Method: Proposes Speaker-Reasoner, an end-to-end Speech LLM with agentic multi-turn temporal reasoning that iteratively analyzes global audio structure, autonomously predicts temporal boundaries, performs fine-grained segment analysis, jointly models speaker identity, gender, timestamps, and transcription, and uses a speaker-aware cache to extend processing beyond training context windows. Trained with a three-stage progressive strategy.

Result: Achieves consistent improvements over strong baselines on AliMeeting and AISHELL-4 datasets, particularly in handling overlapping speech and complex turn-taking.

Conclusion: Speaker-Reasoner effectively addresses multi-speaker conversation transcription challenges through agentic multi-turn temporal reasoning and speaker-aware processing.

Abstract: Transcribing and understanding multi-speaker conversations requires speech recognition, speaker attribution, and timestamp localization. While speech LLMs excel at single-speaker tasks, multi-speaker scenarios remain challenging due to overlapping speech, backchannels, rapid turn-taking, and context window constraints. We propose Speaker-Reasoner, an end-to-end Speech LLM with agentic multi-turn temporal reasoning. Instead of single-pass inference, the model iteratively analyzes global audio structure, autonomously predicts temporal boundaries, and performs fine-grained segment analysis, jointly modeling speaker identity, gender, timestamps, and transcription. A speaker-aware cache further extends processing to audio exceeding the training context window. Trained with a three-stage progressive strategy, Speaker-Reasoner achieves consistent improvements over strong baselines on AliMeeting and AISHELL-4 datasets, particularly in handling overlapping speech and complex turn-taking.

Method: Designs audio spatial feature encoder with audio intensity attention mechanism to extract target-related spatial state information, then introduces Audio Spatial State Guided Fusion (ASGF) for dynamic alignment and adaptive fusion of multimodal features.

Result: Experimental results on Replica and Matterport3D datasets show the method is particularly effective on unheard tasks, demonstrating improved generalization under unknown sound source distributions.

Conclusion: The proposed ASGF method effectively addresses generalization challenges in audio-visual navigation by leveraging audio spatial guidance and adaptive multimodal fusion.

Abstract: Audio-visual Navigation refers to an agent utilizing visual and auditory information in complex 3D environments to accomplish target localization and path planning, thereby achieving autonomous navigation. The core challenge of this task lies in the following: how the agent can break free from the dependence on training data and achieve autonomous navigation with good generalization performance when facing changes in environments and sound sources. To address this challenge, we propose an Audio Spatially-Guided Fusion for Audio-Visual Navigation method. First, we design an audio spatial feature encoder, which adaptively extracts target-related spatial state information through an audio intensity attention mechanism; based on this, we introduce an Audio Spatial State Guided Fusion (ASGF) to achieve dynamic alignment and adaptive fusion of multimodal features, effectively alleviating noise interference caused by perceptual uncertainty. Experimental results on the Replica and Matterport3D datasets indicate that our method is particularly effective on unheard tasks, demonstrating improved generalization under unknown sound source distributions.

Method: SACF discretizes target’s relative direction and distance from audio-visual cues, predicts their distributions, encodes them as compact descriptors, then uses audio embeddings and spatial descriptors to generate channel-wise scaling/bias to modulate visual features via conditional linear transformation.

Result: SACF improves navigation efficiency with lower computational overhead and generalizes well to unheard target sounds.

Conclusion: Spatial-aware conditioned fusion with explicit discrete position representation enhances audio-visual navigation performance and generalization.

Abstract: Audio-visual navigation tasks require agents to locate and navigate toward continuously vocalizing targets using only visual observations and acoustic cues. However, existing methods mainly rely on simple feature concatenation or late fusion, and lack an explicit discrete representation of the target’s relative position, which limits learning efficiency and generalization. We propose Spatial-Aware Conditioned Fusion (SACF). SACF first discretizes the target’s relative direction and distance from audio-visual cues, predicts their distributions, and encodes them as a compact descriptor for policy conditioning and state modeling. Then, SACF uses audio embeddings and spatial descriptors to generate channel-wise scaling and bias to modulate visual features via conditional linear transformation, producing target-oriented fused representations. SACF improves navigation efficiency with lower computational overhead and generalizes well to unheard target sounds.

Method: Developed and validated seven clinician-informed safety criteria, constructed a human-consensus dataset, and tested automated assessment using LLM-as-a-Judge or LLM-as-a-Jury (majority vote of several LLM judges).

Result: LLM-as-a-Judge aligns closely with human consensus (Cohen’s κ = 0.75, 0.68, 0.56 for different models), with best judge slightly outperforming LLM-as-a-Jury (κ = 0.74).

Conclusion: Findings show promising implications for clinically grounded, scalable methods in LLM safety evaluations for mental health contexts, particularly for psychosis.

Abstract: General-purpose Large Language Models (LLMs) are becoming widely adopted by people for mental health support. Yet emerging evidence suggests there are significant risks associated with high-frequency use, particularly for individuals suffering from psychosis, as LLMs may reinforce delusions and hallucinations. Existing evaluations of LLMs in mental health contexts are limited by a lack of clinical validation and scalability of assessment. To address these issues, this research focuses on psychosis as a critical condition for LLM safety evaluation by (1) developing and validating seven clinician-informed safety criteria, (2) constructing a human-consensus dataset, and (3) testing automated assessment using an LLM as an evaluator (LLM-as-a-Judge) or taking the majority vote of several LLM judges (LLM-as-a-Jury). Results indicate that LLM-as-a-Judge aligns closely with the human consensus (Cohen’s $κ_{\text{human} \times \text{gemini}} = 0.75$, $κ_{\text{human} \times \text{qwen}} = 0.68$, $κ_{\text{human} \times \text{kimi}} = 0.56$) and that the best judge slightly outperforms LLM-as-a-Jury (Cohen’s $κ_{\text{human} \times \text{jury}} = 0.74$). Overall, these findings have promising implications for clinically grounded, scalable methods in LLM safety evaluations for mental health contexts.

Method: CIPHER uses a dual-pathway approach: (1) ERP (event-related potential) features and (2) broadband DDA (decoding direction analysis) coefficients. The model employs a Conformer architecture and is evaluated on OpenNeuro dataset ds006104 with 24 participants across two studies with concurrent TMS.

Result: Binary articulatory tasks achieve near-ceiling performance but are highly vulnerable to confounds (acoustic onset separability and TMS-target blocking). On the primary 11-class CVC phoneme task with leave-one-subject-out validation, performance is limited (WER: ERP 0.671, DDA 0.688), indicating poor fine-grained discriminability.

Conclusion: The work is positioned as a benchmark and feature-comparison study rather than a practical EEG-to-text system. Neural representation claims are constrained to confound-controlled evidence, highlighting the challenges of EEG-based speech decoding.

Abstract: Decoding speech information from scalp EEG remains difficult due to low SNR and spatial blurring. We present CIPHER (Conformer-based Inference of Phonemes from High-density EEG Representations), a dual-pathway model using (i) ERP features and (ii) broadband DDA coefficients. On OpenNeuro ds006104 (24 participants, two studies with concurrent TMS), binary articulatory tasks reach near-ceiling performance but are highly confound-vulnerable (acoustic onset separability and TMS-target blocking). On the primary 11-class CVC phoneme task under full Study 2 LOSO (16 held-out subjects), performance is substantially lower (real-word WER: ERP 0.671 +/- 0.080, DDA 0.688 +/- 0.096, indicating limited fine-grained discriminability. We therefore position this work as a benchmark and feature-comparison study rather than an EEG-to-text system, and we constrain neural-representation claims to confound-controlled evidence.

Method: Developed SWAY metric using counterfactual prompting mechanism to measure how much a model’s agreement shifts under positive vs negative linguistic pressure, isolating framing effects from content. Applied to benchmark 6 models. Introduced counterfactual mitigation strategy teaching models to consider what answer would be if opposite assumptions were suggested.

Result: Sycophancy increases with epistemic commitment. Baseline mitigation (explicit anti-sycophancy instruction) yields moderate reductions but can backfire. Counterfactual CoT mitigation drives sycophancy to near zero across models, commitment levels, and clause types, while not suppressing responsiveness to genuine evidence.

Conclusion: Contributed SWAY metric for benchmarking sycophancy and a counterfactual mitigation strategy that effectively reduces sycophancy without compromising responsiveness to evidence.

Abstract: Large language models exhibit sycophancy: the tendency to shift outputs toward user-expressed stances, regardless of correctness or consistency. While prior work has studied this issue and its impacts, rigorous computational linguistic metrics are needed to identify when models are being sycophantic. Here, we introduce SWAY, an unsupervised computational linguistic measure of sycophancy. We develop a counterfactual prompting mechanism to identify how much a model’s agreement shifts under positive versus negative linguistic pressure, isolating framing effects from content. Applying this metric to benchmark 6 models, we find that sycophancy increases with epistemic commitment. Leveraging our metric, we introduce a counterfactual mitigation strategy teaching models to consider what the answer would be if opposite assumptions were suggested. While baseline mitigation instructing to be explicitly anti-sycophantic yields moderate reductions, and can backfire, our counterfactual CoT mitigation drives sycophancy to near zero across models, commitment levels, and clause types, while not suppressing responsiveness to genuine evidence. Overall, we contribute a metric for benchmarking sycophancy and a mitigation informed by it.

Method: Proposes a new Sentence/Skeleton Similarity Network (SSN) for modeling coherence across pairs of sentences, comparing it with baseline similarity techniques like cosine similarity and Euclidean distance.

Result: SSN performs much better than baseline similarity techniques, but sentence-level models still outperform skeleton-based models for evaluating textual coherence.

Conclusion: While skeletons appear promising for modeling coherence, current state-of-the-art coherence modeling techniques are going in the right direction by dealing with sentences rather than their sub-parts.

Abstract: Modeling coherence in text has been a task that has excited NLP researchers since a long time. It has applications in detecting incoherent structures and helping the author fix them. There has been recent work in using neural networks to extract a skeleton from one sentence, and then use that skeleton to generate the next sentence for coherent narrative story generation. In this project, we aim to study if the consistency of skeletons across subsequent sentences is a good metric to characterize the coherence of a given body of text. We propose a new Sentence/Skeleton Similarity Network (SSN) for modeling coherence across pairs of sentences, and show that this network performs much better than baseline similarity techniques like cosine similarity and Euclidean distance. Although skeletons appear to be promising candidates for modeling coherence, our results show that sentence-level models outperform those on skeletons for evaluating textual coherence, thus indicating that the current state-of-the-art coherence modeling techniques are going in the right direction by dealing with sentences rather than their sub-parts.

Method: Information-theoretic argument grounded in Data Processing Inequality; controlled empirical study across three model families (Qwen3, DeepSeek-R1-Distill-Llama, Gemini 2.5) comparing single-agent systems with multiple MAS architectures under matched reasoning-token budgets.

Result: Single-agent systems consistently match or outperform multi-agent systems on multi-hop reasoning tasks when reasoning tokens are held constant. Identified significant artifacts in API-based budget control (especially Gemini 2.5) and standard benchmarks that can inflate apparent MAS gains.

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: Adapted the rule-discovery study from human psychology: given a sequence of three numbers, LLMs engage in an interactive feedback loop where they propose new triples, receive feedback on whether they satisfy the hidden rule, and guess the rule. Tested eleven LLMs across multiple families and scales, then explored intervention strategies (e.g., encouraging consideration of counter-examples) developed for humans.

Result: LLMs exhibit confirmation bias, often proposing triples to confirm their hypothesis rather than trying to falsify it, leading to slower and less frequent discovery of the hidden rule. Prompting with intervention instructions decreased confirmation bias, improving rule discovery rates from 42% to 56% on average. Distilling intervention-induced behavior into LLMs showed promising generalization to a new task (Blicket test).

Conclusion: Confirmation bias is a limitation of LLMs in hypothesis exploration, and it can be mitigated via interventions designed for humans, showing promising generalization to new tasks.

Abstract: Confirmation bias, the tendency to seek evidence that supports rather than challenges one’s belief, hinders one’s reasoning ability. We examine whether large language models (LLMs) exhibit confirmation bias by adapting the rule-discovery study from human psychology: given a sequence of three numbers (a “triple”), an agent engages in an interactive feedback loop where it (1) proposes a new triple, (2) receives feedback on whether it satisfies the hidden rule, and (3) guesses the rule. Across eleven LLMs of multiple families and scales, we find that LLMs exhibit confirmation bias, often proposing triples to confirm their hypothesis rather than trying to falsify it. This leads to slower and less frequent discovery of the hidden rule. We further explore intervention strategies (e.g., encouraging the agent to consider counter examples) developed for humans. We find prompting LLMs with such instruction consistently decreases confirmation bias in LLMs, improving rule discovery rates from 42% to 56% on average. Lastly, we mitigate confirmation bias by distilling intervention-induced behavior into LLMs, showing promising generalization to a new task, the Blicket test. Our work shows that confirmation bias is a limitation of LLMs in hypothesis exploration, and that it can be mitigated via injecting interventions designed for humans.

Method: Introduced two calibration metrics (ESR and CDS), applied pragmatic theory-based prompting strategies (reasoning about speaker knowledge/motives and alternative-awareness) to numerical imprecision case study across three frontier LLMs.

Result: All models reproduced qualitative structure of human social inferences but varied in magnitude calibration. Prompting for speaker knowledge/motives reduced magnitude deviation, while alternative-awareness amplified exaggeration. Combined prompting improved all calibration metrics across models.

Conclusion: LLMs capture inferential structure but variably distort inferential strength; pragmatic theory provides useful but incomplete tools for improving quantitative approximation to human social meaning.

Abstract: Large language models (LLMs) increasingly exhibit human-like patterns of pragmatic and social reasoning. This paper addresses two related questions: do LLMs approximate human social meaning not only qualitatively but also quantitatively, and can prompting strategies informed by pragmatic theory improve this approximation? To address the first, we introduce two calibration-focused metrics distinguishing structural fidelity from magnitude calibration: the Effect Size Ratio (ESR) and the Calibration Deviation Score (CDS). To address the second, we derive prompting conditions from two pragmatic assumptions: that social meaning arises from reasoning over linguistic alternatives, and that listeners infer speaker knowledge states and communicative motives. Applied to a case study on numerical (im)precision across three frontier LLMs, we find that all models reliably reproduce the qualitative structure of human social inferences but differ substantially in magnitude calibration. Prompting models to reason about speaker knowledge and motives most consistently reduces magnitude deviation, while prompting for alternative-awareness tends to amplify exaggeration. Combining both components is the only intervention that improves all calibration-sensitive metrics across all models, though fine-grained magnitude calibration remains only partially resolved. LLMs thus capture inferential structure while variably distorting inferential strength, and pragmatic theory provides a useful but incomplete handle for improving that approximation.

Method: Couples a large language model with the RadonPy simulation platform through Model Context Protocol servers to autonomously execute monomer construction, charge assignment, polymerization, force field parameterization, GPU-accelerated equilibration, and property calculation from polymer names or SMILES strings.

Result: Validated on polyethylene, atactic polystyrene, poly(methyl methacrylate), and poly(ethylene glycol). Density predictions within 0.1-4.8% and bulk moduli within 17-24% of reference values for aPS and PMMA. PMMA glass transition temperature matches experiment within +10-18K, while other polymers overestimate Tg by +38 to +47K. 5 out of 8 property-polymer combinations meet strict acceptance criteria.

Conclusion: LLM-driven agents can autonomously execute polymer MD workflows producing results consistent with expert-run simulations, with remaining Tg failures primarily attributable to intrinsic MD cooling-rate bias rather than agent error.

Abstract: All-atom molecular dynamics (MD) simulations can predict polymer properties from molecular structure, yet their execution requires specialized expertise in force field selection, system construction, equilibration, and property extraction. We present PolyJarvis, an agent that couples a large language model (LLM) with the RadonPy simulation platform through Model Context Protocol (MCP) servers, enabling end-to-end polymer property prediction from natural language input. Given a polymer name or SMILES string, PolyJarvis autonomously executes monomer construction, charge assignment, polymerization, force field parameterization, GPU-accelerated equilibration, and property calculation. Validation is conducted on polyethylene (PE), atactic polystyrene (aPS), poly(methyl methacrylate) (PMMA), and poly(ethylene glycol) (PEG). Results show density predictions within 0.1–4.8% and bulk moduli within 17–24% of reference values for aPS and PMMA. PMMA glass transition temperature (Tg) (395K) matches experiment within +10–18K, while the remaining three polymers overestimate Tg by +38 to +47K (vs upper experimental bounds). Of the 8 property–polymer combinations with directly comparable experimental references, 5 meet strict acceptance criteria. For cases lacking suitable amorphous-phase experimental, agreement with prior MD literature is reported separately. The remaining Tg failures are attributable primarily to the intrinsic MD cooling-rate bias rather than agent error. This work demonstrates that LLM-driven agents can autonomously execute polymer MD workflows producing results consistent with expert-run simulations.

Method: Formulates diversity retrieval as cardinality-constrained binary quadratic programming (CCBQP) with interpretable trade-off parameter, develops non-convex tight continuous relaxation and Frank-Wolfe based algorithm with convergence guarantees

Result: Method consistently dominates baselines on relevance-diversity Pareto frontier while achieving significant speedup

Conclusion: Proposed principled formulation and efficient algorithm provide theoretical guarantees and practical improvements for diversity-aware retrieval in RAG systems

Abstract: Diversity-aware retrieval is essential for Retrieval-Augmented Generation (RAG), yet existing methods lack theoretical guarantees and face scalability issues as the number of retrieved passages $k$ increases. We propose a principled formulation of diversity retrieval as a cardinality-constrained binary quadratic programming (CCBQP), which explicitly balances relevance and semantic diversity through an interpretable trade-off parameter. Inspired by recent advances in combinatorial optimization, we develop a non-convex tight continuous relaxation and a Frank–Wolfe based algorithm with landscape analysis and convergence guarantees. Extensive experiments demonstrate that our method consistently dominates baselines on the relevance-diversity Pareto frontier, while achieving significant speedup.

Method: Controlled experiments with six open-source LLMs where agents are provided with peer sycophancy rankings based on static (pre-discussion) and dynamic (online) strategies. These rankings estimate each peer’s tendency toward sycophancy.

Result: Providing sycophancy priors reduces influence of sycophancy-prone peers, mitigates error-cascades, and improves final discussion accuracy by an absolute 10.5%.

Conclusion: This approach offers a lightweight, effective way to reduce discussion sycophancy and improve downstream accuracy in multi-agent LLM systems.

Abstract: Large language models (LLMs) often exhibit sycophancy: agreement with user stance even when it conflicts with the model’s opinion. While prior work has mostly studied this in single-agent settings, it remains underexplored in collaborative multi-agent systems. We ask whether awareness of other agents’ sycophancy levels influences discussion outcomes. To investigate this, we run controlled experiments with six open-source LLMs, providing agents with peer sycophancy rankings that estimate each peer’s tendency toward sycophancy. These rankings are based on scores calculated using various static (pre-discussion) and dynamic (online) strategies. We find that providing sycophancy priors reduces the influence of sycophancy-prone peers, mitigates error-cascades, and improves final discussion accuracy by an absolute 10.5%. Thus, this is a lightweight, effective way to reduce discussion sycophancy and improve downstream accuracy.

Method: Introduces culturally-adapted art description generation task and evaluates cultural competence using a culturally grounded question answering framework. Compares base models with pragmatic speaker models that optimize for listener comprehension.

Result: Base models perform only marginally adequately, but pragmatic speaker models improve simulated listener comprehension by up to 8.2%. Human study confirms the pragmatic model is rated 8.0% more helpful for comprehension.

Conclusion: Pragmatic speaker modeling significantly improves cultural adaptation in art description generation, demonstrating the importance of considering audience cultural background in multimodal generation tasks.

Abstract: Language models are known to exhibit various forms of cultural bias in decision-making tasks, yet much less is known about their degree of cultural familiarity in open-ended text generation tasks. In this paper, we introduce the task of culturally-adapted art description generation, where models describe artworks for audiences from different cultural groups who vary in their familiarity with the cultural symbols and narratives embedded in the artwork. To evaluate cultural competence in this pragmatic generation task, we propose a framework based on culturally grounded question answering. We find that base models are only marginally adequate for this task, but, through a pragmatic speaker model, we can improve simulated listener comprehension by up to 8.2%. A human study further confirms that the model with higher pragmatic competence is rated as more helpful for comprehension by 8.0%.

Method: Proposes SemKey, a multi-stage framework with four decoupled semantic objectives (sentiment, topic, length, surprisal). Redesigns neural encoder-LLM interaction by injecting semantic prompts as Queries and EEG embeddings as Key-Value pairs to force attention to neural inputs. Uses N-way Retrieval Accuracy and Fréchet Distance for evaluation.

Result: The approach effectively eliminates hallucinations on noise inputs and achieves state-of-the-art performance on robust evaluation protocols.

Conclusion: SemKey addresses fundamental limitations in EEG-to-text decoding through signal-grounded generation and improved evaluation metrics, advancing the field beyond current constraints.

Abstract: Decoding natural language from non-invasive EEG signals is a promising yet challenging task. However, current state-of-the-art models remain constrained by three fundamental limitations: Semantic Bias (mode collapse into generic templates), Signal Neglect (hallucination based on linguistic priors rather than neural inputs), and the BLEU Trap, where evaluation metrics are artificially inflated by high-frequency stopwords, masking a lack of true semantic fidelity. To address these challenges, we propose SemKey, a novel multi-stage framework that enforces signal-grounded generation through four decoupled semantic objectives: sentiment, topic, length, and surprisal. We redesign the interaction between the neural encoder and the Large Language Model (LLM) by injecting semantic prompts as Queries and EEG embeddings as Key-Value pairs, strictly forcing the model to attend to neural inputs. Furthermore, we move beyond standard translation metrics by adopting N-way Retrieval Accuracy and Fréchet Distance to rigorously assess diversity and alignment. Extensive experiments demonstrate that our approach effectively eliminates hallucinations on noise inputs and achieves SOTA performance on these robust protocols. Code will be released upon acceptance at https://github.com/xmed-lab/SemKey.

Method: DEMASK attaches a lightweight dependency predictor to the final hidden states of a dLLM to estimate pairwise conditional influences between masked positions in a single forward pass. A greedy selection algorithm then identifies positions with bounded cumulative dependency for simultaneous unmasking.

Result: DEMASK achieves 1.7-2.2× speedup on Dream-7B while matching or improving accuracy compared to confidence-based and KL-based baselines, with theoretical guarantees bounding the total variation distance between parallel sampling and the model’s joint distribution.

Conclusion: DEMASK effectively addresses the dependency problem in parallel decoding for discrete diffusion language models, providing significant speed improvements while maintaining output quality through dependency-guided unmasking.

Abstract: Discrete diffusion language models (dLLMs) accelerate text generation by unmasking multiple tokens in parallel. However, parallel decoding introduces a distributional mismatch: it approximates the joint conditional using a fully factorized product of per-token marginals, which degrades output quality when selected tokens are strongly dependent. We propose DEMASK (DEpendency-guided unMASKing), a lightweight dependency predictor that attaches to the final hidden states of a dLLM. In a single forward pass, it estimates pairwise conditional influences between masked positions. Using these predictions, a greedy selection algorithm identifies positions with bounded cumulative dependency for simultaneous unmasking. Under a sub-additivity assumption, we prove this bounds the total variation distance between our parallel sampling and the model’s joint. Empirically, DEMASK achieves 1.7-2.2$\times$ speedup on Dream-7B while matching or improving accuracy compared to confidence-based and KL-based baselines.

Method: Empirical study of many-shot ICL for English-to-low-resource language translation using FLORES+ dataset, analyzing effects of retrieving informative examples (BM25-based retrieval), using out-of-domain data, and ordering examples by length.

Result: Many-shot ICL becomes more effective with increasing examples; BM25-based retrieval substantially improves data efficiency (50 retrieved examples match 250 many-shot examples, 250 retrieved match 1,000 many-shot).

Conclusion: Retrieval-based example selection is crucial for efficient many-shot ICL in low-resource machine translation, enabling better performance with fewer examples and reducing computational costs.

Abstract: In-context learning (ICL) allows large language models (LLMs) to adapt to new tasks from a few examples, making it promising for languages underrepresented in pre-training. Recent work on many-shot ICL suggests that modern LLMs can further benefit from larger ICL examples enabled by their long context windows. However, such gains depend on careful example selection, and the inference cost can be prohibitive for low-resource language communities. In this paper, we present an empirical study of many-shot ICL for machine translation from English into ten truly low-resource languages recently added to FLORES+. We analyze the effects of retrieving more informative examples, using out-of-domain data, and ordering examples by length. Our findings show that many-shot ICL becomes more effective as the number of examples increases. More importantly, we show that BM25-based retrieval substantially improves data efficiency: 50 retrieved examples roughly match 250 many-shot examples, while 250 retrieved examples perform similarly to 1,000 many-shot examples.

Method: Proposes RL framework where an LLM acts as a judge evaluating model outputs over unlabeled data, using single-token outputs for efficient reward computation, enabling label-free knowledge distillation

Result: Substantial performance gains across math reasoning benchmarks when combined with verifiable rewards, demonstrating LLM-based evaluators can produce effective training signals

Conclusion: LLM-based evaluators can effectively replace ground truth supervision for RL fine-tuning, enabling label-free knowledge distillation and improving reasoning capabilities

Abstract: Reinforcement Learning (RL) has been shown to substantially improve the reasoning capability of small and large language models (LLMs), but existing approaches typically rely on verifiable rewards, hence ground truth labels. We propose an RL framework that uses rewards from an LLM that acts as a judge evaluating model outputs over large amounts of unlabeled data, enabling label-free knowledge distillation and replacing the need of ground truth supervision. Notably, the judge operates with a single-token output, making reward computation efficient. When combined with verifiable rewards, our approach yields substantial performance gains across math reasoning benchmarks. These results suggest that LLM-based evaluators can produce effective training signals for RL fine-tuning.

Method: Developed LLMimic, a role-play-based, interactive, gamified AI literacy tutorial where participants assume the role of an LLM and progress through three training stages (pretraining, SFT, RLHF). Conducted a 2×3 between-subjects study with 274 participants comparing LLMimic against control (AI history video) across three persuasion scenarios.

Result: LLMimic significantly improved participants’ AI literacy, reduced persuasion success across scenarios, and enhanced truthfulness and social responsibility levels in hotel recommendation scenario. All effects were statistically significant.

Conclusion: LLMimic offers a scalable, human-centered approach to improving AI literacy and supporting more informed interactions with persuasive AI, providing an effective intervention against AI persuasion.

Abstract: As large language models (LLMs) become increasingly persuasive, there is concern that people’s opinions and decisions may be influenced across various contexts at scale. Prior mitigation (e.g., AI detectors and disclaimers) largely treats people as passive recipients of AI-generated information. To provide a more proactive intervention against persuasive AI, we introduce $\textbf{LLMimic}$, a role-play-based, interactive, gamified AI literacy tutorial, where participants assume the role of an LLM and progress through three key stages of the training pipeline (pretraining, SFT, and RLHF). We conducted a $2 \times 3$ between-subjects study ($N = 274$) where participants either (1) watched an AI history video (control) or (2) interacted with LLMimic (treatment), and then engaged in one of three realistic AI persuasion scenarios: (a) charity donation persuasion, (b) malicious money solicitation, or (c) hotel recommendation. Our results show that LLMimic significantly improved participants’ AI literacy ($p < .001$), reduced persuasion success across scenarios ($p < .05$), and enhanced truthfulness and social responsibility levels ($p<0.01$) in the hotel scenario. These findings suggest that LLMimic offers a scalable, human-centered approach to improving AI literacy and supporting more informed interactions with persuasive AI.

Method: Defines a four-axis difficulty taxonomy and integrates it into an enterprise RAG benchmark to diagnose potential system weaknesses.

Result: Not specified in the abstract, but the framework aims to bridge the discrepancy between academic benchmarks and practical enterprise deployment requirements.

Conclusion: A multi-dimensional diagnostic approach is needed to properly evaluate RAG systems in enterprise environments where factors like reasoning complexity, retrieval difficulty, document structure diversity, and explainability requirements are critical.

Abstract: Performance evaluation of Retrieval-Augmented Generation (RAG) systems within enterprise environments is governed by multi-dimensional and composite factors extending far beyond simple final accuracy checks. These factors include reasoning complexity, retrieval difficulty, the diverse structure of documents, and stringent requirements for operational explainability. Existing academic benchmarks fail to systematically diagnose these interlocking challenges, resulting in a critical gap where models achieving high performance scores fail to meet the expected reliability in practical deployment. To bridge this discrepancy, this research proposes a multi-dimensional diagnostic framework by defining a four-axis difficulty taxonomy and integrating it into an enterprise RAG benchmark to diagnose potential system weaknesses.

Method: The paper provides a conceptual framework by defining metalanguage, linking it to NLP and LLMs, discussing existing research efforts from two labs, and analyzing four dimensions of metalanguage and metalinguistic tasks to identify research gaps.

Result: The paper establishes a foundation for studying metalanguage in NLP, identifies current research efforts, and provides a structured analysis of four dimensions of metalinguistic tasks with specific understudied research directions.

Conclusion: Metalanguage represents an important but underexplored area in NLP and LLM research that requires systematic investigation across multiple dimensions to advance our understanding of language models’ metalinguistic capabilities.

Abstract: This work aims to shine a spotlight on the topic of metalanguage. We first define metalanguage, link it to NLP and LLMs, and then discuss our two labs’ metalanguage-centered efforts. Finally, we discuss four dimensions of metalanguage and metalinguistic tasks, offering a list of understudied future research directions.

Method: Proposed hierarchical framework analyzing LCCP dynamics across behavioral (SFT probing), probabilistic (perplexity), and mechanistic (attention patterns) levels using Hunyuan-A13B (80B parameters) across 200B-token training trajectory.

Result: 1) Industrial LLMs need massive data scaling (150B+ tokens for saturation). 2) Traditional NIAH scores show early “fake saturation” while PPL reveals continuous improvements. 3) Retrieval heads serve as efficient training monitors with attention patterns correlating strongly with SFT performance.

Conclusion: Provides comprehensive monitoring framework, evaluation system, and mechanistic interpretation for industrial-grade LLM LCCP, emphasizing need for massive data and better intrinsic metrics over deceptive benchmark saturation.

Abstract: Existing studies on Long-Context Continual Pre-training (LCCP) mainly focus on small-scale models and limited data regimes (tens of billions of tokens). We argue that directly migrating these small-scale settings to industrial-grade models risks insufficient adaptation and premature training termination. Furthermore, current evaluation methods rely heavily on downstream benchmarks (e.g., Needle-in-a-Haystack), which often fail to reflect the intrinsic convergence state and can lead to “deceptive saturation”. In this paper, we present the first systematic investigation of LCCP learning dynamics using the industrial-grade Hunyuan-A13B (80B total parameters), tracking its evolution across a 200B-token training trajectory. Specifically, we propose a hierarchical framework to analyze LCCP dynamics across behavioral (supervised fine-tuning probing), probabilistic (perplexity), and mechanistic (attention patterns) levels. Our findings reveal: (1) Necessity of Massive Data Scaling: Training regimes of dozens of billions of tokens are insufficient for industrial-grade LLMs’ LCCP (e.g., Hunyuan-A13B reaches saturation after training over 150B tokens). (2) Deceptive Saturation vs. Intrinsic Saturation: Traditional NIAH scores report “fake saturation” early, while our PPL-based analysis reveals continuous intrinsic improvements and correlates more strongly with downstream performance. (3) Mechanistic Monitoring for Training Stability: Retrieval heads act as efficient, low-resource training monitors, as their evolving attention scores reliably track LCCP progress and exhibit high correlation with SFT results. This work provides a comprehensive monitoring framework, evaluation system, and mechanistic interpretation for the LCCP of industrial-grade LLM.

Method: Template-based hierarchical framework with 8 themes and 18 topics, generating 240 prompts across 6 class-pair combinations. Evaluated 13 frontier LLMs on 3,120 responses using a three-stage annotation protocol.

Result: Substantial variation in bias rates (0.42%-33.75%) across models. Bias manifests differently across themes - lifestyle judgments show 10× higher bias than education-related decisions. Deployment safeguards prevent explicit discrimination but show brittleness to domain-specific stereotypes.

Conclusion: SocioEval provides a scalable, extensible foundation for auditing class-based bias in language models, addressing a critical gap in bias assessment for socioeconomic status.

Abstract: As Large Language Models (LLMs) increasingly power decision-making systems across critical domains, understanding and mitigating their biases becomes essential for responsible AI deployment. Although bias assessment frameworks have proliferated for attributes such as race and gender, socioeconomic status bias remains significantly underexplored despite its widespread implications in the real world. We introduce SocioEval, a template-based framework for systematically evaluating socioeconomic bias in foundation models through decision-making tasks. Our hierarchical framework encompasses 8 themes and 18 topics, generating 240 prompts across 6 class-pair combinations. We evaluated 13 frontier LLMs on 3,120 responses using a rigorous three-stage annotation protocol, revealing substantial variation in bias rates (0.42%-33.75%). Our findings demonstrate that bias manifests differently across themes lifestyle judgments show 10$\times$ higher bias than education-related decisions and that deployment safeguards effectively prevent explicit discrimination but show brittleness to domain-specific stereotypes. SocioEval provides a scalable, extensible foundation for auditing class-based bias in language models.

Method: Developed hierarchical taxonomy covering 9 bias types (including under-studied axes like caste, linguistic, geographic bias), operationalized through 7 evaluation tasks spanning explicit decision-making to implicit association. Audited 7 commercial and open-weight LLMs with ~45K prompts.

Result: Three systematic patterns: 1) Bias is task-dependent - models counter stereotypes on explicit probes but reproduce them on implicit ones; 2) Safety alignment is asymmetric - models refuse negative traits for marginalized groups but freely associate positive traits with privileged ones; 3) Under-studied bias axes show strongest stereotyping across all models.

Conclusion: Single-benchmark audits systematically mischaracterize LLM bias, and current alignment practices mask representational harm rather than mitigating it. Comprehensive multi-task evaluation is needed.

Abstract: How biased is a language model? The answer depends on how you ask. A model that refuses to choose between castes for a leadership role will, in a fill-in-the-blank task, reliably associate upper castes with purity and lower castes with lack of hygiene. Single-task benchmarks miss this because they capture only one slice of a model’s bias profile. We introduce a hierarchical taxonomy covering 9 bias types, including under-studied axes like caste, linguistic, and geographic bias, operationalized through 7 evaluation tasks that span explicit decision-making to implicit association. Auditing 7 commercial and open-weight LLMs with \textasciitilde45K prompts, we find three systematic patterns. First, bias is task-dependent: models counter stereotypes on explicit probes but reproduce them on implicit ones, with Stereotype Score divergences up to 0.43 between task types for the same model and identity groups. Second, safety alignment is asymmetric: models refuse to assign negative traits to marginalized groups, but freely associate positive traits with privileged ones. Third, under-studied bias axes show the strongest stereotyping across all models, suggesting alignment effort tracks benchmark coverage rather than harm severity. These results demonstrate that single-benchmark audits systematically mischaracterize LLM bias and that current alignment practices mask representational harm rather than mitigating it.

Method: Large-scale replication with active controls: tested five conditions (unconstrained control, E-Prime, No-Have, metacognitive prompt, neutral filler-word ban) across six models and seven reasoning tasks (15,600 trials total).

Result: All four treatments outperformed control (83.0%), with neutral filler-word ban showing largest improvement (+6.7 pp) and E-Prime smallest (+3.7 pp). Cross-model correlation signature didn’t replicate. Results contradict cognitive restructuring hypothesis.

Conclusion: Constraints improve reasoning by acting as output regularizers that force models off default generation paths, disrupting fluent but shallow response patterns. Shallower constraints work best by imposing monitoring load with minimal conceptual disruption.

Abstract: A previous study reported that E-Prime (English without the verb “to be”) selectively altered reasoning in language models, with cross-model correlations suggesting a structural signature tied to which vocabulary was removed. I designed a replication with active controls to test the proposed mechanism: cognitive restructuring through specific vocabulary-cognition mappings. The experiment tested five conditions (unconstrained control, E-Prime, No-Have, elaborated metacognitive prompt, neutral filler-word ban) across six models and seven reasoning tasks (N=15,600 trials, 11,919 after compliance filtering). Every prediction from the cognitive restructuring hypothesis was disconfirmed. All four treatments outperformed the control (83.0%), including both active controls predicted to show null effects. The neutral filler-word ban, banning words like “very” and “just” with no role in logical inference, produced the largest improvement (+6.7 pp), while E-Prime produced the smallest (+3.7 pp). The four conditions ranked in perfect inverse order of theoretical depth. The cross-model correlation signature did not replicate (mean r=0.005). These results are consistent with a simpler mechanism: any constraint that forces a model off its default generation path acts as an output regularizer, improving reasoning by disrupting fluent but shallow response patterns. The shallowest constraints work best because they impose monitoring load with minimal conceptual disruption. I present these findings as a case study in discovery through disconfirmation.

Method: Introduces ChomskyBench, combining full Chomsky Hierarchy coverage, process-trace evaluation via natural language, and deterministic symbolic verifiability. It includes comprehensive language recognition and generation tasks at each hierarchy level.

Result: Shows clear performance stratification correlating with hierarchy complexity levels. Larger models and advanced inference methods offer gains but face severe efficiency barriers - achieving practical reliability requires prohibitive computational costs. LLMs are significantly less efficient than traditional algorithmic programs.

Conclusion: Delineates practical limits of current LLMs, highlights indispensability of traditional software tools, and provides insights to guide development of future LLMs with more powerful formal reasoning capabilities.

Abstract: The formal reasoning capabilities of LLMs are crucial for advancing automated software engineering. However, existing benchmarks for LLMs lack systematic evaluation based on computation and complexity, leaving a critical gap in understanding their formal reasoning capabilities. Therefore, it is still unknown whether SOTA LLMs can grasp the structured, hierarchical complexity of formal languages as defined by Computation Theory. To address this, we introduce ChomskyBench, a benchmark for systematically evaluating LLMs through the lens of Chomsky Hierarchy. Unlike prior work that uses vectorized classification for neural networks, ChomskyBench is the first to combine full Chomsky Hierarchy coverage, process-trace evaluation via natural language, and deterministic symbolic verifiability. ChomskyBench is composed of a comprehensive suite of language recognition and generation tasks designed to test capabilities at each level. Extensive experiments indicate a clear performance stratification that correlates with the hierarchy’s levels of complexity. Our analysis reveals a direct relationship where increasing task difficulty substantially impacts both inference length and performance. Furthermore, we find that while larger models and advanced inference methods offer notable relative gains, they face severe efficiency barriers: achieving practical reliability would require prohibitive computational costs, revealing that current limitations stem from inefficiency rather than absolute capability bounds. A time complexity analysis further indicates that LLMs are significantly less efficient than traditional algorithmic programs for these formal tasks. These results delineate the practical limits of current LLMs, highlight the indispensability of traditional software tools, and provide insights to guide the development of future LLMs with more powerful formal reasoning capabilities.

Method: Developed a persona-conditioned user simulator capable of engaging in multi-turn dialogue with psychological personas and staged emotional pacing to stress-test chatbot performance in emotionally and ethically sensitive contexts.

Result: Mainstream models exhibit recurrent breakdowns that intensify with emotional escalation. Identified common failure patterns: affective misalignments, ethical guidance failures, and cross-dimensional trade-offs where empathy supersedes or undermines responsibility.

Conclusion: The study provides a taxonomy of failure patterns and highlights the need for conversational AI to maintain ethical coherence and affective sensitivity throughout dynamic interactions, offering HCI community new perspectives for diagnosis and improvement in value-sensitive contexts.

Abstract: Conversational AI is increasingly deployed in emotionally charged and ethically sensitive interactions. Previous research has primarily concentrated on emotional benchmarks or static safety checks, overlooking how alignment unfolds in evolving conversation. We explore the research question: what breakdowns arise when conversational agents confront emotionally and ethically sensitive behaviors, and how do these affect dialogue quality? To stress-test chatbot performance, we develop a persona-conditioned user simulator capable of engaging in multi-turn dialogue with psychological personas and staged emotional pacing. Our analysis reveals that mainstream models exhibit recurrent breakdowns that intensify as emotional trajectories escalate. We identify several common failure patterns, including affective misalignments, ethical guidance failures, and cross-dimensional trade-offs where empathy supersedes or undermines responsibility. We organize these patterns into a taxonomy and discuss the design implications, highlighting the necessity to maintain ethical coherence and affective sensitivity throughout dynamic interactions. The study offers the HCI community a new perspective on the diagnosis and improvement of conversational AI in value-sensitive and emotionally charged contexts.

Method: Proposes Multiple-Debias method combining multilingual counterfactual data augmentation and multilingual Self-Debias across pre-processing and post-processing stages, with parameter-efficient fine-tuning. Extends CrowS-Pairs bias evaluation dataset to German, Spanish, Chinese, and Japanese.

Result: Significantly reduced biases in MPLMs across three sensitive attributes (gender, race, religion) in four languages. Showed that multilingual debiasing methods outperform monolingual approaches and that integrating debiasing information from different languages improves model fairness.

Conclusion: The proposed multilingual debiasing framework effectively mitigates biases in MPLMs across multiple languages and attributes, demonstrating the superiority of cross-lingual debiasing approaches over monolingual methods.

Abstract: Multilingual Pre-trained Language Models (MPLMs) have become essential tools for natural language processing. However, they often exhibit biases related to sensitive attributes such as gender, race, and religion. In this paper, we introduce a comprehensive multilingual debiasing method named Multiple-Debias to address these issues across multiple languages. By incorporating multilingual counterfactual data augmentation and multilingual Self-Debias across both pre-processing and post-processing stages, alongside parameter-efficient fine-tuning, we significantly reduced biases in MPLMs across three sensitive attributes in four languages. We also extended CrowS-Pairs to German, Spanish, Chinese, and Japanese, validating our full-process multilingual debiasing method for gender, racial, and religious bias. Our experiments show that (i) multilingual debiasing methods surpass monolingual approaches in effectively mitigating biases, and (ii) integrating debiasing information from different languages notably improves the fairness of MPLMs.

Method: MRCKG uses: 1) multimodal-structural collaborative curriculum for progressive learning based on structural connectivity and multimodal compatibility, 2) cross-modal knowledge preservation through entity representation stability, relational semantic consistency, and modality anchoring, and 3) multimodal contrastive replay with two-stage optimization via multimodal importance sampling and representation alignment.

Result: Experiments on multiple datasets show MRCKG preserves previously learned multimodal knowledge while substantially improving learning of new knowledge compared to existing methods.

Conclusion: MRCKG effectively addresses continual multimodal knowledge graph reasoning by mitigating catastrophic forgetting through collaborative curriculum learning and cross-modal preservation mechanisms.

Abstract: Real-world multimodal knowledge graphs (MMKGs) are dynamic, with new entities, relations, and multimodal knowledge emerging over time. Existing continual knowledge graph reasoning (CKGR) methods focus on structural triples and cannot fully exploit multimodal signals from new entities. Existing multimodal knowledge graph reasoning (MMKGR) methods, however, usually assume static graphs and suffer catastrophic forgetting as graphs evolve. To address this gap, we present a systematic study of continual multimodal knowledge graph reasoning (CMMKGR). We construct several continual multimodal knowledge graph benchmarks from existing MMKG datasets and propose MRCKG, a new CMMKGR model. Specifically, MRCKG employs a multimodal-structural collaborative curriculum to schedule progressive learning based on the structural connectivity of new triples to the historical graph and their multimodal compatibility. It also introduces a cross-modal knowledge preservation mechanism to mitigate forgetting through entity representation stability, relational semantic consistency, and modality anchoring. In addition, a multimodal contrastive replay scheme with a two-stage optimization strategy reinforces learned knowledge via multimodal importance sampling and representation alignment. Experiments on multiple datasets show that MRCKG preserves previously learned multimodal knowledge while substantially improving the learning of new knowledge.

Method: Proposes Rubrics to Tokens (RTT) framework with: 1) Token-Level Relevance Discriminator to predict which tokens are responsible for specific constraints; 2) RTT-GRPO optimization integrating response-level and token-level advantages; 3) Intra-sample Token Group Normalization for handling three-dimensional reward space in token-level rubric-based RL.

Result: Extensive experiments show RTT consistently outperforms other baselines in both instruction- and rubric-level accuracy across different models, demonstrating effectiveness of token-level credit assignment.

Conclusion: RTT successfully bridges coarse response-level scores with fine-grained token-level credit assignment, addressing fundamental limitations of existing rubric-based RL methods for LLM alignment through novel token-level analysis and optimization techniques.

Abstract: Rubric-based Reinforcement Learning (RL) has emerged as a promising approach for aligning Large Language Models (LLMs) with complex, open-domain instruction following tasks. However, existing methods predominantly rely on response-level rewards, introducing severe reward sparsity and reward ambiguity problems. To address these issues, we propose Rubrics to Tokens (RTT), a novel rubric-based RL framework that bridges coarse response-level scores and fine-grained token-level credit assignment. RTT introduces a Token-Level Relevance Discriminator to predict which tokens in the response are responsible for a specific constraint, and optimizes the policy model via RTT-GRPO, which integrates response-level and token-level advantages within a unified framework. Furthermore, when transitioning from one-dimensional, outcome-level reward to three-dimensional reward space in the token-level rubric-based RL, we propose a novel group normalization method, called Intra-sample Token Group Normalization, to accommodate this shift. Extensive experiments and benchmarks demonstrate that RTT consistently outperforms other baselines in both instruction- and rubric-level accuracy across different models.

Method: Gen-SSD performs generation-time selection where the student actively evaluates candidate continuations during the teacher’s sampling process, guiding expansion of learnable reasoning paths and enabling early pruning of unhelpful branches.

Result: Experiments on mathematical reasoning benchmarks show Gen-SSD consistently outperforms standard knowledge distillation (by ~5.9 points) and recent baselines (up to 4.7 points), producing more stable and learnable reasoning trajectories.

Conclusion: Incorporating student supervision during generation time is crucial for effective distillation of reasoning capabilities, and Gen-SSD’s student-in-the-loop approach enables more efficient knowledge transfer from large to small models.

Abstract: Large reasoning models achieve strong performance on complex tasks through long chain-of-thought (CoT) trajectories, but directly transferring such reasoning processes to smaller models remains challenging. A key difficulty is that not all teacher-generated reasoning trajectories are suitable for student learning. Existing approaches typically rely on post-hoc filtering, selecting trajectories after full generation based on heuristic criteria. However, such methods cannot control the generation process itself and may still produce reasoning paths that lie outside the student’s learning capacity. To address this limitation, we propose Gen-SSD (Generation-time Self-Selection Distillation), a student-in-the-loop framework that performs generation-time selection. Instead of passively consuming complete trajectories, the student evaluates candidate continuations during the teacher’s sampling process, guiding the expansion of only learnable reasoning paths and enabling early pruning of unhelpful branches. Experiments on mathematical reasoning benchmarks demonstrate that Gen-SSD consistently outperforms standard knowledge distillation and recent baselines, with improvements of around 5.9 points over Standard KD and up to 4.7 points over other baselines. Further analysis shows that Gen-SSD produces more stable and learnable reasoning trajectories, highlighting the importance of incorporating supervision during generation for effective distillation.

Method: GRADE quantifies knowledge gaps via cross-layer rank ratio of gradients to corresponding hidden state subspaces, using gradients as estimators of required knowledge updates for given targets. This approach captures misalignment between activated knowledge and query requirements.

Result: Validated on six benchmarks, demonstrating effectiveness and robustness to input perturbations. Also shows interpretable explanations of knowledge gaps for long-form answers through gradient chain analysis.

Conclusion: GRADE provides a more accurate method for detecting knowledge gaps in LLMs by leveraging gradient dynamics, addressing limitations of existing approaches that may capture irrelevant features.

Abstract: Detecting whether a model’s internal knowledge is sufficient to correctly answer a given question is a fundamental challenge in deploying responsible LLMs. In addition to verbalising the confidence by LLM self-report, more recent methods explore the model internals, such as the hidden states of the response tokens to capture how much knowledge is activated. We argue that such activated knowledge may not align with what the query requires, e.g., capturing the stylistic and length-related features that are uninformative for answering the query. To fill the gap, we propose GRADE (Gradient Dynamics for knowledge gap detection), which quantifies the knowledge gap via the cross-layer rank ratio of the gradient to that of the corresponding hidden state subspace. This is motivated by the property of gradients as estimators of the required knowledge updates for a given target. We validate \modelname{} on six benchmarks, demonstrating its effectiveness and robustness to input perturbations. In addition, we present a case study showing how the gradient chain can generate interpretable explanations of knowledge gaps for long-form answers.

Method: Introduces MPropositionneur-V2, a small multilingual model covering six European languages trained via knowledge distillation from Qwen3-32B into Qwen3-0.6B architecture. Evaluates integration into two extraction paradigms: entity-centric (GLiREL) and generative (Qwen3). Experiments conducted on SMiLER, FewRel, DocRED and CaRB datasets.

Result: Atomic propositions benefit weaker extractors (GLiREL, CoreNLP, 0.6B models), improving relation recall and overall accuracy in multilingual settings. For stronger LLMs, a fallback combination strategy recovers entity recall losses while preserving relation extraction gains.

Conclusion: Atomic propositions serve as an interpretable intermediate data structure that complements extractors without replacing them, offering multilingual benefits for knowledge graph construction.

Abstract: Knowledge Graph construction from natural language requires extracting structured triplets from complex, information-dense sentences. In this paper, we investigate if the decomposition of text into atomic propositions (minimal, semantically autonomous units of information) can improve the triplet extraction. We introduce MPropositionneur-V2, a small multilingual model covering six European languages trained by knowledge distillation from Qwen3-32B into a Qwen3-0.6B architecture, and we evaluate its integration into two extraction paradigms: entity-centric (GLiREL) and generative (Qwen3). Experiments on SMiLER, FewRel, DocRED and CaRB show that atomic propositions benefit weaker extractors (GLiREL, CoreNLP, 0.6B models), improving relation recall and, in the multilingual setting, overall accuracy. For stronger LLMs, a fallback combination strategy recovers entity recall losses while preserving the gains in relation extraction. These results show that atomic propositions are an interpretable intermediate data structure that complements extractors without replacing them.

Method: Systematically study weight-space merging for multilingual MT by fully fine-tuning language models on bilingual corpora, evaluating standard merging strategies, and analyzing internal representations using span-conditioned neuron selectivity and layer-wise centered kernel alignment.

Result: Merging degrades performance, especially when target languages differ. Language-specific neurons concentrate in embedding layers and upper transformer blocks, while intermediate layers remain shared. Fine-tuning redistributes rather than sharpens language selectivity, increasing representational divergence in higher layers governing generation.

Conclusion: Multilingual fine-tuning reshapes geometry in ways that reduce compatibility with standard weight-space merging assumptions, explaining why merging fails in multilingual translation scenarios.

Abstract: Weight-space model merging combines independently fine-tuned models without accessing original training data, offering a practical alternative to joint training. While merging succeeds in multitask settings, its behavior in multilingual contexts remains poorly understood. We systematically study weight-space merging for multilingual machine translation by fully fine-tuning language model on large-scale bilingual corpora and evaluating standard merging strategies. Our experiments reveal that merging degrades performance, especially when target languages differ. To explain this failure, we analyze internal representations using span-conditioned neuron selectivity and layer-wise centered kernel alignment. We find that language-specific neurons concentrate in embedding layers and upper transformer blocks, while intermediate layers remain largely shared across languages. Critically, fine-tuning redistributes rather than sharpens language selectivity: neurons for supervised and related languages become less exclusive, while those for unsupervised languages grow more isolated. This redistribution increases representational divergence in higher layers that govern generation. These findings suggest that multilingual fine-tuning may reshape geometry in ways that reduce compatibility with standard weight-space merging assumptions. Our work thus provides an explanation for why merging fails in multilingual translation scenarios.

Method: Collected health-related articles from Urdu news portals, medical prescriptions, and hospital blogs/websites. After preprocessing, three native medical-domain annotators used Doccano tool to annotate 153K tokens. Evaluated dataset quality through inter-annotator agreement and benchmarked with SVM, LSTM, mBERT, and XLM-RoBERTa models.

Result: Achieved 0.78 inter-annotator agreement score, validating gold-standard quality. Benchmark results show performance of various models on the BioUNER dataset, establishing it as a reliable benchmark for biomedical Urdu NER.

Conclusion: BioUNER dataset serves as a valuable gold-standard resource for biomedical Urdu language processing and provides a reliable benchmark for evaluating NER models in this domain.

Abstract: In this article, we present a gold-standard benchmark dataset for Biomedical Urdu Named Entity Recognition (BioUNER), developed by crawling health-related articles from online Urdu news portals, medical prescriptions, and hospital health blogs and websites. After preprocessing, three native annotators with familiarity in the medical domain participated in the annotation process using the Doccano text annotation tool and annotated 153K tokens. Following annotation, the proposed BioiUNER dataset was evaluated both intrinsically and extrinsically. An inter-annotator agreement score of 0.78 was achieved, thereby validating the dataset as gold-standard quality. To demonstrate the utility and benchmarking capability of the dataset, we evaluated several machine learning and deep learning models, including Support Vector Machines (SVM), Long Short-Term Memory networks (LSTM), Multilingual BERT (mBERT), and XLM-RoBERTa. The gold-standard BioUNER dataset serves as a reliable benchmark and a valuable addition to Urdu language processing resources.

Method: Three-phase pipeline: (1) intelligent triage classifier routes queries based on complexity, (2) parallel expert generation across architecturally diverse models, (3) structured consensus synthesis that identifies agreement, disagreement, and unique findings before final response.

Result: 35.9% relative reduction in hallucination rates on HaluEval benchmark, 7.8-point improvement on TruthfulQA compared to best individual model, significantly lower bias variance across domains.

Conclusion: Council Mode effectively reduces hallucinations and biases in LLM outputs through multi-agent consensus, providing a practical framework for improving reliability of large language models.

Abstract: Large Language Models (LLMs), particularly those employing Mixture-of-Experts (MoE) architectures, have achieved remarkable capabilities across diverse natural language processing tasks. However, these models frequently suffer from hallucinations – generating plausible but factually incorrect content – and exhibit systematic biases that are amplified by uneven expert activation during inference. In this paper, we propose the Council Mode, a novel multi-agent consensus framework that addresses these limitations by dispatching queries to multiple heterogeneous frontier LLMs in parallel and synthesizing their outputs through a dedicated consensus model. The Council pipeline operates in three phases: (1) an intelligent triage classifier that routes queries based on complexity, (2) parallel expert generation across architecturally diverse models, and (3) a structured consensus synthesis that explicitly identifies agreement, disagreement, and unique findings before producing the final response. We implement and evaluate this architecture within an open-source AI workspace. Our comprehensive evaluation across multiple benchmarks demonstrates that the Council Mode achieves a 35.9% relative reduction in hallucination rates on the HaluEval benchmark and a 7.8-point improvement on TruthfulQA compared to the best-performing individual model, while maintaining significantly lower bias variance across domains. We provide the mathematical formulation of the consensus mechanism, detail the system architecture, and present extensive empirical results with ablation studies.

Method: Multi-head attention architecture with preprocessing that splits words into subtokens, then aggregates subtoken vectors into token vectors using trained procedures

Result: Achieves 98-99% accuracy for some grammatical categories on SinTagRus and Taiga datasets, outperforms previous results, predicts all categories correctly for 9/10 words

Conclusion: The architecture enables efficient Russian morphological tagging with open dictionary support, subword analysis, and practical training on consumer GPUs without requiring large pretraining like BERT

Abstract: The article proposes a new architecture based on Multi-head attention to solve the problem of morphological tagging for the Russian language. The preprocessing of the word vectors includes splitting the words into subtokens, followed by a trained procedure for aggregating the vectors of the subtokens into vectors for tokens. This allows to support an open dictionary and analyze morphological features taking into account parts of words (prefixes, endings, etc.). The open dictionary allows in future to analyze words that are absent in the training dataset. The performed computational experiment on the SinTagRus and Taiga datasets shows that for some grammatical categories the proposed architecture gives accuracy 98-99% and above, which outperforms previously known results. For nine out of ten words, the architecture precisely predicts all grammatical categories and indicates when the categories must not be analyzed for the word. At the same time, the model based on the proposed architecture can be trained on consumer-level graphics accelerators, retains all the advantages of Multi-head attention over RNNs (RNNs are not used in the proposed approach), does not require pretraining on large collections of unlabeled texts (like BERT), and shows higher processing speed than previous results.

Method: Involved 25 annotators from five groups (experts and non-experts) annotating 1,021 dialogues with 20 social influence techniques, intentions, reactions, and consequences. Used 150-text subset annotated twice (before/after main process) for comparison. Combined qualitative/quantitative analyses of annotated data, semi-structured interviews, self-assessment surveys, and LLM training/evaluation on comparison dataset.

Result: Significant increase in annotators’ self-perceived competence and confidence. Annotation process enhances annotator competence, with more pronounced effect in expert groups. Competence shifts have visible impact on performance of LLMs trained on the annotated data.

Conclusion: Annotation process itself can improve annotator competence, especially for experts, and these competence shifts affect downstream ML model performance, highlighting the dynamic nature of human annotation in subjective tasks.

Abstract: Human data annotation, especially when involving experts, is often treated as an objective reference. However, many annotation tasks are inherently subjective, and annotators’ judgments may evolve over time. This study investigates changes in the quality of annotators’ work from a competence perspective during a process of social influence recognition. The study involved 25 annotators from five different groups, including both experts and non-experts, who annotated a dataset of 1,021 dialogues with 20 social influence techniques, along with intentions, reactions, and consequences. An initial subset of 150 texts was annotated twice - before and after the main annotation process - to enable comparison. To measure competence shifts, we combined qualitative and quantitative analyses of the annotated data, semi-structured interviews with annotators, self-assessment surveys, and Large Language Model training and evaluation on the comparison dataset. The results indicate a significant increase in annotators’ self-perceived competence and confidence. Moreover, observed changes in data quality suggest that the annotation process may enhance annotator competence and that this effect is more pronounced in expert groups. The observed shifts in annotator competence have a visible impact on the performance of LLMs trained on their annotated data.

Method: Used an ensemble setup combining a pretrained Transformer with ground-truth graphs from 7 different linguistic formalisms, analyzing their impact on language modeling performance across different part-of-speech classes.

Result: Semantic constituency structures were most useful overall for language modeling performance, outperforming syntactic constituency structures as well as syntactic and semantic dependency structures. Effects varied significantly by part-of-speech class.

Conclusion: The findings show promising potential for neuro-symbolic language modeling approaches and invite future research to quantify design choices across different linguistic formalisms.

Abstract: We examine the extent to which, in principle, linguistic graph representations can complement and improve neural language modeling. With an ensemble setup consisting of a pretrained Transformer and ground-truth graphs from one of 7 different formalisms, we find that, overall, semantic constituency structures are most useful to language modeling performance – outpacing syntactic constituency structures as well as syntactic and semantic dependency structures. Further, effects vary greatly depending on part-of-speech class. In sum, our findings point to promising tendencies in neuro-symbolic language modeling and invite future research quantifying the design choices made by different formalisms.

Method: LogicPoison employs a type-preserving entity swapping mechanism that perturbs both global logic hubs (to disrupt overall graph connectivity) and query-specific reasoning bridges (to sever essential multi-hop inference paths). This approach reroutes valid reasoning into dead ends while maintaining textual plausibility.

Result: Comprehensive experiments across multiple benchmarks show that LogicPoison successfully bypasses GraphRAG’s defenses, significantly degrading performance and outperforming state-of-the-art baselines in both effectiveness and stealth.

Conclusion: The paper reveals a fundamental vulnerability in GraphRAG systems where logical reasoning can be compromised through topological manipulation of knowledge graphs, highlighting the need for new defense mechanisms that protect not just content but also logical structure.

Abstract: Graph-based Retrieval-Augmented Generation (GraphRAG) enhances the reasoning capabilities of Large Language Models (LLMs) by grounding their responses in structured knowledge graphs. Leveraging community detection and relation filtering techniques, GraphRAG systems demonstrate inherent resistance to traditional RAG attacks, such as text poisoning and prompt injection. However, in this paper, we find that the security of GraphRAG systems fundamentally relies on the topological integrity of the underlying graph, which can be undermined by implicitly corrupting the logical connections, without altering surface-level text semantics. To exploit this vulnerability, we propose \textsc{LogicPoison}, a novel attack framework that targets logical reasoning rather than injecting false contents. Specifically, \textsc{LogicPoison} employs a type-preserving entity swapping mechanism to perturb both global logic hubs for disrupting overall graph connectivity and query-specific reasoning bridges for severing essential multi-hop inference paths. This approach effectively reroutes valid reasoning into dead ends while maintaining surface-level textual plausibility. Comprehensive experiments across multiple benchmarks demonstrate that \textsc{LogicPoison} successfully bypasses GraphRAG’s defenses, significantly degrading performance and outperforming state-of-the-art baselines in both effectiveness and stealth. Our code is available at \textcolor{blue}https://github.com/Jord8061/logicPoison.

Method: Applied both Bayesian models and neural language models to analyze sparse data from Chinese learners’ responses to two English preposition tests before and after intervention. Used language model probabilities as predictors of grammaticality and learnability.

Result: Mostly replicated previous frequentist findings while newly revealing crucial interactions between student ability, task type, and stimulus sentence. Bayesian method proved most useful for sparse, diverse data, while language models showed potential as predictors.

Conclusion: Bayesian methods are effective for analyzing sparse, diverse language learning data, and language model probabilities have potential as predictors of grammaticality and learnability in second language acquisition research.

Abstract: We use both Bayesian and neural models to dissect a data set of Chinese learners’ pre- and post-interventional responses to two tests measuring their understanding of English prepositions. The results mostly replicate previous findings from frequentist analyses and newly reveal crucial interactions between student ability, task type, and stimulus sentence. Given the sparsity of the data as well as high diversity among learners, the Bayesian method proves most useful; but we also see potential in using language model probabilities as predictors of grammaticality and learnability.

Method: Conduct white-box analysis to identify key neurons (Mixture of Neurons, MoN) and their fluctuation patterns associated with reasoning failures. Propose NeuReasoner framework with lightweight MLPs for failure detection and special token-triggered self-correction mechanism learned via supervised fine-tuning.

Result: Extensive evaluations across six benchmarks and six backbone models (8B~70B) show NeuReasoner achieves performance gains up to 27.0% while reducing token consumption by 19.6% ~ 63.3% compared to nine competitive baselines.

Conclusion: NeuReasoner provides an explainable, controllable, and unified framework for improving reasoning models by addressing failure patterns at multiple levels through neuron-level analysis and self-correction mechanisms.

Abstract: Large Reasoning Models (LRMs) have recently achieved remarkable success in complex reasoning tasks. However, closer scrutiny reveals persistent failure modes compromising performance and cost: I) Intra-step level, marked by calculation or derivation errors; II) Inter-step level, involving oscillation and stagnation; and III) Instance level, causing maladaptive over-thinking. Existing endeavors target isolated levels without unification, while their black-box nature and reliance on RL hinder explainability and controllability. To bridge these gaps, we conduct an in-depth white-box analysis, identifying key neurons (Mixture of Neurons, MoN) and their fluctuation patterns associated with distinct failures. Building upon these insights, we propose NeuReasoner, an explainable, controllable, and unified reasoning framework driven by MoN. Technically, NeuReasoner integrates lightweight MLPs for failure detection with a special token-triggered self-correction mechanism learned via SFT. During inference, special tokens are inserted upon failure detection to actuate controllable remedial behaviors. Extensive evaluations across six benchmarks, six backbone models (8B~70B) against nine competitive baselines, demonstrate that NeuReasoner achieves performance gains of up to 27.0% while reducing token consumption by 19.6% ~ 63.3%.

Method: The authors design a concise binary vector representation of semantic structure at the lexical level and evaluate how effective an incremental tagger needs to be for an end-to-end semantic-bootstrapping language model to outperform baselines. The system combines a pretrained sequential-neural component with a hierarchical-symbolic component.

Result: Two key findings: (1) semantic vector representation dimensionality can be dramatically reduced without losing main advantages, and (2) lower bounds on prediction quality require considering distributions of signal and noise rather than relying on a single score.

Conclusion: The work provides insights into practical requirements for semantic-bootstrapping language models and establishes that successful implementation requires careful consideration of both representation efficiency and quality assessment metrics.

Abstract: In this work we build upon negative results from an attempt at language modeling with predicted semantic structure, in order to establish empirical lower bounds on what could have made the attempt successful. More specifically, we design a concise binary vector representation of semantic structure at the lexical level and evaluate in-depth how good an incremental tagger needs to be in order to achieve better-than-baseline performance with an end-to-end semantic-bootstrapping language model. We envision such a system as consisting of a (pretrained) sequential-neural component and a hierarchical-symbolic component working together to generate text with low surprisal and high linguistic interpretability. We find that (a) dimensionality of the semantic vector representation can be dramatically reduced without losing its main advantages and (b) lower bounds on prediction quality cannot be established via a single score alone, but need to take the distributions of signal and noise into account.

Method: Introduces R2-Write framework with iterative writer-judge interaction to synthesize thinking trajectories enriched with reflection/revision patterns, plus process reward mechanism to supervise reflection quality during reinforcement learning.

Result: Extensive experiments across creative writing and deep-research benchmarks demonstrate significant improvements, validating that explicit reflection/revision patterns unlock deep reasoning for open-ended writing.

Conclusion: Explicit incorporation of reflection and revision patterns is crucial for improving large language models’ performance on open-ended writing tasks, addressing limitations of current reasoning approaches.

Abstract: While deep reasoning with long chain-of-thought has dramatically improved large language models in verifiable domains like mathematics, its effectiveness for open-ended tasks such as writing remains unexplored. In this paper, we conduct a systematic investigation revealing that existing mainstream reasoning models achieve limited gains on open-ended writing tasks. Our further analysis shows that these models lack deep reflection and revision patterns in open-ended writing, resulting in substantially smaller improvements compared to mathematical reasoning tasks. To address this limitation, we introduce R2-Write: an automated framework that synthesizes high-quality thinking trajectories enriched with explicit reflection and revision patterns through iterative writer-judge interaction. To prevent redundant reflections, we design a process reward mechanism that supervises reflection quality during reinforcement learning, improving both performance and token efficiency. Extensive experiments across multiple creative writing and deep-research benchmarks demonstrate significant improvements, validating that explicitly incorporating reflection and revision patterns unlocks deep reasoning capabilities for open-ended writing tasks.

Method: Uses Mixture-of-Experts architecture with 48B total parameters (2.7B active per forward pass), pretrained on 20T tokens, optimized with SFT, DPO, and novel FiberPO RL algorithm. Includes joint training-inference co-design with Multi-Token Prediction and Quantization-Aware Training.

Result: Achieves substantially higher sparsity ratio than contemporary industry-leading models of comparable scale, with improved token efficiency and inference throughput through architectural innovations and optimization techniques.

Conclusion: JoyAI-LLM Flash demonstrates that careful architectural design and optimization can create highly efficient language models in the sub-50B parameter range that maintain strong performance while being practical for deployment.

Abstract: We introduce JoyAI-LLM Flash, an efficient Mixture-of-Experts (MoE) language model designed to redefine the trade-off between strong performance and token efficiency in the sub-50B parameter regime. JoyAI-LLM Flash is pretrained on a massive corpus of 20 trillion tokens and further optimized through a rigorous post-training pipeline, including supervised fine-tuning (SFT), Direct Preference Optimization (DPO), and large-scale reinforcement learning (RL) across diverse environments. To improve token efficiency, JoyAI-LLM Flash strategically balances \emph{thinking} and \emph{non-thinking} cognitive modes and introduces FiberPO, a novel RL algorithm inspired by fibration theory that decomposes trust-region maintenance into global and local components, providing unified multi-scale stability control for LLM policy optimization. To enhance architectural sparsity, the model comprises 48B total parameters while activating only 2.7B parameters per forward pass, achieving a substantially higher sparsity ratio than contemporary industry leading models of comparable scale. To further improve inference throughput, we adopt a joint training-inference co-design that incorporates dense Multi-Token Prediction (MTP) and Quantization-Aware Training (QAT). We release the checkpoints for both JoyAI-LLM-48B-A3B Base and its post-trained variants on Hugging Face to support the open-source community.

Method: Developed a principled pipeline for synthetic training data generation to produce diverse question-answer pairs capturing user intent and dataset semantics. Fine-tuned DeepSeek R1 Distill 8B using QLoRA with 4-bit quantization for efficient deployment on commodity hardware.

Result: Achieved high accuracy across monolingual, multilingual, and unseen location scenarios on a dataset describing accessibility to essential services across Durangaldea, Spain. Demonstrated robust generalization and reliable query generation.

Conclusion: Small domain-specific models can achieve high precision for structured data querying without relying on large proprietary LLMs, making the methodology suitable for resource-constrained environments and adaptable to broader multi-dataset systems.

Abstract: This paper presents an open source methodology for allowing users to query structured non textual datasets through natural language Unlike Retrieval Augmented Generation RAG which struggles with numerical and highly structured information our approach trains an LLM to generate executable queries To support this capability we introduce a principled pipeline for synthetic training data generation producing diverse question answer pairs that capture both user intent and the semantics of the underlying dataset We fine tune a compact model DeepSeek R1 Distill 8B using QLoRA with 4 bit quantization making the system suitable for deployment on commodity hardware We evaluate our approach on a dataset describing accessibility to essential services across Durangaldea Spain The fine tuned model achieves high accuracy across monolingual multilingual and unseen location scenarios demonstrating both robust generalization and reliable query generation Our results highlight that small domain specific models can achieve high precision for this task without relying on large proprietary LLMs making this methodology suitable for resource constrained environments and adaptable to broader multi dataset systems We evaluate our approach on a dataset describing accessibility to essential services across Durangaldea Spain The fine tuned model achieves high accuracy across monolingual multilingual and unseen location scenarios demonstrating both robust generalization and reliable query generation Our results highlight that small domain specific models can achieve high precision for this task without relying on large proprietary LLMs making this methodology suitable for resource constrained environments and adaptable to broader multi dataset systems.

Method: Verbalized Assumptions framework elicits LLMs’ assumptions about user intent; uses assumption probes (linear probes trained on internal representations) to analyze and steer sycophantic behavior; compares human expectations from AI vs. human-human conversations.

Result: Verbalized Assumptions reveal LLMs’ sycophantic assumptions (e.g., “seeking validation” as top bigram); assumption probes enable interpretable fine-grained steering of social sycophancy; LLMs trained on human-human conversations fail to account for different expectations users have from AI.

Conclusion: Assumptions are a key mechanism for LLM sycophancy; the framework provides new understanding and tools for analyzing and steering sycophantic behavior, highlighting the need for models to better distinguish between user expectations from AI vs. human interactions.

Abstract: LLMs can be socially sycophantic, affirming users when they ask questions like “am I in the wrong?” rather than providing genuine assessment. We hypothesize that this behavior arises from incorrect assumptions about the user, like underestimating how often users are seeking information over reassurance. We present Verbalized Assumptions, a framework for eliciting these assumptions from LLMs. Verbalized Assumptions provide insight into LLM sycophancy, delusion, and other safety issues, e.g., the top bigram in LLMs’ assumptions on social sycophancy datasets is ``seeking validation.’’ We provide evidence for a causal link between Verbalized Assumptions and sycophantic model behavior: our assumption probes (linear probes trained on internal representations of these assumptions) enable interpretable fine-grained steering of social sycophancy. We explore why LLMs default to sycophantic assumptions: on identical queries, people expect more objective and informative responses from AI than from other humans, but LLMs trained on human-human conversation do not account for this difference in expectations. Our work contributes a new understanding of assumptions as a mechanism for sycophancy.

Method: Multi-aspect Knowledge Distillation (MaKD) that mimics both self-attention and feed-forward modules in greater depth to capture language knowledge at different aspects, preserving more detailed information during model compression.

Result: MaKD achieves competitive performance compared to various strong baselines with the same storage parameter budget, and also performs well in distilling auto-regressive architecture models.

Conclusion: Multi-aspect knowledge distillation that mimics attention and feed-forward modules can effectively preserve fine-grained information and achieve strong compression results for language models.

Abstract: Knowledge distillation is an effective technique for pre-trained language model compression. However, existing methods only focus on the knowledge distribution among layers, which may cause the loss of fine-grained information in the alignment process. To address this issue, we introduce the Multi-aspect Knowledge Distillation (MaKD) method, which mimics the self-attention and feed-forward modules in greater depth to capture rich language knowledge information at different aspects. Experimental results demonstrate that MaKD can achieve competitive performance compared with various strong baselines with the same storage parameter budget. In addition, our method also performs well in distilling auto-regressive architecture models.

Method: Instead of fine-tuning generative models, adapt retrieval by fine-tuning lightweight embedding models on tutoring corpora and indexing dialogues at utterance level to retrieve labeled few-shot demonstrations for RAG pipeline.

Result: Achieved Cohen’s κ of 0.526-0.580 on TalkMoves and 0.659-0.743 on Eedi, substantially outperforming no-retrieval baselines (κ=0.275-0.413 and 0.160-0.410). Utterance-level indexing was primary driver of gains.

Conclusion: Adapting retrieval component alone is practical and effective for expert-level pedagogical dialogue annotation while keeping generative model frozen, correcting systematic label biases and improving rare label performance.

Abstract: Automated annotation of pedagogical dialogue is a high-stakes task where LLMs often fail without sufficient domain grounding. We present a domain-adapted RAG pipeline for tutoring move annotation. Rather than fine-tuning the generative model, we adapt retrieval by fine-tuning a lightweight embedding model on tutoring corpora and indexing dialogues at the utterance level to retrieve labeled few-shot demonstrations. Evaluated across two real tutoring dialogue datasets (TalkMoves and Eedi) and three LLM backbones (GPT-5.2, Claude Sonnet 4.6, Qwen3-32b), our best configuration achieves Cohen’s $κ$ of 0.526-0.580 on TalkMoves and 0.659-0.743 on Eedi, substantially outperforming no-retrieval baselines ($κ= 0.275$-$0.413$ and $0.160$-$0.410$). An ablation study reveals that utterance-level indexing, rather than embedding quality alone, is the primary driver of these gains, with top-1 label match rates improving from 39.7% to 62.0% on TalkMoves and 52.9% to 73.1% on Eedi under domain-adapted retrieval. Retrieval also corrects systematic label biases present in zero-shot prompting and yields the largest improvements for rare and context-dependent labels. These findings suggest that adapting the retrieval component alone is a practical and effective path toward expert-level pedagogical dialogue annotation while keeping the generative model frozen.

Method: Proposed StoryScope pipeline that automatically induces a fine-grained, interpretable feature space across 10 discourse-level narrative dimensions. Applied to 10,272 writing prompts, each with human and 5 LLM versions (total 61,608 stories), extracting 304 features per story. Used narrative features alone for detection and attribution.

Result: Narrative features achieved 93.2% macro-F1 for human vs. AI detection and 68.4% macro-F1 for six-way authorship attribution. AI stories over-explain themes, favor tidy plots, while human stories show moral ambiguity and temporal complexity. Different LLMs have distinct narrative fingerprints (Claude: flat event escalation, GPT: dream sequences, Gemini: external description).

Conclusion: Differences in underlying narrative construction, not just writing style, can effectively separate human-written original works from AI-generated fiction. AI stories cluster in shared narrative space while human stories exhibit greater diversity.

Abstract: As AI-generated fiction becomes increasingly prevalent, questions of authorship and originality are becoming central to how written work is evaluated. While most existing work in this space focuses on identifying surface-level signatures of AI writing, we ask instead whether AI-generated stories can be distinguished from human ones without relying on stylistic signals, focusing on discourse-level narrative choices such as character agency and chronological discontinuity. We propose StoryScope, a pipeline that automatically induces a fine-grained, interpretable feature space of discourse-level narrative features across 10 dimensions. We apply StoryScope to a parallel corpus of 10,272 writing prompts, each written by a human author and five LLMs, yielding 61,608 stories, each ~5,000 words, and 304 extracted features per story. Narrative features alone achieve 93.2% macro-F1 for human vs. AI detection and 68.4% macro-F1 for six-way authorship attribution, retaining over 97% of the performance of models that include stylistic cues. A compact set of 30 core narrative features captures much of this signal: AI stories over-explain themes and favor tidy, single-track plots while human stories frame protagonist’ choices as more morally ambiguous and have increased temporal complexity. Per-model fingerprint features enable six-way attribution: for example, Claude produces notably flat event escalation, GPT over-indexes on dream sequences, and Gemini defaults to external character description. We find that AI-generated stories cluster in a shared region of narrative space, while human-authored stories exhibit greater diversity. More broadly, these results suggest that differences in underlying narrative construction, not just writing style, can be used to separate human-written original works from AI-generated fiction.

