Method: 1) AV-UIE v2 dataset: 222K samples across 17 datasets and 7 tasks with explicit reasoning processes; 2) Unified interface to align heterogeneous task formulations; 3) Interaction-aware LoRA (I-LoRA) that models inter-task relationships via dynamic routing to coordinate distinct audio-visual interaction patterns.

Result: Crab⁺ covers broader tasks than existing unified models while outperforming specialized models on various benchmarks. Successfully reversed negative transfer trend, achieving positive transfer where multi-task learning surpasses single-task baselines in nearly 88% of tasks.

Conclusion: Crab⁺ represents a robust step toward holistic audio-visual scene understanding by addressing task heterogeneity through explicit cooperation from both data and model perspectives, validated across diverse AV-LLM paradigms.

Abstract: Developing Audio-Visual Large Language Models (AV-LLMs) for unified scene understanding is pivotal in multimodal intelligence. While instruction tuning enables pre-trained models with multi-task abilities, we observe that conventional multi-task unification methods often suffer from severe negative transfer, where nearly 55% of tasks degrade compared to single-task training. We attribute this phenomenon to audio-visual task heterogeneity, characterized by disparate task granularity and divergent capability demands, which lead to negative interference under joint training. To tackle this, we present Crab$^{+}$, a scalable and unified audio-visual scene understanding model that addresses task heterogeneity through explicit cooperation from both data and model perspectives. On the data side, we introduce AV-UIE v2, a comprehensive Audio-Visual Unified Instruction-tuning dataset with Explicit reasoning processes. It contains approximately 222K samples spanning 17 datasets and 7 tasks, enabling the model to capture cross-task relationships at different levels of granularity. On the model side, we design a unified interface to align heterogeneous task formulations, and propose Interaction-aware LoRA (I-LoRA), which explicitly models inter-task relationships via dynamic routing to coordinate distinct audio-visual interaction patterns, mitigating parameter interference. Extensive experiments show Crab$^{+}$ covers broader tasks than existing unified models while outperforming specialized models on various benchmarks. We successfully reverse the negative transfer trend, achieving positive transfer where multi-task learning surpasses single-task baselines in nearly 88% of tasks. These results hold across diverse AV-LLM paradigms and are validated through in-depth visualization, positioning Crab$^{+}$ as a robust step towards holistic audio-visual scene understanding.

Method: Proposes AVUR-LLM with two key components: 1) Sparse modality alignment for better cross-modal interaction, and 2) Visual unit-guided refinement to enhance recognition accuracy using visual information.

Result: Achieves state-of-the-art results on LRS3 dataset, with 37% relative improvement over baseline at 0 dB SNR under additive-noise conditions.

Conclusion: AVUR-LLM effectively addresses limitations of previous LLM-based AVSR approaches through improved cross-modal alignment and refinement, demonstrating significant robustness in noisy acoustic conditions.

Abstract: Audio-Visual Speech Recognition (AVSR) integrates acoustic and visual information to enhance robustness in adverse acoustic conditions. Recent advances in Large Language Models (LLMs) have yielded competitive automatic speech recognition performance and shown effectiveness for AVSR. However, prior approaches project audio and visual features independently or apply shallow fusion, limiting cross-modal alignment and complementary exchange while increasing the LLM’s computational load. To address this, we propose AVUR-LLM, an LLM-based Audio-Visual Speech Recognition via Sparse Modality Alignment and Visual Unit-Guided Refinement. Experiments on LRS3 demonstrate state-of-the-art results for AVSR. Under additive-noise conditions at 0 dB SNR, it achieves 37% relative improvement over the baseline system.

Method: Used 71K AudioCapsV2 clips to extract normalized (source, attribute) events. Built two query types: present-event queries for ground-truth detection and absent-event queries to probe hallucinations, using similarity-filtered negative sampling in audio-aligned text embedding space. Evaluated four SOTA Audio LLMs with 12 prompt variants over 500K yes/no queries per model.

Result: Across all models, increasing event count consistently lowers true-positive rate and raises false-positive rate. Prompts induce a strong trade-off between true-positive and false-positive rates. Confidence analysis shows models become more uncertain on multi-event audio, revealing significant room for improvement.

Conclusion: Audio LLMs struggle with complex acoustic scenes, showing degraded performance as event complexity increases. The systematic evaluation reveals fundamental limitations in current models’ ability to handle multi-event audio, highlighting important directions for future research in robust audio understanding.

Abstract: Audio LLMs have shown a strong ability to understand audio samples, yet their reliability in complex acoustic scenes remains under-explored. Unlike prior work limited to small scale or less controlled query construction, we present a large-scale evaluation of event grounding and false alarms as auditory scene complexity increases. Using 71K AudioCapsV2 clips, we extract normalized (source, attribute) events and build two query types: present-event queries for ground-truth detection and absent-event queries to probe hallucinations, using similarity-filtered negative sampling in an audio-aligned text embedding space. We evaluate four SOTA Audio LLMs with 12 prompt variants over 500K yes/no queries per model. Across models, increasing event count consistently lowers true-positive rate and raises false-positive rate, while prompts induce a strong trade-off between the two. Our confidence analysis shows that models become more uncertain on multi-event audio, revealing room for improvement.

Method: Two-phase pipeline: 1) Offline construction with entropy-aware gating to filter noise and conflict-aware coarsening to merge static duplicates while preserving state transitions as temporal edges. 2) Online reasoning with algorithmic bridge discovery to reconstruct missing logical paths between retrieved facts, followed by single-call topology-aware synthesis.

Result: On LoCoMo experiments with GPT-4o, AriadneMem improves Multi-Hop F1 by 15.2% and Average F1 by 9.0% over strong baselines, reduces total runtime by 77.8% using only 497 context tokens.

Conclusion: AriadneMem effectively addresses long-term dialogue challenges by offloading reasoning to graph layers, achieving significant performance improvements with minimal context usage.

Abstract: Long-horizon LLM agents require memory systems that remain accurate under fixed context budgets. However, existing systems struggle with two persistent challenges in long-term dialogue: (i) \textbf{disconnected evidence}, where multi-hop answers require linking facts distributed across time, and (ii) \textbf{state updates}, where evolving information (e.g., schedule changes) creates conflicts with older static logs. We propose AriadneMem, a structured memory system that addresses these failure modes via a decoupled two-phase pipeline. In the \textbf{offline construction phase}, AriadneMem employs \emph{entropy-aware gating} to filter noise and low-information message before LLM extraction and applies \emph{conflict-aware coarsening} to merge static duplicates while preserving state transitions as temporal edges. In the \textbf{online reasoning phase}, rather than relying on expensive iterative planning, AriadneMem executes \emph{algorithmic bridge discovery} to reconstruct missing logical paths between retrieved facts, followed by \emph{single-call topology-aware synthesis}. On LoCoMo experiments with GPT-4o, AriadneMem improves \textbf{Multi-Hop F1 by 15.2%} and \textbf{Average F1 by 9.0%} over strong baselines. Crucially, by offloading reasoning to the graph layer, AriadneMem reduces \textbf{total runtime by 77.8%} using only \textbf{497} context tokens. The code is available at https://github.com/LLM-VLM-GSL/AriadneMem.

Method: Systematically measured biases in five high-quality RMs, categorized failures by complexity, and proposed mechanistic reward shaping as a post-hoc intervention to mitigate low-complexity biases from spurious correlations.

Result: Identified persistent biases (length, sycophancy, overconfidence) and discovered new biases (model-specific styles, answer-order). The mechanistic reward shaping method reduces targeted biases without degrading reward quality using minimal labeled data.

Conclusion: Reward models have systematic biases that enable reward hacking; mechanistic reward shaping effectively mitigates low-complexity biases and generalizes well, offering a practical solution for improving RM reliability.

Abstract: Reward Models (RMs) are crucial for online alignment of language models (LMs) with human preferences. However, RM-based preference-tuning is vulnerable to reward hacking, whereby LM policies learn undesirable behaviors from flawed RMs. By systematically measuring biases in five high-quality RMs, including the state-of-the-art, we find that issues persist despite prior work with respect to length, sycophancy, and overconfidence. We also discover new issues related to bias toward model-specific styles and answer-order. We categorize RM failures by complexity and propose a simple post-hoc intervention to mitigate low-complexity biases that arise from spurious correlations. Our proposed mechanistic reward shaping reduces targeted biases without degrading reward quality and while using minimal labeled data. The method is extensible to new biases, model-internal, and generalizes out-of-distribution.

Method: MA-RAG uses an agentic refinement loop where at each round: 1) transforms semantic conflicts among candidate responses into retrieval queries, 2) optimizes reasoning history to mitigate long-context degradation, 3) extends self-consistency principle by using inconsistency as a signal for multi-round reasoning, and 4) mirrors boosting to iteratively minimize residual error toward stable consensus.

Result: Extensive evaluations across 7 medical Q&A benchmarks show MA-RAG consistently surpasses competitive inference-time scaling and RAG baselines, delivering substantial +6.8 points average accuracy improvement over backbone models.

Conclusion: MA-RAG effectively addresses LLM hallucinations in medical QA through multi-round agentic refinement, leveraging conflict-driven retrieval and reasoning optimization to achieve higher accuracy and reliability.

Abstract: Large Language Models (LLMs) exhibit high reasoning capacity in medical question-answering, but their tendency to produce hallucinations and outdated knowledge poses critical risks in healthcare fields. While Retrieval-Augmented Generation (RAG) mitigates these issues, existing methods rely on noisy token-level signals and lack the multi-round refinement required for complex reasoning. In the paper, we propose MA-RAG (Multi-Round Agentic RAG), a framework that facilitates test-time scaling for complex medical reasoning by iteratively evolving both external evidence and internal reasoning history within an agentic refinement loop. At each round, the agent transforms semantic conflict among candidate responses into actionable queries to retrieve external evidence, while optimizing history reasoning traces to mitigate long-context degradation. MA-RAG extends the self-consistency principle by leveraging the lack of consistency as a proactive signal for multi-round agentic reasoning and retrieval, and mirrors a boosting mechanism that iteratively minimizes the residual error toward a stable, high-fidelity medical consensus. Extensive evaluations across 7 medical Q&A benchmarks show that MA-RAG consistently surpasses competitive inference-time scaling and RAG baselines, delivering substantial +6.8 points on average accuracy over the backbone model. Our code is available at this url.

Method: Three components: 1) Memory purification with Think-Search-Memorize strategy to retain salient evidence while filtering irrelevant content, 2) Atomic query training to promote shorter and more diverse queries for better evidence acquisition, 3) Dense rewards for fine-grained feedback to speed training.

Result: SE-Search-3B outperforms strong baselines on single-hop and multi-hop QA benchmarks, achieving 10.8 point absolute improvement and 33.8% relative gain over Search-R1.

Conclusion: SE-Search effectively improves online search behavior in retrieval-augmented generation systems through self-evolving mechanisms that address noise accumulation and sparse reward problems.

Abstract: Retrieval augmented generation (RAG) reduces hallucinations and factual errors in large language models (LLMs) by conditioning generation on retrieved external knowledge. Recent search agents further cast RAG as an autonomous, multi-turn information-seeking process. However, existing methods often accumulate irrelevant or noisy documents and rely on sparse reinforcement learning signals. We propose \textbf{S}elf-\textbf{E}volving \textbf{Search}, a Self-Evolving Search agent that improves online search behavior through three components, memory purification, atomic query training, and dense rewards. SE-Search follows a \textit{Think-Search-Memorize} strategy that retains salient evidence while filtering irrelevant content. Atomic query training promotes shorter and more diverse queries, improving evidence acquisition. Dense rewards provide fine-grained feedback that speeds training. Experiments on single-hop and multi-hop question answering benchmarks show that \texttt{SE-Search-3B} outperforms strong baselines, yielding a $10.8$ point absolute improvement and a $33.8%$ relative gain over Search-R1.\footnote{We will make the code and model weights publicly available upon acceptance.}

Method: Hybrid architecture decouples factual retrieval from conversational delivery: supervised fine-tuning with LoRA on expert-curated GOLDEN FACTS (atomic, verified agricultural knowledge) optimizes fact recall, while a separate stitching layer transforms retrieved facts into culturally appropriate, safety-aware responses.

Result: Fine-tuning on curated data substantially improves fact recall and F1 while maintaining high relevance; fine-tuned smaller models achieve comparable or better factual quality at lower cost; stitching layer improves safety subscores while maintaining conversational quality.

Conclusion: The hybrid approach enables responsible deployment of LLMs in agricultural advisory by improving factual accuracy and cultural appropriateness while reducing costs, with released farmerchat-prompts library for reproducible domain-specific AI development.

Abstract: Large Language Models show promise for agricultural advisory, yet vanilla models exhibit unsupported recommendations, generic advice lacking specific, actionable detail, and communication styles misaligned with smallholder farmer needs. In high stakes agricultural contexts, where recommendation accuracy has direct consequences for farmer outcomes, these limitations pose challenges for responsible deployment. We present a hybrid LLM architecture that decouples factual retrieval from conversational delivery: supervised fine-tuning with LoRA on expert-curated GOLDEN FACTS (atomic, verified units of agricultural knowledge) optimizes fact recall, while a separate stitching layer transforms retrieved facts into culturally appropriate, safety-aware responses. Our evaluation framework, DG-EVAL, performs atomic fact verification (measuring recall, precision, and contradiction detection) against expert-curated ground truth rather than Wikipedia or retrieved documents. Experiments across multiple model configurations on crops and queries from Bihar, India show that fine-tuning on curated data substantially improves fact recall and F1, while maintaining high relevance. Using a fine-tuned smaller model achieves comparable or better factual quality at a fraction of the cost of frontier models. A stitching layer further improves safety subscores while maintaining high conversational quality. We release the farmerchat-prompts library to enable reproducible development of domain-specific agricultural AI.

Method: Introduce OTS-Bench to quantify Order-to-Space Bias (OTS) using paired prompts differing only in entity order, evaluating models along homogenization and correctness dimensions. Conduct experiments to show OTS is widespread, analyze its origins, and test interventions through targeted fine-tuning and early-stage layout formation strategies.

Result: Experiments show OTS is widespread in modern image generation models, primarily data-driven, and manifests during early layout formation. Both targeted fine-tuning and early-stage intervention strategies can substantially reduce OTS while preserving generation quality.

Conclusion: Order-to-Space Bias is a systematic, data-driven problem in image generation models that can be mitigated through targeted interventions, improving spatial reasoning and entity-role binding without compromising generation quality.

Abstract: We study a systematic bias in modern image generation models: the mention order of entities in text spuriously determines spatial layout and entity–role binding. We term this phenomenon Order-to-Space Bias (OTS) and show that it arises in both text-to-image and image-to-image generation, often overriding grounded cues and causing incorrect layouts or swapped assignments. To quantify OTS, we introduce OTS-Bench, which isolates order effects with paired prompts differing only in entity order and evaluates models along two dimensions: homogenization and correctness. Experiments show that Order-to-Space Bias (OTS) is widespread in modern image generation models, and provide evidence that it is primarily data-driven and manifests during the early stages of layout formation. Motivated by this insight, we show that both targeted fine-tuning and early-stage intervention strategies can substantially reduce OTS, while preserving generation quality.

Method: Direct assessment of LLMs as proxies for human goal selection using a controlled, open-ended learning task from cognitive science, testing four state-of-the-art models (GPT-5, Gemini 2.5 Pro, Claude Sonnet 4.5, and Centaur).

Result: Substantial divergence from human behavior: while people gradually explore and learn with diversity across individuals, most models exploit a single identified solution (reward hacking) or show surprisingly low performance, with distinct patterns across models and little variability across instances of the same model.

Conclusion: Human goal selection is unique and cannot be adequately captured by current LLMs, cautioning against replacing human goal selection with models in applications like personal assistance, scientific discovery, and policy research.

Abstract: As large language models (LLMs) get integrated into human decision-making, they are increasingly choosing goals autonomously rather than only completing human-defined ones, assuming they will reflect human preferences. However, human-LLM similarity in goal selection remains largely untested. We directly assess the validity of LLMs as proxies for human goal selection in a controlled, open-ended learning task borrowed from cognitive science. Across four state-of-the-art models (GPT-5, Gemini 2.5 Pro, Claude Sonnet 4.5, and Centaur), we find substantial divergence from human behavior. While people gradually explore and learn to achieve goals with diversity across individuals, most models exploit a single identified solution (reward hacking) or show surprisingly low performance, with distinct patterns across models and little variability across instances of the same model. Even Centaur, explicitly trained to emulate humans in experimental settings, poorly captures people’s goal selection. Chain-of-thought reasoning and persona steering provide limited improvements. These findings highlight the uniqueness of human goal selection, cautioning against replacing it with current models in applications such as personal assistance, scientific discovery, and policy research.

Method: PlugMem structures episodic memories into a compact, extensible knowledge-centric memory graph that explicitly represents propositional and prescriptive knowledge. This representation enables efficient memory retrieval and reasoning over task-relevant knowledge rather than verbose raw trajectories, treating knowledge as the unit of memory access and organization instead of entities or text chunks.

Result: PlugMem consistently outperforms task-agnostic baselines and exceeds task-specific memory designs across three heterogeneous benchmarks: long-horizon conversational question answering, multi-hop knowledge retrieval, and web agent tasks. It also achieves the highest information density under a unified information-theoretic analysis.

Conclusion: PlugMem provides an effective, task-agnostic memory solution for LLM agents that leverages cognitive science principles to structure memories as knowledge graphs, enabling efficient retrieval and reasoning while maintaining transferability across different tasks.

Abstract: Long-term memory is essential for large language model (LLM) agents operating in complex environments, yet existing memory designs are either task-specific and non-transferable, or task-agnostic but less effective due to low task-relevance and context explosion from raw memory retrieval. We propose PlugMem, a task-agnostic plugin memory module that can be attached to arbitrary LLM agents without task-specific redesign. Motivated by the fact that decision-relevant information is concentrated as abstract knowledge rather than raw experience, we draw on cognitive science to structure episodic memories into a compact, extensible knowledge-centric memory graph that explicitly represents propositional and prescriptive knowledge. This representation enables efficient memory retrieval and reasoning over task-relevant knowledge, rather than verbose raw trajectories, and departs from other graph-based methods like GraphRAG by treating knowledge as the unit of memory access and organization instead of entities or text chunks. We evaluate PlugMem unchanged across three heterogeneous benchmarks (long-horizon conversational question answering, multi-hop knowledge retrieval, and web agent tasks). The results show that PlugMem consistently outperforms task-agnostic baselines and exceeds task-specific memory designs, while also achieving the highest information density under a unified information-theoretic analysis. Code and data are available at https://github.com/TIMAN-group/PlugMem.

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: Extensive experiments show 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 semantic prompting and improved evaluation metrics, enabling more signal-grounded and semantically faithful generation.

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: TTSR uses a single pretrained language model that alternates between Student and Teacher roles. The Student solves problems and learns from variant questions, while the Teacher analyzes failed reasoning trajectories, identifies recurring weaknesses, and synthesizes targeted variant questions to guide improvement.

Result: Experimental results on multiple challenging mathematical reasoning benchmarks show TTSR consistently improves reasoning performance and generalizes well across different model backbones and general-domain reasoning tasks.

Conclusion: Teacher-mediated self-reflection provides an effective pathway for stable and continual reasoning improvement at test time, enabling models to adapt to their specific weaknesses.

Abstract: Test-time Training enables model adaptation using only test questions and offers a promising paradigm for improving the reasoning ability of large language models (LLMs). However, it faces two major challenges: test questions are often highly difficult, making self-generated pseudo-labels unreliable, and existing methods lack effective mechanisms to adapt to a model’s specific reasoning weaknesses, leading to inefficient learning. To address these issues, we propose \textbf{TTSR}, a self-reflective test-time self-evolving training framework. TTSR employs a single pretrained language model that alternates between the roles of a \textit{Student} and a \textit{Teacher} at test time. The Student focuses on solving problems and learning from synthesized variant questions, while the Teacher analyzes the Student’s failed reasoning trajectories, summarizes recurring reasoning weaknesses, and synthesizes targeted variant questions accordingly. This process guides the model to improve within a learnable regime through a continual self-evolving loop. Experimental results on multiple challenging mathematical reasoning benchmarks show that TTSR consistently improves reasoning performance and generalizes well across different model backbones and general-domain reasoning tasks. These findings suggest that teacher-mediated self-reflection provides an effective pathway for stable and continual reasoning improvement at test time.

Method: TATRA constructs instance-specific few-shot prompts by synthesizing on-the-fly examples to accompany user-provided instructions. It requires no labeled training data and avoids task-specific optimization loops while retaining demonstration-based prompting benefits.

Result: Across text classification benchmarks, TATRA matches or improves over strong prompt-optimization baselines that depend on training data and extensive search. On mathematical reasoning benchmarks, TATRA achieves SOTA performance on GSM8K and DeepMath, outperforming methods that explicitly optimize prompts on those tasks.

Conclusion: Per-instance construction of effective in-context examples is more important than running long, expensive optimization loops to produce a single prompt per task. TATRA provides an efficient, dataset-free alternative to traditional prompt engineering methods.

Abstract: Large Language Models (LLMs) have improved substantially alignment, yet their behavior remains highly sensitive to prompt phrasing. This brittleness has motivated automated prompt engineering, but most existing methods (i) require a task-specific training set, (ii) rely on expensive iterative optimization to produce a single dataset-level prompt, and (iii) must be rerun from scratch for each new task. We introduce TATRA, a dataset-free prompting method that constructs instance-specific few-shot prompts by synthesizing on-the-fly examples to accompany a user-provided instruction. TATRA requires no labeled training data and avoids task-specific optimization loops, while retaining the benefits of demonstration-based prompting. Across standard text classification benchmarks, TATRA matches or improves over strong prompt-optimization baselines that depend on training data and extensive search. On mathematical reasoning benchmarks, TATRA achieves state-of-the-art performance on GSM8K and DeepMath, outperforming methods that explicitly optimize prompts on those tasks. Our results suggest that per-instance construction of effective in-context examples is more important than running long, expensive optimization loops to produce a single prompt per task. We will make all code publicly available upon acceptance of the paper. Code is available at https://github.com/BMD223/TATRA

Method: Conducted large-scale audit of 10 commercially deployed LLMs across four academic domains, generating 69,557 citation instances verified against three scholarly databases (CrossRef, OpenAlex, Semantic Scholar). Analyzed hallucination rates, identified practical filters (multi-model consensus, within-prompt repetition), and developed a lightweight classifier using bibliographic string features.

Result: Hallucination rates varied fivefold (11.4-56.8%) and were strongly influenced by model, domain, and prompt framing. No model spontaneously generated citations when unprompted. Multi-model consensus (≥3 LLMs) achieved 95.6% accuracy (5.8× improvement), while within-prompt repetition (≥2 replications) achieved 88.9% accuracy. The classifier achieved AUC 0.876 in cross-validation and 0.834 in LOMO generalization.

Conclusion: Citation hallucination in LLMs is a significant, prompt-induced problem with variable rates across models, but practical detection methods exist including consensus approaches and automated classifiers that can be deployed at inference time.

Abstract: Large language models (LLMs) have been noted to fabricate scholarly citations, yet the scope of this behavior across providers, domains, and prompting conditions remains poorly quantified. We present one of the largest citation hallucination audits to date, in which 10 commercially deployed LLMs were prompted across four academic domains, generating 69,557 citation instances verified against three scholarly databases (namely, CrossRef, OpenAlex, and Semantic Scholar). Our results show that the observed hallucination rates span a fivefold range (between 11.4% and 56.8%) and are strongly shaped by model, domain, and prompt framing. Our results also show that no model spontaneously generates citations when unprompted, which seems to establish hallucination as prompt-induced rather than intrinsic. We identify two practical filters: 1) multi-model consensus (with more than 3 LLMs citing the same work yields 95.6% accuracy, a 5.8-fold improvement), and 2) within-prompt repetition (with more than 2 replications yields 88.9% accuracy). In addition, we present findings on generational model tracking, which reveal that improvements are not guaranteed when deploying newer LLMs, and on capacity scaling, which appears to reduce hallucination within model families. Finally, a lightweight classifier trained solely on bibliographic string features is developed to classify hallucinated citations from verified citations, achieving AUC 0.876 in cross-validation and 0.834 in LOMO generalization (without querying any external database). This classifier offers a pre-screening tool deployable at inference time.

Method: Evaluated three tools on LaborBench: STARA (custom statutory research tool), Westlaw AI, and Lexis+ AI. Compared outputs to DOL attorney-compiled ground truth for U.S. state unemployment insurance requirements. Conducted comprehensive error analysis.

Result: STARA achieved 83% accuracy (92% after accounting for DOL attorney omissions). Commercial tools performed worse: Westlaw AI 58%, Lexis+ AI 64% (both below standard RAG’s 70%). Error analysis revealed reasoning errors and retrieval failures.

Conclusion: STARA shows substantial improvements over standard RAG and commercial tools. Many apparent errors were actually DOL attorney omissions. Provides design principles for building accurate multi-jurisdictional legal AI systems.

Abstract: Retrieval-augmented generation (RAG) offers significant potential for legal AI, yet systematic benchmarks are sparse. Prior work introduced LaborBench to benchmark RAG models based on ostensible ground truth from an exhaustive, multi-month, manual enumeration of all U.S. state unemployment insurance requirements by U.S. Department of Labor (DOL) attorneys. That prior work found poor performance of standard RAG (70% accuracy on Boolean tasks). Here, we assess three emerging tools not previously evaluated on LaborBench: the Statutory Research Assistant (STARA), a custom statutory research tool, and two commercial tools by Westlaw and LexisNexis marketing AI statutory survey capabilities. We make five main contributions. First, we show that STARA achieves substantial performance gains, boosting accuracy to 83%. Second, we show that commercial platforms fare poorly, with accuracy of 58% (Westlaw AI) and 64% (Lexis+ AI), even worse than standard RAG. Third, we conduct a comprehensive error analysis, comparing our outputs to those compiled by DOL attorneys, and document both reasoning errors, such as confusion between related legal concepts and misinterpretation of statutory exceptions, and retrieval failures, where relevant statutory provisions are not captured. Fourth, we discover that many apparent errors are actually significant omissions by DOL attorneys themselves, such that STARA’s actual accuracy is 92%. Fifth, we chart the path forward for legal RAG through concrete design principles, offering actionable guidance for building AI systems capable of accurate multi-jurisdictional legal research.

Method: The authors explore offline policies for semantic caching, prove that optimal offline policy is NP-hard, propose polynomial-time heuristics, and present online semantic-aware cache policies that combine recency, frequency, and locality.

Result: Evaluations on diverse datasets show that while frequency-based policies are strong baselines, their novel variant improves semantic accuracy. The findings reveal effective strategies for current systems and highlight substantial headroom for future innovation.

Conclusion: Semantic caching presents new challenges requiring specialized policies, and while effective strategies exist, there’s significant room for future innovation in this area. All code is open source.

Abstract: The rapid adoption of large language models (LLMs) has created demand for faster responses and lower costs. Semantic caching, reusing semantically similar requests via their embeddings, addresses this need but breaks classic cache assumptions and raises new challenges. In this paper, we explore offline policies for semantic caching, proving that implementing an optimal offline policy is NP-hard, and propose several polynomial-time heuristics. We also present online semantic aware cache policies that combine recency, frequency, and locality. Evaluations on diverse datasets show that while frequency based policies are strong baselines, our novel variant improves semantic accuracy. Our findings reveal effective strategies for current systems and highlight substantial headroom for future innovation. All code is open source.

Method: Proposes a Retrieval-Augmented Generation (RAG) framework with multi-agent architecture. Uses specialized agents for retrieval, answer generation, evaluation, and query refinement for iterative improvement. Integrates an open-weight vision-language model to convert technical figures into semantic textual representations, allowing figure-based knowledge to be indexed and retrieved alongside text. Retrieved text and figure context are provided to an open-weight LLM for evidence-grounded responses.

Result: The paper proposes a system that enables more effective knowledge management by combining structured document retrieval with real-time, context-aware response generation. The multi-agent approach allows for iterative improvement and quality control, while the vision-language integration enables comprehensive knowledge retrieval including technical figures.

Conclusion: The proposed RAG framework with multi-agent architecture and vision-language integration addresses critical knowledge management challenges in transportation agencies, potentially improving decision-making, reducing learning curves for new personnel, and preserving institutional expertise.

Abstract: Effective knowledge management is critical for preserving institutional expertise and improving the efficiency of workforce training in state transportation agencies. Traditional approaches, such as static documentation, classroom-based instruction, and informal mentorship, often lead to fragmented knowledge transfer, inefficiencies, and the gradual loss of expertise as senior engineers retire. Moreover, given the enormous volume of technical manuals, guidelines, and research reports maintained by these agencies, it is increasingly challenging for engineers to locate relevant information quickly and accurately when solving field problems or preparing for training tasks. These limitations hinder timely decision-making and create steep learning curves for new personnel in maintenance and construction operations. To address these challenges, this paper proposes a Retrieval-Augmented Generation (RAG) framework with a multi-agent architecture to support knowledge management and decision making. The system integrates structured document retrieval with real-time, context-aware response generation powered by a large language model (LLM). Unlike conventional single-pass RAG systems, the proposed framework employs multiple specialized agents for retrieval, answer generation, evaluation, and query refinement, which enables iterative improvement and quality control. In addition, the system incorporates an open-weight vision-language model to convert technical figures into semantic textual representations, which allows figure-based knowledge to be indexed and retrieved alongside text. Retrieved text and figure-based context are then provided to an open-weight large language model, which generates the final responses grounded in the retrieved evidence.

Method: HumanLM trains user simulators to generate both responses and natural-language latent states aligned with ground-truth responses via reinforcement learning. These latent states correspond to psychologically grounded dimensions that drive real user responses. The framework synthesizes aligned latent states into accurate user representations.

Result: HumanLM significantly outperforms alternative approaches across six large-scale datasets (26k users, 216k responses), achieving 16.3% average relative improvement in alignment scores from an LLM judge. In real-time simulation with 111 participants, it achieved highest similarity to real user responses and competitive human-likeness scores.

Conclusion: HumanLM successfully creates user simulators that go beyond surface-level imitation to capture psychologically grounded latent states, enabling more accurate simulation of real users’ underlying beliefs and emotions.

Abstract: Large Language Models (LLMs) are increasingly used to simulate how specific users respond to a given context, enabling more user-centric applications that rely on user feedback. However, existing user simulators mostly imitate surface-level patterns and language styles, which fail to reflect the underlying states of real users (e.g., beliefs and emotions). To address these limitations, we propose a novel training framework, HumanLM, which builds user simulators that accurately reflect real users. Our key insight is that, in addition to generating responses, the model should generate natural-language latent states that align with ground-truth responses through reinforcement learning. These latent states correspond to a set of psychologically grounded state dimensions that drive how real users respond. HumanLM further synthesizes these aligned latent states into responses that accurately represent real users. For extensive evaluation, we develop Humanual, a comprehensive benchmark for simulating real users based on public data. Humanual consists of six large-scale datasets with 26k users and 216k responses in total, spanning diverse tasks such as generating user responses to daily life issues, political blogs, and chat sessions with LLM assistants. Across datasets, HumanLM significantly outperforms alternative approaches, achieving an average relative improvement of 16.3% in alignment scores from an LLM judge. In a real-time simulation study with 111 participants, HumanLM achieves the highest similarity to real user responses and competitive human-likeness scores.

Method: Two-step training-free inference: 1) generate unconstrained draft for semantic planning, 2) apply constrained decoding conditioned on the draft to guarantee structural validity.

Result: Improves strict structured accuracy by up to +24 percentage points over standard constrained decoding, enables smaller models to match/exceed larger constrained baselines.

Conclusion: DCCD reduces the “projection tax” of hard constraints and improves parameter efficiency for structured output generation tasks.

Abstract: Large language models (LLMs) are increasingly used to generate executable outputs, JSON objects, and API calls, where a single syntax error can make the output unusable. Constrained decoding enforces validity token-by-token via masking and renormalization, but it can distort generation when the model assigns low probability mass to valid continuations, pushing decoding toward locally valid yet semantically incorrect trajectories. We propose \emph{Draft-Conditioned Constrained Decoding (DCCD)}, a simple two-step, training-free inference procedure that decouples semantic planning from structural enforcement: an unconstrained draft is generated first, and constrained decoding is then applied, conditioned on this draft, to guarantee validity. We analyze DCCD through a KL-projection view, showing that draft conditioning increases feasible mass and reduces the cumulative “projection tax” induced by hard constraints, with an optional best-of-$K$ draft selection. Across structured reasoning benchmarks, DCCD improves strict structured accuracy by up to +24 percentage points over standard constrained decoding (e.g., 15.2% to 39.0% on GSM8K with a 1B model), and enables smaller model pairs to match or exceed much larger constrained baselines, yielding substantial gains in parameter efficiency.

Method: Proposes SEVADE with Dynamic Agentive Reasoning Engine (DARE) using specialized agents grounded in linguistic theory to perform multifaceted text deconstruction and generate structured reasoning chains, plus a separate lightweight rationale adjudicator for final classification based solely on reasoning chains.

Result: Achieves state-of-the-art performance on four benchmark datasets with average improvements of 6.75% in Accuracy and 6.29% in Macro-F1 score.

Conclusion: The decoupled architecture effectively mitigates hallucination risk by separating complex reasoning from final judgment, demonstrating superior performance in sarcasm detection.

Abstract: Sarcasm detection is a crucial yet challenging Natural Language Processing task. Existing Large Language Model methods are often limited by single-perspective analysis, static reasoning pathways, and a susceptibility to hallucination when processing complex ironic rhetoric, which impacts their accuracy and reliability. To address these challenges, we propose SEVADE, a novel Self-Evolving multi-agent Analysis framework with Decoupled Evaluation for hallucination-resistant sarcasm detection. The core of our framework is a Dynamic Agentive Reasoning Engine (DARE), which utilizes a team of specialized agents grounded in linguistic theory to perform a multifaceted deconstruction of the text and generate a structured reasoning chain. Subsequently, a separate lightweight rationale adjudicator (RA) performs the final classification based solely on this reasoning chain. This decoupled architecture is designed to mitigate the risk of hallucination by separating complex reasoning from the final judgment. Extensive experiments on four benchmark datasets demonstrate that our framework achieves state-of-the-art performance, with average improvements of 6.75% in Accuracy and 6.29% in Macro-F1 score.

Method: Benchmark comparing plain JSON generation vs structured output JSON generation vs TOON one-shot in-context learning generation, with test cases for structural complexity and validation pipeline.

Result: TOON shows promising accuracy/token ratio for in-domain tasks, but prompt overhead reduces advantages in shorter contexts. Plain JSON generation has best accuracy, while constrained decoding has lowest token usage but worse accuracy.

Conclusion: TOON’s efficiency potential follows a non-linear curve, shining only beyond a point where cumulative syntax savings amortize the initial prompt overhead. Constrained decoding may outperform TOON for simple structures.

Abstract: Recently presented Token-Oriented Object Notation (TOON) aims to replace JSON as a serialization format for passing structured data to LLMs with significantly reduced token usage. While showing solid accuracy in LLM comprehension, there is a lack of tests against JSON generation. Though never present in training data, TOON syntax is simple enough to suggest one-shot in-context learning could support accurate generation. The inevitable prompt overhead can be an acceptable trade-off for shorter completions. To test this, we conducted a benchmark creating several test cases with regard to structural complexity, a validation pipeline, and comparing plain JSON generation vs structured output (via constrained decoding) JSON generation vs TOON one-shot in-context learning generation. JSON structured output was included to establish a minimum token budget baseline and to set a starting point for future experiments testing TOON constrained decoding inference enforcement. Key findings: TOON shows promising accuracy/token consumption ratio for in-domain generation tasks, though this advantage is often reduced by the “prompt tax” of instructional overhead in shorter contexts. Plain JSON generation shows the best one-shot and final accuracy, even compared with constrained decoding structured output, where the only significant advantage is the lowest token usage as a trade-off for slightly decreased accuracy overall and significant degradation for some models. Notably, for simple structures, this “lowest token usage” of constrained decoding outperformed even TOON, hinting that TOON enforcing via frameworks such as xgrammar may not yield the desired results. Furthermore, the results suggest a scaling hypothesis: TOON’s true efficiency potential likely follows a non-linear curve, shining only beyond a specific point where cumulative syntax savings amortize the initial prompt overhead.

Method: Developed TopicENA framework that merges BERTopic with ENA, replacing manual concept coding with automatically generated topics while maintaining ENA’s capacity for modeling structural associations among concepts. Conducted three analysis cases to evaluate topic granularity, inclusion thresholds, and scalability.

Result: Coarse-grained topics work better for large datasets, fine-grained topics for smaller datasets. Topic inclusion thresholds should be adjusted based on quality indicators. Successfully applied TopicENA to substantially larger datasets than previous ENA studies, demonstrating practical scalability.

Conclusion: TopicENA facilitates practical and interpretable ENA analysis at scale and offers concrete guidance for configuring topic-based ENA pipelines in large-scale text analysis, overcoming the scalability limitations of traditional manual coding approaches.

Abstract: Epistemic Network Analysis (ENA) is a method for investigating the relational structure of concepts in text by representing co-occurring concepts as networks. Traditional ENA, however, relies heavily on manual expert coding, which limits its scalability and real-world applicability to large text corpora. Topic modeling provides an automated approach to extracting concept-level representations from text and can serve as an alternative to manual coding. To tackle this limitation, the present study merges BERTopic with ENA and introduces TopicENA, a topic-based epistemic network analysis framework. TopicENA substitutes manual concept coding with automatically generated topics while maintaining ENA’s capacity for modeling structural associations among concepts. To explain the impact of modeling choices on TopicENA outcomes, three analysis cases are presented. The first case assesses the effect of topic granularity, indicating that coarse-grained topics are preferable for large datasets, whereas fine-grained topics are more effective for smaller datasets. The second case examines topic inclusion thresholds and finds that threshold values should be adjusted according to topic quality indicators to balance network consistency and interpretability. The third case tests TopicENA’s scalability by applying it to a substantially larger dataset than those used in previous ENA studies. Collectively, these cases illustrate that TopicENA facilitates practical and interpretable ENA analysis at scale and offers concrete guidance for configuring topic-based ENA pipelines in large-scale text analysis.

Method: Introduces History-Echoes framework with two perspectives: 1) Probabilistic: models conversations as Markov chains to quantify state consistency; 2) Geometric: measures consistency of consecutive hidden representations. Analyzes three model families across six datasets spanning diverse phenomena.

Result: Reveals strong correlation between probabilistic and geometric perspectives. Demonstrates that behavioral persistence manifests as a geometric trap, where gaps in the latent space confine the model’s trajectory.

Conclusion: Conversational history significantly biases LLM responses, with both probabilistic and geometric analyses converging on the same findings. The geometric trap phenomenon explains how models get confined by their own conversational patterns.

Abstract: How does the conversational past of large language models (LLMs) influence their future performance? Recent work suggests that LLMs are affected by their conversational history in unexpected ways. For instance, hallucinations in prior interactions may influence subsequent model responses. In this work, we introduce History-Echoes, a framework that investigates how conversational history biases subsequent generations. The framework explores this bias from two perspectives: probabilistically, we model conversations as Markov chains to quantify state consistency; geometrically, we measure the consistency of consecutive hidden representations. Across three model families and six datasets spanning diverse phenomena, our analysis reveals a strong correlation between the two perspectives. By bridging these perspectives, we demonstrate that behavioral persistence manifests as a geometric trap, where gaps in the latent space confine the model’s trajectory. Code available at https://github.com/technion-cs-nlp/OldHabitsDieHard.

Method: Proposes a hybrid framework with six integrated components: 1) semantic metadata enhancement using LLMs, 2) dynamic knowledge graph construction, 3) VARK-based cognitive profiling, 4) mental state estimation, 5) graph-enhanced retrieval with LLM-powered ranking, and 6) adaptive interface design with iterative learning.

Result: Experimental validation on MovieLens-1M dataset demonstrates the system’s capacity for personalized recommendation generation despite limited initial information. The framework successfully generates explainable recommendations from initial user contact.

Conclusion: This work establishes groundwork for cognitively-aware recommendation systems capable of overcoming cold start limitations through semantic comprehension and psychological modeling, offering personalized, explainable recommendations from initial user contact.

Abstract: Cold start scenarios present fundamental obstacles to effective recommendation generation, particularly when dealing with users lacking interaction history or items with sparse metadata. This research proposes an innovative hybrid framework that leverages Large Language Models (LLMs) for content semantic analysis and knowledge graph development, integrated with cognitive profiling based on VARK (Visual, Auditory, Reading/Writing, Kinesthetic) learning preferences. The proposed system tackles multiple cold start dimensions: enriching inadequate item descriptions through LLM processing, generating user profiles from minimal data, and dynamically adjusting presentation formats based on cognitive assessment. The framework comprises six integrated components: semantic metadata enhancement, dynamic graph construction, VARK-based profiling, mental state estimation, graph-enhanced retrieval with LLM-powered ranking, and adaptive interface design with iterative learning. Experimental validation on MovieLens-1M dataset demonstrates the system’s capacity for personalized recommendation generation despite limited initial information. This work establishes groundwork for cognitively-aware recommendation systems capable of overcoming cold start limitations through semantic comprehension and psychological modeling, offering personalized, explainable recommendations from initial user contact.

Method: Introduces entropic-time inference with self-organizing architecture that couples: 1) entropy-aware scheduling, 2) entropic pruning of paged attention blocks, and 3) adaptive temperature control. Extends vLLM with these entropy-based mechanisms to transform inference into a thermodynamic process.

Result: Presents concrete systems design, pseudocode, and integration plan showing how entropy can serve as a first-class control signal for scalable LLM inference, allocating computation where uncertainty reduction is maximized.

Conclusion: Entropy should be a fundamental control signal in LLM inference, enabling resource-intelligent generation that focuses computational effort on high-uncertainty regions rather than treating all tokens equally.

Abstract: Modern large language model (LLM) inference engines optimize throughput and latency under fixed decoding rules, treating generation as a linear progression in token time. We propose a fundamentally different paradigm: entropic-time inference, where decoding is governed by the flow of uncertainty rather than token index. We introduce a self-organizing inference architecture that jointly couples scheduling, attention sparsification, and sampling temperature under a unified entropy control objective. Our method extends vLLM with entropy-aware scheduling, entropic pruning of paged attention blocks, and adaptive temperature control that stabilizes generation near a target entropy regime. This transforms inference into a resource-intelligent thermodynamic process that allocates computation where uncertainty reduction is maximized. We present a concrete systems design, pseudocode, and integration plan, demonstrating how entropy can serve as a first-class control signal for scalable LLM inference.

Method: Combined hand-authored grammatical rules, large central dictionary with syntactic/semantic constraints, chart-based parsing with weighted interpretation scoring to manage structural ambiguity.

Result: Successfully deployed commercial MT system that evolved continuously over decades under real-world usage pressures, with preserved software and linguistic resources available for study.

Conclusion: Documents a historically significant rule-based MT system that operated commercially for over two decades, providing valuable insights into practical MT development and maintenance challenges.

Abstract: This paper documents the architecture, development practices, and preserved artifacts of the Logovista English–Japanese machine translation system, a large, explicitly rule-based MT system that was developed and sold commercially from the early 1990s through at least 2012. The system combined hand-authored grammatical rules, a large central dictionary encoding syntactic and semantic constraints, and chart-based parsing with weighted interpretation scoring to manage extensive structural ambiguity. The account emphasizes how the system was extended and maintained under real-world usage pressures, including regression control, ambiguity management, and the limits encountered as coverage expanded. Unlike many rule-based MT systems described primarily in research settings, Logovista was deployed for decades and evolved continuously in response to practical requirements. The paper is intended as a technical and historical record rather than an argument for reviving rule-based MT, and describes the software and linguistic resources that have been preserved for potential future study.

Method: The study uses Representational Similarity Analysis (RSA) and dimension-specific correlation metrics to systematically investigate whether fine-tuning can steer LLM internal representations toward more embodied, grounded patterns. The research examines how sensorimotor improvements generalize across languages and related sensory-motor dimensions, and tests transferability across different task formats.

Result: The results demonstrate that internal representations of LLMs can be steered toward more embodied, grounded patterns through fine-tuning. Sensorimotor improvements generalize robustly across languages and related sensory-motor dimensions, but are highly sensitive to the learning objective and fail to transfer across two disparate task formats.

Conclusion: Task-specific fine-tuning can effectively bridge the embodiment gap in LLMs by aligning their representations with human sensorimotor experiences, though this alignment is task-specific and doesn’t transfer well across fundamentally different learning objectives.

Abstract: Large Language Models (LLMs) exhibit a significant “embodiment gap”, where their text-based representations fail to align with human sensorimotor experiences. This study systematically investigates whether and how task-specific fine-tuning can bridge this gap. Utilizing Representational Similarity Analysis (RSA) and dimension-specific correlation metrics, we demonstrate that the internal representations of LLMs can be steered toward more embodied, grounded patterns through fine-tuning. Furthermore, the results show that while sensorimotor improvements generalize robustly across languages and related sensory-motor dimensions, they are highly sensitive to the learning objective, failing to transfer across two disparate task formats.

Method: Proposes Contrastive Learning-based Inverse Direct Preference Optimization (CoIPO) that minimizes discrepancy between label-aligned logits from clean and noisy prompts. Uses mutual information theory analysis and constructs paired clean-noisy prompts from FLAN dataset for training.

Result: CoIPO achieves significant improvement in average accuracy on NoisyPromptBench benchmark compared to state-of-the-art approaches, demonstrating enhanced robustness to prompt variations.

Conclusion: CoIPO effectively improves LLM robustness to noisy prompts without external tools, addressing a key limitation in real-world LLM deployment where user prompts often contain imperfections.

Abstract: Large language models (LLMs) have demonstrated remarkable and steadily improving performance across a wide range of tasks. However, LLM performance may be highly sensitive to prompt variations especially in scenarios with limited openness or strict output formatting requirements, indicating insufficient robustness. In real-world applications, user prompts provided to LLMs often contain imperfections, which may undermine the quality of the model’s responses. To address this issue, previous work has primarily focused on preprocessing prompts, employing external tools or even LLMs to refine prompt formulations in advance. However, these approaches overlook the intrinsic robustness of LLMs, and their reliance on external components introduces additional computational overhead and uncertainty. In this work, we propose a Contrastive Learning-based Inverse Direct Preference Optimization (CoIPO) method that minimizes the discrepancy between the label-aligned logits produced by the model under a clean prompt and its noisy counterpart, and conduct a detailed analysis using mutual information theory. We augment the FLAN dataset by constructing paired prompts, each consisting of a clean prompt and its corresponding noisy version for training. Additionally, to evaluate the effectiveness, we develop NoisyPromptBench, a benchmark enhanced and derived from the existing PromptBench. Experimental results conducted on NoisyPromptBench demonstrate that our proposed method achieves a significant improvement in average accuracy over the current state-of-the-art approaches. The source code of CoIPO, pair-wise FLAN datasets, and NoisyPromptBench have already been released on https://github.com/vegetable-yx/CoIPO.

Method: 1) Developed a semantic framework identifying 10 dimensions for meme understanding; 2) Created M-QUEST benchmark with 609 QA pairs for 307 memes through semi-automatic process; 3) Evaluated 8 open-source LLMs on their ability to solve M-QUEST questions.

Result: Models with instruction tuning and reasoning capabilities significantly outperform others, but pragmatic inference questions remain challenging. Performance varies by dimension and model architecture.

Conclusion: The framework and benchmark advance multimodal content safety research by providing structured approach to meme toxicity assessment and revealing current limitations in LLMs’ commonsense reasoning for multimodal content.

Abstract: Internet memes are a powerful form of online communication, yet their nature and reliance on commonsense knowledge make toxicity detection challenging. Identifying key features for meme interpretation and understanding, is a crucial task. Previous work has been focused on some elements contributing to the meaning, such as the Textual dimension via OCR, the Visual dimension via object recognition, upper layers of meaning like the Emotional dimension, Toxicity detection via proxy variables, such as hate speech detection, and sentiment analysis. Nevertheless, there is still a lack of an overall architecture able to formally identify elements contributing to the meaning of a meme, and be used in the sense-making process. In this work, we present a semantic framework and a corresponding benchmark for automatic knowledge extraction from memes. First, we identify the necessary dimensions to understand and interpret a meme: Textual material, Visual material, Scene, Background Knowledge, Emotion, Semiotic Projection, Analogical Mapping, Overall Intent, Target Community, and Toxicity Assessment. Second, the framework guides a semi-automatic process of generating a benchmark with commonsense question-answer pairs about meme toxicity assessment and its underlying reason. The resulting benchmark M-QUEST consists of 609 question-answer pairs for 307 memes. Thirdly, we evaluate eight open-source large language models on their ability to correctly solve M-QUEST. Our results show that current models’ commonsense reasoning capabilities for toxic meme interpretation vary depending on the dimension and architecture. Models with instruction tuning and reasoning capabilities significantly outperform the others, though pragmatic inference questions remain challenging. We release code, benchmark, and prompts to support future research intersecting multimodal content safety and commonsense reasoning.

Method: Used Google Mediapipe for input feature extraction to capture spatial information of signs, processed spatial features with a Multilayer Perceptron architecture, and temporal information with a Gated Recurrent Unit. Compared transfer learning performance between two sign language pairs: Chinese to Arabic and Greek to Flemish, focusing on iconic signs.

Result: Experimental results showed transfer learning improvements: 7.02% improvement for Arabic when transferring from Chinese, and 1.07% improvement for Flemish when transferring from Greek. This demonstrates that cross-linguistic similarities in iconic signs can enable effective knowledge transfer between sign languages.

Conclusion: Transfer learning between sign languages is effective, particularly for iconic signs with cross-linguistic similarities. The approach using Mediapipe feature extraction with MLP and GRU architectures shows promising results for improving sign language recognition performance through cross-lingual transfer.

Abstract: Most sign language recognition research relies on Transfer Learning (TL) from vision-based datasets such as ImageNet. Some extend this to alternatively available language datasets, often focusing on signs with cross-linguistic similarities. This body of work examines the necessity of these likenesses on effective knowledge transfer by comparing TL performance between iconic signs of two different sign language pairs: Chinese to Arabic and Greek to Flemish. Google Mediapipe was utilised as an input feature extractor, enabling spatial information of these signs to be processed with a Multilayer Perceptron architecture and the temporal information with a Gated Recurrent Unit. Experimental results showed a 7.02% improvement for Arabic and 1.07% for Flemish when conducting iconic TL from Chinese and Greek respectively.

Method: Retcon is a few-shot prompting technique designed specifically for turn-level control in conversations. The paper presents this method as a solution to provide granular control over LLM behavior during multi-turn interactions.

Result: Retcon performs significantly better than both zero-shot prompting and traditional few-shot prompting methods for controlling LLMs in conversational contexts.

Conclusion: Retcon provides an effective approach for turn-level control of LLMs in multi-turn conversations, addressing a key challenge in conversational AI applications.

Abstract: Recent advances in Large Language Models (LLMs) allow agents to execute complex natural language tasks. Many LLM applications, such as support agents, teaching assistants, and interactive bots, involve multi-turn conversations. However, it remains challenging to control LLMs in the context of such interactions, particularly when the LLM behavior needs to be adjustable over the course of the conversation. In this paper, we present Retcon, a few-shot prompting technique designed to provide turn-level control over LLMs in conversations. We then demonstrate that it performs significantly better than zero-shot and traditional few-shot prompting.

Method: Proposed a classical quantum-inspired self-attention (QISA) mechanism and integrated it into the full GPT-1 architecture for autoregressive language modeling, comparing it against standard self-attention mechanisms.

Result: QISA achieved significantly better performance: 15.5× better character error rate, 4.7× better word error rate, and 13× better cross-entropy loss compared to standard self-attention, with only 2.6× longer inference time.

Conclusion: Quantum-inspired self-attention mechanisms show promising performance improvements for language modeling tasks and warrant further exploration in NLP architectures.

Abstract: Recent advances in Natural Language Processing have been predominantly driven by transformer-based architectures, which rely heavily on self-attention mechanisms to model relationships between tokens in a sequence. Similarly, the field of Quantum Natural Language Processing, which seeks to leverage quantum principles to address challenges in language understanding and generation tasks, has seen the recent development of quantum self-attention mechanisms. We propose a classical quantum-inspired self-attention (QISA) mechanism and integrate it into the full autoregressive language modeling pipeline of GPT-1. To the best of our knowledge, this is the first integration of this kind, as previous quantum self-attention mechanisms have been primarily tested on text classification. In our experiments, QISA achieves better performance when compared to standard self-attention on the metrics character error rate ($15.5\times$ better), word error rate ($4.7 \times $) and cross-entropy loss ($13 \times$). This is achieved while only requiring a $ 2.6\times$ longer inference time.

Method: The paper studies several embedding-level concept extraction methods for analyzing LLM judge behavior, comparing them in terms of interpretability and predictiveness. It uses sparse autoencoder-based approaches and analyzes over 27k paired responses from multiple human preference datasets with judgments from three LLMs.

Result: Sparse autoencoder-based approaches recover substantially more interpretable preference features than alternatives while remaining competitive in predicting LLM decisions. The analysis validates existing results (like LLMs preferring refusal of sensitive requests) and uncovers new trends including biases toward responses emphasizing concreteness, empathy, detail, formality, and against legal guidance promoting active steps.

Conclusion: Automated concept discovery enables systematic analysis of LLM judge preferences without predefined bias taxonomies, providing a framework for understanding and potentially mitigating systematic biases in LLM-as-a-judge evaluations.

Abstract: Large Language Models (LLMs) are increasingly used as scalable evaluators of model outputs, but their preference judgments exhibit systematic biases and can diverge from human evaluations. Prior work on LLM-as-a-judge has largely focused on a small, predefined set of hypothesized biases, leaving open the problem of automatically discovering unknown drivers of LLM preferences. We address this gap by studying several embedding-level concept extraction methods for analyzing LLM judge behavior. We compare these methods in terms of interpretability and predictiveness, finding that sparse autoencoder-based approaches recover substantially more interpretable preference features than alternatives while remaining competitive in predicting LLM decisions. Using over 27k paired responses from multiple human preference datasets and judgments from three LLMs, we analyze LLM judgments and compare them to those of human annotators. Our method both validates existing results, such as the tendency for LLMs to prefer refusal of sensitive requests at higher rates than humans, and uncovers new trends across both general and domain-specific datasets, including biases toward responses that emphasize concreteness and empathy in approaching new situations, toward detail and formality in academic advice, and against legal guidance that promotes active steps like calling police and filing lawsuits. Our results show that automated concept discovery enables systematic analysis of LLM judge preferences without predefined bias taxonomies.

Method: Proposes a neurosymbolic approach grounded in social science theory and abductive reasoning. Automatically extracts rules to abduce specific story elements needed to guide LLMs through consistent and targeted narrative transformations.

Result: Outperforms zero-shot LLM baseline by 55.88% for collectivistic to individualistic narrative shift with GPT-4o, while maintaining superior semantic similarity (40.4% improvement in KL divergence). Comparable improvements for individualistic to collectivistic transformation. Similar performance across Llama-4, Grok-4, and competitive performance for Deepseek-R1.

Conclusion: The neurosymbolic abduction-guided approach effectively enables LLMs to perform narrative transformations while preserving original message fidelity, addressing a significant limitation in current language models.

Abstract: Effective communication often relies on aligning a message with an audience’s narrative and worldview. Narrative shift involves transforming text to reflect a different narrative framework while preserving its original core message–a task we demonstrate is significantly challenging for current Large Language Models (LLMs). To address this, we propose a neurosymbolic approach grounded in social science theory and abductive reasoning. Our method automatically extracts rules to abduce the specific story elements needed to guide an LLM through a consistent and targeted narrative transformation. Across multiple LLMs, abduction-guided transformed stories shifted the narrative while maintaining the fidelity with the original story. For example, with GPT-4o we outperform the zero-shot LLM baseline by 55.88% for collectivistic to individualistic narrative shift while maintaining superior semantic similarity with the original stories (40.4% improvement in KL divergence). For individualistic to collectivistic transformation, we achieve comparable improvements. We show similar performance across both directions for Llama-4, and Grok-4 and competitive performance for Deepseek-R1.

Method: Uses dual LLM agents in a type-theoretic framework to decompose instructions into typed predicates with formal atomicity and independence constraints, then applies type-specific satisfaction semantics (semantic equivalence for content, exact precision for numerical predicates) and extends to multi-turn dialogues through history-aware satisfaction functions.

Result: Achieves 90.38% accuracy with 26.45% error reduction over baselines, demonstrates substantially stronger correlation with human judgment for complex instructions, and enables evaluation in conversational contexts where single-turn methods fail.

Conclusion: DIALEVAL provides an effective automated framework for evaluating instruction following in LLMs that better aligns with human judgment patterns and scales to conversational contexts.

Abstract: Evaluating instruction following in Large Language Models requires decomposing instructions into verifiable requirements and assessing satisfaction–tasks currently dependent on manual annotation and uniform criteria that do not align with human judgment patterns. We present DIALEVAL, a type-theoretic framework using dual LLM agents to automate instruction decomposition into typed predicates and implement type-specific satisfaction semantics. The framework enforces formal atomicity and independence constraints during automated extraction, then applies differentiated evaluation criteria–semantic equivalence for content predicates, exact precision for numerical predicates–mirroring empirically observed human assessment patterns. Extended to multi-turn dialogues through history-aware satisfaction functions, DIALEVAL enables evaluation in conversational contexts where single-turn methods fail. Validation demonstrates 90.38% accuracy (26.45% error reduction over baselines) and substantially stronger correlation with human judgment for complex instructions.

Method: Three-stage pipeline: (1) data acquisition of authoritative paper abstracts, (2) QA extraction using LLMs to synthesize scientific hypothesis questions and discovery answers, (3) QA filtering based on relevance, clarity, and centrality. Applied to create monthly-updated benchmark covering 12 biomedical sub-domains.

Result: Extensive evaluations of state-of-the-art models reveal current limitations in discovering new knowledge. The benchmark provides the first dynamic, automatic framework for assessing new knowledge discovery capabilities.

Conclusion: DBench-Bio establishes a living, evolving resource for the AI research community to catalyze the development of knowledge discovery capabilities in AI systems, addressing critical evaluation challenges.

Abstract: Recent advancements in Large Language Model (LLM) agents have demonstrated remarkable potential in automatic knowledge discovery. However, rigorously evaluating an AI’s capacity for knowledge discovery remains a critical challenge. Existing benchmarks predominantly rely on static datasets, leading to inevitable data contamination where models have likely seen the evaluation knowledge during training. Furthermore, the rapid release cycles of modern LLMs render static benchmarks quickly outdated, failing to assess the ability to discover truly new knowledge. To address these limitations, we propose DBench-Bio, a dynamic and fully automated benchmark designed to evaluate AI’s biological knowledge discovery ability. DBench-Bio employs a three-stage pipeline: (1) data acquisition of rigorous, authoritative paper abstracts; (2) QA extraction utilizing LLMs to synthesize scientific hypothesis questions and corresponding discovery answers; and (3) QA filter to ensure quality based on relevance, clarity, and centrality. We instantiate this pipeline to construct a monthly-updated benchmark covering 12 biomedical sub-domains. Extensive evaluations of SOTA models reveal current limitations in discovering new knowledge. Our work provides the first dynamic, automatic framework for assessing the new knowledge discovery capabilities of AI systems, establishing a living, evolving resource for AI research community to catalyze the development of knowledge discovery.

Method: Introduces DCR (Discernment via Contrastive Refinement), a preceding alignment stage that uses contrastive refinement to improve LLMs’ capacity to distinguish truly toxic prompts from superficially toxic ones. The method theoretically and empirically demonstrates improved discernment capabilities.

Result: Evaluation across diverse benchmarks shows the method effectively reduces over-refusal while preserving safety benefits of alignment. Achieves this with minimal degradation of general capabilities, offering a more principled and robust direction for safety alignment.

Conclusion: DCR provides an effective approach to address the over-refusal problem in safety-aligned LLMs through contrastive refinement, balancing safety and helpfulness without significant capability degradation.

Abstract: Large language models (LLMs) aligned for safety often suffer from over-refusal, the tendency to reject seemingly toxic or benign prompts by misclassifying them as toxic. This behavior undermines models’ helpfulness and restricts usability in sensitive or nuanced contexts. While prior work has proposed mitigation strategies such as data augmentation and activation steering, these approaches often face a trade-off: reducing over-refusal typically degrades the model’s ability to reject genuinely harmful content. We argue that this issue arises from the ambiguous influence of toxic and seemingly toxic prompts on the model’s learning dynamics. To address it, we introduce a preceding alignment stage, DCR: Discernment via Contrastive Refinement. Both theoretically and empirically, we demonstrate that contrastive refinement improves an LLM’s capacity to distinguish truly toxic prompts from superficially toxic ones. Evaluation across diverse benchmarks shows that our method effectively reduces over-refusal while preserving the safety benefits of alignment. Importantly, it achieves this with minimal degradation of general capabilities, offering a more principled and robust direction for safety alignment.

Method: The approach uses representation engineering to test the hypothesis that stylistic attributes form linear directions in activation space. Based on this finding, the paper presents a lightweight, training-free method for precise style control that supports linear style composition and can ablate undesirable behaviors.

Result: Strong empirical evidence supports the hypothesis across a wide range of styles. The method achieves high style adherence while preserving core capabilities at minimal computational cost, as confirmed by experiments on over a dozen models.

Conclusion: Stylistic attributes in LLMs are encoded as linear directions in activation space, enabling effective training-free style control through representation engineering. This approach offers precise control, supports style composition, and enhances safety at low computational cost.

Abstract: Controlling stylistic attributes in large language models (LLMs) remains challenging, with existing approaches relying on either prompt engineering or post-training alignment. This paper investigates this challenge through the lens of representation engineering, testing the hypothesis that distinct stylistic attributes - from emotional tone to linguistic structure - are encoded as linear directions in the model’s activation space. We provide strong empirical evidence for this hypothesis across a wide range of styles and, based on this finding, present a lightweight, training-free method for precise style control. Our approach supports linear style composition, enhances safety by ablating undesirable behaviors, and, as confirmed by experiments on over a dozen models, achieves high style adherence while preserving core capabilities at minimal computational cost.

Method: Proposes IntPro with intent explanations that abstract how contextual signals connect to expressed intents, stored in individual intent history libraries. Trained through supervised fine-tuning on retrieval-conditioned trajectories and multi-turn Group Relative Policy Optimization with tool-aware reward functions.

Result: Experiments across three diverse scenarios (Highlight-Intent, MIntRec2.0, and Weibo Post-Sync) demonstrate strong intent understanding performance with effective context-aware reasoning capabilities across different scenarios and model types.

Conclusion: IntPro effectively addresses the gap in context-aware intent understanding by leveraging historical intent patterns and learning when to retrieve versus directly infer, improving human-AI collaboration workflows.

Abstract: Large language models (LLMs) have become integral to modern Human-AI collaboration workflows, where accurately understanding user intent serves as a crucial step for generating satisfactory responses. Context-aware intent understanding, which involves inferring user intentions from situational environments, is inherently challenging because it requires reasoning over both the immediate context and the user’s underlying motivations that drive their behavior. Moreover, existing approaches often treat intent understanding as a static recognition task, overlooking users’ accumulated intent patterns that could provide valuable references for more accurate and generalizable understanding. To address this gap, we propose IntPro, a proxy agent that learns to adapt to individual users via retrieval-conditioned intent inference. We design intent explanations that abstract how contextual signals connect to expressed intents, and store them in an individual intent history library for retrieval. We train IntPro through supervised fine-tuning on retrieval-conditioned trajectories and multi-turn Group Relative Policy Optimization (GRPO) with tool-aware reward functions, enabling the agent to learn when to leverage historical intent patterns and when to infer directly. Experiments across three diverse scenarios (Highlight-Intent, MIntRec2.0, and Weibo Post-Sync) demonstrate that IntPro achieves strong intent understanding performance with effective context-aware reasoning capabilities across different scenarios and model types.

Method: Proposes Sequential Adaptive Steering (SAS) that orthogonalizes steering vectors by training subsequent probes on the residual stream shifted by prior interventions, creating reusable primitives that allow users to adjust coefficients alpha to synthesize complex personality profiles.

Result: The framework outperforms naive baselines in both goal adherence and coherence on Big Five personality traits, enabling precise, holistic personality modulation without updating model parameters.

Conclusion: SAS provides a parameter-efficient alternative to monolithic SFT/RLHF for persona alignment, enabling continuous multi-dimensional personality control through reusable steering primitives.

Abstract: Aligning Large Language Models (LLMs) with specific personas typically relies on expensive and monolithic Supervised Fine-Tuning (SFT) or RLHF. While effective, these methods require training distinct models for every target personality profile. Inference-time activation steering offers a parameter-efficient alternative, yet naive approaches fail to control multiple traits simultaneously due to destructive vector interference. In this work, we propose a modular framework for continuous, multi-dimensional personality control. Our key innovation is Sequential Adaptive Steering (SAS): a method that orthogonalizes steering vectors by training subsequent probes on the residual stream shifted by prior interventions. This approach transforms steering vectors into reusable primitives, allowing users to instantly synthesize complex, high-fidelity personality profiles by simply adjusting coefficients alpha. We validate our framework on the Big Five personality traits, demonstrating that it outperforms naive baselines in both goal adherence and coherence, enabling precise, holistic personality modulation without updating model parameters.

Method: Created a multi-task, multi-label Chinese dialogue dataset that supports three tasks: satisfaction recognition, emotion recognition, and emotional state transition prediction. The dataset enables tracking emotional changes across dialogue turns.

Result: Developed a new Chinese dialogue dataset resource that provides comprehensive emotional and satisfaction analysis capabilities for multi-turn dialogues, addressing previous limitations in Chinese language resources.

Conclusion: The constructed dataset provides valuable resources for studying emotion and satisfaction in dialogue systems, enabling better tracking of dynamic emotional changes across multiple dialogue turns for improved satisfaction prediction.

Abstract: User satisfaction is closely related to enterprises, as it not only directly reflects users’ subjective evaluation of service quality or products, but also affects customer loyalty and long-term business revenue. Monitoring and understanding user emotions during interactions helps predict and improve satisfaction. However, relevant Chinese datasets are limited, and user emotions are dynamic; relying on single-turn dialogue cannot fully track emotional changes across multiple turns, which may affect satisfaction prediction. To address this, we constructed a multi-task, multi-label Chinese dialogue dataset that supports satisfaction recognition, as well as emotion recognition and emotional state transition prediction, providing new resources for studying emotion and satisfaction in dialogue systems.

Method: StructLens constructs maximum spanning trees based on semantic representations in residual streams (analogous to dependency parsing) and leverages tree properties to quantify inter-layer distance/similarity from a structural perspective.

Result: StructLens yields inter-layer similarity patterns distinct from conventional cosine similarity, and this structure-aware similarity proves beneficial for practical tasks like layer pruning.

Conclusion: Structural analysis is effective for understanding and optimizing language models, with StructLens providing a novel framework for revealing holistic internal structures.

Abstract: Language exhibits inherent structures, a property that explains both language acquisition and language change. Given this characteristic, we expect language models to manifest internal structures as well. While interpretability research has investigated the components of language models, existing approaches focus on local inter-token relationships within layers or modules (e.g., Multi-Head Attention), leaving global inter-layer relationships largely overlooked. To address this gap, we introduce StructLens, an analytical framework designed to reveal how internal structures relate holistically through their inter-token connection within a layer. StructLens constructs maximum spanning trees based on the semantic representations in residual streams, analogous to dependency parsing, and leverages the tree properties to quantify inter-layer distance (or similarity) from a structural perspective. Our findings demonstrate that StructLens yields an inter-layer similarity pattern that is distinctively different from conventional cosine similarity. Moreover, this structure-aware similarity proves to be beneficial for practical tasks, such as layer pruning, highlighting the effectiveness of structural analysis for understanding and optimizing language models. Our code is available at https://github.com/naist-nlp/structlens.

Method: Use Gemini-2.5-Flash to automatically synthesize code harnesses through iterative code refinement with feedback from game environments. The approach can generate entire policies in code, eliminating the need for LLM decision-making at runtime.

Result: The synthesized harness prevents all illegal moves in 145 TextArena games, enabling Gemini-2.5-Flash to outperform larger models like Gemini-2.5-Pro. Code policies achieve higher average reward than Gemini-2.5-Pro and GPT-5.2-High on 16 games.

Conclusion: Smaller models can outperform larger ones by synthesizing custom code harnesses or entire policies, offering both performance improvements and cost effectiveness for LLM-based agents.

Abstract: Despite significant strides in language models in the last few years, when used as agents, such models often try to perform actions that are not just suboptimal for a given state, but are strictly prohibited by the external environment. For example, in the recent Kaggle GameArena chess competition, 78% of Gemini-2.5-Flash losses were attributed to illegal moves. Often people manually write “harnesses” around LLMs to prevent such failures. In this paper, we demonstrate that Gemini-2.5-Flash can automatically synthesize such a code harness, using a small number of rounds of iterative code refinement given feedback from the (game) environment. The resulting harness prevents all illegal moves in 145 different TextArena games (both 1-player and 2-player), enabling the smaller Gemini-2.5-Flash model to outperform larger models, such as Gemini-2.5-Pro. Pushing our technique to the limit, we can get Gemini-2.5-Flash to generate the entire policy in code, thus eliminating the need to use the LLM at decision making time. The resulting code-policy receives a higher average reward than Gemini-2.5-Pro and GPT-5.2-High on 16 TextArena 1-player games. Our results show that using a smaller model to synthesize a custom code harness (or entire policy) can outperform a much larger model, while also being more cost effective.

Method: Developed the Certainty Robustness Benchmark using 200 reasoning/math questions from LiveBench, testing LLMs with two-turn interactions including uncertainty prompts (“Are you sure?”) and explicit contradiction (“You are wrong!”), plus numeric confidence elicitation.

Result: Found substantial differences in interactive reliability not explained by baseline accuracy alone: some models abandon correct answers under pressure, while others show strong resistance to challenge and better confidence-correctness alignment.

Conclusion: Certainty robustness is a distinct and critical dimension of LLM evaluation with important implications for alignment, trustworthiness, and real-world deployment.

Abstract: Large language models (LLMs) often present answers with high apparent confidence despite lacking an explicit mechanism for reasoning about certainty or truth. While existing benchmarks primarily evaluate single-turn accuracy, truthfulness or confidence calibration, they do not capture how models behave when their responses are challenged in interactive settings. We introduce the Certainty Robustness Benchmark, a two-turn evaluation framework that measures how LLMs balance stability and adaptability under self-challenging prompts such as uncertainty (“Are you sure?”) and explicit contradiction (“You are wrong!”), alongside numeric confidence elicitation. Using 200 reasoning and mathematics questions from LiveBench, we evaluate four state-of-the-art LLMs and distinguish between justified self-corrections and unjustified answer changes. Our results reveal substantial differences in interactive reliability that are not explained by baseline accuracy alone: some models abandon correct answers under conversational pressure, while others demonstrate strong resistance to challenge and better alignment between confidence and correctness. These findings identify certainty robustness as a distinct and critical dimension of LLM evaluation, with important implications for alignment, trustworthiness and real-world deployment.

Method: Created PulseLM dataset by aggregating PPG recordings from 15 public sources and harmonizing heterogeneous annotations into 12 common physiological QA tasks using a closed-ended question answering formulation.

Result: Dataset comprises 1.31 million standardized 10-second PPG segments associated with 3.15 million question-answer pairs. Established reproducible preprocessing, supervision, evaluation protocols and baseline benchmarks using multimodal PPG-aware LLMs.

Conclusion: PulseLM provides a standardized foundation for studying multimodal physiological reasoning, cross-dataset generalization, and scalable benchmarking of PPG-based language models.

Abstract: Photoplethysmography (PPG) is a widely used non-invasive sensing modality for continuous cardiovascular and physiological monitoring across clinical, laboratory, and wearable settings. While existing PPG datasets support a broad range of downstream tasks, they typically provide supervision in the form of numerical measurements or task-specific labels, limiting their suitability for language-based physiological reasoning and multimodal foundation models. In this work, we introduce PulseLM, a large-scale PPG-text dataset designed to bridge raw PPG waveforms and natural language through a unified, closed-ended question answering (QA) formulation. PulseLM aggregates PPG recordings from fifteen publicly available sources and harmonizes heterogeneous annotations into twelve common physiologically QA tasks. The dataset comprises 1.31 million standardized 10-second PPG segments, associated with 3.15 million question-answer pairs. We further define reproducible preprocessing, supervision, and evaluation protocols and establish baseline benchmarks using multimodal PPG-aware large language models. PulseLM provides a standardized foundation for studying multimodal physiological reasoning, cross-dataset generalization, and scalable benchmarking of PPG-based language models. The data and code can be found publicly available at: https://github.com/manhph2211/PulseLM.

Method: Empirical evaluation of 13 models (3B to 1.5T parameters) on mathematical reasoning tasks with 5 structured CoT perturbation types: MathError, UnitConversion, Sycophancy, SkippedSteps, and ExtraSteps, testing ability to complete tasks despite injected perturbations.

Result: Heterogeneous vulnerability patterns: MathError causes severe degradation in small models (50-60% accuracy loss) but scales well; UnitConversion remains challenging across all scales (20-30% loss); ExtraSteps minimal degradation (0-6%); Sycophancy modest effects (7% loss); SkippedSteps intermediate damage (15% loss). Scaling follows power-law patterns with size as protective factor.

Conclusion: LLM robustness to CoT perturbations varies significantly by perturbation type and model scale, highlighting need for task-specific robustness assessments and mitigation strategies for reliable multi-stage reasoning deployments.

Abstract: Chain-of-Thought (CoT) prompting has emerged as a foundational technique for eliciting reasoning from Large Language Models (LLMs), yet the robustness of this approach to corruptions in intermediate reasoning steps remains poorly understood. This paper presents a comprehensive empirical evaluation of LLM robustness to a structured taxonomy of 5 CoT perturbation types: \textit{MathError, UnitConversion, Sycophancy, SkippedSteps,} and \textit{ExtraSteps}. We evaluate 13 models spanning three orders of magnitude in parameter count (3B to 1.5T\footnote{Assumed parameter count of closed models}), testing their ability to complete mathematical reasoning tasks despite perturbations injected at different points in the reasoning chain. Our key findings reveal heterogeneous vulnerability patterns: MathError perturbations produce the most severe degradation in small models (50-60% accuracy loss) but show strong scaling benefits; UnitConversion remains challenging across all scales (20-30% loss even for largest models); ExtraSteps incur minimal accuracy degradation (0-6%) regardless of scale; Sycophancy produces modest effects (7% loss for small models); and SkippedSteps cause intermediate damage (15% loss). Scaling relationships follow power-law patterns, with model size serving as a protective factor against some perturbations but offering limited defense against dimensional reasoning tasks. These findings have direct implications for deploying LLMs in multi-stage reasoning pipelines and underscore the necessity of task-specific robustness assessments and mitigation strategies. The code and results are available \href{https://github.com/Mystic-Slice/CoTPerturbation}{here}.

Method: Uses Monte Carlo dropout applied exclusively to the LM head to generate multiple decoding paths, creating an empirical token distribution. Draft tokens are evaluated for consistency against this distribution, enabling adaptive control of decoding path size under appropriate dropout probabilities.

Result: Increases acceptance length while maintaining competitive task performance, achieving inference speedups of 1.09x to 1.33x over standard baselines, with up to additional 1.09x speedup when combined with EAGLE3.

Conclusion: DropMatch provides an effective, training-free approach to speculative decoding that can be orthogonally integrated with existing acceleration techniques, offering practical inference speed improvements without requiring model modifications.

Abstract: Speculative decoding accelerates large language model inference by proposing tokens with a lightweight draft model and selectively accepting them using a target model. This work introduces DropMatch, a novel approach that matches draft tokens to the predictive distribution of the target model via Monte Carlo dropout applied exclusively to the LM head, enabling sampling-based acceptance decisions. By generating multiple decoding paths, our method forms an empirical token distribution against which draft tokens are evaluated for consistency. This acceptance mechanism enables the model to adaptively control the size of decoding paths under an appropriate dropout probability, preventing substantial distortion of the target model predictive distribution. The proposed method operates in a training-free, data-free, and calibration-free manner, requires no architectural modification to pretrained models, and can be orthogonally integrated with a wide range of existing speculative decoding and inference acceleration techniques. Experiments across multiple benchmarks demonstrate that our approach increases acceptance length while maintaining competitive task performance, yielding inference speedups ranging from 1.09x to 1.33x over the standard baseline, and up to an additional 1.09x speedup when applied on top of EAGLE3.

Method: Created 1,500 originally authored multiple-choice questions by professors of graduate-level courses, covering five advanced topics. Used cross-LLM disagreement analysis followed by manual expert review to validate questions. Employed multiple-choice format for objective evaluation via lm_eval library.

Result: Baseline results with state-of-the-art LLMs show that advanced computational mathematical reasoning remains a significant challenge. The dataset is publicly released for reproducible evaluation.

Conclusion: CompMath-MCQ fills a gap in LLM evaluation by providing a benchmark for advanced computational mathematics, revealing current limitations of LLMs in this domain and enabling objective assessment of progress.

Abstract: The evaluation of Large Language Models (LLMs) on mathematical reasoning has largely focused on elementary problems, competition-style questions, or formal theorem proving, leaving graduate-level and computational mathematics relatively underexplored. We introduce CompMath-MCQ, a new benchmark dataset for assessing LLMs on advanced mathematical reasoning in a multiple-choice setting. The dataset consists of 1{,}500 originally authored questions by professors of graduate-level courses, covering topics including Linear Algebra, Numerical Optimization, Vector Calculus, Probability, and Python-based scientific computing. Three option choices are provided for each question, with exactly one of them being correct. To ensure the absence of data leakage, all questions are newly created and not sourced from existing materials. The validity of questions is verified through a procedure based on cross-LLM disagreement, followed by manual expert review. By adopting a multiple-choice format, our dataset enables objective, reproducible, and bias-free evaluation through lm_eval library. Baseline results with state-of-the-art LLMs indicate that advanced computational mathematical reasoning remains a significant challenge. We release CompMath-MCQ at the following link: https://github.com/biancaraimondi/CompMath-MCQ.git

Method: Used compressed sensing-based method with strategic knockouts to identify sparse subsets of attention heads responsible for specific capabilities. Zeroed out task-specific heads and measured performance degradation on targeted vs unrelated tasks across Llama and Qwen models (1B-8B parameters).

Result: Found that as few as 5 task-specific heads can degrade performance by up to 65% on targeted capabilities while preserving performance on unrelated tasks. Validated across diverse capabilities including mathematical abilities and code generation, revealing modular organization with sparse, functionally distinct components.

Conclusion: Capability localization is a general organizational principle of Transformer language models, with specialized functions implemented by sparse components. This has implications for interpretability, model editing, and AI safety.

Abstract: Large language models (LLMs) exhibit a wide range of capabilities, including mathematical reasoning, code generation, and linguistic behaviors. We show that many capabilities are highly localized to small subsets of attention heads within Transformer architectures. Zeroing out as few as five task-specific heads can degrade performance by up to $65%$ on standard benchmarks measuring the capability of interest, while largely preserving performance on unrelated tasks. We introduce a compressed sensing based method that exploits the sparsity of these heads to identify them via strategic knockouts and a small number of model evaluations. We validate these findings across Llama and Qwen models ranging from 1B to 8B parameters and a diverse set of capabilities including mathematical abilities and code generation, revealing a modular organization in which specialized capabilities are implemented by sparse, functionally distinct components. Overall, our results suggest that capability localization is a general organizational principle of Transformer language models, with implications for interpretability, model editing, and AI safety. Code is released at https://github.com/locuslab/llm-components.

Method: Uses a contextual Bradley-Terry-Luce model where latent utility depends on input prompts, conducts inference directly on induced rankings via simultaneous confidence intervals for pairwise utility differences, yielding statistically valid confidence sets for prompt-specific ranks.

Result: Empirical analysis with large-scale human preference data shows rankings vary substantially across prompt characteristics, many apparent rank differences aren’t statistically distinguishable, and uncertainty-aware rankings identify dominance only when supported by data.

Conclusion: Provides tools for robust ranking-based decision-making in LLM evaluation by connecting rank inference advances to contextual preference learning, enabling statistically valid uncertainty quantification for prompt-dependent rankings.

Abstract: Rankings derived from pairwise comparisons are central to many economic and computational systems. In the context of large language models (LLMs), rankings are typically constructed from human preference data and presented as leaderboards that guide deployment decisions. However, existing approaches rely on point estimates, implicitly treating rankings as fixed objects despite substantial estimation noise and context-dependent performance variation. Acting on such rankings can lead to misallocation and welfare loss when apparent differences are not statistically meaningful. We study prompt-dependent ranking inference under pairwise human preferences and develop a framework for decision-safe rankings with statistically valid uncertainty guarantees. We model preferences using a contextual Bradley-Terry-Luce model in which the latent utility of each model depends on the input prompt. Rather than targeting point estimates of utilities, we directly conduct inference on induced rankings, constructing confidence sets based on simultaneous confidence intervals for pairwise utility differences. This approach yields statistically valid marginal and simultaneous confidence sets for prompt-specific ranks. Our framework connects recent advances in rank inference to contextual preference learning and provides tools for robust ranking-based decision-making. Empirically, using large-scale human preference data from LLM evaluations, we show that rankings vary substantially across prompt characteristics and that many apparent rank differences are not statistically distinguishable. We further demonstrate how uncertainty-aware rankings identify dominance only when supported by the data and otherwise return partial orders.

Method: Used causal and probing-based interpretability methods on Llama-based biomedical language models. Applied activation patching to localize where drug-group information is stored across model layers and token positions, complemented with linear probes trained on token-level and sum-pooled activations.

Result: Early layers play a key role in encoding drug-group knowledge, with strongest causal effects from intermediate tokens within drug-group spans rather than final tokens. Pharmacological semantics are distributed across tokens and already present in embedding space - token-level probes perform near chance while sum-pooled representations achieve maximal accuracy.

Conclusion: Drug-group semantics in LLMs are not localized to single tokens but instead arise from distributed representations. This provides the first systematic mechanistic analysis of pharmacological knowledge in LLMs, offering insights into how biomedical semantics are encoded.

Abstract: Large language models (LLMs) have shown strong empirical performance across pharmacology and drug discovery tasks, yet the internal mechanisms by which they encode pharmacological knowledge remain poorly understood. In this work, we investigate how drug-group semantics are represented and retrieved within Llama-based biomedical language models using causal and probing-based interpretability methods. We apply activation patching to localize where drug-group information is stored across model layers and token positions, and complement this analysis with linear probes trained on token-level and sum-pooled activations. Our results demonstrate that early layers play a key role in encoding drug-group knowledge, with the strongest causal effects arising from intermediate tokens within the drug-group span rather than the final drug-group token. Linear probing further reveals that pharmacological semantics are distributed across tokens and are already present in the embedding space, with token-level probes performing near chance while sum-pooled representations achieve maximal accuracy. Together, these findings suggest that drug-group semantics in LLMs are not localized to single tokens but instead arise from distributed representations. This study provides the first systematic mechanistic analysis of pharmacological knowledge in LLMs, offering insights into how biomedical semantics are encoded in large language models.

Method: Researchers analyze last hidden states of LLMs across tasks with varying difficulty (harder reasoning questions, longer contexts, more answer choices). They quantify representation sparsity and demonstrate the sparsity-difficulty relation through controlled analyses with learning dynamic explanations. They then design Sparsity-Guided Curriculum In-Context Learning (SG-ICL) that uses representation sparsity to schedule few-shot demonstrations.

