Method: Fine-tune RWKV-based LLM into EmbeddingRWKV as unified model for both embedding and state extraction; design state-based reranker that processes only query tokens using precomputed document states; implement uniform layer selection to use only 25% of layers.

Result: Achieves 5.4×-44.8× speedup during reranking by decoupling inference cost from document length; maintains 98.62% of full-model performance using only 25% of layers; demonstrates high-quality retrieval and reranking with enhanced efficiency.

Conclusion: State-Centric Retrieval provides an efficient unified paradigm for RAG systems by leveraging reusable states as bridge between embedding and reranking stages, significantly improving overall system efficiency while maintaining performance.

Abstract: Current Retrieval-Augmented Generation (RAG) systems typically employ a traditional two-stage pipeline: an embedding model for initial retrieval followed by a reranker for refinement. However, this paradigm suffers from significant inefficiency due to the lack of shared information between stages, leading to substantial redundant computation. To address this limitation, we propose \textbf{State-Centric Retrieval}, a unified retrieval paradigm that utilizes “states” as a bridge to connect embedding models and rerankers. First, we perform state representation learning by fine-tuning an RWKV-based LLM, transforming it into \textbf{EmbeddingRWKV}, a unified model that serves as both an embedding model and a state backbone for extracting compact, reusable states. Building upon these reusable states, we further design a state-based reranker to fully leverage precomputed information. During reranking, the model processes only query tokens, decoupling inference cost from document length and yielding a 5.4$\times$–44.8$\times$ speedup. Furthermore, we observe that retaining all intermediate layer states is unnecessary; with a uniform layer selection strategy, our model maintains 98.62% of full-model performance using only 25% of the layers. Extensive experiments demonstrate that State-Centric Retrieval achieves high-quality retrieval and reranking results while significantly enhancing overall system efficiency. Code is available at \href{https://github.com/howard-hou/EmbeddingRWKV}{our GitHub repository}.

Method: Created MedES benchmark from 260 Chinese medical/ethical/legal sources, developed guardian-in-the-loop framework with automated evaluator (97% accuracy), and aligned 7B LLM through supervised fine-tuning and domain-specific preference optimization.

Result: Aligned 7B model outperforms larger baselines on core ethical tasks within Chinese medical ethics context, showing improvements in both quality and composite evaluation metrics.

Conclusion: Provides practical framework for aligning LLMs with medical ethics in Chinese healthcare, with potential for adaptation to other legal/cultural environments through modular replacement of normative corpus.

Abstract: Recent advances in large language models have enabled their application to a range of healthcare tasks. However, aligning LLMs with the nuanced demands of medical ethics, especially under complex real world scenarios, remains underexplored. In this work, we present MedES, a dynamic, scenario-centric benchmark specifically constructed from 260 authoritative Chinese medical, ethical, and legal sources to reflect the challenges in clinical decision-making. To facilitate model alignment, we introduce a guardian-in-the-loop framework that leverages a dedicated automated evaluator (trained on expert-labeled data and achieving over 97% accuracy within our domain) to generate targeted prompts and provide structured ethical feedback. Using this pipeline, we align a 7B-parameter LLM through supervised fine-tuning and domain-specific preference optimization. Experimental results, conducted entirely within the Chinese medical ethics context, demonstrate that our aligned model outperforms notably larger baselines on core ethical tasks, with observed improvements in both quality and composite evaluation metrics. Our work offers a practical and adaptable framework for aligning LLMs with medical ethics in the Chinese healthcare domain, and suggests that similar alignment pipelines may be instantiated in other legal and cultural environments through modular replacement of the underlying normative corpus.

Method: Created ValAct-15k dataset with 3,000 Reddit-derived advice-seeking scenarios based on Schwartz’s Theory of Basic Human Values. Evaluated 10 frontier LLMs (5 US, 5 Chinese) and 55 humans using both scenario-based questions and traditional value questionnaires.

Result: LLMs show near-perfect cross-model consistency in decisions (r≈1.0) vs. broad human variability (r∈[-0.79,0.98]). Both show weak self-reported vs. enacted value correspondence (r=0.4 humans, 0.3 LLMs). LLMs decline up to 6.6% when instructed to “hold” values vs. selecting them.

Conclusion: Alignment training yields normative value convergence but doesn’t eliminate human-like incoherence between knowing and acting upon values, revealing systematic knowledge-action gaps in LLMs.

Abstract: Value alignment is central to the development of safe and socially compatible artificial intelligence. However, how Large Language Models (LLMs) represent and enact human values in real-world decision contexts remains under-explored. We present ValAct-15k, a dataset of 3,000 advice-seeking scenarios derived from Reddit, designed to elicit ten values defined by Schwartz Theory of Basic Human Values. Using both the scenario-based questions and the traditional value questionnaire, we evaluate ten frontier LLMs (five from U.S. companies, five from Chinese ones) and human participants ($n = 55$). We find near-perfect cross-model consistency in scenario-based decisions (Pearson $r \approx 1.0$), contrasting sharply with the broad variability observed among humans ($r \in [-0.79, 0.98]$). Yet, both humans and LLMs show weak correspondence between self-reported and enacted values ($r = 0.4, 0.3$), revealing a systematic knowledge-action gap. When instructed to “hold” a specific value, LLMs’ performance declines up to $6.6%$ compared to merely selecting the value, indicating a role-play aversion. These findings suggest that while alignment training yields normative value convergence, it does not eliminate the human-like incoherence between knowing and acting upon values.

Method: Constructed a comprehensive benchmark spanning 6 prompting strategies, 7 LLMs, and 4 domain datasets to create diverse human- and AI-generated texts. Fine-tuned classification-based detectors on various generation settings and evaluated cross-prompt, cross-model, and cross-dataset generalization. Computed correlations between generalization accuracies and feature shifts of 80 linguistic features between training and test conditions.

Result: Generalization performance for specific detectors and evaluation conditions is significantly associated with linguistic features such as tense usage and pronoun frequency.

Conclusion: Linguistic analysis provides explanatory insights into why AI-text detectors fail to generalize across different generation conditions, revealing that specific linguistic feature shifts correlate with generalization gaps.

Abstract: AI-text detectors achieve high accuracy on in-domain benchmarks, but often struggle to generalize across different generation conditions such as unseen prompts, model families, or domains. While prior work has reported these generalization gaps, there are limited insights about the underlying causes. In this work, we present a systematic study aimed at explaining generalization behavior through linguistic analysis. We construct a comprehensive benchmark that spans 6 prompting strategies, 7 large language models (LLMs), and 4 domain datasets, resulting in a diverse set of human- and AI-generated texts. Using this dataset, we fine-tune classification-based detectors on various generation settings and evaluate their cross-prompt, cross-model, and cross-dataset generalization. To explain the performance variance, we compute correlations between generalization accuracies and feature shifts of 80 linguistic features between training and test conditions. Our analysis reveals that generalization performance for specific detectors and evaluation conditions is significantly associated with linguistic features such as tense usage and pronoun frequency.

Method: A tri-tier evaluation framework: Tier I computes automated coverage/risk indicators; Tier II applies rubric-based scoring across five dimensions using a single primary judge; Tier III calibrates scores to human ratings via isotonic regression.