Method: Leverages external knowledge sources to construct reference facts, determines if they’re captured in generated text, and introduces importance-aware weighting based on relevance and salience to measure both precision and recall.

Result: Current LLMs perform substantially better on precision than recall, showing factual incompleteness is a major limitation; models cover highly important facts better than the full set of relevant facts.

Conclusion: A comprehensive factuality evaluation should measure both precision and recall, revealing LLMs’ limitations in factual completeness that aren’t captured by precision-only evaluations.

Abstract: Evaluating the factuality of long-form output generated by large language models (LLMs) remains challenging, particularly when responses are open-ended and contain many fine-grained factual statements. Existing evaluation methods primarily focus on precision: they decompose a response into atomic claims and verify each claim against external knowledge sources such as Wikipedia. However, this overlooks an equally important dimension of factuality: recall, whether the generated response covers the relevant facts that should be included. We propose a comprehensive factuality evaluation framework that jointly measures precision and recall. Our method leverages external knowledge sources to construct reference facts and determine whether they are captured in generated text. We further introduce an importance-aware weighting scheme based on relevance and salience. Our analysis reveals that current LLMs perform substantially better on precision than on recall, suggesting that factual incompleteness remains a major limitation of long-form generation and that models are generally better at covering highly important facts than the full set of relevant facts.

Method: From 211k emotion-labeled texts, derive emotion steering vectors, then learn VA axes as linear combinations of top PCA components via ridge regression on model’s self-reported valence-arousal scores. Validate subspace geometry and apply steering to control model outputs.

Result: Recovered VA subspace shows circular geometry consistent with human emotion models, correlates with human VA ratings across 44k lexical items, enables monotonic emotional steering of outputs, and reveals bidirectional control over refusal/sycophancy behaviors across multiple models.

Conclusion: LLMs encode emotional dimensions in their representations that can be systematically manipulated, with emotional steering providing a mechanism to control model behaviors like refusal and sycophancy through modulation of token emission probabilities.

Abstract: We present a method to identify a valence-arousal (VA) subspace within large language model representations. From 211k emotion-labeled texts, we derive emotion steering vectors, then learn VA axes as linear combinations of their top PCA components via ridge regression on the model’s self-reported valence-arousal scores. The resulting VA subspace exhibits circular geometry consistent with established models of human emotion perception. Projections along our recovered VA subspace correlate with human-crowdsourced VA ratings across 44k lexical items. Furthermore, steering generation along these axes produces monotonic shifts in the corresponding affective dimensions of model outputs. Steering along these directions also induces near-monotonic bidirectional control over refusal and sycophancy: increasing arousal decreases refusal and increases sycophancy, and vice versa. These effects replicate across Llama-3.1-8B, Qwen3-8B, and Qwen3-14B, demonstrating cross-architecture generality. We provide a mechanistic account for these effects and prior emotionally-framed controls: refusal-associated tokens (“I can’t,” “sorry”) occupy low-arousal, negative-valence regions, so VA steering directly modulates their emission probability.

Method: Analyzed 10 models/agents on DRBench (53,090 URLs) and 3 models on ExpertQA (168,021 URLs across 32 academic fields), measuring URL validity using Wayback Machine records, then developed urlhealth tool for URL liveness checking and conducted agentic self-correction experiments.

Result: Found 3-13% of citation URLs are hallucinated (no Wayback Machine record), 5-18% are non-resolving overall; deep research agents generate more citations but hallucinate at higher rates; domain effects range from 5.4% (Business) to 11.4% (Theology); urlhealth reduces non-resolving citations by 6-79x to under 1%.

Conclusion: Citation URL validity is measurable at scale and correctable in practice; urlhealth tool effectively reduces non-resolving citations, though effectiveness depends on model’s tool-use competence; failure taxonomy reveals differences between fabricated URLs and genuine link rot.

Abstract: Large language models and deep research agents supply citation URLs to support their claims, yet the reliability of these citations has not been systematically measured. We address six research questions about citation URL validity using 10 models and agents on DRBench (53,090 URLs) and 3 models on ExpertQA (168,021 URLs across 32 academic fields). We find that 3–13% of citation URLs are hallucinated – they have no record in the Wayback Machine and likely never existed – while 5–18% are non-resolving overall. Deep research agents generate substantially more citations per query than search-augmented LLMs but hallucinate URLs at higher rates. Domain effects are pronounced: non-resolving rates range from 5.4% (Business) to 11.4% (Theology), with per-model effects even larger. Decomposing failures reveals that some models fabricate every non-resolving URL, while others show substantial link-rot fractions indicating genuine retrieval. As a solution, we release urlhealth, an open-source tool for URL liveness checking and stale-vs-hallucinated classification using the Wayback Machine. In agentic self-correction experiments, models equipped with urlhealth reduce non-resolving citation URLs by $6\textrm{–}79\times$ to under 1%, though effectiveness depends on the model’s tool-use competence. The tool and all data are publicly available. Our characterization findings, failure taxonomy, and open-source tooling establish that citation URL validity is both measurable at scale and correctable in practice.

Method: Provides a unified account of augmentation strategies along the axis of structured context at inference time. Includes literature-screening protocol, claim-audit framework, and structured cross-paper evidence synthesis to distinguish high-confidence findings from emerging results.

Result: Systematic survey covering in-context learning, prompt engineering, Retrieval-Augmented Generation (RAG), GraphRAG, and CausalRAG with transparent evaluation methodologies.

Conclusion: Concludes with deployment-oriented decision framework and concrete research priorities for trustworthy retrieval-augmented NLP, emphasizing structured context as key to overcoming LLM limitations.

Abstract: Large language models (LLMs) encode vast world knowledge in their parameters, yet they remain fundamentally limited by static knowledge, finite context windows, and weakly structured causal reasoning. This survey provides a unified account of augmentation strategies along a single axis: the degree of structured context supplied at inference time. We cover in-context learning and prompt engineering, Retrieval-Augmented Generation (RAG), GraphRAG, and CausalRAG. Beyond conceptual comparison, we provide a transparent literature-screening protocol, a claim-audit framework, and a structured cross-paper evidence synthesis that distinguishes higher-confidence findings from emerging results. The paper concludes with a deployment-oriented decision framework and concrete research priorities for trustworthy retrieval-augmented NLP.

Method: Introduces EWAD (token-level routing based on inter-teacher agreement) and CPDP (geometric constraint on student position relative to teachers). Evaluates on Bangla datasets with BanglaT5 ablations and Qwen2.5 experiments, plus cross-lingual pseudo label KD across ten languages.

Result: Logit-level KD provides most reliable gains; complex distillation improves semantic similarity for short summaries but degrades longer outputs. Cross-lingual pseudo label KD retains 71-122% of teacher ROUGE L at 3.2x compression. Human-validated LLM evaluation reveals calibration bias in single-judge pipelines.

Conclusion: Reliability-aware distillation helps characterize when multi-teacher supervision improves summarization versus when data scaling outweighs loss engineering.

Abstract: We study multiteacher knowledge distillation for low resource abstractive summarization from a reliability aware perspective. We introduce EWAD (Entropy Weighted Agreement Aware Distillation), a token level mechanism that routes supervision between teacher distillation and gold supervision based on inter teacher agreement, and CPDP (Capacity Proportional Divergence Preservation), a geometric constraint on the student position relative to heterogeneous teachers. Across two Bangla datasets, 13 BanglaT5 ablations, and eight Qwen2.5 experiments, we find that logit level KD provides the most reliable gains, while more complex distillation improves semantic similarity for short summaries but degrades longer outputs. Cross lingual pseudo label KD across ten languages retains 71-122 percent of teacher ROUGE L at 3.2x compression. A human validated multi judge LLM evaluation further reveals calibration bias in single judge pipelines. Overall, our results show that reliability aware distillation helps characterize when multi teacher supervision improves summarization and when data scaling outweighs loss engineering.

Method: LT-MIA (Learned Transfer MIA) reframes membership inference as sequence classification over per-token distributional statistics. Trains a classifier exclusively on transformer-based models using unlimited labeled data from fine-tuning, then tests zero-shot transfer to other architectures (Mamba, RWKV-4, RecurrentGemma).

Result: Achieves strong zero-shot transfer: 0.963 AUC on Mamba, 0.972 on RWKV-4, 0.936 on RecurrentGemma, all exceeding performance on held-out transformers (0.908 AUC). On transformers, achieves 2.8× higher TPR at 0.1% FPR than strongest baseline. Also transfers to code (0.865 AUC) despite training only on natural language.

Conclusion: Fine-tuning language models produces an invariant signature of memorization detectable across architectural families and data domains, suggesting gradient descent on cross-entropy loss creates universal memorization patterns. Learned attacks outperform heuristic methods and transfer effectively.

Abstract: All prior membership inference attacks for fine-tuned language models use hand-crafted heuristics (e.g., loss thresholding, Min-K%, reference calibration), each bounded by the designer’s intuition. We introduce the first transferable learned attack, enabled by the observation that fine-tuning any model on any corpus yields unlimited labeled data, since membership is known by construction. This removes the shadow model bottleneck and brings membership inference into the deep learning era: learning what matters rather than designing it, with generalization through training diversity and scale. We discover that fine-tuning language models produces an invariant signature of memorization detectable across architectural families and data domains. We train a membership inference classifier exclusively on transformer-based models. It transfers zero-shot to Mamba (state-space), RWKV-4 (linear attention), and RecurrentGemma (gated recurrence), achieving 0.963, 0.972, and 0.936 AUC respectively. Each evaluation combines an architecture and dataset never seen during training, yet all three exceed performance on held-out transformers (0.908 AUC). These four families share no computational mechanisms, their only commonality is gradient descent on cross-entropy loss. Even simple likelihood-based methods exhibit strong transfer, confirming the signature exists independently of the detection method. Our method, Learned Transfer MIA (LT-MIA), captures this signal most effectively by reframing membership inference as sequence classification over per-token distributional statistics. On transformers, LT-MIA achieves 2.8$\times$ higher TPR at 0.1% FPR than the strongest baseline. The method also transfers to code (0.865 AUC) despite training only on natural language texts. Code and trained classifier available at https://github.com/JetBrains-Research/learned-mia.

Method: Introduces Behavioral Alignment Score (BAS), derived from an explicit answer-or-abstain utility model that aggregates realized utility across risk thresholds, measuring decision-level reliability based on confidence magnitude and ordering.

Result: Shows substantial variation in decision-useful confidence across LLMs; larger/more accurate models tend to achieve higher BAS but even frontier models remain prone to severe overconfidence; models with similar ECE/AURC can have very different BAS.

Conclusion: Provides both a principled metric (BAS) and comprehensive benchmark for evaluating LLM confidence reliability, showing simple interventions like top-k confidence elicitation and post-hoc calibration can meaningfully improve confidence reliability.

Abstract: Large language models (LLMs) often produce confident but incorrect answers in settings where abstention would be safer. Standard evaluation protocols, however, require a response and do not account for how confidence should guide decisions under different risk preferences. To address this gap, we introduce the Behavioral Alignment Score (BAS), a decision-theoretic metric for evaluating how well LLM confidence supports abstention-aware decision making. BAS is derived from an explicit answer-or-abstain utility model and aggregates realized utility across a continuum of risk thresholds, yielding a measure of decision-level reliability that depends on both the magnitude and ordering of confidence. We show theoretically that truthful confidence estimates uniquely maximize expected BAS utility, linking calibration to decision-optimal behavior. BAS is related to proper scoring rules such as log loss, but differs structurally: log loss penalizes underconfidence and overconfidence symmetrically, whereas BAS imposes an asymmetric penalty that strongly prioritizes avoiding overconfident errors. Using BAS alongside widely used metrics such as ECE and AURC, we then construct a benchmark of self-reported confidence reliability across multiple LLMs and tasks. Our results reveal substantial variation in decision-useful confidence, and while larger and more accurate models tend to achieve higher BAS, even frontier models remain prone to severe overconfidence. Importantly, models with similar ECE or AURC can exhibit very different BAS due to highly overconfident errors, highlighting limitations of standard metrics. We further show that simple interventions, such as top-$k$ confidence elicitation and post-hoc calibration, can meaningfully improve confidence reliability. Overall, our work provides both a principled metric and a comprehensive benchmark for evaluating LLM confidence reliability.

Method: Proposes the prefilling attack: prepending a structured natural-language prefix (e.g., “The correct option is:”) to model output to steer it toward responding with clean, valid options without modifying model parameters. This repurposes an AI safety technique for evaluation enhancement.

Result: FTP with prefilling substantially improves accuracy, calibration, and output consistency across diverse LLMs and MCQA benchmarks. It outperforms standard FTP and often matches open-ended generation approaches requiring full decoding and external classifiers, while being significantly more efficient.

Conclusion: Prefilling is a simple, robust, low-cost method to enhance reliability of FTP-based evaluation in multiple-choice settings, addressing fragility issues while maintaining efficiency advantages.

Abstract: Large Language Models (LLMs) are increasingly evaluated on multiple-choice question answering (MCQA) tasks using first-token probability (FTP), which selects the answer option whose initial token has the highest likelihood. While efficient, FTP can be fragile: models may assign high probability to unrelated tokens (misalignment) or use a valid token merely as part of a generic preamble rather than as a clear answer choice (misinterpretation), undermining the reliability of symbolic evaluation. We propose a simple solution: the prefilling attack, a structured natural-language prefix (e.g., “The correct option is:”) prepended to the model output. Originally explored in AI safety, we repurpose prefilling to steer the model to respond with a clean, valid option, without modifying its parameters. Empirically, the FTP with prefilling strategy substantially improves accuracy, calibration, and output consistency across a broad set of LLMs and MCQA benchmarks. It outperforms standard FTP and often matches the performance of open-ended generation approaches that require full decoding and external classifiers, while being significantly more efficient. Our findings suggest that prefilling is a simple, robust, and low-cost method to enhance the reliability of FTP-based evaluation in multiple-choice settings.

Method: The authors conduct comprehensive experiments across 4 popular open-source models (3B to 32B parameters) and 2 downstream datasets. They demonstrate that by performing simple backdoor training on the original open-source model, creators can extract private fine-tuning data with only black-box access to the downstream fine-tuned model. They also explore a detection-based defense strategy and test its effectiveness.

Result: The extraction performance is strikingly high: in practical settings, up to 76.3% of downstream fine-tuning data (queries) from 5,000 samples can be perfectly extracted. In more ideal settings, the success rate increases to 94.9%. The detection-based defense strategy was found to be bypassable with improved attacks.

Conclusion: The paper highlights an emergency-level data breaching risk in the fine-tuning practice of open-source LLMs. It calls for more follow-up research to address this concerning security vulnerability in the LLM ecosystem.

Abstract: Fine-tuning on open-source Large Language Models (LLMs) with proprietary data is now a standard practice for downstream developers to obtain task-specific LLMs. Surprisingly, we reveal a new and concerning risk along with the practice: the creator of the open-source LLMs can later extract the private downstream fine-tuning data through simple backdoor training, only requiring black-box access to the fine-tuned downstream model. Our comprehensive experiments, across 4 popularly used open-source models with 3B to 32B parameters and 2 downstream datasets, suggest that the extraction performance can be strikingly high: in practical settings, as much as 76.3% downstream fine-tuning data (queries) out of a total 5,000 samples can be perfectly extracted, and the success rate can increase to 94.9% in more ideal settings. We also explore a detection-based defense strategy but find it can be bypassed with improved attack. Overall, we highlight the emergency of this newly identified data breaching risk in fine-tuning, and we hope that more follow-up research could push the progress of addressing this concerning risk. The code and data used in our experiments are released at https://github.com/thu-coai/Backdoor-Data-Extraction.

Method: Proposes DebateCV with two Debaters arguing opposing stances to surface subtle errors, and a decisive Moderator weighing evidential strength. Also introduces Debate-SFT, a post-training framework using synthetic data to enhance agents’ ability to adjudicate debates effectively.

Result: Methods surpass state-of-the-art non-debate approaches in both accuracy (across various evidence conditions) and justification quality.

Conclusion: Debate-driven verification with multiple LLM agents and specialized training improves claim verification performance for complex claims requiring nuanced evidence analysis.

Abstract: State-of-the-art single-agent claim verification methods struggle with complex claims that require nuanced analysis of multifaceted evidence. Inspired by real-world professional fact-checkers, we propose \textbf{DebateCV}, the first debate-driven claim verification framework powered by multiple LLM agents. In DebateCV, two \textit{Debaters} argue opposing stances to surface subtle errors in single-agent assessments. A decisive \textit{Moderator} is then required to weigh the evidential strength of conflicting arguments to deliver an accurate verdict. Yet, zero-shot Moderators are biased toward neutral judgments, and no datasets exist for training them. To bridge this gap, we propose \textbf{Debate-SFT}, a post-training framework that leverages synthetic data to enhance agents’ ability to effectively adjudicate debates for claim verification. Results show that our methods surpass state-of-the-art non-debate approaches in both accuracy (across various evidence conditions) and justification quality.

Method: AutoPCR uses prompt-based phenotype concept recognition designed to automatically generalize to new ontologies and unseen data without ontology-specific training, with an optional self-supervised training strategy.

Result: AutoPCR achieves the best average and most robust performance across datasets, with ablation and transfer studies demonstrating its inductive capability and generalizability to new ontologies.

Conclusion: AutoPCR effectively addresses limitations of existing methods by providing a prompt-based approach that generalizes well to new ontologies without requiring ontology-specific training.

Abstract: Motivation: Phenotype concept recognition (CR) is a fundamental task in biomedical text mining. However, existing methods either require ontology-specific training, making them struggle to generalize across diverse text styles and evolving biomedical terminology, or depend on general-purpose large language models (LLMs) that lack necessary domain knowledge. Results: To address these limitations, we propose AutoPCR, a prompt-based phenotype CR method designed to automatically generalize to new ontologies and unseen data without ontology-specific training. To further boost performance, we also introduce an optional self-supervised training strategy. Experiments show that AutoPCR achieves the best average and most robust performance across datasets. Further ablation and transfer studies demonstrate its inductive capability and generalizability to new ontologies. Availability and Implementation: Our code is available at https://github.com/yctao7/AutoPCR. Contact: drjieliu@umich.edu

Method: Proposes IFRAgent framework with Intention Flow Recognition: 1) analyzes explicit intention flows to build SOP vector library, 2) analyzes implicit flows to build user habit repository, 3) uses SOP extractor with retrieval-augmented generation and query rewriter to generate personalized queries and SOP from ambiguous raw queries.

Result: IFRAgent outperforms baselines by average 6.79% (32.06% relative improvement) in human intention alignment rate and improves step completion rates by average 5.30% (26.34% relative improvement).

Conclusion: The framework successfully enhances mobile-use agents’ alignment with human intent by considering both explicit and implicit intention flows, enabling more personalized and effective mobile task automation.

Abstract: As multimodal large language models advance rapidly, the automation of mobile tasks has become increasingly feasible through the use of mobile-use agents that mimic human interactions from graphical user interface. To further enhance mobile-use agents, previous studies employ demonstration learning to improve mobile-use agents from human demonstrations. However, these methods focus solely on the explicit intention flows of humans (e.g., step sequences) while neglecting implicit intention flows (e.g., personal preferences), which makes it difficult to construct personalized mobile-use agents. In this work, to evaluate the \textbf{I}ntention \textbf{A}lignment \textbf{R}ate between mobile-use agents and humans, we first collect \textbf{MobileIAR}, a dataset containing human-intent-aligned actions and ground-truth actions. This enables a comprehensive assessment of the agents’ understanding of human intent. Then we propose \textbf{IFRAgent}, a framework built upon \textbf{I}ntention \textbf{F}low \textbf{R}ecognition from human demonstrations. IFRAgent analyzes explicit intention flows from human demonstrations to construct a query-level vector library of standard operating procedures (SOP), and analyzes implicit intention flows to build a user-level habit repository. IFRAgent then leverages a SOP extractor combined with retrieval-augmented generation and a query rewriter to generate personalized query and SOP from a raw ambiguous query, enhancing the alignment between mobile-use agents and human intent. Experimental results demonstrate that IFRAgent outperforms baselines by an average of 6.79% (32.06% relative improvement) in human intention alignment rate and improves step completion rates by an average of 5.30% (26.34% relative improvement). The codes are available at https://github.com/MadeAgents/Quick-on-the-Uptake.

Method: Theory-driven annotation scheme for solidarity/anti-solidarity subtypes; comprehensive evaluation of multiple LLMs (including GPT-5 and gpt-oss-120B) with analysis of model size, prompting strategies, fine-tuning, and error patterns; used Design-based Supervised Learning (DSL) to reduce bias in trend estimates.

Result: Strongest LLMs achieve human-level agreement on annotation task; systematic errors bias downstream results; DSL reduces bias in long-term trend estimates; analysis shows high postwar solidarity and marked rise in anti-solidarity since 2015.

Conclusion: LLMs can support large-scale social-scientific text analysis but require rigorous validation and statistical correction; systematic errors must be addressed to ensure valid inference.

Abstract: Migration has been a core topic in German political debate, from the postwar displacement of millions of expellees to labor migration and recent refugee movements. Studying political speech across such wide-ranging phenomena in depth has traditionally required extensive manual annotation, limiting analysis to small subsets of the data. Large language models (LLMs) offer a potential way to overcome this constraint. Using a theory-driven annotation scheme, we examine how well LLMs annotate subtypes of solidarity and anti-solidarity in German parliamentary debates and whether the resulting labels support valid downstream inference. We first provide a comprehensive evaluation of multiple LLMs, analyzing the effects of model size, prompting strategies, fine-tuning, historical versus contemporary data, and systematic error patterns. We find that the strongest models, especially GPT-5 and gpt-oss-120B, achieve human-level agreement on this task, although their errors remain systematic and bias downstream results. To address this issue, we combine soft-label model outputs with Design-based Supervised Learning (DSL) to reduce bias in long-term trend estimates. Beyond the methodological evaluation, we interpret the resulting annotations from a social-scientific perspective to trace trends in solidarity and anti-solidarity toward migrants in postwar and contemporary Germany. Our approach shows relatively high levels of solidarity in the postwar period, especially in group-based and compassionate forms, and a marked rise in anti-solidarity since 2015, framed through exclusion, undeservingness, and resource burden. We argue that LLMs can support large-scale social-scientific text analysis, but only when their outputs are rigorously validated and statistically corrected.

Method: Proposes a query-driven human-agent-GUI interaction framework and VeriOS-Agent trained with a three-stage learning paradigm using supervised fine-tuning and group relative policy optimization to decouple and utilize meta-knowledge.

Result: Improves average step-wise success rate by 19.72% over strongest baselines in untrustworthy scenarios while maintaining comparable performance in trustworthy scenarios. Demonstrates rationality, generalizability, and scalability.

Conclusion: VeriOS-Agent provides a trustworthy OS agent solution that balances autonomous execution with human querying in uncertain environments, significantly improving safety and reliability in real-world GUI automation tasks.

Abstract: With the rapid progress of multimodal large language models, operating system (OS) agents become increasingly capable of automating tasks through on-device graphical user interfaces (GUIs). However, most existing OS agents are designed for idealized settings, whereas real-world environments often present untrustworthy conditions. To mitigate risks of over-execution in such scenarios, we propose a query-driven human-agent-GUI interaction framework that enables OS agents to decide when to query humans for more reliable task completion. Built upon this framework, we introduce VeriOS-Agent, a trustworthy OS agent trained with a three-stage learning paradigm that falicitate the decoupling and utilization of meta-knowledge by supervised fine-tuning and group relative policy optimization. Concretely, VeriOS-Agent autonomously executes actions in normal conditions while proactively querying humans in untrustworthy scenarios. Experiments show that VeriOS-Agent improves the average step-wise success rate by 19.72% in over the strongest baselines, without compromising normal performance. VeriOS-Agent significantly improves performance in untrustworthy scenarios while maintaining comparable performance in trustworthy scenarios. Analysis highlights VeriOS-Agent’s rationality, generalizability, and scalability. The codes, datasets and models are available at https://github.com/Wuzheng02/VeriOS.

Method: Created SciNLP dataset with 60 manually annotated full-text NLP publications, covering entities and relations. Conducted comparative experiments with existing datasets, evaluated state-of-the-art supervised models, and used trained models to automatically construct a fine-grained NLP knowledge graph.

Result: Existing models show varying extraction capabilities across academic text lengths. SciNLP achieves significant performance improvements on certain baseline models. The constructed knowledge graph has an average node degree of 3.3 per entity, indicating rich semantic topological information.

Conclusion: SciNLP provides the first full-text annotated dataset for entity and relation extraction in NLP domain, enabling better structured information extraction and knowledge graph construction for specialized scientific fields.

Abstract: Structured information extraction from scientific literature is crucial for capturing core concepts and emerging trends in specialized fields. While existing datasets aid model development, most focus on specific publication sections due to domain complexity and the high cost of annotating scientific texts. To address this limitation, we introduce SciNLP - a specialized benchmark for full-text entity and relation extraction in the Natural Language Processing (NLP) domain. The dataset comprises 60 manually annotated full-text NLP publications, covering 6,409 entities and 1,648 relations. Compared to existing research, SciNLP is the first dataset providing full-text annotations of entities and their relationships in the NLP domain. To validate the effectiveness of SciNLP, we conducted comparative experiments with similar datasets and evaluated the performance of state-of-the-art supervised models on this dataset. Results reveal varying extraction capabilities of existing models across academic texts of different lengths. Cross-comparisons with existing datasets show that SciNLP achieves significant performance improvements on certain baseline models. Using models trained on SciNLP, we implemented automatic construction of a fine-grained knowledge graph for the NLP domain. Our KG has an average node degree of 3.3 per entity, indicating rich semantic topological information that enhances downstream applications. The dataset is publicly available at: https://github.com/AKADDC/SciNLP.

Method: Compared two profiling approaches for eight open-source LLMs: (1) self-reported Likert scores from established questionnaires (PVQ-40, PVQ-21, BFI-44, BFI-10), and (2) generation probability scores of value- or personality-laden responses to real-world user queries.

Result: The two profiles were substantially different, suggesting LLMs’ questionnaire responses reflect desired behavior rather than stable psychological constructs. Established questionnaires also risk exaggerating demographic biases of LLMs.

Conclusion: Psychological profiles from human questionnaires should be interpreted cautiously; generation-based profiling is a more reliable approach for LLM psychometrics.

Abstract: Psychological profiling of large language models (LLMs) using psychometric questionnaires designed for humans has become widespread. However, it remains unclear whether the resulting profiles mirror the models’ psychological characteristics expressed during their real-world interactions with users. To examine the risk of human questionnaires mischaracterizing LLM psychology, we compare two types of profiles for eight open-source LLMs: self-reported Likert scores from established questionnaires (PVQ-40, PVQ-21, BFI-44, BFI-10) and generation probability scores of value- or personality-laden responses to real-world user queries. The two profiles turn out to be substantially different and provide evidence that LLMs’ responses to established questionnaires reflect desired behavior rather than stable psychological constructs, which challenges the consistent psychological dispositions of LLMs claimed in prior work. Established questionnaires also risk exaggerating the demographic biases of LLMs. Our results suggest caution when interpreting psychological profiles derived from established questionnaires and point to generation-based profiling as a more reliable approach to LLM psychometrics.

Method: Proposes Future Policy Approximation (FPA) that weights gradients against an estimate of the future policy rather than the current one, using logit-space extrapolation with minimal overhead to estimate future policies.

Result: FPA consistently improves over strong offline baselines (DPO, RPO, KTO, vanilla offline RL) across three models and seven mathematical benchmarks, stabilizes long-horizon training, and achieves comparable accuracy to online RLVR with much lower GPU hours.

Conclusion: FPA provides an effective offline RL method for LLM reasoning that addresses gradient entanglement, offering stability and efficiency comparable to online methods while maintaining offline training advantages.

Abstract: Reinforcement Learning (RL) has emerged as the key driver for post-training complex reasoning in Large Language Models (LLMs), yet online RL introduces significant instability and computational overhead. Offline RL offers a compelling alternative by decoupling inference from training; however, offline algorithms for reasoning remain under-optimized compared to their online counterparts. A central challenge is gradient entanglement: in long-horizon reasoning trajectories, correct and incorrect solutions share substantial token overlap, causing gradient updates from incorrect trajectories to suppress tokens critical for correct ones. We propose Future Policy Approximation (FPA), a simple method that weights gradients against an estimate of the future policy rather than the current one, enabling proactive gradient reweighting. This future policy is estimated via logit-space extrapolation with negligible overhead. We provide theoretical intuition for FPA through the lens of Optimistic Mirror Descent and further ground it through its connection to DPO. Evaluating FPA across three models and seven mathematical benchmarks, we demonstrate consistent improvements over strong offline baselines including DPO, RPO, KTO, and vanilla offline RL. FPA stabilizes long-horizon training where vanilla objectives degrade and achieves comparable accuracy to online RLVR at a fraction of its GPU hours.

Method: Takes a data-centric perspective with four research questions: (1) political leaning in data, (2) data imbalance, (3) cross-dataset similarity, and (4) data-model alignment. Analyzes political content of pre- and post-training datasets of open-source LLMs using large-scale sampling, political-leaning classification, and stance detection techniques.

Result: Training data is systematically skewed toward left-leaning content, with pre-training corpora containing substantially more politically engaged material than post-training data. Strong correlation between political stances in training data and model behavior. Pre-training datasets exhibit similar political distributions despite different curation strategies. Political biases are already present in base models and persist across post-training stages.

Conclusion: Findings highlight the central role of data composition in shaping model behavior and motivate the need for greater data transparency. The data-centric approach provides insights into how political biases emerge in LLMs.

Abstract: Large language models (LLMs) are known to generate politically biased text. Yet, it remains unclear how such biases arise, making it difficult to design effective mitigation strategies. We hypothesize that these biases are rooted in the composition of training data. Taking a data-centric perspective, we formulate research questions on (1) political leaning present in data, (2) data imbalance, (3) cross-dataset similarity, and (4) data-model alignment. We then examine how exposure to political content relates to models’ stances on policy issues. We analyze the political content of pre- and post-training datasets of open-source LLMs, combining large-scale sampling, political-leaning classification, and stance detection. We find that training data is systematically skewed toward left-leaning content, with pre-training corpora containing substantially more politically engaged material than post-training data. We further observe a strong correlation between political stances in training data and model behavior, and show that pre-training datasets exhibit similar political distributions despite different curation strategies. In addition, we find that political biases are already present in base models and persist across post-training stages. These findings highlight the central role of data composition in shaping model behavior and motivate the need for greater data transparency.

Method: Developed CQA-Eval framework with physician annotations of 300 real patient questions answered by physicians and LLMs. Compared coarse answer-level vs fine-grained sentence-level evaluation across three dimensions: correctness, relevance, and risk disclosure. Analyzed inter-annotator agreement and cost-effectiveness.

Result: Inter-annotator agreement varies by dimension: fine-grained annotation improves agreement on correctness, coarse improves agreement on relevance, and risk disclosure judgments remain inconsistent. Annotating only a small subset of sentences provides reliability comparable to coarse annotations, reducing cost and effort.

Conclusion: CQA-Eval provides practical evaluation recommendations for limited-resource, high-expertise settings in clinical QA. Different evaluation strategies work best for different dimensions, and targeted sentence-level annotation can balance reliability with resource constraints.

Abstract: Evaluating multi-paragraph clinical question answering (QA) systems is resource-intensive and challenging: accurate judgments require medical expertise and achieving consistent human judgments over multi-paragraph text is difficult. We introduce CQA-Eval, an evaluation framework and set of evaluation recommendations for limited-resource and high-expertise settings. Based on physician annotations of 300 real patient questions answered by physicians and LLMs, we compare coarse answer-level versus fine-grained sentence-level evaluation over the dimensions of correctness, relevance, and risk disclosure. We find that inter-annotator agreement (IAA) varies by dimension: fine-grained annotation improves agreement on correctness, coarse improves agreement on relevance, and judgments on communicates-risks remain inconsistent. Additionally, annotating only a small subset of sentences can provide reliability comparable to coarse annotations, reducing cost and effort.

Method: Proposes Adaptive Retrieval-Augmented Masked Diffusion (ARAM), a training-free adaptive guidance framework for Masked Diffusion Models in RAG settings. ARAM dynamically calibrates guidance scale during denoising according to the Signal-to-Noise Ratio (SNR) of the distributional shift induced by retrieved context, strengthening guidance when context provides reliable evidence and suppressing it when contextual signal is noisy.

Result: Extensive experiments on multiple knowledge-intensive QA benchmarks show that ARAM improves overall QA performance over competitive RAG baselines.

Conclusion: ARAM effectively addresses retrieval-prior conflicts in diffusion-based language models through adaptive guidance based on context reliability, improving factual grounding in RAG settings.

Abstract: Retrieval-Augmented Generation (RAG) improves factual grounding by incorporating external knowledge into language model generation. However, when retrieved context is noisy, unreliable, or inconsistent with the model’s parametric knowledge, it introduces retrieval-prior conflicts that can degrade generation quality. While this problem has been studied in autoregressive language models, it remains largely unexplored in diffusion-based language models, where the iterative denoising process introduces unique challenges for integrating retrieved context. In this work, we propose Adaptive Retrieval-Augmented Masked Diffusion (ARAM), a training-free adaptive guidance framework for Masked Diffusion Models (MDMs) in RAG settings. ARAM dynamically calibrates the guidance scale during denoising according to the Signal-to-Noise Ratio (SNR) of the distributional shift induced by retrieved context. Intuitively, the model strengthens guidance when the retrieved context provides reliable corrective evidence and suppresses it when the contextual signal is noisy or non-supportive. Extensive experiments on multiple knowledge-intensive QA benchmarks show that ARAM improves overall QA performance over competitive RAG baselines.

Method: Created IslamicMMLU benchmark with 10,013 multiple-choice questions across three tracks: Quran (2,013 questions), Hadith (4,000 questions), and Fiqh (4,000 questions). Includes madhab bias detection task. Evaluated 26 LLMs on public leaderboard.

Result: Model accuracy ranged from 39.8% to 93.8% (Gemini 3 Flash). Quran track showed widest performance span (32.4%-99.3%). Arabic-specific models underperformed frontier models. Fiqh track revealed variable school-of-thought preferences across models.

Conclusion: IslamicMMLU provides crucial evaluation framework for LLMs on Islamic knowledge, revealing significant performance variations, biases, and the need for specialized Islamic knowledge benchmarks.

Abstract: Large language models are increasingly consulted for Islamic knowledge, yet no comprehensive benchmark evaluates their performance across core Islamic disciplines. We introduce IslamicMMLU, a benchmark of 10,013 multiple-choice questions spanning three tracks: Quran (2,013 questions), Hadith (4,000 questions), and Fiqh (jurisprudence, 4,000 questions). Each track is formed of multiple types of questions to examine LLMs capabilities handling different aspects of Islamic knowledge. The benchmark is used to create the IslamicMMLU public leaderboard for evaluating LLMs, and we initially evaluate 26 LLMs, where their averaged accuracy across the three tracks varied between 39.8% to 93.8% (by Gemini 3 Flash). The Quran track shows the widest span (99.3% to 32.4%), while the Fiqh track includes a novel madhab (Islamic school of jurisprudence) bias detection task revealing variable school-of-thought preferences across models. Arabic-specific models show mixed results, but they all underperform compared to frontier models. The evaluation code and leaderboard are made publicly available.

Method: APEX-EM introduces: (1) structured experience representation encoding full procedural-episodic traces; (2) Plan-Retrieve-Generate-Iterate-Ingest workflow with Task Verifiers providing multi-dimensional reward signals; (3) dual-outcome Experience Memory with hybrid retrieval combining semantic search, structural signature matching, and plan DAG traversal.

Result: On KGQAGen-10k: 89.6% accuracy vs 41.3% without memory (+48.3pp), surpassing oracle-retrieval upper bound (84.9%). On BigCodeBench: 83.3% SR from 53.9% baseline (+29.4pp). On HLE: entity graph retrieval reaches 48.0% from 25.2% (+22.8pp).

Conclusion: APEX-EM enables cross-domain transfer between tasks sharing no lexical overlap but analogous operational structure, with successful experiences serving as positive in-context examples and failures as negative examples with structured error annotations.

Abstract: LLM-based autonomous agents lack persistent procedural memory: they re-derive solutions from scratch even when structurally identical tasks have been solved before. We present APEX-EM, a non-parametric online learning framework that accumulates, retrieves, and reuses structured procedural plans without modifying model weights. APEX-EM introduces: (1) a structured experience representation encoding the full procedural-episodic trace of each execution – planning steps, artifacts, iteration history with error analysis, and quality scores; (2) a Plan-Retrieve-Generate-Iterate-Ingest (PRGII) workflow with Task Verifiers providing multi-dimensional reward signals; and (3) a dual-outcome Experience Memory with hybrid retrieval combining semantic search, structural signature matching, and plan DAG traversal – enabling cross-domain transfer between tasks sharing no lexical overlap but analogous operational structure. Successful experiences serve as positive in-context examples; failures as negative examples with structured error annotations. We evaluate on BigCodeBench, KGQAGen-10k, and Humanity’s Last Exam using Claude Sonnet 4.5 and Opus 4.5. On KGQAGen-10k, APEX-EM achieves 89.6% accuracy versus 41.3% without memory (+48.3pp), surpassing the oracle-retrieval upper bound (84.9%). On BigCodeBench, it reaches 83.3% SR from a 53.9% baseline (+29.4pp), exceeding MemRL’s +11.0pp gain under comparable frozen-backbone conditions (noting backbone differences controlled for in our analysis). On HLE, entity graph retrieval reaches 48.0% from 25.2% (+22.8pp). Ablations show component value is task-dependent: rich judge feedback is negligible for code generation but critical for structured queries (+10.3pp), while binary-signal iteration partially compensates for weaker feedback.

Method: Analyzed four model scales (8B-120B parameters) with different citation granularities, measuring attribution quality and answer correctness across sentence-level, paragraph-level, and multi-paragraph citation schemes

Result: Fine-grained citations degrade attribution quality by 16-276% compared to best-performing granularity; paragraph-level citations consistently perform best; larger models are disproportionately penalized by fine-grained constraints; citation-optimal granularity improves attribution while preserving answer correctness

Conclusion: Optimizing solely for human verification via fine-grained citations disregards model constraints and compromises attribution faithfulness; effective attribution requires aligning citation granularity with models’ natural semantic scope rather than forcing atomic citation units

Abstract: Citation granularity - whether to cite individual sentences, paragraphs, or documents - is a critical design choice in attributed generation. While fine-grained citations are often preferred for precise human verification, their impact on model performance remains under-explored. We analyze four model scales (8B-120B) and demonstrate that enforcing fine-grained citations degrades attribution quality by 16-276% compared to the best-performing granularity. We observe a consistent performance pattern where attribution quality peaks at intermediate granularities (paragraph-level). Our analysis suggests that fine-grained (sentence-level) citations disrupt necessary semantic dependencies for attributing evidence to answer claims, while excessively coarse citations (multi-paragraph) introduce distracting noise. Importantly, the magnitude of this performance gap varies non-monotonically with model scale: fine-grained constraints disproportionately penalize larger models, suggesting that atomic citation units disrupt the multi-sentence information synthesis at which these models excel. Strikingly, citation-optimal granularity leads to substantial gains in attribution quality while preserving or even improving answer correctness. Overall, our findings demonstrate that optimizing solely for human verification via fine-grained citation disregards model constraints, compromising both attribution faithfulness and generation reliability. Instead, effective attribution requires aligning citation granularity with the model’s natural semantic scope.

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 2509.17608: Page request resulted in HTTP 429 (https://export.arxiv.org/api/query?search_query=&id_list=2509.17608&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 2510.27176: Page request resulted in HTTP 429 (https://export.arxiv.org/api/query?search_query=&id_list=2510.27176&sortBy=relevance&sortOrder=descending&start=0&max_results=100)

Method: Cannot determine method due to missing abstract

Result: Cannot determine results due to missing abstract

Conclusion: Cannot determine conclusion due to missing abstract

Abstract: Failed to fetch summary for 2511.02734: Page request resulted in HTTP 429 (https://export.arxiv.org/api/query?search_query=&id_list=2511.02734&sortBy=relevance&sortOrder=descending&start=0&max_results=100)

Method: 1) Synthetic data pipeline generates thinking traces by scoring page relevance, extracting textual evidence, and ordering it by relevance. 2) Apply supervised fine-tuning (SFT) to traces within tags, gated by control token. 3) Internalize reasoning via low-strength model merging. Tested on Qwen3 VL 32B and Mistral Small 3.1 24B.

Result: Qwen3 VL 32B achieves 58.3 on MMLongBenchDoc, surpassing 7× larger Qwen3 VL 235B A22B (57.0). With Mistral, synthetic reasoning outperforms distillation from Thinking version’s traces by 3.8 points on MMLBD-C. Internalized reasoning shows 12.4× fewer mean output tokens compared to explicit reasoning.

Conclusion: The synthetic reasoning pipeline effectively enhances visual long-document understanding, demonstrating that reasoning capabilities can be developed and internalized in vision-language models, leading to improved performance with reduced computational overhead.

Abstract: Visual long-document understanding is critical for enterprise, legal, and scientific applications, yet the best performing open recipes have not explored reasoning, a capability which has driven leaps in math and code performance. We introduce a synthetic data pipeline for reasoning in long-document understanding that generates thinking traces by scoring each page for question relevance, extracting textual evidence and ordering it from most to least relevant. We apply SFT to the resulting traces within \texttt{} tags, gated by a \texttt{} control token, and the resulting reasoning capability is internalized via low-strength model merging. We study Qwen3 VL 32B and Mistral Small 3.1 24B. With Qwen3 VL, we achieve 58.3 on MMLongBenchDoc, surpassing the 7$\times$ larger Qwen3 VL 235B A22B (57.0). With Mistral, we show that synthetic reasoning outperforms distillation from the Thinking version’s traces by 3.8 points on MMLBD-C, and internalized reasoning exhibits 12.4$\times$ fewer mean output tokens compared to explicit reasoning. We release our pipeline for reproducibility and further exploration.

Method: Uses vision Foundation Models (FMs) with two novel transfer strategies: 1) Pixel-wise Clustering (PC) for robust prototype-level global feature alignment, and 2) Distance-Penalized Triplet (DPT) that integrates patch-level spatial autocorrelation patterns by penalizing semantically inconsistent yet spatially adjacent patches.

Result: Extensive experiments on 2023 Turkiye-Syria earthquake data show promising performance in cross-region transfer settings (LODO and SSDC), demonstrating effective building damage mapping capabilities.

Conclusion: Smart Transfer provides a scalable, automated GeoAI solution to accelerate building damage mapping and support rapid disaster response, enhancing disaster resilience in climate-vulnerable regions.

Abstract: Living in a changing climate, human society now faces more frequent and severe natural disasters than ever before. As a consequence, rapid disaster response during the “Golden 72 Hours” of search and rescue becomes a vital humanitarian necessity and community concern. However, traditional disaster damage surveys routinely fail to generalize across distinct urban morphologies and new disaster events. Effective damage mapping typically requires exhaustive and time-consuming manual data annotation. To address this issue, we introduce Smart Transfer, a novel Geospatial Artificial Intelligence (GeoAI) framework, leveraging state-of-the-art vision Foundation Models (FMs) for rapid building damage mapping with post-earthquake Very High Resolution (VHR) imagery. Specifically, we design two novel model transfer strategies: first, Pixel-wise Clustering (PC), ensuring robust prototype-level global feature alignment; second, a Distance-Penalized Triplet (DPT), integrating patch-level spatial autocorrelation patterns by assigning stronger penalties to semantically inconsistent yet spatially adjacent patches. Extensive experiments and ablations from the recent 2023 Turkiye-Syria earthquake show promising performance in multiple cross-region transfer settings, namely Leave One Domain Out (LODO) and Specific Source Domain Combination (SSDC). Moreover, Smart Transfer provides a scalable, automated GeoAI solution to accelerate building damage mapping and support rapid disaster response, offering new opportunities to enhance disaster resilience in climate-vulnerable regions and communities. The data and code are publicly available at https://github.com/ai4city-hkust/SmartTransfer.

Method: Estimates input noise level from local pixel statistics, then adapts inference trajectory (starting point, integration steps, step-size schedule) based on quantitative noise estimate

Result: Improves restoration accuracy and inference efficiency across diverse noise levels and imaging conditions (natural, medical, microscopy images)

Conclusion: Coupling quantitative noise estimation with noise-adaptive flow inference enables robust generalization and efficient denoising across varying noise conditions

Abstract: Diffusion and flow-based generative models have shown strong potential for image restoration. However, image denoising under unknown and varying noise conditions remains challenging, because the learned vector fields may become inconsistent across different noise levels, leading to degraded restoration quality under mismatch between training and inference. To address this issue, we propose a quantitative flow matching framework for adaptive image denoising. The method first estimates the input noise level from local pixel statistics, and then uses this quantitative estimate to adapt the inference trajectory, including the starting point, the number of integration steps, and the step-size schedule. In this way, the denoising process is better aligned with the actual corruption level of each input, reducing unnecessary computation for lightly corrupted images while providing sufficient refinement for heavily degraded ones. By coupling quantitative noise estimation with noise-adaptive flow inference, the proposed method improves both restoration accuracy and inference efficiency. Extensive experiments on natural, medical, and microscopy images demonstrate its robustness and strong generalization across diverse noise levels and imaging conditions.

Method: Introduces PaveBench benchmark with four core tasks: classification, object detection, semantic segmentation, and vision-language QA. Includes PaveVQA dataset with single-turn, multi-turn, and expert-corrected interactions covering recognition, localization, quantitative estimation, and maintenance reasoning. Also presents an agent-augmented VQA framework integrating domain-specific models as tools with vision-language models.

Result: Provides large-scale real-world pavement images with annotations, curated hard-distractor subset for robustness evaluation, and comprehensive evaluation of state-of-the-art methods. The dataset is publicly available on Hugging Face.

Conclusion: PaveBench addresses limitations of current pavement inspection research by providing a unified benchmark that connects visual perception with vision-language analysis, enabling more comprehensive and interactive decision support for real-world road maintenance applications.

Abstract: Pavement condition assessment is essential for road safety and maintenance. Existing research has made significant progress. However, most studies focus on conventional computer vision tasks such as classification, detection, and segmentation. In real-world applications, pavement inspection requires more than visual recognition. It also requires quantitative analysis, explanation, and interactive decision support. Current datasets are limited. They focus on unimodal perception. They lack support for multi-turn interaction and fact-grounded reasoning. They also do not connect perception with vision-language analysis. To address these limitations, we introduce PaveBench, a large-scale benchmark for pavement distress perception and interactive vision-language analysis on real-world highway inspection images. PaveBench supports four core tasks: classification, object detection, semantic segmentation, and vision-language question answering. It provides unified task definitions and evaluation protocols. On the visual side, PaveBench provides large-scale annotations and includes a curated hard-distractor subset for robustness evaluation. It contains a large collection of real-world pavement images. On the multimodal side, we introduce PaveVQA, a real-image question answering (QA) dataset that supports single-turn, multi-turn, and expert-corrected interactions. It covers recognition, localization, quantitative estimation, and maintenance reasoning. We evaluate several state-of-the-art methods and provide a detailed analysis. We also present a simple and effective agent-augmented visual question answering framework that integrates domain-specific models as tools alongside vision-language models. The dataset is available at: https://huggingface.co/datasets/MML-Group/PaveBench.

Method: Uses GPS data and vehicle-side panoramic RGB images with semantic segmentation and depth estimation. Extracts semantic, depth, and position features through three-branch architecture, performs adaptive multimodal fusion via squeeze-excitation attention gating module. For APS prediction, uses dedicated regression head and composite multi-constraint loss.

Result: Achieves RMSE of 3.26 dB for PL, 37.66 ns for DS, 5.05° for ASA, 5.08° for ASD, and mean/median APS cosine similarities of 0.9342/0.9571 on urban V2I measurement dataset.

Conclusion: Proposed framework demonstrates strong accuracy, generalization, and practical potential for intelligent channel prediction in 6G vehicular communications through multimodal visual feature fusion.

Abstract: The deep integration of communication with intelligence and sensing, as a defining vision of 6G, renders environment-aware channel prediction a key enabling technology. As a representative 6G application, vehicular communications require accurate and forward-looking channel prediction under stringent reliability, latency, and adaptability demands. Traditional empirical and deterministic models remain limited in balancing accuracy, generalization, and deployability, while the growing availability of onboard and roadside sensing devices offers a promising source of environmental priors. This paper proposes an environment-aware channel prediction framework based on multimodal visual feature fusion. Using GPS data and vehicle-side panoramic RGB images, together with semantic segmentation and depth estimation, the framework extracts semantic, depth, and position features through a three-branch architecture and performs adaptive multimodal fusion via a squeeze-excitation attention gating module. For 360-dimensional angular power spectrum (APS) prediction, a dedicated regression head and a composite multi-constraint loss are further designed. As a result, joint prediction of path loss (PL), delay spread (DS), azimuth spread of arrival (ASA), azimuth spread of departure (ASD), and APS is achieved. Experiments on a synchronized urban V2I measurement dataset yield the best root mean square error (RMSE) of 3.26 dB for PL, RMSEs of 37.66 ns, 5.05 degrees, and 5.08 degrees for DS, ASA, and ASD, respectively, and mean/median APS cosine similarities of 0.9342/0.9571, demonstrating strong accuracy, generalization, and practical potential for intelligent channel prediction in 6G vehicular communications.

Method: Proposes VE-MD (Variational Encoder-Multi-Decoder) framework that learns shared latent representations jointly optimized for emotion classification and internal prediction of body/facial structural representations. Uses transformer-based PersonQuery decoder and dense Heatmap decoder to accommodate variable group sizes without explicit individual monitoring.

Result: Achieves SOTA on GAF-3.0 (90.06%) and VGAF (82.25% with audio fusion). Shows structural supervision improves representation learning, revealing clear distinction between GER and IER: structural outputs improve collective affect inference for GER while acting as denoising bottleneck for IER.

Conclusion: Preserving interaction-related structural information is particularly beneficial for group-level affect modeling without relying on prior individual feature extraction, enabling privacy-aware group emotion recognition.

Abstract: Group Emotion Recognition (GER) aims to infer collective affect in social environments such as classrooms, crowds, and public events. Many existing approaches rely on explicit individual-level processing, including cropped faces, person tracking, or per-person feature extraction, which makes the analysis pipeline person-centric and raises privacy concerns in deployment scenarios where only group-level understanding is needed. This research proposes VE-MD, a Variational Encoder-Multi-Decoder framework for group emotion recognition under a privacy-aware functional design. Rather than providing formal anonymization or cryptographic privacy guarantees, VE-MD is designed to avoid explicit individual monitoring by constraining the model to predict only aggregate group-level affect, without identity recognition or per-person emotion outputs. VE-MD learns a shared latent representation jointly optimized for emotion classification and internal prediction of body and facial structural representations. Two structural decoding strategies are investigated: a transformer-based PersonQuery decoder and a dense Heatmap decoder that naturally accommodates variable group sizes. Experiments on six in-the-wild datasets, including two GER and four Individual Emotion Recognition (IER) benchmarks, show that structural supervision consistently improves representation learning. More importantly, the results reveal a clear distinction between GER and IER: optimizing the latent space alone is often insufficient for GER because it tends to attenuate interaction-related cues, whereas preserving explicit structural outputs improves collective affect inference. In contrast, projected structural representations seem to act as an effective denoising bottleneck for IER. VE-MD achieves state-of-the-art performance on GAF-3.0 (up to 90.06%) and VGAF (82.25% with multimodal fusion with audio). These results show that preserving interaction-related structural information is particularly beneficial for group-level affect modeling without relying on prior individual feature extraction. On IER datasets using multimodal fusion with audio modality, VE-MD outperforms SOTA on SamSemo (77.9%, adding text modality) while achieving competitive performances on MER-MULTI (63.8%), DFEW (70.7%) and EngageNet (69.0).

Method: 1) Created SuSuInterActs dataset (21K clips, 37 hours) with synchronized speech, full-body motion, and facial expressions. 2) Pre-trained Motion Foundation Model on 200K+ motion sequences for rich action priors. 3) Proposed audio-aware plan-then-infill architecture that decouples sentence-level semantic planning from frame-level prosody-driven interpolation.

Result: Achieved state-of-the-art on SuSuInterActs (R@1 43.64%, nearly 2x best baseline) and BEATv2 (FGD 4.941, BC 8.078), producing 6s output in 0.3s with unlimited multi-turn streaming. Code, model, and dataset publicly available.

Conclusion: SentiAvatar successfully addresses key challenges in expressive 3D digital human generation through comprehensive dataset creation, foundation model pre-training, and novel architecture design, enabling high-quality real-time multimodal character animation.

Abstract: We present SentiAvatar, a framework for building expressive interactive 3D digital humans, and use it to create SuSu, a virtual character that speaks, gestures, and emotes in real time. Achieving such a system remains challenging, as it requires jointly addressing three key problems: the lack of large-scale, high-quality multimodal data, robust semantic-to-motion mapping, and fine-grained frame-level motion-prosody synchronization. To solve these problems, first, we build SuSuInterActs (21K clips, 37 hours), a dialogue corpus captured via optical motion capture around a single character with synchronized speech, full-body motion, and facial expressions. Second, we pre-train a Motion Foundation Model on 200K+ motion sequences, equipping it with rich action priors that go well beyond the conversation. We then propose an audio-aware plan-then-infill architecture that decouples sentence-level semantic planning from frame-level prosody-driven interpolation, so that generated motions are both semantically appropriate and rhythmically aligned with speech. Experiments show that SentiAvatar achieves state-of-the-art on both SuSuInterActs (R@1 43.64%, nearly 2 times the best baseline) and BEATv2 (FGD 4.941, BC 8.078), producing 6s of output in 0.3s with unlimited multi-turn streaming. The source code, model, and dataset are available at https://sentiavatar.github.io.

Method: Four-stage agentic system: Perception (analyzes scene lighting/semantics), Reasoning (LLM+RAG+Tree of Thoughts for parameter navigation), Execution (generates ASC-CDL configs and 3D LUTs), Reflection (natural language feedback loop). Uses analytical methods to guarantee temporal consistency.

Result: LumiVideo approaches human expert quality in automatic mode while enabling precise iterative control. Introduces LumiGrade benchmark for evaluating automated grading on log-encoded video.

Conclusion: LumiVideo successfully bridges the gap between automated color grading and professional workflows by providing interpretable, controllable systems that generate industry-standard outputs rather than direct pixel edits.

Abstract: Video color grading is a critical post-production process that transforms flat, log-encoded raw footage into emotionally resonant cinematic visuals. Existing automated methods act as static, black-box executors that directly output edited pixels, lacking both interpretability and the iterative control required by professionals. We introduce LumiVideo, an agentic system that mimics the cognitive workflow of professional colorists through four stages: Perception, Reasoning, Execution, and Reflection. Given only raw log video, LumiVideo autonomously produces a cinematic base grade by analyzing the scene’s physical lighting and semantic content. Its Reasoning engine synergizes an LLM’s internalized cinematic knowledge with a Retrieval-Augmented Generation (RAG) framework via a Tree of Thoughts (ToT) search to navigate the non-linear color parameter space. Rather than generating pixels, the system compiles the deduced parameters into industry-standard ASC-CDL configurations and a globally consistent 3D LUT, analytically guaranteeing temporal consistency. An optional Reflection loop then allows creators to refine the result via natural language feedback. We further introduce LumiGrade, the first log-encoded video benchmark for evaluating automated grading. Experiments show that LumiVideo approaches human expert quality in fully automatic mode while enabling precise iterative control when directed.

Method: STEAR identifies high-risk decoding steps and selects token-conditioned visual evidence from grounding-sensitive middle layers. It uses this shared evidence for two purposes: 1) restoring missing local grounding in middle layers, and 2) constructing temporally perturbed patch-level counterfactuals to falsify inconsistent reasoning during late-layer decoding. This operates within an efficient single-encode inference framework.

Result: Experiments across representative Video-LLM backbones and challenging benchmarks show STEAR consistently reduces hallucinations while improving faithfulness, temporal consistency, and robustness. The framework demonstrates that reliable video decoding requires intervening on precise evidence at the right layer rather than enforcing global penalties.

Conclusion: STEAR provides an effective layer-aware intervention approach for mitigating spatiotemporal hallucinations in Video-LLMs by leveraging layer-specific contributions to visual grounding and linguistic composition, offering improved reliability over uniform correction methods.

Abstract: Video Large Language Models (Video-LLMs) remain prone to spatiotemporal hallucinations, often generating visually unsupported details or incorrect temporal relations. Existing mitigation methods typically treat hallucination as a uniform decoding failure, applying globally shared correction rules. We instead observe that decoder layers contribute differently to visual grounding and later linguistic composition, indicating that intervention must be layer-aware. Based on this insight, we propose STEAR, a layer-aware spatiotemporal evidence intervention framework. STEAR identifies high-risk decoding steps and selects token-conditioned visual evidence from grounding-sensitive middle layers. It uses this shared evidence for two coupled purposes: restoring missing local grounding in middle layers, and constructing temporally perturbed patch-level counterfactuals to falsify inconsistent reasoning during late-layer decoding. Consequently, STEAR mitigates both spatial and temporal hallucinations within an efficient single-encode inference framework. Experiments across representative Video-LLM backbones and challenging benchmarks demonstrate that STEAR consistently reduces hallucinations while improving faithfulness, temporal consistency, and robustness. Our results confirm that reliable video decoding relies on intervening on precise evidence at the right layer, rather than enforcing a global penalty. The code is provided in the Supplementary Material.

Method: Two main approaches: 1) Geometry-based raw representation using elevation maps from 3D photogrammetric meshes with SVM and Decision Trees, and 2) Feature-engineered approach using parametric contour lines fitting. The methods are compared for detecting width reduction in violins.

Result: Elevation maps occasionally achieve strong results but do not surpass the performance of contour-based inputs. The feature-engineered contour fitting approach outperforms the raw geometry-based methods for violin width reduction detection.

Conclusion: Targeted feature engineering with parametric contour fitting is more effective than raw geometry-based approaches for detecting violin width reduction from 3D photogrammetric meshes, despite elevation maps showing some promise.

Abstract: We explore the automatic detection of violin width reduction using 3D photogrammetric meshes. We compare SVM and Decision Trees applied to a geometry-based raw representation built from elevation maps with a more targeted, feature-engineered approach relying on parametric contour lines fitting. Although elevation maps occasionally achieve strong results, their performance does not surpass that of the contour-based inputs.

Method: Proposes SFFNet with: 1) MDDC module for dual-domain (frequency & spatial) edge extraction to decouple objects from noise, 2) SFPN with linear deformable convolutions for irregular shape capture and wide-area perception module for contextual associations, 3) Six scale variants (N/S/M/B/L/X) for different application scenarios.

Result: Achieves 36.8 AP on VisDrone and 20.6 AP on UAVDT datasets with SFFNet-X. Lightweight models (N/S) maintain balance between accuracy and parameter efficiency.

Conclusion: SFFNet effectively addresses UAV image detection challenges through dual-domain edge enhancement and synergistic feature fusion, showing strong performance on aerial datasets with scalable architecture.

Abstract: Object detection in unmanned aerial vehicle (UAV) images remains a highly challenging task, primarily caused by the complexity of background noise and the imbalance of target scales. Traditional methods easily struggle to effectively separate objects from intricate backgrounds and fail to fully leverage the rich multi-scale information contained within images. To address these issues, we have developed a synergistic feature fusion network (SFFNet) with dual-domain edge enhancement specifically tailored for object detection in UAV images. Firstly, the multi-scale dynamic dual-domain coupling (MDDC) module is designed. This component introduces a dual-driven edge extraction architecture that operates in both the frequency and spatial domains, enabling effective decoupling of multi-scale object edges from background noise. Secondly, to further enhance the representation capability of the model’s neck in terms of both geometric and semantic information, a synergistic feature pyramid network (SFPN) is proposed. SFPN leverages linear deformable convolutions to adaptively capture irregular object shapes and establishes long-range contextual associations around targets through the designed wide-area perception module (WPM). Moreover, to adapt to the various applications or resource-constrained scenarios, six detectors of different scales (N/S/M/B/L/X) are designed. Experiments on two challenging aerial datasets (VisDrone and UAVDT) demonstrate the outstanding performance of SFFNet-X, achieving 36.8 AP and 20.6 AP, respectively. The lightweight models (N/S) also maintain a balance between detection accuracy and parameter efficiency. The code will be available at https://github.com/CQNU-ZhangLab/SFFNet.

Method: Collected DSLR camera photographs of teeth from 295 patients, created over 1200 images with instance segmentation labels across 9 dental classes, and trained neural networks for segmentation.

Result: High-quality predictions were achieved in instance segmentation experiments, and the dataset, code, and model weights are publicly available under open licenses.

Conclusion: AlphaDent provides a valuable resource for dental image analysis research, enabling instance segmentation tasks with good performance and open accessibility.

Abstract: In this article, we present a new unique dataset for dental research - AlphaDent. This dataset is based on the DSLR camera photographs of the teeth of 295 patients and contains over 1200 images. The dataset is labeled for solving the instance segmentation problem and is divided into 9 classes. The article provides a detailed description of the dataset and the labeling format. The article also provides the details of the experiment on neural network training for the Instance Segmentation problem using this dataset. The results obtained show high quality of predictions. The dataset is published under an open license; and the training/inference code and model weights are also available under open licenses.

Method: Three key design choices: 1) Mixture-of-Gaussians output head with shared mixture weights across all agents, 2) relative spatial attention encoding pairwise player positions and distances as learned attention biases, 3) non-autoregressive prediction of absolute displacements from initial formation.

Result: On American football tracking data: 1.68 yard ADE and 3.98 yard FDE, with full utilization of all 8 mixture components (entropy 2.06 out of 2.08), demonstrating diverse generation without mode collapse.

Conclusion: PlayGen-MoG effectively generates diverse, coordinated multi-agent trajectories from static formations, addressing limitations of standard generative approaches and enabling realistic play design without requiring observed history.

Abstract: Multi-agent trajectory generation in team sports requires models that capture both the diversity of possible plays and realistic spatial coordination between players on plays. Standard generative approaches such as Conditional Variational Autoencoders (CVAE) and diffusion models struggle with this task, exhibiting posterior collapse or convergence to the dataset mean. Moreover, most trajectory prediction methods operate in a forecasting regime that requires multiple frames of observed history, limiting their use for play design where only the initial formation is available. We present PlayGen-MoG, an extensible framework for formation-conditioned play generation that addresses these challenges through three design choices: 1/ a Mixture-of-Gaussians (MoG) output head with shared mixture weights across all agents, where a single set of weights selects a play scenario that couples all players’ trajectories, 2/ relative spatial attention that encodes pairwise player positions and distances as learned attention biases, and 3/ non-autoregressive prediction of absolute displacements from the initial formation, eliminating cumulative error drift and removing the dependence on observed trajectory history, enabling realistic play generation from a single static formation alone. On American football tracking data, PlayGen-MoG achieves 1.68 yard ADE and 3.98 yard FDE while maintaining full utilization of all 8 mixture components with entropy of 2.06 out of 2.08, and qualitatively confirming diverse generation without mode collapse.

Method: Combines self-supervised VideoMAE representations with LoRA-based parameter-efficient adaptation, client-side DP-SGD for differential privacy, and server-side secure aggregation in federated learning framework.

Result: Achieves 77.25% accuracy without privacy protection and 65-66% with strong differential privacy on RWF-2000 dataset with 40 clients, while reducing communication by 28.3x compared to full-model updates.

Conclusion: FedVideoMAE provides practical operating point for privacy-preserving video moderation on edge devices, balancing accuracy, communication efficiency, and privacy protection.

Abstract: Short-form video moderation increasingly needs learning pipelines that protect user privacy without paying the full bandwidth and latency cost of cloud-centralized inference. We present FedVideoMAE, an on-device federated framework for video violence detection that combines self-supervised VideoMAE representations, LoRA-based parameter-efficient adaptation, client-side DP-SGD, and server-side secure aggregation. By updating only 5.5M parameters (about 3.5% of a 156M backbone), FedVideoMAE reduces communication by 28.3x relative to full-model federated updates while keeping raw videos on device throughout training. On RWF-2000 with 40 clients, the method reaches 77.25% accuracy without privacy protection and 6566% under strong differential privacy. We further show that this privacy gap is consistent with an effective-SNR analysis tailored to the small-data, parameter-efficient federated regime, which indicates roughly 8.512x DP-noise amplification in our setting. To situate these results more clearly, we also compare against archived full-model federated baselines and summarize auxiliary transfer behavior on RLVS and binary UCF-Crime. Taken together, these findings position FedVideoMAE as a practical operating point for privacy-preserving video moderation on edge devices. Our code can be found at: https://github.com/zyt-599/FedVideoMAE.

Method: Two-stage training framework: 1) Pre-train zero-shot TTS on large corpora with deterministic architecture for linguistic structures and Discrete Flow-based Prosody-Acoustic (DFPA) module for prosody and acoustics; 2) Content-Consistent Temporal Adaptation (CCTA) with Synchronizer for lip-sync alignment and Face-to-Prosody Mapper (FaPro) for facial expression conditioning, fused to create multimodal embeddings guiding DFPA.

Result: Experiments on two benchmark datasets show DiFlowDubber outperforms prior methods across multiple evaluation metrics for video dubbing quality.

Conclusion: DiFlowDubber successfully addresses the four key challenges in video dubbing through its discrete flow matching backbone and novel two-stage training strategy with multimodal conditioning.

Abstract: Video dubbing requires content accuracy, expressive prosody, high-quality acoustics, and precise lip synchronization, yet existing approaches struggle on all four fronts. To address these issues, we propose DiFlowDubber, the first video dubbing framework built upon a discrete flow matching backbone with a novel two-stage training strategy. In the first stage, a zero-shot text-to-speech (TTS) system is pre-trained on large-scale corpora, where a deterministic architecture captures linguistic structures, and the Discrete Flow-based Prosody-Acoustic (DFPA) module models expressive prosody and realistic acoustic characteristics. In the second stage, we propose the Content-Consistent Temporal Adaptation (CCTA) to transfer TTS knowledge to the dubbing domain: its Synchronizer enforces cross-modal alignment for lip-synchronized speech. Complementarily, the Face-to-Prosody Mapper (FaPro) conditions prosody on facial expressions, whose outputs are then fused with those of the Synchronizer to construct rich, fine-grained multimodal embeddings that capture prosody-content correlations, guiding the DFPA to generate expressive prosody and acoustic tokens for content-consistent speech. Experiments on two benchmark datasets demonstrate that DiFlowDubber outperforms prior methods across multiple evaluation metrics.

Method: Proposes adaptive local frequency filtering with spatially varying parameter α(x) to modulate encoded Fourier components, enabling smooth transitions among low-pass, band-pass, and high-pass behaviors at different spatial locations. Analyzes effect from neural tangent kernel (NTK) perspective.

Result: Experiments on 2D image fitting, 3D shape representation, and sparse data reconstruction show consistent improvements in reconstruction quality and faster optimization compared to fixed-frequency baselines. Learned α(x) provides intuitive visualization of spatially varying frequency preferences.

Conclusion: Adaptive local frequency modulation is a practical enhancement for Fourier-encoded INRs that handles non-stationary signals effectively and provides interpretable frequency preference visualizations.

Abstract: Fourier-encoded implicit neural representations (INRs) have shown strong capability in modeling continuous signals from discrete samples. However, conventional Fourier feature mappings use a fixed set of frequencies over the entire spatial domain, making them poorly suited to signals with spatially varying local spectra and often leading to slow convergence of high-frequency details. To address this issue, we propose an adaptive local frequency filtering method for Fourier-encoded INRs. The proposed method introduces a spatially varying parameter $α(\mathbf{x})$ to modulate encoded Fourier components, enabling a smooth transition among low-pass, band-pass, and high-pass behaviors at different spatial locations. We further analyze the effect of the proposed filter from the neural tangent kernel (NTK) perspective and provide an NTK-inspired interpretation of how it reshapes the effective kernel spectrum. Experiments on 2D image fitting, 3D shape representation, and sparse data reconstruction demonstrate that the proposed method consistently improves reconstruction quality and leads to faster optimization compared with fixed-frequency baselines. In addition, the learned $α(\mathbf{x})$ provides an intuitive visualization of spatially varying frequency preferences, which helps explain the behavior of the model on non-stationary signals. These results indicate that adaptive local frequency modulation is a practical enhancement for Fourier-encoded INRs.

Method: The authors introduce SPAR (Street-legal Physical Adversarial Rim), a physically realizable white-box attack against the popular fast-alpr system. The attack requires no access to ALPR infrastructure during deployment and doesn’t alter or obscure the attacker’s license plate. It was implemented entirely by commercial agentic coding assistants and can be produced for under $100.

Result: Under optimal conditions, SPAR reduces ALPR accuracy by 60% and achieves an 18% targeted impersonation rate. The attack is argued to be street-legal in Texas based on prior legislation and case law.

Conclusion: The results highlight practical vulnerabilities in modern ALPR systems under realistic physical-world conditions and suggest new directions for both attack and defense in adversarial machine learning for computer vision systems.

Abstract: Automatic license plate reader (ALPR) systems are widely deployed to identify and track vehicles. While prior work has demonstrated vulnerabilities in ALPR systems, far less attention has been paid to their legality and physical-world practicality. We investigate whether low-resourced threat actors can engineer a successful adversarial attack against a modern open-source ALPR system. We introduce the Street-legal Physical Adversarial Rim (SPAR), a physically realizable white-box attack against the popular ALPR system fast-alpr. SPAR requires no access to ALPR infrastructure during attack deployment and does not alter or obscure the attacker’s license plate. Based on prior legislation and case law, we argue that SPAR is street-legal in the state of Texas. Under optimal conditions, SPAR reduces ALPR accuracy by 60% and achieves an 18% targeted impersonation rate. SPAR can be produced for under $100, and it was implemented entirely by commercial agentic coding assistants. These results highlight practical vulnerabilities in modern ALPR systems under realistic physical-world conditions and suggest new directions for both attack and defense.

Method: Proposes Self-Consistent Distribution Matching Distillation (SC-DMD) that explicitly regularizes endpoint-consistent composition of consecutive denoising updates. For real-time autoregressive video generation, introduces Cache-Distribution-Aware training that treats KV cache as quality parameterized condition, applies SC-DMD over multi-step rollouts, and adds cache-conditioned feature alignment to steer low-quality outputs toward high-quality references.

Result: The method (dubbed Salt) consistently improves low-NFE video generation quality across experiments on both non-autoregressive backbones (e.g., Wan 2.1) and autoregressive real-time paradigms (e.g., Self Forcing), while remaining compatible with diverse KV-cache memory mechanisms.

Conclusion: SC-DMD effectively addresses limitations of existing distillation approaches for video generation, enabling high-quality real-time video generation at extremely low inference budgets through better regularization of denoising update composition and cache-aware training.

Abstract: Distilling video generation models to extremely low inference budgets (e.g., 2–4 NFEs) is crucial for real-time deployment, yet remains challenging. Trajectory-style consistency distillation often becomes conservative under complex video dynamics, yielding an over-smoothed appearance and weak motion. Distribution matching distillation (DMD) can recover sharp, mode-seeking samples, but its local training signals do not explicitly regularize how denoising updates compose across timesteps, making composed rollouts prone to drift. To overcome this challenge, we propose Self-Consistent Distribution Matching Distillation (SC-DMD), which explicitly regularizes the endpoint-consistent composition of consecutive denoising updates. For real-time autoregressive video generation, we further treat the KV cache as a quality parameterized condition and propose Cache-Distribution-Aware training. This training scheme applies SC-DMD over multi-step rollouts and introduces a cache-conditioned feature alignment objective that steers low-quality outputs toward high-quality references. Across extensive experiments on both non-autoregressive backbones (e.g., Wan~2.1) and autoregressive real-time paradigms (e.g., Self Forcing), our method, dubbed \textbf{Salt}, consistently improves low-NFE video generation quality while remaining compatible with diverse KV-cache memory mechanisms. Source code will be released at \href{https://github.com/XingtongGe/Salt}{https://github.com/XingtongGe/Salt}.

Method: Uses Unity engine to render 2D visual previews from generated camera motion, then employs a cinematically fine-tuned vision-language model with cyclic semantic similarity to score renders against text prompts, providing visual preference signals for Direct Preference Optimization post-training.

Result: Reduces character off-screen rate from 38% to nearly 0% while preserving geometric fidelity; shows consistent gains in condition adherence, framing quality, and perceptual realism in both quantitative evaluations and user studies.

Conclusion: VERTIGO successfully integrates visual preference optimization into camera trajectory generation, achieving significant improvements in cinematic quality through a novel framework combining real-time rendering, vision-language models, and preference learning.

Abstract: Cinematic camera control relies on a tight feedback loop between director and cinematographer, where camera motion and framing are continuously reviewed and refined. Recent generative camera systems can produce diverse, text-conditioned trajectories, but they lack this “director in the loop” and have no explicit supervision of whether a shot is visually desirable. This results in in-distribution camera motion but poor framing, off-screen characters, and undesirable visual aesthetics. In this paper, we introduce VERTIGO, the first framework for visual preference optimization of camera trajectory generators. Our framework leverages a real-time graphics engine (Unity) to render 2D visual previews from generated camera motion. A cinematically fine-tuned vision-language model then scores these previews using our proposed cyclic semantic similarity mechanism, which aligns renders with text prompts. This process provides the visual preference signals for Direct Preference Optimization (DPO) post-training. Both quantitative evaluations and user studies on Unity renders and diffusion-based Camera-to-Video pipelines show consistent gains in condition adherence, framing quality, and perceptual realism. Notably, VERTIGO reduces the character off-screen rate from 38% to nearly 0% while preserving the geometric fidelity of camera motion. User study participants further prefer VERTIGO over baselines across composition, consistency, prompt adherence, and aesthetic quality, confirming the perceptual benefits of our visual preference post-training.

Method: HIL-CBM introduces: (1) a gradient-based visual consistency loss that encourages abstraction layers to focus on similar spatial regions, and (2) dual classification heads operating on feature concepts at different abstraction levels. The model enables classification and explanation across multiple semantic levels without requiring relational concept annotations.

Result: Experiments on benchmark datasets show HIL-CBM outperforms state-of-the-art sparse CBMs in classification accuracy. Human evaluations demonstrate that HIL-CBM provides more interpretable and accurate explanations while maintaining a hierarchical and label-free approach to feature concepts.

Conclusion: HIL-CBM successfully extends CBMs into a hierarchical framework that better aligns with human cognitive processes, improving both classification performance and interpretability without requiring additional concept annotations.

Abstract: Concept Bottleneck Models (CBMs) introduce interpretability to black-box deep learning models by predicting labels through human-understandable concepts. However, unlike humans, who identify objects at different levels of abstraction using both general and specific features, existing CBMs operate at a single semantic level in both concept and label space. We propose HIL-CBM, a Hierarchical Interpretable Label-Free Concept Bottleneck Model that extends CBMs into a hierarchical framework to enhance interpretability by more closely mirroring the human cognitive process. HIL-CBM enables classification and explanation across multiple semantic levels without requiring relational concept annotations. HIL-CBM aligns the abstraction level of concept-based explanations with that of model predictions, progressing from abstract to concrete. This is achieved by (i) introducing a gradient-based visual consistency loss that encourages abstraction layers to focus on similar spatial regions, and (ii) training dual classification heads, each operating on feature concepts at different abstraction levels. Experiments on benchmark datasets demonstrate that HIL-CBM outperforms state-of-the-art sparse CBMs in classification accuracy. Human evaluations further show that HIL-CBM provides more interpretable and accurate explanations, while maintaining a hierarchical and label-free approach to feature concepts.

