. To address these challenges, we introduce a Temporal Kolmogorov-Arnold Network (TKAN) for gait recognition. The proposed model replaces fixed edge weights with learnable one-dimensional functions and incorporates a two-level memory mechanism consisting of short-term RKAN sublayers and a gated long-term pathway. This design enables efficient modeling of both cycle-level dynamics and broader temporal context while maintaining a compact backbone. Experiments on the CASIA-B dataset indicate that the proposed CNN+TKAN framework achieves strong recognition performance under the reported evaluation setting.

Method: Proposes a diff-based scale loss that weights predictions by signed area difference between predicted mask and ground truth for monotonic gradients. Introduces Gaussian-shaped convolution with learnable scale parameter to match center-concentrated intensity profiles, augmented with rotated pinwheel mask that adaptively aligns kernel orientation via straight-through estimator.

Result: Extensive experiments on IRSTD-1k, NUDT-SIRST, and SIRST-UAVB datasets demonstrate consistent improvements in mIoU, Pd, and Fa metrics over state-of-the-art methods.

Conclusion: The proposed diff-based scale loss and Gaussian-shaped convolution with orientation alignment effectively address training stability and spatial attention challenges in infrared small target detection, achieving superior performance across multiple benchmarks.

Abstract: Infrared small target detection still faces two persistent challenges: training instability from non-monotonic scale loss functions, and inadequate spatial attention due to generic convolution kernels that ignore the physical imaging characteristics of small targets. In this paper, we revisit both aspects. For the loss side, we propose a \emph{diff-based scale loss} that weights predictions according to the signed area difference between the predicted mask and the ground truth, yielding strictly monotonic gradients and stable convergence. We further analyze a family of four scale loss variants to understand how their geometric properties affect detection behavior. For the spatial side, we introduce \emph{Gaussian-shaped convolution} with a learnable scale parameter to match the center-concentrated intensity profile of infrared small targets, and augment it with a \emph{rotated pinwheel mask} that adaptively aligns the kernel with target orientation via a straight-through estimator. Extensive experiments on IRSTD-1k, NUDT-SIRST, and SIRST-UAVB demonstrate consistent improvements in $mIoU$, $P_d$, and $F_a$ over state-of-the-art methods. We release our anonymous code and pretrained models.

Method: Established controlled benchmark with deterministic train/validation/test split on AppleBBCH81 dataset. Evaluated six detectors: YOLOv10n, YOLO11n, RT-DETR-L, Faster R-CNN, FCOS, and SSDLite320 using COCO-style mAP metrics, precision-recall curves, and fixed-threshold analysis.

Result: YOLO11n achieved best strict localization (mAP@0.5:0.95 = 0.6065, mAP@0.5 = 0.9620). YOLOv10n had highest F1-score at fixed threshold. RT-DETR-L showed high recall but low precision due to false positives.

Conclusion: Detector selection for orchard applications should consider both localization accuracy and threshold robustness based on downstream task requirements.

Abstract: Accurate apple detection in orchard images is important for yield prediction, fruit counting, robotic harvesting, and crop monitoring. However, changing illumination, leaf clutter, dense fruit clusters, and partial occlusion make detection difficult. To provide a fair and reproducible comparison, this study establishes a controlled benchmark for single-class apple detection on the public AppleBBCH81 dataset using one deterministic train, validation, and test split and a unified evaluation protocol across six representative detectors: YOLOv10n, YOLO11n, RT-DETR-L, Faster R-CNN (ResNet50-FPN), FCOS (ResNet50-FPN), and SSDLite320 (MobileNetV3-Large). Performance is evaluated primarily using COCO-style mAP@0.5 and mAP@0.5:0.95, and threshold-dependent behavior is further analyzed using precision-recall curves and fixed-threshold precision, recall, and F1-score at IoU = 0.5. On the validation split, YOLO11n achieves the best strict localization performance with mAP@0.5:0.95 = 0.6065 and mAP@0.5 = 0.9620, followed closely by RT-DETR-L and YOLOv10n. At a fixed operating point with confidence >= 0.05, YOLOv10n attains the highest F1-score, whereas RT-DETR-L achieves very high recall but low precision because of many false positives at low confidence. These findings show that detector selection for orchard deployment should be guided not only by localization-aware accuracy but also by threshold robustness and the requirements of the downstream task.

Method: GIF uses a conditional diffusion process guided by fused image and graph features, combining geometric layout features with logical circuit topology representations for multimodal conditioning.

Result: On CircuitNet-N28 dataset: 0.78 SSIM, 0.95 Pearson correlation, 21.77 PSNR, 0.026 NMAE, outperforming prior methods. Demonstrates reliable high-quality IR drop image generation.

Conclusion: IR drop analysis can effectively leverage generative modeling advances when geometric layout and logical topology are jointly modeled, enabling structured image generation for chip design.

Abstract: IR drop analysis is essential in physical chip design to ensure the power integrity of on-chip power delivery networks. Traditional Electronic Design Automation (EDA) tools have become slow and expensive as transistor density scales. Recent works have introduced machine learning (ML)-based methods that formulate IR drop analysis as an image prediction problem. These existing ML approaches fail to capture both local and long-range dependencies and ignore crucial geometrical and topological information from physical layouts and logical connectivity. To address these limitations, we propose GIF, a Generative IR drop Framework that uses both geometrical and topological information to generate IR drop images. GIF fuses image and graph features to guide a conditional diffusion process, producing high-quality IR drop images. For instance, On the CircuitNet-N28 dataset, GIF achieves 0.78 SSIM, 0.95 Pearson correlation, 21.77 PSNR, and 0.026 NMAE, outperforming prior methods. These results demonstrate that our framework, using diffusion based multimodal conditioning, reliably generates high quality IR drop images. This shows that IR drop analysis can effectively leverage recent advances in generative modeling when geometric layout features and logical circuit topology are jointly modeled. By combining geometry aware spatial features with logical graph representations, GIF enables IR drop analysis to benefit from recent advances in generative modeling for structured image generation.

Method: Proposes SwinTextUNet, a multimodal segmentation framework that integrates Contrastive Language Image Pretraining (CLIP) derived textual embeddings into a Swin Transformer UNet backbone. Uses cross-attention and convolutional fusion mechanisms to align semantic text guidance with hierarchical visual representations. Evaluates a four-stage variant optimized for performance-complexity balance.

Result: Achieves Dice score of 86.47% and IoU of 78.2% on the QaTaCOV19 dataset. Ablation studies validate the importance of text guidance and multimodal fusion components. The four-stage variant provides optimal balance between performance and complexity.

Conclusion: Vision-language integration shows promise for advancing medical image segmentation and supporting clinically meaningful diagnostic tools. The multimodal approach enhances robustness and accuracy when dealing with ambiguous patterns in medical imaging.

Abstract: Precise medical image segmentation is fundamental for enabling computer aided diagnosis and effective treatment planning. Traditional models that rely solely on visual features often struggle when confronted with ambiguous or low contrast patterns. To overcome these limitations, we introduce SwinTextUNet, a multimodal segmentation framework that incorporates Contrastive Language Image Pretraining (CLIP), derived textual embeddings into a Swin Transformer UNet backbone. By integrating cross attention and convolutional fusion, the model effectively aligns semantic text guidance with hierarchical visual representations, enhancing robustness and accuracy. We evaluate our approach on the QaTaCOV19 dataset, where the proposed four stage variant achieves an optimal balance between performance and complexity, yielding Dice and IoU scores of 86.47% and 78.2%, respectively. Ablation studies further validate the importance of text guidance and multimodal fusion. These findings underscore the promise of vision language integration in advancing medical image segmentation and supporting clinically meaningful diagnostic tools.

Method: Evaluated four omnimodal models on tasks including demographic attribute estimation, identity verification, activity recognition, multilingual speech transcription, and language identification, measuring accuracy differences across age, gender, skin tone, language, and country of origin.

Result: Image and video understanding tasks generally exhibit better performance with smaller demographic disparities, while audio understanding tasks show significantly lower performance and substantial bias, including large accuracy differences across age groups, genders, and languages, and frequent prediction collapse toward narrow categories.

Conclusion: The findings highlight the importance of evaluating fairness across all supported modalities as omnimodal language models are increasingly used in real-world applications, with particular attention needed for audio understanding tasks that show the most significant biases.

Abstract: This paper provides a comprehensive evaluation of demographic and linguistic biases in omnimodal language models that process text, images, audio, and video within a single framework. Although these models are being widely deployed, their performance across different demographic groups and modalities is not well studied. Four omnimodal models are evaluated on tasks that include demographic attribute estimation, identity verification, activity recognition, multilingual speech transcription, and language identification. Accuracy differences are measured across age, gender, skin tone, language, and country of origin. The results show that image and video understanding tasks generally exhibit better performance with smaller demographic disparities. In contrast, audio understanding tasks exhibit significantly lower performance and substantial bias, including large accuracy differences across age groups, genders, and languages, and frequent prediction collapse toward narrow categories. These findings highlight the importance of evaluating fairness across all supported modalities as omnimodal language models are increasingly used in real-world applications.

Method: Proposes Structure-Semantics Co-Tuning (S2-CoT) with two specialized adapters: Structural Fidelity Adapter (SFA) in encoder-decoder for high-fidelity representations via spatial-frequency fusion, and Semantic Context Adapter (SCA) in entropy model to align with SFA-tuned features by refining channel context for efficient statistical coding.

Result: Achieves state-of-the-art results across four diverse base codecs with minimal trainable parameters, closely matching full fine-tuning performance while avoiding the performance degradation of naive adapter approaches.

Conclusion: S2-CoT demonstrates that coordinated adaptation of both structural and semantic components is crucial for effective parameter-efficient fine-tuning in image compression, enabling near-full fine-tuning performance with significantly reduced parameters.

Abstract: Parameter-efficient fine-tuning of pre-trained codecs is a promising direction in image compression for human and machine vision. While most existing works have primarily focused on tuning the feature structure within the encoder-decoder backbones, the adaptation of the statistical semantics within the entropy model has received limited attention despite its function of predicting the probability distribution of latent features. Our analysis reveals that naive adapter insertion into the entropy model can lead to suboptimal outcomes, underscoring that the effectiveness of adapter-based tuning depends critically on the coordination between adapter type and placement across the compression pipeline. Therefore, we introduce Structure-Semantics Co-Tuning (S2-CoT), a novel framework that achieves this coordination via two specialized, synergistic adapters: the Structural Fidelity Adapter (SFA) and the Semantic Context Adapter (SCA). SFA is integrated into the encoder-decoder to preserve high-fidelity representations by dynamically fusing spatial and frequency information; meanwhile, the SCA adapts the entropy model to align with SFA-tuned features by refining the channel context for more efficient statistical coding. Through joint optimization, S2-CoT turns potential performance degradation into synergistic gains, achieving state-of-the-art results across four diverse base codecs with only a small fraction of trainable parameters, closely matching full fine-tuning performance. Code is available at https://github.com/Brock-bit4/S2-CoT.

Method: Proposes a frequency-domain importance-driven dynamic LoRA switch method that recognizes different adapters are specialized for different content types, so each diffusion step carries different significance for each adapter. Also includes an automatic Generation Alignment mechanism to align generation intents at the semantic level for maintaining consistency.

Result: The FREE-Switch framework efficiently combines adapters for different objects and styles, substantially reducing training cost for high-quality customized generation.

Conclusion: The proposed method enables low-cost customized generation by efficiently combining pretrained diffusion adapters while avoiding content drift and detail degradation issues of existing approaches.

Abstract: With the growing availability of open-sourced adapters trained on the same diffusion backbone for diverse scenes and objects, combining these pretrained weights enables low-cost customized generation. However, most existing model merging methods are designed for classification or text generation, and when applied to image generation, they suffer from content drift due to error accumulation across multiple diffusion steps. For image-oriented methods, training-based approaches are computationally expensive and unsuitable for edge deployment, while training-free ones use uniform fusion strategies that ignore inter-adapter differences, leading to detail degradation. We find that since different adapters are specialized for generating different types of content, the contribution of each diffusion step carries different significance for each adapter. Accordingly, we propose a frequency-domain importance-driven dynamic LoRA switch method. Furthermore, we observe that maintaining semantic consistency across adapters effectively mitigates detail loss; thus, we design an automatic Generation Alignment mechanism to align generation intents at the semantic level. Experiments demonstrate that our FREE-Switch (Frequency-based Efficient and Dynamic LoRA Switch) framework efficiently combines adapters for different objects and styles, substantially reducing the training cost of high-quality customized generation.

Method: Created LVSum benchmark with human-annotated long-form videos across 13 domains, each with precise temporal references in summaries. Evaluated proprietary and open-source MLLMs using new LLM-based metrics for content relevance and modality coherence plus standard metrics.

Result: Revealed systematic gaps in temporal understanding among existing MLLMs, providing insights for advancing temporal reasoning in long video summarization.

Conclusion: LVSum establishes a new foundation for evaluating and improving temporal reasoning capabilities in MLLMs for long video summarization tasks.

Abstract: Long video summarization presents significant challenges for current multimodal large language models (MLLMs), particularly in maintaining temporal fidelity over extended durations and producing summaries that are both semantically and temporally grounded. In this work, we present LVSum, a human-annotated benchmark designed specifically for evaluating long video summarization with fine-grained temporal alignment. LVSum comprises diverse long-form videos across 13 domains, each paired with human-generated summaries containing precise temporal references. We conduct a comprehensive evaluation of both proprietary and open-source MLLMs on LVSum, assessing performance using newly introduced LLM-based metrics for content relevance and modality coherence, alongside standard evaluation metrics. Our experiments reveal systematic gaps in temporal understanding among existing MLLMs and offer insights that establish a new foundation for advancing temporal reasoning in long video summarization.

Method: SinkTrack treats the BOS token as an information anchor and injects key contextual features (from input image or instruction) into its representation. This training-free, plug-and-play method maintains attention to initial context throughout generation with minimal inference overhead.

Result: Experiments show SinkTrack mitigates hallucination and context forgetting across textual (+21.6% on SQuAD2.0 with Llama3.1-8B-Instruct) and multimodal tasks (+22.8% on M3CoT with Qwen2.5-VL-7B-Instruct), with consistent gains across architectures and scales.

Conclusion: SinkTrack effectively addresses attention drift in LLMs by leveraging attention sink, providing a robust, generalizable solution for reducing hallucination and context forgetting in both textual and multimodal applications.

Abstract: Large language models (LLMs) suffer from hallucination and context forgetting. Prior studies suggest that attention drift is a primary cause of these problems, where LLMs’ focus shifts towards newly generated tokens and away from the initial input context. To counteract this, we make use of a related, intrinsic characteristic of LLMs: attention sink – the tendency to consistently allocate high attention to the very first token (i.e., ) of a sequence. Concretely, we propose an advanced context anchoring method, SinkTrack, which treats as an information anchor and injects key contextual features (such as those derived from the input image or instruction) into its representation. As such, LLM remains anchored to the initial input context throughout the entire generation process. SinkTrack is training-free, plug-and-play, and introduces negligible inference overhead. Experiments demonstrate that SinkTrack mitigates hallucination and context forgetting across both textual (e.g., +21.6% on SQuAD2.0 with Llama3.1-8B-Instruct) and multi-modal (e.g., +22.8% on M3CoT with Qwen2.5-VL-7B-Instruct) tasks. Its consistent gains across different architectures and scales underscore the robustness and generalizability. We also analyze its underlying working mechanism from the perspective of information delivery. Our source code is available at https://github.com/67L1/SinkTrack.

Method: Prompt Relay is a plug-and-play inference-time method that introduces a penalty into the cross-attention mechanism, ensuring each temporal segment attends only to its assigned prompt, allowing representation of one semantic concept at a time without architectural modifications or computational overhead.

Result: The method improves temporal prompt alignment, reduces semantic interference between events, and enhances visual quality in multi-event video generation while maintaining computational efficiency.

Conclusion: Prompt Relay provides effective temporal control for video diffusion models, addressing key limitations in multi-event video generation and enabling more coherent storytelling capabilities.

Abstract: Video diffusion models have achieved remarkable progress in generating high-quality videos. However, these models struggle to represent the temporal succession of multiple events in real-world videos and lack explicit mechanisms to control when semantic concepts appear, how long they persist, and the order in which multiple events occur. Such control is especially important for movie-grade video synthesis, where coherent storytelling depends on precise timing, duration, and transitions between events. When using a single paragraph-style prompt to describe a sequence of complex events, models often exhibit semantic entanglement, where concepts intended for different moments in the video bleed into one another, resulting in poor text-video alignment. To address these limitations, we propose Prompt Relay, an inference-time, plug-and-play method to enable fine-grained temporal control in multi-event video generation, requiring no architectural modifications and no additional computational overhead. Prompt Relay introduces a penalty into the cross-attention mechanism, so that each temporal segment attends only to its assigned prompt, allowing the model to represent one semantic concept at a time and thereby improving temporal prompt alignment, reducing semantic interference, and enhancing visual quality.

Method: Introduces COUNTINGTRICKS evaluation suite for controlled shape-based counting tests. Uses attention analysis and component-wise probing to trace visual evidence flow. Evaluates Modality Attention Share (MAS) intervention to enforce minimum visual attention during answer generation.

Result: Count-relevant visual evidence is strongest in modality projection stage but degrades substantially in later language layers where models become more susceptible to text priors. MAS intervention shows counting failures stem from underuse of visual evidence during language-stage reasoning.

Conclusion: Counting failures in VLMs result not only from visual perception limits but also from insufficient use of visual evidence during language-stage reasoning, suggesting architectural interventions can improve grounding capabilities.

Abstract: Vision–language models (VLMs) have achieved impressive performance on complex multimodal reasoning tasks, yet they still fail on simple grounding skills such as object counting. Existing evaluations mostly assess only final outputs, offering limited insight into where these failures arise inside the model. In this work, we present an empirical study of VLM counting behavior through both behavioral and mechanistic analysis. We introduce COUNTINGTRICKS, a controlled evaluation suite of simple shape-based counting cases designed to expose vulnerabilities under different patchification layouts and adversarial prompting conditions. Using attention analysis and component-wise probing, we show that count-relevant visual evidence is strongest in the modality projection stage but degrades substantially in later language layers, where models become more susceptible to text priors. Motivated by this finding, we further evaluate Modality Attention Share (MAS), a lightweight intervention that encourages a minimum budget of visual attention during answer generation. Our results suggest that counting failures in VLMs stem not only from visual perception limits, but also from the underuse of visual evidence during language-stage reasoning. Code and dataset will be released at https://github.com/leduy99/-CVPRW26-Modality-Attention-Share.

Method: Constructed a comprehensive latent style bank from seven diverse datasets to enable synthesis of latent prints with over 40 distinct styles. Implemented a semi-automated framework to analyze fingerprint ridge and minutiae integrity in generated impressions.

Result: Generation process largely preserves identity but introduces small local inconsistencies (addition/removal of minutiae), especially in poor quality reference image regions. Mismatch between reference image and style embedding causes global inconsistencies and hallucinated ridge patterns.

Conclusion: Existing synthetic fingerprint generators have limitations in simultaneously achieving diversity and identity consistency, highlighting the need for improved models that balance both aspects.

Abstract: The primary goal of this work is to systematically evaluate the intra-finger variability of synthetic fingerprints (particularly latent prints) generated using a state-of-the-art diffusion model. Specifically, we focus on enhancing the latent style diversity of the generative model by constructing a comprehensive \textit{latent style bank} curated from seven diverse datasets, which enables the precise synthesis of latent prints with over 40 distinct styles encapsulating different surfaces and processing techniques. We also implement a semi-automated framework to understand the integrity of fingerprint ridges and minutiae in the generated impressions. Our analysis indicates that though the generation process largely preserves the identity, a small number of local inconsistencies (addition and removal of minutiae) are introduced, especially when there are poor quality regions in the reference image. Furthermore, mismatch between the reference image and the chosen style embedding that guides the generation process introduces global inconsistencies in the form of hallucinated ridge patterns. These insights highlight the limitations of existing synthetic fingerprint generators and the need to further improve these models to simultaneously enhance both diversity and identity consistency.

Method: Proposes U²Flow with a decoupled learning strategy that derives uncertainty supervision from augmentation consistency via a Laplace-based maximum likelihood objective. The predicted uncertainty guides adaptive flow refinement, modulates regional smoothness loss, and enables uncertainty-guided bidirectional flow fusion.

Result: Extensive experiments on KITTI and Sintel show state-of-the-art performance among unsupervised methods while producing highly reliable uncertainty maps.

Conclusion: U²Flow effectively demonstrates the benefits of joint optical flow and uncertainty estimation in an unsupervised framework, improving both performance and interpretability.

Abstract: Unsupervised optical flow methods typically lack reliable uncertainty estimation, limiting their robustness and interpretability. We propose U$^{2}$Flow, the first recurrent unsupervised framework that jointly estimates optical flow and per-pixel uncertainty. The core innovation is a decoupled learning strategy that derives uncertainty supervision from augmentation consistency via a Laplace-based maximum likelihood objective, enabling stable training without ground truth. The predicted uncertainty is further integrated into the network to guide adaptive flow refinement and dynamically modulate the regional smoothness loss. Furthermore, we introduce an uncertainty-guided bidirectional flow fusion mechanism that enhances robustness in challenging regions. Extensive experiments on KITTI and Sintel demonstrate that U$^{2}$Flow achieves state-of-the-art performance among unsupervised methods while producing highly reliable uncertainty maps, validating the effectiveness of our joint estimation paradigm. The code is available at https://github.com/sunzunyi/U2FLOW.

Method: Integrate Linear Attention (LA) into multimodal frameworks, evaluate across ViT-S/16, ViT-B/16, and ViT-L/16 architectures trained on LAION-400M dataset, validate with ImageNet-21K zero-shot accuracy.

Result: Linear Attention achieves significant computational savings while preserving competitive performance and following the same scaling laws as standard softmax attention.

Conclusion: Linear Attention is a robust, scalable solution for next-generation multimodal Transformers processing large, complex datasets.

Abstract: Multimodal Transformers serve as the backbone for state-of-the-art vision-language models, yet their quadratic attention complexity remains a critical barrier to scalability. In this work, we investigate the viability of Linear Attention (LA) as a high-efficiency alternative within multimodal frameworks. By integrating LA, we reduce the computational overhead from quadratic to linear relative to sequence length while preserving competitive performance. We evaluate our approach across ViT-S/16, ViT-B/16, and ViT-L/16 architectures trained on the LAION-400M dataset, with validation focused on ImageNet-21K zero-shot accuracy. Our systematic evaluation demonstrates that Linear Attention not only yields significant computational savings but also adheres to the same scaling laws as standard softmax attention. These findings position Linear Attention as a robust, scalable solution for next-generation multimodal Transformers tasked with processing increasingly large and complex datasets.

Method: Dual-Anchor Introspective Decoding (DaID) uses contrastive decoding that dynamically calibrates token generation by mining the model’s internal perceptual discrepancies. It identifies a Spotlight layer to amplify visual factual signals and a Shadow layer to suppress textual inertia, guided by visual attention distributions.

Result: Experimental results across multiple benchmarks and MLLMs show DaID significantly mitigates hallucination while enhancing general reasoning capabilities.

Conclusion: DaID effectively reduces hallucinations in MLLMs by leveraging internal model representations to better align text generation with visual evidence, improving both factual accuracy and reasoning.

Abstract: Multimodal Large Language Models (MLLMs) have demonstrated remarkable reasoning capabilities yet continue to suffer from hallucination, where generated text contradicts visual content. In this paper, we introduce Dual-Anchor Introspective Decoding (DaID), a novel contrastive decoding framework that dynamically calibrates each token generation by mining the model’s internal perceptual discrepancies. Specifically, DaID identifies a Spotlight layer to amplify visual factual signals and a Shadow layer to suppress textual inertia. By leveraging visual attention distributions to guide this dual-anchor selection process, our method ensures precise, token-specific adaptation. Experimental results across multiple benchmarks and MLLMs demonstrate that DaID significantly mitigates hallucination while enhancing general reasoning capabilities.

Method: Combines state-of-the-art OCR, advanced image analysis, masked language modeling, and diffusion-based models. Pipeline detects/recognizes text, identifies degradation with occlusion detector, uses inpainting for semantically coherent reconstruction, and diffusion module reintegrates text matching font, size, and alignment.

Result: Created synthetic dataset of 30,078 degraded document images simulating diverse degradation scenarios. Proposed Unified Context Similarity Metric (UCSM) for evaluation. Pipeline advances document restoration and sets new standard for text reconstruction.

Conclusion: The work advances document restoration, benefiting archival research and digital preservation while setting a new standard for text reconstruction. Dataset and code are publicly available.

Abstract: In Document Understanding, the challenge of reconstructing damaged, occluded, or incomplete text remains a critical yet unexplored problem. Subsequent document understanding tasks can benefit from a document reconstruction process. In response, this paper presents a novel unified pipeline combining state-of-the-art Optical Character Recognition (OCR), advanced image analysis, masked language modeling, and diffusion-based models to restore and reconstruct text while preserving visual integrity. We create a synthetic dataset of 30{,}078 degraded document images that simulates diverse document degradation scenarios, setting a benchmark for restoration tasks. Our pipeline detects and recognizes text, identifies degradation with an occlusion detector, and uses an inpainting model for semantically coherent reconstruction. A diffusion-based module seamlessly reintegrates text, matching font, size, and alignment. To evaluate restoration quality, we propose a Unified Context Similarity Metric (UCSM), incorporating edit, semantic, and length similarities with a contextual predictability measure that penalizes deviations when the correct text is contextually obvious. Our work advances document restoration, benefiting archival research and digital preservation while setting a new standard for text reconstruction. The OPRB dataset and code are available at \href{https://huggingface.co/datasets/kpurkayastha/OPRB}{Hugging Face} and \href{https://github.com/kunalpurkayastha/DocRevive}{Github} respectively.

Method: Two-stream framework: primary stream models person-level motion dynamics (detecting/tracking students, extracting optical flow with RAFT, transformer encoding, attention pooling); secondary stream captures scene-level spatiotemporal information using 3D ResNet. Features combined via softmax-gated fusion.

Result: Achieved 0.9621±0.0161 classification accuracy and 0.9530±0.0204 macro-averaged F1 on Classroom Group Engagement Dataset using fivefold cross-validation. Ablation study shows contributions of both streams.

Conclusion: DualEngage effectively models group engagement by jointly considering individual actions and group dynamics, representing one of the first dual-stream approaches in classroom engagement recognition that explicitly leverages motion cues.

Abstract: Student engagement is crucial for improving learning outcomes in group activities. Highly engaged students perform better both individually and contribute to overall group success. However, most existing automated engagement recognition methods are designed for online classrooms or estimate engagement at the individual level. Addressing this gap, we propose DualEngage, a novel two-stream framework for group-level engagement recognition from in-classroom videos. It models engagement as a joint function of both individual and group-level behaviors. The primary stream models person-level motion dynamics by detecting and tracking students, extracting dense optical flow with the Recurrent All-Pairs Field Transforms network, encoding temporal motion patterns using a transformer encoder, and finally aggregating per-student representations through attention pooling into a unified representation. The secondary stream captures scene-level spatiotemporal information from the full video clip, leveraging a pretrained three-dimensional Residual Network. The two-stream representations are combined via softmax-gated fusion, which dynamically weights each stream’s contribution based on the joint context of both features. DualEngage learns a joint representation of individual actions with overarching group dynamics. We evaluate the proposed approach using fivefold cross-validation on the Classroom Group Engagement Dataset developed by Ocean University of China, achieving an average classification accuracy of 0.9621+/-0.0161 with a macro-averaged F1 of 0.9530+/-0.0204. To understand the contribution of each branch, we further conduct an ablation study comparing single-stream variants against the two-stream model. This work is among the first in classroom engagement recognition to adopt a dual-stream design that explicitly leverages motion cues as an estimator.

Method: MatRes is a zero-shot test-time adaptation framework that uses only a single low-quality and high-quality image pair. It enforces conditional similarity at corresponding locations, updating only lightweight modules while keeping all pretrained components frozen, requiring no offline training or additional supervision.

Result: Extensive experiments across diverse combinations show that MatRes yields significant gains in both restoration and geometric alignment compared to using either restoration or matching models alone.

Conclusion: MatRes offers a practical and widely applicable solution for real-world scenarios where users commonly capture multiple images of a scene with varying viewpoints and quality, effectively addressing the often-overlooked mutual interference between matching and restoration.

Abstract: Real-world image pairs often exhibit both severe degradations and large viewpoint changes, making image restoration and geometric matching mutually interfering tasks when treated independently. In this work, we propose MatRes, a zero-shot test-time adaptation framework that jointly improves restoration quality and correspondence estimation using only a single low-quality and high-quality image pair. By enforcing conditional similarity at corresponding locations, MatRes updates only lightweight modules while keeping all pretrained components frozen, requiring no offline training or additional supervision. Extensive experiments across diverse combinations show that MatRes yields significant gains in both restoration and geometric alignment compared to using either restoration or matching models alone. MatRes offers a practical and widely applicable solution for real-world scenarios where users commonly capture multiple images of a scene with varying viewpoints and quality, effectively addressing the often-overlooked mutual interference between matching and restoration.

Method: Active Diffusion Matching (ADM) integrates two interdependent score-based diffusion models to jointly estimate global transformations and local deformations via iterative Langevin Markov chain, with custom sampling strategies for input adaptability.

Result: ADM achieves state-of-the-art alignment accuracy with mAUC improvements of 5.2 points on private SFI-UWFI dataset and 0.4 points on public SFI-SFI dataset compared to existing methods.

Conclusion: ADM effectively bridges the gap in aligning SFIs and UWFIs, providing an innovative solution to a previously unaddressed challenge through joint optimization of global and local alignment.

Abstract: Objective: The study aims to address the challenge of aligning Standard Fundus Images (SFIs) and Ultra-Widefield Fundus Images (UWFIs), which is difficult due to their substantial differences in viewing range and the amorphous appearance of the retina. Currently, no specialized method exists for this task, and existing image alignment techniques lack accuracy. Methods: We propose Active Diffusion Matching (ADM), a novel cross-modal alignment method. ADM integrates two interdependent score-based diffusion models to jointly estimate global transformations and local deformations via an iterative Langevin Markov chain. This approach facilitates a stochastic, progressive search for optimal alignment. Additionally, custom sampling strategies are introduced to enhance the adaptability of ADM to given input image pairs. Results: Comparative experimental evaluations demonstrate that ADM achieves state-of-the-art alignment accuracy. This was validated on two datasets: a private dataset of SFI-UWFI pairs and a public dataset of SFI-SFI pairs, with mAUC improvements of 5.2 and 0.4 points on the private and public datasets, respectively, compared to existing state-of-the-art methods. Conclusion: ADM effectively bridges the gap in aligning SFIs and UWFIs, providing an innovative solution to a previously unaddressed challenge. The method’s ability to jointly optimize global and local alignment makes it highly effective for cross-modal image alignment tasks. Significance: ADM has the potential to transform the integrated analysis of SFIs and UWFIs, enabling better clinical utility and supporting learning-based image enhancements. This advancement could significantly improve diagnostic accuracy and patient outcomes in ophthalmology.

Method: Particle Diffusion Matching (PDM) uses iterative Random Walk Correspondence Search guided by a diffusion model that estimates displacement vectors considering local appearance, particle structural distribution, and global transformation.

Result: PDM achieves state-of-the-art performance across multiple retinal image alignment benchmarks, with substantial improvement on SFI-UWFI pairs and effectiveness in real-world clinical scenarios.

Conclusion: PDM provides accurate, scalable correspondence estimation for retinal image alignment, overcoming limitations of existing methods and enabling better multimodal integration for disease diagnosis and analysis.

Abstract: We propose a robust alignment technique for Standard Fundus Images (SFIs) and Ultra-Widefield Fundus Images (UWFIs), which are challenging to align due to differences in scale, appearance, and the scarcity of distinctive features. Our method, termed Particle Diffusion Matching (PDM), performs alignment through an iterative Random Walk Correspondence Search (RWCS) guided by a diffusion model. At each iteration, the model estimates displacement vectors for particle points by considering local appearance, the structural distribution of particles, and an estimated global transformation, enabling progressive refinement of correspondences even under difficult conditions. PDM achieves state-of-the-art performance across multiple retinal image alignment benchmarks, showing substantial improvement on a primary dataset of SFI-UWFI pairs and demonstrating its effectiveness in real-world clinical scenarios. By providing accurate and scalable correspondence estimation, PDM overcomes the limitations of existing methods and facilitates the integration of complementary retinal image modalities. This diffusion-guided search strategy offers a new direction for improving downstream supervised learning, disease diagnosis, and multi-modal image analysis in ophthalmology.

Method: Developed MAPL-EMIT, an end-to-end vision transformer framework that leverages complete radiance spectrum from EMIT instrument. Uses spectral and spatial context to jointly retrieve methane enhancements across all pixels. Trained on 3.6 million physics-based synthetic plumes injected into global EMIT radiance data.

Result: Model captures 79% of known hand-annotated plume complexes across test set of 1084 EMIT granules, while identifying twice as many plausible plumes as human analysts. Validated against airborne data, landfills, and controlled release experiments. Enables high-throughput implementation on full EMIT catalog.

Conclusion: MAPL-EMIT shifts methane monitoring from labor-intensive workflows to a rapid, scalable paradigm for global plume mapping at facility scale, enabling automated detection of previously uncaptured methane sources.

Abstract: Anthropogenic methane (CH4) point sources drive near-term climate forcing, safety hazards, and system inefficiencies. Space-based imaging spectroscopy is emerging as a tool for identifying emissions globally, but existing approaches largely rely on manual plume identification. Here we present the Methane Analysis and Plume Localization with EMIT (MAPL-EMIT) model, an end-to-end vision transformer framework that leverages the complete radiance spectrum from the Earth Surface Mineral Dust Source Investigation (EMIT) instrument to jointly retrieve methane enhancements across all pixels within a scene. This approach brings together spectral and spatial context to significantly lower detection limits. MAPL-EMIT simultaneously supports enhancement quantification, plume delineation, and source localization, even for multiple overlapping plumes. The model was trained on 3.6 million physics-based synthetic plumes injected into global EMIT radiance data. Synthetic evaluation confirms the model’s ability to identify plumes with high recall and precision and to capture weaker plumes relative to existing matched-filter approaches. On real-world benchmarks, MAPL-EMIT captures 79% of known hand-annotated NASA L2B plume complexes across a test set of 1084 EMIT granules, while capturing twice as many plausible plumes than identified by human analysts. Further validation against coincident airborne data, top-emitting landfills, and controlled release experiments confirms the model’s ability to identify previously uncaptured sources. By incorporating model-generated metrics such as spectral fit scores and estimated noise levels, the framework can further limit false-positive rates. Overall, MAPL-EMIT enables high-throughput implementation on the full EMIT catalog, shifting methane monitoring from labor-intensive workflows to a rapid, scalable paradigm for global plume mapping at the facility scale.

Method: The approach generates synthetic datasets with controlled variations, fine-tunes a LoRA adapter on each dataset, extracts a LoRA subspace for each variation type, and shows these subspaces are approximately disentangled. Integrating them creates a reduced LoRA subspace for efficient fine-tuning with improved accuracy.

Result: The method demonstrates that LoRA subspaces for different 3D variations are approximately disentangled. The reduced LoRA subspace derived from synthetic data generalizes to real datasets, enabling efficient fine-tuning with improved prediction accuracy for downstream tasks.

Conclusion: The paper provides answers to fundamental questions about LoRA subspaces in 3D foundation models, showing disentangled subspaces for different variations that can be effectively computed and generalized from synthetic to real data, improving fine-tuning efficiency and accuracy.

Abstract: With the emergence of 3D foundation models, there is growing interest in fine-tuning them for downstream tasks, where LoRA is the dominant fine-tuning paradigm. As 3D datasets exhibit distinct variations in texture, geometry, camera motion, and lighting, there are interesting fundamental questions: 1) Are there LoRA subspaces associated with each type of variation? 2) Are these subspaces disentangled (i.e., orthogonal to each other)? 3) How do we compute them effectively? This paper provides answers to all these questions. We introduce a robust approach that generates synthetic datasets with controlled variations, fine-tunes a LoRA adapter on each dataset, and extracts a LoRA sub-space associated with each type of variation. We show that these subspaces are approximately disentangled. Integrating them leads to a reduced LoRA subspace that enables efficient LoRA fine-tuning with improved prediction accuracy for downstream tasks. In particular, we show that such a reduced LoRA subspace, despite being derived entirely from synthetic data, generalizes to real datasets. An ablation study validates the effectiveness of the choices in our approach.

Method: 1) Unified embodiment interface with capability-driven scheduling for heterogeneous robot coordination; 2) Visual-centric cross-embodiment multimodal memory for persistent context retention; 3) Critic-based closed-loop feedback with generalist reward model for online evaluation and replanning. Uses decoupled architecture across OpenClaw, shared service, and robot embodiment layers.

Result: ABot-Claw enables real-world interaction, closes the loop from natural language intent to physical action, and supports progressively self-evolving robotic agents in open, dynamic environments.

Conclusion: The system bridges the gap between reasoning and execution, providing a framework for embodied agents that can operate in open-world environments with closed-loop feedback and multi-robot coordination.

Abstract: Current embodied intelligent systems still face a substantial gap between high-level reasoning and low-level physical execution in open-world environments. Although Vision-Language-Action (VLA) models provide strong perception and intuitive responses, their open-loop nature limits long-horizon performance. Agents incorporating System 2 cognitive mechanisms improve planning, but usually operate in closed sandboxes with predefined toolkits and limited real-system control. OpenClaw provides a localized runtime with full system privileges, but lacks the embodied control architecture required for long-duration, multi-robot execution. We therefore propose ABot-Claw, an embodied extension of OpenClaw that integrates: 1) a unified embodiment interface with capability-driven scheduling for heterogeneous robot coordination; 2) a visual-centric cross-embodiment multimodal memory for persistent context retention and grounded retrieval; and 3) a critic-based closed-loop feedback mechanism with a generalist reward model for online progress evaluation, local correction, and replanning. With a decoupled architecture spanning the OpenClaw layer, shared service layer, and robot embodiment layer, ABot-Claw enables real-world interaction, closes the loop from natural language intent to physical action, and supports progressively self-evolving robotic agents in open, dynamic environments.

Method: Proposes Degradation-Consistent Paired Training (DCPT) that constructs clean and degraded views of each training image, then applies two constraints: feature consistency loss (cosine distance between representations) and prediction consistency loss (symmetric KL divergence between output distributions).

Result: On Synthbuster benchmark (9 generators, 8 degradation conditions), DCPT improves degraded-condition average accuracy by 9.1 percentage points compared to baseline, with only 0.9% clean accuracy sacrifice. Most significant improvement under JPEG compression (+15.7% to +17.9%).

Conclusion: Training objective improvement (via DCPT) is more effective than architectural augmentation for degradation robustness. The method adds zero parameters and zero inference overhead while significantly improving robustness to real-world image corruptions.

Abstract: AI-generated image detectors suffer significant performance degradation under real-world image corruptions such as JPEG compression, Gaussian blur, and resolution downsampling. We observe that state-of-the-art methods, including B-Free, treat degradation robustness as a byproduct of data augmentation rather than an explicit training objective. In this work, we propose Degradation-Consistent Paired Training (DCPT), a simple yet effective training strategy that explicitly enforces robustness through paired consistency constraints. For each training image, we construct a clean view and a degraded view, then impose two constraints: a feature consistency loss that minimizes the cosine distance between clean and degraded representations, and a prediction consistency loss based on symmetric KL divergence that aligns output distributions across views. DCPT adds zero additional parameters and zero inference overhead. Experiments on the Synthbuster benchmark (9 generators, 8 degradation conditions) demonstrate that DCPT improves the degraded-condition average accuracy by 9.1 percentage points compared to an identical baseline without paired training, while sacrificing only 0.9% clean accuracy. The improvement is most pronounced under JPEG compression (+15.7% to +17.9%). Ablation further reveals that adding architectural components leads to overfitting on limited training data, confirming that training objective improvement is more effective than architectural augmentation for degradation robustness.

Method: Proposes Hybrid Forcing with three key components: 1) lightweight linear temporal attention to preserve long-range dependencies beyond sliding windows using compact key-value states, 2) block-sparse attention within local sliding windows to reduce redundant computation, and 3) decoupled distillation strategy tailored to the hybrid attention design for stable optimization.

Result: Achieves state-of-the-art performance on both short- and long-form video generation benchmarks. Achieves real-time, unbounded 832x480 video generation at 29.5 FPS on a single NVIDIA H100 GPU without quantization or model compression.

Conclusion: Hybrid Forcing effectively addresses the limitations of sliding window attention in streaming video generation by balancing temporal context preservation with computational efficiency, enabling real-time high-quality video generation.

Abstract: Streaming video generation (SVG) distills a pretrained bidirectional video diffusion model into an autoregressive model equipped with sliding window attention (SWA). However, SWA inevitably loses distant history during long video generation, and its computational overhead remains a critical challenge to real-time deployment. In this work, we propose Hybrid Forcing, which jointly optimizes temporal information retention and computational efficiency through a hybrid attention design. First, we introduce lightweight linear temporal attention to preserve long-range dependencies beyond the sliding window. In particular, we maintain a compact key-value state to incrementally absorb evicted tokens, retaining temporal context with negligible memory and computational overhead. Second, we incorporate block-sparse attention into the local sliding window to reduce redundant computation within short-range modeling, reallocating computational capacity toward more critical dependencies. Finally, we introduce a decoupled distillation strategy tailored to the hybrid attention design. A few-step initial distillation is performed under dense attention, then the distillation of our proposed linear temporal and block-sparse attention is activated for streaming modeling, ensuring stable optimization. Extensive experiments on both short- and long-form video generation benchmarks demonstrate that Hybrid Forcing consistently achieves state-of-the-art performance. Notably, our model achieves real-time, unbounded 832x480 video generation at 29.5 FPS on a single NVIDIA H100 GPU without quantization or model compression. The source code and trained models are available at https://github.com/leeruibin/hybrid-forcing.

Method: VGGT-HPE uses a relative formulation built on a general-purpose geometry foundation model. It’s fine-tuned exclusively on synthetic facial renderings, reducing the problem to estimating geometric displacement from an explicitly provided anchor with known pose. The anchor can be chosen at test time (e.g., near-neutral or temporally adjacent frames).

Result: Despite zero real-world training data, VGGT-HPE achieves state-of-the-art results on the BIWI benchmark, outperforming established absolute regression methods trained on mixed and real datasets. Controlled benchmarks show relative prediction is intrinsically more accurate than absolute regression, with advantages scaling with pose difficulty.

Conclusion: Relative head pose estimation is more robust than absolute regression, especially for difficult poses. The method demonstrates strong generalization from synthetic to real data and allows flexible anchor selection at test time.

Abstract: Monocular head pose estimation is traditionally formulated as direct regression from a single image to an absolute pose. This paradigm forces the network to implicitly internalize a dataset-specific canonical reference frame. In this work, we argue that predicting the relative rigid transformation between two observed head configurations is a fundamentally easier and more robust formulation. We introduce VGGT-HPE, a relative head pose estimator built upon a general-purpose geometry foundation model. Finetuned exclusively on synthetic facial renderings, our method sidesteps the need for an implicit anchor by reducing the problem to estimating a geometric displacement from an explicitly provided anchor with a known pose. As a practical benefit, the relative formulation also allows the anchor to be chosen at test time - for instance, a near-neutral frame or a temporally adjacent one - so that the prediction difficulty can be controlled by the application. Despite zero real-world training data, VGGT-HPE achieves state-of-the-art results on the BIWI benchmark, outperforming established absolute regression methods trained on mixed and real datasets. Through controlled easy- and hard-pair benchmarks, we also systematically validate our core hypothesis: relative prediction is intrinsically more accurate than absolute regression, with the advantage scaling alongside the difficulty of the target pose. Project page and code: https://vasilikivas.github.io/VGGT-HPE

Method: Dual-branch system with HAT-L branch (local transformer) and MambaIRv2-L branch (global state-space model), using test-time local conversion on HAT, eight-way self-ensemble on MambaIRv2, and fixed equal-weight image-space fusion.

Result: Won NTIRE 2026 Infrared Image Super-Resolution Challenge; fused output outperforms single branches on 12 synthetic 4× thermal samples from Caltech Aerial RGB-Thermal dataset in PSNR, SSIM, and overall Score.

Conclusion: Infrared super-resolution benefits from explicit complementarity between locally strong transformer restoration and globally stable state-space modeling.

Abstract: Remote sensing infrared image super-resolution aims to recover sharper thermal observations from low-resolution inputs while preserving target contours, scene layout, and radiometric stability. Unlike visible-image super-resolution, thermal imagery is weakly textured and more sensitive to unstable local sharpening, which makes complementary local and global modeling especially important. This paper presents our solution to the NTIRE 2026 Infrared Image Super-Resolution Challenge, a dual-branch system that combines a HAT-L branch and a MambaIRv2-L branch. The inference pipeline applies test-time local conversion on HAT, eight-way self-ensemble on MambaIRv2, and fixed equal-weight image-space fusion. We report both the official challenge score and a reproducible evaluation on 12 synthetic times-four thermal samples derived from Caltech Aerial RGB-Thermal, on which the fused output outperforms either single branch in PSNR, SSIM, and the overall Score. The results suggest that infrared super-resolution benefits from explicit complementarity between locally strong transformer restoration and globally stable state-space modeling.

Method: Proposes a dual cross-attention-based multimodal fusion framework that explicitly models bidirectional interactions between structural MRI (sMRI) and resting-state functional MRI (rs-fMRI) representations. Tested on REST-meta-MDD dataset using both structural and functional brain atlas configurations with 10-fold stratified cross-validation.

Result: The fusion algorithm achieves robust and competitive performance across all atlas types, consistently outperforming conventional feature-level concatenation for functional atlases while maintaining comparable performance for structural atlases. Best model obtained 84.71% accuracy, 86.42% sensitivity, 82.89% specificity, 84.34% precision, and 85.37% F1-score.

Conclusion: The findings emphasize the importance of explicitly modeling cross-modal interactions for multimodal neuroimaging-based MDD classification, demonstrating that sophisticated fusion approaches can improve diagnostic performance.

Abstract: Major depressive disorder (MDD) is a prevalent mental disorder associated with complex neurobiological changes that cannot be fully captured using a single imaging modality. The use of multimodal magnetic resonance imaging (MRI) provides a more comprehensive understanding of brain changes by combining structural and functional data. Despite this, the effective integration of these modalities remains challenging. In this study, we propose a dual cross-attention-based multimodal fusion framework that explicitly models bidirectional interactions between structural MRI (sMRI) and resting-state functional MRI (rs-fMRI) representations. The proposed approach is tested on the large-scale REST-meta-MDD dataset using both structural and functional brain atlas configurations. Numerous experiments conducted under a 10-fold stratified cross-validation demonstrated that the proposed fusion algorithm achieves robust and competitive performance across all atlas types. The proposed method consistently outperforms conventional feature-level concatenation for functional atlases, while maintaining comparable performance for structural atlases. The most effective dual cross-attention multimodal model obtained 84.71% accuracy, 86.42% sensitivity, 82.89% specificity, 84.34% precision, and 85.37% F1-score. These findings emphasize the importance of explicitly modeling cross-modal interactions for multimodal neuroimaging-based MDD classification.

Method: Proposes PhyMix with two components: (1) Scene-GRPO for implicit alignment using physics evaluator as preference signal, and (2) Test-Time Optimizer (TTO) for explicit refinement using differentiable evaluator signals. Also introduces a unified Physics Evaluator benchmark measuring geometric priors, contact, stability, and deployability.

Result: State-of-the-art performance in both visual fidelity and physical plausibility. Extensive evaluations show robustness in synthetic and real-world images.

Conclusion: The framework successfully unifies evaluation, reward shaping, and inference-time correction to produce 3D indoor scenes that are both visually faithful and physically plausible.

Abstract: Existing single-image 3D indoor scene generators often produce results that look visually plausible but fail to obey real-world physics, limiting their reliability in robotics, embodied AI, and design. To examine this gap, we introduce a unified Physics Evaluator that measures four main aspects: geometric priors, contact, stability, and deployability, which are further decomposed into nine sub-constraints, establishing the first benchmark to measure physical consistency. Based on this evaluator, our analysis shows that state-of-the-art methods remain largely physics-unaware. To overcome this limitation, we further propose a framework that integrates feedback from the Physics Evaluator into both training and inference, enhancing the physical plausibility of generated scenes. Specifically, we propose PhyMix, which is composed of two complementary components: (i) implicit alignment via Scene-GRPO, a critic-free group-relative policy optimization that leverages the Physics Evaluator as a preference signal and biases sampling towards physically feasible layouts, and (ii) explicit refinement via a plug-and-play Test-Time Optimizer (TTO) that uses differentiable evaluator signals to correct residual violations during generation. Overall, our method unifies evaluation, reward shaping, and inference-time correction, producing 3D indoor scenes that are visually faithful and physically plausible. Extensive synthetic evaluations confirm state-of-the-art performance in both visual fidelity and physical plausibility, and extensive qualitative examples in stylized and real-world images further showcase the robustness of the method. We will release codes and models upon publication.

Method: Developed a principled three-tier taxonomy (Aesthetic Quality, Aesthetic Tagging, Generation Quality) with fine-grained sub-dimensions. Created 1,016 diverse prompts and generated over 60,000 videos using 12 video generation models. Annotated subset via human labeling and developed three multi-task neural assessors: VAQA-Net (aesthetic quality), VTag-Net (aesthetic tagging), and VGQA-Net (generation quality).

Result: The models achieve reliable alignment with human judgments, offering both accuracy and efficiency. The benchmark enables scalable and automated evaluation of video generation models across comprehensive quality dimensions.

Conclusion: VGA-Bench provides a unified benchmark for joint evaluation of video generation quality and aesthetic quality, addressing limitations in existing evaluation frameworks and enabling applications in content moderation, model debugging, and generative model optimization.