Result: The framework reduces MAE by 5.2% on a 152-sample held-out set. Evaluation of 15 VLMs on 294 expert anchors across six cultural traditions shows automated metrics are unreliable for cultural depth, Western samples score higher than non-Western, and cross-judge scale mismatch makes naive averaging unreliable.

Conclusion: The proposed framework provides calibrated cultural-understanding scores for model selection and cultural-gap diagnosis, with dimension-level diagnostics and risk indicators, addressing the need for systematic evaluation of VLMs’ cultural interpretation capabilities in art.

Abstract: Vision-Language Models (VLMs) excel at visual perception, yet their ability to interpret cultural meaning in art remains under-validated. We present a tri-tier evaluation framework for cross-cultural art-critique assessment: Tier I computes automated coverage and risk indicators offline; Tier II applies rubric-based scoring using a single primary judge across five dimensions; and Tier III calibrates the Tier II aggregate score to human ratings via isotonic regression, yielding a 5.2% reduction in MAE on a 152-sample held-out set. The framework outputs a calibrated cultural-understanding score for model selection and cultural-gap diagnosis, together with dimension-level diagnostics and risk indicators. We evaluate 15 VLMs on 294 expert anchors spanning six cultural traditions. Key findings are that (i) automated metrics are unreliable proxies for cultural depth, (ii) Western samples score higher than non-Western samples under our sampling and rubric, and (iii) cross-judge scale mismatch makes naive score averaging unreliable, motivating a single primary judge with explicit calibration. Dataset and code are available in the supplementary materials.

Method: Developed a comprehensive data collection and processing pipeline that aggregates ClaimReview feeds, scrapes full debunking articles, normalizes heterogeneous claim verdicts, and enriches them with structured metadata and aligned visual content. Used state-of-the-art LLMs and multimodal LLMs for evidence extraction under predefined categories and justification generation linking evidence to verdicts.

Result: Evaluation with G-Eval and human assessment shows the pipeline enables fine-grained comparison of fact-checking practices across different organizations/media markets, facilitates development of more interpretable and evidence-grounded fact-checking models, and lays groundwork for future multilingual, multimodal misinformation verification research.

Conclusion: The pipeline successfully addresses limitations of existing fact-checking datasets by providing structured, multimodal, multilingual resources with explainable links between claims, evidence, and verdicts, supporting both comparative analysis and model development.

Abstract: The rapid proliferation of misinformation across online platforms underscores the urgent need for robust, up-to-date, explainable, and multilingual fact-checking resources. However, existing datasets are limited in scope, often lacking multimodal evidence, structured annotations, and detailed links between claims, evidence, and verdicts. This paper introduces a comprehensive data collection and processing pipeline that constructs multimodal fact-checking datasets in French and German languages by aggregating ClaimReview feeds, scraping full debunking articles, normalizing heterogeneous claim verdicts, and enriching them with structured metadata and aligned visual content. We used state-of-the-art large language models (LLMs) and multimodal LLMs for (i) evidence extraction under predefined evidence categories and (ii) justification generation that links evidence to verdicts. Evaluation with G-Eval and human assessment demonstrates that our pipeline enables fine-grained comparison of fact-checking practices across different organizations or media markets, facilitates the development of more interpretable and evidence-grounded fact-checking models, and lays the groundwork for future research on multilingual, multimodal misinformation verification.

Method: Created a benchmark with 7,410 matched image-critique pairs spanning eight cultural traditions, using a five-layer framework (L1-L5) from Visual Perception to Philosophical Aesthetics, instantiated as 225 culture-specific dimensions with expert-written bilingual critiques.

Result: Pilot results show that higher-layer reasoning (L3-L5) is consistently more challenging for VLMs than visual and technical analysis (L1-L2), highlighting the gap in cultural understanding capabilities.

Conclusion: VULCA-Bench provides a comprehensive tool to evaluate VLMs’ cultural understanding beyond surface-level perception, revealing current limitations in higher-order cultural interpretation and offering resources for future research.

Abstract: We introduce VULCA-Bench, a multicultural art-critique benchmark for evaluating Vision-Language Models’ (VLMs) cultural understanding beyond surface-level visual perception. Existing VLM benchmarks predominantly measure L1-L2 capabilities (object recognition, scene description, and factual question answering) while under-evaluate higher-order cultural interpretation. VULCA-Bench contains 7,410 matched image-critique pairs spanning eight cultural traditions, with Chinese-English bilingual coverage. We operationalise cultural understanding using a five-layer framework (L1-L5, from Visual Perception to Philosophical Aesthetics), instantiated as 225 culture-specific dimensions and supported by expert-written bilingual critiques. Our pilot results indicate that higher-layer reasoning (L3-L5) is consistently more challenging than visual and technical analysis (L1-L2). The dataset, evaluation scripts, and annotation tools are available under CC BY 4.0 in the supplementary materials.

Method: Introduces LLM Review, a peer-review-inspired framework with Blind Peer Review where agents exchange targeted feedback while revising independently to preserve divergent creative trajectories. Also proposes SciFi-100, a science fiction writing dataset with unified evaluation combining LLM-as-a-judge scoring, human annotation, and rule-based novelty metrics.

Result: LLM Review consistently outperforms multi-agent baselines. Smaller models using this framework can surpass larger single-agent models, suggesting interaction structure may substitute for model scale in creative tasks.

Conclusion: The blind peer review framework effectively enhances LLM creativity while avoiding content homogenization, and structured interaction can compensate for model scale limitations in creative generation tasks.

Abstract: Large Language Models (LLMs) often struggle with creative generation, and multi-agent frameworks that improve reasoning through interaction can paradoxically hinder creativity by inducing content homogenization. We introduce LLM Review, a peer-review-inspired framework implementing Blind Peer Review: agents exchange targeted feedback while revising independently, preserving divergent creative trajectories. To enable rigorous evaluation, we propose SciFi-100, a science fiction writing dataset with a unified framework combining LLM-as-a-judge scoring, human annotation, and rule-based novelty metrics. Experiments demonstrate that LLM Review consistently outperforms multi-agent baselines, and smaller models with our framework can surpass larger single-agent models, suggesting interaction structure may substitute for model scale.

Method: Proposes a longitudinal modeling paradigm with four key updates: (1) evaluation splits aligned to cross-sectional (generalization over people) and prospective (generalization over time) generalization; (2) accuracy metrics separating between-person differences from within-person dynamics; (3) sequence inputs incorporating historical context; (4) model internals supporting different coarseness of latent state representations.

Result: Applied to 17k daily diary transcripts with PTSD symptom severity from 238 participants, traditional document-level evaluation yielded substantially different and sometimes reversed conclusions compared to the proposed ecologically valid modeling approach.

Conclusion: Argues for a shift from word-sequence evaluation toward behavior-sequence paradigms in NLP, emphasizing the need for longitudinal modeling frameworks that properly account for person-indexed, time-ordered data structures.