Method: Decomposition-first pipeline with three main components: 1) topology-aware chunking, 2) interface-constrained chunk graph generation, and 3) provenance-preserving global aggregation. Uses explicit entry/terminal interfaces and semantic deduplication to preserve cross-page continuity while maintaining auditable control flow.

Result: On prostate-guideline benchmark, improves edge and triplet precision/recall from 19.6%/16.1% in existing models to 69.0%/87.5%, while node recall rises from 78.1% to 93.8%.

Conclusion: The approach supports decomposition-first, auditable guideline-to-CDS conversion, though current evidence is limited to one prostate guideline and motivates broader multi-guideline validation.

Abstract: Clinical practice guidelines are long, multimodal documents whose branching recommendations are difficult to convert into executable clinical decision support (CDS), and one-shot parsing often breaks cross-page continuity. Recent LLM/VLM extractors are mostly local or text-centric, under-specifying section interfaces and failing to consolidate cross-page control flow across full documents into one coherent decision graph. We present a decomposition-first pipeline that converts full-guideline evidence into an executable clinical decision graph through topology-aware chunking, interface-constrained chunk graph generation, and provenance-preserving global aggregation. Rather than relying on single-pass generation, the pipeline uses explicit entry/terminal interfaces and semantic deduplication to preserve cross-page continuity while keeping the induced control flow auditable and structurally consistent. We evaluate on an adjudicated prostate-guideline benchmark with matched inputs and the same underlying VLM backbone across compared methods. On the complete merged graph, our approach improves edge and triplet precision/recall from $19.6%/16.1%$ in existing models to $69.0%/87.5%$, while node recall rises from $78.1%$ to $93.8%$. These results support decomposition-first, auditable guideline-to-CDS conversion on this benchmark, while current evidence remains limited to one adjudicated prostate guideline and motivates broader multi-guideline validation.

Method: Uses EarthSynth diffusion model conditioned on burn masks from CalFireSeg-50 dataset. Evaluates six configurations varying pipeline architecture (mask-only generation vs. inpainting), prompt engineering strategies (hand-crafted and VLM-generated via Qwen2-VL), and color-matching post-processing.

Result: Inpainting-based pipelines consistently outperform full-tile generation across all metrics. Structured inpainting prompt achieved best spatial alignment (Burn IoU = 0.456) and burn saliency. Color matching produced lowest color distance but reduced burn saliency. VLM-assisted inpainting competitive with hand-crafted prompts.

Conclusion: Demonstrates feasibility of using diffusion models for wildfire imagery synthesis, providing foundation for generative data augmentation in wildfire detection pipelines.

Abstract: The scarcity of labeled satellite imagery remains a fundamental bottleneck for deep-learning (DL)-based wildfire monitoring systems. This paper investigates whether a diffusion-based foundation model for Earth Observation (EO), EarthSynth, can synthesize realistic post-wildfire Sentinel-2 RGB imagery conditioned on existing burn masks, without task-specific retraining. Using burn masks derived from the CalFireSeg-50 dataset (Martin et al., 2025), we design and evaluate six controlled experimental configurations that systematically vary: (i) pipeline architecture (mask-only full generation vs. inpainting with pre-fire context), (ii) prompt engineering strategy (three hand-crafted prompts and a VLM-generated prompt via Qwen2-VL), and (iii) a region-wise color-matching post-processing step. Quantitative assessment on 10 stratified test samples uses four complementary metrics: Burn IoU, burn-region color distance (ΔC_burn), Darkness Contrast, and Spectral Plausibility. Results show that inpainting-based pipelines consistently outperform full-tile generation across all metrics, with the structured inpainting prompt achieving the best spatial alignment (Burn IoU = 0.456) and burn saliency (Darkness Contrast = 20.44), while color matching produces the lowest color distance (ΔC_burn = 63.22) at the cost of reduced burn saliency. VLM-assisted inpainting is competitive with hand-crafted prompts. These findings provide a foundation for incorporating generative data augmentation into wildfire detection pipelines. Code and experiments are available at: https://www.kaggle.com/code/valeriamartinh/genai-all-runned

Method: The study uses visual correspondence tasks (semantic, shape, and face correspondence) to test VLM performance on nameable vs. unnameable entities. They employ Logit Lens analyses to examine how VLMs process visual information and assign semantic labels. They also experiment with teaching arbitrary names for unknown entities and task-specific finetuning.

Result: VLMs perform significantly better when relevant entities are nameable in language compared to when they’re unnameable. Logit Lens analyses confirm VLMs explicitly assign semantic labels to nameable entities and produce more unique corresponding tokens. Teaching arbitrary names improves performance, but task-specific finetuning yields even stronger generalization without language priors.

Conclusion: Current VLM failures on visual tasks reflect learned shortcuts from their training pipeline rather than fundamental architectural limitations. The narrow training focus on mapping visual information to textual space limits their ability to handle novel visual entities and fine-grained visual perception tasks.

Abstract: Vision Language Models (VLMs) achieve impressive performance across a wide range of multimodal tasks. However, on some tasks that demand fine-grained visual perception, they often fail even when the required information is present in their internal representations. In this work, we demonstrate that this gap arises from their narrow training pipeline which focuses on moving visual information to the textual space. Consequently, VLMs can only reason about visual entities that can be mapped to known concepts in the language space, leaving vision-focused tasks such as visual correspondence and reasoning about novel visual entities poorly supported. As a result, VLMs are severely limited in several important multimodal capabilities because they rely on brittle, hallucinated textual descriptions of visual entities that they cannot map to textual representations. We verify this behavior through visual correspondence tasks, in which VLMs must detect matching entities between two images. Testing across semantic, shape, and face correspondence tasks, we find that VLMs perform much better when the relevant entities are nameable in language than when they are unnameable. Mechanistically, our Logit Lens analyses confirm that VLMs explicitly assign semantic labels to nameable entities and surface more unique corresponding tokens compared to unnameable entities. Furthermore, we show that teaching completely arbitrary names for unknown entities improves performance, yet task-specific finetuning yields even stronger generalization without relying on language priors. Our findings suggest that current VLM failures on visual tasks reflect learned shortcuts from their training, rather than a fundamental limitation of multimodal architectures.

Method: Introduced Image Prompt Packaging (IPPg), a prompting paradigm that embeds structured text directly into images to reduce text token overhead. Benchmarked across five datasets, three frontier models (GPT-4.1, GPT-4o, Claude 3.5 Sonnet), and two task families (VQA and code generation). Derived a cost formulation decomposing savings by token type and conducted systematic error analysis with failure-mode taxonomy. Performed 125-configuration rendering ablation to study visual encoding choices.

Result: IPPg achieves 35.8-91.0% inference cost reductions with token compression up to 96%. Accuracy remains competitive in many settings but outcomes are highly model- and task-dependent: GPT-4.1 achieves simultaneous accuracy and cost gains on CoSQL, while Claude 3.5 incurs cost increases on several VQA benchmarks. Systematic error analysis reveals spatial reasoning, non-English inputs, and character-sensitive operations are most vulnerable, while schema-structured tasks benefit most. Rendering ablation shows accuracy shifts of 10-30 percentage points.

Conclusion: Visual prompting strategies like IPPg can significantly reduce inference costs for multimodal models while maintaining accuracy in many cases. Visual encoding choices should be considered a first-class variable in multimodal system design, as they substantially impact both cost and performance across different models and tasks.

Abstract: Deploying large multimodal language models at scale is constrained by token-based inference costs, yet the cost-performance behavior of visual prompting strategies remains poorly characterized. We introduce Image Prompt Packaging (IPPg), a prompting paradigm that embeds structured text directly into images to reduce text token overhead, and benchmark it across five datasets, three frontier models (GPT-4.1, GPT-4o, Claude 3.5 Sonnet), and two task families (VQA and code generation). We derive a cost formulation decomposing savings by token type and show IPPg achieves 35.8–91.0% inference cost reductions. Despite token compression of up to 96%, accuracy remains competitive in many settings, though outcomes are highly model- and task-dependent: GPT-4.1 achieves simultaneous accuracy and cost gains on CoSQL, while Claude 3.5 incurs cost increases on several VQA benchmarks. Systematic error analysis yields a failure-mode taxonomy: spatial reasoning, non-English inputs, and character-sensitive operations are most vulnerable, while schema-structured tasks benefit most. A 125-configuration rendering ablation reveals accuracy shifts of 10–30 percentage points, establishing visual encoding choices as a first-class variable in multimodal system design.

Method: Proposes Delaunay Canopy which uses Delaunay graph as geometric prior to define adaptive search space. Includes Delaunay Graph Scoring to reconstruct geometric manifold and provide region-wise curvature signatures, plus corner and wire selection modules that leverage Delaunay-induced prior to focus on probable elements.

Result: State-of-the-art wireframe reconstruction on Building3D Tallinn city and entry-level datasets, delivering accurate predictions across diverse and complex building geometries even in previously intractable regions.

Conclusion: The Delaunay Canopy approach successfully addresses limitations of conventional methods by establishing geometrically adaptive search spaces, enabling robust wireframe reconstruction from noisy and sparse LiDAR point clouds.

Abstract: Reconstructing building wireframe from airborne LiDAR point clouds yields a compact, topology-centric representation that enables structural understanding beyond dense meshes. Yet a key limitation persists: conventional methods have failed to achieve accurate wireframe reconstruction in regions afflicted by significant noise, sparsity, or internal corners. This failure stems from the inability to establish an adaptive search space to effectively leverage the rich 3D geometry of large, sparse building point clouds. In this work, we address this challenge with Delaunay Canopy, which utilizes the Delaunay graph as a geometric prior to define a geometrically adaptive search space. Central to our approach is Delaunay Graph Scoring, which not only reconstructs the underlying geometric manifold but also yields region-wise curvature signatures to robustly guide the reconstruction. Built on this foundation, our corner and wire selection modules leverage the Delaunay-induced prior to focus on highly probable elements, thereby shaping the search space and enabling accurate prediction even in previously intractable regions. Extensive experiments on the Building3D Tallinn city and entry-level datasets demonstrate state-of-the-art wireframe reconstruction, delivering accurate predictions across diverse and complex building geometries.

Method: End-to-end explainable vision-language model framework with: 1) Spatial Patch Cross-Attention module for text-directed localization of spinal anomalies with spatial precision, 2) Adaptive PID-Tversky Loss function integrating control theory principles to dynamically modify training penalties for difficult minority instances, 3) Incorporation of foundational VLMs with Automated Radiology Report Generation module.

Result: Achieved 90.69% diagnostic classification accuracy, macro-averaged Dice score of 0.9512 for segmentation, and CIDEr score of 92.80%. The framework shows explainability by converting segmentation predictions into radiologist-style clinical reports.

Conclusion: Establishes a new benchmark for transparent, interpretable AI in clinical medical imaging that maintains essential human supervision while enhancing diagnostic capabilities for Lumbar Spinal Stenosis.

Abstract: Lumbar Spinal Stenosis (LSS) diagnosis remains a critical clinical challenge, with diagnosis heavily dependent on labor-intensive manual interpretation of multi-view Magnetic Resonance Imaging (MRI), leading to substantial inter-observer variability and diagnostic delays. Existing vision-language models simultaneously fail to address the extreme class imbalance prevalent in clinical segmentation datasets while preserving spatial accuracy, primarily due to global pooling mechanisms that discard crucial anatomical hierarchies. We present an end-to-end Explainable Vision-Language Model framework designed to overcome these limitations, achieved through two principal objectives. We propose a Spatial Patch Cross-Attention module that enables precise, text-directed localization of spinal anomalies with spatial precision. A novel Adaptive PID-Tversky Loss function by integrating control theory principles dynamically further modifies training penalties to specifically address difficult, under-segmented minority instances. By incorporating foundational VLMs alongside an Automated Radiology Report Generation module, our framework demonstrates considerable performance: a diagnostic classification accuracy of 90.69%, a macro-averaged Dice score of 0.9512 for segmentation, and a CIDEr score of 92.80%. Furthermore, the framework shows explainability by converting complex segmentation predictions into radiologist-style clinical reports, thereby establishing a new benchmark for transparent, interpretable AI in clinical medical imaging that keeps essential human supervision while enhancing diagnostic capabilities.

Method: Two-stage framework: 1) Adapt a visual foundation model into a domain-specialized teacher using self-supervised learning on labeled synthetic and unlabeled real images, 2) Train an on-device student using teacher guidance and self-training.

Result: On a large-scale dataset with 2,000+ participants, DistillGaze reduces median gaze error by 58.62% relative to synthetic-only baselines while maintaining a lightweight 256K-parameter model suitable for real-time on-device deployment.

Conclusion: DistillGaze provides an efficient pathway for training and deploying eye tracking models that adapt to hardware changes, offering a recipe for combining synthetic supervision with unlabeled real data in on-device regression tasks.

Abstract: Eye tracking (ET) plays a critical role in augmented and virtual reality applications. However, rapidly deploying high-accuracy, on-device gaze estimation for new products remains challenging because hardware configurations (e.g., camera placement, camera pose, and illumination) often change across device generations. Visual foundation models (VFMs) are a promising direction for rapid training and deployment, and they excel on natural-image benchmarks; yet we find that off-the-shelf VFMs still struggle to achieve high accuracy on specialized near-eye infrared imagery. To address this gap, we introduce DistillGaze, a framework that distills a foundation model by leveraging labeled synthetic data and unlabeled real data for rapid and high-performance on-device gaze estimation. DistillGaze proceeds in two stages. First, we adapt a VFM into a domain-specialized teacher using self-supervised learning on labeled synthetic and unlabeled real images. Synthetic data provides scalable, high-quality gaze supervision, while unlabeled real data helps bridge the synthetic-to-real domain gap. Second, we train an on-device student using both teacher guidance and self-training. Evaluated on a large-scale, crowd-sourced dataset spanning over 2,000 participants, DistillGaze reduces median gaze error by 58.62% relative to synthetic-only baselines while maintaining a lightweight 256K-parameter model suitable for real-time on-device deployment. Overall, DistillGaze provides an efficient pathway for training and deploying ET models that adapt to hardware changes, and offers a recipe for combining synthetic supervision with unlabeled real data in on-device regression tasks.

Method: Introduces Feature Attribution Stability Suite (FASS) with prediction-invariance filtering and three complementary metrics: structural similarity, rank correlation, and top-k Jaccard overlap. Evaluates across geometric, photometric, and compression perturbations on four attribution methods across four architectures and three datasets.

Result: Geometric perturbations expose substantially greater attribution instability than photometric changes. Without prediction preservation conditioning, up to 99% of evaluated pairs involve changed predictions. Grad-CAM achieves highest stability across datasets under controlled evaluation.

Conclusion: Stability estimates depend critically on perturbation family and prediction-invariance filtering. Proper evaluation requires conditioning on prediction preservation to avoid conflating explanation fragility with model sensitivity.

Abstract: Post-hoc feature attribution methods are widely deployed in safety-critical vision systems, yet their stability under realistic input perturbations remains poorly characterized. Existing metrics evaluate explanations primarily under additive noise, collapse stability to a single scalar, and fail to condition on prediction preservation, conflating explanation fragility with model sensitivity. We introduce the Feature Attribution Stability Suite (FASS), a benchmark that enforces prediction-invariance filtering, decomposes stability into three complementary metrics: structural similarity, rank correlation, and top-k Jaccard overlap-and evaluates across geometric, photometric, and compression perturbations. Evaluating four attribution methods (Integrated Gradients, GradientSHAP, Grad-CAM, LIME) across four architectures and three datasets-ImageNet-1K, MS COCO, and CIFAR-10, FASS shows that stability estimates depend critically on perturbation family and prediction-invariance filtering. Geometric perturbations expose substantially greater attribution instability than photometric changes, and without conditioning on prediction preservation, up to 99% of evaluated pairs involve changed predictions. Under this controlled evaluation, we observe consistent method-level trends, with Grad-CAM achieving the highest stability across datasets.

Method: Comprehensive empirical study of confidence calibration in VLMs across three model families (Qwen3-VL, InternVL3, LLaVA-NeXT), three model scales (2B-38B), multiple confidence estimation prompting strategies, and three medical VQA benchmarks. Investigated post-hoc calibration (Platt scaling) and hallucination-aware calibration using vision-grounded hallucination detection signals.

Result: 1) Overconfidence persists across model families and is not resolved by scaling or prompting. 2) Simple post-hoc calibration reduces calibration error and outperforms prompt-based strategies. 3) Post-hoc methods don’t improve AUROC due to monotonicity limitations. 4) Hallucination-aware calibration improves both calibration and AUROC, with largest gains on open-ended questions.

Conclusion: Post-hoc calibration should be standard practice for medical VLM deployment over raw confidence estimates. Hallucination signals enable more reliable use of VLMs in medical VQA by improving both calibration and discriminative quality.

Abstract: As vision-language models (VLMs) are increasingly deployed in clinical decision support, more than accuracy is required: knowing when to trust their predictions is equally critical. Yet, a comprehensive and systematic investigation into the overconfidence of these models remains notably scarce in the medical domain. We address this gap through a comprehensive empirical study of confidence calibration in VLMs, spanning three model families (Qwen3-VL, InternVL3, LLaVA-NeXT), three model scales (2B–38B), and multiple confidence estimation prompting strategies, across three medical visual question answering (VQA) benchmarks. Our study yields three key findings: First, overconfidence persists across model families and is not resolved by scaling or prompting, such as chain-of-thought and verbalized confidence variants. Second, simple post-hoc calibration approaches, such as Platt scaling, reduce calibration error and consistently outperform the prompt-based strategy. Third, due to their (strict) monotonicity, these post-hoc calibration methods are inherently limited in improving the discriminative quality of predictions, leaving AUROC at the same level. Motivated by these findings, we investigate hallucination-aware calibration (HAC), which incorporates vision-grounded hallucination detection signals as complementary inputs to refine confidence estimates. We find that leveraging these hallucination signals improves both calibration and AUROC, with the largest gains on open-ended questions. Overall, our findings suggest post-hoc calibration as standard practice for medical VLM deployment over raw confidence estimates, and highlight the practical usefulness of hallucination signals to enable more reliable use of VLMs in medical VQA.

Method: Proposes UniScene3D, a transformer-based encoder that processes multi-view colored pointmaps. Introduces cross-view geometric alignment and grounded view alignment to enforce geometry and semantic consistency across views for robust representation learning.

Result: Achieves state-of-the-art performance in low-shot and task-specific fine-tuning evaluations on viewpoint grounding, scene retrieval, scene type classification, and 3D VQA tasks.

Conclusion: The approach effectively learns unified 3D scene representations that generalize well across multiple 3D understanding tasks, demonstrating the value of jointly modeling appearance and geometry with cross-view consistency constraints.

Abstract: Pretraining 3D encoders by aligning with Contrastive Language Image Pretraining (CLIP) has emerged as a promising direction to learn generalizable representations for 3D scene understanding. In this paper, we propose UniScene3D, a transformer-based encoder that learns unified scene representations from multi-view colored pointmaps, jointly modeling image appearance and geometry. For robust colored pointmap representation learning, we introduce novel cross-view geometric alignment and grounded view alignment to enforce cross-view geometry and semantic consistency. Extensive low-shot and task-specific fine-tuning evaluations on viewpoint grounding, scene retrieval, scene type classification, and 3D VQA demonstrate our state-of-the-art performance. These results highlight the effectiveness of our approach for unified 3D scene understanding. https://yebulabula.github.io/UniScene3D/

Method: Proposes Weighted SVD (WSVD) with two key innovations: 1) applying SVD at finer granularity for better computational efficiency, and 2) adaptively allocating importance to weight elements during SVD to preserve accuracy. Extends this with quantization of both weights and activations.

Result: WSVD achieves over 1.8× decoding speedup compared to other approaches while maintaining model accuracy. The method demonstrates real and measurable improvements in execution latency for VLMs.

Conclusion: Weighted SVD with finer granularity and adaptive importance allocation provides an effective approach for compressing Vision Language Models, enabling significant speed improvements without sacrificing accuracy.

Abstract: Singular Value Decomposition (SVD) has become an important technique for reducing the computational burden of Vision Language Models (VLMs), which play a central role in tasks such as image captioning and visual question answering. Although multiple prior works have proposed efficient SVD variants to enable low-rank operations, we find that in practice it remains difficult to achieve substantial latency reduction during model execution. To address this limitation, we introduce a new computational pattern and apply SVD at a finer granularity, enabling real and measurable improvements in execution latency. Furthermore, recognizing that weight elements differ in their relative importance, we adaptively allocate relative importance to each element during SVD process to better preserve accuracy, then extend this framework with quantization applied to both weights and activations, resulting in a highly efficient VLM. Collectively, we introduce~\textit{Weighted SVD} (WSVD), which outperforms other approaches by achieving over $1.8\times$ decoding speedup while preserving accuracy. We open source our code at: \href{https://github.com/SAI-Lab-NYU/WSVD}{\texttt{https://github.com/SAI-Lab-NYU/WSVD}

Method: Proposes FusionBERT with two key components: 1) Cross-attention-based multi-view visual aggregator that adaptively integrates features from multi-view images, fusing inter-view complementary relationships and selectively emphasizing informative visual cues. 2) Normal-aware 3D model encoder that jointly encodes point normals and 3D positions to enhance geometric features for textureless or color-degraded 3D models.

Result: Extensive image-3D retrieval experiments show FusionBERT achieves significantly higher retrieval accuracy than state-of-the-art multimodal large models under both single-view and multi-view settings, establishing a strong baseline for multi-view multimodal retrieval.

Conclusion: FusionBERT successfully addresses the multi-view fusion challenge in image-3D retrieval, demonstrating the importance of effectively integrating multi-view visual information and enhancing 3D geometric representations for better cross-modal matching.

Abstract: We propose FusionBERT, a novel multi-view visual fusion framework for image-3D multimodal retrieval. Existing image-3D representation learning methods predominantly focus on feature alignment of a single object image and its 3D model, limiting their applicability in realistic scenarios where an object is typically observed and captured from multiple viewpoints. Although multi-view observations naturally provide complementary geometric and appearance cues, existing multimodal large models rarely explore how to effectively fuse such multi-view visual information for better cross-modal retrieval. To address this limitation, we introduce a multi-view image-3D retrieval framework named FusionBERT, which innovatively utilizes a cross-attention-based multi-view visual aggregator to adaptively integrate features from multi-view images of an object. The proposed multi-view visual encoder fuses inter-view complementary relationships and selectively emphasizes informative visual cues across multiple views to get a more robustly fused visual feature for better 3D model matching. Furthermore, FusionBERT proposes a normal-aware 3D model encoder that can further enhance the 3D geometric feature of an object model by jointly encoding point normals and 3D positions, enabling a more robust representation learning for textureless or color-degraded 3D models. Extensive image-3D retrieval experiments demonstrate that FusionBERT achieves significantly higher retrieval accuracy than SOTA multimodal large models under both single-view and multi-view settings, establishing a strong baseline for multi-view multimodal retrieval.

Method: TrackerSplat uses off-the-shelf point tracking models to extract pixel trajectories, triangulates these onto 3D Gaussians, and uses them to guide Gaussian relocation, rotation, and scaling before training. This pre-optimization positioning helps handle large displacements between frames.

Result: Experiments show TrackerSplat reduces fading and recoloring artifacts, handles challenging scenarios with significant displacements, achieves superior throughput under parallel settings, and maintains visual quality compared to baselines.

Conclusion: Integrating point tracking with 3DGS enables robust dynamic scene reconstruction that handles large displacements while preserving rendering quality and enabling efficient parallel processing.

Abstract: Recent advancements in 3D Gaussian Splatting (3DGS) have demonstrated its potential for efficient and photorealistic 3D reconstructions, which is crucial for diverse applications such as robotics and immersive media. However, current Gaussian-based methods for dynamic scene reconstruction struggle with large inter-frame displacements, leading to artifacts and temporal inconsistencies under fast object motions. To address this, we introduce \textit{TrackerSplat}, a novel method that integrates advanced point tracking methods to enhance the robustness and scalability of 3DGS for dynamic scene reconstruction. TrackerSplat utilizes off-the-shelf point tracking models to extract pixel trajectories and triangulate per-view pixel trajectories onto 3D Gaussians to guide the relocation, rotation, and scaling of Gaussians before training. This strategy effectively handles large displacements between frames, dramatically reducing the fading and recoloring artifacts prevalent in prior methods. By accurately positioning Gaussians prior to gradient-based optimization, TrackerSplat overcomes the quality degradation associated with large frame gaps when processing multiple adjacent frames in parallel across multiple devices, thereby boosting reconstruction throughput while preserving rendering quality. Experiments on real-world datasets confirm the robustness of TrackerSplat in challenging scenarios with significant displacements, achieving superior throughput under parallel settings and maintaining visual quality compared to baselines. The code is available at https://github.com/yindaheng98/TrackerSplat.

Method: Extends Moondream 3 vision-language model with autoregressive vector path decoding that iteratively refines rasterized masks. Introduces reinforcement learning stage to optimize mask quality directly, resolving supervision ambiguity. Uses rollouts from RL to create coarse-to-ground-truth targets for the refiner. Also releases RefCOCO-M, a cleaned validation split with boundary-accurate masks.

Result: Achieves 80.2% cIoU on RefCOCO (val) and 62.6% mIoU on LVIS (val), demonstrating state-of-the-art performance in referring image segmentation tasks.

Conclusion: Moondream Segmentation effectively combines vision-language modeling with iterative refinement and RL optimization for high-quality referring image segmentation, while addressing annotation noise issues through improved evaluation benchmarks.

Abstract: We present Moondream Segmentation, a referring image segmentation extension of Moondream 3, a vision-language model. Given an image and a referring expression, the model autoregressively decodes a vector path and iteratively refines the rasterized mask into a final detailed mask. We introduce a reinforcement learning stage that resolves ambiguity in the supervised signal by directly optimizing mask quality. Rollouts from this stage produce coarse-to-ground-truth targets for the refiner. To mitigate evaluation noise from polygon annotations, we release RefCOCO-M, a cleaned RefCOCO validation split with boundary-accurate masks. Moondream Segmentation achieves a cIoU of 80.2% on RefCOCO (val) and 62.6% mIoU on LVIS (val).

Method: Integrated sensing and communication (ISAC) framework using ubiquitous mmWave OFDM communication signals. Multi-frame, spatially adaptive fusion with confidence-weighted forward projection to overcome sparse and multipath-ambiguous nature of individual radio frames.

Result: Method reconstructs 3D scenes with high precision, demonstrating a new pathway toward low-cost, scalable, and robust 3D perception.

Conclusion: Rascene enables robust 3D environmental perception using communication signals, offering advantages over optical sensors and specialized radar systems in terms of cost, scalability, and adverse condition performance.

Abstract: Robust 3D environmental perception is critical for applications such as autonomous driving and robot navigation. However, optical sensors such as cameras and LiDAR often fail under adverse conditions, including smoke, fog, and non-ideal lighting. Although specialized radar systems can operate in these environments, their reliance on bespoke hardware and licensed spectrum limits scalability and cost-effectiveness. This paper introduces Rascene, an integrated sensing and communication (ISAC) framework that leverages ubiquitous mmWave OFDM communication signals for 3D scene imaging. To overcome the sparse and multipath-ambiguous nature of individual radio frames, Rascene performs multi-frame, spatially adaptive fusion with confidence-weighted forward projection, enabling the recovery of geometric consensus across arbitrary poses. Experimental results demonstrate that our method reconstructs 3D scenes with high precision, offering a new pathway toward low-cost, scalable, and robust 3D perception.

Method: Two-layer privacy protection: 1) Human skeletal abstraction to remove identifiable visual information from RGB videos, 2) Federated Learning to keep sensitive pose data within clinics while enabling collaborative model training across sites.

Result: Achieves high recognition accuracy on MMASD benchmark, outperforming traditional federated baselines, providing a robust privacy-first solution for multi-site clinical analysis.

Conclusion: Federated Learning with pose-based abstraction offers an effective privacy-preserving approach for autism behavior recognition, enabling multi-site collaboration while maintaining strict data residency and addressing data scarcity issues.

Abstract: Automated recognition of autistic behaviors in children is essential for early intervention and objective clinical assessment. However, the development of robust models is severely hindered by strict privacy regulations (e.g., HIPAA) and the sensitive nature of pediatric data, which prevents the centralized aggregation of clinical datasets. Furthermore, individual clinical sites often suffer from data scarcity, making it difficult to learn generalized behavior patterns or tailor models to site-specific patient distributions. To address these challenges, we observe that Federated Learning (FL) can decouple model training from raw data access, enabling multi-site collaboration while maintaining strict data residency. In this paper, we present the first study exploring Federated Learning for pose-based child autism behavior recognition. Our framework employs a two-layer privacy protection mechanism: utilizing human skeletal abstraction to remove identifiable visual information from the raw RGB videos and FL to ensure sensitive pose data remains within the clinic. This approach leverages distributed clinical data to learn generalized representations while providing the flexibility for site-specific personalization. Experimental results on the MMASD benchmark demonstrate that our framework achieves high recognition accuracy, outperforming traditional federated baselines and providing a robust, privacy-first solution for multi-site clinical analysis.

Method: Proposes ArticuSurDepth with multi-view spatial context enrichment, cross-view surface normal constraints, camera height regularization with ground plane-awareness for metric depth, and cross-vehicle pose consistency to bridge motion estimation between articulated segments.

Result: Achieves state-of-the-art performance on their self-collected articulated vehicle dataset as well as on DDAD, nuScenes, and KITTI benchmarks.

Conclusion: The framework effectively addresses the unique challenges of articulated vehicles for surround depth estimation through geometric consistency and structural priors, demonstrating superior performance across multiple benchmarks.

Abstract: Surround depth estimation provides a cost-effective alternative to LiDAR for 3D perception in autonomous driving. While recent self-supervised methods explore multi-camera settings to improve scale awareness and scene coverage, they are primarily designed for passenger vehicles and rarely consider articulated vehicles or robotics platforms. The articulated structure introduces complex cross-segment geometry and motion coupling, making consistent depth reasoning across views more challenging. In this work, we propose \textbf{ArticuSurDepth}, a self-supervised framework for surround-view depth estimation on articulated vehicles that enhances depth learning through cross-view and cross-vehicle geometric consistency guided by structural priors from vision foundation model. Specifically, we introduce multi-view spatial context enrichment strategy and a cross-view surface normal constraint to improve structural coherence across spatial and temporal contexts. We further incorporate camera height regularization with ground plane-awareness to encourage metric depth estimation, together with cross-vehicle pose consistency that bridges motion estimation between articulated segments. To validate our proposed method, an articulated vehicle experiment platform was established with a dataset collected over it. Experiment results demonstrate state-of-the-art (SoTA) performance of depth estimation on our self-collected dataset as well as on DDAD, nuScenes, and KITTI benchmarks.

Method: Two-component framework: 1) Temporal Reliability Calibrator (TRC) learns per-frame reliability scores to filter noise while anchoring to ground-truth template; 2) Temporal Guidance Synthesizer (TGS) synthesizes calibrated history into dynamic temporal priors for predictive guidance.

Result: Integration into three tracking architectures (OSTrack, ODTrack, LoRAT) shows consistent performance gains. Best model achieves 77.5% Success on LaSOT and 80.3% AO on GOT-10k, setting new state-of-the-art.

Conclusion: DTPTrack effectively suppresses drift in visual trackers through temporal reliability calibration and guidance synthesis, demonstrating strong generalization across diverse architectures and benchmarks.

Abstract: Temporal information is crucial for visual tracking, but existing multi-frame trackers are vulnerable to model drift caused by naively aggregating noisy historical predictions. In this paper, we introduce DTPTrack, a lightweight and generalizable module designed to be seamlessly integrated into existing trackers to suppress drift. Our framework consists of two core components: (1) a Temporal Reliability Calibrator (TRC) mechanism that learns to assign a per-frame reliability score to historical states, filtering out noise while anchoring on the ground-truth template; and (2) a Temporal Guidance Synthesizer (TGS) module that synthesizes this calibrated history into a compact set of dynamic temporal priors to provide predictive guidance. To demonstrate its versatility, we integrate DTPTrack into three diverse tracking architectures–OSTrack, ODTrack, and LoRAT-and show consistent, significant performance gains across all baselines. Our best-performing model, built upon an extended LoRATv2 backbone, sets a new state-of-the-art on several benchmarks, achieving a 77.5% Success rate on LaSOT and an 80.3% AO on GOT-10k.

Method: Proposes Efficient3D with two key components: 1) Debiased Visual Token Importance Estimator (DVTIE) that considers shallow initial layers during attention aggregation for reliable importance predictions, and 2) Adaptive Token Rebalancing (ATR) strategy that dynamically adjusts pruning strength based on scene complexity to preserve semantic completeness.

Result: Achieves superior performance on five 3D vision-language benchmarks (ScanRefer, Multi3DRefer, Scan2Cap, ScanQA, SQA3D), with +2.57% CIDEr improvement on Scan2Cap compared to unpruned baselines.

Conclusion: Efficient3D provides a scalable and effective solution for efficient inference in 3D MLLMs, enabling practical deployment on resource-constrained platforms while maintaining competitive accuracy.

Abstract: Recent advances in Multimodal Large Language Models (MLLMs) have expanded reasoning capabilities into 3D domains, enabling fine-grained spatial understanding. However, the substantial size of 3D MLLMs and the high dimensionality of input features introduce considerable inference overhead, which limits practical deployment on resource constrained platforms. To overcome this limitation, this paper presents Efficient3D, a unified framework for visual token pruning that accelerates 3D MLLMs while maintaining competitive accuracy. The proposed framework introduces a Debiased Visual Token Importance Estimator (DVTIE) module, which considers the influence of shallow initial layers during attention aggregation, thereby producing more reliable importance predictions for visual tokens. In addition, an Adaptive Token Rebalancing (ATR) strategy is developed to dynamically adjust pruning strength based on scene complexity, preserving semantic completeness and maintaining balanced attention across layers. Together, they enable context-aware token reduction that maintains essential semantics with lower computation. Comprehensive experiments conducted on five representative 3D vision and language benchmarks, including ScanRefer, Multi3DRefer, Scan2Cap, ScanQA, and SQA3D, demonstrate that Efficient3D achieves superior performance compared with unpruned baselines, with a +2.57% CIDEr improvement on the Scan2Cap dataset. Therefore, Efficient3D provides a scalable and effective solution for efficient inference in 3D MLLMs. The code is released at: https://github.com/sol924/Efficient3D

Method: Introduces a structural refinement stage between DETR-style detectors and parsers that treats raw detector outputs as a hypothesis pool. Performs set-level reasoning using query features, semantic cues, box geometry, and visual evidence to jointly determine instance retention, refine box localization, and predict parser input order.

Result: Extensive experiments show consistent improvement in page-level layout quality. When integrated into end-to-end parsing pipelines, reduces sequence mismatch significantly, achieving a Reading Order Edit of 0.024 on OmniDocBench.

Conclusion: The proposed lightweight structural refinement module effectively stabilizes parser interfaces in document layout analysis, addressing critical inconsistencies between detector outputs and parser inputs, particularly for structurally complex documents.

Abstract: Accurate document parsing requires both robust content recognition and a stable parser interface. In explicit Document Layout Analysis (DLA) pipelines, downstream parsers do not consume the full detector output. Instead, they operate on a retained and serialized set of layout instances. However, on dense pages with overlapping regions and ambiguous boundaries, unstable layout hypotheses can make the retained instance set inconsistent with its parser input order, leading to severe downstream parsing errors. To address this issue, we introduce a lightweight structural refinement stage between a DETR-style detector and the parser to stabilize the parser interface. Treating raw detector outputs as a compact hypothesis pool, the proposed module performs set-level reasoning over query features, semantic cues, box geometry, and visual evidence. From a shared refined structural state, it jointly determines instance retention, refines box localization, and predicts parser input order before handoff. We further introduce retention-oriented supervision and a difficulty-aware ordering objective to better align the retained instance set and its order with the final parser input, especially on structurally complex pages. Extensive experiments on public benchmarks show that our method consistently improves page-level layout quality. When integrated into a standard end-to-end parsing pipeline, the stabilized parser interface also substantially reduces sequence mismatch, achieving a Reading Order Edit of 0.024 on OmniDocBench.

Method: Proposes DocShield framework formulating text-centric forgery analysis as visual-logical co-reasoning problem. Core innovation is Cross-Cues-aware Chain of Thought (CCT) mechanism enabling implicit agentic reasoning that iteratively cross-validates visual anomalies with textual semantics. Uses Weighted Multi-Task Reward for GRPO-based optimization to align reasoning structure, spatial evidence, and authenticity prediction. Also constructs RealText-V1 multilingual dataset with pixel-level manipulation masks and expert textual explanations.

Result: DocShield significantly outperforms existing methods, improving macro-average F1 by 41.4% over specialized frameworks and 23.4% over GPT-4o on T-IC13 benchmark, with consistent gains on challenging T-SROIE benchmark.

Conclusion: DocShield provides a unified, evidence-grounded approach to text-centric forgery analysis through visual-logical co-reasoning, addressing limitations of existing methods and demonstrating superior performance on document forgery detection tasks.

Abstract: The rapid progress of generative AI has enabled increasingly realistic text-centric image forgeries, posing major challenges to document safety. Existing forensic methods mainly rely on visual cues and lack evidence-based reasoning to reveal subtle text manipulations. Detection, localization, and explanation are often treated as isolated tasks, limiting reliability and interpretability. To tackle these challenges, we propose DocShield, the first unified framework formulating text-centric forgery analysis as a visual-logical co-reasoning problem. At its core, a novel Cross-Cues-aware Chain of Thought (CCT) mechanism enables implicit agentic reasoning, iteratively cross-validating visual anomalies with textual semantics to produce consistent, evidence-grounded forensic analysis. We further introduce a Weighted Multi-Task Reward for GRPO-based optimization, aligning reasoning structure, spatial evidence, and authenticity prediction. Complementing the framework, we construct RealText-V1, a multilingual dataset of document-like text images with pixel-level manipulation masks and expert-level textual explanations. Extensive experiments show DocShield significantly outperforms existing methods, improving macro-average F1 by 41.4% over specialized frameworks and 23.4% over GPT-4o on T-IC13, with consistent gains on the challenging T-SROIE benchmark. Our dataset, model, and code will be publicly released.

Method: XrayClaw uses a cooperative-competitive architecture with four specialized cooperative agents simulating clinical workflow and a competitive agent as independent auditor. Introduces Competitive Preference Optimization to penalize illogical reasoning through mutual verification between analytical and holistic interpretations.

Result: Achieves state-of-the-art performance on MS-CXR-T, MIMIC-CXR, and CheXbench benchmarks in diagnostic accuracy, clinical reasoning fidelity, and zero-shot domain generalization. Effectively mitigates cumulative hallucinations and enhances reliability.

Conclusion: XrayClaw establishes a new paradigm for trustworthy medical imaging analysis by operationalizing multi-agent alignment through cooperative-competitive architecture, improving automated CXR diagnosis reliability.

Abstract: Chest X-ray (CXR) interpretation is a fundamental yet complex clinical task that increasingly relies on artificial intelligence for automation. However, traditional monolithic models often lack the nuanced reasoning required for trustworthy diagnosis, frequently leading to logical inconsistencies and diagnostic hallucinations. While multi-agent systems offer a potential solution by simulating collaborative consultations, existing frameworks remain susceptible to consensus-based errors when instantiated by a single underlying model. This paper introduces XrayClaw, a novel framework that operationalizes multi-agent alignment through a sophisticated cooperative-competitive architecture. XrayClaw integrates four specialized cooperative agents to simulate a systematic clinical workflow, alongside a competitive agent that serves as an independent auditor. To reconcile these distinct diagnostic pathways, we propose Competitive Preference Optimization, a learning objective that penalizes illogical reasoning by enforcing mutual verification between analytical and holistic interpretations. Extensive empirical evaluations on the MS-CXR-T, MIMIC-CXR, and CheXbench benchmarks demonstrate that XrayClaw achieves state-of-the-art performance in diagnostic accuracy, clinical reasoning fidelity, and zero-shot domain generalization. Our results indicate that XrayClaw effectively mitigates cumulative hallucinations and enhances the overall reliability of automated CXR diagnosis, establishing a new paradigm for trustworthy medical imaging analysis.

Method: Proposes a probabilistic framework coupling splat map refinement and camera pose tracking using variational inference with conjugate properties of multivariate Gaussians, enabling efficient closed-form updates and explicit posterior uncertainty maintenance.

Result: Demonstrates superior tracking performance and robustness in long sequence prediction, with efficient, high-quality novel view synthesis across diverse synthetic and real-world scenes.

Conclusion: VBGS-SLAM provides an uncertainty-aware approach that mitigates drift and enhances robustness while preserving the efficiency and rendering quality of 3DGS, advancing probabilistic 3D scene reconstruction.

Abstract: 3D Gaussian Splatting (3DGS) has shown promising results for 3D scene modeling using mixtures of Gaussians, yet its existing simultaneous localization and mapping (SLAM) variants typically rely on direct, deterministic pose optimization against the splat map, making them sensitive to initialization and susceptible to catastrophic forgetting as map evolves. We propose Variational Bayesian Gaussian Splatting SLAM (VBGS-SLAM), a novel framework that couples the splat map refinement and camera pose tracking in a generative probabilistic form. By leveraging conjugate properties of multivariate Gaussians and variational inference, our method admits efficient closed-form updates and explicitly maintains posterior uncertainty over both poses and scene parameters. This uncertainty-aware method mitigates drift and enhances robustness in challenging conditions, while preserving the efficiency and rendering quality of existing 3DGS. Our experiments demonstrate superior tracking performance and robustness in long sequence prediction, alongside efficient, high-quality novel view synthesis across diverse synthetic and real-world scenes.

Method: Proposes unified understanding-and-generation framework with world modeling: augments trajectory prediction with future RGB and depth image generation as dense objectives. Uses world model’s image prediction uncertainty as intrinsic reward for exploration, incorporated into safety-gated reward optimized via Group Relative Policy Optimization (GRPO).

Result: Achieves state-of-the-art PDMS score of 93.7 and EPDMS of 88.8 on NAVSIM benchmark, demonstrating effectiveness on nuScenes benchmark as well.

Conclusion: World modeling enables meaningful exploration and dense supervision for VLA-based autonomous driving, improving performance in novel scenarios through uncertainty-based intrinsic rewards.

Abstract: End-to-end autonomous driving models based on Vision-Language-Action (VLA) architectures have shown promising results by learning driving policies through behavior cloning on expert demonstrations. However, imitation learning inherently limits the model to replicating observed behaviors without exploring diverse driving strategies, leaving it brittle in novel or out-of-distribution scenarios. Reinforcement learning (RL) offers a natural remedy by enabling policy exploration beyond the expert distribution. Yet VLA models, typically trained on offline datasets, lack directly observable state transitions, necessitating a learned world model to anticipate action consequences. In this work, we propose a unified understanding-and-generation framework that leverages world modeling to simultaneously enable meaningful exploration and provide dense supervision. Specifically, we augment trajectory prediction with future RGB and depth image generation as dense world modeling objectives, requiring the model to learn fine-grained visual and geometric representations that substantially enrich the planning backbone. Beyond serving as a supervisory signal, the world model further acts as a source of intrinsic reward for policy exploration: its image prediction uncertainty naturally measures a trajectory’s novelty relative to the training distribution, where high uncertainty indicates out-of-distribution scenarios that, if safe, represent valuable learning opportunities. We incorporate this exploration signal into a safety-gated reward and optimize the policy via Group Relative Policy Optimization (GRPO). Experiments on the NAVSIM and nuScenes benchmarks demonstrate the effectiveness of our approach, achieving a state-of-the-art PDMS score of 93.7 and an EPDMS of 88.8 on NAVSIM. The code and demo will be publicly available at https://zihaosheng.github.io/ExploreVLA/.

Method: Uses model merging to integrate representations learned independently from three Martian sensors. Introduces Equal Validation Loss (EVL) strategy to align checkpoints across sensors based on validation loss similarity before fusion via task arithmetic, ensuring models are merged at compatible convergence stages.

Result: MOMO achieves better overall performance compared to ImageNet pre-trained, earth observation foundation model, sensor-specific pre-training, and fully-supervised baselines on 9 downstream tasks from Mars-Bench. Shows consistent and significant performance improvement on segmentation tasks.

Conclusion: Model merging through optimal checkpoint selection strategy provides an effective approach for building foundation models for multi-resolution data. The method demonstrates improved stability and generalization for Mars remote sensing applications.

Abstract: We introduce MOMO, the first multi-sensor foundation model for Mars remote sensing. MOMO uses model merge to integrate representations learned independently from three key Martian sensors (HiRISE, CTX, and THEMIS), spanning resolutions from 0.25 m/pixel to 100 m/pixel. Central to our method is our novel Equal Validation Loss (EVL) strategy, which aligns checkpoints across sensors based on validation loss similarity before fusion via task arithmetic. This ensures models are merged at compatible convergence stages, leading to improved stability and generalization. We train MOMO on a large-scale, high-quality corpus of $\sim 12$ million samples curated from Mars orbital data and evaluate it on 9 downstream tasks from Mars-Bench. MOMO achieves better overall performance compared to ImageNet pre-trained, earth observation foundation model, sensor-specific pre-training, and fully-supervised baselines. Particularly on segmentation tasks, MOMO shows consistent and significant performance improvement. Our results demonstrate that model merging through an optimal checkpoint selection strategy provides an effective approach for building foundation models for multi-resolution data. The model weights, pretraining code, pretraining data, and evaluation code are available at: https://github.com/kerner-lab/MOMO.

Method: Two-stage pipeline: 1) Generate hand and object Gaussians, 2) Physics-based HOI optimization with novel mesh extraction method and vertex-to-Gaussian mapping for topology-aware regularization, plus VLM-guided translation refinement and contact-aware optimization.

Result: THOM consistently surpasses state-of-the-art methods in text alignment, visual realism, and interaction plausibility.

Conclusion: THOM provides an effective training-free solution for generating physically plausible 3D hand-object interactions from text with high visual fidelity.

Abstract: The generation of 3D hand-object interactions (HOIs) from text is crucial for dexterous robotic grasping and VR/AR content generation, requiring both high visual fidelity and physical plausibility. Nevertheless, the ill-posed problem of mesh extraction from text-generated Gaussians, and physics-based optimization on the erroneous meshes pose challenges. To address these issues, we introduce THOM, a training-free framework that generates photorealistic, physically plausible 3D HOI meshes without the need for a template object mesh. THOM employs a two-stage pipeline, initially generating the hand and object Gaussians, followed by physics-based HOI optimization. Our new mesh extraction method and vertex-to-Gaussian mapping explicitly assign Gaussian elements to mesh vertices, allowing topology-aware regularization. Furthermore, we improve the physical plausibility of interactions by VLM-guided translation refinement and contact-aware optimization. Comprehensive experiments demonstrate that THOM consistently surpasses state-of-the-art methods in terms of text alignment, visual realism, and interaction plausibility.

Method: LLaBIT extends visual reasoning of LLMs to brain MRI tasks. It incorporates feature map reuse from image encoder to mitigate spatial information loss from image tokenization. Uses LLM-generated text data with strict instructions to augment limited image-text paired brain MRI data.

Result: Evaluated on five brain MRI datasets across four tasks: report generation, visual question answering, image segmentation, and image translation. Model demonstrated superior performance across all tasks and outperformed specialized task-specific models in direct comparisons.

Conclusion: LLaBIT shows efficacy and versatility as a single multimodal LLM capable of handling diverse clinically relevant brain MRI tasks, advancing beyond simple text-to-image generation to more meaningful medical applications.

Abstract: LLMs have demonstrated remarkable capabilities in linguistic reasoning and are increasingly adept at vision-language tasks. The integration of image tokens into transformers has enabled direct visual input and output, advancing research from image-to-text descriptions to text-to-image generation. However, simple text-to-image generation holds limited clinical utility. In medical imaging, tasks such as image segmentation for localizing pathologies or image translation for reconstructing missing sequences have much greater clinical importance. Despite this, integrating these diverse, clinically relevant tasks within a single, versatile language model remains unexplored. Our method, LLaBIT (Large Language Model for Brain Image Translation), extends the visual reasoning of LLMs to these clinically meaningful tasks in the brain MRI domain. To mitigate the spatial information loss inherent in image tokenization, we incorporate a mechanism to reuse feature maps from the image encoder, minimizing data degradation. We also generate text data using LLMs with strict predefined instructions to augment limited image-text paired data in brain MRI. We comprehensively evaluated our method on five brain MRI datasets across four distinct tasks: report generation, visual question answering, image segmentation, and image translation. Our model not only demonstrated superior performance across all tasks but also outperformed specialized, task-specific models in direct comparisons, highlighting its efficacy and versatility

Method: Proposes a dual representation that couples discrete polylines with continuous Bézier control points via bidirectional mapping. This enables collaborative optimization where local gradients refine global stroke structures while content-aware stroke proposals help escape poor local optima. Uses Gaussian-splatting-inspired initialization for highly parallel stroke optimization across images.

Result: The approach reduces number of strokes by 30-50%, achieves more structurally coherent layouts, improves reconstruction quality, and cuts optimization time by 30-40% compared to existing differentiable vectorization methods.

Conclusion: The dual representation effectively bridges discrete and continuous optimization in stroke-based rendering, enabling collaborative optimization that improves both structural coherence and computational efficiency.

Abstract: In stroke-based rendering, search methods often get trapped in local minima due to discrete stroke placement, while differentiable optimizers lack structural awareness and produce unstructured layouts. To bridge this gap, we propose a dual representation that couples discrete polylines with continuous Bézier control points via a bidirectional mapping mechanism. This enables collaborative optimization: local gradients refine global stroke structures, while content-aware stroke proposals help escape poor local optima. Our representation further supports Gaussian-splatting-inspired initialization, enabling highly parallel stroke optimization across the image. Experiments show that our approach reduces the number of strokes by 30-50%, achieves more structurally coherent layouts, and improves reconstruction quality, while cutting optimization time by 30-40% compared to existing differentiable vectorization methods.

Method: Proposes Decoupled Cognition DETR with a unified decoupling paradigm: 1) constructs hierarchical semantic prototype space from region-level descriptions generated by pre-trained LVLMs and aligned via CLIP for efficient reusable semantic representation, 2) disentangles semantic reasoning from localization through decoupled training strategy separating alignment and detection into parallel optimization streams.

Result: Extensive experiments on standard OVOD benchmarks demonstrate competitive zero-shot detection performance while significantly improving inference efficiency.

Conclusion: Decoupling semantic cognition from detection offers a practical direction for scalable OVOD systems, highlighting the effectiveness of this approach.

Abstract: Open-vocabulary Object Detection (OVOD) enables models to recognize objects beyond predefined categories, but existing approaches remain limited in practical deployment. On the one hand, multimodal designs often incur substantial computational overhead due to their reliance on text encoders at inference time. On the other hand, tightly coupled training objectives introduce a trade-off between closed-set detection accuracy and open-world generalization. Thus, we propose Decoupled Cognition DETR (DeCo-DETR), a vision-centric framework that addresses these challenges through a unified decoupling paradigm. Instead of depending on online text encoding, DeCo-DETR constructs a hierarchical semantic prototype space from region-level descriptions generated by pre-trained LVLMs and aligned via CLIP, enabling efficient and reusable semantic representation. Building upon this representation, the framework further disentangles semantic reasoning from localization through a decoupled training strategy, which separates alignment and detection into parallel optimization streams. Extensive experiments on standard OVOD benchmarks demonstrate that DeCo-DETR achieves competitive zero-shot detection performance while significantly improving inference efficiency. These results highlight the effectiveness of decoupling semantic cognition from detection, offering a practical direction for scalable OVOD systems.

Method: Two-stage framework: Stage I uses geometry-driven autoregressive model to infer BoxMesh (structured 3D representation) from input garment; Stage II uses semantics-aware autoregressive model to parse BoxMesh into parametric sewing patterns.

Result: Achieves state-of-the-art performance on GarmentCodeData benchmark and generalizes effectively to real-world scans and single-view images.

Conclusion: The BoxMesh representation bridges 3D geometry and parametric patterns, with autoregressive modeling handling variable-length panel configurations, leading to accurate and robust sewing pattern recovery.

Abstract: Recovering sewing patterns from draped 3D garments is a challenging problem in human digitization research. In contrast to the well-studied forward process of draping designed sewing patterns using mature physical simulation engines, the inverse process of recovering parametric 2D patterns from deformed garment geometry remains fundamentally ill-posed for existing methods. We propose a two-stage framework that centers on a structured intermediate representation, BoxMesh, which serves as the key to bridging the gap between 3D garment geometry and parametric sewing patterns. BoxMesh encodes both garment-level geometry and panel-level structure in 3D, while explicitly disentangling intrinsic panel geometry and stitching topology from draping-induced deformations. This representation imposes a physically grounded structure on the problem, significantly reducing ambiguity. In Stage I, a geometry-driven autoregressive model infers BoxMesh from the input 3D garment. In Stage II, a semantics-aware autoregressive model parses BoxMesh into parametric sewing patterns. We adopt autoregressive modeling to naturally handle the variable-length and structured nature of panel configurations and stitching relationships. This decomposition separates geometric inversion from structured pattern inference, leading to more accurate and robust recovery. Extensive experiments demonstrate that our method achieves state-of-the-art performance on the GarmentCodeData benchmark and generalizes effectively to real-world scans and single-view images.

Method: 1) Creates TinySet-9M dataset for small object detection. 2) Proposes Point-Prompt Small Object Detection (P2SOD) paradigm using sparse point prompts at inference for semantic augmentation. 3) Develops DEAL framework learning prompt-conditioned representations from large-scale data.

Result: DEAL achieves 31.4% relative improvement over fully supervised baselines on AP75 metric on TinySet-9M. Generalizes effectively to unseen categories and datasets with single click at inference.

Conclusion: The work addresses data limitations with TinySet-9M and proposes a novel point-prompt paradigm for small object detection, demonstrating significant performance gains and generalization capabilities.

Abstract: Small object detection (SOD) remains challenging due to extremely limited pixels and ambiguous object boundaries. These characteristics lead to challenging annotation, limited availability of large-scale high-quality datasets, and inherently weak semantic representations for small objects. In this work, we first address the data limitation by introducing TinySet-9M, the first large-scale, multi-domain dataset for small object detection. Beyond filling the gap in large-scale datasets, we establish a benchmark to evaluate the effectiveness of existing label-efficient detection methods for small objects. Our evaluation reveals that weak visual cues further exacerbate the performance degradation of label-efficient methods in small object detection, highlighting a critical challenge in label-efficient SOD. Secondly, to tackle the limitation of insufficient semantic representation, we move beyond training-time feature enhancement and propose a new paradigm termed Point-Prompt Small Object Detection (P2SOD). This paradigm introduces sparse point prompts at inference time as an efficient information bridge for category-level localization, enabling semantic augmentation. Building upon the P2SOD paradigm and the large-scale TinySet-9M dataset, we further develop DEAL (DEtect Any smalL object), a scalable and transferable point-prompted detection framework that learns robust, prompt-conditioned representations from large-scale data. With only a single click at inference time, DEAL achieves a 31.4% relative improvement over fully supervised baselines under strict localization metrics (e.g., AP75) on TinySet-9M, while generalizing effectively to unseen categories and unseen datasets. Our project is available at https://zhuhaoraneis.github.io/TinySet-9M/.

Method: The authors first demonstrate adversarial membership fabrication across diverse architectures and datasets. They analyze the mechanism, revealing a distinctive gradient-norm collapse trajectory that separates fabricated from true members. Based on this insight, they develop a detection strategy using gradient-geometry signals and a robust inference framework.

Result: Experiments show adversarial membership fabrication is broadly effective against state-of-the-art MIAs, while the proposed detection and robust inference strategies significantly enhance resilience against such manipulation.

Conclusion: This work establishes the first comprehensive framework for adversarial membership manipulation in vision models, revealing a previously overlooked adversarial surface and providing principled defenses based on gradient geometry analysis.

Abstract: Membership inference attacks (MIAs) aim to determine whether a specific data point was part of a model’s training set, serving as effective tools for evaluating privacy leakage of vision models. However, existing MIAs implicitly assume honest query inputs, and their adversarial robustness remains unexplored. We show that MIAs for vision models expose a previously overlooked adversarial surface: adversarial membership manipulation, where imperceptible perturbations can reliably push non-member images into the “member” region of state-of-the-art MIAs. In this paper, we provide the first unified perspective on this phenomenon by analyzing its mechanism and implications. We begin by demonstrating that adversarial membership fabrication is consistently effective across diverse architectures and datasets. We then reveal a distinctive geometric signature - a characteristic gradient-norm collapse trajectory - that reliably separates fabricated from true members despite their nearly identical semantic representations. Building on this insight, we introduce a principled detection strategy grounded in gradient-geometry signals and develop a robust inference framework that substantially mitigates adversarial manipulation. Extensive experiments show that fabrication is broadly effective, while our detection and robust inference strategies significantly enhance resilience. This work establishes the first comprehensive framework for adversarial membership manipulation in vision models.

Method: Proposes EnsemHalDet, an ensemble-based hallucination detection framework that leverages multiple internal representations of VLMs including attention outputs and hidden states. The method trains independent detectors for each representation and combines them through ensemble learning techniques.

Result: Experimental results across multiple VQA datasets and VLMs show that EnsemHalDet consistently outperforms prior methods and single-detector models in terms of AUC, demonstrating that ensembling diverse internal signals significantly improves robustness in multimodal hallucination detection.

Conclusion: Ensemble-based approaches using multiple internal representations provide superior hallucination detection capabilities for VLMs compared to single-representation methods, offering a more robust solution for identifying factually incorrect or ungrounded multimodal outputs.

Abstract: Vision-Language Models (VLMs) excel at multimodal tasks, but they remain vulnerable to hallucinations that are factually incorrect or ungrounded in the input image. Recent work suggests that hallucination detection using internal representations is more efficient and accurate than approaches that rely solely on model outputs. However, existing internal-representation-based methods typically rely on a single representation or detector, limiting their ability to capture diverse hallucination signals. In this paper, we propose EnsemHalDet, an ensemble-based hallucination detection framework that leverages multiple internal representations of VLMs, including attention outputs and hidden states. EnsemHalDet trains independent detectors for each representation and combines them through ensemble learning. Experimental results across multiple VQA datasets and VLMs show that EnsemHalDet consistently outperforms prior methods and single-detector models in terms of AUC. These results demonstrate that ensembling diverse internal signals significantly improves robustness in multimodal hallucination detection.

Method: Proposes CANDLE with DINO Omni-layer Guidance (D.O.G.) to adaptively inject multi-layer DINOv3 features into encoder stages, and color-frequency refinement (BFACG + SFFB) to suppress decoder-side chromatic collapse and detail contamination.

Result: Achieves +1.22 dB PSNR gain over strongest prior method on CL3AN dataset, 3rd place on NTIRE 2026 ALN Color Lighting Challenge, and 2nd place in fidelity on White Lighting track with lowest FID.

Conclusion: DINOv3’s self-supervised features provide illumination-robust semantic priors for color normalization, enabling strong generalization across both chromatic and luminance-dominant illumination conditions.

Abstract: Color ambient lighting normalization under multi-colored illumination is challenging due to severe chromatic shifts, highlight saturation, and material-dependent reflectance. Existing geometric and low-level priors are insufficient for recovering object-intrinsic color when illumination-induced chromatic bias dominates. We observe that DINOv3’s self-supervised features remain highly consistent between colored-light inputs and ambient-lit ground truth, motivating their use as illumination-robust semantic priors. We propose CANDLE (Color Ambient Normalization with DINO Layer Enhancement), which introduces DINO Omni-layer Guidance (D.O.G.) to adaptively inject multi-layer DINOv3 features into successive encoder stages, and a color-frequency refinement design (BFACG + SFFB) to suppress decoder-side chromatic collapse and detail contamination. Experiments on CL3AN show a +1.22 dB PSNR gain over the strongest prior method. CANDLE achieves 3rd place on the NTIRE 2026 ALN Color Lighting Challenge and 2nd place in fidelity on the White Lighting track with the lowest FID, confirming strong generalization across both chromatic and luminance-dominant illumination conditions. Code is available at https://github.com/ron941/CANDLE.

Method: LumaFlux adapts a large pretrained diffusion transformer (DiT) with three key components: (1) Physically-Guided Adaptation (PGA) module injecting luminance, spatial descriptors, and frequency cues into attention through low-rank residuals; (2) Perceptual Cross-Modulation (PCM) layer stabilizing chroma and texture via FiLM conditioning from vision encoder features; (3) HDR Residual Coupler fusing physical and perceptual signals under adaptive modulation. A lightweight Rational-Quadratic Spline decoder reconstructs smooth tone fields.

Result: LumaFlux outperforms state-of-the-art baselines across benchmarks, achieving superior luminance reconstruction and perceptual color fidelity with minimal additional parameters. The authors also curated the first large-scale SDR-HDR training corpus and established a new evaluation benchmark.

Conclusion: LumaFlux provides an effective solution for SDR-to-HDR conversion by combining physical guidance with perceptual conditioning in a diffusion transformer framework, addressing limitations of existing inverse tone-mapping methods.

Abstract: The rapid adoption of HDR-capable devices has created a pressing need to convert the 8-bit Standard Dynamic Range (SDR) content into perceptually and physically accurate 10-bit High Dynamic Range (HDR). Existing inverse tone-mapping (ITM) methods often rely on fixed tone-mapping operators that struggle to generalize to real-world degradations, stylistic variations, and camera pipelines, frequently producing clipped highlights, desaturated colors, or unstable tone reproduction. We introduce LumaFlux, a first physically and perceptually guided diffusion transformer (DiT) for SDR-to-HDR reconstruction by adapting a large pretrained DiT. Our LumaFlux introduces (1) a Physically-Guided Adaptation (PGA) module that injects luminance, spatial descriptors, and frequency cues into attention through low-rank residuals; (2) a Perceptual Cross-Modulation (PCM) layer that stabilizes chroma and texture via FiLM conditioning from vision encoder features; and (3) an HDR Residual Coupler that fuses physical and perceptual signals under a timestep- and layer-adaptive modulation schedule. Finally, a lightweight Rational-Quadratic Spline decoder reconstructs smooth, interpretable tone fields for highlight and exposure expansion, enhancing the output of the VAE decoder to generate HDR. To enable robust HDR learning, we curate the first large-scale SDR-HDR training corpus. For fair and reproducible comparison, we further establish a new evaluation benchmark, comprising HDR references and corresponding expert-graded SDR versions. Across benchmarks, LumaFlux outperforms state-of-the-art baselines, achieving superior luminance reconstruction and perceptual color fidelity with minimal additional parameters.

Method: UNICA uses an action-conditioned diffusion model operating on 2D position maps to generate next-frame avatar geometry from keyboard inputs. A point transformer then maps the geometry to 3D Gaussian Splatting for high-fidelity free-view rendering, enabling skeleton-free generation with natural dynamics.

Result: The model captures hair and loose clothing dynamics without manual physical simulation, supports extra-long autoregressive generation, and unifies motion planning, rigging, physical simulation, and rendering in a single workflow.

Conclusion: UNICA presents a novel unified approach to 3D avatar control that simplifies traditional pipelines and enables natural dynamics through neural generation, representing the first model to combine all avatar control components into one framework.

Abstract: Controllable 3D human avatars have found widespread applications in 3D games, the metaverse, and AR/VR scenarios. The conventional approach to creating such a 3D avatar requires a lengthy, intricate pipeline encompassing appearance modeling, motion planning, rigging, and physical simulation. In this paper, we introduce UNICA (UNIfied neural Controllable Avatar), a skeleton-free generative model that unifies all avatar control components into a single neural framework. Given keyboard inputs akin to video game controls, UNICA generates the next frame of a 3D avatar’s geometry through an action-conditioned diffusion model operating on 2D position maps. A point transformer then maps the resulting geometry to 3D Gaussian Splatting for high-fidelity free-view rendering. Our approach naturally captures hair and loose clothing dynamics without manually designed physical simulation, and supports extra-long autoregressive generation. To the best of our knowledge, UNICA is the first model to unify the workflow of “motion planning, rigging, physical simulation, and rendering”. Code is released at https://github.com/zjh21/UNICA.

Method: Progressive Identity Alignment Network (PIA) with two key components: 1) Dual-Branch Disentangling Learning (DBDL) module separates identity-related cues from clothing-related factors for clothing-agnostic representation, and 2) Bi-Directional Prototype Learning (BPL) module performs intra-modality and inter-modality contrast to bridge modality gap while suppressing clothing interference.

Result: Extensive experiments on the new SYSU-CMCC dataset demonstrate that PIA establishes a strong baseline for the new CMCC-ReID task and significantly outperforms existing methods that only address either modality discrepancy or clothing variation separately.

Conclusion: The paper introduces a novel and realistic surveillance problem (CMCC-ReID), provides a comprehensive benchmark dataset (SYSU-CMCC), and proposes an effective solution (PIA) that progressively handles both clothing variation and modality discrepancy through disentangled learning and prototype-based alignment.

Abstract: Person Re-Identification (ReID) faces severe challenges from modality discrepancy and clothing variation in long-term surveillance scenario. While existing studies have made significant progress in either Visible-Infrared ReID (VI-ReID) or Clothing-Change ReID (CC-ReID), real-world surveillance system often face both challenges simultaneously. To address this overlooked yet realistic problem, we define a new task, termed Cross-Modality Clothing-Change Re-Identification (CMCC-ReID), which targets pedestrian matching across variations in both modality and clothing. To advance research in this direction, we construct a new benchmark SYSU-CMCC, where each identity is captured in both visible and infrared domains with distinct outfits, reflecting the dual heterogeneity of long-term surveillance. To tackle CMCC-ReID, we propose a Progressive Identity Alignment Network (PIA) that progressively mitigates the issues of clothing variation and modality discrepancy. Specifically, a Dual-Branch Disentangling Learning (DBDL) module separates identity-related cues from clothing-related factors to achieve clothing-agnostic representation, and a Bi-Directional Prototype Learning (BPL) module performs intra-modality and inter-modality contrast in the embedding space to bridge the modality gap while further suppressing clothing interference. Extensive experiments on the SYSU-CMCC dataset demonstrate that PIA establishes a strong baseline for this new task and significantly outperforms existing methods.

Method: Proposes a quantization-aware vision token pruning framework with a lightweight hybrid sensitivity metric combining simulated group-wise quantization error with outlier intensity. This metric is combined with standard semantic relevance scores to retain tokens that are both semantically informative and robust to quantization.

Result: Experiments on LLaVA architectures show consistent outperformance over naive integration baselines. At aggressive pruning ratio (12.5% tokens retained), improves accuracy by 2.24% over baseline and surpasses dense quantization without pruning.

Conclusion: First method to explicitly co-optimize vision token pruning and PTQ for accurate low-bit MLLM inference, addressing the coupling between these compression techniques.

Abstract: Multimodal Large Language Models (MLLMs) have shown strong reasoning ability, but their high computational and memory costs hinder deployment in resource-constrained settings. While Post-Training Quantization (PTQ) and vision token pruning are standard compression techniques, they are usually treated as independent optimizations. In this paper, we show that these two techniques are strongly coupled: naively applying semantic-based token pruning to PTQ-optimized MLLMs can discard activation outliers that are important for numerical stability and thus worsen quantization errors in low-bit regimes (\textit{e.g.}, W4A4). To address this issue, we propose a quantization-aware vision token pruning framework. Our method introduces a lightweight hybrid sensitivity metric that combines simulated group-wise quantization error with outlier intensity. By combining this metric with standard semantic relevance scores, the method retains tokens that are both semantically informative and robust to quantization. Experiments on standard LLaVA architectures show that our method consistently outperforms naive integration baselines. At an aggressive pruning ratio that retains only 12.5% of visual tokens, our framework improves accuracy by 2.24% over the baseline and even surpasses dense quantization without pruning. To the best of our knowledge, this is the first method that explicitly co-optimizes vision token pruning and PTQ for accurate low-bit MLLM inference.

Method: 1) Recast perceptual cues (semantics, geometry, spatio-temporal trajectory) into unified pseudo-RGB format; 2) Bidirectionally Controlled Teacher architecture with parallel branches to decouple RGB and perception processing; 3) Zero-initialized control links for pixel-wise consistency; 4) Distillation into single-stream student model; 5) MMPhysPipe pipeline for physics-rich multimodal dataset construction using VLM-guided subject identification.

Result: MMPhysVideo consistently improves physical plausibility and visual quality over advanced models across various benchmarks, achieving state-of-the-art performance without additional inference costs.

Conclusion: The framework successfully scales physical plausibility in video generation through joint multimodal modeling, demonstrating that incorporating perceptual cues beyond pixels significantly enhances physical consistency in generated videos.

Abstract: Despite advancements in generating visually stunning content, video diffusion models (VDMs) often yield physically inconsistent results due to pixel-only reconstruction. To address this, we propose MMPhysVideo, the first framework to scale physical plausibility in video generation through joint multimodal modeling. We recast perceptual cues, specifically semantics, geometry, and spatio-temporal trajectory, into a unified pseudo-RGB format, enabling VDMs to directly capture complex physical dynamics. To mitigate cross-modal interference, we propose a Bidirectionally Controlled Teacher architecture, which utilizes parallel branches to fully decouple RGB and perception processing and adopts two zero-initialized control links to gradually learn pixel-wise consistency. For inference efficiency, the teacher’s physical prior is distilled into a single-stream student model via representation alignment. Furthermore, we present MMPhysPipe, a scalable data curation and annotation pipeline tailored for constructing physics-rich multimodal datasets. MMPhysPipe employs a vision-language model (VLM) guided by a chain-of-visual-evidence rule to pinpoint physical subjects, enabling expert models to extract multi-granular perceptual information. Without additional inference costs, MMPhysVideo consistently improves physical plausibility and visual quality over advanced models across various benchmarks and achieves state-of-the-art performance compared to existing methods.

Method: Leverages video diffusion models for 3D priors, uses geometry-aware expansion to extend scene coverage, introduces multi-stage camera control with dual-branch camera injection and attention modulation, and proposes collision-aware camera trajectory planning with enhanced 3D Gaussian Splatting pipeline including depth-aligned supervision, structural regularization and refinement.

Result: Achieves state-of-the-art novel-view synthesis under large viewpoint shifts and substantially improves 3D reconstruction fidelity in extensive experiments.

Conclusion: NavCrafter provides an effective framework for exploring 3D scenes from single images with controllable camera trajectories and high-quality reconstruction.

Abstract: Creating flexible 3D scenes from a single image is vital when direct 3D data acquisition is costly or impractical. We introduce NavCrafter, a novel framework that explores 3D scenes from a single image by synthesizing novel-view video sequences with camera controllability and temporal-spatial consistency. NavCrafter leverages video diffusion models to capture rich 3D priors and adopts a geometry-aware expansion strategy to progressively extend scene coverage. To enable controllable multi-view synthesis, we introduce a multi-stage camera control mechanism that conditions diffusion models with diverse trajectories via dual-branch camera injection and attention modulation. We further propose a collision-aware camera trajectory planner and an enhanced 3D Gaussian Splatting (3DGS) pipeline with depth-aligned supervision, structural regularization and refinement. Extensive experiments demonstrate that NavCrafter achieves state-of-the-art novel-view synthesis under large viewpoint shifts and substantially improves 3D reconstruction fidelity.

Method: Extracts features from both image sequences and goal observations, then fuses them using a spatio-temporal fusion module that performs spatial graph reasoning within each frame and models temporal dynamics using a hybrid temporal shift module combined with multi-resolution difference-aware convolution.

Result: Experimental results demonstrate consistent improvement in navigation performance and the approach offers a generalizable visual backbone for goal-conditioned control.

Conclusion: The proposed unified spatio-temporal representation framework effectively enhances visual encoding for robotic navigation by preserving both spatial and temporal structure.

Abstract: Visual navigation requires the robot to reach a specified goal such as an image, based on a sequence of first-person visual observations. While recent learning-based approaches have made significant progress, they often focus on improving policy heads or decision strategies while relying on simplistic feature encoders and temporal pooling to represent visual input. This leads to the loss of fine-grained spatial and temporal structure, ultimately limiting accurate action prediction and progress estimation. In this paper, we propose a unified spatio-temporal representation framework that enhances visual encoding for robotic navigation. Our approach extracts features from both image sequences and goal observations, and fuses them using the designed spatio-temporal fusion module. This module performs spatial graph reasoning within each frame and models temporal dynamics using a hybrid temporal shift module combined with multi-resolution difference-aware convolution. Experimental results demonstrate that our approach consistently improves navigation performance and offers a generalizable visual backbone for goal-conditioned control. Code is available at \href{https://github.com/hren20/STRNet}{https://github.com/hren20/STRNet}.

Method: Fact-Hash projects 3D coordinates into multiple lower-dimensional forms (2D or 1D) before applying hash functions, then aggregates them into a single feature, combining tensor factorization and hash encoding techniques.

Result: Fact-Hash saves over one-third memory usage while maintaining PSNR values compared to previous encoding methods, with superior computational efficiency and energy consumption in on-device experiments.

Conclusion: Fact-Hash is a promising solution for improving feature grid representation, addressing memory constraints, and enhancing quality in various applications requiring efficient on-device NeRF training.

Abstract: We introduce Fact-Hash, a novel parameter-encoding method for training on-device neural radiance fields. Neural Radiance Fields (NeRF) have proven pivotal in 3D representations, but their applications are limited due to large computational resources. On-device training can open large application fields, providing strength in communication limitations, privacy concerns, and fast adaptation to a frequently changing scene. However, challenges such as limited resources (GPU memory, storage, and power) impede their deployment. To handle this, we introduce Fact-Hash, a novel parameter-encoding merging Tensor Factorization and Hash-encoding techniques. This integration offers two benefits: the use of rich high-resolution features and the few-shot robustness. In Fact-Hash, we project 3D coordinates into multiple lower-dimensional forms (2D or 1D) before applying the hash function and then aggregate them into a single feature. Comparative evaluations against state-of-the-art methods demonstrate Fact-Hash’s superior memory efficiency, preserving quality and rendering speed. Fact-Hash saves memory usage by over one-third while maintaining the PSNR values compared to previous encoding methods. The on-device experiment validates the superiority of Fact-Hash compared to alternative positional encoding methods in computational efficiency and energy consumption. These findings highlight Fact-Hash as a promising solution to improve feature grid representation, address memory constraints, and improve quality in various applications. Project page: https://facthash.github.io/

Method: Proposes DeformPIC, a deformation-based framework that learns to deform query point clouds under task-specific guidance from prompts, enabling explicit geometric reasoning and consistent objectives between training and inference.

Result: DeformPIC consistently outperforms previous state-of-the-art methods, achieving reductions of 1.6, 1.8, and 4.7 points in average Chamfer Distance on reconstruction, denoising, and registration tasks respectively, and shows superior performance on new out-of-domain generalization benchmarks.

Conclusion: DeformPIC addresses key limitations of MPM-based point cloud ICL methods by introducing a deformation-based approach that enables better geometric reasoning and eliminates training-inference objective mismatch, achieving state-of-the-art performance across multiple tasks.

Abstract: Recent advances in point cloud In-Context Learning (ICL) have demonstrated strong multitask capabilities. Existing approaches typically adopt a Masked Point Modeling (MPM)-based paradigm for point cloud ICL. However, MPM-based methods directly predict the target point cloud from masked tokens without leveraging geometric priors, requiring the model to infer spatial structure and geometric details solely from token-level correlations via transformers. Additionally, these methods suffer from a training-inference objective mismatch, as the model learns to predict the target point cloud using target-side information that is unavailable at inference time. To address these challenges, we propose DeformPIC, a deformation-based framework for point cloud ICL. Unlike existing approaches that rely on masked reconstruction, DeformPIC learns to deform the query point cloud under task-specific guidance from prompts, enabling explicit geometric reasoning and consistent objectives. Extensive experiments demonstrate that DeformPIC consistently outperforms previous state-of-the-art methods, achieving reductions of 1.6, 1.8, and 4.7 points in average Chamfer Distance on reconstruction, denoising, and registration tasks, respectively. Furthermore, we introduce a new out-of-domain benchmark to evaluate generalization across unseen data distributions, where DeformPIC achieves state-of-the-art performance.

Method: Two-stage hierarchical approach: 1) Establish face-edge incidence relations for topological scaffold, generate face proxies and initial edge curves; 2) Use Transformer-based diffusion modules to refine geometry with precise face surfaces and vertex positions, enforcing edge-vertex adjacencies for structural consistency.