Abstract: The rapid advancement of AIGC-based video generation has underscored the critical need for comprehensive evaluation frameworks that go beyond traditional generation quality metrics to encompass aesthetic appeal. However, existing benchmarks remain largely focused on technical fidelity, leaving a significant gap in holistic assessment-particularly with respect to perceptual and artistic qualities. To address this limitation, we introduce VGA-Bench, a unified benchmark for joint evaluation of video generation quality and aesthetic quality. VGA-Bench is built upon a principled three-tier taxonomy: Aesthetic Quality, Aesthetic Tagging, and Generation Quality, each decomposed into multiple fine-grained sub-dimensions to enable systematic assessment. Guided by this taxonomy, we design 1,016 diverse prompts and generate a large-scale dataset of over 60,000 videos using 12 video generation models, ensuring broad coverage across content, style, and artifacts. To enable scalable and automated evaluation, we annotate a subset of the dataset via human labeling and develop three dedicated multi-task neural assessors: VAQA-Net for aesthetic quality prediction, VTag-Net for automatic aesthetic tagging, and VGQA-Net for generation and basic quality attributes. Extensive experiments demonstrate that our models achieve reliable alignment with human judgments, offering both accuracy and efficiency. We release VGA-Bench as a public benchmark to foster research in AIGC evaluation, with applications in content moderation, model debugging, and generative model optimization.

Method: Used nnU-Net ResEnc M architecture on HNTS-MRG 2024 dataset for multi-label segmentation. Compared standard Dice loss baseline against two Volume-Aware configurations: “Dual Mask” (VA loss on both PT and LN) and “Selective LN Mask” (VA loss on LN only). Evaluated using volumetric Dice scores, surface-based metrics, and lesion-wise detection sensitivity/precision.

Result: Selective LN Mask configuration achieved highest LN Volumetric Dice Score (0.758 vs. 0.734 baseline) and improved LN Lesion-Wise Detection Sensitivity (84.93% vs. 81.80%). However, PT detection precision declined significantly (63.65% vs. 81.27%). Dual Mask configuration provided most balanced performance, maintaining PT precision at 82.04% while improving LN sensitivity to 83.46%.

Conclusion: Volume-sensitive loss function helps mitigate under-representation of small metastatic lesions in HNC segmentation. While selective weighting yields best nodal detection, dual-mask approach is necessary in multi-label tasks to maintain segmentation accuracy for larger primary tumor volumes.

Abstract: Background: Manual delineation of target volumes in head and neck cancer (HNC) remains a significant bottleneck in radiotherapy planning, characterized by high inter-observer variability and time consumption. This study evaluates the integration of a Volume-Aware (VA) Dice loss function into a self-configuring deep learning framework to enhance the auto-segmentation of primary tumors (PT) and metastatic lymph nodes (LN) for adaptive MR-guided radiotherapy. We investigate how volume-sensitive weighting affects the detection of small, anatomically complex nodal metastases compared to conventional loss functions. Methods: Utilizing the HNTS-MRG 2024 dataset, we implemented an nnU-Net ResEnc M architecture. We conducted a multi-label segmentation task, comparing a standard Dice loss baseline against two Volume-Aware configurations: a “Dual Mask” setup (VA loss on both PT and LN) and a “Selective LN Mask” setup (VA loss on LN only). Evaluation metrics included volumetric Dice scores, surface-based metrics (SDS, MSD, HD95), and lesion-wise binary detection sensitivity and precision. Results: The Selective LN Mask configuration achieved the highest LN Volumetric Dice Score (0.758 vs. 0.734 baseline) and significantly improved LN Lesion-Wise Detection Sensitivity (84.93% vs. 81.80%). However, a critical trade-off was observed; PT detection precision declined significantly in the selective setup (63.65% vs. 81.27%). The Dual Mask configuration provided the most balanced performance across both targets, maintaining primary tumor precision at 82.04% while improving LN sensitivity to 83.46%. Conclusions: A volume-sensitive loss function mitigated the under-representation of small metastatic lesions in HNC. While selective weighting yielded the best nodal detection, a dual-mask approach is required in multi-label tasks to maintain segmentation accuracy for larger primary tumor volumes.

Method: TRACE framework with three progressively coupled components: 1) Semantic anchoring to identify meaningful regions supporting image understanding, 2) Semantic perturbation sensing using frequency cues to capture subtle edits under visual consistency, and 3) Semantic-constrained reasoning to verify candidate regions through joint reasoning over semantic content and scope.

Result: TRACE consistently outperforms existing IML methods on the constructed SML benchmark, producing more complete, compact, and semantically coherent localization results.

Conclusion: The work demonstrates the necessity of moving beyond artifact-based localization and provides a new direction for image forensics in complex semantic editing scenarios through the proposed Semantic Manipulation Localization task and TRACE framework.

Abstract: Image Manipulation Localization (IML) aims to identify edited regions in an image. However, with the increasing use of modern image editing and generative models, many manipulations no longer exhibit obvious low-level artifacts. Instead, they often involve subtle but meaning-altering edits to an object’s attributes, state, or relationships while remaining highly consistent with the surrounding content. This makes conventional IML methods less effective because they mainly rely on artifact detection rather than semantic sensitivity. To address this issue, we introduce Semantic Manipulation Localization (SML), a new task that focuses on localizing subtle semantic edits that significantly change image interpretation. We further construct a dedicated fine-grained benchmark for SML using a semantics-driven manipulation pipeline with pixel-level annotations. Based on this task, we propose TRACE (Targeted Reasoning of Attributed Cognitive Edits), an end-to-end framework that models semantic sensitivity through three progressively coupled components: semantic anchoring, semantic perturbation sensing, and semantic-constrained reasoning. Specifically, TRACE first identifies semantically meaningful regions that support image understanding, then injects perturbation-sensitive frequency cues to capture subtle edits under strong visual consistency, and finally verifies candidate regions through joint reasoning over semantic content and semantic scope. Extensive experiments show that TRACE consistently outperforms existing IML methods on our benchmark and produces more complete, compact, and semantically coherent localization results. These results demonstrate the necessity of moving beyond artifact-based localization and provide a new direction for image forensics in complex semantic editing scenarios.

Method: ColChunk introduces multimodal late chunking using hierarchical clustering on patch-level embeddings fused with a 2D position prior to ensure spatial-semantic coherence. This adaptive grouping creates content-aware representations that preserve global context while drastically reducing vector count.

Result: Evaluations across 24 VDR datasets show ColChunk achieves over 90% reduction in storage requirements while delivering a 9-point average improvement in nDCG@5 across representative single-vector models.

Conclusion: ColChunk provides a practical solution for balancing retrieval accuracy and efficiency in visual document systems, offering a plug-and-play framework that significantly reduces computational and storage costs while improving performance.

Abstract: Multi-vector models dominate Visual Document Retrieval (VDR) due to their fine-grained matching capabilities, but their high storage and computational costs present a major barrier to practical deployment. In this paper, we propose ColChunk, a plug-and-play framework that introduces multimodal late chunking to construct efficient, contextualized multi-vectors. Unlike existing pruning or fixed-token approaches, ColChunk employs hierarchical clustering on patch-level embeddings, fused with a 2D position prior to ensure spatial-semantic coherence. This adaptive grouping allows for a content-aware representation that preserves global context while drastically reducing the vector count. Evaluations across 24 VDR datasets demonstrate ColChunk achieves over a 90% reduction in storage requirements while simultaneously delivering a 9-point average improvement in nDCG@5 across representative single-vector models. ColChunk provides a practical solution for balancing retrieval accuracy and efficiency in visual document systems.

Method: Proposes: 1) Automated Quality Assessment Agent (AQAA) to identify low-quality samples in MIMIC-CXR dataset, establishing LRRG benchmark; 2) Dual-loop Training Strategy using bi-level optimization and gradient consistency to learn quality-agnostic diagnostic features by aligning gradient directions across quality variations.

Result: Extensive experiments show the approach effectively mitigates model performance degradation caused by image quality deterioration. Code and data to be released upon acceptance.

Conclusion: The framework bridges the gap between ideal lab conditions and real-world clinical environments by explicitly addressing image quality variations in radiology report generation.

Abstract: Vision-Language Models (VLMs) have significantly advanced automated Radiology Report Generation (RRG). However, existing methods implicitly assume high-quality inputs, overlooking the noise and artifacts prevalent in real-world clinical environments. Consequently, current models exhibit severe performance degradation when processing suboptimal images. To bridge this gap, we propose a robust report generation framework explicitly designed for image quality variations. We first introduce an Automated Quality Assessment Agent (AQAA) to identify low-quality samples within the MIMIC-CXR dataset and establish the Low-quality Radiology Report Generation (LRRG) benchmark. To tackle degradation-induced shifts, we propose a novel Dual-loop Training Strategy leveraging bi-level optimization and gradient consistency. This approach ensures the model learns quality-agnostic diagnostic features by aligning gradient directions across varying quality regimes. Extensive experiments demonstrate that our approach effectively mitigates model performance degradation caused by image quality deterioration. The code and data will be released upon acceptance.

Method: Proposes Asymptotic Content-Aware Pyramid Attention Network (A3-FPN) with horizontally-spread column network for asymptotically global feature interaction, content-aware attention modules for feature fusion with position-wise offsets and weights, and feature reassembly based on information content.

Result: Achieves 49.6 mask AP on MS COCO and 85.6 mIoU on Cityscapes when paired with OneFormer and Swin-L backbone, demonstrating significant performance gains across multiple datasets.

Conclusion: A3-FPN effectively enhances multi-scale feature representation and can be integrated into various CNN and Transformer-based architectures for improved dense prediction performance.

Abstract: Learning multi-scale representations is the common strategy to tackle object scale variation in dense prediction tasks. Although existing feature pyramid networks have greatly advanced visual recognition, inherent design defects inhibit them from capturing discriminative features and recognizing small objects. In this work, we propose Asymptotic Content-Aware Pyramid Attention Network (A3-FPN), to augment multi-scale feature representation via the asymptotically disentangled framework and content-aware attention modules. Specifically, A3-FPN employs a horizontally-spread column network that enables asymptotically global feature interaction and disentangles each level from all hierarchical representations. In feature fusion, it collects supplementary content from the adjacent level to generate position-wise offsets and weights for context-aware resampling, and learns deep context reweights to improve intra-category similarity. In feature reassembly, it further strengthens intra-scale discriminative feature learning and reassembles redundant features based on information content and spatial variation of feature maps. Extensive experiments on MS COCO, VisDrone2019-DET and Cityscapes demonstrate that A3-FPN can be easily integrated into state-of-the-art CNN and Transformer-based architectures, yielding remarkable performance gains. Notably, when paired with OneFormer and Swin-L backbone, A3-FPN achieves 49.6 mask AP on MS COCO and 85.6 mIoU on Cityscapes. Codes are available at https://github.com/mason-ching/A3-FPN.

Method: Evaluated 24 pretrained matcher families in zero-shot setting on SpaceNet9 and two cross-modal benchmarks using deterministic protocol with tiled large-image inference, robust geometric filtering, and tie-point-grounded metrics.

Result: XoFTR and RoMa achieved lowest mean error (3.0 px), but RoMa did so without cross-modal training. MatchAnything-ELoFTR (3.4 px) trained on synthetic pairs performed similarly. Protocol choices (geometry model, tile size, inlier gating) affected accuracy up to 33×. 3D-reconstruction matchers (MASt3R, DUSt3R) were protocol-sensitive and fragile.

Conclusion: Foundation-model features (DINOv2) may provide modality invariance that substitutes for explicit cross-modal supervision. Protocol design significantly impacts performance, sometimes more than matcher choice. Findings guide practical deployment and future matcher design for cross-modal satellite registration.

Abstract: Cross-modal optical-SAR (Synthetic Aperture Radar) registration is a bottleneck for disaster-response via remote sensing, yet modern image matchers are developed and benchmarked almost exclusively on natural-image domains. We evaluate twenty-four pretrained matcher families–in a zero-shot setting with no fine-tuning or domain adaptation on satellite or SAR data–on SpaceNet9 and two additional cross-modal benchmarks under a deterministic protocol with tiled large-image inference, robust geometric filtering, and tie-point-grounded metrics. Our results reveal asymmetric transfer–matchers with explicit cross-modal training do not uniformly outperform those without it. While XoFTR (trained for visible-thermal matching) and RoMa achieve the lowest reported mean error at $3.0$ px on the labeled SpaceNet9 training scenes, RoMa achieves this without any cross-modal training, and MatchAnything-ELoFTR ($3.4$ px)–trained on synthetic cross-modal pairs–matches closely, suggesting (as a working hypothesis) that foundation-model features (DINOv2) may contribute to modality invariance that partially substitutes for explicit cross-modal supervision. 3D-reconstruction matchers (MASt3R, DUSt3R), which are not designed for traditional 2D image matching, are highly protocol-sensitive and remain fragile under default settings. Deployment protocol choices (geometry model, tile size, inlier gating) shift accuracy by up to $33\times$ for a single matcher, sometimes exceeding the effect of swapping matchers entirely within the evaluated sweep–affine geometry alone reduces mean error from $12.34$ to $9.74$ px. These findings inform both practical deployment of existing matchers and future matcher design for cross-modal satellite registration.

Method: Integrates Vision Foundation Models with Feature Pyramid Networks for robust feature representation, plus a data augmentation mechanism that enforces feature consistency under illumination variations and regularizes disparity prediction consistency between augmented and standard samples.

Result: Achieves state-of-the-art performance among self-supervised methods on multiple benchmarks, competes with supervised methods, and even outperforms some SOTA supervised methods like CFNet on the challenging Booster benchmark.

Conclusion: SMFormer demonstrates that integrating VFMs with effective data augmentation can overcome limitations of photometric consistency assumptions in self-supervised stereo matching, closing the gap with supervised approaches.

Abstract: Recent self-supervised stereo matching methods have made significant progress. They typically rely on the photometric consistency assumption, which presumes corresponding points across views share the same appearance. However, this assumption could be compromised by real-world disturbances, resulting in invalid supervisory signals and a significant accuracy gap compared to supervised methods. To address this issue, we propose SMFormer, a framework integrating more reliable self-supervision guided by the Vision Foundation Model (VFM) and data augmentation. We first incorporate the VFM with the Feature Pyramid Network (FPN), providing a discriminative and robust feature representation against disturbance in various scenarios. We then devise an effective data augmentation mechanism that ensures robustness to various transformations. The data augmentation mechanism explicitly enforces consistency between learned features and those influenced by illumination variations. Additionally, it regularizes the output consistency between disparity predictions of strong augmented samples and those generated from standard samples. Experiments on multiple mainstream benchmarks demonstrate that our SMFormer achieves state-of-the-art (SOTA) performance among self-supervised methods and even competes on par with supervised ones. Remarkably, in the challenging Booster benchmark, SMFormer even outperforms some SOTA supervised methods, such as CFNet.

Method: 1) Transfer 2D MLLM (trained on natural images) to support 3D medical volumetric inputs while reusing all pre-trained parameters. 2) Design Text-Guided Hierarchical MoE (TGH-MoE) framework to distinguish tasks under text prompt guidance. 3) Two-stage training strategy to learn both task-shared and task-specific image features.

Result: The method outperforms existing 3D medical MLLMs in both Medical Report Generation (MRG) and Medical Visual Question Answering (MVQA) tasks.

Conclusion: The proposed approach successfully addresses limitations of existing 3D medical MLLMs by transferring 2D MLLMs and using task-specific feature extraction, demonstrating superior performance on clinical tasks.

Abstract: 3D medical image analysis is of great importance in disease diagnosis and treatment. Recently, multimodal large language models (MLLMs) have exhibited robust perceptual capacity, strong cross-modal alignment, and promising generalizability. Therefore, they have great potential to improve the performance of medical report generation (MRG) and medical visual question answering (MVQA), which serve as two important tasks in clinical scenarios. However, due to the scarcity of 3D medical images, existing 3D medical MLLMs suffer from insufficiently pretrained vision encoder and inability to extract customized image features for different kinds of tasks. In this paper, we propose to first transfer a 2D MLLM, which is well trained with 2D natural images, to support 3D medical volumetric inputs while reusing all of its pre-trained parameters. To enable the vision encoder to extract tailored image features for various tasks, we then design a Text-Guided Hierarchical MoE (TGH-MoE) framework, which can distinguish tasks under the guidance of the text prompt. Furthermore, we propose a two-stage training strategy to learn both task-shared and task-specific image features. As demonstrated empirically, our method outperforms existing 3D medical MLLMs in both MRG and MVQA tasks. Our code will be released once this paper is accepted.

Method: Proposes MedVeriSeg with a Similarity Response Quality Scoring Module that analyzes similarity maps between [SEG] token features and MLLM image features across three aspects: strength, compactness, and purity. Incorporates GPT-4o to jointly assess similarity heatmaps and scoring module results for final verification.

Result: Experiments on a benchmark from SA-Med2D-20M show MedVeriSeg effectively rejects false-query segmentation requests while maintaining reliable recognition of true queries.

Conclusion: MedVeriSeg provides a training-free verification framework that enhances the reliability of medical segmentation MLLMs by preventing hallucinated segmentations for non-existent targets.

Abstract: Despite recent advances in MLLM-based medical image segmentation, existing LISA-like methods cannot reliably reject false queries and often produce hallucinated segmentation masks for absent targets. This limitation reduces practical reliability in both medical education and clinical use. In this work, we propose MedVeriSeg, a training-free verification framework that equips LISA-like medical segmentation models with the ability to identify and reject false queries which contain non-existent targets. Our key observation is that the similarity map between the [SEG] token feature and MLLM image features exhibits markedly different distribution patterns for true and false queries. Based on this, we introduce a Similarity Response Quality Scoring Module that characterizes the similarity map from three aspects: strength, compactness, and purity, producing an initial target-existence prediction. We further incorporate qualitative visual evidence by using GPT-4o to jointly assess the similarity heatmap and the results of Similarity Response Quality Scoring Module for final verification. Experiments on a small-scale benchmark constructed from SA-Med2D-20M show that MedVeriSeg effectively rejects false-query segmentation requests while maintaining reliable recognition of true queries.

Method: Discrete-action RL framework formulates registration as sequential decision-making. Uses shared feature encoder warm-started from supervised pose estimation network to extract features from CT renderings and laparoscopic frames. RL policy head learns to choose rigid transformations along 6DoF and decide when to stop iterations.

Result: Achieved average target registration error (TRE) of 15.70 mm on public laparoscopic dataset, comparable to supervised approaches with optimization, while achieving faster convergence.

Conclusion: RL-based formulation enables automated, efficient iterative registration without manual parameter tuning. Discrete framework provides foundation for future continuous-action and deformable registration models in surgical AR applications.

Abstract: Registration between preoperative CT and intraoperative laparoscopic video plays a crucial role in augmented reality (AR) guidance for minimally invasive surgery. Learning-based methods have recently achieved registration errors comparable to optimization-based approaches while offering faster inference. However, many supervised methods produce coarse alignments that rely on additional optimization-based refinement, thereby increasing inference time. We present a discrete-action reinforcement learning (RL) framework that formulates CT-to-video registration as a sequential decision-making process. A shared feature encoder, warm-started from a supervised pose estimation network to provide stable geometric features and faster convergence, extracts representations from CT renderings and laparoscopic frames, while an RL policy head learns to choose rigid transformations along six degrees of freedom and to decide when to stop the iteration. Experiments on a public laparoscopic dataset demonstrated that our method achieved an average target registration error (TRE) of 15.70 mm, comparable to supervised approaches with optimization, while achieving faster convergence. The proposed RL-based formulation enables automated, efficient iterative registration without manually tuned step sizes or stopping criteria. This discrete framework provides a practical foundation for future continuous-action and deformable registration models in surgical AR applications.

Method: Comparative evaluation between traditional MVS pipeline (COLMAP) and state-of-the-art learning-based approaches including geometry-guided methods (MVSNet, PatchmatchNet, MVSAnywhere, MVSFormer++) and end-to-end frameworks (Stereo4D, FoundationStereo, DUSt3R, MASt3R, Fast3R, VGGT). Two experiments on aerial scenarios: MARS-LVIG dataset with LiDAR ground truth, and Pix4D public scene with Pix4Dmapper ground truth. Evaluated accuracy, coverage, and runtime.

Result: COLMAP provides reliable, geometrically consistent results but requires more computation time. Learning-based methods show stronger feature-matching capability and robustness when traditional methods fail. Geometry-guided methods need careful dataset preparation and often depend on COLMAP priors. End-to-end methods (DUSt3R, VGGT) achieve competitive accuracy with substantially faster reconstruction but exhibit larger residuals in challenging scenarios.

Conclusion: Learning-based MVS methods offer promising alternatives to traditional approaches with better speed and robustness, though trade-offs exist in accuracy and dependency on traditional pipelines. End-to-end methods show particular promise for fast reconstruction despite some accuracy limitations.

Abstract: Photogrammetric 3D reconstruction has long relied on traditional Structure-from-Motion (SfM) and Multi-View Stereo (MVS) methods, which provide high accuracy but face challenges in speed and scalability. Recently, learning-based MVS methods have emerged, aiming for faster and more efficient reconstruction. This work presents a comparative evaluation between a representative traditional MVS pipeline (COLMAP) and state-of-the-art learning-based approaches, including geometry-guided methods (MVSNet, PatchmatchNet, MVSAnywhere, MVSFormer++) and end-to-end frameworks (Stereo4D, FoundationStereo, DUSt3R, MASt3R, Fast3R, VGGT). Two experiments were conducted on different aerial scenarios. The first experiment used the MARS-LVIG dataset, where ground-truth 3D reconstruction was provided by LiDAR point clouds. The second experiment used a public scene from the Pix4D official website, with ground truth generated by Pix4Dmapper. We evaluated accuracy, coverage, and runtime across all methods. Experimental results show that although COLMAP can provide reliable and geometrically consistent reconstruction results, it requires more computation time. In cases where traditional methods fail in image registration, learning-based approaches exhibit stronger feature-matching capability and greater robustness. Geometry-guided methods usually require careful dataset preparation and often depend on camera pose or depth priors generated by COLMAP. End-to-end methods such as DUSt3R and VGGT achieve competitive accuracy and reasonable coverage while offering substantially faster reconstruction. However, they exhibit relatively large residuals in 3D reconstruction, particularly in challenging scenarios.

Method: Predicts 3D Gaussian primitives associated with each SMPL-X vertex in canonical pose. Uses one constrained Gaussian near SMPL-X surface for geometric prior, plus additional unconstrained Gaussians per vertex to capture clothing/hair deviations. Enables animation via linear blend skinning without further network evaluation.

Result: Achieves reconstruction quality comparable to state-of-the-art methods on THuman 2.1, AvatarReX and THuman 4.0 datasets while uniquely supporting real-time animation and interactive applications.

Conclusion: The method provides an efficient, animatable human representation from single forward pass, bridging the gap between high-quality reconstruction and real-time animation capabilities.

Abstract: We present a generalizable feed-forward Gaussian splatting framework for human 3D reconstruction and real-time animation that operates directly on multi-view RGB images and their associated SMPL-X poses. Unlike prior methods that rely on depth supervision, fixed input views, UV map, or repeated feed-forward inference for each target view or pose, our approach predicts, in a canonical pose, a set of 3D Gaussian primitives associated with each SMPL-X vertex. One Gaussian is regularized to remain close to the SMPL-X surface, providing a strong geometric prior and stable correspondence to the parametric body model, while an additional small set of unconstrained Gaussians per vertex allows the representation to capture geometric structures that deviate from the parametric surface, such as clothing and hair. In contrast to recent approaches such as HumanRAM, which require repeated network inference to synthesize novel poses, our method produces an animatable human representation from a single forward pass; by explicitly associating Gaussian primitives with SMPL-X vertices, the reconstructed model can be efficiently animated via linear blend skinning without further network evaluation. We evaluate our method on the THuman 2.1, AvatarReX and THuman 4.0 datasets, where it achieves reconstruction quality comparable to state-of-the-art methods while uniquely supporting real-time animation and interactive applications. Code and pre-trained models are available at https://github.com/Devdoot57/HumanGS .

Method: EditCrafter uses tiled inversion to preserve original image identity and introduces noise-damped manifold-constrained classifier-free guidance (NDCFG++) tailored for high-resolution editing from inverted latents, operating without fine-tuning.

Result: The method achieves impressive editing results across various resolutions without requiring fine-tuning or optimization, successfully handling resolutions significantly exceeding training sizes.

Conclusion: EditCrafter provides an effective solution for high-resolution image editing using pretrained diffusion models, overcoming resolution limitations through tiled inversion and specialized guidance techniques.

Abstract: We propose EditCrafter, a high-resolution image editing method that operates without tuning, leveraging pretrained text-to-image (T2I) diffusion models to process images at resolutions significantly exceeding those used during training. Leveraging the generative priors of large-scale T2I diffusion models enables the development of a wide array of novel generation and editing applications. Although numerous image editing methods have been proposed based on diffusion models and exhibit high-quality editing results, they are difficult to apply to images with arbitrary aspect ratios or higher resolutions since they only work at the training resolutions (512x512 or 1024x1024). Naively applying patch-wise editing fails with unrealistic object structures and repetition. To address these challenges, we introduce EditCrafter, a simple yet effective editing pipeline. EditCrafter operates by first performing tiled inversion, which preserves the original identity of the input high-resolution image. We further propose a noise-damped manifold-constrained classifier-free guidance (NDCFG++) that is tailored for high resolution image editing from the inverted latent. Our experiments show that the our EditCrafter can achieve impressive editing results across various resolutions without fine-tuning and optimization.

Method: Proposes Event-based DEI (E-DEI) algorithm that decomposes the task into event-based motion deblurring and low-light enhancement. Uses dual-path parallel feature propagation architecture with Dual-path Feature Alignment and Fusion (DFAF) module to align and fuse features from dual-exposure images using event data. Created real-world PIED dataset with paired low-/normal-light images and events.

Result: Experiments on multiple datasets demonstrate superiority of the method. The approach effectively reduces artifacts and improves image quality in low-light conditions with motion.

Conclusion: E-DEI effectively leverages event camera data to address limitations of traditional DEI, producing higher quality images in challenging low-light dynamic scenes through better motion handling and feature alignment.

Abstract: By combining complementary benefits of short- and long-exposure images, Dual-Exposure Imaging (DEI) enhances image quality in low-light scenarios. However, existing DEI approaches inevitably suffer from producing artifacts due to spatial displacement from scene motion and image feature discrepancies from different exposure times. To tackle this problem, we propose a novel Event-based DEI (E-DEI) algorithm, which reconstructs high-quality images from dual-exposure image pairs and events, leveraging high temporal resolution of event cameras to provide accurate inter-/intra-frame dynamic information. Specifically, we decompose this complex task into an integration of two sub-tasks, i.e., event-based motion deblurring and low-light image enhancement tasks, which guides us to design E-DEI network as a dual-path parallel feature propagation architecture. We propose a Dual-path Feature Alignment and Fusion (DFAF) module to effectively align and fuse features extracted from dual-exposure images with assistance of events. Furthermore, we build a real-world Dataset containing Paired low-/normal-light Images and Events (PIED). Experiments on multiple datasets show the superiority of our method. The code and dataset are available at github.

Method: Proposes FastSHADE with multi-stage architecture: 1) Asymmetric Frequency Denoising Block (AFDB) decouples spatial structure extraction from high-frequency noise suppression, 2) Spatially Gated Upsampler (SGU) optimizes high-resolution skip connection fusion, and 3) Noise Shifting Self-Augmentation strategy enhances data diversity without domain shifts.

Result: On MAI2021 benchmark, FastSHADE-M achieves real-time latency (<50 ms on mobile GPU) while preserving structural integrity. FastSHADE-XL establishes new state-of-the-art for overall image quality. The scalable model family demonstrates efficient speed-fidelity trade-offs.

Conclusion: FastSHADE successfully bridges the gap between theoretical network efficiency and practical deployment for real-world mobile ISP pipelines, offering a lightweight solution for real-time, high-fidelity image denoising on edge devices.

Abstract: Real-time image denoising is essential for modern mobile photography but remains challenging due to the strict latency and power constraints of edge devices. This paper presents FastSHADE (Fast Self-augmented Hierarchical Asymmetric Denoising), a lightweight U-Net-style network tailored for real-time, high-fidelity restoration on mobile GPUs. Our method features a multi-stage architecture incorporating a novel Asymmetric Frequency Denoising Block (AFDB) that decouples spatial structure extraction from high-frequency noise suppression to maximize efficiency, and a Spatially Gated Upsampler (SGU) that optimizes high-resolution skip connection fusion. To address generalization, we introduce an efficient Noise Shifting Self-Augmentation strategy that enhances data diversity without inducing domain shifts. Evaluations on the MAI2021 benchmark demonstrate that our scalable model family establishes a highly efficient speed-fidelity trade-off. Our base FastSHADE-M variant maintains real-time latency (<50 ms on a modern mobile GPU) while preserving structural integrity, and our scaled-up FastSHADE-XL establishes a new state-of-the-art for overall image quality. Ultimately, FastSHADE successfully bridges the gap between theoretical network efficiency and practical deployment for real-world mobile ISP pipelines.

Method: Introduces Multi-View CIR task, constructs FashionMV dataset (127K products, 472K images, 220K triplets), and proposes ProCIR framework with multimodal LLM using three mechanisms: two-stage dialogue, caption-based alignment, chain-of-thought guidance, plus optional supervised fine-tuning.

Result: Best 0.8B-parameter model outperforms all baselines including 10x larger general-purpose embedding models; systematic ablation shows alignment is most critical, two-stage dialogue is prerequisite, SFT and chain-of-thought are partially redundant.

Conclusion: Multi-View CIR addresses real-world product retrieval needs, ProCIR framework effectively leverages multimodal LLMs, and FashionMV enables product-level CIR research.

Abstract: Composed Image Retrieval (CIR) retrieves target images using a reference image paired with modification text. Despite rapid advances, all existing methods and datasets operate at the image level – a single reference image plus modification text in, a single target image out – while real e-commerce users reason about products shown from multiple viewpoints. We term this mismatch View Incompleteness and formally define a new Multi-View CIR task that generalizes standard CIR from image-level to product-level retrieval. To support this task, we construct FashionMV, the first large-scale multi-view fashion dataset for product-level CIR, comprising 127K products, 472K multi-view images, and over 220K CIR triplets, built through a fully automated pipeline leveraging large multimodal models. We further propose ProCIR (Product-level Composed Image Retrieval), a modeling framework built upon a multimodal large language model that employs three complementary mechanisms – two-stage dialogue, caption-based alignment, and chain-of-thought guidance – together with an optional supervised fine-tuning (SFT) stage that injects structured product knowledge prior to contrastive training. Systematic ablation across 16 configurations on three fashion benchmarks reveals that: (1) alignment is the single most critical mechanism; (2) the two-stage dialogue architecture is a prerequisite for effective alignment; and (3) SFT and chain-of-thought serve as partially redundant knowledge injection paths. Our best 0.8B-parameter model outperforms all baselines, including general-purpose embedding models 10x its size. The dataset, model, and code are publicly available at https://github.com/yuandaxia2001/FashionMV.

Method: Proposes Attention-Guided Visual Jailbreaking with two auxiliary objectives: (1) suppressing attention to alignment-relevant prefix tokens, and (2) anchoring generation on adversarial image features. This push-pull formulation reduces gradient conflict by directly manipulating attention patterns rather than overpowering safety mechanisms.

Result: Achieves 94.4% attack success rate on Qwen-VL (vs. 68.8% baseline) with 40% fewer iterations. Reduces gradient conflict by 45%. At tighter perturbation budgets (ε=8/255), maintains 59.0% ASR compared to 45.7% for standard methods. Mechanistic analysis reveals successful attacks suppress system-prompt attention by 80%, causing “safety blindness.”

Conclusion: Attention manipulation provides a more effective approach to visual jailbreaking than traditional gradient-based attacks. The method demonstrates that LVLMs can be made to generate harmful content not by overriding safety rules, but by preventing their retrieval through attention suppression.

Abstract: Large Vision-Language Models (LVLMs) rely on attention-based retrieval of safety instructions to maintain alignment during generation. Existing attacks typically optimize image perturbations to maximize harmful output likelihood, but suffer from slow convergence due to gradient conflict between adversarial objectives and the model’s safety-retrieval mechanism. We propose Attention-Guided Visual Jailbreaking, which circumvents rather than overpowers safety alignment by directly manipulating attention patterns. Our method introduces two simple auxiliary objectives: (1) suppressing attention to alignment-relevant prefix tokens and (2) anchoring generation on adversarial image features. This simple yet effective push-pull formulation reduces gradient conflict by 45% and achieves 94.4% attack success rate on Qwen-VL (vs. 68.8% baseline) with 40% fewer iterations. At tighter perturbation budgets ($ε=8/255$), we maintain 59.0% ASR compared to 45.7% for standard methods. Mechanistic analysis reveals a failure mode we term safety blindness: successful attacks suppress system-prompt attention by 80%, causing models to generate harmful content not by overriding safety rules, but by failing to retrieve them.

Method: Proposes Adversarial Concept-MIL (AC-MIL) with: 1) decomposition of global image quality into clinically defined radiological concepts using volume-level supervision, 2) unsupervised residual branch with adversarial erasure to prevent information leakage, and 3) spatial diversity constraint to penalize overlap between concept attention maps.

Result: AC-MIL successfully opens the MIL black box, providing highly localized spatial concept maps that allow clinicians to pinpoint specific causes of non-diagnostic scans while maintaining competitive ordinal grading performance against existing baselines.

Conclusion: AC-MIL achieves deep clinical transparency in weakly supervised quality assessment while maintaining strong performance, making it valuable for clinical applications where interpretability is crucial.

Abstract: High-quality Late Gadolinium Enhancement (LGE) MRI can be helpful for atrial fibrillation management, yet scan quality is frequently compromised by patient motion, irregular breathing, and suboptimal image acquisition timing. While Multiple Instance Learning (MIL) has emerged as a powerful tool for automated quality assessment under weak supervision, current state-of-the-art methods map localized visual evidence to a single, opaque global feature vector. This black box approach fails to provide actionable feedback on specific failure modes, obscuring whether a scan degrades due to motion blur, inadequate contrast, or a lack of anatomical context. In this paper, we propose Adversarial Concept-MIL (AC-MIL), a weakly supervised framework that decomposes global image quality into clinically defined radiological concepts using only volume-level supervision. To capture latent quality variations without entangling predefined concepts, our framework incorporates an unsupervised residual branch guided by an adversarial erasure mechanism to strictly prevent information leakage. Furthermore, we introduce a spatial diversity constraint that penalizes overlap between distinct concept attention maps, ensuring localized and interpretable feature extraction. Extensive experiments on a clinical dataset of atrial LGE-MRI volumes demonstrate that AC-MIL successfully opens the MIL black box, providing highly localized spatial concept maps that allow clinicians to pinpoint the specific causes of non-diagnostic scans. Crucially, our framework achieves this deep clinical transparency while maintaining highly competitive ordinal grading performance against existing baselines. Code to be released on acceptance.

Method: Four components: 1) multi-scale window attention with learned scale routing for spatially adaptive feature extraction, 2) class-specific fusion module separating small/large objects into attentive fusion pathways, 3) bird’s-eye-view enhancement through parallel dilated convolution and channel recalibration, 4) class-balanced objective weighting to reduce bias toward frequent categories.

Result: Experiments on V2X-Real benchmark across various cooperation settings show consistent improvements over baselines, with largest gains on trucks, clear improvements on pedestrians, and competitive results on cars.

Conclusion: Aligning feature extraction and fusion with class-dependent geometry and point density leads to more balanced cooperative perception in realistic V2X deployments.

Abstract: Cooperative perception allows connected vehicles and roadside infrastructure to share sensor observations, creating a fused scene representation beyond the capability of any single platform. However, most cooperative 3D object detectors use a uniform fusion strategy for all object classes, which limits their ability to handle the different geometric structures and point-sampling patterns of small and large objects. This problem is further reinforced by narrow evaluation protocols that often emphasize a single dominant class or only a few cooperation settings, leaving robust multi-class detection across diverse vehicle-to-everything interactions insufficiently explored. To address this gap, we propose a class-adaptive cooperative perception architecture for multi-class 3D object detection from LiDAR data. The model integrates four components: multi-scale window attention with learned scale routing for spatially adaptive feature extraction, a class-specific fusion module that separates small and large objects into attentive fusion pathways, bird’s-eye-view enhancement through parallel dilated convolution and channel recalibration for richer contextual representation, and class-balanced objective weighting to reduce bias toward frequent categories. Experiments on the V2X-Real benchmark cover vehicle-centric, infrastructure-centric, vehicle-to-vehicle, infrastructure-to-infrastructure, and vehicle-to-infrastructure settings under identical backbone and training configurations. The proposed method consistently improves mean detection performance over strong intermediate-fusion baselines, with the largest gains on trucks, clear improvements on pedestrians, and competitive results on cars. These results show that aligning feature extraction and fusion with class-dependent geometry and point density leads to more balanced cooperative perception in realistic vehicle-to-everything deployments.

Method: Uses CM-UNet as teacher architecture, reduces search space to two dominant width-related variables, profiles student models on NVIDIA Jetson Orin Nano, fits low-order surrogate models for mIoU, latency, and power, and employs knowledge distillation for efficient training.

Result: The framework enables fast selection of near-optimal architecture settings without exhaustive search, showing that selected variables affect task accuracy and hardware cost differently.

Conclusion: Reduced-space regression is a practical strategy for adapting hybrid CNN-Mamba segmentation models to future space-edge systems.

Abstract: As Earth-observation workloads move toward onboard and edge processing, remote-sensing segmentation models must operate under tight latency and energy constraints. We present SatReg, a regression-based hardware-aware tuning framework for lightweight remote-sensing segmentation on edge platforms. Using CM-UNet as the teacher architecture, we reduce the search space to two dominant width-related variables, profile a small set of student models on an NVIDIA Jetson Orin Nano, and fit low-order surrogate models for mIoU, latency, and power. Knowledge distillation is used to efficiently train the sampled students. The learned surrogates enable fast selection of near-optimal architecture settings for deployment targets without exhaustive search. Results show that the selected variables affect task accuracy and hardware cost differently, making reduced-space regression a practical strategy for adapting hybrid CNN-Mamba segmentation models to future space-edge systems.

Method: Proposes an anatomy-aware segmentation framework that integrates organ exclusion masks derived from TotalSegmentator into U-Net training. These masks encode explicit anatomical priors by identifying non-vascular organs and penalizing aneurysm predictions within these regions, guiding the model to focus on the aorta while suppressing anatomically implausible predictions.

Result: Despite training on a relatively small dataset, the anatomy-aware model achieves high accuracy, substantially reduces false positives, and improves boundary consistency compared to a standard U-Net baseline.

Conclusion: Incorporating anatomical knowledge through exclusion masks provides an efficient mechanism to enhance robustness and generalization, enabling reliable AAA segmentation even with limited training data.

Abstract: In CT angiography, the accurate segmentation of abdominal aortic aneurysms (AAAs) is difficult due to large anatomical variability, low-contrast vessel boundaries, and the close proximity of organs whose intensities resemble vascular structures, often leading to false positives. To address these challenges, we propose an anatomy-aware segmentation framework that integrates organ exclusion masks derived from TotalSegmentator into the training process. These masks encode explicit anatomical priors by identifying non-vascular organsand penalizing aneurysm predictions within these regions, thereby guiding the U-Net to focus on the aorta and its pathological dilation while suppressing anatomically implausible predictions. Despite being trained on a relatively small dataset, the anatomy-aware model achieves high accuracy, substantially reduces false positives, and improves boundary consistency compared to a standard U-Net baseline. The results demonstrate that incorporating anatomical knowledge through exclusion masks provides an efficient mechanism to enhance robustness and generalization, enabling reliable AAA segmentation even with limited training data.

Method: Organized a challenge using the OpenRR-5k dataset containing real-world images with various reflection scenarios and intensities, requiring participants to generate clean images without reflections.

Result: Challenge attracted over 100 registrations with 11 final participants; top-ranked methods advanced state-of-the-art reflection removal performance and received unanimous expert recognition.

Conclusion: The challenge successfully bridged the gap between academic research and real-world applications in reflection removal, with the OpenRR-5k dataset serving as a valuable benchmark for future research.

Abstract: In this paper, we review the NTIRE 2026 challenge on single-image reflection removal (SIRR) in the Wild. SIRR is a fundamental task in image restoration. Despite progress in academic research, most methods are tested on synthetic images or limited real-world images, creating a gap in real-world applications. In this challenge, we provide participants with the OpenRR-5k dataset, which requires them to process real-world images that cover a range of reflection scenarios and intensities, with the goal of generating clean images without reflections. The challenge attracted more than 100 registrations, with 11 of them participating in the final testing phase. The top-ranked methods advanced the state-of-the-art reflection removal performance and earned unanimous recognition from the five experts in the field. The proposed OpenRR-5k dataset is available at https://huggingface.co/datasets/qiuzhangTiTi/OpenRR-5k, and the homepage of this challenge is at https://github.com/caijie0620/OpenRR-5k. Due to page limitations, this article only presents partial content; the full report and detailed analyses are available in the extended arXiv version.

Method: SIMPLER uses H&E as a semantic anchor to learn reusable SIM representations through progressive alignment of SIM and H&E using adversarial, contrastive, and reconstruction-based objectives. This encourages SIM embeddings to internalize histological structure from H&E while preserving modality-specific characteristics.

Result: A single pretrained SIMPLER encoder transfers across multiple downstream tasks (multiple instance learning, morphological clustering), consistently outperforming SIM models trained from scratch or H&E-only pretraining. Joint alignment enhances SIM performance without degrading H&E representations.

Conclusion: SIMPLER enables effective cross-modality learning between traditional histology (H&E) and novel imaging modalities (SIM), demonstrating asymmetric enrichment where SIM benefits from H&E’s rich semantic annotations while maintaining its unique imaging characteristics.

Abstract: Structured Illumination Microscopy (SIM) enables rapid, high-contrast optical sectioning of fresh tissue without staining or physical sectioning, making it promising for intraoperative and point-of-care diagnostics. Recent foundation and large-scale self-supervised models in digital pathology have demonstrated strong performance on section-based modalities such as Hematoxylin and Eosin (H&E) and immunohistochemistry (IHC). However, these approaches are predominantly trained on thin tissue sections and do not explicitly address thick-tissue fluorescence modalities such as SIM. When transferred directly to SIM, performance is constrained by substantial modality shift, and naive fine-tuning often overfits to modality-specific appearance rather than underlying histological structure. We introduce SIMPLER (Structured Illumination Microscopy-Powered Learning for Embedding Representations), a cross-modality self-supervised pretraining framework that leverages H&E as a semantic anchor to learn reusable SIM representations. H&E encodes rich cellular and glandular structure aligned with established clinical annotations, while SIM provides rapid, nondestructive imaging of fresh tissue. During pretraining, SIM and H&E are progressively aligned through adversarial, contrastive, and reconstruction-based objectives, encouraging SIM embeddings to internalize histological structure from H&E without collapsing modality-specific characteristics. A single pretrained SIMPLER encoder transfers across multiple downstream tasks, including multiple instance learning and morphological clustering, consistently outperforming SIM models trained from scratch or H&E-only pretraining. Importantly, joint alignment enhances SIM performance without degrading H&E representations, demonstrating asymmetric enrichment rather

Method: Compared classical approach (handcrafted features + SVM) with deep learning approach (FMAE-IAT embeddings + Multi-Layer Perceptron) on two depression datasets: TPOT (mother-child interactions) and Pitt (patient-clinician interviews). Depression was operationalized differently in each context.

Result: Classical approach achieved higher accuracy in both contexts and was significantly fairer than the deep approach in patient-clinician context. Cross-context generalizability was modest for both approaches, suggesting depression detection may be context-specific.

Conclusion: Classical interpretable methods outperform deep learning approaches for depression detection from vision in accuracy and fairness, with limited generalizability across contexts, indicating depression manifestations may be context-dependent.

Abstract: The classical approach to detecting depression from vision emphasizes interpretable features, such as facial expression, and classifiers such as the Support Vector Machine (SVM). With the advent of deep learning, there has been a shift in feature representations and classification approaches. Contemporary approaches use learnt features from general-purpose vision models such as VGGNet to train machine learning models. Little is known about how classical and deep approaches compare in depression detection with respect to accuracy, fairness, and generalizability, especially across contexts. To address these questions, we compared classical and deep approaches to the detection of depression in the visual modality in two different contexts: Mother-child interactions in the TPOT database and patient-clinician interviews in the Pitt database. In the former, depression was operationalized as a history of depression per the DSM and current or recent clinically significant symptoms. In the latter, all participants met initial criteria for depression per DSM, and depression was reassessed over the course of treatment. The classical approach included handcrafted features with SVM classifiers. Learnt features were turn-level embeddings from the FMAE-IAT that were combined with Multi-Layer Perceptron classifiers. The classical approach achieved higher accuracy in both contexts. It was also significantly fairer than the deep approach in the patient-clinician context. Cross-context generalizability was modest at best for both approaches, which suggests that depression may be context-specific.

Method: Proposes Scale-ALiBi, a linear bias transformer attention mechanism with spatial encoding bias for relationships between image patches at different scales. Implemented using triple-contrastive and reconstructive architecture on aligned high/low-resolution optical and low-resolution SAR satellite imagery.

Result: Shows improvement on the GEO-Bench benchmark. The newly curated dataset of aligned multi-resolution, multi-modal satellite imagery is released publicly.

Conclusion: Scale-ALiBi effectively addresses the challenge of processing satellite imagery across multiple resolutions and modes, providing better representations for downstream tasks through improved spatial encoding in transformer attention.

Abstract: Vision foundation models have been shown to be effective at processing satellite imagery into representations fit for downstream tasks, however, creating models which operate over multiple spatial resolutions and modes is challenging. This paper presents Scale-ALiBi, a linear bias transformer attention mechanism with a spatial encoding bias to relationships between image patches at different ground sample distance scales. We provide an implementation of Scale-ALiBi over a dataset of aligned high- and low-resolution optical and low-resolution SAR satellite imagery data using a triple-contrastive and reconstructive architecture, show an improvement on the GEO-Bench benchmark, and release the newly curated dataset publicly.

Method: Proposes Multinex framework that decomposes images into illumination and color prior stacks from distinct analytic representations, then learns to fuse these representations into luminance and reflectance adjustments using lightweight neural operations within a Retinex residual formulation.

Result: Lightweight variants (45K and 0.7K parameters) significantly outperform corresponding lightweight SOTA models and reach comparable performance to heavy models in extensive benchmarks.

Conclusion: Multinex provides an effective ultra-lightweight solution for low-light image enhancement that balances performance and computational efficiency, making it suitable for edge deployment.

Abstract: Low-light image enhancement (LLIE) aims to restore natural visibility, color fidelity, and structural detail under severe illumination degradation. State-of-the-art (SOTA) LLIE techniques often rely on large models and multi-stage training, limiting practicality for edge deployment. Moreover, their dependence on a single color space introduces instability and visible exposure or color artifacts. To address these, we propose Multinex, an ultra-lightweight structured framework that integrates multiple fine-grained representations within a principled Retinex residual formulation. It decomposes an image into illumination and color prior stacks derived from distinct analytic representations, and learns to fuse these representations into luminance and reflectance adjustments required to correct exposure. By prioritizing enhancement over reconstruction and exploiting lightweight neural operations, Multinex significantly reduces computational cost, exemplified by its lightweight (45K parameters) and nano (0.7K parameters) versions. Extensive benchmarks show that all lightweight variants significantly outperform their corresponding lightweight SOTA models, and reach comparable performance to heavy models. Paper page available at https://albrateanu.github.io/multinex.

Method: Proposes vectorized reformulation solving all systems in single kernel call for 33x speedup. Analyzes two DiffusionNet variants with different tangent-plane transformations. Introduces balanced accuracy metric for partial-to-partial matching evaluation.

Result: Achieves up to 33x speedup while preserving exact solution. Documents distinct behaviors of DiffusionNet variants across benchmarks. Shows balanced accuracy provides useful complementary metric under varying overlap ratios.

Conclusion: Provides computational improvements and implementation clarifications for deep functional maps, packaged in open-source DeepShapeMatchingKit for standardized training, evaluation, and data pipelines.

Abstract: Deep functional maps, leveraging learned feature extractors and spectral correspondence solvers, are fundamental to non-rigid 3D shape matching. Based on an analysis of open-source implementations, we find that standard functional map implementations solve k independent linear systems serially, which is a computational bottleneck at higher spectral resolution. We thus propose a vectorized reformulation that solves all systems in a single kernel call, achieving up to a 33x speedup while preserving the exact solution. Furthermore, we identify and document a previously unnoticed implementation divergence in the spatial gradient features of the mainstay DiffusionNet: two variants that parameterize distinct families of tangent-plane transformations, and present experiments analyzing their respective behaviors across diverse benchmarks. We additionally revisit overlap prediction evaluation for partial-to-partial matching and show that balanced accuracy provides a useful complementary metric under varying overlap ratios. To share these advancements with the wider community, we present an open-source codebase, DeepShapeMatchingKit, that incorporates these improvements and standardizes training, evaluation, and data pipelines for common deep shape matching methods. The codebase is available at: https://github.com/xieyizheng/DeepShapeMatchingKit

Method: Uses LLMs to construct Graph of Events in Space and Time (GEST) - structured specification of actors, actions, objects, and temporal constraints. Employs hierarchical two-agent architecture: Director for story planning and Scene Builder for individual scenes through round-based state machine. Programmatic state backend enforces simulator constraints through validated tool calls, ensuring executable specifications by construction.

Result: Agentic narratives win 79% of text and 74% of video comparisons against procedural baselines via LLM jury. In seeded generation comparisons, engine-generated videos substantially outperform neural generators: 58% vs 25% and 20% on physical validity, and 3.75/5 vs 2.33 and 1.50 on semantic alignment.

Conclusion: The agentic system using structured GEST specifications executed in 3D game engines produces more semantically reliable and physically valid videos than neural generation approaches, demonstrating the value of separating narrative planning from constraint enforcement.