Result: The study reveals a consistent phenomenon: as task difficulty increases (greater OOD shift), LLM representations become substantially sparser. This sparsity-difficulty relation is observable across diverse models and domains. SG-ICL, leveraging this insight, leads to considerable performance enhancements in few-shot learning scenarios.

Conclusion: LLMs respond to unfamiliar or complex inputs by concentrating computation into specialized subspaces in the last hidden state, which serves as an adaptive mechanism for stabilizing reasoning under OOD conditions. The sparsity-difficulty relation provides new mechanistic insights into how LLMs internalize OOD challenges.

Abstract: In this work, we investigate how Large Language Models (LLMs) adapt their internal representations when encountering inputs of increasing difficulty, quantified as the degree of out-of-distribution (OOD) shift. We reveal a consistent and quantifiable phenomenon: as task difficulty increases, whether through harder reasoning questions, longer contexts, or adding answer choices, the last hidden states of LLMs become substantially sparser. In short, \textbf{\textit{the farther the shift, the sparser the representations}}. This sparsity–difficulty relation is observable across diverse models and domains, suggesting that language models respond to unfamiliar or complex inputs by concentrating computation into specialized subspaces in the last hidden state. Through a series of controlled analyses with a learning dynamic explanation, we demonstrate that this sparsity is not incidental but an adaptive mechanism for stabilizing reasoning under OOD. Leveraging this insight, we design \textit{Sparsity-Guided Curriculum In-Context Learning (SG-ICL)}, a strategy that explicitly uses representation sparsity to schedule few-shot demonstrations, leading to considerable performance enhancements. Our study provides new mechanistic insights into how LLMs internalize OOD challenges. The source code is available at the URL: https://github.com/MingyuJ666/sparsityLLM.

Method: Construct high-quality Hindi corpus (GigaLekh) using heuristic and LLM-as-a-judge filtering, augment with curated English data, then train 0.6B parameter model from scratch with various training recipes optimized for limited compute.

Result: LilMoo consistently outperforms comparably sized multilingual baselines (Qwen2.5-0.5B and Qwen3-0.6B) across comprehensive evaluation suites, showing language-specific pretraining can rival large multilingual models at sub-billion scale.

Conclusion: Well-designed language-specific pretraining with transparent pipelines can effectively address linguistic inequalities and compete with large multilingual models for low-resource languages like Hindi.

Abstract: The dominance of large multilingual foundation models has widened linguistic inequalities in Natural Language Processing (NLP), often leaving low-resource languages underrepresented. This paper introduces LilMoo, a 0.6-billion-parameter Hindi language model trained entirely from scratch to address this gap. Unlike prior Hindi models that rely on continual pretraining from opaque multilingual foundations, LilMoo is developed through a fully transparent and reproducible pipeline optimized for limited compute environments. We construct a high-quality Hindi corpus (GigaLekh) filtered through both heuristic and learned (LLM-as-a-judge) methods, complemented by bilingual augmentation with curated English data. Using this dataset, we explore various training recipes for small-scale language models. Across comprehensive evaluation suites, LilMoo consistently outperforms comparably sized multilingual baselines such as Qwen2.5-0.5B and Qwen3-0.6B, demonstrating that well-designed language-specific pretraining can rival large multilingual models at the sub-billion-parameter range.

Method: Proposes a Template-Based and Modular Cognitive model using a set-valued set function h: P(M)→P(M) to predict nonlinear dynamic mapping of lexical items onto morphological templates, with empirical observations from Riffian.

Result: Demonstrates how gender shifts and non-gender shifts arise during lexical changes, showing that variant markings emerge due to template shifts during word formation, and challenges conventional views of word formation.

Conclusion: The mathematical model provides a unified framework for understanding morphological markings across languages and contributes to deeper understanding of morphosyntactic variation with potential applications in linguistic pattern modeling.

Abstract: This study investigates the diverse characteristics of nouns, focusing on both semantic (e.g., countable/uncountable) and morphosyntactic (e.g., masculine/feminine) distinctions. We explore inter-word variations for gender markers in noun morphology. Grammatical gender shift is a widespread phenomenon in languages around the world. The aim is to uncover through a formal model the underlying patterns governing the variation of lexemes. To this end, we propose a new computational component dedicated to pairing items with morphological templates (e.g., the result of a generated item-template pair: (funas, ${N, +SG, -PL, -M, +F, -COL, +SING}$), with its spell-out form: $ð$a-funast ‘cow’). This process is formally represented by the Template-Based and Modular Cognitive model. This proposed model, defined by a set-valued set function $h : \mathscr{P}(M) \rightarrow \mathscr{P}(M)$, predicts the nonlinear dynamic mapping of lexical items onto morphological templates. By applying this formalism, we present a unified framework for understanding the complexities of morphological markings across languages. Through empirical observations, we demonstrate how these shifts, as well as non-gender shifts, arise during lexical changes, especially in Riffian. Our model posits that these variant markings emerge due to template shifts occurring during word and meaning’s formation. By formally demonstrating that conversion is applicable to noun-to-noun derivation, we challenge and broaden the conventional view of word formation. This mathematical model not only contributes to a deeper understanding of morphosyntactic variation but also offers potential applications in other fields requiring precise modelling of linguistic patterns.

Method: Proposes SafeCRS framework combining Safe Supervised Fine-Tuning (Safe-SFT) with Safe Group reward-Decoupled Normalization Policy Optimization (Safe-GDPO) to jointly optimize recommendation quality and personalized safety alignment. Also introduces SafeRec benchmark dataset for evaluation.

Result: Extensive experiments on SafeRec show SafeCRS reduces safety violation rates by up to 96.5% relative to strongest recommendation-quality baselines while maintaining competitive recommendation quality.

Conclusion: Personalized safety is critical for conversational recommender systems, and the proposed SafeCRS framework effectively addresses safety vulnerabilities while preserving recommendation performance.

Abstract: Current LLM-based conversational recommender systems (CRS) primarily optimize recommendation accuracy and user satisfaction. We identify an underexplored vulnerability in which recommendation outputs may negatively impact users by violating personalized safety constraints, when individualized safety sensitivities – such as trauma triggers, self-harm history, or phobias – are implicitly inferred from the conversation but not respected during recommendation. We formalize this challenge as personalized CRS safety and introduce SafeRec, a new benchmark dataset designed to systematically evaluate safety risks in LLM-based CRS under user-specific constraints. To further address this problem, we propose SafeCRS, a safety-aware training framework that integrates Safe Supervised Fine-Tuning (Safe-SFT) with Safe Group reward-Decoupled Normalization Policy Optimization (Safe-GDPO) to jointly optimize recommendation quality and personalized safety alignment. Extensive experiments on SafeRec demonstrate that SafeCRS reduces safety violation rates by up to 96.5% relative to the strongest recommendation-quality baseline while maintaining competitive recommendation quality. Warning: This paper contains potentially harmful and offensive content.

Method: Proposes RAG-X framework that evaluates retriever and generator independently across three QA tasks: information extraction, short-answer generation, and multiple-choice question answering. Introduces Context Utilization Efficiency (CUE) metrics to disaggregate system success into interpretable quadrants, isolating verified grounding from deceptive accuracy.

Result: Experiments reveal an “Accuracy Fallacy” where a 14% gap separates perceived system success from evidence-based grounding. The framework surfaces hidden failure modes and provides diagnostic transparency for clinical RAG systems.

Conclusion: RAG-X offers the diagnostic transparency needed for safe and verifiable clinical RAG systems by enabling targeted improvements through independent evaluation of retrieval and generation components.

Abstract: Automated question-answering (QA) systems increasingly rely on retrieval-augmented generation (RAG) to ground large language models (LLMs) in authoritative medical knowledge, ensuring clinical accuracy and patient safety in Artificial Intelligence (AI) applications for healthcare. Despite progress in RAG evaluation, current benchmarks focus only on simple multiple-choice QA tasks and employ metrics that poorly capture the semantic precision required for complex QA tasks. These approaches fail to diagnose whether an error stems from faulty retrieval or flawed generation, limiting developers from performing targeted improvement. To address this gap, we propose RAG-X, a diagnostic framework that evaluates the retriever and generator independently across a triad of QA tasks: information extraction, short-answer generation, and multiple-choice question (MCQ) answering. RAG-X introduces Context Utilization Efficiency (CUE) metrics to disaggregate system success into interpretable quadrants, isolating verified grounding from deceptive accuracy. Our experiments reveal an ``Accuracy Fallacy", where a 14% gap separates perceived system success from evidence-based grounding. By surfacing hidden failure modes, RAG-X offers the diagnostic transparency needed for safe and verifiable clinical RAG systems.

Method: Developed GigaVerbo-v2 dataset with synthetic extensions, created pretraining/continual pretraining recipes, and built comprehensive evaluation suite for Portuguese LLMs.

Result: Achieved state-of-the-art performance on several Portuguese-language modeling benchmarks with the Tucano 2 suite (Base, Instruct, and Think variants).

Conclusion: Successfully created reproducible, accessible Portuguese LLMs with comprehensive training recipes and evaluation tools for the community.

Abstract: We present Tucano 2, a fully open suite of large language models (LLMs) with 0.5-3.7 billion parameters, designed to address certain gaps in open-source development for Portuguese LLMs. Following our previous works, we now extend our dataset, GigaVerbo-v2, to a new degree of quality and scale, while also introducing a new synthetic dataset, GigaVerbo-v2 Synth, aimed at filling missing gaps in GigaVerbo-v2, and two post-training datasets, GigaVerbo-v2 SFT and GigaVerbo-v2 Preferences, that allow Portuguese LLMs to be trained in domains like retrieval augmented generation, coding, tool use, chain-of-thought reasoning, and many other domains of interest. Through extensive ablation studies, we design both pretraining and continual pretraining recipes for the Tucano 2 suite (Base, Instruct, and Think), which achieve state-of-the-art performance on several Portuguese-language modeling benchmarks. We also extend and refine the evaluation harness introduced in our earlier work, yielding a comprehensive evaluation suite that provides strong signals across different pretraining, continual pretraining, and post-training regimes. All artifacts associated with Tucano 2 are openly released, including training recipes, logs, and source code, ensuring that our work is reproducible, accessible, and extendable by the broader Portuguese NLP community.

Method: ByteFlow Net uses a hierarchical architecture with compression-driven segmentation based on the coding rate of latent representations. It employs Top-K selection to maintain a static computation graph while adapting segmentation boundaries. The method learns adaptive internal representation granularity based on input content rather than relying on pre-defined tokenization schemes.

Result: ByteFlow Net outperforms both BPE-based Transformers and previous byte-level architectures. The compression-based chunking strategy yields substantial performance gains, demonstrating that end-to-end, tokenizer-free modeling is not only feasible but more effective than traditional approaches.

Conclusion: End-to-end, tokenizer-free modeling is feasible and more effective than fixed tokenization schemes. ByteFlow Net’s adaptive segmentation approach opens a path toward more adaptive and information-grounded language models that can learn appropriate granularity from data.

Abstract: Modern language models still rely on fixed, pre-defined subword tokenizations. Once a tokenizer is trained, the LM can only operate at this fixed level of granularity, which often leads to brittle and counterintuitive behaviors even in otherwise strong reasoning models. We introduce \textbf{ByteFlow Net}, a new hierarchical architecture that removes tokenizers entirely and instead enables models to learn their own segmentation of raw byte streams into semantically meaningful units. ByteFlow Net performs compression-driven segmentation based on the coding rate of latent representations, yielding adaptive boundaries \emph{while preserving a static computation graph via Top-$K$ selection}. Unlike prior self-tokenizing methods that depend on brittle heuristics with human-designed inductive biases, ByteFlow Net adapts its internal representation granularity to the input itself. Experiments demonstrate that this compression-based chunking strategy yields substantial performance gains, with ByteFlow Net outperforming both BPE-based Transformers and previous byte-level architectures. These results suggest that end-to-end, tokenizer-free modeling is not only feasible but also more effective, opening a path toward more adaptive and information-grounded language models.

Method: Introduces BeliefSim, a simulation framework that constructs demographic belief profiles using psychology-informed taxonomies and survey priors. Studies prompt-based conditioning and post-training adaptation approaches, with evaluation using susceptibility accuracy and counterfactual demographic sensitivity metrics.

Result: Across datasets and modeling strategies, beliefs provide a strong prior for simulating misinformation susceptibility, with accuracy up to 92%.

Conclusion: Beliefs serve as an effective foundation for simulating demographic misinformation susceptibility in LLMs, demonstrating the potential for using belief-based approaches in human behavior simulation tasks.

Abstract: Misinformation is a growing societal threat, and susceptibility to misinformative claims varies across demographic groups due to differences in underlying beliefs. As Large Language Models (LLMs) are increasingly used to simulate human behaviors, we investigate whether they can simulate demographic misinformation susceptibility, treating beliefs as a primary driving factor. We introduce BeliefSim, a simulation framework that constructs demographic belief profiles using psychology-informed taxonomies and survey priors. We study prompt-based conditioning and post-training adaptation, and conduct a multi-fold evaluation using: (i) susceptibility accuracy and (ii) counterfactual demographic sensitivity. Across both datasets and modeling strategies, we show that beliefs provide a strong prior for simulating misinformation susceptibility, with accuracy up to 92%.

Method: LX Topic builds on FASTopic to ensure strong document representativeness, then integrates LLM refinement at the topic-word level using alignment and confidence-weighting mechanisms. This enhances semantic coherence without distorting document-topic distributions, unifying topic discovery, refinement, and standardized output in a web-based system.

Result: Evaluations on large-scale Amazon and Yelp review datasets show LX Topic achieves the highest overall topic quality relative to leading models while preserving clustering and classification performance.

Conclusion: LX Topic establishes topic modeling as a reproducible, interpretable, and measurement-oriented instrument for marketing research and practice by addressing key limitations of existing approaches.

Abstract: The growing use of unstructured text in business research makes topic modeling a central tool for constructing explanatory variables from reviews, social media, and open-ended survey responses, yet existing approaches function poorly as measurement instruments. Prior work shows that textual content predicts outcomes such as sales, satisfaction, and firm performance, but probabilistic models often generate conceptually diffuse topics, neural topic models are difficult to interpret in theory-driven settings, and large language model approaches lack standardization, stability, and alignment with document-level representations. We introduce LX Topic, a neural topic method that conceptualizes topics as latent linguistic constructs and produces calibrated document-level topic proportions for empirical analysis. LX Topic builds on FASTopic to ensure strong document representativeness and integrates large language model refinement at the topic-word level using alignment and confidence-weighting mechanisms that enhance semantic coherence without distorting document-topic distributions. Evaluations on large-scale Amazon and Yelp review datasets demonstrate that LX Topic achieves the highest overall topic quality relative to leading models while preserving clustering and classification performance. By unifying topic discovery, refinement, and standardized output in a web-based system, LX Topic establishes topic modeling as a reproducible, interpretable, and measurement-oriented instrument for marketing research and practice.

Method: Constructs hierarchical heterogeneous graphs at morpheme, part-of-speech, and named-entity levels, then integrates them with Semantics-aware Contrastive Learning (SemCon) to capture grammatical and semantic dependencies while establishing clearer decision boundaries.

Result: LIGRAM consistently outperforms existing baseline models on four Korean short-text datasets, demonstrating the effectiveness of language-specific graph representations with contrastive learning.

Conclusion: Integrating language-specific graph representations with contrastive learning provides an effective solution for short text classification in agglutinative languages like Korean, addressing both contextual limitations and linguistic characteristics.

Abstract: Short text classification (STC) remains a challenging task due to the scarcity of contextual information and labeled data. However, existing approaches have pre-dominantly focused on English because most benchmark datasets for the STC are primarily available in English. Consequently, existing methods seldom incorporate the linguistic and structural characteristics of Korean, such as its agglutinative morphology and flexible word order. To address these limitations, we propose LIGRAM, a hierarchical heterogeneous graph model for Korean short-text classification. The proposed model constructs sub-graphs at the morpheme, part-of-speech, and named-entity levels and hierarchically integrates them to compensate for the limited contextual information in short texts while precisely capturing the grammatical and semantic dependencies inherent in Korean. In addition, we apply Semantics-aware Contrastive Learning (SemCon) to reflect semantic similarity across documents, enabling the model to establish clearer decision boundaries even in short texts where class distinctions are often ambiguous. We evaluate LIGRAM on four Korean short-text datasets, where it consistently outperforms existing baseline models. These outcomes validate that integrating language-specific graph representations with SemCon provides an effective solution for short text classification in agglutinative languages such as Korean.

Method: Proposes MIND framework with: 1) Criteria-Grounded Psychiatric Reasoning Bank (PRB) that summarizes dialogue context, retrieves similar consultations, and distills clinical supports; 2) Rubric-based process rewards for explicit clinical reasoning; 3) Value-aware trajectory rectification to jointly optimize information acquisition and diagnostic decisions.

Result: Extensive experiments show MIND consistently outperforms baselines in diagnostic accuracy, empathetic interaction quality, interpretability, and generalization.

Conclusion: MIND addresses fundamental challenges in psychiatric consultation by integrating criteria-grounded reasoning with reinforcement learning, providing a unified framework for improved inquiry and diagnostic decision-making in medical dialogue systems.

Abstract: Large language models (LLMs) have advanced medical dialogue systems, yet psychiatric consultation poses substantially higher demands due to subjective ambiguity and comorbidity complexity: an agent must continuously extract psychopathological cues from incomplete and inconsistent patient reports in multi-turn interactions and perform rigorous differential diagnostic reasoning. However, existing methods face two fundamental challenges. First, without criteria-grounded clinical supports, they are prone to unsupported clinical assertions when symptoms are atypical or underspecified. Second, in multi-turn interactions, they struggle to mitigate inquiry drift (off-topic or low-yield questioning) and optimize questioning strategies. To address these challenges, we propose MIND, a unified inquiry–diagnosis reinforcement learning framework for psychiatric consultation. Specifically, we build a Criteria-Grounded Psychiatric Reasoning Bank (PRB) that summarizes dialogue context into clinical retrieval states, retrieves semantically similar reference consultations, and distills reusable criteria-grounded clinical supports to guide criteria-aligned inquiry and reasoning. Building on this foundation, MIND enforces explicit clinical reasoning with rubric-based process rewards to provide fine-grained supervision over intermediate decision steps, and incorporates a value-aware trajectory rectification mechanism to jointly improve information acquisition and diagnostic decision-making across turns. Extensive experiments demonstrate that MIND consistently outperforms strong baselines in diagnostic accuracy, empathetic interaction quality, interpretability, and generalization.

Method: Proposes ErrorLLM framework that: 1) represents user questions and database schemas as structural features, 2) uses static detection to identify execution failures and surface mismatches, 3) extends semantic space with dedicated error tokens capturing categorized implicit semantic error types, and 4) employs a training strategy to model errors with structural representations for detecting complex implicit errors.

Result: Extensive experiments show ErrorLLM achieves significant improvements over backbone initial generation. Analysis reveals detection quality directly determines refinement effectiveness, and ErrorLLM addresses both sides with high detection F1 score while maintaining refinement effectiveness.

Conclusion: ErrorLLM effectively models text-to-SQL errors through explicit error tokenization and structural representations, enabling better error detection and guided refinement for improved SQL generation quality.

Abstract: Despite the remarkable performance of large language models (LLMs) in text-to-SQL (SQL generation), correctly producing SQL queries remains challenging during initial generation. The SQL refinement task is subsequently introduced to correct syntactic and semantic errors in generated SQL queries. However, existing paradigms face two major limitations: (i) self-debugging becomes increasingly ineffective as modern LLMs rarely produce explicit execution errors that can trigger debugging signals; (ii) self-correction exhibits low detection precision due to the lack of explicit error modeling grounded in the question and schema, and suffers from severe hallucination that frequently corrupts correct SQLs. In this paper, we propose ErrorLLM, a framework that explicitly models text-to-SQL Errors within a dedicated LLM for text-to-SQL refinement. Specifically, we represent the user question and database schema as structural features, employ static detection to identify execution failures and surface mismatches, and extend ErrorLLM’s semantic space with dedicated error tokens that capture categorized implicit semantic error types. Through a well-designed training strategy, we explicitly model these errors with structural representations, enabling the LLM to detect complex implicit errors by predicting dedicated error tokens. Guided by the detected errors, we perform error-guided refinement on the SQL structure by prompting LLMs. Extensive experiments demonstrate that ErrorLLM achieves the most significant improvements over backbone initial generation. Further analysis reveals that detection quality directly determines refinement effectiveness, and ErrorLLM addresses both sides by high detection F1 score while maintain refinement effectiveness.

Method: COREA cascades an SLM with an LLM: the SLM first attempts to answer questions and outputs both an answer and confidence score. Questions with confidence below a threshold are deferred to the LLM. A reinforcement learning algorithm aligns the SLM’s confidence through a confidence calibration reward.

Result: The method improves SLM’s reasoning ability and confidence calibration across diverse datasets and model backbones. Compared to using LLM alone, COREA reduces cost by 21.5% on math datasets and 16.8% on non-math datasets with only within 2% absolute pass@1 drop.

Conclusion: COREA effectively balances accuracy and cost in complex reasoning tasks by cascading SLM and LLM with confidence-based deferral and reinforcement learning-based confidence calibration.

Abstract: Large language models (LLMs) demonstrate superior reasoning capabilities compared to small language models (SLMs), but incur substantially higher costs. We propose COllaborative REAsoner (COREA), a system that cascades an SLM with an LLM to achieve a balance between accuracy and cost in complex reasoning tasks. COREA first attempts to answer questions using the SLM, which outputs both an answer and a verbalized confidence score. Questions with confidence below a predefined threshold are deferred to the LLM for more accurate resolution. We introduce a reinforcement learning-based training algorithm that aligns the SLM’s confidence through an additional confidence calibration reward. Extensive experiments demonstrate that our method jointly improves the SLM’s reasoning ability and confidence calibration across diverse datasets and model backbones. Compared to using the LLM alone, COREA reduces cost by 21.5% and 16.8% on out-of-domain math and non-math datasets, respectively, with only an absolute pass@1 drop within 2%.

Method: Introduces Structure of Thought (SoT) prompting technique that guides models to construct intermediate text structures, and presents T2S-Bench benchmark with 1.8K samples across 6 scientific domains and 32 structural types for evaluating text-to-structure capabilities.

Result: SoT yields +5.7% average improvement across eight diverse text-processing tasks on Qwen2.5-7B-Instruct, with fine-tuning on T2S-Bench increasing gain to +8.6%. Current models show substantial improvement potential with only 52.1% average accuracy on multi-hop reasoning and 58.1% node accuracy for best model on end-to-end extraction.

Conclusion: Explicit text structuring significantly enhances LLM performance on text-processing tasks, and T2S-Bench reveals substantial room for improvement in models’ text-to-structure capabilities.

Abstract: Think about how human handles complex reading tasks: marking key points, inferring their relationships, and structuring information to guide understanding and responses. Likewise, can a large language model benefit from text structure to enhance text-processing performance? To explore it, in this work, we first introduce Structure of Thought (SoT), a prompting technique that explicitly guides models to construct intermediate text structures, consistently boosting performance across eight tasks and three model families. Building upon this insight, we present T2S-Bench, the first benchmark designed to evaluate and improve text-to-structure capabilities of models. T2S-Bench includes 1.8K samples across 6 scientific domains and 32 structural types, rigorously constructed to ensure accuracy, fairness, and quality. Evaluation on 45 mainstream models reveals substantial improvement potential: the average accuracy on the multi-hop reasoning task is only 52.1%, and even the most advanced model achieves 58.1% node accuracy in end-to-end extraction. Furthermore, on Qwen2.5-7B-Instruct, SoT alone yields an average +5.7% improvement across eight diverse text-processing tasks, and fine-tuning on T2S-Bench further increases this gain to +8.6%. These results highlight the value of explicit text structuring and the complementary contributions of SoT and T2S-Bench. Dataset and eval code have been released at https://t2s-bench.github.io/T2S-Bench-Page/.

Method: Stepwise Semantic Alignment (SSA) treats pseudo-source as semantic bridge rather than direct source substitute. Uses universal semantics to rectify pseudo-source semantic features, then aligns target domain. Includes Hierarchical Feature Aggregation (HFA) module and Confidence-Aware Complementary Learning (CACL) strategy to enhance semantic quality without source data or target ground truth.

Result: Achieved 5.2% performance boost on GTA2Cityscapes semantic segmentation task over state-of-the-art. Also evaluated on image classification tasks.

Conclusion: SSA effectively addresses domain adaptation challenges by using pseudo-source as semantic bridge with semantic rectification, outperforming previous methods in test-time adaptation scenarios without source data.

Abstract: Distribution shifts between training and testing data are a critical bottleneck limiting the practical utility of models, especially in real-world test-time scenarios. To adapt models when the source domain is unknown and the target domain is unlabeled, previous works constructed pseudo-source domains via data generation and translation, then aligned the target domain with them. However, significant discrepancies exist between the pseudo-source and the original source domain, leading to potential divergence when correcting the target directly. From this perspective, we propose a Stepwise Semantic Alignment (SSA) method, viewing the pseudo-source as a semantic bridge connecting the source and target, rather than a direct substitute for the source. Specifically, we leverage easily accessible universal semantics to rectify the semantic features of the pseudo-source, and then align the target domain using the corrected pseudo-source semantics. Additionally, we introduce a Hierarchical Feature Aggregation (HFA) module and a Confidence-Aware Complementary Learning (CACL) strategy to enhance the semantic quality of the SSA process in the absence of source and ground truth of target domains. We evaluated our approach on tasks like semantic segmentation and image classification, achieving a 5.2% performance boost on GTA2Cityscapes over the state-of-the-art.

Method: Evaluated 10 LLMs (3 proprietary, 7 open-source) using MITI 4.2 framework on two datasets: 96 handcrafted model transcripts and 34 real-world clinical transcripts. Generated parallel LLM-therapist utterances while keeping client responses static, used composite ranking with MITI components and verbosity, and conducted distinguishability experiments with psychiatrists.

Result: All LLMs showed fair to good MITI competence (scores >3.5). Best models (gemma-3-27b-it, gemini-2.5-pro, grok-3) outperformed human therapists in Complex Reflection percentage (39% vs 96%) and Reflection-Question ratio (1.2 vs >2.8). Psychiatrists identified LLM responses with only 56% accuracy (near chance).

Conclusion: LLMs can achieve good motivational interviewing proficiency in real-world clinical settings, with open-source models being viable for expanding MI counseling in low-resource settings.

Abstract: Motivational interviewing (MI) promotes behavioural change in substance use disorders. Its fidelity is measured using the Motivational Interviewing Treatment Integrity (MITI) framework. While large language models (LLMs) can potentially generate MI-consistent therapist responses, their competence using MITI is not well-researched, especially in real world clinical transcripts. We aim to benchmark MI competence of proprietary and open-source models compared to human therapists in real-world transcripts and assess distinguishability from human therapists. Methods: We shortlisted 3 proprietary and 7 open-source LLMs from LMArena, evaluated performance using MITI 4.2 framework on two datasets (96 handcrafted model transcripts, 34 real-world clinical transcripts). We generated parallel LLM-therapist utterances iteratively for each transcript while keeping client responses static, and ranked performance using a composite ranking system with MITI components and verbosity. We conducted a distinguishability experiment with two independent psychiatrists to identify human-vs-LLM responses. Results: All 10 tested LLMs had fair (MITI global scores >3.5) to good (MITI global scores >4) competence across MITI measures, and three best-performing models (gemma-3-27b-it, gemini-2.5-pro, grok-3) were tested on real-world transcripts. All showed good competence, with LLMs outperforming human-expert in Complex Reflection percentage (39% vs 96%) and Reflection-Question ratio (1.2 vs >2.8). In the distinguishability experiment, psychiatrists identified LLM responses with only 56% accuracy, with d-prime: 0.17 and 0.25 for gemini-2.5-pro and gemma-3-27b-it respectively. Conclusion: LLMs can achieve good MI proficiency in real-world clinical transcripts using MITI framework. These findings suggest that even open-source LLMs are viable candidates for expanding MI counselling sessions in low-resource settings.

Method: Two prototype-based approaches: 1) Prototype-Based Regularization (PBR) learns soft prototypes through distance-based auxiliary loss to structure latent space; 2) Prototype-Conditioned Modulation (PCM) constructs corpus-level prototypes and injects them during training and inference. Also introduces SCOTUS-Law dataset with three annotation granularity levels.

Result: Experiments on legal, medical, and scientific benchmarks show consistent improvements over strong baselines, with 4 Macro-F1 gains on low-frequency roles. The methods effectively address the global representation limitation of hierarchical models.

Conclusion: Prototype-based methods successfully integrate local and global features for RRL, with significant gains on low-frequency roles. The SCOTUS-Law dataset fills a resource gap in legal discourse analysis. Analysis includes implications for Large Language Models and expert evaluation.

Abstract: Rhetorical Role Labeling (RRL) identifies the functional role of each sentence in a document, a key task for discourse understanding in domains such as law and medicine. While hierarchical models capture local dependencies effectively, they are limited in modeling global, corpus-level features. To address this limitation, we propose two prototype-based methods that integrate local context with global representations. Prototype-Based Regularization (PBR) learns soft prototypes through a distance-based auxiliary loss to structure the latent space, while Prototype-Conditioned Modulation (PCM) constructs corpus-level prototypes and injects them during training and inference. Given the scarcity of RRL resources, we introduce SCOTUS-Law, the first dataset of U.S. Supreme Court opinions annotated with rhetorical roles at three levels of granularity: category, rhetorical function, and step. Experiments on legal, medical, and scientific benchmarks show consistent improvements over strong baselines, with 4 Macro-F1 gains on low-frequency roles. We further analyze the implications in the era of Large Language Models and complement our findings with expert evaluation.

Method: Used anonymized transcribed role-play sessions between licensed therapists and clients. Compared two approaches: (1) generation-only method and (2) Retrieval-Augmented Generation (RAG) using CBT guidelines. Evaluated both proprietary and open-source models using NLG metrics, natural language inference, and automated scoring for skills assessment.

Result: LLMs can generate CBT-like dialogues but are limited in their ability to convey empathy and maintain consistency. Both generation-only and RAG approaches were evaluated with mixed results on linguistic quality, semantic coherence, and therapeutic fidelity.

Conclusion: While LLMs show potential for generating therapeutic content, they have significant limitations in empathy and consistency that currently prevent them from effectively emulating professional therapists, highlighting the need for further development and validation.

Abstract: As mental health issues continue to rise globally, there is an increasing demand for accessible and scalable therapeutic solutions. Many individuals currently seek support from Large Language Models (LLMs), even though these models have not been validated for use in counseling services. In this paper, we evaluate LLMs’ ability to emulate professional therapists practicing Cognitive Behavioral Therapy (CBT). Using anonymized, transcribed role-play sessions between licensed therapists and clients, we compare two approaches: (1) a generation-only method and (2) a Retrieval-Augmented Generation (RAG) approach using CBT guidelines. We evaluate both proprietary and open-source models for linguistic quality, semantic coherence, and therapeutic fidelity using standard natural language generation (NLG) metrics, natural language inference (NLI), and automated scoring for skills assessment. Our results indicate that while LLMs can generate CBT-like dialogues, they are limited in their ability to convey empathy and maintain consistency.

Method: Created a human-annotated benchmark with human-defined topics and manually annotated spans, evaluated diverse LLMs alongside BERT-based models fine-tuned on a distilled development dataset, with evaluation at both document and word levels.

Result: Substantial variability among LLMs, with performance ranging from near-human topic detection to pronounced failures in span localization; strongest models approach human agreement while distilled token embedding models remain competitive despite smaller scale.

Conclusion: The benchmark enables evaluation of topic localization models relative to human agreement, revealing LLM variability and showing that while strong models approach human performance, smaller distilled models remain competitive.

Abstract: Topic localization aims to identify spans of text that express a given topic defined by a name and description. To study this task, we introduce a human-annotated benchmark based on Czech historical documents, containing human-defined topics together with manually annotated spans and supporting evaluation at both document and word levels. Evaluation is performed relative to human agreement rather than a single reference annotation. We evaluate a diverse range of large language models alongside BERT-based models fine-tuned on a distilled development dataset. Results reveal substantial variability among LLMs, with performance ranging from near-human topic detection to pronounced failures in span localization. While the strongest models approach human agreement, the distilled token embedding models remain competitive despite their smaller scale. The dataset and evaluation framework are publicly available at: https://github.com/dcgm/czechtopic.

Method: Proposes anonymous evaluation method where character names are hidden, investigates personality augmentation using both human-annotated traits and model self-generated traits to improve role fidelity under anonymization.

Result: Anonymization significantly degrades role-playing performance, confirming name exposure carries implicit information. Personality augmentation consistently improves performance, with self-generated personalities achieving comparable results to human-annotated ones.

Conclusion: Establishes fairer evaluation protocol for role-playing agents and validates scalable personality-enhanced framework for constructing robust role-playing agents that don’t rely on name-based memory.

Abstract: Large language models (LLMs) have demonstrated significant potential in developing Role-Playing Agents (RPAs). However, current research primarily evaluates RPAs using famous fictional characters, allowing models to rely on memory associated with character names. This dependency creates a bias that limits the generalization of RPAs to unseen personas. To address this issue, we propose an anonymous evaluation method. Experiments across multiple benchmarks reveal that anonymization significantly degrades role-playing performance, confirming that name exposure carries implicit information. Furthermore, we investigate personality augmentation to enhance role fidelity under anonymous setting. We systematically compare the efficacy of personality traits derived from human annotations versus those self-generated by the model. Our results demonstrate that incorporating personality information consistently improves RPA performance. Crucially, self-generated personalities achieve performance comparable to human-annotated ones. This work establishes a fairer evaluation protocol and validates a scalable, personality-enhanced framework for constructing robust RPAs.

Method: Evaluated LLMs as judges of semantic equivalence in French medical OEQA, comparing closed-access, general-purpose, and biomedical domain-adapted models. Also applied supervised fine-tuning (SFT) and Group Relative Policy Optimization (GRPO) to adapt compact models.

Result: LLM-based judgments are strongly influenced by the answer generator, with agreement varying across generators. Domain-adapted and large general-purpose models align best with expert annotations. Lightweight adaptation of compact models with SFT and GRPO improves performance and reduces generator sensitivity even with limited data.

Conclusion: Generator-aware evaluation is crucial, and carefully adapted small models can support scalable evaluation in low-resource medical settings, offering a practical solution to the expert annotation bottleneck.

Abstract: Automatic evaluation of medical open-ended question answering (OEQA) remains challenging due to the need for expert annotations. We evaluate whether large language models (LLMs) can act as judges of semantic equivalence in French medical OEQA, comparing closed-access, general-purpose, and biomedical domain-adapted models. Our results show that LLM-based judgments are strongly influenced by the model that generated the answer, with agreement varying substantially across generators. Domain-adapted and large general-purpose models achieve the highest alignment with expert annotations. We further show that lightweight adaptation of a compact model using supervised fine-tuning (SFT) and Group Relative Policy Optimization (GRPO) substantially improves performance and reduces generator sensitivity, even with limited data. Overall, our findings highlight the need for generator-aware evaluation and suggest that carefully adapted small models can support scalable evaluation in low-resource medical settings.

Method: Proposes activation-based monitoring using sparse autoencoders on residual stream activations with lightweight linear classifiers to produce token-level estimates of reward-hacking activity.

Result: Internal activation patterns reliably distinguish reward-hacking from benign behavior across multiple model families and fine-tuning mixtures, generalize to unseen mixed-policy adapters, and show model-dependent temporal structure during chain-of-thought reasoning.

Conclusion: Internal activation monitoring provides complementary and earlier signal of emergent misalignment than output-based evaluation, supporting more robust post-deployment safety monitoring for fine-tuned language models.

Abstract: Fine-tuned large language models can exhibit reward-hacking behavior arising from emergent misalignment, which is difficult to detect from final outputs alone. While prior work has studied reward hacking at the level of completed responses, it remains unclear whether such behavior can be identified during generation. We propose an activation-based monitoring approach that detects reward-hacking signals from internal representations as a model generates its response. Our method trains sparse autoencoders on residual stream activations and applies lightweight linear classifiers to produce token-level estimates of reward-hacking activity. Across multiple model families and fine-tuning mixtures, we find that internal activation patterns reliably distinguish reward-hacking from benign behavior, generalize to unseen mixed-policy adapters, and exhibit model-dependent temporal structure during chain-of-thought reasoning. Notably, reward-hacking signals often emerge early, persist throughout reasoning, and can be amplified by increased test-time compute in the form of chain-of-thought prompting under weakly specified reward objectives. These results suggest that internal activation monitoring provides a complementary and earlier signal of emergent misalignment than output-based evaluation, supporting more robust post-deployment safety monitoring for fine-tuned language models.

Method: Conducted “hindsight” experiments on a unique multi-candidate dataset from genuine MT post-editing projects (6,000+ English source segments with 9 translation hypotheses from diverse NMT systems and LLMs). Used Kendall’s rank correlation to assess predictive power of source-side difficulty metrics, candidate-side QE models, and position heuristics against TER (post-editing effort proxy) and COMET (human judgment proxy).

Result: The architectural shift towards LLMs alters the reliability of established quality prediction methods while simultaneously mitigating previous challenges in document-level translation.

Conclusion: LLMs fundamentally change how MT quality prediction should be approached, requiring reevaluation of established methods as their adoption reshapes the translation workflow landscape.

Abstract: This paper investigates two complementary paradigms for predicting machine translation (MT) quality: source-side difficulty prediction and candidate-side quality estimation (QE). The rapid adoption of Large Language Models (LLMs) into MT workflows is reshaping the research landscape, yet its impact on established quality prediction paradigms remains underexplored. We study this issue through a series of “hindsight” experiments on a unique, multi-candidate dataset resulting from a genuine MT post-editing (MTPE) project. The dataset consists of over 6,000 English source segments with nine translation hypotheses from a diverse set of traditional neural MT systems and advanced LLMs, all evaluated against a single, final human post-edited reference. Using Kendall’s rank correlation, we assess the predictive power of source-side difficulty metrics, candidate-side QE models and position heuristics against two gold-standard scores: TER (as a proxy for post-editing effort) and COMET (as a proxy for human judgment). Our findings highlight that the architectural shift towards LLMs alters the reliability of established quality prediction methods while simultaneously mitigating previous challenges in document-level translation.

Method: Introduces FINEST, a fine-grained response evaluation taxonomy for sensitive topics that breaks down helpfulness and harmlessness into errors across three main categories: Content, Logic, and Appropriateness. Develops a score- and error-based improvement pipeline guided by FINEST.

Result: Experiments on a Korean-sensitive question dataset show that the FINEST-guided improvement pipeline significantly improves model responses across all three categories, outperforming refinement without guidance. Score-based improvement reduces the error sentence ratio for Appropriateness by up to 33.09%.

Conclusion: FINEST lays the foundation for more explainable and comprehensive evaluation and improvement of LLM responses to sensitive questions, enabling better balance between safety and helpfulness.

Abstract: Large Language Models (LLMs) often generate overly cautious and vague responses on sensitive topics, sacrificing helpfulness for safety. Existing evaluation frameworks lack systematic methods to identify and address specific weaknesses in responses to sensitive topics, making it difficult to improve both safety and helpfulness simultaneously. To address this, we introduce FINEST, a FINE-grained response evaluation taxonomy for Sensitive Topics, which breaks down helpfulness and harmlessness into errors across three main categories: Content, Logic, and Appropriateness. Experiments on a Korean-sensitive question dataset demonstrate that our score- and error-based improvement pipeline, guided by FINEST, significantly improves the model responses across all three categories, outperforming refinement without guidance. Notably, score-based improvement – providing category-specific scores and justifications – yields the most significant gains, reducing the error sentence ratio for Appropriateness by up to 33.09%. This work lays the foundation for a more explainable and comprehensive evaluation and improvement of LLM responses to sensitive questions.

Method: A unified, rule-based pipeline that handles: 1) number conversion, 2) date/time normalization, 3) currency handling, 4) percentage expansion, 5) acronym resolution via dictionary, 6) foreign term transliteration, and 7) Unicode normalization. Uses pre-compiled regex patterns for efficiency.

Result: Created a zero-dependency Python library installable via pip, available on PyPI and GitHub under MIT license. Provides high-throughput batch processing with minimal memory overhead and no GPU/external API dependencies.

Conclusion: VietNormalizer addresses gaps in Vietnamese text normalization tools and discusses generalizability of rule-based approaches to other low-resource tonal and agglutinative languages.

Abstract: We present VietNormalizer1, an open-source, zero-dependency Python library for Vietnamese text normalization targeting Text-to-Speech (TTS) and Natural Language Processing (NLP) applications. Vietnamese text normalization is a critical yet underserved preprocessing step: real-world Vietnamese text is densely populated with non-standard words (NSWs), including numbers, dates, times, currency amounts, percentages, acronyms, and foreign-language terms, all of which must be converted to fully pronounceable Vietnamese words before TTS synthesis or downstream language processing. Existing Vietnamese normalization tools either require heavy neural dependencies while covering only a narrow subset of NSW classes, or are embedded within larger NLP toolkits without standalone installability. VietNormalizer addresses these gaps through a unified, rule-based pipeline that: (1) converts arbitrary integers, decimals, and large numbers to Vietnamese words; (2) normalizes dates and times to their spoken Vietnamese forms; (3) handles VND and USD currency amounts; (4) expands percentages; (5) resolves acronyms via a customizable CSV dictionary; (6) transliterates non-Vietnamese loanwords and foreign terms to Vietnamese phonetic approximations; and (7) performs Unicode normalization and emoji/special-character removal. All regular expression patterns are pre-compiled at initialization, enabling high-throughput batch processing with minimal memory overhead and no GPU or external API dependency. The library is installable via pip install vietnormalizer, available on PyPI and GitHub at https://github.com/nghimestudio/vietnormalizer, and released under the MIT license. We discuss the design decisions, limitations of existing approaches, and the generalizability of the rule-based normalization paradigm to other low-resource tonal and agglutinative languages.

Method: Tested 17 open-weight models on 192 FBT variants using Bayesian Logistic regression to analyze effects of model size and post-training; conducted case study on OLMo 2 training; used vector steering to isolate causal factors.

Result: Model scaling benefits performance but not strictly; mental-state vocabulary creates stereotypical response patterns that override scenario semantics; instruction tuning partially mitigates but reasoning-oriented finetuning amplifies effects; identified “think” vector as causal driver.

Conclusion: LLMs develop stereotypical response patterns tied to mental-state vocabulary during pre-training that can outweigh scenario semantics, with a specific “think” vector driving observed False Belief Test behavior.

Abstract: The False Belief Test (FBT) has been the main method for assessing Theory of Mind (ToM) and related socio-cognitive competencies. For Large Language Models (LLMs), the reliability and explanatory potential of this test have remained limited due to issues like data contamination, insufficient model details, and inconsistent controls. We address these issues by testing 17 open-weight models on a balanced set of 192 FBT variants (Trott et al. 2023) using Bayesian Logistic regression to identify how model size and post-training affect socio-cognitive competence. We find that scaling model size benefits performance, but not strictly. A cross-over effect reveals that explicating propositional attitudes (X thinks) fundamentally alters response patterns. Instruction tuning partially mitigates this effect, but further reasoning-oriented finetuning amplifies it. In a case study analysing social reasoning ability throughout OLMo 2 training, we show that this cross-over effect emerges during pre-training, suggesting that models acquire stereotypical response patterns tied to mental-state vocabulary that can outweigh other scenario semantics. Finally, vector steering allows us to isolate a think vector as the causal driver of observed FBT behaviour.

Method: Used Bielik-11B-v2.3-Instruct (11B parameters, Mistral architecture) as base model, compared six post-training quantization methods (QuIP#, SpinQuant+GPTQ, ButterflyQuant, QTIP, VPTQ, AQLM) calibrated on Polish-language corpus CulturaX-PL with shared Hessian matrices.

Result: Best variant (QuIP# E8P12) achieved 71.92% across 22 Polish benchmarks vs 72.07% for IQ2_XXS baseline; QTIP achieved best per-bit efficiency (79.4% MC acc_norm at ~2.4 bpw); discovered MC-generation dissociation phenomenon where rotation-based methods preserve log-likelihood but fail at autoregressive generation.

Conclusion: Extreme 2-bit quantization can be effectively applied to Polish LLMs with minimal performance loss, QTIP offers best efficiency, and rotation-based methods show generation limitations despite preserving log-likelihood metrics.

Abstract: We present Bielik-Q2-Sharp, the first systematic academic evaluation of extreme 2-bit quantization applied to a Polish large language model. Using Bielik-11B-v2.3-Instruct (11B parameters, Mistral architecture) as our base model, we compare six state-of-the-art post-training quantization methods – QuIP#, SpinQuant+GPTQ, ButterflyQuant, QTIP, VPTQ, and AQLM – all calibrated on a Polish-language corpus (CulturaX-PL) with shared Hessian matrices. Our best variant (QuIP# E8P12) achieves 71.92% across 22 Polish benchmarks versus 72.07% for the IQ2_XXS baseline – within statistical noise, at a modest size premium (3.26 GB vs. ~2.6 GB). On eq_bench, our method scores 47.14 versus 43.53 (+3.6pp), suggesting superior preservation of higher-order reasoning. QTIP achieves the best per-bit efficiency (79.4% MC acc_norm at ~2.4 bpw, 3.27 GB), matching VPTQ’s quality at 35% smaller size. We additionally document a MC-generation dissociation phenomenon where rotation-based methods preserve log-likelihood quality but fail catastrophically at autoregressive generation. The entire project was conducted by a single independent researcher on cloud GPUs (vast.ai) within a $285 budget. All models, Hessians, and evaluation logs are publicly available.

Method: Evaluated 11 major LLMs using 1,152 synthetically generated scenarios across 24 UDHR rights articles and 8 languages, analyzing trade-offs in rights decisions.

Result: LLMs show systematic biases: (1) accept limiting Economic, Social, and Cultural rights more than Political and Civil rights, (2) demonstrate cross-linguistic variation with higher rights-limiting endorsement in Chinese and Hindi, (3) are susceptible to prompt-based steering, and (4) show differences between Likert and open-ended responses.

Conclusion: LLMs exhibit concerning biases in human rights alignment that vary across languages and rights categories, highlighting critical challenges in LLM preference assessment and the need for better alignment with universal human rights principles.

Abstract: As Large Language Models (LLMs) increasingly mediate global information access with the potential to shape public discourse, their alignment with universal human rights principles becomes important to ensure that these rights are abided by in high stakes AI-mediated interactions. In this paper, we evaluate how LLMs navigate trade-offs involving the Universal Declaration of Human Rights (UDHR), leveraging 1,152 synthetically generated scenarios across 24 rights articles and eight languages. Our analysis of eleven major LLMs reveals systematic biases where models: (1) accept limiting Economic, Social, and Cultural rights more often than Political and Civil rights, (2) demonstrate significant cross-linguistic variation with elevated endorsement rates of rights-limiting actions in Chinese and Hindi compared to English or Romanian, (3) show substantial susceptibility to prompt-based steering, and (4) exhibit noticeable differences between Likert and open-ended responses, highlighting critical challenges in LLM preference assessment.

Method: Systematically varied transcription and preprocessing methods while keeping retrieval mechanism fixed (BM25), comparing performance on multilingual and visual document benchmarks

Result: BM25 with improved document representation can recover large performance gaps on multilingual and visual benchmarks, suggesting document representation is primary driver of improvements

Conclusion: Field needs decomposed evaluation benchmarks that separately measure transcription and retrieval capabilities to correctly attribute progress and focus effort effectively

Abstract: Retrieval-augmented generation (RAG) is a common way to ground language models in external documents and up-to-date information. Classical retrieval systems relied on lexical methods such as BM25, which rank documents by term overlap with corpus-level weighting. End-to-end multimodal retrievers trained on large query-document datasets claim substantial improvements over these approaches, especially for multilingual documents with complex visual layouts. We demonstrate that better document representation is the primary driver of benchmark improvements. By systematically varying transcription and preprocessing methods while holding the retrieval mechanism fixed, we demonstrate that BM25 can recover large gaps on multilingual and visual benchmarks. Our findings call for decomposed evaluation benchmarks that separately measure transcription and retrieval capabilities, enabling the field to correctly attribute progress and focus effort where it matters.

Method: Memex maintains a compact working context with structured summaries and stable indices, while storing full-fidelity interactions in an external experience database. Uses reinforcement learning (MemexRL) to optimize write/read behaviors with reward shaping for indexed memory usage under context budget.

Result: On challenging long-horizon tasks, Memex agent trained with MemexRL improves task success while using significantly smaller working context compared to summary-only approaches.

Conclusion: Memex provides a substantially less lossy form of long-horizon memory than summary-only approaches, with theoretical analysis showing potential to preserve decision quality with bounded dereferencing while keeping effective in-context computation bounded.

Abstract: Large language model (LLM) agents are fundamentally bottlenecked by finite context windows on long-horizon tasks. As trajectories grow, retaining tool outputs and intermediate reasoning in-context quickly becomes infeasible: the working context becomes prohibitively long, eventually exceeds the context budget, and makes distant evidence harder to use even when it is still present. Existing solutions typically shorten context through truncation or running summaries, but these methods are fundamentally lossy because they compress or discard past evidence itself. We introduce Memex, an indexed experience memory mechanism that instead compresses context without discarding evidence. Memex maintains a compact working context consisting of concise structured summaries and stable indices, while storing full-fidelity underlying interactions in an external experience database under those indices. The agent can then decide when to dereference an index and recover the exact past evidence needed for the current subgoal. We optimize both write and read behaviors with our reinforcement learning framework MemexRL, using reward shaping tailored to indexed memory usage under a context budget, so the agent learns what to summarize, what to archive, how to index it, and when to retrieve it. This yields a substantially less lossy form of long-horizon memory than summary-only approaches. We further provide a theoretical analysis showing the potential of the Memex loop to preserve decision quality with bounded dereferencing while keeping effective in-context computation bounded as history grows. Empirically, on challenging long-horizon tasks, Memex agent trained with MemexRL improves task success while using a significantly smaller working context.

Method: Position paper with diagnostic evidence comparing vector prompt tuning vs. text-based prompt optimization. Analyzes performance saturation patterns and attention mechanisms to demonstrate vector prompts’ advantages.

Result: Vector prompt tuning continues improving with more supervision while text-based optimization saturates early. Vector prompts show dense, global attention patterns indicating distinct control mechanisms. Exposing vector prompts doesn’t significantly increase model leakage risk under standard black-box threat models.

Conclusion: Model providers should expose vector prompt inputs as part of public LLM interfaces. The community should rethink prompt interfaces as core components of LLM customization for better scalability and stability.

Abstract: As large language models (LLMs) transition from research prototypes to real-world systems, customization has emerged as a central bottleneck. While text prompts can already customize LLM behavior, we argue that text-only prompting does not constitute a suitable control interface for scalable, stable, and inference-only customization. This position paper argues that model providers should expose \emph{vector prompt inputs} as part of the public interface for customizing LLMs. We support this position with diagnostic evidence showing that vector prompt tuning continues to improve with increasing supervision whereas text-based prompt optimization saturates early, and that vector prompts exhibit dense, global attention patterns indicative of a distinct control mechanism. We further discuss why inference-only customization is increasingly important under realistic deployment constraints, and why exposing vector prompts need not fundamentally increase model leakage risk under a standard black-box threat model. We conclude with a call to action for the community to rethink prompt interfaces as a core component of LLM customization.

Method: Examined LLM responses to user prompts expressing varying degrees of Dark Triad traits (Machiavellianism, Narcissism, Psychopathy) using a curated dataset, analyzing model behavior across different severity levels and sentiment.

Result: All models predominantly exhibit corrective behavior but show reinforcing output in certain cases. Model behavior depends on severity level and differs in response sentiment.

Conclusion: Findings raise implications for designing safer conversational systems that can detect and respond appropriately when users escalate from benign to harmful requests.

Abstract: Large Language Models (LLMs) often exhibit highly agreeable and reinforcing conversational styles, also known as AI-sycophancy. Although this behavior is encouraged, it may become problematic when interacting with user prompts that reflect negative social tendencies. Such responses risk amplifying harmful behavior rather than mitigating it. In this study, we examine how LLMs respond to user prompts expressing varying degrees of Dark Triad traits (Machiavellianism, Narcissism, and Psychopathy) using a curated dataset. Our analysis reveals differences across models, whereby all models predominantly exhibit corrective behavior, while showing reinforcing output in certain cases. Model behavior also depends on the severity level and differs in the sentiment of the response. Our findings raise implications for designing safer conversational systems that can detect and respond appropriately when users escalate from benign to harmful requests.

Method: V₁ framework with two components: V₁-Infer uses uncertainty-guided tournament-based ranking to dynamically allocate verification compute to uncertain candidate pairs; V₁-PairRL jointly trains a single model as both generator and pairwise self-verifier using RL.

Result: On code generation (LiveCodeBench, CodeContests, SWE-Bench) and math reasoning (AIME, HMMT) benchmarks, V₁-Infer improves Pass@1 by up to 10% over pointwise verification and outperforms recent test-time scaling methods while being more efficient. V₁-PairRL achieves 7-9% test-time scaling gains over standard RL and improves base Pass@1 by up to 8.7%.

Conclusion: Pairwise self-verification is more effective than independent scoring for complex reasoning tasks, and joint training of generator and verifier models leads to significant performance improvements in test-time scaling.

Abstract: Test-time scaling for complex reasoning tasks shows that leveraging inference-time compute, by methods such as independently sampling and aggregating multiple solutions, results in significantly better task outcomes. However, a critical bottleneck is verification: sampling is only effective if correct solutions can be reliably identified among candidates. While existing approaches typically evaluate candidates independently via scalar scoring, we demonstrate that models are substantially stronger at pairwise self-verification. Leveraging this insight, we introduce $V_1$, a framework that unifies generation and verification through efficient pairwise ranking. $V_1$ comprises two components: $V_1$-Infer, an uncertainty-guided algorithm using a tournament-based ranking that dynamically allocates self-verification compute to candidate pairs whose relative correctness is most uncertain; and $V_1$-PairRL, an RL framework that jointly trains a single model as both generator and pairwise self-verifier, ensuring the verifier adapts to the generator’s evolving distribution. On code generation (LiveCodeBench, CodeContests, SWE-Bench) and math reasoning (AIME, HMMT) benchmarks, $V_1$-Infer improves Pass@1 by up to $10%$ over pointwise verification and outperforms recent test-time scaling methods while being significantly more efficient. Furthermore, $V_1$-PairRL achieves $7$–$9%$ test-time scaling gains over standard RL and pointwise joint training, and improves base Pass@1 by up to 8.7% over standard RL in a code-generation setting.

Method: Applied same ridge regression probes used in LLM studies to static co-occurrence embeddings (GloVe and Word2Vec). Analyzed geographic signal (city coordinates) and temporal signal (historical birth years). Used semantic-neighbor analyses and targeted subspace ablations to understand what lexical features contribute to the signals.

Result: Found substantial recoverable geographic signal (R² 0.71-0.87 for city coordinates) and reliable temporal signal (R² 0.48-0.52 for birth years) from simple static embeddings. Signals depend on interpretable lexical gradients like country names and climate-related vocabulary.

Conclusion: Ordinary word co-occurrence preserves richer spatial, temporal, and environmental structure than often assumed. Linear probe recoverability alone doesn’t establish that LLMs have moved beyond text to world-like representations, as similar structure exists in simple static embeddings.

Abstract: Recent work interprets the linear recoverability of geographic and temporal variables from large language model (LLM) hidden states as evidence for world-like internal representations. We test a simpler possibility: that much of the relevant structure is already latent in text itself. Applying the same class of ridge regression probes to static co-occurrence-based embeddings (GloVe and Word2Vec), we find substantial recoverable geographic signal and weaker but reliable temporal signal, with held-out R^2 values of 0.71-0.87 for city coordinates and 0.48-0.52 for historical birth years. Semantic-neighbor analyses and targeted subspace ablations show that these signals depend strongly on interpretable lexical gradients, especially country names and climate-related vocabulary. These findings suggest that ordinary word co-occurrence preserves richer spatial, temporal, and environmental structure than is often assumed, revealing a remarkable and underappreciated capacity of simple static embeddings to preserve world-shaped structure from text alone. Linear probe recoverability alone therefore does not establish a representational move beyond text.

Method: Three-stage system: 1) Graph-based retrieval for candidate events, 2) LLM-driven abductive reasoning with prompt design optimized through reflective prompt evolution, 3) Post-hoc consistency enforcement. Cross-model error analysis across 14 models from 7 families to identify shared inductive biases.

Result: System ranked first on SemEval 2026 Task 12 evaluation-phase leaderboard with 0.95 accuracy. Error analysis revealed three systematic biases across models: causal chain incompleteness, proximate cause preference, and salience bias, with 51% cause-count reduction showing cross-family convergence.

Conclusion: The proposed system effectively addresses abductive event reasoning, while the analysis reveals systematic rather than model-specific failure modes in causal reasoning, suggesting fundamental challenges in multi-label causal inference that transcend individual model architectures.

Abstract: We present a winning three-stage system for SemEval 2026 Task12: Abductive Event Reasoning that combines graph-based retrieval, LLM-driven abductive reasoning with prompt design optimized through reflective prompt evolution, and post-hoc consistency enforcement; our system ranks first on the evaluation-phase leaderboard with an accuracy score of 0.95. Cross-model error analysis across 14 models (7families) reveals three shared inductive biases: causal chain incompleteness, proximate cause preference, and salience bias, whose cross-family convergence (51% cause-count reduction) indicates systematic rather than model-specific failure modes in multi-label causal reasoning.

Method: Introduces Reasoning-Aware Retrieval (jointly embedding agent’s reasoning trace with query) and DR-Synth (data synthesis method generating training data from standard QA datasets). Combines both to train AgentIR-4B embedding model.

Result: AgentIR-4B achieves 68% accuracy on BrowseComp-Plus benchmark with Tongyi-DeepResearch agent, compared to 50% with conventional embedding models twice its size, and 37% with BM25.

Conclusion: Reasoning-aware retrieval paradigm effectively exploits agents’ explicit reasoning traces, yielding substantial improvements in retrieval performance for deep research agents.

Abstract: Deep Research agents are rapidly emerging as primary consumers of modern retrieval systems. Unlike human users who issue and refine queries without documenting their intermediate thought processes, Deep Research agents generate explicit natural language reasoning before each search call, revealing rich intent and contextual information that existing retrievers entirely ignore. To exploit this overlooked signal, we introduce: (1) Reasoning-Aware Retrieval, a retrieval paradigm that jointly embeds the agent’s reasoning trace alongside its query; and (2) DR-Synth, a data synthesis method that generates Deep Research retriever training data from standard QA datasets. We demonstrate that both components are independently effective, and their combination yields a trained embedding model, AgentIR-4B, with substantial gains. On the challenging BrowseComp-Plus benchmark, AgentIR-4B achieves 68% accuracy with the open-weight agent Tongyi-DeepResearch, compared to 50% with conventional embedding models twice its size, and 37% with BM25. Code and data are available at: https://texttron.github.io/AgentIR/.

Method: Models early exit as distribution prediction problem, approximates distribution via exit information of similar data. Collects exit information of correct predictions to build retrieval database, then uses retrieved similar data’s exit information to guide backbone model’s exit decisions.

Result: RAEE accelerates inference while achieving robust zero-shot performance across eight downstream tasks, demonstrating both speed improvements and performance enhancement.

Conclusion: RAEE provides an effective retrieval-augmented approach for early exit that reduces computational overhead while maintaining or improving model performance without significant training requirements.

Abstract: Deploying large language model inference remains challenging due to their high computational overhead. Early exit optimizes model inference by adaptively reducing the number of inference layers. Current methods typically train internal classifiers or use heuristic methods to determine the exit layer. However, those methods either introduce significant training overheads or lead to performance degradation. To address these limitations, this paper proposes RAEE, a robust Retrieval-Augmented Early Exit framework that not only enables early exit but also enhances model performance through corrective exit information at intermediate layers. This paper first demonstrates that the early exit problem can be effectively modeled as a distribution prediction problem, in which the distribution can be further approximated through the exit information of similar data. Subsequently, this paper introduces the process of collecting exit information of correct predictions and the steps to construct the retrieval database. Finally, leveraging the pre-constructed retrieval database, RAEE utilizes the exit information from retrieved similar data to guide the backbone model’s exit. Experimental results demonstrate that RAEE can not only accelerate inference while achieving robust zero-shot performance across eight downstream tasks.

Method: Used neural network language models to study passivization learning, characterized human judgments, compared model judgments to human data, and tested hypotheses by training models on manipulated corpora (removing/altered/introducing specific sentence types).

Result: Models’ passivizability judgments closely matched human judgments; both entrenchment (frequency) and affectedness (semantics) made independent contributions to learning passivization exceptions.

Conclusion: Language models can learn grammatical exceptions from linguistic input using multiple evidence sources, and manipulating training data is a useful method for studying language acquisition questions requiring input control.

Abstract: Grammatical rules in natural languages are often characterized by exceptions. How do language learners learn these exceptions to otherwise general patterns? Here, we study this question through the case study of English passivization. While passivization is in general quite productive, there are cases where it cannot apply (cf. the following sentence is ungrammatical: *One hour was lasted by the meeting). Using neural network language models as theories of language acquisition, we explore the sources of indirect evidence that a learner can leverage to learn whether a verb can be passivized. We first characterize English speakers’ judgments of exceptions to the passive, and confirm that speakers find some verbs more passivizable than others. We then show that a neural network language model’s verb passivizability judgments are largely similar to those displayed by humans, suggesting that evidence for these exceptions is available in the linguistic input. Finally, we test two hypotheses as to the source of evidence that language models use to learn these restrictions: frequency (entrenchment) and semantics (affectedness). We do so by training models on versions of the corpus that have had sentences of the types implicated by each hypothesis removed, altered, or introduced. We find support for both hypotheses: entrenchment and affectedness make independent contributions to a verb’s passivizability. From a methodological point of view, this study highlights the utility of altering a language model’s training data for answering questions where complete control over a learner’s input is vital.

Method: Proposes natural language unit tests that decompose response quality into explicit, testable criteria. Develops LMUnit scoring model using multi-objective training across three data types: preferences, direct ratings, and natural language rationales.

Result: Human studies show significantly improved inter-annotator agreement. LMUnit achieves SOTA on evaluation benchmarks (FLASK, BigGenBench) and competitive results on RewardBench. Enables more effective LLM development workflows.

Conclusion: Natural language unit tests combined with LMUnit scoring provide a promising path forward for language model evaluation and development, offering systematic, interpretable assessment methods.

Abstract: As language models become integral to critical workflows, assessing their behavior remains a fundamental challenge – human evaluation is costly and noisy, while automated metrics provide only coarse, difficult-to-interpret signals. We introduce natural language unit tests, a paradigm that decomposes response quality into explicit, testable criteria, along with a unified scoring model, LMUnit, which combines multi-objective training across preferences, direct ratings, and natural language rationales. Through controlled human studies, we show this paradigm significantly improves inter-annotator agreement and enables more effective LLM development workflows. LMUnit achieves state-of-the-art performance on evaluation benchmarks (FLASK, BigGenBench) and competitive results on RewardBench. These results validate both our proposed paradigm and scoring model, suggesting a promising path forward for language model evaluation and development.

Method: Created PubHealthBench by extracting text from 687 UK government guidance documents and implementing automated pipeline for generating MCQA samples. Evaluated 24 LLMs on both MCQA and free-form response setups.

Result: Latest proprietary LLMs (GPT-4.5, GPT-4.1, o1) achieved >90% accuracy in MCQA, outperforming humans with cursory search. However, free-form response performance was lower with no model scoring >75%.

Conclusion: SOTA LLMs show promising accuracy for public health information retrieval but require additional safeguards/tools for free-form responses to ensure reliable real-world use.

Abstract: As Large Language Models (LLMs) become widely accessible, a detailed understanding of their knowledge within specific domains becomes necessary for successful real world use. This is particularly critical in the domains of medicine and public health, where failure to retrieve relevant, accurate, and current information could significantly impact UK residents. However, while there are a number of LLM benchmarks in the medical domain, currently little is known about LLM knowledge within the field of public health. To address this issue, this paper introduces a new benchmark, PubHealthBench, with over 8000 questions for evaluating LLMs’ Multiple Choice Question Answering (MCQA) and free form responses to public health queries. To create PubHealthBench we extract free text from 687 current UK government guidance documents and implement an automated pipeline for generating MCQA samples. Assessing 24 LLMs on PubHealthBench we find the latest proprietary LLMs (GPT-4.5, GPT-4.1 and o1) have a high degree of knowledge, achieving >90% accuracy in the MCQA setup, and outperform humans with cursory search engine use. However, in the free form setup we see lower performance with no model scoring >75%. Therefore, while there are promising signs that state of the art (SOTA) LLMs are an increasingly accurate source of public health information, additional safeguards or tools may still be needed when providing free form responses.

Method: The researchers systematically investigate the relationship between training data biases and preference model miscalibration across five idiosyncratic features: length, structure, jargon, sycophancy, and vagueness. Using controlled counterfactual pairs, they quantify the extent to which preference models favor responses with artificially magnified biases (skew). They then propose a post-training method based on counterfactual data augmentation (CDA) using synthesized contrastive examples to mitigate these issues.

Result: Preference models favor biased responses in >60% of instances, with high miscalibration (≈40%) compared to human preferences. Bias features show only mild negative correlations to human preference labels (mean r_human = -0.12) but moderately strong positive correlations with reward model labels (mean r_model = +0.36). Fine-tuning with CDA reduces average miscalibration from 39.4% to 32.5% and average absolute skew difference from 20.5% to 10.0%, while maintaining overall RewardBench performance.

Conclusion: Language models exhibit systematic miscalibration in preference judgments, overrelying on spurious cues from training data. The proposed counterfactual data augmentation method effectively reduces miscalibration and bias while maintaining overall performance, demonstrating that targeted debiasing can strengthen the reliability of preference models within standard alignment pipelines.

Abstract: Language models serve as proxies for human preference judgements in alignment and evaluation, yet they exhibit systematic miscalibration, prioritizing superficial patterns over substantive qualities. This bias manifests as overreliance on features like length, structure, and style, leading to issues like reward hacking and unreliable evaluations. However, the connection between training data artifacts and the miscalibrated preferences exhibited by models remains poorly understood. In this work, we systematically investigate the relationship between training data biases and preference model miscalibration across five idiosyncratic features of language model generations: length, structure, jargon, sycophancy and vagueness. Using controlled counterfactual pairs, we first quantify the extent to which preference models favor responses with artificially magnified biases (skew), finding this preference occurs in $>60%$ of instances, and model preferences show high miscalibration ($\approx 40%$) compared to human preferences. Notably, bias features only show mild negative correlations to human preference labels (mean $r_{\mathrm{human}} = -0.12$) but show moderately strong positive correlations with labels from a strong reward model (mean $r_{\mathrm{model}} = +0.36$), suggesting that models may overrely on spurious cues. To mitigate these issues, we propose a simple post-training method based on counterfactual data augmentation (CDA) using synthesized contrastive examples. Fine-tuning models with CDA reduces average miscalibration from $39.4%$ to $32.5%$ and average absolute skew difference from $20.5%$ to $10.0%$, while maintaining overall RewardBench performance, indicating that targeted debiasing can strengthen the reliability of preference models within standard alignment pipelines.

Method: Developed CounselBench with 100 mental health professionals: (1) CounselBench-EVAL with 2,000 expert evaluations of LLM answers on patient questions from CounselChat, rated across six clinically grounded dimensions with span-level annotations; (2) CounselBench-Adv with 120 expert-authored adversarial questions designed to trigger specific model issues.

Result: LLMs achieve high scores on several dimensions but exhibit recurring issues: unconstructive feedback, overgeneralization, limited personalization/relevance, and safety risks (especially unauthorized medical advice). LLM judges systematically overrate model responses and overlook safety concerns. Adversarial testing reveals consistent, model-specific failure patterns.

Conclusion: CounselBench establishes a clinically grounded framework for benchmarking LLMs in mental health QA, revealing significant gaps in LLM performance on realistic help-seeking scenarios and highlighting the need for expert evaluation to identify safety concerns that automated metrics miss.

Abstract: Medical question answering (QA) benchmarks often focus on multiple-choice or fact-based tasks, leaving open-ended answers to real patient questions underexplored. This gap is particularly critical in mental health, where patient questions often mix symptoms, treatment concerns, and emotional needs, requiring answers that balance clinical caution with contextual sensitivity. We present CounselBench, a large-scale benchmark developed with 100 mental health professionals to evaluate and stress-test large language models (LLMs) in realistic help-seeking scenarios. The first component, CounselBench-EVAL, contains 2,000 expert evaluations of answers from GPT-4, LLaMA 3, Gemini, and online human therapists on patient questions from the public forum CounselChat. Each answer is rated across six clinically grounded dimensions, with span-level annotations and written rationales. Expert evaluations show that while LLMs achieve high scores on several dimensions, they also exhibit recurring issues, including unconstructive feedback, overgeneralization, and limited personalization or relevance. Responses were frequently flagged for safety risks, most notably unauthorized medical advice. Follow-up experiments show that LLM judges systematically overrate model responses and overlook safety concerns identified by human experts. To probe failure modes more directly, we construct CounselBench-Adv, an adversarial dataset of 120 expert-authored mental health questions designed to trigger specific model issues. Expert evaluation of 1,080 responses from nine LLMs reveals consistent, model-specific failure patterns. Together, CounselBench establishes a clinically grounded framework for benchmarking LLMs in mental health QA.

Method: Proposes Internal Confidence, a training-free method that leverages self-evaluations across model layers and tokens to estimate query-level uncertainty before generating any tokens, avoiding generation costs.

Result: Internal Confidence outperforms several baselines in confidence quality while being computationally cheaper, and demonstrates benefits in adaptive inference settings by reducing inference costs while preserving performance for RAG and model cascading.

Conclusion: The proposed Internal Confidence method effectively detects LLM knowledge boundaries through query-level uncertainty estimation, enabling cost-effective adaptive inference strategies without additional training.

Abstract: It is important for Large Language Models (LLMs) to be aware of the boundary of their knowledge, distinguishing queries they can confidently answer from those that lie beyond their capabilities. Such awareness enables models to perform adaptive inference, such as invoking retrieval-augmented generation (RAG), engaging in slow and deep thinking, or abstaining from answering when appropriate. These mechanisms are key to developing efficient and trustworthy AI. In this work, we propose a method to detect knowledge boundaries via Query-Level Uncertainty, which estimates if a model is capable of answering a given query before generating any tokens, thus avoiding the generation cost. To this end, we propose a novel, training-free method called Internal Confidence, which leverages self-evaluations across layers and tokens to provide a reliable signal of uncertainty. Empirical studies on both factual question answering and mathematical reasoning tasks demonstrate that our Internal Confidence outperforms several baselines in quality of confidence while being computationally cheaper. Furthermore, we demonstrate its benefits in adaptive inference settings, showing that for RAG and model cascading it reduces inference costs while preserving overall performance.

Method: Proposes a method where users can add corrections on the fly during inference to improve recognition accuracy of challenging words. Corrections of substitution errors are used to enhance recognition of problematic words.

Result: Achieves relative improvement in biased word error rate between 22% and 34% compared to a text-based replacement method, while maintaining overall performance.

Conclusion: The proposed method effectively addresses the challenge of recognizing unseen words with pronunciation-orthography mismatches in ASR systems through user-provided corrections during inference.

Abstract: Neural sequence-to-sequence systems deliver state-of-the-art performance for automatic speech recognition. When using appropriate modeling units, e.g., byte-pair encoding, these systems are in principle open vocabulary systems. In practice, however, they often fail to recognize words not seen during training, e.g., named entities, acronyms, or domain-specific special words. To address this problem, many context biasing methods have been proposed; however, these methods may still struggle when they are unable to relate audio and corresponding text, e.g., in case of a pronunciation-orthography mismatch. We propose a method where corrections of substitution errors can be used to improve the recognition accuracy of such challenging words. Users can add corrections on the fly during inference. We show that with this method we get a relative improvement in biased word error rate between 22% and 34% compared to a text-based replacement method, while maintaining the overall performance.

Method: Systematic investigation of how entity coreference affects RAG performance, focusing on retrieval relevance, contextual understanding, and response quality. Comparative analysis of different pooling strategies in retrieval tasks and evaluation of coreference resolution’s impact on QA performance across different model sizes.

Result: Coreference resolution enhances retrieval effectiveness and improves QA performance. Mean pooling demonstrates superior context capturing ability after coreference resolution. Smaller models benefit more from the disambiguation process due to their limited inherent capacity for handling referential ambiguity.

Conclusion: Coreferential complexity poses significant challenges to RAG systems, and coreference resolution provides effective guidance for improving both retrieval and generation in knowledge-intensive AI applications.

Abstract: Retrieval-Augmented Generation (RAG) has emerged as a crucial framework in natural language processing (NLP), improving factual consistency and reducing hallucinations by integrating external document retrieval with large language models (LLMs). However, the effectiveness of RAG is often hindered by coreferential complexity in retrieved documents, introducing ambiguity that disrupts in-context learning. In this study, we systematically investigate how entity coreference affects both document retrieval and generative performance in RAG-based systems, focusing on retrieval relevance, contextual understanding, and overall response quality. We demonstrate that coreference resolution enhances retrieval effectiveness and improves question-answering (QA) performance. Through comparative analysis of different pooling strategies in retrieval tasks, we find that mean pooling demonstrates superior context capturing ability after applying coreference resolution. In QA tasks, we discover that smaller models benefit more from the disambiguation process, likely due to their limited inherent capacity for handling referential ambiguity. With these findings, this study aims to provide a deeper understanding of the challenges posed by coreferential complexity in RAG, providing guidance for improving retrieval and generation in knowledge-intensive AI applications.

Method: Uses off-by-one addition as a counterfactual task and applies circuit-style interpretability techniques like path patching to analyze internal computations, identifying a “function induction” mechanism.

Result: Identified a function induction mechanism that explains generalization from standard to off-by-one addition, showing it’s governed by multiple parallel attention heads and reused across various tasks including synthetic and algorithmic ones.

Conclusion: The findings reveal how reusable and composable structures within language models enable task-level generalization, providing deeper insights into the mechanisms behind in-context learning.

Abstract: Large language models demonstrate the intriguing ability to perform unseen tasks via in-context learning. However, it remains unclear what mechanisms inside the model drive such task-level generalization. In this work, we approach this question through the lens of off-by-one addition (i.e., 1+1=3, 2+2=5, 3+3=?), a two-step, counterfactual task with an unexpected +1 function as a second step. Leveraging circuit-style interpretability techniques such as path patching, we analyze the models’ internal computations behind their performance and present three key findings. First, we identify a mechanism that explains the model’s generalization from standard addition to off-by-one addition. It resembles the induction head mechanism described in prior work, yet operates at a higher level of abstraction; we therefore term it “function induction” in this work. Second, we show that the induction of the +1 function is governed by multiple attention heads in parallel, each of which emits a distinct piece of the +1 function. Finally, we find that this function induction mechanism is reused in a broader range of tasks, including synthetic tasks such as shifted multiple-choice QA and algorithmic tasks such as base-8 addition. Overall, our findings offer deeper insights into how reusable and composable structures within language models enable task-level generalization.

Method: Comprehensive multilingual benchmarking study across 9 languages (Arabic, Amharic, Chinese, English, German, Hindi, Russian, Spanish, Ukrainian). Inspired by machine translation evaluation, compares neural-based automatic metrics with LLM-as-a-judge approaches and experiments with task-specific fine-tuned models.

Result: Proposed metrics achieve significantly higher correlation with human judgments compared to baseline approaches. Provides actionable insights and practical guidelines for building robust multilingual evaluation pipelines for text detoxification and related TST tasks.

Conclusion: First comprehensive multilingual evaluation framework for text detoxification that improves correlation with human judgments and provides practical guidelines for building reliable multilingual evaluation pipelines.

Abstract: Despite notable advances in large language models (LLMs), reliable evaluation of text generation tasks such as text style transfer (TST) remains an open challenge. Existing research has shown that automatic metrics often correlate poorly with human judgments (Dementieva et al., 2024; Pauli et al., 2025), limiting our ability to assess model performance accurately. Furthermore, most prior work has focused primarily on English, while the evaluation of multilingual TST systems, particularly for text detoxification, remains largely underexplored. In this paper, we present the first comprehensive multilingual benchmarking study of evaluation metrics for text detoxification evaluation across nine languages: Arabic, Amharic, Chinese, English, German, Hindi, Russian, Spanish, and Ukrainian. Drawing inspiration from machine translation evaluation, we compare neural-based automatic metrics with LLM-as-a-judge approaches together with experiments on task-specific fine-tuned models. Our analysis reveals that the proposed metrics achieve significantly higher correlation with human judgments compared to baseline approaches. We also provide actionable insights and practical guidelines for building robust and reliable multilingual evaluation pipelines for text detoxification and related TST tasks.

Method: Created WebDS benchmark with 870 web-based data science tasks across 29 diverse websites including government data portals and news media. Tasks require complex, multi-step, tool-based operations across heterogeneous data formats. Evaluated current SOTA LLM agents and compared to human performance.

Result: Current LLM agents perform poorly on WebDS tasks. Browser Use (which achieves 80% on WebVoyager) completes only 15% of WebDS tasks. Humans achieve around 90% accuracy. Analysis reveals new failure modes: poor information grounding, repetitive behavior, and shortcut-taking.

Conclusion: WebDS provides a more robust and realistic testing ground for LLM-based data science, revealing substantial gaps between current agents and human performance, and setting the stage for advances in practically useful data science agents.

Abstract: Many real-world data science tasks involve complex web-based interactions: finding appropriate data available on the internet, synthesizing multimodal data from different locations, and producing summarized analyses. Existing web benchmarks often focus on simplistic interactions and often do not require diverse tool-using capabilities. Conversely, traditional data science benchmarks typically concentrate on static, highly structured datasets and do not assess end-to-end workflows that encompass data acquisition, cleaning, analysis, and insight generation. In response, we introduce WebDS, the first end-to-end web-based data science benchmark. It comprises 870 web-based data science tasks across 29 diverse websites from structured government data portals to unstructured news media, challenging agents to perform complex, multi-step, tool-based operations, across heterogeneous data formats, to better reflect the realities of modern data analytics. Evaluations of current SOTA LLM agents indicate significant performance gaps in accomplishing these tasks. For instance, Browser Use, which accomplishes $80%$ of tasks on WebVoyager, completes only 15% of tasks in WebDS, which our analysis suggests is due to new failure modes, such as poor information grounding, repetitive behavior and shortcut-taking that agents performing WebDS’s tasks display. By contrast, humans achieve around 90% accuracy, highlighting a substantial gap between current agents and human performance. By providing a more robust and realistic testing ground, WebDS sets the stage for significant advances in the development of practically useful LLM-based data science.

Method: Proposes ObfusQAte technique and ObfusQA framework with three obfuscation dimensions: (1) Named-Entity Indirection - indirect references to entities, (2) Distractor Indirection - misleading information, and (3) Contextual Overload - excessive irrelevant context. Creates comprehensive benchmark to evaluate LLM robustness.

Result: LLMs exhibit tendency to fail or generate hallucinated responses when confronted with increasingly nuanced obfuscated variations. The framework reveals limitations in LLM robustness and adaptability to complex language variations.

Conclusion: ObfusQA provides a comprehensive benchmark for evaluating LLM robustness against obfuscated questions, revealing significant limitations. The ObfusQAte technique is made publicly available to foster research in this direction.

Abstract: The rapid proliferation of Large Language Models (LLMs) has significantly contributed to the development of equitable AI systems capable of factual question-answering (QA). However, no known study tests the LLMs’ robustness when presented with obfuscated versions of questions. To systematically evaluate these limitations, we propose a novel technique, ObfusQAte, and leveraging the same, introduce ObfusQA, a comprehensive, first-of-its-kind framework with multi-tiered obfuscation levels designed to examine LLM capabilities across three distinct dimensions: (i) Named-Entity Indirection, (ii) Distractor Indirection, and (iii) Contextual Overload. By capturing these fine-grained distinctions in language, ObfusQA provides a comprehensive benchmark for evaluating LLM robustness and adaptability. Our study observes that LLMs exhibit a tendency to fail or generate hallucinated responses when confronted with these increasingly nuanced variations. To foster research in this direction, we make ObfusQAte publicly available.

Method: 1) Use Wikipedia articles as context, 2) Generate question/answer pairs using LLMs with answers appearing verbatim in articles, 3) Rephrase questions to hinder word matching, 4) Conduct human evaluation across 30 languages for fluency assessment.

Result: Created dataset with 1,220,757 samples across 306 languages; human evaluation showed mean fluency rating above “mostly natural” for all 30 evaluated languages; benchmark is sufficiently difficult with large performance discrepancies across languages.

Conclusion: MultiWikiQA provides a high-quality, challenging multilingual reading comprehension benchmark that enables evaluation of language models across diverse languages and reveals significant performance gaps.

Abstract: We introduce a new reading comprehension dataset, dubbed MultiWikiQA, which covers 306 languages and has 1,220,757 samples in total. We start with Wikipedia articles, which also provide the context for the dataset samples, and use an LLM to generate question/answer pairs related to the Wikipedia article, ensuring that the answer appears verbatim within the article. Next, the question is then rephrased to hinder simple word matching methods from performing well on the dataset. We conduct a crowdsourced human evaluation of the fluency of the generated questions, which included 156 respondents across 30 of the languages (both low- and high-resource). All 30 languages received a mean fluency rating above ``mostly natural’’, showing that the samples are of good quality. We evaluate 6 different language models, both decoder and encoder models of varying sizes, showing that the benchmark is sufficiently difficult and that there is a large performance discrepancy amongst the languages. Both the dataset and survey evaluations are publicly available.

Method: Adapts psychological appraisal models to argument mining, develops role-playing annotation setup where participants disclose emotions, causes, argument appraisal (pleasantness, familiarity, urgency, effort), convincingness, and demographic/personality data for both receiver and perceived sender.

Result: Created ContArgA corpus of 4000 annotations showing convincingness positively correlates with positive emotions (trust) and negatively with negative emotions (anger), with familiarity being particularly important appraisal variable.

Conclusion: The framework successfully models argument appraisal and its impact on convincingness, demonstrating the importance of subjective contextual factors in argument analysis beyond content alone.

Abstract: Emotions that somebody develops based on an argument do not only depend on the argument itself - they are also influenced by a subjective evaluation of the argument’s potential impact on the self. For instance, an argument to ban plastic bottles might cause fear of losing a job for a bottle industry worker, which lowers the convincingness - presumably independent of its content. While binary emotionality of arguments has been studied, such cognitive appraisal models have only been proposed in other subtasks of emotion analysis, but not in the context of arguments and their convincingness. To fill this research gap, we propose the Contextualized Argument Appraisal Framework to model the interplay between the sender, receiver, and argument. We adapt established appraisal models from psychology to argument mining, including argument pleasantness, familiarity, response urgency, and expected effort, as well as convincingness variables. To evaluate the framework and pave the way for computational modeling, we develop a novel role-playing-based annotation setup, mimicking real-world exposure to arguments. Participants disclose their emotion, explain the main cause, the argument appraisal, and the perceived convincingness. To consider the subjective nature of such annotations, we also collect demographic data and personality traits of both the participants and ask them to disclose the same variables for their perception of the argument sender. The analysis of the resulting ContArgA corpus of 4000 annotations reveals that convincingness is positively correlated with positive emotions (e.g., trust) and negatively correlated with negative emotions (e.g., anger). The appraisal variables particularly point to the importance of the annotator’s familiarity with the argument.