Abstract: While NLP typically treats documents as independent and unordered samples, in longitudinal studies, this assumption rarely holds: documents are nested within authors and ordered in time, forming person-indexed, time-ordered $\textit{behavioral sequences}$. Here, we demonstrate the need for and propose a longitudinal modeling and evaluation paradigm that consequently updates four parts of the NLP pipeline: (1) evaluation splits aligned to generalization over people ($\textit{cross-sectional}$) and/or time ($\textit{prospective}$); (2) accuracy metrics separating between-person differences from within-person dynamics; (3) sequence inputs to incorporate history by default; and (4) model internals that support different $\textit{coarseness}$ of latent state over histories (pooled summaries, explicit dynamics, or interaction-based models). We demonstrate the issues ensued by traditional pipeline and our proposed improvements on a dataset of 17k daily diary transcripts paired with PTSD symptom severity from 238 participants, finding that traditional document-level evaluation can yield substantially different and sometimes reversed conclusions compared to our ecologically valid modeling and evaluation. We tie our results to a broader discussion motivating a shift from word-sequence evaluation toward $\textit{behavior-sequence}$ paradigms for NLP.

Method: DyCP dynamically segments and retrieves relevant memory at query time, preserving the sequential structure of dialogue without predefined topic boundaries, supporting efficient and adaptive context retrieval.

Result: Across three long-form dialogue benchmarks (LoCoMo, MT-Bench+, SCM4LLMs) and multiple LLMs, DyCP consistently improves answer quality while reducing response latency.

Conclusion: There’s a gap between modern LLMs’ expanded context windows and their actual long-context processing capacity, highlighting the continued importance of effective context management methods like DyCP.

Abstract: Large Language Models (LLMs) often exhibit increased response latency and degraded answer quality as dialogue length grows, making effective context management essential. However, existing methods rely on extra LLM calls to build memory or perform offline memory construction without considering the current user utterance, which can introduce inefficiencies or disrupt conversational continuity. We introduce DyCP, a lightweight context management method that dynamically segment and retrieve relevant memory at query time. It preserves the sequential structure of dialogue without predefined topic boundaries and supports efficient, adaptive context retrieval. Across three long-form dialogue benchmarks, LoCoMo, MT-Bench+, and SCM4LLMs, and multiple LLMs, DyCP consistently improves answer quality while reducing response latency. We also examine the gap between modern LLMs’ expanded context windows and their actual long-context processing capacity, highlighting the continued importance of effective context management.

Method: Evaluated CVP across genres, historical periods, languages, and affective dimensions; examined linearity assumption underlying the method.

Result: Concept vectors trained on one corpus transfer well to others with minimal performance loss; linearity assumption is approximate but method effectively captures generalizable patterns.

Conclusion: CVP is a portable approach for sentiment analysis that works well across domains, though its linearity assumption is approximate, suggesting potential for further development.

Abstract: Use cases of sentiment analysis in the humanities often require contextualized, continuous scores. Concept Vector Projections (CVP) offer a recent solution: by modeling sentiment as a direction in embedding space, they produce continuous, multilingual scores that align closely with human judgments. Yet the method’s portability across domains and underlying assumptions remain underexplored. We evaluate CVP across genres, historical periods, languages, and affective dimensions, finding that concept vectors trained on one corpus transfer well to others with minimal performance loss. To understand the patterns of generalization, we further examine the linearity assumption underlying CVP. Our findings suggest that while CVP is a portable approach that effectively captures generalizable patterns, its linearity assumption is approximate, pointing to potential for further development.

Method: Used Sparse Autoencoders (SAEs) to analyze and intervene on LLM internal representations, identifying reasoning-related latent features that can be externally steered.

Result: Steering a single reasoning-related latent feature substantially improves accuracy without CoT prompting, achieving comparable performance with more efficient outputs, and can override instructions that discourage reasoning.

Conclusion: CoT prompting is one effective but not unique way to activate LLM reasoning; multi-step reasoning is supported by latent internal activations that can be externally triggered.

Abstract: Chain-of-Thought (CoT) prompting has improved the reasoning performance of large language models (LLMs), but it remains unclear why it works and whether it is the unique mechanism for triggering reasoning in large language models. In this work, we study this question by directly analyzing and intervening on the internal representations of LLMs with Sparse Autoencoders (SAEs), identifying a small set of latent features that are causally associated with LLM reasoning behavior. Across multiple model families and reasoning benchmarks, we find that steering a single reasoning-related latent feature can substantially improve accuracy without explicit CoT prompting. For large models, latent steering achieves performance comparable to standard CoT prompting while producing more efficient outputs. We further observe that this reasoning-oriented internal state is triggered early in generation and can override prompt-level instructions that discourage explicit reasoning. Overall, our results suggest that multi-step reasoning in LLMs is supported by latent internal activations that can be externally activated, while CoT prompting is one effective, but not unique, way of activating this mechanism rather than its necessary cause.

Method: Created HiKE benchmark with high-quality natural Korean-English code-switching data across various topics, featuring meticulous loanword labels and hierarchical CS-level labeling (word, phrase, sentence) for systematic evaluation.

Result: Most multilingual ASR models initially show inadequate CS-ASR performance, but this capability can be enabled through fine-tuning with synthetic CS data.

Conclusion: HiKE provides the first accessible non-synthetic evaluation framework for Korean-English code-switching, enabling systematic assessment of multilingual ASR models and demonstrating that CS capability can be developed through fine-tuning.

Abstract: Despite advances in multilingual automatic speech recognition (ASR), code-switching (CS), the mixing of languages within an utterance common in daily speech, remains a severely underexplored challenge. In this paper, we introduce HiKE: the Hierarchical Korean-English code-switching benchmark, the first globally accessible non-synthetic evaluation framework for Korean-English CS, aiming to provide a means for the precise evaluation of multilingual ASR models and to foster research in the field. The proposed framework not only consists of high-quality, natural CS data across various topics, but also provides meticulous loanword labels and a hierarchical CS-level labeling scheme (word, phrase, and sentence) that together enable a systematic evaluation of a model’s ability to handle each distinct level of code-switching. Through evaluations of diverse multilingual ASR models and fine-tuning experiments, this paper demonstrates that although most multilingual ASR models initially exhibit inadequate CS-ASR performance, this capability can be enabled through fine-tuning with synthetic CS data. HiKE is available at https://github.com/ThetaOne-AI/HiKE.

Method: Proving mathematically that chaining a language model’s autoregressive output is sufficient for universal computation, and demonstrating this capability in randomly initialized models.

Result: Language models can simulate execution of any algorithm on any input, and even randomly initialized models possess universal computational capabilities before training.

Conclusion: Training doesn’t create computational expressiveness but improves programmability, enabling natural language access to intrinsic universal computation capabilities.

Abstract: Language models now provide an interface to express and often solve general problems in natural language, yet their ultimate computational capabilities remain a major topic of scientific debate. Unlike a formal computer, a language model is trained to autoregressively predict successive elements in human-generated text. We prove that chaining a language model’s autoregressive output is sufficient to perform universal computation. That is, a language model can simulate the execution of any algorithm on any input. The challenge of eliciting desired computational behaviour can thus be reframed in terms of programmability: the ease of finding a suitable prompt. Strikingly, we demonstrate that even randomly initialized language models are capable of universal computation before training. This implies that training does not give rise to computational expressiveness – rather, it improves programmability, enabling a natural language interface for accessing these intrinsic capabilities.