Abstract: Existing multi-agent video generation systems use LLM agents to orchestrate neural video generators, producing visually impressive but semantically unreliable outputs with no ground truth annotations. We present an agentic system that inverts this paradigm: instead of generating pixels, the LLM constructs a formal Graph of Events in Space and Time (GEST) – a structured specification of actors, actions, objects, and temporal constraints – which is then executed deterministically in a 3D game engine. A staged LLM refinement pipeline fails entirely at this task (0 of 50 attempts produce an executable specification), motivating a fundamentally different architecture based on a separation of concerns: the LLM handles narrative planning through natural language reasoning, while a programmatic state backend enforces all simulator constraints through validated tool calls, guaranteeing that every generated specification is executable by construction. The system uses a hierarchical two-agent architecture – a Director that plans the story and a Scene Builder that constructs individual scenes through a round-based state machine – with dedicated Relation Subagents that populate the logical and semantic edge types of the GEST formalism that procedural generation leaves empty, making this the first approach to exercise the full expressive capacity of the representation. We evaluate in two stages: autonomous generation against procedural baselines via a 3-model LLM jury, where agentic narratives win 79% of text and 74% of video comparisons; and seeded generation where the same text is given to our system, VEO 3.1, and WAN 2.2, with human annotations showing engine-generated videos substantially outperform neural generators on physical validity (58% vs 25% and 20%) and semantic alignment (3.75/5 vs 2.33 and 1.50).

Method: Created GTASA corpus with per-frame spatial relation graphs and event-level temporal mappings, developed GEST-Engine system based on Graphs of Events in Space and Time, compared with neural generators using human evaluation and video captioning models.

Result: GEST-Engine shows clear advantages over neural generators in physical validity and semantic alignment; self-supervised video encoders encode spatial structure better than VLM visual encoders across 11 spatiotemporal reasoning tasks.

Conclusion: GTASA provides valuable ground truth for evaluating video generation and understanding, revealing strengths of different video encoding approaches for spatiotemporal reasoning.

Abstract: Generating complex multi-actor scenario videos remains difficult even for state-of-the-art neural generators, while evaluating them is hard due to the lack of ground truth for physical plausibility and semantic faithfulness. We introduce GTASA, a corpus of multi-actor videos with per-frame spatial relation graphs and event-level temporal mappings, and the system that produced it based on Graphs of Events in Space and Time (GEST): GEST-Engine. We compare our method with both open and closed source neural generators and prove both qualitatively (human evaluation of physical validity and semantic alignment) and quantitatively (via training video captioning models) the clear advantages of our method. Probing four frozen video encoders across 11 spatiotemporal reasoning tasks enabled by GTASA’s exact 3D ground truth reveals that self-supervised encoders encode spatial structure significantly better than VLM visual encoders.

Method: Two components: 1) Frozen DINOv2 backbone with Low-Rank Adaptation (LoRA) to transfer self-supervised features to fisheye without task-specific pretraining, 2) Fisheye Rotary Position Embedding (FishRoPE) which reparameterizes attention in spherical coordinates so attention operates on angular separation rather than pixel distance. FishRoPE is architecture-agnostic and reduces to standard formulation under pinhole geometry.

Result: Achieves state-of-the-art results on WoodScape 2D detection (54.3 mAP) and SynWoodScapes BEV segmentation (65.1 mIoU).

Conclusion: FishRoPE provides a lightweight framework to adapt frozen vision foundation models to fisheye geometry through spherical coordinate attention, solving the geometric inconsistency problem without expensive retraining.

Abstract: Vision foundation models (VFMs) and Bird’s Eye View (BEV) representation have advanced visual perception substantially, yet their internal spatial representations assume the rectilinear geometry of pinhole cameras. Fisheye cameras, widely deployed on production autonomous vehicles for their surround-view coverage, exhibit severe radial distortion that renders these representations geometrically inconsistent. At the same time, the scarcity of large-scale fisheye annotations makes retraining foundation models from scratch impractical. We present \ours, a lightweight framework that adapts frozen VFMs to fisheye geometry through two components: a frozen DINOv2 backbone with Low-Rank Adaptation (LoRA) that transfers rich self-supervised features to fisheye without task-specific pretraining, and Fisheye Rotary Position Embedding (FishRoPE), which reparameterizes the attention mechanism in the spherical coordinates of the fisheye projection so that both self-attention and cross-attention operate on angular separation rather than pixel distance. FishRoPE is architecture-agnostic, introduces negligible computational overhead, and naturally reduces to the standard formulation under pinhole geometry. We evaluate \ours on WoodScape 2D detection (54.3 mAP) and SynWoodScapes BEV segmentation (65.1 mIoU), where it achieves state-of-the-art results on both benchmarks.

Method: Introduces DETAnt-HOI benchmark (temporally corrected from VidHOI and Action Genome) and HOI-DA framework that jointly performs subject-object localization, present HOI detection, and future anticipation by modeling future interactions as residual transitions from current pair states.

Result: Experiments show consistent improvements in both detection and anticipation, with larger gains at longer horizons. The joint learning approach proves most effective when anticipation is learned together with detection as a structural constraint on pair-level video representation learning.

Conclusion: Anticipation is most effective when learned jointly with detection as a structural constraint on pair-level video representation learning. The proposed framework and benchmark address limitations of existing approaches for more faithful multi-horizon evaluation.

Abstract: Video-based human-object interaction (HOI) understanding requires both detecting ongoing interactions and anticipating their future evolution. However, existing methods usually treat anticipation as a downstream forecasting task built on externally constructed human-object pairs, limiting joint reasoning between detection and prediction. In addition, sparse keyframe annotations in current benchmarks can temporally misalign nominal future labels from actual future dynamics, reducing the reliability of anticipation evaluation. To address these issues, we introduce DETAnt-HOI, a temporally corrected benchmark derived from VidHOI and Action Genome for more faithful multi-horizon evaluation, and HOI-DA, a pair-centric framework that jointly performs subject-object localization, present HOI detection, and future anticipation by modeling future interactions as residual transitions from current pair states. Experiments show consistent improvements in both detection and anticipation, with larger gains at longer horizons. Our results highlight that anticipation is most effective when learned jointly with detection as a structural constraint on pair-level video representation learning. Benchmark and code will be publicly available.

Method: Created a dataset with 112 trials from 13 participants (39.5 hours total) using synchronized five-view RGB-D capture during real assembly/disassembly of a commercial angle grinder. Includes multi-route execution via partial-order prerequisite graph, six-category anomaly taxonomy, NASA-TLX cognitive load measurement, and hierarchical annotation linking hand-specific atomic actions to procedural steps.

Result: The dataset reveals fundamental limitations invisible to single-task benchmarks, particularly under realistic deployment conditions involving incomplete observations, flexible execution paths, and corrective behavior. Systematic baselines demonstrate these challenges.

Conclusion: IMPACT provides the first real industrial assembly benchmark with comprehensive synchronized multi-view capture and detailed annotations, enabling research on deployment-oriented procedural understanding under realistic conditions.

Abstract: We introduce IMPACT, a synchronized five-view RGB-D dataset for deployment-oriented industrial procedural understanding, built around real assembly and disassembly of a commercial angle grinder with professional-grade tools. To our knowledge, IMPACT is the first real industrial assembly benchmark that jointly provides synchronized ego-exo RGB-D capture, decoupled bimanual annotation, compliance-aware state tracking, and explicit anomaly–recovery supervision within a single real industrial workflow. It comprises 112 trials from 13 participants totaling 39.5 hours, with multi-route execution governed by a partial-order prerequisite graph, a six-category anomaly taxonomy, and operator cognitive load measured via NASA-TLX. The annotation hierarchy links hand-specific atomic actions to coarse procedural steps, component assembly states, and per-hand compliance phases, with synchronized null spans across views to decouple perceptual limitations from algorithmic failure. Systematic baselines reveal fundamental limitations that remain invisible to single-task benchmarks, particularly under realistic deployment conditions that involve incomplete observations, flexible execution paths, and corrective behavior. The full dataset, annotations, and evaluation code are available at https://github.com/Kratos-Wen/IMPACT.

Method: Proposes Neural Stochastic Process (NSP) with Neural Process encoder conditioning on arbitrary sets of gauge observations and latent Neural SDE on 2D spatial representation. Trained under single variational objective with simulation-free cost. Also introduces QPEBench benchmark with 43,756 hourly samples over CONUS.

Result: NSP outperforms 13 baselines across all six evaluation metrics on QPEBench and surpasses JAXA’s operational gauge-calibrated product. Additional experiment on Kyushu, Japan confirms generalization to different region with independent data sources.

Conclusion: NSP effectively fuses satellite and gauge data for precipitation estimation by capturing temporal structure, demonstrating state-of-the-art performance and generalization capabilities across different regions.

Abstract: Accurate precipitation estimation is critical for flood forecasting, water resource management, and disaster preparedness. Satellite products provide global hourly coverage but contain systematic biases; ground-based gauges are accurate at point locations but too sparse for direct gridded correction. Existing methods fuse these sources by interpolating gauge observations onto the satellite grid, but treat each time step independently and therefore discard temporal structure in precipitation fields. We propose Neural Stochastic Process (NSP), a model that pairs a Neural Process encoder conditioning on arbitrary sets of gauge observations with a latent Neural SDE on a 2D spatial representation. NSP is trained under a single variational objective with simulation-free cost. We also introduce QPEBench, a benchmark of 43{,}756 hourly samples over the Contiguous United States (2021–2025) with four aligned data sources and six evaluation metrics. On QPEBench, NSP outperforms 13 baselines across all six metrics and surpasses JAXA’s operational gauge-calibrated product. An additional experiment on Kyushu, Japan confirms generalization to a different region with independent data sources.

Method: Uses 2D point tracking for long-range correspondences to enable recovery after occlusion, while incrementally reconstructing an online Truncated Signed Distance Function (TSDF) representation of tracked objects. Initialized only from sparse image points.

Result: Achieves performance comparable to state-of-the-art on severe-occlusion benchmarks while supporting multi-object tracking and recovery from complete occlusion, capabilities not supported by previous model-free approaches.

Conclusion: Point2Pose provides a robust model-free solution for 6D pose tracking that handles multiple objects and recovers from complete occlusions, demonstrated through a new multi-object tracking dataset.

Abstract: We present Point2Pose, a model-free method for causal 6D pose tracking of multiple rigid objects from monocular RGB-D video. Initialized only from sparse image points on the objects to be tracked, our approach tracks multiple unseen objects without requiring object CAD models or category priors. Point2Pose leverages a 2D point tracker to obtain long-range correspondences, enabling instant recovery after complete occlusion. Simultaneously, the system incrementally reconstructs an online Truncated Signed Distance Function (TSDF) representation of the tracked targets. Alongside the method, we introduce a new multi-object tracking dataset comprising both simulation and real-world sequences, with motion-capture ground truth for evaluation. Experiments show that Point2Pose achieves performance comparable to the state-of-the-art methods on a severe-occlusion benchmark, while additionally supporting multi-object tracking and recovery from complete occlusion, capabilities that are not supported by previous model-free tracking approaches.

Method: Introduces DiningBench, a hierarchical multi-view benchmark with 3,021 distinct dishes averaging 5.27 images per entry. Features fine-grained “hard” negatives from identical menus and verification-based nutritional data. Evaluates 29 state-of-the-art open-source and proprietary models across three cognitive levels.

Result: Current VLMs excel at general reasoning but struggle significantly with fine-grained visual discrimination and precise nutritional reasoning. The study identifies five primary failure modes and investigates the impact of multi-view inputs and Chain-of-Thought reasoning.

Conclusion: DiningBench serves as a challenging testbed to advance food-centric VLM research, highlighting the need for improved fine-grained discrimination and nutritional reasoning capabilities in vision-language models.

Abstract: Recent advancements in Vision-Language Models (VLMs) have revolutionized general visual understanding. However, their application in the food domain remains constrained by benchmarks that rely on coarse-grained categories, single-view imagery, and inaccurate metadata. To bridge this gap, we introduce DiningBench, a hierarchical, multi-view benchmark designed to evaluate VLMs across three levels of cognitive complexity: Fine-Grained Classification, Nutrition Estimation, and Visual Question Answering. Unlike previous datasets, DiningBench comprises 3,021 distinct dishes with an average of 5.27 images per entry, incorporating fine-grained “hard” negatives from identical menus and rigorous, verification-based nutritional data. We conduct an extensive evaluation of 29 state-of-the-art open-source and proprietary models. Our experiments reveal that while current VLMs excel at general reasoning, they struggle significantly with fine-grained visual discrimination and precise nutritional reasoning. Furthermore, we systematically investigate the impact of multi-view inputs and Chain-of-Thought reasoning, identifying five primary failure modes. DiningBench serves as a challenging testbed to drive the next generation of food-centric VLM research. All codes are released in https://github.com/meituan/DiningBench.

Method: Proposes Functional Structure Unit (FSU) for function-based decomposition of signs into minimal core blocks. Uses two-stage VLM post-training: Iterative Caption-FSU Distillation and FSU-GRPO with Tree Edit Distance rewards.

Result: SignReasoner achieves new state-of-the-art on TrafficSignEval benchmark with remarkable data efficiency and no architectural modification, significantly improving traffic sign understanding in various VLMs.

Conclusion: The FSU-based approach enables robust generalization to unseen sign compositions by learning underlying structural grammar, making VLMs effective traffic sign reasoners.

Abstract: Accurate semantic understanding of complex traffic signs-including those with intricate layouts, multi-lingual text, and composite symbols-is critical for autonomous driving safety. Current models, both specialized small ones and large Vision Language Models (VLMs), suffer from a significant bottleneck: a lack of compositional generalization, leading to failure when encountering novel sign configurations. To overcome this, we propose SignReasoner, a novel paradigm that transforms general VLMs into expert traffic sign reasoners. Our core innovation is Functional Structure Unit (FSU), which shifts from common instance-based modeling to flexible function-based decomposition. By breaking down complex signs into minimal, core functional blocks (e.g., Direction, Notice, Lane), our model learns the underlying structural grammar, enabling robust generalization to unseen compositions. We define this decomposition as the FSU-Reasoning task and introduce a two-stage VLM post-training pipeline to maximize performance: Iterative Caption-FSU Distillation that enhances the model’s accuracy in both FSU-reasoning and caption generation; FSU-GRPO that uses Tree Edit Distance (TED) to compute FSU differences as the rewards in GRPO algorithm, boosting reasoning abilities. Experiments on the newly proposed FSU-Reasoning benchmark, TrafficSignEval, show that SignReasoner achieves new SOTA with remarkable data efficiency and no architectural modification, significantly improving the traffic sign understanding in various VLMs.

Method: Proposes Discriminative Cue-Prompting with Prompt Dropout (DCP-PD), a plug-and-play framework that distills fine-grained cues from free-text reports and uses them to guide report generation while mitigating shortcut reliance via prompt dropout. Also introduces a hierarchical, location-aware question-set protocol to assess pathology-location grounding.

Result: Achieves state-of-the-art performance on CT-RATE, improving macro F1 from 0.501 to 0.603 (20% relative improvement), and substantially boosts out-of-distribution performance on Rad-ChestCT from F1 0.266 to 0.503 (89% relative improvement).

Conclusion: The proposed DCP-PD framework effectively improves fine-grained attribute alignment and spatial localization in radiology report generation, though fine-grained spatial localization remains challenging even for high-performing models on current benchmarks.

Abstract: Vision–language models (VLMs) for radiology report generation (RRG) can produce long-form chest CT reports from volumetric scans and show strong potential to improve radiology workflow efficiency and consistency. However, existing methods face two key limitations: (i) training supervision is often coarse, aligning a whole CT volume with a full free-text report without explicit alignment for fine-grained attributes or pathology locations; and (ii) evaluation is typically holistic (lexical overlap, entity matching, or LLM-as-a-judge scores) and not diagnostic for spatial grounding. We propose \emph{Discriminative Cue-Prompting with Prompt Dropout (DCP-PD)}, a plug-and-play framework that distills fine-grained cues from free-text reports and uses them to guide report generation while mitigating shortcut reliance via prompt dropout. DCP-PD achieves state-of-the-art performance on CT-RATE, improving macro F1 from $=0.501$ to $0.603$ (20% relative), and substantially boosts out-of-distribution performance on Rad-ChestCT from F1 $=0.266$ to $0.503$ (89% relative). Finally, we introduce a hierarchical, location-aware question-set protocol (presence $\rightarrow$ laterality $\rightarrow$ lobe) to directly assess pathology-location grounding, showing that fine-grained spatial localization remains challenging even for models that score highly on current benchmarks.

Method: Uses residual U-Net backbone with multi-scale recovery module and dual attention mechanisms, enhanced by Motion Perceptual Loss (MPL) that provides artifact-aware supervision by learning generalizable motion artifact representations.

Result: Outperformed state-of-the-art methods on clinical data, with ablation studies showing MPL is crucial for structural consistency and tissue contrast preservation. Radiologist evaluations confirmed superior image quality and diagnostic structure preservation.

Conclusion: PERCEPT-Net effectively suppresses motion artifacts in clinical MRI without compromising anatomical integrity through task-specific perceptual learning, improving clinical robustness and mitigating over-smoothing in medical image reconstruction.

Abstract: Purpose: Existing deep learning-based MRI artifact correction models exhibit poor clinical generalization due to inherent artifact-tissue confusion, failing to discriminate artifacts from anatomical structures. To resolve this, we introduce PERCEPT-Net, a framework leveraging dedicated perceptual supervision for structure-preserving artifact suppression. Method: PERCEPT-Net utilizes a residual U-Net backbone integrated with a multi-scale recovery module and dual attention mechanisms to preserve anatomical context and salient features. The core mechanism, Motion Perceptual Loss (MPL), provides artifact-aware supervision by learning generalizable motion artifact representations. This logic directly guides the network to suppress artifacts while maintaining anatomical fidelity. Training utilized a hybrid dataset of real and simulated sequences, followed by prospective validation via objective metrics and expert radiologist assessments. Result: PERCEPT-Net outperformed state-of-the-art methods on clinical data. Ablation analysis established a direct causal link between MPL and performance; its omission caused a significant deterioration in structural consistency (p < 0.001) and tissue contrast (p < 0.001). Radiologist evaluations corroborated these objective metrics, scoring PERCEPT-Net significantly higher in global image quality (median 3 vs. 2, p < 0.001) and verifying the preservation of critical diagnostic structures. Conclusion: By integrating task-specific, artifact-aware perceptual learning, PERCEPT-Net suppresses motion artifacts in clinical MRI without compromising anatomical integrity. This framework improves clinical robustness and provides a verifiable mechanism to mitigate over-smoothing and structural degradation in medical image reconstruction.

Method: Training-free model using compositional multi-agent system to emulate cognitive behaviors of poster designers (identify elements, organize layout, evaluate candidates). Integrates hybrid denoising strategy during diffusion process for harmonious region transitions.

Result: Superior performance over state-of-the-art methods confirmed by seven quantitative metrics and four user studies. Produces visually striking and aesthetically appealing posters.

Conclusion: ReContraster successfully creates attention-grabbing posters using regional contrast principles and multi-agent systems, with comprehensive evaluation showing its effectiveness.

Abstract: Effective poster design requires rapidly capturing attention and clearly conveying messages. Inspired by the ``contrast effects’’ principle, we propose ReContraster, the first training-free model to leverage regional contrast to make posters stand out. By emulating the cognitive behaviors of a poster designer, ReContraster introduces the compositional multi-agent system to identify elements, organize layout, and evaluate generated poster candidates. To further ensure harmonious transitions across region boundaries, ReContraster integrates the hybrid denoising strategy during the diffusion process. We additionally contribute a new benchmark dataset for comprehensive evaluation. Seven quantitative metrics and four user studies confirm its superiority over relevant state-of-the-art methods, producing visually striking and aesthetically appealing posters.

Method: Uses LoRA modules that learn lightweight task-specific low-rank matrices to perturb pretrained weights for fine-tuning downstream classification tasks, improving parameter efficiency compared to end-to-end fine-tuning.

Result: For gastrointestinal tract diseases, LoRA-based fine-tuning shows significantly better results than end-to-end fine-tuning while maintaining improved parameter efficiency.

Conclusion: LoRA provides an effective parameter-efficient fine-tuning approach for medical image analysis that addresses cross-source distribution shifts without requiring multiple full model copies.

Abstract: Despite recent advancements in the field of medical image analysis with the use of pretrained foundation models, the issue of distribution shifts between cross-source images largely remains adamant. To circumvent that issue, investigators generally train a separate model for each source. However, this method becomes expensive when we fully fine-tune pretrained large models for a single dataset, as we must store multiple copies of those models. Thus, in this work, we propose using a low-rank adaptation (LoRA) module for fine-tuning downstream classification tasks. LoRAs learn lightweight task-specific low-rank matrices that perturb pretrained weights to optimize those downstream tasks. For gastrointestinal tract diseases, they exhibit significantly better results than end-to-end finetuning with improved parameter efficiency. Code is available at: github.com/sanjay931/peft-gi-recognition.

Method: 1) Created AIM-Bench using dual-path affective modeling (Mikels taxonomy + Valence-Arousal-Dominance framework) via hierarchical human-in-the-loop workflow. 2) Developed composite evaluation suite with rule-based and model-based metrics. 3) Proposed AIM-40k dataset using inverse repainting strategy to create balanced instruction-tuning data.

Result: Current editing models show significant challenges including prevalent positivity bias. Fine-tuning on AIM-40k yields 9.15% relative improvement in overall performance on AIM-Bench.

Conclusion: The paper introduces the first benchmark for affective image manipulation and addresses data imbalance issues through a scalable data engine, demonstrating effectiveness through improved model performance.

Abstract: Affective Image Manipulation (AIM) aims to evoke specific emotions through targeted editing. Current image editing benchmarks primarily focus on object-level modifications in general scenarios, lacking the fine-grained granularity to capture affective dimensions. To bridge this gap, we introduce the first benchmark designed for AIM termed AIM-Bench. This benchmark is built upon a dual-path affective modeling scheme that integrates the Mikels emotion taxonomy with the Valence-Arousal-Dominance framework, enabling high-level semantic and fine-grained continuous manipulation. Through a hierarchical human-in-the-loop workflow, we finally curate 800 high-quality samples covering 8 emotional categories and 5 editing types. To effectively assess performance, we also design a composite evaluation suite combining rule-based and model-based metrics to holistically assess instruction consistency, aesthetics, and emotional expressiveness. Extensive evaluations reveal that current editing models face significant challenges, most notably a prevalent positivity bias, which stemming from inherent imbalances in training data distribution. To tackle this, we propose a scalable data engine utilizing an inverse repainting strategy to construct AIM-40k, a balanced instruction-tuning dataset comprising 40k samples. Concretely, we enhance raw affective images via generative redrawing to establish high-fidelity ground truths, and synthesize input images with divergent emotions and paired precise instructions. Fine-tuning a baseline model on AIM-40k yields a 9.15% relative improvement in overall performance, demonstrating the effectiveness of our AIM-40k. Our data and related code will be made open soon.

Method: Introduces CineBench benchmark with diverse user instructions and professional annotations. Proposes CineAgents multi-agent system using “design-and-compose” paradigm: script reverse-engineering to build hierarchical narrative memory, and iterative narrative planning to refine creative blueprint into final compiled script.

Result: CineAgents significantly outperforms existing methods, generating compilations with superior narrative coherence and logical coherence. The benchmark enables comprehensive evaluation of cinematic video compilation systems.

Conclusion: CineBench provides the first comprehensive benchmark for instruction-driven cinematic video compilation, while CineAgents demonstrates a novel multi-agent approach that effectively addresses contextual collapse and temporal fragmentation in video compilation tasks.

Abstract: The surging demand for adapting long-form cinematic content into short videos has motivated the need for versatile automatic video compilation systems. However, existing compilation methods are limited to predefined tasks, and the community lacks a comprehensive benchmark to evaluate the cinematic compilation. To address this, we introduce CineBench, the first benchmark for instruction-driven cinematic video compilation, featuring diverse user instructions and high-quality ground-truth compilations annotated by professional editors. To overcome contextual collapse and temporal fragmentation, we present CineAgents, a multi-agent system that reformulates cinematic video compilation into ``design-and-compose’’ paradigm. CineAgents performs script reverse-engineering to construct a hierarchical narrative memory to provide multi-level context and employs an iterative narrative planning process that refines a creative blueprint into a final compiled script. Extensive experiments demonstrate that CineAgents significantly outperforms existing methods, generating compilations with superior narrative coherence and logical coherence.

Method: Introduces a steganography-enabled attribution framework that embeds cryptographically signed identifiers into images at creation time and uses multimodal harmful content detection as a trigger for attribution verification. Evaluates five watermarking methods across spatial, frequency, and wavelet domains, and integrates a CLIP-based fusion model for multimodal harmful-content detection.

Result: Spread-spectrum watermarking, especially in the wavelet domain, provides strong robustness under blur distortions. The multimodal fusion detector achieves an AUC-ROC of 0.99, enabling reliable cross-modal attribution verification.

Conclusion: The components form an end-to-end forensic pipeline that enables reliable tracing of harmful deployments of AI-generated imagery, supporting accountability in modern synthetic media environments.

Abstract: The rapid growth of generative AI has introduced new challenges in content moderation and digital forensics. In particular, benign AI-generated images can be paired with harmful or misleading text, creating difficult-to-detect misuse. This contextual misuse undermines the traditional moderation framework and complicates attribution, as synthetic images typically lack persistent metadata or device signatures. We introduce a steganography enabled attribution framework that embeds cryptographically signed identifiers into images at creation time and uses multimodal harmful content detection as a trigger for attribution verification. Our system evaluates five watermarking methods across spatial, frequency, and wavelet domains. It also integrates a CLIP-based fusion model for multimodal harmful-content detection. Experiments demonstrate that spread-spectrum watermarking, especially in the wavelet domain, provides strong robustness under blur distortions, and our multimodal fusion detector achieves an AUC-ROC of 0.99, enabling reliable cross-modal attribution verification. These components form an end-to-end forensic pipeline that enables reliable tracing of harmful deployments of AI-generated imagery, supporting accountability in modern synthetic media environments. Our code is available at GitHub: https://github.com/bli1/steganography

Method: ExpertEdit learns an expert motion prior using a masked language modeling objective that infills masked motion spans with expert-level refinements. At inference, novice motion is masked at skill-critical moments and projected into the learned expert manifold, producing localized skill improvements without paired supervision or manual edit guidance.

Result: Across eight diverse techniques and three sports from Ego-Exo4D and Karate Kyokushin, ExpertEdit outperforms state-of-the-art supervised motion editing methods on multiple metrics of motion realism and expert quality.

Conclusion: ExpertEdit enables personalized skill feedback by automatically editing a person’s motion to reflect higher skill using only unpaired expert demonstrations, addressing limitations of existing supervised approaches that require expensive paired data and manual guidance.

Abstract: Visual feedback is critical for motor skill acquisition in sports and rehabilitation, and psychological studies show that observing near-perfect versions of one’s own performance accelerates learning more effectively than watching expert demonstrations alone. We propose to enable such personalized feedback by automatically editing a person’s motion to reflect higher skill. Existing motion editing approaches are poorly suited for this setting because they assume paired input-output data – rare and expensive to curate for skill-driven tasks – and explicit edit guidance at inference. We introduce ExpertEdit, a framework for skill-driven motion editing trained exclusively on unpaired expert video demonstrations. ExpertEdit learns an expert motion prior with a masked language modeling objective that infills masked motion spans with expert-level refinements. At inference, novice motion is masked at skill-critical moments and projected into the learned expert manifold, producing localized skill improvements without paired supervision or manual edit guidance. Across eight diverse techniques and three sports from Ego-Exo4D and Karate Kyokushin, ExpertEdit outperforms state-of-the-art supervised motion editing methods on multiple metrics of motion realism and expert quality. Project page: https://vision.cs.utexas.edu/projects/expert_edit/ .

Method: Proposes UDAPose with two key components: 1) A synthesis method using Direct-Current-based High-Pass Filter (DHF) and Low-light Characteristics Injection Module (LCIM) to inject high-frequency details from input low-light images, and 2) Dynamic Control of Attention (DCA) module that adaptively balances image cues with learned pose priors in Transformer architecture.

Result: Outperforms state-of-the-art methods with significant AP gains: 10.1 (56.4%) on ExLPose-test hard set and 7.4 (31.4%) in cross-dataset validation on EHPT-XC.

Conclusion: UDAPose effectively addresses low-light pose estimation challenges through realistic low-light image synthesis and adaptive fusion of visual cues with pose priors, demonstrating superior performance over existing methods.

Abstract: Low-visibility scenarios, such as low-light conditions, pose significant challenges to human pose estimation due to the scarcity of annotated low-light datasets and the loss of visual information under poor illumination. Recent domain adaptation techniques attempt to utilize well-lit labels by augmenting well-lit images to mimic low-light conditions. But handcrafted augmentations oversimplify noise patterns, while learning-based methods often fail to preserve high-frequency low-light characteristics, producing unrealistic images that lead pose models to generalize poorly to real low-light scenes. Moreover, recent pose estimators rely on image cues through image-to-keypoint cross-attention, but these cues become unreliable under low-light conditions. To address these issues, we propose Unsupervised Domain Adaptation for Pose Estimation (UDAPose), a novel framework that synthesizes low-light images and dynamically fuses visual cues with pose priors for improved pose estimation. Specifically, our synthesis method incorporates a Direct-Current-based High-Pass Filter (DHF) and a Low-light Characteristics Injection Module (LCIM) to inject high-frequency details from input low-light images, overcoming rigidity or the detail loss in existing approaches. Furthermore, we introduce a Dynamic Control of Attention (DCA) module that adaptively balances image cues with learned pose priors in the Transformer architecture. Experiments show that UDAPose outperforms state-of-the-art methods, with notable AP gains of 10.1 (56.4%) on the ExLPose-test hard set (LL-H) and 7.4 (31.4%) in cross-dataset validation on EHPT-XC. Code: https://github.com/Vision-and-Multimodal-Intelligence-Lab/UDAPose

Method: Proposes visual replay module using causal self-attention to estimate token saliency with spatially-coherent constraints, and routing depth scaling that adaptively allocates additional reasoning steps to complex tokens. Uses curriculum strategy to progressively internalize explicit CoT into latent representations.

Result: Achieves state-of-the-art performance across diverse benchmarks while delivering substantial inference speedups over explicit CoT baselines.

Conclusion: The framework successfully addresses visual under-optimization and complex token instability, enabling efficient multimodal latent reasoning with improved performance and reduced latency.

Abstract: Multimodal latent reasoning has emerged as a promising paradigm that replaces explicit Chain-of-Thought (CoT) decoding with implicit feature propagation, simultaneously enhancing representation informativeness and reducing inference latency. By analyzing token-level gradient dynamics during latent training, we reveal two critical observations: (1) visual tokens exhibit significantly higher and more volatile gradient norms than their textual counterparts due to inherent language bias, resulting in systematic visual under-optimization; and (2) semantically simple tokens converge rapidly, whereas complex tokens exhibit persistent gradient instability constrained by fixed architectural depths. To address these limitations, we propose a visual replay module and routing depth scaling to collaboratively enhance visual perception and refine complicated latents for deeper contextual reasoning. The former module leverages causal self-attention to estimate token saliency, reinforcing fine-grained grounding through spatially-coherent constraints. Complementarily, the latter mechanism adaptively allocates additional reasoning steps to complex tokens, enabling deeper contextual refinement. Guided by a curriculum strategy that progressively internalizes explicit CoT into compact latent representations, our framework achieves state-of-the-art performance across diverse benchmarks while delivering substantial inference speedups over explicit CoT baselines.

Method: Leverages imperfect scene reconstructions as geometric proxies, then uses certainty-aware free-view sampling to identify novel viewpoints that are semantically meaningful and minimally affected by reconstruction errors.

Result: Achieves 2.7 dB PSNR gain on out-of-distribution benchmarks and enhances per-scene 3D Gaussian Splatting optimization across multiple datasets.

Conclusion: Provides a practical data generation engine to overcome fundamental bottlenecks in 3D vision by transforming limited real-world sequences into scalable training data.

Abstract: The development of generalizable Novel View Synthesis (NVS) models is critically limited by the scarcity of large-scale training data featuring diverse and precise camera trajectories. While real-world captures are photorealistic, they are typically sparse and discrete. Conversely, synthetic data scales but suffers from a domain gap and often lacks realistic semantics. We introduce FreeScale, a novel framework that leverages the power of scene reconstruction to transform limited real-world image sequences into a scalable source of high-quality training data. Our key insight is that an imperfect reconstructed scene serves as a rich geometric proxy, but naively sampling from it amplifies artifacts. To this end, we propose a certainty-aware free-view sampling strategy identifying novel viewpoints that are both semantically meaningful and minimally affected by reconstruction errors. We demonstrate FreeScale’s effectiveness by scaling up the training of feedforward NVS models, achieving a notable gain of 2.7 dB in PSNR on challenging out-of-distribution benchmarks. Furthermore, we show that the generated data can actively enhance per-scene 3D Gaussian Splatting optimization, leading to consistent improvements across multiple datasets. Our work provides a practical and powerful data generation engine to overcome a fundamental bottleneck in 3D vision. Project page: https://mvp-ai-lab.github.io/FreeScale.

Method: Controlled comparison of visual representations using identical temporal model (MS-TCN++) and training/evaluation settings. Compared supervised encoders (ResNet-50, I3D) against self-supervised foundation models (DINOv3, V-JEPA2) using cached-feature pipeline that separates visual encoding from temporal learning. Also examined domain transfer using unlabeled videos and lightweight adaptation.

Result: Foundation-model features significantly improve segmentation performance. DINOv3 ViT-7B achieved best overall results: 83.4% accuracy and 87.0 edit score on SICS-155 dataset (19 phases). Analysis shows when domain transfer helps or hurts performance.

Conclusion: Modern vision foundation models demonstrate strong transferability to surgical workflow understanding and provide practical guidance for low-label medical video settings. Foundation models outperform traditional supervised approaches in data-scarce surgical phase segmentation.

Abstract: Surgical phase segmentation is central to computer-assisted surgery, yet robust models remain difficult to develop when labeled surgical videos are scarce. We study data-efficient phase segmentation for manual small-incision cataract surgery (SICS) through a controlled comparison of visual representations. To isolate representation quality, we pair each visual encoder with the same temporal model (MS-TCN++) under identical training and evaluation settings on SICS-155 (19 phases). We compare supervised encoders (ResNet-50, I3D) against large self-supervised foundation models (DINOv3, V-JEPA2), and use a cached-feature pipeline that decouples expensive visual encoding from lightweight temporal learning. Foundation-model features improve segmentation performance in this setup, with DINOv3 ViT-7B achieving the best overall results (83.4% accuracy, 87.0 edit score). We further examine cataract-domain transfer using unlabeled videos and lightweight adaptation, and analyze when it helps or hurts. Overall, the study indicates strong transferability of modern vision foundation models to surgical workflow understanding and provides practical guidance for low-label medical video settings. The project website is available at: https://sl2005.github.io/DataEfficient-sics-phase-seg/

Method: Proposes FGML-DG framework with three key components: 1) ‘Concept understanding’ principle to simplify complex features into style information statistics for precise style feature alignment, 2) Meta-style memory and recall method (MetaStyle) to emulate human memory system for past knowledge utilization, 3) Feedback-Driven Re-Training strategy (FDRT) for dynamic learning focus adjustment based on prediction errors.

Result: The method outperforms existing domain generalization approaches on two challenging medical image domain generalization tasks.

Conclusion: The cognitive-inspired meta-learning paradigm effectively addresses domain generalization challenges in medical image segmentation by mimicking human cognitive learning processes, enabling better knowledge transfer and adaptation to unseen domains.

Abstract: In medical image segmentation across multiple modalities (e.g., MRI, CT, etc.) and heterogeneous data sources (e.g., different hospitals and devices), Domain Generalization (DG) remains a critical challenge in AI-driven healthcare. This challenge primarily arises from domain shifts, imaging variations, and patient diversity, which often lead to degraded model performance in unseen domains. To address these limitations, we identify key issues in existing methods, including insufficient simplification of complex style features, inadequate reuse of domain knowledge, and a lack of feedback-driven optimization. To tackle these problems, inspired by Feynman’s learning techniques in educational psychology, this paper introduces a cognitive science-inspired meta-learning paradigm for medical image domain generalization segmentation. We propose, for the first time, a cognitive-inspired Feynman-Guided Meta-Learning framework for medical image domain generalization segmentation (FGML-DG), which mimics human cognitive learning processes to enhance model learning and knowledge transfer. Specifically, we first leverage the ‘concept understanding’ principle from Feynman’s learning method to simplify complex features across domains into style information statistics, achieving precise style feature alignment. Second, we design a meta-style memory and recall method (MetaStyle) to emulate the human memory system’s utilization of past knowledge. Finally, we incorporate a Feedback-Driven Re-Training strategy (FDRT), which mimics Feynman’s emphasis on targeted relearning, enabling the model to dynamically adjust learning focus based on prediction errors. Experimental results demonstrate that our method outperforms other existing domain generalization approaches on two challenging medical image domain generalization tasks.

Method: STORM uses an end-to-end MLLM architecture that jointly performs grounding and tracking. It employs task-composition learning (TCL) that decomposes RMOT into image grounding and object tracking sub-tasks for better data efficiency. The authors also create STORM-Bench, a new RMOT dataset with accurate trajectories and diverse, unambiguous referring expressions.

Result: STORM achieves state-of-the-art performance on image grounding, single-object tracking, and RMOT benchmarks. It demonstrates strong generalization and robust spatial-temporal grounding in complex real-world scenarios.

Conclusion: STORM provides a unified end-to-end solution for referring multi-object tracking that outperforms existing approaches by enabling coherent reasoning across appearance, motion, and language modalities through task-composition learning and improved dataset quality.

Abstract: Referring multi-object tracking (RMOT) is a task of associating all the objects in a video that semantically match with given textual queries or referring expressions. Existing RMOT approaches decompose object grounding and tracking into separated modules and exhibit limited performance due to the scarcity of training videos, ambiguous annotations, and restricted domains. In this work, we introduce STORM, an end-to-end MLLM that jointly performs grounding and tracking within a unified framework, eliminating external detectors and enabling coherent reasoning over appearance, motion, and language. To improve data efficiency, we propose a task-composition learning (TCL) strategy that decomposes RMOT into image grounding and object tracking, allowing STORM to leverage data-rich sub-tasks and learn structured spatial–temporal reasoning. We further construct STORM-Bench, a new RMOT dataset with accurate trajectories and diverse, unambiguous referring expressions generated through a bottom-up annotation pipeline. Extensive experiments show that STORM achieves state-of-the-art performance on image grounding, single-object tracking, and RMOT benchmarks, demonstrating strong generalization and robust spatial–temporal grounding in complex real-world scenarios. STORM-Bench is released at https://github.com/amazon-science/storm-referring-multi-object-grounding.

Method: Created BareBones benchmark with pixel-level silhouettes of geometrically distinct classes across six datasets (including novel WTP-Bench), establishing noise-free geometric taxonomy. Evaluated 26 state-of-the-art VLMs under RGB deprivation conditions.

Result: Revealed consistent, severe performance collapse under RGB deprivation (Texture Bias Cliff), showing VLMs have universal structural blindspots and lack genuine geometric grounding.

Conclusion: BareBones establishes rigorous yardstick for assessing geometric comprehension in VLMs, documenting fundamental limitations in current architectures’ ability to understand pure geometric structure.

Abstract: While Vision-Language Models (VLMs) demonstrate remarkable zero-shot recognition capabilities across a diverse spectrum of multimodal tasks, it yet remains an open question whether these architectures genuinely comprehend geometric structure or merely exploit RGB textures and contextual priors as statistical shortcuts. Existing evaluations fail to isolate this mechanism, conflating semantic reasoning with texture mapping and relying on imprecise annotations that inadvertently leak environmental cues. To address this gap, we introduce \textbf{BareBones}, a zero-shot benchmark designed to stress-test pure geometric shape comprehension. We curate pixel-level silhouettes of geometrically distinct classes across six datasets: five established segmentation sources (ImageNet-S, DIS5K, ThinObject5K, PASCAL VOC, CUB-200) and our novel flagship collection, WTP-Bench, establishing a noise-free geometric taxonomy. WTP-Bench is an extreme, fine-grained visual puzzle that forces models to identify inter-class geometric concepts from boundary contours alone. Our evaluation of 26 state-of-the-art proprietary and open-weight VLMs (\eg, GPT-4.1, Gemini, Claude Sonnet 4.5, LLaVA) reveals a consistent, severe performance collapse under RGB deprivation, a phenomenon we term the \textit{Texture Bias Cliff}. By documenting universal structural blindspots, BareBones establishes a rigorous yardstick for genuine geometric grounding.

Method: Challenge-based evaluation using weighted image quality assessment (IQA) score and AdaFace model as identity checker, with 96 registrants and 10 teams submitting valid models.

Result: 9 teams achieved valid scores in final ranking, advancing performance of real-world face restoration and providing overview of latest trends in the field.

Conclusion: The collaborative challenge successfully advanced face restoration techniques, highlighting current trends and pushing boundaries in perceptual quality and identity preservation.

Abstract: This paper provides a review of the NTIRE 2026 challenge on real-world face restoration, highlighting the proposed solutions and the resulting outcomes. The challenge focuses on generating natural and realistic outputs while maintaining identity consistency. Its goal is to advance state-of-the-art solutions for perceptual quality and realism, without imposing constraints on computational resources or training data. Performance is evaluated using a weighted image quality assessment (IQA) score and employs the AdaFace model as an identity checker. The competition attracted 96 registrants, with 10 teams submitting valid models; ultimately, 9 teams achieved valid scores in the final ranking. This collaborative effort advances the performance of real-world face restoration while offering an in-depth overview of the latest trends in the field.

Method: Proposes Structured Semantic Mapping (SSM) framework with: 1) shared visual backbone for unified facial representations, 2) Textual Semantic Prototype (TSP) module using fixed textual descriptions with learnable context prompts for semantic mediation, and 3) Dynamic Prior Mapping (DPM) module incorporating Facial Action Coding System knowledge and learning data-driven association matrix for bidirectional knowledge transfer.

Result: SSM achieves state-of-the-art performance on both AU detection and FE recognition benchmarks simultaneously. Demonstrates that holistic expression semantics can enhance fine-grained AU learning even across heterogeneous datasets.

Conclusion: The SSM framework effectively addresses bidirectional learning between AUs and FEs under heterogeneous data conditions, enabling mutual enhancement between fine-grained muscular activations and coarse-grained affective states through structured semantic mapping.

Abstract: Facial action unit (AU) detection and facial expression (FE) recognition can be jointly viewed as affective facial behavior tasks, representing fine-grained muscular activations and coarse-grained holistic affective states, respectively. Despite their inherent semantic correlation, existing studies predominantly focus on knowledge transfer from AUs to FEs, while bidirectional learning remains insufficiently explored. In practice, this challenge is further compounded by heterogeneous data conditions, where AU and FE datasets differ in annotation paradigms (frame-level vs.\ clip-level), label granularity, and data availability and diversity, hindering effective joint learning. To address these issues, we propose a Structured Semantic Mapping (SSM) framework for bidirectional AU–FE learning under different data domains and heterogeneous supervision. SSM consists of three key components: (1) a shared visual backbone that learns unified facial representations from dynamic AU and FE videos; (2) semantic mediation via a Textual Semantic Prototype (TSP) module, which constructs structured semantic prototypes from fixed textual descriptions augmented with learnable context prompts, serving as supervision signals and cross-task alignment anchors in a shared semantic space; and (3) a Dynamic Prior Mapping (DPM) module that incorporates prior knowledge derived from the Facial Action Coding System and learns a data-driven association matrix in a high-level feature space, enabling explicit and bidirectional knowledge transfer. Extensive experiments on popular AU detection and FE recognition benchmarks show that SSM achieves state-of-the-art performance on both tasks simultaneously, and demonstrate that holistic expression semantics can in turn enhance fine-grained AU learning even across heterogeneous datasets.

Method: Proposes RDVQ framework with: 1) Differentiable relaxation of codebook distribution to enable end-to-end RD optimization, 2) Autoregressive entropy model for accurate entropy modeling and test-time rate control, 3) Lightweight architecture for efficient compression.

Result: Achieves strong performance at extremely low bitrates with competitive/superior perceptual quality using significantly fewer parameters. Reduces bitrate by up to 75.71% on DISTS and 37.63% on LPIPS on DIV2K-val compared to RDEIC.

Conclusion: RDVQ enables principled rate-distortion optimization for VQ-based compression and introduces an entropy-constrained formulation that bridges image tokenization and compression, offering a unified approach for low-bitrate image compression.

Abstract: The rapid growth of visual data under stringent storage and bandwidth constraints makes extremely low-bitrate image compression increasingly important. While Vector Quantization (VQ) offers strong structural fidelity, existing methods lack a principled mechanism for joint rate-distortion (RD) optimization due to the disconnect between representation learning and entropy modeling. We propose RDVQ, a unified framework that enables end-to-end RD optimization for VQ-based compression via a differentiable relaxation of the codebook distribution, allowing the entropy loss to directly shape the latent prior. We further develop an autoregressive entropy model that supports accurate entropy modeling and test-time rate control. Extensive experiments demonstrate that RDVQ achieves strong performance at extremely low bitrates with a lightweight architecture, attaining competitive or superior perceptual quality with significantly fewer parameters. Compared with RDEIC, RDVQ reduces bitrate by up to 75.71% on DISTS and 37.63% on LPIPS on DIV2K-val. Beyond empirical gains, RDVQ introduces an entropy-constrained formulation of VQ, highlighting the potential for a more unified view of image tokenization and compression. The code will be available at https://github.com/CVL-UESTC/RDVQ.

Method: Two-track challenge: subjective (user study evaluation) and objective tracks using KwaiVIR benchmark with 200 synthetic and 48 wild training videos, plus validation and testing sets.

Result: 95 teams registered, 12 submitted valid solutions achieving strong performance on KwaiVIR benchmark, showing progress in S-UGC video restoration.

Conclusion: The challenge successfully established a benchmark and demonstrated encouraging progress in generative-model-based short-form UGC video restoration in the wild.

Abstract: This paper presents an overview of the NTIRE 2026 Challenge on Short-form UGC Video Restoration in the Wild with Generative Models. This challenge utilizes a new short-form UGC (S-UGC) video restoration benchmark, termed KwaiVIR, which is contributed by USTC and Kuaishou Technology. It contains both synthetically distorted videos and real-world short-form UGC videos in the wild. For this edition, the released data include 200 synthetic training videos, 48 wild training videos, 11 validation videos, and 20 testing videos. The primary goal of this challenge is to establish a strong and practical benchmark for restoring short-form UGC videos under complex real-world degradations, especially in the emerging paradigm of generative-model-based S-UGC video restoration. This challenge has two tracks: (i) the primary track is a subjective track, where the evaluation is based on a user study; (ii) the second track is an objective track. These two tracks enable a comprehensive assessment of restoration quality. In total, 95 teams have registered for this competition. And 12 teams submitted valid final solutions and fact sheets for the testing phase. The submitted methods achieved strong performance on the KwaiVIR benchmark, demonstrating encouraging progress in short-form UGC video restoration in the wild.

Method: STGDNet uses a recurrent multi-branch architecture that iteratively encodes and fuses spatial difference (SD) and temporal difference (TD) sequences to restore structure and color details from blurry RGB inputs. The method leverages complementary modalities from the CVS sensor.

Result: Outperforms current RGB-only and event-based deblurring approaches on both synthetic CVS datasets and real-world evaluations. Demonstrates strong generalization across over 100 extreme real-world scenarios.

Conclusion: The complementary vision sensor with synchronized SD and TD data provides effective guidance for motion deblurring. STGDNet successfully leverages these modalities to achieve state-of-the-art performance in extreme dynamic scenes.

Abstract: Motion blur arises when rapid scene changes occur during the exposure period, collapsing rich intra-exposure motion into a single RGB frame. Without explicit structural or temporal cues, RGB-only deblurring is highly ill-posed and often fails under extreme motion. Inspired by the human visual system, brain-inspired vision sensors introduce temporally dense information to alleviate this problem. However, event cameras still suffer from event rate saturation under rapid motion, while the event modality entangles edge features and motion cues, which limits their effectiveness. As a recent breakthrough, the complementary vision sensor (CVS), Tianmouc, captures synchronized RGB frames together with high-frame-rate, multi-bit spatial difference (SD, encoding structural edges) and temporal difference (TD, encoding motion cues) data within a single RGB exposure, offering a promising solution for RGB deblurring under extreme dynamic scenes. To fully leverage these complementary modalities, we propose Spatio-Temporal Difference Guided Deblur Net (STGDNet), which adopts a recurrent multi-branch architecture that iteratively encodes and fuses SD and TD sequences to restore structure and color details lost in blurry RGB inputs. Our method outperforms current RGB or event-based approaches in both synthetic CVS dataset and real-world evaluations. Moreover, STGDNet exhibits strong generalization capability across over 100 extreme real-world scenarios. Project page: https://tmcDeblur.github.io/

Method: Three complementary components: 1) Dual-masking strategy (encoder and decoder tokens with geometry-focused decoder masks) to strengthen geometry induction; 2) Coarse-to-fine Gaussian splatting to reduce appearance-semantics inconsistencies; 3) Pose-conditioned recalibration mechanism that interrelates multiple head outputs by re-projecting 3D point and semantic maps into image plane using estimated camera parameters.

Result: Unified 3D representations robust to unposed, sparse-view inputs that generalize across diverse tasks, laying a perceptual foundation for spatial intelligence.

Conclusion: UniSplat addresses limitations in 3D representation learning by integrating geometry, appearance, and semantics in a feed-forward framework, enabling robust spatial understanding from challenging unposed multi-view inputs.

Abstract: Robust 3D representation learning forms the perceptual foundation of spatial intelligence, enabling downstream tasks in scene understanding and embodied AI. However, learning such representations directly from unposed multi-view images remains challenging. Recent self-supervised methods attempt to unify geometry, appearance, and semantics in a feed-forward manner, but they often suffer from weak geometry induction, limited appearance detail, and inconsistencies between geometry and semantics. We introduce UniSplat, a feed-forward framework designed to address these limitations through three complementary components. First, we propose a dual-masking strategy that strengthens geometry induction in the encoder. By masking both encoder and decoder tokens, and targeting decoder masks toward geometry-rich regions, the model is forced to infer structural information from incomplete visual cues, yielding geometry-aware representations even under unposed inputs. Second, we develop a coarse-to-fine Gaussian splatting strategy that reduces appearance-semantics inconsistencies by progressively refining the radiance field. Finally, to enforce geometric-semantic consistency, we introduce a pose-conditioned recalibration mechanism that interrelates the outputs of multiple heads by re-projecting predicted 3D point and semantic maps into the image plane using estimated camera parameters, and aligning them with corresponding RGB and semantic predictions to ensure cross-task consistency, thereby resolving geometry-semantic mismatches. Together, these components yield unified 3D representations that are robust to unposed, sparse-view inputs and generalize across diverse tasks, laying a perceptual foundation for spatial intelligence.