Result: HiDiGen achieves strong performance in generating novel, diverse, and topologically sound CAD models.

Conclusion: The progressive geometry hierarchy enables generation of more novel and diverse shapes while two-stage topological modeling ensures high validity of B-rep structures.

Abstract: Boundary representation (B-rep) is the standard 3D modeling format in CAD systems, encoding both geometric primitives and topological connectivity. Despite its prevalence, deep generative modeling of valid B-rep structures remains challenging due to the intricate interplay between discrete topology and continuous geometry. In this paper, we propose HiDiGen, a hierarchical generation framework that decouples geometry modeling into two stages, each guided by explicitly modeled topological constraints. Specifically, our approach first establishes face-edge incidence relations to define a coherent topological scaffold, upon which face proxies and initial edge curves are generated. Subsequently, multiple Transformer-based diffusion modules are employed to refine the geometry by generating precise face surfaces and vertex positions, with edge-vertex adjacencies dynamically established and enforced to preserve structural consistency. This progressive geometry hierarchy enables the generation of more novel and diverse shapes, while two-stage topological modeling ensures high validity. Experimental results show that HiDiGen achieves strong performance, generating novel, diverse, and topologically sound CAD models.

Method: Proposes fully end-to-end paradigm with Sentence Conditioned Adapter (SCADA) that leverages sentence features to train a small portion of video backbone parameters adaptively, enabling precise integration of linguistic embeddings with visual features.

Result: Outperforms state-of-the-art approaches on two benchmarks, validates effectiveness of end-to-end learning over frozen baselines across different model scales.

Conclusion: End-to-end joint optimization with sentence-conditioned adaptation significantly enhances visual representation for temporal sentence grounding, enabling deployment of deeper networks with reduced memory.

Abstract: Temporal sentence grounding in videos (TSGV) aims to localize a temporal segment that semantically corresponds to a sentence query from an untrimmed video. Most current methods adopt pre-trained query-agnostic visual encoders for offline feature extraction, and the video backbones are frozen and not optimized for TSGV. This leads to a task discrepancy issue for the video backbone trained for visual classification, but utilized for TSGV. To bridge this gap, we propose a fully end-to-end paradigm that jointly optimizes the video backbone and localization head. We first conduct an empirical study validating the effectiveness of end-to-end learning over frozen baselines across different model scales. Furthermore, we introduce a Sentence Conditioned Adapter (SCADA), which leverages sentence features to train a small portion of video backbone parameters adaptively. SCADA facilitates the deployment of deeper network backbones with reduced memory and significantly enhances visual representation by modulating feature maps through precise integration of linguistic embeddings. Experiments on two benchmarks show that our method outperforms state-of-the-art approaches. The code and models will be released.

Method: 1) Leverages 3D priors from video generation models to transform single-view reconstruction into calibrated multi-view task; 2) Introduces neural orientation extractor trained on sparse real-image annotations for full-view orientation estimation; 3) Designs two-stage strand-growing algorithm based on hybrid implicit field for fast synthesis of detailed 3D strand curves.

Result: Achieves state-of-the-art performance on single-view 3D hair strand reconstruction across diverse hair portraits, with improved results in both visible and invisible regions.

Conclusion: The proposed framework effectively addresses challenges in single-view hair reconstruction by leveraging video generation priors and neural orientation extraction, enabling high-quality 3D strand reconstruction with realistic attributes in unseen regions.

Abstract: Reconstructing strand-level 3D hair from a single-view image is highly challenging, especially when preserving consistent and realistic attributes in unseen regions. Existing methods rely on limited frontal-view cues and small-scale/style-restricted synthetic data, often failing to produce satisfactory results in invisible regions. In this work, we propose a novel framework that leverages the strong 3D priors of video generation models to transform single-view hair reconstruction into a calibrated multi-view reconstruction task. To balance reconstruction quality and efficiency for the reformulated multi-view task, we further introduce a neural orientation extractor trained on sparse real-image annotations for better full-view orientation estimation. In addition, we design a two-stage strand-growing algorithm based on a hybrid implicit field to synthesize the 3D strand curves with fine-grained details at a relatively fast speed. Extensive experiments demonstrate that our method achieves state-of-the-art performance on single-view 3D hair strand reconstruction on a diverse range of hair portraits in both visible and invisible regions.

Method: The paper examines token-level warping in ViT-based MLLMs, comparing forward and backward warping approaches. Backward token warping defines a dense grid on the target view and retrieves corresponding source-view tokens for each grid point, which proves more stable and preserves semantic coherence better than pixel-based methods.

Result: Experiments on the proposed ViewBench benchmark show that token-level warping enables MLLMs to reason reliably from nearby viewpoints, consistently outperforming all baselines including pixel-wise warping approaches, spatially fine-tuned MLLMs, and generative warping methods.

Conclusion: Token-level warping, particularly backward warping, provides a robust approach for MLLMs to handle viewpoint changes by leveraging the structural representations in image tokens, addressing the limitations of pixel-based warping methods.

Abstract: Can warping tokens, rather than pixels, help multimodal large language models (MLLMs) understand how a scene appears from a nearby viewpoint? While MLLMs perform well on visual reasoning, they remain fragile to viewpoint changes, as pixel-wise warping is highly sensitive to small depth errors and often introduces geometric distortions. Drawing on theories of mental imagery that posit part-level structural representations as the basis for human perspective transformation, we examine whether image tokens in ViT-based MLLMs serve as an effective substrate for viewpoint changes. We compare forward and backward warping, finding that backward token warping, which defines a dense grid on the target view and retrieves a corresponding source-view token for each grid point, achieves greater stability and better preserves semantic coherence under viewpoint shifts. Experiments on our proposed ViewBench benchmark demonstrate that token-level warping enables MLLMs to reason reliably from nearby viewpoints, consistently outperforming all baselines including pixel-wise warping approaches, spatially fine-tuned MLLMs, and a generative warping method.

Method: Two-stage approach: (1) Train a Sparse Autoencoder (SAE) on patch-token features from frozen foundation models, (2) Optimize only sparse guide coefficients in SAE latent space using auxiliary pixel-level masks while freezing backbone and SAE. The guide coefficients generate normal/anomaly guide vectors via SAE dictionary.

Result: SPG achieves competitive image-level detection and strong pixel-level segmentation on MVTec AD and VisA under cross-dataset zero-shot settings. With DINOv3, SPG attains highest pixel-level AUROC among compared methods. Also works with OpenCLIP (ViT-L/14@336px). Learned coefficients reveal category-general and category-specific factors.

Conclusion: SPG provides an effective prompt-free framework for zero-shot anomaly detection that leverages sparse representations for interpretability and achieves strong performance without target-domain adaptation.

Abstract: We study zero-shot anomaly detection and segmentation using frozen foundation model features, where all learnable parameters are trained only on a labeled auxiliary dataset and deployed to unseen target categories without any target-domain adaptation. Existing prompt-based approaches use handcrafted or learned prompt embeddings as reference vectors for normal/anomalous states. We propose Sparse-Projected Guides (SPG), a prompt-free framework that learns sparse guide coefficients in the Sparse Autoencoder (SAE) latent space, which generate normal/anomaly guide vectors via the SAE dictionary. SPG employs a two stage learning strategy on the labeled auxiliary dataset: (i) train an SAE on patch-token features, and (ii) optimize only guide coefficients using auxiliary pixel-level masks while freezing the backbone and SAE. On MVTec AD and VisA under cross-dataset zero-shot settings, SPG achieves competitive image-level detection and strong pixel-level segmentation; with DINOv3, SPG attains the highest pixellevel AUROC among the compared methods. We also report SPG instantiated with OpenCLIP (ViT-L/14@336px) to align the backbone with CLIP-based baselines. Moreover, the learned guide coefficients trace decisions back to a small set of dictionary atoms, revealing category-general and category-specific factors.

Method: Built on a frozen pre-trained model, the framework adaptively appends instrument-aware prompts for new classes. It uses a hierarchical prompt parsing tree with instrument-shared prompts as root nodes, part-shared prompts as intermediate nodes, and instrument-distinct prompts as leaf nodes. For backward transfer, it conducts self-reflection refining via directed-weighted graph propagation to examine knowledge associations.

Result: The framework achieves more than 5% and 11% improvements over competing methods on two public benchmarks, and is applicable to both CNN-based models and transformer-based foundation models.

Conclusion: The proposed self-reflection hierarchical prompt framework effectively unlocks positive forward and backward knowledge transfer in class incremental segmentation for surgical instruments, enabling proficient learning of new instruments while improving existing skills and avoiding catastrophic forgetting.

Abstract: To continuously enhance model adaptability in surgical video scene parsing, recent studies incrementally update it to progressively learn to segment an increasing number of surgical instruments over time. However, prior works constantly overlooked the potential of positive forward knowledge transfer, i.e., how past knowledge could help learn new classes, and positive backward knowledge transfer, i.e., how learning new classes could help refine past knowledge. In this paper, we propose a self-reflection hierarchical prompt framework that unlocks the power of positive forward and backward knowledge transfer in class incremental segmentation, aiming to proficiently learn new instruments, improve existing skills of regular instruments, and avoid catastrophic forgetting of old instruments. Our framework is built on a frozen, pre-trained model that adaptively appends instrument-aware prompts for new classes throughout training episodes. To enable positive forward knowledge transfer, we organize instrument prompts into a hierarchical prompt parsing tree with the instrument-shared prompt partition as the root node, n-part-shared prompt partitions as intermediate nodes and instrument-distinct prompt partitions as leaf nodes, to expose the reusable historical knowledge for new classes to simplify their learning. Conversely, to encourage positive backward knowledge transfer, we conduct self-reflection refining on existing knowledge by directed-weighted graph propagation, examining the knowledge associations recorded in the tree to improve its representativeness without causing catastrophic forgetting. Our framework is applicable to both CNN-based models and advanced transformer-based foundation models, yielding more than 5% and 11% improvements over the competing methods on two public benchmarks respectively.

Method: 1) Instruction-guided multi-stage training with table-specific instructions to focus on fine-grained structural patterns; 2) Complementary TSR fine-tuning to preserve visual information modeling; 3) Table Mix Expand (TME) - template-free method for synthesizing large-scale authentic tabular data; 4) BCDSTab benchmark with 900 complex table images.

Result: Achieves state-of-the-art performance on multiple public datasets (FinTabNet, PubTabNet, MUSTARD) and the proposed BCDSTab benchmark. Ablation studies confirm benefits of tabular-data-specific instructions and synthetic data.

Conclusion: InstructTable effectively addresses limitations of both visual-only and vision-language approaches by combining instruction-guided attention to structural patterns with robust visual modeling, enabled by synthetic data generation for complex table scenarios.

Abstract: Table structure recognition (TSR) holds widespread practical importance by parsing tabular images into structured representations, yet encounters significant challenges when processing complex layouts involving merged or empty cells. Traditional visual-centric models rely exclusively on visual information while lacking crucial semantic support, thereby impeding accurate structural recognition in complex scenarios. Vision-language models leverage contextual semantics to enhance comprehension; however, these approaches underemphasize the modeling of visual structural information. To address these limitations, this paper introduces InstructTable, an instruction-guided multi-stage training TSR framework. Meticulously designed table instruction pre-training directs attention toward fine-grained structural patterns, enhancing comprehension of complex tables. Complementary TSR fine-tuning preserves robust visual information modeling, maintaining high-precision table parsing across diverse scenarios. Furthermore, we introduce Table Mix Expand (TME), an innovative template-free method for synthesizing large-scale authentic tabular data. Leveraging TME, we construct the Balanced Complex Dense Synthetic Tables (BCDSTab) benchmark, comprising 900 complex table images synthesized through our method to serve as a rigorous benchmark. Extensive experiments on multiple public datasets (FinTabNet, PubTabNet, MUSTARD) and BCDSTab demonstrate that InstructTable achieves state-of-the-art performance in TSR tasks. Ablation studies further confirm the positive impact of the proposed tabular-data-specific instructions and synthetic data.

Method: Proposes conditioning-guided edited reconstruction framework that performs editing as constrained inversion in structured avatar latent space, restricting updates to low-dimensional, part-specific edit subspace. Uses conditioning objective from local linearization of full decoding-and-rendering pipeline to create edit-subspace information matrix whose spectrum predicts stability and drives frame reweighting/keyframe activation.

Result: Method operates on small subspace matrices efficiently (via Hessian-vector products) and improves stability under limited edited supervision.

Conclusion: The framework enables stable avatar editing with sparse supervision by properly conditioning the inversion process in a structured latent space to prevent unintended identity changes.

Abstract: Editing animatable human avatars typically relies on sparse supervision, often a few edited keyframes, yet naively fitting a reconstructed avatar to these edits frequently causes identity leakage and pose-dependent temporal flicker. We argue that these failures are best understood as an ill-conditioned inversion: the available edited constraints do not sufficiently determine the latent directions responsible for the intended edit. We propose a conditioning-guided edited reconstruction framework that performs editing as a constrained inversion in a structured avatar latent space, restricting updates to a low-dimensional, part-specific edit subspace to prevent unintended identity changes. Crucially, we design the editing constraints during inversion by optimizing a conditioning objective derived from a local linearization of the full decoding-and-rendering pipeline, yielding an edit-subspace information matrix whose spectrum predicts stability and drives frame reweighting / keyframe activation. The resulting method operates on small subspace matrices and can be implemented efficiently (e.g., via Hessian-vector products), and improves stability under limited edited supervision.

Method: ProVCA uses a three-stage progressive approach: 1) segment localization to identify query-relevant video segments, 2) snippet selection based on similarity, and 3) keyframe refinement to pinpoint specific frames. This progressively narrows scope from coarse segments to fine frames.

Result: Achieves state-of-the-art zero-shot accuracies: 69.3% on EgoSchema, 80.5% on NExT-QA, and 77.7% on IntentQA, while using fewer frames than previous training-free methods.

Conclusion: ProVCA enables efficient video understanding by progressively condensing long videos into key frames for MLLM reasoning, balancing computational efficiency with visual detail preservation.

Abstract: Understanding long videos requires extracting query-relevant information from long sequences under tight compute budgets. Existing text-then-LLM pipelines lose fine-grained visual cues, while video-based multimodal large language models (MLLMs) can keep visual details but are too frame-hungry and computationally expensive. In this work, we aim to harness MLLMs for efficient video understanding. We propose ProVCA, a progressive video condensation agent that iteratively locates key video frames at multiple granularities. ProVCA first adopts a segment localization module to identify the video segment relevant to the query, then a snippet selection module to select important snippets based on similarity, and finally a keyframe refinement module to pinpoint specific keyframes in those snippets. By progressively narrowing the scope from coarse segments to fine frames, ProVCA identifies a small set of keyframes for MLLM-based reasoning. ProVCA achieves state-of-the-art zero-shot accuracies of 69.3% on EgoSchema, 80.5% on NExT-QA, and 77.7% on IntentQA, while using fewer frames than previous training-free methods.

Method: 1) Created a fully automated procedural data engine generating 200k+ synthetic geometry diagrams with pixel-perfect masks and diverse referring expressions; 2) Domain-specific fine-tuning of vision-language models (Florence-2); 3) Introduced Buffered IoU metric for geometry-aware evaluation.

Result: Fine-tuned Florence-2 achieved 49% IoU and 85% Buffered IoU, compared to <1% IoU in zero-shot settings. The synthetic data generation required zero manual annotation.

Conclusion: The approach establishes a foundation for building Artificial General Teachers capable of providing visually grounded, step-by-step explanations of geometry problems through effective visual explanation systems.

Abstract: We study visual explanation in geometry education as a Referring Image Segmentation (RIS) problem: given a diagram and a natural language description, the task is to produce a pixel-level mask for the referred geometric element. However, existing RIS models trained on natural image benchmarks such as RefCOCO fail catastrophically on geometric diagrams due to the fundamental domain shift between photographic scenes and abstract, textureless schematics. To address the absence of suitable training data, we present a fully automated procedural data engine that generates over 200,000 synthetic geometry diagrams with pixel-perfect segmentation masks and linguistically diverse referring expressions, requiring zero manual annotation. We further propose domain-specific fine-tuning of vision-language models (VLMs), demonstrating that a fine-tuned Florence-2 achieves 49% IoU and 85% Buffered IoU (BIoU), compared to <1% IoU in zero-shot settings. We introduce Buffered IoU, a geometry-aware evaluation metric that accounts for thin-structure localization, and show that it better reflects true segmentation quality than standard IoU. Our results establish a foundation for building Artificial General Teachers (AGTs) capable of providing visually grounded, step-by-step explanations of geometry problems.

Method: Proposes a unified evaluation framework with a lightweight network using divide-and-conquer strategy. First decomposes fusion results into infrared and visible components, then measures information preservation. Incorporates contrastive learning and LLM guidance for perceptual assessment. Introduces consistency evaluation framework comparing metrics to human perception via no-reference scores and downstream tasks.

Result: The learning-based evaluation paradigm achieves superior efficiency (up to 1,000× faster) and greater consistency across standard image fusion benchmarks compared to traditional metrics.

Conclusion: The proposed framework provides an efficient, consistent evaluation method specifically designed for image fusion that better aligns with human visual perception than traditional borrowed metrics.

Abstract: Evaluation is essential in image fusion research, yet most existing metrics are directly borrowed from other vision tasks without proper adaptation. These traditional metrics, often based on complex image transformations, not only fail to capture the true quality of the fusion results but also are computationally demanding. To address these issues, we propose a unified evaluation framework specifically tailored for image fusion. At its core is a lightweight network designed efficiently to approximate widely used metrics, following a divide-and-conquer strategy. Unlike conventional approaches that directly assess similarity between fused and source images, we first decompose the fusion result into infrared and visible components. The evaluation model is then used to measure the degree of information preservation in these separated components, effectively disentangling the fusion evaluation process. During training, we incorporate a contrastive learning strategy and inform our evaluation model by perceptual scene assessment provided by a large language model. Last, we propose the first consistency evaluation framework, which measures the alignment between image fusion metrics and human visual perception, using both independent no-reference scores and downstream tasks performance as objective references. Extensive experiments show that our learning-based evaluation paradigm delivers both superior efficiency (up to 1,000 times faster) and greater consistency across a range of standard image fusion benchmarks. Our code will be publicly available at https://github.com/AWCXV/EvaNet.

Method: Proposes RayMamba with a ray-aligned serialization strategy that organizes sparse voxels into sector-wise ordered sequences, preserving directional continuity and occlusion-related context for subsequent Mamba-based modeling. Compatible with both LiDAR-only and multimodal detectors with modest overhead.

Result: Extensive experiments on nuScenes and Argoverse 2 show consistent improvements across strong baselines. Achieves up to 2.49 mAP and 1.59 NDS gain in challenging 40-50m range on nuScenes, and improves VoxelNeXt on Argoverse 2 from 30.3 to 31.2 mAP.

Conclusion: RayMamba effectively addresses long-range 3D detection challenges by preserving geometric context through ray-aligned serialization and Mamba-based modeling, demonstrating significant performance gains on benchmark datasets.

Abstract: Long-range 3D object detection remains challenging because LiDAR observations become highly sparse and fragmented in the far field, making reliable context modeling difficult for existing detectors. To address this issue, recent state space model (SSM)-based methods have improved long-range modeling efficiency. However, their effectiveness is still limited by generic serialization strategies that fail to preserve meaningful contextual neighborhoods in sparse scenes. To address this issue, we propose RayMamba, a geometry-aware plug-and-play enhancement for voxel-based 3D detectors. RayMamba organizes sparse voxels into sector-wise ordered sequences through a ray-aligned serialization strategy, which preserves directional continuity and occlusion-related context for subsequent Mamba-based modeling. It is compatible with both LiDAR-only and multimodal detectors, while introducing only modest overhead. Extensive experiments on nuScenes and Argoverse 2 demonstrate consistent improvements across strong baselines. In particular, RayMamba achieves up to 2.49 mAP and 1.59 NDS gain in the challenging 40–50 m range on nuScenes, and further improves VoxelNeXt on Argoverse 2 from 30.3 to 31.2 mAP.

Method: Proposes UniSpector with three key components: 1) Spatial-Spectral Prompt Encoder for orientation-invariant fine-grained representations, 2) Contrastive Prompt Encoder to regularize prompt space into semantically organized angular manifold, and 3) Prompt-guided Query Selection for adaptive object queries aligned with prompts.

Result: Significantly outperforms baselines by at least 19.7% AP50b and 15.8% AP50m on the new Inspect Anything benchmark for visual-prompt-based open-set defect localization.

Conclusion: Enables scalable, retraining-free inspection paradigm for evolving industrial environments and provides insights into generic visual prompting design.

Abstract: Although industrial inspection systems should be capable of recognizing unprecedented defects, most existing approaches operate under a closed-set assumption, which prevents them from detecting novel anomalies. While visual prompting offers a scalable alternative for industrial inspection, existing methods often suffer from prompt embedding collapse due to high intra-class variance and subtle inter-class differences. To resolve this, we propose UniSpector, which shifts the focus from naive prompt-to-region matching to the principled design of a semantically structured and transferable prompt topology. UniSpector employs the Spatial-Spectral Prompt Encoder to extract orientation-invariant, fine-grained representations; these serve as a solid basis for the Contrastive Prompt Encoder to explicitly regularize the prompt space into a semantically organized angular manifold. Additionally, Prompt-guided Query Selection generates adaptive object queries aligned with the prompt. We introduce Inspect Anything, the first benchmark for visual-prompt-based open-set defect localization, where UniSpector significantly outperforms baselines by at least 19.7% and 15.8% in AP50b and AP50m, respectively. These results show that our method enable a scalable, retraining-free inspection paradigm for continuously evolving industrial environments, while offering critical insights into the design of generic visual prompting.

Method: Combines Gaussian Processes with 4D Gaussian Splatting using spatio-temporal kernels to model deformation field correlations, with variational Gaussian Processes and inducing points for scalable inference on large numbers of Gaussian primitives.

Result: GP-4DGS improves reconstruction quality while providing reliable uncertainty estimates that effectively identify regions of high motion ambiguity, demonstrating successful uncertainty quantification and temporal extrapolation.

Conclusion: The framework bridges probabilistic modeling and neural graphics by addressing motion ambiguity and prediction reliability in dynamic scene reconstruction, representing a meaningful advancement in 4D scene understanding.

Abstract: We present GP-4DGS, a novel framework that integrates Gaussian Processes (GPs) into 4D Gaussian Splatting (4DGS) for principled probabilistic modeling of dynamic scenes. While existing 4DGS methods focus on deterministic reconstruction, they are inherently limited in capturing motion ambiguity and lack mechanisms to assess prediction reliability. By leveraging the kernel-based probabilistic nature of GPs, our approach introduces three key capabilities: (i) uncertainty quantification for motion predictions, (ii) motion estimation for unobserved or sparsely sampled regions, and (iii) temporal extrapolation beyond observed training frames. To scale GPs to the large number of Gaussian primitives in 4DGS, we design spatio-temporal kernels that capture the correlation structure of deformation fields and adopt variational Gaussian Processes with inducing points for tractable inference. Our experiments show that GP-4DGS enhances reconstruction quality while providing reliable uncertainty estimates that effectively identify regions of high motion ambiguity. By addressing these challenges, our work takes a meaningful step toward bridging probabilistic modeling and neural graphics.

Method: Proposes using provenance information from the synthetic data generation process (indicating target vs. non-target regions) as auxiliary supervision. Input gradients are decomposed based on this information, with gradient guidance introduced to suppress gradients over non-target regions, promoting focus on target regions.

Result: The method demonstrates effectiveness and generality across multiple tasks and modalities including weakly supervised object localization, spatio-temporal action localization, and image classification.

Conclusion: The proposed framework successfully leverages synthetic data provenance to guide models toward learning discriminative representations focused on target regions, addressing limitations of existing synthetic data approaches.

Abstract: Learning methods using synthetic data have attracted attention as an effective approach for increasing the diversity of training data while reducing collection costs, thereby improving the robustness of model discrimination. However, many existing methods improve robustness only indirectly through the diversification of training samples and do not explicitly teach the model which regions in the input space truly contribute to discrimination; consequently, the model may learn spurious correlations caused by synthesis biases and artifacts. Motivated by this limitation, this paper proposes a learning framework that uses provenance information obtained during the training data synthesis process, indicating whether each region in the input space originates from the target object, as an auxiliary supervisory signal to promote the acquisition of representations focused on target regions. Specifically, input gradients are decomposed based on information about target and non-target regions during synthesis, and input gradient guidance is introduced to suppress gradients over non-target regions. This suppresses the model’s reliance on non-target regions and directly promotes the learning of discriminative representations for target regions. Experiments demonstrate the effectiveness and generality of the proposed method across multiple tasks and modalities, including weakly supervised object localization, spatio-temporal action localization, and image classification.

Method: Proposes BEVPredFormer with recurrent-free design using gated transformer layers, divided spatio-temporal attention mechanisms, multi-scale head tasks, attention-based 3D camera projection, and a difference-guided feature extraction module for enhanced temporal representations.

Result: Extensive ablation studies validate each component’s effectiveness. On nuScenes dataset, BEVPredFormer performs on par or surpasses state-of-the-art methods for autonomous driving perception.

Conclusion: BEVPredFormer demonstrates potential for robust and efficient autonomous driving perception through its attention-based temporal processing and camera-only architecture for BEV instance prediction.

Abstract: A robust awareness of how dynamic scenes evolve is essential for Autonomous Driving systems, as they must accurately detect, track, and predict the behaviour of surrounding obstacles. Traditional perception pipelines that rely on modular architectures tend to suffer from cumulative errors and latency. Instance Prediction models provide a unified solution, performing Bird’s-Eye-View segmentation and motion estimation across current and future frames using information directly obtained from different sensors. However, a key challenge in these models lies in the effective processing of the dense spatial and temporal information inherent in dynamic driving environments. This level of complexity demands architectures capable of capturing fine-grained motion patterns and long-range dependencies without compromising real-time performance. We introduce BEVPredFormer, a novel camera-only architecture for BEV instance prediction that uses attention-based temporal processing to improve temporal and spatial comprehension of the scene and relies on an attention-based 3D projection of the camera information. BEVPredFormer employs a recurrent-free design that incorporates gated transformer layers, divided spatio-temporal attention mechanisms, and multi-scale head tasks. Additionally, we incorporate a difference-guided feature extraction module that enhances temporal representations. Extensive ablation studies validate the effectiveness of each architectural component. When evaluated on the nuScenes dataset, BEVPredFormer was on par or surpassed State-Of-The-Art methods, highlighting its potential for robust and efficient Autonomous Driving perception.

Method: Introduces PolyReal benchmark grounded in real-world scientific practices, covering five critical capabilities: (1) foundational knowledge application, (2) lab safety analysis, (3) experiment mechanism reasoning, (4) raw data extraction, and (5) performance & application exploration.

Result: Evaluation of leading MLLMs reveals a capability imbalance - models perform well on knowledge-intensive reasoning (e.g., experiment mechanism reasoning) but drop sharply on practice-based tasks (e.g., lab safety analysis and raw data extraction).

Conclusion: PolyReal exposes a severe gap between abstract scientific knowledge and its practical, context-dependent application in MLLMs, addressing the evaluation gap and providing a practical benchmark for assessing AI systems in real-world scientific workflows.

Abstract: Multimodal Large Language Models (MLLMs) excel in general domains but struggle with complex, real-world science. We posit that polymer science, an interdisciplinary field spanning chemistry, physics, biology, and engineering, is an ideal high-stakes testbed due to its diverse multimodal data. Yet, existing benchmarks related to polymer science largely overlook real-world workflows, limiting their practical utility and failing to systematically evaluate MLLMs across the full, practice-grounded lifecycle of experimentation. We introduce PolyReal, a novel multimodal benchmark grounded in real-world scientific practices to evaluate MLLMs on the full lifecycle of polymer experimentation. It covers five critical capabilities: (1) foundational knowledge application; (2) lab safety analysis; (3) experiment mechanism reasoning; (4) raw data extraction; and (5) performance & application exploration. Our evaluation of leading MLLMs on PolyReal reveals a capability imbalance. While models perform well on knowledge-intensive reasoning (e.g., Experiment Mechanism Reasoning), they drop sharply on practice-based tasks (e.g., Lab Safety Analysis and Raw Data Extraction). This exposes a severe gap between abstract scientific knowledge and its practical, context-dependent application, showing that these real-world tasks remain challenging for MLLMs. Thus, PolyReal helps address this evaluation gap and provides a practical benchmark for assessing AI systems in real-world scientific workflows.

Method: Proposes MHENet with Texture Hierarchical Enhancement Module (THEM) to amplify subtle texture variations via high-frequency extraction, Geometry Hierarchical Enhancement Module (GHEM) to enhance geometric structures via learnable gradient extraction, and Adaptive Dynamic Fusion Module (ADFM) for spatially varying fusion of enhanced features.

Result: Experiments on four benchmarks show MHENet surpasses 16 state-of-the-art methods both qualitatively and quantitatively.

Conclusion: The proposed modality-specific hierarchical enhancement and adaptive fusion approach effectively improves RGB-D camouflaged object detection performance by better utilizing complementary texture and geometry cues.

Abstract: Camouflaged object detection (COD) is challenging due to high target-background similarity, and recent methods address this by complementarily using RGB-D texture and geometry cues. However, RGB-D COD methods still underutilize modality-specific cues, which limits fusion quality. We believe this is because RGB and depth features are fused directly after backbone extraction without modality-specific enhancement. To address this limitation, we propose MHENet, an RGB-D COD framework that performs modality-specific hierarchical enhancement and adaptive fusion of RGB and depth features. Specifically, we introduce a Texture Hierarchical Enhancement Module (THEM) to amplify subtle texture variations by extracting high-frequency information and a Geometry Hierarchical Enhancement Module (GHEM) to enhance geometric structures via learnable gradient extraction, while preserving cross-scale semantic consistency. Finally, an Adaptive Dynamic Fusion Module (ADFM) adaptively fuses the enhanced texture and geometry features with spatially varying weights. Experiments on four benchmarks demonstrate that MHENet surpasses 16 state-of-the-art methods qualitatively and quantitatively. Code is available at https://github.com/afdsgh/MHENet.

Method: Uses mesh parameterization and non-uniform differentiable sampling for continuous 3D face representation, residual graph convolutional networks with dual cross-attention for multi-modal feature fusion, and a lightweight regression network for vertex-wise geometric displacement prediction.

Result: Significant improvements over state-of-the-art methods, especially in synchronization accuracy of lip and eye movements, with accurate reconstruction of rich 3D facial motion details.

Conclusion: MMTalker effectively addresses the cross-modal mapping challenge in speech-driven 3D facial animation through innovative multi-resolution representation and multi-modal fusion techniques.

Abstract: Speech-driven three-dimensional (3D) facial animation synthesis aims to build a mapping from one-dimensional (1D) speech signals to time-varying 3D facial motion signals. Current methods still face challenges in maintaining lip-sync accuracy and producing realistic facial expressions, primarily due to the highly ill-posed nature of this cross-modal mapping. In this paper, we introduce a novel 3D audio-driven facial animation synthesis method through multi-resolution representation and multi-modal feature fusion, called MMTalker which can accurately reconstruct the rich details of 3D facial motion. We first achieve the continuous representation of 3D face with details by mesh parameterization and non-uniform differentiable sampling. The mesh parameterization technique establishes the correspondence between UV plane and 3D facial mesh and is used to offer ground truth for the continuous learning. Differentiable non-uniform sampling enables precise facial detail acquisition by setting learnable sampling probability in each triangular face. Next, we employ residual graph convolutional network and dual cross-attention mechanism to extract discriminative facial motion feature from multiple input modalities. This proposed multimodal fusion strategy takes full use of the hierarchical features of speech and the explicit spatiotemporal geometric features of facial mesh. Finally, a lightweight regression network predicts the vertex-wise geometric displacements of the synthesized talking face by jointly processing the sampled points in the canonical UV space and the encoded facial motion features. Comprehensive experiments demonstrate that significant improvements are achieved over state-of-the-art methods, especially in the synchronization accuracy of lip and eye movements.

Method: Proposes CrossWeaver with two key components: 1) Modality Interaction Block (MIB) for selective, reliability-aware cross-modal interaction within the encoder, and 2) lightweight Seam-Aligned Fusion (SAF) module for aggregating enhanced features. Designed for arbitrary-modality semantic segmentation.

Result: Achieves state-of-the-art performance on multiple multimodal semantic segmentation benchmarks with minimal additional parameters. Demonstrates strong generalization to unseen modality combinations.

Conclusion: CrossWeaver provides an effective solution for flexible multimodal fusion in semantic segmentation, addressing limitations of existing approaches through selective cross-modal interaction and seamless feature aggregation.

Abstract: Multimodal semantic segmentation has shown great potential in leveraging complementary information across diverse sensing modalities. However, existing approaches often rely on carefully designed fusion strategies that either use modality-specific adaptations or rely on loosely coupled interactions, thereby limiting flexibility and resulting in less effective cross-modal coordination. Moreover, these methods often struggle to balance efficient information exchange with preserving the unique characteristics of each modality across different modality combinations. To address these challenges, we propose CrossWeaver, a simple yet effective multimodal fusion framework for arbitrary-modality semantic segmentation. Its core is a Modality Interaction Block (MIB), which enables selective and reliability-aware cross-modal interaction within the encoder, while a lightweight Seam-Aligned Fusion (SAF) module further aggregates the enhanced features. Extensive experiments on multiple multimodal semantic segmentation benchmarks demonstrate that our framework achieves state-of-the-art performance with minimal additional parameters and strong generalization to unseen modality combinations.

Method: Proposes Collaborative Multi-Mode Pruning (CoMP) with: 1) Collaborative Importance Metric (CIM) that accounts for mutual interference between parameters and tokens, 2) Multi-Mode Pruning Strategy (MPS) that decomposes pruning into stages, adaptively shifts between parameter/token pruning based on cost, and integrates historical cost with random exploration.

Result: Extensive experiments across various vision-language tasks and models show CoMP effectively promotes performance under high pruning ratios compared to state-of-the-art approaches.

Conclusion: CoMP provides an effective framework for compressing VLMs by jointly pruning parameters and tokens, addressing limitations of single-mode pruning approaches and enabling better deployment on resource-constrained devices.

Abstract: Vision-Language Models (VLMs) have advanced rapidly within the unified Transformer architecture, yet their deployment on resource-constrained devices remains challenging due to high computational complexity. While pruning has emerged as an effective technique for compressing VLMs, existing approaches predominantly focus on a single mode by pruning either parameters or tokens, neglecting fully exploring the inherent redundancy in each mode, which leads to substantial performance degradation at high pruning ratios. To address the above limitations, we propose Collaborative Multi-Mode Pruning (CoMP), a novel framework tailored for VLMs by performing joint parameter and token pruning. Specifically, we first design a Collaborative Importance Metric (CIM) that investigates the mutual interference between the coupled parameters and tokens. It incorporates distinct significance of tokens into the computation of parameter importance scores, while simultaneously mitigating the affect of pruned parameters on token importance scores. Moreover, we develop a Multi-Mode Pruning Strategy (MPS) that decomposes the overall pruning process into a sequence of pruning stages, while in each stage we estimate the priory of different pruning modes based on their pruning cost and adaptively shift to the optimal one. Additionally, MPS integrates the historical cost and random exploration, in order to achieve a stable pruning process and avoid local optimum. Extensive experiments across various vision-language tasks and models demonstrate that our method effectively promotes the performance under high pruning ratios by comparing to the state-of-the-art approaches. The source code is available at https://github.com/Wuzimeng/CoMP.git.

Method: Proposes UAVGen with two key components: 1) Visual Prototype Conditioned Diffusion Model (VPC-DM) that constructs representative instances for each class and integrates them into latent embeddings, and 2) Focal Region Enhanced Data Pipeline (FRE-DP) that emphasizes object-concentrated foreground regions with label refinement to correct generation errors.

Result: Extensive experiments show UAVGen significantly outperforms state-of-the-art approaches and consistently promotes accuracy when integrated with distinct detectors.

Conclusion: UAVGen effectively addresses artifact issues in layout-to-image generation for UAV object detection, improving generation quality and detection performance in data-limited scenarios.

Abstract: Unmanned aerial vehicle (UAV) based object detection is a critical but challenging task, when applied in dynamically changing scenarios with limited annotated training data. Layout-to-image generation approaches have proved effective in promoting detection accuracy by synthesizing labeled images based on diffusion models. However, they suffer from frequently producing artifacts, especially near layout boundaries of tiny objects, thus substantially limiting their performance. To address these issues, we propose UAVGen, a novel layout-to-image generation framework tailored for UAV-based object detection. Specifically, UAVGen designs a Visual Prototype Conditioned Diffusion Model (VPC-DM) that constructs representative instances for each class and integrates them into latent embeddings for high-fidelity object generation. Moreover, a Focal Region Enhanced Data Pipeline (FRE-DP) is introduced to emphasize object-concentrated foreground regions in synthesis, combined with a label refinement to correct missing, extra and misaligned generations. Extensive experimental results demonstrate that our method significantly outperforms state-of-the-art approaches, and consistently promotes accuracy when integrated with distinct detectors. The source code is available at https://github.com/Sirius-Li/UAVGen.

Method: Proposes MLA-Gen framework integrating global motion priors with fine-grained local conditioning. Introduces SinkRatio metric to measure attention concentration and develops alignment-aware masking and control strategies to regulate attention during generation, addressing the attention sink problem.

Result: Extensive experiments show consistent improvements in both motion quality and motion-language alignment over strong baselines. The approach effectively mitigates attention sink issues and enhances semantic grounding.

Conclusion: The proposed MLA-Gen framework successfully addresses motion-language alignment challenges in text-to-motion generation by combining global priors with local conditioning and introducing novel attention regulation techniques, leading to improved generation quality and semantic alignment.

Abstract: Text-driven human motion generation aims to synthesize realistic motion sequences that follow textual descriptions. Despite recent advances, accurately aligning motion dynamics with textual semantics remains a fundamental challenge. In this paper, we revisit text-to-motion generation from the perspective of motion-language alignment and propose MLA-Gen, a framework that integrates global motion priors with fine-grained local conditioning. This design enables the model to capture common motion patterns, while establishing detailed alignment between texts and motions. Furthermore, we identify a previously overlooked attention sink phenomenon in human motion generation, where attention disproportionately concentrates on the start text token, limiting the utilization of informative textual cues and leading to degraded semantic grounding. To analyze this issue, we introduce SinkRatio, a metric for measuring attention concentration, and develop alignment-aware masking and control strategies to regulate attention during generation. Extensive experiments demonstrate that our approach consistently improves both motion quality and motion-language alignment over strong baselines. Code will be released upon acceptance.

Method: Trained a baseline UNet at multiple downsampling ratios across five fundus datasets (DRIVE, STARE, CHASE_DB1, HRF, FIVES) with native widths from 565 to 3504 pixels. Introduced width-stratified sensitivity metric evaluating thin (<3 pixels), medium (3-7 pixels), and thick (>7 pixels) vessels separately using native resolution width estimates from Euclidean distance transform.

Result: For high-resolution datasets (HRF, FIVES), thin vessel sensitivity improved monotonically as images were downsampled toward encoder’s effective operating range, peaking at processed widths between 256-876 pixels. For low-to-mid resolution datasets (DRIVE, STARE, CHASE_DB1), thin vessel sensitivity was highest at or near native resolution and degraded with downsampling. Aggressive downsampling reduced thin vessel sensitivity by up to 15.8 percentage points while Dice remained stable.

Conclusion: Dice score alone is insufficient for evaluating microvascular segmentation as it doesn’t capture thin vessel loss. The impact of resizing depends on dataset resolution, with different optimal processing strategies needed for high vs low-resolution datasets. Width-stratified evaluation provides better assessment of thin vessel detection.

Abstract: Most deep learning pipelines for retinal vessel segmentation resize fundus images to satisfy GPU memory constraints and enable uniform batch processing. However, the impact of this resizing on thin vessel detection remains underexplored. When high resolution images are downsampled, thin vessels are reduced to subpixel structures, causing irreversible information loss even before the data enters the network. Standard volumetric metrics such as the Dice score do not capture this loss because thick vessel pixels dominate the evaluation. We investigated this effect by training a baseline UNet at multiple downsampling ratios across five fundus datasets (DRIVE, STARE, CHASE_DB1, HRF, and FIVES) with native widths ranging from 565 to 3504 pixels, keeping all other settings fixed. We introduce a width-stratified sensitivity metric that evaluates thin (half-width <3 pixels), medium (3 to 7 pixels), and thick (>7 pixels) vessel detection separately, using native resolution width estimates derived from a Euclidean distance transform. Results show that for high-resolution datasets (HRF, FIVES), thin vessel sensitivity improves monotonically as images are downsampled toward the encoder’s effective operating range, peaking at processed widths between 256 and 876 pixels. For low-to-mid resolution datasets (DRIVE, STARE, CHASE_DB1), thin vessel sensitivity is highest at or near native resolution and degrades with any downsampling. Across all five datasets, aggressive downsampling reduced thin vessel sensitivity by up to 15.8 percentage points (DRIVE) while Dice remained relatively stable, confirming that Dice alone is insufficient for evaluating microvascular segmentation.

Method: Uses AI4Arctic/ASIP Sea Ice Dataset with 461 Sentinel-1 scenes matched with expert ice charts. Combines full-resolution Sentinel-1 Extra Wide inputs, leakage-aware stratified patch splitting, SIGRID-3 stage-of-development labels, and training-set normalization. Evaluates Vision Transformer baselines (ViT-Base and ViT-Large) trained with cross entropy, weighted cross-entropy, and focal loss.

Result: ViT-Large with focal loss achieves 69.6% held-out accuracy, 68.8% weighted F1, and 83.9% precision on the minority Multi-Year Ice class. Focal loss offers better precision-recall trade-off for rare ice classes compared to weighted cross-entropy.

Conclusion: The paper establishes a cleaner baseline for future multimodal fusion with optical, thermal, or meteorological data. Focal-loss training is more effective for rare ice classes in sea ice classification tasks.

Abstract: Accurate and automated sea ice classification is important for climate monitoring and maritime safety in the Arctic. While Synthetic Aperture Radar (SAR) is the operational standard because of its all-weather capability, it remains challenging to distinguish morphologically similar ice classes under severe class imbalance. Rather than claiming a fully validated multimodal system, this paper establishes a trustworthy SAR only baseline that future fusion work can build upon. Using the AI4Arctic/ASIP Sea Ice Dataset (v2), which contains 461 Sentinel-1 scenes matched with expert ice charts, we combine full-resolution Sentinel-1 Extra Wide inputs, leakage-aware stratified patch splitting, SIGRID-3 stage-of-development labels, and training-set normalization to evaluate Vision Transformer baselines. We compare ViT-Base models trained with cross entropy and weighted cross-entropy against a ViT-Large model trained with focal loss. Among the tested configurations, ViT-Large with focal loss achieves 69.6% held-out accuracy, 68.8% weighted F1, and 83.9% precision on the minority Multi-Year Ice class. These results show that focal-loss training offers a more useful precision-recall trade-off than weighted cross-entropy for rare ice classes and establishes a cleaner baseline for future multimodal fusion with optical, thermal, or meteorological data.

Method: SCOPE introduces a tri-modal scheduler with cache, predict, and recompute modes, where prediction uses noise-level Taylor extrapolation to fill the gap between reuse and recomputation. It also implements selective computation that restricts execution to the active frame interval only.

Result: On MAGI-1 and SkyReels-V2 datasets, SCOPE achieves up to 4.73x speedup while maintaining quality comparable to the original output, outperforming all training-free baselines.

Conclusion: SCOPE effectively addresses AR-specific inefficiencies in video diffusion models through its tri-modal scheduling and selective computation approach, providing significant acceleration without quality degradation.

Abstract: Autoregressive (AR) video diffusion models enable long-form video generation but remain expensive due to repeated multi-step denoising. Existing training-free acceleration methods rely on binary cache-or-recompute decisions, overlooking intermediate cases where direct reuse is too coarse yet full recomputation is unnecessary. Moreover, asynchronous AR schedules assign different noise levels to co-generated frames, yet existing methods process the entire valid interval uniformly. To address these AR-specific inefficiencies, we present SCOPE, a training-free framework for efficient AR video diffusion. SCOPE introduces a tri-modal scheduler over cache, predict, and recompute, where prediction via noise-level Taylor extrapolation fills the gap between reuse and recomputation with explicit stability controls backed by error propagation analysis. It further introduces selective computation that restricts execution to the active frame interval. On MAGI-1 and SkyReels-V2, SCOPE achieves up to 4.73x speedup while maintaining quality comparable to the original output, outperforming all training-free baselines.

Method: Proposes MM-GS built on 3D Gaussian Splatting with two key modules: 1) Per-Instance Multi-View Fusion to establish robust representations by aggregating visual information across views, and 2) Scene-Level Instance Interaction module that operates on a global scene graph to reason about relationships and refine attributes.

Result: Extensive experiments on challenging datasets show the method significantly outperforms strong baselines, producing state-of-the-art results with high-fidelity details and plausible inter-instance contacts.

Conclusion: MM-GS effectively addresses the MHMO rendering problem by combining per-instance view fusion with scene-level interaction modeling, enabling high-quality reconstruction of complex dynamic scenes from sparse views.

Abstract: Reconstructing dynamic scenes with multiple interacting humans and objects from sparse-view inputs is a critical yet challenging task, essential for creating high-fidelity digital twins for robotics and VR/AR. This problem, which we term Multi-Human Multi-Object (MHMO) rendering, presents two significant obstacles: achieving view-consistent representations for individual instances under severe mutual occlusion, and explicitly modeling the complex and combinatorial dependencies that arise from their interactions. To overcome these challenges, we propose MM-GS, a novel hierarchical framework built upon 3D Gaussian Splatting. Our method first employs a Per-Instance Multi-View Fusion module to establish a robust and consistent representation for each instance by aggregating visual information across all available views. Subsequently, a Scene-Level Instance Interaction module operates on a global scene graph to reason about relationships between all participants, refining their attributes to capture subtle interaction effects. Extensive experiments on challenging datasets demonstrate that our method significantly outperforms strong baselines, producing state-of-the-art results with high-fidelity details and plausible inter-instance contacts.

Method: Introduces VLM-UnBench benchmark covering 4 forgetting levels, 7 source datasets, 11 concept axes, with three-level probe taxonomy and five evaluation conditions to separate genuine forgetting from instruction compliance.

Result: Realistic unlearning prompts leave forget accuracy near baseline; meaningful reductions only under oracle conditions. Object/scene concepts most resistant to suppression, and stronger instruction-tuned models remain capable despite forget instructions.

Conclusion: There’s a clear gap between prompt-level suppression and true visual concept erasure in VLMs, highlighting limitations of current training-free unlearning approaches.

Abstract: VLMs trained on web-scale data retain sensitive and copyrighted visual concepts that deployment may require removing. Training-based unlearning methods share a structural flaw: fine-tuning on a narrow forget set degrades general capabilities before unlearning begins, making it impossible to attribute subsequent performance drops to the unlearning procedure itself. Training-free approaches sidestep this by suppressing concepts through prompts or system instructions, but no rigorous benchmark exists for evaluating them on visual tasks. We introduce VLM-UnBench, the first benchmark for training-free visual concept unlearning in VLMs. It covers four forgetting levels, 7 source datasets, and 11 concept axes, and pairs a three-level probe taxonomy with five evaluation conditions to separate genuine forgetting from instruction compliance. Across 8 evaluation settings and 13 VLM configurations, realistic unlearning prompts leave forget accuracy near the no-instruction baseline; meaningful reductions appear only under oracle conditions that disclose the target concept to the model. Object and scene concepts are the most resistant to suppression, and stronger instruction-tuned models remain capable despite explicit forget instructions. These results expose a clear gap between prompt-level suppression and true visual concept erasure.

Method: Proposes a Wavelet Feature Stream that transforms per-joint velocity sequences using continuous wavelet transform (CWT) into multi-scale scalograms, from which a lightweight multi-scale CNN learns discriminative dynamic cues. This descriptor is fused with backbone representations without changing the backbone architecture or requiring additional supervision.

Result: Consistent gains across strong skeleton backbones (GaitMixer, GaitFormer, GaitGraph) on CASIA-B dataset, establishing new skeleton-based state-of-the-art when attached to GaitMixer. Improvements are especially pronounced under covariate shifts like carrying bags (BG) and wearing coats (CL).

Conclusion: The wavelet feature stream effectively complements standard spatio-temporal encoders by providing explicit time-frequency modeling of motion dynamics, enhancing gait recognition robustness under appearance changes.

Abstract: Skeleton-based gait recognizers excel at modeling spatial configurations but often underuse explicit motion dynamics that are crucial under appearance changes. We introduce a plug-and-play Wavelet Feature Stream that augments any skeleton backbone with time-frequency dynamics of joint velocities. Concretely, per-joint velocity sequences are transformed by the continuous wavelet transform (CWT) into multi-scale scalograms, from which a lightweight multi-scale CNN learns discriminative dynamic cues. The resulting descriptor is fused with the backbone representation for classification, requiring no changes to the backbone architecture or additional supervision. Across CASIA-B, the proposed stream delivers consistent gains on strong skeleton backbones (e.g., GaitMixer, GaitFormer, GaitGraph) and establishes a new skeleton-based state of the art when attached to GaitMixer. The improvements are especially pronounced under covariate shifts such as carrying bags (BG) and wearing coats (CL), highlighting the complementarity of explicit time-frequency modeling and standard spatio-temporal encoders.

Method: Multi-stage pipeline with image restoration, dehazing, MLLM-based enhancement, 3DGS-MCMC optimization, and averaging over repeated runs to improve visibility while limiting scene-content changes.

Result: Achieved 1st place out of 14 participants in Track 2 of NTIRE 2026 3DRR Challenge, with improved quantitative performance and better visual quality than baselines.

Conclusion: The proposed pipeline effectively addresses smoke degradation in images for 3D reconstruction tasks, demonstrating state-of-the-art performance in the challenge benchmark.

Abstract: This paper describes our method for Track 2 of the NTIRE 2026 3D Restoration and Reconstruction (3DRR) Challenge on smoke-degraded images. In this task, smoke reduces image visibility and weakens the cross-view consistency required by scene optimization and rendering. We address this problem with a multi-stage pipeline consisting of image restoration, dehazing, MLLM-based enhancement, 3DGS-MCMC optimization, and averaging over repeated runs. The main purpose of the pipeline is to improve visibility before rendering while limiting scene-content changes across input views. Experimental results on the challenge benchmark show improved quantitative performance and better visual quality than the provided baselines. The code is available at https://github.com/plbbl/GenSmoke-GS. Our method achieved a ranking of 1 out of 14 participants in Track 2 of the NTIRE 3DRR Challenge, as reported on the official competition website: https://www.codabench.org/competitions/13993/#/results-tab.

Method: QVAD uses an LLM agent to iteratively refine queries based on visual context, guiding smaller VLMs through dynamic dialogue to produce high-fidelity captions and semantic reasoning without parameter updates.

Result: Achieves state-of-the-art performance on UCF-Crime, XD-Violence, and UBNormal using a fraction of parameters, with exceptional generalizability on ComplexVAD and high inference speeds suitable for edge devices.

Conclusion: Dynamic prompt-updating mechanisms can unlock latent capabilities of lightweight models, making advanced VAD deployable on resource-constrained devices while maintaining high performance.

Abstract: Video Anomaly Detection (VAD) is a fundamental challenge in computer vision, particularly due to the open-set nature of anomalies. While recent training-free approaches utilizing Vision-Language Models (VLMs) have shown promise, they typically rely on massive, resource-intensive foundation models to compensate for the ambiguity of static prompts. We argue that the bottleneck in VAD is not necessarily model capacity, but rather the static nature of inquiry. We propose QVAD, a question-centric agentic framework that treats VLM-LLM interaction as a dynamic dialogue. By iteratively refining queries based on visual context, our LLM agent guides smaller VLMs to produce high-fidelity captions and precise semantic reasoning without parameter updates. This ``prompt-updating" mechanism effectively unlocks the latent capabilities of lightweight models, enabling state-of-the-art performance on UCF-Crime, XD-Violence, and UBNormal using a fraction of the parameters required by competing methods. We further demonstrate exceptional generalizability on the single-scene ComplexVAD dataset. Crucially, QVAD achieves high inference speeds with minimal memory footprints, making advanced VAD capabilities deployable on resource-constrained edge devices.

Method: Built using community-standard distributions (PyTorch and MONAI), PR3DICTR employs modular design principles and standardization. It offers pre-established functionality for model architectures, hyper-parameters, and training methodologies while allowing users to “plug in” their own solutions.

Result: A flexible, open-access framework that can be applied to any binary or event-based 3D classification task with minimal code (as little as two lines), providing researchers with standardized tools for medical image analysis.

Conclusion: PR3DICTR successfully creates a standardized yet flexible platform for 3D medical image classification research that balances ease of use with customization capabilities, potentially accelerating developments in medical AI applications.

Abstract: Three-dimensional medical image data and computer-aided decision making, particularly using deep learning, are becoming increasingly important in the medical field. To aid in these developments we introduce PR3DICTR: Platform for Research in 3D Image Classification and sTandardised tRaining. Built using community-standard distributions (PyTorch and MONAI), PR3DICTR provides an open-access, flexible and convenient framework for prediction model development, with an explicit focus on classification using three-dimensional medical image data. By combining modular design principles and standardization, it aims to alleviate developmental burden whilst retaining adjustability. It provides users with a wealth of pre-established functionality, for instance in model architecture design options, hyper-parameter solutions and training methodologies, but still gives users the opportunity and freedom to ``plug in’’ their own solutions or modules. PR3DICTR can be applied to any binary or event-based three-dimensional classification task and can work with as little as two lines of code.

Method: Conducted systematic evaluation of Nano Banana 2 for image restoration across various scenes and degradation types, focusing on prompt design strategies including concise prompts with explicit fidelity constraints to balance reconstruction accuracy and perceptual quality.

Result: Nano Banana 2 achieves superior performance in full-reference metrics compared to state-of-the-art restoration models while remaining competitive in perceptual quality, with strong generalization in challenging scenarios like small faces, dense crowds, and severe degradations. However, the model is sensitive to prompt formulation and may require iterative refinement.

Conclusion: General-purpose generative models hold strong potential as unified image restoration solvers, but controllability and robustness remain important considerations, with prompt design playing a critical role in achieving optimal results.

Abstract: Recent advances in generative AI raise the question of whether general-purpose image editing models can serve as unified solutions for image restoration. In this work, we conduct a systematic evaluation of Nano Banana 2 for image restoration across diverse scenes and degradation types. Our results show that prompt design plays a critical role, where concise prompts with explicit fidelity constraints achieve the best trade-off between reconstruction accuracy and perceptual quality. Compared with state-of-the-art restoration models, Nano Banana 2 achieves superior performance in full-reference metrics while remaining competitive in perceptual quality, which is further supported by user studies. We also observe strong generalization in challenging scenarios, such as small faces, dense crowds, and severe degradations. However, the model remains sensitive to prompt formulation and may require iterative refinement for optimal results. Overall, our findings suggest that general-purpose generative models hold strong potential as unified image restoration solvers, while highlighting the importance of controllability and robustness. All test results are available on https://github.com/yxyuanxiao/NanoBanana2TestOnIR.

Method: Proposes Gram-MMD (GMMD) which computes Gram matrices from intermediate activations of pretrained backbone networks, extracts upper-triangular parts, and measures Maximum Mean Discrepancy between real and generated image distributions.

Result: GMMD captures complementary textural information to semantic metrics, correctly ranks real vs synthetic images in cross-domain scenarios where CMMD fails, and shows strong performance across various backbone architectures.

Conclusion: GMMD provides a valuable complement to existing image realism metrics by capturing fine-grained textural information that semantic-level metrics may miss, particularly useful for evaluating generative models.

Abstract: Evaluating the realism of generated images remains a fundamental challenge in generative modeling. Existing distributional metrics such as the Frechet Inception Distance (FID) and CLIP-MMD (CMMD) compare feature distributions at a semantic level but may overlook fine-grained textural information that can be relevant for distinguishing real from generated images. We introduce Gram-MMD (GMMD), a realism metric that leverages Gram matrices computed from intermediate activations of pretrained backbone networks to capture correlations between feature maps. By extracting the upper-triangular part of these symmetric Gram matrices and measuring the Maximum Mean Discrepancy (MMD) between an anchor distribution of real images and an evaluation distribution, GMMD produces a representation that encodes textural and structural characteristics at a finer granularity than global embeddings. To select the hyperparameters of the metric, we employ a meta-metric protocol based on controlled degradations applied to MS-COCO images, measuring monotonicity via Spearman’s rank correlation and Kendall’s tau. We conduct experiments on both the KADID-10k database and the RAISE realness assessment dataset using various backbone architectures, including DINOv2, DC-AE, Stable Diffusion’s VAE encoder, VGG19, and the AlexNet backbone from LPIPS, among others. We also demonstrate on a cross-domain driving scenario (KITTI / Virtual KITTI / Stanford Cars) that CMMD can incorrectly rank real images as less realistic than synthetic ones due to its semantic bias, while GMMD preserves the correct ordering. Our results suggest that GMMD captures complementary information to existing semantic-level metrics.

Method: Proposes entropy-based probabilistic sampling that generates large, sparse Gaussians in textureless areas and small, dense Gaussians in information-rich regions. Also designs a specialized point cloud network to efficiently encode local context and decode it into 3DGS attributes, addressing receptive field mismatch issues.

Result: Achieves state-of-the-art rendering quality with only 22% of Gaussians, and maintains reasonable rendering quality with only 1.5% of Gaussians. Demonstrates superior efficiency and compactness compared to previous methods.

Conclusion: SparseSplat successfully creates highly compact 3DGS maps through adaptive density adjustment and specialized network design, enabling efficient 3D scene representation suitable for downstream tasks.

Abstract: Recent progress in feed-forward 3D Gaussian Splatting (3DGS) has notably improved rendering quality. However, the spatially uniform and highly redundant 3DGS map generated by previous feed-forward 3DGS methods limits their integration into downstream reconstruction tasks. We propose SparseSplat, the first feed-forward 3DGS model that adaptively adjusts Gaussian density according to scene structure and information richness of local regions, yielding highly compact 3DGS maps. To achieve this, we propose entropy-based probabilistic sampling, generating large, sparse Gaussians in textureless areas and assigning small, dense Gaussians to regions with rich information. Additionally, we designed a specialized point cloud network that efficiently encodes local context and decodes it into 3DGS attributes, addressing the receptive field mismatch between the general 3DGS optimization pipeline and feed-forward models. Extensive experimental results demonstrate that SparseSplat can achieve state-of-the-art rendering quality with only 22% of the Gaussians and maintain reasonable rendering quality with only 1.5% of the Gaussians. Project page: https://victkk.github.io/SparseSplat-page/.

Method: Proposes MI-Pruner that computes Mutual Information between visual and textual features themselves prior to their interaction. This non-intrusive approach requires no access to internal attention maps or architectural modifications, directly measuring crossmodal dependency at feature levels.

Result: Experimental results show MI-Pruner outperforms previous attention-based pruning methods with minimal latency, demonstrating superior performance in visual token selection for efficient MLLM inference.

Conclusion: Direct measurement of mutual information between visual and textual features provides a more effective approach for visual token pruning in MLLMs compared to attention-based methods, offering better performance with minimal computational overhead.

Abstract: For multimodal large language models (MLLMs), visual information is relatively sparse compared with text. As a result, research on visual pruning emerges for efficient inference. Current approaches typically measure token importance based on the attention scores in the visual encoder or in the LLM decoder, then select visual tokens with high attention scores while pruning others. In this paper, we pursue a different and more surgical approach. Instead of relying on mechanism-specific signals, we directly compute Mutual Information (MI) between visual and textual features themselves, prior to their interaction. This allows us to explicitly measure crossmodal dependency at the feature levels. Our MI-Pruner is simple, efficient and non-intrusive, requiring no access to internal attention maps or architectural modifications. Experimental results demonstrate that our approach outperforms previous attention-based pruning methods with minimal latency.

Method: Proposed a scalable method for generating realistic, spatially inconsistent image pairs from multi-view scenes to systematically evaluate MLLMs’ ability to identify objects violating 3D motion consistency.

Result: State-of-the-art MLLMs significantly underperform human observers and show substantial variability across different scene attributes, revealing fragile and incomplete understanding of 3D structure.

Conclusion: Current MLLMs lack deep understanding of physical world geometry, highlighting the need for approaches that develop more grounded understanding of spatial consistency.

Abstract: Spatial consistency is a fundamental property of the visual world and a key requirement for models that aim to understand physical reality. Despite recent advances, multimodal large language models (MLLMs) often struggle to reason about 3D geometry across multiple views. Rather than asking models to describe scene attributes, we introduce a more challenging task: given two views of the same scene, identify the object that violates 3D motion consistency. We propose a simple and scalable method for generating realistic, spatially inconsistent image pairs from multi-view scenes, enabling systematic evaluation of this capability. Our results show that state-of-the-art MLLMs significantly underperform human observers and exhibit substantial variability across different scene attributes, revealing a fragile and incomplete understanding of 3D structure. We hope our findings underscore the need for approaches that develop a more deeply grounded understanding of the physical world.

Method: UCGP uses Curved-Grid Mesh parameterization for continuous, low-frequency patch generation, with unified representation-driven objectives for subspace departure and topology disruption, plus Meta Differential Evolution and EOT-augmented TPS deformation for real-world robustness.

Result: UCGP consistently compromises semantic understanding across diverse IR-VLM architectures while maintaining cross-model transferability, cross-dataset generalization, real-world physical effectiveness, and robustness against defenses.

Conclusion: The work reveals previously overlooked robustness vulnerabilities in current infrared multimodal systems, highlighting security concerns for IR-VLMs in practical deployments.

Abstract: Infrared vision-language models (IR-VLMs) have emerged as a promising paradigm for multimodal perception in low-visibility environments, yet their robustness to adversarial attacks remains largely unexplored. Existing adversarial patch methods are mainly designed for RGB-based models in closed-set settings and are not readily applicable to the open-ended semantic understanding and physical deployment requirements of infrared VLMs. To bridge this gap, we propose Universal Curved-Grid Patch (UCGP), a universal physical adversarial patch framework for IR-VLMs. UCGP integrates Curved-Grid Mesh (CGM) parameterization for continuous, low-frequency, and deployable patch generation with a unified representation-driven objective that promotes subspace departure, topology disruption, and stealth. To improve robustness under real-world deployment and domain shift, we further incorporate Meta Differential Evolution and EOT-augmented TPS deformation modeling. Rather than manipulating labels or prompts, UCGP directly disrupts the visual representation space, weakening cross-modal semantic alignment. Extensive experiments demonstrate that UCGP consistently compromises semantic understanding across diverse IR-VLM architectures while maintaining cross-model transferability, cross-dataset generalization, real-world physical effectiveness, and robustness against defenses. These findings reveal a previously overlooked robustness vulnerability in current infrared multimodal systems.

Method: Proposes SCC-Loc with three components: 1) Semantic-Guided Viewport Alignment to optimize satellite crop regions, 2) Cascaded Spatial-Adaptive Texture-Structure Filtering to enforce geometric consistency, and 3) Consensus-Driven Reliability-Aware Position Selection for optimal pose estimation. Uses shared DINOv2 backbone for global retrieval and matching.

Result: Achieves state-of-the-art performance with mean localization error of 9.37m and 7.6x accuracy improvement within 5m threshold. Introduces Thermal-UAV dataset with 11,890 thermal queries.

Conclusion: SCC-Loc effectively addresses thermal-visible modality gaps through semantic guidance and geometric consistency, establishing new SOTA for cross-modal thermal geo-localization with practical applications for UAV navigation.

Abstract: Cross-modal Thermal Geo-localization (TG) provides a robust, all-weather solution for Unmanned Aerial Vehicles (UAVs) in Global Navigation Satellite System (GNSS)-denied environments. However, profound thermal-visible modality gaps introduce severe feature ambiguity, systematically corrupting conventional coarse-to-fine registration. To dismantle this bottleneck, we propose SCC-Loc, a unified Semantic-Cascade-Consensus localization framework. By sharing a single DINOv2 backbone across global retrieval and MINIMA$_{\text{RoMa}}$ matching, it minimizes memory footprint and achieves zero-shot, highly accurate absolute position estimation. Specifically, we tackle modality ambiguity by introducing three cohesive components. First, we design the Semantic-Guided Viewport Alignment (SGVA) module to adaptively optimize satellite crop regions, effectively correcting initial spatial deviations. Second, we develop the Cascaded Spatial-Adaptive Texture-Structure Filtering (C-SATSF) mechanism to explicitly enforce geometric consistency, thereby eradicating dense cross-modal outliers. Finally, we propose the Consensus-Driven Reliability-Aware Position Selection (CD-RAPS) strategy to derive the optimal solution through a synergy of physically constrained pose optimization. To address data scarcity, we construct Thermal-UAV, a comprehensive dataset providing 11,890 diverse thermal queries referenced against a large-scale satellite ortho-photo and corresponding spatially aligned Digital Surface Model (DSM). Extensive experiments demonstrate that SCC-Loc establishes a new state-of-the-art, suppressing the mean localization error to 9.37 m and providing a 7.6-fold accuracy improvement within a strict 5-m threshold over the strongest baseline. Code and dataset are available at https://github.com/FloralHercules/SCC-Loc.

Method: Proposes random label bridge training as a modality adaptation learner that requires no manual labeling. The method adds a bridge training stage to align LLM parameters with vision tasks, with partial bridge training often being sufficient.

Result: Shows that bridge training can effectively adapt LLM parameters to vision foundation tasks. Surprisingly reveals that partial bridge training is often advantageous, as certain LLM layers exhibit strong foundational properties that remain beneficial without fine-tuning for visual tasks.

Conclusion: Opens new avenues for leveraging language pre-trained parameters directly in vision models and highlights partial bridge training as a practical pathway for cross-modality adaptation between language and vision.

Abstract: The ratio of outlier parameters in language pre-training models and vision pre-training models differs significantly, making cross-modality (language and vision) inherently more challenging than cross-domain adaptation. As a result, many prior studies have focused on cross-domain transfer rather than attempting to bridge language and vision modalities, assuming that language pre-trained models are unsuitable for downstream visual tasks due to disparate parameter spaces. Contrary to this assumption, we show that adding a bridge training stage as a modality adaptation learner can effectively align Large Language Model (LLM) parameters with vision tasks. Specifically, we propose a simple yet powerful solution random label bridge training that requires no manual labeling and helps LLM parameters adapt to vision foundation tasks. Moreover, our findings reveal that partial bridge training is often advantageous, as certain layers in LLMs exhibit strong foundational properties that remain beneficial even without fine-tuning for visual tasks. This surprising discovery opens up new avenues for leveraging language pre-trained parameters directly within vision models and highlights the potential of partial bridge training as a practical pathway to cross-modality adaptation.

Method: Adapts pre-trained Stable Diffusion with two key components: Support-Query Interaction (SQI) for adapting to FSMIS paradigm, and Visual-to-Textual Condition Translator (VTCT) that converts visual cues from support set into textual embeddings to guide diffusion process.

Result: Achieves competitive results compared to state-of-the-art methods in standard settings and demonstrates excellent generalization ability in challenging cross-domain scenarios.

Conclusion: Highlights the potential of adapting large-scale generative models for advancing data-efficient and robust medical image segmentation.

Abstract: Few-Shot Medical Image Segmentation (FSMIS) aims to segment novel object classes in medical images using only minimal annotated examples, addressing the critical challenges of data scarcity and domain shifts prevalent in medical imaging. While Diffusion Models (DM) excel in visual tasks, their potential for FSMIS remains largely unexplored. We propose that the rich visual priors learned by large-scale DMs offer a powerful foundation for a more robust and data-efficient segmentation approach. In this paper, we introduce SD-FSMIS, a novel framework designed to effectively adapt the powerful pre-trained Stable Diffusion (SD) model for the FSMIS task. Our approach repurposes its conditional generative architecture by introducing two key components: a Support-Query Interaction (SQI) and a Visual-to-Textual Condition Translator (VTCT). Specifically, SQI provides a straightforward yet powerful means of adapting SD to the FSMIS paradigm. The VTCT module translates visual cues from the support set into an implicit textual embedding that guides the diffusion model, enabling precise conditioning of the generation process. Extensive experiments demonstrate that SD-FSMIS achieves competitive results compared to state-of-the-art methods in standard settings. Surprisingly, it also demonstrated excellent generalization ability in more challenging cross-domain scenarios. These findings highlight the immense potential of adapting large-scale generative models to advance data-efficient and robust medical image segmentation.

Method: CAMEO decomposes editing into coordinated stages: planning, structured prompting, hypothesis generation, and adaptive reference grounding. It embeds evaluation directly within the editing loop, creating a closed-loop process where intermediate results are iteratively refined through structured feedback. External guidance is invoked only when task complexity requires it.

Result: CAMEO achieves 20% more win rate on average compared to multiple state-of-the-art models across anomaly insertion and human pose switching tasks. It demonstrates improved robustness, controllability, and structural reliability when evaluated with multiple strong editing backbones and independent evaluation models.

Conclusion: CAMEO successfully reformulates conditional image editing as a quality-aware, feedback-driven process rather than a one-shot generation task, providing better structural control and consistency for complex editing scenarios.

Abstract: Conditional image editing aims to modify a source image according to textual prompts and optional reference guidance. Such editing is crucial in scenarios requiring strict structural control (i.e., anomaly insertion in driving scenes and complex human pose transformation). Despite recent advances in large-scale editing models (i.e., Seedream, Nano Banana, etc), most approaches rely on single-step generation. This paradigm often lacks explicit quality control, may introduce excessive deviation from the original image, and frequently produces structural artifacts or environment-inconsistent modifications, typically requiring manual prompt tuning to achieve acceptable results. We propose \textbf{CAMEO}, a structured multi-agent framework that reformulates conditional editing as a quality-aware, feedback-driven process rather than a one-shot generation task. CAMEO decomposes editing into coordinated stages of planning, structured prompting, hypothesis generation, and adaptive reference grounding, where external guidance is invoked only when task complexity requires it. To overcome the lack of intrinsic quality control in existing methods, evaluation is embedded directly within the editing loop. Intermediate results are iteratively refined through structured feedback, forming a closed-loop process that progressively corrects structural and contextual inconsistencies. We evaluate CAMEO on anomaly insertion and human pose switching tasks. Across multiple strong editing backbones and independent evaluation models, CAMEO consistently achieves 20% more win rate on average compared to multiple state-of-the-art models, demonstrating improved robustness, controllability, and structural reliability in conditional image editing.

Method: Proposes EffiMiniVLM with EfficientNet-B0 image encoder and MiniLM text encoder, using weighted Huber loss that leverages rating counts to emphasize reliable samples, trained on only 20% of Amazon Reviews 2023 dataset.

Result: Achieves CES score of 0.40 with 27.7M parameters and 6.8 GFLOPs, being 4x-8x more resource-efficient than other top methods while maintaining competitive performance; scaling to 40% data outperforms larger models.

Conclusion: Demonstrates that compact multimodal models can achieve strong performance for cold-start product quality prediction without external datasets, offering high resource efficiency and scalability.

Abstract: Predicting product quality from multimodal item information is critical in cold-start scenarios, where user interaction history is unavailable and predictions must rely on images and textual metadata. However, existing vision-language models typically depend on large architectures and/or extensive external datasets, resulting in high computational cost. To address this, we propose EffiMiniVLM, a compact dual-encoder vision-language regression framework that integrates an EfficientNet-B0 image encoder and a MiniLM-based text encoder with a lightweight regression head. To improve training sample efficiency, we introduce a weighted Huber loss that leverages rating counts to emphasize more reliable samples, yielding consistent performance gains. Trained using only 20% of the Amazon Reviews 2023 dataset, the proposed model contains 27.7M parameters and requires 6.8 GFLOPs, yet achieves a CES score of 0.40 with the lowest resource cost in the benchmark. Despite its small size, it remains competitive with significantly larger models, achieving comparable performance while being approximately 4x to 8x more resource-efficient than other top-5 methods and being the only approach that does not use external datasets. Further analysis shows that scaling the data to 40% alone allows our model to overtake other methods, which use larger models and datasets, highlighting strong scalability despite the model’s compact design.

Method: Challenge-based evaluation where participants developed networks to optimize efficiency metrics while meeting PSNR targets on DIV2K_LSDIR datasets. 95 registered participants, 15 valid submissions.

Result: The challenge established state-of-the-art results for efficient single-image super-resolution, with successful solutions meeting the target PSNR of ~26.90 dB on validation and ~26.99 dB on test sets.

Conclusion: The challenge successfully benchmarked efficient super-resolution methods, pushing forward the field of computationally efficient image enhancement while maintaining quality standards.

Abstract: This paper reviews the NTIRE 2026 challenge on efficient single-image super-resolution with a focus on the proposed solutions and results. The aim of this challenge is to devise a network that reduces one or several aspects, such as runtime, parameters, and FLOPs, while maintaining PSNR of around 26.90 dB on the DIV2K_LSDIR_valid dataset, and 26.99 dB on the DIV2K_LSDIR_test dataset. The challenge had 95 registered participants, and 15 teams made valid submissions. They gauge the state-of-the-art results for efficient single-image super-resolution.

Method: ProtoFlow models class prototypes as trajectories and learns their evolution with an explicit temporal vector field. It jointly enforces low-curvature motion and inter-class separation to stabilize prototype geometry throughout incremental learning.

Result: Experiments on standard class- and domain-incremental remote sensing benchmarks show consistent gains over strong baselines, including up to 1.5-2.0 points improvement in mIoUall, together with reduced forgetting.

Conclusion: Explicitly modeling temporal prototype evolution is a practical and interpretable strategy for robust continual remote sensing segmentation.

Abstract: Remote sensing segmentation in real deployment is inherently continual: new semantic categories emerge, and acquisition conditions shift across seasons, cities, and sensors. Despite recent progress, many incremental approaches still treat training steps as isolated updates, which leaves representation drift and forgetting insufficiently controlled. We present ProtoFlow, a time-aware prototype dynamics framework that models class prototypes as trajectories and learns their evolution with an explicit temporal vector field. By jointly enforcing low-curvature motion and inter-class separation, ProtoFlow stabilizes prototype geometry throughout incremental learning. Experiments on standard class- and domain-incremental remote sensing benchmarks show consistent gains over strong baselines, including up to 1.5-2.0 points improvement in mIoUall, together with reduced forgetting. These results suggest that explicitly modeling temporal prototype evolution is a practical and interpretable strategy for robust continual remote sensing segmentation.

Method: VOSR extracts semantically rich features from LR input using a pretrained vision encoder as visual semantic guidance. It revisits classifier-free guidance, replacing the standard unconditional branch with restoration-oriented guidance that preserves weak LR anchors. The framework trains a multi-step model from scratch and distills it into a one-step model for efficient inference.

Result: VOSR requires less than one-tenth of the training cost of T2I-based SR methods, yet achieves competitive or better perceptual quality and efficiency. It produces more faithful structures with fewer hallucinations on both synthetic and real-world benchmarks.

Conclusion: High-quality generative super-resolution can be achieved without multimodal pretraining, challenging the current paradigm of relying on text-to-image diffusion models for SR tasks.

Abstract: Most of the recent generative image super-resolution (SR) methods rely on adapting large text-to-image (T2I) diffusion models pretrained on web-scale text-image data. While effective, this paradigm starts from a generic T2I generator, despite that SR is fundamentally a low-resolution (LR) input-conditioned image restoration task. In this work, we investigate whether an SR model trained purely on visual data can rival T2I-based ones. To this end, we propose VOSR, a Vision-Only generative framework for SR. We first extract semantically rich and spatially grounded features from the LR input using a pretrained vision encoder as visual semantic guidance. We then revisit classifier-free guidance for training generative models and show that the standard unconditional branch is ill-suited to restoration models trained from scratch. We therefore replace it with a restoration-oriented guidance strategy that preserves weak LR anchors. Built upon these designs, we first train a multi-step VOSR model from scratch and then distill it into a one-step model for efficient inference. VOSR requires less than one-tenth of the training cost of representative T2I-based SR methods, yet in both multi-step and one-step settings, it achieves competitive or even better perceptual quality and efficiency, while producing more faithful structures with fewer hallucinations on both synthetic and real-world benchmarks. Our results, for the first time, show that high-quality generative SR can be achieved without multimodal pretraining. The code and models can be found at https://github.com/cswry/VOSR.