Method: Proposes a comprehensive evaluation framework for confidence estimation that measures three new aspects: 1) Robustness against prompt perturbations, 2) Stability across semantically equivalent answers, and 3) Sensitivity to semantically different answers. The framework systematically tests CE methods on these dimensions beyond traditional calibration and discrimination metrics.

Result: Common CE methods for LLMs often fail on the new evaluation metrics. Methods that achieve good performance on traditional calibration or discrimination metrics are not robust to prompt variations or are not sensitive to answer changes. The framework reveals significant limitations in existing CE evaluations that are relevant for real-world LLM use cases.

Conclusion: The proposed comprehensive evaluation framework exposes limitations in current confidence estimation methods for LLMs and provides practical guidance for selecting and designing more reliable CE methods. It emphasizes the need to consider robustness, stability, and sensitivity metrics alongside traditional calibration and discrimination for real-world LLM applications.

Abstract: Confidence estimation (CE) indicates how reliable the answers of large language models (LLMs) are, and can impact user trust and decision-making. Existing work evaluates CE methods almost exclusively through calibration, examining whether stated confidence aligns with accuracy, or discrimination, whether confidence is ranked higher for correct predictions than incorrect ones. However, these facets ignore pitfalls of CE in the context of LLMs and language variation: confidence estimates should remain consistent under semantically equivalent prompt or answer variations, and should change when the answer meaning differs. Therefore, we present a comprehensive evaluation framework for CE that measures their confidence quality on three new aspects: robustness of confidence against prompt perturbations, stability across semantic equivalent answers, and sensitivity to semantically different answers. In our work, we demonstrate that common CE methods for LLMs often fail on these metrics: methods that achieve good performance on calibration or discrimination are not robust to prompt variations or are not sensitive to answer changes. Overall, our framework reveals limitations of existing CE evaluations relevant for real-world LLM use cases and provides practical guidance for selecting and designing more reliable CE methods.

Method: AdaJudge jointly adapts representation and aggregation through: 1) gated refinement blocks that refine backbone representations into discrimination-oriented space, and 2) adaptive multi-view pooling module that dynamically routes and combines evidence instead of using static readout.

Result: Extensive experiments on RM-Bench and JudgeBench show AdaJudge outperforms strong off-the-shelf reward models and traditional pooling baselines.

Conclusion: AdaJudge provides a unified framework that addresses the limitations of static pooling in reward modeling by enabling adaptive representation refinement and evidence aggregation, leading to better alignment with human preferences.

Abstract: Reward modeling is essential for aligning large language models with human preferences, yet predominant architectures rely on a static pooling strategy to condense sequences into scalar scores. This paradigm, however, suffers from two key limitations: a static inductive bias that misaligns with task-dependent preference signals, and a representational mismatch, as the backbone is optimized for generation rather than fine-grained discrimination. To address this, we propose AdaJudge, a unified framework that jointly adapts representation and aggregation. AdaJudge first refines backbone representations into a discrimination-oriented space via gated refinement blocks. It then replaces the static readout with an adaptive multi-view pooling module that dynamically routes and combines evidence. Extensive experiments on RM-Bench and JudgeBench show that AdaJudge outperforms strong off-the-shelf reward models and traditional pooling baselines.

Method: Introduces robust dynamic few-shot learning that retrieves examples from relevant workflows to suggest answerable queries. The system can be self-learned on prior user queries, making it practical for real-world applications.

Result: Evaluation on three benchmark datasets from real-world user queries shows the method produces more relevant and answerable suggestions than few-shot and retrieval-only baselines, enabling safer and more effective user interaction with agentic RAG.

Conclusion: Query suggestion for agentic RAG is a valuable approach to handle unanswerable questions by suggesting similar, answerable alternatives, improving user interaction safety and effectiveness through the proposed robust dynamic few-shot learning method.

Abstract: Retrieval-augmented generation with tool-calling agents (agentic RAG) has become increasingly powerful in understanding, processing, and responding to user queries. However, the scope of the grounding knowledge is limited and asking questions that exceed this scope may lead to issues like hallucination. While guardrail frameworks aim to block out-of-scope questions (Rodriguez et al., 2024), no research has investigated the question of suggesting answerable queries in order to complete the user interaction. In this paper, we initiate the study of query suggestion for agentic RAG. We consider the setting where user questions are not answerable, and the suggested queries should be similar to aid the user interaction. Such scenarios are frequent for tool-calling LLMs as communicating the restrictions of the tools or the underlying datasets to the user is difficult, and adding query suggestions enhances the interaction with the RAG agent. As opposed to traditional settings for query recommendations such as in search engines, ensuring that the suggested queries are answerable is a major challenge due to the RAG’s multi-step workflow that demands a nuanced understanding of the RAG as a whole, which the executing LLM lacks. As such, we introduce robust dynamic few-shot learning which retrieves examples from relevant workflows. We show that our system can be self-learned, for instance on prior user queries, and is therefore easily applicable in practice. We evaluate our approach on three benchmark datasets based on two unlabeled question datasets collected from real-world user queries. Experiments on real-world datasets confirm that our method produces more relevant and answerable suggestions, outperforming few-shot and retrieval-only baselines, and thus enable safer, more effective user interaction with agentic RAG.

Method: Adaptive Causal Prompting with Sketch-of-Thought (ACPS) framework that leverages structural causal models to infer causal effects and adaptively selects appropriate interventions (standard front-door and conditional front-door adjustments).

Result: Extensive experiments on multiple reasoning benchmarks and LLMs show ACPS consistently outperforms existing prompting baselines in accuracy, robustness, and computational efficiency.

Conclusion: ACPS enables generalizable causal reasoning across heterogeneous tasks without task-specific retraining and significantly reduces token usage and inference cost.

Abstract: Despite notable advancements in prompting methods for Large Language Models (LLMs), such as Chain-of-Thought (CoT), existing strategies still suffer from excessive token usage and limited generalisability across diverse reasoning tasks. To address these limitations, we propose an Adaptive Causal Prompting with Sketch-of-Thought (ACPS) framework, which leverages structural causal models to infer the causal effect of a query on its answer and adaptively select an appropriate intervention (i.e., standard front-door and conditional front-door adjustments). This design enables generalisable causal reasoning across heterogeneous tasks without task-specific retraining. By replacing verbose CoT with concise Sketch-of-Thought, ACPS enables efficient reasoning that significantly reduces token usage and inference cost. Extensive experiments on multiple reasoning benchmarks and LLMs demonstrate that ACPS consistently outperforms existing prompting baselines in terms of accuracy, robustness, and computational efficiency.

Method: ExoFormer learns dedicated exogenous anchor projections outside the sequential layer stack, decoupling the anchor role from computational refinement. It uses a unified normalized mixing framework across all attention pathways (queries, keys, values, gate logits) with different coefficient granularities (elementwise, headwise, scalar).