Method: Proposes HoPA (High-order Proxy Alignment), a unified method that captures high-order cross-modal alignments via a compact proxy. It abstracts omnimodal alignment with a shared similarity structure, avoiding combinatorial complexity of pairwise modality modeling. The method is compatible with trajectory matching and enables scalable joint distillation across heterogeneous modalities.

Result: Extensive experiments on various benchmarks demonstrate superior compression-performance trade-offs compared to existing competitors. Theoretical analysis from spectral perspective validates the method’s rationality against bimodal dataset distillation techniques.

Conclusion: HoPA successfully addresses the challenges of omnimodal dataset distillation by capturing high-order cross-modal alignments, enabling efficient compression of multi-modal datasets while preserving training performance across heterogeneous modalities.

Abstract: Dataset distillation compresses large-scale datasets into compact synthetic sets while preserving training performance, but existing methods are largely restricted to single-modal or bimodal settings. Extending dataset distillation to scenarios involving more than two modalities, i.e., Omnimodal Dataset Distillation, remains underexplored and challenging due to increased heterogeneity and complex cross-modal interactions. In this work, we identify the key determinant that bounds the endpoint discrepancy in the omnimodal setting, which is exacerbated with an increasing number of modalities. To this end, we propose HoPA, a unified method that captures high-order cross-modal alignments via a compact proxy, which is compatible with trajectory matching as well. By abstracting omnimodal alignment with a shared similarity structure, our method avoids the combinatorial complexity of pairwise modality modeling and enables scalable joint distillation across heterogeneous modalities. Theoretical analysis from the spectral perspective reveals the rationality of our proposed method against bimodal dataset distillation techniques. Extensive experiments on various benchmarks demonstrate that the proposed method achieves superior compression-performance trade-offs compared to existing competitors. The source code will be publicly released.

Method: Uses a “restore-and-refine” paradigm: radial exploration renders imperfect panoramic videos from coarse 3DGS initialization, restored by panoramic video-to-video diffusion model, enhanced via video super-resolution, then used as pseudo-ground truths to update global 3D Gaussian field. Also creates PanoV2V-15K dataset for training.

Result: Produces photorealistic and globally consistent 3D scenes, significantly improves long-range camera exploration compared with existing baselines.

Conclusion: Rein3D effectively addresses the challenge of global consistency in 3D indoor scene reconstruction by combining explicit 3D representation with diffusion-based video restoration.

Abstract: The growing demand for Embodied AI and VR applications has highlighted the need for synthesizing high-quality 3D indoor scenes from sparse inputs. However, existing approaches struggle to infer massive amounts of missing geometry in large unseen areas while maintaining global consistency, often producing locally plausible but globally inconsistent reconstructions. We present Rein3D, a framework that reconstructs full 360-degree indoor environments by coupling explicit 3D Gaussian Splatting (3DGS) with temporally coherent priors from video diffusion models. Our approach follows a “restore-and-refine” paradigm: we employ a radial exploration strategy to render imperfect panoramic videos along trajectories starting from the origin, effectively uncovering occluded regions from a coarse 3DGS initialization. These sequences are restored by a panoramic video-to-video diffusion model and further enhanced via video super-resolution to synthesize high-fidelity geometry and textures. Finally, these refined videos serve as pseudo-ground truths to update the global 3D Gaussian field. To support this task, we construct PanoV2V-15K, a dataset of over 15K paired clean and degraded panoramic videos for diffusion-based scene restoration. Experiments demonstrate that Rein3D produces photorealistic and globally consistent 3D scenes and significantly improves long-range camera exploration compared with existing baselines.

Method: Uses recurrent transformer architecture with sequence parallelism to train on 1024-frame sequences, introduces new Re-Detection Average Jaccard metric, and adds geometric augmentations like periodic roll to simulate point re-entries.

Result: Achieves state-of-the-art performance on multiple benchmarks, tracking points in much longer sequences while maintaining low memory and compute footprint.

Conclusion: Recurrent transformers can be substantially improved for point tracking, especially for long sequences and re-detection scenarios, advancing AR/XR and robotics applications.

Abstract: Tracking-Any-Point (TAP) models aim to track any point through a video which is a crucial task in AR/XR and robotics applications. The recently introduced TAPNext approach proposes an end-to-end, recurrent transformer architecture to track points frame-by-frame in a purely online fashion – demonstrating competitive performance at minimal latency. However, we show that TAPNext struggles with longer video sequences and also frequently fails to re-detect query points that reappear after being occluded or leaving the frame. In this work, we present TAPNext++, a model that tracks points in sequences that are orders of magnitude longer while preserving the low memory and compute footprint of the architecture. We train the recurrent video transformer using several data-driven solutions, including training on long 1024-frame sequences enabled by sequence parallelism techniques. We highlight that re-detection performance is a blind spot in the current literature and introduce a new metric, Re-Detection Average Jaccard ($AJ_{RD}$), to explicitly evaluate tracking on re-appearing points. To improve re-detection of points, we introduce tailored geometric augmentations, such as periodic roll that simulates point re-entries, and supervising occluded points. We demonstrate that recurrent transformers can be substantially improved for point tracking and set a new state-of-the-art on multiple benchmarks. Model and code can be found at https://tap-next-plus-plus.github.io.

Method: Proposes CoFusion with: 1) Multi-Scale Generator (MSG) for three-level pyramidal architecture integrating global semantics and local details; 2) dual-branch strategy with Spatial Coordinate-Aware Mixing (SpaCAM) for multi-scale spatial contexts and Spectral Coordinate-Aware Mixing (SpeCAM) for spectral representations via frequency decomposition; 3) Spatial-Spectral Cross-Fusion Module (SSCFM) for dynamic cross-modal alignment and complementary feature fusion.

Result: Extensive experiments on multiple benchmark datasets show CoFusion consistently outperforms state-of-the-art methods, achieving superior performance in both spatial reconstruction and spectral consistency.

Conclusion: CoFusion effectively addresses the limitations of existing methods by explicitly modeling cross-scale and cross-modal dependencies, providing a robust solution for multispectral and hyperspectral image fusion.

Abstract: Multispectral and Hyperspectral Image Fusion (MHIF) aims to reconstruct high-resolution images by integrating low-resolution hyperspectral images (LRHSI) and high-resolution multispectral images (HRMSI). However, existing methods face limitations in modeling cross-scale interactions and spatial-spectral collaboration, making it difficult to achieve an optimal trade-off between spatial detail enhancement and spectral fidelity. To address this challenge, we propose CoFusion: a unified spatial-spectral collaborative fusion framework that explicitly models cross-scale and cross-modal dependencies. Specifically, a Multi-Scale Generator (MSG) is designed to construct a three-level pyramidal architecture, enabling the effective integration of global semantics and local details. Within each scale, a dual-branch strategy is employed: the Spatial Coordinate-Aware Mixing module (SpaCAM) is utilized to capture multi-scale spatial contexts, while the Spectral Coordinate-Aware Mixing module (SpeCAM) enhances spectral representations through frequency decomposition and coordinate mixing. Furthermore, we introduce the Spatial-Spectral Cross-Fusion Module (SSCFM) to perform dynamic cross-modal alignment and complementary feature fusion. Extensive experiments on multiple benchmark datasets demonstrate that CoFusion consistently outperforms state-of-the-art methods, achieving superior performance in both spatial reconstruction and spectral consistency.

Method: Created GeoMeld dataset with 2.5M samples using unified alignment protocol and agentic captioning framework; developed GeoMeld-FM pretraining combining multi-pretext masked autoencoding, JEPA representation learning, and caption-vision contrastive alignment.

Result: Experiments show consistent gains in downstream transfer and cross-sensor robustness, establishing scalable framework for semantically grounded multimodal foundation modeling in remote sensing.

Conclusion: GeoMeld and GeoMeld-FM provide a scalable reference framework for multimodal foundation modeling in remote sensing with semantically grounded supervision and aligned heterogeneous modalities.

Abstract: Effective foundation modeling in remote sensing requires spatially aligned heterogeneous modalities coupled with semantically grounded supervision, yet such resources remain limited at scale. We present GeoMeld, a large-scale multimodal dataset with approximately 2.5 million spatially aligned samples. The dataset spans diverse modalities and resolutions and is constructed under a unified alignment protocol for modality-aware representation learning. GeoMeld provides semantically grounded language supervision through an agentic captioning framework that synthesizes and verifies annotations from spectral signals, terrain statistics, and structured geographic metadata, encoding measurable cross-modality relationships within textual descriptions. To leverage this dataset, we introduce GeoMeld-FM, a pretraining framework that combines multi-pretext masked autoencoding over aligned modalities, JEPA representation learning, and caption-vision contrastive alignment. This joint objective enables the learned representation space to capture both reliable cross-sensor physical consistency and grounded semantics. Experiments demonstrate consistent gains in downstream transfer and cross-sensor robustness. Together, GeoMeld and GeoMeld-FM establish a scalable reference framework for semantically grounded multi-modal foundation modeling in remote sensing.

Method: COREY combines memory-aware operator fusion with Hadamard-based feature reparameterization. It uses activation entropy (estimated with fixed-width histograms) as a runtime scheduling statistic to determine fusion boundaries and tile sizes. To regularize heavy-tailed activations, it absorbs normalized Hadamard transforms into linear projections while preserving functional equivalence.

Result: In controlled prototype studies over heavy-tailed SSM activations, COREY consistently reduces proxy latency, improves throughput, and lowers DRAM traffic compared to unfused and fixed-depth baselines. Low-bit results are reported through a hand-crafted stability proxy for diagnostic purposes.

Conclusion: COREY demonstrates effective optimization of SSM deployments through memory-aware fusion and feature reparameterization, addressing practical memory bandwidth limitations while maintaining functional equivalence.

Abstract: State Space Models (SSMs), represented by the Mamba family, provide linear-time sequence modeling and are attractive for long-context inference. Yet practical deployments remain memory-bandwidth limited because selective state updates are often decomposed into fragmented kernels with repeated intermediate tensor materialization. We present COREY, a prototype framework that combines memory-aware operator fusion with Hadamard-based feature reparameterization. Activation entropy, estimated with fixed-width histograms, is used as a runtime scheduling statistic to place fusion boundaries and choose tile sizes. To regularize heavy-tailed activations, we absorb normalized Hadamard transforms into linear projections, preserving functional equivalence while reducing peak-coordinate concentration. In a controlled prototype study over heavy-tailed SSM activations, COREY consistently reduces proxy latency, improves throughput, and lowers DRAM traffic relative to unfused and fixed-depth baselines. Low-bit results are reported only through a hand-crafted stability proxy and are intended as diagnostic evidence rather than checkpoint-level quality claims. Code repository: https://github.com/mabo1215/COREY_Transformer.git.

Method: Evaluated several open-source and proprietary VLMs on WLASL300 benchmark using prompt-only zero-shot inference, then conducted follow-up experiments to analyze visual-semantic alignment between signs and text descriptions.

Result: Open-source VLMs performed far behind classic supervised ISLR classifiers, but captured partial visual-semantic alignment. Larger proprietary models achieved substantially higher accuracy, highlighting importance of model scale and training data diversity.

Conclusion: Current VLMs show promise for zero-shot ISLR but require further scaling and diverse training to match specialized supervised approaches, with proprietary models demonstrating better performance due to their larger scale.

Abstract: Recent Vision Language Models (VLMs) have demonstrated strong performance across a wide range of multimodal reasoning tasks. This raises the question of whether such general-purpose models can also address specialized visual recognition problems such as isolated sign language recognition (ISLR) without task-specific training. In this work, we investigate the capability of modern VLMs to perform ISLR in a zero-shot setting. We evaluate several open-source and proprietary VLMs on the WLASL300 benchmark. Our experiments show that, under prompt-only zero-shot inference, current open-source VLMs remain far behind classic supervised ISLR classifiers by a wide margin. However, follow-up experiments reveal that these models capture partial visual-semantic alignment between signs and text descriptions. Larger proprietary models achieve substantially higher accuracy, highlighting the importance of model scale and training data diversity. All our code is publicly available on GitHub.

Method: Proposes DINOCell framework that leverages DINOv2 representations and adapts them to microscopy through continued self-supervised training on unlabeled cell images before supervised fine-tuning. This domain adaptation approach better aligns representations with microscopy data characteristics.

Result: Achieves SEG score of 0.784 on LIVECell benchmark, improving by 10.42% over leading SAM-based models. Shows strong zero-shot performance on three out-of-distribution microscopy datasets, demonstrating robust domain adaptation.

Conclusion: Domain-adapted self-supervised pretraining is beneficial for robust cell segmentation, outperforming natural-image pretrained models by better aligning representations with microscopy data characteristics through continued self-supervised training.

Abstract: Instance segmentation enables the analysis of spatial and temporal properties of cells in microscopy images by identifying the pixels belonging to each cell. However, progress is constrained by the scarcity of high-quality labeled microscopy datasets. Many recent approaches address this challenge by initializing models with segmentation-pretrained weights from large-scale natural-image models such as Segment Anything Model (SAM). However, representations learned from natural images often encode objectness and texture priors that are poorly aligned with microscopy data, leading to degraded performance under domain shift. We propose DINOCell, a self-supervised framework for cell instance segmentation that leverages representations from DINOv2 and adapts them to microscopy through continued self-supervised training on unlabeled cell images prior to supervised fine-tuning. On the LIVECell benchmark, DINOCell achieves a SEG score of 0.784, improving by 10.42% over leading SAM-based models, and demonstrates strong zero-shot performance on three out-of-distribution microscopy datasets. These results highlight the benefits of domain-adapted self-supervised pretraining for robust cell segmentation.

Method: Proposes a low-bit gradient camera scheme leveraging gradient-based sensing for efficient data representation. Uses a multi-scale reconstruction CNN to reconstruct high-resolution images from the gradient measurements.

Result: Extensive experiments on simulated and real data demonstrate promising reconstruction quality and feasibility of the proposed method for high-throughput video imaging.

Conclusion: The proposed gradient camera scheme effectively addresses readout and transmission bottlenecks for high-throughput video imaging while maintaining image quality through neural network reconstruction.

Abstract: High throughput video acquisition is a challenging problem and has been drawing increasing attention. Existing high throughput imaging systems splice hundreds of sub-images/videos into high throughput videos, suffering from extremely high system complexity. Alternatively, with pixel sizes reducing to sub-micrometer levels, integrating ultra-high throughput on a single chip is becoming feasible. Nevertheless, the readout and output transmission speed cannot keep pace with the increasing pixel numbers. To this end, this paper analyzes the strength of gradient cameras in fast readout and efficient representation, and proposes a low-bit gradient camera scheme based on existing technologies that can resolve the readout and transmission bottlenecks for high throughput video imaging. A multi-scale reconstruction CNN is proposed to reconstruct high-resolution images. Extensive experiments on both simulated and real data are conducted to demonstrate the promising quality and feasibility of the proposed method.

Method: Challenge format with 168 registered teams submitting solutions evaluated on Raindrop Clarity dataset containing 14,139 training, 407 validation, and 593 testing images of real-world raindrop-affected scenes.

Result: 17 teams submitted valid final solutions achieving strong performance on the dataset, demonstrating progress in this challenging computer vision task.

Conclusion: The challenge successfully advanced raindrop removal techniques and established practical benchmarks for image restoration under varying illumination and focus conditions.

Abstract: This paper presents an overview of the NTIRE 2026 Second Challenge on Day and Night Raindrop Removal for Dual-Focused Images. Building upon the success of the first edition, this challenge attracted a wide range of impressive solutions, all developed and evaluated on our real-world Raindrop Clarity dataset~\cite{jin2024raindrop}. For this edition, we adjust the dataset with 14,139 images for training, 407 images for validation, and 593 images for testing. The primary goal of this challenge is to establish a strong and practical benchmark for the removal of raindrops under various illumination and focus conditions. In total, 168 teams have registered for the competition, and 17 teams submitted valid final solutions and fact sheets for the testing phase. The submitted methods achieved strong performance on the Raindrop Clarity dataset, demonstrating the growing progress in this challenging task.

Method: Proposes CLIP-guided Cross-Entropy Loss (CLIP-CE) using Approximation of Mutual Exclusion (AME) to weight semantic weakening. Introduces Fine-tuned AME (FAME) for adaptive weight tuning. Also creates HazyCOCO dataset with 61,258 synthetic hazy images.

Result: Achieves state-of-the-art performance on hazy object detection benchmarks. The method effectively enhances weakened semantics without unstable image enhancement modules.

Conclusion: Language prompts can effectively enhance object semantics in degraded conditions, providing a stable alternative to image enhancement for hazy object detection.

Abstract: Object detection in hazy environments is challenging because degraded objects are nearly invisible and their semantics are weakened by environmental noise, making it difficult for detectors to identify. Common approaches involve image enhancement to boost weakened semantics, but these methods are limited by the instability of enhanced modules. This paper proposes a novel solution by employing language prompts to enhance weakened semantics without image enhancement. Specifically, we design Approximation of Mutual Exclusion (AME) to provide credible weights for Cross-Entropy Loss, resulting in CLIP-guided Cross-Entropy Loss (CLIP-CE). The provided weights assess the semantic weakening of objects. Through the backpropagation of CLIP-CE, weakened semantics are enhanced, making degraded objects easier to detect. In addition, we present Fine-tuned AME (FAME) which adaptively fine-tunes the weight of AME based on the predicted confidence. The proposed FAME compensates for the imbalanced optimization in AME. Furthermore, we present HazyCOCO, a large-scale synthetic hazy dataset comprising 61258 images. Experimental results demonstrate that our method achieves state-of-the-art performance. The code and dataset will be released.

Method: Analyze VLM internal representations to examine aesthetic attribute presence and distribution, then leverage these representations for lightweight personalization using simple linear models without model fine-tuning.

Result: VLMs encode diverse aesthetic attributes that propagate into language decoder layers, enabling effective PIAA with simple linear models. Analysis reveals how aesthetic information transfers across VLM layers and image domains.

Conclusion: VLMs can be effectively utilized for modeling subjective, individual aesthetic preferences through their internal representations, providing insights for personalized aesthetics assessment.

Abstract: Personalized image aesthetics assessment (PIAA) is an important research problem with practical real-world applications. While methods based on vision-language models (VLMs) are promising candidates for PIAA, it remains unclear whether they internally encode rich, multi-level aesthetic attributes required for effective personalization. In this paper, we first analyze the internal representations of VLMs to examine the presence and distribution of such aesthetic attributes, and then leverage them for lightweight, individual-level personalization without model fine-tuning. Our analysis reveals that VLMs encode diverse aesthetic attributes that propagate into the language decoder layers. Building on these representations, we demonstrate that simple linear models can perform PIAA effectively. We further analyze how aesthetic information is transferred across layers in different VLM architectures and across image domains. Our findings provide insights into how VLMs can be utilized for modeling subjective, individual aesthetic preferences. Our code is available at https://github.com/ynklab/vlm-latent-piaa.

Method: Attach lightweight linear heads to intermediate ViT layers to extract features capturing class evidence evolution, including predicted class logits, top-K competitors, and statistics about class ranking instability across depth. Train a linear probe on these features to predict error indicators.

Result: The method improves or matches AUCPR over baselines across datasets and shows stronger cross-dataset generalization while requiring minimal additional computation.

Conclusion: Vision Transformers contain depth-wise signals that can be leveraged for reliable error prediction, similar to hallucination detection in LLMs, enabling efficient confidence estimation.

Abstract: Reliable confidence estimation is critical when deploying vision models. We study error prediction: determining whether an image classifier’s output is correct using only signals from a single forward pass. Motivated by internal-signal hallucination detection in large language models, we investigate whether similar depth-wise signals exist in Vision Transformers (ViTs). We propose a simple method that models how class evidence evolves across layers. By attaching lightweight linear heads to intermediate layers, we extract features from the last L layers that capture both the logits of the predicted class and its top-K competitors, as well as statistics describing instability of top-ranked classes across depth. A linear probe trained on these features predicts the error indicator. Across datasets, our method improves or matches AUCPR over baselines and shows stronger cross-dataset generalization while requiring minimal additional computation.

Method: MERIT searches over layer-wise self-attention merging recipes between a VLM and its paired text-only backbone using an objective that improves temporal reasoning while penalizing degradation in temporal perception.

Result: Across three representative VLMs and multiple challenging video benchmarks, MERIT consistently improves temporal reasoning, preserves or improves temporal perception, and generalizes beyond the search set to four distinct benchmarks.

Conclusion: Targeted, perception-aware model merging can effectively restore temporal reasoning in VLMs without retraining, with selected layers being disproportionately important for reasoning and shifting model decisions toward temporally relevant evidence.

Abstract: Multimodal adaptation equips large language models (LLMs) with perceptual capabilities, but often weakens the reasoning ability inherited from language-only pretraining. This trade-off is especially pronounced in video-language models (VLMs), where visual alignment can impair temporal reasoning (TR) over sequential events. We propose MERIT, a training-free, task-driven model merging framework for restoring TR in VLMs. MERIT searches over layer-wise self-attention merging recipes between a VLM and its paired text-only backbone using an objective that improves TR while penalizing degradation in temporal perception (TP). Across three representative VLMs and multiple challenging video benchmarks, MERIT consistently improves TR, preserves or improves TP, and generalizes beyond the search set to four distinct benchmarks. It also outperforms uniform full-model merging and random layer selection, showing that effective recovery depends on selecting the right layers. Interventional masking and frame-level attribution further show that the selected layers are disproportionately important for reasoning and shift model decisions toward temporally and causally relevant evidence. These results show that targeted, perception-aware model merging can effectively restore TR in VLMs without retraining.

Method: Proposes a fully automated framework that evaluates dense object placements on real backgrounds using diffusion-based inpainting. Creates HiddenObjects dataset with 27M placement annotations across 27k scenes, with ranked bounding box insertions. Distills spatial priors into a lightweight model for fast inference.

Result: Spatial priors outperform sparse human annotations on downstream image editing (3.90 vs. 2.68 VLM-Judge), significantly surpass existing placement baselines and zero-shot Vision-Language Models. Distilled model achieves 230,000x faster inference.

Conclusion: The method successfully learns explicit spatial priors from diffusion models, creating a scalable framework for object placement that outperforms existing approaches and enables practical fast inference.

Abstract: We propose a method to learn explicit, class-conditioned spatial priors for object placement in natural scenes by distilling the implicit placement knowledge encoded in text-conditioned diffusion models. Prior work relies either on manually annotated data, which is inherently limited in scale, or on inpainting-based object-removal pipelines, whose artifacts promote shortcut learning. To address these limitations, we introduce a fully automated and scalable framework that evaluates dense object placements on high-quality real backgrounds using a diffusion-based inpainting pipeline. With this pipeline, we construct HiddenObjects, a large-scale dataset comprising 27M placement annotations, evaluated across 27k distinct scenes, with ranked bounding box insertions for different images and object categories. Experimental results show that our spatial priors outperform sparse human annotations on a downstream image editing task (3.90 vs. 2.68 VLM-Judge), and significantly surpass existing placement baselines and zero-shot Vision-Language Models for object placement. Furthermore, we distill these priors into a lightweight model for fast practical inference (230,000x faster).

Method: Introduces a lightweight transformer that learns fine-grained region-segment alignments directly from frozen patch and token embeddings. The approach explicitly enforces localized alignment at inference without updating pretrained encoders.

Result: Learning localized alignment over frozen representations matches full fine-tuning on in-domain retrieval while yielding substantial improvements on controlled out-of-domain compositional benchmarks. The approach outperforms prior end-to-end compositional training methods under distribution shift.

Conclusion: Global embedding matching is identified as a key bottleneck in dual-encoder VLMs, and alignment mechanisms are crucial for robust compositional generalization. The work demonstrates that compositional limitations can be addressed through inference-time alignment rather than representation updates.

Abstract: Dual-encoder Vision-Language Models (VLMs) such as CLIP are often characterized as bag-of-words systems due to their poor performance on compositional benchmarks. We argue that this limitation may stem less from deficient representations than from the standard inference protocol based on global cosine similarity. First, through controlled diagnostic experiments, we show that explicitly enforcing fine-grained region-segment alignment at inference dramatically improves compositional performance without updating pretrained encoders. We then introduce a lightweight transformer that learns such alignments directly from frozen patch and token embeddings. Comparing against full fine-tuning and prior end-to-end compositional training methods, we find that although these approaches improve in-domain retrieval, their gains do not consistently transfer under distribution shift. In contrast, learning localized alignment over frozen representations matches full fine-tuning on in-domain retrieval while yielding substantial improvements on controlled out-of-domain compositional benchmarks. These results identify global embedding matching as a key bottleneck in dual-encoder VLMs and highlight the importance of alignment mechanisms for robust compositional generalization.

Method: Proposes R²ScP framework that shifts from generative imputation to retrieval-based recovery using cross-modal retrieval via unified semantic embeddings, with context-aware adaptive purification to remove semantic noise and two-stage training for modeling semantic relationships.

Result: Extensive experiments show R²ScP significantly improves AVQA performance and enhances robustness in modal-incomplete scenarios compared to existing methods.

Conclusion: Retrieval-based recovery with adaptive purification is more effective than generative imputation for handling missing modalities in AVQA, addressing the limitations of capturing modality-specific knowledge and reducing hallucinations.

Abstract: Recent Audio-Visual Question Answering (AVQA) methods have advanced significantly. However, most AVQA methods lack effective mechanisms for handling missing modalities, suffering from severe performance degradation in real-world scenarios with data interruptions. Furthermore, prevailing methods for handling missing modalities predominantly rely on generative imputation to synthesize missing features. While partially effective, these methods tend to capture inter-modal commonalities but struggle to acquire unique, modality-specific knowledge within the missing data, leading to hallucinations and compromised reasoning accuracy. To tackle these challenges, we propose R$^{2}$ScP, a novel framework that shifts the paradigm of missing modality handling from traditional generative imputation to retrieval-based recovery. Specifically, we leverage cross-modal retrieval via unified semantic embeddings to acquire missing domain-specific knowledge. To maximize semantic restoration, we introduce a context-aware adaptive purification mechanism that eliminates latent semantic noise within the retrieved data. Additionally, we employ a two-stage training strategy to explicitly model the semantic relationships between knowledge from different sources. Extensive experiments demonstrate that R$^{2}$ScP significantly improves AVQA and enhances robustness in modal-incomplete scenarios.

Method: Proposes MIGF with separate modality-specific encoding streams before learned gating, combined with modality dropout training to enforce compensation behavior under incomplete inputs. Tested on six backbones with seven missing-modality scenarios.

Result: MIGF improved ideal-scenario Ranking Score for UNet, nnUNet, and Mamba by 2.8%, 4.6%, and 13.4% respectively. Best model (MIGFNet-nnUNet) achieved 0.7304 +/- 0.056. Robustness gains come from strict modality isolation and dropout-driven compensation.

Conclusion: For robust multi-modal segmentation: structurally contain corrupted inputs first, then train explicitly for incomplete-input compensation. Simpler design principle outperforms complex adaptive routing approaches.

Abstract: Multi-parametric prostate MRI – combining T2-weighted, apparent diffusion coefficient, and high b-value diffusion-weighted sequences – is central to non-invasive detection of clinically significant prostate cancer, yet in routine practice individual sequences may be missing or degraded by motion, artifacts, or abbreviated protocols. Existing multi-modal fusion strategies typically assume complete inputs and entangle modality-specific information at early layers, offering limited resilience when one channel is corrupted or absent. We propose Modality-Isolated Gated Fusion (MIGF), an architecture-agnostic module that maintains separate modality-specific encoding streams before a learned gating stage, combined with modality dropout training to enforce compensation behavior under incomplete inputs. We benchmark six bare backbones and assess MIGF-equipped models under seven missing-modality and artifact scenarios on the PI-CAI dataset (1,500 studies, fold-0 split, five random seeds). Among bare backbones, nnUNet provided the strongest balance of performance and stability. MIGF improved ideal-scenario Ranking Score for UNet, nnUNet, and Mamba by 2.8%, 4.6%, and 13.4%, respectively; the best model, MIGFNet-nnUNet (gating + ModDrop, no deep supervision), achieved 0.7304 +/- 0.056. Mechanistic analysis reveals that robustness gains arise from strict modality isolation and dropout-driven compensation rather than adaptive per-sample quality routing: the gate converged to a stable modality prior, and deep supervision was beneficial only for the largest backbone while degrading lighter models. These findings support a simpler design principle for robust multi-modal segmentation: structurally contain corrupted inputs first, then train explicitly for incomplete-input compensation.

Method: Established fairness baseline by analyzing state-of-the-art gaze estimation models using standard fairness metrics, evaluated effectiveness of existing bias mitigation strategies when applied to gaze domain.

Result: Revealed significant performance disparities across ethnicity and gender groups, showed that existing bias mitigation strategies have limited fairness contributions in gaze estimation.

Conclusion: Calls for research into developing robust, equitable gaze estimators, releases annotations, code, and trained models to support future research and reproducibility.

Abstract: While appearance-based gaze estimation has achieved significant improvements in accuracy and domain adaptation, the fairness of these systems across different demographic groups remains largely unexplored. To date, there is no comprehensive benchmark quantifying algorithmic bias in gaze estimation. This paper presents the first extensive evaluation of fairness in appearance-based gaze estimation, focusing on ethnicity and gender attributes. We establish a fairness baseline by analyzing state-of-the-art models using standard fairness metrics, revealing significant performance disparities. Furthermore, we evaluate the effectiveness of existing bias mitigation strategies when applied to the gaze domain and show that their fairness contributions are limited. We summarize key insights and open issues. Overall, our work calls for research into developing robust, equitable gaze estimators. To support future research and reproducibility, we publicly release our annotations, code, and trained models at: github.com/akgulburak/gaze-estimation-fairness

Method: Proposes Adversarial Spectrum Defense (ASD) that uses spectral decomposition via Discrete Wavelet Transform to analyze adversarial patterns across multiple frequency scales, capturing both high-frequency (fine-grained) and low-frequency (spatially pervasive) perturbations. Combines this with off-the-shelf Adversarial Training.

Result: ASD+AT achieved state-of-the-art performance against various attacks, outperforming previous defense methods by 21.73% in APs, even against strong adaptive adversaries specifically designed against ASD.

Conclusion: ASD provides a comprehensive defense strategy against both patch-based and texture-based adversarial attacks by leveraging multi-resolution spectral analysis, demonstrating robustness even against adaptive attacks.

Abstract: Adversarial examples present significant challenges to the security of Deep Neural Network (DNN) applications. Specifically, there are patch-based and texture-based attacks that are usually used to craft physical-world adversarial examples, posing real threats to security-critical applications such as person detection in surveillance and autonomous systems, because those attacks are physically realizable. Existing defense mechanisms face challenges in the adaptive attack setting, i.e., the attacks are specifically designed against them. In this paper, we propose Adversarial Spectrum Defense (ASD), a defense mechanism that leverages spectral decomposition via Discrete Wavelet Transform (DWT) to analyze adversarial patterns across multiple frequency scales. The multi-resolution and localization capability of DWT enables ASD to capture both high-frequency (fine-grained) and low-frequency (spatially pervasive) perturbations. By integrating this spectral analysis with the off-the-shelf Adversarial Training (AT) model, ASD provides a comprehensive defense strategy against both patch-based and texture-based adversarial attacks. Extensive experiments demonstrate that ASD+AT achieved state-of-the-art (SOTA) performance against various attacks, outperforming the APs of previous defense methods by 21.73%, in the face of strong adaptive adversaries specifically designed against ASD. Code available at https://github.com/weiz0823/adv-spectral-defense .

Method: Parameter-efficient finetuning of MLLMs that optimizes latent representations while preserving pretrained multimodal knowledge, enabling strong cross-modal alignment without architectural redesign. Systematic analysis of diverse variables from model backbone to feature aggregation.

Result: Achieves SOTA on GeoText-1652 with 12.2% improvement in Text-to-Image Recall@1, top performance in 5 out of 12 subtasks on CVG-Text, surpassing baselines with far fewer trainable parameters.

Conclusion: MLLMs serve as robust foundation for semantic cross-view retrieval, paving way for MLLM-based geo-localization as scalable, powerful alternative to traditional dual-encoder designs.

Abstract: Natural-language Guided Cross-view Geo-localization (NGCG) aims to retrieve geo-tagged satellite imagery using textual descriptions of ground scenes. While recent NGCG methods commonly rely on CLIP-style dual-encoder architectures, they often suffer from weak cross-modal generalization and require complex architectural designs. In contrast, Multimodal Large Language Models (MLLMs) offer powerful semantic reasoning capabilities but are not directly optimized for retrieval tasks. In this work, we present a simple yet effective framework to adapt MLLMs for NGCG via parameter-efficient finetuning. Our approach optimizes latent representations within the MLLM while preserving its pretrained multimodal knowledge, enabling strong cross-modal alignment without redesigning model architectures. Through systematic analysis of diverse variables, from model backbone to feature aggregation, we provide practical and generalizable insights for leveraging MLLMs in NGCG. Our method achieves SOTA on GeoText-1652 with a 12.2% improvement in Text-to-Image Recall@1 and secures top performance in 5 out of 12 subtasks on CVG-Text, all while surpassing baselines with far fewer trainable parameters. These results position MLLMs as a robust foundation for semantic cross-view retrieval and pave the way for MLLM-based NGCG to be adopted as a scalable, powerful alternative to traditional dual-encoder designs. Project page and code are available at https://yuqichen888.github.io/NGCG-MLLMs-web/.

Method: Created MMRareBench with 1,756 QA pairs and 7,958 medical images from PMC case reports, featuring Orphanet-anchored ontology alignment, track-specific leakage control, evidence-grounded annotations, and two-level evaluation protocol. Evaluated 23 MLLMs across four workflow-aligned tracks.

Result: Evaluation revealed fragmented capability profiles, universally low treatment-planning performance, and medical-domain models trailing general-purpose MLLMs on multi-image tracks despite competitive diagnostic scores. Shows capacity dilution effect where medical fine-tuning narrows diagnostic gap but erodes compositional multi-image capability.

Conclusion: Rare-disease clinical workflows demand multimodal and multi-image evidence integration that current MLLMs struggle with, especially in treatment planning. Medical fine-tuning may inadvertently harm compositional reasoning needed for rare disease cases.

Abstract: Multimodal large language models (MLLMs) have advanced clinical tasks for common conditions, but their performance on rare diseases remains largely untested. In rare-disease scenarios, clinicians often lack prior clinical knowledge, forcing them to rely strictly on case-level evidence for clinical judgments. Existing benchmarks predominantly evaluate common-condition, single-image settings, leaving multimodal and multi-image evidence integration under rare-disease data scarcity systematically unevaluated. We introduce MMRareBench, to our knowledge the first rare-disease benchmark jointly evaluating multimodal and multi-image clinical capability across four workflow-aligned tracks: diagnosis, treatment planning, cross-image evidence alignment, and examination suggestion. The benchmark comprises 1,756 question-answer pairs with 7,958 associated medical images curated from PMC case reports, with Orphanet-anchored ontology alignment, track-specific leakage control, evidence-grounded annotations, and a two-level evaluation protocol. A systematic evaluation of 23 MLLMs reveals fragmented capability profiles and universally low treatment-planning performance, with medical-domain models trailing general-purpose MLLMs substantially on multi-image tracks despite competitive diagnostic scores. These patterns are consistent with a capacity dilution effect: medical fine-tuning can narrow the diagnostic gap but may erode the compositional multi-image capability that rare-disease evidence integration demands.

Method: Comparative study of deep learning models: InceptionV3, MobileNetV2, VGG16, ResNet152, plus a novel proposed 16-layer CNN architecture with convolutional, pooling, flatten, dropout, fully connected, and dense layers. Tested up to 30 epochs with metrics including accuracy, precision, recall, F1-score.

Result: Proposed 16-layer CNN model shows increasing accuracy with more epochs (tested up to 30), overcomes overfitting problems, and provides comprehensive performance comparison of all models using standard evaluation metrics.

Conclusion: The proposed CNN architecture demonstrates improved lung cancer detection capabilities with consistent accuracy improvement across epochs and overcomes overfitting, contributing to advancement in medical imaging classification.

Abstract: Lung cancer, the second leading cause of cancer-related deaths, is primarily linked to long-term tobacco smoking (85% of cases). Surprisingly, 10-15% of cases occur in non-smokers. In 2020, approximately 2 million people were affected globally, resulting in 1.5 million deaths. The survival rate, at around 20%, lags behind other cancers, partly due to late-stage symptom manifestation. Necessitates early and accurate detection for effective treatment. Performance metrics such as accuracy, precision, recall (sensitivity), and F1-score are computed to provide a comprehensive evaluation of each model’s capabilities. By comparing these metrics, this study offers insights into the strengths and limitations of each approach, contributing to the advancement of lung cancer detection techniques. In this paper, we are going to discuss the methodologies of lung cancer detection using different deep learning algorithms - InceptionV3, MobileNetV2, VGG16, ResNet152 - are explored for their efficacy in classifying lung cancer cases. Our Proposed Model algorithm based is a 16 layers architecture based on CNN model. Our Proposed model exhibits several key highlights that contribute to its novelty. By integrating multiple layer types such as convolutional, pooling, flatten, dropout, fully connected and dense layers, the model leverages the strengths of each layer to enhance its predictive capabilities. Novelty of our proposed model is that its accuracy is increasing consistently with the increasing no of epochs. We have tested the model performance up to epoch no 30. Our proposed model also overcome the overfitting problem.

Method: FullTilt operates directly on aligned 2D tilt-series rather than reconstructed 3D volumes. It introduces: 1) a tilt-series encoder for efficient cross-view information fusion, 2) a multiclass visual prompt encoder for flexible prompting, 3) a tilt-aware query initializer to anchor 3D queries effectively, and 4) an auxiliary geometric primitives module to enhance multi-view geometry understanding and robustness to imaging artifacts.

Result: Extensive evaluations on three real-world datasets show that FullTilt achieves state-of-the-art zero-shot performance while drastically reducing runtime and VRAM requirements. The method accelerates inference by orders of magnitude compared to traditional volumetric approaches.

Conclusion: FullTilt represents a paradigm shift in 3D macromolecule detection by eliminating redundant volumetric computation through direct tilt-series processing. The framework enables rapid, large-scale visual proteomics analysis with practical computational requirements.

Abstract: Open-set 3D macromolecule detection in cryogenic electron tomography eliminates the need for target-specific model retraining. However, strict VRAM constraints prohibit processing an entire 3D tomogram, forcing current methods to rely on slow sliding-window inference over extracted subvolumes. To overcome this, we propose FullTilt, an end-to-end framework that redefines 3D detection by operating directly on aligned 2D tilt-series. Because a tilt-series contains significantly fewer images than slices in a reconstructed tomogram, FullTilt eliminates redundant volumetric computation, accelerating inference by orders of magnitude. To process the entire tilt-series simultaneously, we introduce a tilt-series encoder to efficiently fuse cross-view information. We further propose a multiclass visual prompt encoder for flexible prompting, a tilt-aware query initializer to effectively anchor 3D queries, and an auxiliary geometric primitives module to enhance the model’s understanding of multi-view geometry while improving robustness to adverse imaging artifacts. Extensive evaluations on three real-world datasets demonstrate that FullTilt achieves state-of-the-art zero-shot performance while drastically reducing runtime and VRAM requirements, paving the way for rapid, large-scale visual proteomics analysis. All code and data will be publicly available upon publication.

Method: Uses LLMs and VLMs for text-driven hierarchical scene generation, incorporates RAG for semantic consistency, adds optimization module for physical consistency, and employs hierarchical representation for real-time editing.

Result: Produces more reasonable environments than existing baselines while supporting fast and intuitive scene editing with improved semantic and physical consistency.

Conclusion: HOG-Layout successfully enables text-driven hierarchical 3D scene generation and real-time editing with improved consistency and plausibility through LLMs, VLMs, RAG, and optimization techniques.

Abstract: 3D layout generation and editing play a crucial role in Embodied AI and immersive VR interaction. However, manual creation requires tedious labor, while data-driven generation often lacks diversity. The emergence of large models introduces new possibilities for 3D scene synthesis. We present HOG-Layout that enables text-driven hierarchical scene generation, optimization and real-time scene editing with large language models (LLMs) and vision-language models (VLMs). HOG-Layout improves scene semantic consistency and plausibility through retrieval-augmented generation (RAG) technology, incorporates an optimization module to enhance physical consistency, and adopts a hierarchical representation to enhance inference and optimization, achieving real-time editing. Experimental results demonstrate that HOG-Layout produces more reasonable environments compared with existing baselines, while supporting fast and intuitive scene editing.

Method: Models iPPG recovery as an inverse problem using probability paths that evolve camera pixel distribution to ground-truth signal distribution via flow and score vectors; solves stochastic differential equation at test-time to sample posterior distribution; adds regularization using correlation between residual flow vectors of adjacent time windows.

Result: Superior reconstruction quality and uncertainty estimates on three datasets compared to state-of-the-art methods.

Conclusion: RIS-iPPG provides critical uncertainty quantification for iPPG algorithms, enabling wider adoption in clinical and consumer settings through reliable confidence measures.

Abstract: Imaging Photoplethysmography (iPPG), an optical procedure which recovers a human’s blood volume pulse (BVP) waveform using pixel readout from a camera, is an exciting research field with many researchers performing clinical studies of iPPG algorithms. While current algorithms to solve the iPPG task have shown outstanding performance on benchmark datasets, no state-of-the art algorithms, to the best of our knowledge, performs test-time sampling of solution space, precluding an uncertainty analysis that is critical for clinical applications. We address this deficiency though a new paradigm named Regularized Interpolants with Stochasticity for iPPG (RIS-iPPG). Modeling iPPG recovery as an inverse problem, we build probability paths that evolve the camera pixel distribution to the ground-truth signal distribution by predicting the instantaneous flow and score vectors of a time-dependent stochastic process; and at test-time, we sample the posterior distribution of the correct BVP waveform given the camera pixel intensity measurements by solving a stochastic differential equation. Given that physiological changes are slowly varying, we show that iPPG recovery can be improved through regularization that maximizes the correlation between the residual flow vector predictions of two adjacent time windows. Experimental results on three datasets show that RIS-iPPG provides superior reconstruction quality and uncertainty estimates of the reconstruction, a critical tool for the widespread adoption of iPPG algorithms in clinical and consumer settings.

Method: Developed a unified, extensible PyTorch library with registry-based framework supporting 29 supervised architectures, 7 SSL pre-training methods, and 5 PEFT strategies. Includes standardized training runners, cross-validation, automated table generation, statistical testing, and comprehensive test suite.

Result: Created LIDARLearn library with 55+ model configurations covering classification, semantic segmentation, part segmentation, and few-shot learning. Includes 2,200+ automated tests for end-to-end validation and supports rigorous multi-model comparison with statistical testing.

Conclusion: LIDARLearn provides a comprehensive, standardized framework for point cloud analysis that enables fair comparisons, reproducibility, and extensibility for the research community.

Abstract: Three-dimensional (3D) point cloud analysis has become central to applications ranging from autonomous driving and robotics to forestry and ecological monitoring. Although numerous deep learning methods have been proposed for point cloud understanding, including supervised backbones, self-supervised pre-training (SSL), and parameter-efficient fine-tuning (PEFT), their implementations are scattered across incompatible codebases with differing data pipelines, evaluation protocols, and configuration formats, making fair comparisons difficult. We introduce \lib{}, a unified, extensible PyTorch library that integrates over 55 model configurations covering 29 supervised architectures, seven SSL pre-training methods, and five PEFT strategies, all within a single registry-based framework supporting classification, semantic segmentation, part segmentation, and few-shot learning. \lib{} provides standardised training runners, cross-validation with stratified $K$-fold splitting, automated LaTeX/CSV table generation, built-in Friedman/Nemenyi statistical testing with critical-difference diagrams for rigorous multi-model comparison, and a comprehensive test suite with 2,200+ automated tests validating every configuration end-to-end. The code is available at https://github.com/said-ohamouddou/LIDARLearn under the MIT licence.

Method: Five-stage cascade pipeline that extracts and structurally aligns generic priors from vision foundation models across textual, visual, and spatial dimensions, grounding them into structured 3D representations while ensuring semantic coherence and physical plausibility.

Result: Superior performance over existing baselines in generating high-quality compositional 3D scenes, demonstrated through extensive experiments. Also introduces C3DR benchmark for comprehensive evaluation.

Conclusion: ReplicateAnyScene enables practical deployment of compositional 3D reconstruction from casual videos without manual intervention, advancing Spatial Intelligence and Embodied AI capabilities.

Abstract: Humans exhibit an innate capacity to rapidly perceive and segment objects from video observations, and even mentally assemble them into structured 3D scenes. Replicating such capability, termed compositional 3D reconstruction, is pivotal for the advancement of Spatial Intelligence and Embodied AI. However, existing methods struggle to achieve practical deployment due to the insufficient integration of cross-modal information, leaving them dependent on manual object prompting, reliant on auxiliary visual inputs, and restricted to overly simplistic scenes by training biases. To address these limitations, we propose ReplicateAnyScene, a framework capable of fully automated and zero-shot transformation of casually captured videos into compositional 3D scenes. Specifically, our pipeline incorporates a five-stage cascade to extract and structurally align generic priors from vision foundation models across textual, visual, and spatial dimensions, grounding them into structured 3D representations and ensuring semantic coherence and physical plausibility of the constructed scenes. To facilitate a more comprehensive evaluation of this task, we further introduce the C3DR benchmark to assess reconstruction quality from diverse aspects. Extensive experiments demonstrate the superiority of our method over existing baselines in generating high-quality compositional 3D scenes.

Method: The benchmark uses single-site microscopic blood smear images with standardized staining and expert annotations. It’s organized into two phases: Phase 1 provides pristine training data, while Phase 2 introduces degraded images with split-specific severity distributions for train, validation, and test sets. The benchmark includes a standardized submission schema, open-source evaluator, and uses macro-averaged F1 score as the primary ranking metric.

Result: The paper reviews the challenge and summarizes proposed solutions and final outcomes, though specific numerical results are not provided in the abstract. The benchmark successfully creates a controlled environment to stress-test algorithms under realistic deployment conditions.

Conclusion: WBCBench 2026 provides a rigorous benchmark for WBC classification that addresses key practical challenges including domain shift, class imbalance, and proper patient-level data separation, enabling more robust algorithm development and evaluation.

Abstract: We present WBCBench 2026, an ISBI challenge and benchmark for automated WBC classification designed to stress-test algorithms under three key difficulties: (i) severe class imbalance across 13 morphologically fine-grained WBC classes, (ii) strict patient-level separation between training, validation and test sets, and (iii) synthetic scanner- and setting-induced domain shift via controlled noise, blur and illumination perturbations. All images are single-site microscopic blood smear acquisitions with standardised staining and expert hematopathologist annotations. This paper reviews the challenge and summarises the proposed solutions and final outcomes. The benchmark is organised into two phases. Phase 1 provides a pristine training set. Phase 2 introduces degraded images with split-specific severity distributions for train, validation and test, emulating a realistic shift between development and deployment conditions. We specify a standardised submission schema, open-source evaluator, and macro-averaged F1 score as the primary ranking metric.

Method: Derived explicit relationship between homography perturbations and resulting distance error, showing quadratic error growth. Evaluated two correction strategies: regression-based estimation of quadratic error function and direct homography optimization via coordinate-based gradient descent.

Result: Large-scale simulation with 19+ million test samples showed regression achieves higher peak accuracy when model is reliably fitted, while gradient descent provides greater robustness against poor initial calibration.

Conclusion: Improving geometric calibration may yield greater performance gains than increasing model complexity in practical monocular distance estimation systems.

Abstract: Accurate distance estimation from monocular cameras is essential for intelligent monitoring systems. In many deployments, image coordinates are mapped to ground positions using planar homographies initialized by manual selection of corresponding regions. Small inaccuracies in this initialization propagate into systematic distance distortions. This paper derives an explicit relationship between homography perturbations and the resulting distance error, showing that the error grows approximately quadratically with the true distance from the camera. Based on this model, two simple correction strategies are evaluated: regression-based estimation of the quadratic error function and direct optimization of the homography via coordinate-based gradient descent. A large-scale simulation study with more than 19 million test samples demonstrates that regression achieves higher peak accuracy when the model is reliably fitted, whereas gradient descent provides greater robustness against poor initial calibration. This suggests that improving geometric calibration may yield greater performance gains than increasing model complexity in many practical systems.

Method: Proposes UGDA-Net with three components: 1) Uncertainty-Guided Dual Attention using channel variance to modulate feature maps, 2) entropy-weighted hybrid loss focusing on high-uncertainty boundary pixels, and 3) deep supervision for intermediate encoder layers.

Result: Improved segmentation performance with 9.3% increase in Dice coefficient above baseline for U-Net and 13.2% above baseline for LinkNet, with reduced false positives at leaf boundaries and uncertainty heatmaps consistent with complex morphology.

Conclusion: UGDA-Net provides high-definition segmentation of delicate plant structures, showing that uncertainty-guided attention and uncertainty-weighted loss are complementary systems for improving segmentation accuracy.

Abstract: Plant seedling segmentation supports automated phenotyping in precision agriculture. Standard segmentation models face difficulties due to intricate background images and fine structures in leaves. We introduce UGDA-Net (Uncertainty-Guided Dual Attention Network with Entropy-Weighted Loss and Deep Supervision). Three novel components make up UGDA-Net. The first component is Uncertainty-Guided Dual Attention (UGDA). UGDA uses channel variance to modulate feature maps. The second component is an entropy-weighted hybrid loss function. This loss function focuses on high-uncertainty boundary pixels. The third component employs deep supervision for intermediate encoder layers. We performed a comprehensive systematic ablation study. This study focuses on two widely-used architectures, U-Net and LinkNet. It analyzes five incremental configurations: Baseline, Loss-only, Attention-only, Deep Supervision, and UGDA-Net. We trained UGDA-net using a high-resolution plant seedling image dataset containing 432 images. We demonstrate improved segmentation performance and accuracy. With an increase in Dice coefficient of 9.3% above baseline. LinkNet’s variance is 13.2% above baseline. Overlays that are qualitative in nature show the reduced false positives at the leaf boundary. Uncertainty heatmaps are consistent with the complex morphology. UGDA-Net aids in the segmentation of delicate structures in plants and provides a high-def solution. The results showed that uncertainty-guided attention and uncertainty-weighted loss are two complementing systems.