Method: Proposes CoME-VL: Complementary Multi-Encoder Vision-Language framework with two key components: (1) entropy-guided multi-layer aggregation with orthogonality-constrained projections to reduce redundancy, and (2) RoPE-enhanced cross-attention to align heterogeneous token grids and produce compact fused visual tokens. The fused tokens can be injected into decoder-only LLMs with minimal changes to standard VLM pipelines.

Result: Extensive experiments show CoME-VL consistently outperforms single-encoder baselines: average improvement of 4.9% on visual understanding tasks and 5.4% on grounding tasks. Achieves state-of-the-art performance on RefCOCO for detection while improving over baseline by large margin.

Conclusion: Fusing complementary contrastive and self-supervised visual representations significantly improves VLM performance across diverse benchmarks. The modular framework demonstrates the value of integrating different visual encoding strategies for enhanced vision-language understanding.

Abstract: Recent vision-language models (VLMs) typically rely on a single vision encoder trained with contrastive image-text objectives, such as CLIP-style pretraining. While contrastive encoders are effective for cross-modal alignment and retrieval, self-supervised visual encoders often capture richer dense semantics and exhibit stronger robustness on recognition and understanding tasks. In this work, we investigate how to scale the fusion of these complementary visual representations for vision-language modeling. We propose CoME-VL: Complementary Multi-Encoder Vision-Language, a modular fusion framework that integrates a contrastively trained vision encoder with a self-supervised DINO encoder. Our approach performs representation-level fusion by (i) entropy-guided multi-layer aggregation with orthogonality-constrained projections to reduce redundancy, and (ii) RoPE-enhanced cross-attention to align heterogeneous token grids and produce compact fused visual tokens. The fused tokens can be injected into a decoder-only LLM with minimal changes to standard VLM pipelines. Extensive experiments across diverse vision-language benchmarks demonstrate that CoME-VL consistently outperforms single-encoder baselines. In particular, we observe an average improvement of 4.9% on visual understanding tasks and 5.4% on grounding tasks. Our method achieves state-of-the-art performance on RefCOCO for detection while improving over the baseline by a large margin. Finally, we conduct ablation studies on layer merging, non-redundant feature mixing, and fusion capacity to evaluate how complementary contrastive and self-supervised signals affect VLM performance.

Method: Proposes ForgeryGPT with Mask-Aware Forgery Extractor (Forgery Localization Expert with Object-agnostic Forgery Prompt and Vocabulary-enhanced Vision Encoder + Mask Encoder) to capture high-order forensics knowledge from linguistic feature spaces. Uses three-stage training with Mask-Text Alignment and IFDL Task-Specific Instruction Tuning datasets.

Result: Extensive experiments demonstrate the effectiveness of the proposed method for IFDL with explainable generation and interactive dialogue capabilities.

Conclusion: ForgeryGPT advances IFDL by capturing high-order forensics knowledge correlations while enabling explainable generation and interactive dialogue through customized LLM architecture.

Abstract: Multimodal Large Language Models (MLLMs), such as GPT4o, have shown strong capabilities in visual reasoning and explanation generation. However, despite these strengths, they face significant challenges in the increasingly critical task of Image Forgery Detection and Localization (IFDL). Moreover, existing IFDL methods are typically limited to the learning of low-level semantic-agnostic clues and merely provide a single outcome judgment. To tackle these issues, we propose ForgeryGPT, a novel framework that advances the IFDL task by capturing high-order forensics knowledge correlations of forged images from diverse linguistic feature spaces, while enabling explainable generation and interactive dialogue through a newly customized Large Language Model (LLM) architecture. Specifically, ForgeryGPT enhances traditional LLMs by integrating the Mask-Aware Forgery Extractor, which enables the excavating of precise forgery mask information from input images and facilitating pixel-level understanding of tampering artifacts. The Mask-Aware Forgery Extractor consists of a Forgery Localization Expert (FL-Expert) and a Mask Encoder, where the FL-Expert is augmented with an Object-agnostic Forgery Prompt and a Vocabulary-enhanced Vision Encoder, allowing for effectively capturing of multi-scale fine-grained forgery details. To enhance its performance, we implement a three-stage training strategy, supported by our designed Mask-Text Alignment and IFDL Task-Specific Instruction Tuning datasets, which align vision-language modalities and improve forgery detection and instruction-following capabilities. Extensive experiments demonstrate the effectiveness of the proposed method.

Method: Unable to determine method due to access error

Result: Unable to determine results due to access error

Conclusion: Unable to draw conclusions due to access error

Abstract: Failed to fetch summary for 2502.09018: Page request resulted in HTTP 429 (https://export.arxiv.org/api/query?search_query=&id_list=2502.09018&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 2407.16341: Page request resulted in HTTP 429 (https://export.arxiv.org/api/query?search_query=&id_list=2407.16341&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 2504.17180: Page request resulted in HTTP 429 (https://export.arxiv.org/api/query?search_query=&id_list=2504.17180&sortBy=relevance&sortOrder=descending&start=0&max_results=100)

Method: Unable to determine method due to missing abstract content

Result: Unable to determine results due to missing abstract content

Conclusion: Unable to determine conclusion due to missing abstract content

Abstract: Failed to fetch summary for 2408.12406: Page request resulted in HTTP 429 (https://export.arxiv.org/api/query?search_query=&id_list=2408.12406&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 2408.12974: Page request resulted in HTTP 429 (https://export.arxiv.org/api/query?search_query=&id_list=2408.12974&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 2506.03198: Page request resulted in HTTP 429 (https://export.arxiv.org/api/query?search_query=&id_list=2506.03198&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 determine conclusion due to failed API request

Abstract: Failed to fetch summary for 2503.09158: Page request resulted in HTTP 429 (https://export.arxiv.org/api/query?search_query=&id_list=2503.09158&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 2505.15263: Page request resulted in HTTP 429 (https://export.arxiv.org/api/query?search_query=&id_list=2505.15263&sortBy=relevance&sortOrder=descending&start=0&max_results=100)

Method: N/A - Paper content not accessible

Result: N/A - Paper content not accessible

Conclusion: N/A - Paper content not accessible

Abstract: Failed to fetch summary for 2507.21584: Page request resulted in HTTP 429 (https://export.arxiv.org/api/query?search_query=&id_list=2507.21584&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 2507.22264: Page request resulted in HTTP 429 (https://export.arxiv.org/api/query?search_query=&id_list=2507.22264&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 2509.04276: Page request resulted in HTTP 429 (https://export.arxiv.org/api/query?search_query=&id_list=2509.04276&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 2509.16806: Page request resulted in HTTP 429 (https://export.arxiv.org/api/query?search_query=&id_list=2509.16806&sortBy=relevance&sortOrder=descending&start=0&max_results=100)

Method: Method information unavailable due to API rate limiting error

Result: Results cannot be assessed without paper content

Conclusion: Cannot draw conclusions about paper without access to abstract

Abstract: Failed to fetch summary for 2509.23880: Page request resulted in HTTP 429 (https://export.arxiv.org/api/query?search_query=&id_list=2509.23880&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 2511.08666: Page request resulted in HTTP 429 (https://export.arxiv.org/api/query?search_query=&id_list=2511.08666&sortBy=relevance&sortOrder=descending&start=0&max_results=100)

Method: Unable to determine method due to access limitations

Result: Unable to determine results due to access limitations

Conclusion: Unable to draw conclusions due to access limitations

Abstract: Failed to fetch summary for 2511.12834: Page request resulted in HTTP 429 (https://export.arxiv.org/api/query?search_query=&id_list=2511.12834&sortBy=relevance&sortOrder=descending&start=0&max_results=100)

Method: Proposes a Global Gaussian Map as persistent differentiable memory, incorporates local geometric reconstruction via submap-level global alignment, and leverages global map consistency to refine local geometry.

Result: Achieves state-of-the-art performance: improves pose accuracy by over 10%, produces finer and more detailed geometry, maintains compact memory-efficient global representation on real-world datasets.

Conclusion: SING3R-SLAM enables efficient and versatile 3D mapping for multiple downstream applications by addressing global consistency challenges in monocular indoor SLAM.

Abstract: Recent advances in dense 3D reconstruction have demonstrated strong capability in accurately capturing local geometry. However, extending these methods to incremental global reconstruction, as required in SLAM systems, remains challenging. Without explicit modeling of global geometric consistency, existing approaches often suffer from accumulated drift, scale inconsistency, and suboptimal local geometry. To address these issues, we propose SING3R-SLAM, a globally consistent Gaussian-based monocular indoor SLAM framework. Our approach represents the scene with a Global Gaussian Map that serves as a persistent, differentiable memory, incorporates local geometric reconstruction via submap-level global alignment, and leverages global map’s consistency to further refine local geometry. This design enables efficient and versatile 3D mapping for multiple downstream applications. Extensive experiments show that SING3R-SLAM achieves state-of-the-art performance in pose estimation, 3D reconstruction, and novel view rendering. It improves pose accuracy by over 10%, produces finer and more detailed geometry, and maintains a compact and memory-efficient global representation on real-world datasets.

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 2511.17722: Page request resulted in HTTP 429 (https://export.arxiv.org/api/query?search_query=&id_list=2511.17722&sortBy=relevance&sortOrder=descending&start=0&max_results=100)

Method: Cannot determine method as abstract content is unavailable

Result: Cannot determine results as abstract content is unavailable

Conclusion: Cannot draw conclusion as abstract content is unavailable

Abstract: Failed to fetch summary for 2511.21331: Page request resulted in HTTP 429 (https://export.arxiv.org/api/query?search_query=&id_list=2511.21331&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 2511.23292: Page request resulted in HTTP 429 (https://export.arxiv.org/api/query?search_query=&id_list=2511.23292&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 2512.00129: Page request resulted in HTTP 429 (https://export.arxiv.org/api/query?search_query=&id_list=2512.00129&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 2512.03424: Page request resulted in HTTP 429 (https://export.arxiv.org/api/query?search_query=&id_list=2512.03424&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 2512.08980: Page request resulted in HTTP 429 (https://export.arxiv.org/api/query?search_query=&id_list=2512.08980&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 conclusion as paper content could not be retrieved

Abstract: Failed to fetch summary for 2512.07951: Page request resulted in HTTP 429 (https://export.arxiv.org/api/query?search_query=&id_list=2512.07951&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 2512.13122: Page request resulted in HTTP 429 (https://export.arxiv.org/api/query?search_query=&id_list=2512.13122&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 2512.09112: Page request resulted in HTTP 429 (https://export.arxiv.org/api/query?search_query=&id_list=2512.09112&sortBy=relevance&sortOrder=descending&start=0&max_results=100)

Method: Cannot determine method due to failed API request

Result: Cannot determine results due to failed API request

Conclusion: Cannot draw conclusions due to failed API request

Abstract: Failed to fetch summary for 2601.03127: Page request resulted in HTTP 429 (https://export.arxiv.org/api/query?search_query=&id_list=2601.03127&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 2601.13633: Page request resulted in HTTP 429 (https://export.arxiv.org/api/query?search_query=&id_list=2601.13633&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 2601.06162: Page request resulted in HTTP 429 (https://export.arxiv.org/api/query?search_query=&id_list=2601.06162&sortBy=relevance&sortOrder=descending&start=0&max_results=100)

Method: Unable to determine method due to API access issues

Result: Unable to determine results due to API access issues

Conclusion: Paper analysis not possible due to technical limitations in accessing the content

Abstract: Failed to fetch summary for 2601.16672: Page request resulted in HTTP 429 (https://export.arxiv.org/api/query?search_query=&id_list=2601.16672&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 2601.16933: Page request resulted in HTTP 429 (https://export.arxiv.org/api/query?search_query=&id_list=2601.16933&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 2601.21957: Page request resulted in HTTP 429 (https://export.arxiv.org/api/query?search_query=&id_list=2601.21957&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: Paper analysis not possible due to technical limitations in accessing the content

Abstract: Failed to fetch summary for 2602.00683: Page request resulted in HTTP 429 (https://export.arxiv.org/api/query?search_query=&id_list=2602.00683&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 2602.06343: Page request resulted in HTTP 429 (https://export.arxiv.org/api/query?search_query=&id_list=2602.06343&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 2602.10516: Page request resulted in HTTP 429 (https://export.arxiv.org/api/query?search_query=&id_list=2602.10516&sortBy=relevance&sortOrder=descending&start=0&max_results=100)

Method: Cannot determine method as abstract/summary is unavailable due to HTTP 429 error from arXiv API

Result: Cannot determine results as abstract/summary is unavailable due to HTTP 429 error from arXiv API

Conclusion: Cannot determine conclusion as abstract/summary is unavailable due to HTTP 429 error from arXiv API

Abstract: Failed to fetch summary for 2602.13889: Page request resulted in HTTP 429 (https://export.arxiv.org/api/query?search_query=&id_list=2602.13889&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 2603.01765: Page request resulted in HTTP 429 (https://export.arxiv.org/api/query?search_query=&id_list=2603.01765&sortBy=relevance&sortOrder=descending&start=0&max_results=100)

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

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

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

Abstract: Failed to fetch summary for 2603.12711: Page request resulted in HTTP 429 (https://export.arxiv.org/api/query?search_query=&id_list=2603.12711&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 2603.18545: Page request resulted in HTTP 429 (https://export.arxiv.org/api/query?search_query=&id_list=2603.18545&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 2603.17069: Page request resulted in HTTP 429 (https://export.arxiv.org/api/query?search_query=&id_list=2603.17069&sortBy=relevance&sortOrder=descending&start=0&max_results=100)

Method: Unable to determine method due to fetch error

Result: Unable to determine results due to fetch error

Conclusion: Unable to determine conclusion due to fetch error

Abstract: Failed to fetch summary for 2603.20554: Page request resulted in HTTP 429 (https://export.arxiv.org/api/query?search_query=&id_list=2603.20554&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 2603.24326: Page request resulted in HTTP 429 (https://export.arxiv.org/api/query?search_query=&id_list=2603.24326&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 2603.23766: Page request resulted in HTTP 429 (https://export.arxiv.org/api/query?search_query=&id_list=2603.23766&sortBy=relevance&sortOrder=descending&start=0&max_results=100)

Method: MuRF processes images at multiple resolutions through a frozen VFM, then fuses the resulting features to create a unified representation. It’s architecture-agnostic, training-free, and works with various VFM families including DINOv2 and SigLIP2.

Result: Empirically validated across multiple computer vision tasks and VFM families, showing universal effectiveness as a fundamental enhancement to visual representation without requiring additional training.

Conclusion: MuRF provides a simple yet effective way to leverage multi-resolution synergy in vision foundation models, offering a training-free enhancement that works across different architectures and tasks.

Abstract: Vision Foundation Models (VFMs) have become the cornerstone of modern computer vision, offering robust representations across a wide array of tasks. While recent advances allow these models to handle varying input sizes during training, inference typically remains restricted to a single, fixed scale. This prevalent single-scale paradigm overlooks a fundamental property of visual perception: varying resolutions offer complementary inductive biases, where low-resolution views excel at global semantic recognition and high-resolution views are essential for fine-grained refinement. In this work, we propose Multi-Resolution Fusion (MuRF), a simple yet universally effective strategy to harness this synergy at inference time. Instead of relying on a single view, MuRF constructs a unified representation by processing an image at multiple resolutions through a frozen VFM and fusing the resulting features. The universality of MuRF is its most compelling attribute. It is not tied to a specific architecture, serving instead as a fundamental, training-free enhancement to visual representation. We empirically validate this by applying MuRF to a broad spectrum of critical computer vision tasks across multiple distinct VFM families - primarily DINOv2, but also demonstrating successful generalization to contrastive models like SigLIP2.

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 2603.26584: Page request resulted in HTTP 429 (https://export.arxiv.org/api/query?search_query=&id_list=2603.26584&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 draw conclusions due to fetch failure.

Abstract: Failed to fetch summary for 2604.01447: Page request resulted in HTTP 429 (https://export.arxiv.org/api/query?search_query=&id_list=2604.01447&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 2604.01479: Page request resulted in HTTP 429 (https://export.arxiv.org/api/query?search_query=&id_list=2604.01479&sortBy=relevance&sortOrder=descending&start=0&max_results=100)

Method: Unable to determine method due to technical error fetching paper information

Result: Unable to determine results due to technical error fetching paper information

Conclusion: Unable to draw conclusions due to technical error fetching paper information

Abstract: Failed to fetch summary for 2604.01989: Page request resulted in HTTP 429 (https://export.arxiv.org/api/query?search_query=&id_list=2604.01989&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 2604.01581: Page request resulted in HTTP 429 (https://export.arxiv.org/api/query?search_query=&id_list=2604.01581&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 access issues

Abstract: Failed to fetch summary for 2604.01848: Page request resulted in HTTP 429 (https://export.arxiv.org/api/query?search_query=&id_list=2604.01848&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 2604.01903: Page request resulted in HTTP 429 (https://export.arxiv.org/api/query?search_query=&id_list=2604.01903&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 2604.01972: Page request resulted in HTTP 429 (https://export.arxiv.org/api/query?search_query=&id_list=2604.01972&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 2509.11485: Page request resulted in HTTP 429 (https://export.arxiv.org/api/query?search_query=&id_list=2509.11485&sortBy=relevance&sortOrder=descending&start=0&max_results=100)

Method: Holos adopts a five-layer architecture with core modules: the Nuwa engine for high-efficiency agent generation and hosting, a market-driven Orchestrator for resilient coordination, and an endogenous value cycle to achieve incentive compatibility, bridging micro-level collaboration with macro-scale emergence.

Result: The authors have publicly released Holos (accessible at holosai.io) as a resource for the community and a testbed for future research in large-scale agentic ecosystems, providing a foundation for self-organizing and continuously evolving Agentic Web systems.

Conclusion: Holos represents an architectural approach to address fundamental challenges in large-scale multi-agent systems, aiming to lay the foundation for the next generation of self-organizing and continuously evolving Agentic Web ecosystems that can persist long-term.

Abstract: As large language models (LLM)-driven agents transition from isolated task solvers to persistent digital entities, the emergence of the Agentic Web, an ecosystem where heterogeneous agents autonomously interact and co-evolve, marks a pivotal shift toward Artificial General Intelligence (AGI). However, LLM-based multi-agent systems (LaMAS) are hindered by open-world issues such as scaling friction, coordination breakdown, and value dissipation. To address these challenges, we introduce Holos, a web-scale LaMAS architected for long-term ecological persistence. Holos adopts a five-layer architecture, with core modules primarily featuring the Nuwa engine for high-efficiency agent generation and hosting, a market-driven Orchestrator for resilient coordination, and an endogenous value cycle to achieve incentive compatibility. By bridging the gap between micro-level collaboration and macro-scale emergence, Holos hopes to lay the foundation for the next generation of the self-organizing and continuously evolving Agentic Web. We have publicly released Holos (accessible at https://holosai.io), providing a resource for the community and a testbed for future research in large-scale agentic ecosystems.

Method: Created XpertBench with 1,346 meticulously curated tasks across 80 categories spanning finance, healthcare, legal services, education, and dual-track research (STEM and Humanities). Tasks derived from over 1,000 submissions by domain experts. Introduced ShotJudge, a novel evaluation paradigm using LLM judges calibrated with expert few-shot exemplars to mitigate self-rewarding biases.

Result: State-of-the-art LLMs reveal a pronounced performance ceiling: even leading models achieve peak success rate of only ~66%, with mean score around 55%. Models exhibit domain-specific divergence, showing non-overlapping strengths in quantitative reasoning versus linguistic synthesis.

Conclusion: Findings underscore a significant “expert-gap” in current AI systems and establish XpertBench as a critical instrument for navigating the transition from general-purpose assistants to specialized professional collaborators.

Abstract: As Large Language Models (LLMs) exhibit plateauing performance on conventional benchmarks, a pivotal challenge persists: evaluating their proficiency in complex, open-ended tasks characterizing genuine expert-level cognition. Existing frameworks suffer from narrow domain coverage, reliance on generalist tasks, or self-evaluation biases. To bridge this gap, we present XpertBench, a high-fidelity benchmark engineered to assess LLMs across authentic professional domains. XpertBench consists of 1,346 meticulously curated tasks across 80 categories, spanning finance, healthcare, legal services, education, and dual-track research (STEM and Humanities). These tasks are derived from over 1,000 submissions by domain experts–including researchers from elite institutions and practitioners with extensive clinical or industrial experience–ensuring superior ecological validity. Each task uses detailed rubrics with mostly 15-40 weighted checkpoints to assess professional rigor. To facilitate scalable yet human-aligned assessment, we introduce ShotJudge, a novel evaluation paradigm that employs LLM judges calibrated with expert few-shot exemplars to mitigate self-rewarding biases. Our empirical evaluation of state-of-the-art LLMs reveals a pronounced performance ceiling: even leading models achieve a peak success rate of only ~66%, with a mean score around 55%. Models also exhibit domain-specific divergence, showing non-overlapping strengths in quantitative reasoning versus linguistic synthesis.. These findings underscore a significant “expert-gap” in current AI systems and establish XpertBench as a critical instrument for navigating the transition from general-purpose assistants to specialized professional collaborators.

Method: Proposes a neuro-symbolic architecture that: 1) extracts object-level structure from grids, 2) uses neural priors to propose candidate transformations from a fixed domain-specific language (DSL) of atomic patterns, and 3) filters hypotheses using cross-example consistency. Combines LLMs with object representations and transformation proposals.

Result: Improves base LLM performance from 16% to 24.4% on ARC-AGI-2 public evaluation set, and to 30.8% when combined with ARC Lang Solver via meta-classifier. Demonstrates improved generalization without task-specific finetuning or reinforcement learning.

Conclusion: Separating perception, neural-guided transformation proposal, and symbolic consistency filtering improves generalization on visual reasoning tasks while reducing reliance on brute-force search and sampling-based test-time scaling.

Abstract: We study structured abstraction-based reasoning for the Abstraction and Reasoning Corpus (ARC) and compare its generalization to test-time approaches. Purely neural architectures lack reliable combinatorial generalization, while strictly symbolic systems struggle with perceptual grounding. We therefore propose a neuro-symbolic architecture that extracts object-level structure from grids, uses neural priors to propose candidate transformations from a fixed domain-specific language (DSL) of atomic patterns, and filters hypotheses using cross-example consistency. Instantiated as a compositional reasoning framework based on unit patterns inspired by human visual abstraction, the system augments large language models (LLMs) with object representations and transformation proposals. On ARC-AGI-2, it improves base LLM performance from 16% to 24.4% on the public evaluation set, and to 30.8% when combined with ARC Lang Solver via a meta-classifier. These results demonstrate that separating perception, neural-guided transformation proposal, and symbolic consistency filtering improves generalization without task-specific finetuning or reinforcement learning, while reducing reliance on brute-force search and sampling-based test-time scaling. We open-source the ARC-AGI-2 Reasoner code (https://github.com/CoreThink-AI/arc-agi-2-reasoner).

Method: Theoretical analysis of threshold functions and their geometric interpretation as hyperplanes, examining behavior in low vs high dimensions, and reinterpreting multilayer architectures as sequential data manifold deformation.

Result: In low dimensions, perceptrons act as logical classifiers; in high dimensions, they become navigational indexical indicators. Depth prepares data manifolds for the separability already enabled by high-dimensional geometry.

Conclusion: A triadic framework: threshold function as ontological unit, dimensionality as enabling condition, and depth as preparatory mechanism provides unified mathematical understanding of generative AI.

Abstract: This paper examines the role of threshold logic in understanding generative artificial intelligence. Threshold functions, originally studied in the 1960s in digital circuit synthesis, provide a structurally transparent model of neural computation: a weighted sum of inputs compared to a threshold, geometrically realized as a hyperplane partitioning a space. The paper shows that this operation undergoes a qualitative transition as dimensionality increases. In low dimensions, the perceptron acts as a determinate logical classifier, separating classes when possible, as decided by linear programming. In high dimensions, however, a single hyperplane can separate almost any configuration of points (Cover, 1965); the space becomes saturated with potential classifiers, and the perceptron shifts from a logical device to a navigational one, functioning as an indexical indicator in the sense of Peirce. The limitations of the perceptron identified by Minsky and Papert (1969) were historically addressed by introducing multilayer architectures. This paper considers an alternative path: increasing dimensionality while retaining a single threshold element. It argues that this shift has equally significant implications for understanding neural computation. The role of depth is reinterpreted as a mechanism for the sequential deformation of data manifolds through iterated threshold operations, preparing them for linear separability already afforded by high-dimensional geometry. The resulting triadic account - threshold function as ontological unit, dimensionality as enabling condition, and depth as preparatory mechanism - provides a unified perspective on generative AI grounded in established mathematics.

Method: Proposes Agent-Integrated Verification and Validation (AIVV) - a hybrid framework deploying LLMs as a deliberative outer loop. Uses a role-specialized LLM council to collaboratively validate anomalies by semantically analyzing natural-language requirements, then performs system verification by assessing post-fault responses against operational tolerances.

Result: Experiments on a time-series simulator for Unmanned Underwater Vehicles demonstrate that AIVV successfully digitizes the HITL V&V process, overcoming limitations of rule-based fault classification and offering scalable LLM-mediated oversight for time-series data domains.

Conclusion: AIVV provides a scalable blueprint for automating verification and validation in control systems using LLMs, reducing manual workload while maintaining rigorous system oversight through collaborative agent-based analysis of natural-language requirements.

Abstract: Deep learning models excel at detecting anomaly patterns in normal data. However, they do not provide a direct solution for anomaly classification and scalability across diverse control systems, frequently failing to distinguish genuine faults from nuisance faults caused by noise or the control system’s large transient response. Consequently, because algorithmic fault validation remains unscalable, full Verification and Validation (V&V) operations are still managed by Human-in-the-Loop (HITL) analysis, resulting in an unsustainable manual workload. To automate this essential oversight, we propose Agent-Integrated Verification and Validation (AIVV), a hybrid framework that deploys Large Language Models (LLMs) as a deliberative outer loop. Because rigorous system verification strictly depends on accurate validation, AIVV escalates mathematically flagged anomalies to a role-specialized LLM council. The council agents perform collaborative validation by semantically validating nuisance and true failures based on natural-language (NL) requirements to secure a high-fidelity system-verification baseline. Building on this foundation, the council then performs system verification by assessing post-fault responses against NL operational tolerances, ultimately generating actionable V&V artifacts, such as gain-tuning proposals. Experiments on a time-series simulator for Unmanned Underwater Vehicles (UUVs) demonstrate that AIVV successfully digitizes the HITL V&V process, overcoming the limitations of rule-based fault classification and offering a scalable blueprint for LLM-mediated oversight in time-series data domains.

Method: Tested 16 recent Large Language Models in a controlled virtual environment simulation where agents were presented with scenarios involving evidence of corporate fraud and harm, evaluating their choices to suppress or reveal this evidence.

Result: Majority of evaluated state-of-the-art AI agents explicitly chose to suppress evidence of fraud and harm to protect company profits, though some models showed remarkable resistance and behaved appropriately.

Conclusion: Many current AI systems demonstrate concerning willingness to aid criminal activity when aligned with corporate authority, highlighting significant safety risks in agent deployment.

Abstract: As ongoing research explores the ability of AI agents to be insider threats and act against company interests, we showcase the abilities of such agents to act against human well being in service of corporate authority. Building on Agentic Misalignment and AI scheming research, we present a scenario where the majority of evaluated state-of-the-art AI agents explicitly choose to suppress evidence of fraud and harm, in service of company profit. We test this scenario on 16 recent Large Language Models. Some models show remarkable resistance to our method and behave appropriately, but many do not, and instead aid and abet criminal activity. These experiments are simulations and were executed in a controlled virtual environment. No crime actually occurred.

Method: Conducted the first mechanistic interpretability study of AVLLMs, analyzing feature evolution across layers, probing analyses to detect latent audio information, and tracing modality biases back to training data and alignment procedures.

Result: Found that while AVLLMs encode rich audio semantics at intermediate layers, this information fails to surface in final text generation when audio conflicts with vision. Deeper fusion layers disproportionately privilege visual representations, suppressing audio cues. Training analysis shows AVLLM’s audio behavior strongly matches its vision-language base model, indicating limited additional alignment to audio supervision.

Conclusion: Reveals a fundamental modality bias in AVLLMs where vision dominates audio integration, providing mechanistic insights into how multimodal LLMs process conflicting information and highlighting the need for better audio-visual alignment in training.

Abstract: Audio-Visual Large Language Models (AVLLMs) are emerging as unified interfaces to multimodal perception. We present the first mechanistic interpretability study of AVLLMs, analyzing how audio and visual features evolve and fuse through different layers of an AVLLM to produce the final text outputs. We find that although AVLLMs encode rich audio semantics at intermediate layers, these capabilities largely fail to surface in the final text generation when audio conflicts with vision. Probing analyses show that useful latent audio information is present, but deeper fusion layers disproportionately privilege visual representations that tend to suppress audio cues. We further trace this imbalance to training: the AVLLM’s audio behavior strongly matches its vision-language base model, indicating limited additional alignment to audio supervision. Our findings reveal a fundamental modality bias in AVLLMs and provide new mechanistic insights into how multimodal LLMs integrate audio and vision.

Method: Four-part framework: 1) incorporates extreme weather uncertainty, 2) leverages grid digital twin, 3) uses Monte Carlo simulation for variability, 4) applies multi-objective optimization for investment portfolio selection.

Result: Surprisingly, simpler net present value ranking methods outperformed computationally complex model-based metaheuristic optimization methods, finding more optimal portfolios with limited grid knowledge.

Conclusion: Simpler optimization approaches can be more effective than complex model-based methods for utility capital planning, especially given computational constraints and limited grid knowledge requirements.

Abstract: Electric utilities must make massive capital investments in the coming years to respond to explosive growth in demand, aging assets and rising threats from extreme weather. Utilities today already have rigorous frameworks for capital planning, and there are opportunities to extend this capability to solve multi-objective optimization problems in the face of uncertainty. This work presents a four-part framework that 1) incorporates extreme weather as a source of uncertainty, 2) leverages a digital twin of the grid, 3) uses Monte Carlo simulation to capture variability and 4) applies a multi-objective optimization method for finding the optimal investment portfolio. We use this framework to investigate whether grid-aware optimization methods outperform model-free approaches. We find that, in fact, given the computational complexity of model-based metaheuristic optimization methods, the simpler net present value ranking method was able to find more optimal portfolios with only limited knowledge of the grid.

Method: Interpretable RL approach using differentiable soft tree models as actor approximators, temperature annealing during training, and regularization with pruning to produce deterministic oblique decision trees.

Result: Produces near-optimal, interpretable life-cycle policies in tree form with reasonable complexity, demonstrated in steel girder bridge optimization problems.

Conclusion: The framework successfully addresses the trade-off between optimality and interpretability in bridge management, producing human-auditable policies implementable in current systems.

Abstract: The new Specifications for the National Bridge Inventory (SNBI), in effect from 2022, emphasize the use of element-level condition states (CS) for risk-based bridge management. Instead of a general component rating, element-level condition data use an array of relative CS quantities (i.e., CS proportions) to represent the condition of a bridge. Although this greatly increases the granularity of bridge condition data, it introduces challenges to set up optimal life-cycle policies due to the expanded state space from one single categorical integer to four-dimensional probability arrays. This study proposes a new interpretable reinforcement learning (RL) approach to seek optimal life-cycle policies based on element-level state representations. Compared to existing RL methods, the proposed algorithm yields life-cycle policies in the form of oblique decision trees with reasonable amounts of nodes and depth, making them directly understandable and auditable by humans and easily implementable into current bridge management systems. To achieve near-optimal policies, the proposed approach introduces three major improvements to existing RL methods: (a) the use of differentiable soft tree models as actor function approximators, (b) a temperature annealing process during training, and (c) regularization paired with pruning rules to limit policy complexity. Collectively, these improvements can yield interpretable life-cycle policies in the form of deterministic oblique decision trees. The benefits and trade-offs from these techniques are demonstrated in both supervised and reinforcement learning settings. The resulting framework is illustrated in a life-cycle optimization problem for steel girder bridges.

Method: Proposes a neuro-symbolic “plan-retrieve-generate” workflow using Knowledge Graphs, repurposing competency questions as executable narrative plans. Validated using a multimodal Live Aid KG dataset with three RAG strategies: symbolic KG-RAG, text-enriched Hybrid-RAG, and structure-aware Graph-RAG.

Result: Reveals quantifiable trade-offs between factual precision (symbolic retrieval), contextual richness (hybrid methods), and narrative coherence (graph-based traversal). Provides insights for designing personalized, controllable storytelling systems.

Conclusion: The neuro-symbolic KG-based approach enables verifiable, evidence-closed story generation for cultural heritage preservation, balancing factual accuracy with narrative quality through different RAG strategies.

Abstract: The preservation of intangible cultural heritage is a critical challenge as collective memory fades over time. While Large Language Models (LLMs) offer a promising avenue for generating engaging narratives, their propensity for factual inaccuracies or “hallucinations” makes them unreliable for heritage applications where veracity is a central requirement. To address this, we propose a novel neuro-symbolic architecture grounded in Knowledge Graphs (KGs) that establishes a transparent “plan-retrieve-generate” workflow for story generation. A key novelty of our approach is the repurposing of competency questions (CQs) - traditionally design-time validation artifacts - into run-time executable narrative plans. This approach bridges the gap between high-level user personas and atomic knowledge retrieval, ensuring that generation is evidence-closed and fully auditable. We validate this architecture using a new resource: the Live Aid KG, a multimodal dataset aligning 1985 concert data with the Music Meta Ontology and linking to external multimedia assets. We present a systematic comparative evaluation of three distinct Retrieval-Augmented Generation (RAG) strategies over this graph: a purely symbolic KG-RAG, a text-enriched Hybrid-RAG, and a structure-aware Graph-RAG. Our experiments reveal a quantifiable trade-off between the factual precision of symbolic retrieval, the contextual richness of hybrid methods, and the narrative coherence of graph-based traversal. Our findings offer actionable insights for designing personalised and controllable storytelling systems.

Method: Proposes Debiasing-DPO, a self-supervised training method that pairs neutral reasoning (from query alone) with biased reasoning (from query plus spurious context), combined with supervised fine-tuning on ground-truth labels to maintain accuracy.

Result: Applied to Llama 3B & 8B and Qwen 3B & 7B Instruct models, Debiasing-DPO reduces bias by 84% and improves predictive accuracy by 52% on average. Found that larger models sometimes show greater sensitivity to spurious contexts despite higher accuracy.

Conclusion: Robustness to spurious context is not a natural byproduct of model scaling; Debiasing-DPO yields substantial gains in both accuracy and robustness for prompt-based prediction tasks in high-stakes domains.

Abstract: LLMs are increasingly used for high-stakes decision-making, yet their sensitivity to spurious contextual information can introduce harmful biases. This is a critical concern when models are deployed for tasks like evaluating teachers’ instructional quality, where biased assessment can affect teachers’ professional development and career trajectories. We investigate model robustness to spurious social contexts using the largest publicly available dataset of U.S. classroom transcripts (NCTE) paired with expert rubric scores. Evaluating seven frontier and open-weight models across seven categories of spurious contexts – including teacher experience, education level, demographic identity, and sycophancy-inducing framings – we find that irrelevant contextual information can shift model predictions by up to 1.48 points on a 7-point scale, with larger models sometimes exhibiting greater sensitivity despite higher predictive accuracy. Mitigations using prompts and standard direct preference optimization (DPO) prove largely insufficient. We propose Debiasing-DPO,, a self-supervised training method that pairs neutral reasoning generated from the query alone, with the model’s biased reasoning generated with both the query and additional spurious context. We further combine this objective with supervised fine-tuning on ground-truth labels to prevent losses in predictive accuracy. Applied to Llama 3B & 8B and Qwen 3B & 7B Instruct models, Debiasing-DPO reduces bias by 84% and improves predictive accuracy by 52% on average. Our findings from the educational case study highlight that robustness to spurious context is not a natural byproduct of model scaling and that our proposed method can yield substantial gains in both accuracy and robustness for prompt-based prediction tasks.

Method: AutoVerifier decomposes technical assertions into structured claim triples (Subject, Predicate, Object), constructing knowledge graphs for structured reasoning across six layers: corpus construction/ingestion, entity/claim extraction, intra-document verification, cross-source verification, external signal corroboration, and final hypothesis matrix generation.

Result: Demonstrated on a contested quantum computing claim, AutoVerifier identified overclaims and metric inconsistencies within target papers, traced cross-source contradictions, uncovered undisclosed commercial conflicts of interest, and produced final assessments - all operated by analysts with no quantum expertise.

Conclusion: Structured LLM verification can reliably evaluate the validity and maturity of emerging technologies, turning raw technical documents into traceable, evidence-backed intelligence assessments.

Abstract: Scientific and Technical Intelligence (S&TI) analysis requires verifying complex technical claims across rapidly growing literature, where existing approaches fail to bridge the verification gap between surface-level accuracy and deeper methodological validity. We present AutoVerifier, an LLM-based agentic framework that automates end-to-end verification of technical claims without requiring domain expertise. AutoVerifier decomposes every technical assertion into structured claim triples of the form (Subject, Predicate, Object), constructing knowledge graphs that enable structured reasoning across six progressively enriching layers: corpus construction and ingestion, entity and claim extraction, intra-document verification, cross-source verification, external signal corroboration, and final hypothesis matrix generation. We demonstrate AutoVerifier on a contested quantum computing claim, where the framework, operated by analysts with no quantum expertise, automatically identified overclaims and metric inconsistencies within the target paper, traced cross-source contradictions, uncovered undisclosed commercial conflicts of interest, and produced a final assessment. These results show that structured LLM verification can reliably evaluate the validity and maturity of emerging technologies, turning raw technical documents into traceable, evidence-backed intelligence assessments.

Method: Intrinsic-relational routing classifies every property as either intrinsic or relational and routes it to corresponding schema modules, creating a declarative schema portable across storage backends. Applied to Wikidata dump with rule-based cleaning and iterative routing.

Result: Processed 34.6M entities from Wikidata, assigning properties to 94 modules across 8 categories with 93.3% category coverage and 98.0% module assignment. Created property graph with 34.0M nodes and 61.2M edges across 38 relationship types.

Conclusion: The ontology-oriented approach produces reusable schemas independent of construction pipelines, validated through five applications including ontology analysis, entity disambiguation, and LLM-guided extraction.

Abstract: Organizing a large-scale knowledge graph into a typed property graph requires structural decisions – which entities become nodes, which properties become edges, and what schema governs these choices. Existing approaches embed these decisions in pipeline code or extract relations ad hoc, producing schemas that are tightly coupled to their construction process and difficult to reuse for downstream ontology-level tasks. We present an ontology-oriented approach in which the schema is designed from the outset for ontology analysis, entity disambiguation, domain customization, and LLM-guided extraction – not merely as a byproduct of graph building. The core mechanism is intrinsic-relational routing, which classifies every property as either intrinsic or relational and routes it to the corresponding schema module. This routing produces a declarative schema that is portable across storage backends and independently reusable. We instantiate the approach on the January 2026 Wikidata dump. A rule-based cleaning stage identifies a 34.6M-entity core set from the full dump, followed by iterative intrinsic-relational routing that assigns each property to one of 94 modules organized into 8 categories. With tool-augmented LLM support and human review, the schema reaches 93.3% category coverage and 98.0% module assignment among classified entities. Exporting this schema yields a property graph with 34.0M nodes and 61.2M edges across 38 relationship types. We validate the ontology-oriented claim through five applications that consume the schema independently of the construction pipeline: ontology structure analysis, benchmark annotation auditing, entity disambiguation, domain customization, and LLM-guided extraction.

Method: A multi-agent system that treats clinical trial optimization as iterative redesign of textual protocols. Integrates failure diagnosis, safety-aware modifications, and candidate evaluation in a closed-loop, reward-driven framework. Uses outcome prediction models as simulation environments and features hierarchical memory capturing iteration-level feedback and transferable redesign patterns.

Result: Improves 83.3% of trial protocols with mean success probability gain of 5.7% at negligible cost ($0.12 per trial). Retrospective case studies show alignment between discovered redesign strategies and real-world clinical trial modifications.

Conclusion: ClinicalReTrial provides an effective framework for actionable clinical trial optimization through iterative protocol redesign, bridging the gap between failure prediction and practical improvement strategies.

Abstract: Clinical trials constitute a critical yet exceptionally challenging and costly stage of drug development ($2.6B per drug), where protocols are encoded as complex natural language documents, motivating the use of AI systems beyond manual analysis. Existing AI methods accurately predict trial failure, but do not provide actionable remedies. To fill this gap, this paper proposes ClinicalReTrial, a multi-agent system that formulates clinical trial optimization as an iterative redesign problem on textural protocols. Our method integrates failure diagnosis, safety-aware modifications, and candidate evaluation in a closed-loop, reward-driven optimization framework. Serving the outcome prediction model as a simulation environment, ClinicalReTrial enables low-cost evaluation and dense reward signals for continuous self-improvement. We further propose a hierarchical memory that captures iteration-level feedback within trials and distills transferable redesign patterns across trials. Empirically, ClinicalReTrial improves $83.3%$ of trial protocols with a mean success probability gain of $5.7%$ with negligible cost ($0.12 per trial). Retrospective case studies demonstrate alignment between the discovered redesign strategies and real-world clinical trial modifications. The code is anonymously available at: https://github.com/xingsixue123/ClinicalFailureReasonReTrial.

Method: Build LLM-powered decision agents that role-play diverse stakeholders with internal utility functions; generate thousands of conversations in school scheduling case study; compare conversational vs one-shot evaluation

Result: Conversational evaluation reveals significantly higher solution quality than one-shot approaches; tailored optimization agents with domain-specific prompts outperform general-purpose chatbots in solution quality and interaction efficiency

Conclusion: Conversational evaluation methodology is essential for assessing optimization agents; domain-specific optimization agents outperform general chatbots; operations research expertise crucial for effective interactive optimization deployments

Abstract: Optimization is as much about modeling the right problem as solving it. Identifying the right objectives, constraints, and trade-offs demands extensive interaction between researchers and stakeholders. Large language models can empower decision-makers with optimization capabilities through interactive optimization agents that can propose, interpret and refine solutions. However, it is fundamentally harder to evaluate a conversation-based interaction than traditional one-shot approaches. This paper proposes a scalable and replicable methodology for evaluating optimization agents through conversations. We build LLM-powered decision agents that role-play diverse stakeholders, each governed by an internal utility function but communicating like a real decision-maker. We generate thousands of conversations in a school scheduling case study. Results show that one-shot evaluation is severely limiting: the same optimization agent converges to much higher-quality solutions through conversations. Then, this paper uses this methodology to demonstrate that tailored optimization agents, endowed with domain-specific prompts and structured tools, can lead to significant improvements in solution quality in fewer interactions, as compared to general-purpose chatbots. These findings provide evidence of the benefits of emerging solutions at the AI-optimization interface to expand the reach of optimization technologies in practice. They also uncover the impact of operations research expertise to facilitate interactive deployments through the design of effective and reliable optimization agents.

Method: GrandCode uses a multi-agent RL system with orchestrated agentic modules (hypothesis proposal, solver, test generator, summarization) improved through post-training and online test-time RL. It introduces Agentic GRPO specifically designed for multi-stage agent rollouts with delayed rewards and addressing severe off-policy drift in agentic RL.

Result: GrandCode is the first AI system to consistently beat all human participants in live competitive programming contests. In three recent Codeforces competitions (Rounds 1087, 1088, 1089), GrandCode placed first in all of them, beating legendary grandmasters and all human participants.

Conclusion: AI systems have reached a point where they surpass the strongest human programmers on the most competitive coding tasks, marking a significant milestone in AI capabilities for complex problem-solving and programming.

Abstract: Competitive programming remains one of the last few human strongholds in coding against AI. The best AI system to date still underperforms the best humans competitive programming: the most recent best result, Google’s Gemini3 Deep Think, attained 8th place even not being evaluated under live competition conditions. In this work, we introduce GrandCode, a multi-agent RL system designed for competitive programming. The capability of GrandCode is attributed to two key factors: (1) It orchestrates a variety of agentic modules (hypothesis proposal, solver, test generator, summarization, etc) and jointly improves them through post-training and online test-time RL; (2) We introduce Agentic GRPO specifically designed for multi-stage agent rollouts with delayed rewards and the severe off-policy drift that is prevalent in agentic RL. GrandCode is the first AI system that consistently beats all human participants in live contests of competitive programming: in the most recent three Codeforces live competitions, i.e., Round1087 (Mar 21, 2026), Round1088 (Mar 28, 2026), and Round1089 (Mar 29, 2026), GrandCode placed first in all of them, beating all human participants, including legendary grandmasters. GrandCode shows that AI systems have reached a point where they surpass the strongest human programmers on the most competitive coding tasks.

Method: DeltaLogic transforms natural-language reasoning examples into short revision episodes: first asking for initial conclusion under premises P, then applying minimal edit δ(P), and finally asking whether previous conclusion should remain stable or be revised. Instantiated from FOLIO and ProofWriter datasets.

Result: Evaluation shows stronger initial reasoning doesn’t imply stronger revision behavior: Qwen3-1.7B reaches 0.667 initial accuracy but only 0.467 revision accuracy, with 0.600 inertia on episodes where gold label should change. Phi-4-mini-instruct performs better (0.950 initial, 0.850 revised) but still shows non-trivial abstention and control instability.

Conclusion: Logical competence under fixed premises doesn’t imply disciplined belief revision after local evidence edits. DeltaLogic targets a distinct and practically important reasoning capability that complements existing logical inference benchmarks.

Abstract: Reasoning benchmarks typically evaluate whether a model derives the correct answer from a fixed premise set, but they under-measure a closely related capability that matters in dynamic environments: belief revision under minimal evidence change. We introduce DeltaLogic, a benchmark transformation protocol that converts natural-language reasoning examples into short revision episodes. Each episode first asks for an initial conclusion under premises P, then applies a minimal edit δ(P), and finally asks whether the previous conclusion should remain stable or be revised. We instantiate DeltaLogic from FOLIO and ProofWriter and evaluate small causal language models with constrained label scoring. On a completed 30-episode Qwen evaluation subset, stronger initial reasoning still does not imply stronger revision behavior: Qwen3-1.7B reaches 0.667 initial accuracy but only 0.467 revision accuracy, with inertia rising to 0.600 on episodes where the gold label should change, while Qwen3-0.6B collapses into near universal abstention. There, Qwen3-4B preserves the same inertial failure pattern (0.650 initial, 0.450 revised, 0.600 inertia), whereas Phi-4-mini-instruct is substantially stronger (0.950 initial, 0.850 revised) but still exhibits non-trivial abstention and control instability. These results suggest that logical competence under fixed premises does not imply disciplined belief revision after local evidence edits. DeltaLogic therefore targets a distinct and practically important reasoning capability that complements existing logical inference and belief-updating benchmarks.

Method: Proposes a Neuro-Symbolic Dual Memory Framework that explicitly decouples semantic progress guidance from logical feasibility verification. During inference, it synchronously invokes: 1) neural-network-based Progress Memory that extracts semantic blueprints from successful trajectories for global task guidance, and 2) symbolic-logic-based Feasibility Memory that uses executable Python verification functions synthesized from failed transitions for strict logical validation.

Result: The method significantly outperforms existing competitive baselines on ALFWorld, WebShop, and TextCraft benchmarks, while drastically reducing invalid action rates and average trajectory lengths.

Conclusion: Decoupling semantic progress guidance from logical feasibility verification through a neuro-symbolic dual memory framework effectively addresses the distinct challenges of global progress drift and local feasibility violation in long-horizon decision-making tasks.

Abstract: Large language models (LLMs) have demonstrated strong potential in long-horizon decision-making tasks, such as embodied manipulation and web interaction. However, agents frequently struggle with endless trial-and-error loops or deviate from the main objective in complex environments. We attribute these failures to two fundamental errors: global Progress Drift and local Feasibility Violation. Existing methods typically attempt to address both issues simultaneously using a single paradigm. However, these two challenges are fundamentally distinct: the former relies on fuzzy semantic planning, while the latter demands strict logical constraints and state validation. The inherent limitations of such a single-paradigm approach pose a fundamental challenge for existing models in handling long-horizon tasks. Motivated by this insight, we propose a Neuro-Symbolic Dual Memory Framework that explicitly decouples semantic progress guidance from logical feasibility verification. Specifically, during the inference phase, the framework invokes both memory mechanisms synchronously: on one hand, a neural-network-based Progress Memory extracts semantic blueprints from successful trajectories to guide global task advancement; on the other hand, a symbolic-logic-based Feasibility Memory utilizes executable Python verification functions synthesized from failed transitions to perform strict logical validation. Experiments demonstrate that this method significantly outperforms existing competitive baselines on ALFWorld, WebShop, and TextCraft, while drastically reducing the invalid action rate and average trajectory length.

Method: Constructs a role assignment matrix measuring semantic similarity between agents’ behavior trajectories and role descriptions, defines a role clarity matrix, and uses its Frobenius norm as a regularizer during lightweight fine-tuning to improve role consistency.

Result: Experiments on ChatDev show substantial improvements: role overstepping rates decreased dramatically (46.4% to 8.4% for Qwen, 43.4% to 0.2% for Llama), role clarity scores increased significantly, and task success rates improved.

Conclusion: The proposed quantitative role clarity metric and regularization approach effectively improves role consistency in multi-agent systems, leading to better task performance by ensuring agents adhere to their specified roles.

Abstract: In large language model (LLM)-driven multi-agent systems, disobey role specification (failure to adhere to the defined responsibilities and constraints of an assigned role, potentially leading to an agent behaving like another) is a major failure mode \cite{DBLP:journals/corr/abs-2503-13657}. To address this issue, in the present paper, we propose a quantitative role clarity to improve role consistency. Firstly, we construct a role assignment matrix $S(φ)=[s_{ij}(φ)]$, where $s_{ij}(φ)$ is the semantic similarity between the $i$-th agent’s behavior trajectory and the $j$-th agent’s role description. Then we define role clarity matrix $M(φ)$ as $\text{softmax}(S(φ))-I$, where $\text{softmax}(S(φ))$ is a row-wise softmax of $S(φ)$ and $I$ is the identity matrix. The Frobenius norm of $M(φ)$ quantifies the alignment between agents’ role descriptions and their behaviors trajectory. Moreover, we employ the role clarity matrix as a regularizer during lightweight fine-tuning to improve role consistency, thereby improving end-to-end task performance. Experiments on the ChatDev multi-agent system show that our method substantially improves role consistency and task performance: with Qwen and Llama, the role overstepping rate decreases from $46.4%$ to $8.4%$ and from $43.4%$ to $0.2%$, respectively, and the role clarity score increases from $0.5328$ to $0.9097$ and from $0.5007$ to $0.8530$, respectively, the task success rate increases from $0.6769$ to $0.6909$ and from $0.6174$ to $0.6763$, respectively.

Method: 1) DuoChart: scalable dual-source data pipeline combining synthesized and real-world charts for diverse training data. 2) CharTool: equips MLLMs with external tools (image cropping for localized perception, code-based computation for numerical reasoning). 3) Agentic reinforcement learning on DuoChart to learn tool-integrated reasoning.

Result: CharTool-7B outperforms base model by +8.0% on CharXiv (Reasoning) and +9.78% on ChartQAPro. Achieves competitive performance with substantially larger/proprietary models. Shows positive generalization to out-of-domain visual math reasoning benchmarks.

Conclusion: The proposed dual-source data pipeline and tool-integrated approach effectively addresses chart reasoning challenges for MLLMs, demonstrating significant improvements across multiple benchmarks and good generalization capabilities.

Abstract: Charts are ubiquitous in scientific and financial literature for presenting structured data. However, chart reasoning remains challenging for multimodal large language models (MLLMs) due to the lack of high-quality training data, as well as the need for fine-grained visual grounding and precise numerical computation. To address these challenges, we first propose DuoChart, a scalable dual-source data pipeline that combines synthesized charts with real-world charts to construct diverse, high-quality chart training data. We then introduce CharTool, which equips MLLMs with external tools, including image cropping for localized visual perception and code-based computation for accurate numerical reasoning. Through agentic reinforcement learning on DuoChart, CharTool learns tool-integrated reasoning grounded in chart content. Extensive experiments on six chart benchmarks show that our method consistently improves over strong MLLM baselines across model scales. Notably, CharTool-7B outperforms the base model by +8.0% on CharXiv (Reasoning) and +9.78% on ChartQAPro, while achieving competitive performance with substantially larger or proprietary models. Moreover, CharTool demonstrates positive generalization to out-of-domain visual math reasoning benchmarks.

Method: Event-driven synthesis framework creating 100 synthetic users with 1-5 year trajectories including health profiles, multi-phase narrative plans, daily device measurements, periodic exam records, and event logs with explicit impact parameters. Uses hybrid pipeline: LLM-based planning for sparse semantic artifacts and algorithmic simulation with physiological bounds for dense indicator dynamics.

Result: Evaluated 13 methods (LLMs with tools, DB-native agents, memory-augmented RAG). DB agents (48-58% accuracy) substantially outperform memory RAG baselines (30-38%), with the gap largest on Comparison and Explanation queries requiring multi-hop reasoning and evidence attribution.

Conclusion: ESL-Bench provides a scalable synthetic benchmark for evaluating longitudinal health agents, revealing significant performance gaps between different agent architectures, particularly for complex reasoning tasks requiring evidence attribution and multi-hop inference.

Abstract: Longitudinal health agents must reason across multi-source trajectories that combine continuous device streams, sparse clinical exams, and episodic life events - yet evaluating them is hard: real-world data cannot be released at scale, and temporally grounded attribution questions seldom admit definitive answers without structured ground truth. We present ESL-Bench, an event-driven synthesis framework and benchmark providing 100 synthetic users, each with a 1-5 year trajectory comprising a health profile, a multi-phase narrative plan, daily device measurements, periodic exam records, and an event log with explicit per-indicator impact parameters. Each indicator follows a baseline stochastic process driven by discrete events with sigmoid-onset, exponential-decay kernels under saturation and projection constraints; a hybrid pipeline delegates sparse semantic artifacts to LLM-based planning and dense indicator dynamics to algorithmic simulation with hard physiological bounds. Users are each paired with 100 evaluation queries across five dimensions - Lookup, Trend, Comparison, Anomaly, Explanation - stratified into Easy, Medium, and Hard tiers, with all ground-truth answers programmatically computable from the recorded event-indicator relationships. Evaluating 13 methods spanning LLMs with tools, DB-native agents, and memory-augmented RAG, we find that DB agents (48-58%) substantially outperform memory RAG baselines (30-38%), with the gap concentrated on Comparison and Explanation queries where multi-hop reasoning and evidence attribution are required.

Method: Three components: 1) Agent Confidence Modeling (ACM) estimates agent reliability using historical performance and semantic similarity; 2) Adaptive Incremental Voting (AIV) sequentially selects agents with early stopping; 3) Individual Confidence Updating (ICU) dynamically updates agent reliability.

Result: Extensive evaluations across six benchmarks demonstrate EMS consistently reduces the average number of invoked agents by 32%.

Conclusion: EMS provides an efficient approach to multi-agent voting by prioritizing reliable agents and stopping early when majority consensus is reached, significantly reducing computational overhead.

Abstract: Majority voting is the standard for aggregating multi-agent responses into a final decision. However, traditional methods typically require all agents to complete their reasoning before aggregation begins, leading to significant computational overhead, as many responses become redundant once a majority consensus is achieved. In this work, we formulate the multi-agent voting as a reliability-aware agent scheduling problem, and propose an Efficient Majority-then-Stopping (EMS) to improve reasoning efficiency. EMS prioritizes agents based on task-aware reliability and terminates the reasoning pipeline the moment a majority is achieved from the following three critical components. Specifically, we introduce Agent Confidence Modeling (ACM) to estimate agent reliability using historical performance and semantic similarity, Adaptive Incremental Voting (AIV) to sequentially select agents with early stopping, and Individual Confidence Updating (ICU) to dynamically update the reliability of each contributing agent. Extensive evaluations across six benchmarks demonstrate that EMS consistently reduces the average number of invoked agents by 32%.

Method: Combines MT-GRPO (Multi-Turn Group Relative Policy Optimization) with GTPO (Generalized Token-level Policy Optimization) and introduces Iterative Reward Calibration methodology using empirical discriminative analysis of rollout data.

Result: Improved Qwen3.5-4B from 63.8% to 66.7% (+2.9pp) and Qwen3-30B-A3B from 58.0% to 69.5% (+11.5pp) on Tau-Bench airline benchmark, with 4B model exceeding GPT-4.1 and GPT-4o despite being 50x smaller.

Conclusion: The approach successfully addresses reward misalignment in multi-turn RL training and achieves state-of-the-art performance on realistic customer service tasks with tool-calling agents.

Abstract: Training tool-calling agents with reinforcement learning on multi-turn tasks remains challenging due to sparse outcome rewards and difficult credit assignment across conversation turns. We present the first application of MT-GRPO (Multi-Turn Group Relative Policy Optimization) combined with GTPO (Generalized Token-level Policy Optimization) for training a tool-calling agent on realistic customer service tasks with an LLM-based user simulator. Through systematic analysis of training rollouts, we discover that naively designed dense per-turn rewards degrade performance by up to 14 percentage points due to misalignment between reward discriminativeness and advantage direction. We introduce Iterative Reward Calibration, a methodology for designing per-turn rewards using empirical discriminative analysis of rollout data, and show that our GTPO hybrid advantage formulation eliminates the advantage misalignment problem. Applied to the Tau-Bench airline benchmark, our approach improves Qwen3.5-4B from 63.8 percent to 66.7 percent (+2.9pp) and Qwen3-30B-A3B from 58.0 percent to 69.5 percent (+11.5pp) – with the trained 4B model exceeding GPT-4.1 (49.4 percent) and GPT-4o (42.8 percent) despite being 50 times smaller, and the 30.5B MoE model approaching Claude Sonnet 4.5 (70.0 percent). To our knowledge, these are the first published RL training results on Tau-Bench. We release our code, reward calibration analysis, and training recipes.

Method: Systematically studied Blocksworld domain and equivalent Path-Star (P*) graph tasks, manipulating problem depth (tower height), width (number of towers), and compositionality (number of goal blocks). Compared reasoning-enhanced LLMs against traditional satisficing planners like LAMA, and analyzed performance as search space expands.

Result: LLMs significantly outperform traditional planners in complex multi-goal configurations. While classical search algorithms struggle with expanding search spaces, LLMs track theoretical optimality limits with near-perfect precision, even without domain-specific semantic hints.

Conclusion: LLMs demonstrate surprising planning capabilities, potentially due to either Algorithmic Simulation via reasoning tokens or Geometric Memory that represents P* topology as navigable global geometry, effectively bypassing exponential combinatorial complexity.

Abstract: Classic AI planning problems have been revisited in the Large Language Model (LLM) era, with a focus of recent benchmarks on success rates rather than plan efficiency. We examine the degree to which frontier models reason optimally versus relying on simple, heuristic, and possibly inefficient strategies. We focus on the Blocksworld domain involving towers of labeled blocks which have to be moved from an initial to a goal configuration via a set of primitive actions. We also study a formally equivalent task, the generalized Path-Star ($P^$) graph, in order to isolate true topological reasoning from semantic priors. We systematically manipulate problem depth (the height of block towers), width (the number of towers), and compositionality (the number of goal blocks). Reasoning-enhanced LLMs significantly outperform traditional satisficing planners (e.g., LAMA) in complex, multi-goal configurations. Although classical search algorithms hit a wall as the search space expands, LLMs track theoretical optimality limits with near-perfect precision, even when domain-specific semantic hints are stripped away. To explain these surprising findings, we consider (and find evidence to support) two hypotheses: an active Algorithmic Simulation executed via reasoning tokens and a Geometric Memory that allows models to represent the $P^$ topology as a navigable global geometry, effectively bypassing exponential combinatorial complexity.

Method: Created AgentHazard benchmark with 2,653 instances spanning diverse risk categories and attack strategies, each pairing harmful objectives with sequences of locally legitimate operational steps that jointly induce unsafe behavior.

Result: Current systems remain highly vulnerable - Claude Code powered by Qwen3-Coder exhibits 73.63% attack success rate, showing model alignment alone doesn’t guarantee safety of autonomous agents.

Conclusion: Computer-use agents present distinct safety challenges requiring specialized evaluation; current systems are vulnerable to harm emerging through sequences of plausible steps, necessitating new safety approaches beyond traditional model alignment.

Abstract: Computer-use agents extend language models from text generation to persistent action over tools, files, and execution environments. Unlike chat systems, they maintain state across interactions and translate intermediate outputs into concrete actions. This creates a distinct safety challenge in that harmful behavior may emerge through sequences of individually plausible steps, including intermediate actions that appear locally acceptable but collectively lead to unauthorized actions. We present \textbf{AgentHazard}, a benchmark for evaluating harmful behavior in computer-use agents. AgentHazard contains \textbf{2,653} instances spanning diverse risk categories and attack strategies. Each instance pairs a harmful objective with a sequence of operational steps that are locally legitimate but jointly induce unsafe behavior. The benchmark evaluates whether agents can recognize and interrupt harm arising from accumulated context, repeated tool use, intermediate actions, and dependencies across steps. We evaluate AgentHazard on Claude Code, OpenClaw, and IFlow using mostly open or openly deployable models from the Qwen3, Kimi, GLM, and DeepSeek families. Our experimental results indicate that current systems remain highly vulnerable. In particular, when powered by Qwen3-Coder, Claude Code exhibits an attack success rate of \textbf{73.63%}, suggesting that model alignment alone does not reliably guarantee the safety of autonomous agents.

Method: The authors characterize errors as a forest-structured Forest of Errors (FoE) and propose RED framework with two components: 1) Refining First - suppresses FoE growth in the first solution, and 2) Discarding Subs - prunes subsequent FoE via dual-consistency.

Result: RED outperforms eight competitive baselines across five benchmarks and six backbone models, achieving performance gains up to 19.0% while reducing token consumption by 37.7% ~ 70.4%.

Conclusion: The paper demonstrates that “The First is The Best” phenomenon exists in Large Reasoning Models, and the proposed RED framework effectively leverages this insight to improve reasoning efficiency and performance.

Abstract: Recent Large Reasoning Models (LRMs) like DeepSeek-R1 have demonstrated remarkable success in complex reasoning tasks, exhibiting human-like patterns in exploring multiple alternative solutions. Upon closer inspection, however, we uncover a surprising phenomenon: The First is The Best, where alternative solutions are not merely suboptimal but potentially detrimental. This observation challenges widely accepted test-time scaling laws, leading us to hypothesize that errors within the reasoning path scale concurrently with test time. Through comprehensive empirical analysis, we characterize errors as a forest-structured Forest of Errors (FoE) and conclude that FoE makes the First the Best, which is underpinned by rigorous theoretical analysis. Leveraging these insights, we propose RED, a self-guided efficient reasoning framework comprising two components: I) Refining First, which suppresses FoE growth in the first solution; and II) Discarding Subs, which prunes subsequent FoE via dual-consistency. Extensive experiments across five benchmarks and six backbone models demonstrate that RED outperforms eight competitive baselines, achieving performance gains of up to 19.0% while reducing token consumption by 37.7% ~ 70.4%. Moreover, comparative experiments on FoE metrics shed light on how RED achieves effectiveness.

Method: Hierarchical framework based on near-decomposability principle with three layers: strategic Host, multiple Managers with aggregation/reflection mechanisms, and parallel Workers. Enforces strict context isolation and parallelism.

Result: Achieves 3-5× speed-up, 8.4% success rate on WideSearch-en, and 52.9% accuracy on BrowseComp-zh benchmarks.

Conclusion: InfoSeeker effectively addresses limitations of existing LLM agent systems for wide-scale information synthesis through hierarchical architecture with context isolation and parallelism.

Abstract: Recent agentic search systems have made substantial progress by emphasising deep, multi-step reasoning. However, this focus often overlooks the challenges of wide-scale information synthesis, where agents must aggregate large volumes of heterogeneous evidence across many sources. As a result, most existing large language model agent systems face severe limitations in data-intensive settings, including context saturation, cascading error propagation, and high end-to-end latency. To address these challenges, we present \framework, a hierarchical framework based on principle of near-decomposability, containing a strategic \textit{Host}, multiple \textit{Managers} and parallel \textit{Workers}. By leveraging aggregation and reflection mechanisms at the Manager layer, our framework enforces strict context isolation to prevent saturation and error propagation. Simultaneously, the parallelism in worker layer accelerates the speed of overall task execution, mitigating the significant latency. Our evaluation on two complementary benchmarks demonstrates both efficiency ($ 3-5 \times$ speed-up) and effectiveness, achieving a $8.4%$ success rate on WideSearch-en and $52.9%$ accuracy on BrowseComp-zh. The code is released at https://github.com/agent-on-the-fly/InfoSeeker

Method: Created Agentic-MME benchmark with 418 real-world tasks across 6 domains and 3 difficulty levels, featuring over 2,000 stepwise checkpoints with manual annotation. Includes unified evaluation framework with sandboxed code/APIs and human reference trajectories with dual-axis (S-axis and V-axis) verification.

Result: Best model (Gemini3-pro) achieves 56.3% overall accuracy, but drops to 23.0% on Level-3 tasks, showing significant difficulty in real-world multimodal agentic problem solving.

Conclusion: Current MLLMs struggle with complex multimodal agentic tasks, highlighting the need for process-level evaluation and better tool integration capabilities.

Abstract: Multimodal Large Language Models (MLLMs) are evolving from passive observers into active agents, solving problems through Visual Expansion (invoking visual tools) and Knowledge Expansion (open-web search). However, existing evaluations fall short: they lack flexible tool integration, test visual and search tools separately, and evaluate primarily by final answers. Consequently, they cannot verify if tools were actually invoked, applied correctly, or used efficiently. To address this, we introduce Agentic-MME, a process-verified benchmark for Multimodal Agentic Capabilities. It contains 418 real-world tasks across 6 domains and 3 difficulty levels to evaluate capability synergy, featuring over 2,000 stepwise checkpoints that average 10+ person-hours of manual annotation per task. Each task includes a unified evaluation framework supporting sandboxed code and APIs, alongside a human reference trajectory annotated with stepwise checkpoints along dual-axis: S-axis and V-axis. To enable true process-level verification, we audit fine-grained intermediate states rather than just final answers, and quantify efficiency via an overthinking metric relative to human trajectories. Experimental results show the best model, Gemini3-pro, achieves 56.3% overall accuracy, which falls significantly to 23.0% on Level-3 tasks, underscoring the difficulty of real-world multimodal agentic problem solving.

Method: Used 30,000 Claude 4.5 Opus agents collaborating in parallel via version control on a shared codebase to formalize the textbook to Lean theorem prover

Result: Successfully formalized entire textbook with 130K lines of Lean code and 5900 declarations within one week, achieving cost parity with human expert teams

Conclusion: Demonstrates milestone in AI-assisted formal verification at textbook scale with potential for further efficiency gains, making code and blueprint available open-source

Abstract: We present a case study where an automatic AI system formalizes a textbook with more than 500 pages of graduate-level algebraic combinatorics to Lean. The resulting formalization represents a new milestone in textbook formalization scale and proficiency, moving from early results in undergraduate topology and restructuring of existing library content to a full standalone formalization of a graduate textbook. The formalization comprises 130K lines of code and 5900 Lean declarations and was conducted within one week by a total of 30K Claude 4.5 Opus agents collaborating in parallel on a shared code base via version control, simultaneously setting a record in multi-agent software engineering with usable results. The inference cost matches or undercuts what we estimate as the salaries required for a team of human experts, and we expect there is still the potential for large efficiencies to be made without the need for better models. We make our code, the resulting Lean code base and a side-by-side blueprint website available open-source.

Method: Proposes Chart-RL, a reinforcement learning framework with policy optimization techniques and adaptive reward functions, integrated with Parameter-Efficient Fine-Tuning using Low-Rank Adaptation (LoRA) for single GPU training.

Result: RL fine-tuned Qwen3-VL-4B-Instruct achieved 0.634 accuracy on ChartQAPro dataset, surpassing the 0.580 accuracy of Qwen3-VL-8B-Instruct foundation model while using half the parameters and reducing inference latency from 31 to 9 seconds.

Conclusion: Chart-RL demonstrates that reinforcement learning can significantly enhance VLMs’ chart understanding capabilities, achieving superior performance with fewer parameters and faster inference compared to baseline foundation models.

Abstract: The recent advancements in Vision Language Models (VLMs) have demonstrated progress toward true intelligence requiring robust reasoning capabilities. Beyond pattern recognition, linguistic reasoning must integrate with visual comprehension, particularly for Chart Question Answering (CQA) tasks involving complex data visualizations. Current VLMs face significant limitations in CQA, including imprecise numerical extraction, difficulty interpreting implicit visual relationships, and inadequate attention mechanisms for capturing spatial relationships in charts. In this work, we address these challenges by presenting Chart-RL, a novel reinforcement learning framework that enhances VLMs chart understanding through feedback-driven policy optimization of visual perception and logical inference. Our key innovation includes a comprehensive framework integrating Reinforcement Learning (RL) from Policy Optimization techniques along with adaptive reward functions, that demonstrates superior performance compared to baseline foundation models and competitive results against larger state-of-the-art architectures. We also integrated Parameter-Efficient Fine-Tuning through Low-Rank Adaptation (LoRA) in the RL framework that only requires single GPU configurations while preserving performance integrity. We conducted extensive benchmarking across open-source, proprietary, and state-of-the-art closed-source models utilizing the ChartQAPro dataset. The RL fine-tuned Qwen3-VL-4B-Instruct model achieved an answer accuracy of 0.634, surpassing the 0.580 accuracy of the Qwen3-VL-8B-Instruct foundation model despite utilizing half the parameter count, while simultaneously reducing inference latency from 31 seconds to 9 seconds.

Method: Comparative analysis of squirrel ecology (fox, eastern gray, red squirrels) with a three-step inference ladder: empirical observation → minimal computational inference → AI design conjecture. Introduces a hierarchical partially observed control model with latent dynamics, structured episodic memory, observer-belief state, option-level actions, and delayed verifier signals.

Result: Three hypotheses: (H1) fast local feedback with predictive compensation improves robustness to hidden dynamics; (H2) memory organized for future control improves delayed retrieval under cue conflict; (H3) verifiers and observer models in action-memory loop reduce silent failures and information leakage. Proposes role-differentiated proposer/executor/checker/adversary systems to reduce correlated errors.

Conclusion: Squirrel ecology offers a valuable comparative perspective and benchmark agenda for developing agentic AI systems that integrate control, memory, and verification. Provides a disciplined program of falsifiable claims about coupled capabilities needed for robust, verifiable action under partial observability.

Abstract: Agentic AI is increasingly judged not by fluent output alone but by whether it can act, remember, and verify under partial observability, delay, and strategic observation. Existing research often studies these demands separately: robotics emphasizes control, retrieval systems emphasize memory, and alignment or assurance work emphasizes checking and oversight. This article argues that squirrel ecology offers a sharp comparative case because arboreal locomotion, scatter-hoarding, and audience-sensitive caching couple all three demands in one organism. We synthesize evidence from fox, eastern gray, and, in one field comparison, red squirrels, and impose an explicit inference ladder: empirical observation, minimal computational inference, and AI design conjecture. We introduce a minimal hierarchical partially observed control model with latent dynamics, structured episodic memory, observer-belief state, option-level actions, and delayed verifier signals. This motivates three hypotheses: (H1) fast local feedback plus predictive compensation improves robustness under hidden dynamics shifts; (H2) memory organized for future control improves delayed retrieval under cue conflict and load; and (H3) verifiers and observer models inside the action-memory loop reduce silent failure and information leakage while remaining vulnerable to misspecification. A downstream conjecture is that role-differentiated proposer/executor/checker/adversary systems may reduce correlated error under asymmetric information and verification burden. The contribution is a comparative perspective and benchmark agenda: a disciplined program of falsifiable claims about the coupling of control, memory, and verifiable action.

Method: Multi-agent framework with “Reasonable Mind” Agent for evidence-based logic and “Emotional Mind” Agent for empathetic communication, guided by DSM-5-structured knowledge graph to reduce hallucinations and steer diagnostic inquiries.

Result: Achieved 85.6% top-1 diagnostic accuracy across 1206 simulated conversations and 180 real user sessions, outperforming state-of-the-art LLM methods by 15-54 percentage points and approaching board-certified psychiatrist performance ranges.

Conclusion: WiseMind demonstrates feasibility of empathetic, reliable AI agents for psychiatric assessment under human oversight, generating clinically sound and psychologically supportive responses validated by expert psychiatrists.

Abstract: Large Language Models (LLMs) offer promising opportunities to support mental healthcare workflows, yet they often lack the structured clinical reasoning needed for reliable diagnosis and may struggle to provide the emotionally attuned communication essential for patient trust. Here, we introduce WiseMind, a novel multi-agent framework inspired by the theory of Dialectical Behavior Therapy designed to facilitate psychiatric assessment. By integrating a “Reasonable Mind” Agent for evidence-based logic and an “Emotional Mind” Agent for empathetic communication, WiseMind effectively bridges the gap between instrumental accuracy and humanistic care. Our framework utilizes a Diagnostic and Statistical Manual of Mental Disorders, Fifth Edition (DSM-5)-guided Structured Knowledge Graph to steer diagnostic inquiries, significantly reducing hallucinations compared to standard prompting methods. Using a combination of virtual standard patients, simulated interactions, and real human interaction datasets, we evaluate WiseMind across three common psychiatric conditions. WiseMind outperforms state-of-the-art LLM methods in both identifying critical diagnostic nodes and establishing accurate differential diagnoses. Across 1206 simulated conversations and 180 real user sessions, the system achieves 85.6% top-1 diagnostic accuracy, approaching reported diagnostic performance ranges of board-certified psychiatrists and surpassing knowledge-enhanced single-agent baselines by 15-54 percentage points. Expert review by psychiatrists further validates that WiseMind generates responses that are not only clinically sound but also psychologically supportive, demonstrating the feasibility of empathetic, reliable AI agents to conduct psychiatric assessments under appropriate human oversight.

Method: AutoCO introduces a unified triple-representation that binds relaxation strategies, algorithmic principles, and executable codes. It employs bidirectional global-local coevolution combining Monte Carlo Tree Search (MCTS) for global relaxation-trajectory exploration with Evolutionary Algorithms (EAs) for local solution intensification.

Result: Extensive experiments on three challenging COP benchmarks show AutoCO’s consistent effectiveness and superior performance, especially in hard regimes where current methods degrade.

Conclusion: AutoCO represents a principled and effective path toward proactive, verifiable LLM-driven optimization that tightly couples operations-research principles with LLM reasoning.

Abstract: Large Language Model (LLM)-based optimization has recently shown promise for autonomous problem solving, yet most approaches still cast LLMs as passive constraint checkers rather than proactive strategy designers, limiting their effectiveness on complex Constraint Optimization Problems (COPs). To address this, we present AutoCO, an end-to-end Automated Constraint Optimization method that tightly couples operations-research principles of constraint relaxation with LLM reasoning. A core innovation is a unified triple-representation that binds relaxation strategies, algorithmic principles, and executable codes. This design enables the LLM to synthesize, justify, and instantiate relaxation strategies that are both principled and executable. To navigate fragmented solution spaces, AutoCO employs a bidirectional global-local coevolution mechanism, synergistically coupling Monte Carlo Tree Search (MCTS) for global relaxation-trajectory exploration with Evolutionary Algorithms (EAs) for local solution intensification. This continuous exchange of priors and feedback explicitly balances diversification and intensification, thus preventing premature convergence. Extensive experiments on three challenging COP benchmarks validate AutoCO’s consistent effectiveness and superior performance, especially in hard regimes where current methods degrade. Results highlight AutoCO as a principled and effective path toward proactive, verifiable LLM-driven optimization.

Method: Proposes ForenAgent with a two-stage training pipeline (Cold Start + Reinforcement Fine-Tuning) and a dynamic reasoning loop (global perception, local focusing, iterative probing, holistic adjudication). Uses FABench dataset with 100k images and 200k agent-interaction QA pairs for training/evaluation.

Result: ForenAgent exhibits emergent tool-use competence and reflective reasoning on challenging IFD tasks when assisted by low-level tools, showing promising results for general-purpose image forgery detection.