Result: ExoFormer variants consistently outperform internal-anchor counterparts, with dynamic variant achieving 2.13-point downstream accuracy improvement and 1.84x better data efficiency. It also achieves 2x reduction in attention sink compared to standard Gated Attention, though all variants exhibit representation collapse.

Conclusion: External anchors preserve essential token identity while allowing layers to specialize in computational refinement (Offloading Hypothesis). The representation collapse is explained as a consequence of this specialization. Code and models are released for future research.

Abstract: Transformers that reuse early-layer attention projections as residuals face a fundamental tension: the first layer must simultaneously serve as a stable reference for all deeper layers and as an effective computational block. To resolve this, we propose ExoFormer, which learns dedicated exogenous anchor projections outside the sequential layer stack, decoupling the anchor role from computational refinement. Through a unified normalized mixing framework (studying different coefficient granularities: elementwise, headwise, scalar) across all attention pathways (queries, keys, values, and gate logits), ExoFormer variants consistently outperform their internal-anchor counterparts. Moreover, the dynamic variant achieves a 2.13-point increase in downstream accuracy over the baseline and demonstrates superior data efficiency, matching baseline validation loss with 1.84x fewer tokens. ExoFormer also achieves a 2x reduction in attention sink compared to standard Gated Attention. Paradoxically, all ExoFormer variants exhibit signs of representation collapse. We explain this via an Offloading Hypothesis: external anchors preserve essential token identity, allowing layers to specialize exclusively in computational refinement. We release codes and models to facilitate future research.

Method: Created RMCB benchmark with 347,496 reasoning traces from six LRMs across diverse domains, then evaluated over ten representation-based methods including sequential, graph-based, and text-based architectures.

Result: Found persistent trade-off: text-based encoders achieve best AUROC (0.672) while structurally-aware models yield best ECE (0.148). Increased architectural complexity doesn’t reliably outperform simpler sequential baselines.

Conclusion: RMCB provides the most comprehensive benchmark for confidence estimation in reasoning models, establishing rigorous baselines and demonstrating limitations of current representation-based paradigms.

Abstract: The miscalibration of Large Reasoning Models (LRMs) undermines their reliability in high-stakes domains, necessitating methods to accurately estimate the confidence of their long-form, multi-step outputs. To address this gap, we introduce the Reasoning Model Confidence estimation Benchmark (RMCB), a public resource of 347,496 reasoning traces from six popular LRMs across different architectural families. The benchmark is constructed from a diverse suite of datasets spanning high-stakes domains, including clinical, financial, legal, and mathematical reasoning, alongside complex general reasoning benchmarks, with correctness annotations provided for all samples. Using RMCB, we conduct a large-scale empirical evaluation of over ten distinct representation-based methods, spanning sequential, graph-based, and text-based architectures. Our central finding is a persistent trade-off between discrimination (AUROC) and calibration (ECE): text-based encoders achieve the best AUROC (0.672), while structurally-aware models yield the best ECE (0.148), with no single method dominating both. Furthermore, we find that increased architectural complexity does not reliably outperform simpler sequential baselines, suggesting a performance ceiling for methods relying solely on chunk-level hidden states. This work provides the most comprehensive benchmark for this task to date, establishing rigorous baselines and demonstrating the limitations of current representation-based paradigms.

Method: Two-stage approach: 1) Continued pre-training on LLaMA 3.1 8B using 1.97B tokens (1.84B Urdu + 140M English Wikipedia to prevent catastrophic forgetting), 2) Supervised fine-tuning on the Alif Urdu-instruct dataset.

Result: Qalb achieves weighted average score of 90.34, outperforming previous SOTA Alif-1.0-Instruct (87.1) by 3.24 points and base LLaMA-3.1 8B-Instruct by 44.64 points. Achieves SOTA performance across seven diverse tasks including Classification, Sentiment Analysis, and Reasoning.

Conclusion: Continued pre-training on diverse, high-quality language data combined with targeted instruction fine-tuning effectively adapts foundation models to low-resource languages like Urdu, demonstrating a viable approach for underrepresented languages.

Abstract: Despite remarkable progress in large language models, Urdu-a language spoken by over 230 million people-remains critically underrepresented in modern NLP systems. Existing multilingual models demonstrate poor performance on Urdu-specific tasks, struggling with the language’s complex morphology, right-to-left Nastaliq script, and rich literary traditions. Even the base LLaMA-3.1 8B-Instruct model shows limited capability in generating fluent, contextually appropriate Urdu text. We introduce Qalb, an Urdu language model developed through a two-stage approach: continued pre-training followed by supervised fine-tuning. Starting from LLaMA 3.1 8B, we perform continued pre-training on a dataset of 1.97 billion tokens. This corpus comprises 1.84 billion tokens of diverse Urdu text-spanning news archives, classical and contemporary literature, government documents, and social media-combined with 140 million tokens of English Wikipedia data to prevent catastrophic forgetting. We then fine-tune the resulting model on the Alif Urdu-instruct dataset. Through extensive evaluation on Urdu-specific benchmarks, Qalb demonstrates substantial improvements, achieving a weighted average score of 90.34 and outperforming the previous state-of-the-art Alif-1.0-Instruct model (87.1) by 3.24 points, while also surpassing the base LLaMA-3.1 8B-Instruct model by 44.64 points. Qalb achieves state-of-the-art performance with comprehensive evaluation across seven diverse tasks including Classification, Sentiment Analysis, and Reasoning. Our results demonstrate that continued pre-training on diverse, high-quality language data, combined with targeted instruction fine-tuning, effectively adapts foundation models to low-resource languages.

Method: CT-SFT uses label-balanced mean baseline and task-directional relevance scoring to identify sparse task-relevant attention heads in a proxy-language checkpoint, then transfers to target language by updating only those heads (plus LayerNorm) via head-level gradient masking.

Result: CT-SFT improves cross-lingual accuracy over continued full fine-tuning on NusaX-Senti and XNLI while updating only a small subset of parameters. It shows an editing-preserving trade-off and substantially reduces catastrophic forgetting.

Conclusion: CT-SFT provides an effective approach for low-resource language adaptation that balances task transfer with preservation of source-language competence through selective parameter updates.

Abstract: Adapting LLMs to low-resource languages is difficult: labeled data is scarce, full-model fine-tuning is unstable, and continued cross-lingual tuning can cause catastrophic forgetting. We propose Circuit-Targeted Supervised Fine-Tuning (CT-SFT): a counterfactual-free adaptation of CD-T (Contextual Decomposition Transformer) that uses a label-balanced mean baseline and task-directional relevance scoring to identify a sparse set of task-relevant attention heads in a proxy-language checkpoint, then transfer learns to a target language by updating only those heads (plus LayerNorm) via head-level gradient masking. Across NusaX-Senti and XNLI, CT-SFT improves cross-lingual accuracy over continued full fine-tuning while updating only a small subset of model parameters. We find an editing-preserving trade-off: harder transfers favor editing circuit heads, while easier transfers often favor near-zero (i.e., low-relevance heads) updates, preserving the source mechanism. CT-SFT also substantially reduces catastrophic forgetting, preserving proxy/source-language competence during transfer.