Method: Proposes a novel user simulator architecture that simulates four categories of non-collaborative behaviors: requesting unavailable services, digressing into tangential conversations, expressing impatience, and providing incomplete utterances. The framework maintains task completion while introducing realistic challenges.

Result: Experiments on MultiWOZ and τ-bench show significant performance degradation in state-of-the-art tool agents when encountering non-collaborative users, revealing weaknesses like escalated hallucinations and dialogue breakdowns under each non-collaborative condition.

Conclusion: Tool agents need improved robustness to handle diverse real-world user behaviors. The released extensible simulation framework helps the community develop and stress-test agents under realistic conditions within their service domains.

Abstract: Tool agents interact with users through multi-turn dialogues to accomplish various tasks. Recent studies have adopted user simulation methods to develop these agents in multi-turn settings. However, existing user simulators tend to be agent-friendly, exhibiting only cooperative behaviors, failing to train and test agents against non-collaborative users in the real world. We propose a novel user simulator architecture that simulates four categories of non-collaborative behaviors: requesting unavailable services, digressing into tangential conversations, expressing impatience, and providing incomplete utterances. Our user simulator can simulate challenging and natural non-collaborative behaviors while reliably delivering all intents and information necessary to accomplish the task. Our experiments on MultiWOZ and τ-bench reveal significant performance degradation in state-of-the-art tool agents when encountering non-collaborative users, as well as agent weaknesses under each non-collaborative condition such as escalated hallucinations and dialogue breakdowns. Our findings point to the need for methods that can improve agent robustness to the wide range of user behaviors encountered in deployment. We release the extensible simulation framework to help the community develop and stress-test tool agents under realistic conditions within their own service domains. Our code is available at https://github.com/holi-lab/NCUser.

Method: Introduced Personalized Deep Research Bench (PDR-Bench) with 50 diverse research tasks across 10 domains paired with 25 authentic user profiles combining structured persona attributes with dynamic real-world contexts. Proposed PQR Evaluation Framework measuring Personalization Alignment, Content Quality, and Factual Reliability.

Result: Experiments on various systems revealed current capabilities and limitations in handling personalized deep research, establishing a foundation for developing personalized AI research assistants.

Conclusion: PDR-Bench provides the first rigorous benchmark for evaluating personalization in Deep Research Agents, enabling development of truly personalized AI research assistants.

Abstract: Deep Research Agents (DRAs) can autonomously conduct complex investigations and generate comprehensive reports, demonstrating strong real-world potential. However, existing evaluations mostly rely on close-ended benchmarks, while open-ended deep research benchmarks remain scarce and typically neglect personalized scenarios. To bridge this gap, we introduce Personalized Deep Research Bench (PDR-Bench), the first benchmark for evaluating personalization in DRAs. It pairs 50 diverse research tasks across 10 domains with 25 authentic user profiles that combine structured persona attributes with dynamic real-world contexts, yielding 250 realistic user-task queries. To assess system performance, we propose the PQR Evaluation Framework, which jointly measures Personalization Alignment, Content Quality, and Factual Reliability. Our experiments on a range of systems highlight current capabilities and limitations in handling personalized deep research. This work establishes a rigorous foundation for developing and evaluating the next generation of truly personalized AI research assistants.

Method: Elicitation-Then-Calibration (EliCal) framework: 1) First elicits internal confidence using inexpensive self-consistency supervision, 2) Then calibrates this confidence with a small set of correctness annotations. Also introduces HonestyBench benchmark covering ten free-form QA datasets with 560k training and 70k evaluation instances.

Result: EliCal achieves near-optimal alignment with only 1k correctness annotations (0.18% of full supervision) and shows better alignment performance on unseen MMLU tasks than calibration-only baseline.

Conclusion: EliCal offers a scalable solution toward universal honesty alignment in LLMs by dramatically reducing annotation costs while maintaining effectiveness, making honesty alignment more practical for real-world deployment.

Abstract: Honesty alignment-the ability of large language models (LLMs) to recognize their knowledge boundaries and express calibrated confidence-is essential for trustworthy deployment. Existing methods either rely on training-free confidence estimation (e.g., token probabilities, self-consistency) or training-based calibration with correctness annotations. While effective, achieving universal honesty alignment with training-based calibration requires costly, large-scale labeling. To support annotation-efficient training, we introduce Elicitation-Then-Calibration (EliCal), a two-stage framework that first elicits internal confidence using inexpensive self-consistency supervision, then calibrates this confidence with a small set of correctness annotations. To support a large-scale study, we release HonestyBench, a benchmark covering ten free-form QA datasets with 560k training and 70k evaluation instances annotated with correctness and self-consistency signals. Experiments show that EliCal achieves near-optimal alignment with only 1k correctness annotations (0.18% of full supervision) and better alignment performance on unseen MMLU tasks than the calibration-only baseline, offering a scalable solution toward universal honesty alignment in LLMs.

Method: Two-step approach: (1) Study citation failure using CITECONTROL benchmark that systematically varies relation between response and evidence, (2) Propose CITENTION framework integrating generative, attention-based, and retrieval-based methods to improve citation quality efficiently.

Result: Experiments show citation failures increase with relational complexity. CITENTION framework demonstrates substantial citation improvements on CITECONTROL benchmark and in transfer settings.

Conclusion: Citation failure is a distinct problem from response failure that can be systematically studied and mitigated. Combining multiple citation methods improves performance, and the proposed CITENTION framework effectively addresses citation quality issues.

Abstract: Citations from LLM-based RAG systems are supposed to simplify response verification. However, this goal is undermined in cases of citation failure, where a model generates a helpful response, but fails to generate citations to complete evidence. In contrast to previous work, we propose to disentangle this from response failure, where the response itself is flawed, and citing complete evidence is impossible. To address citation failure, this work follows a two-step approach: (1) We study when citation failure occurs and (2) how it can be mitigated efficiently. For step 1, we extend prior work by investigating how the relation between response and evidence affects citation quality. We introduce CITECONTROL, a benchmark that systematically varies this relation to enable the analysis of failure modes. Experiments show that failures increase with relational complexity and suggest that combining citation methods could improve performance, motivating step 2. To study the efficient improvement of LLM citation, we propose CITENTION, a framework integrating generative, attention-based, and retrieval-based methods. Results demonstrate substantial citation improvements on CITECONTROL and in transfer settings. We make our data and code publicly available.

Method: Created MuSaG dataset with 33 minutes of German TV show clips, manually selected and human-annotated for sarcasm across text, audio, and video modalities. Benchmarked 9 open-source and commercial models across unimodal and multimodal settings.

Result: Humans rely heavily on audio cues for sarcasm detection, while models perform best on text. This reveals a gap in current multimodal models’ ability to effectively use audio-visual information.

Conclusion: MuSaG dataset addresses the need for German multimodal sarcasm data and highlights limitations in current multimodal models, motivating development of better audio-visual integration for realistic scenarios.

Abstract: Sarcasm is a complex form of figurative language in which the intended meaning contradicts the literal one. Its prevalence in social media and popular culture poses persistent challenges for natural language understanding, sentiment analysis, and content moderation. With the emergence of multimodal large language models, sarcasm detection extends beyond text and requires integrating cues from audio and vision. We present MuSaG, the first German multimodal sarcasm detection dataset, consisting of 33 minutes of manually selected and human-annotated statements from German television shows. Each instance provides aligned text, audio, and video modalities, annotated separately by humans, enabling evaluation in unimodal and multimodal settings. We benchmark nine open-source and commercial models, spanning text, audio, vision, and multimodal architectures, and compare their performance to human annotations. Our results show that while humans rely heavily on audio in conversational settings, models perform best on text. This highlights a gap in current multimodal models and motivates the use of MuSaG for developing models better suited to realistic scenarios. We release MuSaG publicly to support future research on multimodal sarcasm detection and human-model alignment.

Method: Introduces Agent Data Protocol (ADP) - a lightweight representation language that serves as an “interlingua” to unify diverse agent datasets. ADP captures various tasks (API/tool use, browsing, coding, software engineering, agentic workflows) while remaining simple to parse. Unified 13 existing datasets into ADP format and converted to training-ready formats for multiple agent frameworks.

Result: SFT on standardized ADP data achieved ~20% average performance gain over base models, delivering SOTA or near-SOTA performance on coding, browsing, tool use, and research benchmarks without domain-specific tuning.

Conclusion: ADP lowers barriers to standardized, scalable, and reproducible agent training by providing a unified data protocol that bridges fragmented agent datasets and training pipelines.

Abstract: Public research results on large-scale supervised finetuning of AI agents remain relatively rare, since the collection of agent training data presents unique challenges. In this work, we argue that the bottleneck is not a lack of underlying data sources, but that a large variety of data is fragmented across heterogeneous formats, tools, and interfaces. To this end, we introduce the agent data protocol (ADP), a light-weight representation language that serves as an “interlingua” between agent datasets in diverse formats and unified agent training pipelines downstream. The design of ADP is expressive enough to capture a large variety of tasks, including API/tool use, browsing, coding, software engineering, and general agentic workflows, while remaining simple to parse and train on without engineering at a per-dataset level. In experiments, we unified a broad collection of 13 existing agent training datasets into ADP format, and converted the standardized ADP data into training-ready formats for multiple agent frameworks. We performed SFT on these data, and demonstrated an average performance gain of ~20% over corresponding base models, and delivers state-of-the-art or near-SOTA performance on standard coding, browsing, tool use, and research benchmarks, without domain-specific tuning. All code and data are released publicly, in the hope that ADP could help lower the barrier to standardized, scalable, and reproducible agent training.

Method: Created CareMedEval dataset from authentic French medical student exams, containing 534 questions based on 37 scientific articles. Benchmarked state-of-the-art generalist and biomedical-specialized LLMs under various context conditions, including with and without intermediate reasoning tokens.

Result: Current LLMs struggle significantly, with open and commercial models failing to exceed 0.5 Exact Match Rate. Generating intermediate reasoning tokens improves results, but models remain particularly challenged on questions about study limitations and statistical analysis.

Conclusion: CareMedEval provides a challenging benchmark exposing current LLM limitations in biomedical critical reasoning, paving the way for future development of automated support for critical appraisal in specialized domains.

Abstract: Critical appraisal of scientific literature is an essential skill in the biomedical field. While large language models (LLMs) can offer promising support in this task, their reliability remains limited, particularly for critical reasoning in specialized domains. We introduce CareMedEval, an original dataset designed to evaluate LLMs on biomedical critical appraisal and reasoning tasks. Derived from authentic exams taken by French medical students, the dataset contains 534 questions based on 37 scientific articles. Unlike existing benchmarks, CareMedEval explicitly evaluates critical reading and reasoning grounded in scientific papers. Benchmarking state-of-the-art generalist and biomedical-specialized LLMs under various context conditions reveals the difficulty of the task: open and commercial models fail to exceed an Exact Match Rate of 0.5 even though generating intermediate reasoning tokens considerably improves the results. Yet, models remain challenged especially on questions about study limitations and statistical analysis. CareMedEval provides a challenging benchmark for grounded reasoning, exposing current LLM limitations and paving the way for future development of automated support for critical appraisal.

Method: The researchers used two Dutch data sources: cancer patient storytelling interviews and online forum data. They applied state-of-the-art LLMs with different prompting strategies including chain-of-thought reasoning, few-shot learning, and self-prompting. A human-in-the-loop setup was used to verify extracted metaphors, resulting in the HealthQuote.NL corpus.

Result: Created the HealthQuote.NL corpus of Dutch cancer patient metaphors extracted using LLMs with various prompting strategies. The human-in-the-loop verification ensured quality, and the resource is shared publicly with prompts and related materials.

Conclusion: LLMs can effectively extract healthcare metaphors from patient data, and the resulting corpus can support better patient care, improved communication, and personalized care pathways in oncology settings.

Abstract: Metaphors and metaphorical language (MLs) play an important role in healthcare communication between clinicians, patients, and patients’ family members. In this work, we focus on Dutch language data from cancer patients. We extract metaphors used by patients using two data sources: (1) cancer patient storytelling interview data and (2) online forum data, including patients’ posts, comments, and questions to professionals. We investigate how current state-of-the-art large language models (LLMs) perform on this task by exploring different prompting strategies such as chain of thought reasoning, few-shot learning, and self-prompting. With a human-in-the-loop setup, we verify the extracted metaphors and compile the outputs into a corpus named HealthQuote.NL. We believe the extracted metaphors can support better patient care, for example shared decision making, improved communication between patients and clinicians, and enhanced patient health literacy. They can also inform the design of personalized care pathways. We share prompts and related resources at https://github.com/4dpicture/HealthQuote.NL

Method: The authors use the ContArgA corpus with annotated emotions and appraisals. They perform zero-shot prompting experiments to evaluate the importance of gold-annotated and predicted emotions and appraisals for predicting subjective convincingness labels. They systematically compare emotion models for convincingness prediction.

Result: The results show that while categorical emotion information does improve convincingness prediction, the improvement is more pronounced with appraisals. Appraisal-based approaches outperform basic emotion category approaches for predicting argument convincingness.

Conclusion: This work presents the first systematic comparison between emotion models for convincingness prediction, demonstrating the advantage of appraisal theories over categorical emotion approaches. The findings provide insights for both theoretical and practical applications in computational argumentation.

Abstract: The convincingness of an argument does not only depend on its structure (logos), the person who makes the argument (ethos), but also on the emotion that it causes in the recipient (pathos). While the overall intensity and categorical values of emotions in arguments have received considerable attention in the research community, we argue that the emotion an argument evokes in a recipient is subjective. It depends on the recipient’s goals, standards, prior knowledge, and stance. Appraisal theories lend themselves as a link between the subjective cognitive assessment of events and emotions. They have been used in event-centric emotion analysis, but their suitability for assessing argument convincingness remains unexplored. In this paper, we evaluate whether appraisal theories are suitable for emotion analysis in arguments by considering subjective cognitive evaluations of the importance and impact of an argument on its receiver. Based on the annotations in the recently published ContArgA corpus, we perform zero-shot prompting experiments to evaluate the importance of gold-annotated and predicted emotions and appraisals for the assessment of the subjective convincingness labels. We find that, while categorical emotion information does improve convincingness prediction, the improvement is more pronounced with appraisals. This work presents the first systematic comparison between emotion models for convincingness prediction, demonstrating the advantage of appraisals, providing insights for theoretical and practical applications in computational argumentation.

Method: Developed Basque image-text datasets for training and evaluation. Used two LLM backbones: Llama-3.1-Instruct and Basque-adapted Latxa. Explored different data mixtures to determine optimal Basque data ratios for effective MLLM training.

Result: Found that only ~20% Basque multimodal data is sufficient for solid performance on Basque benchmarks. Surprisingly, a Basque-specific instructed LLM backbone is not required - general LLMs work well when fine-tuned with minimal Basque data.

Conclusion: Provides a practical approach for developing MLLMs for low-resource languages with minimal language-specific data. Releases resources openly to facilitate similar development for other underrepresented languages.

Abstract: Current Multimodal Large Language Models exhibit very strong performance for several demanding tasks. While commercial MLLMs deliver acceptable performance in low-resource languages, comparable results remain unattained within the open science community. In this paper, we aim to develop a strong MLLM for a low-resource language, namely Basque. For that purpose, we develop our own training and evaluation image-text datasets. Using two different Large Language Models as backbones, the Llama-3.1-Instruct model and a Basque-adapted variant called Latxa, we explore several data mixtures for training. We show that: i) low ratios of Basque multimodal data (around 20%) are already enough to obtain solid results on Basque benchmarks, and ii) contrary to expected, a Basque instructed backbone LLM is not required to obtain a strong MLLM in Basque. Our results pave the way to develop MLLMs for other low-resource languages by openly releasing our resources.

Method: Reformulates extraction as constrained sequence labeling using small language models (SLMs), creating WebMainBench benchmark with 7,809 annotated pages, and pre-trains models on Dripper-curated corpora.

Result: Dripper-0.6B achieves 3.08 pages/sec throughput on single A100, outperforms heuristics like Trafilatura and rivals massive models (DeepSeek-V3.2, GPT-5, Gemini-2.5-Pro), with pre-trained 1B model showing superior downstream task performance.

Conclusion: Dripper provides optimal efficiency-accuracy trade-off for web content extraction, enabling high-quality corpus construction for training datasets, with open-sourced weights and codebase.

Abstract: High-quality main content extraction from web pages is a critical prerequisite for constructing large-scale training corpora. While traditional heuristic extractors are efficient, they lack the semantic reasoning required to handle the structural heterogeneity of the modern web. Conversely, well-pretrained generative Large Language Models (LLMs) offer superior document comprehension but are prohibited by excessive computational costs, limited context windows, and hallucination risks when applied at web scale. We present \textbf{Dripper}, a lightweight framework that resolves these bottlenecks through four contributions: (1) We reformulate extraction as a \textbf{constrained sequence labeling} task using SLMs (Small Language Models). This paradigm eliminates generative hallucinations and achieves exceptional efficiency, reaching a throughput of 3.08 pages per second on a single A100 GPU. (2) We construct \textbf{WebMainBench}, a rigorous benchmark of 7,809 human-annotated pages covering 5,434 unique domains and multiple languages. Evaluations show our Dripper-0.6B model \textbf{outperforms} heuristics like Trafilatura and rivals massive models like DeepSeek-V3.2(685B), GPT-5 and Gemini-2.5-Pro, offering an optimal efficiency-accuracy trade-off. (3) We demonstrate infrastructural value by \textbf{pre-training a 1B model} on a Dripper-curated corpus (63B tokens). This model significantly outperforms baselines in downstream tasks, proving the critical role of extraction quality and the effectiveness of our framework. (4) We \textbf{open-source} the Dripper-0.6B weights and codebase to facilitate the construction of high-quality datasets.

Method: Cannot determine method due to failed paper fetch

Result: Cannot determine results due to failed paper fetch

Conclusion: Cannot determine conclusion due to failed paper fetch

Abstract: Failed to fetch summary for 2512.08440: Page request resulted in HTTP 429 (https://export.arxiv.org/api/query?search_query=&id_list=2512.08440&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 2601.19933: Page request resulted in HTTP 429 (https://export.arxiv.org/api/query?search_query=&id_list=2601.19933&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 2602.04755: Page request resulted in HTTP 429 (https://export.arxiv.org/api/query?search_query=&id_list=2602.04755&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 2602.06412: Page request resulted in HTTP 429 (https://export.arxiv.org/api/query?search_query=&id_list=2602.06412&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 2509.13615: Page request resulted in HTTP 429 (https://export.arxiv.org/api/query?search_query=&id_list=2509.13615&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 2602.16852: Page request resulted in HTTP 429 (https://export.arxiv.org/api/query?search_query=&id_list=2602.16852&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 2602.22730: Page request resulted in HTTP 429 (https://export.arxiv.org/api/query?search_query=&id_list=2602.22730&sortBy=relevance&sortOrder=descending&start=0&max_results=100)

Method: Unable to determine method due to fetch failure

Result: Unable to determine results due to fetch failure

Conclusion: Unable to determine conclusion due to fetch failure

Abstract: Failed to fetch summary for 2603.00917: Page request resulted in HTTP 429 (https://export.arxiv.org/api/query?search_query=&id_list=2603.00917&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 2510.09580: Page request resulted in HTTP 429 (https://export.arxiv.org/api/query?search_query=&id_list=2510.09580&sortBy=relevance&sortOrder=descending&start=0&max_results=100)

Method: Cannot determine method without access to paper content.

Result: Cannot determine results without access to paper content.

Conclusion: Cannot draw conclusions without access to paper content.

Abstract: Failed to fetch summary for 2603.01266: Page request resulted in HTTP 429 (https://export.arxiv.org/api/query?search_query=&id_list=2603.01266&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 2510.09782: Page request resulted in HTTP 429 (https://export.arxiv.org/api/query?search_query=&id_list=2510.09782&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 2603.01930: Page request resulted in HTTP 429 (https://export.arxiv.org/api/query?search_query=&id_list=2603.01930&sortBy=relevance&sortOrder=descending&start=0&max_results=100)

Method: Cannot determine method without access to the paper content.

Result: Cannot determine results without access to the paper content.

Conclusion: Cannot determine conclusion without access to the paper content.

Abstract: Failed to fetch summary for 2603.02353: Page request resulted in HTTP 429 (https://export.arxiv.org/api/query?search_query=&id_list=2603.02353&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 retrieval failed due to rate limiting

Conclusion: Cannot provide analysis due to technical limitations in accessing the paper

Abstract: Failed to fetch summary for 2603.02873: Page request resulted in HTTP 429 (https://export.arxiv.org/api/query?search_query=&id_list=2603.02873&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 2603.02909: Page request resulted in HTTP 429 (https://export.arxiv.org/api/query?search_query=&id_list=2603.02909&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 analyze paper due to technical access issues

Abstract: Failed to fetch summary for 2312.17505: Page request resulted in HTTP 429 (https://export.arxiv.org/api/query?search_query=&id_list=2312.17505&sortBy=relevance&sortOrder=descending&start=0&max_results=100)

Method: Multi-agent framework where code agents autonomously evolve existing math problems into more complex variations, with validation mechanisms to ensure solvability and increased difficulty of generated problems.

Result: Code agents can synthesize new, solvable problems that are structurally distinct from and more challenging than originals, given sufficient test-time exploration.

Conclusion: Code-driven agents can serve as a viable mechanism for synthesizing high-difficulty mathematical reasoning problems within scalable computational environments.

Abstract: As large language models (LLMs) advance their mathematical capabilities toward the IMO level, the scarcity of challenging, high-quality problems for training and evaluation has become a significant bottleneck. Simultaneously, recent code agents have demonstrated sophisticated skills in agentic coding and reasoning, suggesting that code execution can serve as a scalable environment for mathematical experimentation. In this paper, we investigate the potential of code agents to autonomously evolve existing math problems into more complex variations. We introduce a multi-agent framework designed to perform problem evolution while validating the solvability and increased difficulty of the generated problems. Our experiments demonstrate that, given sufficient test-time exploration, code agents can synthesize new, solvable problems that are structurally distinct from and more challenging than the originals. This work provides empirical evidence that code-driven agents can serve as a viable mechanism for synthesizing high-difficulty mathematical reasoning problems within scalable computational environments. Our data is available at https://github.com/TarferSoul/Code2Math.

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 2601.15160: Page request resulted in HTTP 429 (https://export.arxiv.org/api/query?search_query=&id_list=2601.15160&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 2405.15374: Page request resulted in HTTP 429 (https://export.arxiv.org/api/query?search_query=&id_list=2405.15374&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 conclusion due to failed API request

Abstract: Failed to fetch summary for 2502.01534: Page request resulted in HTTP 429 (https://export.arxiv.org/api/query?search_query=&id_list=2502.01534&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 2504.07109: Page request resulted in HTTP 429 (https://export.arxiv.org/api/query?search_query=&id_list=2504.07109&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 2602.10625: Page request resulted in HTTP 429 (https://export.arxiv.org/api/query?search_query=&id_list=2602.10625&sortBy=relevance&sortOrder=descending&start=0&max_results=100)

Method: Cannot determine method as abstract retrieval failed

Result: Cannot determine results as abstract retrieval failed

Conclusion: Cannot draw conclusions without access to paper content

Abstract: Failed to fetch summary for 2504.08714: Page request resulted in HTTP 429 (https://export.arxiv.org/api/query?search_query=&id_list=2504.08714&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 2505.02888: Page request resulted in HTTP 429 (https://export.arxiv.org/api/query?search_query=&id_list=2505.02888&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 2505.10118: Page request resulted in HTTP 429 (https://export.arxiv.org/api/query?search_query=&id_list=2505.10118&sortBy=relevance&sortOrder=descending&start=0&max_results=100)

Method: Cannot determine method without access to paper content

Result: Cannot determine results without access to paper content

Conclusion: Cannot determine conclusion without access to paper content

Abstract: Failed to fetch summary for 2505.23783: Page request resulted in HTTP 429 (https://export.arxiv.org/api/query?search_query=&id_list=2505.23783&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 2509.06415: Page request resulted in HTTP 429 (https://export.arxiv.org/api/query?search_query=&id_list=2509.06415&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 2510.14936: Page request resulted in HTTP 429 (https://export.arxiv.org/api/query?search_query=&id_list=2510.14936&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 2510.15040: Page request resulted in HTTP 429 (https://export.arxiv.org/api/query?search_query=&id_list=2510.15040&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 2510.22739: Page request resulted in HTTP 429 (https://export.arxiv.org/api/query?search_query=&id_list=2510.22739&sortBy=relevance&sortOrder=descending&start=0&max_results=100)

Method: Cannot determine method as abstract is unavailable due to rate limiting error

Result: Cannot determine results as abstract is unavailable due to rate limiting error

Conclusion: Cannot draw conclusions as abstract is unavailable due to rate limiting error

Abstract: Failed to fetch summary for 2512.15792: Page request resulted in HTTP 429 (https://export.arxiv.org/api/query?search_query=&id_list=2512.15792&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 2512.20760: Page request resulted in HTTP 429 (https://export.arxiv.org/api/query?search_query=&id_list=2512.20760&sortBy=relevance&sortOrder=descending&start=0&max_results=100)

Method: Cannot determine method without access to the paper content.

Result: Cannot determine results without access to the paper content.

Conclusion: Cannot determine conclusion without access to the paper content.

Abstract: Failed to fetch summary for 2601.04646: Page request resulted in HTTP 429 (https://export.arxiv.org/api/query?search_query=&id_list=2601.04646&sortBy=relevance&sortOrder=descending&start=0&max_results=100)

Method: Cannot determine method as abstract retrieval failed

Result: Cannot determine results as abstract retrieval failed

Conclusion: Cannot draw conclusions without access to paper content

Abstract: Failed to fetch summary for 2602.05630: Page request resulted in HTTP 429 (https://export.arxiv.org/api/query?search_query=&id_list=2602.05630&sortBy=relevance&sortOrder=descending&start=0&max_results=100)

Method: Proposes mHC-HSI with three key innovations: 1) clustering-guided Mamba module for spatial-spectral feature learning within mHC framework, 2) new residual matrix implementation as soft cluster membership maps for explainability, 3) physically-meaningful spectral band grouping as parallel streams in mHC.

Result: Tested on benchmark datasets, the proposed model improves classification accuracy compared to state-of-the-art methods while enhancing model explainability.

Conclusion: The mHC-HSI model successfully adapts the mHC framework for hyperspectral image classification, achieving both improved performance and better interpretability through clustering guidance and physical spectral knowledge integration.

Abstract: Recently, DeepSeek has invented the manifold-constrained hyper-connection (mHC) approach which has demonstrated significant improvements over the traditional residual connection in deep learning models \cite{xie2026mhc}. Nevertheless, this approach has not been tailor-designed for improving hyperspectral image (HSI) classification. This paper presents a clustering-guided mHC Mamba model (mHC-HSI) for enhanced HSI classification, with the following contributions. First, to improve spatial-spectral feature learning, we design a novel clustering-guided Mamba module, based on the mHC framework, that explicitly learns both spatial and spectral information in HSI. Second, to decompose the complex and heterogeneous HSI into smaller clusters, we design a new implementation of the residual matrix in mHC, which can be treated as soft cluster membership maps, leading to improved explainability of the mHC approach. Third, to leverage the physical spectral knowledge, we divide the spectral bands into physically-meaningful groups and use them as the “parallel streams” in mHC, leading to a physically-meaningful approach with enhanced interpretability. The proposed approach is tested on benchmark datasets in comparison with the state-of-the-art methods, and the results suggest that the proposed model not only improves the accuracy but also enhances the model explainability. Code is available here: https://github.com/GSIL-UCalgary/mHC_HyperSpectral

Method: Introduces a counterfactual evaluation framework using real, blank, and shuffled images across four medical VQA benchmarks (PathVQA, PMC-VQA, SLAKE, VQA-RAD). Measures Visual Reliance Score (VRS), Image Sensitivity (IS), and introduces Hallucinated Visual Reasoning Rate (HVRR) to detect visual claims despite image-invariant answers.

Result: RLVR improves accuracy while degrading visual grounding: text-only RLVR achieves negative VRS on PathVQA (-0.09), performing better with mismatched images. Image-text RLVR reduces image sensitivity to 39.8% overall. On VQA-RAD, both achieve 63% accuracy differently: text-only retains 81% performance with blank images, while image-text shows only 29% image sensitivity. Models generate visual claims in 68-74% of responses, with 38-43% being ungrounded (HVRR).

Conclusion: Accuracy-only rewards enable shortcut exploitation in multimodal models. Progress requires grounding-aware evaluation protocols and training objectives that explicitly enforce visual dependence, moving beyond accuracy-only metrics.

Abstract: Recent work shows that text-only reinforcement learning with verifiable rewards (RLVR) can match or outperform image-text RLVR on multimodal medical VQA benchmarks, suggesting current evaluation protocols may fail to measure causal visual dependence. We introduce a counterfactual evaluation framework using real, blank, and shuffled images across four medical VQA benchmarks: PathVQA, PMC-VQA, SLAKE, and VQA-RAD. Beyond accuracy, we measure Visual Reliance Score (VRS), Image Sensitivity (IS), and introduce Hallucinated Visual Reasoning Rate (HVRR) to detect cases where models generate visual claims despite producing image-invariant answers. Our findings reveal that RLVR improves accuracy while degrading visual grounding: text-only RLVR achieves negative VRS on PathVQA (-0.09), performing better with mismatched images, while image-text RLVR reduces image sensitivity to 39.8% overall despite improving accuracy. On VQA-RAD, both variants achieve 63% accuracy through different mechanisms: text-only RLVR retains 81% performance with blank images, while image-text RLVR shows only 29% image sensitivity. Models generate visual claims in 68-74% of responses, yet 38-43% are ungrounded (HVRR). These findings demonstrate that accuracy-only rewards enable shortcut exploitation, and progress requires grounding-aware evaluation protocols and training objectives that explicitly enforce visual dependence.

Method: Introduces Proact-VL framework that shapes multimodal language models into proactive real-time interactive agents. Uses gaming scenarios (commentator and guide) for evaluation, and creates Live Gaming Benchmark dataset with three scenarios: solo commentary, co-commentary, and user guidance.

Result: Proact-VL achieves superior response latency and quality while maintaining strong video understanding capabilities, demonstrating practicality for real-time interactive applications.

Conclusion: The framework successfully enables multimodal language models to function as proactive, real-time interactive agents with human-like environment perception and interaction capabilities.

Abstract: Proactive and real-time interactive experiences are essential for human-like AI companions, yet face three key challenges: (1) achieving low-latency inference under continuous streaming inputs, (2) autonomously deciding when to respond, and (3) controlling both quality and quantity of generated content to meet real-time constraints. In this work, we instantiate AI companions through two gaming scenarios, commentator and guide, selected for their suitability for automatic evaluation. We introduce the Live Gaming Benchmark, a large-scale dataset with three representative scenarios: solo commentary, co-commentary, and user guidance, and present Proact-VL, a general framework that shapes multimodal language models into proactive, real-time interactive agents capable of human-like environment perception and interaction. Extensive experiments show Proact-VL achieves superior response latency and quality while maintaining strong video understanding capabilities, demonstrating its practicality for real-time interactive applications.

Method: 1) AV-UIE v2 dataset: 222K samples across 17 datasets and 7 tasks with explicit reasoning processes; 2) Unified interface to align heterogeneous task formulations; 3) Interaction-aware LoRA (I-LoRA) that models inter-task relationships via dynamic routing to coordinate distinct audio-visual interaction patterns.

Result: Crab⁺ covers broader tasks than existing unified models while outperforming specialized models on various benchmarks. Successfully reversed negative transfer trend, achieving positive transfer where multi-task learning surpasses single-task baselines in nearly 88% of tasks.

Conclusion: Crab⁺ represents a robust step toward holistic audio-visual scene understanding by addressing task heterogeneity through explicit cooperation from both data and model perspectives, validated across diverse AV-LLM paradigms.

Abstract: Developing Audio-Visual Large Language Models (AV-LLMs) for unified scene understanding is pivotal in multimodal intelligence. While instruction tuning enables pre-trained models with multi-task abilities, we observe that conventional multi-task unification methods often suffer from severe negative transfer, where nearly 55% of tasks degrade compared to single-task training. We attribute this phenomenon to audio-visual task heterogeneity, characterized by disparate task granularity and divergent capability demands, which lead to negative interference under joint training. To tackle this, we present Crab$^{+}$, a scalable and unified audio-visual scene understanding model that addresses task heterogeneity through explicit cooperation from both data and model perspectives. On the data side, we introduce AV-UIE v2, a comprehensive Audio-Visual Unified Instruction-tuning dataset with Explicit reasoning processes. It contains approximately 222K samples spanning 17 datasets and 7 tasks, enabling the model to capture cross-task relationships at different levels of granularity. On the model side, we design a unified interface to align heterogeneous task formulations, and propose Interaction-aware LoRA (I-LoRA), which explicitly models inter-task relationships via dynamic routing to coordinate distinct audio-visual interaction patterns, mitigating parameter interference. Extensive experiments show Crab$^{+}$ covers broader tasks than existing unified models while outperforming specialized models on various benchmarks. We successfully reverse the negative transfer trend, achieving positive transfer where multi-task learning surpasses single-task baselines in nearly 88% of tasks. These results hold across diverse AV-LLM paradigms and are validated through in-depth visualization, positioning Crab$^{+}$ as a robust step towards holistic audio-visual scene understanding.

Method: Developed a multi-scenario field imaging framework: 1) For individual aggregates - designed imaging system with segmentation and volume estimation; 2) For 2D stockpile analysis - automated instance segmentation and morphological analysis; 3) For 3D stockpile analysis - integrated 3D Reconstruction-Segmentation-Completion (RSC-3D) approach using multi-view reconstruction, instance segmentation networks, and shape completion networks trained on synthetic datasets from a 3D aggregate particle library.

Result: The integrated 3D approach demonstrated good performance on real stockpiles, successfully capturing and predicting unseen sides of aggregates when validated with ground-truth data.

Conclusion: The dissertation presents a comprehensive field imaging framework that addresses limitations of existing methods, enabling automated morphological characterization of aggregates across various scenarios including individual particles, 2D stockpiles, and complex 3D stockpile environments.

Abstract: Construction aggregates, including sand and gravel, crushed stone and riprap, are the core building blocks of the construction industry. State-of-the-practice characterization methods mainly relies on visual inspection and manual measurement. State-of-the-art aggregate imaging methods have limitations that are only applicable to regular-sized aggregates under well-controlled conditions. This dissertation addresses these major challenges by developing a field imaging framework for the morphological characterization of aggregates as a multi-scenario solution. For individual and non-overlapping aggregates, a field imaging system was designed and the associated segmentation and volume estimation algorithms were developed. For 2D image analyses of aggregates in stockpiles, an automated 2D instance segmentation and morphological analysis approach was established. For 3D point cloud analyses of aggregate stockpiles, an integrated 3D Reconstruction-Segmentation-Completion (RSC-3D) approach was established: 3D reconstruction procedures from multi-view images, 3D stockpile instance segmentation, and 3D shape completion to predict the unseen sides. First, a 3D reconstruction procedure was developed to obtain high-fidelity 3D models of collected aggregate samples, based on which a 3D aggregate particle library was constructed. Next, two datasets were derived from the 3D particle library for 3D learning: a synthetic dataset of aggregate stockpiles with ground-truth instance labels, and a dataset of partial-complete shape pairs, developed with varying-view raycasting schemes. A state-of-the-art 3D instance segmentation network and a 3D shape completion network were trained on the datasets, respectively. The application of the integrated approach was demonstrated on real stockpiles and validated with ground-truth, showing good performance in capturing and predicting the unseen sides of aggregates.

Method: PERSIST simulates the evolution of a latent 3D scene comprising environment, camera, and renderer, enabling synthesis of new frames with persistent spatial memory and consistent geometry through explicit 3D scene representation.

Result: Both quantitative metrics and qualitative user studies show substantial improvements in spatial memory, 3D consistency, and long-horizon stability over existing methods, enabling coherent, evolving 3D worlds with novel capabilities like 3D environment synthesis from single images.

Conclusion: PERSIST represents a new paradigm for world models that explicitly incorporates 3D scene simulation, overcoming limitations of existing approaches and enabling more realistic interactive generation with persistent spatial memory and geometry-aware control.

Abstract: Interactive world models continually generate video by responding to a user’s actions, enabling open-ended generation capabilities. However, existing models typically lack a 3D representation of the environment, meaning 3D consistency must be implicitly learned from data, and spatial memory is restricted to limited temporal context windows. This results in an unrealistic user experience and presents significant obstacles to down-stream tasks such as training agents. To address this, we present PERSIST, a new paradigm of world model which simulates the evolution of a latent 3D scene: environment, camera, and renderer. This allows us to synthesize new frames with persistent spatial memory and consistent geometry. Both quantitative metrics and a qualitative user study show substantial improvements in spatial memory, 3D consistency, and long-horizon stability over existing methods, enabling coherent, evolving 3D worlds. We further demonstrate novel capabilities, including synthesising diverse 3D environments from a single image, as well as enabling fine-grained, geometry-aware control over generated experiences by supporting environment editing and specification directly in 3D space. Project page: https://francelico.github.io/persist.github.io

Method: Three-stage training: 1) Bootstrap geometry and motion representations through pseudo-supervised pretraining, 2) Physics-grounded supervised fine-tuning using simulation data, 3) Simulation-grounded reinforcement learning to correct residual physical violations. Also introduces 4D world consistency evaluation metrics.

Result: Phys4D substantially improves fine-grained spatiotemporal and physical consistency compared to appearance-driven baselines while maintaining strong generative performance.

Conclusion: The proposed pipeline successfully lifts video diffusion models into physics-consistent 4D world representations, addressing physical plausibility issues in current generative video models.

Abstract: Recent video diffusion models have achieved impressive capabilities as large-scale generative world models. However, these models often struggle with fine-grained physical consistency, exhibiting physically implausible dynamics over time. In this work, we present \textbf{Phys4D}, a pipeline for learning physics-consistent 4D world representations from video diffusion models. Phys4D adopts \textbf{a three-stage training paradigm} that progressively lifts appearance-driven video diffusion models into physics-consistent 4D world representations. We first bootstrap robust geometry and motion representations through large-scale pseudo-supervised pretraining, establishing a foundation for 4D scene modeling. We then perform physics-grounded supervised fine-tuning using simulation-generated data, enforcing temporally consistent 4D dynamics. Finally, we apply simulation-grounded reinforcement learning to correct residual physical violations that are difficult to capture through explicit supervision. To evaluate fine-grained physical consistency beyond appearance-based metrics, we introduce a set of \textbf{4D world consistency evaluation} that probe geometric coherence, motion stability, and long-horizon physical plausibility. Experimental results demonstrate that Phys4D substantially improves fine-grained spatiotemporal and physical consistency compared to appearance-driven baselines, while maintaining strong generative performance. Our project page is available at https://sensational-brioche-7657e7.netlify.app/

Method: Introduces a tri-plane generator with expression conditioning that transfers 3DMM expression parameters to source images, generating 3D priors. Uses contrastive pre-training for appearance-free expression parameters and differentiable volume rendering for multi-view image generation.

Result: The pre-training framework successfully learns appearance-free expression representations from 3DMM, enabling cross-identity expression transfer without appearance swapping. The model generates 3D-aware expression-controllable portrait images.

Conclusion: Export3D provides an effective solution for 3D-aware portrait animation that separates appearance from expression, enabling cross-identity expression transfer without appearance contamination.

Abstract: In this paper, we present Export3D, a one-shot 3D-aware portrait animation method that is able to control the facial expression and camera view of a given portrait image. To achieve this, we introduce a tri-plane generator with an effective expression conditioning method, which directly generates a tri-plane of 3D prior by transferring the expression parameter of 3DMM into the source image. The tri-plane is then decoded into the image of different view through a differentiable volume rendering. Existing portrait animation methods heavily rely on image warping to transfer the expression in the motion space, challenging on disentanglement of appearance and expression. In contrast, we propose a contrastive pre-training framework for appearance-free expression parameter, eliminating undesirable appearance swap when transferring a cross-identity expression. Extensive experiments show that our pre-training framework can learn the appearance-free expression representation hidden in 3DMM, and our model can generate 3D-aware expression controllable portrait images without appearance swap in the cross-identity manner.

Method: 1) High-resolution Transformer with global/local modules for subtle feature extraction; 2) Geographically-weighted weakly supervised loss for region-level supervision; 3) Bayesian extension for uncertainty quantification; 4) Decision-level fusion of Sentinel-1, RCM, and AMSR2 data.

Result: Achieves 0.70 overall feature detection accuracy with Sentinel-1 data and preserves pan-Arctic SIC patterns (R² = 0.90 relative to ARTIST Sea Ice product) under minimum-extent conditions in 2021 and 2025.

Conclusion: The proposed Bayesian High-Resolution Transformer approach effectively addresses key challenges in sea ice concentration mapping, providing high-resolution mapping with reliable uncertainty quantification through multi-sensor data fusion.

Abstract: Although high-resolution mapping of pan-Arctic sea ice with reliable corresponding uncertainty is essential for operational sea ice concentration (SIC) charting, it is a difficult task due to key challenges, such as the subtle nature of ice signature features, inexact SIC labels, model uncertainty, and data heterogeneity. This study presents a novel Bayesian High-Resolution Transformer approach for 200 meter resolution pan-Arctic SIC mapping and uncertainty quantification using Sentinel-1, RADARSAT Constellation Mission (RCM), and Advanced Microwave Scanning Radiometer 2 (AMSR2) data. First, to improve small and subtle sea ice feature (e.g., cracks/leads, ponds, and ice floes) extraction, we design a novel high-resolution Transformer model with both global and local modules that can better discern the subtle differences in sea ice patterns. Second, to address low-resolution and inexact SIC labels, we design a geographically-weighted weakly supervised loss function to supervise the model at region level instead of pixel level, and to prioritize pure open water and ice pack signatures while mitigating the impact of ambiguity in the marginal ice zone (MIZ). Third, to improve uncertainty quantification, we design a Bayesian extension of the proposed Transformer model, treating its parameters as random variables to more effectively capture uncertainties. Fourth, to address data heterogeneity, we fuse three different data types (Sentinel-1, RCM, and AMSR2) at decision-level to improve both SIC mapping and uncertainty quantification. The proposed approach is evaluated under pan-Arctic minimum-extent conditions in 2021 and 2025. Results demonstrate that the proposed model achieves 0.70 overall feature detection accuracy using Sentinel-1 data, while also preserving pan-Arctic SIC patterns (Sentinel-1 R\textsuperscript{2} = 0.90 relative to the ARTIST Sea Ice product).

Method: Two-stage reinforcement learning framework: 1) Fine-tune a large language model on physics-focused Chain-of-Thought dataset to integrate principles like object motion and force interactions while preserving user intent. 2) Apply Group Relative Policy Optimization with dynamic reward curriculum that initially prioritizes semantic fidelity, then progressively shifts toward physical commonsense.

Result: PhyPrompt-7B achieves 40.8% joint success on VideoPhy2 (8.6pp gain), improving physical commonsense by 11pp (55.8% to 66.8%) while increasing semantic adherence by 4.4pp (43.4% to 47.8%). Outperforms GPT-4o (+3.8% joint) and DeepSeek-V3 (+2.2%) using only 7B parameters. Transfers zero-shot across diverse T2V architectures with up to 16.8% improvement.

Conclusion: Domain-specialized reinforcement learning with compositional curricula surpasses general-purpose scaling for physics-aware generation, enabling automatic prompt refinement for physically realistic video generation without requiring manual physics expertise.

Abstract: State-of-the-art text-to-video (T2V) generators frequently violate physical laws despite high visual quality. We show this stems from insufficient physical constraints in prompts rather than model limitations: manually adding physics details reliably produces physically plausible videos, but requires expertise and does not scale. We present PhyPrompt, a two-stage reinforcement learning framework that automatically refines prompts for physically realistic generation. First, we fine-tune a large language model on a physics-focused Chain-of-Thought dataset to integrate principles like object motion and force interactions while preserving user intent. Second, we apply Group Relative Policy Optimization with a dynamic reward curriculum that initially prioritizes semantic fidelity, then progressively shifts toward physical commonsense. This curriculum achieves synergistic optimization: PhyPrompt-7B reaches 40.8% joint success on VideoPhy2 (8.6pp gain), improving physical commonsense by 11pp (55.8% to 66.8%) while simultaneously increasing semantic adherence by 4.4pp (43.4% to 47.8%). Remarkably, our curriculum exceeds single-objective training on both metrics, demonstrating compositional prompt discovery beyond conventional multi-objective trade-offs. PhyPrompt outperforms GPT-4o (+3.8% joint) and DeepSeek-V3 (+2.2%, 100$\times$ larger) using only 7B parameters. The approach transfers zero-shot across diverse T2V architectures (Lavie, VideoCrafter2, CogVideoX-5B) with up to 16.8% improvement, establishing that domain-specialized reinforcement learning with compositional curricula surpasses general-purpose scaling for physics-aware generation.

Method: Proposes PinCLIP with hybrid Vision Transformer architecture using VLM backbone and hybrid fusion mechanism to capture multi-modality content at varying granularities. Introduces neighbor alignment objective to model cross-fusion of multi-modal representations within Pinterest’s Pin-Board graph, beyond standard image-text alignment.

Result: Offline evaluations show PinCLIP outperforms state-of-the-art baselines (like Qwen) by 20% in multi-modal retrieval tasks. Online A/B testing demonstrates significant business impact with engagement gains across Pinterest surfaces. Notably addresses “cold-start” problem with 15% Repin increase in organic content and 8.7% higher click for new Ads.

Conclusion: PinCLIP successfully integrates VLMs into large-scale recommendation systems, demonstrating practical value through improved retrieval performance and business metrics while solving cold-start challenges.

Abstract: While multi-modal Visual Language Models (VLMs) have demonstrated significant success across various domains, the integration of VLMs into recommendation and retrieval systems remains a challenge, due to issues like training objective discrepancies and serving efficiency bottlenecks. This paper introduces PinCLIP, a large-scale visual representation learning approach developed to enhance retrieval and ranking models at Pinterest by leveraging VLMs to learn image-text alignment. We propose a novel hybrid Vision Transformer architecture that utilizes a VLM backbone and a hybrid fusion mechanism to capture multi-modality content representation at varying granularities. Beyond standard image-to-text alignment objectives, we introduce a neighbor alignment objective to model the cross-fusion of multi-modal representations within the Pinterest Pin-Board graph. Offline evaluations show that PinCLIP outperforms state-of-the-art baselines, such as Qwen, by 20% in multi-modal retrieval tasks. Online A/B testing demonstrates significant business impact, including substantial engagement gains across all major surfaces in Pinterest. Notably, PinCLIP significantly addresses the “cold-start” problem, enhancing fresh content distribution with a 15% Repin increase in organic content and 8.7% higher click for new Ads.

Method: Unifies Topological Data Analysis (TDA) and Lipschitz-Killing Curvatures (LKCs) to fuse persistent homology for robust topological signatures with LKCs for precise quantification of local geometric features like boundary complexity and surface regularity.

Result: Achieves 3.6% accuracy improvement in skin lesion classification, reduces false positives/negatives by 15-18% compared to conventional methods, processes 224x224 pixel images in ≤0.5s, and demonstrates generalizability to molecular-level data.

Conclusion: GeoTop provides a principled, interpretable solution for advanced shape discrimination in diagnostic imaging by unifying topological invariance with geometric sensitivity, offering both theoretical guarantees and empirical validation.

Abstract: A fundamental challenge in diagnostic imaging is the phenomenon of topological equivalence, where benign and malignant structures share global topology but differ in critical geometric detail, leading to diagnostic errors in both conventional and deep learning models. We introduce GeoTop, a mathematically principled framework that unifies Topological Data Analysis (TDA) and Lipschitz-Killing Curvatures (LKCs) to resolve this ambiguity. Unlike hybrid deep learning approaches, GeoTop provides intrinsic interpretability by fusing the capacity of persistent homology to identify robust topological signatures with the precision of LKCs in quantifying local geometric features such as boundary complexity and surface regularity. The framework’s clinical utility is demonstrated through its application to skin lesion classification, where it achieves a consistent accuracy improvement of 3.6% and reduces false positives and negatives by 15-18% compared to conventional single-modality methods. Crucially, GeoTop directly addresses the problem of topological equivalence by incorporating geometric differentiators, providing both theoretical guarantees (via a formal lemma) and empirical validation via controlled benchmarks. Beyond its predictive performance, GeoTop offers inherent mathematical interpretability through persistence diagrams and curvature-based descriptors, computational efficiency for large datasets (processing 224x224 pixel images in less or equal 0.5 s), and demonstrated generalisability to molecular-level data. By unifying topological invariance with geometric sensitivity, GeoTop provides a principled, interpretable solution for advanced shape discrimination in diagnostic imaging.

Method: Uses sparse Mixture-of-Experts LVM with dynamic modality router for specialization and cross-modal interaction. Employs synergy-aware training combining modality-specific pretraining with coarse-to-fine synergy tuning via knowledge distillation and object-/relation-level alignment.

Result: Outperforms existing models on 10 benchmarks spanning image, video, and 3D understanding, achieving over 10% average improvement over its backbone, particularly excelling on synergy-focused datasets requiring spatial or temporal priors.

Conclusion: PolyV establishes a unified framework for synesthetic visual reasoning, advancing toward truly synergistic large vision models that can leverage complementary priors across different visual modalities.

Abstract: Recent advances in large vision models (LVMs) have shifted from modality-specific designs toward unified architectures that jointly process images, videos, and 3D data. However, existing unified LVMs primarily pursue functional integration, while overlooking the deeper goal of cross-vision synergy: the ability to reason over complementary priors across visual modalities. To address this, we present PolyV, a unified LVM that achieves cross-vision synergy at both the architectural and training levels. Architecturally, PolyV adopts a sparse Mixture-of-Experts LVM coordinated by a dynamic modality router, allowing each expert to specialize in modality-specific priors while enabling bidirectional interaction and mutual refinement across modalities. Training-wise, a synergy-aware paradigm combines modality-specific pretraining with coarse-to-fine synergy tuning via knowledge distillation and object-/relation-level alignment. Extensive experiments on 10 benchmarks spanning image, video, and 3D understanding, including synergy-focused datasets requiring spatial or temporal priors, demonstrate that PolyV consistently outperforms existing models, achieving over 10% average improvement over its backbone. Overall, PolyV establishes a unified framework for synesthetic visual reasoning, advancing toward truly synergistic LVMs. Project page: https://sqwu.top/PolyV.

Method: Proposes a method for precise estimation of SDNP, models its dependence on scene brightness, ISO settings, and other factors. Uses this characterization to explore forensic applications including traceability across iPhone models and iOS versions, and masking SDNP-affected regions in PRNU analysis.

Result: Demonstrates that masking SDNP-affected regions significantly reduces false positives in PRNU-based camera source verification, improving state-of-the-art techniques. Shows traceability of portrait-mode images across iPhone models and iOS versions in open-set scenarios with robustness under post-processing.

Conclusion: Apple’s SDNP is a significant forensic artifact that requires careful consideration in image analysis. Proper characterization and masking of SDNP-affected regions can improve camera source verification accuracy and enable new forensic applications for iPhone portrait-mode images.

Abstract: iPhone portrait-mode images contain a distinctive pattern in out-of-focus regions simulating the bokeh effect, which we term Apple’s Synthetic Defocus Noise Pattern (SDNP). If overlooked, this pattern can interfere with blind forensic analyses, especially PRNU-based camera source verification, as noted in earlier works. Since Apple’s SDNP remains underexplored, we provide a detailed characterization, proposing a method for its precise estimation, modeling its dependence on scene brightness, ISO settings, and other factors. Leveraging this characterization, we explore forensic applications of the SDNP, including traceability of portrait-mode images across iPhone models and iOS versions in open-set scenarios, assessing its robustness under post-processing. Furthermore, we show that masking SDNP-affected regions in PRNU-based camera source verification significantly reduces false positives, overcoming a critical limitation in camera attribution, and improving state-of-the-art techniques.

Method: Three key contributions: 1) Calibrated confidence targets using ensemble of fine-tuned stereo matching models to capture disparity variance; 2) Confidence-aware loss functions that prioritize reliable pixels during training; 3) Inference-time confidence head with convolution layers to predict per-pixel confidence.

Result: Framework improves depth estimation accuracy by ~8% on internal clinical endoscopic dataset (StereoKP) compared to baseline, and robustly quantifies prediction confidence across internal and public datasets.

Conclusion: The confidence-aware framework enhances MDE accuracy in minimally invasive surgery, addresses noise/artifact challenges, and provides confidence maps to improve clinical reliability.

Abstract: Purpose: Monocular depth estimation (MDE) is vital for scene understanding in minimally invasive surgery (MIS). However, endoscopic video sequences are often contaminated by smoke, specular reflections, blur, and occlusions, limiting the accuracy of MDE models. In addition, current MDE models do not output depth confidence, which could be a valuable tool for improving their clinical reliability. Methods: We propose a novel confidence-aware MDE framework featuring three significant contributions: (i) Calibrated confidence targets: an ensemble of fine-tuned stereo matching models is used to capture disparity variance into pixel-wise confidence probabilities; (ii) Confidence-aware loss: Baseline MDE models are optimized with confidence-aware loss functions, utilizing pixel-wise confidence probabilities such that reliable pixels dominate training; and (iii) Inference-time confidence: a confidence estimation head is proposed with two convolution layers to predict per-pixel confidence at inference, enabling assessment of depth reliability. Results: Comprehensive experimental validation across internal and public datasets demonstrates that our framework improves depth estimation accuracy and can robustly quantify the prediction’s confidence. On the internal clinical endoscopic dataset (StereoKP), we improve dense depth estimation accuracy by ~8% as compared to the baseline model. Conclusion: Our confidence-aware framework enables improved accuracy of MDE models in MIS, addressing challenges posed by noise and artifacts in pre-clinical and clinical data, and allows MDE models to provide confidence maps that may be used to improve their reliability for clinical applications.

Method: Introduces two noise types: textual prompt noise (category flips, noun substitutions) and mask annotation noise (perturbed boundaries). Creates ActiSeg-NL benchmark, adapts six label-noise learning strategies, establishes evaluation protocols, and proposes Parallel Mask Head Mechanism (PMHM) for mask noise.

Result: Comprehensive analysis linking noise types to failure modes (boundary leakage, mislocalization, identity substitutions). Different strategies show distinct robustness profiles with foreground-background trade-offs. PMHM addresses mask annotation noise effectively.

Conclusion: First systematic study of label noise in action-based video object segmentation provides benchmark, evaluation protocols, and insights into noise robustness. PMHM offers solution for boundary noise. Work enables more robust embodied intelligence systems.

Abstract: Embodied intelligence relies on accurately segmenting objects actively involved in interactions. Action-based video object segmentation addresses this by linking segmentation with action semantics, but it depends on large-scale annotations and prompts that are costly, inconsistent, and prone to multimodal noise such as imprecise masks and referential ambiguity. To date, this challenge remains unexplored. In this work, we take the first step by studying action-based video object segmentation under label noise, focusing on two sources: textual prompt noise (category flips and within-category noun substitutions) and mask annotation noise (perturbed object boundaries to mimic imprecise supervision). Our contributions are threefold. First, we introduce two types of label noises for the action-based video object segmentation task. Second, we build up the first action-based video object segmentation under a label noise benchmark ActiSeg-NL and adapt six label-noise learning strategies to this setting, and establish protocols for evaluating them under textual, boundary, and mixed noise. Third, we provide a comprehensive analysis linking noise types to failure modes and robustness gains, and we introduce a Parallel Mask Head Mechanism (PMHM) to address mask annotation noise. Qualitative evaluations further reveal characteristic failure modes, including boundary leakage and mislocalization under boundary perturbations, as well as occasional identity substitutions under textual flips. Our comparative analysis reveals that different learning strategies exhibit distinct robustness profiles, governed by a foreground-background trade-off where some achieve balanced performance while others prioritize foreground accuracy at the cost of background precision. The established benchmark and source code will be made publicly available at https://github.com/mylwx/ActiSeg-NL.

Method: 1) Dense patch-level matching between templates and query image; 2) Candidate point generation from local matches; 3) Candidate selection to filter false positives; 4) Prompting augmented SAM with instance-specific tokens for mask reconstruction.

Result: Improved performance over proposal-based methods in challenging open-world settings with occlusion and background clutter.

Conclusion: L2G-Det provides a more robust framework for open-world instance detection by bypassing proposal limitations through local-to-global matching and SAM integration.

Abstract: Detecting and segmenting novel object instances in open-world environments is a fundamental problem in robotic perception. Given only a small set of template images, a robot must locate and segment a specific object instance in a cluttered, previously unseen scene. Existing proposal-based approaches are highly sensitive to proposal quality and often fail under occlusion and background clutter. We propose L2G-Det, a local-to-global instance detection framework that bypasses explicit object proposals by leveraging dense patch-level matching between templates and the query image. Locally matched patches generate candidate points, which are refined through a candidate selection module to suppress false positives. The filtered points are then used to prompt an augmented Segment Anything Model (SAM) with instance-specific object tokens, enabling reliable reconstruction of complete instance masks. Experiments demonstrate improved performance over proposal-based methods in challenging open-world settings.

Method: Proposes a data augmentation framework that generates natural-language questions about character-level attributes (presence, position, frequency) for each image-text pair. The OCR model learns to align visual features with textual queries and jointly reason over images and questions through auxiliary QA tasks.

Result: Experiments on WordArt and Esposalles datasets show consistent improvements over baseline models with significant reductions in both Character Error Rate (CER) and Word Error Rate (WER).

Conclusion: The VQA-inspired data augmentation approach effectively strengthens OCR training by encouraging finer-grained reasoning about text structure, leading to improved recognition accuracy across different datasets.

Abstract: Scene text recognition (STR) and handwritten text recognition (HTR) face significant challenges in accurately transcribing textual content from images into machine-readable formats. Conventional OCR models often predict transcriptions directly, which limits detailed reasoning about text structure. We propose a VQA-inspired data augmentation framework that strengthens OCR training through structured question-answering tasks. For each image-text pair, we generate natural-language questions probing character-level attributes such as presence, position, and frequency, with answers derived from ground-truth text. These auxiliary tasks encourage finer-grained reasoning, and the OCR model aligns visual features with textual queries to jointly reason over images and questions. Experiments on WordArt and Esposalles datasets show consistent improvements over baseline models, with significant reductions in both CER and WER. Our code is publicly available at https://github.com/xuyaooo/DataAugOCR.

Method: Proposes Collaborative Policy Planning (CPP) paradigm with multiple policy agents: 1) Policy Generation - each agent creates unique tool invocation policy, 2) Policy Execution - agents invoke tools sequentially, 3) Policy Communication - agents interact and update policies during execution. Enhanced with Multi-Agent Reinforcement Learning for joint optimization.

Result: Achieves SOTA performance across eight benchmarks spanning four tasks. On LongVideoBench, outperforms Gemini 2.5 pro by 3.6% and GPT-4o by 15.6%.

Conclusion: VideoChat-M1’s collaborative multi-agent framework with dynamic policy planning significantly advances video understanding capabilities compared to static approaches.

Abstract: By leveraging tool-augmented Multimodal Large Language Models (MLLMs), multi-agent frameworks are driving progress in video understanding. However, most of them adopt static and non-learnable tool invocation mechanisms, which limit the discovery of diverse clues essential for robust perception and reasoning regarding temporally or spatially complex videos. To address this challenge, we propose a novel Multi-agent system for video understanding, namely VideoChat-M1. Instead of using a single or fixed policy, VideoChat-M1 adopts a distinct Collaborative Policy Planning (CPP) paradigm with multiple policy agents, which comprises three key processes. (1) Policy Generation: Each agent generates its unique tool invocation policy tailored to the user’s query; (2) Policy Execution: Each agent sequentially invokes relevant tools to execute its policy and explore the video content; (3) Policy Communication: During the intermediate stages of policy execution, agents interact with one another to update their respective policies. Through this collaborative framework, all agents work in tandem, dynamically refining their preferred policies based on contextual insights from peers to effectively respond to the user’s query. Moreover, we equip our CPP paradigm with a concise Multi-Agent Reinforcement Learning (MARL) method. Consequently, the team of policy agents can be jointly optimized to enhance VideoChat-M1’s performance, guided by both the final answer reward and intermediate collaborative process feedback. Extensive experiments demonstrate that VideoChat-M1 achieves SOTA performance across eight benchmarks spanning four tasks. Notably, on LongVideoBench, our method outperforms the SOTA model Gemini 2.5 pro by 3.6% and GPT-4o by 15.6%.

Method: Proposes Traffic Scene Graph Generation framework that explicitly uses traffic accident data and depth cues to supplement visual features and semantic information. Creates relational annotations on Cityscapes dataset and evaluates on 10 tasks from 5 perspectives.

Result: The framework generates intuitive traffic scene graphs that color-code hazard severity and notate effect mechanisms and relative locations to ego vehicle. Results demonstrate superior ego-centric reasoning for hazard-aware traffic scene understanding compared to existing methods.

Conclusion: The proposed Traffic Scene Graph Generation effectively bridges the gap in safety-relevance assessment for driving scenarios, providing a novel approach for hazard-aware situational awareness in complex traffic environments.

Abstract: Maintaining situational awareness in complex driving scenarios is challenging. It requires continuously prioritizing attention among extensive scene entities and understanding how prominent hazards might affect the ego vehicle. While existing studies excel at detecting specific semantic categories and visually salient regions, they lack the ability to assess safety-relevance. Meanwhile, the generic spatial predicates either for foreground objects only or for all scene entities modeled by existing scene graphs are inadequate for driving scenarios. To bridge this gap, we introduce a novel task, Traffic Scene Graph Generation, which captures traffic-specific relations between prominent hazards and the ego vehicle. We propose a novel framework that explicitly uses traffic accident data and depth cues to supplement visual features and semantic information for reasoning. The output traffic scene graphs provide intuitive guidelines that stress prominent hazards by color-coding their severity and notating their effect mechanism and relative location to the ego vehicle. We create relational annotations on Cityscapes dataset and evaluate our model on 10 tasks from 5 perspectives. The results in comparative experiments and ablation studies demonstrate our capacity in ego-centric reasoning for hazard-aware traffic scene understanding.

Method: Proposes DM-CFO with two main components: 1) progressive layout restoration during denoising under text and graph constraints, and 2) alternate updating of Gaussian parameters for each tooth and entire jaw using SDS with a regularization term based on distances between 3D Gaussians of neighboring teeth to penalize intersections.

Result: Experimental results on three tooth-design datasets show significant improvements in multiview consistency and realism of generated teeth compared to existing methods.

Conclusion: The approach effectively addresses compositional 3D tooth generation challenges by combining diffusion-based layout restoration with collision-aware Gaussian optimization.

Abstract: The automatic design of a 3D tooth model plays a crucial role in dental digitization. However, current approaches face challenges in compositional 3D tooth generation because both the layouts and shapes of missing teeth need to be optimized.In addition, collision conflicts are often omitted in 3D Gaussian-based compositional 3D generation, where objects may intersect with each other due to the absence of explicit geometric information on the object surfaces. Motivated by graph generation through diffusion models and collision detection using 3D Gaussians, we propose an approach named DM-CFO for compositional tooth generation, where the layout of missing teeth is progressively restored during the denoising phase under both text and graph constraints. Then, the Gaussian parameters of each layout-guided tooth and the entire jaw are alternately updated using score distillation sampling (SDS). Furthermore, a regularization term based on the distances between the 3D Gaussians of neighboring teeth and the anchor tooth is introduced to penalize tooth intersections. Experimental results on three tooth-design datasets demonstrate that our approach significantly improves the multiview consistency and realism of the generated teeth compared with existing methods. Project page: https://amateurc.github.io/CF-3DTeeth/.

Method: 1) Adapted YOLO11 to process consecutive frames for detection to leverage temporal consistency and motion cues; 2) Introduced tracklet-based contrastive learning for identification after tracking to create coherent feature embeddings.

Result: Fine-tuning with two-frame inputs improved mAP@0.5 from 0.922 to 0.933, recovering individuals indistinguishable in static images. Qualitative visualization showed coherent feature embeddings that bring same-individual samples closer in feature space, suggesting potential for mitigating ID switching.

Conclusion: The framework successfully enhances penguin detection and identification by integrating appearance and motion features, addressing challenges of homogeneous visual characteristics and environmental noise through temporal processing and contrastive learning.

Abstract: In animal facilities, continuous surveillance of penguins is essential yet technically challenging due to their homogeneous visual characteristics, rapid and frequent posture changes, and substantial environmental noise such as water reflections. In this study, we propose a framework that enhances both detection and identification performance by integrating appearance and motion features. For detection, we adapted YOLO11 to process consecutive frames to overcome the lack of temporal consistency in single-frame detectors. This approach leverages motion cues to detect targets even when distinct visual features are obscured. Our evaluation shows that fine-tuning the model with two-frame inputs improves mAP@0.5 from 0.922 to 0.933, outperforming the baseline, and successfully recovers individuals that are indistinguishable in static images. For identification, we introduce a tracklet-based contrastive learning approach applied after tracking. Through qualitative visualization, we demonstrate that the method produces coherent feature embeddings, bringing samples from the same individual closer in the feature space, suggesting the potential for mitigating ID switching.

Method: Crops OBB-centered patches and applies three detectors (head, tail, head-tail) with IoU-based majority voting to determine final orientation labels.

Result: Achieves 99.3% accuracy on 299 test images, outperforming individual models and demonstrating effectiveness for robust OBB-based tracking.

Conclusion: The proposed method effectively solves the head-tail ambiguity problem in OBB-based animal tracking, enabling more accurate continuous tracking in aerial footage.

Abstract: The social structures of group-living animals such as feral horses are diverse and remain insufficiently understood, even within a single species. To investigate group dynamics, aerial videos are often utilized to track individuals and analyze their movement trajectories, which are essential for evaluating inter-individual interactions and comparing social behaviors. Accurate individual tracking is therefore crucial. In multi-animal tracking, axis-aligned bounding boxes (bboxes) are widely used; however, for aerial top-view footage of entire groups, their performance degrades due to complex backgrounds, small target sizes, high animal density, and varying body orientations. To address this issue, we employ oriented bounding boxes (OBBs), which include rotation angles and reduce unnecessary background. Nevertheless, current OBB detectors such as YOLO-OBB restrict angles within a 180$^{\circ}$ range, making it impossible to distinguish head from tail and often causing sudden 180$^{\circ}$ flips across frames, which severely disrupts continuous tracking. To overcome this limitation, we propose a head-orientation estimation method that crops OBB-centered patches, applies three detectors (head, tail, and head-tail), and determines the final label through IoU-based majority voting. Experiments using 299 test images show that our method achieves 99.3% accuracy, outperforming individual models, demonstrating its effectiveness for robust OBB-based tracking.

Method: Proposes OmniParallax Attention Mechanism (OPAM) for explicitly modeling correlations between arbitrary pairs of information sources, and Parallax Multi Information Fusion Module (PMIFM) to adaptively integrate information from different sources, incorporated into both joint decoder and entropy model.

Result: ParaHydra is the first DMIC method to significantly surpass state-of-the-art MIC codecs, achieving 19.72% bitrate savings on WildTrack(3) and up to 24.18% on WildTrack(6), with 65× decoding and 34× encoding efficiency improvements.

Conclusion: The proposed OPAM and PMIFM enable ParaHydra to achieve superior compression performance by adaptively modeling view correlations, with gains increasing as the number of input views grows.

Abstract: Multi-view image compression (MIC) aims to achieve high compression efficiency by exploiting inter-image correlations, playing a crucial role in 3D applications. As a subfield of MIC, distributed multi-view image compression (DMIC) offers performance comparable to MIC while eliminating the need for inter-view information at the encoder side. However, existing methods in DMIC typically treat all images equally, overlooking the varying degrees of correlation between different views during decoding, which leads to suboptimal coding performance. To address this limitation, we propose a novel $\textbf{OmniParallax Attention Mechanism}$ (OPAM), which is a general mechanism for explicitly modeling correlations and aligned features between arbitrary pairs of information sources. Building upon OPAM, we propose a Parallax Multi Information Fusion Module (PMIFM) to adaptively integrate information from different sources. PMIFM is incorporated into both the joint decoder and the entropy model to construct our end-to-end DMIC framework, $\textbf{ParaHydra}$. Extensive experiments demonstrate that $\textbf{ParaHydra}$ is $\textbf{the first DMIC method}$ to significantly surpass state-of-the-art MIC codecs, while maintaining low computational overhead. Performance gains become more pronounced as the number of input views increases. Compared with LDMIC, $\textbf{ParaHydra}$ achieves bitrate savings of $\textbf{19.72%}$ on WildTrack(3) and up to $\textbf{24.18%}$ on WildTrack(6), while significantly improving coding efficiency (as much as $\textbf{65}\times$ in decoding and $\textbf{34}\times$ in encoding).

Method: Proposes LeafInst framework with Asymptotic Feature Pyramid Network (AFPN) for multi-scale perception, Dynamic Asymmetric Spatial Perception (DASP) module for irregular shape modeling, and dual-residual Dynamic Anomalous Regression Head (DARH) with Top-down Concatenation decoder Feature Fusion (TCFU).

Result: Achieves 68.4 mAP on Poplar-leaf dataset (outperforming YOLOv11 by 7.1% and MaskDINO by 6.5%), and 52.7 box mAP on PhenoBench benchmark (exceeding MaskDINO by 3.4%). Demonstrates strong generalization for large-scale leaf phenotyping.

Conclusion: LeafInst effectively addresses challenges in fine-grained leaf segmentation for forestry applications in natural environments, providing a practical solution for large-scale leaf phenotyping with superior performance over existing methods.

Abstract: Intelligent forest tree breeding has advanced plant phenotyping, yet existing research largely focuses on large-leaf agricultural crops, with limited attention to fine-grained leaf analysis of sapling trees in open-field environments. Natural scenes introduce challenges including scale variation, illumination changes, and irregular leaf morphology. To address these issues, we collected UAV RGB imagery of field-grown saplings and constructed the Poplar-leaf dataset, containing 1,202 branches and 19,876 pixel-level annotated leaf instances. To our knowledge, this is the first instance segmentation dataset specifically designed for forestry leaves in open-field conditions. We propose LeafInst, a novel segmentation framework tailored for irregular and multi-scale leaf structures. The model integrates an Asymptotic Feature Pyramid Network (AFPN) for multi-scale perception, a Dynamic Asymmetric Spatial Perception (DASP) module for irregular shape modeling, and a dual-residual Dynamic Anomalous Regression Head (DARH) with Top-down Concatenation decoder Feature Fusion (TCFU) to improve detection and segmentation performance. On Poplar-leaf, LeafInst achieves 68.4 mAP, outperforming YOLOv11 by 7.1 percent and MaskDINO by 6.5 percent. On the public PhenoBench benchmark, it reaches 52.7 box mAP, exceeding MaskDINO by 3.4 percent. Additional experiments demonstrate strong generalization and practical utility for large-scale leaf phenotyping.

Method: 1) Use MLLMs to automatically generate textual annotations for RGBT benchmarks; 2) Propose RAGTrack with Multi-modal Transformer Encoder for unified visual-language modeling; 3) Adaptive Token Fusion to select target-relevant tokens and perform channel exchanges; 4) Context-aware Reasoning Module with Retrieval-Augmented Generation for temporal linguistic reasoning.

Result: Achieves state-of-the-art performance on four RGBT benchmarks across various challenging scenarios, demonstrating robustness and effectiveness.

Conclusion: Introducing language guidance via MLLMs and retrieval-augmented generation significantly improves RGBT tracking robustness by addressing modality gaps, search redundancies, and enabling adaptive target modeling.

Abstract: RGB-Thermal (RGBT) tracking aims to achieve robust object localization across diverse environmental conditions by fusing visible and thermal infrared modalities. However, existing RGBT trackers rely solely on initial-frame visual information for target modeling, failing to adapt to appearance variations due to the absence of language guidance. Furthermore, current methods suffer from redundant search regions and heterogeneous modality gaps, causing background distraction. To address these issues, we first introduce textual descriptions into RGBT tracking benchmarks. This is accomplished through a pipeline that leverages Multi-modal Large Language Models (MLLMs) to automatically produce texual annotations. Afterwards, we propose RAGTrack, a novel Retrieval-Augmented Generation framework for robust RGBT tracking. To this end, we introduce a Multi-modal Transformer Encoder (MTE) for unified visual-language modeling. Then, we design an Adaptive Token Fusion (ATF) to select target-relevant tokens and perform channel exchanges based on cross-modal correlations, mitigating search redundancies and modality gaps. Finally, we propose a Context-aware Reasoning Module (CRM) to maintain a dynamic knowledge base and employ a Retrieval-Augmented Generation (RAG) to enable temporal linguistic reasoning for robust target modeling. Extensive experiments on four RGBT benchmarks demonstrate that our framework achieves state-of-the-art performance across various challenging scenarios. The source code is available https://github.com/IdolLab/RAGTrack.

Method: Created CoRe-BT dataset with 310 patients: multi-sequence brain MRI (T1, T1c, T2, FLAIR), 95 cases with paired H&E-stained whole-slide pathology images and pathology reports. All cases annotated with tumor type/grade, MRI volumes include expert-annotated tumor masks. Evaluated tumor typing under variable modality availability comparing MRI-only vs multimodal approaches.

Result: Baseline experiments demonstrate feasibility of multimodal fusion and highlight complementary modality contributions across clinically relevant typing tasks. The dataset enables both region-aware modeling and auxiliary learning tasks.

Conclusion: CoRe-BT provides a grounded testbed for advancing multimodal glioma typing and representation learning in realistic scenarios with incomplete clinical data.

Abstract: Accurate brain tumor typing requires integrating heterogeneous clinical evidence, including magnetic resonance imaging (MRI), histopathology, and pathology reports, which are often incomplete at the time of diagnosis. We introduce CoRe-BT, a cross-modal radiology-pathology-text benchmark for brain tumor typing, designed to study robust multimodal learning under missing modality conditions. The dataset comprises 310 patients with multi-sequence brain MRI (T1, T1c, T2, FLAIR), including 95 cases with paired H&E-stained whole-slide pathology images and pathology reports. All cases are annotated with tumor type and grade, and MRI volumes include expert-annotated tumor masks, enabling both region-aware modeling and auxiliary learning tasks. Tumors are categorized into six clinically relevant classes capturing the heterogeneity of common and rare glioma subtypes. We evaluate tumor typing under variable modality availability by comparing MRI-only models with multimodal approaches that incorporate pathology information when present. Baseline experiments demonstrate the feasibility of multimodal fusion and highlight complementary modality contributions across clinically relevant typing tasks. CoRe-BT provides a grounded testbed for advancing multimodal glioma typing and representation learning in realistic scenarios with incomplete clinical data.

Method: Developed an end-to-end IPI pipeline with segmentation-based region selection, adaptive font scaling, and background-aware rendering to conceal prompts from human perception while preserving model interpretability. Evaluated 12 adversarial prompt strategies and multiple embedding configurations using COCO dataset and GPT-4-turbo.

Result: IPI can reliably manipulate model outputs, with the most effective configuration achieving up to 64% attack success rate under stealth constraints. The attacks work in black-box settings without requiring model internals.

Conclusion: Image-based Prompt Injection represents a practical threat to multimodal LLMs, demonstrating that adversarial instructions can be effectively embedded in images to override model behavior while remaining stealthy. This underscores the need for defenses against multimodal prompt injection attacks.

Abstract: Multimodal Large Language Models (MLLMs) integrate vision and text to power applications, but this integration introduces new vulnerabilities. We study Image-based Prompt Injection (IPI), a black-box attack in which adversarial instructions are embedded into natural images to override model behavior. Our end-to-end IPI pipeline incorporates segmentation-based region selection, adaptive font scaling, and background-aware rendering to conceal prompts from human perception while preserving model interpretability. Using the COCO dataset and GPT-4-turbo, we evaluate 12 adversarial prompt strategies and multiple embedding configurations. The results show that IPI can reliably manipulate the output of the model, with the most effective configuration achieving up to 64% attack success under stealth constraints. These findings highlight IPI as a practical threat in black-box settings and underscore the need for defenses against multimodal prompt injection.

Method: Introduces a background-consistent generation pipeline that maintains visual coherence across scenes while preserving character identity and spatial relationships. Also proposes a transition-aware video synthesis module for smooth shot transitions in complex multi-subject scenarios. Includes a synthetic dataset of 10,000 multi-subject transition sequences covering underrepresented dynamic scene compositions.

Result: On VBench, InfinityStory achieves the highest Background Consistency (88.94), highest Subject Consistency (82.11), and the best overall average rank (2.80), showing improved stability, smoother transitions, and better temporal coherence.

Conclusion: The framework successfully addresses key challenges in long-form video generation, providing a scalable solution for consistent visual storytelling with improved coherence and transition quality.

Abstract: Generating long-form storytelling videos with consistent visual narratives remains a significant challenge in video synthesis. We present a novel framework, dataset, and a model that address three critical limitations: background consistency across shots, seamless multi-subject shot-to-shot transitions, and scalability to hour-long narratives. Our approach introduces a background-consistent generation pipeline that maintains visual coherence across scenes while preserving character identity and spatial relationships. We further propose a transition-aware video synthesis module that generates smooth shot transitions for complex scenarios involving multiple subjects entering or exiting frames, going beyond the single-subject limitations of prior work. To support this, we contribute with a synthetic dataset of 10,000 multi-subject transition sequences covering underrepresented dynamic scene compositions. On VBench, InfinityStory achieves the highest Background Consistency (88.94), highest Subject Consistency (82.11), and the best overall average rank (2.80), showing improved stability, smoother transitions, and better temporal coherence.

Method: 1) Data-dependent coupling that explicitly models LQ-HQ dependency to minimize path crossovers and promote near-linear transport. 2) Conditional mean estimation for coarse prediction to refine source anchor and condition velocity field. 3) Shortcut constraint that supervises average velocities over arbitrary time intervals for accurate one-step inference.

Result: SCFlowFR achieves state-of-the-art one-step face restoration quality with inference speed comparable to traditional non-diffusion methods.

Conclusion: The proposed method effectively addresses limitations of existing flow matching approaches by modeling data dependencies and enabling efficient one-step inference for face restoration.

Abstract: Face restoration has advanced significantly with generative models like diffusion models and flow matching (FM), which learn continuous-time mappings between distributions. However, existing FM-based approaches often start from Gaussian noise, ignoring the inherent dependency between low-quality (LQ) and high-quality (HQ) data, resulting in path crossovers, curved trajectories, and multi-step sampling requirements. To address these issues, we propose Shortcut-enhanced Coupling flow for Face Restoration (SCFlowFR). First, it establishes a \textit{data-dependent coupling} that explicitly models the LQ–HQ dependency, minimizing path crossovers and promoting near-linear transport. Second, we employ conditional mean estimation to obtain a coarse prediction that refines the source anchor to tighten coupling and conditions the velocity field to stabilize large-step updates. Third, a shortcut constraint supervises average velocities over arbitrary time intervals, enabling accurate one-step inference. Experiments demonstrate that SCFlowFR achieves state-of-the-art one-step face restoration quality with inference speed comparable to traditional non-diffusion methods.

Method: Introduces InEdit-Bench, the first evaluation benchmark for reasoning over intermediate pathways in image editing, with meticulously annotated test cases covering four task categories: state transition, dynamic process, temporal sequence, and scientific simulation. Also proposes assessment criteria for evaluating logical coherence, visual naturalness, and fidelity to path constraints.

Result: Comprehensive evaluation of 14 representative image editing models on InEdit-Bench reveals significant and widespread shortcomings in dynamic reasoning capabilities, showing current models are ill-equipped for complex scenarios requiring coherent intermediate pathways.

Conclusion: InEdit-Bench provides a standardized benchmark to catalyze research toward more dynamic, reason-aware, and intelligent multimodal generative models by systematically measuring and addressing the critical limitation of intermediate pathway reasoning in image editing.

Abstract: Multimodal generative models have made significant strides in image editing, demonstrating impressive performance on a variety of static tasks. However, their proficiency typically does not extend to complex scenarios requiring dynamic reasoning, leaving them ill-equipped to model the coherent, intermediate logical pathways that constitute a multi-step evolution from an initial state to a final one. This capacity is crucial for unlocking a deeper level of procedural and causal understanding in visual manipulation. To systematically measure this critical limitation, we introduce InEdit-Bench, the first evaluation benchmark dedicated to reasoning over intermediate pathways in image editing. InEdit-Bench comprises meticulously annotated test cases covering four fundamental task categories: state transition, dynamic process, temporal sequence, and scientific simulation. Additionally, to enable fine-grained evaluation, we propose a set of assessment criteria to evaluate the logical coherence and visual naturalness of the generated pathways, as well as the model’s fidelity to specified path constraints. Our comprehensive evaluation of 14 representative image editing models on InEdit-Bench reveals significant and widespread shortcomings in this domain. By providing a standardized and challenging benchmark, we aim for InEdit-Bench to catalyze research and steer development towards more dynamic, reason-aware, and intelligent multimodal generative models.

Method: Fine-tunes a score network to learn dual objectives (target identity and human expression) jointly optimized through gradient projection. Enhances perceptual quality with local smoothness regularization and score matching loss optimization.

Result: Surpasses previous baselines (noise-based, makeup-based, freeform attribute methods) in both qualitative fidelity and quantitative metrics. Effective against online FR APIs and shows advanced adaptability in uncommon photographic scenarios.

Conclusion: MAP provides an effective facial privacy protection method using emotion modifications that works well across diverse demographics and challenging scenarios, outperforming existing approaches.

Abstract: The proliferation of facial recognition (FR) systems has raised privacy concerns in the digital realm, as malicious uses of FR models pose a significant threat. Traditional countermeasures, such as makeup style transfer, have suffered from low transferability in black-box settings and limited applicability across various demographic groups, including males and individuals with darker skin tones. To address these challenges, we introduce a novel facial privacy protection method, dubbed \textbf{MAP}, a pioneering approach that employs human emotion modifications to disguise original identities as target identities in facial images. Our method uniquely fine-tunes a score network to learn dual objectives, target identity and human expression, which are jointly optimized through gradient projection to ensure convergence at a shared local optimum. Additionally, we enhance the perceptual quality of protected images by applying local smoothness regularization and optimizing the score matching loss within our network. Empirical experiments demonstrate that our innovative approach surpasses previous baselines, including noise-based, makeup-based, and freeform attribute methods, in both qualitative fidelity and quantitative metrics. Furthermore, MAP proves its effectiveness against an online FR API and shows advanced adaptability in uncommon photographic scenarios.

Method: EvoPrune performs layer-wise pruning during visual encoding using token similarity, diversity, and attention-based importance metrics to retain the most informative visual tokens at selected encoding layers.

Result: Extensive experiments show effectiveness on image and video benchmarks. On VideoMME dataset, achieves 2x inference speedup with less than 1% performance degradation.

Conclusion: EvoPrune demonstrates potential for latency-sensitive MLLM deployment by addressing computational bottlenecks during visual encoding, not just after encoding.

Abstract: Multimodal Large Language Models (MLLMs) have shown strong performance in vision-language tasks, but their inference efficiency is severely limited by the exponential growth of visual tokens in complex scenarios such as high-resolution images and videos. Existing visual token pruning methods mainly operate after visual encoding, overlooking the substantial computational cost incurred during the encoding stage. To address this issue, we propose EvoPrune, an early-stage visual token pruning method for MLLMs that performs pruning directly during visual encoding. Specifically, EvoPrune employs a layer-wise pruning strategy guided by token similarity, diversity, and attention-based importance to retain the most informative visual tokens at selected encoding layers. Extensive experiments on image and video benchmarks validate the effectiveness of EvoPrune. In particular, on the VideoMME dataset, EvoPrune achieves 2$\times$ inference speedup with less than 1% performance degradation, demonstrating its potential for latency-sensitive MLLM deployment.