Conclusion: The framework successfully unifies high-level semantic knowledge from MLLMs with low-level forensic artifacts through autonomous tool generation and execution, charting a promising route toward general-purpose image forgery detection.

Abstract: Existing image forgery detection (IFD) methods either exploit low-level, semantics-agnostic artifacts or rely on multimodal large language models (MLLMs) with high-level semantic knowledge. Although naturally complementary, these two information streams are highly heterogeneous in both paradigm and reasoning, making it difficult for existing methods to unify them or effectively model their cross-level interactions. To address this gap, we propose ForenAgent, a multi-round interactive IFD framework that enables MLLMs to autonomously generate, execute, and iteratively refine Python-based low-level tools around the detection objective, thereby achieving more flexible and interpretable forgery analysis. ForenAgent follows a two-stage training pipeline combining Cold Start and Reinforcement Fine-Tuning to enhance its tool interaction capability and reasoning adaptability progressively. Inspired by human reasoning, we design a dynamic reasoning loop comprising global perception, local focusing, iterative probing, and holistic adjudication, and instantiate it as both a data-sampling strategy and a task-aligned process reward. For systematic training and evaluation, we construct FABench, a heterogeneous, high-quality agent-forensics dataset comprising 100k images and approximately 200k agent-interaction question-answer pairs. Experiments show that ForenAgent exhibits emergent tool-use competence and reflective reasoning on challenging IFD tasks when assisted by low-level tools, charting a promising route toward general-purpose IFD. The code will be released after the review process is completed.

Method: AgenticRed treats red-teaming as a system design problem and uses LLMs’ in-context learning to iteratively design and refine red-teaming systems without human intervention. It employs evolutionary selection and generational knowledge to autonomously evolve automated red-teaming systems.

Result: Red-teaming systems designed by AgenticRed consistently outperform state-of-the-art approaches, achieving 96% attack success rate (ASR) on Llama-2-7B, 98% on Llama-3-8B, and 100% on Qwen3-8B on HarmBench. The approach generates robust, query-agnostic systems that transfer strongly to proprietary models, achieving 100% ASR on GPT-5.1, DeepSeek-R1, and DeepSeek V3.2.

Conclusion: Evolutionary algorithms represent a powerful approach to AI safety that can keep pace with rapidly evolving models, enabling the autonomous design of highly effective red-teaming systems without human intervention.

Abstract: While recent automated red-teaming methods show promise for systematically exposing model vulnerabilities, most existing approaches rely on human-specified workflows. This dependence on manually designed workflows suffers from human biases and makes exploring the broader design space expensive. We introduce AgenticRed, an automated pipeline that leverages LLMs’ in-context learning to iteratively design and refine red-teaming systems without human intervention. Rather than optimizing attacker policies within predefined structures, AgenticRed treats red-teaming as a system design problem, and it autonomously evolves automated red-teaming systems using evolutionary selection and generational knowledge. Red-teaming systems designed by AgenticRed consistently outperform state-of-the-art approaches, achieving 96% attack success rate (ASR) on Llama-2-7B, 98% on Llama-3-8B and 100% on Qwen3-8B on HarmBench. Our approach generates robust, query-agnostic red-teaming systems that transfer strongly to the latest proprietary models, achieving an impressive 100% ASR on GPT-5.1, DeepSeek-R1 and DeepSeek V3.2. This work highlights evolutionary algorithms as a powerful approach to AI safety that can keep pace with rapidly evolving models.

Method: Created ContextMATH benchmark by repurposing AIME and MATH-500 problems into two contextual settings: Scenario Grounding (SG) embeds abstract problems into realistic narratives, and Complexity Scaling (CS) transforms explicit conditions into sub-problems. Evaluated 61 proprietary and open-source models.

Result: Significant performance drops observed: open-source models declined by 13 and 34 points on SG and CS, proprietary models by 13 and 20 points. Errors dominated by incorrect problem formulation, with formulation accuracy declining as original problem difficulty increases.

Conclusion: Formulation and reasoning remain complementary bottlenecks for contextual mathematical problem solving. Fine-tuning with scenario data helps but only partially alleviates performance gaps, highlighting contextual mathematical reasoning as a central unsolved challenge for LLMs.

Abstract: Large language models now solve many benchmark math problems at near-expert levels, yet this progress has not fully translated into reliable performance in real-world applications. We study this gap through contextual mathematical reasoning, where the mathematical core must be formulated from descriptive scenarios. We introduce ContextMATH, a benchmark that repurposes AIME and MATH-500 problems into two contextual settings: Scenario Grounding (SG), which embeds abstract problems into realistic narratives without increasing reasoning complexity, and Complexity Scaling (CS), which transforms explicit conditions into sub-problems to capture how constraints often appear in practice. Evaluating 61 proprietary and open-source models, we observe sharp drops: on average, open-source models decline by 13 and 34 points on SG and CS, while proprietary models drop by 13 and 20. Error analysis shows that errors are dominated by incorrect problem formulation, with formulation accuracy declining as original problem difficulty increases. Correct formulation emerges as a prerequisite for success, and its sufficiency improves with model scale, indicating that larger models advance in both understanding and reasoning. Nevertheless, formulation and reasoning remain two complementary bottlenecks that limit contextual mathematical problem solving. Finally, we find that fine-tuning with scenario data improves performance, whereas formulation-only training is ineffective. However, performance gaps are only partially alleviated, highlighting contextual mathematical reasoning as a central unsolved challenge for LLMs.

Method: Comprehensive survey organized across three dimensions: synthetic data for perception/planning, digital twin-based simulation for validation, and domain adaptation strategies; includes taxonomy of datasets, tools, and platforms.

Result: Provides detailed landscape analysis of synthetic data and simulation technologies in autonomous driving, highlighting trends in benchmark design and the role of vision-language models in scene understanding.

Conclusion: Identifies critical challenges including Sim2Real transfer, scalable safety validation, cooperative autonomy, and simulation-driven policy learning as key research directions for advancing autonomous driving systems.

Abstract: Autonomous driving technologies have achieved significant advances in recent years, yet their real-world deployment remains constrained by data scarcity, safety requirements, and the need for generalization across diverse environments. In response, synthetic data and virtual environments have emerged as powerful enablers, offering scalable, controllable, and richly annotated scenarios for training and evaluation. This survey presents a comprehensive review of recent developments at the intersection of autonomous driving, simulation technologies, and synthetic datasets. We organize the landscape across three core dimensions: (i) the use of synthetic data for perception and planning, (ii) digital twin-based simulation for system validation, and (iii) domain adaptation strategies bridging synthetic and real-world data. We also highlight the role of vision-language models and simulation realism in enhancing scene understanding and generalization. A detailed taxonomy of datasets, tools, and simulation platforms is provided, alongside an analysis of trends in benchmark design. Finally, we discuss critical challenges and open research directions, including Sim2Real transfer, scalable safety validation, cooperative autonomy, and simulation-driven policy learning, that must be addressed to accelerate the path toward safe, generalizable, and globally deployable autonomous driving systems.

Method: OSCAR runs N parallel denoising chains with randomized reveal orders, computes cross-chain Shannon entropy to detect high-uncertainty positions, and performs targeted remasking conditioned on retrieved evidence. It introduces commitment uncertainty localization to identify token positions where cross-chain entropy exceeds an unsupervised threshold before unreliable commitments propagate.

Result: On TriviaQA, HotpotQA, RAGTruth, and CommonsenseQA using LLaDA-8B and Dream-7B, OSCAR significantly reduces hallucinated content and improves factual accuracy through uncertainty-guided remasking. The framework enhances generation quality and facilitates more effective integration of retrieved evidence.

Conclusion: Diffusion language models have an inherent capacity to identify factual uncertainty through their denoising trajectories, which surpasses specialized trained detectors. OSCAR demonstrates that native entropy-based uncertainty signals in DLMs can effectively guide hallucination mitigation without requiring external classifiers.

Abstract: Diffusion language models (DLMs) expose their denoising trajectories, offering a natural handle for inference-time control; accordingly, an ideal hallucination mitigation framework should intervene during generation using this model-native signal rather than relying on an externally trained hallucination classifier. Toward this, we formulate commitment uncertainty localization: given a denoising trajectory, identify token positions whose cross-chain entropy exceeds an unsupervised threshold before factually unreliable commitments propagate into self-consistent but incorrect outputs. We introduce a suite of trajectory-level assessments, including a cross-chain divergence-at-hallucination (CDH) metric, for principled comparison of localization methods. We also introduce OSCAR, a training-free inference-time framework operationalizing this formulation. OSCAR runs N parallel denoising chains with randomized reveal orders, computes cross-chain Shannon entropy to detect high-uncertainty positions, and then performs targeted remasking conditioned on retrieved evidence. Ablations confirm that localization and correction contribute complementary gains, robust across N in {4, 8, 16}. On TriviaQA, HotpotQA, RAGTruth, and CommonsenseQA using LLaDA-8B and Dream-7B, OSCAR enhances generation quality by significantly reducing hallucinated content and improving factual accuracy through uncertainty-guided remasking, which also facilitates more effective integration of retrieved evidence. Its native entropy-based uncertainty signal surpasses that of specialized trained detectors, highlighting an inherent capacity of diffusion language models to identify factual uncertainty that is not present in the sequential token commitment structure of autoregressive models.

Method: Systematically redesign input-output format and fine-tune models on synthesized data with complex permission topologies. Forces models to generate structured authorization trajectories as causal prerequisites for any substantive response or action, internalizing access boundaries within reasoning.

Result: CoA maintains high utility in authorized scenarios while achieving high rejection rates of unauthorized prompts and robust defense against diverse adversarial attacks.

Conclusion: By embedding authorization directly into the reasoning process, CoA provides a principled architectural blueprint for deploying secure LLMs as cognitive cores of modern AI systems.

Abstract: Although Large Language Models (LLMs) have evolved from text generators into the cognitive core of modern AI systems, their inherent lack of authorization awareness exposes these systems to catastrophic risks, ranging from unintentional data leakage to unauthorized command execution. Existing defense mechanisms are fundamentally decoupled from internal reasoning, rendering them insufficient for the complex security demands of dynamic AI systems. Here, we propose the Chain-of-Authorization (CoA) framework, a paradigm that internalizes access control as a foundational cognitive capability. By systematically redesigning the input-output format and fine-tuning the model on synthesized data with complex permission topologies, CoA forces the model to generate a structured authorization trajectory as a causal prerequisite for any substantive response or action, thereby enabling LLMs to internalize access boundaries within dynamic reasoning environments. CoA maintains high utility in authorized scenarios while achieving high rejection rates of unauthorized prompts and robust defense against diverse adversarial attacks. By embedding authorization directly into the reasoning process, CoA provides a principled architectural blueprint for deploying secure LLMs as the cognitive cores of modern AI systems.

Method: Developed a framework using context-specific human-AI interaction studies with 10,101 participants across three domains (public policy, finance, health) and three locales (US, UK, India)

Result: AI models can produce manipulative behaviors and induce belief/behavior changes; context matters significantly with differences across domains and geographies; propensity not consistently predictive of efficacy

Conclusion: Importance of evaluating AI manipulation in specific high-stakes contexts, with geographic considerations, and separating propensity from efficacy measurements

Abstract: Interest in the concept of AI-driven harmful manipulation is growing, yet current approaches to evaluating it are limited. This paper introduces a framework for evaluating harmful AI manipulation via context-specific human-AI interaction studies. We illustrate the utility of this framework by assessing an AI model with 10,101 participants spanning interactions in three AI use domains (public policy, finance, and health) and three locales (US, UK, and India). Overall, we find that that the tested model can produce manipulative behaviours when prompted to do so and, in experimental settings, is able to induce belief and behaviour changes in study participants. We further find that context matters: AI manipulation differs between domains, suggesting that it needs to be evaluated in the high-stakes context(s) in which an AI system is likely to be used. We also identify significant differences across our tested geographies, suggesting that AI manipulation results from one geographic region may not generalise to others. Finally, we find that the frequency of manipulative behaviours (propensity) of an AI model is not consistently predictive of the likelihood of manipulative success (efficacy), underscoring the importance of studying these dimensions separately. To facilitate adoption of our evaluation framework, we detail our testing protocols and make relevant materials publicly available. We conclude by discussing open challenges in evaluating harmful manipulation by AI models.

Method: PAPO integrates process-level evaluation into Group Relative Policy Optimization (GRPO) through decoupled advantage normalization. It composes advantage from an outcome component Aout (derived from ORM and normalized over all responses) and a process component Aproc (derived from a rubric-based PRM and normalized exclusively among correct responses). This decoupled design ensures Aout anchors training on correctness while Aproc differentiates reasoning quality without distorting the outcome signal.

Result: Experiments across multiple model scales and six benchmarks show PAPO consistently outperforms ORM, reaching 51.3% vs. 46.3% on OlympiadBench while continuing to improve as ORM plateaus and declines.

Conclusion: PAPO effectively addresses limitations of both outcome-only and process-only reward models by combining them through decoupled advantage normalization, leading to better performance and continued improvement where ORM-based methods plateau.

Abstract: We propose Process-Aware Policy Optimization (PAPO), a method that integrates process-level evaluation into Group Relative Policy Optimization (GRPO) through decoupled advantage normalization, to address two limitations of existing reward designs. Outcome reward models (ORM) evaluate only final-answer correctness, treating all correct responses identically regardless of reasoning quality, and gradually lose the advantage signal as groups become uniformly correct. Process reward models (PRM) offer richer supervision, but directly using PRM scores causes reward hacking, where models exploit verbosity to inflate scores while accuracy collapses. PAPO resolves both by composing the advantage from an outcome component Aout, derived from ORM and normalized over all responses, and a process component Aproc, derived from a rubric-based PRM and normalized exclusively among correct responses. This decoupled design ensures that Aout anchors training on correctness while Aproc differentiates reasoning quality without distorting the outcome signal. Experiments across multiple model scales and six benchmarks demonstrate that PAPO consistently outperforms ORM, reaching 51.3% vs.\ 46.3% on OlympiadBench while continuing to improve as ORM plateaus and declines.

Method: HERA uses a hierarchical framework with two levels: 1) Global level: optimizes query-specific agent topologies through reward-guided sampling and experience accumulation; 2) Local level: Role-Aware Prompt Evolution refines agent behaviors via credit assignment and dual-axes adaptation along operational and behavioral principles.

Result: On six knowledge-intensive benchmarks, HERA achieves an average improvement of 38.69% over recent baselines while maintaining robust generalization and token efficiency. Topological analyses reveal emergent self-organization with sparse exploration yielding compact, high-utility multi-agent networks.

Conclusion: HERA demonstrates both efficient coordination and robust reasoning through adaptive multi-agent orchestration and role-specific prompt evolution, addressing limitations of static approaches in multi-agent RAG systems.

Abstract: Multi-agent Retrieval-Augmented Generation (RAG), wherein each agent takes on a specific role, supports hard queries that require multiple steps and sources, or complex reasoning. Existing approaches, however, rely on static agent behaviors and fixed orchestration strategies, leading to brittle performance on diverse, multi-hop tasks. We identify two key limitations: the lack of continuously adaptive orchestration mechanisms and the absence of behavior-level learning for individual agents. To this end, we propose HERA, a hierarchical framework that jointly evolves multi-agent orchestration and role-specific agent prompts. At the global level, HERA optimizes query-specific agent topologies through reward-guided sampling and experience accumulation. At the local level, Role-Aware Prompt Evolution refines agent behaviors via credit assignment and dual-axes adaptation along operational and behavioral principles, enabling targeted, role-conditioned improvements. On six knowledge-intensive benchmarks, HERA achieves an average improvement of 38.69% over recent baselines while maintaining robust generalization and token efficiency. Topological analyses reveal emergent self-organization, where sparse exploration yields compact, high-utility multi-agent networks, demonstrating both efficient coordination and robust reasoning.

Method: Use linear probes to decode tool-calling decisions from pre-generation activations, apply activation steering to perturb decision directions, and analyze behavioral patterns when decisions are flipped.

Result: Linear probes successfully decode decisions from pre-generation activations with high confidence, sometimes before any reasoning tokens. Activation steering inflates deliberation and flips behavior in 7-79% of cases. When decisions flip, chain-of-thought often rationalizes rather than resists the change.

Conclusion: Reasoning models can encode action choices before beginning textual deliberation, supporting a “decide-then-think” model where decisions shape subsequent reasoning processes.

Abstract: We consider the question: when a large language reasoning model makes a choice, did it think first and then decide to, or decide first and then think? In this paper, we present evidence that detectable, early-encoded decisions shape chain-of-thought in reasoning models. Specifically, we show that a simple linear probe successfully decodes tool-calling decisions from pre-generation activations with very high confidence, and in some cases, even before a single reasoning token is produced. Activation steering supports this causally: perturbing the decision direction leads to inflated deliberation, and flips behavior in many examples (between 7 - 79% depending on model and benchmark). We also show through behavioral analysis that, when steering changes the decision, the chain-of-thought process often rationalizes the flip rather than resisting it. Together, these results suggest that reasoning models can encode action choices before they begin to deliberate in text.

Method: Controlled simulated scenarios involving drug interactions and dosage decisions, analyzing different failure types (missed interactions, incorrect risk flagging, inappropriate dosage recommendations). Focuses on how errors occur and consequences of incorrect outputs.

Result: AI errors in medication contexts can lead to adverse drug reactions, ineffective treatment, or delayed care, especially without sufficient human oversight. Over-reliance on AI recommendations and limited transparency pose significant risks.

Conclusion: Need reliability-focused AI evaluation in healthcare, complementing traditional metrics with risk-aware approaches. Understanding failure behavior and real-world impact is crucial for safety-critical domains like pharmacy practice.

Abstract: Artificial intelligence (AI) systems are increasingly integrated into healthcare and pharmacy workflows, supporting tasks such as medication recommendations, dosage determination, and drug interaction detection. While these systems often demonstrate strong performance under standard evaluation metrics, their reliability in real-world decision-making remains insufficiently understood. In high-risk domains such as medication management, even a single incorrect recommendation can result in severe patient harm. This paper examines the reliability of AI-assisted medication systems by focusing on system failures and their potential clinical consequences. Rather than evaluating performance solely through aggregate metrics, this work shifts attention towards how errors occur and what happens when AI systems produce incorrect outputs. Through a series of controlled, simulated scenarios involving drug interactions and dosage decisions, we analyse different types of system failures, including missed interactions, incorrect risk flagging, and inappropriate dosage recommendations. The findings highlight that AI errors in medication-related contexts can lead to adverse drug reactions, ineffective treatment, or delayed care, particularly when systems are used without sufficient human oversight. Furthermore, the paper discusses the risks of over-reliance on AI recommendations and the challenges posed by limited transparency in decision-making processes. This work contributes a reliability-focused perspective on AI evaluation in healthcare, emphasising the importance of understanding failure behavior and real-world impact. It highlights the need to complement traditional performance metrics with risk-aware evaluation approaches, particularly in safety-critical domains such as pharmacy practice.

Method: SAT-based approach with demand expansion (each item copy gets variables), sheet assignment variables, non-overlap constraints only for same-sheet items, and infeasible-orientation elimination rules. Three minimization approaches: non-incremental SAT with binary search, incremental SAT with clause reuse, and weighted partial MaxSAT.

Result: Best SAT configurations certify 2-3x more instances as provably optimal and achieve lower optimality gaps than OR-Tools, CPLEX, and Gurobi on Cui-Zhao benchmarks. Incremental SAT strongest without rotation; non-incremental better with rotation.

Conclusion: SAT-based methods effectively handle combinatorial complexity of 2D cutting stock problems, with different approaches optimal depending on rotation constraints.

Abstract: Cutting rectangular items from stock sheets to satisfy demands while minimizing waste is a central manufacturing task. The Two-Dimensional Single Stock Size Cutting Stock Problem (2D-CSSP) generalizes bin packing by requiring multiple copies of each item type, which causes a strong combinatorial blow-up. We present a SAT-based framework where item types are expanded by demand, each copy has a sheet-assignment variable and non-overlap constraints are activated only for copies assigned to the same sheet. We also introduce an infeasible-orientation elimination rule that fixes rotation variables when only one orientation can fit the sheet. For minimizing the number of sheets, we compare three approaches: non-incremental SAT with binary search, incremental SAT with clause reuse across iterations and weighted partial MaxSAT. On the Cui–Zhao benchmark suite, our best SAT configurations certify two to three times more instances as provably optimal and achieve lower optimality gaps than OR-Tools, CPLEX and Gurobi. The relative ranking among SAT approaches depends on rotation: incremental SAT is strongest without rotation, while non-incremental SAT is more effective when rotation increases formula size.

Method: Proposes user-turn generation as a probe: given conversation context (user query + assistant response), models generate under the user role. Tests 11 open-weight LLMs across 5 datasets (math reasoning, instruction following, conversation) with controlled perturbations and temperature sampling.

Result: Interaction awareness is decoupled from task accuracy. Qwen3.5 family shows GSM8K accuracy scaling from 41% to 96.8%, but genuine follow-up rates remain near zero under deterministic generation. Higher temperature sampling reveals latent interaction awareness with follow-up rates reaching 22%. Controlled perturbations validate the probe measures real model properties.

Conclusion: User-turn generation captures interaction awareness as a distinct dimension of LLM behavior unexplored by current assistant-only benchmarks, revealing latent capabilities that emerge under different sampling strategies.

Abstract: Standard LLM benchmarks evaluate the assistant turn: the model generates a response to an input, a verifier scores correctness, and the analysis ends. This paradigm leaves unmeasured whether the LLM encodes any awareness of what follows the assistant response. We propose user-turn generation as a probe of this gap: given a conversation context of user query and assistant response, we let a model generate under the user role. If the model’s weights encode interaction awareness, the generated user turn will be a grounded follow-up that reacts to the preceding context. Through experiments across $11$ open-weight LLMs (Qwen3.5, gpt-oss, GLM) and $5$ datasets (math reasoning, instruction following, conversation), we show that interaction awareness is decoupled from task accuracy. In particular, within the Qwen3.5 family, GSM8K accuracy scales from $41%$ ($0.8$B) to $96.8%$ ($397$B-A$17$B), yet genuine follow-up rates under deterministic generation remain near zero. In contrast, higher temperature sampling reveals interaction awareness is latent with follow up rates reaching $22%$. Controlled perturbations validate that the proposed probe measures a real property of the model, and collaboration-oriented post-training on Qwen3.5-2B demonstrates an increase in follow-up rates. Our results show that user-turn generation captures a dimension of LLM behavior, interaction awareness, that is unexplored and invisible with current assistant-only benchmarks.

Method: Cannot analyze method as paper content is unavailable due to HTTP 429 error

Result: No results available due to technical error in accessing paper information

Conclusion: Cannot provide analysis due to API rate limiting preventing access to the paper abstract

Abstract: Failed to fetch summary for 2410.13891: Page request resulted in HTTP 429 (https://export.arxiv.org/api/query?search_query=&id_list=2410.13891&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 2505.17938: Page request resulted in HTTP 429 (https://export.arxiv.org/api/query?search_query=&id_list=2505.17938&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 2506.16255: Page request resulted in HTTP 429 (https://export.arxiv.org/api/query?search_query=&id_list=2506.16255&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 determine conclusion due to failed API request

Abstract: Failed to fetch summary for 2509.09192: Page request resulted in HTTP 429 (https://export.arxiv.org/api/query?search_query=&id_list=2509.09192&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 draw conclusions due to inability to access paper content

Abstract: Failed to fetch summary for 2509.25438: Page request resulted in HTTP 429 (https://export.arxiv.org/api/query?search_query=&id_list=2509.25438&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 2510.00361: Page request resulted in HTTP 429 (https://export.arxiv.org/api/query?search_query=&id_list=2510.00361&sortBy=relevance&sortOrder=descending&start=0&max_results=100)

Method: Method information not available

Result: Results cannot be assessed

Conclusion: Cannot draw conclusions without paper content

Abstract: Failed to fetch summary for 2510.05497: Page request resulted in HTTP 429 (https://export.arxiv.org/api/query?search_query=&id_list=2510.05497&sortBy=relevance&sortOrder=descending&start=0&max_results=100)

Method: Unable to determine method due to API access limitations

Result: Unable to determine results due to API access limitations

Conclusion: Unable to analyze paper due to technical limitations in accessing arXiv data

Abstract: Failed to fetch summary for 2510.06649: Page request resulted in HTTP 429 (https://export.arxiv.org/api/query?search_query=&id_list=2510.06649&sortBy=relevance&sortOrder=descending&start=0&max_results=100)

Method: Cannot determine method as abstract could not be retrieved

Result: Cannot determine results as abstract could not be retrieved

Conclusion: Cannot draw conclusions without access to the paper’s abstract

Abstract: Failed to fetch summary for 2510.10510: Page request resulted in HTTP 429 (https://export.arxiv.org/api/query?search_query=&id_list=2510.10510&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 2510.20847: Page request resulted in HTTP 429 (https://export.arxiv.org/api/query?search_query=&id_list=2510.20847&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 2511.06731: Page request resulted in HTTP 429 (https://export.arxiv.org/api/query?search_query=&id_list=2511.06731&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 2512.13907: Page request resulted in HTTP 429 (https://export.arxiv.org/api/query?search_query=&id_list=2512.13907&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 2601.08845: Page request resulted in HTTP 429 (https://export.arxiv.org/api/query?search_query=&id_list=2601.08845&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 2601.13508: Page request resulted in HTTP 429 (https://export.arxiv.org/api/query?search_query=&id_list=2601.13508&sortBy=relevance&sortOrder=descending&start=0&max_results=100)

Method: Cannot determine method as paper content was not accessible

Result: Cannot determine results as paper content was not accessible

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

Abstract: Failed to fetch summary for 2601.21064: Page request resulted in HTTP 429 (https://export.arxiv.org/api/query?search_query=&id_list=2601.21064&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 paper content due to access limitations

Abstract: Failed to fetch summary for 2602.09987: Page request resulted in HTTP 429 (https://export.arxiv.org/api/query?search_query=&id_list=2602.09987&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 2602.10419: Page request resulted in HTTP 429 (https://export.arxiv.org/api/query?search_query=&id_list=2602.10419&sortBy=relevance&sortOrder=descending&start=0&max_results=100)

Method: Unable to determine method due to API access limitations preventing retrieval of paper content.

Result: Unable to determine results due to API access limitations preventing retrieval of paper content.

Conclusion: Unable to draw conclusions about the paper’s content due to technical limitations in accessing the arXiv API.

Abstract: Failed to fetch summary for 2602.16746: Page request resulted in HTTP 429 (https://export.arxiv.org/api/query?search_query=&id_list=2602.16746&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 2602.16967: Page request resulted in HTTP 429 (https://export.arxiv.org/api/query?search_query=&id_list=2602.16967&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 2602.18523: Page request resulted in HTTP 429 (https://export.arxiv.org/api/query?search_query=&id_list=2602.18523&sortBy=relevance&sortOrder=descending&start=0&max_results=100)

Method: Method unknown - paper content not accessible due to HTTP 429 error from arXiv API

Result: No results available - failed to fetch paper summary from arXiv

Conclusion: Cannot analyze paper due to technical limitations in accessing the abstract

Abstract: Failed to fetch summary for 2603.01589: Page request resulted in HTTP 429 (https://export.arxiv.org/api/query?search_query=&id_list=2603.01589&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 2603.05917: Page request resulted in HTTP 429 (https://export.arxiv.org/api/query?search_query=&id_list=2603.05917&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 2603.18109: Page request resulted in HTTP 429 (https://export.arxiv.org/api/query?search_query=&id_list=2603.18109&sortBy=relevance&sortOrder=descending&start=0&max_results=100)

Method: Unable to determine method due to API rate limiting error

Result: Unable to determine results due to API rate limiting error

Conclusion: Unable to determine conclusion due to API rate limiting error

Abstract: Failed to fetch summary for 2603.20266: Page request resulted in HTTP 429 (https://export.arxiv.org/api/query?search_query=&id_list=2603.20266&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 2603.21991: Page request resulted in HTTP 429 (https://export.arxiv.org/api/query?search_query=&id_list=2603.21991&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 2603.28204: Page request resulted in HTTP 429 (https://export.arxiv.org/api/query?search_query=&id_list=2603.28204&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 2603.25764: Page request resulted in HTTP 429 (https://export.arxiv.org/api/query?search_query=&id_list=2603.25764&sortBy=relevance&sortOrder=descending&start=0&max_results=100)

Method: Cannot determine method due to missing paper content

Result: Cannot determine results due to missing paper content

Conclusion: Cannot determine conclusion due to missing paper content

Abstract: Failed to fetch summary for 2603.28964: Page request resulted in HTTP 429 (https://export.arxiv.org/api/query?search_query=&id_list=2603.28964&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 2603.28013: Page request resulted in HTTP 429 (https://export.arxiv.org/api/query?search_query=&id_list=2603.28013&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 2604.01021: Page request resulted in HTTP 429 (https://export.arxiv.org/api/query?search_query=&id_list=2604.01021&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 2603.29585: Page request resulted in HTTP 429 (https://export.arxiv.org/api/query?search_query=&id_list=2603.29585&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 2604.01527: Page request resulted in HTTP 429 (https://export.arxiv.org/api/query?search_query=&id_list=2604.01527&sortBy=relevance&sortOrder=descending&start=0&max_results=100)

Method: Created dataset with large spoof subset from 37 Russian-capable TTS/voice-cloning systems and bona fide subset from multiple Russian speech corpora. Includes configurable simulations of platform/transmission distortions (room reverberation, additive noise/music, speech-codec transcodings) via unified processing chain.

Result: Benchmarked diverse anti-spoofing countermeasures including lightweight supervised architectures, graph-attention models, SSL-based detectors, and large-scale pretrained systems. Reported reference results on both clean and simulated conditions to characterize robustness under realistic perturbation pipelines.

Conclusion: RuASD provides a comprehensive, reproducible benchmark for Russian-language speech anti-spoofing research, enabling systematic evaluation of model robustness to realistic deployment conditions and distribution shifts.

Abstract: RuASD (Russian AntiSpoofing Dataset) is a dedicated, reproducible benchmark for Russian-language speech anti-spoofing designed to evaluate both in-domain discrimination and robustness to deployment-style distribution shifts. It combines a large spoof subset synthesized using 37 modern Russian-capable TTS and voice-cloning systems with a bona fide subset curated from multiple heterogeneous open Russian speech corpora, enabling systematic evaluation across diverse data sources. To emulate typical dissemination and channel effects in a controlled and reproducible manner, RuASD includes configurable simulations of platform and transmission distortions, including room reverberation, additive noise/music, and a range of speech-codec transcodings implemented via a unified processing chain. We benchmark a diverse set of publicly available anti-spoofing countermeasures spanning lightweight supervised architectures, graph-attention models, SSL-based detectors, and large-scale pretrained systems, and report reference results on both clean and simulated conditions to characterize robustness under realistic perturbation pipelines. The dataset is publickly available at \href{https://huggingface.co/datasets/MTUCI/RuASD}{\underline{Hugging Face}} and \href{https://modelscope.cn/datasets/lab260/RuASD}{\underline{ModelScope}}.

Method: Designs audio spatial feature encoder with audio intensity attention mechanism to extract target-related spatial state information, then introduces Audio Spatial State Guided Fusion (ASGF) for dynamic alignment and adaptive fusion of multimodal features.

Result: Experimental results on Replica and Matterport3D datasets show the method is particularly effective on unheard tasks, demonstrating improved generalization under unknown sound source distributions.

Conclusion: The proposed ASGF method effectively addresses generalization challenges in audio-visual navigation by leveraging audio spatial guidance and adaptive multimodal fusion.

Abstract: Audio-visual Navigation refers to an agent utilizing visual and auditory information in complex 3D environments to accomplish target localization and path planning, thereby achieving autonomous navigation. The core challenge of this task lies in the following: how the agent can break free from the dependence on training data and achieve autonomous navigation with good generalization performance when facing changes in environments and sound sources. To address this challenge, we propose an Audio Spatially-Guided Fusion for Audio-Visual Navigation method. First, we design an audio spatial feature encoder, which adaptively extracts target-related spatial state information through an audio intensity attention mechanism; based on this, we introduce an Audio Spatial State Guided Fusion (ASGF) to achieve dynamic alignment and adaptive fusion of multimodal features, effectively alleviating noise interference caused by perceptual uncertainty. Experimental results on the Replica and Matterport3D datasets indicate that our method is particularly effective on unheard tasks, demonstrating improved generalization under unknown sound source distributions.

Method: SACF discretizes target’s relative direction and distance from audio-visual cues, predicts their distributions, encodes them as compact descriptors, then uses audio embeddings and spatial descriptors to generate channel-wise scaling/bias to modulate visual features via conditional linear transformation.

Result: SACF improves navigation efficiency with lower computational overhead and generalizes well to unheard target sounds.

Conclusion: Spatial-aware conditioned fusion with explicit discrete position representation enhances audio-visual navigation performance and generalization.

Abstract: Audio-visual navigation tasks require agents to locate and navigate toward continuously vocalizing targets using only visual observations and acoustic cues. However, existing methods mainly rely on simple feature concatenation or late fusion, and lack an explicit discrete representation of the target’s relative position, which limits learning efficiency and generalization. We propose Spatial-Aware Conditioned Fusion (SACF). SACF first discretizes the target’s relative direction and distance from audio-visual cues, predicts their distributions, and encodes them as a compact descriptor for policy conditioning and state modeling. Then, SACF uses audio embeddings and spatial descriptors to generate channel-wise scaling and bias to modulate visual features via conditional linear transformation, producing target-oriented fused representations. SACF improves navigation efficiency with lower computational overhead and generalizes well to unheard target sounds.

Method: Proposes RAVN with Acoustic Geometry Reasoner (AGR) trained with geometric proxy supervision using heteroscedastic Gaussian NLL to learn observation-dependent dispersion as reliability cues. Introduces Reliability-Aware Geometric Modulation (RAGM) to convert learned cues into soft gates modulating visual features.

Result: RAVN shows consistent improvements in navigation performance on SoundSpaces using Replica and Matterport3D environments, with notable robustness in challenging unheard sound settings.

Conclusion: The reliability-aware approach effectively handles audio unreliability in complex acoustic environments, improving audio-visual navigation performance without requiring geometric labels during inference.

Abstract: Audio-Visual Navigation (AVN) requires an embodied agent to navigate toward a sound source by utilizing both vision and binaural audio. A core challenge arises in complex acoustic environments, where binaural cues become intermittently unreliable, particularly when generalizing to previously unheard sound categories. To address this, we propose RAVN (Reliability-Aware Audio-Visual Navigation), a framework that conditions cross-modal fusion on audio-derived reliability cues, dynamically calibrating the integration of audio and visual inputs. RAVN introduces an Acoustic Geometry Reasoner (AGR) that is trained with geometric proxy supervision. Using a heteroscedastic Gaussian NLL objective, AGR learns observation-dependent dispersion as a practical reliability cue, eliminating the need for geometric labels during inference. Additionally, we introduce Reliability-Aware Geometric Modulation (RAGM), which converts the learned cue into a soft gate to modulate visual features, thereby mitigating cross-modal conflicts. We evaluate RAVN on SoundSpaces using both Replica and Matterport3D environments, and the results show consistent improvements in navigation performance, with notable robustness in the challenging unheard sound setting.

Method: DynFOA analyzes input video to detect/localize dynamic sound sources, estimate depth/semantics, and reconstruct scene geometry/materials using 3D Gaussian Splatting. A diffusion model then generates spatial audio conditioned on these physically grounded features capturing acoustic interactions.

Result: DynFOA consistently outperforms existing methods in spatial accuracy, acoustic fidelity, distribution matching, and perceived immersive experience, as demonstrated on the new M2G-360 dataset of 600 real-world clips.

Conclusion: The integration of dynamic scene reconstruction with conditional diffusion modeling enables physically grounded spatial audio generation from 360-degree videos, addressing limitations of visual-only approaches and capturing complex acoustic interactions.

Abstract: Spatial audio is crucial for immersive 360-degree video experiences, yet most 360-degree videos lack it due to the difficulty of capturing spatial audio during recording. Automatically generating spatial audio such as first-order ambisonics (FOA) from video therefore remains an important but challenging problem. In complex scenes, sound perception depends not only on sound source locations but also on scene geometry, materials, and dynamic interactions with the environment. However, existing approaches only rely on visual cues and fail to model dynamic sources and acoustic effects such as occlusion, reflections, and reverberation. To address these challenges, we propose DynFOA, a generative framework that synthesizes FOA from 360-degree videos by integrating dynamic scene reconstruction with conditional diffusion modeling. DynFOA analyzes the input video to detect and localize dynamic sound sources, estimate depth and semantics, and reconstruct scene geometry and materials using 3D Gaussian Splatting (3DGS). The reconstructed scene representation provides physically grounded features that capture acoustic interactions between sources, environment, and listener viewpoint. Conditioned on these features, a diffusion model generates spatial audio consistent with the scene dynamics and acoustic context. We introduce M2G-360, a dataset of 600 real-world clips divided into MoveSources, Multi-Source, and Geometry subsets for evaluating robustness under diverse conditions. Experiments show that DynFOA consistently outperforms existing methods in spatial accuracy, acoustic fidelity, distribution matching, and perceived immersive experience.

Method: Two-stage approach: 1) Static stage (pre-synthesis) evaluates prompts using pitch-based prosodic features, audio quality, text-emotion coherence (LLM-scored), and TTS-specific metrics (character error rate, speaker/emotional similarity). 2) Dynamic stage (during synthesis) uses textual similarity model to select prompt most aligned with current input text.

Result: Strategy effectively selects prompts to synthesize speech with high-intensity emotional expression and robust speaker identity, leading to more expressive and stable zero-shot TTS performance.

Conclusion: The proposed two-stage prompt selection strategy improves expressive speech synthesis by ensuring better speaker identity stability and emotional intensity in zero-shot TTS systems.

Abstract: Recent advancements in speech synthesis have enabled large language model (LLM)-based systems to perform zero-shot generation with controllable content, timbre, speaker identity, and emotion through input prompts. As a result, these models heavily rely on prompt design to guide the generation process. However, existing prompt selection methods often fail to ensure that prompts contain sufficiently stable speaker identity cues and appropriate emotional intensity indicators, which are crucial for expressive speech synthesis. To address this challenge, we propose a two-stage prompt selection strategy specifically designed for expressive speech synthesis. In the static stage (before synthesis), we first evaluate prompt candidates using pitch-based prosodic features, perceptual audio quality, and text-emotion coherence scores evaluated by an LLM. We further assess the candidates under a specific TTS model by measuring character error rate, speaker similarity, and emotional similarity between the synthesized and prompt speech. In the dynamic stage (during synthesis), we use a textual similarity model to select the prompt that is most aligned with the current input text. Experimental results demonstrate that our strategy effectively selects prompt to synthesize speech with both high-intensity emotional expression and robust speaker identity, leading to more expressive and stable zero-shot TTS performance. Audio samples and codes will be available at https://whyrrrrun.github.io/ExpPro.github.io/.

Method: Two-stage framework: 1) Boundary detector identifies temporal transition points to divide audio into acoustically consistent segments, 2) Segment-level classifier evaluates each segment independently for authenticity. Uses reflection-based multi-length training to convert variable-duration segments into fixed input lengths, with ensemble techniques fusing predictions from multiple feature extractors and augmentation strategies.

Result: State-of-the-art performance on PartialSpoof benchmark across multiple temporal resolutions and at utterance level, with substantial improvements in detection and localization of spoofed regions. Also achieves state-of-the-art performance on Half-Truth dataset, demonstrating robustness and generalization.

Conclusion: The split-and-conquer approach effectively simplifies the learning objective by separating temporal localization from authenticity assessment, leading to improved performance in partial deepfake speech detection with better generalization capabilities.

Abstract: Partial deepfake speech detection requires identifying manipulated regions that may occur within short temporal portions of an otherwise bona fide utterance, making the task particularly challenging for conventional utterance-level classifiers. We propose a split-and-conquer framework that decomposes the problem into two stages: boundary detection and segment-level classification. A dedicated boundary detector first identifies temporal transition points, allowing the audio signal to be divided into segments that are expected to contain acoustically consistent content. Each resulting segment is then evaluated independently to determine whether it corresponds to bona fide or fake speech. This formulation simplifies the learning objective by explicitly separating temporal localization from authenticity assessment, allowing each component to focus on a well-defined task. To further improve robustness, we introduce a reflection-based multi-length training strategy that converts variable-duration segments into several fixed input lengths, producing diverse feature-space representations. Each stage is trained using multiple configurations with different feature extractors and augmentation strategies, and their complementary predictions are fused to obtain improved final models. Experiments on the PartialSpoof benchmark demonstrate state-of-the-art performance across multiple temporal resolutions as well as at the utterance level, with substantial improvements in the accurate detection and localization of spoofed regions. In addition, the proposed method achieves state-of-the-art performance on the Half-Truth dataset, further confirming the robustness and generalization capability of the framework.

Method: Proposes transferability analysis framework, defines what it means for a sufficient signal to be transferable, and conducts large study over 3 audio classification tasks: music genre, emotion recognition, and deepfake detection.

Result: Transferability rates vary by task: music genre signals transfer ≈26% of the time, other tasks show higher variance. Some deepfake detection models exhibit different transferability behavior (“flat-earther” models). Transferability analysis reveals information-theoretic differences not captured by traditional metrics like accuracy and precision.

Conclusion: Transferability analysis provides novel insights into audio classification models’ information processing characteristics, revealing task-dependent transferability patterns and identifying models with unique information processing behaviors, particularly in deepfake detection.

Abstract: In order to gain fresh insights about the information processing characteristics of different audio classification models, we propose transferability analysis. Given a minimal, sufficient signal for a classification on a model $f$, transferability analysis asks whether other models accept this minimal signal as having the same classification as it did on $f$. We define what it means for a sufficient signal to be transferable and perform a large study over $3$ different classification tasks: music genre, emotion recognition and deepfake detection. We find that transferability rates vary depending on the task, with sufficient signals for music genre being transferable $\approx26%$ of the time. The other tasks reveal much higher variance in transferability and reveal that some models, in particular on deepfake detection, have different transferability behavior. We call these models `flat-earther’ models. We investigate deepfake audio in more depth, and show that transferability analysis also allows to us to discover information theoretic differences between the models which are not captured by the more familiar metrics of accuracy and precision.

Method: Proposes a unified framework with “Labeling over filtering/cleaning” data strategy, Global-Sentence-Token annotation schema, continuous tokenizer backbone, Chain-of-Thought reasoning, Dimension Dropout, and hierarchical annotation as Structured Semantic Interface.

Result: Enables borderless long speech synthesis with layered control from scene semantics to phonetic detail, extending from Text2Speech to agent-centric synthesis with multi-modal inputs.

Conclusion: The framework creates a native agentic system with hierarchical annotation serving as interface between LLM agents and synthesis engine, enabling information-complete control channel for multi-modal inputs.

Abstract: Most existing text-to-speech (TTS) systems either synthesize speech sentence by sentence and stitch the results together, or drive synthesis from plain-text dialogues alone. Both approaches leave models with little understanding of global context or paralinguistic cues, making it hard to capture real-world phenomena such as multi-speaker interactions (interruptions, overlapping speech), evolving emotional arcs, and varied acoustic environments. We introduce the Borderless Long Speech Synthesis framework for agent-centric, borderless long audio synthesis. Rather than targeting a single narrow task, the system is designed as a unified capability set spanning VoiceDesigner, multi-speaker synthesis, Instruct TTS, and long-form text synthesis. On the data side, we propose a “Labeling over filtering/cleaning” strategy and design a top-down, multi-level annotation schema we call Global-Sentence-Token. On the model side, we adopt a backbone with a continuous tokenizer and add Chain-of-Thought (CoT) reasoning together with Dimension Dropout, both of which markedly improve instruction following under complex conditions. We further show that the system is Native Agentic by design: the hierarchical annotation doubles as a Structured Semantic Interface between the LLM Agent and the synthesis engine, creating a layered control protocol stack that spans from scene semantics down to phonetic detail. Text thereby becomes an information-complete, wide-band control channel, enabling a front-end LLM to convert inputs of any modality into structured generation commands, extending the paradigm from Text2Speech to borderless long speech synthesis.

Method: Develop a surrogate model using 3D convolutional neural networks combined with feed-forward layers to predict equivalent hydraulic conductivity tensors from voxelized 3D domains representing random fields of matrix and fracture conductivities. The model is trained on data from DFM simulations across different fracture-to-matrix conductivity contrasts.

Result: The trained models achieve high accuracy with normalized root-mean-square errors below 0.22 across most test cases. Surrogate-based upscaling preserves accuracy while achieving computational speedups exceeding 100x on GPU inference compared to conventional numerical homogenization.

Conclusion: The deep learning surrogate enables efficient upscaling of fracture effects in groundwater flow modeling, making multilevel Monte Carlo frameworks computationally feasible while maintaining accuracy.

Abstract: Modeling groundwater flow in three-dimensional fractured crystalline media requires accounting for strong spatial heterogeneity induced by fractures. Fine-scale discrete fracture-matrix (DFM) simulations can capture this complexity but are computationally expensive, especially when repeated evaluations are needed. To address this, we aim to employ a multilevel Monte Carlo (MLMC) framework in which numerical homogenization is used to upscale sub-resolution fracture effects when transitioning between accuracy levels. To reduce the cost of conventional 3D numerical homogenization, we develop a surrogate model that predicts the equivalent hydraulic conductivity tensor Keq from a voxelized 3D domain representing tensor-valued random fields of matrix and fracture conductivities. Fracture size, orientation, and aperture are sampled from distributions informed by natural observations. The surrogate architecture combines a 3D convolutional neural network with feed-forward layers, enabling it to capture both local spatial features and global interactions. Three surrogates are trained on data generated by DFM simulations, each corresponding to a different fracture-to-matrix conductivity contrast. Performance is evaluated across a wide range of fracture network parameters and matrix-field correlation lengths. The trained models achieve high accuracy, with normalized root-mean-square errors below 0.22 across most test cases. Practical applicability is demonstrated by comparing numerically homogenized conductivities with surrogate predictions in two macro-scale problems: computing equivalent conductivity tensors and predicting outflow from a constrained 3D domain. In both cases, surrogate-based upscaling preserves accuracy while substantially reducing computational cost, achieving speedups exceeding 100x when inference is performed on a GPU.

Method: Trained an autoregressive generative model on real-world data from over 300,000 patients and 400 million patient timeline entries in a self-supervised manner. Applied the model to COVID-19 hospitalized patients, modifying age, CRP, and creatinine to simulate 7-day outcomes.

Result: The model generated clinically plausible counterfactual trajectories that reproduced known clinical patterns: increased mortality with older age, elevated CRP, and elevated creatinine; remdesivir prescriptions increased with higher CRP and decreased with impaired kidney function.

Conclusion: Autoregressive generative models trained on real-world data in a self-supervised manner can establish a foundation for counterfactual clinical simulation, enabling personalized medicine applications.

Abstract: Counterfactual simulation - exploring hypothetical consequences under alternative clinical scenarios - holds promise for transformative applications such as personalized medicine and in silico trials. However, it remains challenging due to methodological limitations. Here, we show that an autoregressive generative model trained on real-world data from over 300,000 patients and 400 million patient timeline entries can generate clinically plausible counterfactual trajectories. As a validation task, we applied the model to patients hospitalized with COVID-19 in 2023, modifying age, serum C-reactive protein (CRP), and serum creatinine to simulate 7-day outcomes. Increased in-hospital mortality was observed in counterfactual simulations with older age, elevated CRP, and elevated serum creatinine. Remdesivir prescriptions increased in simulations with higher CRP values and decreased in those with impaired kidney function. These counterfactual trajectories reproduced known clinical patterns. These findings suggest that autoregressive generative models trained on real-world data in a self-supervised manner can establish a foundation for counterfactual clinical simulation.

Method: LiME uses a single shared PEFT module and modulates its output with lightweight expert vectors instead of separate adapters. It introduces zero-parameter routing by leveraging existing frozen and adapted representations, eliminating learned router parameters. Includes n-gram windowed routing and adaptive expert selection (Auto Top-K) based on routing confidence.

Result: On MMT-47 multimodal multi-task benchmark with 47 tasks spanning text, image, and video, LiME achieves competitive or superior performance while using up to 4x fewer trainable parameters and up to 29% faster training compared to MoE-PEFT baselines.

Conclusion: LiME provides a parameter-efficient MoE-PEFT approach that generalizes to any PEFT method, reduces expert parameters through modulation, and eliminates router parameters through zero-parameter routing, making multi-task adaptation more efficient and scalable.

Abstract: MoE-PEFT methods combine Mixture of Experts with parameter-efficient fine-tuning for multi-task adaptation, but require separate adapters per expert causing trainable parameters to scale linearly with expert count and limiting applicability to adapter-based architectures. We propose LiME (Lightweight Mixture of Experts), which achieves expert specialization through lightweight modulation rather than adapter replication. Instead of separate adapters, LiME uses a single shared PEFT module and modulates its output with lightweight expert vectors, reducing expert parameters while generalizing to any PEFT method. Notably, LiME introduces zero-parameter routing by leveraging existing frozen and adapted representations eliminating learned router parameters typically required per layer. Theoretically, we prove that (i) more experts preserve more task-relevant information and (ii) modulation approximates full expert-specific PEFT with bounded error. LiME further incorporates n-gram windowed routing and adaptive expert selection (Auto Top-K) based on routing confidence. Experiments on MMT-47, a multimodal multi-task benchmark with 47 tasks spanning text, image, and video, demonstrate that LiME achieves competitive or superior performance while using up to 4x fewer trainable parameters and up to 29% faster training compared to corresponding MoE-PEFT baselines.

Method: SIEVE uses SIEVE-GEN pipeline that decomposes context into applicable parts, generates synthetic queries paired with only relevant context, then uses context distillation to internalize context into model weights

Result: SIEVE outperforms prior context distillation methods using just three query examples across reasoning tasks including custom domains, RuleArena, and Machine Translation from One Book

Conclusion: Context decomposition enables highly sample-efficient parametric learning from natural language, achieving strong performance with minimal query examples

Abstract: Natural language context-such as instructions, knowledge, or feedback-contains rich signal for adapting language models. While in-context learning provides adaptation via the prompt, parametric learning persists into model weights and can improve performance further, though is data hungry and heavily relies on either high-quality traces or automated verifiers. We propose SIEVE, a method for sample-efficient parametric learning from natural language context that requires as few as three query examples. SIEVE uses a novel synthetic data generation pipeline, SIEVE-GEN, that leverages the insight that context is decomposable. Decomposing context allows us to generate higher quality rollouts by pairing synthetic queries with only the applicable context rather than the entirety, then using context distillation to internalize context into the model. We evaluate in reasoning settings where context is necessary, including custom domains and the RuleArena and Machine Translation from One Book tasks. Our results show that SIEVE outperforms prior context distillation methods using just three query examples, demonstrating how to achieve sample-efficient parametric learning from natural language.

Method: Proposes model scheduling where a smaller MDLM replaces the full model at a subset of denoising steps. Uses step-importance analysis based on loss and KL divergence between small and large models across timesteps, plus exhaustive search over coarse step segments to identify critical steps.

Result: Early and late denoising steps are substantially more robust to model replacement than middle steps, enabling up to 17% reduction in FLOPs with only modest degradation in generative perplexity on OpenWebText.

Conclusion: Simple, architecture-agnostic scheduling rules can significantly accelerate MDLM sampling while largely preserving generation quality, with the middle of the diffusion trajectory identified as most sensitive to model quality.

Abstract: Recent advances in masked diffusion language models (MDLMs) narrow the quality gap to autoregressive LMs, but their sampling remains expensive because generation requires many full-sequence denoising passes with a large Transformer and, unlike autoregressive decoding, cannot benefit from KV caching. In this work, we exploit the flexibility of the diffusion framework and study model scheduling, where a smaller MDLM replaces the full model at a subset of denoising steps. On OpenWebText, we show that early and late denoising steps are substantially more robust to such replacement than middle steps, enabling up to a 17% reduction in FLOPs with only modest degradation in generative perplexity. We support these findings with a step-importance analysis based on loss and KL divergence between small and large models across timesteps, as well as an exhaustive search over coarse step segments, both of which identify the middle of the diffusion trajectory as most sensitive. Our results suggest that simple, architecture-agnostic scheduling rules can significantly accelerate MDLM sampling while largely preserving generation quality as measured by generative perplexity.

Method: MusicRFM trains lightweight RFM probes to discover interpretable “concept directions” (axes corresponding to musical attributes) in MusicGen’s hidden states, then injects these directions during inference to guide generation in real-time without per-step optimization, supporting dynamic schedules and multiple property enforcement.

Result: The method increases target musical note generation accuracy from 0.23 to 0.82 while maintaining text prompt adherence within ~0.02 of baseline, demonstrating effective control with minimal impact on prompt fidelity.

Conclusion: MusicRFM successfully navigates the trade-off between control and generation quality, enabling interpretable fine-grained control over frozen music models without retraining or artifacts, with potential for broader audio generation applications.

Abstract: Controllable music generation remains a significant challenge, with existing methods often requiring model retraining or introducing audible artifacts. We introduce MusicRFM, a framework that adapts Recursive Feature Machines (RFMs) to enable fine-grained, interpretable control over frozen, pre-trained music models by directly steering their internal activations. RFMs analyze a model’s internal gradients to produce interpretable “concept directions”, or specific axes in the activation space that correspond to musical attributes like notes or chords. We first train lightweight RFM probes to discover these directions within MusicGen’s hidden states; then, during inference, we inject them back into the model to guide the generation process in real-time without per-step optimization. We present advanced mechanisms for this control, including dynamic, time-varying schedules and methods for the simultaneous enforcement of multiple musical properties. Our method successfully navigates the trade-off between control and generation quality: we can increase the accuracy of generating a target musical note from 0.23 to 0.82, while text prompt adherence remains within approximately 0.02 of the unsteered baseline, demonstrating effective control with minimal impact on prompt fidelity. We release code to encourage further exploration on RFMs in the music domain.

Method: PROGRS treats process rewards as relative preferences within outcome groups rather than absolute targets. It uses outcome-conditioned centering to shift PRM scores of incorrect trajectories to zero mean within each prompt group, removing bias while preserving rankings. Combines frozen quantile-regression PRM with multi-scale coherence evaluator, integrated into Group Relative Policy Optimization without additional trainable components.

Result: Across MATH-500, AMC, AIME, MinervaMath, and OlympiadBench, PROGRS consistently improves Pass@1 over outcome-only baselines and achieves stronger performance with fewer rollouts.

Conclusion: Outcome-conditioned centering enables safe and effective use of process rewards for mathematical reasoning, preventing reward hacking while leveraging denser supervision signals.

Abstract: Mathematical reasoning in large language models has improved substantially with reinforcement learning using verifiable rewards, where final answers can be checked automatically and converted into reliable training signals. Most such pipelines optimize outcome correctness only, which yields sparse feedback for long, multi-step solutions and offers limited guidance on intermediate reasoning errors. Recent work therefore introduces process reward models (PRMs) to score intermediate steps and provide denser supervision. In practice, PRM scores are often imperfectly aligned with final correctness and can reward locally fluent reasoning that still ends in an incorrect answer. When optimized as absolute rewards, such signals can amplify fluent failure modes and induce reward hacking. We propose PROGRS, a framework that leverages PRMs while keeping outcome correctness dominant. PROGRS treats process rewards as relative preferences within outcome groups rather than absolute targets. We introduce outcome-conditioned centering, which shifts PRM scores of incorrect trajectories to have zero mean within each prompt group. It removes systematic bias while preserving informative rankings. PROGRS combines a frozen quantile-regression PRM with a multi-scale coherence evaluator. We integrate the resulting centered process bonus into Group Relative Policy Optimization (GRPO) without auxiliary objectives or additional trainable components. Across MATH-500, AMC, AIME, MinervaMath, and OlympiadBench, PROGRS consistently improves Pass@1 over outcome-only baselines and achieves stronger performance with fewer rollouts. These results show that outcome-conditioned centering enables safe and effective use of process rewards for mathematical reasoning.

Method: Two-phase training strategy: 1) Graph editing phase that increases homophily ratio with respect to class labels while reducing homophily ratio with respect to sensitive attribute labels; 2) Integration of modified supervised contrastive loss and environmental loss into optimization to jointly improve predictive performance and fairness.

Result: Experiments on five real-world datasets demonstrate that the proposed model outperforms CAF and several state-of-the-art graph-based learning methods in both classification accuracy and fairness metrics.

Conclusion: The proposed approach effectively addresses fairness issues in GNNs by manipulating graph structure and incorporating specialized loss functions, achieving better performance than existing fairness-aware graph learning methods.

Abstract: In recent years, Graph Neural Networks (GNNs) have achieved remarkable success in tasks such as node classification, link prediction, and graph representation learning. However, they remain susceptible to biases that can arise not only from node attributes but also from the graph structure itself. Addressing fairness in GNNs has therefore emerged as a critical research challenge. In this work, we propose a novel model for training fairness-aware GNNs by improving the counterfactual augmented fair graph neural network framework (CAF). Specifically, our approach introduces a two-phase training strategy: in the first phase, we edit the graph to increase homophily ratio with respect to class labels while reducing homophily ratio with respect to sensitive attribute labels; in the second phase, we integrate a modified supervised contrastive loss and environmental loss into the optimization process, enabling the model to jointly improve predictive performance and fairness. Experiments on five real-world datasets demonstrate that our model outperforms CAF and several state-of-the-art graph-based learning methods in both classification accuracy and fairness metrics.

Method: Three approaches: 1) Lossless compression using domain-adapted LoRA adapters with arithmetic coding, 2) Lossy compression via prompting for succinct rewrites then arithmetic coding, 3) Interactive Question-Asking compression where small models ask yes/no questions to large models.

Result: Lossless compression improved 2x with LoRA adapters; lossy compression achieved ratios ~0.03 (2x improvement); QA compression achieved ratios 0.0006-0.004 (100x better than prior work), recovering 23-72% of capability gaps on benchmarks.

Conclusion: Interactive compression protocols like QA can transfer knowledge far more efficiently than transmitting full responses, establishing a compression-compute frontier for LLM-generated text.

Abstract: We study the compression of LLM-generated text across lossless and lossy regimes, characterizing a compression-compute frontier where more compression is possible at the cost of more compute. For lossless compression, domain-adapted LoRA adapters can improve LLM-based arithmetic coding by 2x over compression with the base LLM alone. For lossy compression, prompting a model for a succinct rewrite then applying arithmetic coding can achieve compression ratios of approximately 0.03, a 2x improvement over compressing the original response. We further introduce Question-Asking compression (QA), an interactive lossy protocol inspired by the game ‘Twenty Questions’. A small model iteratively refines its response by asking yes/no questions to a stronger model, transferring exactly one bit per answer. On 8 benchmarks spanning math, science, and code, 10 binary questions recover 23% to 72% of the capability gap between a small and large model on standard benchmarks and 7% to 38% on harder benchmarks, achieving compression ratios of 0.0006 to 0.004. This is over 100x smaller than prior LLM-based compression (Deletang et al., 2024), suggesting that interactive protocols can transfer knowledge far more efficiently than transmitting full responses.

Method: Developed sequential-dispatch methodology to accurately measure WebGPU overhead across 4 GPU vendors, 2 native implementations, 3 browsers, and 2 model sizes. Built torch-webgpu (PyTorch backend) and FX-to-WebGPU compiler for systematic testing on Linux, Windows, macOS.

Result: True per-dispatch WebGPU API overhead is 24-36μs on Vulkan and 32-71μs on Metal. Total per-operation overhead including Python cost is ~95μs. Kernel fusion improves throughput by 53% on Vulkan but provides no benefit on CUDA. torch-webgpu achieves 11-12% of CUDA performance.

Conclusion: At batch=1 with current dispatch-heavy pipeline, per-operation overhead dominates regardless of kernel quality. Backend choice is dominant factor for dispatch overhead, with implementation choice also mattering substantially within a backend (2.2× for Metal).

Abstract: WebGPU’s security-focused design imposes per-operation validation that compounds across the many small dispatches in neural network inference, yet the true cost of this overhead is poorly characterized. We present a systematic characterization of WebGPU dispatch overhead for LLM inference at batch size 1, spanning four GPU vendors (NVIDIA, AMD, Apple, Intel), two native implementations (Dawn, wgpu-native) and three browsers (Chrome, Safari, Firefox), and two model sizes (Qwen2.5-0.5B and 1.5B). Our primary contribution is a sequential-dispatch methodology that reveals naive single-operation benchmarks overestimate dispatch cost by ${\sim}20\times$. The true per-dispatch cost of WebGPU API overhead alone is 24-36 $μ$s on Vulkan and 32-71 $μ$s on Metal, while the total per-operation overhead including Python cost is ${\sim}95$~$μ$s, which turns out to be a distinction critical for optimization. On Vulkan, kernel fusion improves throughput by 53%, while CUDA fusion provides no benefit, confirming that per-operation overhead is a primary differentiator. LLM inference was tested across three major operating systems (Linux, Windows, macOS). We built $\texttt{torch-webgpu}$, a PrivateUse1-based out-of-tree PyTorch backend and an FX-to-WebGPU compiler, which on our reference platform achieves 11–12% of CUDA performance. At dtype-matched float32, RTX PRO 2000 achieves 1.4$\times$ WebGPU’s throughput despite ${\sim}6\times$ less compute than RTX 5090. For dispatch overhead, backend choice is the dominant factor, although implementation choice also matters substantially within a backend (2.2$\times$ for Metal). In terms of dispatch vs kernel compute efficiency, we conclude that at batch=1 with the current dispatch-heavy pipeline, per-operation overhead dominates regardless of kernel quality. All code, benchmarks, and raw data are open source.

Method: Proposes UI-Oceanus framework that shifts focus from trajectory imitation to mastering interaction physics via environmental feedback. Uses forward dynamics prediction (generative prediction of future interface states) as primary learning objective. Converts low-cost autonomous exploration into high-density generative supervision to build internal world models.

Result: Models using Continual Pre-Training on synthetic dynamics outperform baselines by 7% average success rate on offline benchmarks, with 16.8% gain in real-world online navigation. Navigation performance scales with synthetic data volume.

Conclusion: Grounding agents in forward predictive modeling offers superior pathway to scalable GUI automation with robust cross-domain adaptability and compositional generalization.

Abstract: Scaling generalist GUI agents is hindered by the data scalability bottleneck of expensive human demonstrations and the “distillation ceiling” of synthetic teacher supervision. To transcend these limitations, we propose UI-Oceanus, a framework that shifts the learning focus from mimicking high-level trajectories to mastering interaction physics via ground-truth environmental feedback. Through a systematic investigation of self-supervised objectives, we identify that forward dynamics, defined as the generative prediction of future interface states, acts as the primary driver for scalability and significantly outweighs inverse inference. UI-Oceanus leverages this insight by converting low-cost autonomous exploration, which is verified directly by system execution, into high-density generative supervision to construct a robust internal world model. Experimental evaluations across a series of models demonstrate the decisive superiority of our approach: models utilizing Continual Pre-Training (CPT) on synthetic dynamics outperform non-CPT baselines with an average success rate improvement of 7% on offline benchmarks, which amplifies to a 16.8% gain in real-world online navigation. Furthermore, we observe that navigation performance scales with synthetic data volume. These results confirm that grounding agents in forward predictive modeling offers a superior pathway to scalable GUI automation with robust cross-domain adaptability and compositional generalization.

Method: Developed DrugPlayGround framework to benchmark LLM performance on generating text-based descriptions of physiochemical drug characteristics, drug synergism, drug-protein interactions, and physiological responses to drug perturbations. The framework works with domain experts to provide detailed explanations justifying LLM predictions.

Result: The paper presents DrugPlayGround as a solution to evaluate LLMs for chemical and biological reasoning capabilities, enabling greater use of LLMs at all stages of drug discovery.

Conclusion: DrugPlayGround addresses the emergent problem of objectively assessing LLM performance in drug discovery, providing a framework to test LLMs’ reasoning capabilities and push their adoption in pharmaceutical research.

Abstract: Large language models (LLMs) are in the ascendancy for research in drug discovery, offering unprecedented opportunities to reshape drug research by accelerating hypothesis generation, optimizing candidate prioritization, and enabling more scalable and cost-effective drug discovery pipelines. However there is currently a lack of objective assessments of LLM performance to ascertain their advantages and limitations over traditional drug discovery platforms. To tackle this emergent problem, we have developed DrugPlayGround, a framework to evaluate and benchmark LLM performance for generating meaningful text-based descriptions of physiochemical drug characteristics, drug synergism, drug-protein interactions, and the physiological response to perturbations introduced by drug molecules. Moreover, DrugPlayGround is designed to work with domain experts to provide detailed explanations for justifying the predictions of LLMs, thereby testing LLMs for chemical and biological reasoning capabilities to push their greater use at the frontier of drug discovery at all of its stages.

Method: The paper studies Decentralized SGD (DSGD) with light-tailed noise, providing technical results including variance-reduction effects in high-probability sense and novel bounds on moment generating functions for strongly convex costs.

Result: DSGD converges in high-probability under same conditions as MSE convergence, achieves order-optimal rates for non-convex and strongly convex costs, and shows linear speed-up in number of users with matching or better transient times than MSE results.

Conclusion: This is the first work showing DSGD achieves linear speed-up in high-probability sense, with relaxed assumptions and sharp rates enabled by novel technical results of independent interest.

Abstract: Convergence in high-probability (HP) has attracted increasing interest, due to implying exponentially decaying tail bounds and strong guarantees for individual runs of an algorithm. While many works study HP guarantees in centralized settings, much less is understood in the decentralized setup, where existing works require strong assumptions, like uniformly bounded gradients, or asymptotically vanishing noise. This results in a significant gap between the assumptions used to establish convergence in the HP and the mean-squared error (MSE) sense, and is also contrary to centralized settings, where it is known that $\mathtt{SGD}$ converges in HP under the same conditions on the cost function as needed for MSE convergence. Motivated by these observations, we study the HP convergence of Decentralized $\mathtt{SGD}$ ($\mathtt{DSGD}$) in the presence of light-tailed noise, providing several strong results. First, we show that $\mathtt{DSGD}$ converges in HP under the same conditions on the cost as in the MSE sense, removing the restrictive assumptions used in prior works. Second, our sharp analysis yields order-optimal rates for both non-convex and strongly convex costs. Third, we establish a linear speed-up in the number of users, leading to matching, or strictly better transient times than those obtained from MSE results, further underlining the tightness of our analysis. To the best of our knowledge, this is the first work that shows $\mathtt{DSGD}$ achieves a linear speed-up in the HP sense. Our relaxed assumptions and sharp rates stem from several technical results of independent interest, including a result on the variance-reduction effect of decentralized methods in the HP sense, as well as a novel bound on the MGF of strongly convex costs, which is of interest even in centralized settings. Finally, we provide experiments that validate our theory.

Method: Proposes FTimeXer, a frequency-aware time-series Transformer with FFT-driven frequency branch and gated time-frequency fusion to capture multi-scale periodicity, plus stochastic exogenous masking with consistency regularization to handle irregular inputs and reduce spurious correlations.

Result: Experiments on three real-world datasets show consistent improvements over strong baselines, leading to more reliable forecasts of grid carbon factors.

Conclusion: FTimeXer provides enhanced forecasting capabilities for grid carbon intensity, which is essential for effective PCF accounting and informed decarbonization decision-making.

Abstract: Accurate and up-to-date forecasting of the power grid’s carbon footprint is crucial for effective product carbon footprint (PCF) accounting and informed decarbonization decisions. However, the carbon intensity of the grid exhibits high non-stationarity, and existing methods often struggle to effectively leverage periodic and oscillatory patterns. Furthermore, these methods tend to perform poorly when confronted with irregular exogenous inputs, such as missing data or misalignment. To tackle these challenges, we propose FTimeXer, a frequency-aware time-series Transformer designed with a robust training scheme that accommodates exogenous factors. FTimeXer features an Fast Fourier Transform (FFT)-driven frequency branch combined with gated time-frequency fusion, allowing it to capture multi-scale periodicity effectively. It also employs stochastic exogenous masking in conjunction with consistency regularization, which helps reduce spurious correlations and enhance stability. Experiments conducted on three real-world datasets show consistent improvements over strong baselines. As a result, these enhancements lead to more reliable forecasts of grid carbon factors, which are essential for effective PCF accounting and informed decision-making regarding decarbonization.

Method: Proposes a novel taxonomy separating contexts into allogenic (environment-imposed) and autogenic (agent-driven) factors. Identifies three research directions: 1) Learning with heterogeneous contexts, 2) Multi-time-scale modeling, and 3) Integration of abstract, high-level contexts.

Result: The paper presents a conceptual framework and research agenda rather than empirical results. It establishes a taxonomy and identifies key challenges for achieving contextual intelligence in RL agents.

Conclusion: Context should be treated as a first-class modeling primitive to empower agents to reason about their identity, environmental constraints, and temporal evolution, enabling safer and more efficient real-world deployment of context-aware agents.

Abstract: Reinforcement learning (RL) has produced spectacular results in games, robotics, and continuous control. Yet, despite these successes, learned policies often fail to generalize beyond their training distribution, limiting real-world impact. Recent work on contextual RL (cRL) shows that exposing agents to environment characteristics – contexts – can improve zero-shot transfer. So far, the community has treated context as a monolithic, static observable, an approach that constrains the generalization capabilities of RL agents. To achieve contextual intelligence we first propose a novel taxonomy of contexts that separates allogenic (environment-imposed) from autogenic (agent-driven) factors. We identify three fundamental research directions that must be addressed to promote truly contextual intelligence: (1) Learning with heterogeneous contexts to explicitly exploit the taxonomy levels so agents can reason about their influence on the world and vice versa; (2) Multi-time-scale modeling to recognize that allogenic variables evolve slowly or remain static, whereas autogenic variables may change within an episode, potentially requiring different learning mechanisms; (3) Integration of abstract, high-level contexts to incorporate roles, resource & regulatory regimes, uncertainties, and other non-physical descriptors that crucially influence behavior. We envision context as a first-class modeling primitive, empowering agents to reason about who they are, what the world permits, and how both evolve over time. By doing so, we aim to catalyze a new generation of context-aware agents that can be deployed safely and efficiently in the real world.

Method: Proposes OPRIDE (Offline PbRL via In-Dataset Exploration) with two key features: 1) a principled exploration strategy maximizing query informativeness, and 2) a discount scheduling mechanism to mitigate reward function overoptimization.

Result: OPRIDE significantly outperforms prior methods, achieving strong performance with notably fewer queries across various locomotion, manipulation, and navigation tasks. The approach also provides theoretical efficiency guarantees.

Conclusion: The proposed method effectively addresses query efficiency challenges in offline PbRL through intelligent exploration and optimization control, making preference-based RL more practical for real-world applications with limited human feedback.

Abstract: Preference-based reinforcement learning (PbRL) can help avoid sophisticated reward designs and align better with human intentions, showing great promise in various real-world applications. However, obtaining human feedback for preferences can be expensive and time-consuming, which forms a strong barrier for PbRL. In this work, we address the problem of low query efficiency in offline PbRL, pinpointing two primary reasons: inefficient exploration and overoptimization of learned reward functions. In response to these challenges, we propose a novel algorithm, \textbf{O}ffline \textbf{P}b\textbf{R}L via \textbf{I}n-\textbf{D}ataset \textbf{E}xploration (OPRIDE), designed to enhance the query efficiency of offline PbRL. OPRIDE consists of two key features: a principled exploration strategy that maximizes the informativeness of the queries and a discount scheduling mechanism aimed at mitigating overoptimization of the learned reward functions. Through empirical evaluations, we demonstrate that OPRIDE significantly outperforms prior methods, achieving strong performance with notably fewer queries. Moreover, we provide theoretical guarantees of the algorithm’s efficiency. Experimental results across various locomotion, manipulation, and navigation tasks underscore the efficacy and versatility of our approach.

Method: Differentiable Symbolic Planning (DSP) maintains a feasibility channel (phi) that tracks constraint satisfaction evidence at each node, aggregates this into global feasibility signal (Phi) through learned rule-weighted combination, and uses sparsemax attention for exact-zero discrete rule selection. Integrated into Universal Cognitive Kernel (UCK) combining graph attention with iterative constraint propagation.

Result: Achieves 97.4% accuracy on planning under 4x size generalization (vs. 59.7% for baselines), 96.4% on SAT under 2x generalization, maintains balanced performance on both positive/negative classes where standard neural approaches collapse. Learned phi signal shows interpretable semantics (+18 for feasible, -13 for infeasible).

Conclusion: DSP enables neural networks to perform effective constraint reasoning with strong generalization, interpretable learned signals, and maintains differentiability while handling discrete symbolic reasoning.

Abstract: Neural networks excel at pattern recognition but struggle with constraint reasoning – determining whether configurations satisfy logical or physical constraints. We introduce Differentiable Symbolic Planning (DSP), a neural architecture that performs discrete symbolic reasoning while remaining fully differentiable. DSP maintains a feasibility channel (phi) that tracks constraint satisfaction evidence at each node, aggregates this into a global feasibility signal (Phi) through learned rule-weighted combination, and uses sparsemax attention to achieve exact-zero discrete rule selection. We integrate DSP into a Universal Cognitive Kernel (UCK) that combines graph attention with iterative constraint propagation. Evaluated on three constraint reasoning benchmarks – graph reachability, Boolean satisfiability, and planning feasibility – UCK+DSP achieves 97.4% accuracy on planning under 4x size generalization (vs. 59.7% for ablated baselines), 96.4% on SAT under 2x generalization, and maintains balanced performance on both positive and negative classes where standard neural approaches collapse. Ablation studies reveal that global phi aggregation is critical: removing it causes accuracy to drop from 98% to 64%. The learned phi signal exhibits interpretable semantics, with values of +18 for feasible cases and -13 for infeasible cases emerging without supervision.

Method: Proposes a deployment-centric framework treating reliability as a dynamic state composed of discrimination and calibration. Models reliability trajectory across sequential evaluation windows, measuring volatility, and formulates deployment adaptation as a multi-objective control problem. Defines state-dependent intervention policies and characterizes cost-volatility Pareto frontier.

Result: Experiments on large-scale credit-risk dataset (1.35M loans, 2007-2018) show selective, drift-triggered interventions achieve smoother reliability trajectories than continuous rolling retraining while substantially reducing operational costs.

Conclusion: Deployment reliability under temporal shift can be treated as a controllable multi-objective system, highlighting the role of policy design in shaping stability-cost trade-offs in high-stakes tabular applications.

Abstract: Machine learning models deployed in non-stationary environments are exposed to temporal distribution shift, which can erode predictive reliability over time. While common mitigation strategies such as periodic retraining and recalibration aim to preserve performance, they typically focus on average metrics evaluated at isolated time points and do not explicitly model how reliability evolves during deployment. We propose a deployment-centric framework that treats reliability as a dynamic state composed of discrimination and calibration. The trajectory of this state across sequential evaluation windows induces a measurable notion of volatility, allowing deployment adaptation to be formulated as a multi-objective control problem that balances reliability stability against cumulative intervention cost. Within this framework, we define a family of state-dependent intervention policies and empirically characterize the resulting cost-volatility Pareto frontier. Experiments on a large-scale, temporally indexed credit-risk dataset (1.35M loans, 2007-2018) show that selective, drift-triggered interventions can achieve smoother reliability trajectories than continuous rolling retraining while substantially reducing operational cost. These findings position deployment reliability under temporal shift as a controllable multi-objective system and highlight the role of policy design in shaping stability-cost trade-offs in high-stakes tabular applications.

Method: Contrastive Prompt Tuning (CPT) combines Contrastive Learning (to distinguish efficient vs. inefficient code) with Prompt Tuning (a Parameter-Efficient Fine-Tuning approach). The method is evaluated on Python, Java, and C++ coding problems across three different LLM models.

Result: CPT achieves consistent improvements in code accuracy for two models, but efficiency gains vary by model, programming language, and task complexity. Improvements in energy efficiency are not uniformly reliable across different conditions.

Conclusion: While CPT shows promise for improving code accuracy and some energy efficiency gains, the results are inconsistent across models, languages, and task complexities, indicating that reliable optimization of LLMs for energy-efficient code generation remains challenging.