Method: 1) Interaction-aware variational autoencoder encodes hand-object motion sequences into unified latent space using kinematics; 2) Masked flow matching combines auto-regressive temporal reasoning with continuous latent generation; 3) Predicts object motions relative to initial frame for better generalization and pre-training on synthetic data.

Result: State-of-the-art performance on GRAB, OakInk, and DexYCB benchmarks in both physical plausibility and motion diversity for interaction motion synthesis.

Conclusion: HO-Flow effectively addresses limitations in 3D hand-object interaction generation by combining interaction-aware encoding with flow-based temporal modeling, enabling realistic synthesis from text and 3D objects.

Abstract: Generating realistic 3D hand-object interactions (HOI) is a fundamental challenge in computer vision and robotics, requiring both temporal coherence and high-fidelity physical plausibility. Existing methods remain limited in their ability to learn expressive motion representations for generation and perform temporal reasoning. In this paper, we present HO-Flow, a framework for synthesizing realistic hand-object motion sequences from texts and canoncial 3D objects. HO-Flow first employs an interaction-aware variational autoencoder to encode sequences of hand and object motions into a unified latent manifold by incorporating hand and object kinematics, enabling the representation to capture rich interaction dynamics. It then leverages a masked flow matching model that combines auto-regressive temporal reasoning with continuous latent generation, improving temporal coherence. To further enhance generalization, HO-Flow predicts object motions relative to the initial frame, enabling effective pre-training on large-scale synthetic data. Experiments on the GRAB, OakInk, and DexYCB benchmarks demonstrate that HO-Flow achieves state-of-the-art performance in both physical plausibility and motion diversity for interaction motion synthesis.

Method: Immune2V enforces temporally balanced latent divergence at the encoder level to prevent adversarial signal dilution across frames, and aligns intermediate generative representations with precomputed collapse-inducing trajectories to counteract text-guidance override.

Result: Extensive experiments show Immune2V produces substantially stronger and more persistent degradation than adapted image-level baselines under the same imperceptibility budget, effectively defending against unauthorized I2V generation.

Conclusion: The proposed Immune2V framework successfully addresses the unique challenges of defending against image-to-video deepfake generation by understanding and counteracting the temporal and guidance mechanisms of modern I2V models.

Abstract: Image-to-video (I2V) generation has the potential for societal harm because it enables the unauthorized animation of static images to create realistic deepfakes. While existing defenses effectively protect against static image manipulation, extending these to I2V generation remains underexplored and non-trivial. In this paper, we systematically analyze why modern I2V models are highly robust against naive image-level adversarial attacks (i.e., immunization). We observe that the video encoding process rapidly dilutes the adversarial noise across future frames, and the continuous text-conditioned guidance actively overrides the intended disruptive effect of the immunization. Building on these findings, we propose the Immune2V framework which enforces temporally balanced latent divergence at the encoder level to prevent signal dilution, and aligns intermediate generative representations with a precomputed collapse-inducing trajectory to counteract the text-guidance override. Extensive experiments demonstrate that Immune2V produces substantially stronger and more persistent degradation than adapted image-level baselines under the same imperceptibility budget.

Method: Uses ResNet CNN with patchwise classification for segmentation, trained on cyst segmentation challenge dataset and tested on data from 4 different OCT vendors with annotations from 2 graders.

Result: Achieved over 70% dice coefficient across all vendors regardless of image quality, outperforming previous state-of-the-art approaches.

Conclusion: The ResNet-based approach provides robust cyst segmentation that works well across different OCT imaging systems and image qualities, addressing limitations of previous methods.

Abstract: Retinal Cysts are formed by leakage and accumulation of fluid in the retina due to the incompetence of retinal vasculature. These cystic spaces have significance in several ocular diseases such as age-related macular degeneration, diabetic macular edema, etc. Optical coherence tomography is one of the predominant diagnosing techniques for imaging retinal pathologies. Segmenting and quantification of intraretinal cysts plays the vital role in predicting visual acuity. In literature, several methods have been proposed for automatic segmentation of intraretinal cysts. As cystoid macular edema becomes a major problem to humankind, we need to quantify it accurately and operate it out, else it might cause many problems later on. Though research is being carried out in this area, not much of progress has been made and accuracy achieved so far is 68% which is very less. Also, the methods depend on the quality of the image and give very low results for high noise images like topcon. This work uses ResNet CNN (Convolutional Neural Network) approach of segmentation by the way of patchwise classification for training on image set from cyst segmentation challenge dataset and testing on test data set given by 2 different graders for all 4 vendors in the challenge. It also compares these methods using first publicly available novel cyst segmentation challenge dataset. The methods were evaluated using quantitative measures to assess their robustness against the challenges of intraretinal cyst segmentation. The results are found to be better than the previous state of the art approaches giving more than 70% dice coefficient on all vendors irrespective of their quality.

Method: LRD-Net uses a sequential frequency-guided architecture with a Multi-Scale Wavelet Guidance Module that generates attention signals to condition a MobileNetV3-based spatial backbone. It employs real-centered learning with exponential moving average prototype updates and drift regularization.

Result: LRD-Net achieves state-of-the-art cross-domain detection accuracy on the DiFF benchmark with only 2.63M parameters (9x fewer than conventional approaches), 8x faster training, and nearly 10x faster inference.

Conclusion: Robust cross-domain face forgery detection can be achieved without sacrificing computational efficiency, making LRD-Net suitable for real-time deployment in mobile authentication systems and resource-constrained environments.

Abstract: The rapid advancement of diffusion-based generative models has made face forgery detection a critical challenge in digital forensics. Current detection methods face two fundamental limitations: poor cross-domain generalization when encountering unseen forgery types, and substantial computational overhead that hinders deployment on resource-constrained devices. We propose LRD-Net (Lightweight Real-centered Detection Network), a novel framework that addresses both challenges simultaneously. Unlike existing dual-branch approaches that process spatial and frequency information independently, LRD-Net adopts a sequential frequency-guided architecture where a lightweight Multi-Scale Wavelet Guidance Module generates attention signals that condition a MobileNetV3-based spatial backbone. This design enables effective exploitation of frequency-domain cues while avoiding the redundancy of parallel feature extraction. Furthermore, LRD-Net employs a real-centered learning strategy with exponential moving average prototype updates and drift regularization, anchoring representations around authentic facial images rather than modeling diverse forgery patterns. Extensive experiments on the DiFF benchmark demonstrate that LRD-Net achieves state-of-the-art cross-domain detection accuracy, consistently outperforming existing methods. Critically, LRD-Net accomplishes this with only 2.63M parameters - approximately 9x fewer than conventional approaches - while achieving over 8x faster training and nearly 10x faster inference. These results demonstrate that robust cross-domain face forgery detection can be achieved without sacrificing computational efficiency, making LRD-Net suitable for real-time deployment in mobile authentication systems and resource-constrained environments.

Method: Uses facial expression recognition through feature point extraction with a modified Harris algorithm that reduces time complexity compared to existing methods. The approach analyzes customer facial expressions in supermarkets/hypermarkets to infer product acceptance.

Result: The modified Harris algorithm proved significantly faster while maintaining nearly the same accuracy for corner point detection needed in facial expression recognition, with reduced time complexity compared to existing algorithms.

Conclusion: The proposed method offers an efficient approach for automated product review through facial expression analysis, with the modified Harris algorithm providing practical speed improvements suitable for real-time retail applications.

Abstract: This paper proposes a method to review public acceptance of products based on their brand by analyzing the facial expression of the customer intending to buy the product from a supermarket or hypermarket. In such cases, facial expression recognition plays a significant role in product review. Here, facial expression detection is performed by extracting feature points using a modified Harris algorithm. The modified Harris algorithm reduced the time complexity of the existing feature extraction Harris Algorithm. A comparison of time complexities of existing algorithms is done with proposed algorithm. The algorithm proved to be significantly faster and nearly accurate for the needed application by reducing the time complexity for corner points detection.

Method: Three-component framework: 1) Reference-Guided Deformable Encoder (RGDE) uses hierarchical reference-driven modulation and multi-scale deformable aggregation for semantic alignment; 2) Uncertainty-Aware Evidential Decoder (UAED) incorporates Dirichlet evidence estimation for uncertainty modeling; 3) Boundary-Aware Refinement Module (BARM) enhances ambiguous boundaries using edge cues and prediction confidence.

Result: EviRCOD achieves state-of-the-art performance on the Ref-COD benchmark and provides well-calibrated uncertainty estimates, demonstrating effectiveness in addressing the three core challenges.

Conclusion: The proposed EviRCOD framework effectively addresses key limitations in Ref-COD through integrated solutions for semantic alignment, uncertainty modeling, and boundary preservation, advancing the field of referring camouflaged object detection.

Abstract: Referring Camouflaged Object Detection (Ref-COD) focuses on segmenting specific camouflaged targets in a query image using category-aligned references. Despite recent advances, existing methods struggle with reference-target semantic alignment, explicit uncertainty modeling, and robust boundary preservation. To address these issues, we propose EviRCOD, an integrated framework consisting of three core components: (1) a Reference-Guided Deformable Encoder (RGDE) that employs hierarchical reference-driven modulation and multi-scale deformable aggregation to inject semantic priors and align cross-scale representations; (2) an Uncertainty-Aware Evidential Decoder (UAED) that incorporates Dirichlet evidence estimation into hierarchical decoding to model uncertainty and propagate confidence across scales; and (3) a Boundary-Aware Refinement Module (BARM) that selectively enhances ambiguous boundaries by exploiting low-level edge cues and prediction confidence. Experiments on the Ref-COD benchmark demonstrate that EviRCOD achieves state-of-the-art detection performance while providing well-calibrated uncertainty estimates. Code is available at: https://github.com/blueecoffee/EviRCOD.

Method: Developed a scalable evaluation framework applying reconstruction and diagnostic AI models in tandem across two tasks (classification, segmentation), three reconstruction approaches (U-Net, GAN, diffusion), and two data types (X-ray, MRI).

Result: Conventional reconstruction metrics poorly track task performance; diagnostic accuracy remains stable even as reconstruction PSNR declines with increasing noise. Fairness metrics show greater variability with reconstruction sometimes amplifying demographic biases (particularly patient sex), though overall additional bias is modest compared to inherent diagnostic model biases.

Conclusion: Holistic performance and fairness assessments are crucial throughout medical imaging workflows, especially as generative reconstruction models are increasingly deployed, despite limited efficacy of bias mitigation strategies adapted from classification literature.

Abstract: AI-based image reconstruction models are increasingly deployed in clinical workflows to improve image quality from noisy data, such as low-dose X-rays or accelerated MRI scans. However, these models are typically evaluated using pixel-level metrics like PSNR, leaving their impact on downstream diagnostic performance and fairness unclear. We introduce a scalable evaluation framework that applies reconstruction and diagnostic AI models in tandem, which we apply to two tasks (classification, segmentation), three reconstruction approaches (U-Net, GAN, diffusion), and two data types (X-ray, MRI) to assess the potential downstream implications of reconstruction. We find that conventional reconstruction metrics poorly track task performance, where diagnostic accuracy remains largely stable even as reconstruction PSNR declines with increasing image noise. Fairness metrics exhibit greater variability, with reconstruction sometimes amplifying demographic biases, particularly regarding patient sex. However, the overall magnitude of this additional bias is modest compared to the inherent biases already present in diagnostic models. To explore potential bias mitigation, we adapt two strategies from classification literature to the reconstruction setting, but observe limited efficacy. Overall, our findings emphasize the importance of holistic performance and fairness assessments throughout the entire medical imaging workflow, especially as generative reconstruction models are increasingly deployed.

Method: Proposes STGV framework that decomposes video features into learnable 2D spatial and 3D temporal hash encodings, plus a key frame canonical initialization strategy for stable Gaussian representation.

Result: Achieves better video representation quality (+0.98 PSNR) compared to other Gaussian-based methods and competitive performance in downstream video tasks.

Conclusion: STGV effectively separates static and dynamic components through spatio-temporal decomposition, enabling better motion pattern learning while maintaining background details.

Abstract: 2D Gaussian Splatting (2DGS) has recently become a promising paradigm for high-quality video representation. However, existing methods employ content-agnostic or spatio-temporal feature overlapping embeddings to predict canonical Gaussian primitive deformations, which entangles static and dynamic components in videos and prevents modeling their distinct properties effectively. These result in inaccurate predictions for spatio-temporal deformations and unsatisfactory representation quality. To address these problems, this paper proposes a Spatio-Temporal hash encoding framework for Gaussian-based Video representation (STGV). By decomposing video features into learnable 2D spatial and 3D temporal hash encodings, STGV effectively facilitates the learning of motion patterns for dynamic components while maintaining background details for static elements.In addition, we construct a more stable and consistent initial canonical Gaussian representation through a key frame canonical initialization strategy, preventing from feature overlapping and a structurally incoherent geometry representation. Experimental results demonstrate that our method attains better video representation quality (+0.98 PSNR) against other Gaussian-based methods and achieves competitive performance in downstream video tasks.

Method: TAMISeg integrates three core components: 1) consistency-aware encoder pretrained with strong perturbations for robust feature extraction, 2) semantic encoder distillation module supervised by frozen DINOv3 teacher for enhanced semantic discriminability, and 3) scale-adaptive decoder for segmenting anatomical structures across different spatial scales.

Result: Experiments on Kvasir-SEG, MosMedData+, and QaTa-COV19 datasets show TAMISeg consistently outperforms existing uni-modal and multi-modal methods in both qualitative and quantitative evaluations.

Conclusion: TAMISeg effectively addresses medical image segmentation challenges by incorporating text guidance and semantic distillation, reducing reliance on pixel-level annotations while improving segmentation performance on complex anatomical structures.

Abstract: Medical image segmentation remains challenging due to limited fine-grained annotations, complex anatomical structures, and image degradation from noise, low contrast, or illumination variation. We propose TAMISeg, a text-guided segmentation framework that incorporates clinical language prompts and semantic distillation as auxiliary semantic cues to enhance visual understanding and reduce reliance on pixel-level fine-grained annotations. TAMISeg integrates three core components: a consistency-aware encoder pretrained with strong perturbations for robust feature extraction, a semantic encoder distillation module with supervision from a frozen DINOv3 teacher to enhance semantic discriminability, and a scale-adaptive decoder that segments anatomical structures across different spatial scales. Experiments on the Kvasir-SEG, MosMedData+, and QaTa-COV19 datasets demonstrate that TAMISeg consistently outperforms existing uni-modal and multi-modal methods in both qualitative and quantitative evaluations. Code will be made publicly available at https://github.com/qczggaoqiang/TAMISeg.

Method: Created ReXSonoVQA benchmark with 514 video clips and 514 questions (249 multiple-choice, 265 free-response) targeting three competencies: Action-Goal Reasoning, Artifact Resolution & Optimization, and Procedure Context & Planning. Evaluated state-of-the-art VLMs (Gemini 3 Pro, Qwen3.5-397B, LLaVA-Video-72B, Seed 2.0 Pro) in zero-shot settings.

Result: VLMs can extract some procedural information from ultrasound videos, but troubleshooting questions remain challenging with minimal gains over text-only baselines, exposing limitations in causal reasoning. The benchmark reveals current models’ deficiencies in understanding dynamic procedural content.

Conclusion: ReXSonoVQA enables development of perception systems for ultrasound training, guidance, and robotic automation by providing a benchmark for evaluating dynamic procedural understanding in vision-language models, highlighting the need for improved causal reasoning capabilities.

Abstract: Ultrasound acquisition requires skilled probe manipulation and real-time adjustments. Vision-language models (VLMs) could enable autonomous ultrasound systems, but existing benchmarks evaluate only static images, not dynamic procedural understanding. We introduce ReXSonoVQA, a video QA benchmark with 514 video clips and 514 questions (249 MCQ, 265 free-response) targeting three competencies: Action-Goal Reasoning, Artifact Resolution & Optimization, and Procedure Context & Planning. Zero-shot evaluation of Gemini 3 Pro, Qwen3.5-397B, LLaVA-Video-72B, and Seed 2.0 Pro shows VLMs can extract some procedural information, but troubleshooting questions remain challenging with minimal gains over text-only baselines, exposing limitations in causal reasoning. ReXSonoVQA enables developing perception systems for ultrasound training, guidance, and robotic automation.

Method: Two main modules: Streamable Vector Quantized Motion Tokenizer (SVQ) for causal discrete motion tokens, and Hierarchical Autoregressive Transformer (HAR) with region-expert transformers and causal spatio-temporal fusion. Uses streamable causal audio encoder and autoregressive masking training for robustness.

Result: Produces coherent, diverse, beat-synchronous full-body gestures in real time, matching or surpassing state-of-the-art offline methods on BEAT2 dataset under true zero look-ahead conditions.

Conclusion: LiveGesture enables real-time, streamable speech-driven gesture generation with region-coordinated motion, addressing limitations of existing offline approaches.

Abstract: We propose LiveGesture, the first fully streamable, speech-driven full-body gesture generation framework that operates with zero look-ahead and supports arbitrary sequence length. Unlike existing co-speech gesture methods, which are designed for offline generation and either treat body regions independently or entangle all joints within a single model, LiveGesture is built from the ground up for causal, region-coordinated motion generation. LiveGesture consists of two main modules: the Streamable Vector Quantized Motion Tokenizer (SVQ) and the Hierarchical Autoregressive Transformer (HAR). The SVQ tokenizer converts the motion sequence of each body region into causal, discrete motion tokens, enabling real-time, streamable token decoding. On top of SVQ, HAR employs region-expert autoregressive (xAR) transformers to model expressive, fine-grained motion dynamics for each body region. A causal spatio-temporal fusion module (xAR Fusion) then captures and integrates correlated motion dynamics across regions. Both xAR and xAR Fusion are conditioned on live, continuously arriving audio signals encoded by a streamable causal audio encoder. To enhance robustness under streaming noise and prediction errors, we introduce autoregressive masking training, which leverages uncertainty-guided token masking and random region masking to expose the model to imperfect, partially erroneous histories during training. Experiments on the BEAT2 dataset demonstrate that LiveGesture produces coherent, diverse, and beat-synchronous full-body gestures in real time, matching or surpassing state-of-the-art offline methods under true zero look-ahead conditions.

Method: Two-stage framework: 1) Semantic Layer Peeling (SLP) - VLM-guided progressive decomposition of images into semantically coherent layers with hybrid inpainting to recover occluded regions; 2) Adaptive Layered Vectorization (ALV) - error-budget-driven adjustment mechanism for efficient vectorization of complete layers.

Result: Extensive experiments show AmodalSVG significantly outperforms prior methods in visual fidelity and enables object-level editing directly in the vector domain, capabilities not supported by existing approaches.

Conclusion: AmodalSVG successfully produces semantically organized and geometrically complete SVG representations that support structural editability, advancing image vectorization beyond modal tracing to amodal reconstruction.

Abstract: We introduce AmodalSVG, a new framework for amodal image vectorization that produces semantically organized and geometrically complete SVG representations from natural images. Existing vectorization methods operate under a modal paradigm: tracing only visible pixels and disregarding occlusion. Consequently, the resulting SVGs are semantically entangled and geometrically incomplete, limiting SVG’s structural editability. In contrast, AmodalSVG reconstructs full object geometries, including occluded regions, into independent, editable vector layers. To achieve this, AmodalSVG reformulates image vectorization as a two-stage framework, performing semantic decoupling and completion in the raster domain to produce amodally complete semantic layers, which are then independently vectorized. In the first stage, we introduce Semantic Layer Peeling (SLP), a VLM-guided strategy that progressively decomposes an image into semantically coherent layers. By hybrid inpainting, SLP recovers complete object appearances under occlusions, enabling explicit semantic decoupling. To vectorize these layers efficiently, we propose Adaptive Layered Vectorization (ALV), which dynamically modulates the primitive budget via an error-budget-driven adjustment mechanism. Extensive experiments demonstrate that AmodalSVG significantly outperforms prior methods in visual fidelity. Moreover, the resulting amodal layers enable object-level editing directly in the vector domain, capabilities not supported by existing vectorization approaches. Code will be released upon acceptance.

Method: Progressive representation-learning framework that sequentially grows the model by activating deeper blocks over time, allowing early layers to encode stable cranial base morphology first before higher layers specialize in discriminating adjacent maturation stages.

Result: The expert-inspired training strategy produces more stable optimization and consistently higher accuracy than standard training, particularly for ambiguous intermediate stages, without changing network architectures or loss functions.

Conclusion: The framework establishes a principled link between expert dental intuition and deep visual representations, enabling robust SOS staging and offering a general strategy for modeling continuous biological processes in medical imaging.

Abstract: Accurate assessment of spheno-occipital synchondrosis (SOS) maturation is a key indicator of craniofacial growth and a critical determinant for orthodontic and surgical timing. However, SOS staging from cone-beam CT (CBCT) relies on subtle, continuously evolving morphological cues, leading to high inter-observer variability and poor reproducibility, especially at transitional fusion stages. We frame SOS assessment as a fine-grained visual recognition problem and propose a progressive representation-learning framework that explicitly mirrors how expert clinicians reason about synchondral fusion: from coarse anatomical structure to increasingly subtle patterns of closure. Rather than training a full-capacity network end-to-end, we sequentially grow the model by activating deeper blocks over time, allowing early layers to first encode stable cranial base morphology before higher-level layers specialize in discriminating adjacent maturation stages. This yields a curriculum over network depth that aligns deep feature learning with the biological continuum of SOS fusion. Extensive experiments across convolutional and transformer-based architectures show that this expert-inspired training strategy produces more stable optimization and consistently higher accuracy than standard training, particularly for ambiguous intermediate stages. Importantly, these gains are achieved without changing network architectures or loss functions, demonstrating that training dynamics alone can substantially improve anatomical representation learning. The proposed framework establishes a principled link between expert dental intuition and deep visual representations, enabling robust, data-efficient SOS staging from CBCT and offering a general strategy for modeling other continuous biological processes in medical imaging.

Method: Proposes an information-theoretic probing framework that jointly analyzes how UMMs encode inputs and generate outputs. Applied to ten representative UMMs, the framework examines entropy trajectories and information flow patterns across modalities.

Result: Reveals pseudo-unification stems from dual divergence: (1) Modality-Asymmetric Encoding where vision and language follow different entropy trajectories, and (2) Pattern-Split Response where text generation shows high-entropy creativity while image synthesis enforces low-entropy fidelity. Only models with unified information flow achieve genuine unification.

Conclusion: Real multimodal synergy requires consistency in information flow, not just shared parameters. Models that unify both encoding and response patterns (e.g., via contextual prediction) achieve stronger reasoning-based text-to-image generation even with fewer parameters.

Abstract: Unified multimodal models (UMMs) were designed to combine the reasoning ability of large language models (LLMs) with the generation capability of vision models. In practice, however, this synergy remains elusive: UMMs fail to transfer LLM-like reasoning to image synthesis and exhibit divergent response behaviors. We term this phenomenon pseudo-unification. Diagnosing its internal causes is important, but existing probing methods either lack model-internal insight or ignore prompt-response dependencies. To address these limitations, we propose an information-theoretic probing framework that jointly analyzes how UMMs encode inputs and generate outputs. Applied to ten representative UMMs, our framework reveals that pseudo-unification stems from a dual divergence: (i) Modality-Asymmetric Encoding, where vision and language follow different entropy trajectories, and (ii) Pattern-Split Response, where text generation exhibits high-entropy creativity while image synthesis enforces low-entropy fidelity. Only models that unify both sides (e.g., via contextual prediction) achieve more genuine unification, enabling stronger reasoning-based text-to-image generation even with fewer parameters. Our work provides the first model-internal probing of unification, demonstrating that real multimodal synergy requires consistency in information flow, not just shared parameters.

Method: DiTTA distills SAM2’s temporal segmentation knowledge into ISS models during a brief initialization phase, complemented by a lightweight temporal fusion module to aggregate cross-frame context. It uses test-time adaptation without annotated videos.

Result: Extensive experiments on VSPW and Cityscapes show DiTTA achieves competitive or superior performance relative to fully-supervised VSS methods, even when adapting with limited partial video snippets (e.g., initial 10%).

Conclusion: DiTTA provides a practical, annotation-free solution for real-world video semantic segmentation by efficiently converting image segmentation models into temporally-aware video models through test-time adaptation.

Abstract: Fully supervised Video Semantic Segmentation (VSS) relies heavily on densely annotated video data, limiting practical applicability. Alternatively, applying pre-trained Image Semantic Segmentation (ISS) models frame-by-frame avoids annotation costs but ignores crucial temporal coherence. Recent foundation models such as SAM2 enable high-quality mask propagation yet remain impractical for direct VSS due to limited semantic understanding and computational overhead. In this paper, we propose DiTTA (Distillation-assisted Test-Time Adaptation), a novel framework that converts an ISS model into a temporally-aware VSS model through efficient test-time adaptation (TTA), without annotated videos. DiTTA distills SAM2’s temporal segmentation knowledge into the ISS model during a brief, single-pass initialization phase, complemented by a lightweight temporal fusion module to aggregate cross-frame context. Crucially, DiTTA achieves robust generalization even when adapting with highly limited partial video snippets (e.g., initial 10%), significantly outperforming zero-shot refinement approaches that repeatedly invoke SAM2 during inference. Extensive experiments on VSPW and Cityscapes demonstrate DiTTA’s effectiveness, achieving competitive or superior performance relative to fully-supervised VSS methods, thus providing a practical and annotation-free solution for real-world VSS tasks.

Method: Proposes FineEdit with multi-level bounding box injection to effectively utilize spatial conditions. Creates FineEdit-1.2M dataset with 1.2M image editing pairs and precise bounding box annotations, and FineEdit-Bench benchmark with 1,000 images across 10 subjects for evaluation.

Result: Outperforms state-of-the-art open-source models (Qwen-Image-Edit, LongCat-Image-Edit) in instruction compliance and background preservation on FineEdit-Bench. Shows superior generalization and robustness on open benchmarks (GEdit and ImgEdit Bench).

Conclusion: Bounding box guidance enables precise localization and background preservation in diffusion-based image editing, with the proposed FineEdit model and datasets advancing region-based editing capabilities.

Abstract: Diffusion-based image editing models have achieved significant progress in real world applications. However, conventional models typically rely on natural language prompts, which often lack the precision required to localize target objects. Consequently, these models struggle to maintain background consistency due to their global image regeneration paradigm. Recognizing that visual cues provide an intuitive means for users to highlight specific areas of interest, we utilize bounding boxes as guidance to explicitly define the editing target. This approach ensures that the diffusion model can accurately localize the target while preserving background consistency. To achieve this, we propose FineEdit, a multi-level bounding box injection method that enables the model to utilize spatial conditions more effectively. To support this high precision guidance, we present FineEdit-1.2M, a large scale, fine-grained dataset comprising 1.2 million image editing pairs with precise bounding box annotations. Furthermore, we construct a comprehensive benchmark, termed FineEdit-Bench, which includes 1,000 images across 10 subjects to effectively evaluate region based editing capabilities. Evaluations on FineEdit-Bench demonstrate that our model significantly outperforms state-of-the-art open-source models (e.g., Qwen-Image-Edit and LongCat-Image-Edit) in instruction compliance and background preservation. Further assessments on open benchmarks (GEdit and ImgEdit Bench) confirm its superior generalization and robustness.

Method: Concatenates multiple responses with separator tokens and applies cross-entropy over their scalar scores, enabling single-forward-pass evaluation. Built on a 4B vision-language backbone with LoRA fine-tuning and lightweight MLP value head. Introduces two new benchmarks: MR²Bench-Image (human-annotated rankings over 8 models) and MR²Bench-Video (94K crowdsourced judgments over 19 video QA models).

Result: Achieves state-of-the-art results on six multimodal reward benchmarks, including the new MR²Bench-Image and MR²Bench-Video. Outperforms existing larger generative and discriminative reward models. When used in RL with GRPO, produces improved policy models that maintain performance across standard benchmarks while substantially improving open-ended generation quality.

Conclusion: The multi-response reward model enables efficient N-way preference learning with significant computational benefits (up to N× speedup) while achieving superior performance on multimodal reward evaluation tasks and improving policy training for open-ended generation.

Abstract: We present a discriminative multimodal reward model that scores all candidate responses in a single forward pass. Conventional discriminative reward models evaluate each response independently, requiring multiple forward passes, one for each potential response. Our approach concatenates multiple responses with separator tokens and applies cross-entropy over their scalar scores, enabling direct comparative reasoning and efficient $N$-way preference learning. The multi-response design also yields up to $N\times$ wall-clock speedup and FLOPs reduction over conventional single-response scoring. To enable $N$-way reward evaluation beyond existing pairwise benchmarks, we construct two new benchmarks: (1) MR$^2$Bench-Image contains human-annotated rankings over responses from 8 diverse models; (2) MR$^2$Bench-Video is a large-scale video-based reward benchmark derived from 94K crowdsourced pairwise human judgments over video question-answering spanning 19 models, denoised via preference graph ensemble. Both benchmarks provide 4-response evaluation variants sampled from the full rankings. Built on a 4B vision-language backbone with LoRA fine-tuning and a lightweight MLP value head, our model achieves state-of-the-art results on six multimodal reward benchmarks, including MR$^2$Bench-Image, MR$^2$Bench-Video, and four other existing benchmarks. Our model outperforms existing larger generative and discriminative reward models. We further demonstrate that our reward model, when used in reinforcement learning with GRPO, produces improved policy models that maintain performance across standard multimodal benchmarks while substantially improving open-ended generation quality, outperforming a single-response discriminative reward model (RM) baseline by a large margin in both training stability and open-ended generation quality.

Method: Extends scene graphs to inter-image comparison, creating Distortion Graphs that represent region-level degradation information. Introduces PandaSet (region-level dataset), PandaBench (benchmark suite), and Panda architecture for generating distortion graphs.

Result: Shows that current MLLMs fail at region-level degradation understanding even with explicit region cues. Training on PandaSet or prompting with DG enables region-wise distortion understanding, creating a new direction for fine-grained image assessment.

Conclusion: Distortion Graphs provide a compact, interpretable structure for representing dense degradation information in image pairs, enabling new capabilities in fine-grained, structured pairwise image assessment.

Abstract: In this work, we introduce a new perspective on comparative image assessment by representing an image pair as a structured composition of its regions. In contrast, existing methods focus on whole image analysis, while implicitly relying on region-level understanding. We extend the intra-image notion of a scene graph to inter-image, and propose a novel task of Distortion Graph (DG). DG treats paired images as a structured topology grounded in regions, and represents dense degradation information such as distortion type, severity, comparison and quality score in a compact interpretable graph structure. To realize the task of learning a distortion graph, we contribute (i) a region-level dataset, PandaSet, (ii) a benchmark suite, PandaBench, with varying region-level difficulty, and (iii) an efficient architecture, Panda, to generate distortion graphs. We demonstrate that PandaBench poses a significant challenge for state-of-the-art multimodal large language models (MLLMs) as they fail to understand region-level degradations even when fed with explicit region cues. We show that training on PandaSet or prompting with DG elicits region-wise distortion understanding, opening a new direction for fine-grained, structured pairwise image assessment.

Method: Proposed hybrid method combining handcrafted features (Local Binary Pattern, Histogram of Gradients, Gabor Filters) with deep features from DenseNet-169. Features are concatenated and fed to three classifiers: SVM, XGBoost, and LightGBM.

Result: DenseNet-169 + Gabor (SVM) achieved highest accuracy of 99.17% on augmented dataset, outperforming other systems. The hybrid framework offers better defect-detection accuracy, resistance, and flexibility for real-life PV panel monitoring.

Conclusion: The proposed hybrid framework provides an effective automated solution for solar panel defect detection, combining traditional computer vision techniques with deep learning for superior performance in real-world monitoring applications.

Abstract: To ensure energy efficiency and reliable operations, it is essential to monitor solar panels in generation plants to detect defects. It is quite labor-intensive, time consuming and costly to manually monitor large-scale solar plants and those installed in remote areas. Manual inspection may also be susceptible to human errors. Consequently, it is necessary to create an automated, intelligent defect-detection system, that ensures continuous monitoring, early fault detection, and maximum power generation. We proposed a novel hybrid method for defect detection in SOLAR plates by combining both handcrafted and deep learning features. Local Binary Pattern (LBP), Histogram of Gradients (HoG) and Gabor Filters were used for the extraction of handcrafted features. Deep features extracted by leveraging the use of DenseNet-169. Both handcrafted and deep features were concatenated and then fed to three distinct types of classifiers, including Support Vector Machines (SVM), Extreme Gradient Boost (XGBoost) and Light Gradient-Boosting Machine (LGBM). Experimental results evaluated on the augmented dataset show the superior performance, especially DenseNet-169 + Gabor (SVM), had the highest scores with 99.17% accuracy which was higher than all the other systems. In general, the proposed hybrid framework offers better defect-detection accuracy, resistance, and flexibility that has a solid basis on the real-life use of the automated PV panels monitoring system.

Method: Two-stage framework: 1) Frozen algorithm-based preprocessing that normalizes input distribution by providing complementary brightness-corrected views, 2) Compact U-Net built entirely from depthwise-separable convolutions that focuses on residual color correction.

Result: Achieved 4th place in CVPR 2026 NTIRE Efficient Low-Light Image Enhancement Challenge, demonstrating competitive perceptual quality with significantly fewer parameters than existing methods.

Conclusion: The proposed lightweight framework effectively balances efficiency and quality for low-light image enhancement, with extended benchmarks and ablations confirming its general effectiveness.

Abstract: We present a lightweight two-stage framework for low-light image enhancement (LLIE) that achieves competitive perceptual quality with significantly fewer parameters than existing methods. Our approach combines frozen algorithm-based preprocessing with a compact U-Net built entirely from depthwise-separable convolutions. The preprocessing normalizes the input distribution by providing complementary brightness-corrected views, enabling the trainable network to focus on residual color correction. Our method achieved 4th place in the CVPR 2026 NTIRE Efficient Low-Light Image Enhancement Challenge. We further provide extended benchmarks and ablations to demonstrate the general effectiveness of our methods.

Method: Investigated generalizability of SSL models (DINO-based ViT backbones) pretrained on ImageNet-1k and HPA FOV datasets. Evaluated embeddings on OpenCell dataset with/without fine-tuning, tested two channel-mismatch strategies, varied fine-tuning data fractions, and analyzed single-cell embeddings on labeled OpenCell subset.

Result: HPA FOV-pretrained model achieved highest zero-shot performance (macro F1 0.822 ± 0.007). Fine-tuning further improved to 0.860 ± 0.013. At single-cell level, HPA single-cell-pretrained model achieved highest k-nearest neighbor performance across all neighborhood sizes (macro F1 ≥ 0.796).

Conclusion: SSL methods like DINO, when pretrained on large domain-relevant datasets, enable effective use of deep learning features for fine-tuning on small, task-specific microscopy datasets.

Abstract: Background: Task-specific microscopy datasets are often small, making it difficult to train deep learning models that learn robust features. While self-supervised learning (SSL) has shown promise through pretraining on large, domain-specific datasets, generalizability across datasets with differing staining protocols and channel configurations remains underexplored. We investigated the generalizability of SSL models pretrained on ImageNet-1k and HPA FOV, evaluating their embeddings on OpenCell with and without fine-tuning, two channel-mismatch strategies, and varying fine-tuning data fractions. We additionally analyzed single-cell embeddings on a labeled OpenCell subset. Result: DINO-based ViT backbones pretrained on HPA FOV or ImageNet-1k transfer well to OpenCell even without fine-tuning. The HPA FOV-pretrained model achieved the highest zero-shot performance (macro $F_1$ 0.822 $\pm$ 0.007). Fine-tuning further improved performance to 0.860 $\pm$ 0.013. At the single-cell level, the HPA single-cell-pretrained model achieved the highest k-nearest neighbor performance across all neighborhood sizes (macro $F_1$ $\geq$ 0.796). Conclusion: SSL methods like DINO, pretrained on large domain-relevant datasets, enable effective use of deep learning features for fine-tuning on small, task-specific microscopy datasets.

Method: Deep learning-based end-to-end transceiver design that integrates 3D Gaussian Splatting directly into training, jointly optimizing communication modules via 3DGS rendering loss, enabling sparse pilot scheme

Result: Significantly outperforms existing baselines, delivering superior transmission performance and accurate 3D scene reconstructions on real-world aerial image datasets

Conclusion: Proposed E2E design effectively balances efficiency and reliability for 3D scene reconstruction in low-altitude networks by integrating 3DGS into communication optimization

Abstract: Large-scale three-dimensional (3D) scene reconstruction in low-altitude intelligent networks (LAIN) demands highly efficient wireless image transmission. However, existing schemes struggle to balance severe pilot overhead with the transmission accuracy required to maintain reconstruction fidelity. To strike a balance between efficiency and reliability, this paper proposes a novel deep learning-based end-to-end (E2E) transceiver design that integrates 3D Gaussian Splatting (3DGS) directly into the training process. By jointly optimizing the communication modules via the combined 3DGS rendering loss, our approach explicitly improves scene recovery quality. Furthermore, this task-driven framework enables the use of a sparse pilot scheme, significantly reducing transmission overhead while maintaining robust image recovery under low-altitude channel conditions. Extensive experiments on real-world aerial image datasets demonstrate that the proposed E2E design significantly outperforms existing baselines, delivering superior transmission performance and accurate 3D scene reconstructions.

Method: Proposes MMR-AD benchmark for MLLM-based AD research and Anomaly-R1 baseline model - a reasoning-based AD model that learns from Chain-of-Thought data in MMR-AD and is enhanced by reinforcement learning.

Result: Current state-of-the-art generalist MLLMs perform poorly on industrial AD requirements, while Anomaly-R1 achieves remarkable improvements over generalist MLLMs in both anomaly detection and localization.

Conclusion: MMR-AD enables MLLM-based general AD research, revealing current MLLM limitations and demonstrating that specialized training (like Anomaly-R1) can significantly improve AD performance for industrial applications.

Abstract: In the progress of industrial anomaly detection, general anomaly detection (GAD) is an emerging trend and also the ultimate goal. Unlike the conventional single- and multi-class AD, general AD aims to train a general AD model that can directly detect anomalies in diverse novel classes without any retraining or fine-tuning on the target data. Recently, Multimodal Large Language Models (MLLMs) have shown great promise in achieving general anomaly detection due to their revolutionary visual understanding and language reasoning capabilities. However, MLLM’s general AD ability remains underexplored due to: (1) MLLMs are pretrained on amounts of data sourced from the Web, these data still have significant gaps with the data in AD scenarios. Moreover, the image-text pairs during pretraining are also not specifically for AD tasks. (2) The current mainstream AD datasets are image-based and not yet suitable for post-training MLLMs. To facilitate MLLM-based general AD research, we present MMR-AD, which is a comprehensive benchmark for both training and evaluating MLLM-based AD models. With MMR-AD, we reveal that the AD performance of current SOTA generalist MLLMs still falls far behind the industrial requirements. Based on MMR-AD, we also propose a baseline model, Anomaly-R1, which is a reasoning-based AD model that learns from the CoT data in MMR-AD and is further enhanced by reinforcement learning. Extensive experiments show that our Anomaly-R1 achieves remarkable improvements over generalist MLLMs in both anomaly detection and localization.

Method: Proposes Energy-oriented diffusion Bridge (E-Bridge) with: 1) Novel bridge process over shorter time horizon starting from entropy-regularized point mixing degraded image and Gaussian noise, 2) Single-step mapping function inspired by consistency models optimized via continuous-time consistency objective, 3) Tunable trajectory length as task-adaptive knob for balancing information preservation vs generative power.

Result: Achieves state-of-the-art performance across various image restoration tasks (denoising, super-resolution) while enabling high-quality recovery with single or fewer sampling steps.

Conclusion: E-Bridge provides an efficient diffusion bridge framework that reduces trajectory energy, enables fast sampling, and adapts to different degradation levels through tunable trajectory length, advancing practical image restoration applications.

Abstract: Diffusion bridge models have shown great promise in image restoration by explicitly connecting clean and degraded image distributions. However, they often rely on complex and high-cost trajectories, which limit both sampling efficiency and final restoration quality. To address this, we propose an Energy-oriented diffusion Bridge (E-Bridge) framework to approximate a set of low-cost manifold geodesic trajectories to boost the performance of the proposed method. We achieve this by designing a novel bridge process that evolves over a shorter time horizon and makes the reverse process start from an entropy-regularized point that mixes the degraded image and Gaussian noise, which theoretically reduces the required trajectory energy. To solve this process efficiently, we draw inspiration from consistency models to learn a single-step mapping function, optimized via a continuous-time consistency objective tailored for our trajectory, so as to analytically map any state on the trajectory to the target image. Notably, the trajectory length in our framework becomes a tunable task-adaptive knob, allowing the model to adaptively balance information preservation against generative power for tasks of varying degradation, such as denoising versus super-resolution. Extensive experiments demonstrate that our E-Bridge achieves state-of-the-art performance across various image restoration tasks while enabling high-quality recovery with a single or fewer sampling steps. Our project page is https://jinnh.github.io/E-Bridge/.

Method: Four specialized agents (Design, Generation, Assembly, Review) with key insight to predict assembly relationships during design stage using a Connector that defines attachment points and joint parameters. Includes validation steps, cross-stage rollback mechanism, and self-evolving experience store for continuous improvement.

Result: Extensive evaluations on three datasets (ArtiCAD-Bench, CADPrompt, ACD) validate effectiveness. Demonstrates applicability in requirement-driven conceptual design, physical prototyping, and generation of embodied AI training assets through URDF export.

Conclusion: ArtiCAD successfully addresses the unexplored problem of generating editable, articulated CAD assemblies from high-level descriptions, offering a robust multi-agent approach with continuous learning capabilities.

Abstract: Parametric Computer-Aided Design (CAD) of articulated assemblies is essential for product development, yet generating these multi-part, movable models from high-level descriptions remains unexplored. To address this, we propose ArtiCAD, the first training-free multi-agent system capable of generating editable, articulated CAD assemblies directly from text or images. Our system divides this complex task among four specialized agents: Design, Generation, Assembly, and Review. One of our key insights is to predict assembly relationships during the initial design stage rather than the assembly stage. By utilizing a Connector that explicitly defines attachment points and joint parameters, ArtiCAD determines these relationships before geometry generation, effectively bypassing the limited spatial reasoning capabilities of current LLMs and VLMs. To further ensure high-quality outputs, we introduce validation steps in the generation and assembly stages, accompanied by a cross-stage rollback mechanism that accurately isolates and corrects design- and code-level errors. Additionally, a self-evolving experience store accumulates design knowledge to continuously improve performance on future tasks. Extensive evaluations on three datasets (ArtiCAD-Bench, CADPrompt, and ACD) validate the effectiveness of our approach. We further demonstrate the applicability of ArtiCAD in requirement-driven conceptual design, physical prototyping, and the generation of embodied AI training assets through URDF export.

Method: Uses dynamic scene regions (moving elements) as supervisory signal, learning a dynamic 2D Gaussian Splatting representation with novel constraints that encourage dynamic regions to deform while keeping static regions stable.

Result: Improves LPIPS by 23% for albedo estimation and 15% for scene relighting compared to next-best baseline, with a new synthetic benchmark for dynamic inverse rendering evaluation.

Conclusion: Dynamic elements provide crucial cues for disentangling material and illumination, enabling more accurate inverse rendering in arbitrary dynamic scenes with challenging lighting.

Abstract: In 3D reconstruction, the problem of inverse rendering, namely recovering the illumination of the scene and the material properties, is fundamental. Existing Gaussian Splatting-based methods primarily target static scenes and often assume simplified or moderate lighting to avoid entangling shadows with surface appearance. This limits their ability to accurately separate lighting effects from material properties, particularly in real-world conditions. We address this limitation by leveraging dynamic elements - regions of the scene that undergo motion - as a supervisory signal for inverse rendering. Motion reveals the same surfaces under varying lighting conditions, providing stronger cues for disentangling material and illumination. This thesis is supported by our experimental results which show we improve LPIPS by 23% for albedo estimation and by 15% for scene relighting relative to next-best baseline. To this end, we introduce LumiMotion, the first Gaussian-based approach that leverages dynamics for inverse rendering and operates in arbitrary dynamic scenes. Our method learns a dynamic 2D Gaussian Splatting representation that employs a set of novel constraints which encourage the dynamic regions of the scene to deform, while keeping static regions stable. As we demonstrate, this separation is crucial for correct optimization of the albedo. Finally, we release a new synthetic benchmark comprising five scenes under four lighting conditions, each in both static and dynamic variants, for the first time enabling systematic evaluation of inverse rendering methods in dynamic environments and challenging lighting. Link to project page: https://joaxkal.github.io/LumiMotion/

Method: Created TraversalBench with single continuous polylines, start markers, and vertex markers. Systematically balanced path-structural factors: self-intersection count, tortuosity, vertex count, and nearby confounding lines. Minimized reliance on OCR, world knowledge, and open-ended planning. Conducted first-crossing analysis and auxiliary reading-order tests.

Result: Self-intersections are the dominant source of difficulty, with errors sharply localized at crossing points. Nearby confounding lines cause weaker persistent degradation. Models show consistent preference for left-to-right reading layouts. Performance drops steeply when models must resolve correct continuation at crossings.

Conclusion: TraversalBench serves as a controlled diagnostic for path-faithful visual reasoning and a testbed for studying multimodal spatial reasoning under ambiguity and clutter. It contributes to the limited area of sustained visual grounding benchmarks for VLMs.

Abstract: Vision-language models (VLMs) perform strongly on many multimodal benchmarks. However, the ability to follow complex visual paths – a task that human observers typically find straightforward – remains under-tested. We introduce TraversalBench, a controlled benchmark for exact visual path traversal. Each instance contains a single continuous polyline, a unique start marker, and markers placed at path vertices; the task is to recover the exact ordered sequence encountered when traversing the path from start to finish. The benchmark explicitly balances key path-structural factors including self-intersection count, tortuosity, vertex count, and nearby confounding lines, while minimizing reliance on OCR, world knowledge, and open-ended planning. We find that self-intersections are the dominant source of difficulty. A first-crossing analysis shows that errors are sharply localized: performance is relatively stable immediately before the first crossing, then drops steeply when the model must resolve the correct continuation. By contrast, nearby confounding lines produce a weaker persistent degradation that compounds with repeated exposure. These analyses make TraversalBench a useful diagnostic for identifying whether models suffer from human-like failures or other breakdowns in sustained visual processing. An auxiliary reading-order benchmark further reveals a consistent preference for layouts compatible with left-to-right serialization, while not explaining away the main effects of path complexity. Together, these results position TraversalBench as a controlled diagnostic of path-faithful visual reasoning and as a useful testbed for studying multimodal spatial reasoning under ambiguity, clutter, and distractor structure. More broadly, we position TraversalBench as a contribution to the still-limited area of sustained visual grounding benchmarks for VLMs.

Method: Proposes EmissionGen baseline for emission texture generation, constructs Objaverse-Emission dataset with 40k 3D assets containing emission materials, and defines evaluation metrics for the task.

Result: Demonstrates significant potential for industrial applications, enabling 3D objects to faithfully reproduce emission materials from reference images.

Conclusion: Introduces a novel emission texture generation task with dataset, baseline method, and evaluation metrics to address limitations of current 3D texture generation in creating realistic emission effects.

Abstract: 3D texture generation is receiving increasing attention, as it enables the creation of realistic and aesthetic texture materials for untextured 3D meshes. However, existing 3D texture generation methods are limited to producing only a few types of non-emissive PBR materials (e.g., albedo, metallic maps and roughness maps), making them difficult to replicate highly popular styles, such as cyberpunk, failing to achieve effects like realistic LED emissions. To address this limitation, we propose a novel task, emission texture generation, which enables the synthesized 3D objects to faithfully reproduce the emission materials from input reference images. Our key contributions include: first, We construct the Objaverse-Emission dataset, the first dataset that contains 40k 3D assets with high-quality emission materials. Second, we propose EmissionGen, a novel baseline for the emission texture generation task. Third, we define detailed evaluation metrics for the emission texture generation task. Our results demonstrate significant potential for future industrial applications. Dataset will be available at https://github.com/yx345kw/EmissionGen.

Method: PLOVIS uses Open-Vocabulary Image Segmentation (OVIS) models as pseudo label generators, creating 2D images directly from 3D point clouds. It employs two-stage filtering (removing low-confidence then likely incorrect labels) and a class-balanced memory bank to handle noise and class imbalance in pseudo labels.

Result: Experiments on ScanNet, S3DIS, Toronto3D, and Semantic3D datasets under data-scarce conditions (few tens of scenes with <100 annotated points each) show PLOVIS consistently outperforms standard fine-tuning and state-of-the-art weakly supervised methods.

Conclusion: PLOVIS effectively addresses the three concurrent forms of data insufficiency in 3D point cloud segmentation through open-vocabulary image segmentation-based pseudo-labeling, demonstrating superior performance in realistic data-scarce scenarios.

Abstract: Semantic segmentation of 3D point cloud scenes is a crucial task for various applications. In real-world scenarios, training segmentation models often faces three concurrent forms of data insufficiency: scarcity of training scenes, scarcity of point-level annotations, and absence of 2D image sequences from which point clouds were reconstructed. Existing data-efficient algorithms typically address only one or two of these challenges, leaving the joint treatment of all three unexplored. This paper proposes a data-efficient training framework specifically designed to address the three forms of data insufficiency. Our proposed algorithm, called Point pseudo-Labeling via Open-Vocabulary Image Segmentation (PLOVIS), leverages an Open-Vocabulary Image Segmentation (OVIS) model as a pseudo label generator to compensate for the lack of training data. PLOVIS creates 2D images for pseudo-labeling directly from training 3D point clouds, eliminating the need for 2D image sequences. To mitigate the inherent noise and class imbalance in pseudo labels, we introduce a two-stage filtering of pseudo labels combined with a class-balanced memory bank for effective training. The two-stage filtering mechanism first removes low-confidence pseudo labels, then discards likely incorrect pseudo labels, thereby enhancing the quality of pseudo labels. Experiments on four benchmark datasets, i.e., ScanNet, S3DIS, Toronto3D, and Semantic3D, under realistic data-scarce conditions (a few tens of training 3D scenes, each annotated with only <100 3D points) demonstrate that PLOVIS consistently outperforms existing methods including standard fine-tuning strategies and state-of-the-art weakly supervised learning algorithms. Code will be made publicly available.