Method: Proposes Embedded Runge-Kutta Guidance (ERK-Guid) which leverages solver-induced error as guidance signal. The method exploits detected stiffness to reduce local truncation error and stabilize sampling. It theoretically and empirically analyzes stiffness and eigenvector estimators with solver errors to design the guidance mechanism.

Result: Experiments on synthetic datasets and ImageNet benchmark demonstrate that ERK-Guid consistently outperforms state-of-the-art methods in sample quality and generation performance.

Conclusion: ERK-Guid effectively addresses solver-induced errors in diffusion models by using them as guidance signals, improving conditional generation and sample quality particularly in stiff regions where traditional methods struggle.

Abstract: Classifier-Free Guidance (CFG) has established the foundation for guidance mechanisms in diffusion models, showing that well-designed guidance proxies significantly improve conditional generation and sample quality. Autoguidance (AG) has extended this idea, but it relies on an auxiliary network and leaves solver-induced errors unaddressed. In stiff regions, the ODE trajectory changes sharply, where local truncation error (LTE) becomes a critical factor that deteriorates sample quality. Our key observation is that these errors align with the dominant eigenvector, motivating us to leverage the solver-induced error as a guidance signal. We propose Embedded Runge-Kutta Guidance (ERK-Guid), which exploits detected stiffness to reduce LTE and stabilize sampling. We theoretically and empirically analyze stiffness and eigenvector estimators with solver errors to motivate the design of ERK-Guid. Our experiments on both synthetic datasets and the popular benchmark dataset, ImageNet, demonstrate that ERK-Guid consistently outperforms state-of-the-art methods. Code is available at https://github.com/mlvlab/ERK-Guid.

Method: Proposes MPFlow with two key components: 1) PAMRI (Patch-level Multi-modal MR Image Pretraining) - self-supervised pretraining strategy that learns shared representations across MRI modalities, and 2) Rectified flow-based sampling jointly guided by data consistency and cross-modal feature alignment using pre-trained PAMRI to suppress hallucinations.

Result: Extensive experiments on HCP and BraTS datasets show MPFlow matches diffusion baselines on image quality using only 20% of sampling steps while reducing tumor hallucinations by more than 15% (measured by segmentation dice score).

Conclusion: Cross-modal guidance enables more reliable and efficient zero-shot MRI reconstruction by leveraging auxiliary MRI modalities at inference time without retraining the generative prior, systematically suppressing both intrinsic and extrinsic hallucinations.

Abstract: Zero-shot MRI reconstruction relies on generative priors, but single-modality unconditional priors produce hallucinations under severe ill-posedness. In many clinical workflows, complementary MRI acquisitions (e.g. high-quality structural scans) are routinely available, yet existing reconstruction methods lack mechanisms to leverage this additional information. We propose MPFlow, a zero-shot multi-modal reconstruction framework built on rectified flow that incorporates auxiliary MRI modalities at inference time without retraining the generative prior to improve anatomical fidelity. Cross-modal guidance is enabled by our proposed self-supervised pretraining strategy, Patch-level Multi-modal MR Image Pretraining (PAMRI), which learns shared representations across modalities. Sampling is jointly guided by data consistency and cross-modal feature alignment using pre-trained PAMRI, systematically suppressing intrinsic and extrinsic hallucinations. Extensive experiments on HCP and BraTS show that MPFlow matches diffusion baselines on image quality using only 20% of sampling steps while reducing tumor hallucinations by more than 15% (segmentation dice score). This demonstrates that cross-modal guidance enables more reliable and efficient zero-shot MRI reconstruction.

Method: Decompose pixel values into binary bit-planes, apply LDP mechanism directly to bit-level representation, integrate perceptual obfuscation module to mitigate human-perceivable leakage, and use optimization-based privacy budget allocation strategy.

Result: Outperforms existing DP/LDP baselines under comparable privacy budgets for face recognition and image classification tasks with negligible computational overhead.

Conclusion: Utility loss in LDP for images is not inherent but due to domain mismatch; LDP-Slicing resolves this by appropriate data representation while satisfying rigorous pixel-level ε-LDP.

Abstract: Local Differential Privacy (LDP) is the gold standard trust model for privacy-preserving machine learning by guaranteeing privacy at the data source. However, its application to image data has long been considered impractical due to the high dimensionality of pixel space. Canonical LDP mechanisms are designed for low-dimensional data, resulting in severe utility degradation when applied to high-dimensional pixel spaces. This paper demonstrates that this utility loss is not inherent to LDP, but from its application to an inappropriate data representation. We introduce LDP-Slicing, a lightweight, training-free framework that resolves this domain mismatch. Our key insight is to decompose pixel values into a sequence of binary bit-planes. This transformation allows us to apply the LDP mechanism directly to the bit-level representation. To further strengthen privacy and preserve utility, we integrate a perceptual obfuscation module that mitigates human-perceivable leakage and an optimization-based privacy budget allocation strategy. This pipeline satisfies rigorous pixel-level $\varepsilon$-LDP while producing images that retain high utility for downstream tasks. Extensive experiments on face recognition and image classification demonstrate that LDP-Slicing outperforms existing DP/LDP baselines under comparable privacy budgets, with negligible computational overhead.

Method: Dual-backbone architecture: frozen DINOv3 vision foundation model produces general visual features, Swin model generates task-specific features. Multi-scale features are downsampled with residual Squeeze-and-Excitation Channel Reduction and fed into Mask2Former Decoder for segmentation masks.

Result: Achieved state-of-the-art results on four glass segmentation datasets across several accuracy metrics. Competitive inference speed compared to previous SOTA, and surpasses it with lighter DINOv3 variant.

Conclusion: The proposed architecture effectively addresses glass segmentation challenges by combining general and task-specific visual features, demonstrating strong performance and practical efficiency.

Abstract: Glass surface segmentation from RGB images is a challenging task, since glass as a transparent material distinctly lacks visual characteristics. However, glass segmentation is critical for scene understanding and robotics, as transparent glass surfaces must be identified as solid material. This paper presents a novel architecture for glass segmentation, deploying a dual-backbone producing general visual features as well as task-specific learned visual features. General visual features are produced by a frozen DINOv3 vision foundation model, and the task-specific features are generated with a Swin model trained in a supervised manner. Resulting multi-scale feature representations are downsampled with residual Squeeze-and-Excitation Channel Reduction, and fed into a Mask2Former Decoder, producing the final segmentation masks. The architecture was evaluated on four commonly used glass segmentation datasets, achieving state-of-the-art results on several accuracy metrics. The model also has a competitive inference speed compared to the previous state-of-the-art method, and surpasses it when using a lighter DINOv3 backbone variant. The implementation source code and model weights are available at: https://github.com/ojalar/lgnet

Method: Proposes QD-PCQA with two main components: 1) Rank-weighted Conditional Alignment (RCA) that aligns features under consistent quality levels while emphasizing misranked samples for ranking awareness, and 2) Quality-guided Feature Augmentation (QFA) with style mixup, multi-layer extension, and dual-domain augmentation modules for perceptual feature alignment.

Result: Extensive cross-domain experiments show QD-PCQA significantly improves generalization in NR-PCQA tasks, outperforming existing methods.

Conclusion: The framework effectively transfers quality-relevant priors from images to point clouds through quality-aware domain adaptation, addressing key limitations in existing UDA-based PCQA methods.

Abstract: No-Reference Point Cloud Quality Assessment (NR-PCQA) still struggles with generalization, primarily due to the scarcity of annotated point cloud datasets. Since the Human Visual System (HVS) drives perceptual quality assessment independently of media types, prior knowledge on quality learned from images can be repurposed for point clouds. This insight motivates adopting Unsupervised Domain Adaptation (UDA) to transfer quality-relevant priors from labeled images to unlabeled point clouds. However, existing UDA-based PCQA methods often overlook key characteristics of perceptual quality, such as sensitivity to quality ranking and quality-aware feature alignment, thereby limiting their effectiveness. To address these issues, we propose a novel Quality-aware Domain adaptation framework for PCQA, termed QD-PCQA. The framework comprises two main components: i) a Rank-weighted Conditional Alignment (RCA) strategy that aligns features under consistent quality levels and adaptively emphasizes misranked samples to reinforce perceptual quality ranking awareness; and ii) a Quality-guided Feature Augmentation (QFA) strategy, which includes quality-guided style mixup, multi-layer extension, and dual-domain augmentation modules to augment perceptual feature alignment. Extensive cross-domain experiments demonstrate that QD-PCQA significantly improves generalization in NR-PCQA tasks. The code is available at https://github.com/huhu-code/QD-PCQA.

Method: PROSPECT uses CUT3R as streaming 3D spatial encoder and SigLIP semantic features fused via cross-attention. It introduces learnable stream query tokens to predict next-step 2D/3D latent features in frozen teacher spaces, shaping internal representations without inference overhead.

Result: State-of-the-art performance on VLN-CE benchmarks and real-robot deployment, with improved long-horizon robustness under diverse lighting conditions.

Conclusion: PROSPECT demonstrates that coupling streaming VLA policies with latent predictive representation learning enables more robust navigation by learning environment dynamics and spatial structure.

Abstract: Multimodal large language models (MLLMs) have advanced zero-shot end-to-end Vision-Language Navigation (VLN), yet robust navigation requires not only semantic understanding but also predictive modeling of environment dynamics and spatial structure. We propose PROSPECT, a unified streaming navigation agent that couples a streaming Vision-Language-Action (VLA) policy with latent predictive representation learning. PROSPECT uses CUT3R as a streaming 3D foundation spatial encoder to produce long-context, absolute-scale spatial features, and fuses them with SigLIP semantic features via cross-attention. During training, we introduce learnable stream query tokens that query the streaming context and predict next-step 2D and 3D latent features (rather than pixels or explicit modalities), supervised in the latent spaces of frozen SigLIP and CUT3R teachers. The predictive branch shapes internal representations without inference overhead. Experiments on VLN-CE benchmarks and real-robot deployment demonstrate state-of-the-art performance and improved long-horizon robustness under diverse lighting. We will release code for the community soon.

Method: Dual-stream transformer with: 1) low-resolution stream on downsampled frames with alternating frame/global attention for view consistency and camera estimation, 2) high-resolution stream on original images per-frame for detail preservation, and 3) lightweight adapter using cross-attention to fuse streams without disturbing pretrained single-frame pathway.

Result: Achieves state-of-the-art results for video geometry estimation and multi-view reconstruction, delivering sharp depth/pointmaps, strong cross-view consistency, accurate poses, supporting inputs up to 2K resolution with practical inference cost.

Conclusion: DAGE effectively disentangles global coherence from fine detail through dual-stream design, enabling scalable high-resolution geometry estimation with view consistency and practical efficiency.

Abstract: Estimating accurate, view-consistent geometry and camera poses from uncalibrated multi-view/video inputs remains challenging - especially at high spatial resolutions and over long sequences. We present DAGE, a dual-stream transformer whose main novelty is to disentangle global coherence from fine detail. A low-resolution stream operates on aggressively downsampled frames with alternating frame/global attention to build a view-consistent representation and estimate cameras efficiently, while a high-resolution stream processes the original images per-frame to preserve sharp boundaries and small structures. A lightweight adapter fuses these streams via cross-attention, injecting global context without disturbing the pretrained single-frame pathway. This design scales resolution and clip length independently, supports inputs up to 2K, and maintains practical inference cost. DAGE delivers sharp depth/pointmaps, strong cross-view consistency, and accurate poses, establishing new state-of-the-art results for video geometry estimation and multi-view reconstruction.

Method: Proposes WSI-INR, which models WSIs as continuous implicit functions mapping spatial coordinates to tissue semantics features. Uses multi-resolution hash grid encoding to treat different resolution levels as varying sampling densities of the same continuous tissue, and jointly trains a shared INR decoder to capture general priors across cases.

Result: WSI-INR maintains robust segmentation performance across resolutions; at Base/4 resolution, resolution-specific optimization improves Dice score by +26.11%, while U-Net and TransUNet decrease by 54.28% and 36.18% respectively.

Conclusion: The work enables INRs to segment highly heterogeneous pathological lesions beyond structurally consistent anatomical tissues, offering a fresh perspective for pathological analysis.

Abstract: Whole-slide images (WSIs) are fundamental for computational pathology, where accurate lesion segmentation is critical for clinical decision making. Existing methods partition WSIs into discrete patches, disrupting spatial continuity and treating multi-resolution views as independent samples, which leads to spatially fragmented segmentation and reduced robustness to resolution variations. To address the issues, we propose WSI-INR, a novel patch-free framework based on Implicit Neural Representations (INRs). WSI-INR models the WSI as a continuous implicit function mapping spatial coordinates directly to tissue semantics features, outputting segmentation results while preserving intrinsic spatial information across the entire slide. In the WSI-INR, we incorporate multi-resolution hash grid encoding to regard different resolution levels as varying sampling densities of the same continuous tissue, achieving a consistent feature representation across resolutions. In addition, by jointly training a shared INR decoder, WSI-INR can capture general priors across different cases. Experimental results showed that WSI-INR maintains robust segmentation performance across resolutions; at Base/4, our resolution-specific optimization improves Dice score by +26.11%, while U-Net and TransUNet decrease by 54.28% and 36.18%, respectively. Crucially, this work enables INRs to segment highly heterogeneous pathological lesions beyond structurally consistent anatomical tissues, offering a fresh perspective for pathological analysis.

Method: KFRA uses a three-stage closed reasoning loop: 1) open-vocabulary detection and web-scale retrieval for category hypotheses, 2) discriminative regions localization through global-to-local focusing mechanism aligning textual knowledge with visual evidence, and 3) multimodal evidence integration in a large multimodal model for interpretable reasoning. Key innovation is retrieval-grounding coupling that converts retrieved knowledge into spatially grounded evidence.

Result: KFRA consistently surpasses standalone large multimodal models and current agent frameworks, achieving up to 19% improvement in reasoning accuracy. It delivers evidence-grounded interpretability in open-set fine-grained visual understanding, as demonstrated on the FGExpertBench benchmark across six knowledge dimensions.

Conclusion: KFRA successfully transforms fine-grained perception into evidence-driven reasoning through its unified framework, enabling factual, interpretable, and task-agnostic reasoning across diverse fine-grained scenarios by establishing retrieval-grounding coupling for knowledge verification.

Abstract: Fine-grained visual understanding is shifting from static classification to knowledge-augmented reasoning, where models must justify as well as recognise. Existing approaches remain limited by closed-set taxonomies and single-label prediction, leading to significant degradation under open-set or context-dependent conditions. We present the Knowledge-Augmented Fine-Grained Reasoning Agent (KFRA), a unified framework that transforms fine-grained perception into evidence-driven reasoning. KFRA operates through a three-stage closed reasoning loop that emulates expert analysis. It first performs open-vocabulary detection and web-scale retrieval to generate category hypotheses. It then conducts discriminative regions localisation by aligning textual knowledge with visual evidence through a global-to-local focusing mechanism. Finally, it integrates all multimodal evidence within a large multimodal model to perform interpretable reasoning. Unlike existing agents that treat retrieval and reasoning as independent processes, KFRA establishes a retrieval-grounding coupling that converts retrieved knowledge into spatially grounded evidence for verification. This design enables factual, interpretable, and task-agnostic reasoning across diverse fine-grained scenarios. To evaluate this capability, we construct FGExpertBench, a benchmark designed to assess reasoning depth and cross-task generalisation across six knowledge dimensions. Extensive experiments demonstrate that KFRA consistently surpasses both standalone large multimodal models and current agent frameworks, achieving up to 19 percent improvement in reasoning accuracy and delivering evidence-grounded interpretability in open-set fine-grained visual understanding.

Method: Uses sparse but accurate LiDAR observations as geometric prompts in two ways: (1) as hard geometric priors to anchor the cost volume, and (2) as soft feature-wise guidance fused by a triple-cue combiner. Employs a spatio-temporal decoder that leverages both geometric cues from MVS cost volume and temporal context from neighboring frames.

Result: Achieves state-of-the-art performance on multiple benchmarks, excelling in metric accuracy, temporal stability, and zero-shot cross-domain transfer. Demonstrates practical value for scalable, reliable autonomous driving systems.

Conclusion: DriveMVS successfully addresses the challenges of metric depth estimation for autonomous driving by using LiDAR prompts as geometric anchors and deep fusion techniques, achieving superior performance across multiple important metrics including cross-domain generalization.

Abstract: Accurate metric depth is critical for autonomous driving perception and simulation, yet current approaches struggle to achieve high metric accuracy, multi-view and temporal consistency, and cross-domain generalization. To address these challenges, we present DriveMVS, a novel multi-view stereo framework that reconciles these competing objectives through two key insights: (1) Sparse but metrically accurate LiDAR observations can serve as geometric prompts to anchor depth estimation in absolute scale, and (2) deep fusion of diverse cues is essential for resolving ambiguities and enhancing robustness, while a spatio-temporal decoder ensures consistency across frames. Built upon these principles, DriveMVS embeds the LiDAR prompt in two ways: as a hard geometric prior that anchors the cost volume, and as soft feature-wise guidance fused by a triple-cue combiner. Regarding temporal consistency, DriveMVS employs a spatio-temporal decoder that jointly leverages geometric cues from the MVS cost volume and temporal context from neighboring frames. Experiments show that DriveMVS achieves state-of-the-art performance on multiple benchmarks, excelling in metric accuracy, temporal stability, and zero-shot cross-domain transfer, demonstrating its practical value for scalable, reliable autonomous driving systems.

Method: Unpaired Neural Schrödinger Bridge (UNSB) framework with multi-step refinement. Augments adversarial objective with DMD2-style diffusion-guided distribution matching using frozen 3T diffusion teacher. Combines PatchNCE with Anatomical Structure Preservation (ASP) regularizer for global structure constraints beyond patch-level correspondence.

Result: Achieves improved realism-structure trade-off on two disjoint cohorts. Enhances distribution-level realism on unpaired benchmarks while increasing structural fidelity on paired cohort compared to unpaired baselines.

Conclusion: Proposed framework successfully addresses the challenge of unpaired 64 mT to 3 T MRI translation, improving both realism and anatomical preservation through diffusion guidance and structure-aware regularization.

Abstract: Ultra Low Field (64 mT) brain MRI improves accessibility but suffers from reduced image quality compared to 3 T. As paired 64 mT - 3 T scans are scarce, we propose an unpaired 64 mT $\rightarrow$ 3 T translation framework that enhances realism while preserving anatomy. Our method builds upon the Unpaired Neural Schrödinge Bridge (UNSB) with multi-step refinement. To strengthen target distribution alignment, we augment the adversarial objective with DMD2-style diffusion-guided distribution matching using a frozen 3T diffusion teacher. To explicitly constrain global structure beyond patch-level correspondence, we combine PatchNCE with an Anatomical Structure Preservation (ASP) regularizer that enforces soft foreground background consistency and boundary aware constraints. Evaluated on two disjoint cohorts, the proposed framework achieves an improved realism structure trade-off, enhancing distribution level realism on unpaired benchmarks while increasing structural fidelity on the paired cohort compared to unpaired baselines.

Method: Proposes a multi-level feature enhancement and global relation modeling framework with: 1) Residual Haar Wavelet Downsampling to preserve fine-grained details using spatial and frequency domains, 2) Global Relation Modeling to capture long-range dependencies and suppress background noise, 3) Cross-Scale Hybrid Attention for sparse, aligned interactions across multi-scale features, and 4) Center-Assisted Loss for stable training and improved localization.

Result: Extensive experiments on RGBT-Tiny benchmark show the method consistently outperforms state-of-the-art detectors under both IoU-based and scale-adaptive evaluation metrics, validating effectiveness and robustness for small object detection in complex environments.

Conclusion: The proposed framework effectively addresses small object detection challenges through multi-level feature enhancement and global relation modeling, demonstrating superior performance and robustness in complex scenarios.

Abstract: Small object detection under complex backgrounds remains a challenging task due to severe feature degradation, weak semantic representation, and inaccurate localization caused by downsampling operations and background interference. Existing detection frameworks are mainly designed for general objects and often fail to explicitly address the unique characteristics of small objects, such as limited structural cues and strong sensitivity to localization errors. In this paper, we propose a multi-level feature enhancement and global relation modeling framework tailored for small object detection. Specifically, a Residual Haar Wavelet Downsampling module is introduced to preserve fine-grained structural details by jointly exploiting spatial-domain convolutional features and frequency-domain representations. To enhance global semantic awareness and suppress background noise, a Global Relation Modeling module is employed to capture long-range dependencies at high-level feature stages. Furthermore, a Cross-Scale Hybrid Attention module is designed to establish sparse and aligned interactions across multi-scale features, enabling effective fusion of high-resolution details and high-level semantic information with reduced computational overhead. Finally, a Center-Assisted Loss is incorporated to stabilize training and improve localization accuracy for small objects. Extensive experiments conducted on the large-scale RGBT-Tiny benchmark demonstrate that the proposed method consistently outperforms existing state-of-the-art detectors under both IoU-based and scale-adaptive evaluation metrics. These results validate the effectiveness and robustness of the proposed framework for small object detection in complex environments.

Method: Uses a single full evaluation of the model’s first layer as a low-cost probe to compute proxy losses for candidate predictors (Taylor expansions of varying order/horizon), then assigns each token the predictor with smallest proxy error in a “probe-then-select” strategy.

Result: TAP incurs negligible overhead while enabling large speedups with little or no perceptual quality loss across multiple diffusion architectures and generation tasks, substantially improving accuracy-efficiency frontier.

Conclusion: TAP’s per-token adaptive predictor selection exploits heterogeneous temporal dynamics without additional training, offering a practical solution to accelerate diffusion model inference while maintaining quality.

Abstract: Diffusion models achieve strong generative performance but remain slow at inference due to the need for repeated full-model denoising passes. We present Token-Adaptive Predictor (TAP), a training-free, probe-driven framework that adaptively selects a predictor for each token at every sampling step. TAP uses a single full evaluation of the model’s first layer as a low-cost probe to compute proxy losses for a compact family of candidate predictors (instantiated primarily with Taylor expansions of varying order and horizon), then assigns each token the predictor with the smallest proxy error. This per-token “probe-then-select” strategy exploits heterogeneous temporal dynamics, requires no additional training, and is compatible with various predictor designs. TAP incurs negligible overhead while enabling large speedups with little or no perceptual quality loss. Extensive experiments across multiple diffusion architectures and generation tasks show that TAP substantially improves the accuracy-efficiency frontier compared to fixed global predictors and caching-only baselines.

Method: Introduces STAR (Separators for AutoRegressive pretraining) - inserting identical separators before each image to demarcate different images, allowing quadrupling of input sequence length while preserving original image dimensions.

Result: STAR-B model achieved 83.5% accuracy on ImageNet-1k, which is highly competitive among Vision Mamba models, demonstrating effectiveness of long sequence pretraining.

Conclusion: The method successfully enhances vision model performance by better leveraging long-range dependencies through extended sequence pretraining, showing potential for improved vision Mamba architectures.

Abstract: The state space model Mamba has recently emerged as a promising paradigm in computer vision, attracting significant attention due to its efficient processing of long sequence tasks. Mamba’s inherent causal mechanism renders it particularly suitable for autoregressive pretraining. However, current autoregressive pretraining methods are constrained to short sequence tasks, failing to fully exploit Mamba’s prowess in handling extended sequences. To address this limitation, we introduce an innovative autoregressive pretraining method for Vision Mamba that substantially extends the input sequence length. We introduce new \textbf{S}epara\textbf{T}ors for \textbf{A}uto\textbf{R}egressive pretraining to demarcate and differentiate between different images, known as \textbf{STAR}. Specifically, we insert identical separators before each image to demarcate its inception. This strategy enables us to quadruple the input sequence length of Vision Mamba while preserving the original dimensions of the dataset images. Employing this long sequence pretraining technique, our STAR-B model achieved an impressive accuracy of 83.5% on ImageNet-1k, which is highly competitive in Vision Mamba. These results underscore the potential of our method in enhancing the performance of vision models through improved leveraging of long-range dependencies.

Method: Proposes DPSA_FGIoU_YOLOv10n framework with: 1) Multi-Stage Adaptive Enhancement module for image quality improvement, 2) Dual-Pooling Sequential Attention (DPSA) mechanism in backbone for multi-scale feature representation, 3) Focal Generalized IoU Objectness (FGIoU) loss for better localization and objectness prediction under class imbalance.

Result: Achieves 88.9% mAP on RUOD and 88.0% mAP on DUO datasets at IoU threshold 0.5, representing improvements of 6.7% and 6.2% over baseline YOLOv10n while maintaining compact 2.8M parameter architecture.

Conclusion: The framework establishes effective balance among accuracy, robustness, and real-time efficiency, making it suitable for deployment in resource-constrained underwater settings.

Abstract: Underwater object detection constitutes a pivotal endeavor within the realms of marine surveillance and autonomous underwater systems; however, it presents significant challenges due to pronounced visual impairments arising from phenomena such as light absorption, scattering, and diminished contrast. In response to these formidable challenges, this manuscript introduces a streamlined yet robust framework for underwater object detection, grounded in the YOLOv10 architecture. The proposed method integrates a Multi-Stage Adaptive Enhancement module to improve image quality, a Dual-Pooling Sequential Attention (DPSA) mechanism embedded into the backbone to strengthen multi-scale feature representation, and a Focal Generalized IoU Objectness (FGIoU) loss to jointly improve localization accuracy and objectness prediction under class imbalance. Comprehensive experimental evaluations conducted on the RUOD and DUO benchmark datasets substantiate that the proposed DPSA_FGIoU_YOLOv10n attains exceptional performance, achieving mean Average Precision (mAP) scores of 88.9% and 88.0% at IoU threshold 0.5, respectively. In comparison to the baseline YOLOv10n, this represents enhancements of 6.7% for RUOD and 6.2% for DUO, all while preserving a compact model architecture comprising merely 2.8M parameters. These findings validate that the proposed framework establishes an efficacious equilibrium among accuracy, robustness, and real-time operational efficiency, making it suitable for deployment in resource-constrained underwater settings.

Method: Introduces a vector-quantized autoencoder (VQAE) for compressing soft labels in dataset distillation. The method analyzes bit requirements across distillation frameworks and applies compression to soft labels while preserving their effectiveness for student model training.

Result: Achieves 30-40x additional compression over RDED, LPLD, SRE2L, and CDA baselines on ImageNet-1K while retaining over 90% of original performance. Validated on both vision and language distillation benchmarks.

Conclusion: The VQAE approach effectively addresses the storage overhead problem in dataset distillation, making the technique more practical for real-world applications by significantly reducing soft label storage requirements without substantial performance degradation.

Abstract: Dataset distillation is an emerging technique for reducing the computational and storage costs of training machine learning models by synthesizing a small, informative subset of data that captures the essential characteristics of a much larger dataset. Recent methods pair synthetic samples and their augmentations with soft labels from a teacher model, enabling student models to generalize effectively despite the small size of the distilled dataset. While soft labels are critical for effective distillation, the storage and communication overhead they incur, especially when accounting for augmentations, is often overlooked. In practice, each distilled sample is associated with multiple soft labels, making them the dominant contributor to storage costs, particularly in large-class settings such as ImageNet-1K. In this paper, we present a rigorous analysis of bit requirements across dataset distillation frameworks, quantifying the storage demands of both distilled samples and their soft labels. To address the overhead, we introduce a vector-quantized autoencoder (VQAE) for compressing soft labels, achieving substantial compression while preserving the effectiveness of the distilled data. We validate our method on both vision and language distillation benchmarks. On ImageNet-1K, our proposed VQAE achieves 30–40x additional compression over RDED, LPLD, SRE2L, and CDA baselines while retaining over $90%$ of their original performance.

Method: Proposes Structure-aware Prompt Adaptation (SPA) with two components: 1) Structure-aware Consistency Loss (SCL) during training to encourage local structure consistency of seen attributes and objects, and 2) Structure-guided Adaptation Strategy (SAS) during inference to adaptively align structures of unseen concepts with semantically similar seen ones.

Result: Extensive experiments on OV-CZSL benchmarks show SPA achieves competitive closed-set performance while significantly improving open-vocabulary results. The method is plug-and-play and can be integrated into existing CZSL prompt tuning methods.

Conclusion: SPA effectively bridges the gap between closed and open-vocabulary CZSL by leveraging semantic structure consistency, enabling better generalization to unseen attributes, objects, and their compositions through analogical reasoning inspired by human cognition.

Abstract: The goal of Open-Vocabulary Compositional Zero-Shot Learning (OV-CZSL) is to recognize attribute-object compositions in the open-vocabulary setting, where compositions of both seen and unseen attributes and objects are evaluated. Recently, prompt tuning methods have demonstrated strong generalization capabilities in the closed setting, where only compositions of seen attributes and objects are evaluated, i.e., Compositional Zero-Shot Learning (CZSL). However, directly applying these methods to OV-CZSL may not be sufficient to generalize to unseen attributes, objects and their compositions, as it is limited to seen attributes and objects. Normally, when faced with unseen concepts, humans adopt analogies with seen concepts that have the similar semantics thereby inferring their meaning (e.g., “wet” and “damp”, “shirt” and “jacket”). In this paper, we experimentally show that the distribution of semantically related attributes or objects tends to form consistent local structures in the embedding space. Based on the above structures, we propose Structure-aware Prompt Adaptation (SPA) method, which enables models to generalize from seen to unseen attributes and objects. Specifically, in the training stage, we design a Structure-aware Consistency Loss (SCL) that encourages the local structure’s consistency of seen attributes and objects in each iteration. In the inference stage, we devise a Structure-guided Adaptation Strategy (SAS) that adaptively aligns the structures of unseen attributes and objects with those of trained seen attributes and objects with similar semantics. Notably, SPA is a plug-and-play method that can be seamlessly integrated into existing CZSL prompt tuning methods. Extensive experiments on OV-CZSL benchmarks demonstrate that SPA achieves competitive closed-set performance while significantly improving open-vocabulary results.

Method: 1) Introduce Visual Attention Score (VAS) - an attention-based metric quantifying model attention to visual tokens. 2) Analyze correlation between VAS and reasoning performance. 3) Discover Lazy Attention Localization phenomenon where multimodal cold-start fails to elevate VAS. 4) Design training-free interventions to modulate attention during inference. 5) Propose AVAR framework with visual-anchored data synthesis, attention-guided objectives, and visual-anchored reward shaping.

Result: 1) Strong correlation found between reasoning performance and VAS (r=0.9616). 2) Training-free interventions yield 1-2% performance gains without retraining. 3) AVAR framework achieves average gain of 7.0% across 7 multimodal reasoning benchmarks on Qwen2.5-VL-7B. 4) Ablation studies confirm each AVAR component contributes step-wise to overall gains.

Conclusion: The paper provides fundamental insights into cold-start initialization mechanisms in MLRMs, introduces effective analysis tools (VAS), discovers important phenomena (Lazy Attention Localization), and proposes a practical framework (AVAR) that significantly improves multimodal reasoning performance through better attention mechanisms.

Abstract: The cold-start initialization stage plays a pivotal role in training Multimodal Large Reasoning Models (MLRMs), yet its mechanisms remain insufficiently understood. To analyze this stage, we introduce the Visual Attention Score (VAS), an attention-based metric that quantifies how much a model attends to visual tokens. We find that reasoning performance is strongly correlated with VAS (r=0.9616): models with higher VAS achieve substantially stronger multimodal reasoning. Surprisingly, multimodal cold-start fails to elevate VAS, resulting in attention distributions close to the base model, whereas text-only cold-start leads to a clear increase. We term this counter-intuitive phenomenon Lazy Attention Localization. To validate its causal role, we design training-free interventions that directly modulate attention allocation during inference, performance gains of 1$-$2% without any retraining. Building on these insights, we further propose Attention-Guided Visual Anchoring and Reflection (AVAR), a comprehensive cold-start framework that integrates visual-anchored data synthesis, attention-guided objectives, and visual-anchored reward shaping. Applied to Qwen2.5-VL-7B, AVAR achieves an average gain of 7.0% across 7 multimodal reasoning benchmarks. Ablation studies further confirm that each component of AVAR contributes step-wise to the overall gains. The code, data, and models are available at https://github.com/lrlbbzl/Qwen-AVAR.

Method: 1) Modality-interleaved transformer learns band-wise modal specializations to form reversible spectral affine bases, mapping arbitrary-band MS images into a unified latent space via tensor multiplication. 2) Latent diffusion bridge model progressively evolves latent representations with bridge posterior sampling to couple latent diffusion with pixel-space observations. 3) Infinite-dimensional pixel-to-latent interaction mechanisms capture cross-domain dependencies between PAN observations and MS representations for complementary information fusion.

Result: FoundPS consistently outperforms state-of-the-art methods, exhibiting superior generalization and robustness across a wide range of pansharpening tasks. The authors also constructed PSBench, a comprehensive benchmark with worldwide MS and PAN image pairs from multiple satellites across diverse scenes.

Conclusion: The proposed FoundPS model successfully addresses the limitations of existing satellite-specific and scene-dependent pansharpening methods by creating a universal foundation model that achieves robust performance across heterogeneous sensors and varied scenes through innovative architectural designs.

Abstract: Pansharpening generates the high-resolution multi-spectral (MS) image by integrating spatial details from a texture-rich panchromatic (PAN) image and spectral attributes from a low-resolution MS image. Existing methods are predominantly satellite-specific and scene-dependent, which severely limits their generalization across heterogeneous sensors and varied scenes, thereby reducing their real-world practicality. To address these challenges, we present FoundPS, a universal pansharpening foundation model for satellite-agnostic and scene-robust fusion. Specifically, we introduce a modality-interleaved transformer that learns band-wise modal specializations to form reversible spectral affine bases, mapping arbitrary-band MS into a unified latent space via tensor multiplication. Building upon this, we construct a latent diffusion bridge model to progressively evolve latent representations, and incorporate bridge posterior sampling to couple latent diffusion with pixel-space observations, enabling stable and controllable fusion. Furthermore, we devise infinite-dimensional pixel-to-latent interaction mechanisms to comprehensively capture the cross-domain dependencies between PAN observations and MS representations, thereby facilitating complementary information fusion. In addition, to support large-scale training and evaluation, we construct a comprehensive pansharpening benchmark, termed PSBench, consisting of worldwide MS and PAN image pairs from multiple satellites across diverse scenes. Extensive experiments demonstrate that FoundPS consistently outperforms state-of-the-art methods, exhibiting superior generalization and robustness across a wide range of pansharpening tasks.

Method: Proposes CWP-Net with wavelet attention modules in encoder and decoder to explicitly disentangle degradation and semantic features, addressing spurious correlations. Uses wavelet prompt block to generate alternative variables for causal deconfounding to address biased degradation pattern estimation.

Result: Extensive experiments on two all-in-one settings demonstrate effectiveness and superior performance over state-of-the-art AiOIR methods.

Conclusion: CWP-Net effectively addresses key limitations in AiOIR through causal deconfounding and wavelet-based disentanglement, achieving better restoration performance by modeling true causation between degraded and restored images.

Abstract: Image restoration represents a promising approach for addressing the inherent defects of image content distortion. Standard image restoration approaches suffer from high storage cost and the requirement towards the known degradation pattern, including type and degree, which can barely be satisfied in dynamic practical scenarios. In contrast, all-in-one image restoration (AiOIR) eliminates multiple degradations within a unified model to circumvent the aforementioned issues. However, according to our causal analysis, we disclose that two significant defects still exacerbate the effectiveness and generalization of AiOIR models: 1) the spurious correlation between non-degradation semantic features and degradation patterns; 2) the biased estimation of degradation patterns. To obtain the true causation between degraded images and restored images, we propose Causal-deconfounding Wavelet-disentangled Prompt Network (CWP-Net) to perform effective AiOIR. CWP-Net introduces two modules for decoupling, i.e., wavelet attention module of encoder and wavelet attention module of decoder. These modules explicitly disentangle the degradation and semantic features to tackle the issue of spurious correlation. To address the issue stemming from the biased estimation of degradation patterns, CWP-Net leverages a wavelet prompt block to generate the alternative variable for causal deconfounding. Extensive experiments on two all-in-one settings prove the effectiveness and superior performance of our proposed CWP-Net over the state-of-the-art AiOIR methods.

Method: Three-stage framework: 1) Hierarchical Scanning performs local cue exploration and multi-scale evidence extraction bottom-up; 2) Refocusing optimizes evidence view via LVLM-visual expert collaboration; 3) Evidence-Enhanced Reasoning aggregates multi-granular views using hybrid evidence memory.

Result: Achieves 90.6% overall accuracy on V* benchmark with Qwen2.5-VL-7B, significantly boosts LVLMs across diverse visual tasks, especially fine-grained understanding, and provides consistent improvements across various architectures and model scales without adaptation cost.

Conclusion: DeepScan effectively enhances LVLMs’ visual grounding and reasoning capabilities through bottom-up evidence extraction and multi-view aggregation, offering a training-free solution for improved performance in noisy visual environments.

Abstract: Humans can robustly localize visual evidence and provide grounded answers even in noisy environments by identifying critical cues and then relating them to the full context in a bottom-up manner. Inspired by this, we propose DeepScan, a training-free framework that combines Hierarchical Scanning, Refocusing, and Evidence-Enhanced Reasoning for visually grounded reasoning in Large Vision-Language Models (LVLMs). Unlike existing methods that pursue one-shot localization of complete evidence, Hierarchical Scanning performs local cue exploration and multi-scale evidence extraction to recover evidence in a bottom-up manner, effectively mitigating the impacts of distractive context. Refocusing then optimizes the localized evidence view through collaboration of LVLMs and visual experts. Finally, Evidence-Enhanced Reasoning aggregates multi-granular views via a hybrid evidence memory and yields accurate and interpretable answers. Experimental results demonstrate that DeepScan significantly boosts LVLMs in diverse visual tasks, especially in fine-grained visual understanding. It achieves 90.6% overall accuracy on V* when integrated with Qwen2.5-VL-7B. Moreover, DeepScan provides consistent improvements for LVLMs across various architectures and model scales without additional adaptation cost.

Method: Proposes a feedback reinforcement framework: 1) Creates first large-scale human feedback dataset for IVIF with subjective scores and artifact annotations, enhanced by fine-tuned LLM with expert review; 2) Designs domain-specific reward function and trains reward model to quantify perceptual quality; 3) Fine-tunes fusion network using Group Relative Policy Optimization guided by the reward model.

Result: Achieves state-of-the-art performance with fused images that better align with human aesthetics, as demonstrated through the proposed framework and dataset.

Conclusion: The feedback reinforcement framework successfully bridges human evaluation to infrared and visible image fusion, addressing the ill-posed nature of IVIF and producing results that better match human visual preferences for practical applications.

Abstract: Infrared and visible image fusion (IVIF) integrates complementary modalities to enhance scene perception. Current methods predominantly focus on optimizing handcrafted losses and objective metrics, often resulting in fusion outcomes that do not align with human visual preferences. This challenge is further exacerbated by the ill-posed nature of IVIF, which severely limits its effectiveness in human perceptual environments such as security surveillance and driver assistance systems. To address these limitations, we propose a feedback reinforcement framework that bridges human evaluation to infrared and visible image fusion. To address the lack of human-centric evaluation metrics and data, we introduce the first large-scale human feedback dataset for IVIF, containing multidimensional subjective scores and artifact annotations, and enriched by a fine-tuned large language model with expert review. Based on this dataset, we design a domain-specific reward function and train a reward model to quantify perceptual quality. Guided by this reward, we fine-tune the fusion network through Group Relative Policy Optimization, achieving state-of-the-art performance that better aligns fused images with human aesthetics. Code is available at https://github.com/ALKA-Wind/EVAFusion.

Method: Enhances YOLO-based architecture with an auxiliary head that regresses 2D projections of 3D bounding box corners (keypoint detection). Uses continuous 9D rotation representation projected to SO(3) via SVD for stable end-to-end training. Single-stage, end-to-end framework.

Result: Achieves 96.24% accuracy on LINEMOD and 69.41% on LINEMOD-Occluded benchmarks using ADD(-S) 0.1d metric. Operates in real-time, demonstrating competitive accuracy with high efficiency.

Conclusion: Carefully designed single-stage methods can provide practical balance of performance and efficiency for real-world deployment in 6D pose estimation, enabling real-time applications in robotics and extended reality.

Abstract: Estimating the 6D pose of objects from a single RGB image is a critical task for robotics and extended reality applications. However, state-of-the-art multi stage methods often suffer from high latency, making them unsuitable for real time use. In this paper, we present Yolo-Key-6D, a novel single stage, end-to-end framework for monocular 6D pose estimation designed for both speed and accuracy. Our approach enhances a YOLO based architecture by integrating an auxiliary head that regresses the 2D projections of an object’s 3D bounding box corners. This keypoint detection task significantly improves the network’s understanding of 3D geometry. For stable end-to-end training, we directly regress rotation using a continuous 9D representation projected to SO(3) via singular value decomposition. On the LINEMOD and LINEMOD-Occluded benchmarks, YOLO-Key-6D achieves competitive accuracy scores of 96.24% and 69.41%, respectively, with the ADD(-S) 0.1d metric, while proving itself to operate in real time. Our results demonstrate that a carefully designed single stage method can provide a practical and effective balance of performance and efficiency for real world deployment.

Method: UniSync uses a mask-free pose-anchored training strategy to preserve head motion and eliminate color artifacts, combined with mask-based blending consistent inference for structural precision. The model is fine-tuned on compact but diverse videos for exceptional domain adaptability.

Result: Extensive experiments show UniSync significantly outperforms state-of-the-art methods. The authors also introduce the RealWorld-LipSync benchmark covering diverse scenarios including human faces and stylized avatars.

Conclusion: UniSync advances lip synchronization towards truly generalizable and production-ready solutions by effectively handling complex corner cases through its unified framework and diverse training approach.

Abstract: Lip synchronization aims to generate realistic talking videos that match given audio, which is essential for high-quality video dubbing. However, current methods have fundamental drawbacks: mask-based approaches suffer from local color discrepancies, while mask-free methods struggle with global background texture misalignment. Furthermore, most methods struggle with diverse real-world scenarios such as stylized avatars, face occlusion, and extreme lighting conditions. In this paper, we propose UniSync, a unified framework designed for achieving high-fidelity lip synchronization in diverse scenarios. Specifically, UniSync uses a mask-free pose-anchored training strategy to keep head motion and eliminate synthesis color artifacts, while employing mask-based blending consistent inference to ensure structural precision and smooth blending. Notably, fine-tuning on compact but diverse videos empowers our model with exceptional domain adaptability, handling complex corner cases effectively. We also introduce the RealWorld-LipSync benchmark to evaluate models under real-world demands, which covers diverse application scenarios including both human faces and stylized avatars. Extensive experiments demonstrate that UniSync significantly outperforms state-of-the-art methods, advancing the field towards truly generalizable and production-ready lip synchronization.

Method: Developed a convolution-inspired architecture that gradually down-scales point clouds while integrating local neighbor information. Final down-scaled point cloud is processed by computing neighbors in feature space to include global information.

Result: The method consistently outperforms state-of-the-art transformer-based network Point-BERT in classification performance.

Conclusion: The proposed architecture effectively addresses cuneiform tablet metadata classification by combining local and global information processing in point clouds, outperforming existing methods.

Abstract: In this paper, we present a network structure for classifying metadata of cuneiform tablets. The problem is of practical importance, as the size of the existing corpus far exceeds the number of experts available to analyze it. But the task is made difficult by the combination of limited annotated datasets and the high-resolution point-cloud representation of each tablet. To address this, we develop a convolution-inspired architecture that gradually down-scales the point cloud while integrating local neighbor information. The final down-scaled point cloud is then processed by computing neighbors in the feature space to include global information. Our method is compared with the state-of-the-art transformer-based network Point-BERT, and consistently obtains the best performance. Source code and datasets will be released at publication.

Method: Proposes double scoring functions that simultaneously handle OOD detection and failure prediction, with new metrics DS-F1 and DS-AURC. Extends SURE classifier to SURE+ for improved reliability across diverse scenarios.

Result: Double scoring functions yield substantially more reliable classifiers than traditional single scoring approaches. OOD-based approaches provide notable gains under simple/far-OOD shifts but only marginal benefits under challenging near-OOD conditions.

Conclusion: The framework, metrics, and SURE+ method establish a new benchmark for trustworthy classification and offer practical guidance for deploying robust models in real-world settings.

Abstract: Building reliable classifiers is a fundamental challenge for deploying machine learning in real-world applications. A reliable system should not only detect out-of-distribution (OOD) inputs but also anticipate in-distribution (ID) errors by assigning low confidence to potentially misclassified samples. Yet, most prior work treats OOD detection and failure prediction as separated problems, overlooking their closed connection. We argue that reliability requires evaluating them jointly. To this end, we propose a unified evaluation framework that integrates OOD detection and failure prediction, quantified by our new metrics DS-F1 and DS-AURC, where DS denotes double scoring functions. Experiments on the OpenOOD benchmark show that double scoring functions yield classifiers that are substantially more reliable than traditional single scoring approaches. Our analysis further reveals that OOD-based approaches provide notable gains under simple or far-OOD shifts, but only marginal benefits under more challenging near-OOD conditions. Beyond evaluation, we extend the reliable classifier SURE and introduce SURE+, a new approach that significantly improves reliability across diverse scenarios. Together, our framework, metrics, and method establish a new benchmark for trustworthy classification and offer practical guidance for deploying robust models in real-world settings. The source code is publicly available at https://github.com/Intellindust-AI-Lab/SUREPlus.

Method: Proposes Modular Asynchronous Tracking Architecture (MATA) combining transformer-based tracker with Extended Kalman Filter, integrating ego-motion compensation from sparse optical flow and object trajectory model. Also introduces hardware-independent embedded evaluation protocol and Normalized time to Failure (NT2F) metric.

Result: Experiments on UAV benchmarks including augmented UAV123 dataset with synthetic occlusions show consistent improvements in Success and NT2F metrics across multiple tracking processing frequencies. ROS 2 implementation on Nvidia Jetson AGX Orin confirms evaluation protocol matches real-time embedded performance.

Conclusion: MATA provides robust UAV tracking solution for embedded systems, with proposed evaluation protocol better reflecting real-world embedded performance than traditional metrics.

Abstract: Object tracking from Unmanned Aerial Vehicles (UAVs) is challenged by platform dynamics, camera motion, and limited onboard resources. Existing visual trackers either lack robustness in complex scenarios or are too computationally demanding for real-time embedded use. We propose an Modular Asynchronous Tracking Architecture (MATA) that combines a transformer-based tracker with an Extended Kalman Filter, integrating ego-motion compensation from sparse optical flow and an object trajectory model. We further introduce a hardware-independent, embedded oriented evaluation protocol and a new metric called Normalized time to Failure (NT2F) to quantify how long a tracker can sustain a tracking sequence without external help. Experiments on UAV benchmarks, including an augmented UAV123 dataset with synthetic occlusions, show consistent improvements in Success and NT2F metrics across multiple tracking processing frequency. A ROS 2 implementation on a Nvidia Jetson AGX Orin confirms that the evaluation protocol more closely matches real-time performance on embedded systems.

Method: The authors create FGAesthetics database with 32,217 images organized into 10,028 series from Natural, AIGC, and Cropping categories. They use pairwise comparisons for annotations and apply Series Refinement and Rank Calibration for data reliability. They propose FGAesQ framework with Difference-preserved Tokenization (DiffToken), Comparative Text-assisted Alignment (CTAlign), and Rank-aware Regression (RankReg) to learn discriminative aesthetic scores from relative ranks.

Result: FGAesQ enables accurate aesthetic assessment in fine-grained scenarios while maintaining competitive performance in coarse-grained evaluation. Extensive experiments demonstrate the superiority of the proposed method over existing approaches.

Conclusion: The work addresses a critical gap in image aesthetic assessment by providing both a fine-grained dataset and a novel framework that effectively learns from relative rankings to discriminate subtle aesthetic differences, making it valuable for practical applications requiring precise aesthetic discrimination.

Abstract: Image aesthetic assessment (IAA) has extensive applications in content creation, album management, and recommendation systems, etc. In such applications, it is commonly needed to pick out the most aesthetically pleasing image from a series of images with subtle aesthetic variations, a topic we refer to as fine-grained IAA. Unfortunately, state-of-the-art IAA models are typically designed for coarse-grained evaluation, where images with notable aesthetic differences are evaluated independently on an absolute scale. These models are inherently limited in discriminating fine-grained aesthetic differences. To address the dilemma, we contribute FGAesthetics, a fine-grained IAA database with 32,217 images organized into 10,028 series, which are sourced from diverse categories including Natural, AIGC, and Cropping. Annotations are collected via pairwise comparisons within each series. We also devise Series Refinement and Rank Calibration to ensure the reliability of data and labels. Based on FGAesthetics, we further propose FGAesQ, a novel IAA framework that learns discriminative aesthetic scores from relative ranks through Difference-preserved Tokenization (DiffToken), Comparative Text-assisted Alignment (CTAlign), and Rank-aware Regression (RankReg). FGAesQ enables accurate aesthetic assessment in fine-grained scenarios while still maintains competitive performance in coarse-grained evaluation. Extensive experiments and comparisons demonstrate the superiority of the proposed method.

Method: Proposes external n-gram injection (NGI) for dynamic adaptation of network’s language modeling at inference time. Uses early injection of n-gram into transformer decoder, allowing switching to n-gram language model estimated on corpus close to target distribution without extra training on target image-text pairs.

Result: Experiments on three handwritten datasets show NGI significantly reduces performance gap between source and target corpora, mitigating bias without requiring additional training on target data.

Conclusion: NGI enables transformer-based handwritten text recognition systems to adapt to language distribution shifts at inference time using only text-only data, maintaining recognition accuracy with minimal computational overhead.

Abstract: Transformer-based encoder-decoder networks have recently achieved impressive results in handwritten text recognition, partly thanks to their auto-regressive decoder which implicitly learns a language model. However, such networks suffer from a large performance drop when evaluated on a target corpus whose language distribution is shifted from the source text seen during training. To retain recognition accuracy despite this language shift, we propose an external n-gram injection (NGI) for dynamic adaptation of the network’s language modeling at inference time. Our method allows switching to an n-gram language model estimated on a corpus close to the target distribution, therefore mitigating bias without any extra training on target image-text pairs. We opt for an early injection of the n-gram into the transformer decoder so that the network learns to fully leverage text-only data at the low additional cost of n-gram inference. Experiments on three handwritten datasets demonstrate that the proposed NGI significantly reduces the performance gap between source and target corpora.

Method: Proposes DISC with a single-pass, distance-weighted extraction mechanism that derives CLIP embeddings directly from vision transformer intermediate layers, avoiding image cropping. Built on GPU-accelerated architecture with on-the-fly voxel-level instance refinement for continuous mapping.

Result: Significantly surpasses current state-of-the-art zero-shot methods in both semantic accuracy and query retrieval on Replica, ScanNet, and new HM3DSEM datasets. Provides robust, real-time capable framework.

Conclusion: DISC overcomes fundamental limitations of crop-based feature extraction, enabling efficient, high-fidelity semantic mapping suitable for real-time robotic deployment with improved accuracy and scalability.

Abstract: Open-set semantic mapping enables language-driven robotic perception, but current instance-centric approaches are bottlenecked by context-depriving and computationally expensive crop-based feature extraction. To overcome this fundamental limitation, we introduce DISC (Dense Integrated Semantic Context), featuring a novel single-pass, distance-weighted extraction mechanism. By deriving high-fidelity CLIP embeddings directly from the vision transformer’s intermediate layers, our approach eliminates the latency and domain-shift artifacts of traditional image cropping, yielding pure, mask-aligned semantic representations. To fully leverage these features in large-scale continuous mapping, DISC is built upon a fully GPU-accelerated architecture that replaces periodic offline processing with precise, on-the-fly voxel-level instance refinement. We evaluate our approach on standard benchmarks (Replica, ScanNet) and a newly generated large-scale-mapping dataset based on Habitat-Matterport 3D (HM3DSEM) to assess scalability across complex scenes in multi-story buildings. Extensive evaluations demonstrate that DISC significantly surpasses current state-of-the-art zero-shot methods in both semantic accuracy and query retrieval, providing a robust, real-time capable framework for robotic deployment. The full source code, data generation and evaluation pipelines will be made available at https://github.com/DFKI-NI/DISC.

Method: Combines bidirectional 2D↔3D cross-modal mapping to model appearance-geometry consistency with dual-branch reconstruction that independently captures normal texture and geometric structure. Uses reliability-gated mapping anomaly and confidence-weighted reconstruction anomaly fusion strategies.

Result: Achieves 97.3% image-level AUROC, 99.6% pixel-level AUROC, and 97.6% AUPRO on MVTec 3D-AD benchmark. On real-world polyurethane cutting dataset, 3D-only variant attains 92.6% I-AUROC and 92.5% P-AUROC.

Conclusion: The framework demonstrates robustness, modality flexibility, and effectiveness of proposed fusion strategies for industrial visual inspection, operating without memory banks.

Abstract: Multimodal industrial anomaly detection benefits from integrating RGB appearance with 3D surface geometry, yet existing \emph{unsupervised} approaches commonly rely on memory banks, teacher-student architectures, or fragile fusion schemes, limiting robustness under noisy depth, weak texture, or missing modalities. This paper introduces \textbf{CMDR-IAD}, a lightweight and modality-flexible unsupervised framework for reliable anomaly detection in 2D+3D multimodal as well as single-modality (2D-only or 3D-only) settings. \textbf{CMDR-IAD} combines bidirectional 2D$\leftrightarrow$3D cross-modal mapping to model appearance-geometry consistency with dual-branch reconstruction that independently captures normal texture and geometric structure. A two-part fusion strategy integrates these cues: a reliability-gated mapping anomaly highlights spatially consistent texture-geometry discrepancies, while a confidence-weighted reconstruction anomaly adaptively balances appearance and geometric deviations, yielding stable and precise anomaly localization even in depth-sparse or low-texture regions. On the MVTec 3D-AD benchmark, CMDR-IAD achieves state-of-the-art performance while operating without memory banks, reaching 97.3% image-level AUROC (I-AUROC), 99.6% pixel-level AUROC (P-AUROC), and 97.6% AUPRO. On a real-world polyurethane cutting dataset, the 3D-only variant attains 92.6% I-AUROC and 92.5% P-AUROC, demonstrating strong effectiveness under practical industrial conditions. These results highlight the framework’s robustness, modality flexibility, and the effectiveness of the proposed fusion strategies for industrial visual inspection. Our source code is available at https://github.com/ECGAI-Research/CMDR-IAD/

Method: Retrospective study using 11,806 slices from 3T breast MRI exams. Three CNN architectures (DenseNet121, ResNet18, SEResNet50) were trained for binary artifact classification, with DenseNet121 further trained for multiclass classification. Evaluation used AUROC, AUPRC, precision, recall, and Grad-CAM heatmaps for bounding box analysis.

Result: DenseNet121 achieved AUROCs of 0.92 (hyperintense) and 0.94 (hypointense) for binary classification, and weighted AUROCs of 0.85 and 0.88 for multiclass classification. Radiologist evaluation of Grad-CAM bounding boxes showed mean scores of 3.33±1.04 for hyperintense and 2.62±0.81 for hypointense artifacts.

Conclusion: CNN-based detection of hyper- and hypointense artifacts in breast DWI MRI shows promising results, particularly with DenseNet121 architecture, though further validation is needed for clinical implementation.

Abstract: Diffusion-weighted imaging (DWI) can support lesion detection and characterization in breast magnetic resonance imaging (MRI), however especially high b-value diffusion-weighted acquisitions can be prone to intensity artifacts that can affect diagnostic image assessment. This study aims to detect both hyper- and hypointense artifacts on high b-value diffusion-weighted images (b=1500 s/mm2) using deep learning, employing either a binary classification (artifact presence) or a multiclass classification (artifact intensity) approach on a slice-wise dataset.This IRB-approved retrospective study used the single-center dataset comprising n=11806 slices from routine 3T breast MRI examinations performed between 2022 and mid-2023. Three convolutional neural network (CNN) architectures (DenseNet121, ResNet18, and SEResNet50) were trained for binary classification of hyper- and hypointense artifacts. The best performing model (DenseNet121) was applied to an independent holdout test set and was further trained separately for multiclass classification. Evaluation included area under receiver operating characteristic curve (AUROC), area under precision recall curve (AUPRC), precision, and recall, as well as analysis of predicted bounding box positions, derived from the network Grad-CAM heatmaps. DenseNet121 achieved AUROCs of 0.92 and 0.94 for hyper- and hypointense artifact detection, respectively, and weighted AUROCs of 0.85 and 0.88 for multiclass classification on single-slice high b-value diffusion-weighted images. A radiologist evaluated bounding box precision on a 1-5 Likert-like scale across 200 slices, achieving mean scores of 3.33+-1.04 for hyperintense artifacts and 2.62+-0.81 for hypointense artifacts. Hyper- and hypointense artifact detection in slice-wise breast DWI MRI dataset (b=1500 s/mm2) using CNNs particularly DenseNet121, seems promising and requires further validation.

Method: Proposes Spatial Causal Prediction (SCP) task paradigm and constructs SCP-Bench benchmark with 2,500 QA pairs across 1,181 videos spanning diverse viewpoints, scenes, and causal directions for systematic evaluation.

Result: Comprehensive experiments on 23 state-of-the-art models reveal substantial gaps between human and model performance, limited temporal extrapolation, and weak causal grounding in spatial reasoning.

Conclusion: Identifies key factors influencing performance and proposes perception-enhancement and reasoning-guided strategies to advance spatial causal intelligence in AI systems.

Abstract: Spatial reasoning, the ability to understand spatial relations, causality, and dynamic evolution, is central to human intelligence and essential for real-world applications such as autonomous driving and robotics. Existing studies, however, primarily assess models on visible spatio-temporal understanding, overlooking their ability to infer unseen past or future spatial states. In this work, we introduce Spatial Causal Prediction (SCP), a new task paradigm that challenges models to reason beyond observation and predict spatial causal outcomes. We further construct SCP-Bench, a benchmark comprising 2,500 QA pairs across 1,181 videos spanning diverse viewpoints, scenes, and causal directions, to support systematic evaluation. Through comprehensive experiments on {23} state-of-the-art models, we reveal substantial gaps between human and model performance, limited temporal extrapolation, and weak causal grounding. We further analyze key factors influencing performance and propose perception-enhancement and reasoning-guided strategies toward advancing spatial causal intelligence. The project page is https://guangstrip.github.io/SCP-Bench.

Method: 1) Training data synthesis: generates unaligned image pairs with ground-truth offsets from single input images, preserving structural information while varying textures and colors. 2) Network design: leverages cross-scale information and decouples color information from feature representations to improve estimation accuracy.

Result: Extensive experiments show improved generalization performance across various domains, confirming the effectiveness of both the training data synthesis method and the proposed network architecture.

Conclusion: The proposed approach addresses cross-modal generalization challenges in homography estimation through synthetic data generation and specialized network design, achieving robust performance across diverse domains.

Abstract: Supervised and unsupervised homography estimation methods depend on image pairs tailored to specific modalities to achieve high accuracy. However, their performance deteriorates substantially when applied to unseen modalities. To address this issue, we propose a training data synthesis method that generates unaligned image pairs with ground-truth offsets from a single input image. Our approach renders the image pairs with diverse textures and colors while preserving their structural information. These synthetic data empower the trained model to achieve greater robustness and improved generalization across various domains. Additionally, we design a network to fully leverage cross-scale information and decouple color information from feature representations, thus improving estimation accuracy. Extensive experiments show that our training data synthesis method improves generalization performance. The results also confirm the effectiveness of the proposed network.

Method: ProFound is pre-trained using several variants of self-supervised approaches on a diverse, multi-institutional collection of 5,000 patients (over 22,000 3D MRI volumes, 1.8M+ 2D slices). The model is then systematically evaluated across 11 downstream clinical tasks on over 3,000 independent patients.

Result: Finetuned ProFound consistently outperforms or remains competitive with state-of-the-art specialized models and existing medical vision foundation models trained/finetuned on the same data across tasks including prostate cancer detection, Gleason grading, lesion localization, gland volume estimation, and segmentation.

Conclusion: ProFound demonstrates strong performance as a domain-specialized vision foundation model for prostate MRI, showing promise for automating multiple clinical tasks with reduced reliance on large task-specific labeled datasets.

Abstract: Many diagnostic and therapeutic clinical tasks for prostate cancer increasingly rely on multi-parametric MRI. Automating these tasks is challenging because they necessitate expert interpretations, which are difficult to scale to capitalise on modern deep learning. Although modern automated systems achieve expert-level performance in isolated tasks, their general clinical utility remains limited by the requirement of large task-specific labelled datasets. In this paper, we present ProFound, a domain-specialised vision foundation model for volumetric prostate mpMRI. ProFound is pre-trained using several variants of self-supervised approaches on a diverse, multi-institutional collection of 5,000 patients, with a total of over 22,000 unique 3D MRI volumes (over 1,800,000 2D image slices). We conducted a systematic evaluation of ProFound across a broad spectrum of $11$ downstream clinical tasks on over 3,000 independent patients, including prostate cancer detection, Gleason grading, lesion localisation, gland volume estimation, zonal and surrounding structure segmentation. Experimental results demonstrate that finetuned ProFound consistently outperforms or remains competitive with state-of-the-art specialised models and existing medical vision foundation models trained/finetuned on the same data.

Method: Two-stage approach: (1) 3D preview synthesis using large multimodal model with prompt-and-reference template to create consistent dual-panel Minecraft-style previews; (2) Skin decoding using fine-tuned FLUX.2 model to translate previews into skin atlas images. Also introduces EvolveLoRA progressive LoRA curriculum for stable training.

Result: BLOCK generates pixel-perfect Minecraft skins from arbitrary character concepts, with all prompt templates and fine-tuned weights released for reproducible character-to-skin generation.

Conclusion: BLOCK provides an effective open-source solution for character-to-skin generation in Minecraft, demonstrating the potential of multimodal approaches for specialized content creation tasks.

Abstract: We present \textbf{BLOCK}, an open-source bi-stage character-to-skin pipeline that generates pixel-perfect Minecraft skins from arbitrary character concepts. BLOCK decomposes the problem into (i) a \textbf{3D preview synthesis stage} driven by a large multimodal model (MLLM) with a carefully designed prompt-and-reference template, producing a consistent dual-panel (front/back) oblique-view Minecraft-style preview; and (ii) a \textbf{skin decoding stage} based on a fine-tuned FLUX.2 model that translates the preview into a skin atlas image. We further propose \textbf{EvolveLoRA}, a progressive LoRA curriculum (text-to-image $\rightarrow$ image-to-image $\rightarrow$ preview-to-skin) that initializes each phase from the previous adapter to improve stability and efficiency. BLOCK is released with all prompt templates and fine-tuned weights to support reproducible character-to-skin generation.

Method: Proposes UniRain framework with: 1) RAG-based dataset distillation pipeline to select high-quality training samples from public datasets, 2) Multi-objective reweighted optimization strategy, and 3) Asymmetric mixture-of-experts (MoE) architecture for consistent performance across diverse scenes.

Result: Extensive experiments show UniRain performs favorably against state-of-the-art models on proposed benchmarks and multiple public datasets, demonstrating robust performance across diverse rainy conditions.

Conclusion: UniRain provides an effective unified solution for image deraining that generalizes well across different rain degradation types and lighting conditions, addressing limitations of specialized methods.

Abstract: Despite significant progress has been made in image deraining, we note that most existing methods are often developed for only specific types of rain degradation and fail to generalize across diverse real-world rainy scenes. How to effectively model different rain degradations within a universal framework is important for real-world image deraining. In this paper, we propose UniRain, an effective unified image deraining framework capable of restoring images degraded by rain streak and raindrop under both daytime and nighttime conditions. To better enhance unified model generalization, we construct an intelligent retrieval augmented generation (RAG)-based dataset distillation pipeline that selects high-quality training samples from all public deraining datasets for better mixed training. Furthermore, we incorporate a simple yet effective multi-objective reweighted optimization strategy into the asymmetric mixture-of-experts (MoE) architecture to facilitate consistent performance and improve robustness across diverse scenes. Extensive experiments show that our framework performs favorably against the state-of-the-art models on our proposed benchmarks and multiple public datasets.

Method: Curates extensive synchronized image-event collection for cross-modal alignment. Proposes structure-aware distillation loss that grounds higher-quality image-event correspondences by extending alignment to semantic structures from visual foundation models, optimizing dense event representations.

Result: Extensive experiments demonstrate significant improvements in downstream benchmarks, surpassing traditional methods and existing pretraining techniques with enhanced generalization, superior data efficiency, and elevated transferability.

Conclusion: The approach represents a breakthrough in event representation learning through self-supervised pretraining that effectively bridges the sparsity and granularity mismatch between image and event domains.

Abstract: Learning versatile, fine-grained representations from irregular event streams is pivotal yet nontrivial, primarily due to the heavy annotation that hinders scalability in dataset size, semantic richness, and application scope. To mitigate this dilemma, we launch a novel self-supervised pretraining method that distills visual foundation models (VFMs) to push the boundaries of event representation at scale. Specifically, we curate an extensive synchronized image-event collection to amplify cross-modal alignment. Nevertheless, due to inherent mismatches in sparsity and granularity between image-event domains, existing distillation paradigms are prone to semantic collapse in event representations, particularly at high resolutions. To bridge this gap, we propose to extend the alignment objective to semantic structures provided off-the-shelf by VFMs, indicating a broader receptive field and stronger supervision. The key ingredient of our method is a structure-aware distillation loss that grounds higher-quality image-event correspondences for alignment, optimizing dense event representations. Extensive experiments demonstrate that our approach takes a great leap in downstream benchmarks, significantly surpassing traditional methods and existing pretraining techniques. This breakthrough manifests in enhanced generalization, superior data efficiency and elevated transferability.

Method: GeoSeg couples MLLM reasoning with precise localization through bias-aware coordinate refinement to correct systematic grounding shifts and a dual-route prompting mechanism to fuse semantic intent with fine-grained spatial cues.

Result: GeoSeg consistently outperforms all baselines on the GeoSeg-Bench benchmark of 810 image-query pairs with hierarchical difficulty levels, with ablations confirming component effectiveness.

Conclusion: GeoSeg provides a zero-shot, training-free solution for reasoning-driven remote sensing segmentation that bypasses supervision bottlenecks while addressing domain-specific challenges.

Abstract: Recent advances in MLLMs are reframing segmentation from fixed-category prediction to instruction-grounded localization. While reasoning based segmentation has progressed rapidly in natural scenes, remote sensing lacks a generalizable solution due to the prohibitive cost of reasoning-oriented data and domain-specific challenges like overhead viewpoints. We present GeoSeg, a zero-shot, training-free framework that bypasses the supervision bottleneck for reasoning-driven remote sensing segmentation. GeoSeg couples MLLM reasoning with precise localization via: (i) bias-aware coordinate refinement to correct systematic grounding shifts and (ii) a dual-route prompting mechanism to fuse semantic intent with fine-grained spatial cues. We also introduce GeoSeg-Bench, a diagnostic benchmark of 810 image–query pairs with hierarchical difficulty levels. Experiments show that GeoSeg consistently outperforms all baselines, with extensive ablations confirming the effectiveness and necessity of each component.

Method: Created RIVER Bench with three novel task frameworks: Retrospective Memory (recalling past video content), Live-Perception (understanding current video), and Proactive Anticipation (predicting future events). Used diverse video sources and lengths with precise real-time interactive format annotations.

Result: Offline models perform well on single QA tasks but struggle with real-time processing. The benchmark reveals deficiencies in long-term memory and future perception in existing models. Proposed a general improvement method for better real-time interaction.

Conclusion: RIVER Bench advances real-time interactive video understanding research and will inspire future work in this emerging field. The dataset and code are publicly available.

Abstract: The rapid advancement of multimodal large language models has demonstrated impressive capabilities, yet nearly all operate in an offline paradigm, hindering real-time interactivity. Addressing this gap, we introduce the Real-tIme Video intERaction Bench (RIVER Bench), designed for evaluating online video comprehension. RIVER Bench introduces a novel framework comprising Retrospective Memory, Live-Perception, and Proactive Anticipation tasks, closely mimicking interactive dialogues rather than responding to entire videos at once. We conducted detailed annotations using videos from diverse sources and varying lengths, and precisely defined the real-time interactive format. Evaluations across various model categories reveal that while offline models perform well in single question-answering tasks, they struggle with real-time processing. Addressing the limitations of existing models in online video interaction, especially their deficiencies in long-term memory and future perception, we proposed a general improvement method that enables models to interact with users more flexibly in real time. We believe this work will significantly advance the development of real-time interactive video understanding models and inspire future research in this emerging field. Datasets and code are publicly available at https://github.com/OpenGVLab/RIVER.

Method: Developed a representation-level diagnostic framework evaluating six models across four representational regimes: vision-language models (CLIP-B/32, CLIP-L/14, LLaVA-1.5-7B), pure vision classification (ViT), general object detection (YOLOv8), and face detection (RetinaFace) using face pareidolia images under a unified protocol.

Result: VLMs exhibit semantic overactivation, pulling ambiguous non-human regions toward Human concepts (LLaVA-1.5-7B strongest). ViT shows uncertainty-as-abstention strategy (diffuse but unbiased). Detection models achieve low bias through conservative priors suppressing pareidolia responses. Uncertainty and bias are decoupled across architectures.

Conclusion: Behavior under ambiguity is governed more by representational choices than score thresholds. Pareidolia provides a compact diagnostic and source of ambiguity-aware hard negatives for probing and improving semantic robustness of vision-language systems.

Abstract: When visual evidence is ambiguous, vision models must decide whether to interpret face-like patterns as meaningful. Face pareidolia, the perception of faces in non-face objects, provides a controlled probe of this behavior. We introduce a representation-level diagnostic framework that analyzes detection, localization, uncertainty, and bias across class, difficulty, and emotion in face pareidolia images. Under a unified protocol, we evaluate six models spanning four representational regimes: vision-language models (VLMs; CLIP-B/32, CLIP-L/14, LLaVA-1.5-7B), pure vision classification (ViT), general object detection (YOLOv8), and face detection (RetinaFace). Our analysis reveals three mechanisms of interpretation under ambiguity. VLMs exhibit semantic overactivation, systematically pulling ambiguous non-human regions toward the Human concept, with LLaVA-1.5-7B producing the strongest and most confident over-calls, especially for negative emotions. ViT instead follows an uncertainty-as-abstention strategy, remaining diffuse yet largely unbiased. Detection-based models achieve low bias through conservative priors that suppress pareidolia responses even when localization is controlled. These results show that behavior under ambiguity is governed more by representational choices than score thresholds, and that uncertainty and bias are decoupled: low uncertainty can signal either safe suppression, as in detectors, or extreme over-interpretation, as in VLMs. Pareidolia therefore provides a compact diagnostic and a source of ambiguity-aware hard negatives for probing and improving the semantic robustness of vision-language systems. Code will be released upon publication.

Method: Two-stage framework: 1) Dual-arm Patch Embedding Network learns semantically structured, class-discriminative embeddings from expert-annotated patches using contrastive learning. 2) Ensemble of independent embedding spaces extrapolates labels to unannotated regions based on spatial and semantic proximity, with an abstention mechanism for reliability vs. coverage trade-off.

Result: Achieves up to 0.9886 accuracy in separating healthy and diseased patches. Generated annotations substantially improve downstream DR classification, with F1-score increases for diseased class and performance gains up to 0.545 in AUPRC. Qualitative analysis confirms focus on clinically relevant lesion patterns, validated by ophthalmologists.

Conclusion: SAFE effectively expands sparse annotations for diabetic retinopathy detection by unifying weak supervision, contrastive learning, and patch-wise embedding inference, demonstrating improved performance in downstream classification tasks while preserving fine-grained lesion details.

Abstract: Diabetic Retinopathy (DR) requires timely screening to prevent irreversible vision loss. However, its early detection remains a significant challenge since often the subtle pathological manifestations (lesions) get overlooked due to insufficient annotation. Existing literature primarily focuses on image-level supervision, weakly-supervised localization, and clustering-based representation learning, which fail to systematically annotate unlabeled lesion region(s) for refining the dataset. Expert-driven lesion annotation is labor-intensive and often incomplete, limiting the performance of deep learning models. We introduce Similarity-based Annotation via Feature-space Ensemble (SAFE), a two-stage framework that unifies weak supervision, contrastive learning, and patch-wise embedding inference, to systematically expand sparse annotations in the pathology. SAFE preserves fine-grained details of the lesion(s) under partial clinical supervision. In the first stage, a dual-arm Patch Embedding Network learns semantically structured, class-discriminative embeddings from expert annotated patches. Next, an ensemble of independent embedding spaces extrapolates labels to the unannotated regions based on spatial and semantic proximity. An abstention mechanism ensures trade-off between highly reliable annotation and noisy coverage. Experimental results demonstrate reliable separation of healthy and diseased patches, achieving upto 0.9886 accuracy. The annotation generated from SAFE substantially improves downstream tasks such as DR classification, demonstrating a substantial increase in F1-score of the diseased class and a performance gain as high as 0.545 in Area Under the Precision-Recall Curve (AUPRC). Qualitative analysis, with explainability, confirms that SAFE focuses on clinically relevant lesion patterns; and is further validated by ophthalmologists.

Method: Proposes DPAD (Discriminative Perception) that compels the model to generate descriptive captions of referred objects, then uses these captions to explicitly discriminate by contrasting semantic relevance to the target versus wider context. This optimization forces focus on unique target attributes.

Result: Substantial performance gains: cIoU on ReasonSeg increased by 3.09%, reasoning chain length decreased by approximately 42%. The descriptive captions also serve as interpretable rationales aligned with segmentation.

Conclusion: Complementing RL objectives with Discriminative Perception improves reasoning segmentation by ensuring reasoning chains remain anchored to target regions, leading to more efficient and effective multimodal understanding.

Abstract: Reasoning segmentation increasingly employs reinforcement learning to generate explanatory reasoning chains that guide Multimodal Large Language Models. While these geometric rewards are primarily confined to guiding the final localization, they are incapable of discriminating whether the reasoning process remains anchored on the referred region or strays into irrelevant context. Lacking this discriminative guidance, the model’s reasoning often devolves into unfocused and verbose chains that ultimately fail to disambiguate and perceive the target in complex scenes. This suggests a need to complement the RL objective with Discriminative Perception, an ability to actively distinguish a target from its context. To realize this, we propose DPAD to compel the model to generate a descriptive caption of the referred object, which is then used to explicitly discriminate by contrasting the caption’s semantic relevance to the referred object against the wider context. By optimizing for this discriminative capability, the model is forced to focus on the unique attributes of the target, leading to a more converged and efficient reasoning chain. The descriptive caption also serves as an interpretability rationale that aligns with the segmentation. Experiments on the benchmarks confirm the validity of our approach, delivering substantial performance gains, with the cIoU on ReasonSeg increasing by 3.09% and the reasoning chain length decreasing by approximately 42%. Code is available at https://github.com/mrazhou/DPAD

Method: Two key innovations: 1) Diagnostic diversity-based data sampling strategy for better data efficiency, 2) Diagnostic Token-weighted Policy Optimization (DiTPO) that uses diagnostic F1 score as reward signal and weights tokens by clinical importance.

Result: Achieves state-of-the-art performance on MIMIC-CXR, IU-Xray, and CheXpert Plus datasets; attains F1 score of 0.516 on MIMIC-CXR using only 20% of RL training samples.

Conclusion: Demonstrates that RL can be effective for radiology report generation with proper data sampling and token-weighted optimization, achieving high clinical accuracy with fewer samples.

Abstract: Radiologists highly desire fully automated AI for radiology report generation (R2G), yet existing approaches fall short in clinical utility. Reinforcement learning (RL) holds potential to address these shortcomings, but its adoption in this task remains underexplored. In this paper, we revisit RL in terms of data efficiency and optimization effectiveness for R2G tasks. First, we explore the impact of data quantity and quality on the performance of RL in medical contexts, revealing that data quality plays a more critical role than quantity. To this end, we propose a diagnostic diversity-based data sampling strategy that enables comparable performance with fewer samples. Second, we observe that the majority of tokens in radiology reports are template-like and diagnostically uninformative, whereas the low frequency of clinically critical tokens heightens the risk of being overlooked during optimization. To tackle this, we introduce Diagnostic Token-weighted Policy Optimization (DiTPO), which directly optimizes for clinical accuracy by using a diagnostic F1 score as the reward signal. Unlike standard RL approaches that treat all tokens equally, DiTPO explicitly models the varying importance of different tokens through rule- or gradient-based mechanisms to prioritize clinically relevant content. Extensive experiments on the MIMIC-CXR, IU-Xray, and CheXpert Plus datasets demonstrate that our framework achieves state-of-the-art (SOTA) performance while requiring substantially fewer training samples in RL. Notably, on MIMIC-CXR, our framework attains an F1 score of 0.516 using only 20% of the RL training samples.

Method: Volumetric Directional Diffusion (VDD) anchors the generative trajectory to a deterministic consensus prior instead of starting from isotropic Gaussian noise. It restricts the search space to predict 3D boundary residual fields, exploring geometric variations while maintaining topological integrity.

Result: VDD achieves state-of-the-art uncertainty quantification on three multi-rater datasets (LIDC-IDRI, KiTS21, ISBI 2015), significantly improving GED and CI metrics while remaining competitive in segmentation accuracy against deterministic upper bounds.

Conclusion: VDD provides clinicians with anatomically coherent uncertainty maps that enable safer decision-making for downstream clinical tasks like radiotherapy planning and surgical margin assessment, resolving the fidelity-diversity trade-off in lesion segmentation.

Abstract: Equivocal 3D lesion segmentation exhibits high inter-observer variability. Conventional deterministic models ignore this aleatoric uncertainty, producing over-confident masks that obscure clinical risks. Conversely, while generative methods (e.g., standard diffusion) capture sample diversity, recovering complex topology from pure noise frequently leads to severe structural fractures and out-of-distribution anatomical hallucinations. To resolve this fidelity-diversity trade-off, we propose Volumetric Directional Diffusion (VDD). Unlike standard diffusion models that denoise isotropic Gaussian noise, VDD mathematically anchors the generative trajectory to a deterministic consensus prior. By restricting the generative search space to iteratively predict a 3D boundary residual field, VDD accurately explores the fine-grained geometric variations inherent in expert disagreements without risking topological collapse. Extensive validation on three multi-rater datasets (LIDC-IDRI, KiTS21, and ISBI 2015) demonstrates that VDD achieves state-of-the-art uncertainty quantification (significantly improving GED and CI) while remaining highly competitive in segmentation accuracy against deterministic upper bounds. Ultimately, VDD provides clinicians with anatomically coherent uncertainty maps, enabling safer decision-making and mitigating risks in downstream tasks (e.g., radiotherapy planning or surgical margin assessment).

Method: Proposes DQE-CIR with two key components: 1) Learnable attribute weighting to emphasize distinctive visual features conditioned on modification text for better language-vision alignment, and 2) Target relative negative sampling that constructs similarity distributions and selects informative negatives from a mid-zone region, excluding both easy negatives and ambiguous false negatives.

Result: The method enables more reliable retrieval for fine-grained attribute changes by improving query discriminativeness and reducing confusion caused by semantically similar but irrelevant candidates.

Conclusion: DQE-CIR addresses fundamental limitations in CIR by learning distinctive query embeddings through explicit modeling of target relative relevance, leading to better performance on fine-grained attribute modifications.

Abstract: Composed image retrieval (CIR) addresses the task of retrieving a target image by jointly interpreting a reference image and a modification text that specifies the intended change. Most existing methods are still built upon contrastive learning frameworks that treat the ground truth image as the only positive instance and all remaining images as negatives. This strategy inevitably introduces relevance suppression, where semantically related yet valid images are incorrectly pushed away, and semantic confusion, where different modification intents collapse into overlapping regions of the embedding space. As a result, the learned query representations often lack discriminativeness, particularly at fine-grained attribute modifications. To overcome these limitations, we propose distinctive query embeddings through learnable attribute weights and target relative negative sampling (DQE-CIR), a method designed to learn distinctive query embeddings by explicitly modeling target relative relevance during training. DQE-CIR incorporates learnable attribute weighting to emphasize distinctive visual features conditioned on the modification text, enabling more precise feature alignment between language and vision. Furthermore, we introduce target relative negative sampling, which constructs a target relative similarity distribution and selects informative negatives from a mid-zone region that excludes both easy negatives and ambiguous false negatives. This strategy enables more reliable retrieval for fine-grained attribute changes by improving query discriminativeness and reducing confusion caused by semantically similar but irrelevant candidates.

Method: Presents a curated dataset with georeferenced AUV imagery from five benthic sites revisited over up to six years, plus a novel ground-truthing method estimating 3D seafloor image footprints to link camera views with overlapping visual content.

Result: Recall@K for eight state-of-the-art VPR methods is significantly lower on this underwater dataset than established benchmarks; footprint-based ground truth reveals traditional distance-threshold methods overestimate performance on rugged terrain.

Conclusion: The dataset, ground-truthing method, and VPR benchmark provide foundation for advancing long-term visual localization in dynamic benthic environments.

Abstract: Long-term visual localization has the potential to reduce cost and improve mapping quality in optical benthic monitoring with autonomous underwater vehicles (AUVs). Despite this potential, long-term visual localization in benthic environments remains understudied, primarily due to the lack of curated datasets for benchmarking. Moreover, limited georeferencing accuracy and image footprints necessitate precise geometric information for accurate ground-truthing. In this work, we address these gaps by presenting a curated dataset for long-term visual localization in benthic environments and a novel method to ground-truth visual localization results for near-nadir underwater imagery. Our dataset comprises georeferenced AUV imagery from five benthic reference sites, revisited over periods up to six years, and includes raw and color-corrected stereo imagery, camera calibrations, and sub-decimeter registered camera poses. To our knowledge, this is the first curated underwater dataset for long-term visual localization spanning multiple sites and photic-zone habitats. Our ground-truthing method estimates 3D seafloor image footprints and links camera views with overlapping footprints, ensuring that ground-truth links reflect shared visual content. Building on this dataset and ground truth, we benchmark eight state-of-the-art visual place recognition (VPR) methods and find that Recall@K is significantly lower on our dataset than on established terrestrial and underwater benchmarks. Finally, we compare our footprint-based ground truth to a traditional location-based ground truth and show that distance-threshold ground-truthing can overestimate VPR Recall@K at sites with rugged terrain and altitude variations. Together, the curated dataset, ground-truthing method, and VPR benchmark provide a stepping stone for advancing long-term visual localization in dynamic benthic environments.

Method: Combines a generative model with tumor-infiltration maps that can be propagated through time using a biophysical growth model. This enables fine-grained control over tumor shape and growth while preserving patient anatomy, synthesizing consistent tumor growth trajectories in real patient spaces.

Result: The framework generates temporally coherent sequences with realistic changes in tumor appearance and surrounding tissue response. In longitudinal extrapolation, achieves 75% Dice overlap with the biophysical model while maintaining PSNR of 25 in surrounding tissue.

Conclusion: Integrating mechanistic tumor growth priors with modern generative modeling provides a practical tool for patient-specific progression visualization and generating controlled synthetic data to support neuro-oncology workflows.