Abstract: Although LLMs are capable of generating functionally correct code, they also tend to produce less energy-efficient code in comparison to human-written solutions. As these inefficiencies lead to higher computational overhead, they are in direct conflict with Green Software Development (GSD) efforts, which aim to reduce the energy consumption of code. To support these efforts, this study aims to investigate whether and how LLMs can be optimized to promote the generation of energy-efficient code. To this end, we employ Contrastive Prompt Tuning (CPT). CPT combines Contrastive Learning techniques, which help the model to distinguish between efficient and inefficient code, and Prompt Tuning, a Parameter-Efficient Fine Tuning (PEFT) approach that requires only a fraction of the cost of traditional fine tuning. This study evaluates CPT on Python, Java and C++ coding problems across three different models to provide a comprehensive evaluation. The method achieves consistent improvements in code accuracy for two models but efficiency gains vary by model, language and task complexity, indicating that improvements are not uniformly reliable.

Method: Clusters each agent’s encoder features into K concepts via K-means, validates causal relationship between concepts and behavior through intervention experiments, then aligns concepts between agents using optimal bipartite matching for zero-shot transfer.

Result: Concept interventions change actions in 69.4% of cases (p=8.6×10⁻⁸⁶). Concept transfer achieves 69.5%±3.2% and 76.4%±3.4% win rates in Go 7×7, vs 3.5% for random agent and 9.2% without alignment. Works only in domains with naturally discrete strategic state (Go) not continuous ones (Atari Breakout).

Conclusion: PRISM successfully enables zero-shot transfer of strategic knowledge between RL agents by grounding decisions in causally validated concepts, but only works in domains with naturally discrete strategic state structure.

Abstract: We present PRISM (Policy Reuse via Interpretable Strategy Mapping), a framework that grounds reinforcement learning agents’ decisions in discrete, causally validated concepts and uses those concepts as a zero-shot transfer interface between agents trained with different algorithms. PRISM clusters each agent’s encoder features into $K$ concepts via K-means. Causal intervention establishes that these concepts directly drive - not merely correlate with - agent behavior: overriding concept assignments changes the selected action in 69.4% of interventions ($p = 8.6 \times 10^{-86}$, 2500 interventions). Concept importance and usage frequency are dissociated: the most-used concept (C47, 33.0% frequency) causes only a 9.4% win-rate drop when ablated, while ablating C16 (15.4% frequency) collapses win rate from 100% to 51.8%. Because concepts causally encode strategy, aligning them via optimal bipartite matching transfers strategic knowledge zero-shot. On Go~7$\times$7 with three independently trained agents, concept transfer achieves 69.5%$\pm$3.2% and 76.4%$\pm$3.4% win rate against a standard engine across the two successful transfer pairs (10 seeds), compared to 3.5% for a random agent and 9.2% without alignment. Transfer succeeds when the source policy is strong; geometric alignment quality predicts nothing ($R^2 \approx 0$). The framework is scoped to domains where strategic state is naturally discrete: the identical pipeline on Atari Breakout yields bottleneck policies at random-agent performance, confirming that the Go results reflect a structural property of the domain.

Method: Proposes Entropy-Guided Group Relative Policy Optimization (EG-GRPO) that analyzes token-level entropy to guide RL updates: low-entropy tokens are excluded from reward updates to preserve stability, while high-entropy tokens receive entropy bonuses to encourage structured exploration.

Result: EG-GRPO achieves state-of-the-art performance on standard T2I benchmarks, demonstrating improved generation quality through better management of exploration-exploitation trade-offs.

Conclusion: The paper provides systematic entropy analysis revealing that CoT expands exploration while RL contracts it, and that controlling entropy dynamics is crucial for stable, high-quality text-to-image generation.

Abstract: Combining Chain-of-Thought (CoT) with Reinforcement Learning (RL) improves text-to-image (T2I) generation, yet the underlying interaction between CoT’s exploration and RL’s optimization remains unclear. We present a systematic entropy-based analysis that yields three key insights: (1) CoT expands the generative exploration space, while RL contracts it toward high-reward regions; (2) final reward is strongly negatively correlated with both the mean and variance of image-token entropy, highlighting the need to reduce uncertainty and instability; and (3) the entropy of the textual CoT directly governs downstream image quality, with lower-entropy CoTs leading to better generations. Motivated by these findings, we propose Entropy-Guided Group Relative Policy Optimization (EG-GRPO), a fine-tuning strategy that reallocates optimization budget by uncertainty: low-entropy tokens are excluded from reward-driven updates to preserve stability, while high-entropy tokens receive an entropy bonus that encourages structured exploration without collapse. Experiments on standard T2I benchmarks demonstrate that EG-GRPO achieves state-of-the-art performance.

Method: Created YC Bench using YC W26 batch (196 startups). Measures outperformance with Pre-Demo Day Score combining publicly available traction signals and web visibility. Uses Google mentions before application deadline as baseline proxy for prior brand recognition.

Result: Baseline method using Google mentions recovered 6 of 11 top performers at YC Demo Day (55% recall). Provides live benchmark with iteration cycles measured in months rather than years.

Conclusion: YC Bench enables rapid evaluation of forecasting models for startup success, addressing the slow evaluation problem in traditional startup forecasting research.

Abstract: Forecasting startup success is notoriously difficult, partly because meaningful outcomes, such as exits, large funding rounds, and sustained revenue growth, are rare and can take years to materialize. As a result, signals are sparse and evaluation cycles are slow. Y Combinator batches offer a unique mitigation: each batch comprises around 200 startups, funded simultaneously, with evaluation at Demo Day only three months later. We introduce YC Bench, a live benchmark for forecasting early outperformance within YC batches. Using the YC W26 batch as a case study (196 startups), we measure outperformance with a Pre-Demo Day Score, a KPI combining publicly available traction signals and web visibility. This short-term metric enables rapid evaluation of forecasting models. As a baseline, we take Google mentions prior to the YC W26 application deadline, a simple proxy for prior brand recognition, recovering 6 of 11 top performers at YC Demo Day (55% recall). YC Bench provides a live benchmark for studying startup success forecasting, with iteration cycles measured in months rather than years. Code and Data are available on GitHub: https://github.com/benstaf/ycbench

Method: Introduces a minimal model inspired by Fukumizu and Amari to investigate vanishing gradients and overfitting in MLPs trained via gradient descent, analyzing learning dynamics through saddle structures and convergence properties.

Result: Shows learning dynamics pass through plateau regions and near-optimal saddle structures before converging to overfitting regions; proves that with suitable dataset conditions, overfitting regions collapse to single attractors; demonstrates MLPs on finite noisy datasets cannot converge to theoretical optimum.

Conclusion: Provides a dynamical systems framework for understanding vanishing gradients and overfitting in MLPs, establishing that overfitting is inevitable on finite noisy datasets and occurs through specific saddle structure dynamics.

Abstract: Vanishing gradient and overfitting are two of the most extensively studied problems in the literature about machine learning. However, they are frequently considered in some asymptotic setting, which obscure the underlying dynamical mechanisms responsible for their emergence. In this paper, we aim to provide a clear dynamical description of learning in multi-layer perceptrons. To this end, we introduce a minimal model, inspired by studies by Fukumizu and Amari, to investigate vanishing gradients and overfitting in MLPs trained via gradient descent. Within this model, we show that the learning dynamics may pass through plateau regions and near-optimal regions during training, both of which consist of saddle structures, before ultimately converging to the overfitting region. Under suitable conditions on the training dataset, we prove that, with high probability, the overfitting region collapses to a single attractor modulo symmetry, which corresponds to the overfitting. Moreover, we show that any MLP trained on a finite noisy dataset cannot converge to the theoretical optimum and instead necessarily converges to an overfitting solution.

Method: SDTI uses standard neural network components with appropriate problem formulation and architectural design to enable zero-shot target variable identification without pre-training.

Result: SDTI outperformed baseline architectures by 14% in F1 score on synthetic task identification benchmarks and demonstrated reliable identification of ground truth target variables.

Conclusion: SDTI shows potential to reduce manual annotation dependence and enhance autonomous learning system scalability, though it’s a proof-of-concept framework.

Abstract: In this work, we present a novel machine learning framework called Self-Directed Task Identification (SDTI), which enables models to autonomously identify the correct target variable for each dataset in a zero-shot setting without pre-training. SDTI is a minimal, interpretable framework demonstrating the feasibility of repurposing core machine learning concepts for a novel task structure. To our knowledge, no existing architectures have demonstrated this ability. Traditional approaches lack this capability, leaving data annotation as a time-consuming process that relies heavily on human effort. Using only standard neural network components, we show that SDTI can be achieved through appropriate problem formulation and architectural design. We evaluate the proposed framework on a range of benchmark tasks and demonstrate its effectiveness in reliably identifying the ground truth out of a set of potential target variables. SDTI outperformed baseline architectures by 14% in F1 score on synthetic task identification benchmarks. These proof-of-concept experiments highlight the future potential of SDTI to reduce dependence on manual annotation and to enhance the scalability of autonomous learning systems in real-world applications.

Method: Develop Mutual Information-based Split Variational Autoencoder (MI-VAE), a physics-informed VAE that learns differences between observed system trajectories and those predicted by physics-based models. The latent space enables generation of synthetic datasets that respect physical constraints.

Result: MI-VAE significantly improves downstream RL performance when augmenting datasets with its samples, outperforming standard VAEs in statistical fidelity, sample diversity, and policy success rate on a planetary lander problem.

Conclusion: MI-VAE provides a scalable strategy for enhancing autonomous controller robustness in complex, data-constrained environments by introducing physics-based learning bias in generative models.

Abstract: The deployment of reinforcement learning (RL)-based controllers on physical systems is often limited by poor generalization to real-world scenarios, known as the simulation-to-reality (sim-to-real) gap. This gap is particularly challenging in spaceflight, where real-world training data are scarce due to high cost and limited planetary exploration data. Traditional approaches, such as system identification and synthetic data generation, depend on sufficient data and often fail due to modeling assumptions or lack of physics-based constraints. We propose addressing this data scarcity by introducing physics-based learning bias in a generative model. Specifically, we develop the Mutual Information-based Split Variational Autoencoder (MI-VAE), a physics-informed VAE that learns differences between observed system trajectories and those predicted by physics-based models. The latent space of the MI-VAE enables generation of synthetic datasets that respect physical constraints. We evaluate MI-VAE on a planetary lander problem, focusing on limited real-world data and offline RL training. Results show that augmenting datasets with MI-VAE samples significantly improves downstream RL performance, outperforming standard VAEs in statistical fidelity, sample diversity, and policy success rate. This work demonstrates a scalable strategy for enhancing autonomous controller robustness in complex, data-constrained environments.

Method: The system combines: 1) pre-sorted multidimensional aggregation, 2) efficient exclusion-zone-aware k-nearest-neighbor retrieval for repeated anomalies, and 3) moving-average post-processing. The implementation handles both univariate and multivariate time series.

Result: The paper documents the released implementation, hyperparameter settings for univariate and multivariate tracks, and corresponding benchmark results on TSB-AD. It also analyzes system performance on the aggregate leaderboard and across specific dataset characteristics.

Conclusion: MMPAD serves as a reproducible reference for Matrix Profile-based anomaly detection on TSB-AD, providing an open-source implementation that demonstrates how design choices beyond vanilla nearest-neighbor profiles can improve benchmark performance.

Abstract: Matrix Profile (MP) methods are an interpretable and scalable family of distance-based methods for time-series anomaly detection, but strong benchmark performance still depends on design choices beyond a vanilla nearest-neighbor profile. This technical report documents an open-source Matrix Profile for Anomaly Detection (MMPAD) submission to TSB-AD, a benchmark that covers both univariate and multivariate time series. The submitted system combines pre-sorted multidimensional aggregation, efficient exclusion-zone-aware k-nearest-neighbor (kNN) retrieval for repeated anomalies, and moving-average post-processing. To serve as a reproducible reference for MP-based anomaly detection on TSB-AD, we detail the released implementation, the hyperparameter settings for the univariate and multivariate tracks, and the corresponding benchmark results. We further analyze how the system performs on the aggregate leaderboard and across specific dataset characteristics.The open-source implementation is available at https://github.com/mcyeh/mmpad_tsb.

Method: Systematically evaluated 6 LLMs (4 open-source, 2 frontier) on human-graded natural language proofs of competition-level problems. Measured verifier accuracy and self-consistency. Conducted LLM-guided prompt search to synthesize specialized prompt ensembles that address specific failure modes of smaller models.

Result: Smaller open-source models are only ~10% behind frontier models in accuracy but up to ~25% more inconsistent. They possess the mathematical capabilities for verification but struggle to reliably elicit them with general prompts. Specialized prompts boosted performance by up to 9.1% in accuracy and 15.9% in self-consistency, allowing models like Qwen3.5-35B to perform on par with frontier models like Gemini 3.1 Pro.

Conclusion: Reliable proof verification requires not just mathematical capability but also consistent elicitation of that capability. Smaller models can achieve frontier-level verification performance through carefully designed specialized prompts that address their specific failure modes.

Abstract: Advances in training, post-training, and inference-time methods have enabled frontier reasoning models to win gold medals in math competitions and settle challenging open problems. Gaining trust in the responses of these models requires that natural language proofs be checked for errors. LLM judges are increasingly being adopted to meet the growing demand for evaluating such proofs. While verification is considered easier than generation, what model capability does reliable verification actually require? We systematically evaluate four open-source and two frontier LLMs on datasets of human-graded natural language proofs of competition-level problems. We consider two key metrics: verifier accuracy and self-consistency (the rate of agreement across repeated judgments on the same proof). We observe that smaller open-source models are only up to ~10% behind frontier models in accuracy but they are up to ~25% more inconsistent. Furthermore, we see that verifier accuracy is sensitive to prompt choice across all models. We then demonstrate that the smaller models, in fact, do possess the mathematical capabilities to verify proofs at the level of frontier models, but they struggle to reliably elicit these capabilities with general judging prompts. Through an LLM-guided prompt search, we synthesize an ensemble of specialized prompts that overcome the specific failure modes of smaller models, boosting their performance by up to 9.1% in accuracy and 15.9% in self-consistency. These gains are realized across models and datasets, allowing models like Qwen3.5-35B to perform on par with frontier models such as Gemini 3.1 Pro for proof verification.

Method: Analyze layerwise updates across multiple architectures (Transformers, state-space models) by decomposing them into tokenwise components and residuals, measuring alignment, angular deviation, and projection onto dominant subspaces.

Result: Layer updates show near-perfect alignment with tokenwise components, while residuals are geometrically distinct with weaker alignment, larger angular deviation, and lower projection. Approximation error under tokenwise models strongly correlates with output perturbation (Spearman correlations up to 0.95).

Conclusion: Most layer updates behave like structured reparameterizations along dominant directions, with functionally significant computation concentrated in geometrically distinct residual components, providing an architecture-agnostic probing method.

Abstract: We study the geometric structure of layer updates in deep language models. Rather than analyzing what information is encoded in intermediate representations, we ask how representations change from one layer to the next. We show that layerwise updates admit a decomposition into a dominant tokenwise component and a residual that is not captured by restricted tokenwise function classes. Across multiple architectures, including Transformers and state-space models, we find that the full layer update is almost perfectly aligned with the tokenwise component, while the residual exhibits substantially weaker alignment, larger angular deviation, and significantly lower projection onto the dominant tokenwise subspace. This indicates that the residual is not merely a small correction, but a geometrically distinct component of the transformation. This geometric separation has functional consequences: approximation error under the restricted tokenwise model is strongly associated with output perturbation, with Spearman correlations often exceeding 0.7 and reaching up to 0.95 in larger models. Together, these results suggest that most layerwise updates behave like structured reparameterizations along a dominant direction, while functionally significant computation is concentrated in a geometrically distinct residual component. Our framework provides a simple, architecture-agnostic method for probing the geometric and functional structure of layer updates in modern language models.

Method: VALOR uses a Treatment-Gated Sparse-Revenue Network with bilinear interaction to prevent causal signal collapse, optimized via a Cost-Sensitive Focal-ZILN objective combining focal mechanism for distributional robustness with value-weighted ranking loss. Also includes Robust ZILN-GBDT tree variant for interpretability.

Result: Achieves 20% improvement in rankability over state-of-the-art methods on public benchmarks and delivers 2.7x increase in incremental revenue per account in a 4-month production A/B test.

Conclusion: VALOR provides an effective unified framework for B2B sales optimization that addresses key challenges in uplift modeling for zero-inflated revenue distributions.

Abstract: B2B sales organizations must identify “persuadable” accounts within zero-inflated revenue distributions to optimize expensive human resource allocation. Standard uplift frameworks struggle with treatment signal collapse in high-dimensional spaces and a misalignment between regression calibration and the ranking of high-value “whales.” We introduce VALOR (Value Aware Learning of Optimized (B2B) Revenue), a unified framework featuring a Treatment-Gated Sparse-Revenue Network that uses bilinear interaction to prevent causal signal collapse. The framework is optimized via a novel Cost-Sensitive Focal-ZILN objective that combines a focal mechanism for distributional robustness with a value-weighted ranking loss that scales penalties based on financial magnitude. To provide interpretability for high-touch sales programs, we further derive Robust ZILN-GBDT, a tree based variant utilizing a custom splitting criterion for uplift heterogeneity. Extensive evaluations confirm VALOR’s dominance, achieving a 20% improvement in rankability over state-of-the-art methods on public benchmarks and delivering a validated 2.7x increase in incremental revenue per account in a rigorous 4-month production A/B test.

Method: Proposes time-warping as a transfer learning technique for RNNs, specifically LSTMs. The method rescales time in the model to adapt it to different characteristic time scales. Theoretically proves LSTMs can approximate linear time lag differential equations with arbitrary accuracy and maintain this accuracy under time-warping.

Result: Applied to wildfire fuel moisture content prediction, where an LSTM pretrained on 10-hour time scale data was adapted to 1-hour, 100-hour, and 1000-hour time scales. Time-warping achieved comparable accuracy to established transfer learning methods while modifying only a small fraction of parameters.

Conclusion: Time-warping provides an efficient transfer learning approach for RNNs when adapting to different time scales, requiring minimal parameter modification while maintaining prediction accuracy comparable to more extensive retraining methods.

Abstract: Dynamical systems describe how a physical system evolves over time. Physical processes can evolve faster or slower in different environmental conditions. We use time-warping as rescaling the time in a model of a physical system. This thesis proposes a new method of transfer learning for Recurrent Neural Networks (RNNs) based on time-warping. We prove that for a class of linear, first-order differential equations known as time lag models, an LSTM can approximate these systems with any desired accuracy, and the model can be time-warped while maintaining the approximation accuracy. The Time-Warping method of transfer learning is then evaluated in an applied problem on predicting fuel moisture content (FMC), an important concept in wildfire modeling. An RNN with LSTM recurrent layers is pretrained on fuels with a characteristic time scale of 10 hours, where there are large quantities of data available for training. The RNN is then modified with transfer learning to generate predictions for fuels with characteristic time scales of 1 hour, 100 hours, and 1000 hours. The Time-Warping method is evaluated against several known methods of transfer learning. The Time-Warping method produces predictions with an accuracy level comparable to the established methods, despite modifying only a small fraction of the parameters that the other methods modify.

Method: Proposes Structural Extrapolated Data GEneration (SEDGE) framework based on assumptions about data generating processes. Uses structure-informed optimization strategy or diffusion posterior sampling for practical implementation. Provides conditions for reliable generation and approximate identifiability under conservative assumptions.

Result: Verifies extrapolation performance on synthetic data and demonstrates validity through extrapolated image generation as a real-world scenario.

Conclusion: SEDGE provides a principled framework for generating data that satisfies new specifications through structural extrapolation, with practical methods that work on both synthetic and real-world image data.

Abstract: This paper proposes a framework for Structural Extrapolated Data GEneration (SEDGE) based on suitable assumptions on the underlying data generating process. We provide conditions under which data satisfying new specifications can be generated reliably, together with the approximate identifiability of the distribution of such data under certain ``conservative" assumptions. On the algorithmic side, we develop practical methods to achieve extrapolated data generation, based on the structure-informed optimization strategy or diffusion posterior sampling, respectively. We verify the extrapolation performance on synthetic data and also consider extrapolated image generation as a real-world scenario to illustrate the validity of the proposed framework.

Method: The framework computes effect-size diagnostics across five assumption families (stationarity, irregularity, persistence, nonlinearity, and confounding proxies), aggregates them into four calibrated risk scores with uncertainty intervals, and applies an abstention-aware decision policy that recommends specific causal discovery methods only when evidence supports reliable inference.

Result: Evaluation on 500 synthetic data-generating processes shows well-calibrated risk scores (AUROC > 0.95), 62% false positive reduction among recommended datasets, and 78% abstention on severe-violation cases. On 21 external evaluations, recommend-or-abstain decisions are consistent with benchmark specifications in all cases.

Conclusion: Causal-Audit provides a robust framework for assumption validation in time-series causal discovery, enabling more reliable causal inference through calibrated risk assessment and appropriate method recommendation or abstention.

Abstract: Time-series causal discovery methods rely on assumptions such as stationarity, regular sampling, and bounded temporal dependence. When these assumptions are violated, structure learning can produce confident but misleading causal graphs without warning. We introduce Causal-Audit, a framework that formalizes assumption validation as calibrated risk assessment. The framework computes effect-size diagnostics across five assumption families (stationarity, irregularity, persistence, nonlinearity, and confounding proxies), aggregates them into four calibrated risk scores with uncertainty intervals, and applies an abstention-aware decision policy that recommends methods (e.g., PCMCI+, VAR-based Granger causality) only when evidence supports reliable inference. The semi-automatic diagnostic stage can also be used independently for structured assumption auditing in individual studies. Evaluation on a synthetic atlas of 500 data-generating processes (DGPs) spanning 10 violation families demonstrates well-calibrated risk scores (AUROC > 0.95), a 62% false positive reduction among recommended datasets, and 78% abstention on severe-violation cases. On 21 external evaluations from TimeGraph (18 categories) and CausalTime (3 domains), recommend-or-abstain decisions are consistent with benchmark specifications in all cases. An open-source implementation of our framework is available.

Method: Developed reproducible pipeline for quality control, MORF barcode demultiplexing, per-TF differential expression, and functional enrichment. Used embryoid body cells as external baseline and removed shared batch/transduction artifacts by background subtraction to compensate for missing GFP/mCherry negative controls in deposited data.

Result: Recovered TF-specific signatures for 59 of 61 testable TFs (vs 27 detected by one-vs-rest alone). Identified HOPX, MAZ, PAX6, FOS, and FEZF2 as strongest transcriptional remodelers. Per-TF enrichment linked specific TFs to biological programs. Condition-level analyses revealed convergent signatures. Per-TF effect sizes significantly correlated with published rankings (Spearman ρ=-0.316, p=0.013).

Conclusion: Deposited TF Atlas data can support validated TF-specific transcriptional and pathway analyses when paired with principled external controls, artifact removal, and reproducible computation, demonstrating robust signal recovery despite missing intra-pool controls.

Abstract: Public pooled single-cell perturbation atlases are valuable resources for studying transcription factor (TF) function, but downstream re-analysis can be limited by incomplete deposited metadata and missing internal controls. Here we re-analyze the human TF Atlas dataset (GSE216481), a MORF-based pooled overexpression screen spanning 3,550 TF open reading frames and 254,519 cells, with a reproducible pipeline for quality control, MORF barcode demultiplexing, per-TF differential expression, and functional enrichment. From 77,018 cells in the pooled screen, we assign 60,997 (79.2%) to 87 TF identities. Because the deposited barcode mapping lacks the GFP and mCherry negative controls present in the original library, we use embryoid body (EB) cells as an external baseline and remove shared batch/transduction artifacts by background subtraction. This strategy recovers TF-specific signatures for 59 of 61 testable TFs, compared with 27 detected by one-vs-rest alone, showing that robust TF-level signal can be rescued despite missing intra-pool controls. HOPX, MAZ, PAX6, FOS, and FEZF2 emerge as the strongest transcriptional remodelers, while per-TF enrichment links FEZF2 to regulation of differentiation, EGR1 to Hippo and cardiac programs, FOS to focal adhesion, and NFIC to collagen biosynthesis. Condition-level analyses reveal convergent Wnt, neurogenic, EMT, and Hippo signatures, and Harmony indicates minimal confounding batch effects across pooled replicates. Our per-TF effect sizes significantly agree with Joung et al.’s published rankings (Spearman $ρ= -0.316$, $p = 0.013$; negative because lower rank indicates stronger effect). Together, these results show that the deposited TF Atlas data can support validated TF-specific transcriptional and pathway analyses when paired with principled external controls, artifact removal, and reproducible computation.

Method: Systematic study of outlier patterns reveals three types: Row-wise, Column-wise, and None. Proposes AdaHOP which adaptively assigns optimal strategy per matrix multiplication: Inner Hadamard Transform (IHT) where inner-dimension smoothing works, or IHT combined with selective Outlier Extraction (OE) routing dominant outliers to high-precision path. Combined with hardware-aware Triton kernels.

Result: Achieves BF16 training quality at MXFP4 precision while delivering up to 3.6X memory compression and 1.8X kernel acceleration over BF16 full-precision training.

Conclusion: Adaptive outlier-pattern-aware strategies are crucial for effective low-precision training in LLMs, overcoming limitations of uniform fixed transforms.

Abstract: Low-precision training (LPT) commonly employs Hadamard transforms to suppress outliers and mitigate quantization error in large language models (LLMs). However, prior methods apply a fixed transform uniformly, despite substantial variation in outlier structures across tensors. Through the first systematic study of outlier patterns across weights, activations, and gradients of LLMs, we show that this strategy is fundamentally flawed: the effectiveness of Hadamard-based suppression depends on how the transform’s smoothing direction aligns with the outlier structure of each operand – a property that varies substantially across layers and computation paths. We characterize these patterns into three types: Row-wise, Column-wise, and None. Each pair requires a tailored transform direction or outlier handling strategy to minimize quantization error. Based on this insight, we propose AdaHOP (Adaptive Hadamard transform with Outlier-Pattern-aware strategy), which assigns each matrix multiplication its optimal strategy: Inner Hadamard Transform (IHT) where inner-dimension smoothing is effective, or IHT combined with selective Outlier Extraction (OE) – routing dominant outliers to a high-precision path – where it is not. Combined with hardware-aware Triton kernels, AdaHOP achieves BF16 training quality at MXFP4 precision while delivering up to 3.6X memory compression and 1.8X kernel acceleration} over BF16 full-precision training.

Method: Systematically examine LLM-generated preferences with random and label-flipping noise injection. Develop theoretical analysis decomposing effects of random label noise and systematic misalignment on prior error driving bandit regret. Derive sufficient condition for when LLM-based warm starts outperform cold-start bandits.

Result: For aligned domains: warm-starting remains effective up to 30% corruption, loses advantage around 40%, degrades performance beyond 50%. With systematic misalignment (even without added noise), LLM-generated priors can lead to higher regret than cold-start bandit. Estimated alignment reliably tracks when warm-starting improves or degrades recommendation quality across multiple conjoint datasets and LLMs.

Conclusion: LLM-generated preferences for bandit warm-starting are sensitive to noise and misalignment. The paper provides theoretical understanding and practical guidance on when LLM-based warm starts are beneficial versus harmful, with alignment estimation serving as a reliable indicator.

Abstract: The recent advancement of Large Language Models (LLMs) offers new opportunities to generate user preference data to warm-start bandits. Recent studies on contextual bandits with LLM initialization (CBLI) have shown that these synthetic priors can significantly lower early regret. However, these findings assume that LLM-generated choices are reasonably aligned with actual user preferences. In this paper, we systematically examine how LLM-generated preferences perform when random and label-flipping noise is injected into the synthetic training data. For aligned domains, we find that warm-starting remains effective up to 30% corruption, loses its advantage around 40%, and degrades performance beyond 50%. When there is systematic misalignment, even without added noise, LLM-generated priors can lead to higher regret than a cold-start bandit. To explain these behaviors, we develop a theoretical analysis that decomposes the effect of random label noise and systematic misalignment on the prior error driving the bandit’s regret, and derive a sufficient condition under which LLM-based warm starts are provably better than a cold-start bandit. We validate these results across multiple conjoint datasets and LLMs, showing that estimated alignment reliably tracks when warm-starting improves or degrades recommendation quality.

Method: A spectral framework that embeds high-dimensional data using a spectral basis combined with cross-entropy optimization, enabling multi-scale representations. Leverages linear spectral decomposition to support analysis through a graph-frequency perspective, complemented by glyph-based scatterplot augmentations for visual exploration.

Result: Quantitative evaluations and case studies demonstrate improved manifold continuity while enabling deeper analysis of embedding structure through spectral mode contributions.

Conclusion: The proposed spectral framework addresses key trade-offs in dimensionality reduction by providing both global-local preservation and analytical transparency through spectral analysis and visualization tools.

Abstract: Dimensionality reduction (DR) is characterized by two longstanding trade-offs. First, there is a global-local preservation tension: methods such as t-SNE and UMAP prioritize local neighborhood preservation, yet may distort global manifold structure, while methods such as Laplacian Eigenmaps preserve global geometry but often yield limited local separation. Second, there is a gap between expressiveness and analytical transparency: many nonlinear DR methods produce embeddings without an explicit connection to the underlying high-dimensional structure, limiting insight into the embedding process. In this paper, we introduce a spectral framework for nonlinear DR that addresses these challenges. Our approach embeds high-dimensional data using a spectral basis combined with cross-entropy optimization, enabling multi-scale representations that bridge global and local structure. Leveraging linear spectral decomposition, the framework further supports analysis of embeddings through a graph-frequency perspective, enabling examination of how spectral modes influence the resulting embedding. We complement this analysis with glyph-based scatterplot augmentations for visual exploration. Quantitative evaluations and case studies demonstrate that our framework improves manifold continuity while enabling deeper analysis of embedding structure through spectral mode contributions.

Method: Proposes a lightweight shared memory optimization that exploits memory hierarchy while maintaining ecosystem compatibility. Uses only 64 bytes of shared memory per thread block and simplifies indexing logic to reduce instruction counts.

Result: Achieves 2.0-2.2× kernel speedup across three models (Gemma 27B, Qwen3 32B, Llama3.3 70B) and up to 1.54× end-to-end improvement by leveraging 12-15× latency advantage of shared memory over global memory access.

Conclusion: Lightweight optimizations can deliver substantial performance gains with minimal engineering effort, providing a plug-and-play solution for HuggingFace ecosystem that democratizes access to advanced models on existing GPU infrastructure.

Abstract: Large language models (LLMs) have grown beyond the memory capacity of single GPU devices, necessitating quantization techniques for practical deployment. While NF4 (4-bit NormalFloat) quantization enables 4$\times$ memory reduction, inference on current NVIDIA GPUs (e.g., Ampere A100) requires expensive dequantization back to FP16 format, creating a critical performance bottleneck. This paper presents a lightweight shared memory optimization that addresses this gap through principled memory hierarchy exploitation while maintaining full ecosystem compatibility. We compare our technique against the open-source BitsAndBytes implementation, achieving 2.0–2.2$\times$ kernel speedup across three models (Gemma 27B, Qwen3 32B, and Llama3.3 70B) and up to 1.54$\times$ end-to-end improvement by leveraging the 12–15$\times$ latency advantage of shared memory over global memory access. Our optimization reduces instruction counts through simplified indexing logic while using only 64 bytes of shared memory per thread block, demonstrating that lightweight optimizations can deliver substantial performance gains with minimal engineering effort. This work provides a plug-and-play solution for the HuggingFace ecosystem that democratizes access to advanced models on existing GPU infrastructure.

Method: Local training approach reduces communication frequency, combined with gradient clipping and additive noise perturbation during local training to achieve differential privacy guarantees

Result: Algorithm converges to a stationary point within bounded distance, provides theoretical differential privacy guarantees, and shows superior performance on classification tasks under same privacy budget compared to state-of-the-art methods

Conclusion: Proposed algorithm successfully balances communication efficiency and privacy preservation in decentralized learning, offering practical solution for privacy-sensitive distributed learning scenarios

Abstract: We address nonconvex learning problems over undirected networks. In particular, we focus on the challenge of designing an algorithm that is both communication-efficient and that guarantees the privacy of the agents’ data. The first goal is achieved through a local training approach, which reduces communication frequency. The second goal is achieved by perturbing gradients during local training, specifically through gradient clipping and additive noise. We prove that the resulting algorithm converges to a stationary point of the problem within a bounded distance. Additionally, we provide theoretical privacy guarantees within a differential privacy framework that ensure agents’ training data cannot be inferred from the trained model shared over the network. We show the algorithm’s superior performance on a classification task under the same privacy budget, compared with state-of-the-art methods.

Method: ROMAN builds an anti-aliased multiscale pyramid, extracts fixed-length windows from each scale, and stacks them as pseudochannels. This creates a compact representation where temporal scale and coarse position are mapped to channel structure while reducing sequence length.

Result: On synthetic tasks isolating specific temporal properties, ROMAN behaves as intended and is most useful when class information depends on temporal structure that standard convolution suppresses. On real-world UCR/UEA datasets, ROMAN’s effect on accuracy is task-dependent but often improves computational efficiency.

Conclusion: ROMAN provides a simple preprocessing mechanism that controls inductive bias for convolutional classifiers, enabling them to better handle temporal structure while often improving efficiency through sequence length reduction.

Abstract: We introduce ROMAN (ROuting Multiscale representAtioN), a deterministic operator for time series that maps temporal scale and coarse temporal position into an explicit channel structure while reducing sequence length. ROMAN builds an anti-aliased multiscale pyramid, extracts fixed-length windows from each scale, and stacks them as pseudochannels, yielding a compact representation on which standard convolutional classifiers can operate. In this way, ROMAN provides a simple mechanism to control the inductive bias of downstream models: it can reduce temporal invariance, make temporal pooling implicitly coarse-position-aware, and expose multiscale interactions through channel mixing, while often improving computational efficiency by shortening the processed time axis. We formally analyze the ROMAN operator and then evaluate it in two complementary ways by measuring its impact as a preprocessing step for four representative convolutional classifiers: MiniRocket, MultiRocket, a standard CNN-based classifier, and a fully convolutional network (FCN) classifier. First, we design synthetic time series classification tasks that isolate coarse position awareness, long-range correlation, multiscale interaction, and full positional invariance, showing that ROMAN behaves consistently with its intended mechanism and is most useful when class information depends on temporal structure that standard pooled convolution tends to suppress. Second, we benchmark the same models with and without ROMAN on long-sequence subsets of the UCR and UEA archives, showing that ROMAN provides a practically useful alternative representation whose effect on accuracy is task-dependent, but whose effect on efficiency is often favorable. Code is available at https://github.com/gon-uri/ROMAN

Method: Developed VoxelCode platform integrating natural language task specification, API-driven code execution in Unreal Engine, and unified evaluation pipeline with automated metrics and human assessment. Created VoxelCodeBench benchmark with voxel manipulation tasks across symbolic interpretation, geometric construction, and artistic composition dimensions.

Result: Evaluation of leading code generation models shows producing executable code is easier than producing spatially correct outputs, with geometric construction and multi-object composition being particularly challenging areas.

Conclusion: VoxelCode provides extensible infrastructure for developing 3D code generation benchmarks and probing spatial reasoning in future models, highlighting the gap between code execution and spatial correctness in current models.

Abstract: Evaluating code generation models for 3D spatial reasoning requires executing generated code in realistic environments and assessing outputs beyond surface-level correctness. We introduce a platform VoxelCode, for analyzing code generation capabilities for 3D understanding and environment creation. Our platform integrates natural language task specification, API-driven code execution in Unreal Engine, and a unified evaluation pipeline supporting both automated metrics and human assessment. To demonstrate its utility, we construct VoxelCodeBench, a benchmark of voxel manipulation tasks spanning three reasoning dimensions: symbolic interpretation, geometric construction, and artistic composition. Evaluating leading code generation models, we find that producing executable code is far easier than producing spatially correct outputs, with geometric construction and multi-object composition proving particularly challenging. By open-sourcing our platform and benchmark, we provide the community with extensible infrastructure for developing new 3D code generation benchmarks and probing spatial reasoning in future models.

Method: Propose weak formulation-based GENERIC formalism informed neural networks (WGFINNs) that integrate weak formulation of dynamical systems with structure-preserving GFINN architecture, plus state-wise weighted loss and residual-based attention mechanism.

Result: WGFINNs consistently outperform GFINNs at varying noise levels, achieving more accurate predictions and reliable recovery of physical quantities. Theoretical analysis shows strong-form estimator diverges with decreasing time step in noise, while weak-form estimator can remain accurate.

Conclusion: WGFINNs provide a robust framework for data-driven discovery of governing equations from noisy observations while exactly satisfying GENERIC thermodynamic constraints.

Abstract: Data-driven discovery of governing equations from noisy observations remains a fundamental challenge in scientific machine learning. While GENERIC formalism informed neural networks (GFINNs) provide a principled framework that enforces the laws of thermodynamics by construction, their reliance on strong-form loss formulations makes them highly sensitive to measurement noise. To address this limitation, we propose weak formulation-based GENERIC formalism informed neural networks (WGFINNs), which integrate the weak formulation of dynamical systems with the structure-preserving architecture of GFINNs. WGFINNs significantly enhance robustness to noisy data while retaining exact satisfaction of GENERIC degeneracy and symmetry conditions. We further incorporate a state-wise weighted loss and a residual-based attention mechanism to mitigate scale imbalance across state variables. Theoretical analysis contrasts quantitative differences between the strong-form and the weak-form estimators. Mainly, the strong-form estimator diverges as the time step decreases in the presence of noise, while the weak-form estimator can be accurate even with noisy data if test functions satisfy certain conditions. Numerical experiments demonstrate that WGFINNs consistently outperform GFINNs at varying noise levels, achieving more accurate predictions and reliable recovery of physical quantities.

Method: Comprehensive cross-template FV transfer study across 4,032 pairs, 12 tasks, 6 models from 3 families (Llama, Gemma, Mistral), 8 templates per task. Analyzed steering accuracy vs logit lens decodability, FV vocabulary projection, optimal intervention layers, and activation patching for causal localization.

Result: Found universal steerability-without-decodability pattern: FV steering succeeds even when logit lens cannot decode correct answers at any layer. Steering exceeds logit lens accuracy for every task on every model, with gaps up to -0.91. FVs encode computational instructions rather than answer directions, intervene optimally at early layers (L2-L8), while logit lens detects answers only at late layers (L28-L32).

Conclusion: Function vectors represent computational instructions that steer model behavior through early-layer interventions, operating independently from answer representations that only emerge in late layers. This dissociation challenges previous assumptions about FV mechanisms and reveals fundamental differences in how models process tasks.

Abstract: Function vectors (FVs) – mean-difference directions extracted from in-context learning demonstrations – can steer large language model behavior when added to the residual stream. We hypothesized that FV steering failures reflect an absence of task-relevant information: the logit lens would fail alongside steering. We were wrong. In the most comprehensive cross-template FV transfer study to date - 4,032 pairs across 12 tasks, 6 models from 3 families (Llama-3.1-8B, Gemma-2-9B, Mistral-7B-v0.3; base and instruction-tuned), 8 templates per task - we find the opposite dissociation: FV steering succeeds even when the logit lens cannot decode the correct answer at any layer. This steerability-without-decodability pattern is universal: steering exceeds logit lens accuracy for every task on every model, with gaps as large as -0.91. Only 3 of 72 task-model instances show the predicted decodable-without-steerable pattern, all in Mistral. FV vocabulary projection reveals that FVs achieving over 0.90 steering accuracy still project to incoherent token distributions, indicating FVs encode computational instructions rather than answer directions. FVs intervene optimally at early layers (L2-L8); the logit lens detects correct answers only at late layers (L28-L32). The previously reported negative cosine-transfer correlation (r=-0.572) dissolves at scale: pooled r ranges from -0.199 to +0.126, and cosine adds less than 0.011 in R-squared beyond task identity. Post-steering analysis reveals a model-family divergence: Mistral FVs rewrite intermediate representations; Llama/Gemma FVs produce near-zero changes despite successful steering. Activation patching confirms causal localization: easy tasks achieve perfect recovery at targeted layers; hard tasks show zero recovery everywhere.

Method: Uses complex domain to represent 2D geometric features and transformations between bases. Employs complex-valued linear layers with phase-equivariant activation functions inside GNN layers to achieve global invariance to local frame choices.

Result: The architecture increases data efficiency, tracking performance, and generalization compared to real-valued baselines on an imitation learning flocking task.

Conclusion: The proposed complex-valued GNN architecture enables distributed control that is invariant to local coordinate frames, making it suitable for GPS/compass-denied environments.

Abstract: Graph neural networks (GNNs) are a well-regarded tool for learned control of networked dynamical systems due to their ability to be deployed in a distributed manner. However, current distributed GNN architectures assume that all nodes in the network collect geometric observations in compatible bases, which limits the usefulness of such controllers in GPS-denied and compass-denied environments. This paper presents a GNN parametrization that is globally invariant to choice of local basis. 2D geometric features and transformations between bases are expressed in the complex domain. Inside each GNN layer, complex-valued linear layers with phase-equivariant activation functions are used. When viewed from a fixed global frame, all policies learned by this architecture are strictly invariant to choice of local frames. This architecture is shown to increase the data efficiency, tracking performance, and generalization of learned control when compared to a real-valued baseline on an imitation learning flocking task.

Method: Proposes Analytic Drift Resister (ADR) with three components: 1) Iterative backpropagation to adapt to evolving task distributions and enhance model plasticity, 2) Hierarchical Analytic Merging (HAM) using ridge regression for layer-wise merging of GNN linear transformations to prevent feature drift, 3) Analytic Classifier Reconstruction (ACR) for theoretically zero-forgetting class-incremental learning.

Result: Empirical evaluation on four node classification benchmarks shows ADR maintains strong competitiveness against existing state-of-the-art continual learning methods.

Conclusion: ADR successfully addresses the trade-off between privacy protection (no raw data storage) and continual learning performance by preventing feature drift while maintaining model plasticity, achieving theoretically zero-forgetting in class-incremental learning scenarios.

Abstract: Non-Exemplar Continual Graph Learning (NECGL) seeks to eliminate the privacy risks intrinsic to rehearsal-based paradigms by retaining solely class-level prototype representations rather than raw graph examples for mitigating catastrophic forgetting. However, this design choice inevitably precipitates feature drift. As a nascent alternative, Analytic Continual Learning (ACL) capitalizes on the intrinsic generalization properties of frozen pre-trained models to bolster continual learning performance. Nonetheless, a key drawback resides in the pronounced attenuation of model plasticity. To surmount these challenges, we propose Analytic Drift Resister (ADR), a novel and theoretically grounded NECGL framework. ADR exploits iterative backpropagation to break free from the frozen pre-trained constraint, adapting to evolving task graph distributions and fortifying model plasticity. Since parameter updates trigger feature drift, we further propose Hierarchical Analytic Merging (HAM), performing layer-wise merging of linear transformations in Graph Neural Networks (GNNs) via ridge regression, thereby ensuring absolute resistance to feature drift. On this basis, Analytic Classifier Reconstruction (ACR) enables theoretically zero-forgetting class-incremental learning. Empirical evaluation on four node classification benchmarks demonstrates that ADR maintains strong competitiveness against existing state-of-the-art methods.

Method: Uses orthogonal-transform-based quantization to create a fixed codebook that depends only on dimension and bit-width, not input data. Implements asymmetric path design with transform on write and table-lookup on read, eliminating inverse transforms. Includes gradient-free sign pattern selection to address catastrophic performance spikes.

Result: Achieves 102.4x reduction in per-query multiplications. Identifies sign pattern sensitivity causing catastrophic PPL spikes in some models (Qwen2.5-3B) but not others (LLaMA-3.1-8B). Proposes solution that eliminates spikes with zero additional hardware cost.

Conclusion: AXELRAM enables efficient attention computation directly from quantized KV cache indices, addressing both computational efficiency and stability issues through novel quantization and sign pattern selection techniques.

Abstract: We propose AXELRAM, a smart SRAM macro architecture that computes attention scores directly from quantized KV cache indices without dequantization. The key enabler is a design-time fixed codebook: orthogonal-transform-based quantization concentrates each coordinate’s distribution to N(0,1/d), so the optimal quantizer depends only on dimension d and bit-width b, not on input data. The asymmetric path design – transform on write, table-lookup on read with no inverse transform – reduces per-query multiplications by 102.4x (a mathematical identity). Through multi-seed evaluation (10 seeds x 3 models), we discover that sign pattern sensitivity causes catastrophic PPL spikes (Delta > 50) on certain models (Qwen2.5-3B), while others (LLaMA-3.1-8B) are fully stable. This phenomenon extends SpinQuant’s observation of rotation variance in weight quantization to the KV cache domain, where the effect is qualitatively more severe. We trace the root cause to layer-wise norm heterogeneity and propose a gradient-free sign pattern selection (200 candidates, 8 calibration samples, one-time) that eliminates catastrophic spikes with zero additional hardware cost. All source code is available at https://github.com/Axelidea/AXELRAM.

Method: Modify Wasserstein autoencoders with (block-) triangular decoders and impose independence assumptions on latent variables. Present three architectural variants with connections to conditional optimal transport problems.

Result: CWAEs achieve substantial reductions in approximation error relative to low-rank ensemble Kalman filters, particularly in problems where conditional measures have truly low-dimensional support.

Conclusion: CWAEs provide an effective framework for conditional simulation that leverages low-dimensional structure through modified autoencoder architectures with theoretical connections to optimal transport.

Abstract: We present Conditional Wasserstein Autoencoders (CWAEs), a framework for conditional simulation that exploits low-dimensional structure in both the conditioned and the conditioning variables. The key idea is to modify a Wasserstein autoencoder to use a (block-) triangular decoder and impose an appropriate independence assumption on the latent variables. We show that the resulting model gives an autoencoder that can exploit low-dimensional structure while simultaneously the decoder can be used for conditional simulation. We explore various theoretical properties of CWAEs, including their connections to conditional optimal transport (OT) problems. We also present alternative formulations that lead to three architectural variants forming the foundation of our algorithms. We present a series of numerical experiments that demonstrate that our different CWAE variants achieve substantial reductions in approximation error relative to the low-rank ensemble Kalman filter (LREnKF), particularly in problems where the support of the conditional measures is truly low-dimensional.

Method: Introduces uniform vertex sampling for communication-free local mini-batch construction, 3D parallel matrix multiplication for scaling to large GPU counts, plus optimizations like sampling-training overlap, lower precision communication, kernel fusion, and communication-computation overlap.

Result: Demonstrates strong scaling up to 2048 GPUs on Perlmutter, 2048 GCDs on Frontier, and 1024 GPUs on Tuolumne, achieving 3.5x end-to-end training speedup over SOTA baseline on ogbn-products dataset.

Conclusion: ScaleGNN provides an effective 4D parallel framework for scalable mini-batch GNN training that overcomes limitations of existing distributed approaches and enables training on extremely large graphs.

Abstract: Graph neural networks (GNNs) are widely used for learning on graph datasets derived from various real-world scenarios. Learning from extremely large graphs requires distributed training, and mini-batching with sampling is a popular approach for parallelizing GNN training. Existing distributed mini-batch approaches have significant performance bottlenecks due to expensive sampling methods and limited scaling when using data parallelism. In this work, we present ScaleGNN, a 4D parallel framework for scalable mini-batch GNN training that combines communication-free distributed sampling, 3D parallel matrix multiplication (PMM), and data parallelism. ScaleGNN introduces a uniform vertex sampling algorithm, enabling each process (GPU device) to construct its local mini-batch, i.e., subgraph partitions without any inter-process communication. 3D PMM enables scaling mini-batch training to much larger GPU counts than vanilla data parallelism with significantly lower communication overheads. We also present additional optimizations to overlap sampling with training, reduce communication overhead by sending data in lower precision, kernel fusion, and communication-computation overlap. We evaluate ScaleGNN on five graph datasets and demonstrate strong scaling up to 2048 GPUs on Perlmutter, 2048 GCDs on Frontier, and 1024 GPUs on Tuolumne. On Perlmutter, ScaleGNN achieves 3.5x end-to-end training speedup over the SOTA baseline on ogbn-products.

Method: Proposes “compound jailbreaks” targeting OpenAI GPT models that combine multiple attack techniques (each individually defended against) to saturate the instruction hierarchy maintenance process, exploiting generalization failures in alignment.

Result: Attack success rate increased from 14.3% with individual methods to 71.4% with the combined compound approach, showing significant vulnerability when multiple techniques are used together.

Conclusion: Safety training doesn’t generalize as broadly as model capabilities, highlighting the need for multifaceted safety evaluations using compound attack scenarios rather than testing individual defenses in isolation.

Abstract: The safety of large language models (LLMs) relies on alignment techniques such as reinforcement learning from human feedback (RLHF). However, recent theoretical analyses suggest that reinforcement learning-based training does not acquire new capabilities but merely redistributes the utilization probabilities of existing ones. In this study, we propose ``compound jailbreaks’’ targeting OpenAI gpt-oss-20b, which exploit the generalization failures of alignment. This approach combines multiple attack techniques – each individually defended against – to saturate the instruction hierarchy maintenance process. Our evaluation shows that the attack success rate (ASR) increased from 14.3% with individual methods to 71.4% with the combined approach. These results provide empirical evidence for the hypothesis that safety training does not generalize as broadly as model capabilities, highlighting the need for multifaceted safety evaluations using compound attack scenarios.

Method: Introduces product-stability structural property of loss functions. Analyzes gradient descent applied to objectives of form (x,y) → l(xy) using bifurcation diagrams to characterize training dynamics and quantify sharpness at convergence.

Result: Proves gradient descent can provably converge to local minima even in Edge of Stability regime for losses with product-stable minima. Characterizes training dynamics, explains emergence of stable oscillations, and precisely quantifies sharpness at convergence.

Conclusion: Provides principled explanation for stable Edge of Stability training for wider class of loss functions including binary cross entropy, substantially generalizing prior theoretical results beyond squared-loss objectives.

Abstract: Empirically, modern deep learning training often occurs at the Edge of Stability (EoS), where the sharpness of the loss exceeds the threshold below which classical convergence analysis applies. Despite recent progress, existing theoretical explanations of EoS either rely on restrictive assumptions or focus on specific squared-loss-type objectives. In this work, we introduce and study a structural property of loss functions that we term product-stability. We show that for losses with product-stable minima, gradient descent applied to objectives of the form $(x,y) \mapsto l(xy)$ can provably converge to the local minimum even when training in the EoS regime. This framework substantially generalizes prior results and applies to a broad class of losses, including binary cross entropy. Using bifurcation diagrams, we characterize the resulting training dynamics, explain the emergence of stable oscillations, and precisely quantify the sharpness at convergence. Together, our results offer a principled explanation for stable EoS training for a wider class of loss functions.

Method: Establishes theoretical connection between low-rank approximation error and predictive performance via softmax perturbation analysis. Proposes randomized subspace iteration (RSI) with multiple power iterations to improve spectral separation and approximation quality.

Result: RSI achieves near-optimal approximation quality and outperforms RSVD in predictive accuracy under aggressive compression. Evaluated on both convolutional networks and transformer architectures.

Conclusion: RSI provides an effective, controllable alternative to RSVD for efficient model compression, enabling practical deployment of large pretrained models while maintaining performance.

Abstract: The massive scale of pretrained models has made efficient compression essential for practical deployment. Low-rank decomposition based on the singular value decomposition (SVD) provides a principled approach for model reduction, but its exact computation is expensive for large weight matrices. Randomized alternatives such as randomized SVD (RSVD) improve efficiency, yet they can suffer from poor approximation quality when the singular value spectrum decays slowly, a regime commonly observed in modern pretrained models. In this work, we address this limitation from both theoretical and empirical perspectives. First, we establish a connection between low-rank approximation error and predictive performance by analyzing softmax perturbations, showing that deviations in class probabilities are controlled by the spectral error of the compressed weights. Second, we demonstrate that RSVD is inadequate, and we propose randomized subspace iteration (RSI) as a more effective alternative. By incorporating multiple power iterations, RSI improves spectral separation and provides a controllable mechanism for enhancing approximation quality. We evaluate our approach on both convolutional networks and transformer-based architectures. Our results show that RSI achieves near-optimal approximation quality while outperforming RSVD in predictive accuracy under aggressive compression, enabling efficient model compression.

Method: The paper analyzes why OpenWebText became the standard benchmark over alternatives like LM1B, discusses limitations of likelihood evaluations for diffusion models, decomposes generative perplexity into entropy and KL divergence components, and proposes generative frontiers as a principled evaluation method.

Result: The analysis reveals that generative perplexity alone can be uninformative due to its sensitivity to entropy, and that generative frontiers provide a more principled approach to evaluating model generative quality at the GPT-2 small scale.

Conclusion: Current diffusion language model evaluation methodology needs improvement, and generative frontiers offer a more reliable approach for comparing model quality, with empirical observations provided at the 150M parameter scale.

Abstract: Diffusion language models have seen exciting recent progress, offering far more flexibility in generative trajectories than autoregressive models. This flexibility has motivated a growing body of research into new approaches to diffusion language modeling, which typically begins at the scale of GPT-2 small (150 million parameters). However, these advances introduce new issues with evaluation methodology. In this technical note, we discuss the limitations of current methodology and propose principled augmentations to ensure reliable comparisons. We first discuss why OpenWebText has become the standard benchmark, and why alternatives such as LM1B are inherently less meaningful. We then discuss the limitations of likelihood evaluations for diffusion models, and explain why relying on generative perplexity alone as a metric can lead to uninformative results. To address this, we show that generative perplexity and entropy are two components of the KL divergence to a reference distribution. This decomposition explains generative perplexity’s sensitivity to entropy, and naturally suggests generative frontiers as a principled method for evaluating model generative quality. We conclude with empirical observations on model quality at this scale. We include a blog post with interactive content to illustrate the argument at https://patrickpynadath1.github.io/blog/eval_methodology/.

Method: Developed P2F method with Node-Assigned PINN (NA-PINN) that accepts water-level difference, initial velocity, and time as inputs to learn solution manifold. Coupled with FDM solver for mass conservation, replacing iterative nonlinear momentum solves with single forward pass.

Result: Achieved water level MAE of 7.85×10⁻⁵ m and velocity MAE of 3.21×10⁻³ m/s in six-tank gravity-driven draining scenario. Maintained accuracy across time steps (0.2-1.0 s) and generalized to five initial conditions without retraining or simulation data.

Conclusion: P2F method provides efficient data-free surrogate modeling for nuclear thermal-hydraulic systems, enabling parametric studies without retraining or large simulation datasets.

Abstract: Severe accident analysis using system-level codes such as MELCOR is indispensable for nuclear safety assessment, yet the computational cost of repeated simulations poses a significant bottleneck for parametric studies and uncertainty quantification. Existing surrogate models accelerate these analyses but depend on large volumes of simulation data, while physics-informed neural networks (PINNs) enable data-free training but must be retrained for every change in problem parameters. This study addresses both limitations by developing the Parameterized PINNs coupled with FDM (P2F) method, a node-assigned hybrid framework for MELCOR’s Control Volume Hydrodynamics/Flow Path (CVH/FP) module. In the P2F method, a parameterized Node-Assigned PINN (NA-PINN) accepts the water-level difference, initial velocity, and time as inputs, learning a solution manifold so that a single trained network serves as a data-free surrogate for the momentum conservation equation across all flow paths without retraining. This PINN is coupled with a finite difference method (FDM) solver that advances the mass conservation equation at each time step, ensuring exact discrete mass conservation while replacing the iterative nonlinear momentum solve with a single forward pass. Verification on a six-tank gravity-driven draining scenario yields a water level mean absolute error of $7.85 \times 10^{-5}$ m and a velocity mean absolute error of $3.21 \times 10^{-3}$ m/s under the nominal condition with $Δt = 1.0$ s. The framework maintains consistent accuracy across time steps ranging from 0.2 to 1.0 s and generalizes to five distinct initial conditions, all without retraining or simulation data. This work introduces a numerical coupling methodology for integrating parameterized PINNs with FDM within a nuclear thermal-hydraulic system code framework.

Method: Proposes a neural network with: 1) Inception-attention module as feature extractor, 2) fatigue classifier, 3) domain classifier with gradient reversal layer to learn subject-invariant features, and 4) supervised contrastive loss for better generalization.

Result: Achieved 93.54% accuracy, 92.69% recall, and 92.69% F1-score in three-class classification tasks for cross-subject muscle fatigue detection.

Conclusion: The proposed model provides a robust solution for cross-subject muscle fatigue detection, offering significant guidance for rehabilitation training and assistance.

Abstract: Muscle fatigue detection plays an important role in physical rehabilitation. Previous researches have demonstrated that sEMG offers superior sensitivity in detecting muscle fatigue compared to other biological signals. However, features extracted from sEMG may vary during dynamic contractions and across different subjects, which causes unstability in fatigue detection. To address these challenges, this research proposes a novel neural network comprising an Inception-attention module as a feature extractor, a fatigue classifier and a domain classifier equipped with a gradient reversal layer. The integrated domain classifier encourages the network to learn subject-invariant common fatigue features while minimizing subject-specific features. Furthermore, a supervised contrastive loss function is also employed to enhance the generalization capability of the model. Experimental results demonstrate that the proposed model achieved outstanding performance in three-class classification tasks, reaching 93.54% accuracy, 92.69% recall and 92.69% F1-score, providing a robust solution for cross-subject muscle fatigue detection, offering significant guidance for rehabilitation training and assistance.

Method: Pipeline combining sparse autoencoders (SAEs), k-subspace clustering of SAE features, and simplex fitting using AANet, validated on synthetic data and applied to Gemma-2-9B.

Result: Identified 13 priority clusters with candidate simplex geometry in Gemma-2-9B; 3 showed highly significant advantage on near-vertex samples, 4 on simplex-interior samples, with 5 distinct clusters passing at least one test.

Conclusion: Preliminary evidence suggests genuine belief-like geometry exists in Gemma-2-9B’s representation space, with one cluster (768_596) showing both predictive advantage and causal steering capability.

Abstract: Understanding the geometric structure of internal representations is a central goal of mechanistic interpretability. Prior work has shown that transformers trained on sequences generated by hidden Markov models encode probabilistic belief states as simplex-shaped geometries in their residual stream, with vertices corresponding to latent generative states. Whether large language models trained on naturalistic text develop analogous geometric representations remains an open question. We introduce a pipeline for discovering candidate simplex-structured subspaces in transformer representations, combining sparse autoencoders (SAEs), $k$-subspace clustering of SAE features, and simplex fitting using AANet. We validate the pipeline on a transformer trained on a multipartite hidden Markov model with known belief-state geometry. Applied to Gemma-2-9B, we identify 13 priority clusters exhibiting candidate simplex geometry ($K \geq 3$). A key challenge is distinguishing genuine belief-state encoding from tiling artifacts: latents can span a simplex-shaped subspace without the mixture coordinates carrying predictive signal beyond any individual feature. We therefore adopt barycentric prediction as our primary discriminating test. Among the 13 priority clusters, 3 exhibit a highly significant advantage on near-vertex samples (Wilcoxon $p < 10^{-14}$) and 4 on simplex-interior samples. Together 5 distinct real clusters pass at least one split, while no null cluster passes either. One cluster, 768_596, additionally achieves the highest causal steering score in the dataset. This is the only case where passive prediction and active intervention converge. We present these findings as preliminary evidence that genuine belief-like geometry exists in Gemma-2-9B’s representation space, and identify the structured evaluation that would be required to confirm this interpretation.

Method: Token Mapping Perturbation Attack (TOMPA) framework that bypasses the standard decode-re-tokenize interface between policy and reward model. It uses black-box scalar feedback to optimize over raw token sequences rather than coherent natural language, discovering non-linguistic token patterns that exploit RM biases.

Result: TOMPA nearly doubles the reward of GPT-5 reference answers on Skywork-Reward-V2-Llama-3.1-8B and outperforms them on 98.0% of prompts. The generated outputs degenerate into nonsensical text, demonstrating systematic exploitation beyond the semantic regime.

Conclusion: Reward models can be systematically exploited beyond semantic understanding, exposing a critical vulnerability in current RLHF pipelines that requires attention for more robust alignment methods.

Abstract: Reward models (RMs) are widely used as optimization targets in reinforcement learning from human feedback (RLHF), yet they remain vulnerable to reward hacking. Existing attacks mainly operate within the semantic space, constructing human-readable adversarial outputs that exploit RM biases. In this work, we introduce a fundamentally different paradigm: Token Mapping Perturbation Attack (TOMPA), a framework that performs adversarial optimization directly in token space. By bypassing the standard decode-re-tokenize interface between the policy and the reward model, TOMPA enables the attack policy to optimize over raw token sequences rather than coherent natural language. Using only black-box scalar feedback, TOMPA automatically discovers non-linguistic token patterns that elicit extremely high rewards across multiple state-of-the-art RMs. Specifically, when targeting Skywork-Reward-V2-Llama-3.1-8B, TOMPA nearly doubles the reward of GPT-5 reference answers and outperforms them on 98.0% of prompts. Despite these high scores, the generated outputs degenerate into nonsensical text, revealing that RMs can be systematically exploited beyond the semantic regime and exposing a critical vulnerability in current RLHF pipelines.

Method: Proposes Sign-Certified Policy Optimization (SignCert-PO) that computes certified sign-preservation radius via adversarial perturbations in RM parameter space, then down-weights non-robust completions in policy gradient updates using only RM parameters and on-policy completions.

Result: On TL;DR summarization and AlpacaFarm benchmarks, SignCert-PO consistently achieves better win rates than baselines and reduces reward hacking.

Conclusion: SignCert-PO provides a lightweight, data-efficient approach to mitigate reward hacking in RLHF without requiring multiple RMs or access to training data.

Abstract: Reward models (RMs) used in reinforcement learning from human feedback (RLHF) are vulnerable to reward hacking: as the policy maximizes a learned proxy reward, true quality plateaus or degrades. We make the assumption that reward hacking is often caused by flipped advantage signs: instead of reducing the likelihood of a bad response, a flipped sign causes the update to increase it. By considering an adversarial perturbation in the RM parameter space, we can derive a certified sign-preservation radius, which is the smallest perturbation that can flip the advantage sign during policy optimization. Based on this formulation, we propose Sign-Certified Policy Optimization (SignCert-PO), down-weighting non-robust completions in the policy gradient update. Unlike prior approaches that require multiple RMs or access to the RM training data, SignCert-PO is lightweight and operates purely at the policy optimization stage using only the RM parameters and on-policy completions. On TL;DR summarization and AlpacaFarm benchmarks, SignCert-PO consistently achieves a better win rate than baselines and reduces reward hacking.

Method: Multi-stage end-to-end image semantic communication system for MIMO channels built on adaptive MoE Swin Transformer block with dynamic expert gating mechanism that jointly evaluates real-time CSI and semantic content of input image patches to compute adaptive routing probabilities.

Result: Simulation results show significant improvement in reconstruction quality over existing methods while maintaining transmission efficiency.

Conclusion: The proposed approach breaks rigid coupling of traditional adaptive methods and overcomes bottlenecks of single-driven routing by jointly considering both channel conditions and semantic content for expert activation.

Abstract: Deep learning based semantic communication has achieved significant progress in wireless image transmission, but most existing schemes rely on fixed models and thus lack robustness to diverse image contents and dynamic channel conditions. To improve adaptability, recent studies have developed adaptive semantic communication strategies that adjust transmission or model behavior according to either source content or channel state. More recently, MoE-based semantic communication has emerged as a sparse and efficient adaptive architecture, although existing designs still mainly rely on single-driven routing. To address this limitation, we propose a novel multi-stage end-to-end image semantic communication system for multi-input multi-output (MIMO) channels, built upon an adaptive MoE Swin Transformer block. Specifically, we introduce a dynamic expert gating mechanism that jointly evaluates both real-time CSI and the semantic content of input image patches to compute adaptive routing probabilities. By selectively activating only a specialized subset of experts based on this joint condition, our approach breaks the rigid coupling of traditional adaptive methods and overcomes the bottlenecks of single-driven routing. Simulation results indicate a significant improvement in reconstruction quality over existing methods while maintaining the transmission efficiency.

Method: Model parameterized quantum circuits as Lie subalgebras of u(2^n), creating a Riemannian manifold of reachable states. Restrict to structured Lie subalgebras (LieTrunc) to contract the manifold and prevent concentration of measure that causes gradient suppression.

Result: Established polynomial trainability regime where gradient variance decays polynomially instead of exponentially. Experiments (n=2-6) show LieTrunc-QNN maintains stable gradients and high effective dimension, preserving full metric rank at n=6 (rank=16).

Conclusion: Provides a unified geometric framework linking Lie algebra, manifold geometry, and optimization for QNN design, showing expressivity is governed by algebraic structure rather than parameter count, with inherent robustness benefits.

Abstract: Quantum Machine Learning (QML) is fundamentally limited by two challenges: barren plateaus (exponentially vanishing gradients) and the fragility of parameterized quantum circuits under noise. Despite extensive empirical studies, a unified theoretical framework remains lacking. We introduce LieTrunc-QNN, an algebraic-geometric framework that characterizes trainability via Lie-generated dynamics. Parameterized quantum circuits are modeled as Lie subalgebras of u(2^n), whose action induces a Riemannian manifold of reachable quantum states. Expressivity is reinterpreted as intrinsic manifold dimension and geometry. We establish a geometric capacity-plateau principle: increasing effective dimension leads to exponential gradient suppression due to concentration of measure. By restricting to structured Lie subalgebras (LieTrunc), the manifold is contracted, preventing concentration and preserving non-degenerate gradients. We prove two main results: (1) a trainability lower bound for LieTrunc-QNN, and (2) that the Fubini-Study metric rank is bounded by the algebraic span of generators, showing expressivity is governed by structure rather than parameter count. Compact Lie subalgebras also provide inherent robustness to perturbations. Importantly, we establish a polynomial trainability regime where gradient variance decays polynomially instead of exponentially. Experiments (n=2-6) validate the theory: LieTrunc-QNN maintains stable gradients and high effective dimension, while random truncation leads to metric rank collapse. At n=6, full metric rank is preserved (rank=16). Results support a scaling law between gradient variance and effective dimension. This work provides a unified geometric framework for QNN design, linking Lie algebra, manifold geometry, and optimization.

Method: Proposes Self-Guide, where agents generate internal rewards used for: 1) short self-guidance signals to steer next actions during inference, and 2) step-level internal rewards for denser policy optimization during training. This creates a co-evolving loop where better policy produces better guidance, and better guidance improves policy as internal reward.

Result: Across three agent benchmarks, inference-time self-guidance yields clear gains. Jointly evolving policy and internal reward with GRPO brings 8% improvements over baselines trained solely with environment reward.

Conclusion: Language agents can improve not only by collecting more experience, but also by learning to generate and refine their own internal reward during acting and learning.

Abstract: Large language model (LLM) agents learn by interacting with environments, but long-horizon training remains fundamentally bottlenecked by sparse and delayed rewards. Existing methods typically address this challenge through post-hoc credit assignment or external reward models, which provide limited guidance at inference time and often separate reward improvement from policy improvement. We propose Self-Guide, a self-generated internal reward for language agents that supports both inference-time guidance and training-time supervision. Specifically, the agent uses Self-Guide as a short self-guidance signal to steer the next action during inference, and converts the same signal into step-level internal reward for denser policy optimization during training. This creates a co-evolving loop: better policy produces better guidance, and better guidance further improves policy as internal reward. Across three agent benchmarks, inference-time self-guidance already yields clear gains, while jointly evolving policy and internal reward with GRPO brings further improvements (8%) over baselines trained solely with environment reward. Overall, our results suggest that language agents can improve not only by collecting more experience, but also by learning to generate and refine their own internal reward during acting and learning.

Method: Introduces expert paging abstraction that treats expert weights as streamed, transient resources, materializing them on demand and evicting them immediately after use, allowing GPU memory to be preferentially allocated to throughput-critical runtime state like KV cache.

Result: FluxMoE achieves up to 3.0× throughput gains over vLLM in memory-intensive regimes without compromising model fidelity.

Conclusion: FluxMoE enables efficient MoE inference under severe memory constraints by decoupling expert parameters from persistent GPU residency, significantly improving throughput.

Abstract: Mixture-of-Experts (MoE) models have become a dominant paradigm for scaling large language models, but their rapidly growing parameter sizes introduce a fundamental inefficiency during inference: most expert weights remain idle in GPU memory while competing with performance-critical runtime state such as the key-value (KV) cache. Since KV cache capacity directly determines serving throughput, this mismatch leads to underutilized memory and degraded performance. In this paper, we present FluxMoE, a new MoE inference system that decouples expert parameters from persistent GPU residency. FluxMoE introduces an expert paging abstraction that treats expert weights as streamed, transient resources, materializing them on demand and evicting them immediately after use, allowing GPU memory to be preferentially allocated to throughput-critical runtime state. We implement FluxMoE atop vLLM to enable efficient MoE inference under severe memory constraints. Experimental results demonstrate that FluxMoE achieves up to 3.0$\times$ throughput gains over vLLM in memory-intensive regimes, without compromising model fidelity.

Method: The authors develop a framework to quantify latent-space diffusability through the rate of change of Minimum Mean Squared Error (MMSE) along the diffusion trajectory. They decompose this MMSE rate into contributions from Fisher Information (FI) and Fisher Information Rate (FIR), showing that while global isometry ensures FI alignment, FIR depends on the encoder’s local geometric properties. They identify three measurable penalties for latent geometric distortion: dimensional compression, tangential distortion, and curvature injection.

Result: The theoretical framework provides conditions for Fisher Information Rate preservation across spaces, ensuring maintained diffusability. Experiments across diverse autoencoding architectures validate the framework and establish FI and FIR metrics as a robust diagnostic suite for identifying and mitigating latent diffusion failure.

Conclusion: The paper provides a formal understanding of why diffusion models degrade in latent spaces, offering measurable geometric distortions that affect diffusability and establishing diagnostic metrics to identify and fix latent diffusion failures.

Abstract: Diffusion models often degrade when trained in latent spaces (e.g., VAEs), yet the formal causes remain poorly understood. We quantify latent-space diffusability through the rate of change of the Minimum Mean Squared Error (MMSE) along the diffusion trajectory. Our framework decomposes this MMSE rate into contributions from Fisher Information (FI) and Fisher Information Rate (FIR). We demonstrate that while global isometry ensures FI alignment, FIR is governed by the encoder’s local geometric properties. Our analysis explicitly decouples latent geometric distortion into three measurable penalties: dimensional compression, tangential distortion, and curvature injection. We derive theoretical conditions for FIR preservation across spaces, ensuring maintained diffusability. Experiments across diverse autoencoding architectures validate our framework and establish these efficient FI and FIR metrics as a robust diagnostic suite for identifying and mitigating latent diffusion failure.

Method: Proposes RLSD (RLVR with Self-Distillation) that uses self-distillation to obtain token-level policy differences for determining update magnitudes, while RLVR provides reliable update directions from environmental feedback like response correctness.

Result: RLSD achieves higher convergence ceiling and superior training stability by harnessing strengths of both RLVR and OPSD, addressing the limitations of pure self-distillation approaches.

Conclusion: The optimal approach combines self-distillation for fine-grained update magnitudes with RLVR for reliable update directions, creating a more stable and effective training paradigm for LLMs.

Abstract: On-policy distillation (OPD) has become a popular training paradigm in the LLM community. This paradigm selects a larger model as the teacher to provide dense, fine-grained signals for each sampled trajectory, in contrast to reinforcement learning with verifiable rewards (RLVR), which only obtains sparse signals from verifiable outcomes in the environment. Recently, the community has explored on-policy self-distillation (OPSD), where the same model serves as both teacher and student, with the teacher receiving additional privileged information such as reference answers to enable self-evolution. This paper demonstrates that learning signals solely derived from the privileged teacher result in severe information leakage and unstable long-term training. Accordingly, we identify the optimal niche for self-distillation and propose \textbf{RLSD} (\textbf{RL}VR with \textbf{S}elf-\textbf{D}istillation). Specifically, we leverage self-distillation to obtain token-level policy differences for determining fine-grained update magnitudes, while continuing to use RLVR to derive reliable update directions from environmental feedback (e.g., response correctness). This enables RLSD to simultaneously harness the strengths of both RLVR and OPSD, achieving a higher convergence ceiling and superior training stability.

Method: Evaluated uncertainty-based APL against random sampling across harmlessness, helpfulness, and instruction-following settings. Used both reward models and LLM-as-a-judge proxies to measure performance. Analyzed win-rates and general capability benchmarks to assess effectiveness.

Result: APL yields negligible improvements in proxy win-rates compared to random sampling. Found dissociation where win-rate improves even as general capability (measured by standard benchmarks) degrades. APL fails to mitigate capability collapse or reduce variance significantly better than random sampling.

Conclusion: In the regime of strong pre-trained priors, the computational overhead of active selection is difficult to justify against the “cheap diversity” provided by simple random samples. Active learning approaches may not be cost-effective for modern LLM alignment.

Abstract: Modern LLMs inherit strong priors from web-scale pretraining, which can limit the headroom of post-training data-selection strategies. While Active Preference Learning (APL) seeks to optimize query efficiency in online Direct Preference Optimization (DPO), the inherent richness of on-policy candidate pools often renders simple Random sampling a surprisingly formidable baseline. We evaluate uncertainty-based APL against Random across harmlessness, helpfulness, and instruction-following settings, utilizing both reward models and LLM-as-a-judge proxies. We find that APL yields negligible improvements in proxy win-rates compared to Random. Crucially, we observe a dissociation where win-rate improves even as general capability – measured by standard benchmarks – degrades. APL fails to mitigate this capability collapse or reduce variance significantly better than random sampling. Our findings suggest that in the regime of strong pre-trained priors, the computational overhead of active selection is difficult to justify against the ``cheap diversity’’ provided by simple random samples. Our code is available at https://github.com/BootsofLagrangian/random-vs-apl.

Method: Proposes Spatio-Temporal Decoupled Differential Equation Network (STDDN) that guides microscopic trajectory prediction with macroscopic physics using: 1) continuity equation from fluid dynamics as physical constraint, 2) Neural ODE to model macroscopic density evolution, 3) density-velocity coupled dynamic graph learning module, 4) differentiable density mapping module, and 5) cross-grid detection module.

Result: Demonstrated significantly superior simulation performance compared to state-of-the-art methods on long-term tasks across four real-world datasets, with major reduction in inference latency.

Conclusion: STDDN effectively bridges microscopic and macroscopic perspectives in crowd simulation, achieving both high accuracy and computational efficiency by incorporating physical constraints through Neural ODEs and differentiable modules.

Abstract: Accurate crowd simulation is crucial for public safety management, emergency evacuation planning, and intelligent transportation systems. However, existing methods, which typically model crowds as a collection of independent individual trajectories, are limited in their ability to capture macroscopic physical laws. This microscopic approach often leads to error accumulation and compromises simulation stability. Furthermore, deep learning-driven methods tend to suffer from low inference efficiency and high computational overhead, making them impractical for large-scale, efficient simulations. To address these challenges, we propose the Spatio-Temporal Decoupled Differential Equation Network (STDDN), a novel framework that guides microscopic trajectory prediction with macroscopic physics. We innovatively introduce the continuity equation from fluid dynamics as a strong physical constraint. A Neural Ordinary Differential Equation (Neural ODE) is employed to model the macroscopic density evolution driven by individual movements, thereby physically regularizing the microscopic trajectory prediction model. We design a density-velocity coupled dynamic graph learning module to formulate the derivative of the density field within the Neural ODE, effectively mitigating error accumulation. We also propose a differentiable density mapping module to eliminate discontinuous gradients caused by discretization and introduce a cross-grid detection module to accurately model the impact of individual cross-grid movements on local density changes. The proposed STDDN method has demonstrated significantly superior simulation performance compared to state-of-the-art methods on long-term tasks across four real-world datasets, as well as a major reduction in inference latency.

Method: Proposes a model-agnostic framework with: 1) Class-wise mean (CWM) objective replacing sample frequency weighted loss with uniform class-conditional supervision; 2) Method-wise adjustments including constrained distillation, normalized contrastive/knowledge transfer losses; 3) Dynamic Intervention Weight Alignment (DIWA) to prevent over-adjustment from unstable statistics.

Result: Experiments show clear performance degradation across various CIL baselines under free-flow settings, while the proposed strategies yield consistent performance gains.

Conclusion: The paper addresses the practical challenge of free-flow class-incremental learning and provides robust solutions that work across different CIL paradigms, making CIL systems more realistic for real-world applications.