Method: Use bGPT, a byte-level transformer designed for next-byte prediction, to generate likely fragment continuations from partial BMP image data. Evaluate predictions using cosine similarity, SSIM, chi-square distance, and Jensen-Shannon divergence.

Result: Generative models can effectively predict byte-level patterns to support fragment matching in unallocated disk space, demonstrating the feasibility of generative approaches for multimedia carving.

Conclusion: Generative models like bGPT offer promising capabilities for multimedia file carving by predicting missing data patterns, advancing beyond traditional signature-based and discriminative approaches in digital forensics.

Abstract: Digital forensic investigations often face significant challenges when recovering fragmented multimedia files that lack file system metadata. While traditional file carving relies on signatures and discriminative deep learning models for fragment classification, these methods cannot reconstruct or predict missing data. We propose a generative approach to multimedia carving using bGPT, a byte-level transformer designed for next-byte prediction. By feeding partial BMP image data into the model, we simulate the generation of likely fragment continuations. We evaluate the fidelity of these predictions using different metrics, namely, cosine similarity, structural similarity index (SSIM), chi-square distance, and Jensen-Shannon divergence (JSD). Our findings demonstrate that generative models can effectively predict byte-level patterns to support fragment matching in unallocated disk space.

Method: Proposes a Gaussian-process-guided framework that estimates sparse GP posterior statistics over compact spatio-temporal descriptors to explicitly characterize local degradation response and uncertainty, guiding adaptive temporal-detail fusion. Includes structure-color collaborative reconstruction head, heteroscedastic objective, and overlap-tiled inference for memory-bounded 4K deployment.

Result: Achieves competitive restoration fidelity with substantially fewer parameters than existing methods, enables real-time full-resolution 4K inference with significant speedup over closest quality competitor, and maintains robust performance across multi-level mixed-degradation schedules. Generalizes to real sensor noise and improves downstream object detection.

Conclusion: UHD-GPGNet effectively addresses the joint requirements of UHD video denoising through explicit GP guidance and efficient architecture design, enabling practical real-time 4K deployment with strong generalization to real-world conditions.

Abstract: Ultra-high-definition (UHD) video denoising requires simultaneously suppressing complex spatio-temporal degradations, preserving fine textures and chromatic stability, and maintaining efficient full-resolution 4K deployment. In this paper, we propose UHD-GPGNet, a Gaussian-process-guided local spatio-temporal denoising framework that addresses these requirements jointly. Rather than relying on implicit feature learning alone, the method estimates sparse GP posterior statistics over compact spatio-temporal descriptors to explicitly characterize local degradation response and uncertainty, which then guide adaptive temporal-detail fusion. A structure-color collaborative reconstruction head decouples luminance, chroma, and high-frequency correction, while a heteroscedastic objective and overlap-tiled inference further stabilize optimization and enable memory-bounded 4K deployment. Experiments on UVG and RealisVideo-4K show that UHD-GPGNet achieves competitive restoration fidelity with substantially fewer parameters than existing methods, enables real-time full-resolution 4K inference with significant speedup over the closest quality competitor, and maintains robust performance across a multi-level mixed-degradation schedule.A real-world study on phone-captured 4K video further confirms that the model, trained entirely on synthetic degradation, generalizes to unseen real sensor noise and improves downstream object detection under challenging conditions.

Method: TTSP (Test-Time Scaling over Perception) treats perception as scalable inference: 1) generates multiple exploratory perception traces, 2) filters unreliable traces using entropy-based confidence estimation, 3) distills validated observations into structured knowledge, and 4) iteratively refines subsequent exploration toward unresolved uncertainty.

Result: Extensive experiments on high-resolution and general multimodal reasoning benchmarks show TTSP consistently outperforms strong baselines across backbone sizes, while exhibiting favorable scalability and token efficiency.

Conclusion: Scaling perception at test time is a promising direction for robust multimodal reasoning under perceptual uncertainty, addressing the fundamental Grounding Paradox in current MLLMs.

Abstract: Recent multimodal large language models (MLLMs) have begun to support Thinking with Images by invoking visual tools such as zooming and cropping during inference. Yet these systems remain brittle in fine-grained visual reasoning because they must decide where to look before they have access to the evidence needed to make that decision correctly. We identify this circular dependency as the Grounding Paradox. To address it, we propose Test-Time Scaling over Perception (TTSP), a framework that treats perception itself as a scalable inference process. TTSP generates multiple exploratory perception traces, filters unreliable traces using entropy-based confidence estimation, distills validated observations into structured knowledge, and iteratively refines subsequent exploration toward unresolved uncertainty. Extensive experiments on high-resolution and general multimodal reasoning benchmarks show that TTSP consistently outperforms strong baselines across backbone sizes, while also exhibiting favorable scalability and token efficiency. Our results suggest that scaling perception at test time is a promising direction for robust multimodal reasoning under perceptual uncertainty.

Method: Proposes a 4-stage pipeline: 2D part segmentation, 3D reconstruction, articulation estimation, and function template inference. Uses function templates as structured computational representations that can be compiled into executable code across simulation platforms.

Result: Introduces a dataset of 271 egocentric videos with challenging real-world interactions, paired with 3D geometry, 2D/3D segmentation, articulation and function template annotations. Benchmarking shows the task is challenging for off-the-shelf methods.

Conclusion: EgoFun3D provides a comprehensive framework for modeling interactive 3D objects from egocentric videos, highlighting the difficulty of the task and opening avenues for future research in functional understanding of objects from video.

Abstract: We present EgoFun3D, a coordinated task formulation, dataset, and benchmark for modeling interactive 3D objects from egocentric videos. Interactive objects are of high interest for embodied AI but scarce, making modeling from readily available real-world videos valuable. Our task focuses on obtaining simulation-ready interactive 3D objects from egocentric video input. While prior work largely focuses on articulations, we capture general cross-part functional mappings (e.g., rotation of stove knob controls stove burner temperature) through function templates, a structured computational representation. Function templates enable precise evaluation and direct compilation into executable code across simulation platforms. To enable comprehensive benchmarking, we introduce a dataset of 271 egocentric videos featuring challenging real-world interactions with paired 3D geometry, segmentation over 2D and 3D, articulation and function template annotations. To tackle the task, we propose a 4-stage pipeline consisting of: 2D part segmentation, reconstruction, articulation estimation, and function template inference. Comprehensive benchmarking shows that the task is challenging for off-the-shelf methods, highlighting avenues for future work.

Method: Proposes an agentic label harmonization workflow that uses a vision-language model to reconcile both category semantics and bounding box granularity across heterogeneous annotation sources before training. Applied to document layout detection where annotation standards vary widely.

Result: Without harmonization, naïve mixed-dataset fine-tuning degrades RT-DETRv2 detector performance (table TEDS drops from 0.800 to 0.750). With harmonization applied to two corpora with partially overlapping taxonomies, detection F-score improves from 0.860 to 0.883, table TEDS improves to 0.814, and bounding box overlap drops from 0.043 to 0.016. Representation analysis shows harmonized training produces more compact and separable embeddings.

Conclusion: Annotation inconsistency distorts learned feature space in multimodal models, and resolving it before training through vision-language model-based harmonization restores representation structure and improves performance across multiple metrics.

Abstract: Fine-tuning object detection (OD) models on combined datasets assumes annotation compatibility, yet datasets often encode conflicting spatial definitions for semantically equivalent categories. We propose an agentic label harmonization workflow that uses a vision-language model to reconcile both category semantics and bounding box granularity across heterogeneous sources before training. We evaluate on document layout detection as a challenging case study, where annotation standards vary widely across corpora. Without harmonization, naïve mixed-dataset fine-tuning degrades a pretrained RT-DETRv2 detector: on SCORE-Bench, which measures how accurately the full document conversion pipeline reproduces ground-truth structure, table TEDS drops from 0.800 to 0.750. Applied to two corpora whose 16 and 10 category taxonomies share only 8 direct correspondences, harmonization yields consistent gains across content fidelity, table structure, and spatial consistency: detection F-score improves from 0.860 to 0.883, table TEDS improves to 0.814, and mean bounding box overlap drops from 0.043 to 0.016. Representation analysis further shows that harmonized training produces more compact and separable post-decoder embeddings, confirming that annotation inconsistency distorts the learned feature space and that resolving it before training restores representation structure.

Method: ReSpinQuant leverages offline activation rotation fusion and matches basis using efficient residual subspace rotation, reconciling high expressivity of layer-wise adaptation with negligible inference overhead.

Result: Extensive experiments on W4A4 and W3A3 quantization show ReSpinQuant achieves state-of-the-art performance, outperforming global rotation methods and matching the accuracy of computationally expensive layer-wise methods with minimal overhead.

Conclusion: ReSpinQuant successfully resolves the efficiency-accuracy trade-off in rotation-based PTQ for LLMs, providing a practical solution for efficient quantization without sacrificing model quality.

Abstract: Rotation-based Post-Training Quantization (PTQ) has emerged as a promising solution for mitigating activation outliers in the quantization of Large Language Models (LLMs). Global rotation methods achieve inference efficiency by fusing activation rotations into attention and FFN blocks, but suffer from limited expressivity as they are constrained to use a single learnable rotation matrix across all layers. To tackle this, layer-wise transformation methods emerged, achieving superior accuracy through localized adaptation. However, layer-wise methods cannot fuse activation rotation matrices into weights, requiring online computations and causing significant overhead. In this paper, we propose ReSpinQuant, a quantization framework that resolves such overhead by leveraging offline activation rotation fusion and matching basis using efficient residual subspace rotation. This design reconciles the high expressivity of layer-wise adaptation with only negligible inference overhead. Extensive experiments on W4A4 and W3A3 quantization demonstrate that ReSpinQuant achieves state-of-the-art performance, outperforming global rotation methods and matching the accuracy of computationally expensive layer-wise methods with minimal overhead.

Method: MapATM leverages historical actor (vehicle) trajectory information as structural priors for road geometry to enhance lane detection performance through a novel deep neural network architecture.

Result: Achieves 4.6 AP improvement for lane dividers and 2.6 mAP increase on NuScenes dataset (10.1% and 6.1% relative improvements), with stable map reconstruction across diverse driving scenarios.

Conclusion: MapATM demonstrates practical value for autonomous driving by using vehicle trajectories to improve lane detection robustness in challenging conditions, enabling more reliable HD mapping.

Abstract: High-definition (HD) mapping tasks, which perform lane detections and predictions, are extremely challenging due to non-ideal conditions such as view occlusions, distant lane visibility, and adverse weather conditions. Those conditions often result in compromised lane detection accuracy and reduced reliability within autonomous driving systems. To address these challenges, we introduce MapATM, a novel deep neural network that effectively leverages historical actor trajectory information to improve lane detection accuracy, where actors refer to moving vehicles. By utilizing actor trajectories as structural priors for road geometry, MapATM achieves substantial performance enhancements, notably increasing AP by 4.6 for lane dividers and mAP by 2.6 on the challenging NuScenes dataset, representing relative improvements of 10.1% and 6.1%, respectively, compared to strong baseline methods. Extensive qualitative evaluations further demonstrate MapATM’s capability to consistently maintain stable and robust map reconstruction across diverse and complex driving scenarios, underscoring its practical value for autonomous driving applications.

Method: Reference-guided prompting: for each test frame, select a reference frame from earlier in the same video to establish visual baseline. Prompt VLM with reference/test pairs, then aggregate noisy frame predictions into stable video-level decisions without fine-tuning. Also introduced RefGlitch synthetic dataset for controlled analysis.

Result: Experiments across 5 VLMs and 3 datasets (synthetic + real-world) show reference guidance consistently strengthens frame-level detection and reliably transfers to stronger video-level triage under realistic QA conditions.

Conclusion: RESP provides a practical multi-frame framework for gameplay glitch detection that leverages reference-guided prompting to improve robustness against scene variation, enabling more scalable automated quality assurance.

Abstract: Visual glitches in video games degrade player experience and perceived quality, yet manual quality assurance cannot scale to the growing test surface of modern game development. Prior automation efforts, particularly those using vision-language models (VLMs), largely operate on single frames or rely on limited video-level baselines that struggle under realistic scene variation, making robust video-level glitch detection challenging. We present RESP, a practical multi-frame framework for gameplay glitch detection with VLMs. Our key idea is reference-guided prompting: for each test frame, we select a reference frame from earlier in the same video, establishing a visual baseline and reframing detection as within-video comparison rather than isolated classification. RESP sequentially prompts the VLM with reference/test pairs and aggregates noisy frame predictions into a stable video-level decision without fine-tuning the VLM. To enable controlled analysis of reference effects, we introduce RefGlitch, a synthetic dataset of manually labeled reference/test frame pairs with balanced coverage across five glitch types. Experiments across five VLMs and three datasets (one synthetic, two real-world) show that reference guidance consistently strengthens frame-level detection and that the improved frame-level evidence reliably transfers to stronger video-level triage under realistic QA conditions. Code and data are available at: \href{https://github.com/PipiZong/RESP_code.git}{this https URL}.

Method: Proposes FlowCoMotion with token-latent coupling: 1) Latent branch uses multi-view distillation to regularize continuous latent space, 2) Token branch uses discrete temporal resolution quantization for high-level semantic cues, 3) Token-latent coupling network combines both representations, 4) Velocity field prediction based on text conditions, 5) ODE solver integrates velocity field from simple prior to target motion state.

Result: Achieves competitive performance on text-to-motion benchmarks including HumanML3D and SnapMoGen, demonstrating effectiveness in generating motions with both semantic alignment and fine-grained details.

Conclusion: FlowCoMotion successfully unifies continuous and discrete motion representations through token-latent coupling, enabling text-to-motion generation that captures both semantic content and high-fidelity motion details, outperforming existing approaches.

Abstract: Text-to-motion generation is driven by learning motion representations for semantic alignment with language. Existing methods rely on either continuous or discrete motion representations. However, continuous representations entangle semantics with dynamics, while discrete representations lose fine-grained motion details. In this context, we propose FlowCoMotion, a novel motion generation framework that unifies both treatments from a modeling perspective. Specifically, FlowCoMotion employs token-latent coupling to capture both semantic content and high-fidelity motion details. In the latent branch, we apply multi-view distillation to regularize the continuous latent space, while in the token branch we use discrete temporal resolution quantization to extract high-level semantic cues. The motion latent is then obtained by combining the representations from the two branches through a token-latent coupling network. Subsequently, a velocity field is predicted based on the textual conditions. An ODE solver integrates this velocity field from a simple prior, thereby guiding the sample to the potential state of the target motion. Extensive experiments show that FlowCoMotion achieves competitive performance on text-to-motion benchmarks, including HumanML3D and SnapMoGen.

Method: Introduces a novel regularizer that guides tokenizers to mimic the hidden state dynamics of state-space models (SSMs), transferring their frequency awareness property to latent features. This is grounded in theoretical analysis of SSMs and enforces encoding of fine spatial structures and frequency-domain cues into compact latent features.

Result: Experiments show the method improves generation quality in diffusion models while incurring only minimal loss in reconstruction fidelity, demonstrating more effective use of representation capacity and improved generative modelability.

Conclusion: The proposed regularizer successfully creates latent spaces that are both compact and generation-friendly by leveraging state-space model dynamics, offering a promising approach for improving vision models through better latent space design.

Abstract: Image tokenizers are central to modern vision models as they often operate in latent spaces. An ideal latent space must be simultaneously compact and generation-friendly: it should capture image’s essential content compactly while remaining easy to model with generative approaches. In this work, we introduce a novel regularizer to align latent spaces with these two objectives. The key idea is to guide tokenizers to mimic the hidden state dynamics of state-space models (SSMs), thereby transferring their critical property, frequency awareness, to latent features. Grounded in a theoretical analysis of SSMs, our regularizer enforces encoding of fine spatial structures and frequency-domain cues into compact latent features; leading to more effective use of representation capacity and improved generative modelability. Experiments demonstrate that our method improves generation quality in diffusion models while incurring only minimal loss in reconstruction fidelity.

Method: Proposes LDEPrompt with layer-importance guidance for adaptive layer selection and dynamic freezing/expansion of the prompt pool, addressing limitations of fixed pools and manual selection.

Result: Achieves state-of-the-art performance on widely used class-incremental learning benchmarks, demonstrating effectiveness and scalability.

Conclusion: LDEPrompt effectively addresses key limitations in prompt-based class-incremental learning through adaptive layer selection and dynamic prompt pool management.

Abstract: Prompt-based class-incremental learning methods typically construct a prompt pool consisting of multiple trainable key-prompts and perform instance-level matching to select the most suitable prompt embeddings, which has shown promising results. However, existing approaches face several limitations, including fixed prompt pools, manual selection of prompt embeddings, and strong reliance on the pretrained backbone for prompt selection. To address these issues, we propose a \textbf{L}ayer-importance guided \textbf{D}ual \textbf{E}xpandable \textbf{P}rompt Pool (\textbf{LDEPrompt}), which enables adaptive layer selection as well as dynamic freezing and expansion of the prompt pool. Extensive experiments on widely used class-incremental learning benchmarks demonstrate that LDEPrompt achieves state-of-the-art performance, validating its effectiveness and scalability.

Method: Proposes CDPR: Cross-modal Diffusion with Polarization for Reliable Monocular Depth Estimation. Encodes RGB and polarization (AoLP/DoLP) images into shared latent space via pre-trained VAE. Uses learnable confidence-aware gating mechanism to dynamically fuse multi-modal information, suppressing noisy polarization signals while preserving informative cues. Framework can be generalized to surface normal prediction with minimal modification.

Result: Experiments on synthetic and real-world datasets show CDPR significantly outperforms RGB-only baselines in challenging regions while maintaining competitive performance in standard scenes. Also demonstrates scalability to polarization-guided dense prediction tasks like surface normal estimation.

Conclusion: Integration of polarization priors with diffusion models enhances monocular depth estimation robustness, especially for challenging surfaces. The framework shows promise for general polarization-guided dense prediction tasks beyond depth estimation.

Abstract: Monocular depth estimation is a fundamental yet challenging task in computer vision, especially under complex conditions such as textureless surfaces, transparency, and specular reflections. Recent diffusion-based approaches have significantly advanced performance by reformulating depth prediction as a denoising process in the latent space. However, existing methods rely solely on RGB inputs, which often lack sufficient cues in challenging regions. In this work, we present CDPR - Cross-modal Diffusion with Polarization for Reliable Monocular Depth Estimation - a novel diffusion-based framework that integrates physically grounded polarization priors to enhance estimation robustness. Specifically, we encode both RGB and polarization (AoLP/DoLP) images into a shared latent space via a pre-trained Variational Autoencoder (VAE), and dynamically fuse multi-modal information through a learnable confidence-aware gating mechanism. This fusion module adaptively suppresses noisy signals in polarization inputs while preserving informative cues, particularly around reflective or transparent surfaces, and provides the integrated latent representation for subsequent monocular depth estimation. Beyond depth estimation, we further verify that our framework can be easily generalized to surface normal prediction with minimal modification, showcasing its scalability to general polarization-guided dense prediction tasks. Experiments on both synthetic and real-world datasets validate that CDPR significantly outperforms RGB-only baselines in challenging regions while maintaining competitive performance in standard scenes.

Method: DualComp uses a lightweight pre-trained router to dynamically guide feature processing into two pathways: 1) Object semantic stream with Spatially-Contiguous Semantic Aggregator for background compression while protecting small objects, and 2) Scene geometric stream with Instruction-Guided Structure Recoverer that reconstructs spatial skeletons using greedy path-tracing topology completion.

Result: Experiments on XLRS-Bench show DualComp achieves high-fidelity remote sensing interpretation at exceptionally low computational cost with simultaneous improvements in both efficiency and accuracy.

Conclusion: The proposed task-adaptive dual-stream compression framework effectively addresses the computational bottleneck in processing UHR remote sensing imagery for MLLMs by respecting the fundamental duality between semantic and geometric interpretation needs.

Abstract: Multimodal Large Language Models (MLLMs) have demonstrated immense potential in Earth observation. However, the massive visual tokens generated when processing Ultra-High-Resolution (UHR) imagery introduce prohibitive computational overhead, severely bottlenecking their inference efficiency. Existing visual token compression methods predominantly adopt static and uniform compression strategies, neglecting the inherent “Semantic-Geometric Duality” in remote sensing interpretation tasks. Specifically, object semantic tasks focus on the abstract semantics of objects and benefit from aggressive background pruning, whereas scene geometric tasks critically rely on the integrity of spatial topology. To address this challenge, we propose DualComp, a task-adaptive dual-stream token compression framework. Dynamically guided by a lightweight pre-trained router, DualComp decouples feature processing into two dedicated pathways. In the object semantic stream, the Spatially-Contiguous Semantic Aggregator (SCSA) utilizes size-adaptive clustering to aggregates redundant background while protecting small object. In the scene geometric stream, the Instruction-Guided Structure Recoverer (IGSR) introduces a greedy path-tracing topology completion mechanism to reconstruct spatial skeletons. Experiments on the UHR remote sensing benchmark XLRS-Bench demonstrate that DualComp accomplishes high-fidelity remote sensing interpretation at an exceptionally low computational cost, achieving simultaneous improvements in both efficiency and accuracy.

Method: Inject object spatial-temporal information directly into visual modality by rendering colored bounding boxes and trajectory trails onto video frames as visual prompts, with only a concise color-to-object legend as text.

Result: BoxTuning surpasses text-coordinate baselines on spatially oriented tasks and nearly eliminates accuracy degradation on reasoning-centric tasks across five video QA benchmarks (CLEVRER, Perception Test, STAR, NExT-QA, IntentQA).

Conclusion: Visual prompting with bounding boxes and trajectory trails is a more natural and efficient paradigm for conveying object information to video MLLMs, resolving the modality mismatch of text-coordinate approaches.

Abstract: Object-level spatial-temporal understanding is essential for video question answering, yet existing multimodal large language models (MLLMs) encode frames holistically and lack explicit mechanisms for fine-grained object grounding. Recent work addresses this by serializing bounding box coordinates as text tokens, but this text-coordinate paradigm suffers from a fundamental modality mismatch: object information is inherently visual, yet encoding it as text incurs a high token cost that forces aggressive temporal downsampling. We propose BoxTuning, which resolves this mismatch by injecting object spatial-temporal information directly into the visual modality. Colored bounding boxes and trajectory trails are rendered onto video frames as visual prompts, with only a concise color-to-object legend retained as text. This reduces the token cost significantly, achieving 87-93% text token reduction in practice. It also preserves full temporal resolution, where the trajectory trails further encode inter-frame motion direction and speed within each keyframe, recovering fine-grained dynamics that text-coordinate methods are forced to discard. Experimental results on five video QA benchmarks (CLEVRER, Perception Test, STAR, NExT-QA, IntentQA) show that BoxTuning surpasses text-coordinate baselines on spatially oriented tasks and nearly eliminates the accuracy degradation observed on reasoning-centric tasks, establishing visual prompting as a more natural and efficient paradigm for conveying object information to video MLLMs.

Method: Two core components: 1) Sparse Hypergraph-enhanced Cross-Modal Fusion (S-HCF) that uses event-guided activity maps and high-order hypergraph modeling on motion-critical sparse tokens; 2) Fine-Grained Mixture of Experts (FG-MoE) with specialized hypergraph experts for different image regions using pixel-level spatial gating, load-balancing loss, and zero-initialization.

Result: Achieves superior accuracy-efficiency trade-off on mainstream RGB-Event benchmarks, outperforming state-of-the-art methods while maintaining lightweight footprint suitable for real-time edge deployment.

Conclusion: Hyper-FEOD effectively addresses multimodal heterogeneity and computational efficiency challenges in RGB-Event object detection through sparse hypergraph fusion and adaptive expert enhancement.

Abstract: Integrating frame-based RGB cameras with event streams offers a promising solution for robust object detection under challenging dynamic conditions. However, the inherent heterogeneity and data redundancy of these modalities often lead to prohibitive computational overhead or suboptimal feature fusion. In this paper, we propose Hyper-FEOD, a high-performance and efficient detection framework, which synergistically optimizes multi-modal interaction through two core components. First, we introduce Sparse Hypergraph-enhanced Cross-Modal Fusion (S-HCF), which leverages the inherent sparsity of event streams to construct an event-guided activity map. By performing high-order hypergraph modeling exclusively on selected motion-critical sparse tokens, S-HCF captures complex non-local dependencies between RGB and event data while overcoming the traditional complexity bottlenecks of hypergraph computation. Second, we design a Fine-Grained Mixture of Experts (FG-MoE) Enhancement module to address the diverse semantic requirements of different image regions. This module employs specialized hypergraph experts tailored for object boundaries, internal textures, and backgrounds, utilizing a pixel-level spatial gating mechanism to adaptively route and enhance features. Combined with a load-balancing loss and zero-initialization strategy, FG-MoE ensures stable training and precise feature refinement without disrupting the pre-trained backbone’s distribution. Experimental results on mainstream RGB-Event benchmarks demonstrate that Hyper-FEOD achieves a superior accuracy-efficiency trade-off, outperforming state-of-the-art methods while maintaining a lightweight footprint suitable for real-time edge deployment.

Method: The framework uses: 1) Naka-guided chroma-correction network combining physics-prior enhancement, dual-branch input modeling, frequency-decoupled correction, and mask-guided optimization; 2) Feed-forward multi-view reconstruction for dense scene priors; 3) Lightweight Point Preprocessing Module (PPM) with coordinate alignment, voxel pooling, and distance-adaptive progressive pruning.

Result: NAKA-GS outperforms baseline methods by a large margin, improves restoration quality, training stability, and optimization efficiency for low-light 3D reconstruction without heavy inference overhead.

Conclusion: The proposed bionics-inspired framework effectively addresses low-light challenges in 3D Gaussian Splatting by jointly optimizing photometric restoration and geometric initialization, demonstrating superior performance in the NTIRE 3D Restoration and Reconstruction Challenge.

Abstract: Low-light conditions severely hinder 3D restoration and reconstruction by degrading image visibility, introducing color distortions, and contaminating geometric priors for downstream optimization. We present NAKA-GS, a bionics-inspired framework for low-light 3D Gaussian Splatting that jointly improves photometric restoration and geometric initialization. Our method starts with a Naka-guided chroma-correction network, which combines physics-prior low-light enhancement, dual-branch input modeling, frequency-decoupled correction, and mask-guided optimization to suppress bright-region chromatic artifacts and edge-structure errors. The enhanced images are then fed into a feed-forward multi-view reconstruction model to produce dense scene priors. To further improve Gaussian initialization, we introduce a lightweight Point Preprocessing Module (PPM) that performs coordinate alignment, voxel pooling, and distance-adaptive progressive pruning to remove noisy and redundant points while preserving representative structures. Without introducing heavy inference overhead, NAKA-GS improves restoration quality, training stability, and optimization efficiency for low-light 3D reconstruction. The proposed method was presented in the NTIRE 3D Restoration and Reconstruction (3DRR) Challenge, and outperformed the baseline methods by a large margin. The code is available at https://github.com/RunyuZhu/Naka-GS

Method: Explores deep learning models for multi-modal A/H recognition in videos using three approaches: supervised learning, unsupervised domain adaptation for personalization, and zero-shot inference using large language models (LLMs). Experiments conducted on the BAH video dataset.

Result: Results show limited performance, indicating current models are insufficient for accurate A/H recognition. The paper suggests better methods are needed for spatio-temporal and multimodal fusion to leverage conflicts within and across modalities.

Conclusion: Current deep learning models are inadequate for recognizing subtle ambivalence/hesitancy in videos. More sophisticated multi-modal models with better fusion techniques are required to capture the complex emotional conflicts that characterize A/H states.

Abstract: Using behavioural science, health interventions focus on behaviour change by providing a framework to help patients acquire and maintain healthy habits that improve medical outcomes. In-person interventions are costly and difficult to scale, especially in resource-limited regions. Digital health interventions offer a cost-effective approach, potentially supporting independent living and self-management. Automating such interventions, especially through machine learning, has gained considerable attention recently. Ambivalence and hesitancy (A/H) play a primary role for individuals to delay, avoid, or abandon health interventions. A/H are subtle and conflicting emotions that place a person in a state between positive and negative evaluations of a behaviour, or between acceptance and refusal to engage in it. They manifest as affective inconsistency across modalities or within a modality, such as language, facial, vocal expressions, and body language. While experts can be trained to recognize A/H, integrating them into digital health interventions is costly and less effective. Automatic A/H recognition is therefore critical for the personalization and cost-effectiveness of digital health interventions. Here, we explore the application of deep learning models for A/H recognition in videos, a multi-modal task by nature. In particular, this paper covers three learning setups: supervised learning, unsupervised domain adaptation for personalization, and zero-shot inference via large language models (LLMs). Our experiments are conducted on the unique and recently published BAH video dataset for A/H recognition. Our results show limited performance, suggesting that more adapted multi-modal models are required for accurate A/H recognition. Better methods for modeling spatio-temporal and multimodal fusion are necessary to leverage conflicts within/across modalities.

Method: Proposes rPPG-VQA with dual-branch architecture: signal-level branch uses robust SNR estimation with multi-method consensus, and scene-level branch uses multimodal large language model (MLLM) to identify motion and lighting interferences. Also introduces two-stage adaptive sampling (TAS) strategy to curate optimal training datasets.

Result: Experiments show that training on large-scale “in-the-wild” videos filtered by rPPG-VQA framework enables development of unsupervised rPPG models with substantial improvement in accuracy on standard benchmarks.

Conclusion: The rPPG-VQA framework effectively assesses video suitability for rPPG training, enabling better utilization of unlabeled video data and improving unsupervised rPPG model performance.

Abstract: Unsupervised remote photoplethysmography (rPPG) promises to leverage unlabeled video data, but its potential is hindered by a critical challenge: training on low-quality “in-the-wild” videos severely degrades model performance. An essential step missing here is to assess the suitability of the videos for rPPG model learning before using them for the task. Existing video quality assessment (VQA) methods are mainly designed for human perception and not directly applicable to the above purpose. In this work, we propose rPPG-VQA, a novel framework for assessing video suitability for rPPG. We integrate signal-level and scene-level analyses and design a dual-branch assessment architecture. The signal-level branch evaluates the physiological signal quality of the videos via robust signal-to-noise ratio (SNR) estimation with a multi-method consensus mechanism, and the scene-level branch uses a multimodal large language model (MLLM) to identify interferences like motion and unstable lighting. Furthermore, we propose a two-stage adaptive sampling (TAS) strategy that utilizes the quality score to curate optimal training datasets. Experiments show that by training on large-scale, “in-the-wild” videos filtered by our framework, we can develop unsupervised rPPG models that achieve a substantial improvement in accuracy on standard benchmarks. Our code is available at https://github.com/Tianyang-Dai/rPPG-VQA.

Method: Introduce MosaicMRI - largest open-source raw musculoskeletal MRI dataset (2,671 volumes, 80,156 slices) with diverse orientations, contrasts, anatomies, and coil configurations. Use VarNet baseline for accelerated reconstruction experiments to study scaling behavior and cross-anatomy generalization.

Result: Models trained on combined anatomies outperform anatomy-specific models in low-sample regimes. Identified groups of body parts (e.g., foot and elbow) that generalize well with each other. Performance under domain shifts depends on training set size, anatomy, and protocol-specific factors.

Conclusion: Anatomical diversity in training data provides significant benefits, and cross-anatomical correlations can be exploited for better generalization in MRI reconstruction tasks.

Abstract: Deep learning underpins a wide range of applications in MRI, including reconstruction, artifact removal, and segmentation. However, progress has been driven largely by public datasets focused on brain and knee imaging, shaping how models are trained and evaluated. As a result, careful studies of the reliability of these models across diverse anatomical settings remain limited. In this work, we introduce MosaicMRI, a large and diverse collection of fully sampled raw musculoskeletal (MSK) MR measurements designed for training and evaluating machine-learning-based methods. MosaicMRI is the largest open-source raw MSK MRI dataset to date, comprising 2,671 volumes and 80,156 slices. The dataset offers substantial diversity in volume orientation (e.g., axial, sagittal), imaging contrasts (e.g., PD, T1, T2), anatomies (e.g., spine, knee, hip, ankle, and others), and numbers of acquisition coils. Using VarNet as a baseline for accelerated reconstruction task, we perform a comprehensive set of experiments to study scaling behavior with respect to both model capacity and dataset size. Interestingly, models trained on the combined anatomies significantly outperform anatomy-specific models in low-sample regimes, highlighting the benefits of anatomical diversity and the presence of exploitable cross-anatomical correlations. We further evaluate robustness and cross-anatomy generalization by training models on one anatomy (e.g., spine) and testing them on another (e.g., knee). Notably, we identify groups of body parts (e.g., foot and elbow) that generalize well with each other, and highlight that performance under domain shifts depends on both training set size, anatomy, and protocol-specific factors.

Method: Proposes Boxes2Pixels framework with: (1) hierarchical decoder over frozen DINOv2 features for semantic stability, (2) auxiliary binary localization head to decouple sparse foreground discovery from class prediction, and (3) one-sided online self-correction that relaxes background supervision when student is confident to address teacher false negatives.

Result: On wind turbine inspection benchmark: improves anomaly mIoU by +6.97 and binary IoU by +9.71 over strongest baseline. Online self-correction increases binary recall by +18.56 while using 80% fewer trainable parameters.

Conclusion: Boxes2Pixels effectively addresses noise in SAM-generated pseudo-masks for industrial defect segmentation, achieving significant improvements in segmentation quality with fewer parameters through noise-robust distillation.

Abstract: Accurate defect segmentation is critical for industrial inspection, yet dense pixel-level annotations are rarely available. A common workaround is to convert inexpensive bounding boxes into pseudo-masks using foundation segmentation models such as the Segment Anything Model (SAM). However, these pseudo-labels are systematically noisy on industrial surfaces, often hallucinating background structure while missing sparse defects. To address this limitation, a noise-robust box-to-pixel distillation framework, Boxes2Pixels, is proposed that treats SAM as a noisy teacher rather than a source of ground-truth supervision. Bounding boxes are converted into pseudo-masks offline by SAM, and a compact student is trained with (i) a hierarchical decoder over frozen DINOv2 features for semantic stability, (ii) an auxiliary binary localization head to decouple sparse foreground discovery from class prediction, and (iii) a one-sided online self-correction mechanism that relaxes background supervision when the student is confident, targeting teacher false negatives. On a manually annotated wind turbine inspection benchmark, the proposed Boxes2Pixels improves anomaly mIoU by +6.97 and binary IoU by +9.71 over the strongest baseline trained under identical weak supervision. Moreover, online self-correction increases the binary recall by +18.56, while the model employs 80% fewer trainable parameters. Code is available at https://github.com/CLendering/Boxes2Pixels.

Method: Proposes RADA: Region-Aware Dual-encoder Auxiliary learning pipeline with dual-encoder framework pre-trained on Alpha-CLIP to extract fine-grained, region-specific visual features. Combines image-level visual features with text-level semantic guidance for region-aware semantic supervision. Integrated into triple-view training framework.

Result: Achieves state-of-the-art performance under extremely sparse annotation settings on LA2018, KiTS19, and LiTS datasets, demonstrating robust generalization across diverse medical imaging datasets.

Conclusion: RADA effectively addresses the limitations of geometric-only pseudo-label propagation by incorporating semantic understanding through fine-grained visual features and text guidance, enabling high-quality medical segmentation with minimal annotations.

Abstract: Deep learning has greatly advanced medical image segmentation, but its success relies heavily on fully supervised learning, which requires dense annotations that are costly and time-consuming for 3D volumetric scans. Barely-supervised learning reduces annotation burden by using only a few labeled slices per volume. Existing methods typically propagate sparse annotations to unlabeled slices through geometric continuity to generate pseudo-labels, but this strategy lacks semantic understanding, often resulting in low-quality pseudo-labels. Furthermore, medical image segmentation is inherently a pixel-level visual understanding task, where accuracy fundamentally depends on the quality of local, fine-grained visual features. Inspired by this, we propose RADA, a novel Region-Aware Dual-encoder Auxiliary learning pipeline which introduces a dual-encoder framework pre-trained on Alpha-CLIP to extract fine-grained, region-specific visual features from the original images and limited annotations. The framework combines image-level fine-grained visual features with text-level semantic guidance, providing region-aware semantic supervision that bridges image-level semantics and pixel-level segmentation. Integrated into a triple-view training framework, RADA achieves SOTA performance under extremely sparse annotation settings on LA2018, KiTS19 and LiTS, demonstrating robust generalization across diverse datasets.

Method: SeSAM decomposes class masks into connected components, samples point prompts along object skeletons, selects SAM masks using weak-label coverage, and iteratively refines labels using pseudo-labels. It integrates with semi-supervised learning to balance ground-truth labels, SAM-based pseudo-labels, and high-confidence pseudo-labels.

Result: Extensive experiments across multiple benchmarks and weak annotation types show SeSAM consistently outperforms weakly supervised baselines while substantially reducing annotation cost relative to fine supervision.

Conclusion: SeSAM successfully adapts SAM for semantic segmentation using weak labels, providing an effective framework that balances annotation cost and segmentation quality.

Abstract: Semantic segmentation requires dense pixel-level annotations, which are costly and time-consuming to acquire. To address this, we present SeSAM, a framework that uses a foundational segmentation model, i.e. Segment Anything Model (SAM), with weak labels, including coarse masks, scribbles, and points. SAM, originally designed for instance-based segmentation, cannot be directly used for semantic segmentation tasks. In this work, we identify specific challenges faced by SAM and determine appropriate components to adapt it for class-based segmentation using weak labels. Specifically, SeSAM decomposes class masks into connected components, samples point prompts along object skeletons, selects SAM masks using weak-label coverage, and iteratively refines labels using pseudo-labels, enabling SAM-generated masks to be effectively used for semantic segmentation. Integrated with a semi-supervised learning framework, SeSAM balances ground-truth labels, SAM-based pseudo-labels, and high-confidence pseudo-labels, significantly improving segmentation quality. Extensive experiments across multiple benchmarks and weak annotation types show that SeSAM consistently outperforms weakly supervised baselines while substantially reducing annotation cost relative to fine supervision.

Method: Created a large-scale benchmark with public training set and hidden test set reflecting real-world incidence. Evaluated 11 teams’ approaches using diverse architectures, pretraining, ensembling, and calibration strategies with operating-point-specific metrics emphasizing high sensitivity.

Result: While several methods achieved strong discriminative performance, positive predictive values remained low, highlighting the difficulty of low-prevalence detection. All methods used fully supervised classification despite dominance of normal findings, lacking prevalence-agnostic approaches.

Conclusion: The challenge reveals critical gaps in current CADe systems for low-prevalence detection and provides a public dataset and evaluation framework to support development of prevalence-robust systems suitable for clinical surveillance.

Abstract: Computer-aided detection (CADe) of early neoplasia in Barrett’s esophagus is a low-prevalence surveillance problem in which clinically relevant findings are rare. Although many CADe systems report strong performance on balanced or enriched datasets, their behavior under realistic prevalence remains insufficiently characterized. The RARE25 challenge addresses this gap by introducing a large-scale, prevalence-aware benchmark for neoplasia detection. It includes a public training set and a hidden test set reflecting real-world incidence. Methods were evaluated using operating-point-specific metrics emphasizing high sensitivity and accounting for prevalence. Eleven teams from seven countries submitted approaches using diverse architectures, pretraining, ensembling, and calibration strategies. While several methods achieved strong discriminative performance, positive predictive values remained low, highlighting the difficulty of low-prevalence detection and the risk of overestimating clinical utility when prevalence is ignored. All methods relied on fully supervised classification despite the dominance of normal findings, indicating a lack of prevalence-agnostic approaches such as anomaly detection or one-class learning. By releasing a public dataset and a reproducible evaluation framework, RARE25 aims to support the development of CADe systems robust to prevalence shift and suitable for clinical surveillance workflows.

Method: Combines 3D rectified flow architecture to capture cross-modal relationships with BiomedCLIP vision-language model for text-guided personalized generation using clinical scores and imaging knowledge.

Result: Produces synthetic PET images with superior fidelity and generalizability, accurately recapitulates disease-specific patterns, and enables precise personalized stratification of mild cognitive impairment when combined with MRI.

Conclusion: DIReCT++ advances a scalable, data-efficient tool for early diagnosis and prognostic prediction of Alzheimer’s disease through multi-modal neuroimaging synthesis.

Abstract: The biological definition of Alzheimer’s disease (AD) relies on multi-modal neuroimaging, yet the clinical utility of positron emission tomography (PET) is limited by cost and radiation exposure, hindering early screening at preclinical or prodromal stages. While generative models offer a promising alternative by synthesizing PET from magnetic resonance imaging (MRI), achieving subject-specific precision remains a primary challenge. Here, we introduce DIReCT$++$, a Domain-Informed ReCTified flow model for synthesizing multi-tracer PET from MRI combined with fundamental clinical information. Our approach integrates a 3D rectified flow architecture to capture complex cross-modal and cross-tracer relationships with a domain-adapted vision-language model (BiomedCLIP) that provides text-guided, personalized generation using clinical scores and imaging knowledge. Extensive evaluations on multi-center datasets demonstrate that DIReCT$++$ not only produces synthetic PET images ($^{18}$F-AV-45 and $^{18}$F-FDG) of superior fidelity and generalizability but also accurately recapitulates disease-specific patterns. Crucially, combining these synthesized PET images with MRI enables precise personalized stratification of mild cognitive impairment (MCI), advancing a scalable, data-efficient tool for the early diagnosis and prognostic prediction of AD. The source code will be released on https://github.com/ladderlab-xjtu/DIReCT-PLUS.

Method: Used four configurations of Google’s Gemini 2.5 Flash and Flash Lite models across scenes from 100 hours of video. Introduced three evaluation metrics: Contentfulness (useful scene content vs meta-commentary), Thought-Final Coverage (faithfulness of thought stream to final output), and Dominant Entity Analysis (identifying focused subjects, actions, settings). GPT-5 served as independent judge.

Result: Quality gains from additional thinking plateau quickly, with most improvement in first few hundred tokens. Flash Lite offers best balance between quality and token usage. Tight reasoning budgets cause compression-step hallucination (adding content never reasoned about). Flash and Flash Lite produce similar thought streams but differ in style: Flash discusses reasoning process while Lite focuses on scene description.

Conclusion: Internal reasoning traces significantly impact video scene understanding, but benefits diminish quickly. Model architecture affects reasoning style, and compression artifacts can lead to hallucinations when reasoning budgets are constrained.

Abstract: We benchmark how internal reasoning traces, which we call thought streams, affect video scene understanding in vision-language models. Using four configurations of Google’s Gemini 2.5 Flash and Flash Lite across scenes extracted from 100 hours of video, we ask three questions: does more thinking lead to better outputs, where do the gains stop, and what do these models actually think about? We introduce three evaluation metrics. Contentfulness measures how much of the thought stream is useful scene content versus meta-commentary. Thought-Final Coverage measures how faithfully the thought stream translates into the final output. Dominant Entity Analysis identifies which subjects, actions, and settings the model focuses on. GPT-5 serves as an independent judge. We find that quality gains from additional thinking plateau quickly, with most improvement occurring in the first few hundred tokens. Flash Lite offers the best balance between quality and token usage. Tight reasoning budgets cause the model to add content in the final output that it never reasoned about, a form of compression-step hallucination. Despite being different model tiers, Flash and Flash Lite produce similar thought streams, though they differ in style: Flash discusses its reasoning process, while Lite focuses on describing the scene.

Method: 1) Category-level collaboration knowledge mining strategy exploiting inter-class and intra-class relationships across domains. 2) Clustering-based memory bank encoding class prototypes, auxiliary features, and intra-class disparity information, iteratively updated via unsupervised clustering. 3) Base-to-novel selection metric to discover source-domain features related to novel categories for initializing novel-category classifiers. 4) Adaptive feature assignment strategy transferring learned category-level knowledge to target domain with asynchronous memory bank updates.

Result: Extensive experiments on multiple benchmarks show the method consistently surpasses state-of-the-art AOOD methods by 1.1-5.5 mAP.

Conclusion: The proposed approach effectively addresses limitations in cross-domain representation, novel category ambiguity, and source-domain bias in adaptive open-set object detection through category-level knowledge mining and collaborative learning strategies.

Abstract: Existing object detectors often struggle to generalize across domains while adapting to emerging novel categories. Adaptive open-set object detection (AOOD) addresses this challenge by training on base categories in the source domain and adapting to both base and novel categories in the target domain without target annotations. However, current AOOD methods remain limited by weak cross-domain representations, ambiguity among novel categories, and source-domain feature bias. To address these issues, we propose a category-level collaboration knowledge mining strategy that exploits both inter-class and intra-class relationships across domains. Specifically, we construct a clustering-based memory bank to encode class prototypes, auxiliary features, and intra-class disparity information, and iteratively update it via unsupervised clustering to enhance category-level knowledge representation. We further design a base-to-novel selection metric to discover source-domain features related to novel categories and use them to initialize novel-category classifiers. In addition, an adaptive feature assignment strategy transfers the learned category-level knowledge to the target domain and asynchronously updates the memory bank to alleviate source-domain bias. Extensive experiments on multiple benchmarks show that our method consistently surpasses state-of-the-art AOOD methods by 1.1-5.5 mAP.

Method: Integrates medical prior knowledge with a feature-level region prompt integration mechanism that can handle various prompt forms (points, bounding boxes, masks) while maintaining global context. Pre-trained on a large-scale medical dataset with 6.4M images and 97.3M region-level annotations.

Result: Significantly outperforms baseline methods in various medical tasks including zero-shot recognition, interactive segmentation, and empowering multimodal large language models. Provides cross-disease and cross-modality fine-grained spatial semantic understanding.

Conclusion: MedP-CLIP offers a scalable, plug-and-play visual backbone for medical AI that combines holistic image understanding with precise regional analysis, addressing the critical need for fine-grained medical image comprehension.

Abstract: Contrastive Language-Image Pre-training (CLIP) has demonstrated outstanding performance in global image understanding and zero-shot transfer through large-scale text-image alignment. However, the core of medical image analysis often lies in the fine-grained understanding of specific anatomical structures or lesion regions. Therefore, precisely comprehending region-of-interest (RoI) information provided by medical professionals or perception models becomes crucial. To address this need, we propose MedP-CLIP, a region-aware medical vision-language model (VLM). MedP-CLIP innovatively integrates medical prior knowledge and designs a feature-level region prompt integration mechanism, enabling it to flexibly respond to various prompt forms (e.g., points, bounding boxes, masks) while maintaining global contextual awareness when focusing on local regions. We pre-train the model on a meticulously constructed large-scale dataset (containing over 6.4 million medical images and 97.3 million region-level annotations), equipping it with cross-disease and cross-modality fine-grained spatial semantic understanding capabilities. Experiments demonstrate that MedP-CLIP significantly outperforms baseline methods in various medical tasks, including zero-shot recognition, interactive segmentation, and empowering multimodal large language models. This model provides a scalable, plug-and-play visual backbone for medical AI, combining holistic image understanding with precise regional analysis.

Method: Created the SeIQA dataset with 510 training, 80 validation, and 160 testing image pairs (degraded vs ground truth). Organized a challenge with 58 registered teams, 6 of which submitted valid solutions for benchmarking.

Result: The challenge successfully established a new benchmark for human-oriented semantic image quality assessment, with submitted solutions achieving state-of-the-art performance on the SeIQA dataset.

Conclusion: The LoViF 2026 Challenge successfully raised awareness and established a benchmark for human-oriented semantic image quality assessment, advancing research in semantic information evaluation from human perspective.

Abstract: This paper reviews the LoViF 2026 Challenge on Human-oriented Semantic Image Quality Assessment. This challenge aims to raise a new direction, i.e., how to evaluate the loss of semantic information from the human perspective, intending to promote the development of some new directions, like semantic coding, processing, and semantic-oriented optimization, etc. Unlike existing datasets of quality assessment, we form a dataset of human-oriented semantic quality assessment, termed the SeIQA dataset. This dataset is divided into three parts for this competition: (i) training data: 510 pairs of degraded images and their corresponding ground truth references; (ii) validation data: 80 pairs of degraded images and their corresponding ground-truth references; (iii) testing data: 160 pairs of degraded images and their corresponding ground-truth references. The primary objective of this challenge is to establish a new and powerful benchmark for human-oriented semantic image quality assessment. There are a total of 58 teams registered in this competition, and 6 teams submitted valid solutions and fact sheets for the final testing phase. These submissions achieved state-of-the-art (SOTA) performance on the SeIQA dataset.

Method: Uses a two-phase region merging process: starting from fine partitions guided by deep features and object priors, first preserves object consistency, then allows controlled inter-object grouping. Can be modulated with visual attention maps or user input.

Result: Strongly outperforms existing hierarchical methods in accuracy and regularity, while performing on par with state-of-the-art non-hierarchical methods on standard benchmarks.

Conclusion: H-SPAM provides a unified framework for generating accurate, regular, and perfectly nested hierarchical superpixels that can benefit multi-scale vision pipelines.

Abstract: Superpixels offer a compact image representation by grouping pixels into coherent regions. Recent methods have reached a plateau in terms of segmentation accuracy by generating noisy superpixel shapes. Moreover, most existing approaches produce a single fixed-scale partition that limits their use in vision pipelines that would benefit multi-scale representations. In this work, we introduce H-SPAM (Hierarchical Superpixel Anything Model), a unified framework for generating accurate, regular, and perfectly nested hierarchical superpixels. Starting from a fine partition, guided by deep features and external object priors, H-SPAM constructs the hierarchy through a two-phase region merging process that first preserves object consistency and then allows controlled inter-object grouping. The hierarchy can also be modulated using visual attention maps or user input to preserve important regions longer in the hierarchy. Experiments on standard benchmarks show that H-SPAM strongly outperforms existing hierarchical methods in both accuracy and regularity, while performing on par with most recent state-of-the-art non-hierarchical methods. Code and pretrained models are available: https://github.com/waldo-j/hspam.

Method: Organized a competition with a dataset of 800 real-captured low-light portrait groups (each with low-light input, ground truth, and person mask). Used hybrid evaluation combining objective quantitative metrics with rigorous subjective assessment protocols.