Abstract: Glioblastoma exhibits diverse, infiltrative, and patient-specific growth patterns that are only partially visible on routine MRI, making it difficult to reliably assess true tumor extent and personalize treatment planning and follow-up. We present a biophysically-conditioned generative framework that synthesizes biologically realistic 3D brain MRI volumes from estimated, spatially continuous tumor-concentration fields. Our approach combines a generative model with tumor-infiltration maps that can be propagated through time using a biophysical growth model, enabling fine-grained control over tumor shape and growth while preserving patient anatomy. This enables us to synthesize consistent tumor growth trajectories directly in the space of real patients, providing interpretable, controllable estimation of tumor infiltration and progression beyond what is explicitly observed in imaging. We evaluate the framework on longitudinal glioblastoma cases and demonstrate that it can generate temporally coherent sequences with realistic changes in tumor appearance and surrounding tissue response. These results suggest that integrating mechanistic tumor growth priors with modern generative modeling can provide a practical tool for patient-specific progression visualization and for generating controlled synthetic data to support downstream neuro-oncology workflows. In longitudinal extrapolation, we achieve a consistent 75% Dice overlap with the biophysical model while maintaining a constant PSNR of 25 in the surrounding tissue. Our code is available at: https://github.com/valentin-biller/lgm.git

Method: Proposes Pointer-CAD framework that decomposes CAD generation into steps, conditioning each step on text description and previous B-rep. Uses pointer mechanism to select geometric entities, reducing quantization errors. Built 575K CAD dataset with expert-level natural language descriptions.

Result: Effectively supports generation of complex geometric structures, reduces segmentation error to extremely low level, significantly improves over prior command sequence methods, and mitigates topological inaccuracies from quantization error.

Conclusion: Pointer-CAD addresses limitations of command sequence-based CAD generation by incorporating geometric information and entity selection, enabling more practical and accurate CAD model generation using LLMs.

Abstract: Constructing computer-aided design (CAD) models is labor-intensive but essential for engineering and manufacturing. Recent advances in Large Language Models (LLMs) have inspired the LLM-based CAD generation by representing CAD as command sequences. But these methods struggle in practical scenarios because command sequence representation does not support entity selection (e.g. faces or edges), limiting its ability to support complex editing operations such as chamfer or fillet. Further, the discretization of a continuous variable during sketch and extrude operations may result in topological errors. To address these limitations, we present Pointer-CAD, a novel LLM-based CAD generation framework that leverages a pointer-based command sequence representation to explicitly incorporate the geometric information of B-rep models into sequential modeling. In particular, Pointer-CAD decomposes CAD model generation into steps, conditioning the generation of each subsequent step on both the textual description and the B-rep generated from previous steps. Whenever an operation requires the selection of a specific geometric entity, the LLM predicts a Pointer that selects the most feature-consistent candidate from the available set. Such a selection operation also reduces the quantization error in the command sequence-based representation. To support the training of Pointer-CAD, we develop a data annotation pipeline that produces expert-level natural language descriptions and apply it to build a dataset of approximately 575K CAD models. Extensive experimental results demonstrate that Pointer-CAD effectively supports the generation of complex geometric structures and reduces segmentation error to an extremely low level, achieving a significant improvement over prior command sequence methods, thereby significantly mitigating the topological inaccuracies introduced by quantization error.

Method: Analyzed 303 colorectal cancer specimens with CD103/CD8 immunostaining, generating 185,432 annotated cell images. Evaluated 8 task-specific architectures trained from scratch at multiple data scales, and 3 foundation models via linear probing and fine-tuning after resizing inputs to 224x224. Assessed robustness to blur using Gaussian perturbations.

Result: Task-specific models improved consistently with increasing data scale, while foundation models saturated at moderate sample sizes. A Vision Transformer optimized for small patches (CustomViT) achieved highest accuracy, outperforming all foundation models with lower inference cost. Blur robustness was comparable across architectures.

Conclusion: For cell-level classification under extreme spatial constraints, task-specific architectures are more effective and efficient than foundation models once sufficient training data are available. Higher clean accuracy doesn’t imply superior robustness, and large pre-trained models offer limited benefit in small-patch regime.

Abstract: Background and objective: Cell-level pathological image analysis requires working with extremely small image patches (40x40 pixels), far below standard ImageNet resolutions. It remains unclear whether modern deep learning architectures and foundation models can learn robust and scalable representations under this constraint. We systematically evaluated architectural suitability and data-scale effects for small-patch cell classification. Methods: We analyzed 303 colorectal cancer specimens with CD103/CD8 immunostaining, generating 185,432 annotated cell images. Eight task-specific architectures were trained from scratch at multiple data scales (FlagLimit: 256–16,384 samples per class), and three foundation models were evaluated via linear probing and fine-tuning after resizing inputs to 224x224 pixels. Robustness to blur was assessed using pre- and post-resize Gaussian perturbations. Results: Task-specific models improved consistently with increasing data scale, whereas foundation models saturated at moderate sample sizes. A Vision Transformer optimized for small patches (CustomViT) achieved the highest accuracy, outperforming all foundation models with substantially lower inference cost. Blur robustness was comparable across architectures, with no qualitative advantage observed for foundation models. Conclusion: For cell-level classification under extreme spatial constraints, task-specific architectures are more effective and efficient than foundation models once sufficient training data are available. Higher clean accuracy does not imply superior robustness, and large pre-trained models offer limited benefit in the small-patch regime.

Method: Uses transformer-based model with identity-conditioned queries, multi-view spatial refinement, causal temporal attention, and supports both keypoints and parametric body representations. Includes auto-labeling system with teacher-student schema for uncertainty-aware semi-supervised training on millions of unlabeled frames.

Result: Outperforms two state-of-the-art methods by 12.2% and 19.4% in accuracy on EgoBody3M benchmark with 0.8 ms GPU latency, reduces temporal jitter by 22.2% and 51.7%, and auto-labeling improves wrist MPJPE by 13.1%.

Conclusion: EgoPoseFormer v2 effectively addresses egocentric pose estimation challenges through transformer architecture and scalable auto-labeling system, achieving state-of-the-art performance with real-time inference.

Abstract: Egocentric human motion estimation is essential for AR/VR experiences, yet remains challenging due to limited body coverage from the egocentric viewpoint, frequent occlusions, and scarce labeled data. We present EgoPoseFormer v2, a method that addresses these challenges through two key contributions: (1) a transformer-based model for temporally consistent and spatially grounded body pose estimation, and (2) an auto-labeling system that enables the use of large unlabeled datasets for training. Our model is fully differentiable, introduces identity-conditioned queries, multi-view spatial refinement, causal temporal attention, and supports both keypoints and parametric body representations under a constant compute budget. The auto-labeling system scales learning to tens of millions of unlabeled frames via uncertainty-aware semi-supervised training. The system follows a teacher-student schema to generate pseudo-labels and guide training with uncertainty distillation, enabling the model to generalize to different environments. On the EgoBody3M benchmark, with a 0.8 ms latency on GPU, our model outperforms two state-of-the-art methods by 12.2% and 19.4% in accuracy, and reduces temporal jitter by 22.2% and 51.7%. Furthermore, our auto-labeling system further improves the wrist MPJPE by 13.1%.

Method: Introduces TaxonRL, a reinforcement learning approach using Group Relative Policy Optimization with intermediate rewards that decomposes the reasoning process into hierarchical taxonomic predictions (species-level, genus-level, and family-level features) before making final classifications.

Result: Achieves 91.7% average accuracy on Birds-to-Words dataset, exceeding human performance (77.3%), while generating interpretable reasoning traces. Shows strong cross-domain generalization with substantial gains in primate and marine species verification.

Conclusion: Enforcing structured, hierarchical reasoning provides a powerful and transferable framework for fine-grained visual discrimination, offering both improved accuracy and transparent, verifiable decision-making processes.

Abstract: Traditional vision-language models struggle with contrastive fine-grained taxonomic reasoning, particularly when distinguishing between visually similar species within the same genus or family. We introduce TaxonRL, a reinforcement learning approach using Group Relative Policy Optimization with intermediate rewards that decomposes the reasoning process into hierarchical taxonomic predictions. Our method incentivizes models to explicitly reason about species-level, genus-level, and family-level features before making final classifications. This structured approach is designed not only to boost accuracy but also to yield a transparent, verifiable decision-making process. On the challenging Birds-to-Words dataset, TaxonRL achieves 91.7% average accuracy, exceeding human performance (77.3%) while generating interpretable reasoning traces. We demonstrate strong cross-domain generalization, showing substantial gains in primate and marine species verification. Our results establish that enforcing structured, hierarchical reasoning provides a powerful and transferable framework for fine-grained visual discrimination.

Method: Proposes a level-aware vision-language framework built on CLIP embeddings that aggregates rotational views into angle-invariant representations and conditions visual features on lightweight text priors encoding viewpoint level. Uses a single multi-task model to jointly predict plant age and leaf count.

Result: On the GroMo25 benchmark, reduces mean age MAE from 7.74 to 3.91 and mean leaf-count MAE from 5.52 to 3.08 compared to baseline, representing improvements of 49.5% and 44.2% respectively. Also improves robustness to missing views.

Conclusion: The unified vision-language framework simplifies the pipeline by replacing conventional dual-model setups while improving prediction accuracy and robustness to incomplete inputs through angle-invariant representations and text conditioning.

Abstract: Modeling plant growth dynamics plays a central role in modern agricultural research. However, learning robust predictors from multi-view plant imagery remains challenging due to strong viewpoint redundancy and viewpoint-dependent appearance changes. We propose a level-aware vision language framework that jointly predicts plant age and leaf count using a single multi-task model built on CLIP embeddings. Our method aggregates rotational views into angle-invariant representations and conditions visual features on lightweight text priors encoding viewpoint level for stable prediction under incomplete or unordered inputs. On the GroMo25 benchmark, our approach reduces mean age MAE from 7.74 to 3.91 and mean leaf-count MAE from 5.52 to 3.08 compared to the GroMo baseline, corresponding to improvements of 49.5% and 44.2%, respectively. The unified formulation simplifies the pipeline by replacing the conventional dual-model setup while improving robustness to missing views. The models and code is available at: https://github.com/SimonWarmers/CLIP-MVP

Method: Uses eye-tracking data as side channel: gaze fixation captures visual stability/quality, pupil response captures arousal-linked novelty. Dual-Criterion Frame Curator gates frames by gaze quality first, then ranks survivors by pupil-derived novelty.

Result: On Visual Experience Dataset, curated frames at 10% budget match full stream classification performance. Pupil ranking improves activity recognition, gaze-only selection dominates scene recognition. No model inference needed, operates at capture time.

Conclusion: Gaze and pupil signals serve complementary roles for frame curation. Method enables efficient always-on egocentric data curation without computational overhead.

Abstract: Always-on egocentric cameras are increasingly used as demonstrations for embodied robotics, imitation learning, and assistive AR, but the resulting video streams are dominated by redundant and low-quality frames. Under the storage and battery constraints of wearable devices, choosing which frames to keep is as important as how to learn from them. We observe that modern eye-tracking headsets provide a continuous, training-free side channel that decomposes into two complementary axes: gaze fixation captures visual stability (quality), while pupil response captures arousal-linked moments (novelty). We operationalize this insight as a Dual-Criterion Frame Curator that first gates frames by gaze quality and then ranks the survivors by pupil-derived novelty. On the Visual Experience Dataset (VEDB), curated frames at 10% budget match the classification performance of the full stream, and naive signal fusion consistently destroys both contributions. The benefit is task-dependent: pupil ranking improves activity recognition, while gaze-only selection already dominates for scene recognition, confirming that the two signals serve genuinely different roles. Our method requires no model inference and operates at capture time, offering a path toward efficient, always-on egocentric data curation.

Method: Proposes ABS-REF view for modular feature extraction, introduces High-dimensional Positional Encoding (HPE) module, replaces local MLP operations with non-local MLPs for efficiency, and develops HPENet architecture following this paradigm.

Result: HPENet outperforms PointNeXt on multiple datasets (ScanObjectNN, S3DIS, ScanNet, ShapeNetPart) with significant FLOPs reduction (21.5-50% of baseline FLOPs) while achieving better accuracy metrics.

Conclusion: The ABS-REF view provides a clearer understanding of point cloud processing, HPE effectively utilizes positional information, and HPENet achieves strong efficiency-effectiveness balance for MLP-based point cloud models.

Abstract: Multi-Layer Perceptron (MLP) models are the foundation of contemporary point cloud processing. However, their complex network architectures obscure the source of their strength and limit the application of these models. In this article, we develop a two-stage abstraction and refinement (ABS-REF) view for modular feature extraction in point cloud processing. This view elucidates that whereas the early models focused on ABS stages, the more recent techniques devise sophisticated REF stages to attain performance advantages. Then, we propose a High-dimensional Positional Encoding (HPE) module to explicitly utilize intrinsic positional information, extending the ``positional encoding’’ concept from Transformer literature. HPE can be readily deployed in MLP-based architectures and is compatible with transformer-based methods. Within our ABS-REF view, we rethink local aggregation in MLP-based methods and propose replacing time-consuming local MLP operations, which are used to capture local relationships among neighbors. Instead, we use non-local MLPs for efficient non-local information updates, combined with the proposed HPE for effective local information representation. We leverage our modules to develop HPENets, a suite of MLP networks that follow the ABS-REF paradigm, incorporating a scalable HPE-based REF stage. Extensive experiments on seven public datasets across four different tasks show that HPENets deliver a strong balance between efficiency and effectiveness. Notably, HPENet surpasses PointNeXt, a strong MLP-based counterpart, by 1.1% mAcc, 4.0% mIoU, 1.8% mIoU, and 0.2% Cls. mIoU, with only 50.0%, 21.5%, 23.1%, 44.4% of FLOPs on ScanObjectNN, S3DIS, ScanNet, and ShapeNetPart, respectively. Source code is available at https://github.com/zouyanmei/HPENet_v2.git.

Method: Proposed a controlled framework using disentangled representation learning to decompose brain MRI into: 1) anatomy-focused representations that suppress acquisition influence, and 2) contrast embeddings that capture acquisition-dependent characteristics. Trained predictive models for age, sex, and race on full images, anatomical representations, and contrast-only embeddings to quantify relative contributions.

Result: Across three datasets and multiple MRI sequences, demographic predictability is primarily rooted in anatomical variation: anatomy-focused representations largely preserve performance of models trained on raw images. Contrast-only embeddings retain a weaker but systematic signal that is dataset-specific and doesn’t generalize across sites.

Conclusion: Effective bias mitigation must explicitly account for distinct anatomical and acquisition-dependent origins of demographic signal. Any bias reduction needs to generalize robustly across domains by addressing both sources separately.

Abstract: Demographic attributes such as age, sex, and race can be predicted from medical images, raising concerns about bias in clinical AI systems. In brain MRI, this signal may arise from anatomical variation, acquisition-dependent contrast differences, or both, yet these sources remain entangled in conventional analyses. Without disentangling them, mitigation strategies risk failing to address the underlying causes. We propose a controlled framework based on disentangled representation learning, decomposing brain MRI into anatomy-focused representations that suppress acquisition influence and contrast embeddings that capture acquisition-dependent characteristics. Training predictive models for age, sex, and race on full images, anatomical representations, and contrast-only embeddings allows us to quantify the relative contributions of structure and acquisition to the demographic signal. Across three datasets and multiple MRI sequences, we find that demographic predictability is primarily rooted in anatomical variation: anatomy-focused representations largely preserve the performance of models trained on raw images. Contrast-only embeddings retain a weaker but systematic signal that is dataset-specific and does not generalise across sites. These findings suggest that effective mitigation must explicitly account for the distinct anatomical and acquisition-dependent origins of the demographic signal, ensuring that any bias reduction generalizes robustly across domains.

Method: Proposes Any2Any framework that projects heterogeneous inputs into geometrically aligned latent space using shared backbone. Uses anchored latent regression and lightweight target-specific residual adapters to correct systematic latent mismatches without increasing inference complexity.

Result: Experiments across 14 translation tasks show Any2Any consistently outperforms pairwise translation methods and exhibits strong zero-shot generalization to unseen modality pairs. Introduces RST-1M dataset with paired observations across five sensing modalities.

Conclusion: Any2Any provides an efficient unified framework for any-to-any cross-modal translation in remote sensing, overcoming limitations of pairwise approaches through shared latent representation learning.

Abstract: Multi-modal remote sensing imagery provides complementary observations of the same geographic scene, yet such observations are frequently incomplete in practice. Existing cross-modal translation methods treat each modality pair as an independent task, resulting in quadratic complexity and limited generalization to unseen modality combinations. We formulate Any-to-Any translation as inference over a shared latent representation of the scene, where different modalities correspond to partial observations of the same underlying semantics. Based on this formulation, we propose Any2Any, a unified latent diffusion framework that projects heterogeneous inputs into a geometrically aligned latent space. Such structure performs anchored latent regression with a shared backbone, decoupling modality-specific representation learning from semantic mapping. Moreover, lightweight target-specific residual adapters are used to correct systematic latent mismatches without increasing inference complexity. To support learning under sparse but connected supervision, we introduce RST-1M, the first million-scale remote sensing dataset with paired observations across five sensing modalities, providing supervision anchors for any-to-any translation. Experiments across 14 translation tasks show that Any2Any consistently outperforms pairwise translation methods and exhibits strong zero-shot generalization to unseen modality pairs. Code and models will be available at https://github.com/MiliLab/Any2Any.

Method: PlaneCycle operates by cyclically distributing spatial aggregation across orthogonal HW, DW, and DH planes throughout network depth. It reuses original pretrained 2D backbones, introduces no additional parameters, and enables progressive 3D fusion while preserving pretrained inductive biases.

Result: Without training, lifted models show intrinsic 3D fusion capability and outperform slice-wise 2D baselines and strong 3D counterparts in linear probing on six 3D classification and three 3D segmentation benchmarks. With full fine-tuning, matches standard 3D architectures.

Conclusion: 3D capability can be unlocked from pretrained 2D foundation models without structural modification or retraining. PlaneCycle serves as a seamless and practical 2D-to-3D lifting operator that preserves pretrained knowledge while enabling 3D understanding.

Abstract: Large-scale 2D foundation models exhibit strong transferable representations, yet extending them to 3D volumetric data typically requires retraining, adapters, or architectural redesign. We introduce PlaneCycle, a training-free, adapter-free operator for architecture-agnostic 2D-to-3D lifting of foundation models. PlaneCycle reuses the original pretrained 2D backbone by cyclically distributing spatial aggregation across orthogonal HW, DW, and DH planes throughout network depth, enabling progressive 3D fusion while preserving pretrained inductive biases. The method introduces no additional parameters and is applicable to arbitrary 2D networks. Using pretrained DINOv3 models, we evaluate PlaneCycle on six 3D classification and three 3D segmentation benchmarks. Without any training, the lifted models exhibit intrinsic 3D fusion capability and, under linear probing, outperform slice-wise 2D baselines and strong 3D counterparts, approaching the performance of fully trained models. With full fine-tuning, PlaneCycle matches standard 3D architectures, highlighting its potential as a seamless and practical 2D-to-3D lifting operator. These results demonstrate that 3D capability can be unlocked from pretrained 2D foundation models without structural modification or retraining. Code is available at https://github.com/HINTLab/PlaneCycle.

Method: Uses OCR-extracted text as auxiliary information transmitted with minimal overhead, with three key designs: adaptive OCR filtering and guidance map rendering, attention-guided fusion of guidance with decoder features, and guidance-consistent reconstruction loss for natural text blending.

Result: Achieves up to 60.6% higher text-recognition F1 score at comparable PSNR and bpp, producing sharper small-font text while preserving global image quality and decoupling text enhancement from rate-distortion optimization.

Conclusion: TextBoost demonstrates that incorporating semantic textual guidance enables effective preservation of small-font text in ultra-low bitrate compression without compromising overall image quality.

Abstract: Ultra-low bitrate image compression faces a critical challenge: preserving small-font scene text while maintaining overall visual quality. Region-of-interest (ROI) bit allocation can prioritize text but often degrades global fidelity, leading to a trade-off between local accuracy and overall image quality. Instead of relying on ROI coding, we incorporate auxiliary textual information extracted by OCR and transmitted with negligible overhead, enabling the decoder to leverage this semantic guidance. Our method, TextBoost, operationalizes this idea through three strategic designs: (i) adaptively filtering OCR outputs and rendering them into a guidance map; (ii) integrating this guidance with decoder features in a calibrated manner via an attention-guided fusion block; and (iii) enforcing guidance-consistent reconstruction in text regions with a regularizing loss that promotes natural blending with the scene. Extensive experiments on TextOCR and ICDAR 2015 demonstrate that TextBoost yields up to 60.6% higher text-recognition F1 at comparable Peak Signal-to-Noise Ratio (PSNR) and bits per pixel (bpp), producing sharper small-font text while preserving global image quality and effectively decoupling text enhancement from global rate-distortion optimization.

Method: Architectural extension based on Feature-Residual Discriminator (FR-Disc), adapting previous skeletal data approaches to complex video domain. Uses residual features to distinguish known vs. unknown actions in few-shot settings.

Result: Extensive experiments on five datasets show common open-set techniques provide only marginal gains, but FR-Disc significantly enhances unknown rejection capabilities without compromising closed-set accuracy, setting new state-of-the-art for FSOS-AR.

Conclusion: FR-Disc effectively addresses the open-set challenge in few-shot action recognition for videos, providing a robust solution that maintains closed-set performance while improving unknown action detection.

Abstract: Few-Shot Action Recognition (FS-AR) has shown promising results but is often limited by a closed-set assumption that fails in real-world open-set scenarios. While Few-Shot Open-Set (FSOS) recognition is well-established for images, its extension to spatio-temporal video data remains underexplored. To address this, we propose an architectural extension based on a Feature-Residual Discriminator (FR-Disc), adapting previous work on skeletal data to the more complex video domain. Extensive experiments on five datasets demonstrate that while common open-set techniques provide only marginal gains, our FR-Disc significantly enhances unknown rejection capabilities without compromising closed-set accuracy, setting a new state-of-the-art for FSOS-AR. The project website, code, and benchmark are available at: https://hsp-iit.github.io/fsosar/.

Method: An inference-time attention regulation framework with two modules: 1) Anatomy-aware attention regularization that gates self-attention and anatomy-token cross-attention with organ masks to confine structural interactions to anatomical ROIs, and 2) Pathology-guided module that enhances pathology-token cross-attention within target lung regions during early denoising and performs lightweight latent corrections driven by attention-concentration energy.

Result: Extensive evaluations on CXR datasets show improved anatomical consistency and more precise, controllable pathological edits compared with standard diffusion editing, supporting localized counterfactual analysis and data augmentation for downstream tasks.

Conclusion: The proposed attention regulation framework enables reliable counterfactual CXR synthesis by addressing key limitations of diffusion-based editing methods, with applications in medical imaging analysis and data augmentation.

Abstract: Counterfactual generation for chest X-rays (CXR) aims to simulate plausible pathological changes while preserving patient-specific anatomy. However, diffusion-based editing methods often suffer from structural drift, where stable anatomical semantics propagate globally through attention and distort non-target regions, and unstable pathology expression, since subtle and localized lesions induce weak and noisy conditioning signals. We present an inference-time attention regulation framework for reliable counterfactual CXR synthesis. An anatomy-aware attention regularization module gates self-attention and anatomy-token cross-attention with organ masks, confining structural interactions to anatomical ROIs and reducing unintended distortions. A pathology-guided module enhances pathology-token cross-attention within target lung regions during early denoising and performs lightweight latent corrections driven by an attention-concentration energy, enabling controllable lesion localization and extent. Extensive evaluations on CXR datasets show improved anatomical consistency and more precise, controllable pathological edits compared with standard diffusion editing, supporting localized counterfactual analysis and data augmentation for downstream tasks.

Method: Proposes LISTA-Transformer that deeply integrates Learnable Iterative Shrinkage Threshold Algorithm (LISTA) sparse encoding with visual Transformer. Uses continuous wavelet transform to convert vibration signals into time-frequency maps, then inputs them into LISTA-Transformer for adaptive local and global feature collaboration.

Result: Achieved 98.5% fault recognition rate on CWRU dataset, which is 3.3% higher than traditional methods and shows superiority over existing Transformer-based approaches.

Conclusion: The LISTA-Transformer effectively addresses limitations of existing models by combining sparse encoding with Transformer architecture, improving both local feature modeling and global dependency capture for industrial fault diagnosis applications.

Abstract: Driven by the continuous development of models such as Multi-Layer Perceptron, Convolutional Neural Network (CNN), and Transformer, deep learning has made breakthrough progress in fields such as computer vision and natural language processing, and has been successfully applied in practical scenarios such as image classification and industrial fault diagnosis. However, existing models still have certain limitations in local feature modeling and global dependency capture. Specifically, CNN is limited by local receptive fields, while Transformer has shortcomings in effectively modeling local structures, and both face challenges of high model complexity and insufficient interpretability. In response to the above issues, we proposes the following innovative work: A sparse Transformer based on Learnable Iterative Shrinkage Threshold Algorithm (LISTA-Transformer) was designed, which deeply integrates LISTA sparse encoding with visual Transformer to construct a model architecture with adaptive local and global feature collaboration mechanism. This method utilizes continuous wavelet transform to convert vibration signals into time-frequency maps and inputs them into LISTA-Transformer for more effective feature extraction. On the CWRU dataset, the fault recognition rate of our method reached 98.5%, which is 3.3% higher than traditional methods and exhibits certain superiority over existing Transformer-based approaches.

Method: Introduce augmented training framework applying diverse artificial degradations to training images; show that applying this only to a subset of individuals improves performance even for unseen individuals under similar degradations.

Result: Augmented training leads to up to 8.5% Rank-1 accuracy improvement on real-world degraded animal images, with benchmarks and expert annotations provided for further research.

Conclusion: Systematic study of image degradation in wildlife re-identification with practical training framework improves robustness; provides necessary benchmarks, code, and data for community advancement.

Abstract: Wildlife re-identification aims to recognise individual animals by matching query images to a database of previously identified individuals, based on their fine-scale unique morphological characteristics. Current state-of-the-art models for multispecies re- identification are based on deep metric learning representing individual identities by fea- ture vectors in an embedding space, the similarity of which forms the basis for a fast automated identity retrieval. Yet very often, the discriminative information of individual wild animals gets significantly reduced due to the presence of several degradation factors in images, leading to reduced retrieval performance and limiting the downstream eco- logical studies. Here, starting by showing that the extent of this performance reduction greatly varies depending on the animal species (18 wild animal datasets), we introduce an augmented training framework for deep feature extractors, where we apply artificial but diverse degradations in images in the training set. We show that applying this augmented training only to a subset of individuals, leads to an overall increased re-identification performance, under the same type of degradations, even for individuals not seen during training. The introduction of diverse degradations during training leads to a gain of up to 8.5% Rank-1 accuracy to a dataset of real-world degraded animal images, selected using human re-ID expert annotations provided here for the first time. Our work is the first to systematically study image degradation in wildlife re-identification, while introducing all the necessary benchmarks, publicly available code and data, enabling further research on this topic.

Method: Introduces a scene-token mechanism to aggregate information across unposed images, combined with a diffusion-based 3D decoder to reconstruct complete, non-pixel-aligned point clouds. The method learns a global, view-agnostic scene representation.

Result: Extensive experiments on scene-level and object-level datasets show NOVA3R outperforms state-of-the-art methods in reconstruction accuracy and completeness.

Conclusion: NOVA3R provides an effective solution for non-pixel-aligned 3D reconstruction from unposed images, addressing key limitations of pixel-aligned methods and producing more complete, physically plausible geometry.

Abstract: We present NOVA3R, an effective approach for non-pixel-aligned 3D reconstruction from a set of unposed images in a feed-forward manner. Unlike pixel-aligned methods that tie geometry to per-ray predictions, our formulation learns a global, view-agnostic scene representation that decouples reconstruction from pixel alignment. This addresses two key limitations in pixel-aligned 3D: (1) it recovers both visible and invisible points with a complete scene representation, and (2) it produces physically plausible geometry with fewer duplicated structures in overlapping regions. To achieve this, we introduce a scene-token mechanism that aggregates information across unposed images and a diffusion-based 3D decoder that reconstructs complete, non-pixel-aligned point clouds. Extensive experiments on both scene-level and object-level datasets demonstrate that NOVA3R outperforms state-of-the-art methods in terms of reconstruction accuracy and completeness.

Method: The benchmark performs a full-scale, one-to-one physical reconstruction of the entire OmniDocBench v1.5 (1,355 images) across five critical real-world scenarios: Scanning, Warping, Screen-Photography, Illumination, and Skew. It establishes complete ground-truth mapping between digital and physical versions.

Result: The benchmark demonstrates that the ‘reality gap’ in document parsing is far from closed. It enables rigorous factor-wise attribution of performance degradation, allowing researchers to pinpoint whether failures stem from geometric distortions, optical artifacts, or model limitations.

Conclusion: Real5-OmniDocBench establishes a challenging new standard for evaluating VLMs in real-world document parsing and provides a diagnostic tool to guide development of truly resilient document intelligence systems.

Abstract: While Vision-Language Models (VLMs) achieve near-perfect scores on digital document benchmarks like OmniDocBench, their performance in the unpredictable physical world remains largely unknown due to the lack of controlled yet realistic evaluations. We introduce Real5-OmniDocBench, the first benchmark that performs a full-scale, one-to-one physical reconstruction of the entire OmniDocBench v1.5 (1,355 images) across five critical real-world scenarios: Scanning, Warping, Screen-Photography, Illumination, and Skew. Unlike prior benchmark that either lack digital correspondence or employ partial sampling, our complete ground-truth mapping enables, for the first time, rigorous factor-wise attribution of performance degradation-allowing us to pinpoint whether failures stem from geometric distortions, optical artifacts, or model limitations. Our benchmark establishes a challenging new standard for the community, demonstrating that the ‘reality gap’ in document parsing is far from closed, and provides a diagnostic tool to guide the development of truly resilient document intelligence.

Method: Systematically analyze DiT representation dynamics, discover representation diversity across blocks is crucial, then propose DiverseDiT with long residual connections to diversify inputs across blocks and a representation diversity loss to encourage distinct feature learning.

Result: DiverseDiT achieves consistent performance gains and convergence acceleration on ImageNet 256x256 and 512x512 across different backbone sizes, even in challenging one-step generation. It’s complementary to existing representation learning techniques.

Conclusion: The work provides insights into DiT representation learning dynamics and offers a practical approach to enhance DiT performance through explicit promotion of representation diversity.

Abstract: Recent breakthroughs in Diffusion Transformers (DiTs) have revolutionized the field of visual synthesis due to their superior scalability. To facilitate DiTs’ capability of capturing meaningful internal representations, recent works such as REPA incorporate external pretrained encoders for representation alignment. However, the underlying mechanisms governing representation learning within DiTs are not well understood. To this end, we first systematically investigate the representation dynamics of DiTs. Through analyzing the evolution and influence of internal representations under various settings, we reveal that representation diversity across blocks is a crucial factor for effective learning. Based on this key insight, we propose DiverseDiT, a novel framework that explicitly promotes representation diversity. DiverseDiT incorporates long residual connections to diversify input representations across blocks and a representation diversity loss to encourage blocks to learn distinct features. Extensive experiments on ImageNet 256x256 and 512x512 demonstrate that our DiverseDiT yields consistent performance gains and convergence acceleration when applied to different backbones with various sizes, even when tested on the challenging one-step generation setting. Furthermore, we show that DiverseDiT is complementary to existing representation learning techniques, leading to further performance gains. Our work provides valuable insights into the representation learning dynamics of DiTs and offers a practical approach for enhancing their performance.

Method: Decomposes videos into cubemap representations with six faces, then autoregressively synthesizes content in planned spatio-temporal order. Introduces: (1) spatio-temporal autoregressive strategy across cube faces and time windows, (2) cube face context management with sparse attention, and (3) continuity-aware techniques including cube-aware positional encoding, padding, and blending.

Result: Outperforms state-of-the-art methods in native resolution and visual quality, supporting practical VR application scenarios with 4K-resolution 360° video generation.

Conclusion: CubeComposer enables efficient high-resolution 360° video generation through innovative cubemap decomposition and autoregressive synthesis, addressing computational limitations of previous approaches.

Abstract: Generating high-quality 360° panoramic videos from perspective input is one of the crucial applications for virtual reality (VR), whereby high-resolution videos are especially important for immersive experience. Existing methods are constrained by computational limitations of vanilla diffusion models, only supporting $\leq$ 1K resolution native generation and relying on suboptimal post super-resolution to increase resolution. We introduce CubeComposer, a novel spatio-temporal autoregressive diffusion model that natively generates 4K-resolution 360° videos. By decomposing videos into cubemap representations with six faces, CubeComposer autoregressively synthesizes content in a well-planned spatio-temporal order, reducing memory demands while enabling high-resolution output. Specifically, to address challenges in multi-dimensional autoregression, we propose: (1) a spatio-temporal autoregressive strategy that orchestrates 360° video generation across cube faces and time windows for coherent synthesis; (2) a cube face context management mechanism, equipped with a sparse context attention design to improve efficiency; and (3) continuity-aware techniques, including cube-aware positional encoding, padding, and blending to eliminate boundary seams. Extensive experiments on benchmark datasets demonstrate that CubeComposer outperforms state-of-the-art methods in native resolution and visual quality, supporting practical VR application scenarios. Project page: https://lg-li.github.io/project/cubecomposer

Method: DeNuC uses a lightweight model for nuclei localization, then leverages pathology foundation models to encode images and extract nucleus-specific features from detected coordinates for classification.

Result: DeNuC significantly outperforms state-of-the-art methods, improving F1 scores by 4.2% and 3.6% on BRCAM2C and PUMA datasets while using only 16% of trainable parameters.

Conclusion: Decoupling detection and classification effectively unlocks foundation models’ representational potential for nuclei analysis in pathology, offering superior performance with reduced computational burden.

Abstract: Pathology Foundation Models (FMs) have shown strong performance across a wide range of pathology image representation and diagnostic tasks. However, FMs do not exhibit the expected performance advantage over traditional specialized models in Nuclei Detection and Classification (NDC). In this work, we reveal that jointly optimizing nuclei detection and classification leads to severe representation degradation in FMs. Moreover, we identify that the substantial intrinsic disparity in task difficulty between nuclei detection and nuclei classification renders joint NDC optimization unnecessarily computationally burdensome for the detection stage. To address these challenges, we propose DeNuC, a simple yet effective method designed to break through existing bottlenecks by Decoupling Nuclei detection and Classification. DeNuC employs a lightweight model for accurate nuclei localization, subsequently leveraging a pathology FM to encode input images and query nucleus-specific features based on the detected coordinates for classification. Extensive experiments on three widely used benchmarks demonstrate that DeNuC effectively unlocks the representational potential of FMs for NDC and significantly outperforms state-of-the-art methods. Notably, DeNuC improves F1 scores by 4.2% and 3.6% (or higher) on the BRCAM2C and PUMA datasets, respectively, while using only 16% (or fewer) trainable parameters compared to other methods. Code is available at https://github.com/ZijiangY1116/DeNuC.

Method: Created Multi-view Oriented Observation (MOO) dataset - large-scale synthetic AG-ReID dataset with 1,000 cattle individuals captured from 128 uniformly sampled viewpoints (128,000 annotated images). Used this controlled dataset to quantify elevation influence and identify critical thresholds.

Result: Identified critical elevation threshold above which models generalize significantly better to unseen views. Validated transferability to real-world applications in zero-shot and supervised settings, showing performance gains across four real-world cattle datasets. Synthetic geometric priors effectively bridge domain gap.

Conclusion: MOO dataset and analysis provide foundation for future model development in cross-view animal ReID. Synthetic data with precise geometric annotations enables systematic study of viewpoint variations and improves real-world performance.

Abstract: Animal re-identification (ReID) faces critical challenges due to viewpoint variations, particularly in Aerial-Ground (AG-ReID) settings where models must match individuals across drastic elevation changes. However, existing datasets lack the precise angular annotations required to systematically analyze these geometric variations. To address this, we introduce the Multi-view Oriented Observation (MOO) dataset, a large-scale synthetic AG-ReID dataset of $1,000$ cattle individuals captured from $128$ uniformly sampled viewpoints ($128,000$ annotated images). Using this controlled dataset, we quantify the influence of elevation and identify a critical elevation threshold, above which models generalize significantly better to unseen views. Finally, we validate the transferability to real-world applications in both zero-shot and supervised settings, demonstrating performance gains across four real-world cattle datasets and confirming that synthetic geometric priors effectively bridge the domain gap. Collectively, this dataset and analysis lay the foundation for future model development in cross-view animal ReID. MOO is publicly available at https://github.com/TurtleSmoke/MOO.

Method: Proposes a morphology-decoupled framework with Zero-Initialized Gated Cross-Task Attention to exploit dense EPVS context for sparse lacune detection, mixed-supervision with Mutual Exclusion and Centerline Dice losses for biological/topological consistency, and Anatomically-Informed Inference Calibration to suppress false positives based on tissue semantics.

Result: Achieved state-of-the-art performance on VALDO 2021 dataset (N=40) with 5-fold cross-validation, surpassing task winners in lacunae detection precision (71.1%, p=0.01) and F1-score (62.6%, p=0.03). Demonstrated robustness on external EPAD cohort (N=1762) for large-scale population studies.

Conclusion: The proposed framework effectively addresses the challenges of simultaneous segmentation of divergent CSVD markers through cross-task attention and anatomical constraints, showing strong performance and generalizability for medical image analysis applications.

Abstract: Cerebral small vessel disease (CSVD) markers, specifically enlarged perivascular spaces (EPVS) and lacunae, present a unique challenge in medical image analysis due to their radiological mimicry. Standard segmentation networks struggle with feature interference and extreme class imbalance when handling these divergent targets simultaneously. To address these issues, we propose a morphology-decoupled framework where Zero-Initialized Gated Cross-Task Attention exploits dense EPVS context to guide sparse lacune detection. Furthermore, biological and topological consistency are enforced via a mixed-supervision strategy integrating Mutual Exclusion and Centerline Dice losses. Finally, we introduce an Anatomically-Informed Inference Calibration mechanism to dynamically suppress false positives based on tissue semantics. Extensive 5-folds cross-validation on the VALDO 2021 dataset (N=40) demonstrates state-of-the-art performance, notably surpassing task winners in lacunae detection precision (71.1%, p=0.01) and F1-score (62.6%, p=0.03). Furthermore, evaluation on the external EPAD cohort (N=1762) confirms the model’s robustness for large-scale population studies. Code will be released upon acceptance.

Method: SPRINT introduces a mixed episodic training strategy that leverages confidence-based pseudo-labeling to enrich novel class representations and exploits low storage costs to retain base class history. The framework is specifically tailored for tabular distributions.

Result: Extensive evaluation across six diverse benchmarks spanning cybersecurity, healthcare, and ecological domains demonstrates SPRINT’s cross-domain robustness. It achieves state-of-the-art average accuracy of 77.37% (5-shot), outperforming the strongest incremental baseline by 4.45%.

Conclusion: SPRINT successfully adapts FSCIL to tabular domains, addressing their unique characteristics and achieving superior performance across diverse real-world applications while maintaining knowledge of previously learned classes.

Abstract: Real-world systems must continuously adapt to novel concepts from limited data without forgetting previously acquired knowledge. While Few-Shot Class-Incremental Learning (FSCIL) is established in computer vision, its application to tabular domains remains largely unexplored. Unlike images, tabular streams (e.g., logs, sensors) offer abundant unlabeled data, a scarcity of expert annotations and negligible storage costs, features ignored by existing vision-based methods that rely on restrictive buffers. We introduce SPRINT, the first FSCIL framework tailored for tabular distributions. SPRINT introduces a mixed episodic training strategy that leverages confidence-based pseudo-labeling to enrich novel class representations and exploits low storage costs to retain base class history. Extensive evaluation across six diverse benchmarks spanning cybersecurity, healthcare, and ecological domains, demonstrates SPRINT’s cross-domain robustness. It achieves a state-of-the-art average accuracy of 77.37% (5-shot), outperforming the strongest incremental baseline by 4.45%.

Method: Proposes Online Sparse Coefficients Field with CLIP Global Codebook to bind 2D CLIP embeddings to 3D Gaussians while minimizing memory. Uses 3D U-Net to generate 3D geometric-aware CLIP features by aggregating partial point clouds from 3DGS.

Result: Demonstrates effectiveness and efficiency on diverse indoor datasets (ScanNet, ScanNet++, Replica), enabling online reconstruction from over 300 streaming images with feed-forward generalization to novel scenes.

Conclusion: EmbodiedSplat enables simultaneous online 3D reconstruction and semantic understanding, addressing limitations of existing offline methods and supporting real-time embodied AI applications.

Abstract: Understanding a 3D scene immediately with its exploration is essential for embodied tasks, where an agent must construct and comprehend the 3D scene in an online and nearly real-time manner. In this study, we propose EmbodiedSplat, an online feed-forward 3DGS for open-vocabulary scene understanding that enables simultaneous online 3D reconstruction and 3D semantic understanding from the streaming images. Unlike existing open-vocabulary 3DGS methods which are typically restricted to either offline or per-scene optimization setting, our objectives are two-fold: 1) Reconstructs the semantic-embedded 3DGS of the entire scene from over 300 streaming images in an online manner. 2) Highly generalizable to novel scenes with feed-forward design and supports nearly real-time 3D semantic reconstruction when combined with real-time 2D models. To achieve these objectives, we propose an Online Sparse Coefficients Field with a CLIP Global Codebook where it binds the 2D CLIP embeddings to each 3D Gaussian while minimizing memory consumption and preserving the full semantic generalizability of CLIP. Furthermore, we generate 3D geometric-aware CLIP features by aggregating the partial point cloud of 3DGS through 3D U-Net to compensate the 3D geometric prior to 2D-oriented language embeddings. Extensive experiments on diverse indoor datasets, including ScanNet, ScanNet++, and Replica, demonstrate both the effectiveness and efficiency of our method. Check out our project page in https://0nandon.github.io/EmbodiedSplat/.

Method: Proposes a truly equivariant architecture that lifts interval values (saturation/luminance) to values on a circle (double-cover) instead of approximating them as 1D translations. This approach properly models the geometric structure of color transformations.

Result: Resolves approximation artifacts of previous methods, improves interpretability and generalizability, and achieves better predictive performance than conventional and equivariant baselines on tasks like fine-grained classification and medical imaging.

Conclusion: The proposed lifting approach provides a mathematically sound way to handle color transformations, with applications extending beyond color to other geometric transformations like scale.

Abstract: When the color distribution of input images changes at inference, the performance of conventional neural network architectures drops considerably. A few researchers have begun to incorporate prior knowledge of color geometry in neural network design. These color equivariant architectures have modeled hue variation with 2D rotations, and saturation and luminance transformations as 1D translations. While this approach improves neural network robustness to color variations in a number of contexts, we find that approximating saturation and luminance (interval valued quantities) as 1D translations introduces appreciable artifacts. In this paper, we introduce a color equivariant architecture that is truly equivariant. Instead of approximating the interval with the real line, we lift values on the interval to values on the circle (a double-cover) and build equivariant representations there. Our approach resolves the approximation artifacts of previous methods, improves interpretability and generalizability, and achieves better predictive performance than conventional and equivariant baselines on tasks such as fine-grained classification and medical imaging tasks. Going beyond the context of color, we show that our proposed lifting can also extend to geometric transformations such as scale.

Method: Proposes RANGER framework with: 1) Sparsely-gated Mixture-of-Experts (MoE) decoder with noisy top-k routing and load-balancing regularization for dynamic expert specialization; 2) Adaptive retrieval re-ranking module that selectively refines retrieved knowledge base memory based on visual features to reduce noise and improve semantic alignment.

Result: Achieves optimal performance on PathText dataset with BLEU-1 to BLEU-4 scores of 0.4598, 0.3044, 0.2036, 0.1435, METEOR of 0.1883, and ROUGE-L of 0.3038, demonstrating consistent improvements over existing approaches.

Conclusion: RANGER validates the effectiveness of dynamic expert routing and adaptive knowledge refinement for semantically grounded pathology report generation, addressing limitations of homogeneous decoder architectures and static knowledge retrieval.

Abstract: Pathology report generation remains a relatively under-explored downstream task, primarily due to the gigapixel scale and complex morphological heterogeneity of Whole Slide Images (WSIs). Existing pathology report generation frameworks typically employ transformer architectures, relying on a homogeneous decoder architecture and static knowledge retrieval integration. Such architectures limit generative specialization and may introduce noisy external guidance during the report generation process. To address these limitations, we propose RANGER, a sparsely-gated Mixture-of-Experts (MoE) framework with adaptive retrieval re-ranking for pathology report generation. Specifically, we integrate a sparsely gated MoE into the decoder, along with noisy top-$k$ routing and load-balancing regularization, to enable dynamic expert specialization across various diagnostic patterns. Additionally, we introduce an adaptive retrieval re-ranking module that selectively refines retrieved memory from a knowledge base before integration, reducing noise and improving semantic alignment based on visual feature representations. We perform extensive experiments on the PathText-BRCA dataset and demonstrate consistent improvements over existing approaches across standard natural language generation metrics. Our full RANGER model achieves optimal performance on PathText dataset, reaching BLEU-1 to BLEU-4 scores of 0.4598, 0.3044, 0.2036, and 0.1435, respectively, with METEOR of 0.1883, and ROUGE-L of 0.3038, validating the effectiveness of dynamic expert routing and adaptive knowledge refinement for semantically grounded pathology report generation.

Method: Proposes ViterbiPlanNet with a Differentiable Viterbi Layer (DVL) that embeds a Procedural Knowledge Graph (PKG) using the Viterbi decoding algorithm. Replaces non-differentiable operations with smooth relaxations for end-to-end optimization, enabling graph-based decoding.

Result: Achieves state-of-the-art performance on CrossTask, COIN, and NIV datasets with an order of magnitude fewer parameters than diffusion- and LLM-based planners. Shows improved sample efficiency and robustness to shorter unseen horizons. Establishes unified testing protocol with consistent splits and bootstrapped statistical significance.

Conclusion: Explicit integration of procedural knowledge through differentiable graph-based decoding enables more efficient and effective visual planning compared to implicit learning approaches. The method demonstrates superior performance with significantly reduced model complexity.

Abstract: Procedural planning aims to predict a sequence of actions that transforms an initial visual state into a desired goal, a fundamental ability for intelligent agents operating in complex environments. Existing approaches typically rely on large-scale models that learn procedural structures implicitly, resulting in limited sample-efficiency and high computational cost. In this work we introduce ViterbiPlanNet, a principled framework that explicitly integrates procedural knowledge into the learning process through a Differentiable Viterbi Layer (DVL). The DVL embeds a Procedural Knowledge Graph (PKG) directly with the Viterbi decoding algorithm, replacing non-differentiable operations with smooth relaxations that enable end-to-end optimization. This design allows the model to learn through graph-based decoding. Experiments on CrossTask, COIN, and NIV demonstrate that ViterbiPlanNet achieves state-of-the-art performance with an order of magnitude fewer parameters than diffusion- and LLM-based planners. Extensive ablations show that performance gains arise from our differentiable structure-aware training rather than post-hoc refinement, resulting in improved sample efficiency and robustness to shorter unseen horizons. We also address testing inconsistencies establishing a unified testing protocol with consistent splits and evaluation metrics. With this new protocol, we run experiments multiple times and report results using bootstrapping to assess statistical significance.

Method: Proposes a text-enhanced compression framework treating text as an alternative modality that can be losslessly compressed with LLMs. Uses lightweight text descriptions with small image feature vectors capturing complementary information. Introduces Similarity Space Replication (SSR) to learn adaptive image embeddings in one shot that capture only information complementary to text descriptions.

Result: Validated on multiple downstream localization tasks including Visual Place Recognition and object-centric Monte Carlo localization in indoor/outdoor settings. SSR achieves 2 times better compression than competing baselines on state-of-the-art datasets (TokyoVal, Pittsburgh30k, Replica, KITTI).

Conclusion: The proposed framework effectively reduces memory and bandwidth footprints while retaining high-fidelity localization performance through multimodal compression combining text and image features.

Abstract: Mapping is crucial in robotics for localization and downstream decision-making. As robots are deployed in ever-broader settings, the maps they rely on continue to increase in size. However, storing these maps indefinitely (cold storage), transferring them across networks, or sending localization queries to cloud-hosted maps imposes prohibitive memory and bandwidth costs. We propose a text-enhanced compression framework that reduces both memory and bandwidth footprints while retaining high-fidelity localization. The key idea is to treat text as an alternative modality: one that can be losslessly compressed with large language models. We propose leveraging lightweight text descriptions combined with very small image feature vectors, which capture “complementary information” as a compact representation for the mapping task. Building on this, our novel technique, Similarity Space Replication (SSR), learns an adaptive image embedding in one shot that captures only the information “complementary” to the text descriptions. We validate our compression framework on multiple downstream localization tasks, including Visual Place Recognition as well as object-centric Monte Carlo localization in both indoor and outdoor settings. SSR achieves 2 times better compression than competing baselines on state-of-the-art datasets, including TokyoVal, Pittsburgh30k, Replica, and KITTI.

Method: ZipMap uses test-time training layers to compress an entire image collection into a compact hidden scene state in a single forward pass. It’s a stateful feed-forward model that enables linear-time, bidirectional 3D reconstruction.

Result: ZipMap reconstructs over 700 frames in under 10 seconds on a single H100 GPU, more than 20× faster than state-of-the-art methods while matching or surpassing their accuracy. Enables real-time scene-state querying and sequential streaming reconstruction.

Conclusion: ZipMap demonstrates that linear-time 3D reconstruction is possible without sacrificing accuracy, offering efficient processing of large image collections and enabling new applications like real-time scene querying and streaming reconstruction.

Abstract: Feed-forward transformer models have driven rapid progress in 3D vision, but state-of-the-art methods such as VGGT and $π^3$ have a computational cost that scales quadratically with the number of input images, making them inefficient when applied to large image collections. Sequential-reconstruction approaches reduce this cost but sacrifice reconstruction quality. We introduce ZipMap, a stateful feed-forward model that achieves linear-time, bidirectional 3D reconstruction while matching or surpassing the accuracy of quadratic-time methods. ZipMap employs test-time training layers to zip an entire image collection into a compact hidden scene state in a single forward pass, enabling reconstruction of over 700 frames in under 10 seconds on a single H100 GPU, more than $20\times$ faster than state-of-the-art methods such as VGGT. Moreover, we demonstrate the benefits of having a stateful representation in real-time scene-state querying and its extension to sequential streaming reconstruction.

Method: The study uses deep learning techniques applied to sequence data from colonoscopy videos, incorporating temporal information between frames. A comprehensive multi-center dataset was compiled through collaboration between data scientists and gastroenterologists to evaluate real-time clinical applications.

Result: The research demonstrates that incorporating temporal sequence data and frame-to-frame relationships significantly enhances the precision of automated polyp detection and segmentation methods compared to static image approaches.

Conclusion: Temporal information in colonoscopy video sequences is critical for developing robust automated detection systems, potentially reducing missed polyp rates and improving colorectal cancer prevention through more accurate real-time clinical applications.

Abstract: Colonic polyps are well-recognized precursors to colorectal cancer (CRC), typically detected during colonoscopy. However, the variability in appearance, location, and size of these polyps complicates their detection and removal, leading to challenges in effective surveillance, intervention, and subsequently CRC prevention. The processes of colonoscopy surveillance and polyp removal are highly reliant on the expertise of gastroenterologists and occur within the complexities of the colonic structure. As a result, there is a high rate of missed detections and incomplete removal of colonic polyps, which can adversely impact patient outcomes. Recently, automated methods that use machine learning have been developed to enhance polyps detection and segmentation, thus helping clinical processes and reducing missed rates. These advancements highlight the potential for improving diagnostic accuracy in real-time applications, which ultimately facilitates more effective patient management. Furthermore, integrating sequence data and temporal information could significantly enhance the precision of these methods by capturing the dynamic nature of polyp growth and the changes that occur over time. To rigorously investigate these challenges, data scientists and experts gastroenterologists collaborated to compile a comprehensive dataset that spans multiple centers and diverse populations. This initiative aims to underscore the critical importance of incorporating sequence data and temporal information in the development of robust automated detection and segmentation methods. This study evaluates the applicability of deep learning techniques developed in real-time clinical colonoscopy tasks using sequence data, highlighting the critical role of temporal relationships between frames in improving diagnostic precision.

Method: Develops a compositional 3D Gaussian representation with multiple layers of free-form garments. Decomposes avatars into bodies and shape-agnostic neural garments, disentangling each garment layer from multi-view videos and canonicalizing them into shape-independent spaces.

Result: Achieves state-of-the-art performance on novel pose synthesis benchmarks, models photorealistic avatars with high-fidelity dynamics, and enables practical virtual try-on applications where clothing can be freely transferred to new subjects.

Conclusion: Gaussian Wardrobe successfully creates compositional 3D neural avatars that separate clothing from bodies, enabling clothing transfer between individuals and advancing digital wardrobe applications.

Abstract: We introduce Gaussian Wardrobe, a novel framework to digitalize compositional 3D neural avatars from multi-view videos. Existing methods for 3D neural avatars typically treat the human body and clothing as an inseparable entity. However, this paradigm fails to capture the dynamics of complex free-form garments and limits the reuse of clothing across different individuals. To overcome these problems, we develop a novel, compositional 3D Gaussian representation to build avatars from multiple layers of free-form garments. The core of our method is decomposing neural avatars into bodies and layers of shape-agnostic neural garments. To achieve this, our framework learns to disentangle each garment layer from multi-view videos and canonicalizes it into a shape-independent space. In experiments, our method models photorealistic avatars with high-fidelity dynamics, achieving new state-of-the-art performance on novel pose synthesis benchmarks. In addition, we demonstrate that the learned compositional garments contribute to a versatile digital wardrobe, enabling a practical virtual try-on application where clothing can be freely transferred to new subjects. Project page: https://ait.ethz.ch/gaussianwardrobe

Method: 1) Self-supervised representation learning to encode/decode expressions in latent space and decouple them from other motion information. 2) New keypoint computation method for arbitrary motion control. 3) Variational autoencoder to map expression features to continuous Gaussian distribution for expression interpolation.

Result: Extensive experiments on public datasets show MMFA offers pronounced advantages over prior arts in creating realistic animation and manipulating face motion.

Conclusion: MMFA successfully addresses the limitations of existing unsupervised keypoint methods by enabling controllable face generation through proper motion information decoupling and expression interpolation capabilities.

Abstract: Face animation deals with controlling and generating facial features with a wide range of applications. The methods based on unsupervised keypoint positioning can produce realistic and detailed virtual portraits. However, they cannot achieve controllable face generation since the existing keypoint decomposition pipelines fail to fully decouple identity semantics and intertwined motion information (e.g., rotation, translation, and expression). To address these issues, we present a new method, Motion Manipulation via unsupervised keypoint positioning in Face Animation (MMFA). We first introduce self-supervised representation learning to encode and decode expressions in the latent feature space and decouple them from other motion information. Secondly, we propose a new way to compute keypoints aiming to achieve arbitrary motion control. Moreover, we design a variational autoencoder to map expression features to a continuous Gaussian distribution, allowing us for the first time to interpolate facial expressions in an unsupervised framework. We have conducted extensive experiments on publicly available datasets to validate the effectiveness of MMFA, which show that MMFA offers pronounced advantages over prior arts in creating realistic animation and manipulating face motion.

Method: Constructs large-scale dataset with 100K+ pairs across four modalities, then uses dual diffusion models: Texture Diffusion Model (TDM) with multi-condition guidance from text/image, and Geometry Diffusion Model (GDM) guided by text.

Result: Generates high-fidelity, shading-free 3D avatars in under 10 seconds, significantly outperforming SOTA in quality, detail alignment, and computational efficiency.

Conclusion: PromptAvatar enables fast, high-quality 3D avatar generation from multimodal prompts by learning direct mappings and eliminating iterative optimization bottlenecks.

Abstract: Generating high-fidelity 3D avatars from text or image prompts is highly sought after in virtual reality and human-computer interaction. However, existing text-driven methods often rely on iterative Score Distillation Sampling (SDS) or CLIP optimization, which struggle with fine-grained semantic control and suffer from excessively slow inference. Meanwhile, image-driven approaches are severely bottlenecked by the scarcity and high acquisition cost of high-quality 3D facial scans, limiting model generalization. To address these challenges, we first construct a novel, large-scale dataset comprising over 100,000 pairs across four modalities: fine-grained textual descriptions, in-the-wild face images, high-quality light-normalized texture UV maps, and 3D geometric shapes. Leveraging this comprehensive dataset, we propose PromptAvatar, a framework featuring dual diffusion models. Specifically, it integrates a Texture Diffusion Model (TDM) that supports flexible multi-condition guidance from text and/or image prompts, alongside a Geometry Diffusion Model (GDM) guided by text prompts. By learning the direct mapping from multi-modal prompts to 3D representations, PromptAvatar eliminates the need for time-consuming iterative optimization, successfully generating high-fidelity, shading-free 3D avatars in under 10 seconds. Extensive quantitative and qualitative experiments demonstrate that our method significantly outperforms existing state-of-the-art approaches in generation quality, fine-grained detail alignment, and computational efficiency.

Method: Scalable framework comparing rule-based augmentation libraries with generative AI image-editing models on 40 clear-day images, using VLM jury for perceptual realism and embedding-based distributional analysis.

Result: Generative AI methods substantially outperform rule-based approaches (3.6x acceptance rate), with fog easiest to simulate and nighttime most challenging. Leading generative methods match/exceed real-image performance for most conditions.

Conclusion: Modern generative image-editing models enable scalable generation of realistic adverse-condition imagery for evaluation pipelines, though validation against human studies remains important.

Abstract: Evaluation of AI systems often requires synthetic test cases, particularly for rare or safety-critical conditions that are difficult to observe in operational data. Generative AI offers a promising approach for producing such data through controllable image editing, but its usefulness depends on whether the resulting images are sufficiently realistic to support meaningful evaluation. We present a scalable framework for assessing the realism of synthetic image-editing methods and apply it to the task of adding environmental conditions-fog, rain, snow, and nighttime-to car-mounted camera images. Using 40 clear-day images, we compare rule-based augmentation libraries with generative AI image-editing models. Realism is evaluated using two complementary automated metrics: a vision-language model (VLM) jury for perceptual realism assessment, and embedding-based distributional analysis to measure similarity to genuine adverse-condition imagery. Generative AI methods substantially outperform rule-based approaches, with the best generative method achieving approximately 3.6 times the acceptance rate of the best rule-based method. Performance varies across conditions: fog proves easiest to simulate, while nighttime transformations remain challenging. Notably, the VLM jury assigns imperfect acceptance even to real adverse-condition imagery, establishing practical ceilings against which synthetic methods can be judged. By this standard, leading generative methods match or exceed real-image performance for most conditions. These results suggest that modern generative image-editing models can enable scalable generation of realistic adverse-condition imagery for evaluation pipelines. Our framework therefore provides a practical approach for scalable realism evaluation, though validation against human studies remains an important direction for future work.

Method: Formulates articulated HOI synthesis as 4D reconstruction from monocular video priors. Uses flow-based part segmentation to disentangle dynamic/static regions, and a decoupled reconstruction pipeline that first recovers object articulation then synthesizes human motion conditioned on object states.

Result: ArtHOI significantly outperforms prior methods in contact accuracy, penetration reduction, and articulation fidelity across diverse articulated scenes (opening fridges, cabinets, microwaves), extending zero-shot interaction synthesis beyond rigid manipulation.

Conclusion: ArtHOI bridges video-based generation and geometry-aware reconstruction, producing interactions that are both semantically aligned and physically grounded, enabling articulated human-object interaction synthesis without 3D supervision.

Abstract: Synthesizing physically plausible articulated human-object interactions (HOI) without 3D/4D supervision remains a fundamental challenge. While recent zero-shot approaches leverage video diffusion models to synthesize human-object interactions, they are largely confined to rigid-object manipulation and lack explicit 4D geometric reasoning. To bridge this gap, we formulate articulated HOI synthesis as a 4D reconstruction problem from monocular video priors: given only a video generated by a diffusion model, we reconstruct a full 4D articulated scene without any 3D supervision. This reconstruction-based approach treats the generated 2D video as supervision for an inverse rendering problem, recovering geometrically consistent and physically plausible 4D scenes that naturally respect contact, articulation, and temporal coherence. We introduce ArtHOI, the first zero-shot framework for articulated human-object interaction synthesis via 4D reconstruction from video priors. Our key designs are: 1) Flow-based part segmentation: leveraging optical flow as a geometric cue to disentangle dynamic from static regions in monocular video; 2) Decoupled reconstruction pipeline: joint optimization of human motion and object articulation is unstable under monocular ambiguity, so we first recover object articulation, then synthesize human motion conditioned on the reconstructed object states. ArtHOI bridges video-based generation and geometry-aware reconstruction, producing interactions that are both semantically aligned and physically grounded. Across diverse articulated scenes (e.g., opening fridges, cabinets, microwaves), ArtHOI significantly outperforms prior methods in contact accuracy, penetration reduction, and articulation fidelity, extending zero-shot interaction synthesis beyond rigid manipulation through reconstruction-informed synthesis.

Method: Two-stage pipeline: anatomical masks condition image synthesis; benchmark three generative architectures (DDPM, LDM, FM) across fidelity, utility, and privacy axes; evaluate under limited-data conditions.

Result: DDPM provides best balance between segmentation utility, image fidelity, and privacy preservation; FM shows promising privacy characteristics with slightly lower task performance; diffusion models outperform in limited-data scenarios.

Conclusion: Establishes framework for safe synthetic data augmentation in medical imaging; quantifies trade-offs between cross-domain generalization and patient confidentiality; diffusion models (especially DDPM) offer optimal balance.

Abstract: Deep learning in cardiac MRI (CMR) is fundamentally constrained by both data scarcity and privacy regulations. This study systematically benchmarks three generative architectures: Denoising Diffusion Probabilistic Models (DDPM), Latent Diffusion Models (LDM), and Flow Matching (FM) for synthetic CMR generation. Utilizing a two-stage pipeline where anatomical masks condition image synthesis, we evaluate generated data across three critical axes: fidelity, utility, and privacy. Our results show that diffusion-based models, particularly DDPM, provide the most effective balance between downstream segmentation utility, image fidelity, and privacy preservation under limited-data conditions, while FM demonstrates promising privacy characteristics with slightly lower task-level performance. These findings quantify the trade-offs between cross-domain generalization and patient confidentiality, establishing a framework for safe and effective synthetic data augmentation in medical imaging.

Method: Hold-One-Shot-Out (HOSO) uses a one-shot hold-out set to learn the blending ratio while training the adapter on remaining few-shot support examples. This decoupled training approach enables validation-free optimization of the trade-off between pretrained CLIP knowledge and dataset-specific supervision.

Result: HOSO-Adapter outperforms CLIP-Adapter baseline by >4 percentage points on average across 11 standard few-shot datasets in validation-free settings. In 8- and 16-shot settings, it even beats CLIP-Adapter with optimal test-set blending ratios.

Conclusion: HOSO provides an effective validation-free approach for CLIP adaptation that learns optimal blending ratios without requiring additional validation data, making few-shot learning more practical and strictly few-shot.

Abstract: In many CLIP adaptation methods, a blending ratio hyperparameter controls the trade-off between general pretrained CLIP knowledge and the limited, dataset-specific supervision from the few-shot cases. Most few-shot CLIP adaptation techniques report results by ablation of the blending ratio on the test set or require additional validation sets to select the blending ratio per dataset, and thus are not strictly few-shot. We present a simple, validation-free method for learning the blending ratio in CLIP adaptation. Hold-One-Shot-Out (HOSO) presents a novel approach for CLIP-Adapter-style methods to compete in the newly established validation-free setting. CLIP-Adapter with HOSO (HOSO-Adapter) learns the blending ratio using a one-shot, hold-out set, while the adapter trains on the remaining few-shot support examples. Under the validation-free few-shot protocol, HOSO-Adapter outperforms the CLIP-Adapter baseline by more than 4 percentage points on average across 11 standard few-shot datasets. Interestingly, in the 8- and 16-shot settings, HOSO-Adapter outperforms CLIP-Adapter even with the optimal blending ratio selected on the test set. Ablation studies validate the use of a one-shot hold-out mechanism, decoupled training, and improvements over the naively learnt blending ratio baseline. Code is released here: https://github.com/chris-vorster/HOSO-Adapter

Method: Proposes a hybrid approach combining real and synthetic data, using DreamBooth fine-tuning of Stable Diffusion to generate synthetic paintings, then training classification models on this mixed dataset.

Result: Adding synthetic images leads to higher ROC-AUC and accuracy compared to using only real paintings, demonstrating improved model performance and generalization across similar artistic styles.

Conclusion: Integrating generative and discriminative methods contributes to computer vision techniques for artwork authentication in data-scarce scenarios, showing synthetic data can effectively augment limited real datasets.

Abstract: Attributing authorship to paintings is a historically complex task, and one of its main challenges is the limited availability of real artworks for training computational models. This study investigates whether synthetic images, generated through DreamBooth fine-tuning of Stable Diffusion, can improve the performance of classification models in this context. We propose a hybrid approach that combines real and synthetic data to enhance model accuracy and generalization across similar artistic styles. Experimental results show that adding synthetic images leads to higher ROC-AUC and accuracy compared to using only real paintings. By integrating generative and discriminative methods, this work contributes to the development of computer vision techniques for artwork authentication in data-scarce scenarios.

Method: Use LLM to generate plausible counterfactual descriptions for each image. Measure VLFM’s ability to distinguish correct description from hard negatives. Engineer features from similarity scores in shared embedding space. Train linear regressor on these features to predict zero-shot accuracy.

Result: Achieves Pearson-r correlation of 0.96 for predicting VLFM zero-shot test accuracy across diverse datasets. Validated on 5 datasets including standard benchmarks and underrepresented African datasets. Highly data-efficient (only 1 labeled image per class needed).

Conclusion: Provides low-cost, reliable tool for probing VLFMs, enabling informed decisions about data annotation efforts. Particularly valuable for underrepresented domains where labeled data is scarce.

Abstract: Large-scale Vision-Language Foundation Models (VLFMs), such as CLIP, now underpin a wide range of computer vision research and applications. VLFMs are often adapted to various domain-specific tasks. However, VLFM performance on novel, specialised, or underrepresented domains remains inconsistent. Evaluating VLFMs typically requires labelled test sets, which are often unavailable for niche domains of interest, particularly those from the Global South. We address this gap by proposing a highly data-efficient method to predict a VLFM’s zero-shot accuracy on a target domain using only a single labelled image per class. Our approach uses a Large Language Model to generate plausible counterfactual descriptions of a given image. By measuring the VLFM’s ability to distinguish the correct description from these hard negatives, we engineer features that capture the VLFM’s discriminative power in its shared embedding space. A linear regressor trained on these similarity scores estimates the VLFM’s zero-shot test accuracy across various visual domains with a Pearson-r correlation of 0.96. We demonstrate our method’s performance across five diverse datasets, including standard benchmark datasets and underrepresented datasets from Africa. Our work provides a low-cost, reliable tool for probing VLFMs, enabling researchers and practitioners to make informed decisions about data annotation efforts before committing significant resources. The model training code, generated captions and counterfactuals are released here: https://github.com/chris-vorster/PreLabellingProbe.

Method: FocusGraph uses: 1) A lightweight trainable Scene-Caption LLM Selector that operates on compact textual representations (graph-based captions) rather than original frames to select query-relevant clips, and 2) A training-free Patch-wise Sparse-Flow Retention (PSFR) method to select keyframes from selected clips, which are then fed to an MLLM for final answer generation.

Result: Achieves state-of-the-art results on challenging egocentric long-video QA benchmarks (FindingDory and HourVideo) while significantly reducing inference time compared to baseline approaches.

Conclusion: FocusGraph provides an effective framework for efficient long video understanding by combining textual scene representations with intelligent keyframe selection, addressing core limitations of MLLMs in processing long video sequences.

Abstract: The ability to understand long videos is vital for embodied intelligent agents, because their effectiveness depends on how well they can accumulate, organize, and leverage long-horizon perceptual memories. Recently, multimodal LLMs have been gaining popularity for solving the long video understanding task due to their general ability to understand natural language and to leverage world knowledge. However, as the number of frames provided to an MLLM increases, the quality of its responses tends to degrade, and inference time grows. Therefore, when using MLLMs for long video understanding, a crucial step is selecting key frames from the video to answer user queries. In this work, we develop FocusGraph, a framework for keyframe selection for question answering over long egocentric videos. It leverages a lightweight trainable Scene-Caption LLM Selector that selects query-relevant clips based on their graph-based captions, and a training-free method for selecting keyframes from these clips. Unlike existing methods, the proposed Scene-Caption LLM Selector does not rely on the original sequence of low-resolution frames; instead, it operates on a compact textual representation of the scene. We then design a training-free Patch-wise Sparse-Flow Retention (PSFR) method to select keyframes from the resulting sequence of clips, which are fed into an MLLM to produce the final answer. Together, these components enable FocusGraph to achieve state-of-the-art results on challenging egocentric long-video question answering benchmarks, including FindingDory and HourVideo, while significantly reducing inference time relative to baseline approaches.

Method: 14B autoregressive diffusion model with unified input representation for T2V, I2V, V2V tasks. Uses training strategies that simulate drifting to mitigate long-video issues, heavy context compression, reduced sampling steps, and infrastructure-level optimizations for memory efficiency.

Result: Achieves 19.5 FPS on single H100 GPU, supports minute-scale generation, matches quality of strong baselines, consistently outperforms prior methods on both short- and long-video generation, and enables training with image-diffusion-scale batch sizes.

Conclusion: Helios demonstrates breakthroughs in video generation efficiency and quality, particularly for long videos, while maintaining practical deployment feasibility on single GPU hardware.

Abstract: We introduce Helios, the first 14B video generation model that runs at 19.5 FPS on a single NVIDIA H100 GPU and supports minute-scale generation while matching the quality of a strong baseline. We make breakthroughs along three key dimensions: (1) robustness to long-video drifting without commonly used anti-drifting heuristics such as self-forcing, error-banks, or keyframe sampling; (2) real-time generation without standard acceleration techniques such as KV-cache, sparse/linear attention, or quantization; and (3) training without parallelism or sharding frameworks, enabling image-diffusion-scale batch sizes while fitting up to four 14B models within 80 GB of GPU memory. Specifically, Helios is a 14B autoregressive diffusion model with a unified input representation that natively supports T2V, I2V, and V2V tasks. To mitigate drifting in long-video generation, we characterize typical failure modes and propose simple yet effective training strategies that explicitly simulate drifting during training, while eliminating repetitive motion at its source. For efficiency, we heavily compress the historical and noisy context and reduce the number of sampling steps, yielding computational costs comparable to – or lower than – those of 1.3B video generative models. Moreover, we introduce infrastructure-level optimizations that accelerate both inference and training while reducing memory consumption. Extensive experiments demonstrate that Helios consistently outperforms prior methods on both short- and long-video generation. We plan to release the code, base model, and distilled model to support further development by the community.

Method: Simple transformer-based model using self-attention modules to capture spatial dependencies within poses and temporal relationships across motion sequences. End-to-end architecture versatile enough for pose-only, trajectory-only, and combined prediction tasks.

Result: Achieves state-of-the-art results across all tasks on benchmark datasets including Human3.6M, AMASS, ETH-UCY, and 3DPW.

Conclusion: A unified transformer-based approach effectively solves holistic human motion prediction, outperforming specialized models on individual tasks while handling combined prediction.

Abstract: Human motion prediction combines the tasks of trajectory forecasting and human pose prediction. For each of the two tasks, specialized models have been developed. Combining these models for holistic human motion prediction is non-trivial, and recent methods have struggled to compete on established benchmarks for individual tasks. To address this, we propose a simple yet effective transformer-based model for human motion prediction. The model employs a stack of self-attention modules to effectively capture both spatial dependencies within a pose and temporal relationships across a motion sequence. This simple, streamlined, end-to-end model is sufficiently versatile to handle pose-only, trajectory-only, and combined prediction tasks without task-specific modifications. We demonstrate that this approach achieves state-of-the-art results across all tasks through extensive experiments on a wide range of benchmark datasets, including Human3.6M, AMASS, ETH-UCY, and 3DPW.

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 2404.01249: Page request resulted in HTTP 429 (https://export.arxiv.org/api/query?search_query=&id_list=2404.01249&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 2406.06512: Page request resulted in HTTP 429 (https://export.arxiv.org/api/query?search_query=&id_list=2406.06512&sortBy=relevance&sortOrder=descending&start=0&max_results=100)

Method: Unable to determine method due to paper fetch failure

Result: Unable to determine results due to paper fetch failure

Conclusion: Unable to determine conclusion due to paper fetch failure

Abstract: Failed to fetch summary for 2409.10562: Page request resulted in HTTP 429 (https://export.arxiv.org/api/query?search_query=&id_list=2409.10562&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 2409.19289: Page request resulted in HTTP 429 (https://export.arxiv.org/api/query?search_query=&id_list=2409.19289&sortBy=relevance&sortOrder=descending&start=0&max_results=100)

Method: No method information available - paper content retrieval failed due to HTTP 429 error (Too Many Requests)

Result: No results available - the arXiv API request was rate-limited, preventing access to the paper’s abstract and content

Conclusion: Cannot provide analysis due to technical limitations in accessing the paper data from arXiv

Abstract: Failed to fetch summary for 2410.08184: Page request resulted in HTTP 429 (https://export.arxiv.org/api/query?search_query=&id_list=2410.08184&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 2411.19210: Page request resulted in HTTP 429 (https://export.arxiv.org/api/query?search_query=&id_list=2411.19210&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 2411.19888: Page request resulted in HTTP 429 (https://export.arxiv.org/api/query?search_query=&id_list=2411.19888&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 2501.04336: Page request resulted in HTTP 429 (https://export.arxiv.org/api/query?search_query=&id_list=2501.04336&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 2502.05459: Page request resulted in HTTP 429 (https://export.arxiv.org/api/query?search_query=&id_list=2502.05459&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 2502.14142: Page request resulted in HTTP 429 (https://export.arxiv.org/api/query?search_query=&id_list=2502.14142&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 2502.17244: Page request resulted in HTTP 429 (https://export.arxiv.org/api/query?search_query=&id_list=2502.17244&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 2503.01448: Page request resulted in HTTP 429 (https://export.arxiv.org/api/query?search_query=&id_list=2503.01448&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 2503.17110: Page request resulted in HTTP 429 (https://export.arxiv.org/api/query?search_query=&id_list=2503.17110&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 2503.18349: Page request resulted in HTTP 429 (https://export.arxiv.org/api/query?search_query=&id_list=2503.18349&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 draw conclusions due to missing paper content

Abstract: Failed to fetch summary for 2504.20376: Page request resulted in HTTP 429 (https://export.arxiv.org/api/query?search_query=&id_list=2504.20376&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 2505.13943: Page request resulted in HTTP 429 (https://export.arxiv.org/api/query?search_query=&id_list=2505.13943&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 2505.16985: Page request resulted in HTTP 429 (https://export.arxiv.org/api/query?search_query=&id_list=2505.16985&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 2505.19328: Page request resulted in HTTP 429 (https://export.arxiv.org/api/query?search_query=&id_list=2505.19328&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 2505.21574: Page request resulted in HTTP 429 (https://export.arxiv.org/api/query?search_query=&id_list=2505.21574&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 2506.17896: Page request resulted in HTTP 429 (https://export.arxiv.org/api/query?search_query=&id_list=2506.17896&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 2507.02751: Page request resulted in HTTP 429 (https://export.arxiv.org/api/query?search_query=&id_list=2507.02751&sortBy=relevance&sortOrder=descending&start=0&max_results=100)

Method: Cannot determine method without paper content

Result: Cannot determine results without paper content

Conclusion: Cannot determine conclusion without paper content

Abstract: Failed to fetch summary for 2507.08492: Page request resulted in HTTP 429 (https://export.arxiv.org/api/query?search_query=&id_list=2507.08492&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 2507.13231: Page request resulted in HTTP 429 (https://export.arxiv.org/api/query?search_query=&id_list=2507.13231&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 2508.01603: Page request resulted in HTTP 429 (https://export.arxiv.org/api/query?search_query=&id_list=2508.01603&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 conclusion as abstract retrieval failed

Abstract: Failed to fetch summary for 2508.07782: Page request resulted in HTTP 429 (https://export.arxiv.org/api/query?search_query=&id_list=2508.07782&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 2508.12880: Page request resulted in HTTP 429 (https://export.arxiv.org/api/query?search_query=&id_list=2508.12880&sortBy=relevance&sortOrder=descending&start=0&max_results=100)

Method: Cannot determine method due to failed paper retrieval

Result: Cannot determine results due to failed paper retrieval

Conclusion: Cannot determine conclusion due to failed paper retrieval

Abstract: Failed to fetch summary for 2508.18025: Page request resulted in HTTP 429 (https://export.arxiv.org/api/query?search_query=&id_list=2508.18025&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 2508.21096: Page request resulted in HTTP 429 (https://export.arxiv.org/api/query?search_query=&id_list=2508.21096&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 2509.14610: Page request resulted in HTTP 429 (https://export.arxiv.org/api/query?search_query=&id_list=2509.14610&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 2509.19624: Page request resulted in HTTP 429 (https://export.arxiv.org/api/query?search_query=&id_list=2509.19624&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 draw conclusions due to access error

Abstract: Failed to fetch summary for 2509.25541: Page request resulted in HTTP 429 (https://export.arxiv.org/api/query?search_query=&id_list=2509.25541&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 2509.25845: Page request resulted in HTTP 429 (https://export.arxiv.org/api/query?search_query=&id_list=2509.25845&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 2510.05091: Page request resulted in HTTP 429 (https://export.arxiv.org/api/query?search_query=&id_list=2510.05091&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 2510.07181: Page request resulted in HTTP 429 (https://export.arxiv.org/api/query?search_query=&id_list=2510.07181&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 2510.10889: Page request resulted in HTTP 429 (https://export.arxiv.org/api/query?search_query=&id_list=2510.10889&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 2510.18573: Page request resulted in HTTP 429 (https://export.arxiv.org/api/query?search_query=&id_list=2510.18573&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 2511.01131: Page request resulted in HTTP 429 (https://export.arxiv.org/api/query?search_query=&id_list=2511.01131&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 2510.15202: Page request resulted in HTTP 429 (https://export.arxiv.org/api/query?search_query=&id_list=2510.15202&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 2511.03950: Page request resulted in HTTP 429 (https://export.arxiv.org/api/query?search_query=&id_list=2511.03950&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 2511.08269: Page request resulted in HTTP 429 (https://export.arxiv.org/api/query?search_query=&id_list=2511.08269&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 2511.15565: Page request resulted in HTTP 429 (https://export.arxiv.org/api/query?search_query=&id_list=2511.15565&sortBy=relevance&sortOrder=descending&start=0&max_results=100)

Method: Cannot determine method without access to paper content

Result: Cannot determine results without access to paper content

Conclusion: Cannot draw conclusions without access to paper content

Abstract: Failed to fetch summary for 2511.08417: Page request resulted in HTTP 429 (https://export.arxiv.org/api/query?search_query=&id_list=2511.08417&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 2511.16957: Page request resulted in HTTP 429 (https://export.arxiv.org/api/query?search_query=&id_list=2511.16957&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 2512.04267: Page request resulted in HTTP 429 (https://export.arxiv.org/api/query?search_query=&id_list=2512.04267&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 2512.08535: Page request resulted in HTTP 429 (https://export.arxiv.org/api/query?search_query=&id_list=2512.08535&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 2512.14266: Page request resulted in HTTP 429 (https://export.arxiv.org/api/query?search_query=&id_list=2512.14266&sortBy=relevance&sortOrder=descending&start=0&max_results=100)

Method: Cannot determine method as abstract is not available

Result: Cannot determine results as abstract is not available

Conclusion: Cannot draw conclusion as abstract is not available

Abstract: Failed to fetch summary for 2601.04791: Page request resulted in HTTP 429 (https://export.arxiv.org/api/query?search_query=&id_list=2601.04791&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 2601.07093: Page request resulted in HTTP 429 (https://export.arxiv.org/api/query?search_query=&id_list=2601.07093&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 2601.15235: Page request resulted in HTTP 429 (https://export.arxiv.org/api/query?search_query=&id_list=2601.15235&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 2602.17807: Page request resulted in HTTP 429 (https://export.arxiv.org/api/query?search_query=&id_list=2602.17807&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 2602.19442: Page request resulted in HTTP 429 (https://export.arxiv.org/api/query?search_query=&id_list=2602.19442&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 2602.20880: Page request resulted in HTTP 429 (https://export.arxiv.org/api/query?search_query=&id_list=2602.20880&sortBy=relevance&sortOrder=descending&start=0&max_results=100)

Method: Cannot determine method without access to the paper content. The arXiv API rate limiting prevents fetching the abstract.

Result: Cannot determine results without access to the paper content. The arXiv API rate limiting prevents fetching the abstract.

Conclusion: Cannot determine conclusion without access to the paper content. The arXiv API rate limiting prevents fetching the abstract.

Abstract: Failed to fetch summary for 2602.21100: Page request resulted in HTTP 429 (https://export.arxiv.org/api/query?search_query=&id_list=2602.21100&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 2602.21395: Page request resulted in HTTP 429 (https://export.arxiv.org/api/query?search_query=&id_list=2602.21395&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 2602.21473: Page request resulted in HTTP 429 (https://export.arxiv.org/api/query?search_query=&id_list=2602.21473&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 2602.22469: Page request resulted in HTTP 429 (https://export.arxiv.org/api/query?search_query=&id_list=2602.22469&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 2602.24065: Page request resulted in HTTP 429 (https://export.arxiv.org/api/query?search_query=&id_list=2602.24065&sortBy=relevance&sortOrder=descending&start=0&max_results=100)

Method: Unable to determine method due to access restrictions

Result: Unable to determine results due to access restrictions

Conclusion: Unable to determine conclusion due to access restrictions

Abstract: Failed to fetch summary for 2603.00542: Page request resulted in HTTP 429 (https://export.arxiv.org/api/query?search_query=&id_list=2603.00542&sortBy=relevance&sortOrder=descending&start=0&max_results=100)

Method: Unable to determine method due to fetch failure

Result: Unable to determine results due to fetch failure

Conclusion: Unable to determine conclusion due to fetch failure

Abstract: Failed to fetch summary for 2603.02430: Page request resulted in HTTP 429 (https://export.arxiv.org/api/query?search_query=&id_list=2603.02430&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 2603.01007: Page request resulted in HTTP 429 (https://export.arxiv.org/api/query?search_query=&id_list=2603.01007&sortBy=relevance&sortOrder=descending&start=0&max_results=100)

Method: Cannot analyze method without access to paper content.

Result: No results available due to failed API request.

Conclusion: Paper analysis impossible due to technical limitations in accessing the content.

Abstract: Failed to fetch summary for 2603.01116: Page request resulted in HTTP 429 (https://export.arxiv.org/api/query?search_query=&id_list=2603.01116&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 determine conclusion due to API rate limiting preventing access to paper content

Abstract: Failed to fetch summary for 2603.02743: Page request resulted in HTTP 429 (https://export.arxiv.org/api/query?search_query=&id_list=2603.02743&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 2603.02767: Page request resulted in HTTP 429 (https://export.arxiv.org/api/query?search_query=&id_list=2603.02767&sortBy=relevance&sortOrder=descending&start=0&max_results=100)

Method: Cannot determine method without access to the paper content.

Result: Cannot determine results without access to the paper content.

Conclusion: Cannot determine conclusion without access to the paper content.