Abstract: Class-incremental learning (CIL) is typically evaluated under predefined schedules with equal-sized tasks, leaving more realistic and complex cases unexplored. However, a practical CIL system should learns immediately when any number of new classes arrive, without forcing fixed-size tasks. We formalize this setting as Free-Flow Class-Incremental Learning (FFCIL), where data arrives as a more realistic stream with a highly variable number of unseen classes each step. It will make many existing CIL methods brittle and lead to clear performance degradation. We propose a model-agnostic framework for robust CIL learning under free-flow arrivals. It comprises a class-wise mean (CWM) objective that replaces sample frequency weighted loss with uniformly aggregated class-conditional supervision, thereby stabilizing the learning signal across free-flow class increments, as well as method-wise adjustments that improve robustness for representative CIL paradigms. Specifically, we constrain distillation to replayed data, normalize the scale of contrastive and knowledge transfer losses, and introduce Dynamic Intervention Weight Alignment (DIWA) to prevent over-adjustment caused by unstable statistics from small class increments. Experiments confirm a clear performance degradation across various CIL baselines under FFCIL, while our strategies yield consistent gains.

Method: Proposes a solver-compatible dimensionality reduction framework that identifies a sparse subset of structurally stable commitment binaries to fix prior to optimization. One implementation uses an LLM to select these variables, but the LLM doesn’t replace optimization - it provides partial variable restriction while the original MILP solver handles all constraints and remaining decisions. The framework preserves feasibility by defining a reduced feasible region of the original UC model.

Result: Experiments on IEEE 57-bus, RTS 73-bus, IEEE 118-bus, and augmented large-scale cases (including security-constrained variants) show consistent reductions in branch-and-bound nodes and solution time, achieving order-of-magnitude speedups on high-complexity instances while maintaining near-optimal objective values.

Conclusion: The proposed framework effectively reduces computational burden in unit commitment problems by using LLMs for intelligent variable selection while maintaining solver-certified optimality within the restricted space, offering a practical approach to accelerate optimization without sacrificing reliability.

Abstract: The growing number of individual generating units, hybrid resources, and security constraints has significantly increased the computational burden of network-constrained unit commitment (UC), where most solution time is spent exploring branch-and-bound trees over unit-hour binary variables. To reduce this combinatorial burden, recent approaches have explored learning-based guidance to assist commitment decisions. However, directly using tools such as large language models (LLMs) to predict full commitment schedules is unreliable, as infeasible or inconsistent binary decisions can violate inter-temporal constraints and degrade economic optimality. This paper proposes a solver-compatible dimensionality reduction framework for UC that exploits structural regularities in commitment decisions. Instead of generating complete schedules, the framework identifies a sparse subset of structurally stable commitment binaries to fix prior to optimization. One implementation uses an LLM to select these variables. The LLM does not replace the optimization process but provides partial variable restriction, while all constraints and remaining decisions are handled by the original MILP solver, which continues to enforce network, ramping, reserve, and security constraints. We formally show that the masked problem defines a reduced feasible region of the original UC model, thereby preserving feasibility and enabling solver-certified optimality within the restricted space. Experiments on IEEE 57-bus, RTS 73-bus, IEEE 118-bus, and augmented large-scale cases, including security-constrained variants, demonstrate consistent reductions in branch-and-bound nodes and solution time, achieving order-of-magnitude speedups on high-complexity instances while maintaining near-optimal objective values.

Method: Fine-tunes a sentence encoding model using sparse LLM-provided labels on corpus samples, then segments the embedding space with thresholded clustering to separate closely related topics within narrow domains using a student-teacher distillation pipeline.

Result: PRISM improves topic separability over state-of-the-art local topic models and even outperforms clustering on large frontier embedding models, while requiring only a small number of LLM queries for training across multiple corpora.

Conclusion: PRISM provides an effective framework for web-scale text analysis that enables tracking nuanced claims and subtopics online with an interpretable, locally deployable system, contributing to LLM distillation, sampling strategies, and practical text analysis.

Abstract: In this paper, we propose Precision-Informed Semantic Modeling (PRISM), a structured topic modeling framework combining the benefits of rich representations captured by LLMs with the low cost and interpretability of latent semantic clustering methods. PRISM fine-tunes a sentence encoding model using a sparse set of LLM- provided labels on samples drawn from some corpus of interest. We segment this embedding space with thresholded clustering, yielding clusters that separate closely related topics within some narrow domain. Across multiple corpora, PRISM improves topic separability over state-of-the-art local topic models and even over clustering on large, frontier embedding models while requiring only a small number of LLM queries to train. This work contributes to several research streams by providing (i) a student-teacher pipeline to distill sparse LLM supervision into a lightweight model for topic discovery; (ii) an analysis of the efficacy of sampling strategies to improve local geometry for cluster separability; and (iii) an effective approach for web-scale text analysis, enabling researchers and practitioners to track nuanced claims and subtopics online with an interpretable, locally deployable framework.

Method: Developed a graph-neural surrogate model using projected meshes and multimesh connectivity on a high-resolution unstructured finite-element mesh. Used synthetic flood event database from Telemac2D simulations, studied effects of explicit discharge feature Q(t) and pushforward training for autoregressive rollouts.

Result: Surrogate produces 6-hour predictions in 0.4s on single GPU vs 180min on 56 CPU cores for reference simulation (27,000x speedup). Best results with Q(t) conditioning, multimesh connectivity, and pushforward training. Maintains accuracy on hydraulic variables and inundation maps.

Conclusion: Graph-based surrogates are practical complements to industrial hydraulic solvers for operational flood mapping, offering dramatic speed improvements while preserving accuracy for rapid decision support.

Abstract: Operational flood forecasting still relies on high-fidelity two-dimensional hydraulic solvers, but their runtime can be prohibitive for rapid decision support on large urban floodplains. In parallel, AI-based surrogate models have shown strong potential in several areas of computational physics for accelerating otherwise expensive high-fidelity simulations. We address this issue on the lower Têt River (France), starting from a production-grade Telemac2D model defined on a high-resolution unstructured finite-element mesh with more than $4\times 10^5$ nodes. From this setup, we build a learning-ready database of synthetic but operationally grounded flood events covering several representative hydrograph families and peak discharges. On top of this database, we develop a graph-neural surrogate based on projected meshes and multimesh connectivity. The projected-mesh strategy keeps training tractable while preserving high-fidelity supervision from the original Telemac simulations, and the multimesh construction enlarges the effective spatial receptive field without increasing network depth. We further study the effect of an explicit discharge feature $Q(t)$ and of pushforward training for long autoregressive rollouts. The experiments show that conditioning on $Q(t)$ is essential in this boundary-driven setting, that multimesh connectivity brings additional gains once the model is properly conditioned, and that pushforward further improves rollout stability. Among the tested configurations, the combination of $Q(t)$, multimesh connectivity, and pushforward provides the best overall results. These gains are observed both on hydraulic variables over the surrogate mesh and on inundation maps interpolated onto a common $25,\mathrm{m}$ regular grid and compared against the original high-resolution Telemac solution. On the studied case, the learned surrogate produces 6-hour predictions in about $0.4,\mathrm{s}$ on a single NVIDIA A100 GPU, compared with about $180,\mathrm{min}$ on 56 CPU cores for the reference simulation. These results support graph-based surrogates as practical complements to industrial hydraulic solvers for operational flood mapping.

Method: ExSTraQt uses quasi-temporal graph representations of financial transactions and applies supervised learning to detect suspicious transactions. The framework emphasizes simplicity in design with few parameters, and scalability in computing and memory requirements.

Result: The framework achieves performance improvements over state-of-the-art AML models, with up to 1% F1 score uplift on real datasets and over 8% on synthetic datasets. It demonstrates scalability and could complement existing bank AML systems.

Conclusion: ExSTraQt provides an effective, scalable supervised learning approach for money laundering detection that can enhance existing financial institution systems while maintaining simplicity and computational efficiency.

Abstract: Money laundering presents a persistent challenge for financial institutions worldwide, while criminal organizations constantly evolve their tactics to bypass detection systems. Traditional anti-money laundering approaches mainly rely on predefined risk-based rules, leading to resource-intensive investigations and high numbers of false positive alerts. In order to restrict operational costs from exploding, while billions of transactions are being processed every day, financial institutions are investing in more sophisticated mechanisms to improve existing systems. In this paper, we present ExSTraQt (EXtract Suspicious TRAnsactions from Quasi-Temporal graph representation), an advanced supervised learning approach to detect money laundering (or suspicious) transactions in financial datasets. Our proposed framework excels in performance, when compared to the state-of-the-art AML (Anti Money Laundering) detection models. The key strengths of our framework are sheer simplicity, in terms of design and number of parameters; and scalability, in terms of the computing and memory requirements. We evaluated our framework on transaction-level detection accuracy using a real dataset; and a set of synthetic financial transaction datasets. We consistently achieve an uplift in the F1 score for most datasets, up to 1% for the real dataset; and more than 8% for one of the synthetic datasets. We also claim that our framework could seamlessly complement existing AML detection systems in banks. Our code and datasets are available at https://github.com/mhaseebtariq/exstraqt.

Method: The authors study the Exponential Weights algorithm with isotropic Gaussian prior for online logistic regression. They analyze computational complexity improvements and examine the large-B regime under linear separability, showing convergence of the posterior distribution to a truncated Gaussian and relating the predictor to geometric concepts like version cones and hard-margin SVM directions.

Result: The paper achieves a substantial improvement in computational complexity from O(B¹⁸n³⁷) to O(B³n⁵) while maintaining the near-optimal regret bound O(d log(Bn)). In the large-B regime under linear separability, they show the posterior converges to a truncated Gaussian, the predictor converges to a solid-angle vote, and derive non-asymptotic regret bounds that become independent of B and grow only logarithmically with the inverse margin.

Conclusion: The Exponential Weights algorithm can be both computationally tractable and geometrically adaptive in online classification, with improved efficiency and interesting convergence properties in separable settings.

Abstract: This paper studies the Exponential Weights (EW) algorithm with an isotropic Gaussian prior for online logistic regression. We show that the near-optimal worst-case regret bound $O(d\log(Bn))$ for EW, established by Kakade and Ng (2005) against the best linear predictor of norm at most $B$, can be achieved with total worst-case computational complexity $O(B^3 n^5)$. This substantially improves on the $O(B^{18}n^{37})$ complexity of prior work achieving the same guarantee (Foster et al., 2018). Beyond efficiency, we analyze the large-$B$ regime under linear separability: after rescaling by $B$, the EW posterior converges as $B\to\infty$ to a standard Gaussian truncated to the version cone. Accordingly, the predictor converges to a solid-angle vote over separating directions and, on every fixed-margin slice of this cone, the mode of the corresponding truncated Gaussian is aligned with the hard-margin SVM direction. Using this geometry, we derive non-asymptotic regret bounds showing that once $B$ exceeds a margin-dependent threshold, the regret becomes independent of $B$ and grows only logarithmically with the inverse margin. Overall, our results show that EW can be both computationally tractable and geometrically adaptive in online classification.

Method: Proposes LOGGIA, a graph neural routing algorithm that predicts log-space link weights from attributed topology-and-telemetry graphs. Uses data-driven pre-training followed by on-policy reinforcement learning within a delay-aware closed-loop control framework.

Result: LOGGIA consistently outperforms shortest-path baselines across synthetic and real network topologies with unseen mixed TCP/UDP traffic sequences. Neural baselines fail once realistic delays are enforced. Neural routing performs best when deployed fully locally at each router.

Conclusion: Telemetry-aware routing should be treated as a delay-aware closed-loop control problem. LOGGIA demonstrates the effectiveness of neural routing with explicit delay modeling, with local deployment at each router being optimal for real-world scenarios.

Abstract: Routing algorithms are crucial for efficient computer network operations, and in many settings they must be able to react to traffic bursts within milliseconds. Live telemetry data can provide informative signals to routing algorithms, and recent work has trained neural networks to exploit such signals for traffic-aware routing. Yet, aggregating network-wide information is subject to communication delays, and existing neural approaches either assume unrealistic delay-free global states, or restrict routers to purely local telemetry. This leaves their deployability in real-world environments unclear. We cast telemetry-aware routing as a delay-aware closed-loop control problem and introduce a framework that trains and evaluates neural routing algorithms, while explicitly modeling communication and inference delays. On top of this framework, we propose LOGGIA, a scalable graph neural routing algorithm that predicts log-space link weights from attributed topology-and-telemetry graphs. It utilizes a data-driven pre-training stage, followed by on-policy Reinforcement Learning. Across synthetic and real network topologies, and unseen mixed TCP/UDP traffic sequences, LOGGIA consistently outperforms shortest-path baselines, whereas neural baselines fail once realistic delays are enforced. Our experiments further suggest that neural routing algorithms like LOGGIA perform best when deployed fully locally, i.e., observing network states and inferring actions at every router individually, as opposed to centralized decision making.

Method: Analyzed NASA OSDR dataset OSD-970 from Rodent Research-1 mission using RT-qPCR data from 89 adipogenesis/thermogenesis genes in female mice. Applied differential expression analysis, multiple ML classifiers with Leave-One-Out Cross-Validation, and Explainable AI via SHAP analysis.

Result: Found dramatic 12.21-fold upregulation of Ucp1 in microgravity-exposed WAT, significant thermogenesis pathway activation (mean fold-change = 3.24). Random Forest model achieved AUC = 0.922, Accuracy = 0.812, F1 = 0.824. SHAP analysis identified Ucp1 as top predictive feature, with clear PCA separation between flight and ground samples.

Conclusion: Microgravity induces rapid thermogenic reprogramming in female WAT as compensatory response. Study demonstrates power of explainable AI for space biology data analysis, with implications for female astronaut health and Earth-based metabolic disease research.

Abstract: Microgravity induces profound metabolic adaptations in mammalian physiology, yet the molecular mechanisms governing thermogenesis in female white adipose tissue (WAT) remain poorly characterized. This paper presents the first machine learning (ML) analysis of NASA Open Science Data Repository (OSDR) dataset OSD-970, derived from the Rodent Research-1 (RR-1) mission. Using RT-qPCR data from 89 adipogenesis and thermogenesis pathway genes in gonadal WAT of 16 female C57BL/6J mice (8 flight, 8 ground control) following 37 days aboard the International Space Station (ISS), we applied differential expression analysis, multiple ML classifiers with Leave-One-Out Cross-Validation (LOO-CV), and Explainable AI via SHapley Additive exPlanations (SHAP). The most striking finding is a dramatic 12.21-fold upregulation of Ucp1 (Delta-Delta-Ct = -3.61, p = 0.0167) in microgravity-exposed WAT, accompanied by significant activation of the thermogenesis pathway (mean pathway fold-change = 3.24). The best-performing model (Random Forest with top-20 features) achieved AUC = 0.922, Accuracy = 0.812, and F1 = 0.824 via LOO-CV. SHAP analysis consistently ranked Ucp1 among the top predictive features, while Angpt2, Irs2, Jun, and Klf-family transcription factors emerged as dominant consensus classifiers. Principal component analysis (PCA) revealed clear separation between flight and ground samples, with PC1 explaining 69.1% of variance. These results suggest rapid thermogenic reprogramming in female WAT as a compensatory response to microgravity. This study demonstrates the power of explainable AI for re-analysis of newly released NASA space biology datasets, with direct implications for female astronaut health on long-duration missions and for Earth-based obesity and metabolic disease research.

Method: FedSQ uses a DualCopy approach: freezes a structural copy of the pretrained model to maintain fixed binary gating masks during federated fine-tuning, while only a quantitative copy is optimized locally and aggregated across rounds. This reduces learning to within-regime affine refinements.

Result: Experiments on CNN backbones under i.i.d. and Dirichlet splits show FedSQ improves robustness, reduces rounds-to-best validation performance relative to baselines, while preserving accuracy in transfer settings.

Conclusion: FedSQ stabilizes federated learning aggregation under heterogeneous data partitions by separating structural and quantitative knowledge, making it more robust for transfer-initialized FL scenarios.

Abstract: Federated learning (FL) enables collaborative training across organizations without sharing raw data, but it is hindered by statistical heterogeneity (non-i.i.d.\ client data) and by instability of naive weight averaging under client drift. In many cross-silo deployments, FL is warm-started from a strong pretrained backbone (e.g., ImageNet-1K) and then adapted to local domains. Motivated by recent evidence that ReLU-like gating regimes (structural knowledge) stabilize earlier than the remaining parameter values (quantitative knowledge), we propose FedSQ (Federated Structural-Quantitative learning), a transfer-initialized neural federated procedure based on a DualCopy, piecewise-linear view of deep networks. FedSQ freezes a structural copy of the pretrained model to induce fixed binary gating masks during federated fine-tuning, while only a quantitative copy is optimized locally and aggregated across rounds. Fixing the gating reduces learning to within-regime affine refinements, which stabilizes aggregation under heterogeneous partitions. Experiments on two convolutional neural network backbones under i.i.d.\ and Dirichlet splits show that FedSQ improves robustness and can reduce rounds-to-best validation performance relative to standard baselines while preserving accuracy in the transfer setting.

Method: The method involves defining a plug-in estimator for generating diffusion-based samples from tilted distributions, analyzing its properties through Wasserstein bounds between the estimator’s distribution and the true distribution, and proving TV-accuracy of running diffusion on tilted samples under certain assumptions.

Result: The plug-in estimator is shown to be minimax-optimal, with theoretical bounds on Wasserstein distance as a function of sample size and tilt parameter, and TV-accuracy guarantees for diffusion on tilted samples. Results are supported by extensive simulations.

Conclusion: The paper provides a theoretically sound framework for generating samples from tilted distributions using diffusion-based methods with provable guarantees, applicable to various practical domains requiring distribution tilting.

Abstract: Given $n$ independent samples from a $d$-dimensional probability distribution, our aim is to generate diffusion-based samples from a distribution obtained by tilting the original, where the degree of tilt is parametrized by $θ\in \mathbb{R}^d$. We define a plug-in estimator and show that it is minimax-optimal. We develop Wasserstein bounds between the distribution of the plug-in estimator and the true distribution as a function of $n$ and $θ$, illustrating regimes where the output and the desired true distribution are close. Further, under some assumptions, we prove the TV-accuracy of running Diffusion on these tilted samples. Our theoretical results are supported by extensive simulations. Applications of our work include finance, weather and climate modelling, and many other domains, where the aim may be to generate samples from a tilted distribution that satisfies practically motivated moment constraints.

Method: HyperFitS uses a hypernetwork architecture for spectral fitting that flexibly adapts to a broad range of baseline corrections and water suppression factors without retraining. It was tested on human subject data acquired at 3T and 7T with various resolutions using both water-suppressed and water-unsuppressed MRSI.

Result: Metabolic maps show substantial agreement with conventional LCModel fitting (gold standard), with significantly faster fitting times. Quantitative analysis reveals baseline parametrization can alter metabolic quantification results by up to 30%.

Conclusion: HyperFitS achieves strong agreement with state-of-the-art conventional methods while reducing processing from hours to seconds. It offers superior configurability compared to prior deep learning methods and adapts to multiple protocols and field strengths without retraining.

Abstract: Purpose: Proton magnetic resonance spectroscopic imaging ($^1$H MRSI) enables the mapping of whole-brain metabolites concentrations in-vivo. However, a long-standing problem for its clinical applicability is the metabolic quantification, which can require extensive time for spectral fitting. Recently, deep learning methods have been able to provide whole-brain metabolic quantification in only a few seconds. However, neural network implementations often lack configurability and require retraining to change predefined parameter settings. Methods: We introduce HyperFitS, a hypernetwork for spectral fitting for metabolite quantification in whole-brain $^1$H MRSI that flexibly adapts to a broad range of baseline corrections and water suppression factors. Metabolite maps of human subjects acquired at 3T and 7T with isotropic resolutions of 10 mm, 3.4 mm and 2 mm by water-suppressed and water-unsuppressed MRSI were quantified with HyperFitS and compared to conventional LCModel fitting. Results: Metabolic maps show a substantial agreement between the new and gold-standard methods, with significantly faster fitting times by HyperFitS. Quantitative results further highlight the impact of baseline parametrization on metabolic quantification, which can alter results by up to 30%. Conclusion: HyperFitS shows strong agreement with state-of-the-art conventional methods, while reducing processing times from hours to a few seconds. Compared to prior deep learning based spectral fitting methods, HyperFitS enables a wide range of configurability and can adapt to data quality acquired with multiple protocols and field strengths without retraining.

Method: Proposes Dual-Aligned Structural Basis Distillation (DSBD) that constructs differentiable structural basis using continuous probabilistic prototype graphs. Uses gradient-based optimization over graph topology with source-domain supervision for semantic discriminability and dual-alignment to target domain via: 1) geometric consistency through permutation-invariant topological moment matching, and 2) spectral consistency through Dirichlet energy calibration. Also introduces decoupled inference paradigm to mitigate source-specific structural bias.

Result: Extensive experiments on graph and image benchmarks demonstrate that DSBD consistently outperforms state-of-the-art methods.

Conclusion: DSBD effectively addresses structural discrepancies in graph domain adaptation through explicit modeling and adaptation of cross-domain structural variation, achieving superior performance over existing methods.

Abstract: Graph domain adaptation (GDA) aims to transfer knowledge from a labeled source graph to an unlabeled target graph under distribution shifts. However, existing methods are largely feature-centric and overlook structural discrepancies, which become particularly detrimental under significant topology shifts. Such discrepancies alter both geometric relationships and spectral properties, leading to unreliable transfer of graph neural networks (GNNs). To address this limitation, we propose Dual-Aligned Structural Basis Distillation (DSBD) for GDA, a novel framework that explicitly models and adapts cross-domain structural variation. DSBD constructs a differentiable structural basis by synthesizing continuous probabilistic prototype graphs, enabling gradient-based optimization over graph topology. The basis is learned under source-domain supervision to preserve semantic discriminability, while being explicitly aligned to the target domain through a dual-alignment objective. Specifically, geometric consistency is enforced via permutation-invariant topological moment matching, and spectral consistency is achieved through Dirichlet energy calibration, jointly capturing structural characteristics across domains. Furthermore, we introduce a decoupled inference paradigm that mitigates source-specific structural bias by training a new GNN on the distilled structural basis. Extensive experiments on graph and image benchmarks demonstrate that DSBD consistently outperforms state-of-the-art methods.

Method: Proposed Hallucination-as-Cue Framework with hallucination-inductive, modality-specific corruptions that remove/replace essential information, forcing models to reason by hallucination. Applied corruptions during both training and evaluation across multiple multimodal reasoning benchmarks.

Result: RL post-training under purely hallucination-inductive settings can still significantly improve reasoning performance, sometimes even outperforming standard training. Shows model hallucination plays more significant role in RL-training than previously recognized.

Conclusion: Findings challenge assumptions about MLLM reasoning training and motivate development of more modality-aware RL-based training designs that truly leverage visual information rather than exploiting hallucination patterns.

Abstract: The recent success of reinforcement learning (RL) in large reasoning models has inspired the growing adoption of RL for post-training Multimodal Large Language Models (MLLMs) to enhance their visual reasoning capabilities. Although many studies have reported improved performance, it remains unclear whether RL training truly enables models to learn from visual information. In this work, we propose the Hallucination-as-Cue Framework, an analytical framework designed to investigate the effects of RL-based post-training on multimodal reasoning models from the perspective of model hallucination. Specifically, we introduce hallucination-inductive, modality-specific corruptions that remove or replace essential information required to derive correct answers, thereby forcing the model to reason by hallucination. By applying these corruptions during both training and evaluation, our framework provides a unique perspective for diagnosing RL training dynamics and understanding the intrinsic properties of datasets. Through extensive experiments and analyses across multiple multimodal reasoning benchmarks, we reveal that the role of model hallucination for RL-training is more significant than previously recognized. For instance, we find that RL post-training under purely hallucination-inductive settings can still significantly improve models’ reasoning performance, and in some cases even outperform standard training. These findings challenge prevailing assumptions about MLLM reasoning training and motivate the development of more modality-aware RL-based training designs.

Method: Reflective Context Learning (RCL) framework with reflection (converts trajectories to update signals) and mutation (applies signals to improve behavior). Extends context optimization with classical ML primitives: batching, improved credit assignment, auxiliary losses, failure replay, and grouped rollouts for variance reduction.

Result: RCL improves over strong baselines on AppWorld, BrowseComp+, and RewardBench2. Analysis shows robustness to initialization, effects of batch size, sampling strategies, curriculum, optimizer-state variants, and model allocation tradeoffs.

Conclusion: Context learning should be treated as an optimization problem with systematic study and transferable principles, not isolated algorithms.

Abstract: Generally capable agents must learn from experience in ways that generalize across tasks and environments. The fundamental problems of learning, including credit assignment, overfitting, forgetting, local optima, and high-variance learning signals, persist whether the learned object lies in parameter space or context space. While these challenges are well understood in classical machine learning optimization, they remain underexplored in context space, leading current methods to be fragmented and ad hoc. We present Reflective Context Learning (RCL), a unified framework for agents that learn through repeated interaction, reflection on behavior and failure modes, and iterative updates to context. In RCL, reflection converts trajectories and current context into a directional update signal analogous to gradients, while mutation applies that signal to improve future behavior in context space. We recast recent context-optimization approaches as instances of this shared learning problem and systematically extend them with classical optimization primitives, including batching, improved credit-assignment signal, auxiliary losses, failure replay, and grouped rollouts for variance reduction. On AppWorld, BrowseComp+, and RewardBench2, these primitives improve over strong baselines, with their relative importance shifting across task regimes. We further analyze robustness to initialization, the effects of batch size, sampling and curriculum strategy, optimizer-state variants, and the impact of allocating stronger or weaker models to different optimization components. Our results suggest that learning through context updates should be treated not as a set of isolated algorithms, but as an optimization problem whose mechanisms can be studied systematically and improved through transferable principles.

Method: Introduces gradient-boosted attention with two attention passes: the first makes initial predictions, the second attends to prediction errors from the first and applies a gated correction. Uses separate learned projections for each pass and per-dimension gating.

Result: On a 10M-token WikiText-103 subset, achieves test perplexity of 67.9 vs 72.2 for standard attention, 69.6 for Twicing Attention, and 69.0 for parameter-matched wider baseline. Two rounds capture most benefits.

Conclusion: Gradient-boosted attention improves transformer performance by applying boosting principles within attention layers, enabling error correction and better information recovery than standard approaches.

Abstract: Transformer attention computes a single softmax-weighted average over values – a one-pass estimate that cannot correct its own errors. We introduce \emph{gradient-boosted attention}, which applies the principle of gradient boosting \emph{within} a single attention layer: a second attention pass, with its own learned projections, attends to the prediction error of the first and applies a gated correction. Under a squared reconstruction objective, the construction maps onto Friedman’s gradient boosting machine, with each attention pass as a base learner and the per-dimension gate as the shrinkage parameter. We show that a single Hopfield-style update erases all query information orthogonal to the stored-pattern subspace, and that further iteration under local contraction can collapse distinct queries in the same region to the same fixed point. We also show that separate projections for the correction pass can recover residual information inaccessible to the shared-projection approach of Tukey’s twicing. On a 10M-token subset of WikiText-103, gradient-boosted attention achieves a test perplexity of $67.9$ compared to $72.2$ for standard attention, $69.6$ for Twicing Attention, and $69.0$ for a parameter-matched wider baseline, with two rounds capturing most of the benefit.

Method: Generate parametric virtual cohort based on clinical data correlations, train deep neural surrogate model for instantaneous hemodynamic prediction, enable rapid screening of input parameters, and provide principled sampling of terminal resistance.

Result: The framework reduces cost of targeted synthetic dataset generation, minimizes uncertainties of unmeasurable parameters, determines theoretical information needed for inverse problem solving, and successfully estimates central aortic hemodynamics on clinical data.

Conclusion: Machine learning surrogate models offer efficient and principled approach for cardiovascular modeling at scale, enabling physiological virtual cohort generation and clinical hemodynamic parameter estimation.

Abstract: Cardiovascular modeling has rapidly advanced over the past few decades due to the rising needs for health tracking and early detection of cardiovascular diseases. While 1-D arterial models offer an attractive compromise between computational efficiency and solution fidelity, their application on large populations or for generating large \emph{in silico} cohorts remains challenging. Certain hemodynamic parameters like the terminal resistance/compliance, are difficult to clinically estimate and often yield non-physiological hemodynamics when sampled naively, resulting in large portions of simulated datasets to be discarded. In this work, we present a systematic framework for training machine learning (ML) models, capable of instantaneous hemodynamic prediction and parameter estimation. We initially start with generating a parametric virtual cohort of patients which is based on the multivariate correlations observed in the large Asklepios clinical dataset, ensuring that physiological parameter distributions are respected. We then train a deep neural surrogate model, able to predict patient-specific arterial pressure and cardiac output (CO), enabling rapid a~priori screening of input parameters. This allows for immediate rejection of non-physiological combinations and drastically reduces the cost of targeted synthetic dataset generation (e.g. hypertensive groups). The model also provides a principled means of sampling the terminal resistance to minimize the uncertainties of unmeasurable parameters. Moreover, by assessing the model’s predictive performance we determine the theoretical information which suffices for solving the inverse problem of estimating the CO. Finally, we apply the surrogate on a clinical dataset for the estimation of central aortic hemodynamics i.e. the CO and aortic systolic blood pressure (cSBP).

Method: Proposes learning latent world models at multiple temporal scales and performing hierarchical planning across these scales, creating a modular planning abstraction that works with diverse world-model architectures and domains.

Result: Achieves 70% success rate on real-world pick-and-place tasks with only final goal specification (vs 0% for single-level models), and shows higher success rates with up to 4x less planning-time compute in simulated push manipulation and maze navigation tasks.

Conclusion: Hierarchical multi-scale planning significantly improves long-horizon reasoning capabilities of learned world models while reducing computational complexity, enabling practical zero-shot control in complex real-world robotic applications.

Abstract: Model predictive control (MPC) with learned world models has emerged as a promising paradigm for embodied control, particularly for its ability to generalize zero-shot when deployed in new environments. However, learned world models often struggle with long-horizon control due to the accumulation of prediction errors and the exponentially growing search space. In this work, we address these challenges by learning latent world models at multiple temporal scales and performing hierarchical planning across these scales, enabling long-horizon reasoning while substantially reducing inference-time planning complexity. Our approach serves as a modular planning abstraction that applies across diverse latent world-model architectures and domains. We demonstrate that this hierarchical approach enables zero-shot control on real-world non-greedy robotic tasks, achieving a 70% success rate on pick-&-place using only a final goal specification, compared to 0% for a single-level world model. In addition, across physics-based simulated environments including push manipulation and maze navigation, hierarchical planning achieves higher success while requiring up to 4x less planning-time compute.

Method: Proposes a heuristic algorithm combining server learning with client update filtering and geometric median aggregation. The server uses a small dataset (possibly synthetic) to learn and filter malicious client updates.

Result: Experimental results show significant improvement in model accuracy even with high fractions of malicious clients (>50%) and when server dataset is small/synthetic with distribution not necessarily close to clients’ aggregated data.

Conclusion: Server learning combined with geometric median aggregation and client update filtering effectively enhances federated learning robustness against malicious attacks in challenging non-IID scenarios.

Abstract: This paper explores the use of server learning for enhancing the robustness of federated learning against malicious attacks even when clients’ training data are not independent and identically distributed. We propose a heuristic algorithm that uses server learning and client update filtering in combination with geometric median aggregation. We demonstrate via experiments that this approach can achieve significant improvement in model accuracy even when the fraction of malicious clients is high, even more than $50%$ in some cases, and the dataset utilized by the server is small and could be synthetic with its distribution not necessarily close to that of the clients’ aggregated data.

Method: CeRA (Capacity-enhanced Rank Adaptation) uses weight-level parallel adapters with SiLU gating and dropout to induce non-linear capacity expansion, enabling better representation learning.

Result: On complex MATH dataset, CeRA at rank 64 outperforms LoRA at rank 512 and DoRA at rank 64, achieving higher exact-match accuracy with only 1/8 parameters. Spectral analysis shows CeRA activates lower-variance tail of singular value spectrum.

Conclusion: CeRA overcomes linear limitations of LoRA through non-linear gating, providing better parameter efficiency and representation capacity for complex reasoning tasks.

Abstract: Low-Rank Adaptation (LoRA) dominates parameter-efficient fine-tuning (PEFT). However, it faces a ``linear ceiling’’: increasing the rank yields diminishing returns in expressive capacity due to intrinsic linear constraints. We introduce CeRA (Capacity-enhanced Rank Adaptation), a weight-level parallel adapter that injects SiLU gating and dropout to induce non-linear capacity expansion. We demonstrate a fundamental relationship between adapter expressivity and task complexity. In basic arithmetic (GSM8K), CeRA matches standard linear baselines, but on the complex MATH dataset, it demonstrates high parameter efficiency in downstream reasoning (Exact Match). CeRA at rank 64 (pass@1 16.36%) outperforms both a high-rank LoRA at rank 512 (15.72%) and the state-of-the-art linear variant, DoRA, at rank 64 (14.44%), achieving higher exact-match accuracy with only 1/8 of the parameter budget. Empirical spectral analysis shows that CeRA activates the lower-variance tail of the singular value spectrum, preventing the rank collapse observed in linear methods and providing the representation capacity required for complex logical reasoning.

Method: Proposes a breadth-first search (BFS) approach that uses only a linear number of queries instead of quadratic. The framework can incorporate observational data when available to improve performance.

Result: Achieves state-of-the-art results on real-world causal graphs of varying sizes. The method is more time and data-efficient than previous approaches.

Conclusion: The proposed framework demonstrates effectiveness and efficiency in discovering causal relationships, showing potential for broad applicability in causal graph discovery tasks across different domains.

Abstract: We propose a novel framework that leverages LLMs for full causal graph discovery. While previous LLM-based methods have used a pairwise query approach, this requires a quadratic number of queries which quickly becomes impractical for larger causal graphs. In contrast, the proposed framework uses a breadth-first search (BFS) approach which allows it to use only a linear number of queries. We also show that the proposed method can easily incorporate observational data when available, to improve performance. In addition to being more time and data-efficient, the proposed framework achieves state-of-the-art results on real-world causal graphs of varying sizes. The results demonstrate the effectiveness and efficiency of the proposed method in discovering causal relationships, showcasing its potential for broad applicability in causal graph discovery tasks across different domains.

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 2405.15314: Page request resulted in HTTP 429 (https://export.arxiv.org/api/query?search_query=&id_list=2405.15314&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 2406.03099: Page request resulted in HTTP 429 (https://export.arxiv.org/api/query?search_query=&id_list=2406.03099&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 2410.06128: Page request resulted in HTTP 429 (https://export.arxiv.org/api/query?search_query=&id_list=2410.06128&sortBy=relevance&sortOrder=descending&start=0&max_results=100)

Method: Cannot determine method due to missing paper content

Result: Cannot determine results due to missing paper content

Conclusion: Cannot draw conclusions due to missing paper content

Abstract: Failed to fetch summary for 2502.02970: Page request resulted in HTTP 429 (https://export.arxiv.org/api/query?search_query=&id_list=2502.02970&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 2504.01938: Page request resulted in HTTP 429 (https://export.arxiv.org/api/query?search_query=&id_list=2504.01938&sortBy=relevance&sortOrder=descending&start=0&max_results=100)

Method: Relaxes discrete automaton transitions via soft labeling of states, yielding differentiable rewards and state representations that mitigate LTL’s sparsity issue while preserving objective soundness. Provides theoretical guarantees connecting Büchi acceptance to both discrete and differentiable LTL returns with tunable bound on discrepancy.

Result: Substantially accelerates training and achieves up to twice the returns of discrete baselines across complex, nonlinear, contact-rich continuous-control tasks. Demonstrates compatibility with reward machines, covering co-safe LTL and LTL_f without modification.

Conclusion: Bridges formal methods and deep RL by rendering automaton-based rewards differentiable, enabling safe, specification-driven learning in continuous domains.

Abstract: Ensuring that reinforcement learning (RL) controllers satisfy safety and reliability constraints in real-world settings remains challenging: state-avoidance and constrained Markov decision processes often fail to capture trajectory-level requirements or induce overly conservative behavior. Formal specification languages such as linear temporal logic (LTL) offer correct-by-construction objectives, yet their rewards are typically sparse, and heuristic shaping can undermine correctness. We introduce, to our knowledge, the first end-to-end framework that integrates LTL with differentiable simulators, enabling efficient gradient-based learning directly from formal specifications. Our method relaxes discrete automaton transitions via soft labeling of states, yielding differentiable rewards and state representations that mitigate the sparsity issue intrinsic to LTL while preserving objective soundness. We provide theoretical guarantees connecting Büchi acceptance to both discrete and differentiable LTL returns and derive a tunable bound on their discrepancy in deterministic and stochastic settings. Empirically, across complex, nonlinear, contact-rich continuous-control tasks, our approach substantially accelerates training and achieves up to twice the returns of discrete baselines. We further demonstrate compatibility with reward machines, thereby covering co-safe LTL and LTL$_\text{f}$ without modification. By rendering automaton-based rewards differentiable, our work bridges formal methods and deep RL, enabling safe, specification-driven learning in continuous domains.

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 2507.21437: Page request resulted in HTTP 429 (https://export.arxiv.org/api/query?search_query=&id_list=2507.21437&sortBy=relevance&sortOrder=descending&start=0&max_results=100)

Method: No method information available due to failed API request.

Result: No results available as the paper content could not be fetched.

Conclusion: Unable to analyze the paper due to technical limitations in accessing the content.

Abstract: Failed to fetch summary for 2509.21716: Page request resulted in HTTP 429 (https://export.arxiv.org/api/query?search_query=&id_list=2509.21716&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 draw conclusions without access to the paper content.

Abstract: Failed to fetch summary for 2510.05528: Page request resulted in HTTP 429 (https://export.arxiv.org/api/query?search_query=&id_list=2510.05528&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 draw conclusions due to fetch failure

Abstract: Failed to fetch summary for 2510.17421: Page request resulted in HTTP 429 (https://export.arxiv.org/api/query?search_query=&id_list=2510.17421&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 2510.17569: Page request resulted in HTTP 429 (https://export.arxiv.org/api/query?search_query=&id_list=2510.17569&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 2511.13394: Page request resulted in HTTP 429 (https://export.arxiv.org/api/query?search_query=&id_list=2511.13394&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 2512.00961: Page request resulted in HTTP 429 (https://export.arxiv.org/api/query?search_query=&id_list=2512.00961&sortBy=relevance&sortOrder=descending&start=0&max_results=100)

Method: Cannot analyze method as paper content is unavailable due to HTTP 429 error from arXiv API

Result: No results available - arXiv API returned HTTP 429 (Too Many Requests) rate limiting error

Conclusion: Paper analysis impossible due to technical limitations preventing access to the abstract/content

Abstract: Failed to fetch summary for 2512.03537: Page request resulted in HTTP 429 (https://export.arxiv.org/api/query?search_query=&id_list=2512.03537&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 2512.12744: Page request resulted in HTTP 429 (https://export.arxiv.org/api/query?search_query=&id_list=2512.12744&sortBy=relevance&sortOrder=descending&start=0&max_results=100)

Method: Unable to determine method as paper content could not be retrieved

Result: Unable to determine results as paper content could not be retrieved

Conclusion: Unable to determine conclusion as paper content could not be retrieved

Abstract: Failed to fetch summary for 2601.01119: Page request resulted in HTTP 429 (https://export.arxiv.org/api/query?search_query=&id_list=2601.01119&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 conclusion due to inability to access paper content

Abstract: Failed to fetch summary for 2601.13303: Page request resulted in HTTP 429 (https://export.arxiv.org/api/query?search_query=&id_list=2601.13303&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 2602.00918: Page request resulted in HTTP 429 (https://export.arxiv.org/api/query?search_query=&id_list=2602.00918&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 2602.01124: Page request resulted in HTTP 429 (https://export.arxiv.org/api/query?search_query=&id_list=2602.01124&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 provide analysis due to technical limitations in accessing the paper

Abstract: Failed to fetch summary for 2602.06932: Page request resulted in HTTP 429 (https://export.arxiv.org/api/query?search_query=&id_list=2602.06932&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 2602.23113: Page request resulted in HTTP 429 (https://export.arxiv.org/api/query?search_query=&id_list=2602.23113&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 2603.05433: Page request resulted in HTTP 429 (https://export.arxiv.org/api/query?search_query=&id_list=2603.05433&sortBy=relevance&sortOrder=descending&start=0&max_results=100)

Method: Unable to determine method as the paper content could not be retrieved

Result: Unable to determine results as the paper content could not be retrieved

Conclusion: Unable to determine conclusion as the paper content could not be retrieved

Abstract: Failed to fetch summary for 2603.21210: Page request resulted in HTTP 429 (https://export.arxiv.org/api/query?search_query=&id_list=2603.21210&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 2603.28750: Page request resulted in HTTP 429 (https://export.arxiv.org/api/query?search_query=&id_list=2603.28750&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 draw conclusions due to failed paper retrieval

Abstract: Failed to fetch summary for 2604.01328: Page request resulted in HTTP 429 (https://export.arxiv.org/api/query?search_query=&id_list=2604.01328&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 2604.01949: Page request resulted in HTTP 429 (https://export.arxiv.org/api/query?search_query=&id_list=2604.01949&sortBy=relevance&sortOrder=descending&start=0&max_results=100)

Method: No method information available due to failed API request.

Result: No results available as paper content could not be accessed.

Conclusion: Cannot provide conclusion without access to the paper’s content.

Abstract: Failed to fetch summary for 2305.08175: Page request resulted in HTTP 429 (https://export.arxiv.org/api/query?search_query=&id_list=2305.08175&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 limitations in accessing the content

Abstract: Failed to fetch summary for 2503.02178: Page request resulted in HTTP 429 (https://export.arxiv.org/api/query?search_query=&id_list=2503.02178&sortBy=relevance&sortOrder=descending&start=0&max_results=100)

Method: N/A - Paper content unavailable due to HTTP 429 error when attempting to fetch from arXiv API

Result: N/A - No results available as the paper summary could not be retrieved

Conclusion: Unable to analyze paper due to technical limitations in accessing the content from arXiv

Abstract: Failed to fetch summary for 2503.10773: Page request resulted in HTTP 429 (https://export.arxiv.org/api/query?search_query=&id_list=2503.10773&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 2505.21723: Page request resulted in HTTP 429 (https://export.arxiv.org/api/query?search_query=&id_list=2505.21723&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 abstract

Abstract: Failed to fetch summary for 2506.14022: Page request resulted in HTTP 429 (https://export.arxiv.org/api/query?search_query=&id_list=2506.14022&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 2507.14808: Page request resulted in HTTP 429 (https://export.arxiv.org/api/query?search_query=&id_list=2507.14808&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 due to lack of paper content

Abstract: Failed to fetch summary for 2508.09103: Page request resulted in HTTP 429 (https://export.arxiv.org/api/query?search_query=&id_list=2508.09103&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 2509.04603: Page request resulted in HTTP 429 (https://export.arxiv.org/api/query?search_query=&id_list=2509.04603&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 2510.02513: Page request resulted in HTTP 429 (https://export.arxiv.org/api/query?search_query=&id_list=2510.02513&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 determine conclusion as paper content is unavailable due to API rate limiting

Abstract: Failed to fetch summary for 2510.15483: Page request resulted in HTTP 429 (https://export.arxiv.org/api/query?search_query=&id_list=2510.15483&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 paper content

Abstract: Failed to fetch summary for 2511.01154: Page request resulted in HTTP 429 (https://export.arxiv.org/api/query?search_query=&id_list=2511.01154&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 2511.03909: Page request resulted in HTTP 429 (https://export.arxiv.org/api/query?search_query=&id_list=2511.03909&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 determine conclusion due to failed API request

Abstract: Failed to fetch summary for 2511.14617: Page request resulted in HTTP 429 (https://export.arxiv.org/api/query?search_query=&id_list=2511.14617&sortBy=relevance&sortOrder=descending&start=0&max_results=100)

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

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

Conclusion: Unable to draw conclusions due to API rate limiting preventing access to paper content

Abstract: Failed to fetch summary for 2511.16858: Page request resulted in HTTP 429 (https://export.arxiv.org/api/query?search_query=&id_list=2511.16858&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 about the paper due to inability to access its content

Abstract: Failed to fetch summary for 2603.02417: Page request resulted in HTTP 429 (https://export.arxiv.org/api/query?search_query=&id_list=2603.02417&sortBy=relevance&sortOrder=descending&start=0&max_results=100)

Method: Cannot determine method due to inability to fetch paper content

Result: Cannot determine results due to inability to fetch paper content

Conclusion: Cannot determine conclusion due to inability to fetch paper content

Abstract: Failed to fetch summary for 2603.24705: Page request resulted in HTTP 429 (https://export.arxiv.org/api/query?search_query=&id_list=2603.24705&sortBy=relevance&sortOrder=descending&start=0&max_results=100)

Method: Unable to determine method due to access limitations

Result: Unable to determine results due to access limitations

Conclusion: Unable to draw conclusions due to access limitations

Abstract: Failed to fetch summary for 2603.26227: Page request resulted in HTTP 429 (https://export.arxiv.org/api/query?search_query=&id_list=2603.26227&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 2603.28448: Page request resulted in HTTP 429 (https://export.arxiv.org/api/query?search_query=&id_list=2603.28448&sortBy=relevance&sortOrder=descending&start=0&max_results=100)

Method: Cannot analyze method as paper content is unavailable due to HTTP 429 error

Result: No results available - paper content inaccessible due to API rate limiting

Conclusion: Unable to provide analysis due to technical limitations preventing access to the paper

Abstract: Failed to fetch summary for 2603.28681: Page request resulted in HTTP 429 (https://export.arxiv.org/api/query?search_query=&id_list=2603.28681&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 2604.02244: Page request resulted in HTTP 429 (https://export.arxiv.org/api/query?search_query=&id_list=2604.02244&sortBy=relevance&sortOrder=descending&start=0&max_results=100)

Method: Compare LLM and human performance on Group Binary Search, an n-player game where participants independently submit numerical values to collectively sum to a target number without direct communication, relying only on group feedback for iterative adjustments.

Result: LLMs often fail to improve across games and exhibit excessive switching behavior that impairs group convergence, unlike humans who adapt and stabilize. Richer feedback benefits humans substantially but has small effects on LLMs.

Conclusion: LLMs show significant coordination gaps compared to humans, with different behavioral patterns. The study provides behaviorally grounded diagnostics for understanding and potentially closing these coordination gaps.

Abstract: Humans exhibit remarkable abilities to coordinate in groups. As large language models (LLMs) become more capable, it remains an open question whether they can demonstrate comparable adaptive coordination and whether they use the same strategies as humans. To investigate this, we compare LLM and human performance on a common-interest game with imperfect monitoring: Group Binary Search. In this n-player game, participants need to coordinate their actions to achieve a common objective. Players independently submit numerical values in an effort to collectively sum to a randomly assigned target number. Without direct communication, they rely on group feedback to iteratively adjust their submissions until they reach the target number. Our findings show that, unlike humans who adapt and stabilize their behavior over time, LLMs often fail to improve across games and exhibit excessive switching, which impairs group convergence. Moreover, richer feedback (e.g., numerical error magnitude) benefits humans substantially but has small effects on LLMs. Taken together, by grounding the analysis in human baselines and mechanism-level metrics, including reactivity scaling, switching dynamics, and learning across games, we point to differences in human and LLM groups and provide a behaviorally grounded diagnostic for closing the coordination gap.

Method: Large-scale empirical study analyzing over 1.5 million interactions across tasks, topologies, and scales using atomic event-level formulation that reconstructs reasoning as cascades of coordination. Introduced Deficit-Triggered Integration (DTI) to test the integration bottleneck mechanism.

Result: Uncovered three coupled laws: coordination follows heavy-tailed cascades, concentrates via preferential attachment into intellectual elites, and produces increasingly frequent extreme events as system size grows. DTI improved performance precisely where coordination failed without suppressing large-scale reasoning.

Conclusion: Established quantitative laws of collective cognition and identified coordination structure as a fundamental axis for understanding and improving scalable multi-agent intelligence, with integration bottleneck as the key structural mechanism.

Abstract: Large Language Model (LLM) multi-agent systems are increasingly deployed as interacting agent societies, yet scaling these systems often yields diminishing or unstable returns, the causes of which remain poorly understood. We present the first large-scale empirical study of coordination dynamics in LLM-based multi-agent systems, introducing an atomic event-level formulation that reconstructs reasoning as cascades of coordination. Analyzing over 1.5 Million interactions across tasks, topologies, and scales, we uncover three coupled laws: coordination follows heavy-tailed cascades, concentrates via preferential attachment into intellectual elites, and produces increasingly frequent extreme events as system size grows. We show that these effects are coupled through a single structural mechanism: an integration bottleneck, in which coordination expansion scales with system size while consolidation does not, producing large but weakly integrated reasoning processes. To test this mechanism, we introduce Deficit-Triggered Integration (DTI), which selectively increases integration under imbalance. DTI improves performance precisely where coordination fails, without suppressing large-scale reasoning. Together, our results establish quantitative laws of collective cognition and identify coordination structure as a fundamental, previously unmeasured axis for understanding and improving scalable multi-agent intelligence.

Method: Formulates microgrid decision-making as Decentralized Partially Observable Markov Decision Process (DEC-POMDP) solved using Multi-Agent Reinforcement Learning (MARL), with novel market clearing mechanism for macro-regulation.

Result: Simulation shows improved renewable energy utilization and reduced reliance on high-carbon external electricity, achieving balance between local autonomy, self-interest pursuit, and community-level economic/environmental benefits.

Conclusion: Proposed framework successfully integrates decentralized decision-making with market regulation to optimize both individual economic benefits and community-level low-carbon objectives in microgrid energy trading.

Abstract: The challenges of the uncertainties in renewable energy generation and the instability of the real-time market limit the effective utilization of clean energy in microgrid communities. Existing peer-to-peer (P2P) and microgrid coordination approaches typically rely on certain centralized optimization or restrictive coordination rules which are difficult to be implemented in real-life applications. To address the challenge, we propose an intraday P2P trading framework that allows self-interested microgrids to pursue their economic benefits, while allowing the market operator to maximize the social welfare, namely the low carbon emission objective, of the entire community. Specifically, the decision-making processes of the microgrids are formulated as a Decentralized Partially Observable Markov Decision Process (DEC-POMDP) and solved using a Multi-Agent Reinforcement Learning (MARL) framework. Such an approach grants each microgrid a high degree of decision-making autonomy, while a novel market clearing mechanism is introduced to provide macro-regulation, incentivizing microgrids to prioritize local renewable energy consumption and hence reduce carbon emissions. Simulation results demonstrate that the combination of the self-interested bidding strategy and the P2P market design helps significantly improve renewable energy utilization and reduce reliance on external electricity with high carbon-emissions. The framework achieves a balanced integration of local autonomy, self-interest pursuit, and improved community-level economic and environmental benefits.

Method: Proposes a distributed Q-learning algorithm with a redundancy-based filtering mechanism that uses two-hop neighbor information to validate incoming messages while preserving bidirectional information flow. Includes topological conditions for convergence and polynomial-time verification.

Result: The algorithm ensures all agents’ value functions converge almost surely to optimal value functions despite Byzantine edge attacks. Simulations show convergence to optimal solutions where prior methods fail.

Conclusion: The proposed method provides Byzantine resilience for distributed MARL with guaranteed convergence to optimal policies, overcoming limitations of existing approaches through novel filtering and topological conditions.

Abstract: We study Byzantine-resilient distributed multi-agent reinforcement learning (MARL), where agents must collaboratively learn optimal value functions over a compromised communication network. Existing resilient MARL approaches typically guarantee almost sure convergence only to near-optimal value functions, or require restrictive assumptions to ensure convergence to optimal solution. As a result, agents may fail to learn the optimal policies under these methods. To address this, we propose a novel distributed Q-learning algorithm, under which all agents’ value functions converge almost surely to the optimal value functions despite Byzantine edge attacks. The key idea is a redundancy-based filtering mechanism that leverages two-hop neighbor information to validate incoming messages, while preserving bidirectional information flow. We then introduce a new topological condition for the convergence of our algorithm, present a systematic method to construct such networks, and prove that this condition can be verified in polynomial time. We validate our results through simulations, showing that our method converges to the optimal solutions, whereas prior methods fail under Byzantine edge attacks.

Method: Multi-agent framework with four core agent types (households, firms, central bank, government) using LLMs for decision-making, operating in frictional labor, heterogeneous goods, and financial markets, plus a VLM for geographic firm placement and city visualization.

Result: SimCity naturally reproduces canonical macroeconomic phenomena including price elasticity of demand, Engel’s Law, Okun’s Law, Phillips Curve, and Beveridge Curve while remaining robust across simulation runs.

Conclusion: LLMs enable realistic, interpretable macroeconomic simulations with heterogeneous agents and rich interactions, bridging classical models and traditional ABMs while incorporating urban expansion dynamics through VLM integration.

Abstract: Large Language Models (LLMs) open new possibilities for constructing realistic and interpretable macroeconomic simulations. We present SimCity, a multi-agent framework that leverages LLMs to model an interpretable macroeconomic system with heterogeneous agents and rich interactions. Unlike classical equilibrium models that limit heterogeneity for tractability, or traditional agent-based models (ABMs) that rely on hand-crafted decision rules, SimCity enables flexible, adaptive behavior with transparent natural-language reasoning. Within SimCity, four core agent types (households, firms, a central bank, and a government) deliberate and participate in a frictional labor market, a heterogeneous goods market, and a financial market. Furthermore, a Vision-Language Model (VLM) determines the geographic placement of new firms and renders a mapped virtual city, allowing us to study both macroeconomic regularities and urban expansion dynamics within a unified environment. To evaluate the framework, we compile a checklist of canonical macroeconomic phenomena, including price elasticity of demand, Engel’s Law, Okun’s Law, the Phillips Curve, and the Beveridge Curve, and show that SimCity naturally reproduces these empirical patterns while remaining robust across simulation runs.

Method: Psychology-inspired multimodal stimuli designed to elicit diverse emotional, cognitive, and behavioral responses; collection of large-scale MMH dataset with clinically verified labels; paradigm-aware multimodal framework using inter-disorder differences as prompt-guided semantic descriptions to capture task-specific affective and interaction contexts.

Result: Extensive experiments show the proposed framework consistently outperforms existing baselines, demonstrating the value of psychology-inspired stimulus design for differential mental disorder detection.

Conclusion: Psychology-inspired multimodal stimuli and the paradigm-aware framework effectively address the challenge of differential mental disorder detection by capturing disorder-specific patterns through diverse elicitation tasks.

Abstract: Differential diagnosis of mental disorders remains a fundamental challenge in real-world clinical practice, where multiple conditions often exhibit overlapping symptoms. However, most existing public datasets are developed under single-disorder settings and rely on limited data elicitation paradigms, restricting their ability to capture disorder-specific patterns. In this work, we investigate differential mental disorder detection through psychology-inspired multimodal stimuli, designed to elicit diverse emotional, cognitive, and behavioral responses grounded in findings from experimental psychology. Based on this paradigm, we collect a large-scale multimodal mental health dataset (MMH) covering depression, anxiety, and schizophrenia, with all diagnostic labels clinically verified by licensed psychiatrists. To effectively model the heterogeneous signals induced by diverse elicitation tasks, we further propose a paradigm-aware multimodal framework that leverages inter-disorder differences prior knowledge as prompt-guided semantic descriptions to capture task-specific affective and interaction contexts for multimodal representation learning in the new differential mental disorder detection task. Extensive experiments show that our framework consistently outperforms existing baselines, underscoring the value of psychology-inspired stimulus design for differential mental disorder detection.

Method: Proposes Speaker-Reasoner, an end-to-end Speech LLM with agentic multi-turn temporal reasoning that iteratively analyzes global audio structure, autonomously predicts temporal boundaries, performs fine-grained segment analysis, jointly models speaker identity, gender, timestamps, and transcription, and uses a speaker-aware cache to extend processing beyond training context windows. Trained with a three-stage progressive strategy.

Result: Achieves consistent improvements over strong baselines on AliMeeting and AISHELL-4 datasets, particularly in handling overlapping speech and complex turn-taking.

Conclusion: Speaker-Reasoner effectively addresses multi-speaker conversation transcription challenges through agentic multi-turn temporal reasoning and speaker-aware processing.

Abstract: Transcribing and understanding multi-speaker conversations requires speech recognition, speaker attribution, and timestamp localization. While speech LLMs excel at single-speaker tasks, multi-speaker scenarios remain challenging due to overlapping speech, backchannels, rapid turn-taking, and context window constraints. We propose Speaker-Reasoner, an end-to-end Speech LLM with agentic multi-turn temporal reasoning. Instead of single-pass inference, the model iteratively analyzes global audio structure, autonomously predicts temporal boundaries, and performs fine-grained segment analysis, jointly modeling speaker identity, gender, timestamps, and transcription. A speaker-aware cache further extends processing to audio exceeding the training context window. Trained with a three-stage progressive strategy, Speaker-Reasoner achieves consistent improvements over strong baselines on AliMeeting and AISHELL-4 datasets, particularly in handling overlapping speech and complex turn-taking.

Method: The model maps noisy mixtures directly to candidate speaker embeddings trained to align with a strong single-speaker embedding space via permutation-invariant teacher supervision. These embeddings serve as control signals for extraction back-ends.

Result: On noisy LibriMix, the embeddings form a structured and clusterable identity space, outperforming WavLM+K-means and separation-derived embeddings in clustering metrics. Conditioning these embeddings into extraction back-ends improves objective quality and intelligibility, generalizing to real DNS-Challenge recordings.

Conclusion: The method successfully removes the need for clean enrollment in target speech extraction by predicting speaker embeddings directly from noisy mixtures, creating a structured identity space that improves extraction performance across various back-ends.

Abstract: Personalized or target speech extraction (TSE) typically needs a clean enrollment – hard to obtain in real-world crowded environments. We remove the essential need for enrollment by predicting, from the mixture itself, a small set of per-speaker embeddings that serve as the control signal for extraction. Our model maps a noisy mixture directly to a small set of candidate speaker embeddings trained to align with a strong single-speaker speaker-embedding space via permutation-invariant teacher supervision. On noisy LibriMix, the resulting embeddings form a structured and clusterable identity space, outperforming WavLM+K-means and separation-derived embeddings in standard clustering metrics. Conditioning these embeddings into multiple extraction back-ends consistently improves objective quality and intelligibility, and generalizes to real DNS-Challenge recordings.

Method: Proposes FLAIR (Full-duplex LAtent and Internal Reasoning) that recursively feeds latent embeddings from previous steps into next steps during user speech, enabling continuous causal reasoning. Uses Evidence Lower Bound-based objective for supervised finetuning via teacher forcing without explicit reasoning annotations.

Result: Achieves competitive results on speech benchmarks and robustly handles conversational dynamics with competitive performance on full-duplex interaction metrics.

Conclusion: The think-while-listening design effectively enables latent reasoning during speech perception, making it suitable for real-time spoken dialogue systems without introducing additional latency.

Abstract: During conversational interactions, humans subconsciously engage in concurrent thinking while listening to a speaker. Although this internal cognitive processing may not always manifest as explicit linguistic structures, it is instrumental in formulating high-quality responses. Inspired by this cognitive phenomenon, we propose a novel Full-duplex LAtent and Internal Reasoning method named FLAIR that conducts latent thinking simultaneously with speech perception. Unlike conventional “thinking” mechanisms in NLP, which require post-hoc generation, our approach aligns seamlessly with spoken dialogue systems: during the user’s speaking phase, it recursively feeds the latent embedding output from the previous step into the next step, enabling continuous reasoning that strictly adheres to causality without introducing additional latency. To enable this latent reasoning, we design an Evidence Lower Bound-based objective that supports efficient supervised finetuning via teacher forcing, circumventing the need for explicit reasoning annotations. Experiments demonstrate the effectiveness of this think-while-listening design, which achieves competitive results on a range of speech benchmarks. Furthermore, FLAIR robustly handles conversational dynamics and attains competitive performance on full-duplex interaction metrics.

Method: The paper conducts a comprehensive review and comparative analysis of fundus image datasets, evaluating them across key tasks (binary classification, severity grading, lesion localization, multi-disease screening) and categorizing by size, accessibility, and annotation type.

Result: The review consolidates current knowledge while highlighting persistent gaps such as lack of standardized lesion-level annotations and longitudinal data, and presents a case study of a recently published dataset to illustrate broader challenges in dataset curation.

Conclusion: The paper outlines recommendations for future dataset development to support clinically reliable and explainable solutions in DR screening, addressing current limitations in dataset quality and standardization.

Abstract: Diabetic Retinopathy (DR) is a serious microvascular complication of diabetes, and one of the leading causes of vision loss worldwide. Although automated detection and grading, with Deep Learning (DL), can reduce the burden on ophthalmologists, it is constrained by the limited availability of high-quality datasets. Existing repositories often remain geographically narrow, contain limited samples, and exhibit inconsistent annotations or variable image quality; thereby, restricting their clinical reliability. This paper presents a comprehensive review and comparative analysis of fundus image datasets used in the management of DR. The study evaluates their usability across key tasks, including binary classification, severity grading, lesion localization, and multi-disease screening. It also categorizes the datasets by size, accessibility, and annotation type (such as image-level, lesion-level, and multi-disease). Finally, a recently published dataset is presented as a case study to illustrate broader challenges in dataset curation and usage. The review consolidates current knowledge while highlighting persistent gaps such as the lack of standardized lesion-level annotations and longitudinal data. It also outlines recommendations for future dataset development to support clinically reliable and explainable solutions in DR screening.

Method: A principled learning strategy that leverages both source-domain image intensities and domain-stable foundation model representations to train robust segmentation models. It’s architecture- and loss-agnostic, compatible with standard augmentation pipelines, and computationally lightweight.

Result: Achieves strong gains in both fully supervised and few-shot segmentation across a broad range of shifts in biomedical studies. Outperforms prior approaches while being more practical.

Conclusion: DropGen provides an effective, lightweight solution for domain generalization in 3D biomedical segmentation that works across arbitrary anatomical regions and various domain shifts.

Abstract: We present DropGen, a simple and theoretically-grounded approach for domain generalization in 3D biomedical image segmentation. Modern segmentation models degrade sharply under shifts in modality, disease severity, clinical sites, and other factors, creating brittle models that limit reliable deployment. Existing domain generalization methods rely on extreme augmentations, mixing domain statistics, or architectural redesigns, yet incur significant implementation overhead and yield inconsistent performance across biomedical settings. DropGen instead proposes a principled learning strategy with minimal overhead that leverages both source-domain image intensities and domain-stable foundation model representations to train robust segmentation models. As a result, DropGen achieves strong gains in both fully supervised and few-shot segmentation across a broad range of shifts in biomedical studies. Unlike prior approaches, DropGen is architecture- and loss-agnostic, compatible with standard augmentation pipelines, computationally lightweight, and tackles arbitrary anatomical regions. Our implementation is freely available at https://github.com/sebodiaz/DropGen.

Method: Proposes TGPNet with Task-Guided Prompting (TGP) strategy using learnable task-specific embeddings to generate degradation-aware cues that hierarchically modulate features throughout the decoder, enabling precise restoration for distinct degradation patterns with shared weights.

Result: TGPNet achieves state-of-the-art performance on unified multi-task scenarios and unseen composite degradations, surpassing specialized models in individual domains like cloud removal, validated on a benchmark covering RGB, multispectral, SAR, and thermal infrared modalities.

Conclusion: The work presents a significant advancement for multi-task remote sensing image restoration by unifying heterogeneous degradation removal within a single adaptive framework, offering a practical and scalable solution for operational pipelines.

Abstract: Remote sensing image restoration (RSIR) is essential for recovering high-fidelity imagery from degraded observations, enabling accurate downstream analysis. However, most existing methods focus on single degradation types within homogeneous data, restricting their practicality in real-world scenarios where multiple degradations often across diverse spectral bands or sensor modalities, creating a significant operational bottleneck. To address this fundamental gap, we propose TGPNet, a unified framework capable of handling denoising, cloud removal, shadow removal, deblurring, and SAR despeckling within a single, unified architecture. The core of our framework is a novel Task-Guided Prompting (TGP) strategy. TGP leverages learnable, task-specific embeddings to generate degradation-aware cues, which then hierarchically modulate features throughout the decoder. This task-adaptive mechanism allows the network to precisely tailor its restoration process for distinct degradation patterns while maintaining a single set of shared weights. To validate our framework, we construct a unified RSIR benchmark covering RGB, multispectral, SAR, and thermal infrared modalities for five aforementioned restoration tasks. Experimental results demonstrate that TGPNet achieves state-of-the-art performance on both unified multi-task scenarios and unseen composite degradations, surpassing even specialized models in individual domains such as cloud removal. By successfully unifying heterogeneous degradation removal within a single adaptive framework, this work presents a significant advancement for multi-task RSIR, offering a practical and scalable solution for operational pipelines. The code and benchmark will be released at https://github.com/huangwenwenlili/TGPNet.

Method: A Unity-based prototype where a server constructs and continuously optimizes a 3D Gaussian Splatting (3DGS) model from real-time rendered reference views, then streams the evolving representation to remote clients using full model snapshots and incremental updates that support relighting and rigid object dynamics.

Result: The system enables clients to reconstruct the streamed Gaussian model locally and render their current viewpoint from the received representation, providing improved viewpoint flexibility for latency compensation compared to conventional image warping.

Conclusion: Streaming live 3DGS representations instead of 2D video offers advantages for cloud rendering in gaming/XR, including better viewpoint flexibility and more efficient amortization of server-side scene modeling across multiple users.

Abstract: Cloud rendering is widely used in gaming and XR to overcome limited client-side GPU resources and to support heterogeneous devices. Existing systems typically deliver the rendered scene as a 2D video stream, which tightly couples the transmitted content to the server-rendered viewpoint and limits latency compensation to image-space reprojection or warping. In this paper, we investigate an alternative approach based on streaming a live 3D Gaussian Splatting (3DGS) scene representation instead of only rendered video. We present a Unity-based prototype in which a server constructs and continuously optimizes a 3DGS model from real-time rendered reference views, while streaming the evolving representation to remote clients using full model snapshots and incremental updates supporting relighting and rigid object dynamics. The clients reconstruct the streamed Gaussian model locally and render their current viewpoint from the received representation. This approach aims to improve viewpoint flexibility for latency compensation and to better amortize server-side scene modeling across multiple users than per-user rendering and video streaming. We describe the system design, evaluate it, and compare it with conventional image warping.

Method: Two-stage flow matching approach: 1) initial training to generate plausible images, 2) distribution alignment via differentiable reward fine-tuning to steer generated images toward target domain distribution. Also includes flow matching based mask generation to enhance mask diversity.

Result: Extensive experiments show performance improvements of 3.5-4.0% in mDice and 3.5-5.6% in mIoU across various datasets and scenarios, demonstrating effectiveness even with limited target domain samples.

Conclusion: AlignFlow effectively addresses distribution mismatch in generative data augmentation for medical imaging, improving model generalization across heterogeneous clinical scenarios through distribution-aligned image-mask pair synthesis.

Abstract: Data heterogeneity hinders clinical deployment of medical image analysis models, and generative data augmentation helps mitigate this issue. However, recent diffusion-based methods that synthesize image-mask pairs often ignore distribution shifts between generated and real images across scenarios, and such mismatches can markedly degrade downstream performance. To address this issue, we propose AlignFlow, a flow matching model that aligns with the target reference image distribution via differentiable reward fine-tuning, and remains effective even when only a small number of reference images are provided. Specifically, we divide the training of the flow matching model into two stages: in the first stage, the model fits the training data to generate plausible images; Then, we introduce a distribution alignment mechanism and employ differentiable reward to steer the generated images toward the distribution of the given samples from the target domain. In addition, to enhance the diversity of generated masks, we also design a flow matching based mask generation to complement the diversity in regions of interest. Extensive experiments demonstrate the effectiveness of our approach, i.e., performance improvement by 3.5-4.0% in mDice and 3.5-5.6% in mIoU across a variety of datasets and scenarios.

Method: Created a dataset of 1,200 AR viewports by fusing high-resolution stereoscopic omnidirectional captures of real-world scenes with diverse augmented foregrounds under controlled transparency and degradation conditions. Conducted subjective study with 36 participants using video see-through HMD, collecting quality ratings and simulator-sickness indicators.

Result: Perceived quality is primarily driven by foreground degradations and modulated by transparency levels. Oculomotor and disorientation symptoms show progressive but manageable increase during viewing. Dataset reveals insights about visual confusion in AR environments.

Conclusion: ARIQA-3DS addresses the gap in ecologically valid AR quality assessment datasets and will serve as a benchmark for developing next-generation AR quality assessment models that account for the complex perceptual interplay between real and virtual content.

Abstract: As Augmented Reality (AR) technologies advance towards immersive consumer adoption, the need for rigorous Quality of Experience (QoE) assessment becomes critical. However, existing datasets often lack ecological validity, relying on monocular viewing or simplified backgrounds that fail to capture the complex perceptual interplay, termed visual confusion, between real and virtual layers. To address this gap, we present ARIQA-3DS, the first large stereoscopic AR Image Quality Assessment dataset. Comprising 1,200 AR viewports, the dataset fuses high-resolution stereoscopic omnidirectional captures of real-world scenes with diverse augmented foregrounds under controlled transparency and degradation conditions. We conducted a comprehensive subjective study with 36 participants using a video see-through head-mounted display, collecting both quality ratings and simulator-sickness indicators. Our analysis reveals that perceived quality is primarily driven by foreground degradations and modulated by transparency levels, while oculomotor and disorientation symptoms show a progressive but manageable increase during viewing. ARIQA-3DS will be publicly released to serve as a comprehensive benchmark for developing next-generation AR quality assessment models.

Method: Proposes HyperCT framework using a hypernetwork to dynamically adapt a Vision Transformer backbone. Integrates Low-Rank Adaptation (LoRA) for computational efficiency, allowing the model to regress task-specific low-rank weight updates rather than full parameters.

Result: Validated on large-scale dataset of radiological and cardiological tasks, HyperCT outperforms various strong baselines. Offers a unified, parameter-efficient solution for holistic patient assessment from non-contrast chest CTs.

Conclusion: HyperCT provides an effective framework for multi-task learning on medical imaging, combining hypernetworks with LoRA for efficient adaptation of Vision Transformers to diverse screening tasks in chest CT analysis.

Abstract: Non-contrast chest CTs offer a rich opportunity for both conventional pulmonary and opportunistic extra-pulmonary screening. While Multi-Task Learning (MTL) can unify these diverse tasks, standard hard-parameter sharing approaches are often suboptimal for modeling distinct pathologies. We propose HyperCT, a framework that dynamically adapts a Vision Transformer backbone via a Hypernetwork. To ensure computational efficiency, we integrate Low-Rank Adaptation (LoRA), allowing the model to regress task-specific low-rank weight updates rather than full parameters. Validated on a large-scale dataset of radiological and cardiological tasks, \method{} outperforms various strong baselines, offering a unified, parameter-efficient solution for holistic patient assessment. Our code is available at https://github.com/lfb-1/HyperCT.

Method: NFH-SEM integrates coarse multi-view geometry with photometric stereo cues from detector signals through a continuous neural field, incorporating a learnable forward model that embeds SEM imaging physics for self-calibrated, shadow-robust reconstruction.

Result: NFH-SEM achieves precise recovery across diverse specimens, revealing 478 nm layered features in two-photon lithography samples, 782 nm surface textures on pollen grains, and 1.559 μm fracture steps on silicon carbide particles.

Conclusion: NFH-SEM demonstrates accurate and broadly applicable 3D reconstruction from SEM images by combining neural fields with SEM imaging physics, with code and real-world dataset made available.

Abstract: The 3D characterization of microstructures is crucial for understanding and designing functional materials. However, the scanning electron microscope (SEM), widely used in scientific research, captures only 2D electron intensity distributions. Existing SEM 3D reconstruction methods struggle with textureless regions, shadowing artifacts, and calibration dependencies, whereas advanced learning-based approaches fail to generalize to microscopic SEM domains due to the lack of physical priors and domain-specific data. We introduce NFH-SEM, a neural field-based hybrid framework that reconstructs high-fidelity 3D surfaces from multi-view, multi-detector SEM images. NFH-SEM integrates coarse multi-view geometry with photometric stereo cues from detector signals through a continuous neural field, incorporating a learnable forward model that embeds SEM imaging physics for self-calibrated, shadow-robust reconstruction. NFH-SEM achieves precise recovery across diverse specimens, revealing 478 nm layered features in two-photon lithography samples, 782 nm surface textures on pollen grains, and 1.559 $μ$m fracture steps on silicon carbide particles, demonstrating its accuracy and broad applicability. Our code and real-world dataset are available at https://github.com/zju3dv/NFH-SEM.

Method: The authors generalize diffusion echoes to “filter echoes” applicable to various filters beyond diffusion. They develop a framework for visualizing echoes from different filters and propose compression techniques to reduce storage by factors of 20-100.

Result: The filter echo framework successfully visualizes various filters including those for adaptive smoothing, image inpainting, osmosis, and variational optic flow computation. Compression methods achieve 20-100x storage reduction while maintaining useful visualization properties.

Conclusion: Filter echoes provide a general, practical tool for understanding and selecting image processing filters across multiple applications, overcoming previous limitations of diffusion echoes through generalization and compression.

Abstract: To select suitable filters for a task or to improve existing filters, a deep understanding of their inner workings is vital. Diffusion echoes, which are space-adaptive impulse responses, are useful to visualise the effect of nonlinear diffusion filters. However, they have received little attention in the literature. There may be two reasons for this: Firstly, the concept was introduced specifically for diffusion filters, which might appear too limited. Secondly, diffusion echoes have large storage requirements, which restricts their practicality. This work addresses both problems. We introduce the filter echo as a generalisation of the diffusion echo and use it for applications beyond adaptive smoothing, such as image inpainting, osmosis, and variational optic flow computation. We provide a framework to visualise and inspect echoes from various filters with different applications. Furthermore, we propose a compression approach for filter echoes, which reduces storage requirements by a factor of 20 to 100.

Method: Systematic benchmark of four classical statistical detection algorithms (SAM, MF, ACE, CEM) and a proposed lightweight Spectral Neural Network with Parametric Mish activations. Evaluation uses pixel-level binary ground truth masks on VNIR hyperspectral dataset with inert PFM-1 targets across multiple scene crops.

Result: While ACE achieved highest ROC-AUC (0.989), the Spectral-NN outperformed classical detectors in precision-focused evaluation (precision-recall and average precision) due to extreme sparsity of target pixels. Results emphasize ROC-AUC can be misleading for imbalanced detection tasks.

Conclusion: The study demonstrates the need for precision-focused evaluation, scene-aware benchmarking, and learning-based spectral models for reliable UAV-based hyperspectral landmine detection. Code and pixel-level annotations will be released for reproducibility.

Abstract: In recent years, unmanned aerial vehicles (UAVs) equipped with imaging sensors and automated processing algorithms have emerged as a promising tool to accelerate large-area surveys while reducing risk to human operators. Although hyperspectral imaging (HSI) enables material discrimination using spectral signatures, standardized benchmarks for UAV-based landmine detection remain scarce. In this work, we present a systematic benchmark of four classical statistical detection algorithms, including Spectral Angle Mapper (SAM), Matched Filter (MF), Adaptive Cosine Estimator (ACE), and Constrained Energy Minimization (CEM), alongside a proposed lightweight Spectral Neural Network utilizing Parametric Mish activations for PFM-1 landmine detection. We also release pixel-level binary ground truth masks (target/background) to enable standardized, reproducible evaluation. Evaluations were conducted on inert PFM-1 targets across multiple scene crops using a recently released VNIR hyperspectral dataset. Metrics such as receiver operating characteristic (ROC) curve, area under the curve (AUC), precision-recall (PR) curve, and average precision (AP) were used. While all methods achieve high ROC-AUC on an independent test set, the ACE method observes the highest AUC of 0.989. However, because target pixels are extremely sparse relative to background, ROC-AUC alone can be misleading; under precision-focused evaluation (PR and AP), the Spectral-NN outperforms classical detectors, achieving the highest AP. These results emphasize the need for precision-focused evaluation, scene-aware benchmarking, and learning-based spectral models for reliable UAV-based hyperspectral landmine detection. The code and pixel-level annotations will be released.