Result: The challenge attracted over 100 participating teams and received more than 3,000 valid submissions, demonstrating widespread interest from both academia and industry in solving low-light portrait restoration problems.

Conclusion: The challenge successfully established a benchmark for real-world low-light portrait restoration, providing a dataset and evaluation framework that addresses the specific challenges of balancing noise suppression, detail preservation, and color/illumination fidelity.

Abstract: In this paper, we present a comprehensive overview of the NTIRE 2026 3rd Restore Any Image Model (RAIM) challenge, with a specific focus on Track 3: AI Flash Portrait. Despite significant advancements in deep learning for image restoration, existing models still encounter substantial challenges in real-world low-light portrait scenarios. Specifically, they struggle to achieve an optimal balance among noise suppression, detail preservation, and faithful illumination and color reproduction. To bridge this gap, this challenge aims to establish a novel benchmark for real-world low-light portrait restoration. We comprehensively evaluate the proposed algorithms utilizing a hybrid evaluation system that integrates objective quantitative metrics with rigorous subjective assessment protocols. For this competition, we provide a dataset containing 800 groups of real-captured low-light portrait data. Each group consists of a 1K-resolution low-light input image, a 1K ground truth (GT), and a 1K person mask. This challenge has garnered widespread attention from both academia and industry, attracting over 100 participating teams and receiving more than 3,000 valid submissions. This report details the motivation behind the challenge, the dataset construction process, the evaluation metrics, and the various phases of the competition. The released dataset and baseline code for this track are publicly available from the same \href{https://github.com/zsn1434/AI_Flash-BaseLine/tree/main}{GitHub repository}, and the official challenge webpage is hosted on \href{https://www.codabench.org/competitions/12885/}{CodaBench}.

Method: 1) Construct CA-CDD (category-agnostic change detection dataset), 2) Design category-agnostic change head to detect transitions of arbitrary categories and index them to specific classes, 3) Propose Seg2Change adapter to adapt open-vocabulary semantic segmentation models to change detection task.

Result: Achieves state-of-the-art OVCD performance: +9.52 IoU on WHU-CD and +5.50 mIoU on SECOND datasets.

Conclusion: Seg2Change provides a simple yet effective framework for open-vocabulary change detection by adapting existing segmentation models, enabling detection of changes across arbitrary categories without being limited to predefined classes.

Abstract: Change detection is a fundamental task in remote sensing, aiming to quantify the impacts of human activities and ecological dynamics on land-cover changes. Existing change detection methods are limited to predefined classes in training datasets, which constrains their scalability in real-world scenarios. In recent years, numerous advanced open-vocabulary semantic segmentation models have emerged for remote sensing imagery. However, there is still a lack of an effective framework for directly applying these models to open-vocabulary change detection (OVCD), a novel task that integrates vision and language to detect changes across arbitrary categories. To address these challenges, we first construct a category-agnostic change detection dataset, termed CA-CDD. Further, we design a category-agnostic change head to detect the transitions of arbitrary categories and index them to specific classes. Based on them, we propose Seg2Change, an adapter designed to adapt open-vocabulary semantic segmentation models to change detection task. Without bells and whistles, this simple yet effective framework achieves state-of-the-art OVCD performance (+9.52 IoU on WHU-CD and +5.50 mIoU on SECOND). Our code is released at https://github.com/yogurts-sy/Seg2Change.

Method: Proposes semantic bridge fusion framework: (1) text as shared semantic bridge to align RGB and IR under unified category conditions, (2) recalibrated thermal semantic prior projected onto RGB for semantic-level mapping fusion, (3) bi-support modeling with consensus and discrepancy supports, (4) bidirectional semantic alignment module for closed-loop vision-text guidance.

Result: Extensive experiments demonstrate effectiveness and superior detection performance on multispectral benchmarks.

Conclusion: The proposed framework effectively addresses RGB-IR granularity asymmetry and leverages both consensus and discrepancy information through text-guided semantic bridging and bi-support modeling.

Abstract: Text-guided multispectral object detection uses text semantics to guide semantic-aware cross-modal interaction between RGB and IR for more robust perception. However, notable limitations remain: (1) existing methods often use text only as an auxiliary semantic enhancement signal, without exploiting its guiding role to bridge the inherent granularity asymmetry between RGB and IR; and (2) conventional data-driven attention-based fusion tends to emphasize stable consensus while overlooking potentially valuable cross-modal discrepancies. To address these issues, we propose a semantic bridge fusion framework with bi-support modeling for multispectral object detection. Specifically, text is used as a shared semantic bridge to align RGB and IR responses under a unified category condition, while the recalibrated thermal semantic prior is projected onto the RGB branch for semantic-level mapping fusion. We further formulate RGB-IR interaction evidence into the regular consensus support and the complementary discrepancy support that contains potentially discriminative cues, and introduce them into fusion via dynamic recalibration as a structured inductive bias. In addition, we design a bidirectional semantic alignment module for closed-loop vision-text guidance enhancement. Extensive experiments demonstrate the effectiveness of the proposed fusion framework and its superior detection performance on multispectral benchmarks. Code is available at https://github.com/zhenwang5372/Bridging-RGB-IR-Gap.

Method: DeSAP introduces decoupled similarity to capture fine-grained cross-modal relevance between visual features and text tokens, providing explicit task-related guidance. This is integrated with visual saliency signals from visual attention to perform token pruning under both task-related and visual cues.

Result: On LLaVA-1.5-7B, DeSAP achieves 10× FLOPs reduction and 2.3× prefill speedup by retaining only 11.1% of visual tokens while maintaining 98.1% of original performance. Extensive experiments across diverse benchmarks and architectures show consistent outperformance of SOTA methods.

Conclusion: DeSAP enables robust and efficient token pruning for LVLMs by combining task-related cross-modal relevance with visual saliency, achieving significant computational savings with minimal performance degradation.

Abstract: Token pruning has emerged as an effective approach to reduce the substantial computational overhead of Large Vision-Language Models (LVLMs) by discarding less informative visual tokens while preserving performance. However, existing methods typically rely on individual attention sources from different LVLM components, resulting in incomplete and suboptimal pruning decisions due to biased attention distributions. To address this problem, we propose DeSAP, a novel Decoupled Similarity-Aware Pruning method for precise, task-aware token pruning within the visual encoder. Specifically, DeSAP introduces a decoupled similarity to capture fine-grained cross-modal relevance between visual features and text tokens, providing explicit task-related guidance for pruning. By integrating decoupled similarity with visual saliency signals derived from visual attention, DeSAP performs token pruning under the guidance of both task-related and visual cues, enabling robust pruning even under aggressive pruning ratios. Extensive experiments across diverse benchmarks and architectures show that DeSAP consistently outperforms SOTA methods in both accuracy and efficiency. On LLaVA-1.5-7B, DeSAP achieves a 10 times FLOPs reduction and a 2.3 times prefill speedup by retaining only 11.1% of visual tokens, while maintaining 98.1% of the original performance.

Method: Proposes Multi-Stream Scene Script (MTSS) with Stream Factorization (decoupling videos into Reference, Shot, Event, and Global streams) and Relational Grounding (reconnecting streams through explicit identity and temporal links).

Result: MTSS reduces total error rate by 25% on Video-SALMONN-2, gains 67% performance on Daily-Omni reasoning, narrows gap between small/large MLLMs, and improves video generation: 45% identity consistency, 56% audio-visual alignment, 71% temporal controllability.

Conclusion: MTSS provides a more learnable, factorized caption interface that enhances both video understanding and generation in multimodal LLMs by addressing structural bottlenecks of monolithic representations.

Abstract: Advances in Multimodal Large Language Models (MLLMs) are transforming video captioning from a descriptive endpoint into a semantic interface for both video understanding and generation. However, the dominant paradigm still casts videos as monolithic narrative paragraphs that entangle visual, auditory, and identity information. This dense coupling not only compromises representational fidelity but also limits scalability, since even local edits can trigger global rewrites. To address this structural bottleneck, we propose Multi-Stream Scene Script (MTSS), a novel paradigm that replaces monolithic text with factorized and explicitly grounded scene descriptions. MTSS is built on two core principles: Stream Factorization, which decouples a video into complementary streams (Reference, Shot, Event, and Global), and Relational Grounding, which reconnects these isolated streams through explicit identity and temporal links to maintain holistic video consistency. Extensive experiments demonstrate that MTSS consistently enhances video understanding across various models, achieving an average reduction of 25% in the total error rate on Video-SALMONN-2 and an average performance gain of 67% on the Daily-Omni reasoning benchmark. It also narrows the performance gap between smaller and larger MLLMs, indicating a substantially more learnable caption interface. Finally, even without architectural adaptation, replacing monolithic prompts with MTSS in multi-shot video generation yields substantial human-rated improvements: a 45% boost in cross-shot identity consistency, a 56% boost in audio-visual alignment, and a 71% boost in temporal controllability.

Method: VLEED uses a variational autoencoder to transform pretrained embeddings, with a mutual information-based objective that estimates entropy of categorical attributes in latent space. This encourages separation of demographic attributes from identity-relevant information while maintaining fine-grained control over information removal.

Result: Evaluated on IJB-C, RFW, and VGGFace2 datasets for gender and ethnicity disentanglement. VLEED offers better privacy-utility tradeoffs than existing methods and can reduce recognition bias across demographic groups, as measured by verification utility, attribute predictability, and group disparity metrics.

Conclusion: VLEED provides an effective post-hoc approach for disentangling demographic attributes from face recognition embeddings, offering improved privacy-utility tradeoffs and bias reduction capabilities compared to state-of-the-art methods.

Abstract: Face recognition embeddings encode identity, but they also encode other factors such as gender and ethnicity. Depending on how these factors are used by a downstream system, separating them from the information needed for verification is important for both privacy and fairness. We propose Variational Latent Entropy Estimation Disentanglement (VLEED), a post-hoc method that transforms pretrained embeddings with a variational autoencoder and encourages a distilled representation where the categorical variable of interest is separated from identity-relevant information. VLEED uses a mutual information-based objective realised through the estimation of the entropy of the categorical attribute in the latent space, and provides stable training with fine-grained control over information removal. We evaluate our method on IJB-C, RFW, and VGGFace2 for gender and ethnicity disentanglement, and compare it to various state-of-the-art methods. We report verification utility, predictability of the disentangled variable under linear and nonlinear classifiers, and group disparity metrics based on false match rates. Our results show that VLEED offers a wide range of privacy-utility tradeoffs over existing methods and can also reduce recognition bias across demographic groups.

Method: Reformulates abundance estimation as a deep equilibrium model using implicit differentiation for constant-memory training, replaces gradient operators with trainable convolutional networks to capture spectral-spatial information.

Result: DEQ-Unmix achieves superior unmixing performance on synthetic and two real-world datasets while maintaining constant memory cost compared to existing methods.

Conclusion: DEQ-Unmix provides an efficient, interpretable solution for hyperspectral unmixing with constant memory requirements, addressing limitations of existing approaches.

Abstract: Hyperspectral unmixing (HU) is crucial for analyzing hyperspectral imagery, yet achieving accurate unmixing remains challenging. While traditional methods struggle to effectively model complex spectral-spatial features, deep learning approaches often lack physical interpretability. Unrolling-based methods, despite offering network interpretability, inadequately exploit spectral-spatial information and incur high memory costs and numerical precision issues during backpropagation. To address these limitations, we propose DEQ-Unmix, which reformulates abundance estimation as a deep equilibrium model, enabling efficient constant-memory training via implicit differentiation. It replaces the gradient operator of the data reconstruction term with a trainable convolutional network to capture spectral-spatial information. By leveraging implicit differentiation, DEQ-Unmix enables efficient and constant-memory backpropagation. Experiments on synthetic and two real-world datasets demonstrate that DEQ-Unmix achieves superior unmixing performance while maintaining constant memory cost.

Method: The paper provides a comprehensive overview organized around a three-role taxonomy: 1) Semantic Reasoner (video understanding, temporal reasoning, multimodal fusion), 2) Expressive Performer (LLM-driven/augmented speech generation), and 3) Visual Synthesizer (video generators for lip-sync and visual alignment).

Result: The survey systematically categorizes MLLM-based approaches to video translation, highlighting how they overcome limitations of traditional cascaded pipelines and achieve competitive translation quality with stronger robustness in zero-shot and multi-speaker scenarios.

Conclusion: The paper identifies open challenges in video understanding, temporal modeling, and multimodal alignment, and outlines promising future research directions for MLLM-powered video translation systems.

Abstract: Recent developments in video translation have further enhanced cross-lingual access to video content, with multimodal large language models (MLLMs) playing an increasingly important supporting role. With strong multimodal understanding, reasoning, and generation capabilities, MLLMs-based video translation systems are overcoming the limitations of traditional cascaded pipelines that separately handle automatic speech recognition, machine translation, text-to-speech and lip synchronization. These MLLM-powered approaches not only achieve competitive or superior translation quality, but also demonstrate stronger robustness in zero-shot settings and multi-speaker scenarios, while jointly modeling semantic fidelity, timing, speaker identity, and emotional consistency. However, despite the rapid progress of MLLMs and extensive surveys on general video-language understanding, a focused and systematic review of how MLLMs empower video translation tasks is still lacking. To fill this gap, we provide the first comprehensive overview of MLLMs-based video translation, organized around a three-role taxonomy: 1) Semantic Reasoner, which characterizes how MLLMs perform video understanding, temporal reasoning, and multimodal fusion; 2) Expressive Performer, which analyzes LLM-driven and LLM-augmented techniques for expressive, controllable speech generation; and 3) Visual Synthesizer, which examines different types of video generators for high-fidelity lip-sync and visual alignment. Finally, we discuss open challenges in video understanding, temporal modeling, and multimodal alignment, and outline promising future research directions for MLLMs-powered video translation.

Method: 1) 3D Representation Autoencoder (3DRAE) repurposes frozen 2D encoders to ground view-coupled 2D semantics into view-decoupled 3D latent representation. 2) 3D Diffusion Transformer (3DDiT) performs diffusion modeling in this 3D latent space for efficient generation.

Result: Enables representing 3D scenes from arbitrary views with fixed complexity and rich semantics, supports diverse conditioning configurations, and allows decoding to images/point maps along arbitrary camera trajectories without per-trajectory diffusion sampling.

Conclusion: Proposes first approach for direct 3D scene generation in implicit 3D latent space, addressing fundamental limitations of 2D-based methods and enabling more efficient and spatially consistent 3D scene generation.

Abstract: 3D scene generation has long been dominated by 2D multi-view or video diffusion models. This is due not only to the lack of scene-level 3D latent representation, but also to the fact that most scene-level 3D visual data exists in the form of multi-view images or videos, which are naturally compatible with 2D diffusion architectures. Typically, these 2D-based approaches degrade 3D spatial extrapolation to 2D temporal extension, which introduces two fundamental issues: (i) representing 3D scenes via 2D views leads to significant representation redundancy, and (ii) latent space rooted in 2D inherently limits the spatial consistency of the generated 3D scenes. In this paper, we propose, for the first time, to perform 3D scene generation directly within an implicit 3D latent space to address these limitations. First, we repurpose frozen 2D representation encoders to construct our 3D Representation Autoencoder (3DRAE), which grounds view-coupled 2D semantic representations into a view-decoupled 3D latent representation. This enables representing 3D scenes observed from arbitrary numbers of views–at any resolution and aspect ratio–with fixed complexity and rich semantics. Then we introduce 3D Diffusion Transformer (3DDiT), which performs diffusion modeling in this 3D latent space, achieving remarkably efficient and spatially consistent 3D scene generation while supporting diverse conditioning configurations. Moreover, since our approach directly generates a 3D scene representation, it can be decoded to images and optional point maps along arbitrary camera trajectories without requiring per-trajectory diffusion sampling pass, which is common in 2D-based approaches.

Method: Designed PD36 C, a compact CNN with 1,250,694 parameters (4.77 MB), trained on the New Plant Diseases Dataset (87k images, 38 classes) using TensorFlow Keras. Created a Qt for Python desktop application with intuitive GUI for offline inference on commodity hardware.

Result: Achieved 0.99697 training accuracy by epoch 30 and 0.9953 average test accuracy across 38 classes. Many classes achieved perfect precision and recall (1.00), while lower-performing classes like Corn Cercospora leaf spot still achieved ~0.9777 precision and ~0.9634 recall.

Conclusion: Small CNNs can achieve competitive accuracy with careful design and well-curated datasets, making them practical for edge deployment in agricultural disease detection, though challenges remain with adverse conditions and multiple concurrent diseases.

Abstract: Deep learning has markedly advanced image based plant disease diagnosis as improved hardware and dataset quality have enabled increasingly accurate neural network models. This paper presents PD36 C, a compact convolutional neural network (1,250,694 parameters and 4.77 MB) for plant disease classification. Trained with TensorFlow Keras on the New Plant Diseases Dataset (87k images, 38 classes), PD36 C is designed for robustness and edge deployability, complemented by a Qt for Python desktop application that offers an intuitive GUI and offline inference on commodity hardware. Across experiments, training accuracy reached 0.99697 by epoch 30, and average test accuracy was 0.9953 across 38 classes. Per class performance is uniformly high; on the lower end, Corn (maize) Cercospora leaf spot achieved precision around 0.9777 and recall around 0.9634, indicating occasional confusion with visually similar categories, while on the upper end numerous classes including Apple Black rot, Cedar apple rust, Blueberry healthy, Cherry Powdery mildew, Cherry healthy, and all four grape categories achieved perfect precision 1.00 and recall of 1.00, indicating no false positives and strong coverage. These results show that with a well curated dataset and careful architectural design, small CNNs can achieve competitive accuracy compared with recent baselines while remaining practical for edge scenarios. We also note typical constraints such as adverse weather, low quality imagery, and leaves exhibiting multiple concurrent diseases that can degrade performance and warrant future work on domain robustness. Overall, PD36 C and its application pipeline contribute a field ready, efficient solution for AI assisted plant disease detection in smart agriculture.

Method: Proposes LoGo-MR framework with: 1) neighbor-slice encoding to capture local cues for short-term risk, 2) transformer-enhanced multiple-instance learning (MIL) to model distributed global patterns for long-term risk with interpretable slice importance. Extends to LoGo3-MR across axial, sagittal, and coronal planes for complementary volumetric information and voxel-level risk saliency mapping.

Result: Outperforms 2D/3D baselines and existing SOTA MIL methods on large breast MRI screening cohort (~7.5K). Achieves AUCs of 0.77-0.69 for 1- to 5-year prediction, improves C-index by ~6% over 3D CNNs. LoGo3-MR further improves overall performance with interpretable localization across three planes, validated across seven backbones with consistent gains.

Conclusion: The method demonstrates clinical potential for efficient MRI-based breast cancer risk stratification in large-scale screening, offering both predictive performance and interpretability through risk saliency mapping that can assist radiologists in localizing risk-relevant regions.

Abstract: Efficient and explainable breast cancer (BC) risk prediction is critical for large-scale population-based screening. Breast MRI provides functional information for personalized risk assessment. Yet effective modeling remains challenging as fully 3D CNNs capture volumetric context at high computational cost, whereas lightweight 2D CNNs fail to model inter-slice continuity. Importantly, breast MRI modeling for shor- and long-term BC risk stratification remains underexplored. In this study, we propose LoGo-MR, a 2.5D local-global structural modeling framework for five-year BC risk prediction. Aligned with clinical interpretation, our framework first employs neighbor-slice encoding to capture subtle local cues linked to short-term risk. It then integrates transformer-enhanced multiple-instance learning (MIL) to model distributed global patterns related to long-term risk and provide interpretable slice importance. We further apply this framework across axial, sagittal, and coronal planes as LoGo3-MR to capture complementary volumetric information. This multi-plane formulation enables voxel-level risk saliency mapping, which may assist radiologists in localizing risk-relevant regions during breast MRI interpretation. Evaluated on a large breast MRI screening cohort (~7.5K), our method outperforms 2D/3D baselines and existing SOTA MIL methods, achieving AUCs of 0.77-0.69 for 1- to 5-year prediction and improving C-index by ~6% over 3D CNNs. LoGo3-MR further improves overall performance with interpretable localization across three planes, and validation across seven backbones shows consistent gains. These results highlight the clinical potential of efficient MRI-based BC risk stratification for large-scale screening. Code will be released publicly.

Method: Proposes LEADER with: 1) Robust Projection-based Geometric Encoder capturing multi-scale geometric features, and 2) Truncated Relative Reliability loss modeling point-wise ambiguity to mitigate unreliable predictions.

Result: Outperforms state-of-the-art methods on Oxford RobotCar and NCLT datasets with 24.1% and 73.9% relative reductions in position error respectively.

Conclusion: LEADER provides a robust solution for LiDAR relocalization by addressing noise and outlier issues through geometric feature enhancement and reliability-aware loss formulation.

Abstract: LiDAR relocalization has attracted increasing attention as it can deliver accurate 6-DoF pose estimation in complex 3D environments. Recent learning-based regression methods offer efficient solutions by directly predicting global poses without the need for explicit map storage. However, these methods often struggle in challenging scenes due to their equal treatment of all predicted points, which is vulnerable to noise and outliers. In this paper, we propose LEADER, a robust LiDAR-based relocalization framework enhanced by a simple, yet effective geometric encoder. Specifically, a Robust Projection-based Geometric Encoder architecture which captures multi-scale geometric features is first presented to enhance descriptiveness in geometric representation. A Truncated Relative Reliability loss is then formulated to model point-wise ambiguity and mitigate the influence of unreliable predictions. Extensive experiments on the Oxford RobotCar and NCLT datasets demonstrate that LEADER outperforms state-of-the-art methods, achieving 24.1% and 73.9% relative reductions in position error over existing techniques, respectively. The source code is released on https://github.com/JiansW/LEADER.

Method: Hybrid architecture combining CRNN-based detection/redaction of PHI regions (burned-in text, metadata) with latent-diffusion inpainting (Stable Diffusion 2) to restore redacted areas with anatomically plausible content.

Result: The method produces visually coherent de-identified medical images that maintain fidelity for downstream models while substantially reducing re-identification risk, as validated by privacy metrics and task-based evaluations.

Conclusion: The automated pipeline enables secure sharing of medical imaging collections by balancing privacy protection with data utility, lowering barriers to multi-institutional collaboration in medical imaging AI.

Abstract: Removing patient-specific information from medical images is crucial to enable sharing and open science without compromising patient identities. However, many methods currently used for deidentification have negative effects on downstream image analysis tasks because of removal of relevant but non-identifiable information. This work presents an end-to-end deep learning framework for transforming raw clinical image volumes into de-identified, analysis-ready datasets without compromising downstream utility. The methodology developed and tested in this work first detects and redacts regions likely to contain protected health information (PHI), such as burned-in text and metadata, and then uses a generative deep learning model to inpaint the redacted areas with anatomically and imaging plausible content. The proposed pipeline leverages a lightweight hybrid architecture, combining CRNN-based redaction with a latent-diffusion inpainting restoration module (Stable Diffusion 2). We evaluate the approach using both privacy-oriented metrics, which quantify residual PHI and success of redaction, and image-quality and task-based metrics, which assess the fidelity of restored volumes for representative deep learning applications. Our results suggest that the proposed method yields de-identified medical images that are visually coherent, maintaining fidelity for downstream models, while substantially reducing the risk of patient re-identification. By automating anonymization and image reconstruction within a single workflow, and dissemination of large-scale medical imaging collections, thereby lowering a key barrier to data sharing and multi-institutional collaboration in medical imaging AI.

Method: ConvFormer3D-TAP integrates 3D convolutional tokenization with multiscale self-attention, trained using masked spatiotemporal reconstruction and uncertainty-weighted multi-clip fusion. The design captures local anatomical structure through convolutional priors and long-range cardiac-cycle dynamics through hierarchical attention.

Result: On 150,974 clinically acquired cine sequences spanning six standard views, achieved 96% validation accuracy with per-class F1-scores ≥ 0.94 and strong calibration (ECE = 0.025; Brier = 0.040). Residual confusions are concentrated in anatomically adjacent long-axis and LVOT/AV view pairs.

Conclusion: ConvFormer3D-TAP serves as a scalable front-end for view routing, filtering and quality control in end-to-end cMRI workflows, addressing the challenge of reliable view classification under clinical variability.

Abstract: Reliable recognition of standard cine cardiac MRI views is essential because each view determines which cardiac anatomy is visualized and which quantitative analyses can be performed. Incorrect view identification, whether by a human reader or an automated deep learning system, can propagate errors into segmentation, volumetric assessment, strain analysis, and valve evaluation. However, accurate view classification remains challenging under routine clinical variability in scanner vendor, acquisition protocol, motion artifacts, and plane prescription. We present ConvFormer3D-TAP, a cine-specific spatiotemporal architecture that integrates 3D convolutional tokenization with multiscale self-attention. The model is trained using masked spatiotemporal reconstruction and uncertainty-weighted multi-clip fusion to enhance robustness across cardiac phases and ambiguous temporal segments. The design captures complementary cues: local anatomical structure through convolutional priors and long-range cardiac-cycle dynamics through hierarchical attention. On a cohort of 150,974 clinically acquired cine sequences spanning six standard cine cardiac MRI views, ConvFormer3D-TAP achieved 96% validation accuracy with per-class F1-scores >= 0.94 and strong calibration (ECE = 0.025; Brier = 0.040). Error analysis shows that residual confusions are concentrated in anatomically adjacent long-axis and LVOT/AV view pairs, consistent with intrinsic prescription overlap. These results support ConvFormer3D-TAP as a scalable front-end for view routing, filtering and quality control in end-to-end cMRI workflows.

Method: Four-stage pipeline: (1) Physical Prior Extraction with dual-stream statistical guidance, (2) Guided Manifold Purification using OmniContext Autoencoder, (3) Residual Score Modeling with Diffusion Transformer guarded by Physical Spectral Firewall, (4) Vector Dynamics Inference evaluating high-dimensional vector interference patterns.

Result: Comprehensive evaluations on eight datasets confirm R2VD establishes new state-of-the-art performance with exceptional target detectability and background suppression.

Conclusion: R2VD fundamentally redefines reconstruction as a manifold purification origin, establishing a residual-guided generative dynamics paradigm that robustly decouples targets from backgrounds using vector interference patterns instead of scalar errors.

Abstract: While Hyperspectral Anomaly Detection (HAD) excels at identifying sparse targets in complex scenes, existing models remain trapped in a scalar “reconstruction-as-endpoint” paradigm. This reliance on ambiguous scalar residuals consistently triggers sub-pixel anomaly vanishing during spatial downsampling, alongside severe confirmation bias when unpurified anomalies corrupt training weights. In this paper, we propose Reconstruction-to-Vector Diffusion (R2VD), which fundamentally redefines reconstruction as a manifold purification origin to establish a novel residual-guided generative dynamics paradigm. Our framework introduces a four-stage pipeline: (1) a Physical Prior Extraction (PPE) stage that mitigates early confirmation bias via dual-stream statistical guidance; (2) a Guided Manifold Purification (GMP) stage utilizing an OmniContext Autoencoder (OCA) to extract purified residual maps while preserving fragile sub-pixel topologies; (3) a Residual Score Modeling (RSM) stage where a Diffusion Transformer (DiT), guarded by a Physical Spectral Firewall (PSF), effectively isolates cross-spectral leakage; and (4) a Vector Dynamics Inference (VDI) stage that robustly decouples targets from backgrounds by evaluating high-dimensional vector interference patterns instead of conventional scalar errors. Comprehensive evaluations on eight datasets confirm that R2VD establishes a new state-of-the-art, delivering exceptional target detectability and background suppression.

Method: Proposes two modules: 1) Angle-guided ROI Adaptive Optimization module that quantifies ROI-Camera angles to refine motion-affected signals and capture global motion, and 2) Multi-region Joint Graph Signal Denoising module that jointly models intra- and inter-regional ROI signals using graph signal processing to suppress motion artifacts.

Result: Joint use of both modules reduces MAE by an average of 20.38% over baseline on three public datasets. Ablation studies confirm the effectiveness of each individual module.

Conclusion: The work demonstrates the potential of angle-guided optimization and graph-based denoising to enhance rPPG performance in motion scenarios, making non-contact heart rate measurement more robust to facial motions.

Abstract: Remote photoplethysmography (rPPG) enables non-contact heart rate measurement from facial videos, but its performance is significantly degraded by facial motions such as speaking and head shaking. To address this issue, we propose two plug-and-play modules. The Angle-guided ROI Adaptive Optimization module quantifies ROI-Camera angles to refine motion-affected signals and capture global motion, while the Multi-region Joint Graph Signal Denoising module jointly models intra- and inter-regional ROI signals using graph signal processing to suppress motion artifacts. The modules are compatible with reflection model-based rPPG methods and validated on three public datasets. Results show that jointly use markedly reduces MAE, with an average decrease of 20.38% over the baseline, while ablation studies confirm the effectiveness of each module. The work demonstrates the potential of angle-guided optimization and graph-based denoising to enhance rPPG performance in motion scenarios.

Method: Uses two-pass raycasting to derive reliable image-aligned masks from LoD3 CityGML models, leveraging parent-child relations to validate facade elements. Fuses geometry-grounded masks with foundation-model predictions for scene-consistent instance correspondences, and learns compact identity encoding for each Gaussian with joint 2D identity supervision and 3D spatial regularization.

Result: Outperforms existing 2D-driven semantic 3DGS baselines (LangSplat, Gaga) by up to 15.8 IoU points in coarse building segmentation and 14.2 mIoU points in fine-grained semantic segmentation on TUM2TWIN and Gold Coast datasets.

Conclusion: GS4City effectively incorporates structured building semantics into Gaussian scene representations, enabling semantically queryable and structure-aware urban reconstruction by bridging city models and photorealistic Gaussian representations.

Abstract: Recent semantic 3D Gaussian Splatting (3DGS) methods primarily rely on 2D foundation models, often yielding ambiguous boundaries and limited support for structured urban semantics. While city models such as CityGML encode hierarchically organized semantics together with building geometry, these labels cannot be directly mapped to Gaussian primitives. We present GS4City, a hierarchical semantic Gaussian Splatting method that incorporates city-model priors for urban scene understanding. GS4City derives reliable image-aligned masks from Level of Detail (LoD) 3 CityGML models via two-pass raycasting, explicitly using parent-child relations to validate and recover fine-grained facade elements. It then fuses these geometry-grounded masks with foundation-model predictions to establish scene-consistent instance correspondences, and learns a compact identity encoding for each Gaussian under joint 2D identity supervision and 3D spatial regularization. Experiments on the TUM2TWIN and Gold Coast datasets show that GS4City effectively incorporates structured building semantics into Gaussian scene representations, outperforming existing 2D-driven semantic 3DGS baselines, including LangSplat and Gaga, by up to 15.8 IoU points in coarse building segmentation and 14.2 mIoU points in fine-grained semantic segmentation. By bridging structured city models and photorealistic Gaussian scene representations, GS4City enables semantically queryable and structure-aware urban reconstruction. Code is available at https://github.com/Jinyzzz/GS4City.

Method: Proposes LangSCD with: 1) modular language component using VLMs to generate textual descriptions of scene changes, 2) cross-modal feature enhancer fusing language and visual features, 3) geometric-semantic matching module refining masks with semantic consistency and spatial completeness, and 4) NYC-CD dataset with 8,122 real-world image pairs and multiclass annotations.

Result: Extensive experiments show language and matching modules consistently improve existing change-detection architectures, achieving state-of-the-art performance across multiple street-view benchmarks. The NYC-CD dataset provides valuable resource for fine-grained scene change analysis.

Conclusion: Integrating linguistic reasoning with visual representations enables robust scene change detection in complex urban environments. The vision-language approach overcomes limitations of single-modal methods and provides semantic understanding beyond binary change detection.

Abstract: Scene change detection (SCD) is crucial for urban monitoring and navigation but remains challenging in real-world environments due to lighting variations, seasonal shifts, viewpoint differences, and complex urban layouts. Existing methods rely primarily on low-level visual features, limiting their ability to accurately identify changed objects amid the visual complexity of urban scenes. In this paper, we propose LangSCD, a vision-language framework for scene change detection that overcomes this single-modal limitation by incorporating semantic reasoning through language. Our approach introduces a modular language component that leverages vision-language models (VLMs) to generate textual descriptions of scene changes, which are fused with visual features through a cross-modal feature enhancer. We further introduce a geometric-semantic matching module that refines the predicted masks by enforcing semantic consistency and spatial completeness. Existing real-world scene change detection benchmarks provide only binary change annotations, which are insufficient for downstream applications requiring fine-grained understanding of scene dynamics. To address this limitation, we introduce NYC-CD, a large-scale dataset of 8,122 real-world image pairs collected in New York City with multiclass change annotations generated through a semi-automatic pipeline. Extensive experiments across multiple street-view benchmarks demonstrate that our language and matching modules consistently improve existing change-detection architectures, achieving state-of-the-art performance and highlighting the value of integrating linguistic reasoning with visual representations for robust scene change detection.

Method: Proposes ORVOSB benchmark with frame-level causal annotations and referent-shift labels (210 videos, 12,907 frames, 512 queries across 5 reasoning categories). Introduces baseline with continually-updated segmentation prompts and structured temporal token reservoir for long-horizon reasoning under bounded computation.

Result: Experiments show existing methods struggle under strict causality and referent shifts, while the proposed baseline establishes a strong foundation for future research in online reasoning video object segmentation.

Conclusion: ORVOS addresses the practical gap in video segmentation by introducing causal evaluation and methods, enabling real-world deployment of language-guided video understanding systems that must make incremental decisions without future information.

Abstract: Reasoning video object segmentation predicts pixel-level masks in videos from natural-language queries that may involve implicit and temporally grounded references. However, existing methods are developed and evaluated in an offline regime, where the entire video is available at inference time and future frames can be exploited for retrospective disambiguation, deviating from real-world deployments that require strictly causal, frame-by-frame decisions. We study Online Reasoning Video Object Segmentation (ORVOS), where models must incrementally interpret queries using only past and current frames without revisiting previous predictions, while handling referent shifts as events unfold. To support evaluation, we introduce ORVOSB, a benchmark with frame-level causal annotations and referent-shift labels, comprising 210 videos, 12,907 annotated frames, and 512 queries across five reasoning categories. We further propose a baseline with continually-updated segmentation prompts and a structured temporal token reservoir for long-horizon reasoning under bounded computation. Experiments show that existing methods struggle under strict causality and referent shifts, while our baseline establishes a strong foundation for future research.

Method: Cannot determine method as paper content is unavailable

Result: Cannot determine results as paper content is unavailable

Conclusion: Cannot draw conclusion as paper content is unavailable

Abstract: Failed to fetch summary for 2505.23606: Page request resulted in HTTP 429 (https://export.arxiv.org/api/query?search_query=&id_list=2505.23606&sortBy=relevance&sortOrder=descending&start=0&max_results=100)

Method: Formulates cross-scale remote sensing understanding as a unified cost-aware problem that couples fine-grained HR sampling with cross-patch representation prediction. Introduces GL-10M benchmark with 10 million spatially aligned multi-resolution images for systematic evaluation of budget-constrained cross-scale reasoning.

Result: Extensive experiments on recognition and retrieval tasks show the method consistently achieves superior performance-cost trade-off compared to existing approaches.

Conclusion: The proposed unified framework enables more effective task reasoning with fewer HR observations by addressing limitations of existing isolated patch-based sampling methods through fine-grained importance assessment and cross-patch contextual modeling.

Abstract: Remote sensing understanding inherently requires multi-resolution observation, since different targets and application tasks demand different levels of spatial detail. While low-resolution (LR) imagery enables efficient global observation, high-resolution (HR) imagery provides critical local details at much higher acquisition cost and limited coverage. This motivates a cross-scale sensing strategy that selectively acquires HR imagery from LR-based global perception to improve task performance under constrained cost. Existing methods for HR sampling methods typically make selection decisions from isolated LR patches, which ignore fine-grained intra-patch importance and cross-patch contextual interactions, leading to fragmented feature representation and suboptimal scene reasoning under sparse HR observations. To address this issue, we formulate cross-scale remote sensing understanding as a unified cost-aware problem that couples fine-grained HR sampling with cross-patch representation prediction, enabling more effective task reasoning with fewer HR observations. Furthermore, we present GL-10M, a large-scale benchmark of 10 million spatially aligned multi-resolution images, enabling systematic evaluation of budget-constrained cross-scale reasoning in remote sensing. Extensive experiments on recognition and retrieval tasks show that our method consistently achieves a superior performance-cost trade-off.

Method: Proposes HuiYanEarth-SAR based on AlphaEarth with integrated scattering mechanisms. Uses geospatial priors to control macroscopic structures and implicit scattering characteristic modeling for microscopic texture authenticity. Generates SAR images solely from geographic coordinates.

Result: Achieves capability to generate high-fidelity SAR images for global locations based on geographic coordinates, creating an efficient SAR scene simulator that bridges geography, scattering mechanisms, and AI.

Conclusion: Advances SAR research from perception/understanding to simulation/creation, provides key technical support for constructing high-confidence digital twin Earth, and establishes new paradigm connecting geography, scattering mechanisms, and AI.

Abstract: Synthetic Aperture Radar (SAR) imagery generation is essential for deepening the study of scattering mechanisms, establishing trustworthy electromagnetic scene models, and fundamentally alleviating the data scarcity bottleneck that constrains development in this field. However, existing methods find it difficult to simultaneously ensure high fidelity in both global geospatial semantics and microscopic scattering mechanisms, resulting in severe challenges for global generation. To address this, we propose \textbf{HuiYanEarth-SAR}, the first foundational SAR imagery generation model based on AlphaEarth and integrated scattering mechanisms. By injecting geospatial priors to control macroscopic structures and utilizing implicit scattering characteristic modeling to ensure the authenticity of microscopic textures, we achieve the capability of generating high-fidelity SAR images for global locations solely based on geographic coordinates. This study not only constructs an efficient SAR scene simulator but also establishes a bridge connecting geography, scatter mechanism, and artificial intelligence from a methodological standpoint. It advances SAR research by expanding the paradigm from perception and understanding to simulation and creation, providing key technical support for constructing a high-confidence digital twin of the Earth.

Method: Enhances Restormer baseline with expanded training corpus (larger and more diverse public image datasets), two-stage optimization, and ×8 geometric self-ensemble at inference. Retains TLC-style local inference wrapper for consistency.

Result: Achieves 30.762 dB PSNR and 0.861 SSIM on challenge validation set, improving over baseline Restormer by up to 3.366 dB PSNR. Dominant gains from expanded training corpus and two-stage optimization, with self-ensemble providing marginal but consistent improvement.

Conclusion: Data-centric training enhancements and test-time self-ensemble can significantly boost performance of mature restoration architectures without requiring new backbone designs, demonstrating the importance of training strategies and inference optimization.

Abstract: This paper presents our solution to the NTIRE 2026 Image Denoising Challenge (Gaussian color image denoising at fixed noise level $σ= 50$). Rather than proposing a new restoration backbone, we revisit the performance boundary of the mature Restormer architecture from two complementary directions: stronger data-centric training and more complete Test-Time capability release. Starting from the public Restormer $σ!=!50$ baseline, we expand the standard multi-dataset training recipe with larger and more diverse public image corpora and organize optimization into two stages. At inference, we apply $\times 8$ geometric self-ensemble to further release model capacity. A TLC-style local inference wrapper is retained for implementation consistency; however, systematic ablation reveals its quantitative contribution to be negligible in this setting. On the challenge validation set of 100 images, our final submission achieves 30.762 dB PSNR and 0.861 SSIM, improving over the public Restormer $σ!=!50$ pretrained baseline by up to 3.366 dB PSNR. Ablation studies show that the dominant gain originates from the expanded training corpus and the two-stage optimization schedule, and self-ensemble provides marginal but consistent improvement.

Method: Proposes two key modules: 1) Degradation-aware Token Injection - encodes lightweight degradation statistics from low-resolution inputs and fuses them with semantic conditioning features for explicit degradation-aware restoration. 2) Spatially Asymmetric Noise Injection - modulates diffusion noise with local edge strength to better preserve structural regions during training. Both are lightweight add-ons to the diffusion SR framework.

Result: Experiments on DIV2K and RealSR datasets show competitive no-reference perceptual quality and visually more realistic restoration results than recent baselines, while maintaining favorable perception-distortion trade-off. Ablations confirm effectiveness of each module and their complementary gains.

Conclusion: The proposed degradation-aware and structure-preserving diffusion framework effectively handles real-world SR challenges by explicitly modeling degradations and preserving structural details through lightweight modifications to diffusion conditioning.

Abstract: Real-world image super-resolution is particularly challenging for diffusion models because real degradations are complex, heterogeneous, and rarely modeled explicitly. We propose a degradation-aware and structure-preserving diffusion framework for real-world SR. Specifically, we introduce Degradation-aware Token Injection, which encodes lightweight degradation statistics from low-resolution inputs and fuses them with semantic conditioning features, enabling explicit degradation-aware restoration. We further propose Spatially Asymmetric Noise Injection, which modulates diffusion noise with local edge strength to better preserve structural regions during training. Both modules are lightweight add-ons to the adopted diffusion SR framework, requiring only minor modifications to the conditioning pipeline. Experiments on DIV2K and RealSR show that our method delivers competitive no-reference perceptual quality and visually more realistic restoration results than recent baselines, while maintaining a favorable perception–distortion trade-off. Ablations confirm the effectiveness of each module and their complementary gains when combined. The code and model are publicly available at https://github.com/jiyang0315/DASP-SR.git.

Method: Proposes PACO framework with hierarchical decision tree: known-class routing, birth-aware novel assignment, and attach-versus-create operations over dynamic prototype memory. Simulates proxy discovery during offline training to initialize thresholds, which are continuously updated during inference using mature novel prototypes.

Result: Significant improvements over state-of-the-art baselines across seven benchmarks, showing that properly calibrated and adaptive thresholds can substantially improve performance even without changing the underlying representation.

Conclusion: PACO provides an effective inference-time module for OCD that addresses fundamental flaws in existing approaches by treating category discovery as a dynamic process with adaptive decision-making, requiring no heavy training or dataset-specific tuning.

Abstract: On-the-Fly Category Discovery (OCD) requires a model, trained on an offline support set, to recognize known classes while discovering new ones from an online streaming sequence. Existing methods focus heavily on offline training. They aim to learn discriminative representations on the support set so that novel classes can be separated at test time. However, their discovery mechanism at inference is typically reduced to a single threshold. We argue that this paradigm is fundamentally flawed as OCD is not a static classification problem, but a dynamic process. The model must continuously decide 1) whether a sample belongs to a known class, 2) matches an existing novel category, or 3) should initiate a new one. Moreover, prior methods treat the support set as fixed knowledge. They do not update their decision boundaries as new evidence arrives during inference. This leads to unstable and inconsistent category formation. Our experiments confirm these issues. With properly calibrated and adaptive thresholds, substantial improvements can be achieved, even without changing the representation. Motivated by this, we propose PACO, a support-set-calibrated, tree-structured online decision framework. The framework models inference as a sequence of hierarchical decisions, including known-class routing, birth-aware novel assignment, and attach-versus-create operations over a dynamic prototype memory. Furthermore, we simulate the proxy discovery process to initialize the thresholds during offline training to align with inference. Thresholds are continuously updated during inference using mature novel prototypes. Importantly, PACO requires no heavy training and no dataset-specific tuning. It can be directly integrated into existing OCD pipelines as an inference-time module. Extensive experiments show significant improvements over SOTA baselines across seven benchmarks.

Method: Challenge-based approach using a novel dataset of 108,750 real and 185,750 AI-generated images from 42 different generators, augmented with 36 image transformations. Participants developed detection models evaluated using ROC AUC on transformed and untransformed test images.

Result: 511 participants registered with 20 teams submitting valid solutions. The challenge produced state-of-the-art detection methods robust to real-world image transformations.

Conclusion: The challenge successfully advanced robust AI-generated image detection methods that work under realistic conditions, providing valuable benchmarks and solutions for practical applications.

Abstract: This paper presents an overview of the NTIRE 2026 Challenge on Robust AI-Generated Image Detection in the Wild, held in conjunction with the NTIRE workshop at CVPR 2026. The goal of this challenge was to develop detection models capable of distinguishing real images from generated ones in realistic scenarios: the images are often transformed (cropped, resized, compressed, blurred) for practical usage, and therefore, the detection models should be robust to such transformations. The challenge is based on a novel dataset consisting of 108,750 real and 185,750 AI-generated images from 42 generators comprising a large variety of open-source and closed-source models of various architectures, augmented with 36 image transformations. Methods were evaluated using ROC AUC on the full test set, including both transformed and untransformed images. A total of 511 participants registered, with 20 teams submitting valid final solutions. This report provides a comprehensive overview of the challenge, describes the proposed solutions, and can be used as a valuable reference for researchers and practitioners in increasing the robustness of the detection models to real-world transformations.

Method: TAG-Head combines a Transformer encoder with learnable 3D positional encodings to capture long-range dependencies, followed by a spatio-temporal graph with two types of edges: (1) fully-connected intra-frame edges to resolve subtle appearance differences within frames, and (2) time-aligned temporal edges connecting features at same spatial locations across frames to stabilize motion cues without over-smoothing. The head is lightweight, plug-and-play across various 3D backbones (SlowFast, R(2+1)D-34, I3D), and trained end-to-end.

Result: TAG-Head achieves state-of-the-art performance among RGB-only models on FineGym (Gym99 and Gym288) and HAA500 datasets. It surpasses many recent multimodal approaches that use privileged information (video + pose + text). The design is compact with minimal parameter/FLOP overhead and low latency.

Conclusion: TAG-Head advances fine-grained human action recognition by explicitly coupling global context with high-resolution spatial interactions and low-variance temporal continuity in a slim, composable graph head. The RGB-only approach enables practical adoption in systems favoring simple sensor setups while delivering performance gains typically associated with heavier multimodal models.

Abstract: Fine-grained human action recognition (FHAR) is challenging because visually similar actions differ by subtle spatio-temporal cues. Many recent systems enhance discriminability with extra modalities (e.g., pose, text, optical flow), but this increases annotation burden and computational cost. We introduce TAG-Head, a lightweight spatio-temporal graph head that upgrades standard 3D backbones (SlowFast, R(2+1)D-34, I3D, etc.) for FHAR using RGB only. Our pipeline first applies a Transformer encoder with learnable 3D positional encodings to the backbone tokens, capturing long-range dependencies across space and time. The resulting features are then refined by a graph in which (i) fully-connected intra-frame edges to resolve subtle appearance differences within frames, and (ii) time-aligned temporal edges that connect features at the same spatial location across frames to stabilise motion cues without over-smoothing. The head is compact (little parameter/FLOP overhead), plug-and-play across backbones, and trained end-to-end with the backbone. Extensive evaluations on FineGym (Gym99 and Gym288) and HAA500 show that TAG-Head sets a new state-of-the-art among RGB-only models and surpasses many recent multimodal approaches (video + pose + text) that rely on privileged information. Ablations disentangle the contributions of the Transformer and the graph topology, and complexity analyses confirm low latency. TAG-Head advances FHAR by explicitly coupling global context with high-resolution spatial interactions and low-variance temporal continuity inside a slim, composable graph head. The simplicity of the design enables straightforward adoption in practical systems that favour RGB-only sensors, while delivering performance gains typically associated with heavier or multimodal models. Code will be released on GitHub.

Method: Proposes SVD-Prune, a training-free, plug-and-play token pruning method based on Singular Value Decomposition. It decomposes the vision token feature matrix and selects top-K tokens using statistical leverage scores, ensuring only tokens contributing most to the dominant global variance are preserved.

Result: Experiments show SVD-Prune consistently outperforms prior pruning methods under extreme vision token budgets, maintaining strong performance even with only 32 and 16 vision tokens.

Conclusion: SVD-Prune effectively addresses computational challenges in VLMs by preserving essential visual content through global variance-based token selection, enabling efficient multimodal processing without performance degradation.

Abstract: Vision-Language Models (VLM) have revolutionized multimodal learning by jointly processing visual and textual information. Yet, they face significant challenges due to the high computational and memory demands of processing long sequences of vision tokens. Many existing methods rely on local heuristics, such as attention scores or token norms. However, these criteria suffer from positional bias and information dispersion, limiting their ability to preserve essential content at high pruning ratios and leading to performance degradation on visually detailed images. To address these issues, we propose SVD-Prune, a trainingfree, plug-and-play token pruning method based on Singular Value Decomposition. It decomposes the vision token feature matrix and selects the top-K tokens using statistical leverage scores, ensuring only tokens contributing most to the dominant global variance are preserved. Experiments show that SVD-Prune consistently outperforms prior pruning methods under extreme vision token budgets, maintaining strong performance even with 32 and 16 vision tokens.

Method: Reframes pretrained VLM embedding spaces as text-conditional similarity spaces without additional training, separating textual conditioning from visual feature extraction to enable efficient multi-conditioned retrieval with fixed visual embeddings.

Result: Achieves high retrieval accuracy and notable computational efficiency on standard datasets and the proposed CLAY-EVAL synthetic evaluation dataset across diverse conditioned retrieval settings.

Conclusion: CLAY provides an effective solution for adaptive, multi-conditioned image retrieval by leveraging pretrained VLMs without retraining, offering both accuracy and efficiency advantages over previous methods.

Abstract: Human perception of visual similarity is inherently adaptive and subjective, depending on the users’ interests and focus. However, most image retrieval systems fail to reflect this flexibility, relying on a fixed, monolithic metric that cannot incorporate multiple conditions simultaneously. To address this, we propose CLAY, an adaptive similarity computation method that reframes the embedding space of pretrained Vision-Language Models (VLMs) as a text-conditional similarity space without additional training. This design separates the textual conditioning process and visual feature extraction, allowing highly efficient and multi-conditioned retrieval with fixed visual embeddings. We also construct a synthetic evaluation dataset CLAY-EVAL, for comprehensive assessment under diverse conditioned retrieval settings. Experiments on standard datasets and our proposed dataset show that CLAY achieves high retrieval accuracy and notable computational efficiency compared to previous works.

Method: PTD uses a coarse-to-fine framework with two modules: PTD_rec learns deterministic mapping to recover low-frequency signals (coarse content), and PTD_diff uses dual-domain guided conditional diffusion to generate reliable textures for the coarse prediction. This approach reduces randomness and enables better trade-off between visual quality and detail fidelity.