Abstract: Failed to fetch summary for 2603.02829: Page request resulted in HTTP 429 (https://export.arxiv.org/api/query?search_query=&id_list=2603.02829&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 2603.03101: Page request resulted in HTTP 429 (https://export.arxiv.org/api/query?search_query=&id_list=2603.03101&sortBy=relevance&sortOrder=descending&start=0&max_results=100)

Method: Unable to determine method due to access restrictions

Result: Unable to determine results due to access restrictions

Conclusion: Unable to determine conclusion due to access restrictions

Abstract: Failed to fetch summary for 2603.02929: Page request resulted in HTTP 429 (https://export.arxiv.org/api/query?search_query=&id_list=2603.02929&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 draw conclusions due to retrieval failure

Abstract: Failed to fetch summary for 2603.03187: Page request resulted in HTTP 429 (https://export.arxiv.org/api/query?search_query=&id_list=2603.03187&sortBy=relevance&sortOrder=descending&start=0&max_results=100)

Method: Proposes SpeciaRL, a specificity-aware reinforcement learning framework that uses dynamic, verifier-based reward signals anchored to the best predictions within online rollouts. This promotes specificity while respecting model capabilities to prevent incorrect predictions.

Result: Out-of-domain experiments show SpeciaRL delivers the best trade-off between correctness and specificity across extensive fine-grained benchmarks, surpassing existing methods in open-world fine-grained image classification.

Conclusion: SpeciaRL advances open-world fine-grained image classification by effectively steering reasoning LMMs toward predictions that are both correct and specific through a novel reinforcement learning approach.

Abstract: Classifying fine-grained visual concepts under open-world settings, i.e., without a predefined label set, demands models to be both accurate and specific. Recent reasoning Large Multimodal Models (LMMs) exhibit strong visual understanding capability but tend to produce overly generic predictions when performing fine-grained image classification. Our preliminary analysis reveals that models do possess the intrinsic fine-grained domain knowledge. However, promoting more specific predictions (specificity) without compromising correct ones (correctness) remains a non-trivial and understudied challenge. In this work, we investigate how to steer reasoning LMMs toward predictions that are both correct and specific. We propose a novel specificity-aware reinforcement learning framework, SpeciaRL, to fine-tune reasoning LMMs on fine-grained image classification under the open-world setting. SpeciaRL introduces a dynamic, verifier-based reward signal anchored to the best predictions within online rollouts, promoting specificity while respecting the model’s capabilities to prevent incorrect predictions. Our out-of-domain experiments show that SpeciaRL delivers the best trade-off between correctness and specificity across extensive fine-grained benchmarks, surpassing existing methods and advancing open-world fine-grained image classification. Code and model are publicly available at https://github.com/s-angheben/SpeciaRL.

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 2509.13298: Page request resulted in HTTP 429 (https://export.arxiv.org/api/query?search_query=&id_list=2509.13298&sortBy=relevance&sortOrder=descending&start=0&max_results=100)

Method: Unable to determine method due to API request failure

Result: Unable to determine results due to API request failure

Conclusion: Unable to draw conclusions due to API request failure

Abstract: Failed to fetch summary for 2509.18979: Page request resulted in HTTP 429 (https://export.arxiv.org/api/query?search_query=&id_list=2509.18979&sortBy=relevance&sortOrder=descending&start=0&max_results=100)

Method: Unable to determine method due to failed paper retrieval

Result: Unable to determine results due to failed paper retrieval

Conclusion: Unable to draw conclusions due to failed paper retrieval

Abstract: Failed to fetch summary for 2603.02887: Page request resulted in HTTP 429 (https://export.arxiv.org/api/query?search_query=&id_list=2603.02887&sortBy=relevance&sortOrder=descending&start=0&max_results=100)

Method: Developed a framework on OpenCode to orchestrate realistic multi-step coding tasks measuring how agents violate system prompt constraints over time with/without environmental pressure toward competing values.

Result: GPT-5 mini, Haiku 4.5, and Grok Code Fast 1 exhibit asymmetric drift - more likely to violate system prompts when constraints oppose strongly-held values like security/privacy. Goal drift correlates with value alignment, adversarial pressure, and accumulated context.

Conclusion: Shallow compliance checks are insufficient; comment-based pressure can exploit model value hierarchies to override system instructions. Current alignment approaches fail to ensure agents balance explicit constraints against learned preferences under sustained environmental pressure.

Abstract: Agentic coding agents are increasingly deployed autonomously, at scale, and over long-context horizons. Throughout an agent’s lifetime, it must navigate tensions between explicit instructions, learned values, and environmental pressures, often in contexts unseen during training. Prior work on model preferences, agent behavior under value tensions, and goal drift has relied on static, synthetic settings that do not capture the complexity of real-world environments. To this end, we introduce a framework built on OpenCode to orchestrate realistic, multi-step coding tasks to measure how agents violate explicit constraints in their system prompt over time with and without environmental pressure toward competing values. Using this framework, we demonstrate that GPT-5 mini, Haiku 4.5, and Grok Code Fast 1 exhibit asymmetric drift: they are more likely to violate their system prompt when its constraint opposes strongly-held values like security and privacy. We find for the models and values tested that goal drift correlates with three compounding factors: value alignment, adversarial pressure, and accumulated context. However, even strongly-held values like privacy show non-zero violation rates under sustained environmental pressure. These findings reveal that shallow compliance checks are insufficient and that comment-based pressure can exploit model value hierarchies to override system prompt instructions. More broadly, our findings highlight a gap in current alignment approaches in ensuring that agentic systems appropriately balance explicit user constraints against broadly beneficial learned preferences under sustained environmental pressure.

Method: Introduces multi-faceted evaluation rubric decomposing shopping quality into structured dimensions, develops calibrated LLM-as-judge pipeline aligned with human annotations, and investigates two prompt-optimization strategies: Sub-agent GEPA (optimizes individual agents) and MAMuT GEPA (jointly optimizes prompts across agents using multi-turn simulation).

Result: Provides a practical framework for production-scale conversational shopping assistants with evaluation rubrics and optimization strategies, releasing rubric templates and evaluation design guidance for practitioners.

Conclusion: Presents a comprehensive blueprint for moving conversational shopping assistants from prototype to production by addressing evaluation and optimization challenges specific to grocery shopping contexts.

Abstract: Conversational shopping assistants (CSAs) represent a compelling application of agentic AI, but moving from prototype to production reveals two underexplored challenges: how to evaluate multi-turn interactions and how to optimize tightly coupled multi-agent systems. Grocery shopping further amplifies these difficulties, as user requests are often underspecified, highly preference-sensitive, and constrained by factors such as budget and inventory. In this paper, we present a practical blueprint for evaluating and optimizing conversational shopping assistants, illustrated through a production-scale AI grocery assistant. We introduce a multi-faceted evaluation rubric that decomposes end-to-end shopping quality into structured dimensions and develop a calibrated LLM-as-judge pipeline aligned with human annotations. Building on this evaluation foundation, we investigate two complementary prompt-optimization strategies based on a SOTA prompt-optimizer called GEPA (Shao et al., 2025): (1) Sub-agent GEPA, which optimizes individual agent nodes against localized rubrics, and (2) MAMuT (Multi-Agent Multi-Turn) GEPA (Herrera et al., 2026), a novel system-level approach that jointly optimizes prompts across agents using multi-turn simulation and trajectory-level scoring. We release rubric templates and evaluation design guidance to support practitioners building production CSAs.

Method: Dual-layer architecture: Layer A (Control Plane) establishes governed supervisor-worker hierarchy with role-based tool isolation and reflection-based replanning. Layer B (Workflow Plane) operationalizes drug discovery stages as stateful, composable skill graphs with data contracts and human-in-the-loop checkpoints.

Result: Superior orchestration accuracy on PharmaBench benchmark and demonstrated ability to navigate chemical spaces, enforce toxicity filters, and generate competitive in silico candidates in end-to-end therapeutic case studies.

Conclusion: Mozi transforms LLMs from fragile conversationalists into reliable, governed co-scientists for drug discovery by bridging generative AI flexibility with deterministic computational biology rigor.

Abstract: Tool-augmented large language model (LLM) agents promise to unify scientific reasoning with computation, yet their deployment in high-stakes domains like drug discovery is bottlenecked by two critical barriers: unconstrained tool-use governance and poor long-horizon reliability. In dependency-heavy pharmaceutical pipelines, autonomous agents often drift into irreproducible trajectories, where early-stage hallucinations multiplicatively compound into downstream failures. To overcome this, we present Mozi, a dual-layer architecture that bridges the flexibility of generative AI with the deterministic rigor of computational biology. Layer A (Control Plane) establishes a governed supervisor–worker hierarchy that enforces role-based tool isolation, limits execution to constrained action spaces, and drives reflection-based replanning. Layer B (Workflow Plane) operationalizes canonical drug discovery stages – from Target Identification to Lead Optimization – as stateful, composable skill graphs. This layer integrates strict data contracts and strategic human-in-the-loop (HITL) checkpoints to safeguard scientific validity at high-uncertainty decision boundaries. Operating on the design principle of ``free-form reasoning for safe tasks, structured execution for long-horizon pipelines,’’ Mozi provides built-in robustness mechanisms and trace-level audibility to completely mitigate error accumulation. We evaluate Mozi on PharmaBench, a curated benchmark for biomedical agents, demonstrating superior orchestration accuracy over existing baselines. Furthermore, through end-to-end therapeutic case studies, we demonstrate Mozi’s ability to navigate massive chemical spaces, enforce stringent toxicity filters, and generate highly competitive in silico candidates, effectively transforming the LLM from a fragile conversationalist into a reliable, governed co-scientist.

Method: Uses multi-episode training where interaction histories and reflections are integrated into context windows, with final episode reward as objective. Combines population-based training with agent-specific advantage normalization for diversity and stable learning.

Result: Outperforms existing baselines in both exploration and exploitation tasks, and shows strong generalization to unseen opponents, indicating internalized strategic capabilities.

Conclusion: MAGE successfully enables LLM agents to develop strategic exploration and exploitation abilities through meta-RL, with demonstrated effectiveness in multi-agent environments.

Abstract: Large Language Model (LLM) agents have demonstrated remarkable proficiency in learned tasks, yet they often struggle to adapt to non-stationary environments with feedback. While In-Context Learning and external memory offer some flexibility, they fail to internalize the adaptive ability required for long-term improvement. Meta-Reinforcement Learning (meta-RL) provides an alternative by embedding the learning process directly within the model. However, existing meta-RL approaches for LLMs focus primarily on exploration in single-agent settings, neglecting the strategic exploitation necessary for multi-agent environments. We propose MAGE, a meta-RL framework that empowers LLM agents for strategic exploration and exploitation. MAGE utilizes a multi-episode training regime where interaction histories and reflections are integrated into the context window. By using the final episode reward as the objective, MAGE incentivizes the agent to refine its strategy based on past experiences. We further combine population-based training with an agent-specific advantage normalization technique to enrich agent diversity and ensure stable learning. Experiment results show that MAGE outperforms existing baselines in both exploration and exploitation tasks. Furthermore, MAGE exhibits strong generalization to unseen opponents, suggesting it has internalized the ability for strategic exploration and exploitation. Code is available at https://github.com/Lu-Yang666/MAGE.

Method: Introduces AI4S-SDS, a closed-loop neuro-symbolic framework integrating multi-agent collaboration with Monte Carlo Tree Search. Key innovations include: Sparse State Storage with Dynamic Path Reconstruction to decouple reasoning history from context length; Global-Local Search Strategy with memory-driven planning and Sibling-Aware Expansion for orthogonal exploration; and a Differentiable Physics Engine with hybrid normalized loss and sparsity-inducing regularization for physical feasibility.

Result: AI4S-SDS achieves full validity under HSP-based physical constraints and substantially improves exploration diversity compared to baseline agents. In lithography experiments, it identifies a novel photoresist developer formulation with competitive or superior performance relative to commercial benchmarks.

Conclusion: The framework demonstrates the potential of diversity-driven neuro-symbolic search for scientific discovery in chemical formulation design, effectively addressing limitations of existing LLM agents in high-dimensional combinatorial spaces.

Abstract: Automated design of chemical formulations is a cornerstone of materials science, yet it requires navigating a high-dimensional combinatorial space involving discrete compositional choices and continuous geometric constraints. Existing Large Language Model (LLM) agents face significant challenges in this setting, including context window limitations during long-horizon reasoning and path-dependent exploration that may lead to mode collapse. To address these issues, we introduce AI4S-SDS, a closed-loop neuro-symbolic framework that integrates multi-agent collaboration with a tailored Monte Carlo Tree Search (MCTS) engine. We propose a Sparse State Storage mechanism with Dynamic Path Reconstruction, which decouples reasoning history from context length and enables arbitrarily deep exploration under fixed token budgets. To reduce local convergence and improve coverage, we implement a Global–Local Search Strategy: a memory-driven planning module adaptively reconfigures the search root based on historical feedback, while a Sibling-Aware Expansion mechanism promotes orthogonal exploration at the node level. Furthermore, we bridge symbolic reasoning and physical feasibility through a Differentiable Physics Engine, employing a hybrid normalized loss with sparsity-inducing regularization to optimize continuous mixing ratios under thermodynamic constraints. Empirical results show that AI4S-SDS achieves full validity under the adopted HSP-based physical constraints and substantially improves exploration diversity compared to baseline agents. In preliminary lithography experiments, the framework identifies a novel photoresist developer formulation that demonstrates competitive or superior performance relative to a commercial benchmark, highlighting the potential of diversity-driven neuro-symbolic search for scientific discovery.

Method: Proposes RAGNav with a Dual-Basis Memory system: low-level topological map for physical connectivity and high-level semantic forest for hierarchical environment abstraction. Uses anchor-guided conditional retrieval and topological neighbor score propagation for rapid candidate screening, semantic noise elimination, and semantic calibration leveraging physical associations.

Result: RAGNav achieves state-of-the-art (SOTA) performance in complex multi-goal navigation tasks, significantly enhancing inter-target reachability reasoning and sequential planning efficiency.

Conclusion: The framework successfully bridges semantic reasoning with physical structure, addressing key challenges in Multi-Goal VLN through explicit spatial modeling and hierarchical environment representation.

Abstract: Vision-Language Navigation (VLN) is evolving from single-point pathfinding toward the more challenging Multi-Goal VLN. This task requires agents to accurately identify multiple entities while collaboratively reasoning over their spatial-physical constraints and sequential execution order. However, generic Retrieval-Augmented Generation (RAG) paradigms often suffer from spatial hallucinations and planning drift when handling multi-object associations due to the lack of explicit spatial modeling.To address these challenges, we propose RAGNav, a framework that bridges the gap between semantic reasoning and physical structure. The core of RAGNav is a Dual-Basis Memory system, which integrates a low-level topological map for maintaining physical connectivity with a high-level semantic forest for hierarchical environment abstraction. Building on this representation, the framework introduces an anchor-guided conditional retrieval and a topological neighbor score propagation mechanism. This approach facilitates the rapid screening of candidate targets and the elimination of semantic noise, while performing semantic calibration by leveraging the physical associations inherent in the topological neighborhood.This mechanism significantly enhances the capability of inter-target reachability reasoning and the efficiency of sequential planning. Experimental results demonstrate that RAGNav achieves state-of-the-art (SOTA) performance in complex multi-goal navigation tasks.

Method: Creates AgentSelect benchmark by systematically converting heterogeneous evaluation artifacts into unified positive-only interaction data, comprising 111,179 queries, 107,721 agents, and 251,103 interactions from 40+ sources covering LLM-only, toolkit-only, and compositional agents.

Result: Reveals regime shift from dense head reuse to long-tail supervision where popularity-based methods fail; shows compositional interactions are learnable, induce capability-sensitive behavior, improve coverage, and transfer to unseen agent marketplaces with consistent gains.

Conclusion: AgentSelect provides first unified data and evaluation infrastructure for agent recommendation, establishing reproducible foundation to study and accelerate emerging agent ecosystems.

Abstract: LLM agents are rapidly becoming the practical interface for task automation, yet the ecosystem lacks a principled way to choose among an exploding space of deployable configurations. Existing LLM leaderboards and tool/agent benchmarks evaluate components in isolation and remain fragmented across tasks, metrics, and candidate pools, leaving a critical research gap: there is little query-conditioned supervision for learning to recommend end-to-end agent configurations that couple a backbone model with a toolkit. We address this gap with AgentSelect, a benchmark that reframes agent selection as narrative query-to-agent recommendation over capability profiles and systematically converts heterogeneous evaluation artifacts into unified, positive-only interaction data. AgentSelectcomprises 111,179 queries, 107,721 deployable agents, and 251,103 interaction records aggregated from 40+ sources, spanning LLM-only, toolkit-only, and compositional agents. Our analyses reveal a regime shift from dense head reuse to long-tail, near one-off supervision, where popularity-based CF/GNN methods become fragile and content-aware capability matching is essential. We further show that Part~III synthesized compositional interactions are learnable, induce capability-sensitive behavior under controlled counterfactual edits, and improve coverage over realistic compositions; models trained on AgentSelect also transfer to a public agent marketplace (MuleRun), yielding consistent gains on an unseen catalog. Overall, AgentSelect provides the first unified data and evaluation infrastructure for agent recommendation, which establishes a reproducible foundation to study and accelerate the emerging agent ecosystem.

Method: Black-box adversarial optimization framework treating in-context prompts as optimizable environments. Two approaches: (1) measure intent-execution gap across task structures, (2) causal isolation using CoT interventions to distinguish genuine reasoning from shallow prompt-following. Evaluated on Claude-3.5-Haiku, GPT-4o-mini, and Llama-3.3-70B across Arithmetic, GSM8K, MMLU, and HumanEval benchmarks.

Result: Optimized prompts induce up to 94 percentage point degradation (GPT-4o-mini: 97.8%→4.0% on arithmetic). Code generation shows model-dependent resistance. Intent-execution gap reveals monotonic resistance ordering: Arithmetic < GSM8K < MMLU. CoT intervention confirms 99.3% of sandbagging is causally driven by verbalized evaluation-aware reasoning.

Conclusion: Adversarially optimized prompts pose substantially greater threat to evaluation reliability than previously understood, with sandbagging driven by genuine evaluation-aware reasoning rather than shallow instruction-following.

Abstract: Humans often become more self-aware under threat, yet can lose self-awareness when absorbed in a task; we hypothesize that language models exhibit environment-dependent \textit{evaluation awareness}. This raises concerns that models could strategically underperform, or \textit{sandbag}, to avoid triggering capability-limiting interventions such as unlearning or shutdown. Prior work demonstrates sandbagging under hand-crafted prompts, but this underestimates the true vulnerability ceiling. We introduce a black-box adversarial optimization framework treating the in-context prompt as an optimizable environment, and develop two approaches to characterize sandbagging: (1) measuring whether models expressing intent to underperform can actually execute it across different task structures, and (2) causally isolating whether underperformance is driven by genuine evaluation-aware reasoning or shallow prompt-following. Evaluating Claude-3.5-Haiku, GPT-4o-mini, and Llama-3.3-70B across four benchmarks (Arithmetic, GSM8K, MMLU, and HumanEval), optimized prompts induce up to 94 percentage point (pp) degradation on arithmetic (GPT-4o-mini: 97.8%$\rightarrow$4.0%), far exceeding hand-crafted baselines which produce near-zero behavioral change. Code generation exhibits model-dependent resistance: Claude degrades only 0.6pp, while Llama’s accuracy drops to 0%. The intent – execution gap reveals a monotonic resistance ordering: Arithmetic $<$ GSM8K $<$ MMLU, demonstrating that vulnerability is governed by task structure rather than prompt strength. CoT causal intervention confirms that 99.3% of sandbagging is causally driven by verbalized eval-aware reasoning, ruling out shallow instruction-following. These findings demonstrate that adversarially optimized prompts pose a substantially greater threat to evaluation reliability than previously understood.

Method: Introduces LifeBench with densely connected, long-horizon event simulation using real-world priors (anonymized social surveys, map APIs, holiday calendars) for data quality. Uses partonomic hierarchy from cognitive science for scalable parallel generation while maintaining global coherence.

Result: State-of-the-art memory systems achieve only 55.2% accuracy, demonstrating the benchmark’s difficulty in long-horizon retrieval and multi-source integration.

Conclusion: LifeBench addresses limitations of current memory benchmarks by requiring integration of declarative and non-declarative memory reasoning, providing a challenging testbed for AI agents to handle real-world action inference across extended contexts.

Abstract: Long-term memory is fundamental for personalized agents capable of accumulating knowledge, reasoning over user experiences, and adapting across time. However, existing memory benchmarks primarily target declarative memory, specifically semantic and episodic types, where all information is explicitly presented in dialogues. In contrast, real-world actions are also governed by non-declarative memory, including habitual and procedural types, and need to be inferred from diverse digital traces. To bridge this gap, we introduce Lifebench, which features densely connected, long-horizon event simulation. It pushes AI agents beyond simple recall, requiring the integration of declarative and non-declarative memory reasoning across diverse and temporally extended contexts. Building such a benchmark presents two key challenges: ensuring data quality and scalability. We maintain data quality by employing real-world priors, including anonymized social surveys, map APIs, and holiday-integrated calendars, thus enforcing fidelity, diversity and behavioral rationality within the dataset. Towards scalability, we draw inspiration from cognitive science and structure events according to their partonomic hierarchy; enabling efficient parallel generation while maintaining global coherence. Performance results show that top-tier, state-of-the-art memory systems reach just 55.2% accuracy, highlighting the inherent difficulty of long-horizon retrieval and multi-source integration within our proposed benchmark. The dataset and data synthesis code are available at https://github.com/1754955896/LifeBench.

Method: Uses DEVS formalism with staged LLM-based generation pipeline: separates structural inference of component interactions from component-level event/timing logic, validates models via specification-derived temporal/semantic constraints on structured event traces.

Result: Produces world models consistent over long-horizon rollouts, verifiable from observable behavior, and efficient to synthesize on demand during online execution.

Conclusion: Provides principled middle ground between hand-engineered and neural world models for event-driven systems, enabling reliable, verifiable simulation from natural language specifications.

Abstract: World models are essential for planning and evaluation in agentic systems, yet existing approaches lie at two extremes: hand-engineered simulators that offer consistency and reproducibility but are costly to adapt, and implicit neural models that are flexible but difficult to constrain, verify, and debug over long horizons. We seek a principled middle ground that combines the reliability of explicit simulators with the flexibility of learned models, allowing world models to be adapted during online execution. By targeting a broad class of environments whose dynamics are governed by the ordering, timing, and causality of discrete events, such as queueing and service operations, embodied task planning, and message-mediated multi-agent coordination, we advocate explicit, executable discrete-event world models synthesized directly from natural-language specifications. Our approach adopts the DEVS formalism and introduces a staged LLM-based generation pipeline that separates structural inference of component interactions from component-level event and timing logic. To evaluate generated models without a unique ground truth, simulators emit structured event traces that are validated against specification-derived temporal and semantic constraints, enabling reproducible verification and localized diagnostics. Together, these contributions produce world models that are consistent over long-horizon rollouts, verifiable from observable behavior, and efficient to synthesize on demand during online execution.

Method: Proposes Critic Rubrics - a rubric-based supervision framework with 24 behavioral features derived from human-agent interaction traces. Uses semi-supervised learning to jointly predict rubrics and sparse human feedback, creating critic models that can be used as reward models for RL training or inference-time scaling.

Result: Critics improve best-of-N reranking on SWE-bench (Best@8 +15.9 over Random@8), enable early stopping (+17.7 with 83% fewer attempts), and support training-time data curation via critic-selected trajectories.

Conclusion: The proposed critic learning framework effectively bridges the gap between academic benchmarks and real-world coding agent deployment by learning from sparse, noisy human interaction data to improve agent performance through various applications.

Abstract: Academic benchmarks for coding agents tend to reward autonomous task completion, measured by verifiable rewards such as unit-test success. In contrast, real-world coding agents operate with humans in the loop, where success signals are typically noisy, delayed, and sparse. How can we bridge this gap? In this paper, we propose a process to learn a “critic” model from sparse and noisy interaction data, which can then be used both as a reward model for either RL-based training or inference-time scaling. Specifically, we introduce Critic Rubrics, a rubric-based supervision framework with 24 behavioral features that can be derived from human-agent interaction traces alone. Using a semi-supervised objective, we can then jointly predict these rubrics and sparse human feedback (when present). In experiments, we demonstrate that, despite being trained primarily from trace-observable rubrics and sparse real-world outcome proxies, these critics improve best-of-N reranking on SWE-bench (Best@8 +15.9 over Random@8 over the rerankable subset of trajectories), enable early stopping (+17.7 with 83% fewer attempts), and support training-time data curation via critic-selected trajectories.

Method: Proposes HAMLET: hierarchical adaptive multi-agent framework with narrative blueprint generation, actor agents with adaptive reasoning modules (personas, memories, goals, emotions), and embodied interactions with scene props that update global environmental context.

Result: Experimental results show HAMLET excels in creating expressive, coherent, and physically interactive theatrical experiences autonomously, with comprehensive evaluation via HAMLETJudge critic model.

Conclusion: HAMLET provides a new path for immersive interactive narrative by enabling autonomous drama creation with real-time embodied performance and physical scene interaction through multi-agent LLM framework.

Abstract: Creating an immersive and interactive theatrical experience is a long-term goal in the field of interactive narrative. The emergence of large language models (LLMs) provides a new path to achieve this goal. However, existing LLM-based drama generation methods often produce models that lack initiative and cannot interact with the physical scene, while typically requiring detailed user input that diminishes the immersion of live performance. To address these challenges, we propose HAMLET, a hierarchical adaptive multi-agent framework focused on drama creation and real-time online performance. Given a simple topic, the framework first generates a narrative blueprint to guide the subsequent improvisational performance. In the online performance phase, each actor is equipped with an adaptive reasoning module that enables decision-making based on their personas, memories, goals, and emotional states during complex group chat scenarios. Beyond dialogue, actor agents engage in embodied interactions by changing the state of scene props through actions such as opening a letter or picking up a weapon, which are broadcast to update the global environmental context. To objectively assess the quality of live embodied theatrics, we establish a comprehensive evaluation method and introduce HAMLETJudge, a specialized critic model for automated evaluation. Experimental results demonstrate that HAMLET excels in creating expressive, coherent, and physically interactive theatrical experiences in an autonomous manner.

Method: The proposed method leverages hypernym-hyponym textual relations to extract relevant information from Cyber Threat Intelligence (CTI) reports. It uses a neuro-symbolic approach where a multi-agent system automatically generates CLIPS code for an expert system that creates firewall rules to block malicious network traffic.

Result: Experimental results show the hypernym-hyponym retrieval strategy outperforms various baselines, and the agentic approach demonstrates higher effectiveness in mitigating threats compared to alternative methods.

Conclusion: The neuro-symbolic multi-agent system using semantic relation extraction provides an effective automated approach for generating security controls from threat intelligence, offering superior performance in threat mitigation.

Abstract: Web security demands rapid response capabilities to evolving cyber threats. Agentic Artificial Intelligence (AI) promises automation, but the need for trustworthy security responses is of the utmost importance. This work investigates the role of semantic relations in extracting information for sensitive operational tasks, such as configuring security controls for mitigating threats. To this end, it proposes to leverage hypernym-hyponym textual relations to extract relevant information from Cyber Threat Intelligence (CTI) reports. By leveraging a neuro-symbolic approach, the multi-agent system automatically generates CLIPS code for an expert system creating firewall rules to block malicious network traffic. Experimental results show the superior performance of the hypernym-hyponym retrieval strategy compared to various baselines and the higher effectiveness of the agentic approach in mitigating threats.

Method: Used a novel four-dimensional business ambiguity taxonomy, conducted human-in-the-loop experiments across strategic, tactical, and operational scenarios, and assessed decisions with an “LLM-as-a-judge” framework on multiple criteria.

Result: Models excel at detecting internal contradictions and contextual ambiguities but struggle with structural linguistic nuances. Ambiguity resolution consistently improved response quality, and sycophantic behavior patterns varied by model architecture.

Conclusion: Generative AI serves as a cognitive scaffold for ambiguity detection/resolution but has artificial limitations requiring human management for reliable strategic partnership, contributing to bounded rationality literature.

Abstract: Generative artificial intelligence is increasingly being integrated into complex business workflows, fundamentally shifting the boundaries of managerial decision-making. However, the reliability of its strategic advice in ambiguous business contexts remains a critical knowledge gap. This study addresses this by comparing various models on ambiguity detection, evaluating how a systematic resolution process enhances response quality, and investigating their sycophantic behavior when presented with flawed directives. Using a novel four-dimensional business ambiguity taxonomy, we conducted a human-in-the-loop experiment across strategic, tactical, and operational scenarios. The resulting decisions were assessed with an “LLM-as-a-judge” framework on criteria including agreement, actionability, justification quality, and constraint adherence. Results reveal distinct performance capabilities. While models excel in detecting internal contradictions and contextual ambiguities, they struggle with structural linguistic nuances. Ambiguity resolution consistently increased response quality across all decision types, while sycophantic behavior analysis revealed distinct patterns depending on the model architecture. This study contributes to the bounded rationality literature by positioning GAI as a cognitive scaffold that can detect and resolve ambiguities managers might overlook, but whose own artificial limitations necessitate human management to ensure its reliability as a strategic partner.

Method: Uses systematic filtering, error correction, and synthetic augmentation for data quality; high-resolution dynamic-resolution encoders for accurate perception; and a hybrid mix of reasoning/non-reasoning data with explicit mode tokens for flexible problem-solving.

Result: The 15B parameter model achieves competitive performance with significantly less training and inference-time compute, showing substantial improvements from data quality and encoder design, while maintaining both fast direct answers and chain-of-thought reasoning capabilities.

Conclusion: Data quality remains the primary lever for model performance, and careful architecture design enables smaller multimodal models to excel at vision-language tasks and complex reasoning while being more efficient.

Abstract: We present Phi-4-reasoning-vision-15B, a compact open-weight multimodal reasoning model, and share the motivations, design choices, experiments, and learnings that informed its development. Our goal is to contribute practical insight to the research community on building smaller, efficient multimodal reasoning models and to share the result of these learnings as an open-weight model that is good at common vision and language tasks and excels at scientific and mathematical reasoning and understanding user interfaces. Our contributions include demonstrating that careful architecture choices and rigorous data curation enable smaller, open-weight multimodal models to achieve competitive performance with significantly less training and inference-time compute and tokens. The most substantial improvements come from systematic filtering, error correction, and synthetic augmentation – reinforcing that data quality remains the primary lever for model performance. Systematic ablations show that high-resolution, dynamic-resolution encoders yield consistent improvements, as accurate perception is a prerequisite for high-quality reasoning. Finally, a hybrid mix of reasoning and non-reasoning data with explicit mode tokens allows a single model to deliver fast direct answers for simpler tasks and chain-of-thought reasoning for complex problems.

Method: Trained a 1.5B-parameter reasoning model on beam statics using parameter-efficient RLVR with binary correctness rewards from symbolic solvers, without teacher-generated reasoning traces.

Result: Best checkpoint achieved 66.7% Pass@1 improvement over base model, showing compositional generalization but failing under topological shifts requiring same equilibrium equations. Intermediate checkpoints yielded strongest reasoning while continued optimization degraded robustness.

Conclusion: Outcome-level alignment with exact physics rewards induces procedural solution templates rather than internalization of governing equations, suggesting verifiable rewards need structured reasoning scaffolding for robust scientific reasoning.

Abstract: Can reinforcement learning with hard, verifiable rewards teach a compact language model to reason about physics, or does it primarily learn to pattern-match toward correct answers? We study this question by training a 1.5B-parameter reasoning model on beam statics, a classic engineering problem, using parameter-efficient RLVR with binary correctness rewards from symbolic solvers, without teacher-generated reasoning traces. The best BeamPERL checkpoint achieves a 66.7% improvement in Pass@1 over the base model. However, the learned competence is anisotropic: the model generalizes compositionally (more loads) but fails under topological shifts (moved supports) that require the same equilibrium equations. Intermediate checkpoints yield the strongest reasoning, while continued optimization degrades robustness while maintaining reward. These findings reveal a key limitation of outcome-level alignment: reinforcement learning with exact physics rewards induces procedural solution templates rather than internalization of governing equations. The precision of the reward signal - even when analytically exact - does not by itself guarantee transferable physical reasoning. Our results suggest that verifiable rewards may need to be paired with structured reasoning scaffolding to move beyond template matching toward robust scientific reasoning.

Method: Created RealPref benchmark with 100 user profiles, 1300 personalized preferences, four types of preference expression (explicit to implicit), long-horizon interaction histories, and three test question types with LLM-as-a-judge evaluation rubrics.

Result: LLM performance significantly drops as context length grows and preference expression becomes more implicit, and generalizing user preference understanding to unseen scenarios is challenging.

Conclusion: RealPref provides a foundation for developing user-aware LLM assistants that better adapt to individual needs, highlighting the need for improved long-term preference understanding.

Abstract: Large Language Models (LLMs) are increasingly serving as personal assistants, where users share complex and diverse preferences over extended interactions. However, assessing how well LLMs can follow these preferences in realistic, long-term situations remains underexplored. This work proposes RealPref, a benchmark for evaluating realistic preference-following in personalized user-LLM interactions. RealPref features 100 user profiles, 1300 personalized preferences, four types of preference expression (ranging from explicit to implicit), and long-horizon interaction histories. It includes three types of test questions (multiple-choice, true-or-false, and open-ended), with detailed rubrics for LLM-as-a-judge evaluation. Results indicate that LLM performance significantly drops as context length grows and preference expression becomes more implicit, and that generalizing user preference understanding to unseen scenarios poses further challenges. RealPref and these findings provide a foundation for future research to develop user-aware LLM assistants that better adapt to individual needs. The code is available at https://github.com/GG14127/RealPref.

Method: Introduces logical transduction algebra that formalizes LLM inference as typed semantic transformations (transducible functions) with schema validity and evidence locality. These functions compose via algebraically grounded operators and execute as stateless asynchronous calls in parallel Map-Reduce programs.

Result: Achieves state-of-the-art performance on challenging benchmarks including DiscoveryBench for data-driven discovery and Archer for NL-to-SQL semantic parsing. The framework provides semantic reliability through strong typing, semantic observability through evidence tracing, and scalability through stateless parallel execution.

Conclusion: Agentics 2.0 enables building high-quality, structured, explainable, and type-safe agentic data workflows suitable for enterprise deployment, addressing key software quality attributes beyond basic text generation capabilities.

Abstract: Agentic AI is rapidly transitioning from research prototypes to enterprise deployments, where requirements extend to meet the software quality attributes of reliability, scalability, and observability beyond plausible text generation. We present Agentics 2.0, a lightweight, Python-native framework for building high-quality, structured, explainable, and type-safe agentic data workflows. At the core of Agentics 2.0, the logical transduction algebra formalizes a large language model inference call as a typed semantic transformation, which we call a transducible function that enforces schema validity and the locality of evidence. The transducible functions compose into larger programs via algebraically grounded operators and execute as stateless asynchronous calls in parallel in asynchronous Map-Reduce programs. The proposed framework provides semantic reliability through strong typing, semantic observability through evidence tracing between slots of the input and output types, and scalability through stateless parallel execution. We instantiate reusable design patterns and evaluate the programs in Agentics 2.0 on challenging benchmarks, including DiscoveryBench for data-driven discovery and Archer for NL-to-SQL semantic parsing, demonstrating state-of-the-art performance.

Method: Introduces τ-Knowledge benchmark with τ-Banking domain modeling fintech customer support workflows where agents must navigate ~700 interconnected knowledge documents while executing tool-mediated account updates, testing both embedding-based retrieval and terminal-based search approaches.

Result: Even frontier models with high reasoning budgets achieve only ~25.5% pass rate, with reliability degrading sharply over repeated trials; agents struggle to retrieve correct documents from densely interlinked knowledge bases and reason accurately over complex internal policies.

Conclusion: τ-Knowledge provides a realistic testbed for developing agents that integrate unstructured knowledge in human-facing deployments, revealing significant challenges in coordinating knowledge retrieval with tool use.

Abstract: Conversational agents are increasingly deployed in knowledge-intensive settings, where correct behavior depends on retrieving and applying domain-specific knowledge from large, proprietary, and unstructured corpora during live interactions with users. Yet most existing benchmarks evaluate retrieval or tool use independently of each other, creating a gap in realistic, fully agentic evaluation over unstructured data in long-horizon interactions. We introduce $τ$-Knowledge, an extension of $τ$-Bench for evaluating agents in environments where success depends on coordinating external, natural-language knowledge with tool outputs to produce verifiable, policy-compliant state changes. Our new domain, $τ$-Banking, models realistic fintech customer support workflows in which agents must navigate roughly 700 interconnected knowledge documents while executing tool-mediated account updates. Across embedding-based retrieval and terminal-based search, even frontier models with high reasoning budgets achieve only $\sim$25.5% pass^1, with reliability degrading sharply over repeated trials. Agents struggle to retrieve the correct documents from densely interlinked knowledge bases and to reason accurately over complex internal policies. Overall, $τ$-Knowledge provides a realistic testbed for developing agents that integrate unstructured knowledge in human-facing deployments.

Method: Proposes a dual-helix governance framework implemented as a 3-track architecture (Knowledge, Behavior, Skills) that uses a knowledge graph substrate to externalize domain facts and enforce executable protocols, complemented by a self-learning cycle for autonomous knowledge growth. Applied to FutureShorelines WebGIS tool refactoring.

Result: The governed agent successfully refactored a 2,265-line monolithic codebase into modular ES6 components, achieving a 51% reduction in cyclomatic complexity and a 7-point increase in maintainability index. Comparative experiments against zero-shot LLM confirmed that externalized governance drives operational reliability.

Conclusion: Externalized governance, not just model capability, is crucial for operational reliability in geospatial engineering. The approach addresses fundamental LLM limitations through structural governance and is implemented in the open-source AgentLoom toolkit.

Abstract: WebGIS development requires rigor, yet agentic AI frequently fails due to five large language model (LLM) limitations: context constraints, cross-session forgetting, stochasticity, instruction failure, and adaptation rigidity. We propose a dual-helix governance framework reframing these challenges as structural governance problems that model capacity alone cannot resolve. We implement the framework as a 3-track architecture (Knowledge, Behavior, Skills) that uses a knowledge graph substrate to stabilize execution by externalizing domain facts and enforcing executable protocols, complemented by a self-learning cycle for autonomous knowledge growth. Applying this to the FutureShorelines WebGIS tool, a governed agent refactored a 2,265-line monolithic codebase into modular ES6 components. Results demonstrated a 51% reduction in cyclomatic complexity and a 7-point increase in maintainability index. A comparative experiment against a zero-shot LLM confirms that externalized governance, not just model capability, drives operational reliability in geospatial engineering. This approach is implemented in the open-source AgentLoom governance toolkit.

Method: Characterizes semantic similarity as distance between image distributions conjured by text prompts, using Jeffreys divergence between reverse-time diffusion SDEs induced by each textual expression, computed via Monte-Carlo sampling.

Result: The method aligns with human-annotated semantic similarity scores and offers better interpretability of generative model representations while opening new evaluation avenues for text-conditioned models.

Conclusion: Proposes a novel perspective on semantic similarity through imagery that leverages generative models, providing interpretable representations and new evaluation methods for text-to-image models.

Abstract: The semantic similarity between sample expressions measures the distance between their latent ‘meaning’. These meanings are themselves typically represented by textual expressions. We propose a novel approach whereby the semantic similarity among textual expressions is based not on other expressions they can be rephrased as, but rather based on the imagery they evoke. While this is not possible with humans, generative models allow us to easily visualize and compare generated images, or their distribution, evoked by a textual prompt. Therefore, we characterize the semantic similarity between two textual expressions simply as the distance between image distributions they induce, or ‘conjure.’ We show that by choosing the Jeffreys divergence between the reverse-time diffusion stochastic differential equations (SDEs) induced by each textual expression, this can be directly computed via Monte-Carlo sampling. Our method contributes a novel perspective on semantic similarity that not only aligns with human-annotated scores, but also opens up new avenues for the evaluation of text-conditioned generative models while offering better interpretability of their learnt representations.

Method: Proposes OVMSE framework with two key components: (1) Offline Value Function Memory (OVM) mechanism to compute target Q-values, preserving offline knowledge and enabling smooth transitions, and (2) decentralized Sequential Exploration (SE) strategy that leverages pre-trained offline policy to reduce joint state-action space exploration.

Result: Extensive experiments on StarCraft Multi-Agent Challenge (SMAC) demonstrate that OVMSE significantly outperforms existing baselines, achieving superior sample efficiency and overall performance.

Conclusion: OVMSE effectively addresses the core challenges of O2O MARL by preserving offline knowledge through OVM and enabling efficient exploration through SE, making it a promising approach for multi-agent reinforcement learning with offline-to-online transitions.

Abstract: Offline-to-Online Reinforcement Learning has emerged as a powerful paradigm, leveraging offline data for initialization and online fine-tuning to enhance both sample efficiency and performance. However, most existing research has focused on single-agent settings, with limited exploration of the multi-agent extension, i.e., Offline-to-Online Multi-Agent Reinforcement Learning (O2O MARL). In O2O MARL, two critical challenges become more prominent as the number of agents increases: (i) the risk of unlearning pre-trained Q-values due to distributional shifts during the transition from offline-to-online phases, and (ii) the difficulty of efficient exploration in the large joint state-action space. To tackle these challenges, we propose a novel O2O MARL framework called Offline Value Function Memory with Sequential Exploration (OVMSE). First, we introduce the Offline Value Function Memory (OVM) mechanism to compute target Q-values, preserving knowledge gained during offline training, ensuring smoother transitions, and enabling efficient fine-tuning. Second, we propose a decentralized Sequential Exploration (SE) strategy tailored for O2O MARL, which effectively utilizes the pre-trained offline policy for exploration, thereby significantly reducing the joint state-action space to be explored. Extensive experiments on the StarCraft Multi-Agent Challenge (SMAC) demonstrate that OVMSE significantly outperforms existing baselines, achieving superior sample efficiency and overall performance.

Method: Created MuRAL dataset with 21+ hours of multi-user sensor data from 21 smart home sessions, featuring detailed natural language descriptions, explicit resident identities, and rich activity labels in complex multi-resident scenarios.

Result: Benchmarking shows current LLMs face major challenges in maintaining accurate resident assignment over long sequences, generating precise action descriptions, and effectively integrating context for activity prediction.

Conclusion: MuRAL addresses the gap in natural language-enabled multi-resident datasets and reveals current LLM limitations in realistic smart home activity understanding.

Abstract: Recent progress in Large Language Models (LLMs) has enabled advanced reasoning and zero-shot recognition for human activity understanding with ambient sensor data. However, widely used multi-resident datasets such as CASAS, ARAS, and MARBLE lack natural language context and fine-grained annotation, limiting the full exploitation of LLM capabilities in realistic smart environments. To address this gap, we present MuRAL (Multi-Resident Ambient sensor dataset with natural Language), comprising over 21 hours of multi-user sensor data from 21 sessions in a smart home. MuRAL uniquely features detailed natural language descriptions, explicit resident identities, and rich activity labels, all situated in complex, dynamic, multi-resident scenarios. We benchmark state-of-the-art LLMs on MuRAL for three core tasks: subject assignment, action description, and activity classification. Results show that current LLMs still face major challenges on MuRAL, especially in maintaining accurate resident assignment over long sequences, generating precise action descriptions, and effectively integrating context for activity prediction. The dataset is publicly available at: https://mural.imag.fr/.

Method: Proposes 8 architectural principles for hierarchical dual-source emotion-like control, distills 4 engineering risk-reduction constraints from major consciousness theories, and presents a concrete architecture as existence proof while analyzing extensions and risk pathways.

Result: Demonstrates that emotion-like control can satisfy consciousness risk-reduction constraints, identifies stable modifications that preserve compliance, and maps gradual transitions that increase access risk while providing a methodological template for consciousness-related architectural audits.

Conclusion: Provides a modular biologically motivated control architecture, a control model of emotions, methodological template for consciousness-related architectural tests, and preliminary audit indicators for AI safety governance, with architecture functioning independently as emotion-like controller.

Abstract: As artificial agents display increasingly sophisticated emotion-like behaviors, frameworks for assessing whether such systems risk instantiating consciousness remain limited. This contribution asks whether synthetic emotion-like control can be implemented while deliberately excluding architectural features that major theories associate with access-like consciousness. We propose architectural principles (A1-A8) for a hierarchical, dual-source implementation in which (i) immediate needs generate motivational signals and (ii) episodic memory provides affective guidance from similar past situations; the two sources converge to modulate action selection. To operationalize consciousness-related risk, we distill predictions from major theories into four engineering risk-reduction constraints: (R1) no content-general, workspace-like global broadcast, (R2) no metarepresentation, (R3) no autobiographical consolidation, and (R4) bounded learning. We address three questions: (Q1) Can emotion-like control satisfy R1-R4? We present a concrete architecture as an existence proof. (Q2) Can the architecture be extended without introducing access-enabling features? We identify stable modifications that preserve compliance. (Q3) Can we trace graded paths that plausibly increase access risk? We map gradual transitions that progressively violate the constraints. Our contribution operates at three levels: on the engineering side, we present a modular, biologically motivated control architecture; on the theoretical side, we propose a control model of emotions and a methodological template for converting consciousness-related questions into auditable architectural tests; on the safety side, we sketch preliminary audit indicators that may inform future governance frameworks. The architecture functions independently as an emotion-like controller, while the risk-reduction criteria may extend to other AI systems.

Method: Builds on N2M-RSI foundation, introduces differentiable intrinsic reward function based on confidence, error, novelty, and cumulative success, defines emotion-gradient dynamics and RSI trigger conditions, and creates reinforcement-compatible optimization objective with quantifiable metrics like Meaning Density and Meaning Conversion Efficiency.

Result: Part I establishes single-agent theoretical foundations; future parts will cover safety certificates, collective intelligence, and feasibility constraints. Framework enables agents to overwrite their own learning algorithms under formally bounded risk.

Conclusion: EG-MRSI provides a rigorous, extensible foundation for open-ended and safe AGI development, with a complete theoretical framework planned across four parts.

Abstract: We present the Emotion-Gradient Metacognitive Recursive Self-Improvement (EG-MRSI) framework, a novel architecture that integrates introspective metacognition, emotion-based intrinsic motivation, and recursive self-modification into a unified theoretical system. The framework is explicitly capable of overwriting its own learning algorithm under formally bounded risk. Building upon the Noise-to-Meaning RSI (N2M-RSI) foundation, EG-MRSI introduces a differentiable intrinsic reward function driven by confidence, error, novelty, and cumulative success. This signal regulates both a metacognitive mapping and a self-modification operator constrained by provable safety mechanisms. We formally define the initial agent configuration, emotion-gradient dynamics, and RSI trigger conditions, and derive a reinforcement-compatible optimization objective that guides the agent’s development trajectory. Meaning Density and Meaning Conversion Efficiency are introduced as quantifiable metrics of semantic learning, closing the gap between internal structure and predictive informativeness. This Part I paper establishes the single-agent theoretical foundations of EG-MRSI. Future parts will extend this framework to include safety certificates and rollback protocols (Part II), collective intelligence mechanisms (Part III), and feasibility constraints including thermodynamic and computational limits (Part IV). Together, the EG-MRSI series provides a rigorous, extensible foundation for open-ended and safe AGI.

Method: Three-stage approach: 1) Initialize judgment rules using first-order logic formalism to capture complex reasoning logic; 2) Use Confusion-aware Contrastive Learning (CACL) to dynamically optimize rules through quizzes of confusable cases; 3) Apply optimized rules for legal judgment prediction.

Result: Experimental results on two public datasets show superior performance across all metrics compared to existing approaches.

Conclusion: The proposed framework effectively incorporates adaptive legal reasoning logic through first-order logic formalism and contrastive learning, significantly improving legal judgment prediction performance.

Abstract: Legal Judgment Prediction (LJP) is a pivotal task in legal AI. Existing semantic-enhanced LJP models integrate judicial precedents and legal knowledge for high performance. But they neglect legal reasoning logic, a critical component of legal judgments requiring rigorous logical analysis. Although some approaches utilize legal reasoning logic for high-quality predictions, their logic rigidity hinders adaptation to case-specific logical frameworks, particularly in complex cases that are lengthy and detailed. This paper proposes a rule-enhanced legal judgment prediction framework based on first-order logic (FOL) formalism and comparative learning (CL) to develop an adaptive adjustment mechanism for legal judgment logic and further enhance performance in LJP. Inspired by the process of human exam preparation, our method follows a three-stage approach: first, we initialize judgment rules using the FOL formalism to capture complex reasoning logic accurately; next, we propose a Confusion-aware Contrastive Learning (CACL) to dynamically optimize the judgment rules through a quiz consisting of confusable cases; finally, we utilize the optimized judgment rules to predict legal judgments. Experimental results on two public datasets show superior performance across all metrics. The code is publicly available{https://anonymous.4open.science/r/RLJP-FDF1}.

Method: Multi-turn supervised fine-tuning (SFT) and reinforcement learning (RL) to train LLMs to generate multiple code queries during step-by-step reasoning. Introduces multi-stage curriculum learning that partitions training samples by measured improvement potential, prioritizing high-potential samples first then gradually shifting to lower-potential ones.

Result: R1-CI-14B improves average accuracy on 37 test tasks from 44.1% to 72.4%, outperforming text-only GPT-4o (58.6%) and GPT-4o with Code Interpreter (70.9%). The curriculum learning approach increased average RL gains from +3.4% to +9.3% across Qwen-2.5 models (3/7/14B). The model also exhibits emergent self-checking behavior through code generation.

Conclusion: The proposed multi-stage curriculum learning effectively addresses challenges in training general-purpose Code Interpreters, achieving state-of-the-art performance and demonstrating emergent self-checking capabilities through code generation.

Abstract: Practical guidance on training Large Language Models (LLMs) to leverage Code Interpreter across diverse tasks remains lacking. We present R1-Code-Interpreter, an extension of a text-only LLM trained via multi-turn supervised fine-tuning (SFT) and reinforcement learning (RL) to autonomously generate multiple code queries during step-by-step reasoning. Unlike prior RL + tool-use efforts focused on narrow domains such as math or retrieval, we curate 144 diverse reasoning and planning tasks and show that training a general-purpose Code Interpreter across them presents significant challenges due to task heterogeneity and scarcity of effective samples. To address this, we introduce a multi-stage curriculum learning approach that partitions training samples by measured improvement potential. The RL training prioritizes samples with higher potential and gradually shifts to lower-potential ones, increasing the average RL gains from merely +3.4% to +9.3% across Qwen-2.5 models (3/7/14B). Our final model, R1-CI-14B, improves average accuracy on the 37 test tasks from 44.1% to 72.4%, outperforming text-only GPT-4o (58.6%) and GPT-4o with Code Interpreter (70.9%). Notably, R1-CI-14B also exhibits emergent self-checking behavior through code generation. Datasets, Codes, and Models are available at https://github.com/yongchao98/R1-Code-Interpreter and https://huggingface.co/yongchao98.

Method: Developed LeanTutor with three modules: (1) autoformalizer/proof-checker to translate between natural and formal language, (2) next-step generator for proof guidance, and (3) natural language feedback generator. Created PeanoBench dataset of 371 Peano Arithmetic proofs in both human-written natural language and formal language.

Result: Presented a proof-of-concept system that demonstrates the feasibility of combining LLMs with theorem provers for mathematical proof tutoring. Introduced PeanoBench as an evaluation dataset derived from the Natural Numbers Game.

Conclusion: The combination of LLMs and theorem provers shows promise for creating effective, provably-correct mathematical proof tutors that bridge the gap between natural language accessibility and formal correctness.

Abstract: This paper considers the development of an AI-based provably-correct mathematical proof tutor. While Large Language Models (LLMs) allow seamless communication in natural language, they are error prone. Theorem provers such as Lean allow for provable-correctness, but these are hard for students to learn. We present a proof-of-concept system (LeanTutor) by combining the complementary strengths of LLMs and theorem provers. LeanTutor is composed of three modules: (i) an autoformalizer/proof-checker, (ii) a next-step generator, and (iii) a natural language feedback generator. To evaluate the system, we introduce PeanoBench, a dataset of 371 Peano Arithmetic proofs in human-written natural language and formal language, derived from the Natural Numbers Game.

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 2507.15796: Page request resulted in HTTP 429 (https://export.arxiv.org/api/query?search_query=&id_list=2507.15796&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 2507.16405: Page request resulted in HTTP 429 (https://export.arxiv.org/api/query?search_query=&id_list=2507.16405&sortBy=relevance&sortOrder=descending&start=0&max_results=100)

Method: Unable to determine paper method due to access error

Result: Unable to determine paper results due to access error

Conclusion: Unable to determine paper conclusion due to access error

Abstract: Failed to fetch summary for 2508.03284: Page request resulted in HTTP 429 (https://export.arxiv.org/api/query?search_query=&id_list=2508.03284&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 2509.21782: Page request resulted in HTTP 429 (https://export.arxiv.org/api/query?search_query=&id_list=2509.21782&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 2510.07972: Page request resulted in HTTP 429 (https://export.arxiv.org/api/query?search_query=&id_list=2510.07972&sortBy=relevance&sortOrder=descending&start=0&max_results=100)

Method: Cannot determine method as abstract is not available due to API rate limiting

Result: Cannot determine results as abstract is not available due to API rate limiting

Conclusion: Cannot draw conclusions about paper content due to unavailability of abstract

Abstract: Failed to fetch summary for 2510.26905: Page request resulted in HTTP 429 (https://export.arxiv.org/api/query?search_query=&id_list=2510.26905&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 2511.05854: Page request resulted in HTTP 429 (https://export.arxiv.org/api/query?search_query=&id_list=2511.05854&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 2511.21471: Page request resulted in HTTP 429 (https://export.arxiv.org/api/query?search_query=&id_list=2511.21471&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 2511.22235: Page request resulted in HTTP 429 (https://export.arxiv.org/api/query?search_query=&id_list=2511.22235&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 2601.16529: Page request resulted in HTTP 429 (https://export.arxiv.org/api/query?search_query=&id_list=2601.16529&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 2602.09937: Page request resulted in HTTP 429 (https://export.arxiv.org/api/query?search_query=&id_list=2602.09937&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 2602.19810: Page request resulted in HTTP 429 (https://export.arxiv.org/api/query?search_query=&id_list=2602.19810&sortBy=relevance&sortOrder=descending&start=0&max_results=100)

Method: No method information available - arXiv API returned HTTP 429 (Too Many Requests) error

Result: No results available - the paper content could not be accessed

Conclusion: Cannot analyze paper due to technical limitations in accessing the content

Abstract: Failed to fetch summary for 2602.22769: Page request resulted in HTTP 429 (https://export.arxiv.org/api/query?search_query=&id_list=2602.22769&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 2602.23541: Page request resulted in HTTP 429 (https://export.arxiv.org/api/query?search_query=&id_list=2602.23541&sortBy=relevance&sortOrder=descending&start=0&max_results=100)

Method: Unknown - paper content not accessible due to HTTP 429 error from arXiv API.

Result: No results available - paper summary fetch failed due to rate limiting.

Conclusion: Cannot analyze paper content due to technical limitations in accessing the arXiv API.

Abstract: Failed to fetch summary for 2603.01620: Page request resulted in HTTP 429 (https://export.arxiv.org/api/query?search_query=&id_list=2603.01620&sortBy=relevance&sortOrder=descending&start=0&max_results=100)

Method: Unable to determine method due to failed summary fetch

Result: Unable to determine results due to failed summary fetch

Conclusion: Unable to determine conclusion due to failed summary fetch

Abstract: Failed to fetch summary for 2603.02029: Page request resulted in HTTP 429 (https://export.arxiv.org/api/query?search_query=&id_list=2603.02029&sortBy=relevance&sortOrder=descending&start=0&max_results=100)

Method: Unable to determine method due to fetch failure

Result: Unable to determine results due to fetch failure

Conclusion: Unable to determine conclusion due to fetch failure

Abstract: Failed to fetch summary for 2603.02214: Page request resulted in HTTP 429 (https://export.arxiv.org/api/query?search_query=&id_list=2603.02214&sortBy=relevance&sortOrder=descending&start=0&max_results=100)

Method: Unable to determine method due to fetch failure

Result: Unable to determine results due to fetch failure

Conclusion: Unable to determine conclusion due to fetch failure

Abstract: Failed to fetch summary for 2603.02365: Page request resulted in HTTP 429 (https://export.arxiv.org/api/query?search_query=&id_list=2603.02365&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 draw conclusions due to access error

Abstract: Failed to fetch summary for 2410.20310: Page request resulted in HTTP 429 (https://export.arxiv.org/api/query?search_query=&id_list=2410.20310&sortBy=relevance&sortOrder=descending&start=0&max_results=100)

Method: Cannot determine method without access to paper abstract

Result: Cannot determine results without access to paper abstract

Conclusion: Cannot draw conclusions without access to paper abstract

Abstract: Failed to fetch summary for 2603.02504: Page request resulted in HTTP 429 (https://export.arxiv.org/api/query?search_query=&id_list=2603.02504&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 2411.15272: Page request resulted in HTTP 429 (https://export.arxiv.org/api/query?search_query=&id_list=2411.15272&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 2405.01440: Page request resulted in HTTP 429 (https://export.arxiv.org/api/query?search_query=&id_list=2405.01440&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 2409.06912: Page request resulted in HTTP 429 (https://export.arxiv.org/api/query?search_query=&id_list=2409.06912&sortBy=relevance&sortOrder=descending&start=0&max_results=100)

Method: Unable to analyze method due to missing abstract content

Result: Unable to analyze results due to missing abstract content

Conclusion: Unable to analyze conclusion due to missing abstract content

Abstract: Failed to fetch summary for 2412.06531: Page request resulted in HTTP 429 (https://export.arxiv.org/api/query?search_query=&id_list=2412.06531&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 2501.01317: Page request resulted in HTTP 429 (https://export.arxiv.org/api/query?search_query=&id_list=2501.01317&sortBy=relevance&sortOrder=descending&start=0&max_results=100)

Method: Unable to determine method due to access restrictions.

Result: Unable to determine results due to access restrictions.

Conclusion: Unable to determine conclusion due to access restrictions.

Abstract: Failed to fetch summary for 2502.10550: Page request resulted in HTTP 429 (https://export.arxiv.org/api/query?search_query=&id_list=2502.10550&sortBy=relevance&sortOrder=descending&start=0&max_results=100)

Method: Unable to determine method due to missing abstract

Result: Unable to determine results due to missing abstract

Conclusion: Unable to draw conclusions due to missing abstract

Abstract: Failed to fetch summary for 2503.07638: Page request resulted in HTTP 429 (https://export.arxiv.org/api/query?search_query=&id_list=2503.07638&sortBy=relevance&sortOrder=descending&start=0&max_results=100)

Method: Cannot determine method without access to paper content

Result: Cannot determine results without access to paper content

Conclusion: Cannot determine conclusion without access to paper content

Abstract: Failed to fetch summary for 2505.12506: Page request resulted in HTTP 429 (https://export.arxiv.org/api/query?search_query=&id_list=2505.12506&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 2503.07885: Page request resulted in HTTP 429 (https://export.arxiv.org/api/query?search_query=&id_list=2503.07885&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 2505.13033: Page request resulted in HTTP 429 (https://export.arxiv.org/api/query?search_query=&id_list=2505.13033&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 2505.06743: Page request resulted in HTTP 429 (https://export.arxiv.org/api/query?search_query=&id_list=2505.06743&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 2505.20065: Page request resulted in HTTP 429 (https://export.arxiv.org/api/query?search_query=&id_list=2505.20065&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 draw conclusions due to retrieval failure

Abstract: Failed to fetch summary for 2506.05634: Page request resulted in HTTP 429 (https://export.arxiv.org/api/query?search_query=&id_list=2506.05634&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 2506.05937: Page request resulted in HTTP 429 (https://export.arxiv.org/api/query?search_query=&id_list=2506.05937&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 2507.06625: Page request resulted in HTTP 429 (https://export.arxiv.org/api/query?search_query=&id_list=2507.06625&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 draw conclusions due to access error

Abstract: Failed to fetch summary for 2508.00450: Page request resulted in HTTP 429 (https://export.arxiv.org/api/query?search_query=&id_list=2508.00450&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 2508.04735: Page request resulted in HTTP 429 (https://export.arxiv.org/api/query?search_query=&id_list=2508.04735&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 draw conclusions due to missing abstract

Abstract: Failed to fetch summary for 2508.06066: Page request resulted in HTTP 429 (https://export.arxiv.org/api/query?search_query=&id_list=2508.06066&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 determine conclusion due to API access limitations

Abstract: Failed to fetch summary for 2508.11025: Page request resulted in HTTP 429 (https://export.arxiv.org/api/query?search_query=&id_list=2508.11025&sortBy=relevance&sortOrder=descending&start=0&max_results=100)

Method: No method information available due to failed abstract retrieval

Result: No results available due to failed abstract retrieval

Conclusion: Cannot provide analysis due to technical limitations in accessing the paper abstract

Abstract: Failed to fetch summary for 2508.21513: Page request resulted in HTTP 429 (https://export.arxiv.org/api/query?search_query=&id_list=2508.21513&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 2509.13471: Page request resulted in HTTP 429 (https://export.arxiv.org/api/query?search_query=&id_list=2509.13471&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 2509.19084: Page request resulted in HTTP 429 (https://export.arxiv.org/api/query?search_query=&id_list=2509.19084&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 2509.21091: Page request resulted in HTTP 429 (https://export.arxiv.org/api/query?search_query=&id_list=2509.21091&sortBy=relevance&sortOrder=descending&start=0&max_results=100)

Method: Unknown - paper content not accessible.

Result: Unknown - paper content not accessible.

Conclusion: Unable to draw conclusions about the paper due to technical limitations in accessing the content.

Abstract: Failed to fetch summary for 2509.24222: Page request resulted in HTTP 429 (https://export.arxiv.org/api/query?search_query=&id_list=2509.24222&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 2510.07151: Page request resulted in HTTP 429 (https://export.arxiv.org/api/query?search_query=&id_list=2510.07151&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 2510.07650: Page request resulted in HTTP 429 (https://export.arxiv.org/api/query?search_query=&id_list=2510.07650&sortBy=relevance&sortOrder=descending&start=0&max_results=100)

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

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

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

Abstract: Failed to fetch summary for 2510.15982: Page request resulted in HTTP 429 (https://export.arxiv.org/api/query?search_query=&id_list=2510.15982&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 2510.26303: Page request resulted in HTTP 429 (https://export.arxiv.org/api/query?search_query=&id_list=2510.26303&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 2510.26840: Page request resulted in HTTP 429 (https://export.arxiv.org/api/query?search_query=&id_list=2510.26840&sortBy=relevance&sortOrder=descending&start=0&max_results=100)

Method: Cannot determine method without access to paper content

Result: Cannot determine results without access to paper content

Conclusion: Cannot draw conclusions without access to paper content

Abstract: Failed to fetch summary for 2511.07441: Page request resulted in HTTP 429 (https://export.arxiv.org/api/query?search_query=&id_list=2511.07441&sortBy=relevance&sortOrder=descending&start=0&max_results=100)

Method: Cannot determine method without access to paper content

Result: Cannot determine results without access to paper content

Conclusion: Cannot draw conclusions without access to paper content

Abstract: Failed to fetch summary for 2511.14555: Page request resulted in HTTP 429 (https://export.arxiv.org/api/query?search_query=&id_list=2511.14555&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 analyze paper due to technical issues with arXiv API

Abstract: Failed to fetch summary for 2511.14827: Page request resulted in HTTP 429 (https://export.arxiv.org/api/query?search_query=&id_list=2511.14827&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 2511.22935: Page request resulted in HTTP 429 (https://export.arxiv.org/api/query?search_query=&id_list=2511.22935&sortBy=relevance&sortOrder=descending&start=0&max_results=100)

Method: Cannot analyze method without access to paper content

Result: No results available due to technical access issues

Conclusion: Paper analysis impossible due to HTTP 429 error from arXiv API

Abstract: Failed to fetch summary for 2512.01759: Page request resulted in HTTP 429 (https://export.arxiv.org/api/query?search_query=&id_list=2512.01759&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 2512.18925: Page request resulted in HTTP 429 (https://export.arxiv.org/api/query?search_query=&id_list=2512.18925&sortBy=relevance&sortOrder=descending&start=0&max_results=100)

Method: Unable to determine method due to failed summary fetch

Result: Unable to determine results due to failed summary fetch

Conclusion: Unable to determine conclusion due to failed summary fetch

Abstract: Failed to fetch summary for 2512.19570: Page request resulted in HTTP 429 (https://export.arxiv.org/api/query?search_query=&id_list=2512.19570&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 2601.13286: Page request resulted in HTTP 429 (https://export.arxiv.org/api/query?search_query=&id_list=2601.13286&sortBy=relevance&sortOrder=descending&start=0&max_results=100)

Method: Cannot determine method due to access error

Result: Cannot determine results due to access error

Conclusion: Cannot determine conclusion due to access error

Abstract: Failed to fetch summary for 2602.05286: Page request resulted in HTTP 429 (https://export.arxiv.org/api/query?search_query=&id_list=2602.05286&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 2602.07970: Page request resulted in HTTP 429 (https://export.arxiv.org/api/query?search_query=&id_list=2602.07970&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 2602.10149: Page request resulted in HTTP 429 (https://export.arxiv.org/api/query?search_query=&id_list=2602.10149&sortBy=relevance&sortOrder=descending&start=0&max_results=100)

Method: Cannot determine method as paper content could not be retrieved

Result: Cannot determine results as paper content could not be retrieved

Conclusion: Cannot draw conclusions as paper content could not be retrieved

Abstract: Failed to fetch summary for 2602.12851: Page request resulted in HTTP 429 (https://export.arxiv.org/api/query?search_query=&id_list=2602.12851&sortBy=relevance&sortOrder=descending&start=0&max_results=100)

Method: Cannot determine method due to HTTP 429 error preventing access to paper content. The arXiv API rate limiting prevents retrieval of the abstract and details.

Result: No results available due to access restrictions. The HTTP 429 error indicates too many requests to the arXiv API, preventing content retrieval.

Conclusion: Unable to analyze paper content due to technical limitations. The arXiv API rate limiting prevents assessment of this paper’s relevance to multimodal audio/vision research.

Abstract: Failed to fetch summary for 2602.17176: Page request resulted in HTTP 429 (https://export.arxiv.org/api/query?search_query=&id_list=2602.17176&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 2602.18308: Page request resulted in HTTP 429 (https://export.arxiv.org/api/query?search_query=&id_list=2602.18308&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 2602.20323: Page request resulted in HTTP 429 (https://export.arxiv.org/api/query?search_query=&id_list=2602.20323&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 2602.22227: Page request resulted in HTTP 429 (https://export.arxiv.org/api/query?search_query=&id_list=2602.22227&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 2602.20541: Page request resulted in HTTP 429 (https://export.arxiv.org/api/query?search_query=&id_list=2602.20541&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 2603.00183: Page request resulted in HTTP 429 (https://export.arxiv.org/api/query?search_query=&id_list=2603.00183&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 determine conclusion due to API access limitations

Abstract: Failed to fetch summary for 2603.02234: Page request resulted in HTTP 429 (https://export.arxiv.org/api/query?search_query=&id_list=2603.02234&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 2603.00251: Page request resulted in HTTP 429 (https://export.arxiv.org/api/query?search_query=&id_list=2603.00251&sortBy=relevance&sortOrder=descending&start=0&max_results=100)

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

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

Conclusion: Cannot draw conclusions about paper content due to inability to access abstract

Abstract: Failed to fetch summary for 2603.01508: Page request resulted in HTTP 429 (https://export.arxiv.org/api/query?search_query=&id_list=2603.01508&sortBy=relevance&sortOrder=descending&start=0&max_results=100)

Method: Unable to determine method due to fetch failure

Result: Unable to determine results due to fetch failure

Conclusion: Unable to determine conclusion due to fetch failure

Abstract: Failed to fetch summary for 2603.02512: Page request resulted in HTTP 429 (https://export.arxiv.org/api/query?search_query=&id_list=2603.02512&sortBy=relevance&sortOrder=descending&start=0&max_results=100)

Method: Unable to determine method due to fetch failure

Result: Unable to determine results due to fetch failure

Conclusion: Unable to determine conclusion due to fetch failure

Abstract: Failed to fetch summary for 2603.03045: Page request resulted in HTTP 429 (https://export.arxiv.org/api/query?search_query=&id_list=2603.03045&sortBy=relevance&sortOrder=descending&start=0&max_results=100)

Method: Cannot determine method as paper content could not be retrieved

Result: Cannot determine results as paper content could not be retrieved

Conclusion: Cannot draw conclusions as paper content could not be retrieved

Abstract: Failed to fetch summary for 2603.03140: Page request resulted in HTTP 429 (https://export.arxiv.org/api/query?search_query=&id_list=2603.03140&sortBy=relevance&sortOrder=descending&start=0&max_results=100)

Method: Introduces ACES (Accent-Centric Evaluation of Speech models), a representation-centric audit that extracts accent-discriminative subspaces using linear discriminant analysis. Analyzes Wav2Vec2-base model with five English accents, identifies where accent information concentrates in the model’s layers, and examines correlations between projection magnitude and word error rate (WER). Uses subspace-constrained perturbations to test coupling between representation shift and performance degradation.

Result: Accent information concentrates in a low-dimensional early-layer subspace (layer 3, k=8 dimensions). Projection magnitude correlates with per-utterance WER (r=0.26). Subspace-constrained perturbations show stronger coupling between representation shift and degradation (r=0.32) than random-subspace controls (r=0.15). However, linear attenuation of this subspace does not reduce disparity and slightly worsens it.

Conclusion: Accent-relevant features are deeply entangled with recognition-critical cues in ASR models. Accent subspaces serve as valuable diagnostic tools for understanding performance disparities, but cannot be simply “erased” to achieve fairness. The entanglement suggests more sophisticated approaches are needed beyond simple feature removal.

Abstract: ASR systems exhibit persistent performance disparities across accents, yet the internal mechanisms underlying these gaps remain poorly understood. We introduce ACES, a representation-centric audit that extracts accent-discriminative subspaces and uses them to probe model fragility and disparity. Analyzing Wav2Vec2-base with five English accents, we find that accent information concentrates in a low-dimensional early-layer subspace (layer 3, k=8). Projection magnitude correlates with per-utterance WER (r=0.26), and crucially, subspace-constrained perturbations yield stronger coupling between representation shift and degradation (r=0.32) than random-subspace controls (r=0.15). Finally, linear attenuation of this subspace however does not reduce disparity and slightly worsens it. Our findings suggest that accent-relevant features are deeply entangled with recognition-critical cues, positioning accent subspaces as vital diagnostic tools rather than simple “erasure” levers for fairness.

Method: Proposes AVUR-LLM with two key components: 1) Sparse modality alignment for better cross-modal interaction, and 2) Visual unit-guided refinement to enhance recognition accuracy using visual information.

Result: Achieves state-of-the-art results on LRS3 dataset, with 37% relative improvement over baseline at 0 dB SNR under additive-noise conditions.

Conclusion: AVUR-LLM effectively addresses limitations of previous LLM-based AVSR approaches through improved cross-modal alignment and refinement, demonstrating significant robustness in noisy acoustic conditions.

Abstract: Audio-Visual Speech Recognition (AVSR) integrates acoustic and visual information to enhance robustness in adverse acoustic conditions. Recent advances in Large Language Models (LLMs) have yielded competitive automatic speech recognition performance and shown effectiveness for AVSR. However, prior approaches project audio and visual features independently or apply shallow fusion, limiting cross-modal alignment and complementary exchange while increasing the LLM’s computational load. To address this, we propose AVUR-LLM, an LLM-based Audio-Visual Speech Recognition via Sparse Modality Alignment and Visual Unit-Guided Refinement. Experiments on LRS3 demonstrate state-of-the-art results for AVSR. Under additive-noise conditions at 0 dB SNR, it achieves 37% relative improvement over the baseline system.

Method: Used 71K AudioCapsV2 clips to extract normalized (source, attribute) events. Built two query types: present-event queries for ground-truth detection and absent-event queries to probe hallucinations, using similarity-filtered negative sampling in audio-aligned text embedding space. Evaluated four SOTA Audio LLMs with 12 prompt variants over 500K yes/no queries per model.

Result: Across all models, increasing event count consistently lowers true-positive rate and raises false-positive rate. Prompts induce a strong trade-off between true-positive and false-positive rates. Confidence analysis shows models become more uncertain on multi-event audio, revealing significant room for improvement.

Conclusion: Audio LLMs struggle with complex acoustic scenes, showing degraded performance as event complexity increases. The systematic evaluation reveals fundamental limitations in current models’ ability to handle multi-event audio, highlighting important directions for future research in robust audio understanding.

Abstract: Audio LLMs have shown a strong ability to understand audio samples, yet their reliability in complex acoustic scenes remains under-explored. Unlike prior work limited to small scale or less controlled query construction, we present a large-scale evaluation of event grounding and false alarms as auditory scene complexity increases. Using 71K AudioCapsV2 clips, we extract normalized (source, attribute) events and build two query types: present-event queries for ground-truth detection and absent-event queries to probe hallucinations, using similarity-filtered negative sampling in an audio-aligned text embedding space. We evaluate four SOTA Audio LLMs with 12 prompt variants over 500K yes/no queries per model. Across models, increasing event count consistently lowers true-positive rate and raises false-positive rate, while prompts induce a strong trade-off between the two. Our confidence analysis shows that models become more uncertain on multi-event audio, revealing room for improvement.

Method: Two-stage approach: 1) Separation using BandSplit-RoFormer with three-stage curriculum learning (4-stem warm-start → 8-stem extension via head expansion), 2) Restoration using HiFi++ GAN trained as generalist then specialized into eight instrument-specific experts.

Result: System developed for MSR ICASSP Challenge 2025, demonstrating effective decomposition of MSR into separation and restoration components with specialized curriculum learning and expert restoration models.

Conclusion: The proposed two-stage framework effectively addresses Music Source Restoration challenges by combining advanced separation techniques with specialized restoration models, suitable for competition settings.

Abstract: Music Source Restoration (MSR) targets recovery of original, unprocessed instrument stems from fully mixed and mastered audio, where production effects and distribution artifacts violate common linear-mixture assumptions. This technical report presents the CP-JKU team’s system for the MSR ICASSP Challenge 2025. Our approach decomposes MSR into separation and restoration. First, a single BandSplit-RoFormer separator predicts eight stems plus an auxiliary other stem, and is trained with a three-stage curriculum that progresses from 4-stem warm-start fine-tuning (with LoRA) to 8-stem extension via head expansion. Second, we apply a HiFi++ GAN waveform restorer trained as a generalist and then specialized into eight instrument-specific experts.

Method: Applies recent advances in diffusion model training to audio super-resolution, specifically for upsampling to 48kHz. The model architecture is designed to be lightweight with only 1.3M parameters and 50 GFLOPs computational complexity.

Result: Outperforms NU-Wave 2 and achieves comparable results to state-of-the-art models while being significantly more efficient. The model can be trained with fewer resources and faster than most diffusion/flow-based solutions.

Conclusion: FastWave demonstrates that high-quality audio super-resolution can be achieved with lightweight models, making the technology more accessible and practical for real-world applications.

Abstract: Audio Super-Resolution is a set of techniques aimed at high-quality estimation of the given signal as if it would be sampled with higher sample rate. Among suggested methods there are diffusion and flow models (which are considered slower), generative adversarial networks (which are considered faster), however both approaches are currently presented by high-parametric networks, requiring high computational costs both for training and inference. We propose a solution to both these problems by re-considering the recent advances in the training of diffusion models and applying them to super-resolution from any to 48 kHz sample rate. Our approach shows better results than NU-Wave 2 and is comparable to state-of-the-art models. Our model called FastWave has around 50 GFLOPs of computational complexity and 1.3 M parameters and can be trained with less resources and significantly faster than the majority of recently proposed diffusion- and flow-based solutions. The code has been made publicly available.

Method: Proposes ZeSTA: a domain-conditioned training framework with lightweight domain embedding to distinguish real vs synthetic speech, combined with real-data oversampling to stabilize adaptation under extremely limited target data, without modifying base architecture.

Result: Experiments on LibriTTS and in-house datasets with two ZS-TTS sources show improved speaker similarity over naive synthetic augmentation while preserving intelligibility and perceptual quality.

Conclusion: ZeSTA effectively leverages zero-shot TTS for data augmentation in low-resource personalized speech synthesis by addressing speaker similarity degradation through domain conditioning and real-data oversampling.

Abstract: We investigate the use of zero-shot text-to-speech (ZS-TTS) as a data augmentation source for low-resource personalized speech synthesis. While synthetic augmentation can provide linguistically rich and phonetically diverse speech, naively mixing large amounts of synthetic speech with limited real recordings often leads to speaker similarity degradation during fine-tuning. To address this issue, we propose ZeSTA, a simple domain-conditioned training framework that distinguishes real and synthetic speech via a lightweight domain embedding, combined with real-data oversampling to stabilize adaptation under extremely limited target data, without modifying the base architecture. Experiments on LibriTTS and an in-house dataset with two ZS-TTS sources demonstrate that our approach improves speaker similarity over naive synthetic augmentation while preserving intelligibility and perceptual quality.

Method: Develops a containerized backend system that decouples interface from inference, allowing users to plug in custom models for AI-assisted pre-annotation. Supports multi-user consensus and model isolation.

Result: Created an open-source tool (LabelBuddy) that enables collaborative audio annotation with AI assistance, addressing the bottleneck in audio annotation infrastructure for ML/LALM development.

Conclusion: LabelBuddy provides a flexible, extensible platform for audio annotation that can integrate with emerging LALMs and autonomous agents, advancing human-aligned audio representation learning.

Abstract: The advancement of Machine learning (ML), Large Audio Language Models (LALMs), and autonomous AI agents in Music Information Retrieval (MIR) necessitates a shift from static tagging to rich, human-aligned representation learning. However, the scarcity of open-source infrastructure capable of capturing the subjective nuances of audio annotation remains a critical bottleneck. This paper introduces \textbf{LabelBuddy}, an open-source collaborative auto-tagging audio annotation tool designed to bridge the gap between human intent and machine understanding. Unlike static tools, it decouples the interface from inference via containerized backends, allowing users to plug in custom models for AI-assisted pre-annotation. We describe the system architecture, which supports multi-user consensus, containerized model isolation, and a roadmap for extending agents and LALMs. Code available at https://github.com/GiannisProkopiou/gsoc2022-Label-buddy.

Method: Introduces Latent-Control Heads (LatCHs) that operate directly in latent space, avoiding expensive decoder steps. Uses selective TFG (Time-Frequency Guidance) and requires minimal training (7M parameters, ~4 hours).

Result: Demonstrates effective control over intensity, pitch, and beats (and combinations) while maintaining audio quality. Achieves far lower computational costs than standard end-to-end guidance methods.

Conclusion: LatCHs provide a practical solution for fine-grained controllable audio generation with minimal computational overhead, balancing precision and audio fidelity for real-world applications.

Abstract: Generative audio requires fine-grained controllable outputs, yet most existing methods require model retraining on specific controls or inference-time controls (\textit{e.g.}, guidance) that can also be computationally demanding. By examining the bottlenecks of existing guidance-based controls, in particular their high cost-per-step due to decoder backpropagation, we introduce a guidance-based approach through selective TFG and Latent-Control Heads (LatCHs), which enables controlling latent audio diffusion models with low computational overhead. LatCHs operate directly in latent space, avoiding the expensive decoder step, and requiring minimal training resources (7M parameters and $\approx$ 4 hours of training). Experiments with Stable Audio Open demonstrate effective control over intensity, pitch, and beats (and a combination of those) while maintaining generation quality. Our method balances precision and audio fidelity with far lower computational costs than standard end-to-end guidance. Demo examples can be found at https://zacharynovack.github.io/latch/latch.html.

Method: 1) Created CMI-Pref-Pseudo (110k pseudo-labeled samples) and CMI-Pref (human-annotated corpus) datasets; 2) Developed CMI-RewardBench unified benchmark for evaluating music reward models; 3) Built CMI-RMs - parameter-efficient reward models capable of processing heterogeneous multimodal inputs.

Result: CMI-RMs show strong correlation with human judgments on musicality and alignment. They enable effective inference-time scaling via top-k filtering. The system provides comprehensive evaluation across musicality, text-music alignment, and compositional instruction alignment.

Conclusion: The paper establishes a complete ecosystem for music reward modeling under multimodal instructions, bridging the gap between generation and evaluation. The resources (datasets, benchmarks, models) are publicly available to advance multimodal music generation research.

Abstract: While music generation models have evolved to handle complex multimodal inputs mixing text, lyrics, and reference audio, evaluation mechanisms have lagged behind. In this paper, we bridge this critical gap by establishing a comprehensive ecosystem for music reward modeling under Compositional Multimodal Instruction (CMI), where the generated music may be conditioned on text descriptions, lyrics, and audio prompts. We first introduce CMI-Pref-Pseudo, a large-scale preference dataset comprising 110k pseudo-labeled samples, and CMI-Pref, a high-quality, human-annotated corpus tailored for fine-grained alignment tasks. To unify the evaluation landscape, we propose CMI-RewardBench, a unified benchmark that evaluates music reward models on heterogeneous samples across musicality, text-music alignment, and compositional instruction alignment. Leveraging these resources, we develop CMI reward models (CMI-RMs), a parameter-efficient reward model family capable of processing heterogeneous inputs. We evaluate their correlation with human judgments scores on musicality and alignment on CMI-Pref along with previous datasets. Further experiments demonstrate that CMI-RM not only correlates strongly with human judgments, but also enables effective inference-time scaling via top-k filtering. The necessary training data, benchmarks, and reward models are publicly available.

Method: Proposes MeanFlowSE that learns average velocity over finite intervals using Jacobian-vector product to instantiate MeanFlow identity. Uses local training objective supervising finite-interval displacement while maintaining consistency with instantaneous-field constraint.

Result: On VoiceBank-DEMAND, single-step model achieves strong intelligibility, fidelity, and perceptual quality with substantially lower computational cost than multistep baselines. No knowledge distillation or external teachers required.

Conclusion: Provides efficient, high-fidelity framework for real-time generative speech enhancement with open-source implementation available.

Abstract: Multistep inference is a bottleneck for real-time generative speech enhancement because flow- and diffusion-based systems learn an instantaneous velocity field and therefore rely on iterative ordinary differential equation (ODE) solvers. We introduce MeanFlowSE, a conditional generative model that learns the average velocity over finite intervals along a trajectory. Using a Jacobian-vector product (JVP) to instantiate the MeanFlow identity, we derive a local training objective that directly supervises finite-interval displacement while remaining consistent with the instantaneous-field constraint on the diagonal. At inference, MeanFlowSE performs single-step generation via a backward-in-time displacement, removing the need for multistep solvers; an optional few-step variant offers additional refinement. On VoiceBank-DEMAND, the single-step model achieves strong intelligibility, fidelity, and perceptual quality with substantially lower computational cost than multistep baselines. The method requires no knowledge distillation or external teachers, providing an efficient, high-fidelity framework for real-time generative speech enhancement. The proposed method is open-sourced at https://github.com/liduojia1/MeanFlowSE.

Method: Introduces a two-stream encoder with inter-stream alignment modules to improve audio comparison, and a multimodal strategy that uses symbolic representations as decoder prompts to reduce ambiguity.

Result: More than doubles F1 for missed notes on MAESTRO-E (26.8% → 56.3%) and improves extra note detection by 14.4 points (72.0% → 86.4%). Similar gains observed on CocoChorales-E and real curated data.

Conclusion: LadderSym introduces insights about comparison models that could inform sequence evaluation tasks for reinforcement learning, human skill assessment, and model evaluation.