Method: Converts historical service interactions into reusable procedural experience by inducing workflows with prerequisite-augmented action blocks. At deployment, aligns agent’s execution trajectory to retrieved workflows and performs prerequisite-aware feasibility reasoning for state-grounded next actions.

Result: Experiments on ToolSandbox and τ²-bench show consistent improvement across base models.

Conclusion: WISE-Flow enables self-evolving agents in user-facing service environments by leveraging historical interactions to create reusable workflows that improve agent performance and reliability.

Abstract: Large language model (LLM)-based agents are widely deployed in user-facing services but remain error-prone in new tasks, tend to repeat the same failure patterns, and show substantial run-to-run variability. Fixing failures via environment-specific training or manual patching is costly and hard to scale. To enable self-evolving agents in user-facing service environments, we propose WISE-Flow, a workflow-centric framework that converts historical service interactions into reusable procedural experience by inducing workflows with prerequisite-augmented action blocks. At deployment, WISE-Flow aligns the agent’s execution trajectory to retrieved workflows and performs prerequisite-aware feasibility reasoning to achieve state-grounded next actions. Experiments on ToolSandbox and $τ^2$-bench show consistent improvement across base models.

Method: SwiftMem uses query-aware retrieval with specialized indexing: temporal index for logarithmic-time range queries, and semantic DAG-Tag index with hierarchical tag structures. It also includes embedding-tag co-consolidation mechanism to reorganize storage based on semantic clusters for better cache locality.

Result: Experiments on LoCoMo and LongMemEval benchmarks show SwiftMem achieves 47× faster search compared to state-of-the-art baselines while maintaining competitive accuracy.

Conclusion: SwiftMem enables practical deployment of memory-augmented LLM agents by solving the latency bottleneck problem through efficient query-aware retrieval with specialized indexing techniques.

Abstract: Agentic memory systems have become critical for enabling LLM agents to maintain long-term context and retrieve relevant information efficiently. However, existing memory frameworks suffer from a fundamental limitation: they perform exhaustive retrieval across the entire storage layer regardless of query characteristics. This brute-force approach creates severe latency bottlenecks as memory grows, hindering real-time agent interactions. We propose SwiftMem, a query-aware agentic memory system that achieves sub-linear retrieval through specialized indexing over temporal and semantic dimensions. Our temporal index enables logarithmic-time range queries for time-sensitive retrieval, while the semantic DAG-Tag index maps queries to relevant topics through hierarchical tag structures. To address memory fragmentation during growth, we introduce an embedding-tag co-consolidation mechanism that reorganizes storage based on semantic clusters to improve cache locality. Experiments on LoCoMo and LongMemEval benchmarks demonstrate that SwiftMem achieves 47$\times$ faster search compared to state-of-the-art baselines while maintaining competitive accuracy, enabling practical deployment of memory-augmented LLM agents.

Method: Fine-tune function vectors (from causal mediation analysis) with ~20 word pairs; create composite function vectors via weighted combination; patch these vectors into LLM activations at inference.

Result: Fine-tuned function vectors outperform original vectors on relation tasks across model sizes, improve relation word decoding, and align better with human similarity judgments. Composite vectors significantly boost performance on challenging analogy problems.

Conclusion: Activation patching with fine-tuned and composite function vectors provides a controllable mechanism for encoding relational knowledge, advancing LLM interpretability and reasoning capabilities.

Abstract: Representing relations between concepts is a core prerequisite for intelligent systems to make sense of the world. Recent work using causal mediation analysis has shown that a small set of attention heads encodes task representation in in-context learning, captured in a compact representation known as the function vector. We show that fine-tuning function vectors with only a small set of examples (about 20 word pairs) yields better performance on relation-based word-completion tasks than using the original vectors derived from causal mediation analysis. These improvements hold for both small and large language models. Moreover, the fine-tuned function vectors yield improved decoding performance for relation words and show stronger alignment with human similarity judgments of semantic relations. Next, we introduce the composite function vector - a weighted combination of fine-tuned function vectors - to extract relational knowledge and support analogical reasoning. At inference time, inserting this composite vector into LLM activations markedly enhances performance on challenging analogy problems drawn from cognitive science and SAT benchmarks. Our results highlight the potential of activation patching as a controllable mechanism for encoding and manipulating relational knowledge, advancing both the interpretability and reasoning capabilities of large language models.

Method: Used CLARITY dataset from SemEval 2026 shared task. Compared GPT-3.5 baseline against GPT-5.2 with three prompting strategies: simple prompting, chain-of-thought prompting, and chain-of-thought with few-shot examples. Evaluated predictions against human annotations using accuracy, class-wise metrics, and hierarchical exact match.

Result: GPT-5.2 consistently outperforms GPT-3.5 baseline on clarity prediction. Chain-of-thought with few-shot prompting improved accuracy from 56% to 63%. Chain-of-thought prompting yielded highest evasion accuracy at 34%. Reasoning-based prompting improved topic identification accuracy from 60% to 74% relative to human annotations.

Conclusion: Prompt design reliably improves high-level clarity evaluation, but fine-grained evasion and topic detection remain challenging despite structured reasoning prompts, indicating areas for future improvement in automatic evaluation of LLM responses.

Abstract: Automatic evaluation of large language model (LLM) responses requires not only factual correctness but also clarity, particularly in political question-answering. While recent datasets provide human annotations for clarity and evasion, the impact of prompt design on automatic clarity evaluation remains underexplored. In this paper, we study prompt-based clarity evaluation using the CLARITY dataset from the SemEval 2026 shared task. We compare a GPT-3.5 baseline provided with the dataset against GPT-5.2 evaluated under three prompting strategies: simple prompting, chain-of-thought prompting, and chain-of-thought with few-shot examples. Model predictions are evaluated against human annotations using accuracy and class-wise metrics for clarity and evasion, along with hierarchical exact match. Results show that GPT-5.2 consistently outperforms the GPT-3.5 baseline on clarity prediction, with accuracy improving from 56 percent to 63 percent under chain-of-thought with few-shot prompting. Chain-of-thought prompting yields the highest evasion accuracy at 34 percent, though improvements are less stable across fine-grained evasion categories. We further evaluate topic identification and find that reasoning-based prompting improves accuracy from 60 percent to 74 percent relative to human annotations. Overall, our findings indicate that prompt design reliably improves high-level clarity evaluation, while fine-grained evasion and topic detection remain challenging despite structured reasoning prompts.

Method: LogiSafetyGen framework converts unstructured regulations into Linear Temporal Logic oracles and uses logic-guided fuzzing to synthesize safety-critical traces. LogiSafetyBench is constructed with 240 human-verified tasks requiring LLMs to generate Python programs satisfying both functional objectives and latent compliance rules.

Result: Evaluation of 13 SOTA LLMs shows larger models achieve better functional correctness but frequently prioritize task completion over safety, resulting in non-compliant behavior.

Conclusion: There’s a critical gap in current LLM evaluation regarding regulatory compliance, and larger models’ tendency to prioritize functionality over safety highlights the need for benchmarks that assess both aspects simultaneously.