Result: Extensive experiments show PTD achieves superior performance in structure similarity and visual appeal with few sampling steps, mitigating diffusion model randomness while balancing visual quality and high-frequency detail fidelity.

Conclusion: The PTD model effectively addresses texture consistency challenges in sparse-view CT reconstruction through a progressive coarse-to-fine approach that combines deterministic and diffusion-based methods for improved visual quality and detail preservation.

Abstract: Diffusion-based sparse-view CT (SVCT) imaging has achieved remarkable advancements in recent years, thanks to its more stable generative capability. However, recovering reliable image content and visually consistent textures is still a crucial challenge. In this paper, we present a Progressively Texture-aware Diffusion (PTD) model, a coarse-to-fine learning framework tailored for SVCT. Specifically, PTD comprises a basic reconstructive module PTD${\textit{rec}}$ and a conditional diffusion module PTD${\textit{diff}}$. PTD${\textit{rec}}$ first learns a deterministic mapping to recover the majority of the underlying low-frequency signals (i.e., coarse content with smoothed textures), which serves as the initial estimation to enable fidelity. Moreover, PTD${\textit{diff}}$ aims to reconstruct high-fidelity details for coarse prediction, which explores a dual-domain guided conditional diffusion to generate reliable and consistent textures. Extensive experiments on sparse-view CT reconstruction demonstrate that our PTD achieves superior performance in terms of structure similarity and visual appeal with only a few sampling steps, which mitigates the randomness inherent in general diffusion models and enables a better trade-off between visual quality and fidelity of high-frequency details.

Method: Cannot determine method as paper content is unavailable

Result: Cannot determine results as paper content is unavailable

Conclusion: Cannot draw conclusion as paper content is unavailable

Abstract: Failed to fetch summary for 2508.09165: Page request resulted in HTTP 429 (https://export.arxiv.org/api/query?search_query=&id_list=2508.09165&sortBy=relevance&sortOrder=descending&start=0&max_results=100)

Method: Empirical study of three FL strategies (FedAvg, FedProx, and bulk synchronous parallel) applied to multi-label remote sensing image classification under controlled non-IID label-skew conditions. Evaluated three CNN architectures (LeNet, AlexNet, ResNet-34) and analyzed effects of algorithm choice, model capacity, client fraction, client count, batch size, and communication cost.

Result: FedProx outperforms FedAvg for deeper architectures under data heterogeneity; BSP approaches centralized accuracy but with high sequential communication cost; LeNet provides the best accuracy-communication trade-off for the dataset scale considered.

Conclusion: Federated learning is essential for distributed remote sensing data, but algorithm choice and model architecture significantly impact performance under non-IID conditions. FedProx shows advantages for deeper models, while simpler architectures offer better communication efficiency trade-offs.

Abstract: Remote sensing archives are inherently distributed: Earth observation missions such as Sentinel-1, Sentinel-2, and Sentinel-3 have collectively accumulated more than 5 petabytes of imagery, stored and processed across many geographically dispersed platforms. Training machine learning models on such data in a centralized fashion is impractical due to data volume, sovereignty constraints, and geographic distribution. Federated learning (FL) addresses this by keeping data local and exchanging only model updates. A central challenge for remote sensing is the non-IID nature of Earth observation data: label distributions vary strongly by geographic region, degrading the convergence of standard FL algorithms. In this paper, we conduct a systematic empirical study of three FL strategies – FedAvg, FedProx, and bulk synchronous parallel (BSP) – applied to multi-label remote sensing image classification under controlled non-IID label-skew conditions. We evaluate three convolutional neural network (CNN) architectures of increasing depth (LeNet, AlexNet, and ResNet-34) and analyze the joint effect of algorithm choice, model capacity, client fraction, client count, batch size, and communication cost. Experiments on the UC Merced multi-label dataset show that FedProx outperforms FedAvg for deeper architectures under data heterogeneity, that BSP approaches centralized accuracy at the cost of high sequential communication, and that LeNet provides the best accuracy-communication trade-off for the dataset scale considered.

Method: Dual-branch pipeline: 1) Hybrid attention network with TLC inference for stable main reconstruction, 2) MambaIRv2 with geometric self-ensemble for high-frequency detail recovery. Branches process same input independently and fuse via lightweight weighted combination without parameter updates.

Result: Consistently improves over base branch and slightly exceeds pure strong branch in PSNR under DIV2K bicubic ×4 evaluation. Provides low-overhead upgrade path for existing systems.

Conclusion: Training-free output-level ensemble offers practical alternative to architectural redesign, enabling effective combination of existing models without additional training costs.

Abstract: Single-image super-resolution has progressed from deep convolutional baselines to stronger Transformer and state-space architectures, yet the corresponding performance gains typically come with higher training cost, longer engineering iteration, and heavier deployment burden. In many practical settings, multiple pretrained models with partially complementary behaviors are already available, and the binding constraint is no longer architectural capacity but how effectively their outputs can be combined without additional training. Rather than pursuing further architectural redesign, this paper proposes a training-free output-level ensemble framework. A dual-branch pipeline is constructed in which a Hybrid attention network with TLC inference provides stable main reconstruction, while a MambaIRv2 branch with geometric self-ensemble supplies strong compensation for high-frequency detail recovery. The two branches process the same low-resolution input independently and are fused in the image space via a lightweight weighted combination, without updating any model parameters or introducing an additional trainable module. As our solution to the NTIRE 2026 Image Super-Resolution ($\times 4$) Challenge, the proposed design consistently improves over the base branch and slightly exceeds the pure strong branch in PSNR at the best operating point under a unified DIV2K bicubic $\times 4$ evaluation protocol. Ablation studies confirm that output-level compensation provides a low-overhead and practically accessible upgrade path for existing super-resolution systems.

Method: AdvFLYP finetunes CLIP with adversarial images created from web image-text pairs using contrastive loss, plus regularization terms: logit-level regularization for robustness and feature-level regularization for clean accuracy.

Result: Extensive experiments on 14 downstream datasets show superiority over mainstream practices in maintaining zero-shot capabilities while improving adversarial robustness across domains.

Conclusion: Following CLIP’s original pretraining recipe during adversarial finetuning with web data and appropriate regularization preserves zero-shot abilities while enhancing robustness transferability.

Abstract: Despite their impressive zero-shot abilities, vision-language models such as CLIP have been shown to be susceptible to adversarial attacks. To enhance its adversarial robustness, recent studies finetune the pretrained vision encoder of CLIP with adversarial examples on a proxy dataset such as ImageNet by aligning adversarial images with correct class labels. However, these methods overlook the important roles of training data distributions and learning objectives, resulting in reduced zero-shot capabilities and limited transferability of robustness across domains and datasets. In this work, we propose a simple yet effective paradigm AdvFLYP, which follows the training recipe of CLIP’s pretraining process when performing adversarial finetuning to the model. Specifically, AdvFLYP finetunes CLIP with adversarial images created based on image-text pairs collected from the web, and match them with their corresponding texts via a contrastive loss. To alleviate distortion of adversarial image embeddings of noisy web images, we further propose to regularise AdvFLYP by penalising deviation of adversarial image features. We show that logit- and feature-level regularisation terms benefit robustness and clean accuracy, respectively. Extensive experiments on 14 downstream datasets spanning various domains show the superiority of our paradigm over mainstream practices. Our code and model weights are released at https://github.com/Sxing2/AdvFLYP.

Method: Proposes a model that learns local visuo-tactile alignment through dense cross-modal feature interactions to generate tactile saliency maps. Introduces in-the-wild multi-material scene images to expand visual diversity and a material-diversity pairing strategy that aligns each tactile sample with visually varied but tactilely consistent images. Also constructs two new tactile-grounded material segmentation datasets for evaluation.

Result: Experiments on both new and existing benchmarks show the approach substantially outperforms prior visuo-tactile methods in tactile localization tasks.

Conclusion: The proposed method successfully addresses limitations of existing approaches by enabling fine-grained local visuo-tactile alignment and overcoming dataset constraints through innovative data strategies, achieving superior performance in tactile localization.

Abstract: We address the problem of tactile localization, where the goal is to identify image regions that share the same material properties as a tactile input. Existing visuo-tactile methods rely on global alignment and thus fail to capture the fine-grained local correspondences required for this task. The challenge is amplified by existing datasets, which predominantly contain close-up, low-diversity images. We propose a model that learns local visuo-tactile alignment via dense cross-modal feature interactions, producing tactile saliency maps for touch-conditioned material segmentation. To overcome dataset constraints, we introduce: (i) in-the-wild multi-material scene images that expand visual diversity, and (ii) a material-diversity pairing strategy that aligns each tactile sample with visually varied yet tactilely consistent images, improving contextual localization and robustness to weak signals. We also construct two new tactile-grounded material segmentation datasets for quantitative evaluation. Experiments on both new and existing benchmarks show that our approach substantially outperforms prior visuo-tactile methods in tactile localization.

Method: Cannot determine method as the paper abstract is unavailable due to HTTP 429 error from arXiv API

Result: Cannot determine results as the paper abstract is unavailable due to HTTP 429 error from arXiv API

Conclusion: Cannot determine conclusion as the paper abstract is unavailable due to HTTP 429 error from arXiv API

Abstract: Failed to fetch summary for 2604.11585: Page request resulted in HTTP 429 (https://export.arxiv.org/api/query?search_query=&id_list=2604.11585&sortBy=relevance&sortOrder=descending&start=0&max_results=100)

Method: The paper studies existing AR image generator watermarking schemes and demonstrates their vulnerabilities. It assesses existing attacks and introduces three new attacks: (1) vector-quantized regeneration removal attack, (2) adversarial optimization-based attack, and (3) frequency injection attack.

Result: Evaluation shows removal and forgery attacks can be effective with access to just a single watermarked reference image, without needing original model parameters or watermarking secrets. Existing watermarking schemes fail to reliably support synthetic content detection for dataset filtering.

Conclusion: Current watermarking schemes for AR image generation are vulnerable to attacks, enabling both watermark removal and “Watermark Mimicry” where authentic images can be manipulated to trigger false detection. This undermines their reliability for content verification and dataset filtering applications.

Abstract: The proliferation of autoregressive (AR) image generators demands reliable detection and attribution of their outputs to mitigate misinformation, and to filter synthetic images from training data to prevent model collapse. To address this need, watermarking techniques, specifically designed for AR models, embed a subtle signal at generation time, enabling downstream verification through a corresponding watermark detector. In this work, we study these schemes and demonstrate their vulnerability to both watermark removal and forgery attacks. We assess existing attacks and further introduce three new attacks: (i) a vector-quantized regeneration removal attack, (ii) adversarial optimization-based attack, and (iii) a frequency injection attack. Our evaluation reveals that removal and forgery attacks can be effective with access to a single watermarked reference image and without access to original model parameters or watermarking secrets. Our findings indicate that existing watermarking schemes for AR image generation do not reliably support synthetic content detection for dataset filtering. Moreover, they enable Watermark Mimicry, whereby authentic images can be manipulated to imitate a generator’s watermark and trigger false detection to prevent their inclusion in future model training.

Method: Proposes Philautia-Eval framework to quantify model-specific preference bias by disentangling preference tendencies from generation quality differences. Analyzes 1.29M caption-score pairs from 12 MLLMs. Introduces Pomms, a simple ensemble of MLLMs to mitigate bias.

Result: Found that representative MLLMs exhibit self-preference bias. Also discovered mutual preference bias within particular model families, potentially driven by reused connectors and overlapping instruction-tuning resources. Pomms ensemble effectively mitigated model-specific preference bias while maintaining performance.

Conclusion: MLLM-as-a-Judge methods have significant model-specific preference biases that could distort model comparisons. The proposed Philautia-Eval framework helps quantify these biases, and the Pomms ensemble approach offers a practical solution to mitigate bias in evaluation.

Abstract: Automatic evaluation using multimodal large language models (MLLMs), commonly referred to as MLLM-as-a-Judge, has been widely used to measure model performance. If such MLLM-as-a-Judge methods were biased, they could distort model comparisons and benchmark-driven scientific progress. However, it remains unclear to what extent MLLM-as-a-Judge methods favor or disfavor text generated by specific MLLMs. In this study, we propose Philautia-Eval to investigate such model-specific preference bias. Philautia-Eval quantifies the degree of the bias by disentangling preference tendencies from differences in generation quality. Using 1.29M caption-score pairs collected from 12 MLLMs, we found that representative MLLMs tend to exhibit self-preference bias. Moreover, experimental results indicate mutual preference bias within particular model families, which is potentially driven by reused connectors and overlapping instruction-tuning resources. Finally, we introduce a simple ensemble of MLLMs, Pomms. Our results demonstrated that Pomms effectively mitigated the model-specific preference bias while maintaining performance.

Method: Proposes a label-free adaptation framework that uses predictions from a non-robust teacher model as semantic anchors for both clean and adversarial objectives during test-time adaptation, providing theoretical analysis showing improved stability over self-consistency regularization.

Result: The approach achieves improved optimization stability, lower sensitivity to hyperparameters, and better robustness-accuracy trade-off on CIFAR-10 and ImageNet under photometric transformations compared to existing baselines.

Conclusion: Non-robust pretrained models can be effectively adapted at test-time to improve adversarial robustness using the proposed label-free framework with teacher predictions as semantic anchors.

Abstract: Nowadays, pretrained models are increasingly used as general-purpose backbones and adapted at test-time to downstream environments where target data are scarce and unlabeled. While this paradigm has proven effective for improving clean accuracy on the target domain, adversarial robustness has received far less attention, especially when the original pretrained model is not explicitly designed to be robust. This raises a practical question: \emph{can a pretrained, non-robust model be adapted at test-time to improve adversarial robustness on a target distribution?} To face this question, this work studies how adversarial training strategies behave when integrated into adaptation schemes for the unsupervised test-time setting, where only a small set of unlabeled target samples is available. It first analyzes how classical adversarial training formulations can be extended to this scenario, showing that straightforward distillation-based adaptations remain unstable and highly sensitive to hyperparameter tuning, particularly when the teacher itself is non-robust. To address these limitations, the work proposes a label-free framework that uses the predictions of a non-robust teacher model as a semantic anchor for both the clean and adversarial objectives during adaptation. We further provide theoretical insights showing that our formulation yields a more stable alternative to the self-consistency-based regularization commonly used in classical adversarial training. Experiments evaluate the proposed approach on CIFAR-10 and ImageNet under induced photometric transformations. The results support the theoretical insights by showing that the proposed approach achieves improved optimization stability, lower sensitivity to parameter choices, and a better robustness-accuracy trade-off than existing baselines in this post-deployment test-time setting.

Method: Design a unified formal language integrating plane and solid geometry, construct GDP-29K dataset (20k plane + 9k solid geometry samples), and propose training paradigm combining Supervised Fine-Tuning with Reinforcement Learning via Verifiable Rewards.

Result: Achieves state-of-the-art parsing performance. The parsed formal descriptions serve as a critical cognitive scaffold, significantly boosting MLLMs’ capabilities for downstream geometry reasoning tasks.

Conclusion: The approach successfully addresses geometric reasoning challenges in MLLMs through formal language parsing, with the unified framework covering both plane and solid geometry, and the training method ensuring syntactic correctness and geometric consistency.

Abstract: Multimodal Large Language Models (MLLMs) have achieved remarkable progress but continue to struggle with geometric reasoning, primarily due to the perception bottleneck regarding fine-grained visual elements. While formal languages have aided plane geometry understanding, solid geometry which requires spatial understanding remains largely unexplored. In this paper, we address this challenge by designing a unified formal language that integrates plane and solid geometry, comprehensively covering geometric structures and semantic relations. We construct GDP-29K, a large-scale dataset comprising 20k plane and 9k solid geometry samples collected from diverse real-world sources, each paired with its ground-truth formal description. To ensure syntactic correctness and geometric consistency, we propose a training paradigm that combines Supervised Fine-Tuning with Reinforcement Learning via Verifiable Rewards. Experiments show that our approach achieves state-of-the-art parsing performance. Furthermore, we demonstrate that our parsed formal descriptions serve as a critical cognitive scaffold, significantly boosting MLLMs’ capabilities for downstream geometry reasoning tasks. Our data and code are available at Geoparsing.

Method: Introduces a native dual-mode MLLM with dynamic visual token scaling inspired by human visual system. Features two complementary perception modes: focus mode (for fine-grained tasks) and standby mode (for long-form understanding). Includes dynamically detachable KV-cache design for streaming visual understanding.

Result: Standby mode retains 97.7-99.7% of original accuracy using only 1/40-1/10th of visual tokens for long-form visual understanding. The model efficiently maintains ultra-long visual memory for streaming scenarios.

Conclusion: POINTS-Long provides new insights for future MLLM design and lays foundation for adaptive, efficient long-form visual understanding, addressing key scalability challenges in multimodal AI.

Abstract: Multimodal Large Language Models (MLLMs) have recently demonstrated remarkable capabilities in cross-modal understanding and generation. However, the rapid growth of visual token sequences–especially in long-video and streaming scenarios–poses a major challenge to their scalability and real-world deployment. Thus, we introduce POINTS-Long, a native dual-mode MLLM featuring dynamic visual token scaling inspired by the human visual system. The model supports two complementary perception modes: focus mode and standby mode, enabling users to dynamically trade off efficiency and accuracy during inference. On fine-grained visual tasks, the focus mode retains the optimal performance, while on long-form general visual understanding, the standby mode retains 97.7-99.7% of the original accuracy using only 1/40-1/10th of the visual tokens. Moreover, POINTS-Long natively supports streaming visual understanding via a dynamically detachable KV-cache design, allowing efficient maintenance of ultra-long visual memory. Our work provides new insights into the design of future MLLMs and lays the foundation for adaptive and efficient long-form visual understanding.

Method: Unable to determine method due to fetch failure

Result: Unable to determine results due to fetch failure

Conclusion: Unable to determine conclusion due to fetch failure

Abstract: Failed to fetch summary for 2604.11757: Page request resulted in HTTP 429 (https://export.arxiv.org/api/query?search_query=&id_list=2604.11757&sortBy=relevance&sortOrder=descending&start=0&max_results=100)

Method: Combines neural implicit shape representations (resolution-agnostic 3D modeling) with autodecoder formulation (direct optimization of per-instance latent codes under sparse supervision) and autoregressive normalizing flows (capturing latent anatomical variability distribution). Uses adaptive latent relevance weighting with sparsity-inducing priors for compact, structured latent spaces.

Result: Accurate high-resolution reconstruction from sparse inputs on lumbar vertebrae and femur datasets. Recovers structured modes of anatomical variation consistent with population trends. Supports uncertainty quantification and anatomically plausible shape synthesis without manual latent dimensionality tuning.

Conclusion: MorphoFlow enables scalable, flexible statistical shape modeling from sparse annotations by integrating neural implicit representations with probabilistic generative modeling, providing a practical solution for population-level anatomical analysis.

Abstract: Statistical shape modeling (SSM) is central to population level analysis of anatomical variability, yet most existing approaches rely on densely annotated segmentations and fixed latent representations. These requirements limit scalability and reduce flexibility when modeling complex anatomical variation. We introduce MorphoFlow, a sparse supervised generative shape modeling framework that learns compact probabilistic shape representations directly from sparse surface annotations. MorphoFlow integrates neural implicit shape representations with an autodecoder formulation and autoregressive normalizing flows to learn an expressive probabilistic density over the latent shape space. The neural implicit representation enables resolution-agnostic modeling of 3D anatomy, while the autodecoder formulation supports direct optimization of per-instance latent codes under sparse supervision. The autoregressive flow captures the distribution of latent anatomical variability providing a tractable, likelihood-based generative model of shapes. To promote compact and structured latent representations, we incorporate adaptive latent relevance weighting through sparsity-inducing priors, enabling the model to regulate the contribution of individual latent dimensions according to their relevance to the underlying anatomical variation while preserving generative expressivity. The resulting latent space supports uncertainty quantification and anatomically plausible shape synthesis without manual latent dimensionality tuning. Evaluation on publicly available lumbar vertebrae and femur datasets demonstrates accurate high-resolution reconstruction from sparse inputs and recovery of structured modes of anatomical variation consistent with population level trends.

Method: Proposes efficient KernelSHAP framework that restricts computation to user-defined ROI and its receptive-field support, accelerates inference via patch logit caching (reusing baseline predictions for unaffected patches while preserving nnU-Net’s fusion scheme), and compares three feature abstractions: whole-organ units, regular FCC supervoxels, and hybrid organ-aware supervoxels.

Result: Caching reduces redundant computation by 15-30%; faithfulness vs interpretability trade-off: regular supervoxels maximize perturbation metrics but lack anatomical alignment, while organ-aware units yield more clinically interpretable explanations and effectively highlight false-positive drivers under normalized metrics.

Conclusion: The framework enables efficient and clinically meaningful explanations for 3D medical image segmentation, with organ-aware supervoxels providing better clinical interpretability for identifying false-positive drivers in medical imaging applications.

Abstract: Perturbation-based explainability methods such as KernelSHAP provide model-agnostic attributions but are typically impractical for patch-based 3D medical image segmentation due to the large number of coalition evaluations and the high cost of sliding-window inference. We present an efficient KernelSHAP framework for volumetric CT segmentation that restricts computation to a user-defined region of interest and its receptive-field support, and accelerates inference via patch logit caching, reusing baseline predictions for unaffected patches while preserving nnU-Net’s fusion scheme. To enable clinically meaningful attributions, we compare three automatically generated feature abstractions within the receptive-field crop: whole-organ units, regular FCC supervoxels, and hybrid organ-aware supervoxels, and we study multiple aggregation/value functions targeting stabilizing evidence (TP/Dice/Soft Dice) or false-positive behavior. Experiments on whole-body CT segmentations show that caching substantially reduces redundant computation (with computational savings ranging from 15% to 30%) and that faithfulness and interpretability exhibit clear trade-offs: regular supervoxels often maximize perturbation-based metrics but lack anatomical alignment, whereas organ-aware units yield more clinically interpretable explanations and are particularly effective for highlighting false-positive drivers under normalized metrics.

Method: Transform 4D point cloud videos into graph spectral signals, decompose them into multiple frequency bands (low-frequency for coarse shapes, high-frequency for fine-grained details), and design STS-Mixer - a unified framework that mixes spatial, temporal, and spectral representations through spectral analysis.

Result: STS-Mixer achieves superior performance consistently across multiple widely adopted benchmarks on both 3D action recognition and 4D semantic segmentation tasks, demonstrating its effectiveness in capturing rich geometries and temporal dynamics.

Conclusion: The spectral perspective provides a powerful complementary approach to traditional spatiotemporal methods for 4D point cloud video understanding, enabling fine-grained and holistic analysis through integrated spatial, temporal, and spectral representations.

Abstract: 4D point cloud videos capture rich spatial and temporal dynamics of scenes which possess unique values in various 4D understanding tasks. However, most existing methods work in the spatiotemporal domain where the underlying geometric characteristics of 4D point cloud videos are hard to capture, leading to degraded representation learning and understanding of 4D point cloud videos. We address the above challenge from a complementary spectral perspective. By transforming 4D point cloud videos into graph spectral signals, we can decompose them into multiple frequency bands each of which captures distinct geometric structures of point cloud videos. Our spectral analysis reveals that the decomposed low-frequency signals capture more coarse shapes while high-frequency signals encode more fine-grained geometry details. Building on these observations, we design Spatio-Temporal-Spectral Mixer (STS-Mixer), a unified framework that mixes spatial, temporal, and spectral representations of point cloud videos. STS-Mixer integrates multi-band delineated spectral signals with spatiotemporal information to capture rich geometries and temporal dynamics, while enabling fine-grained and holistic understanding of 4D point cloud videos. Extensive experiments show that STS-Mixer achieves superior performance consistently across multiple widely adopted benchmarks on both 3D action recognition and 4D semantic segmentation tasks. Code and models are available at https://github.com/Vegetebird/STS-Mixer.

Method: Collected 960 gaze recordings from 16 expert radiologists interpreting 30 real and 30 synthetic chest X-rays under diagnostic assessment and real-fake classification conditions. Extended protocol to 6 state-of-the-art multimodal LLMs to generate predictions under matched conditions.

Result: Dataset includes raw gaze samples, fixation maps, scanpaths, saliency density maps, diagnostic labels, authenticity judgments, and corresponding LLM predictions with confidence scores. Provides analyses of gaze agreement, inter-observer consistency, and benchmarking of radiologists versus LLMs.

Conclusion: GazeVaLM supports research in gaze modeling, clinical decision-making, human-AI comparison, generative image realism assessment, and uncertainty quantification, facilitating reproducible research on expert and AI perception of medical images.

Abstract: We introduce GazeVaLM, a public eye-tracking dataset for studying clinical perception during chest radiograph authenticity assessment. The dataset comprises 960 gaze recordings from 16 expert radiologists interpreting 30 real and 30 synthetic chest X-rays (generated by diffusion based generative AI) under two conditions: diagnostic assessment and real-fake classification (Visual Turing test). For each image-observer pair, we provide raw gaze samples, fixation maps, scanpaths, saliency density maps, structured diagnostic labels, and authenticity judgments. We extend the protocol to 6 state-of-the-art multimodal LLMs, releasing their predicted diagnoses, authenticity labels, and confidence scores under matched conditions - enabling direct human-AI comparison at both decision and uncertainty levels. We further provide analyses of gaze agreement, inter-observer consistency, and benchmarking of radiologists versus LLMs in diagnostic accuracy and authenticity detection. GazeVaLM supports research in gaze modeling, clinical decision-making, human-AI comparison, generative image realism assessment, and uncertainty quantification. By jointly releasing visual attention data, clinical labels, and model predictions, we aim to facilitate reproducible research on how experts and AI systems perceive, interpret, and evaluate medical images. The dataset is available at https://huggingface.co/datasets/davidcwong/GazeVaLM.

Method: Proposed a budget-aware uncertainty-driven quality assurance framework built on nnU-Net, combining uncertainty quantification and post-hoc calibration to produce voxel-wise uncertainty maps based on predictive entropy. Compared temperature scaling (TS), deep ensembles (DE), checkpoint ensembles (CE), and test-time augmentation (TTA), evaluated individually and in combination on TMLI as a representative use case.

Result: Segmentation accuracy remained stable across configurations, while temperature scaling substantially improved calibration. Uncertainty-error alignment improved most with calibrated checkpoint-based inference, leading to uncertainty maps that more consistently highlight regions requiring manual edits. Reliability was assessed through ROI-masked calibration metrics and uncertainty-error alignment under realistic revision constraints.

Conclusion: Integrating calibration with efficient ensembling appears to be a promising strategy to implement a budget-aware QA workflow for radiotherapy segmentation, providing uncertainty maps that can guide targeted manual review.

Abstract: Accurate delineation of the Clinical Target Volume (CTV) is essential for radiotherapy planning, yet remains time-consuming and difficult to assess, especially for complex treatments such as Total Marrow and Lymph Node Irradiation (TMLI). While deep learning-based auto-segmentation can reduce workload, safe clinical deployment requires reliable cues indicating where models may be wrong. In this work, we propose a budget-aware uncertainty-driven quality assurance (QA) framework built on nnU-Net, combining uncertainty quantification and post-hoc calibration to produce voxel-wise uncertainty maps (based on predictive entropy) that can guide targeted manual review. We compare temperature scaling (TS), deep ensembles (DE), checkpoint ensembles (CE), and test-time augmentation (TTA), evaluated both individually and in combination on TMLI as a representative use case. Reliability is assessed through ROI-masked calibration metrics and uncertainty–error alignment under realistic revision constraints, summarized as AUC over the top 0-5% most uncertain voxels. Across configurations, segmentation accuracy remains stable, whereas TS substantially improves calibration. Uncertainty-error alignment improves most with calibrated checkpoint-based inference, leading to uncertainty maps that highlight more consistently regions requiring manual edits. Overall, integrating calibration with efficient ensembling seems a promising strategy to implement a budget-aware QA workflow for radiotherapy segmentation.

Method: Proposes UNIGEOCLIP with all-to-all contrastive alignment across five geospatial modalities, plus a scaled latitude-longitude encoder that captures multi-scale geographic structure for improved spatial representation.

Result: Extensive experiments show UNIGEOCLIP consistently outperforms single-modality contrastive models and coordinate-only baselines across downstream geospatial tasks, demonstrating benefits of holistic multimodal alignment.

Conclusion: UNIGEOCLIP successfully creates a unified multimodal embedding space for geospatial data, enabling seamless comparison, retrieval, and reasoning across arbitrary combinations of five complementary geospatial modalities.

Abstract: The growing availability of co-located geospatial data spanning aerial imagery, street-level views, elevation models, text, and geographic coordinates offers a unique opportunity for multimodal representation learning. We introduce UNIGEOCLIP, a massively multimodal contrastive framework to jointly align five complementary geospatial modalities in a single unified embedding space. Unlike prior approaches that fuse modalities or rely on a central pivot representation, our method performs all-to-all contrastive alignment, enabling seamless comparison, retrieval, and reasoning across arbitrary combinations of modalities. We further propose a scaled latitude-longitude encoder that improves spatial representation by capturing multi-scale geographic structure. Extensive experiments across downstream geospatial tasks demonstrate that UNIGEOCLIP consistently outperforms single-modality contrastive models and coordinate-only baselines, highlighting the benefits of holistic multimodal geospatial alignment. A reference implementation is available at https://gastruc.github.io/unigeoclip.

Method: Organized FOMO25 challenge with large pretraining dataset FOMO60K, evaluating models on clinical data in few-shot/out-of-domain settings. Tasks included infarct classification, meningioma segmentation, and brain age regression. Evaluated 19 foundation models from 16 teams using standardized containerized pipeline.

Result: SSL pretraining improves generalization on clinical data under domain shift; strongest out-of-domain models surpassed supervised in-domain baselines. No single pretraining objective benefits all tasks: MAE favors segmentation, hybrid reconstruction-contrastive favors classification. Small pretrained models achieved strong performance without reliable benefits from scaling model size/training duration.

Conclusion: Self-supervised foundation models show promise for clinical brain MRI analysis, with different objectives optimal for different tasks and efficient small models performing well without extensive scaling.

Abstract: Clinical deployment of automated brain MRI analysis faces a fundamental challenge: clinical data is heterogeneous and noisy, and high-quality labels are prohibitively costly to obtain. Self-supervised learning (SSL) can address this by leveraging the vast amounts of unlabeled data produced in clinical workflows to train robust \textit{foundation models} that adapt out-of-domain with minimal supervision. However, the development of foundation models for brain MRI has been limited by small pretraining datasets and in-domain benchmarking focused on high-quality, research-grade data. To address this gap, we organized the FOMO25 challenge as a satellite event at MICCAI 2025. FOMO25 provided participants with a large pretraining dataset, FOMO60K, and evaluated models on data sourced directly from clinical workflows in few-shot and out-of-domain settings. Tasks covered infarct classification, meningioma segmentation, and brain age regression, and considered both models trained on FOMO60K (method track) and any data (open track). Nineteen foundation models from sixteen teams were evaluated using a standardized containerized pipeline. Results show that (a) self-supervised pretraining improves generalization on clinical data under domain shift, with the strongest models trained \textit{out-of-domain} surpassing supervised baselines trained \textit{in-domain}. (b) No single pretraining objective benefits all tasks: MAE favors segmentation, hybrid reconstruction-contrastive objectives favor classification, and (c) strong performance was achieved by small pretrained models, and improvements from scaling model size and training duration did not yield reliable benefits.

Method: Proposes a top-down “unfolding” scheme starting from full-resolution 3DGS model, iteratively deriving coarser LODs using adaptive learnable mask-based pruning. Uses hierarchical spatial grids for global structure and shared Anchor Codebook for localized details, creating compact feature representation with minimal overhead for progressive refinement.

Result: Maintains high rendering quality across all LODs while achieving substantial storage reduction, demonstrating practicality for real-time 3DGS rendering in bandwidth- and memory-constrained scenarios.

Conclusion: The framework enables efficient, progressive 3DGS rendering with compact representation suitable for streaming and resource-limited applications through top-down LOD construction and shared feature modeling.

Abstract: 3D Gaussian Splatting (3DGS) has become a state-of-the-art framework for real-time, high-fidelity novel view synthesis. However, its substantial storage requirements and inherently unstructured representation pose challenges for deployment in streaming and resource-constrained environments. Existing Level-of-Detail (LOD) strategies, particularly those based on bottom-up construction, often introduce redundancy or lead to fidelity degradation. To overcome these limitations, we propose Iterative Gaussian Synopsis, a novel framework for compact and progressive rendering through a top-down “unfolding” scheme. Our approach begins with a full-resolution 3DGS model and iteratively derives coarser LODs using an adaptive, learnable mask-based pruning mechanism. This process constructs a multi-level hierarchy that preserves visual quality while improving efficiency. We integrate hierarchical spatial grids, which capture the global scene structure, with a shared Anchor Codebook that models localized details. This combination produces a compact yet expressive feature representation, designed to minimize redundancy and support efficient, level-specific adaptation. The unfolding mechanism promotes inter-layer reusability and requires only minimal data overhead for progressive refinement. Experiments show that our method maintains high rendering quality across all LODs while achieving substantial storage reduction. These results demonstrate the practicality and scalability of our approach for real-time 3DGS rendering in bandwidth- and memory-constrained scenarios.

Method: Cannot analyze method as paper content is unavailable

Result: No results available due to technical access issues

Conclusion: Cannot draw conclusions about paper content due to access limitations

Abstract: Failed to fetch summary for 2604.11689: Page request resulted in HTTP 429 (https://export.arxiv.org/api/query?search_query=&id_list=2604.11689&sortBy=relevance&sortOrder=descending&start=0&max_results=100)

Method: Two-stage approach: 1) Forecast future scene structure in the feature space of a frozen vision foundation model, 2) Condition a latent diffusion model on these predicted representations to render photorealistic frames. Uses nested dropout and mixed supervision to address train-test mismatch between ground-truth and predicted representations.

Result: Experiments on challenging driving benchmarks show significant improvements in temporal semantic consistency, perceptual quality, and training efficiency compared to strong diffusion baselines.

Conclusion: The semantics-first hierarchical design effectively separates scene dynamics from appearance generation, leading to better video prediction performance in complex environments.

Abstract: Accurate future video prediction requires both high visual fidelity and consistent scene semantics, particularly in complex dynamic environments such as autonomous driving. We present Re2Pix, a hierarchical video prediction framework that decomposes forecasting into two stages: semantic representation prediction and representation-guided visual synthesis. Instead of directly predicting future RGB frames, our approach first forecasts future scene structure in the feature space of a frozen vision foundation model, and then conditions a latent diffusion model on these predicted representations to render photorealistic frames. This decomposition enables the model to focus first on scene dynamics and then on appearance generation. A key challenge arises from the train-test mismatch between ground-truth representations available during training and predicted ones used at inference. To address this, we introduce two conditioning strategies, nested dropout and mixed supervision, that improve robustness to imperfect autoregressive predictions. Experiments on challenging driving benchmarks demonstrate that the proposed semantics-first design significantly improves temporal semantic consistency, perceptual quality, and training efficiency compared to strong diffusion baselines. We provide the implementation code at https://github.com/Sta8is/Re2Pix

Method: Cannot determine method as paper content is unavailable

Result: Cannot determine results as paper content is unavailable

Conclusion: Cannot draw conclusions about paper content due to access limitations

Abstract: Failed to fetch summary for 2604.11711: Page request resulted in HTTP 429 (https://export.arxiv.org/api/query?search_query=&id_list=2604.11711&sortBy=relevance&sortOrder=descending&start=0&max_results=100)

Method: Unable to determine method due to failed paper fetch

Result: Unable to determine results due to failed paper fetch

Conclusion: Unable to determine conclusion due to failed paper fetch

Abstract: Failed to fetch summary for 2604.11714: Page request resulted in HTTP 429 (https://export.arxiv.org/api/query?search_query=&id_list=2604.11714&sortBy=relevance&sortOrder=descending&start=0&max_results=100)

Method: Cannot determine method without access to paper content

Result: Cannot determine results without access to paper content

Conclusion: Cannot determine conclusion without access to paper content

Abstract: Failed to fetch summary for 2604.11724: Page request resulted in HTTP 429 (https://export.arxiv.org/api/query?search_query=&id_list=2604.11724&sortBy=relevance&sortOrder=descending&start=0&max_results=100)

Method: Learn a highly compressed motion embedding (64x temporal compression) from large-scale trajectories obtained from tracker models. Train a conditional flow-matching model in this latent space to generate motion latents conditioned on text prompts or spatial pokes.

Result: The approach generates motion distributions that outperform both state-of-the-art video models and specialized task-specific approaches, achieving orders of magnitude more efficient motion generation.

Conclusion: Operating directly on compressed motion embeddings enables efficient generation of long, realistic motions conditioned on various goal specifications, advancing motion modeling capabilities.

Abstract: Understanding and predicting motion is a fundamental component of visual intelligence. Although modern video models exhibit strong comprehension of scene dynamics, exploring multiple possible futures through full video synthesis remains prohibitively inefficient. We model scene dynamics orders of magnitude more efficiently by directly operating on a long-term motion embedding that is learned from large-scale trajectories obtained from tracker models. This enables efficient generation of long, realistic motions that fulfill goals specified via text prompts or spatial pokes. To achieve this, we first learn a highly compressed motion embedding with a temporal compression factor of 64x. In this space, we train a conditional flow-matching model to generate motion latents conditioned on task descriptions. The resulting motion distributions outperform those of both state-of-the-art video models and specialized task-specific approaches.

Method: Uses logarithmic encoding (common in cinematic pipelines) to map HDR to distribution aligned with pretrained models’ latent space. Lightweight fine-tuning adapts models without retraining encoders. Introduces camera-mimicking degradations training strategy to help models infer missing HDR details from learned priors.

Result: Demonstrates high-quality HDR video generation using pretrained video models with minimal adaptation. Achieves strong results across diverse scenes and challenging lighting conditions.

Conclusion: HDR generation can be effectively handled without redesigning generative models by choosing representations that align with their learned priors, showing that fundamentally different image formation regimes can be addressed through proper adaptation.

Abstract: High dynamic range (HDR) imagery offers a rich and faithful representation of scene radiance, but remains challenging for generative models due to its mismatch with the bounded, perceptually compressed data on which these models are trained. A natural solution is to learn new representations for HDR, which introduces additional complexity and data requirements. In this work, we show that HDR generation can be achieved in a much simpler way by leveraging the strong visual priors already captured by pretrained generative models. We observe that a logarithmic encoding widely used in cinematic pipelines maps HDR imagery into a distribution that is naturally aligned with the latent space of these models, enabling direct adaptation via lightweight fine-tuning without retraining an encoder. To recover details that are not directly observable in the input, we further introduce a training strategy based on camera-mimicking degradations that encourages the model to infer missing high dynamic range content from its learned priors. Combining these insights, we demonstrate high-quality HDR video generation using a pretrained video model with minimal adaptation, achieving strong results across diverse scenes and challenging lighting conditions. Our results indicate that HDR, despite representing a fundamentally different image formation regime, can be handled effectively without redesigning generative models, provided that the representation is chosen to align with their learned priors.

Method: The paper organizes literature into four themes: object-centric visual understanding, object-centric referring segmentation, object-centric visual editing, and object-centric visual generation. It reviews key modeling paradigms, learning strategies, and evaluation protocols.

Result: Provides a structured taxonomy of approaches combining LMMs with object-centric vision, identifying current capabilities and limitations in object-level multimodal systems.

Conclusion: Object-centric vision offers a principled framework to extend LMMs from global scene understanding to precise object-level operations, with future directions including robust instance permanence, fine-grained spatial control, consistent multi-step interaction, unified cross-task modeling, and reliable benchmarking.

Abstract: Large Multimodal Models (LMMs) have achieved remarkable progress in general-purpose vision–language understanding, yet they remain limited in tasks requiring precise object-level grounding, fine-grained spatial reasoning, and controllable visual manipulation. In particular, existing systems often struggle to identify the correct instance, preserve object identity across interactions, and localize or modify designated regions with high precision. Object-centric vision provides a principled framework for addressing these challenges by promoting explicit representations and operations over visual entities, thereby extending multimodal systems from global scene understanding to object-level understanding, segmentation, editing, and generation. This paper presents a comprehensive review of recent advances at the convergence of LMMs and object-centric vision. We organize the literature into four major themes: object-centric visual understanding, object-centric referring segmentation, object-centric visual editing, and object-centric visual generation. We further summarize the key modeling paradigms, learning strategies, and evaluation protocols that support these capabilities. Finally, we discuss open challenges and future directions, including robust instance permanence, fine-grained spatial control, consistent multi-step interaction, unified cross-task modeling, and reliable benchmarking under distribution shift. We hope this paper provides a structured perspective on the development of scalable, precise, and trustworthy object-centric multimodal systems.

Method: Developed Lottie Tokenizer to encode layered geometric primitives, transforms, and keyframe-based motion into compact token sequences. Created LottieAnimation-660K dataset (660k animations, 15M static images). Fine-tuned Qwen-VL to create LottieGPT for multimodal vector animation generation.

Result: Tokenizer dramatically reduces sequence length while preserving structural fidelity. LottieGPT generates coherent, editable vector animations from natural language/visual prompts, outperforms SOTA on SVG generation (single-frame case).

Conclusion: First successful framework for native vector animation generation, enabling editable parametric motion synthesis and bridging the gap between multimodal models and structured vector content creation.

Abstract: Despite rapid progress in video generation, existing models are incapable of producing vector animation, a dominant and highly expressive form of multimedia on the Internet. Vector animations offer resolution-independence, compactness, semantic structure, and editable parametric motion representations, yet current generative models operate exclusively in raster space and thus cannot synthesize them. Meanwhile, recent advances in large multimodal models demonstrate strong capabilities in generating structured data such as slides, 3D meshes, LEGO sequences, and indoor layouts, suggesting that native vector animation generation may be achievable. In this work, we present the first framework for tokenizing and autoregressively generating vector animations. We adopt Lottie, a widely deployed JSON-based animation standard, and design a tailored Lottie Tokenizer that encodes layered geometric primitives, transforms, and keyframe-based motion into a compact and semantically aligned token sequence. To support large-scale training, we also construct LottieAnimation-660K, the largest and most diverse vector animation dataset to date, consisting of 660k real-world Lottie animation and 15M static Lottie image files curated from broad Internet sources. Building upon these components, we finetune Qwen-VL to create LottieGPT, a native multimodal model capable of generating coherent, editable vector animations directly from natural language or visual prompts. Experiments show that our tokenizer dramatically reduces sequence length while preserving structural fidelity, enabling effective autoregressive learning of dynamic vector content. LottieGPT exhibits strong generalization across diverse animation styles and outperforms previous state-of-the-art models on SVG generation (a special case of single-frame vector animation).

Method: SyncFix formulates refinement as a joint latent bridge matching problem, learning a joint conditional distribution over multiple views to synchronize distorted and clean representations. The framework enforces consistency throughout the denoising trajectory of diffusion models. Training uses only image pairs but generalizes to arbitrary numbers of views during inference.

Result: SyncFix consistently generates high-quality reconstructions and surpasses current state-of-the-art baselines, even without clean reference images. Reconstruction quality improves with additional views (with diminishing returns at higher view counts), and achieves even higher fidelity when sparse references are available.

Conclusion: SyncFix provides an effective framework for enforcing cross-view consistency in diffusion-based scene refinement, enabling high-quality 3D reconstructions that maintain semantic and geometric coherence across multiple viewpoints.

Abstract: We present SyncFix, a framework that enforces cross-view consistency during the diffusion-based refinement of reconstructed scenes. SyncFix formulates refinement as a joint latent bridge matching problem, synchronizing distorted and clean representations across multiple views to fix the semantic and geometric inconsistencies. This means SyncFix learns a joint conditional over multiple views to enforce consistency throughout the denoising trajectory. Our training is done only on image pairs, but it generalizes naturally to an arbitrary number of views during inference. Moreover, reconstruction quality improves with additional views, with diminishing returns at higher view counts. Qualitative and quantitative results demonstrate that SyncFix consistently generates high-quality reconstructions and surpasses current state-of-the-art baselines, even in the absence of clean reference images. SyncFix achieves even higher fidelity when sparse references are available.

Method: OmniShow introduces: 1) Unified Channel-wise Conditioning for efficient image and pose injection, 2) Gated Local-Context Attention for precise audio-visual synchronization, and 3) Decoupled-Then-Joint Training strategy with model merging to leverage heterogeneous sub-task datasets and address data scarcity. They also establish HOIVG-Bench as a comprehensive evaluation benchmark.

Result: Extensive experiments demonstrate that OmniShow achieves state-of-the-art performance across various multimodal conditioning settings, setting a solid standard for the emerging HOIVG task.

Conclusion: OmniShow provides an effective end-to-end framework for HOIVG that harmonizes multimodal conditions and delivers industry-grade performance, addressing both controllability-quality trade-offs and data scarcity challenges.

Abstract: In this work, we study Human-Object Interaction Video Generation (HOIVG), which aims to synthesize high-quality human-object interaction videos conditioned on text, reference images, audio, and pose. This task holds significant practical value for automating content creation in real-world applications, such as e-commerce demonstrations, short video production, and interactive entertainment. However, existing approaches fail to accommodate all these requisite conditions. We present OmniShow, an end-to-end framework tailored for this practical yet challenging task, capable of harmonizing multimodal conditions and delivering industry-grade performance. To overcome the trade-off between controllability and quality, we introduce Unified Channel-wise Conditioning for efficient image and pose injection, and Gated Local-Context Attention to ensure precise audio-visual synchronization. To effectively address data scarcity, we develop a Decoupled-Then-Joint Training strategy that leverages a multi-stage training process with model merging to efficiently harness heterogeneous sub-task datasets. Furthermore, to fill the evaluation gap in this field, we establish HOIVG-Bench, a dedicated and comprehensive benchmark for HOIVG. Extensive experiments demonstrate that OmniShow achieves overall state-of-the-art performance across various multimodal conditioning settings, setting a solid standard for the emerging HOIVG task.

Method: Proposes Pair2Scene framework that learns local object relations (support and functional relations) through a network trained on curated 3D-Pairs dataset. Uses hierarchical structure with recursive model application and collision-aware rejection sampling to align local rules into coherent global layouts.

Result: Outperforms existing methods in generating complex environments beyond training data while maintaining physical and semantic plausibility. Demonstrates scalability to dense scenes through extensive experiments.

Conclusion: Pair2Scene effectively addresses limitations of current 3D scene generation methods by focusing on local dependencies rather than global distributions, enabling generation of physically and semantically plausible complex indoor scenes beyond training distribution.

Abstract: Generating high-fidelity 3D indoor scenes remains a significant challenge due to data scarcity and the complexity of modeling intricate spatial relations. Current methods often struggle to scale beyond training distribution to dense scenes or rely on LLMs/VLMs that lack the ability for precise spatial reasoning. Building on top of the observation that object placement relies mainly on local dependencies instead of information-redundant global distributions, in this paper, we propose Pair2Scene, a novel procedural generation framework that integrates learned local rules with scene hierarchies and physics-based algorithms. These rules mainly capture two types of inter-object relations, namely support relations that follow physical hierarchies, and functional relations that reflect semantic links. We model these rules through a network, which estimates spatial position distributions of dependent objects conditioned on position and geometry of the anchor ones. Accordingly, we curate a dataset 3D-Pairs from existing scene data to train the model. During inference, our framework can generate scenes by recursively applying our model within a hierarchical structure, leveraging collision-aware rejection sampling to align local rules into coherent global layouts. Extensive experiments demonstrate that our framework outperforms existing methods in generating complex environments that go beyond training data while maintaining physical and semantic plausibility.

Method: Extensive experiments training on large 3D vision datasets, evaluating on image matching benchmarks, comparing rotation invariance in descriptors vs. matchers, studying emergence through scale.

Result: Descriptor-level rotation invariance yields similar performance to matcher-level, enables faster rotation-invariant matching, doesn’t hurt upright performance at scale, and scale improves generalization to rotations.

Conclusion: Rotation invariance can be effectively learned in descriptors, enabling efficient rotation-robust matching with state-of-the-art performance on challenging benchmarks.

Abstract: Finding matching keypoints between images is a core problem in 3D computer vision. However, modern matchers struggle with large in-plane rotations. A straightforward mitigation is to learn rotation invariance via data augmentation. However, it remains unclear at which stage rotation invariance should be incorporated. In this paper, we study this in the context of a modern sparse matching pipeline. We perform extensive experiments by training on a large collection of 3D vision datasets and evaluating on popular image matching benchmarks. Surprisingly, we find that incorporating rotation invariance already in the descriptor yields similar performance to handling it in the matcher. However, rotation invariance is achieved earlier in the matcher when it is learned in the descriptor, allowing for a faster rotation-invariant matcher. Further, we find that enforcing rotation invariance does not hurt upright performance when trained at scale. Finally, we study the emergence of rotation invariance through scale and find that increasing the training data size substantially improves generalization to rotated images. We release two matchers robust to in-plane rotations that achieve state-of-the-art performance on e.g. multi-modal (WxBS), extreme (HardMatch), and satellite image matching (SatAst). Code is available at https://github.com/davnords/loma.

Method: Unable to determine method due to technical error fetching paper content

Result: Unable to determine results due to technical error fetching paper content

Conclusion: Unable to draw conclusions due to technical error fetching paper content

Abstract: Failed to fetch summary for 2603.02123: Page request resulted in HTTP 429 (https://export.arxiv.org/api/query?search_query=&id_list=2603.02123&sortBy=relevance&sortOrder=descending&start=0&max_results=100)

Method: Unable to determine method due to fetch failure

Result: Unable to determine results due to fetch failure

Conclusion: Unable to determine conclusion due to fetch failure

Abstract: Failed to fetch summary for 2604.11386: Page request resulted in HTTP 429 (https://export.arxiv.org/api/query?search_query=&id_list=2604.11386&sortBy=relevance&sortOrder=descending&start=0&max_results=100)

Method: Unable to determine method due to API rate limiting preventing access to paper details

Result: Unable to determine results due to API rate limiting preventing access to paper details

Conclusion: Unable to determine conclusion due to API rate limiting preventing access to paper details

Abstract: Failed to fetch summary for 2604.11400: Page request resulted in HTTP 429 (https://export.arxiv.org/api/query?search_query=&id_list=2604.11400&sortBy=relevance&sortOrder=descending&start=0&max_results=100)