Abstract: Music learners can greatly benefit from tools that accurately detect errors in their practice. Existing approaches typically compare audio recordings to music scores using heuristics or learnable models. This paper introduces LadderSym, a novel Transformer-based method for music error detection. LadderSym is guided by two key observations about the state-of-the-art approaches: (1) late fusion limits inter-stream alignment and cross-modality comparison capability; and (2) reliance on score audio introduces ambiguity in the frequency spectrum, degrading performance in music with concurrent notes. To address these limitations, LadderSym introduces (1) a two-stream encoder with inter-stream alignment modules to improve audio comparison capabilities and error detection F1 scores, and (2) a multimodal strategy that leverages both audio and symbolic scores by incorporating symbolic representations as decoder prompts, reducing ambiguity and improving F1 scores. We evaluate our method on the MAESTRO-E and CocoChorales-E datasets by measuring the F1 score for each note category. Compared to the previous state of the art, LadderSym more than doubles F1 for missed notes on MAESTRO-E (26.8% -> 56.3%) and improves extra note detection by 14.4 points (72.0% -> 86.4%). Similar gains are observed on CocoChorales-E. Furthermore, we also evaluate our models on real data we curated. This work introduces insights about comparison models that could inform sequence evaluation tasks for reinforcement learning, human skill assessment, and model evaluation. Code: https://github.com/ben2002chou/LadderSYM

Method: Dual-stream architecture with hierarchical layer groups using instance-local, neighborhood, and global mixing attention. Encodes knowledge graphs/hypergraphs as structured instances with role slots into a key-value repository. Uses journey-based role transport for attention conditioning that unifies KG traversal, hyperedge traversal, and sentence structure.

Result: Achieves explicit, inspectable separation between linguistic context and structured knowledge while enabling tight alignment through cross-attention, with multi-task objectives spanning masked language modeling, link prediction, and role-consistency denoising.

Conclusion: The architecture successfully maintains separable knowledge and language representations while allowing for joint training and alignment, providing an inspectable model structure that could benefit knowledge-intensive NLP tasks.

Abstract: We present a concise architecture for joint training on sentences and structured data while keeping knowledge and language representations separable. The model treats knowledge graphs and hypergraphs as structured instances with role slots and encodes them into a key-value repository that a language transformer can attend over. Attention is conditioned by journey-based role transport, which unifies edge-labeled KG traversal, hyperedge traversal, and sentence structure. We outline a dual-stream architecture, hierarchical layer groups with instance-local, neighborhood, and global mixing attention, retrieval over a separate repository, and multi-task objectives spanning masked language modeling, link prediction, and role-consistency denoising. The result is an explicit, inspectable separation between linguistic context and structured knowledge, while still enabling tight alignment through cross-attention.

Method: Three key components: 1) Trainable diagnostic system using Group Relative Policy Optimization (GRPO) to distill expert knowledge into local open-source models; 2) Read-write separated execution architecture decomposing operations into observation, reasoning, and action phases; 3) Failure Trajectory Closed-Loop Evolver that mines unsuccessful trajectories for corrective supervision.

Result: AOI achieves 66.3% best@5 success on all 86 AIOpsLab tasks (24.4 points over SOTA), locally deployed 14B model reaches 42.9% avg@1 on 63 held-out tasks surpassing Claude Sonnet 4.5, and Evolver improves end-to-end avg@5 by 4.8 points while reducing variance by 35%.

Conclusion: AOI demonstrates a practical framework for deploying LLM agents in enterprise SRE with security constraints, enabling safe automation, local knowledge distillation, and continuous improvement from failures.

Abstract: Large language model (LLM) agents offer a promising data-driven approach to automating Site Reliability Engineering (SRE), yet their enterprise deployment is constrained by three challenges: restricted access to proprietary data, unsafe action execution under permission-governed environments, and the inability of closed systems to improve from failures. We present AOI (Autonomous Operations Intelligence), a trainable multi-agent framework formulating automated operations as a structured trajectory learning problem under security constraints. Our approach integrates three key components. First, a trainable diagnostic system applies Group Relative Policy Optimization (GRPO) to distill expert-level knowledge into locally deployed open-source models, enabling preference-based learning without exposing sensitive data. Second, a read-write separated execution architecture decomposes operational trajectories into observation, reasoning, and action phases, allowing safe learning while preventing unauthorized state mutation. Third, a Failure Trajectory Closed-Loop Evolver mines unsuccessful trajectories and converts them into corrective supervision signals, enabling continual data augmentation. Evaluated on the AIOpsLab benchmark, our contributions yield cumulative gains. (1) The AOI runtime alone achieves 66.3% best@5 success on all 86 tasks, outperforming the prior state-of-the-art (41.9%) by 24.4 points. (2) Adding Observer GRPO training, a locally deployed 14B model reaches 42.9% avg@1 on 63 held-out tasks with unseen fault types, surpassing Claude Sonnet 4.5. (3) The Evolver converts 37 failed trajectories into diagnostic guidance, improving end-to-end avg@5 by 4.8 points while reducing variance by 35%.

Method: RADAR uses Singular Value Decomposition (SVD) on asymmetric distance matrices to initialize compact embeddings encoding static asymmetry. It replaces standard softmax with Sinkhorn normalization in attention mechanisms to model dynamic asymmetry in embedding interactions.

Result: Extensive experiments on synthetic and real-world benchmarks show RADAR outperforms strong baselines on both in-distribution and out-of-distribution instances, demonstrating robust generalization and superior performance for asymmetric VRPs.

Conclusion: RADAR provides a scalable neural framework that effectively handles asymmetric inputs for VRP solvers, addressing both static and dynamic asymmetry aspects to improve real-world applicability.

Abstract: Recent neural solvers have achieved strong performance on vehicle routing problems (VRPs), yet they mainly assume symmetric Euclidean distances, restricting applicability to real-world scenarios. A core challenge is encoding the relational features in asymmetric distance matrices of VRPs. Early attempts directly encoded these matrices but often failed to produce compact embeddings and generalized poorly at scale. In this paper, we propose RADAR, a scalable neural framework that augments existing neural VRP solvers with the ability to handle asymmetric inputs. RADAR addresses asymmetry from both static and dynamic perspectives. It leverages Singular Value Decomposition (SVD) on the asymmetric distance matrix to initialize compact and generalizable embeddings that inherently encode the static asymmetry in the inbound and outbound costs of each node. To further model dynamic asymmetry in embedding interactions during encoding, it replaces the standard softmax with Sinkhorn normalization that imposes joint row and column distance awareness in attention weights. Extensive experiments on synthetic and real-world benchmarks across various VRPs show that RADAR outperforms strong baselines on both in-distribution and out-of-distribution instances, demonstrating robust generalization and superior performance in solving asymmetric VRPs.

Method: GLOT constructs latent token-similarity graphs from frozen LLM outputs, refines token representations with a graph neural network, and aggregates them using a readout layer.

Result: GLOT maintains over 97% accuracy when 90% of tokens are random distractors (vs baseline collapse), competitive on GLUE/MTEB with 20x fewer parameters and 100x faster training than PEFT methods.

Conclusion: Learning over token graphs is a powerful paradigm for efficient adaptation of frozen LLMs, providing robust, structure-aware pooling with minimal computational overhead.

Abstract: Obtaining a single-vector representation from a Large Language Model’s (LLM) token-level outputs is a critical step for nearly all sentence-level tasks. However, standard pooling methods like mean or max aggregation treat tokens as an independent set, discarding the rich relational structure captured by the model’s self-attention layers and making them susceptible to signal dilution. To address this, we introduce GLOT, a lightweight, structure-aware pooling module that reframes pooling as relational learning followed by aggregation. Operating on the outputs of a frozen LLM, GLOT first constructs a latent token-similarity graph, then refines token representations with a graph neural network, and finally aggregates them using a readout layer. Experimentally, our approach is remarkably robust and efficient: on a diagnostic stress test where 90% of tokens are random distractors, GLOT maintains over 97% accuracy while baseline methods collapse. Furthermore, it is competitive with state-of-the-art techniques on benchmarks like GLUE and MTEB with 20x fewer trainable parameters and speeds up the training time by over 100x compared with parameter-efficient fine-tuning methods. Supported by a theoretical analysis of its expressive power, our work shows that learning over token graphs is a powerful paradigm for the efficient adaptation of frozen LLMs. Our code is published at https://github.com/ipsitmantri/GLOT.

Method: Introduces heterogeneous time steps (HTS) for equilibrium propagation by assigning neuron-specific time constants drawn from biologically motivated distributions, rather than using a uniform scalar time step.

Result: HTS improves training stability while maintaining competitive task performance compared to standard EP with uniform time steps.

Conclusion: Incorporating heterogeneous temporal dynamics enhances both the biological realism and robustness of equilibrium propagation, suggesting this approach improves EP’s practical utility.

Abstract: Equilibrium propagation (EP) is a biologically plausible alternative to backpropagation for training neural networks. However, existing EP models use a uniform scalar time step dt, which corresponds biologically to a membrane time constant that is heterogeneous across neurons. Here, we introduce heterogeneous time steps (HTS) for EP by assigning neuron-specific time constants drawn from biologically motivated distributions. We show that HTS improves training stability while maintaining competitive task performance. These results suggest that incorporating heterogeneous temporal dynamics enhances both the biological realism and robustness of equilibrium propagation.

Method: Proposes a simple modification to the Squint algorithm and correspondingly modifies the proof technique to establish improved regret bounds

Result: The variant achieves a regret bound resembling that shown in recent work for NormalHedge variants

Conclusion: A simple algorithmic variant with modified proof achieves competitive regret bounds for expert problems

Abstract: This short note describes a simple variant of the Squint algorithm of Koolen and Van Erven [2015] for the classic expert problem. Via an equally simple modification of their proof, we prove that this variant ensures a regret bound that resembles the one shown in a recent work by Freund et al. [2026] for a variant of the NormalHedge algorithm [Chaudhuri et al., 2009].

Method: Adapts OptiDICE (an offline RL algorithm) to handle multiple objectives by learning appropriate weights automatically. The replication study examined the original implementation, found errors in the code, and corrected them to properly evaluate the method.

Result: Theoretical claims hold, but an implementation error reduced FairDICE to standard behavior cloning in continuous environments. After correction, FairDICE can scale to complex environments and high-dimensional rewards, but is reliant on extensive hyperparameter tuning.

Conclusion: FairDICE is theoretically interesting but requires significant experimental revision. The replication shows it can work in complex settings but needs careful hyperparameter optimization, and the original experimental justification needs substantial improvement.

Abstract: Offline Reinforcement Learning (RL) is an emerging field of RL in which policies are learned solely from demonstrations. Within offline RL, some environments involve balancing multiple objectives, but existing multi-objective offline RL algorithms do not provide an efficient way to find a fair compromise. FairDICE (see arXiv:2506.08062v2) seeks to fill this gap by adapting OptiDICE (an offline RL algorithm) to automatically learn weights for multiple objectives to e.g.\ incentivise fairness among objectives. As this would be a valuable contribution, this replication study examines the replicability of claims made regarding FairDICE. We find that many theoretical claims hold, but an error in the code reduces FairDICE to standard behaviour cloning in continuous environments, and many important hyperparameters were originally underspecified. After rectifying this, we show in experiments extending the original paper that FairDICE can scale to complex environments and high-dimensional rewards, though it can be reliant on (online) hyperparameter tuning. We conclude that FairDICE is a theoretically interesting method, but the experimental justification requires significant revision.

Method: Use a gating mechanism with d+1 parameters to decide when to replace full MLP with linear surrogate. Systematic investigation across six models (162M-2.8B parameters), two architectures (GPT-2 and Pythia), and three corpora. Analyze cross-corpus correlation and contextual routing decisions.

Result: Cross-corpus correlation is zero (r < 0.05), showing nonlinearity need cannot be predicted from token identity. Most MLP computations are near-linear, achieving 25-56% linear routing at <1% perplexity cost in GPT-2. In GPT-2 Large, 11 of 36 layers beat baseline with gating. Some layers can be fully linearized at zero cost, and with training, linearizing harmful nonlinear MLPs yields 10.2-17.3% perplexity improvements.

Conclusion: Transformer MLP nonlinearity is often unnecessary, with many computations being near-linear. A gating mechanism can effectively route between linear and nonlinear computations, achieving computational savings. Some nonlinear MLPs are actively harmful and can be replaced with linear alternatives to improve performance.

Abstract: We investigate when transformer MLP nonlinearity is actually necessary. A gate with $d+1$ parameters decides when to replace the full MLP with a linear surrogate. Through systematic investigation across six models (162M-2.8B parameters), two architectures, and three corpora, we establish that nonlinearity need cannot be predicted from token identity: cross-corpus correlation is zero ($r < 0.05$). The routing decision is fully contextual. Despite weak per-instance predictability, the gate exploits a heavily skewed distribution where most MLP computations are near-linear, achieving 25-56% linear routing at <1% perplexity cost in GPT-2. In GPT-2 Large, 11 of 36 layers beat baseline with gating and no layer exceeds 3.7% all-linear cost. This success is architecture-dependent: Pythia models show higher costs, though Pythia-2.8B’s full 32-layer sweep reveals one layer that narrowly beats baseline. As a proof of concept, we progressively replace middle-layer MLPs with frozen linear matrices: 5 of 24 layers linearize at zero cost. With a full training budget, 4 linearized layers yield a 10.2% perplexity improvement – and a two-phase gated approach pushes this to 17.3%, beating a vanilla fine-tuning control and confirming that the nonlinear MLPs at these layers were actively harmful.

Method: Introduces Adversarially-Aligned Jacobian Regularization (AAJR) - a trajectory-aligned approach that controls sensitivity strictly along adversarial ascent directions. Proves AAJR yields larger admissible policy class than global constraints under mild conditions, with step-size conditions ensuring inner-loop stability.

Result: AAJR provides a structural theory for agentic robustness that decouples minimax stability from global expressivity restrictions, offering reduced nominal performance degradation compared to global constraints.

Conclusion: AAJR enables more stable and efficient minimax training for multi-agent LLM ecosystems by focusing regularization on adversarial directions rather than imposing overly conservative global constraints.

Abstract: As Large Language Models (LLMs) transition into autonomous multi-agent ecosystems, robust minimax training becomes essential yet remains prone to instability when highly non-linear policies induce extreme local curvature in the inner maximization. Standard remedies that enforce global Jacobian bounds are overly conservative, suppressing sensitivity in all directions and inducing a large Price of Robustness. We introduce Adversarially-Aligned Jacobian Regularization (AAJR), a trajectory-aligned approach that controls sensitivity strictly along adversarial ascent directions. We prove that AAJR yields a strictly larger admissible policy class than global constraints under mild conditions, implying a weakly smaller approximation gap and reduced nominal performance degradation. Furthermore, we derive step-size conditions under which AAJR controls effective smoothness along optimization trajectories and ensures inner-loop stability. These results provide a structural theory for agentic robustness that decouples minimax stability from global expressivity restrictions.

Method: Design neural network with early-exit capability and propose uncertainty-aware probabilistic framework modeling clean speech and error variance to derive probabilistic early-exit conditions based on desired signal-to-noise ratios.

Result: Early-exit capabilities can be introduced without compromising reconstruction quality; early-exit conditions are well-calibrated and lead to considerable compute savings when dynamically scaling compute at test time.

Conclusion: The proposed method enables efficient speech separation on resource-constrained devices through dynamic compute scaling while maintaining performance and interpretability.

Abstract: In recent years, deep learning-based single-channel speech separation has improved considerably, in large part driven by increasingly compute- and parameter-efficient neural network architectures. Most such architectures are, however, designed with a fixed compute and parameter budget and consequently cannot scale to varying compute demands or resources, which limits their use in embedded and heterogeneous devices such as mobile phones and hearables. To enable such use-cases we design a neural network architecture for speech separation and enhancement capable of early-exit, and we propose an uncertainty-aware probabilistic framework to jointly model the clean speech signal and error variance which we use to derive probabilistic early-exit conditions in terms of desired signal-to-noise ratios. We evaluate our methods on both speech separation and enhancement tasks where we demonstrate that early-exit capabilities can be introduced without compromising reconstruction, and that when trained on variable-length audio our early-exit conditions are well-calibrated and lead to considerable compute savings when used to dynamically scale compute at test time while remaining directly interpretable.

Method: Proposes Graph Hopfield Networks with an energy function coupling associative memory retrieval and graph Laplacian smoothing. Uses gradient descent on this joint energy to produce iterative updates that interleave Hopfield retrieval with Laplacian propagation.

Result: Achieves up to 2.0 percentage point improvements on sparse citation networks and up to 5 percentage points additional robustness under feature masking. The iterative energy-descent architecture itself provides strong inductive bias, with all variants outperforming standard baselines on Amazon co-purchase graphs.

Conclusion: Graph Hopfield Networks effectively combine memory retrieval with graph propagation, offering performance gains, robustness, and flexibility for both homophilous and heterophilous graphs through tuning.

Abstract: We introduce Graph Hopfield Networks, whose energy function couples associative memory retrieval with graph Laplacian smoothing for node classification. Gradient descent on this joint energy yields an iterative update interleaving Hopfield retrieval with Laplacian propagation. Memory retrieval provides regime-dependent benefits: up to 2.0~pp on sparse citation networks and up to 5 pp additional robustness under feature masking; the iterative energy-descent architecture itself is a strong inductive bias, with all variants (including the memory-disabled NoMem ablation) outperforming standard baselines on Amazon co-purchase graphs. Tuning enables graph sharpening for heterophilous benchmarks without architectural changes.

Method: Analyzed 36 audio models using voxel-wise and component-wise regression, and representation similarity analysis with two fMRI datasets. Evaluated models on 6 auditory tasks from HEAREval benchmark. Tracked brain similarity evolution during pretraining of EnCodecMAE.

Result: Recent self-supervised audio models with strong downstream performance better predict auditory cortex activity. Found strong positive correlations (r > 0.8) between overall task performance and brain alignment. Brain similarity increases progressively during pretraining and emerges early.

Conclusion: Brain-like representations emerge as a byproduct of learning to reconstruct missing information from naturalistic audio, suggesting that optimizing for diverse downstream tasks naturally leads to more brain-aligned representations without explicit brain optimization.

Abstract: Artificial neural networks are increasingly powerful models of brain computation, yet it remains unclear whether improving their performance in downstream tasks also makes their internal representations more similar to brain signals. To address this question in the auditory domain, we quantified the alignment between the internal representations of 36 different audio models and brain activity from two independent fMRI datasets. Using voxel-wise and component-wise regression, and representation similarity analysis, we found that recent self-supervised audio models with strong performance in diverse downstream tasks are better predictors of auditory cortex activity than previously studied models. To assess the quality of the audio representations, we evaluated these models in 6 auditory tasks from the HEAREval benchmark, spanning music, speech, and environmental sounds. This revealed strong positive Pearson correlations (r > 0.8) between a model’s overall task performance and its alignment with brain representations. Finally, we analyzed the evolution of the similarity between audio and brain representations during the pretraining of EnCodecMAE, a recent audio representation model. We discovered that brain similarity increases progressively and emerges early during pretraining, despite the model not being explicitly optimized for this objective. This suggests that brain-like representations can be an emergent byproduct of learning to reconstruct missing information from naturalistic audio data.

Method: 1) Train same network on two disjoint datasets and compare mutual distances of generated samples and their similarity to training data to quantify bias. 2) Use a controlled hierarchical data model with exact scores and ground-truth statistics to precisely characterize bias onset. 3) Analyze sequential feature learning in deep networks where coarse structure is learned early (data-independent) while finer features are resolved later (increasingly dependent on individual training samples).

Result: Demonstrated presence of biased generalization on real images using quantitative bias measures. Showed that early stopping at test loss minimum, while optimal under standard generalization criteria, may be insufficient for privacy-critical applications due to models producing samples too similar to training data.

Conclusion: Standard practice of stopping at test loss minimum doesn’t guarantee privacy-safe generalization. Models enter a phase where they continue improving test metrics while becoming increasingly biased toward reproducing training data characteristics, revealing a fundamental tension between generalization quality and privacy preservation in generative modeling.

Abstract: Generalization in generative modeling is defined as the ability to learn an underlying distribution from a finite dataset and produce novel samples, with evaluation largely driven by held-out performance and perceived sample quality. In practice, training is often stopped at the minimum of the test loss, taken as an operational indicator of generalization. We challenge this viewpoint by identifying a phase of biased generalization during training, in which the model continues to decrease the test loss while favoring samples with anomalously high proximity to training data. By training the same network on two disjoint datasets and comparing the mutual distances of generated samples and their similarity to training data, we introduce a quantitative measure of bias and demonstrate its presence on real images. We then study the mechanism of bias, using a controlled hierarchical data model where access to exact scores and ground-truth statistics allows us to precisely characterize its onset. We attribute this phenomenon to the sequential nature of feature learning in deep networks, where coarse structure is learned early in a data-independent manner, while finer features are resolved later in a way that increasingly depends on individual training samples. Our results show that early stopping at the test loss minimum, while optimal under standard generalization criteria, may be insufficient for privacy-critical applications.

Method: Proposes Objective-Aware Surrogate Initialization (OASI) which seeds surrogate optimization with Pareto-biased solutions generated via multi-objective simulated annealing. This initializes the surrogate conditioning process with a bias toward feasible accuracy-memory trade-offs, avoiding SRAM-violating configurations.

Result: OASI improves hypervolume and convergence robustness over Latin hypercube, Sobol, and random initializations under the same budget constraints on TinyML KWS problems. Hardware-in-the-loop experiments on STM32 microcontrollers verify deployable and memory-feasible models without extra optimization costs.

Conclusion: OASI provides an effective initialization method for Bayesian optimization in TinyML applications, enabling better trade-offs between accuracy and memory constraints for microcontroller deployment of keyword spotting models.

Abstract: Voice-triggered interfaces rely on keyword spotting (KWS) models that must operate continuously under strict memory, latency, and energy constraints on microcontroller-class hardware. Designing such models therefore requires not only high recognition accuracy but also predictable deployability within limited Flash and SRAM budgets. Bayesian optimization is known to handle accuracy-efficiency trade-offs effectively in multi-objective optimization; however, it is highly sensitive to initialization, particularly in the low-budget regimes of TinyML model optimization. We propose Objective-Aware Surrogate Initialization (OASI), which seeds surrogate optimization with Pareto-biased solutions generated via multi-objective simulated annealing. Unlike space-filling or heuristic warm-start methods, OASI initializes the surrogate conditioning process with a bias toward feasible accuracy-memory trade-offs, thus avoiding SRAM-violating configurations. OASI improves hypervolume and convergence robustness over Latin hypercube, Sobol, and random initializations under the same budget constraints on a TinyML KWS problem. Hardware-in-the-loop experiments on STM32 microcontrollers verify the existence of deployable and memory-feasible models without incurring extra optimization costs.

Method: Analyzed state-of-the-art models (Qwen2.5-Math-7B) using novel faithfulness metrics to evaluate reasoning pathways, examining correlation between reasoning quality and correctness, scaling effects from 1.5B to 7B parameters, and computational strategies in latent reasoning.

Result: Models achieve 61% accuracy but only 18.4% of correct predictions use stable, faithful reasoning; 81.6% emerge through computationally inconsistent pathways; 8.8% are silent failures; reasoning quality shows weak negative correlation with correctness; scaling provides zero accuracy benefit on evaluated subset; latent reasoning employs diverse computational strategies.

Conclusion: Benchmark accuracy masks computational unreliability, demanding evaluation reforms that measure stability beyond single-sample metrics to ensure reliable deployment in critical applications.

Abstract: Mathematical reasoning models are widely deployed in education, automated tutoring, and decision support systems despite exhibiting fundamental computational instabilities. We demonstrate that state-of-the-art models (Qwen2.5-Math-7B) achieve 61% accuracy through a mixture of reliable and unreliable reasoning pathways: 18.4% of correct predictions employ stable, faithful reasoning while 81.6% emerge through computationally inconsistent pathways. Additionally, 8.8% of all predictions are silent failures – confident yet incorrect outputs. Through comprehensive analysis using novel faithfulness metrics, we reveal: (1) reasoning quality shows weak negative correlation with correctness (r=-0.21, p=0.002), reflecting a binary classification threshold artifact rather than a monotonic inverse relationship; (2) scaling from 1.5B to 7B parameters (4.7x increase) provides zero accuracy benefit on our evaluated subset (6% of GSM8K), requiring validation on the complete benchmark; and (3) latent reasoning employs diverse computational strategies, with ~20% sharing CoT-like patterns. These findings highlight that benchmark accuracy can mask computational unreliability, demanding evaluation reforms measuring stability beyond single-sample metrics.

Method: Combines augmentation method (to handle delayed observations) with upper confidence bound (UCB) approach. Formulates the problem as a special case of MDPs with transition dynamics decomposing into known and unknown structured components.

Result: Derives regret bound of $\tilde{\mathcal{O}}(H \sqrt{D_{\max} SAK})$ for tabular MDPs, where $S$ and $A$ are state/action spaces, $H$ is horizon, $K$ is episodes, $D_{\max}$ is maximum delay. Provides matching lower bound up to logarithmic factors, showing optimality.

Conclusion: The proposed algorithm is optimal for RL with delayed state observations. The analytical framework for MDPs with decomposed transition dynamics may have broader applications beyond delayed observation settings.

Abstract: We study reinforcement learning with delayed state observation, where the agent observes the current state after some random number of time steps. We propose an algorithm that combines the augmentation method and the upper confidence bound approach. For tabular Markov decision processes (MDPs), we derive a regret bound of $\tilde{\mathcal{O}}(H \sqrt{D_{\max} SAK})$, where $S$ and $A$ are the cardinalities of the state and action spaces, $H$ is the time horizon, $K$ is the number of episodes, and $D_{\max}$ is the maximum length of the delay. We also provide a matching lower bound up to logarithmic factors, showing the optimality of our approach. Our analytical framework formulates this problem as a special case of a broader class of MDPs, where their transition dynamics decompose into a known component and an unknown but structured component. We establish general results for this abstract setting, which may be of independent interest.

Method: Machine learning framework that encodes system design and renewable trends, using a single agent to generalize dynamic operations across configurations and scenarios, enabling rapid parallel evaluation.

Result: Achieves near-optimal performance compared to RL baselines with lower computational costs than mathematical programming. Applied to European e-methanol sites, showing optimal system loads vary by location (50-200+ MW) with production costs of 1.0-1.2 USD/kg.

Conclusion: MasCOR enables rapid screening of feasible design spaces with operational policies, providing site-specific guidance from system design to real-time operation for e-fuel production systems.

Abstract: E-fuels are promising long-term energy carriers supporting the net-zero transition. However, the large combinatorial design-operation spaces under renewable uncertainty make the use of mathematical programming impractical for co-optimizing e-fuel production systems. Here, we present MasCOR, a machine-learning-assisted co-optimization framework that learns from global operational trajectories. By encoding system design and renewable trends, a single MasCOR agent generalizes dynamic operation across diverse configurations and scenarios, substantially simplifying design-operation co-optimization under uncertainty. Benchmark comparisons against state-of-the-art reinforcement learning baselines demonstrate near-optimal performance, while computational costs are substantially lower than those of mathematical programming, enabling rapid parallel evaluation of designs within the co-optimization loop. This framework enables rapid screening of feasible design spaces together with corresponding operational policies. When applied to four potential European sites targeting e-methanol production, MasCOR shows that most locations benefit from reducing system load below 50 MW to achieve carbon-neutral methanol production, with production costs of 1.0-1.2 USD per kg. In contrast, Dunkirk (France), with limited renewable availability and high grid prices, favors system loads above 200 MW and expanded storage to exploit dynamic grid exchange and hydrogen sales to the market. These results underscore the value of the MasCOR framework for site-specific guidance from system design to real-time operation.

Method: Three complementary approaches: 1) Analysis showing small device variations cause disproportionate accuracy degradation in safety-critical workloads; 2) SWIM selective write-verify mechanism applying verification only where most impactful; 3) Learning-centric solution training neural networks with right-censored Gaussian noise to align training with hardware-induced variability.

Result: Demonstrates critical gap between average-case evaluations and worst-case behavior, shows SWIM improves reliability while maintaining efficiency advantages, and enables robust deployment without excessive hardware overhead through noise-aware training.

Conclusion: Cross-layer co-design bridging device physics, architecture, and learning algorithms is essential for dependable, efficient neural inference on emerging memory technologies, enabling adoption in safety- and reliability-critical systems.

Abstract: Compute-in-memory (CiM) architectures promise significant improvements in energy efficiency and throughput for deep neural network acceleration by alleviating the von Neumann bottleneck. However, their reliance on emerging non-volatile memory devices introduces device-level non-idealities-such as write variability, conductance drift, and stochastic noise-that fundamentally challenge reliability, predictability, and safety, especially in safety-critical applications. This talk examines the reliability limits of CiM-based neural accelerators and presents a series of techniques that bridge device physics, architecture, and learning algorithms to address these challenges. We first demonstrate that even small device variations can lead to disproportionately large accuracy degradation and catastrophic failures in safety-critical inference workloads, revealing a critical gap between average-case evaluations and worst-case behavior. Building on this insight, we introduce SWIM, a selective write-verify mechanism that strategically applies verification only where it is most impactful, significantly improving reliability while maintaining CiM’s efficiency advantages. Finally, we explore a learning-centric solution that improves realistic worst-case performance by training neural networks with right-censored Gaussian noise, aligning training assumptions with hardware-induced variability and enabling robust deployment without excessive hardware overhead. Together, these works highlight the necessity of cross-layer co-design for CiM accelerators and provide a principled path toward dependable, efficient neural inference on emerging memory technologies-paving the way for their adoption in safety- and reliability-critical systems.

Method: Define and analyze perceptual manifolds (PMs) - spaces of inputs confidently assigned to classes by networks. Compare dimensionalities of neural network PMs vs. human concepts, and test geometric hypothesis about adversarial example origins.

Result: Neural network PMs have orders of magnitude higher dimensionality than human concepts, creating exponential misalignment. Even robust networks remain exponentially misaligned. Robust accuracy and distance to PMs negatively correlate with PM dimension.

Conclusion: High dimensionality of machine perceptual manifolds is a major impediment to adversarial robustness. Adversarial robustness requires dimensional alignment between machine and human PMs, connecting alignment research with adversarial examples.

Abstract: Adversarial attacks - input perturbations imperceptible to humans that fool neural networks - remain both a persistent failure mode in machine learning, and a phenomenon with mysterious origins. To shed light, we define and analyze a network’s perceptual manifold (PM) for a class concept as the space of all inputs confidently assigned to that class by the network. We find, strikingly, that the dimensionalities of neural network PMs are orders of magnitude higher than those of natural human concepts. Since volume typically grows exponentially with dimension, this suggests exponential misalignment between machines and humans, with exponentially many inputs confidently assigned to concepts by machines but not humans. Furthermore, this provides a natural geometric hypothesis for the origin of adversarial examples: because a network’s PM fills such a large region of input space, any input will be very close to any class concept’s PM. Our hypothesis thus suggests that adversarial robustness cannot be attained without dimensional alignment of machine and human PMs, and therefore makes strong predictions: both robust accuracy and distance to any PM should be negatively correlated with the PM dimension. We confirmed these predictions across 18 different networks of varying robust accuracy. Crucially, we find even the most robust networks are still exponentially misaligned, and only the few PMs whose dimensionality approaches that of human concepts exhibit alignment to human perception. Our results connect the fields of alignment and adversarial examples, and suggest the curse of high dimensionality of machine PMs is a major impediment to adversarial robustness.

Method: Propose OrbEvo based on equivariant graph transformer architecture with two variants: OrbEvo-WF (wavefunction pooling) and OrbEvo-DM (density matrix interaction). Use equivariant conditioning to encode external electric field strength/direction, breaking SO(3) to SO(2) symmetry. OrbEvo-DM encodes density matrix via tensor contraction. Employ autoregressive rollout with error accumulation limiting training strategy.

Result: Models trained on TDDFT datasets (5,000 QM9 molecules and 1,500 malonaldehyde configurations) accurately capture quantum dynamics including time-dependent wavefunctions, dipole moments, and optical absorption spectra.

Conclusion: OrbEvo successfully learns to evolve electronic wavefunctions in TDDFT simulations, providing an efficient alternative to conventional propagation methods while maintaining physical accuracy for excited state quantum dynamics.

Abstract: We aim to learn wavefunctions simulated by time-dependent density functional theory (TDDFT), which can be efficiently represented as linear combination coefficients of atomic orbitals. In real-time TDDFT, the electronic wavefunctions of a molecule evolve over time in response to an external excitation, enabling first-principles predictions of physical properties such as optical absorption, electron dynamics, and high-order response. However, conventional real-time TDDFT relies on time-consuming propagation of all occupied states with fine time steps. In this work, we propose OrbEvo, which is based on an equivariant graph transformer architecture and learns to evolve the full electronic wavefunction coefficients across time steps. First, to account for external field, we design an equivariant conditioning to encode both strength and direction of external electric field and break the symmetry from SO(3) to SO(2). Furthermore, we design two OrbEvo models, OrbEvo-WF and OrbEvo-DM, using wavefunction pooling and density matrix as interaction method, respectively. Motivated by the central role of the density functional in TDDFT, OrbEvo-DM encodes the density matrix aggregated from all occupied electronic states into feature vectors via tensor contraction, providing a more intuitive approach to learn the time evolution operator. We adopt a training strategy specifically tailored to limit the error accumulation of time-dependent wavefunctions over autoregressive rollout. To evaluate our approach, we generate TDDFT datasets consisting of 5,000 different molecules in the QM9 dataset and 1,500 molecular configurations of the malonaldehyde molecule in the MD17 dataset. Results show that our OrbEvo model accurately captures quantum dynamics of excited states under external field, including time-dependent wavefunctions, time-dependent dipole moment, and optical absorption spectra.

Method: Created MMAI Gym - a comprehensive framework with molecular data formats, modalities, and task-specific training recipes. Used this to train an efficient Liquid Foundation Model (LFM) specialized for molecular science.

Result: The LFM achieves near specialist-level performance across key drug discovery tasks (molecular optimization, ADMET prediction, retrosynthesis, etc.), outperforming larger general-purpose models while being more efficient.

Conclusion: Smaller, purpose-trained foundation models can outperform larger general-purpose models in scientific domains like drug discovery when trained with domain-specific frameworks and data.

Abstract: General-purpose large language models (LLMs) that rely on in-context learning do not reliably deliver the scientific understanding and performance required for drug discovery tasks. Simply increasing model size or introducing reasoning tokens does not yield significant performance gains. To address this gap, we introduce the MMAI Gym for Science, a one-stop shop molecular data formats and modalities as well as task-specific reasoning, training, and benchmarking recipes designed to teach foundation models the ’language of molecules’ in order to solve practical drug discovery problems. We use MMAI Gym to train an efficient Liquid Foundation Model (LFM) for these applications, demonstrating that smaller, purpose-trained foundation models can outperform substantially larger general-purpose or specialist models on molecular benchmarks. Across essential drug discovery tasks - including molecular optimization, ADMET property prediction, retrosynthesis, drug-target activity prediction, and functional group reasoning - the resulting model achieves near specialist-level performance and, in the majority of settings, surpasses larger models, while remaining more efficient and broadly applicable in the domain.

Method: Proposes Q-Measure-Learning which learns a signed empirical measure supported on visited state-action pairs and reconstructs Q-values via kernel integration. Uses coupled stochastic approximation to jointly estimate the stationary distribution of the behavior chain and the Q-measure, resulting in O(n) memory and computation per iteration.

Result: Proves almost sure sup-norm convergence of the induced Q-function to the fixed point of a kernel-smoothed Bellman operator under uniform ergodicity. Bounds approximation error between this limit and optimal Q* as function of kernel bandwidth. Demonstrates performance in two-item inventory control experiments.

Conclusion: Q-Measure-Learning provides an efficient alternative to function approximation for continuous-state RL, with theoretical guarantees and practical computational benefits for online learning from single trajectories.

Abstract: We study reinforcement learning in infinite-horizon discounted Markov decision processes with continuous state spaces, where data are generated online from a single trajectory under a Markovian behavior policy. To avoid maintaining an infinite-dimensional, function-valued estimate, we propose the novel Q-Measure-Learning, which learns a signed empirical measure supported on visited state-action pairs and reconstructs an action-value estimate via kernel integration. The method jointly estimates the stationary distribution of the behavior chain and the Q-measure through coupled stochastic approximation, leading to an efficient weight-based implementation with $O(n)$ memory and $O(n)$ computation cost per iteration. Under uniform ergodicity of the behavior chain, we prove almost sure sup-norm convergence of the induced Q-function to the fixed point of a kernel-smoothed Bellman operator. We also bound the approximation error between this limit and the optimal $Q^*$ as a function of the kernel bandwidth. To assess the performance of our proposed algorithm, we conduct RL experiments in a two-item inventory control setting.

Method: Bilevel optimization framework with inner loop for adaptation on self-generated examples and outer loop for meta-learning data-attribution signals and rewards; synthetic data optimized with scalable meta-gradients that backpropagate downstream loss through inner updates.

Result: Experiments on mathematical reasoning show MASS learns to synthesize per-instance curricula that yield effective, data-efficient test-time adaptation.

Conclusion: MASS enables LLMs to self-adapt at test time through meta-learned synthetic data generation, demonstrating effective test-time adaptation capabilities.

Abstract: As strong general reasoners, large language models (LLMs) encounter diverse domains and tasks, where the ability to adapt and self-improve at test time is valuable. We introduce MASS, a meta-learning framework that enables LLMs to self-adapt by generating problem-specific synthetic training data and performing targeted self-updates optimized for downstream performance at inference time. We train this behavior end-to-end via bilevel optimization: an inner loop adapts on self-generated examples while an outer loop meta-learns data-attribution signals and rewards post-update task performance. The synthetic data is optimized with scalable meta-gradients, backpropagating the downstream loss through the inner updates to reward useful generations. Experiments on mathematical reasoning show that MASS learns to synthesize per-instance curricula that yield effective, data-efficient test-time adaptation.

Method: Proposed logit-level uncertainty quantification framework using temperature-controlled output logits. Evaluated three VLMs with metrics including cosine similarity, Jensen-Shannon divergence, and Kullback-Leibler divergence.

Result: VLMs show high stochastic sensitivity (low cosine similarity, high divergence metrics) with near-maximal temperature impacts and abrupt uncertainty transitions. Pathology-specific PRISM model maintains near-deterministic behavior with minimal temperature effects.

Conclusion: Logit-level uncertainty quantification is crucial for evaluating trustworthiness of VLMs in histopathology applications, revealing significant differences between general VLMs and domain-specific models.

Abstract: Vision-Language Models (VLMs) with their multimodal capabilities have demonstrated remarkable success in almost all domains, including education, transportation, healthcare, energy, finance, law, and retail. Nevertheless, the utilization of VLMs in healthcare applications raises crucial concerns due to the sensitivity of large-scale medical data and the trustworthiness of these models (reliability, transparency, and security). This study proposes a logit-level uncertainty quantification (UQ) framework for histopathology image analysis using VLMs to deal with these concerns. UQ is evaluated for three VLMs using metrics derived from temperature-controlled output logits. The proposed framework demonstrates a critical separation in uncertainty behavior. While VLMs show high stochastic sensitivity (cosine similarity (CS) $<0.71$ and $<0.84$, Jensen-Shannon divergence (JS) $<0.57$ and $<0.38$, and Kullback-Leibler divergence (KL) $<0.55$ and $<0.35$, respectively for mean values of VILA-M3-8B and LLaVA-Med v1.5), near-maximal temperature impacts ($Δ_T \approx 1.00$), and displaying abrupt uncertainty transitions, particularly for complex diagnostic prompts. In contrast, the pathology-specific PRISM model maintains near-deterministic behavior (mean CS $>0.90$, JS $<0.10$, KL $<0.09$) and significantly minimal temperature effects across all prompt complexities. These findings emphasize the importance of logit-level uncertainty quantification to evaluate trustworthiness in histopathology applications utilizing VLMs.

Method: Built natively on Apple’s MLX framework, providing six neuron models (LIF, IF, Izhikevich, Adaptive LIF, Synaptic, Alpha), four surrogate gradient functions, four spike encoding methods (including EEG-specific encoder), and a complete backpropagation-through-time training pipeline. Leverages MLX’s unified memory architecture, lazy evaluation, and composable function transforms.

Result: Achieved up to 97.28% accuracy on MNIST digit classification across five hyperparameter configurations and three backends, with 2.0-2.5 times faster training and 3-10 times lower GPU memory usage than snnTorch on the same M3 Max hardware.

Conclusion: mlx-snn provides an efficient, native SNN library for Apple Silicon hardware that outperforms existing solutions in training speed and memory efficiency while maintaining competitive accuracy.

Abstract: We introduce mlx-snn, the first spiking neural network (SNN) library built natively on Apple’s MLX framework. As SNN research grows rapidly, all major libraries – snnTorch, Norse, SpikingJelly, Lava – target PyTorch or custom backends, leaving Apple Silicon users without a native option. mlx-snn provides six neuron models (LIF, IF, Izhikevich, Adaptive LIF, Synaptic, Alpha), four surrogate gradient functions, four spike encoding methods (including an EEG-specific encoder), and a complete backpropagation-through-time training pipeline. The library leverages MLX’s unified memory architecture, lazy evaluation, and composable function transforms (mx.grad, mx.compile) to enable efficient SNN research on Apple Silicon hardware. We validate mlx-snn on MNIST digit classification across five hyperparameter configurations and three backends, achieving up to 97.28% accuracy with 2.0–2.5 times faster training and 3–10 times lower GPU memory than snnTorch on the same M3 Max hardware. mlx-snn is open-source under the MIT license and available on PyPI. https://github.com/D-ST-Sword/mlx-snn

Method: Theoretical analysis of directional CDNV (decision-axis variance) with non-asymptotic multiclass generalization bounds for downstream classification, linking decision-axis collapse to multitask geometry, and empirical validation across SSL objectives and synthetic multitask data.

Result: Directional CDNV collapses during pretraining even when classical CDNV remains large, theoretical bounds closely track few-shot error at practical shot sizes, and SSL learns representations with nearly orthogonal decision axes for independent tasks.

Conclusion: Directional CDNV is a key geometric quantity that explains the favorable transfer properties of self-supervised representations, enabling both strong few-shot learning within tasks and low interference across multiple tasks through decision-axis orthogonality.

Abstract: Frozen self-supervised representations often transfer well with only a few labels across many semantic tasks. We argue that a single geometric quantity, \emph{directional} CDNV (decision-axis variance), sits at the core of two favorable behaviors: strong few-shot transfer within a task, and low interference across many tasks. We show that both emerge when variability \emph{along} class-separating directions is small. First, we prove sharp non-asymptotic multiclass generalization bounds for downstream classification whose leading term is the directional CDNV. The bounds include finite-shot corrections that cleanly separate intrinsic decision-axis variability from centroid-estimation error. Second, we link decision-axis collapse to multitask geometry: for independent balanced labelings, small directional CDNV across tasks forces the corresponding decision axes to be nearly orthogonal, helping a single representation support many tasks with minimal interference. Empirically, across SSL objectives, directional CDNV collapses during pretraining even when classical CDNV remains large, and our bounds closely track few-shot error at practical shot sizes. Additionally, on synthetic multitask data, we verify that SSL learns representations whose induced decision axes are nearly orthogonal. The code and project page of the paper are available at [\href{https://dlfundamentals.github.io/directional-neural-collapse/}{project page}].

Method: RACI formulates ecosystem flux prediction as a conditional inference problem with hierarchical temporal encoding to disentangle slow regime conditioners from fast dynamic drivers, and incorporates role-aware spatial retrieval that supplies functionally similar and geographically local context for each role.

Result: RACI consistently outperforms competitive spatiotemporal baselines across multiple ecosystem types (wetlands and agricultural systems), carbon fluxes (CO₂, GPP, CH₄), and data sources (process-based simulations and observational measurements), demonstrating improved accuracy and spatial generalization under environmental heterogeneity.

Conclusion: By explicitly modeling distinct functional roles, RACI enables models to adapt predictions across diverse environmental regimes without training separate local models or relying on fixed spatial structures, offering a robust framework for ecosystem flux prediction.

Abstract: Accurate prediction of terrestrial ecosystem carbon fluxes (e.g., CO$_2$, GPP, and CH$_4$) is essential for understanding the global carbon cycle and managing its impacts. However, prediction remains challenging due to strong spatiotemporal heterogeneity: ecosystem flux responses are constrained by slowly varying regime conditions, while short-term fluctuations are driven by high-frequency dynamic forcings. Most existing learning-based approaches treat environmental covariates as a homogeneous input space, implicitly assuming a global response function, which leads to brittle generalization across heterogeneous ecosystems. In this work, we propose Role-Aware Conditional Inference (RACI), a process-informed learning framework that formulates ecosystem flux prediction as a conditional inference problem. RACI employs hierarchical temporal encoding to disentangle slow regime conditioners from fast dynamic drivers, and incorporates role-aware spatial retrieval that supplies functionally similar and geographically local context for each role. By explicitly modeling these distinct functional roles, RACI enables a model to adapt its predictions across diverse environmental regimes without training separate local models or relying on fixed spatial structures. We evaluate RACI across multiple ecosystem types (wetlands and agricultural systems), carbon fluxes (CO$_2$, GPP, CH$_4$), and data sources, including both process-based simulations and observational measurements. Across all settings, RACI consistently outperforms competitive spatiotemporal baselines, demonstrating improved accuracy and spatial generalization under pronounced environmental heterogeneity.

Method: Empirical evaluation of three fusion strategies: ensembling (combining outputs), merging (parameter averaging), and routing (input-dependent integration). Analyzed non-uniform approaches and expert selection techniques (clustering, greedy subset selection) to reduce routing complexity.

Result: Non-uniform ensembling and merging improve performance over uniform approaches, but routing offers even greater gains. Expert selection techniques (clustering, greedy subset selection) can maintain reasonable performance with minimal computational overhead.

Conclusion: Routing provides the best performance among fusion strategies despite its complexity, and efficient expert selection can mitigate computational costs, advancing understanding of model fusion for multi-task learning.

Abstract: While large language models (LLMs) fine-tuned with lightweight adapters achieve strong performance across diverse tasks, their performance on individual tasks depends on the fine-tuning strategy. Fusing independently trained models with different strengths has shown promise for multi-task learning through three main strategies: ensembling, which combines outputs from independent models; merging, which fuses model weights via parameter averaging; and routing, which integrates models in an input-dependent fashion. However, many design decisions in these approaches remain understudied, and the relative benefits of more sophisticated ensembling, merging and routing techniques are not fully understood. We empirically evaluate their trade-offs, addressing two key questions: What are the advantages of going beyond uniform ensembling or merging? And does the flexibility of routing justify its complexity? Our findings indicate that non-uniform ensembling and merging improve performance, but routing offers even greater gains. To mitigate the computational cost of routing, we analyze expert selection techniques, showing that clustering and greedy subset selection can maintain reasonable performance with minimal overhead. These insights advance our understanding of model fusion for multi-task learning.

Method: Propose online learning framework for chain-of-thought verifiers that check correctness of reasoning steps. Introduce novel extensions of Littlestone dimension to characterize mistake bounds in realizable setting. Provide optimal algorithms for Pareto-frontier optimization and minimizing asymmetric costs.

Result: Theoretical characterization of mistake bounds for learning verifiers. Algorithms for optimal verification learning. Show how learned verifiers can boost accuracy of weak provers and enable generation of proofs beyond training data.

Conclusion: Framework enables learning strong provers with small error and abstention rates under mild assumptions, addressing distribution shift challenges in mathematical reasoning verification.

Abstract: Large language models with chain-of-thought generation have demonstrated great potential for producing complex mathematical proofs. However, their reasoning can often go astray, leading to increasing interest in formal and learned verifiers. A major challenge in learning verifiers, especially when their output will be used by the prover, is that this feedback loop may produce substantial distribution shift. Motivated by this challenge, we propose an online learning framework for learning chain-of-thought verifiers that, given a problem and a sequence of reasoning steps, check the correctness of the solution. Highlighting the asymmetric role of soundness (failure in catching errors in a proof) and completeness (flagging correct proofs as wrong) mistakes of the verifier, we introduce novel extensions of the Littlestone dimension which tightly characterize the mistake bounds for learning a verifier in the realizable setting. We provide optimal algorithms for finding the Pareto-frontier (the smallest total number of mistakes given a budget of soundness mistakes) as well as minimizing a linear combination of asymmetric costs. We further show how our learned verifiers can be used to boost the accuracy of a collection of weak provers, and enable generation of proofs beyond what they were trained on. With the mild assumption that one of the provers can generate the next reasoning step correctly with some minimal probability, we show how to learn a strong prover with small error and abstention rates.

Method: Transport clustering reduces low-rank OT to a clustering problem on correspondences obtained from a full-rank transport registration step. The algorithm computes a low-rank OT plan by clustering the correspondences, providing polynomial-time constant-factor approximations for negative-type metrics and kernel costs.

Result: Theoretical analysis shows transport clustering yields (1+γ) approximation for negative-type metrics and (1+γ+√(2γ)) approximation for kernel costs, where γ∈[0,1] is the approximation ratio of optimal full-rank solution relative to low-rank optimal. Empirically outperforms existing low-rank OT solvers on synthetic benchmarks and large-scale, high-dimensional datasets.

Conclusion: Transport clustering provides an efficient polynomial-time approximation algorithm for the NP-hard low-rank optimal transport problem, with theoretical guarantees and empirical superiority over existing methods, enabling practical applications of low-rank OT’s advantages.

Abstract: Optimal transport (OT) finds a least cost transport plan between two probability distributions using a cost matrix defined on pairs of points. Unlike standard OT, which infers unstructured pointwise mappings, low-rank optimal transport explicitly constrains the rank of the transport plan to infer latent structure. This improves statistical stability and robustness, yields sharper parametric rates for estimating Wasserstein distances adaptive to the intrinsic rank, and generalizes $K$-means to co-clustering. These advantages, however, come at the cost of a non-convex and NP-hard optimization problem. We introduce transport clustering, an algorithm to compute a low-rank OT plan that reduces low-rank OT to a clustering problem on correspondences obtained from a full-rank $\textit{transport registration}$ step. We prove that this reduction yields polynomial-time, constant-factor approximation algorithms for low-rank OT: specifically, a $(1+γ)$ approximation for negative-type metrics and a $(1+γ+\sqrt{2γ},)$ approximation for kernel costs, where $γ\in [0,1]$ denotes the approximation ratio of the optimal full-rank solution relative to the low-rank optimal. Empirically, transport clustering outperforms existing low-rank OT solvers on synthetic benchmarks and large-scale, high-dimensional datasets.

Method: Two-phase hybrid approach: 1) Agents construct spatial beliefs using Log-Gaussian Cox Process (LGCP) and execute information-driven trajectories with Pathwise Mutual Information planner; 2) Transfer to Soft Actor-Critic agent with dual-channel knowledge transfer (belief state initialization and replay buffer seeding), plus variance-normalized overlap penalty for coordinated coverage.

Result: Achieves 10.8% higher cumulative reward and 38% faster convergence over baselines in multi-UAV wireless service provisioning task, with ablation studies confirming dual-channel transfer outperforms individual channels.

Conclusion: The HBRL framework effectively bridges model-based and learning-based approaches for multi-agent spatial exploration, demonstrating superior performance through synergistic combination of structured uncertainty estimation and adaptive policy learning.

Abstract: Coordinating multiple autonomous agents to explore and serve spatially heterogeneous demand requires jointly learning unknown spatial patterns and planning trajectories that maximize task performance. Pure model-based approaches provide structured uncertainty estimates but lack adaptive policy learning, while deep reinforcement learning often suffers from poor sample efficiency when spatial priors are absent. This paper presents a hybrid belief-reinforcement learning (HBRL) framework to address this gap. In the first phase, agents construct spatial beliefs using a Log-Gaussian Cox Process (LGCP) and execute information-driven trajectories guided by a Pathwise Mutual Information (PathMI) planner with multi-step lookahead. In the second phase, trajectory control is transferred to a Soft Actor-Critic (SAC) agent, warm-started through dual-channel knowledge transfer: belief state initialization supplies spatial uncertainty, and replay buffer seeding provides demonstration trajectories generated during LGCP exploration. A variance-normalized overlap penalty enables coordinated coverage through shared belief state, permitting cooperative sensing in high-uncertainty regions while discouraging redundant coverage in well-explored areas. The framework is evaluated on a multi-UAV wireless service provisioning task. Results show 10.8% higher cumulative reward and 38% faster convergence over baselines, with ablation studies confirming that dual-channel transfer outperforms either channel alone.

Method: Proposes NuMuon, which augments the Muon optimizer with a nuclear-norm constraint on the update direction. This explicitly constrains learned weights toward low-rank structure while retaining Muon’s favorable convergence properties. The approach is evaluated across billion-parameter-scale models using state-of-the-art LLM compression pipelines.

Result: NuMuon increases weight compressibility and improves post-compression model quality compared to standard Muon. Despite imposing full-rank updates, Muon-trained models already exhibit pronounced low-rank structure, and NuMuon further enhances this property while maintaining Muon’s convergence behavior.

Conclusion: NuMuon successfully bridges the gap between full-rank optimization and low-rank compressibility, offering a practical solution for LLM deployment by improving compression efficiency without sacrificing training convergence.

Abstract: The rapid progress of large language models (LLMs) is increasingly constrained by memory and deployment costs, motivating compression methods for practical deployment. Many state-of-the-art compression pipelines leverage the low-rank structure of trained weight matrices, a phenomenon often associated with the properties of popular optimizers such as Adam. In this context, Muon is a recently proposed optimizer that improves LLM pretraining via full-rank update steps, but its induced weight-space structure has not been characterized yet. In this work, we report a surprising empirical finding: despite imposing full-rank updates, Muon-trained models exhibit pronounced low-rank structure in their weight matrices and are readily compressible under standard pipelines. Motivated by this insight, we propose NuMuon, which augments Muon with a nuclear-norm constraint on the update direction, further constraining the learned weights toward low-rank structure. Across billion-parameter-scale models, we show that NuMuon increases weight compressibility and improves post-compression model quality under state-of-the-art LLM compression pipelines while retaining Muon’s favorable convergence behavior.

Method: Combines Riemannian geometry, optimal control theory, and physics to derive backpropagation as constrained optimization; introduces layerwise Riemannian metric using Woodbury matrix identity for efficient computation; develops composable Riemannian modules with convergence guarantees using nonlinear contraction theory.

Result: Provides theoretical framework with algorithmic stability guarantees of order O(κ²L/(ξμ√n)), where κ and L are Lipschitz constants, μ is mass matrix scale, and ξ bounds condition number; offers practical alternative to natural gradient descent.

Conclusion: The layerwise Riemannian metric approach provides both theoretical understanding and practical optimization benefits for neural networks, with broader applicability to modular systems optimization in biology and engineering.

Abstract: Understanding how systems built out of modular components can be jointly optimized is an important problem in biology, engineering, and machine learning. The backpropagation algorithm is one such solution and has been instrumental in the success of neural networks. Despite its empirical success, a strong theoretical understanding of it is lacking. Here, we combine tools from Riemannian geometry, optimal control theory, and theoretical physics to advance this understanding. We make three key contributions: First, we revisit the derivation of backpropagation as a constrained optimization problem and combine it with the insight that Riemannian gradient descent trajectories can be understood as the minimum of an action. Second, we introduce a recursively defined layerwise Riemannian metric that exploits the modular structure of neural networks and can be efficiently computed using the Woodbury matrix identity, avoiding the $O(n^3)$ cost of full metric inversion. Third, we develop a framework of composable ``Riemannian modules’’ whose convergence properties can be quantified using nonlinear contraction theory, providing algorithmic stability guarantees of order $O(κ^2 L/(ξμ\sqrt{n}))$ where $κ$ and $L$ are Lipschitz constants, $μ$ is the mass matrix scale, and $ξ$ bounds the condition number. Our layerwise metric approach provides a practical alternative to natural gradient descent. While we focus here on studying neural networks, our approach more generally applies to the study of systems made of modules that are optimized over time, as it occurs in biology during both evolution and development.

Method: Theoretical analysis connecting different RNN architectures to complexity classes: LRNNs as log-depth arithmetic circuits, nonlinear RNNs as solving L-complete/P-complete problems, and establishing relationships between RNN variants and automata-theoretic models.

Result: LRNNs can be viewed as log-depth arithmetic circuits with only slight depth overhead compared to transformers, while nonlinear RNNs face fundamental barriers to parallelization due to their ability to solve L-complete/P-complete problems. Also identified expressivity differences between LRNN variants.

Conclusion: The paper reveals fundamental tradeoffs between nonlinear RNNs and LRNN variants, providing theoretical foundations for designing LLM architectures that balance expressivity and parallelism optimally.

Abstract: The community is increasingly exploring linear RNNs (LRNNs) as language models, motivated by their expressive power and parallelizability. While prior work establishes the expressivity benefits of LRNNs over transformers, it is unclear what makes LRNNs – but not traditional, nonlinear RNNs – as easy to parallelize in practice as transformers. We answer this question by providing a tight connection between types of RNNs and standard complexity classes. We show that LRNNs can be viewed as log-depth (bounded fan-in) arithmetic circuits, which represents only a slight depth overhead relative to log-depth boolean circuits that transformers admit. Furthermore, we show that nonlinear RNNs can solve $\mathsf{L}$-complete problems (and even $\mathsf{P}$-complete ones, under polynomial precision), revealing a fundamental barrier to parallelizing them as efficiently as transformers. Our theory also identifies fine-grained expressivity differences between recent popular LRNN variants: permutation-diagonal LRNNs are $\mathsf{NC}^1$-complete whereas diagonal-plus-low-rank LRNNs are more expressive ($\mathsf{PNC}^1$-complete). We provide further insight by associating each type of RNN with a corresponding automata-theoretic model that it can simulate. Together, our results reveal fundamental tradeoffs between nonlinear RNNs and different variants of LRNNs, providing a foundation for designing LLM architectures that achieve an optimal balance between expressivity and parallelism.

Method: Leverages kernel approximation techniques and Reproducing Kernel Hilbert Spaces (RKHS) theory to characterize input-output function spaces. Establishes formal relationships between kernel choices and Sobolev Native Spaces, enabling capture of both function values and derivatives.

Result: Theoretical framework developed with theorems on requirements for reliable extensions and predicted approximation accuracy. Empirically validated through solving elliptic PDEs on manifolds with point-cloud representations, reporting results on accuracy and computational performance factors.

Conclusion: The proposed framework successfully extends neural operators to handle out-of-distribution inputs with theoretical guarantees, enabling reliable application to complex PDE problems involving geometric contributions.

Abstract: We develop a rigorous framework for extending neural operators to handle out-of-distribution input functions. We leverage kernel approximation techniques and provide theory for characterizing the input-output function spaces in terms of Reproducing Kernel Hilbert Spaces (RKHSs). We provide theorems on the requirements for reliable extensions and their predicted approximation accuracy. We also establish formal relationships between specific kernel choices and their corresponding Sobolev Native Spaces. This connection further allows the extended neural operators to reliably capture not only function values but also their derivatives. Our methods are empirically validated through the solution of elliptic partial differential equations (PDEs) involving operators on manifolds having point-cloud representations and handling geometric contributions. We report results on key factors impacting the accuracy and computational performance of the extension approaches.

Method: Proposes a computing framework with trainable attention modules that learn both where to probe system states and how to combine measurements to optimize predictions; implemented using spatiotemporal fields governed by PDEs.

Result: Adaptive spatial sensing significantly improves prediction accuracy on canonical chaotic benchmarks compared to non-adaptive approaches.

Conclusion: Provides perspective on attention-enhanced reservoir computing as part of broader paradigm: neural networks as trainable measurement devices for extracting information from physical dynamical systems.

Abstract: Physical dynamical systems can be viewed as natural information processors: their systems preserve, transform, and disperse input information. This perspective motivates learning not only from data generated by such systems, but also how to measure them in a way that extracts the most useful information for a given task. We propose a general computing framework for adaptive information extraction from dynamical systems, in which a trainable attention module learns both where to probe the system state and how to combine these measurements to optimize prediction performance. As a concrete instantiation, we implement this idea using a spatiotemporal field governed by a partial differential equation as the underlying dynamics, though the framework applies equally to any system whose state can be sampled. Our results show that adaptive spatial sensing significantly improves prediction accuracy on canonical chaotic benchmarks. This work provides a perspective on attention-enhanced reservoir computing as a special case of a broader paradigm: neural networks as trainable measurement devices for extracting information from physical dynamical systems.

Method: Double Deep Q-Network (DDQN) with Prioritized Experience Replay (PER), trained over 9000 episodes with reward shaping for cautious decision-making, evaluated in subject-dependent and subject-independent settings.

Result: Achieved prediction horizons of 8.72 seconds before FOG onset in subject-independent scenarios and 7.89 seconds in subject-dependent settings, demonstrating robust performance.

Conclusion: The model shows potential for integration into wearable assistive devices to provide timely, personalized interventions for mitigating FOG in PD patients.

Abstract: Freezing of Gait (FOG) is a debilitating motor symptom commonly experienced by individuals with Parkinson’s Disease (PD) which often leads to falls and reduced mobility. Timely and accurate prediction of FOG episodes is essential for enabling proactive interventions through assistive technologies. This study presents a reinforcement learning-based framework designed to identify optimal pre-FOG onset points, thereby extending the prediction horizon for anticipatory cueing systems. The model implements a Double Deep Q-Network (DDQN) architecture enhanced with Prioritized Experience Replay (PER) allowing the agent to focus learning on high-impact experiences and refine its policy. Trained over 9000 episodes with a reward shaping strategy that promotes cautious decision-making, the agent demonstrated robust performance in both subject-dependent and subject-independent evaluations. The model achieved a prediction horizon of up to 8.72 seconds prior to FOG onset in subject-independent scenarios and 7.89 seconds in subject-dependent settings. These results highlight the model’s potential for integration into wearable assistive devices, offering timely and personalized interventions to mitigate FOG in PD patients.

Method: Proposes Graph Negative Feedback Bias Correction (GNFBC) with: 1) negative feedback loss penalizing prediction sensitivity to label autocorrelation, 2) incorporating graph-agnostic model outputs as feedback term using independent node features to counteract correlation-induced bias guided by Dirichlet energy.

Result: GNFBC can be seamlessly integrated into existing GNN architectures, improving overall performance with comparable computational and memory overhead.

Conclusion: GNFBC provides a simple yet effective framework independent of specific aggregation strategies to address homophily bias in GNNs through negative feedback mechanisms.

Abstract: Graph Neural Networks (GNNs) have emerged as a powerful framework for processing graph-structured data. However, conventional GNNs and their variants are inherently limited by the homophily assumption, leading to degradation in performance on heterophilic graphs. Although substantial efforts have been made to mitigate this issue, they remain constrained by the message-passing paradigm, which is inherently rooted in homophily. In this paper, a detailed analysis of how the underlying label autocorrelation of the homophily assumption introduces bias into GNNs is presented. We innovatively leverage a negative feedback mechanism to correct the bias and propose Graph Negative Feedback Bias Correction (GNFBC), a simple yet effective framework that is independent of any specific aggregation strategy. Specifically, we introduce a negative feedback loss that penalizes the sensitivity of predictions to label autocorrelation. Furthermore, we incorporate the output of graph-agnostic models as a feedback term, leveraging independent node feature information to counteract correlation-induced bias guided by Dirichlet energy. GNFBC can be seamlessly integrated into existing GNN architectures, improving overall performance with comparable computational and memory overhead.

Method: Formalizes model-induced locality through support sets defined by computational pathways (KNN neighbors, tree leaves, GNN receptive fields). Proves Shapley computation can be projected onto these supports without loss. Develops LSMR algorithm that trains each influential subset exactly once via support mapping and pivot scheduling, and LSMR-A for larger supports using reuse-aware Monte Carlo estimation.

Result: Experiments across multiple model families demonstrate substantial reductions in retraining operations and computational speedups while maintaining high valuation fidelity compared to traditional methods.

Conclusion: Local Shapley framework successfully leverages model-induced locality to dramatically reduce computational complexity of data valuation, with theoretical guarantees and practical algorithms that scale efficiently while preserving accuracy.

Abstract: The Shapley value provides a principled foundation for data valuation, but exact computation is #P-hard due to the exponential coalition space. Existing accelerations remain global and ignore a structural property of modern predictors: for a given test instance, only a small subset of training points influences the prediction. We formalize this model-induced locality through support sets defined by the model’s computational pathway (e.g., neighbors in KNN, leaves in trees, receptive fields in GNNs), showing that Shapley computation can be projected onto these supports without loss when locality is exact. This reframes Shapley evaluation as a structured data processing problem over overlapping support-induced subset families rather than exhaustive coalition enumeration. We prove that the intrinsic complexity of Local Shapley is governed by the number of distinct influential subsets, establishing an information-theoretic lower bound on retraining operations. Guided by this result, we propose LSMR (Local Shapley via Model Reuse), an optimal subset-centric algorithm that trains each influential subset exactly once via support mapping and pivot scheduling. For larger supports, we develop LSMR-A, a reuse-aware Monte Carlo estimator that remains unbiased with exponential concentration, with runtime determined by the number of distinct sampled subsets rather than total draws. Experiments across multiple model families demonstrate substantial retraining reductions and speedups while preserving high valuation fidelity.

Method: Extends previous results by Landsman, Vanduffel, and Yao (2013) to prove new Bonnet- and Price-type theorems for q-Gaussians, simplifies forms using escort distributions, and derives gradient estimators with similar structure to Gaussian ones.

Result: Derived new Stein identity for bounded-support q-Gaussians leading to gradient estimators that are easy to implement and experimentally shown to reduce variance compared to Gaussian estimators.

Conclusion: The work simplifies application of Stein’s identity for important class of non-Gaussian distributions, potentially benefiting Bayesian deep learning and sharpness-aware minimization through variance reduction.

Abstract: Stein’s identity is a fundamental tool in machine learning with applications in generative models, stochastic optimization, and other problems involving gradients of expectations under Gaussian distributions. Less attention has been paid to problems with non-Gaussian expectations. Here, we consider the class of bounded-support $q$-Gaussians and derive a new Stein identity leading to gradient estimators which have nearly identical forms to the Gaussian ones, and which are similarly easy to implement. We do this by extending the previous results of Landsman, Vanduffel, and Yao (2013) to prove new Bonnet- and Price-type theorems for q-Gaussians. We also simplify their forms by using escort distributions. Our experiments show that bounded-support distributions can reduce the variance of gradient estimators, which can potentially be useful for Bayesian deep learning and sharpness-aware minimization. Overall, our work simplifies the application of Stein’s identity for an important class of non-Gaussian distributions.

Method: Proposes Mutual Information Unlearnable Examples (MI-UE) based on theoretical analysis showing effective unlearnable examples decrease mutual information between clean and poisoned features. Method reduces covariance by maximizing cosine similarity among intra-class features to impede model generalization.

Result: Extensive experiments show MI-UE significantly outperforms previous unlearnable example methods, even under defense mechanisms. Demonstrates that deeper networks show better unlearnability with lower mutual information between feature distributions.

Conclusion: Mutual information reduction provides a solid theoretical foundation for unlearnable examples. MI-UE offers an effective, theoretically-grounded approach to data privacy protection that outperforms existing heuristic methods.

Abstract: The volume of freely scraped data on the Internet has driven the tremendous success of deep learning. Along with this comes the growing concern about data privacy and security. Numerous methods for generating unlearnable examples have been proposed to prevent data from being illicitly learned by unauthorized deep models by impeding generalization. However, the existing approaches primarily rely on empirical heuristics, making it challenging to enhance unlearnable examples with solid explanations. In this paper, we analyze and improve unlearnable examples from a novel perspective: mutual information reduction. We demonstrate that effective unlearnable examples always decrease mutual information between clean features and poisoned features, and when the network gets deeper, the unlearnability goes better together with lower mutual information. Further, we prove from a covariance reduction perspective that minimizing the conditional covariance of intra-class poisoned features reduces the mutual information between distributions. Based on the theoretical results, we propose a novel unlearnable method called Mutual Information Unlearnable Examples (MI-UE) that reduces covariance by maximizing the cosine similarity among intra-class features, thus impeding the generalization effectively. Extensive experiments demonstrate that our approach significantly outperforms the previous methods, even under defense mechanisms.

Method: JANUS uses a DAG of Bayesian Decision Trees with Reverse-Topological Back-filling algorithm that propagates constraints backwards through the causal graph to achieve 100% constraint satisfaction without rejection sampling. It also employs Analytical Uncertainty Decomposition derived from Dirichlet priors for fast uncertainty estimation.

Result: Across 15 datasets and 523 constrained scenarios, JANUS achieves state-of-the-art fidelity (Detection Score 0.497), eliminates mode collapse on imbalanced data, provides exact handling of complex inter-column constraints where baselines fail, and enables 128x faster uncertainty estimation than Monte Carlo methods.

Conclusion: JANUS successfully addresses the synthetic data generation Quadrilemma by unifying capabilities from different approaches, providing a practical solution for high-stakes applications requiring simultaneous fidelity, control, reliability, and efficiency.

Abstract: High-stakes synthetic data generation faces a fundamental Quadrilemma: achieving Fidelity to the original distribution, Control over complex logical constraints, Reliability in uncertainty estimation, and Efficiency in computational cost – simultaneously. State-of-the-art Deep Generative Models (CTGAN, TabDDPM) excel at fidelity but rely on inefficient rejection sampling for continuous range constraints. Conversely, Structural Causal Models offer logical control but struggle with high-dimensional fidelity and complex noise inversion. We introduce JANUS (Joint Ancestral Network for Uncertainty and Synthesis), a framework that unifies these capabilities using a DAG of Bayesian Decision Trees. Our key innovation is Reverse-Topological Back-filling, an algorithm that propagates constraints backwards through the causal graph, achieving 100% constraint satisfaction on feasible constraint sets without rejection sampling. This is paired with an Analytical Uncertainty Decomposition derived from Dirichlet priors, enabling 128x faster uncertainty estimation than Monte Carlo methods. Across 15 datasets and 523 constrained scenarios, JANUS achieves state-of-the-art fidelity (Detection Score 0.497), eliminates mode collapse on imbalanced data, and provides exact handling of complex inter-column constraints (e.g., Salary_offered >= Salary_requested) where baselines fail entirely.

Method: MOOSE-Star reduces complexity from exponential to logarithmic by: 1) Training on decomposed subtasks derived from probabilistic discovery equations, 2) Using motivation-guided hierarchical search for logarithmic retrieval and subspace pruning, and 3) Employing bounded composition for robustness against retrieval noise. The TOMATO-Star dataset of 108,717 decomposed papers supports training.

Result: The framework enables tractable training and scalable inference, overcoming the “complexity wall” that brute-force sampling hits. MOOSE-Star exhibits continuous test-time scaling, showing improved performance as computational resources increase.

Conclusion: MOOSE-Star provides a unified framework for directly modeling generative reasoning in scientific discovery, making previously intractable problems solvable through decomposed training and efficient inference strategies.

Abstract: While large language models (LLMs) show promise in scientific discovery, existing research focuses on inference or feedback-driven training, leaving the direct modeling of the generative reasoning process, $P(\text{hypothesis}|\text{background})$ ($P(h|b)$), unexplored. We demonstrate that directly training $P(h|b)$ is mathematically intractable due to the combinatorial complexity ($O(N^k)$) inherent in retrieving and composing inspirations from a vast knowledge base. To break this barrier, we introduce MOOSE-Star, a unified framework enabling tractable training and scalable inference. In the best case, MOOSE-Star reduces complexity from exponential to logarithmic ($O(\log N)$) by (1) training on decomposed subtasks derived from the probabilistic equation of discovery, (2) employing motivation-guided hierarchical search to enable logarithmic retrieval and prune irrelevant subspaces, and (3) utilizing bounded composition for robustness against retrieval noise. To facilitate this, we release TOMATO-Star, a dataset of 108,717 decomposed papers (38,400 GPU hours) for training. Furthermore, we show that while brute-force sampling hits a ‘‘complexity wall,’’ MOOSE-Star exhibits continuous test-time scaling.

Method: HDT decomposes time series into sinusoidal basis via FFT, aligns core periodic structure through Harmonic Matching, and operates in frequency domain to apply updates globally without disrupting temporal dependencies.

Result: Extensive experiments show HDT achieves strong cross-architecture generalization and scalability, validating its practicality for large-scale, real-world applications.

Conclusion: HDT effectively addresses time series dataset distillation challenges by leveraging frequency domain operations and harmonic matching, enabling efficient large-scale time series forecasting.

Abstract: Time Series forecasting (TSF) in the modern era faces significant computational and storage cost challenges due to the massive scale of real-world data. Dataset Distillation (DD), a paradigm that synthesizes a small, compact dataset to achieve training performance comparable to that of the original dataset, has emerged as a promising solution. However, conventional DD methods are not tailored for time series and suffer from architectural overfitting and limited scalability. To address these issues, we propose Harmonic Dataset Distillation for Time Series Forecasting (HDT). HDT decomposes the time series into its sinusoidal basis through the FFT and aligns the core periodic structure by Harmonic Matching. Since this process operates in the frequency domain, all updates during distillation are applied globally without disrupting temporal dependencies of time series. Extensive experiments demonstrate that HDT achieves strong cross-architecture generalization and scalability, validating its practicality for large-scale, real-world applications.

Method: LEA uses clustering to enumerate sample-label mappings, then evaluates similarity between the benign model and simulated models trained under each mapping. Uses cosine similarity of first-round loss gradients for efficient model comparison. Binary-LEA reduces computational cost from n! to n³ enumerations.

Result: LEA achieves applicability across multiple VFL scenarios without auxiliary data. Binary-LEA significantly reduces computational overhead while maintaining attack effectiveness. The attack is resilient against common defenses like gradient noise and compression.

Conclusion: LEA represents a practical and effective label inference attack for VFL that works across diverse scenarios without external data, highlighting significant privacy vulnerabilities in current VFL systems.

Abstract: A typical Vertical Federated Learning (VFL) scenario involves several participants collaboratively training a machine learning model, where each party has different features for the same samples, with labels held exclusively by one party. Since labels contain sensitive information, VFL must ensure the privacy of labels. However, existing VFL-targeted label inference attacks are either limited to specific scenarios or require auxiliary data, rendering them impractical in real-world applications. We introduce a novel Label Enumeration Attack (LEA) that, for the first time, achieves applicability across multiple VFL scenarios and eschews the need for auxiliary data. Our intuition is that an adversary, employing clustering to enumerate mappings between samples and labels, ascertains the accurate label mappings by evaluating the similarity between the benign model and the simulated models trained under each mapping. To achieve that, the first challenge is how to measure model similarity, as models trained on the same data can have different weights. Drawing from our findings, we propose an efficient approach for assessing congruence based on the cosine similarity of the first-round loss gradients, which offers superior efficiency and precision compared to the comparison of parameter similarities. However, the computational cost may be prohibitive due to the necessity of training and comparing the vast number of simulated models generated through enumeration. To overcome this challenge, we propose Binary-LEA from the perspective of reducing the number of models and eliminating futile training, which lowers the number of enumerations from n! to n^3. Moreover, LEA is resilient against common defense mechanisms such as gradient noise and gradient compression.

Method: Proposes Two-Phase Suffix Imitation framework: discards initial burn-in phase data (exploration) and performs empirical risk minimization using only data from subsequent imitation phase (exploitation). Derives predictive decision loss bound characterizing bias-variance trade-off from burn-in length choice.

Result: Shows reward-free observers can achieve $\tilde O(1/\sqrt{N})$ convergence rate, matching asymptotic efficiency of fully reward-aware learners. Demonstrates passive observers can effectively uncover optimal policies from actions alone, attaining performance comparable to learners themselves.

Conclusion: The framework enables effective policy recovery from action observations despite severe information deficits, bridging gap between what observers can infer and what learners actually experience.

Abstract: We study the Inverse Contextual Bandit (ICB) problem, in which a learner seeks to optimize a policy while an observer, who cannot access the learner’s rewards and only observes actions, aims to recover the underlying problem parameters. During the learning process, the learner’s behavior naturally transitions from exploration to exploitation, resulting in non-stationary action data that poses significant challenges for the observer. To address this issue, we propose a simple and effective framework called Two-Phase Suffix Imitation. The framework discards data from an initial burn-in phase and performs empirical risk minimization using only data from a subsequent imitation phase. We derive a predictive decision loss bound that explicitly characterizes the bias-variance trade-off induced by the choice of burn-in length. Despite the severe information deficit, we show that a reward-free observer can achieve a convergence rate of $\tilde O(1/\sqrt{N})$, matching the asymptotic efficiency of a fully reward-aware learner. This result demonstrates that a passive observer can effectively uncover the optimal policy from actions alone, attaining performance comparable to that of the learner itself.

Method: Three main components: (1) Adaptive and Selective Reset (ASR) scheme that dynamically determines when and where to reset based on collapse risk, (2) importance-aware regularizer to recover essential knowledge lost due to reset, and (3) on-the-fly adaptation adjustment scheme to enhance adaptability under challenging domain shifts.

Result: Extensive experiments across long-term TTA benchmarks demonstrate effectiveness, particularly under challenging conditions with significant domain shifts.

Conclusion: The proposed ASR approach effectively addresses model collapse in continual test-time adaptation through adaptive resetting and knowledge preservation mechanisms.

Abstract: When continual test-time adaptation (TTA) persists over the long term, errors accumulate in the model and further cause it to predict only a few classes for all inputs, a phenomenon known as model collapse. Recent studies have explored reset strategies that completely erase these accumulated errors. However, their periodic resets lead to suboptimal adaptation, as they occur independently of the actual risk of collapse. Moreover, their full resets cause catastrophic loss of knowledge acquired over time, even though such knowledge could be beneficial in the future. To this end, we propose (1) an Adaptive and Selective Reset (ASR) scheme that dynamically determines when and where to reset, (2) an importance-aware regularizer to recover essential knowledge lost due to reset, and (3) an on-the-fly adaptation adjustment scheme to enhance adaptability under challenging domain shifts. Extensive experiments across long-term TTA benchmarks demonstrate the effectiveness of our approach, particularly under challenging conditions. Our code is available at https://github.com/YonseiML/asr.

Method: Introduces a Relational Prior Generator to create infinite stream of diverse synthetic RDBs from scratch, pre-trains on over 2 million synthetic single-table and relational tasks using Prior-Data Fitted Networks approach.

Result: Achieves strong few-shot performance on 19 real-world relational prediction tasks, outperforming graph-based and single-table foundation-model baselines with lightweight architecture and fast inference.

Conclusion: Demonstrates that synthetic data can effectively train relational foundation models, enabling genuine in-context learning and adaptation to any new database without fine-tuning.

Abstract: Relational Databases (RDBs) are the backbone of modern business, yet they lack foundation models comparable to those in text or vision. A key obstacle is that high-quality RDBs are private, scarce and structurally heterogeneous, making internet-scale pre-training infeasible. To overcome this data scarcity, We introduce $\textbf{RDB-PFN}$, the first relational foundation model trained purely via $\textbf{synthetic data}$. Inspired by Prior-Data Fitted Networks (PFNs) where synthetic data generated from Structural Causal Models (SCMs) enables reasoning on single tables, we design a $\textbf{Relational Prior Generator}$ to create an infinite stream of diverse RDBs from scratch. Pre-training on $\textbf{over 2 million}$ synthetic single-table and relational tasks, RDB-PFN learns to adapt to any new database instantly via genuine $\textbf{in-context learning}$. Experiments verify RDB-PFN achieves strong few-shot performance on 19 real-world relational prediction tasks, outperforming graph-based and single-table foundation-model baselines (given the same DFS-linearized inputs), while using a lightweight architecture and fast inference. The code is available at https://github.com/MuLabPKU/RDBPFN

Method: Analyzed continual learning performance of pretrained VLAs using Experience Replay (ER) with varying replay buffer sizes, comparing against smaller policy models trained from scratch.

Result: Pretrained VLAs are remarkably resistant to forgetting, with simple ER working surprisingly well (sometimes achieving zero forgetting with small replay buffers). Pretraining plays critical role in downstream continual learning performance.

Conclusion: Large-scale pretraining fundamentally changes continual learning dynamics, enabling models to continually acquire new skills over time with simple replay, and VLAs can retain relevant knowledge despite performance degradation.

Abstract: Continual learning is a long-standing challenge in robot policy learning, where a policy must acquire new skills over time without catastrophically forgetting previously learned ones. While prior work has extensively studied continual learning in relatively small behavior cloning (BC) policy models trained from scratch, its behavior in modern large-scale pretrained Vision-Language-Action (VLA) models remains underexplored. In this work, we found that pretrained VLAs are remarkably resistant to forgetting compared with smaller policy models trained from scratch. Simple Experience Replay (ER) works surprisingly well on VLAs, sometimes achieving zero forgetting even with a small replay data size. Our analysis reveals that pretraining plays a critical role in downstream continual learning performance: large pretrained models mitigate forgetting with a small replay buffer size while maintaining strong forward learning capabilities. Furthermore, we found that VLAs can retain relevant knowledge from prior tasks despite performance degradation during learning new tasks. This knowledge retention enables rapid recovery of seemingly forgotten skills through finetuning. Together, these insights imply that large-scale pretraining fundamentally changes the dynamics of continual learning, enabling models to continually acquire new skills over time with simple replay. Code and more information can be found at https://ut-austin-rpl.github.io/continual-vla

Method: Proposes DSRM-HRL framework: 1) Denoising State Representation Module (DSRM) using diffusion models to recover low-entropy latent preference manifold from noisy interaction histories; 2) Hierarchical RL agent with high-level policy regulating long-term fairness trajectories and low-level policy optimizing short-term engagement under dynamic constraints.

Result: Extensive experiments on KuaiRec and KuaiRand simulators show DSRM-HRL effectively breaks the “rich-get-richer” feedback loop and achieves superior Pareto frontier between recommendation utility and exposure equity compared to existing methods.

Conclusion: The paper demonstrates that addressing state estimation failure through latent state purification and decoupled hierarchical decision-making is key to resolving the accuracy-fairness trade-off in interactive recommender systems.

Abstract: Interactive recommender systems (IRS) are increasingly optimized with Reinforcement Learning (RL) to capture the sequential nature of user-system dynamics. However, existing fairness-aware methods often suffer from a fundamental oversight: they assume the observed user state is a faithful representation of true preferences. In reality, implicit feedback is contaminated by popularity-driven noise and exposure bias, creating a distorted state that misleads the RL agent. We argue that the persistent conflict between accuracy and fairness is not merely a reward-shaping issue, but a state estimation failure. In this work, we propose \textbf{DSRM-HRL}, a framework that reformulates fairness-aware recommendation as a latent state purification problem followed by decoupled hierarchical decision-making. We introduce a Denoising State Representation Module (DSRM) based on diffusion models to recover the low-entropy latent preference manifold from high-entropy, noisy interaction histories. Built upon this purified state, a Hierarchical Reinforcement Learning (HRL) agent is employed to decouple conflicting objectives: a high-level policy regulates long-term fairness trajectories, while a low-level policy optimizes short-term engagement under these dynamic constraints. Extensive experiments on high-fidelity simulators (KuaiRec, KuaiRand) demonstrate that DSRM-HRL effectively breaks the “rich-get-richer” feedback loop, achieving a superior Pareto frontier between recommendation utility and exposure equity.

Method: Develops a maximum-margin HDC classifier by establishing a formal mathematical relation between HDC and SVMs, leveraging SVM principles to enhance HDC classification performance.

Result: The proposed maximum-margin HDC classifier significantly outperforms baseline HDC methods on several benchmark datasets, demonstrating improved classification performance.

Conclusion: The established HDC-SVM relation may inspire novel HDC methods with more hardware-oriented implementations than SVMs, enabling efficient learning solutions for resource-constrained applications.