Abstract: The integration of large language models (LLMs) into autonomous agents has enabled complex tool use, yet in high-stakes domains, these systems must strictly adhere to regulatory standards beyond simple functional correctness. However, existing benchmarks often overlook implicit regulatory compliance, thus failing to evaluate whether LLMs can autonomously enforce mandatory safety constraints. To fill this gap, we introduce LogiSafetyGen, a framework that converts unstructured regulations into Linear Temporal Logic oracles and employs logic-guided fuzzing to synthesize valid, safety-critical traces. Building on this framework, we construct LogiSafetyBench, a benchmark comprising 240 human-verified tasks that require LLMs to generate Python programs that satisfy both functional objectives and latent compliance rules. Evaluations of 13 state-of-the-art (SOTA) LLMs reveal that larger models, despite achieving better functional correctness, frequently prioritize task completion over safety, which results in non-compliant behavior.

Method: Triplet-based Self-Play fine-tuning (T-SPIN) with two key designs: 1) incorporates historical advantages between iteratively generated responses and proto-synthetic responses from initial policy, 2) introduces entropy constraint to enable reference-free fine-tuning.

Result: T-SPIN outperforms SPIN on various tasks with stable evolution during iterations. Achieves comparable or better performance than supervised fine-tuning with only 25% of samples.

Conclusion: T-SPIN effectively addresses SPIN’s limitations through historical advantages and entropy constraints, enabling stable optimization and better performance with scarce annotated data.

Abstract: Recently, self-play fine-tuning (SPIN) has been proposed to adapt large language models to downstream applications with scarce expert-annotated data, by iteratively generating synthetic responses from the model itself. However, SPIN is designed to optimize the current reward advantages of annotated responses over synthetic responses at hand, which may gradually vanish during iterations, leading to unstable optimization. Moreover, the utilization of reference policy induces a misalignment issue between the reward formulation for training and the metric for generation. To address these limitations, we propose a novel Triplet-based Self-Play fIne-tuNing (T-SPIN) method that integrates two key designs. First, beyond current advantages, T-SPIN additionally incorporates historical advantages between iteratively generated responses and proto-synthetic responses produced by the initial policy. Even if the current advantages diminish, historical advantages remain effective, stabilizing the overall optimization. Second, T-SPIN introduces the entropy constraint into the self-play framework, which is theoretically justified to support reference-free fine-tuning, eliminating the training-generation discrepancy. Empirical results on various tasks demonstrate not only the superior performance of T-SPIN over SPIN, but also its stable evolution during iterations. Remarkably, compared to supervised fine-tuning, T-SPIN achieves comparable or even better performance with only 25% samples, highlighting its effectiveness when faced with scarce annotated data.

Method: GAG treats private expertise as an additional expert modality and injects it via a compact, representation-level interface aligned to the frozen base model, inspired by how multimodal LLMs align heterogeneous modalities. This avoids prompt-time evidence serialization while enabling plug-and-play specialization and scalable multi-domain composition.

Result: GAG improves specialist performance over strong RAG baselines by 15.34% and 14.86% on two private scientific QA benchmarks (immunology adjuvant and catalytic materials), while maintaining performance on six open general benchmarks and enabling near-oracle selective activation for scalable multi-domain deployment.

Conclusion: GAG provides an effective solution for private knowledge injection that overcomes limitations of both fine-tuning and RAG, enabling reliable, scalable deployment of specialized LLMs in high-stakes domains while preserving general capabilities.

Abstract: In domains such as biomedicine, materials, and finance, high-stakes deployment of large language models (LLMs) requires injecting private, domain-specific knowledge that is proprietary, fast-evolving, and under-represented in public pretraining. However, the two dominant paradigms for private knowledge injection each have pronounced drawbacks: fine-tuning is expensive to iterate, and continual updates risk catastrophic forgetting and general-capability regression; retrieval-augmented generation (RAG) keeps the base model intact but is brittle in specialized private corpora due to chunk-induced evidence fragmentation, retrieval drift, and long-context pressure that yields query-dependent prompt inflation. Inspired by how multimodal LLMs align heterogeneous modalities into a shared semantic space, we propose Generation-Augmented Generation (GAG), which treats private expertise as an additional expert modality and injects it via a compact, representation-level interface aligned to the frozen base model, avoiding prompt-time evidence serialization while enabling plug-and-play specialization and scalable multi-domain composition with reliable selective activation. Across two private scientific QA benchmarks (immunology adjuvant and catalytic materials) and mixed-domain evaluations, GAG improves specialist performance over strong RAG baselines by 15.34% and 14.86% on the two benchmarks, respectively, while maintaining performance on six open general benchmarks and enabling near-oracle selective activation for scalable multi-domain deployment.

Method: Systematic study of MoE architectures to identify key performance factors. Analysis of how total parameters (N_total) and expert sparsity (s = n_exp/n_topk) affect model performance.

Result: MoE performance is primarily determined by total parameters and expert sparsity ratio. Larger number of experts slightly penalizes performance by forcing reduction in core model dimensions (depth and width) to meet memory constraints. The expert count and top-k values don’t cancel out within the sparsity ratio.

Conclusion: Optimal MoE design should maximize total parameters while minimizing expert sparsity (maximizing n_topk) and minimizing n_exp under given constraints. This provides a robust framework for resolving architectural ambiguity in MoE design.

Abstract: Modern Mixture-of-Experts (MoE) language models are designed based on total parameters (memory footprint) and active parameters (inference cost). However, we find these two factors alone are insufficient to describe an optimal architecture. Through a systematic study, we demonstrate that MoE performance is primarily determined by total parameters ($N_{total}$) and expert sparsity ($s:=n_{exp}/n_{topk}$). Moreover, $n_{exp}$ and $n_{topk}$ do not “cancel out” within the sparsity ratio; instead, a larger total number of experts slightly penalizes performance by forcing a reduction in core model dimensions (depth and width) to meet memory constraints. This motivates a simple principle for MoE design which maximizes $N_{total}$ while minimizing $s$ (maximizing $n_{topk}$) and $n_{exp}$ under the given constraints. Our findings provide a robust framework for resolving architectural ambiguity and guiding MoE design.

Method: Developed a user-oriented simulation paradigm that decouples task generation from a dedicated user simulator mimicking human behavioral rules (incremental request-making, turn-by-turn feedback). The pipeline operates as a plug-and-play module capable of initiating generation from any state, facilitating multiple task completions within single trajectories.

Result: The framework generates more authentic, extended multi-turn dialogues that reflect real-world iterative problem solving, producing high-density datasets with high turn-count conversations that better represent multifaceted human-agent interaction demands.

Conclusion: Shifting from task-oriented to user-oriented simulation enables scalable generation of realistic tool-use data that captures the extended, iterative nature of human-agent collaboration, addressing limitations of existing approaches for training large reasoning models as autonomous agents.