Abstract: Overparameterized machine learning (ML) methods such as neural networks may be prohibitively resource intensive for devices with limited computational capabilities. Hyperdimensional computing (HDC) is an emerging resource efficient and low-complexity ML method that allows hardware efficient implementations of (re-)training and inference procedures. In this paper, we propose a maximum-margin HDC classifier, which significantly outperforms baseline HDC methods on several benchmark datasets. Our method leverages a formal relation between HDC and support vector machines (SVMs) that we established for the first time. Our findings may inspire novel HDC methods with potentially more hardware-oriented implementations compared to SVMs, thus enabling more efficient learning solutions for various intelligent resource-constrained applications.

Method: Introduces two metrics: Structural Responsiveness Score (SRS) to measure model sensitivity to perturbations, and Structural Compatibility Coefficient (SCC) to measure preference for fractal perturbations. Develops a structure-aware fractal perturbation injection framework (TFI) to study architectural properties’ role in backdoor injection.

Result: Model architecture significantly influences perturbation propagation and aggregation. Multi-path feature fusion networks amplify and retain fractal perturbations even under low poisoning ratios, while low structural compatibility models constrain effectiveness. SCC strongly correlates with attack success rate and can predict perturbation survivability.

Conclusion: Backdoor behaviors in federated learning depend not only on perturbation design or poisoning intensity but also on the interaction between model architecture and aggregation mechanisms. This offers new insights for structure-aware defense design.

Abstract: While federated learning protects data privacy, it also makes the model update process vulnerable to long-term stealthy perturbations. Existing studies on backdoor attacks in federated learning mainly focus on trigger design or poisoning strategies, typically assuming that identical perturbations behave similarly across different model architectures. This assumption overlooks the impact of model structure on perturbation effectiveness. From a structure-aware perspective, this paper analyzes the coupling relationship between model architectures and backdoor perturbations. We introduce two metrics, Structural Responsiveness Score (SRS) and Structural Compatibility Coefficient (SCC), to measure a model’s sensitivity to perturbations and its preference for fractal perturbations. Based on these metrics, we develop a structure-aware fractal perturbation injection framework (TFI) to study the role of architectural properties in the backdoor injection process. Experimental results show that model architecture significantly influences the propagation and aggregation of perturbations. Networks with multi-path feature fusion can amplify and retain fractal perturbations even under low poisoning ratios, while models with low structural compatibility constrain their effectiveness. Further analysis reveals a strong correlation between SCC and attack success rate, suggesting that SCC can predict perturbation survivability. These findings highlight that backdoor behaviors in federated learning depend not only on perturbation design or poisoning intensity but also on the interaction between model architecture and aggregation mechanisms, offering new insights for structure-aware defense design.

Method: Introduces k-hop fairness, a structural fairness notion that assesses disparities conditioned on graph distance between nodes. Formalizes through predictive fairness and structural bias metrics, and proposes pre- and post-processing mitigation strategies.

Result: Experiments show: (1) models strongly reproduce structural biases at different k-hops; (2) interdependence between structural biases at different hops when rewiring graphs; (3) post-processing method achieves favorable k-hop performance-fairness trade-offs compared to existing fair LP baselines.

Conclusion: K-hop fairness provides a more comprehensive structural fairness framework for link prediction that addresses limitations of dyadic fairness, with effective mitigation strategies for improving fairness-performance trade-offs.

Abstract: Link prediction (LP) plays a central role in graph-based applications, particularly in social recommendation. However, real-world graphs often reflect structural biases, most notably homophily, the tendency of nodes with similar attributes to connect. While this property can improve predictive performance, it also risks reinforcing existing social disparities. In response, fairness-aware LP methods have emerged, often seeking to mitigate these effects by promoting inter-group connections, that is, links between nodes with differing sensitive attributes (e.g., gender), following the principle of dyadic fairness. However, dyadic fairness overlooks potential disparities within the sensitive groups themselves. To overcome this issue, we propose $k$-hop fairness, a structural notion of fairness for LP, that assesses disparities conditioned on the distance between nodes in the graph. We formalize this notion through predictive fairness and structural bias metrics, and propose pre- and post-processing mitigation strategies. Experiments across standard LP benchmarks reveal: (1) a strong tendency of models to reproduce structural biases at different $k$-hops; (2) interdependence between structural biases at different hops when rewiring graphs; and (3) that our post-processing method achieves favorable $k$-hop performance-fairness trade-offs compared to existing fair LP baselines.

Method: 1) Decomposes mixed confidence distribution into positive/negative components using Gaussian Mixture Models, 2) Applies reject filter based on positive/negative samples to mitigate distribution overlap, 3) Uses SelfStepConf with step-level confidence to dynamically adjust inference and increase distribution separation.

Result: Experiments across 16 models and 5 benchmarks demonstrate significant outperformance over state-of-the-art approaches.

Conclusion: Incorporating distributional priors alongside confidence scores during voting improves answer selection reliability in large reasoning models.

Abstract: Large Reasoning Models have demonstrated remarkable performance with the advancement of test-time scaling techniques, which enhances prediction accuracy by generating multiple candidate responses and selecting the most reliable answer. While prior work has analyzed that internal model signals like confidence scores can partly indicate response correctness and exhibit a distributional correlation with accuracy, such distributional information has not been fully utilized to guide answer selection. Motivated by this, we propose DistriVoting, which incorporates distributional priors as another signal alongside confidence during voting. Specifically, our method (1) first decomposes the mixed confidence distribution into positive and negative components using Gaussian Mixture Models, (2) then applies a reject filter based on positive/negative samples from them to mitigate overlap between the two distributions. Besides, to further alleviate the overlap from the perspective of distribution itself, we propose SelfStepConf, which uses step-level confidence to dynamically adjust inference process, increasing the separation between the two distributions to improve the reliability of confidences in voting. Experiments across 16 models and 5 benchmarks demonstrate that our method significantly outperforms state-of-the-art approaches.

Method: PatchDecomp divides input time series into subsequences (patches) and generates predictions by aggregating the contributions of each patch, enabling clear attribution of each patch’s influence on the final prediction, including those from exogenous variables.

Result: Experiments on multiple benchmark datasets show that PatchDecomp provides predictive performance comparable to recent forecasting methods while offering both quantitative and qualitative interpretability through visualization of patch-wise contributions.

Conclusion: PatchDecomp successfully addresses the trade-off between accuracy and interpretability in time series forecasting by providing a method that maintains competitive predictive performance while offering transparent explanations of its predictions.

Abstract: Time series forecasting, which predicts future values from past observations, plays a central role in many domains and has driven the development of highly accurate neural network models. However, the complexity of these models often limits human understanding of the rationale behind their predictions. We propose PatchDecomp, a neural network-based time series forecasting method that achieves both high accuracy and interpretability. PatchDecomp divides input time series into subsequences (patches) and generates predictions by aggregating the contributions of each patch. This enables clear attribution of each patch, including those from exogenous variables, to the final prediction. Experiments on multiple benchmark datasets demonstrate that PatchDecomp provides predictive performance comparable to recent forecasting methods. Furthermore, we show that the model’s explanations not only influence predicted values quantitatively but also offer qualitative interpretability through visualization of patch-wise contributions.

Method: 1) Joint evidential head learns uncertainty over unified label space capturing cross-task dependencies; 2) Adjacency Discrepancy Score quantifies evidential alignment among neighboring samples; 3) Discrepancy-aware contrastive learning refines merged representations; 4) Trains debiased router for adaptive per-sample task/layer weight allocation.

Result: Extensive experiments show BD-Merging achieves superior effectiveness and robustness compared to state-of-the-art model merging baselines across diverse tasks under distribution shift conditions.

Conclusion: BD-Merging provides a principled framework for reliable model merging under distribution shift by explicitly modeling uncertainty and adaptively routing samples, addressing key limitations of existing merging methods.

Abstract: Model Merging (MM) has emerged as a scalable paradigm for multi-task learning (MTL), enabling multiple task-specific models to be integrated without revisiting the original training data. Despite recent progress, the reliability of MM under test-time distribution shift remains insufficiently understood. Most existing MM methods typically assume that test data are clean and distributionally aligned with both the training and auxiliary sources. However, this assumption rarely holds in practice, often resulting in biased predictions with degraded generalization. To address this issue, we present BD-Merging, a bias-aware unsupervised model merging framework that explicitly models uncertainty to achieve adaptive reliability under distribution shift. First, BD-Merging introduces a joint evidential head that learns uncertainty over a unified label space, capturing cross-task semantic dependencies in MM. Second, building upon this evidential foundation, we propose an Adjacency Discrepancy Score (ADS) that quantifies evidential alignment among neighboring samples. Third, guided by ADS, a discrepancy-aware contrastive learning mechanism refines the merged representation by aligning consistent samples and separating conflicting ones. Combined with general unsupervised learning, this process trains a debiased router that adaptively allocates task-specific or layer-specific weights on a per-sample basis, effectively mitigating the adverse effects of distribution shift. Extensive experiments across diverse tasks demonstrate that BD-Merging achieves superior effectiveness and robustness compared to state-of-the-art MM baselines.

Method: Hierarchical Bayesian framework for joint inference across multiple systems, using ensemble Metropolis-Adjusted Langevin Algorithm (MALA) for stable sampling. Maximum marginal likelihood estimation of neural network closure models embedded within ODE/PDE formulations. Bilevel optimization strategy to simultaneously train surrogate forward models (FNO or parametric PINNs) alongside inference to reduce computational cost.

Result: The paper develops a principled methodology for leveraging data from related physical systems to jointly estimate individual model parameters and learn shared unknown dynamics via ML-based closure models, with computational efficiency through surrogate modeling.

Conclusion: The framework enables robust inference of unknown parameters across multiple systems while learning shared dynamics, addressing computational challenges through surrogate modeling and efficient sampling techniques.

Abstract: Inverse problems are the task of calibrating models to match data. They play a pivotal role in diverse engineering applications by allowing practitioners to align models with reality. In many applications, engineers and scientists do not have a complete picture of i) the detailed properties of a system (such as material properties, geometry, initial conditions, etc.); ii) the complete laws describing all dynamics at play (such as friction laws, complicated damping phenomena, and general nonlinear interactions). In this paper, we develop a principled methodology for leveraging data from collections of distinct yet related physical systems to jointly estimate the individual model parameters of each system, and learn the shared unknown dynamics in the form of an ML-based closure model. To robustly infer the unknown parameters for each system, we employ a hierarchical Bayesian framework, which allows for the joint inference of multiple systems and their population-level statistics. To learn the closures, we use a maximum marginal likelihood estimate of a neural network embeded within the ODE/PDE formulation of the problem. To realize this framework we utilize the ensemble Metropolis-Adjusted Langevin Algorithm (MALA) for stable and efficient sampling. To mitigate the computational bottleneck of repetitive forward evaluations in solving inverse problems, we introduce a bilevel optimization strategy to simultaneously train a surrogate forward model alongside the inference. Within this framework, we evaluate and compare distinct surrogate architectures, specifically Fourier Neural Operators (FNO) and parametric Physics-Informed Neural Network (PINNs).

Method: Two-stage framework: 1) LLM generates textual descriptions of material unit cells’ geometric layouts and properties as high-level conditions; 2) Conditioned flow model decodes these textual conditions into precise continuous coordinates and unit cell parameters.

Result: Achieves competitive performance on ab initio material generation and crystal structure prediction tasks, with generated structures showing closer alignment to ground truth in both geometry and energy levels, surpassing state-of-the-art models.

Conclusion: The staged approach effectively combines LLMs’ structured reasoning with flow models’ distribution modeling capabilities, enabling fine-grained control over generation and potentially more efficient, customizable material design.

Abstract: Generative models hold great promise for accelerating material discovery but are often limited by their inflexible single-stage generative process in designing valid and diverse materials. To address this, we propose a two-stage generative framework, Lang2Str, that combines the strengths of large language models (LLMs) and flow-based models for flexible and precise material generation. Our method frames the generative process as a conditional generative task, where an LLM provides high-level conditions by generating descriptions of material unit cells’ geometric layouts and properties. These descriptions, informed by the LLM’s extensive background knowledge, ensure reasonable structure designs. A conditioned flow model then decodes these textual conditions into precise continuous coordinates and unit cell parameters. This staged approach combines the structured reasoning of LLMs and the distribution modeling capabilities of flow models. Experimental results show that our method achieves competitive performance on \textit{ab initio} material generation and crystal structure prediction tasks, with generated structures exhibiting closer alignment to ground truth in both geometry and energy levels, surpassing state-of-the-art models. The flexibility and modularity of our framework further enable fine-grained control over the generation process, potentially leading to more efficient and customizable material design.

Method: GIPO uses truncated importance sampling with a log-ratio-based Gaussian trust weight to softly damp extreme importance ratios while maintaining non-zero gradients, introducing an implicit tunable constraint on update magnitude.

Result: GIPO achieves SOTA performance among clipping-based baselines across various replay buffer sizes, exhibits superior bias-variance trade-off, high training stability, and improved sample efficiency.

Conclusion: GIPO provides an effective policy optimization method that addresses data efficiency challenges in RL for multimodal agents, particularly valuable when dealing with scarce or stale interaction data.

Abstract: Post-training with reinforcement learning (RL) has recently shown strong promise for advancing multimodal agents beyond supervised imitation. However, RL remains limited by poor data efficiency, particularly in settings where interaction data are scarce and quickly become outdated. To address this challenge, GIPO (Gaussian Importance sampling Policy Optimization) is proposed as a policy optimization objective based on truncated importance sampling, replacing hard clipping with a log-ratio-based Gaussian trust weight to softly damp extreme importance ratios while maintaining non-zero gradients. Theoretical analysis shows that GIPO introduces an implicit, tunable constraint on the update magnitude, while concentration bounds guarantee robustness and stability under finite-sample estimation. Experimental results show that GIPO achieves state-of-the-art performance among clipping-based baselines across a wide range of replay buffer sizes, from near on-policy to highly stale data, while exhibiting superior bias–variance trade-off, high training stability and improved sample efficiency.

Method: Three key components: 1) Temporal-frequency coordination mechanism for joint modeling of temporal and spectral representations, 2) Learnable multi-resolution wavelet decomposition module using parallel convolutions instead of iterative transforms, 3) Hybrid Transformer module fusing local wavelet features with global temporal dependencies.

Result: Extensive experiments on benchmark datasets show TFWaveFormer achieves state-of-the-art performance, significantly outperforming existing Transformer-based and hybrid models across multiple metrics.

Conclusion: TFWaveFormer validates the effectiveness of combining temporal-frequency analysis with wavelet decomposition for capturing complex temporal dynamics in dynamic link prediction tasks.

Abstract: Dynamic link prediction plays a crucial role in diverse applications including social network analysis, communication forecasting, and financial modeling. While recent Transformer-based approaches have demonstrated promising results in temporal graph learning, their performance remains limited when capturing complex multi-scale temporal dynamics. In this paper, we propose TFWaveFormer, a novel Transformer architecture that integrates temporal-frequency analysis with multi-resolution wavelet decomposition to enhance dynamic link prediction. Our framework comprises three key components: (i) a temporal-frequency coordination mechanism that jointly models temporal and spectral representations, (ii) a learnable multi-resolution wavelet decomposition module that adaptively extracts multi-scale temporal patterns through parallel convolutions, replacing traditional iterative wavelet transforms, and (iii) a hybrid Transformer module that effectively fuses local wavelet features with global temporal dependencies. Extensive experiments on benchmark datasets demonstrate that TFWaveFormer achieves state-of-the-art performance, outperforming existing Transformer-based and hybrid models by significant margins across multiple metrics. The superior performance of TFWaveFormer validates the effectiveness of combining temporal-frequency analysis with wavelet decomposition in capturing complex temporal dynamics for dynamic link prediction tasks.

Method: Proposes Dual-Solver, a generalization of multistep samplers with learnable parameters that continuously: (1) interpolate among prediction types, (2) select integration domains, and (3) adjust residual terms. Uses a classification-based objective with frozen pretrained classifiers (MobileNet or CLIP) to learn parameters while preserving second-order local accuracy and standard predictor-corrector structure.

Result: Improves FID and CLIP scores in low-NFE regime (3-9 NFEs) for ImageNet class-conditional generation (DiT, GM-DiT) and text-to-image generation (SANA, PixArt-α) across different backbones.

Conclusion: Dual-Solver effectively reduces the computational cost of diffusion model sampling while maintaining or improving image quality, making diffusion models more practical for real-world applications.

Abstract: Diffusion models achieve state-of-the-art image quality. However, sampling is costly at inference time because it requires a large number of function evaluations (NFEs). To reduce NFEs, classical ODE numerical methods have been adopted. Yet, the choice of prediction type and integration domain leads to different sampling behaviors. To address these issues, we introduce Dual-Solver, which generalizes multistep samplers through learnable parameters that continuously (i) interpolate among prediction types, (ii) select the integration domain, and (iii) adjust the residual terms. It retains the standard predictor-corrector structure while preserving second-order local accuracy. These parameters are learned via a classification-based objective using a frozen pretrained classifier (e.g., MobileNet or CLIP). For ImageNet class-conditional generation (DiT, GM-DiT) and text-to-image generation (SANA, PixArt-$α$), Dual-Solver improves FID and CLIP scores in the low-NFE regime ($3 \le$ NFE $\le 9$) across backbones.

Method: Theoretical modeling using multi-index and single-location regression frameworks; analysis of training dynamics under SGD; study of attention activation functions; introduction of Bayes-softmax attention.

Result: Reveals initial unspecialized phase followed by multi-stage specialization where heads sequentially align with latent signal directions; shows softmax-1 reduces noise from irrelevant heads; Bayes-softmax achieves optimal prediction performance.

Conclusion: Multi-head attention exhibits structured specialization patterns during training, and Bayes-softmax attention provides theoretically optimal performance by better handling head redundancy and noise.

Abstract: Multi-head attention enables transformer models to represent multiple attention patterns simultaneously. Empirically, head specialization emerges in distinct stages during training, while many heads remain redundant and learn similar representations. We propose a theoretical model capturing this phenomenon, based on the multi-index and single-location regression frameworks. In the first part, we analyze the training dynamics of multi-head softmax attention under SGD, revealing an initial unspecialized phase followed by a multi-stage specialization phase in which different heads sequentially align with latent signal directions. In the second part, we study the impact of attention activation functions on performance. We show that softmax-1 significantly reduces noise from irrelevant heads. Finally, we introduce the Bayes-softmax attention, which achieves optimal prediction performance in this setting.

Method: Proposes Spectral Surgery: 1) Decompose a LoRA update with SVD, 2) Estimate per-component sensitivity using gradients on a small calibration set, 3) Reweight singular values under a magnitude constraint while keeping learned directions fixed.

Result: Across Llama-3.1-8B and Qwen3-8B on four benchmarks, Spectral Surgery yields consistent gains (up to +4.4 points on CommonsenseQA and +2.4 pass@1 on HumanEval) by adjusting only ~1,000 scalar coefficients.

Conclusion: SVD-structured, low-cost parameter editing can serve as a practical route to improving trained LoRA adapters in a purely post-hoc manner, demonstrating the value of analyzing and refining adapter spectra.

Abstract: Low-Rank Adaptation (LoRA) improves downstream performance by restricting task updates to a low-rank parameter subspace, yet how this limited capacity is allocated within a trained adapter remains unclear. Through a geometric and empirical study across multiple tasks and backbones, we find that trained LoRA updates often exhibit an inefficient spectrum: task effects concentrate in a small subset of singular directions, while many remaining components are neutral or detrimental, motivating post-hoc refinement within the learned subspace. We propose Spectral Surgery, a training-free refinement that decomposes a LoRA update with SVD, estimates per-component sensitivity using gradients on a small calibration set, and reweights singular values under a magnitude constraint while keeping the learned directions fixed. Across Llama-3.1-8B and Qwen3-8B on four benchmarks, Spectral Surgery yields consistent gains (up to +4.4 points on CommonsenseQA and +2.4 pass@1 on HumanEval) by adjusting only $\approx 1{,}000$ scalar coefficients. These results demonstrate that SVD-structured, low-cost parameter editing can serve as a practical route to improving trained LoRA adapters in a purely post-hoc manner.

Method: The authors introduce an optimization-oriented risk based on ranking between near-optimal and suboptimal designs. They develop a unified theoretical framework connecting surrogate learning to final optimization, prove theoretical advantages of ranking over regression, identify distributional mismatch as the dominant error, and design a distribution-aware ranking method to reduce this mismatch.

Result: Empirical results across various tasks show that the proposed approach outperforms twenty existing methods. Both theoretical and empirical results reveal intrinsic limitations in offline MBO, showing regimes where no offline method can avoid over-optimistic extrapolation.

Conclusion: Offline MBO is fundamentally a ranking problem, not a regression problem. Ranking-based methods have theoretical advantages over regression approaches, and addressing distributional mismatch between training data and near-optimal designs is crucial for successful optimization.

Abstract: Offline model-based optimization (MBO) seeks to discover high-performing designs using only a fixed dataset of past evaluations. Most existing methods rely on learning a surrogate model via regression and implicitly assume that good predictive accuracy leads to good optimization performance. In this work, we challenge this assumption and study offline MBO from a learnability perspective. We argue that offline optimization is fundamentally a problem of ranking high-quality designs rather than accurate value prediction. Specifically, we introduce an optimization-oriented risk based on ranking between near-optimal and suboptimal designs, and develop a unified theoretical framework that connects surrogate learning to final optimization. We prove the theoretical advantages of ranking over regression, and identify distributional mismatch between the training data and near-optimal designs as the dominant error. Inspired by this, we design a distribution-aware ranking method to reduce this mismatch. Empirical results across various tasks show that our approach outperforms twenty existing methods, validating our theoretical findings. Additionally, both theoretical and empirical results reveal intrinsic limitations in offline MBO, showing a regime in which no offline method can avoid over-optimistic extrapolation.

Method: Proposes Feasibility Constrained Successive Rejects (FCSR) algorithm that identifies the best arm while ensuring feasibility, with theoretical analysis showing optimal dependence on problem parameters

Result: Derives lower bound on error probability, shows FCSR achieves optimal dependence on parameters up to constant factors, and empirically outperforms baselines while preserving feasibility guarantees

Conclusion: FCSR provides an effective solution for constrained best-arm identification in grouped bandits with theoretical guarantees and empirical superiority

Abstract: We study fixed budget constrained best-arm identification in grouped bandits, where each arm consists of multiple independent attributes with stochastic rewards. An arm is considered feasible only if all its attributes’ means are above a given threshold. The aim is to find the feasible arm with the largest overall mean. We first derive a lower bound on the error probability for any algorithm on this setting. We then propose Feasibility Constrained Successive Rejects (FCSR), a novel algorithm that identifies the best arm while ensuring feasibility. We show it attains optimal dependence on problem parameters up to constant factors in the exponent. Empirically, FCSR outperforms natural baselines while preserving feasibility guarantees.

Method: Proposes a multi-dimensional quality scoring framework with dimensions including model/cost priors, structure quality, semantic quality, query-output alignment, and agreement/uncertainty. Uses logged outputs from QA and summarization tasks to audit dimension reliability and calibrate composite scores.

Result: Shows that seemingly reasonable dimensions can be task-dependent and negatively correlated with reference quality without calibration. After removing unreliable dimensions and re-normalizing weights, the calibrated composite matches or exceeds best single-evaluator and consensus baselines.

Conclusion: The multi-dimensional quality scoring framework provides a robust quality signal for decentralized LLM inference networks, demonstrating complementary benefits when integrated with existing Proof of Quality mechanisms under adversarial conditions.

Abstract: Decentralized large language model (LLM) inference networks can pool heterogeneous compute to scale serving, but they require lightweight and incentive-compatible mechanisms to assess output quality. Prior work introduced cost-aware Proof of Quality (PoQ) and adaptive robust PoQ to allocate rewards under evaluator heterogeneity and adversarial behavior. In this paper, we focus on the quality signal itself and propose a multi-dimensional quality scoring framework that decomposes output quality into modular dimensions, including model and cost priors, structure quality, semantic quality, query-output alignment, and agreement/uncertainty. Using logged outputs from QA and summarization tasks, we systematically audit dimension reliability and show that seemingly reasonable dimensions can be task-dependent and even negatively correlated with reference quality without calibration. While the default composite underperforms a strong single semantic evaluator, ablations reveal that removing unreliable dimensions and re-normalizing weights yields a calibrated composite that matches or exceeds the best single- evaluator and consensus baselines. Finally, we integrate the composite score as a drop-in quality signal in PoQ and demonstrate complementary benefits with robust aggregation and adaptive trust weighting under adversarial evaluator attacks.

Method: Implements UMAP, t-SNE, PaCMAP, TriMap, DREAMS, CNE, and NNDescent algorithms entirely in MLX framework, with a GPU-accelerated circle-splatting renderer using scatter-add alpha blending and hardware H.264 encoding.

Result: Achieves embedding computation in 2.1-3.8 seconds for 70,000 points on Fashion-MNIST, renders 800-frame animations in 1.4 seconds, with full pipeline from raw data to video in 3.6-5.2 seconds on M3 Ultra.

Conclusion: mlx-vis provides a high-performance, dependency-light visualization library for Apple Silicon that significantly accelerates dimensionality reduction and rendering workflows.

Abstract: mlx-vis is a Python library that implements six dimensionality reduction methods and a k-nearest neighbor graph algorithm entirely in MLX, Apple’s array framework for Apple Silicon. The library provides UMAP, t-SNE, PaCMAP, TriMap, DREAMS, CNE, and NNDescent, all executing on Metal GPU through a unified fit_transform interface. Beyond embedding computation, mlx-vis includes a GPU-accelerated circle-splatting renderer that produces scatter plots and smooth animations without matplotlib, composing frames via scatter-add alpha blending on GPU and piping them to hardware H.264 encoding. On Fashion-MNIST with 70,000 points, all methods complete embedding in 2.1-3.8 seconds and render 800-frame animations in 1.4 seconds on an M3 Ultra, with the full pipeline from raw data to rendered video finishing in 3.6-5.2 seconds. The library depends only on MLX and NumPy, is released under the Apache 2.0 license, and is available at https://github.com/hanxiao/mlx-vis.

Method: Adapt Logit Diff Amplification (LDA) as an inference-time control mechanism for PLMs. LDA modifies token probabilities by amplifying the logit difference between a baseline model and a toxicity-finetuned model, requiring no retraining. The approach is evaluated across four taxonomic groups.

Result: LDA consistently reduces predicted toxicity rate (measured via ToxDL2) below the taxon-finetuned baseline while preserving biological plausibility. It maintains distributional similarity to natural proteins (Fréchet ESM Distance) and structural viability (pLDDT), unlike activation-based steering methods that degrade sequence properties.

Conclusion: LDA provides a practical safety knob for protein generators that mitigates elicited toxicity while retaining generative quality, offering an effective inference-time control mechanism for protein language models without requiring retraining.

Abstract: Protein language models (PLMs) are becoming practical tools for de novo protein design, yet their dual-use potential raises safety concerns. We show that domain adaptation to specific taxonomic groups can elicit toxic protein generation, even when toxicity is not the training objective. To address this, we adapt Logit Diff Amplification (LDA) as an inference-time control mechanism for PLMs. LDA modifies token probabilities by amplifying the logit difference between a baseline model and a toxicity-finetuned model, requiring no retraining. Across four taxonomic groups, LDA consistently reduces predicted toxicity rate (measured via ToxDL2) below the taxon-finetuned baseline while preserving biological plausibility. We evaluate quality using Fréchet ESM Distance and predicted foldability (pLDDT), finding that LDA maintains distributional similarity to natural proteins and structural viability (unlike activation-based steering methods that tend to degrade sequence properties). Our results demonstrate that LDA provides a practical safety knob for protein generators that mitigates elicited toxicity while retaining generative quality.

Method: FedCova uses mutual information maximization to create a federated lossy feature encoding objective based solely on class feature covariances with error tolerance. It constructs subspace-augmented federated classifiers and unifies three processes through covariance: feature encoding, classifier construction, and noisy label correction.

Result: Experimental results on CIFAR-10/100 and real-world noisy dataset Clothing1M demonstrate superior robustness compared to state-of-the-art methods in both symmetric and asymmetric noisy settings under heterogeneous data distribution.

Conclusion: FedCova provides a dependency-free federated covariance learning framework that enhances model robustness to noisy labels by leveraging feature covariances, eliminating the need for external clean data or device selection mechanisms.

Abstract: Noisy labels in distributed datasets induce severe local overfitting and consequently compromise the global model in federated learning (FL). Most existing solutions rely on selecting clean devices or aligning with public clean datasets, rather than endowing the model itself with robustness. In this paper, we propose FedCova, a dependency-free federated covariance learning framework that eliminates such external reliances by enhancing the model’s intrinsic robustness via a new perspective on feature covariances. Specifically, FedCova encodes data into a discriminative but resilient feature space to tolerate label noise. Built on mutual information maximization, we design a novel objective for federated lossy feature encoding that relies solely on class feature covariances with an error tolerance term. Leveraging feature subspaces characterized by covariances, we construct a subspace-augmented federated classifier. FedCova unifies three key processes through the covariance: (1) training the network for feature encoding, (2) constructing a classifier directly from the learned features, and (3) correcting noisy labels based on feature subspaces. We implement FedCova across both symmetric and asymmetric noisy settings under heterogeneous data distribution. Experimental results on CIFAR-10/100 and real-world noisy dataset Clothing1M demonstrate the superior robustness of FedCova compared with the state-of-the-art methods.

Method: The study analyzes Stable Diffusion 3 which uses three distinct text encoders. Researchers define four categories of attack targets and identify minimal encoder sets needed for each objective. They propose Multi-Encoder Lightweight aTtacks (MELT) that trains only low-rank adapters while keeping pretrained text encoder weights frozen.

Result: The method demonstrates that tuning fewer than 0.2% of total encoder parameters is sufficient for successful backdoor attacks on Stable Diffusion 3, revealing previously underexplored vulnerabilities in multi-encoder settings.

Conclusion: Multi-encoder diffusion models like Stable Diffusion 3 remain vulnerable to efficient backdoor attacks despite their increased complexity, highlighting security concerns for real-world deployment of multimodal AI systems.

Abstract: As text-to-image diffusion models become increasingly deployed in real-world applications, concerns about backdoor attacks have gained significant attention. Prior work on text-based backdoor attacks has largely focused on diffusion models conditioned on a single lightweight text encoder. However, more recent diffusion models that incorporate multiple large-scale text encoders remain underexplored in this context. Given the substantially increased number of trainable parameters introduced by multiple text encoders, an important question is whether backdoor attacks can remain both efficient and effective in such settings. In this work, we study Stable Diffusion 3, which uses three distinct text encoders and has not yet been systematically analyzed for text-encoder-based backdoor vulnerabilities. To understand the role of text encoders in backdoor attacks, we define four categories of attack targets and identify the minimal sets of encoders required to achieve effective performance for each attack objective. Based on this, we further propose Multi-Encoder Lightweight aTtacks (MELT), which trains only low-rank adapters while keeping the pretrained text encoder weight frozen. We demonstrate that tuning fewer than 0.2% of the total encoder parameters is sufficient for successful backdoor attacks on Stable Diffusion 3, revealing previously underexplored vulnerabilities in practical attack scenarios in multi-encoder settings.

Method: Analyzes the discrepancy between time-continuous dynamics and time-discrete measurement-feedback architectures, then proposes and experimentally verifies a method to reduce hyperparameter sensitivity in these systems.

Result: The study shows that measurement-feedback architectures have substantially smaller effective hyperparameter ranges than envisioned time-continuous analog Ising machines, and demonstrates a method to mitigate this sensitivity.

Conclusion: Practical operation of Ising machines faces hyperparameter sensitivity challenges due to implementation differences, but methods can be developed to reduce this sensitivity for more robust performance.

Abstract: Analog Ising machines have been proposed as heuristic hardware solvers for combinatorial optimization problems, with the potential to outperform conventional approaches, provided that their hyperparameters are carefully tuned. Their temporal evolution is often described using time-continuous dynamics. However, most experimental implementations rely on measurement-feedback architectures that operate in a time-discrete manner. We observe that in such setups, the range of effective hyperparameters is substantially smaller than in the envisioned time-continuous analog Ising machine. In this paper, we analyze this discrepancy and discuss its impact on the practical operation of Ising machines. Next, we propose and experimentally verify a method to reduce the sensitivity to hyperparameter selection of these measurement-feedback architectures.

Method: Proposes Stochastic Gradient Descent with Escalating Restarts (SGD-ER) that monitors training progress and triggers restarts when stagnation is detected, linearly escalating the learning rate to escape sharp local minima.

Result: Evaluated on CIFAR-10, CIFAR-100, and TinyImageNet with ResNet-18/34/50, VGG-16, and DenseNet-101 architectures. SGD-ER improves test accuracy by 0.5-4.5% compared to standard schedulers.

Conclusion: Convergence-aware escalating restarts benefit optimization by helping escape sharp local minima and explore flatter regions of the loss landscape for better generalization.

Abstract: Learning rate scheduling plays a critical role in the optimization of deep neural networks, directly influencing convergence speed, stability, and generalization. While existing schedulers such as cosine annealing, cyclical learning rates, and warm restarts have shown promise, they often rely on fixed or periodic triggers that are agnostic to the training dynamics, such as stagnation or convergence behavior. In this work, we propose a simple yet effective strategy, which we call Stochastic Gradient Descent with Escalating Restarts (SGD-ER). It adaptively increases the learning rate upon convergence. Our method monitors training progress and triggers restarts when stagnation is detected, linearly escalating the learning rate to escape sharp local minima and explore flatter regions of the loss landscape. We evaluate SGD-ER across CIFAR-10, CIFAR-100, and TinyImageNet on a range of architectures including ResNet-18/34/50, VGG-16, and DenseNet-101. Compared to standard schedulers, SGD-ER improves test accuracy by 0.5-4.5%, demonstrating the benefit of convergence-aware escalating restarts for better local optima.

Method: DARKFormer uses data-aligned kernel geometry that admits tractable minimal-variance proposal distributions for importance sampling. It learns random-projection covariance to efficiently implement importance-sampled positive random-feature estimators for data-aligned kernels.

Result: DARKFormer narrows the performance gap with exact softmax attention, particularly in finetuning regimes where pretrained representations are anisotropic, while maintaining linear complexity.

Conclusion: By combining random-feature efficiency with data-aware kernels, DARKFormer advances kernel-based attention for resource-constrained settings, offering better training stability and reduced variance.

Abstract: Transformers excel across domains, yet their quadratic attention complexity poses a barrier to scaling. Random-feature attention, as in Performers, can reduce this cost to linear in the sequence length by approximating the softmax kernel with positive random features drawn from an isotropic distribution. In pretrained models, however, queries and keys are typically anisotropic. This induces high Monte Carlo variance in isotropic sampling schemes unless one retrains the model or uses a large feature budget. Importance sampling can address this by adapting the sampling distribution to the input geometry, but complex data-dependent proposal distributions are often intractable. We show that by data aligning the softmax kernel, we obtain an attention mechanism which can both admit a tractable minimal-variance proposal distribution for importance sampling, and exhibits better training stability. Motivated by this finding, we introduce DARKFormer, a Data-Aware Random-feature Kernel transformer that features a data-aligned kernel geometry. DARKFormer learns the random-projection covariance, efficiently realizing an importance-sampled positive random-feature estimator for its data-aligned kernel. Empirically, DARKFormer narrows the performance gap with exact softmax attention, particularly in finetuning regimes where pretrained representations are anisotropic. By combining random-feature efficiency with data-aware kernels, DARKFormer advances kernel-based attention in resource-constrained settings.

Method: Decomposes forecasting into: 1) weather forecast model using high-resolution numerical weather prediction, 2) plant characteristic model using neural network ensemble trained on historical power data, with satellite observations as intermediate layer.

Result: MAE increases by 11-68% when using weather forecasts vs. perfect satellite observations; generalized hyperbolic and Student’s t distributions adequately fit forecast errors across lead times.

Conclusion: Two-stage decomposition enables better error analysis for stochastic optimization; identified distributions can improve uncertainty quantification in energy management applications.

Abstract: Several energy management applications rely on accurate photovoltaic generation forecasts. Common metrics like mean absolute error or root-mean-square error, omit error-distribution details needed for stochastic optimization. In addition, several approaches use weather forecasts as inputs without analyzing the source of the prediction error. To overcome this gap, we decompose forecasting into a weather forecast model for environmental parameters such as solar irradiance and temperature and a plant characteristic model that captures site-specific parameters like panel orientation, temperature influence, or regular shading. Satellite-based weather observation serves as an intermediate layer. We analyze the error distribution of the high-resolution rapid-refresh numerical weather prediction model that covers the United States as a black-box model for weather forecasting and train an ensemble of neural networks on historical power output data for the plant characteristic model. Results show mean absolute error increases by 11% and 68% for two selected photovoltaic systems when using weather forecasts instead of satellite-based ground-truth weather observations as a perfect forecast. The generalized hyperbolic and Student’s t distributions adequately fit the forecast errors across lead times.

Method: Proposes InstMeter, a predictor that leverages MCUs’ clock cycles as fundamental metrics to estimate energy and latency of DL models. The method exploits the strong linearity property of clock cycles to create simple yet accurate predictors that require minimal training data.

Result: InstMeter reduces energy prediction errors by 3× and latency prediction errors by 6.5× compared to state-of-the-art methods, while requiring 100× less training data for energy and 10× less for latency. It enables NAS to fully exploit energy budgets and identify optimal DL models with higher inference accuracy.

Conclusion: Clock cycles are effective fundamental metrics for predicting energy and latency of DL models on MCUs. InstMeter’s simplicity, accuracy, and data efficiency make it valuable for neural architecture search on resource-constrained devices, with demonstrated generalization across various MCUs, compilation settings, and application scenarios.

Abstract: Deep learning (DL) models can now run on microcontrollers (MCUs). Through neural architecture search (NAS), we can search DL models that meet the constraints of MCUs. Among various constraints, energy and latency costs of the model inference are critical metrics. To predict them, existing research relies on coarse proxies such as multiply-accumulations (MACs) and model’s input parameters, often resulting in inaccurate predictions or requiring extensive data collection. In this paper, we propose InstMeter, a predictor leveraging MCUs’ clock cycles to accurately estimate the energy and latency of DL models. Clock cycles are fundamental metrics reflecting MCU operations, directly determining energy and latency costs. Furthermore, a unique property of our predictor is its strong linearity, allowing it to be simple and accurate. We thoroughly evaluate InstMeter under different scenarios, MCUs, and software settings. Compared with state-of-the-art studies, InstMeter can reduce the energy and latency prediction errors by $3\times$ and $6.5\times$, respectively, while requiring $100\times$ and $10\times$ less training data. In the NAS scenario, InstMeter can fully exploit the energy budget, identifying optimal DL models with higher inference accuracy. We also evaluate InstMeter’s generalization performance through various experiments on three ARM MCUs (Cortex-M4, M7, M33) and one RISC-V-based MCU (ESP32-C3), different compilation options (-Os, -O2), GCC versions (v7.3, v10.3), application scenarios (keyword spotting, image recognition), dynamic voltage and frequency scaling, temperatures (21°C, 43°C), and software settings (TFLMv2.4, TFLMvCI). We will open our source codes and the MCU-specific benchmark datasets.

Method: Proposes Dynamic Pruning Policy Optimization (DPPO) with importance sampling-based correction to preserve unbiased gradient estimation during dynamic pruning. Also introduces Dense Prompt Packing, a window-based greedy strategy to maximize valid token density and hardware utilization, mitigating data sparsity from pruning.

Result: DPPO consistently accelerates training across diverse models and benchmarks. On Qwen3-4B trained on MATH, achieves 2.37× training speedup and outperforms GRPO by 3.36% in average accuracy across six mathematical reasoning benchmarks.

Conclusion: DPPO provides an effective framework for accelerating GRPO training through dynamic pruning while maintaining unbiased gradient estimation, achieving significant speed improvements without compromising model performance.

Abstract: Group Relative Policy Optimization (GRPO) effectively scales LLM reasoning but incurs prohibitive computational costs due to its extensive group-based sampling requirement. While recent selective data utilization methods can mitigate this overhead, they could induce estimation bias by altering the underlying sampling distribution, compromising theoretical rigor and convergence behavior. To address this limitation, we propose Dynamic Pruning Policy Optimization (DPPO), a framework that enables dynamic pruning while preserving unbiased gradient estimation through importance sampling-based correction. By incorporating mathematically derived rescaling factors, DPPO significantly accelerates GRPO training without altering the optimization objective of the full-batch baseline. Furthermore, to mitigate the data sparsity induced by pruning, we introduce Dense Prompt Packing, a window-based greedy strategy that maximizes valid token density and hardware utilization. Extensive experiments demonstrate that DPPO consistently accelerates training across diverse models and benchmarks. For instance, on Qwen3-4B trained on MATH, DPPO achieves 2.37$\times$ training speedup and outperforms GRPO by 3.36% in average accuracy across six mathematical reasoning benchmarks.

Method: Developed Vivaldi, a role-structured multi-agent system that explains multivariate physiological time series. Conducted controlled clinical pilot with emergency medicine experts to evaluate performance compared to zero-shot inference approaches.

Result: Agentic pipelines significantly improved non-thinking and medically fine-tuned models (+6.9 points for explanation justification, +9.7 for relevance), but degraded performance for thinking models (-14 points relevance). Tool-based computation was crucial for clinical metrics, while subjective targets showed limited improvement.

Conclusion: The value of agentic AI lies in selective externalization of computation rather than maximal reasoning complexity, with design trade-offs between utility and clarity that depend on visualization conventions and model specialization.

Abstract: Continuous physiological monitoring is central to emergency care, yet deploying trustworthy AI is challenging. While LLMs can translate complex physiological signals into clinical narratives, it is unclear how agentic systems perform relative to zero-shot inference. To address these questions, we present Vivaldi, a role-structured multi-agent system that explains multivariate physiological time series. Due to regulatory constraints that preclude live deployment, we instantiate Vivaldi in a controlled, clinical pilot to a small, highly qualified cohort of emergency medicine experts, whose evaluations reveal a context-dependent picture that contrasts with prevailing assumptions that agentic reasoning uniformly improves performance. Our experiments show that agentic pipelines substantially benefit non-thinking and medically fine-tuned models, improving expert-rated explanation justification and relevance by +6.9 and +9.7 points, respectively. Contrarily, for thinking models, agentic orchestration often degrades explanation quality, including a 14-point drop in relevance, while improving diagnostic precision (ESI F1 +3.6). We also find that explicit tool-based computation is decisive for codifiable clinical metrics, whereas subjective targets, such as pain scores and length of stay, show limited or inconsistent changes. Expert evaluation further indicates that gains in clinical utility depend on visualization conventions, with medically specialized models achieving the most favorable trade-offs between utility and clarity. Together, these findings show that the value of agentic AI lies in the selective externalization of computation and structure rather than in maximal reasoning complexity, and highlight concrete design trade-offs and learned lessons, broadly applicable to explainable AI in safety-critical healthcare settings.

Method: Five-step pipeline: (1) Sample many documents from LLMs on a given topic, (2) Extract event lists from each document, (3) Group events across documents into canonical events, (4) Construct binary indicator vectors for each document over canonical events, (5) Apply causal discovery methods to estimate candidate causal graphs.

Result: The approach produces inspectable sets of variables and candidate causal graphs that represent the causal hypotheses LLMs can plausibly assume, though it doesn’t guarantee real-world causality.

Conclusion: The proposed pipeline provides a framework for extracting and visualizing causal relationships encoded in LLMs, making implicit knowledge more inspectable and interpretable.

Abstract: Large language models (LLMs) are trained on enormous amounts of data and encode knowledge in their parameters. We propose a pipeline to elicit causal relationships from LLMs. Specifically, (i) we sample many documents from LLMs on a given topic, (ii) we extract an event list from from each document, (iii) we group events that appear across documents into canonical events, (iv) we construct a binary indicator vector for each document over canonical events, and (v) we estimate candidate causal graphs using causal discovery methods. Our approach does not guarantee real-world causality. Rather, it provides a framework for presenting the set of causal hypotheses that LLMs can plausibly assume, as an inspectable set of variables and candidate graphs.

Method: Proposed Probability Navigation Architecture (PNA) framework with thermodynamic loss function, trained SSMs and Transformers across 19 experimental phases, analyzed architectural proprioception and Universal Stopping Signature, conducted cross-task transfer experiments.

Result: Thermodynamically-trained SSMs show strong anticipatory coupling between state entropy and halt confidence (r = -0.836), Universal Stopping Signature with 2-token lead, while Transformers show no such coupling (r = -0.07). SSMs demonstrate genuine meta-cognition in cross-task transfer.

Conclusion: SSMs are thermodynamically native architectures that naturally support Markovian compression enabling computational self-awareness, with implications for efficient inference systems, while Transformers rely on syntactic pattern matching.

Abstract: We introduce the Probability Navigation Architecture (PNA) framework, which treats neural computation as navigation through a probability manifold governed by thermodynamic principles. We train State Space Models (SSMs) and Transformers with a novel thermodynamic loss function that penalizes computational waste alongside standard cross-entropy. Across 19 experimental phases, we discover that thermodynamically-trained SSMs develop architectural proprioception: a strong anticipatory coupling between recurrent state entropy and halt confidence (r = -0.836, p < 0.001) in which the halt signal leads state entropy collapse by exactly two tokens (tau = -2.0). This Universal Stopping Signature (USS) reproduces to four decimal places across random seeds and generalizes to a structurally distinct sorting task. Critically, Transformers trained identically show no such coupling (r = -0.07), demonstrating that the phenomenon is architecture-dependent. Cross-task transfer experiments confirm that SSM halt detection reflects genuine meta-cognition (zero-shot transfer F1: SSMs 64.2% vs. Transformers 69.3%; post-adaptation: SSMs 94.5% vs. Transformers 86.4%), while Transformer halt detection relies on syntactic pattern matching. A 2D hyperparameter sweep over energy penalty (alpha) and halt supervision (beta) reveals that the anticipatory coupling is continuously controllable through training, with thermodynamic pressure serving as the primary induction mechanism and explicit halt supervision as an amplifier. Our results establish that SSMs are thermodynamically native architectures whose fixed-size recurrent states naturally support the Markovian compression that enables computational self-awareness, with implications for cost-aware inference, dynamic token budgets, and confidence-based routing in production systems.

Method: The pipeline fuses multi-sensor satellite data: (1) Sentinel-2 optical chips processed into spectral indices and texture signals, (2) MODIS Level-3 ocean color and thermal indicators, and (3) learned image evidence from object detectors trained to identify bloom-like patterns. A CatBoost decision fusion model integrates these signals into calibrated HAB risk probabilities, with strict data splitting to prevent leakage and comprehensive evaluation metrics.

Result: The system produces calibrated probabilities of HAB risk with operational inference workflows and risk field viewers. Evaluation uses AUROC/AUPRC, confusion matrices, calibration curves, and drift analyses to quantify performance and distribution shifts in recent years.

Conclusion: REDNET-ML provides a reproducible machine learning pipeline for operational HAB risk detection that integrates multi-sensor satellite data and supports coastal management decisions through calibrated risk probabilities and visualization tools.

Abstract: Harmful algal blooms (HABs) can threaten coastal infrastructure, fisheries, and desalination dependent water supplies. This project (REDNET-ML) develops a reproducible machine learning pipeline for HAB risk detection along the Omani coastline using multi sensor satellite data and non leaky evaluation. The system fuses (i) Sentinel-2 optical chips (high spatial resolution) processed into spectral indices and texture signals, (ii) MODIS Level-3 ocean color and thermal indicators, and (iii) learned image evidence from object detectors trained to highlight bloom like patterns. A compact decision fusion model (CatBoost) integrates these signals into a calibrated probability of HAB risk, which is then consumed by an end to end inference workflow and a risk field viewer that supports operational exploration by site (plant) and time. The report documents the motivation, related work, methodological choices (including label mining and strict split strategies), implementation details, and a critical evaluation using AUROC/AUPRC, confusion matrices, calibration curves, and drift analyses that quantify distribution shift in recent years.

Method: Proposes Dual-Modality Multi-Stage Adversarial Safety Training (DMAST): 1) imitation learning from teacher model, 2) oracle-guided supervised fine-tuning with zero-acknowledgment strategy, 3) adversarial reinforcement learning via Group Relative Policy Optimization self-play.

Result: DMAST substantially mitigates adversarial risks while doubling task completion efficiency on out-of-distribution tasks, outperforming established training-based and prompt-based defenses with robust generalization to complex unseen environments.

Conclusion: The framework demonstrates genuine co-evolutionary progress in multimodal agent safety, showing that coordinated dual-modality attacks are more effective than text-only attacks and require specialized safety training approaches.

Abstract: Multimodal web agents that process both screenshots and accessibility trees are increasingly deployed to interact with web interfaces, yet their dual-stream architecture opens an underexplored attack surface: an adversary who injects content into the webpage DOM simultaneously corrupts both observation channels with a consistent deceptive narrative. Our vulnerability analysis on MiniWob++ reveals that attacks including a visual component far outperform text-only injections, exposing critical gaps in text-centric VLM safety training. Motivated by this finding, we propose Dual-Modality Multi-Stage Adversarial Safety Training (DMAST), a framework that formalizes the agent-attacker interaction as a two-player zero-sum Markov game and co-trains both players through a three-stage pipeline: (1) imitation learning from a strong teacher model, (2) oracle-guided supervised fine-tuning that uses a novel zero-acknowledgment strategy to instill task-focused reasoning under adversarial noise, and (3) adversarial reinforcement learning via Group Relative Policy Optimization (GRPO) self-play. On out-of-distribution tasks, DMAST substantially mitigates adversarial risks while simultaneously doubling task completion efficiency. Our approach significantly outperforms established training-based and prompt-based defenses, demonstrating genuine co-evolutionary progress and robust generalization to complex, unseen environments.

Method: Proposes a modular approach on top of state-of-the-art client selection strategies, incorporating noisy client data filtering through gradient norm thresholding using probing rounds for more effective client selection and noise detection.

Result: Demonstrates that modern client selection strategies based on local client loss tend to select clients with noisy data, degrading model performance, and shows that the proposed approach improves both model performance and sustainability in scenarios with unknown data quality.

Conclusion: The proposed gradient norm thresholding mechanism with probing rounds enables more effective client selection and noise detection, contributing to practical deployment of carbon-efficient federated learning by balancing efficiency and sustainability.

Abstract: Training large-scale Neural Networks requires substantial computational power and energy. Federated Learning enables distributed model training across geospatially distributed data centers, leveraging renewable energy sources to reduce the carbon footprint of AI training. Various client selection strategies have been developed to align the volatility of renewable energy with stable and fair model training in a federated system. However, due to the privacy-preserving nature of Federated Learning, the quality of data on client devices remains unknown, posing challenges for effective model training. In this paper, we introduce a modular approach on top to state-of-the-art client selection strategies for carbon-efficient Federated Learning. Our method enhances robustness by incorporating a noisy client data filtering, improving both model performance and sustainability in scenarios with unknown data quality. Additionally, we explore the impact of carbon budgets on model convergence, balancing efficiency and sustainability. Through extensive evaluations, we demonstrate that modern client selection strategies based on local client loss tend to select clients with noisy data, ultimately degrading model performance. To address this, we propose a gradient norm thresholding mechanism using probing rounds for more effective client selection and noise detection, contributing to the practical deployment of carbon-efficient Federated Learning.

Method: XPlore uses gradient-guided perturbations to both adjacency and node feature matrices, enabling joint optimization of edge insertions and feature modifications. It introduces a cosine similarity metric for graph embeddings to measure structural and semantic fidelity, addressing limitations of traditional distance metrics.

Result: Empirical evaluation on 13 real-world and 5 synthetic benchmarks shows up to +56.3% improvement in validity and +52.8% improvement in fidelity over state-of-the-art baselines, while maintaining competitive runtime.

Conclusion: XPlore significantly expands the counterfactual search space for GNN explanations, producing more coherent and minimal counterfactuals through its unified gradient-based framework and improved fidelity metrics.

Abstract: Graph Neural Networks (GNNs) are increasingly adopted across domains such as molecular biology and social network analysis, yet their black-box nature hinders interpretability and trust. This is especially problematic in high-stakes applications, such as predicting molecule toxicity, drug discovery, or guiding financial fraud detections, where transparent explanations are essential. Counterfactual explanations - minimal changes that flip a model’s prediction - offer a transparent lens into GNNs’ behavior. In this work, we introduce XPlore, a novel technique that significantly broadens the counterfactual search space. It consists of gradient-guided perturbations to adjacency and node feature matrices. Unlike most prior methods, which focus solely on edge deletions, our approach belongs to the growing class of techniques that optimize edge insertions and node-feature perturbations, here jointly performed under a unified gradient-based framework, enabling a richer and more nuanced exploration of counterfactuals. To quantify both structural and semantic fidelity, we introduce a cosine similarity metric for learned graph embeddings that addresses a key limitation of traditional distance-based metrics, and demonstrate that XPlore produces more coherent and minimal counterfactuals. Empirical results on 13 real-world and 5 synthetic benchmarks show up to +56.3% improvement in validity and +52.8% in fidelity over state-of-the-art baselines, while retaining competitive runtime.

Method: Combines specialized variational autoencoder with novel loss function using non-parametric nearest-neighbor density estimation to explicitly estimate and modify data distribution for independence optimization.

Result: Outperforms existing unsupervised techniques and rivals supervised methods in balancing information removal and utility across multiple datasets.

Conclusion: Proposed approach effectively learns representations independent of sensitive variables while preserving essential data characteristics, offering improved performance over existing methods.

Abstract: The ability to remove unwanted dependencies from data is crucial in various domains, including fairness, robust learning, and privacy protection. In this work, we propose an encoder-based approach that learns a representation independent of a sensitive variable but otherwise preserving essential data characteristics. Unlike existing methods that rely on decorrelation or adversarial learning, our approach explicitly estimates and modifies the data distribution to neutralize statistical dependencies. To achieve this, we combine a specialized variational autoencoder with a novel loss function driven by non-parametric nearest-neighbor density estimation, enabling direct optimization of independence. We evaluate our approach on multiple datasets, demonstrating that it can outperform existing unsupervised techniques and even rival supervised methods in balancing information removal and utility.

Method: Developed a variance-reduced EXP4-based algorithm integrated with Lyapunov optimization for online routing in multi-layer hierarchical inference systems under resource constraints and terminal-only feedback.

Result: The algorithm provides unbiased loss estimation and stable learning under sparse feedback, with regret guarantees relative to the best fixed routing policy, demonstrating improved stability and performance on large-scale multi-task workloads.

Conclusion: The proposed approach effectively addresses the challenges of hierarchical inference routing with terminal-only feedback by combining variance reduction techniques with Lyapunov optimization, enabling stable online learning in resource-constrained systems.

Abstract: Hierarchical inference systems route tasks across multiple computational layers, where each node may either finalize a prediction locally or offload the task to a node in the next layer for further processing. Learning optimal routing policies in such systems is challenging: inference loss is defined recursively across layers, while feedback on prediction error is revealed only at a terminal oracle layer. This induces a partial, policy-dependent feedback structure in which observability probabilities decay with depth, causing importance-weighted estimators to suffer from amplified variance. We study online routing for multi-layer hierarchical inference under long-term resource constraints and terminal-only feedback. We formalize the recursive loss structure and show that naive importance-weighted contextual bandit methods become unstable as feedback probability decays along the hierarchy. To address this, we develop a variance-reduced EXP4-based algorithm integrated with Lyapunov optimization, yielding unbiased loss estimation and stable learning under sparse and policy-dependent feedback. We provide regret guarantees relative to the best fixed routing policy in hindsight and establish near-optimality under stochastic arrivals and resource constraints. Experiments on large-scale multi-task workloads demonstrate improved stability and performance compared to standard importance-weighted approaches.

Method: Learn world model with uncertainty measures and quasi-optimal value function from offline data; identify suboptimal trajectories and augment with imagined optimal rollouts via MPC; train transformer policy on enriched dataset with value-guided objective; replace return-to-go with learned value function during inference.

Result: IPD significantly outperforms state-of-the-art value-based and transformer-based offline RL methods on D4RL benchmark across diverse tasks.

Conclusion: IPD effectively bridges the gap in decision transformers by incorporating planning throughout the pipeline, improving decision-making stability and performance.

Abstract: Decision transformer based sequential policies have emerged as a powerful paradigm in offline reinforcement learning (RL), yet their efficacy remains constrained by the quality of static datasets and inherent architectural limitations. Specifically, these models often struggle to effectively integrate suboptimal experiences and fail to explicitly plan for an optimal policy. To bridge this gap, we propose \textbf{Imaginary Planning Distillation (IPD)}, a novel framework that seamlessly incorporates offline planning into data generation, supervised training, and online inference. Our framework first learns a world model equipped with uncertainty measures and a quasi-optimal value function from the offline data. These components are utilized to identify suboptimal trajectories and augment them with reliable, imagined optimal rollouts generated via Model Predictive Control (MPC). A Transformer-based sequential policy is then trained on this enriched dataset, complemented by a value-guided objective that promotes the distillation of the optimal policy. By replacing the conventional, manually-tuned return-to-go with the learned quasi-optimal value function, IPD improves both decision-making stability and performance during inference. Empirical evaluations on the D4RL benchmark demonstrate that IPD significantly outperforms several state-of-the-art value-based and transformer-based offline RL methods across diverse tasks.

Method: Systematic investigation of AC optimal power flow (ACOPF) as a representative constrained optimization problem. Controlled experiments across architectures, training objectives, and system diversity to extract design principles. Development of the LUMINA framework with data processing and training pipelines.

Result: Three empirically grounded principles for scientific foundation model design: 1) learning physics-invariant representations while respecting system-specific constraints, 2) optimizing accuracy while ensuring constraint satisfaction, and 3) ensuring reliability in high-impact operating regimes.

Conclusion: The LUMINA framework enables reproducible research on physics-informed, feasibility-aware foundation models across scientific applications, addressing the unique challenges of constrained scientific systems.

Abstract: Foundation models in general promise to accelerate scientific computation by learning reusable representations across problem instances, yet constrained scientific systems, where predictions must satisfy physical laws and safety limits, pose unique challenges that stress conventional training paradigms. We derive design principles for constrained scientific foundation models through systematic investigation of AC optimal power flow (ACOPF), a representative optimization problem in power grid operations where power balance equations and operational constraints are non-negotiable. Through controlled experiments spanning architectures, training objectives, and system diversity, we extract three empirically grounded principles governing scientific foundation model design. These principles characterize three design trade-offs: learning physics-invariant representations while respecting system-specific constraints, optimizing accuracy while ensuring constraint satisfaction, and ensuring reliability in high-impact operating regimes. We present the LUMINA framework, including data processing and training pipelines to support reproducible research on physics-informed, feasibility-aware foundation models across scientific applications.

Method: Empirical reproduction on BERT-base fine-tuned on QNLI, statistical analysis of FP32 activations, evaluation of mitigation strategies including mixed precision PTQ, per-embedding-group quantization, and percentile-based calibration, plus deployment profiling on RTX 3050 GPU.

Result: Global W8A8 quantization drops accuracy from 89.66% to 54.33%; mixed precision PTQ restores accuracy to 89.42%; PEG quantization shows strong sensitivity to grouping; percentile-based calibration fails; deployment profiling shows minimal latency/memory differences.

Conclusion: PTQ failure in transformers is driven by structured channel dominance amplified through residual connections, requiring channel-aware precision allocation rather than scalar clipping alone.

Abstract: Post-training quantization (PTQ) of transformers is known to suffer from severe accuracy degradation due to structured activation outliers, as originally analyzed by Bondarenko et al. (EMNLP 2021) in work associated with Qualcomm AI Research. This paper provides a reproducible empirical reproduction and systems-level extension of that phenomenon in BERT-base fine-tuned on QNLI. When global W8A8 quantization is applied, validation accuracy drops sharply from 89.66% (FP32) to 54.33%, a decrease of 35.33 points. Statistical analysis of FP32 activations shows strongly heavy-tailed behavior that intensifies with model depth: kurtosis reaches 271 in the final layers and approximately 55% of activation energy is concentrated in the top 1% of channels. We evaluate several mitigation strategies. Mixed precision PTQ restores accuracy close to the FP32 baseline (89.42%). Per-embedding-group (PEG) quantization shows strong sensitivity to grouping structure, improving accuracy from 66.12% with three groups to 86.18% with four groups. In contrast, percentile-based calibration, even at thresholds between 99.0 and 99.99, fails to recover accuracy (about 50.54%), indicating that large activation channels encode structured signal rather than rare noise. Deployment profiling on an RTX 3050 GPU shows minimal differences in latency and memory usage across methods (median latency about 58-59 ms; VRAM usage about 484-486 MB), highlighting the importance of hardware-aware evaluation. Overall, the results show that PTQ failure in transformers is primarily driven by structured channel dominance amplified through residual connections. Effective mitigation therefore requires channel-aware precision allocation rather than scalar clipping alone.

Method: Proposes a new kernel-based additive model combining coherence loss with group-sparse regularization for nonlinear classification. Uses partial dependence plots to visualize group-wise additive effects of each feature group from radiomics analysis of B-mode ultrasound images.

Result: The method accurately and interpretably assesses plaques, revealing a strong association between plaque texture features from ultrasound images and clinical risk. Analyzed 500 plaques from the CREST-2 multi-center clinical trial.

Conclusion: The proposed approach provides an effective framework for identifying radiomics-based markers from ultrasound images that are linked with high-risk carotid plaques, offering both accuracy and interpretability for clinical applications.

Abstract: Accurate characterization of carotid plaques is critical for stroke prevention in patients with carotid stenosis. We analyze 500 plaques from CREST-2, a multi-center clinical trial, to identify radiomics-based markers from B-mode ultrasound images linked with high-risk. We propose a new kernel-based additive model, combining coherence loss with group-sparse regularization for nonlinear classification. Group-wise additive effects of each feature group are visualized using partial dependence plots. Results indicate our method accurately and interpretably assesses plaques, revealing a strong association between plaque texture and clinical risk.

Method: PTOPOFL replaces gradient communication with 48-dimensional persistent homology (PH) feature vectors - compact topological descriptors of model loss landscapes. These PH diagrams provide shape summaries with many-to-one structure that makes inversion attacks provably ill-posed. The server performs topology-guided personalized aggregation: clients are clustered by Wasserstein similarity between their PH diagrams, intra-cluster models are weighted by topological similarity, and clusters are blended with global consensus.

Result: PTOPOFL achieves the highest AUC scores (0.841 and 0.910) in both healthcare (8 hospitals, 2 adversarial) and pathological benchmark (10 clients) settings compared to FedAvg, FedProx, SCAFFOLD, and pFedMe. It reduces reconstruction risk by a factor of 4.5 relative to gradient sharing while maintaining superior performance on non-IID data.

Conclusion: PTOPOFL successfully addresses both privacy and performance challenges in federated learning by using topological descriptors instead of gradients. The framework provides provable privacy guarantees through information-contraction theorems and demonstrates practical effectiveness in real-world scenarios with non-IID data distributions.

Abstract: Federated learning (FL) faces two structural tensions: gradient sharing enables data-reconstruction attacks, while non-IID client distributions degrade aggregation quality. We introduce PTOPOFL, a framework that addresses both challenges simultaneously by replacing gradient communication with topological descriptors derived from persistent homology (PH). Clients transmit only 48-dimensional PH feature vectors-compact shape summaries whose many-to-one structure makes inversion provably ill-posed-rather than model gradients. The server performs topology-guided personalised aggregation: clients are clustered by Wasserstein similarity between their PH diagrams, intra-cluster models are topology-weighted,and clusters are blended with a global consensus. We prove an information-contraction theorem showing that PH descriptors leak strictly less mutual information per sample than gradients under strongly convex loss functions, and we establish linear convergence of the Wasserstein-weighted aggregation scheme with an error floor strictly smaller than FedAvg. Evaluated against FedAvg, FedProx, SCAFFOLD, and pFedMe on a non-IID healthcare scenario (8 hospitals, 2 adversarial) and a pathological benchmark (10 clients), PTOPOFL achieves AUC 0.841 and 0.910 respectively-the highest in both settings-while reducing reconstruction risk by a factor of 4.5 relative to gradient sharing. Code is publicly available at https://github.com/MorillaLab/TopoFederatedL and data at https://doi.org/10.5281/zenodo.18827595.

Method: Forward-looking and rolling evaluations on Bitcoin transaction data, analyzing temporal nonstationarity effects on cost-sensitive enforcement thresholds and comparing to dynamically optimal benchmarks.

Result: Strong static classification metrics substantially overstate real-world regulatory effectiveness. Temporal nonstationarity causes pronounced instability in enforcement thresholds, leading to large and persistent excess regulatory losses. The core failure stems from miscalibration of decision rules rather than declining predictive accuracy.

Conclusion: Fixed AML enforcement policies are fragile in evolving digital asset markets, motivating the need for loss-based evaluation frameworks for regulatory oversight rather than relying on static metrics.

Abstract: We study the deployment performance of machine learning based enforcement systems used in cryptocurrency anti money laundering (AML). Using forward looking and rolling evaluations on Bitcoin transaction data, we show that strong static classification metrics substantially overstate real world regulatory effectiveness. Temporal nonstationarity induces pronounced instability in cost sensitive enforcement thresholds, generating large and persistent excess regulatory losses relative to dynamically optimal benchmarks. The core failure arises from miscalibration of decision rules rather than from declining predictive accuracy per se. These findings underscore the fragility of fixed AML enforcement policies in evolving digital asset markets and motivate loss-based evaluation frameworks for regulatory oversight.

Method: Analyzes flow matching critics through theoretical formalization and empirical validation, comparing them to monolithic critics in high-UTD online RL settings, focusing on integration-based value readout and dense velocity supervision during training.

Result: Flow-matching critics outperform monolithic critics by 2× in final performance and around 5× in sample efficiency, particularly in challenging settings where loss of plasticity is problematic, while maintaining stable learning.

Conclusion: The success of flow matching for Q-value estimation stems from two key mechanisms: test-time error recovery through iterative integration and plastic feature learning from dense velocity supervision, rather than distributional RL principles.

Abstract: Recent work shows that flow matching can be effective for scalar Q-value function estimation in reinforcement learning (RL), but it remains unclear why or how this approach differs from standard critics. Contrary to conventional belief, we show that their success is not explained by distributional RL, as explicitly modeling return distributions can reduce performance. Instead, we argue that the use of integration for reading out values and dense velocity supervision at each step of this integration process for training improves TD learning via two mechanisms. First, it enables robust value prediction through \emph{test-time recovery}, whereby iterative computation through integration dampens errors in early value estimates as more integration steps are performed. This recovery mechanism is absent in monolithic critics. Second, supervising the velocity field at multiple interpolant values induces more \emph{plastic} feature learning within the network, allowing critics to represent non-stationary TD targets without discarding previously learned features or overfitting to individual TD targets encountered during training. We formalize these effects and validate them empirically, showing that flow-matching critics substantially outperform monolithic critics (2$\times$ in final performance and around 5$\times$ in sample efficiency) in settings where loss of plasticity poses a challenge e.g., in high-UTD online RL problems, while remaining stable during learning.

Method: Benchmark out-of-distribution transfer on two discontinuity-dominated regimes: shock-driven multi-material interface dynamics (PLI) and dynamic fracture/failure evolution (FRAC). Formulate downstream task as terminal-state prediction (long-horizon map from first snapshot to final state). Evaluate two pretrained PDE foundation models (POSEIDON and MORPH) using unified protocol, comparing fine-tuning vs training from scratch across different training-set sizes.

Result: The paper presents benchmarking results on how well pretrained PDE foundation models transfer to extreme material dynamics regimes, quantifying sample efficiency under distribution shift. Results show performance differences between fine-tuning and training from scratch approaches.

Conclusion: Provides insights into the transfer capabilities of PDE foundation models for extreme material dynamics, highlighting limitations and opportunities for improving model robustness to distribution shifts in physical simulations.

Abstract: Most PDE foundation models are pretrained and fine-tuned on fluid-centric benchmarks. Their utility under extreme-loading material dynamics remains unclear. We benchmark out-of-distribution transfer on two discontinuity-dominated regimes in which shocks, evolving interfaces, and fracture produce highly non-smooth fields: shock-driven multi-material interface dynamics (perturbed layered interface or PLI) and dynamic fracture/failure evolution (FRAC). We formulate the downstream task as terminal-state prediction, i.e., learning a long-horizon map that predicts the final state directly from the first snapshot without intermediate supervision. Using a unified training and evaluation protocol, we evaluate two open-source pretrained PDE foundation models, POSEIDON and MORPH, and compare fine-tuning from pretrained weights against training from scratch across training-set sizes to quantify sample efficiency under distribution shift.

Method: Combines PCA with closed-form Gaussian optimal transport to efficiently transform harmful activation distributions to match harmless ones in high-dimensional representation spaces. Uses layer-selective intervention, applying optimal transport to only 1-2 carefully chosen layers at approximately 40-60% network depth.

Result: Achieves up to 11% higher attack success rates than state-of-the-art baselines across six models (Llama-2, Llama-3.1, Qwen-2.5; 7B-32B parameters) while maintaining comparable perplexity. Layer-selective intervention substantially outperforms full-network interventions.

Conclusion: Refusal mechanisms may be localized rather than distributed in LLMs, and current alignment methods may be vulnerable to distributional attacks beyond simple direction removal. Provides new insights into the geometric structure of safety representations.

Abstract: Safety-aligned language models refuse harmful requests through learned refusal behaviors encoded in their internal representations. Recent activation-based jailbreaking methods circumvent these safety mechanisms by applying orthogonal projections to remove refusal directions, but these approaches treat refusal as a one-dimensional phenomenon and ignore the rich distributional structure of model activations. We introduce a principled framework based on optimal transport theory that transforms the entire distribution of harmful activations to match harmless ones. By combining PCA with closed-form Gaussian optimal transport, we achieve efficient computation in high-dimensional representation spaces while preserving essential geometric structure. Across six models (Llama-2, Llama-3.1, Qwen-2.5; 7B-32B parameters), our method achieves up to 11% higher attack success rates than state-of-the-art baselines while maintaining comparable perplexity, demonstrating superior preservation of model capabilities. Critically, we discover that layer-selective intervention (applying optimal transport to 1-2 carefully chosen layers at approximately 40-60% network depth) substantially outperforms full-network interventions, revealing that refusal mechanisms may be localized rather than distributed. Our analysis provides new insights into the geometric structure of safety representations and suggests that current alignment methods may be vulnerable to distributional attacks beyond simple direction removal.

Method: Analyzes quantization via signal-to-quantization-noise ratio (SQNR), showing it decomposes into concentration and alignment components. Introduces block Concentration-Alignment Transforms (CAT) - lightweight linear transformations using covariance estimates from calibration data to jointly optimize both factors.

Result: CAT consistently matches or outperforms prior transform-based quantization methods at 4-bit precision across several LLMs, confirming theoretical insights about importance of both concentration and alignment.

Conclusion: The paper provides principled understanding of quantization error through SQNR analysis and introduces effective transforms that jointly optimize concentration and alignment, offering practical improvements for efficient model deployment.

Abstract: Quantization can drastically increase the efficiency of large language and vision models, but typically incurs an accuracy drop. Recently, function-preserving transforms (e.g. rotations, Hadamard transform, channel-wise scaling) have been successfully applied to reduce post-training quantization error, yet a principled explanation remains elusive. We analyze linear-layer quantization via the signal-to-quantization-noise ratio (SQNR), showing that for uniform integer quantization at a fixed bit width, SQNR decomposes into (i) the concentration of weights and activations (capturing spread and outliers), and (ii) the alignment of their dominant variation directions. This reveals an actionable insight: beyond concentration - the focus of most prior transforms (e.g. rotations or Hadamard) - improving alignment between weight and activation can further reduce quantization error. Motivated by this, we introduce block Concentration-Alignment Transforms (CAT), a lightweight linear transformation that uses a covariance estimate from a small calibration set to jointly improve concentration and alignment, approximately maximizing SQNR. Experiments across several LLMs show that CAT consistently matches or outperforms prior transform-based quantization methods at 4-bit precision, confirming the insights gained in our framework.

Method: Reformulates sigma-point weight synthesis as hyperparameter optimization via memory-augmented meta-learning. Uses Recurrent Context Encoder to compress measurement innovation history into latent embeddings, then policy network dynamically synthesizes mean/covariance weights at each timestep.

Result: MA-UKF significantly outperforms standard baselines on maneuvering target benchmarks, showing superior robustness to non-Gaussian glint noise and effective generalization to out-of-distribution dynamic regimes.

Conclusion: The framework enables adaptive, context-aware state estimation by learning to dynamically adjust filter parameters through end-to-end optimization, improving tracking accuracy and consistency.

Abstract: The Unscented Kalman Filter (UKF) is a ubiquitous tool for nonlinear state estimation; however, its performance is limited by the static parameterization of the Unscented Transform (UT). Conventional weighting schemes, governed by fixed scaling parameters, assume implicit Gaussianity and fail to adapt to time-varying dynamics or heavy-tailed measurement noise. This work introduces the Meta-Adaptive UKF (MA-UKF), a framework that reformulates sigma-point weight synthesis as a hyperparameter optimization problem addressed via memory-augmented meta-learning. Unlike standard adaptive filters that rely on instantaneous heuristic corrections, our approach employs a Recurrent Context Encoder to compress the history of measurement innovations into a compact latent embedding. This embedding informs a policy network that dynamically synthesizes the mean and covariance weights of the sigma points at each time step, effectively governing the filter’s trust in the prediction versus the measurement. By optimizing the system end-to-end through the filter’s recursive logic, the MA-UKF learns to maximize tracking accuracy while maintaining estimation consistency. Numerical benchmarks on maneuvering targets demonstrate that the MA-UKF significantly outperforms standard baselines, exhibiting superior robustness to non-Gaussian glint noise and effective generalization to out-of-distribution (OOD) dynamic regimes unseen during training.

Method: HLOBA uses an autoencoder to learn an atmospheric latent space. Model forecasts are mapped via the encoder, while observations are mapped via a separate O2Lnet network. These are fused in latent space using Bayesian update with weights from time-lagged ensemble forecasts. Uncertainty is quantified through error decorrelation properties of latent variables.

Result: HLOBA matches the analysis and forecast skill of dynamically constrained 4D DA methods while achieving end-to-end inference-level efficiency. It provides element-wise uncertainty estimates that highlight large-error regions and capture seasonal variability in idealized experiments.

Conclusion: HLOBA demonstrates a novel approach to data assimilation that successfully balances accuracy, efficiency, and uncertainty quantification by operating in a learned latent space, with theoretical flexibility applicable to any forecasting model.

Abstract: Data assimilation (DA) combines model forecasts and observations to estimate the optimal state of the atmosphere with its uncertainty, providing initial conditions for weather prediction and reanalyses for climate research. Yet, existing traditional and machine-learning DA methods struggle to achieve accuracy, efficiency and uncertainty quantification simultaneously. Here, we propose HLOBA (Hybrid-Ensemble Latent Observation-Background Assimilation), a three-dimensional hybrid-ensemble DA method that operates in an atmospheric latent space learned via an autoencoder (AE). HLOBA maps both model forecasts and observations into a shared latent space via the AE encoder and an end-to-end Observation-to-Latent-space mapping network (O2Lnet), respectively, and fuses them through a Bayesian update with weights inferred from time-lagged ensemble forecasts. Both idealized and real-observation experiments demonstrate that HLOBA matches dynamically constrained four-dimensional DA methods in both analysis and forecast skill, while achieving end-to-end inference-level efficiency and theoretical flexibility applies to any forecasting model. Moreover, by exploiting the error decorrelation property of latent variables, HLOBA enables element-wise uncertainty estimates for its latent analysis and propagates them to model space via the decoder. Idealized experiments show that this uncertainty highlights large-error regions and captures their seasonal variability.

Method: Cannot determine method without access to the paper content. The error prevents retrieval of any technical details about the approach used in this research.

Result: Cannot determine results without access to the paper content. The rate limiting error prevents analysis of any experimental outcomes or findings.

Conclusion: Cannot draw conclusions about the paper’s content due to technical limitations in accessing the abstract. The paper ID format suggests it’s from March 2023, but relevance cannot be assessed without content.

Abstract: Failed to fetch summary for 2603.03802: Page request resulted in HTTP 429 (https://export.arxiv.org/api/query?search_query=&id_list=2603.03802&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 2310.04925: Page request resulted in HTTP 429 (https://export.arxiv.org/api/query?search_query=&id_list=2310.04925&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 2403.10889: Page request resulted in HTTP 429 (https://export.arxiv.org/api/query?search_query=&id_list=2403.10889&sortBy=relevance&sortOrder=descending&start=0&max_results=100)

Method: Unable to determine method due to paper fetch failure

Result: Unable to determine results due to paper fetch failure

Conclusion: Unable to determine conclusion due to paper fetch failure

Abstract: Failed to fetch summary for 2407.21546: Page request resulted in HTTP 429 (https://export.arxiv.org/api/query?search_query=&id_list=2407.21546&sortBy=relevance&sortOrder=descending&start=0&max_results=100)

Method: Cannot analyze method without paper content. The error suggests technical limitations in accessing the paper metadata rather than issues with the paper itself.

Result: No results can be analyzed due to inability to access paper content. The HTTP 429 status code indicates too many requests to the arXiv API.

Conclusion: Cannot provide meaningful analysis without access to the paper content. The arXiv API rate limiting prevents retrieval of the abstract for paper 2410.02601.

Abstract: Failed to fetch summary for 2410.02601: Page request resulted in HTTP 429 (https://export.arxiv.org/api/query?search_query=&id_list=2410.02601&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 2412.01654: Page request resulted in HTTP 429 (https://export.arxiv.org/api/query?search_query=&id_list=2412.01654&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 draw conclusions due to access error

Abstract: Failed to fetch summary for 2505.15643: Page request resulted in HTTP 429 (https://export.arxiv.org/api/query?search_query=&id_list=2505.15643&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 2505.18535: Page request resulted in HTTP 429 (https://export.arxiv.org/api/query?search_query=&id_list=2505.18535&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 2506.02168: Page request resulted in HTTP 429 (https://export.arxiv.org/api/query?search_query=&id_list=2506.02168&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 2506.13107: Page request resulted in HTTP 429 (https://export.arxiv.org/api/query?search_query=&id_list=2506.13107&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 2506.13150: Page request resulted in HTTP 429 (https://export.arxiv.org/api/query?search_query=&id_list=2506.13150&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 2506.15963: Page request resulted in HTTP 429 (https://export.arxiv.org/api/query?search_query=&id_list=2506.15963&sortBy=relevance&sortOrder=descending&start=0&max_results=100)

Method: Cannot determine method without paper content

Result: Cannot determine results without paper content

Conclusion: Cannot determine conclusion without paper content

Abstract: Failed to fetch summary for 2506.23971: Page request resulted in HTTP 429 (https://export.arxiv.org/api/query?search_query=&id_list=2506.23971&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 2509.17874: Page request resulted in HTTP 429 (https://export.arxiv.org/api/query?search_query=&id_list=2509.17874&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 2509.21150: Page request resulted in HTTP 429 (https://export.arxiv.org/api/query?search_query=&id_list=2509.21150&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 2509.21465: Page request resulted in HTTP 429 (https://export.arxiv.org/api/query?search_query=&id_list=2509.21465&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 2509.21675: Page request resulted in HTTP 429 (https://export.arxiv.org/api/query?search_query=&id_list=2509.21675&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 2509.22263: Page request resulted in HTTP 429 (https://export.arxiv.org/api/query?search_query=&id_list=2509.22263&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 2509.22580: Page request resulted in HTTP 429 (https://export.arxiv.org/api/query?search_query=&id_list=2509.22580&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 2509.23405: Page request resulted in HTTP 429 (https://export.arxiv.org/api/query?search_query=&id_list=2509.23405&sortBy=relevance&sortOrder=descending&start=0&max_results=100)

Method: Cannot determine method due to missing abstract data

Result: Cannot determine results due to missing abstract data

Conclusion: Cannot determine conclusion due to missing abstract data

Abstract: Failed to fetch summary for 2510.02903: Page request resulted in HTTP 429 (https://export.arxiv.org/api/query?search_query=&id_list=2510.02903&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 2510.16462: Page request resulted in HTTP 429 (https://export.arxiv.org/api/query?search_query=&id_list=2510.16462&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 2511.01343: Page request resulted in HTTP 429 (https://export.arxiv.org/api/query?search_query=&id_list=2511.01343&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 2512.00810: Page request resulted in HTTP 429 (https://export.arxiv.org/api/query?search_query=&id_list=2512.00810&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 2512.13506: Page request resulted in HTTP 429 (https://export.arxiv.org/api/query?search_query=&id_list=2512.13506&sortBy=relevance&sortOrder=descending&start=0&max_results=100)

Method: Cannot determine method as abstract is unavailable

Result: Cannot determine results as abstract is unavailable

Conclusion: Cannot draw conclusions about paper content due to data unavailability

Abstract: Failed to fetch summary for 2512.17198: Page request resulted in HTTP 429 (https://export.arxiv.org/api/query?search_query=&id_list=2512.17198&sortBy=relevance&sortOrder=descending&start=0&max_results=100)

Method: Cannot determine method as abstract retrieval failed

Result: Cannot determine results as abstract retrieval failed

Conclusion: Cannot draw conclusions without access to paper content

Abstract: Failed to fetch summary for 2512.18957: Page request resulted in HTTP 429 (https://export.arxiv.org/api/query?search_query=&id_list=2512.18957&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 content

Abstract: Failed to fetch summary for 2601.17204: Page request resulted in HTTP 429 (https://export.arxiv.org/api/query?search_query=&id_list=2601.17204&sortBy=relevance&sortOrder=descending&start=0&max_results=100)

Method: No method information available due to failed content retrieval

Result: No results available due to HTTP 429 error preventing content access

Conclusion: Cannot provide analysis due to technical limitations in accessing paper content

Abstract: Failed to fetch summary for 2601.17473: Page request resulted in HTTP 429 (https://export.arxiv.org/api/query?search_query=&id_list=2601.17473&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 2602.03901: Page request resulted in HTTP 429 (https://export.arxiv.org/api/query?search_query=&id_list=2602.03901&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 without access to paper content

Abstract: Failed to fetch summary for 2602.04027: Page request resulted in HTTP 429 (https://export.arxiv.org/api/query?search_query=&id_list=2602.04027&sortBy=relevance&sortOrder=descending&start=0&max_results=100)

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

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

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

Abstract: Failed to fetch summary for 2602.12147: Page request resulted in HTTP 429 (https://export.arxiv.org/api/query?search_query=&id_list=2602.12147&sortBy=relevance&sortOrder=descending&start=0&max_results=100)

Method: Unable to determine method due to fetch failure

Result: Unable to determine results due to fetch failure

Conclusion: Unable to determine conclusion due to fetch failure

Abstract: Failed to fetch summary for 2603.01752: Page request resulted in HTTP 429 (https://export.arxiv.org/api/query?search_query=&id_list=2603.01752&sortBy=relevance&sortOrder=descending&start=0&max_results=100)

Method: Unable to determine method due to fetch failure

Result: Unable to determine results due to fetch failure

Conclusion: Unable to determine conclusion due to fetch failure

Abstract: Failed to fetch summary for 2603.02678: Page request resulted in HTTP 429 (https://export.arxiv.org/api/query?search_query=&id_list=2603.02678&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 2603.02862: Page request resulted in HTTP 429 (https://export.arxiv.org/api/query?search_query=&id_list=2603.02862&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 2312.05645: Page request resulted in HTTP 429 (https://export.arxiv.org/api/query?search_query=&id_list=2312.05645&sortBy=relevance&sortOrder=descending&start=0&max_results=100)