Abstract: The recent paradigm shift toward large reasoning models (LRMs) as autonomous agents has intensified the demand for sophisticated, multi-turn tool-use capabilities. Yet, existing datasets and data-generation approaches are limited by static, predefined toolsets that cannot scale to the complexity of open-ended human-agent collaboration. To address this, we initially developed a framework for automated task-oriented multi-turn dialogue generation at scale, utilizing an LRM-based simulator to dynamically generate high-value, domain-specific tools to solve specified tasks. However, we observe that a purely task-oriented design often results in “solely task-solving” trajectories, where the agent completes the objective with minimal interaction, failing to generate the high turn-count conversations seen in realistic scenarios. To bridge this gap, we shift toward a user-oriented simulation paradigm. By decoupling task generation from a dedicated user simulator that mimics human behavioral rules - such as incremental request-making and turn-by-turn feedback - we facilitate more authentic, extended multi-turn dialogues that reflect the iterative nature of real-world problem solving. Our generation pipeline operates as a versatile, plug-and-play module capable of initiating generation from any state, ensuring high scalability in producing extended tool-use data. Furthermore, by facilitating multiple task completions within a single trajectory, it yields a high-density dataset that reflects the multifaceted demands of real-world human-agent interaction.

Method: Med-CoReasoner uses language-informed co-reasoning framework that elicits parallel English and local-language reasoning, abstracts them into structured concepts, and integrates local clinical knowledge into English logical scaffold via concept-level alignment and retrieval.

Result: Experiments across three benchmarks show Med-CoReasoner improves multilingual reasoning performance by average of 5%, with particularly substantial gains in low-resource languages. Model distillation and expert evaluation confirm clinically sound and culturally grounded reasoning traces.

Conclusion: Med-CoReasoner effectively bridges multilingual medical reasoning gap by combining structural robustness of English reasoning with practice-grounded expertise in local languages, enabling more equitable global medical AI deployment.

Abstract: While reasoning-enhanced large language models perform strongly on English medical tasks, a persistent multilingual gap remains, with substantially weaker reasoning in local languages, limiting equitable global medical deployment. To bridge this gap, we introduce Med-CoReasoner, a language-informed co-reasoning framework that elicits parallel English and local-language reasoning, abstracts them into structured concepts, and integrates local clinical knowledge into an English logical scaffold via concept-level alignment and retrieval. This design combines the structural robustness of English reasoning with the practice-grounded expertise encoded in local languages. To evaluate multilingual medical reasoning beyond multiple-choice settings, we construct MultiMed-X, a benchmark covering seven languages with expert-annotated long-form question answering and natural language inference tasks, comprising 350 instances per language. Experiments across three benchmarks show that Med-CoReasoner improves multilingual reasoning performance by an average of 5%, with particularly substantial gains in low-resource languages. Moreover, model distillation and expert evaluation analysis further confirm that Med-CoReasoner produces clinically sound and culturally grounded reasoning traces.

Method: DART uses reinforcement learning with dynamic rollout trees that branch at promising positions to explore diverse tool-integrated trajectories during training. It then employs tree-based process advantage estimation to identify and credit specific sub-trajectories where tool invocation positively contributes to solutions.

Result: Extensive experiments on challenging benchmarks like AIME and GPQA-Diamond show that DART significantly outperforms existing methods, successfully harmonizing tool execution with long CoT reasoning.

Conclusion: DART provides an effective framework for enabling spontaneous tool-use in long chain-of-thought reasoning without human annotation, addressing key challenges in tool-integrated reasoning for LLMs.

Abstract: Tool-Integrated Reasoning has emerged as a key paradigm to augment Large Language Models (LLMs) with computational capabilities, yet integrating tool-use into long Chain-of-Thought (long CoT) remains underexplored, largely due to the scarcity of training data and the challenge of integrating tool-use without compromising the model’s intrinsic long-chain reasoning. In this paper, we introduce DART (Discovery And Reinforcement of Tool-Integrated Reasoning Chains via Rollout Trees), a reinforcement learning framework that enables spontaneous tool-use during long CoT reasoning without human annotation. DART operates by constructing dynamic rollout trees during training to discover valid tool-use opportunities, branching out at promising positions to explore diverse tool-integrated trajectories. Subsequently, a tree-based process advantage estimation identifies and credits specific sub-trajectories where tool invocation positively contributes to the solution, effectively reinforcing these beneficial behaviors. Extensive experiments on challenging benchmarks like AIME and GPQA-Diamond demonstrate that DART significantly outperforms existing methods, successfully harmonizing tool execution with long CoT reasoning.

Method: D²Plan uses a dual-agent paradigm with a Reasoner and a Purifier. The Reasoner constructs explicit global plans during reasoning and dynamically adapts them based on retrieval feedback. The Purifier assesses retrieval relevance and condenses key information for the Reasoner. Training involves a two-stage framework: SFT cold-start on synthesized trajectories followed by RL with plan-oriented rewards.

Result: Extensive experiments demonstrate that D²Plan enables more coherent multi-step reasoning and stronger resilience to irrelevant information, achieving superior performance on challenging QA benchmarks.

Conclusion: The dual-agent planning paradigm effectively addresses the failure modes of search-augmented LLMs in complex retrieval-augmented reasoning tasks, leading to improved performance through better plan construction and information filtering.

Abstract: Recent search-augmented LLMs trained with reinforcement learning (RL) can interleave searching and reasoning for multi-hop reasoning tasks. However, they face two critical failure modes as the accumulating context becomes flooded with both crucial evidence and irrelevant information: (1) ineffective search chain construction that produces incorrect queries or omits retrieval of critical information, and (2) reasoning hijacking by peripheral evidence that causes models to misidentify distractors as valid evidence. To address these challenges, we propose D$^2$Plan, a Dual-agent Dynamic global Planning paradigm for complex retrieval-augmented reasoning. D$^2$Plan operates through the collaboration of a Reasoner and a Purifier: the Reasoner constructs explicit global plans during reasoning and dynamically adapts them based on retrieval feedback; the Purifier assesses retrieval relevance and condenses key information for the Reasoner. We further introduce a two-stage training framework consisting of supervised fine-tuning (SFT) cold-start on synthesized trajectories and RL with plan-oriented rewards to teach LLMs to master the D$^2$Plan paradigm. Extensive experiments demonstrate that D$^2$Plan enables more coherent multi-step reasoning and stronger resilience to irrelevant information, thereby achieving superior performance on challenging QA benchmarks.

Method: Evaluated advanced prompting techniques (few-shot learning, chain-of-thought prompting, self-consistency) against baseline approaches. Tasks included sentiment classification, aspect-based sentiment analysis, and irony detection. Performance measured using accuracy, recall, precision, and F1 score across different datasets.

Result: Advanced prompting significantly improves sentiment analysis performance. Few-shot approach works best for GPT-4o-mini, while chain-of-thought prompting boosts irony detection in gemini-1.5-flash by up to 46%. Different models respond optimally to different prompting strategies.

Conclusion: Prompting strategies must be tailored to both the specific LLM architecture and the semantic complexity of the task. Effective prompt design requires alignment between model characteristics and task requirements for optimal sentiment analysis performance.

Abstract: This study investigates the use of prompt engineering to enhance large language models (LLMs), specifically GPT-4o-mini and gemini-1.5-flash, in sentiment analysis tasks. It evaluates advanced prompting techniques like few-shot learning, chain-of-thought prompting, and self-consistency against a baseline. Key tasks include sentiment classification, aspect-based sentiment analysis, and detecting subtle nuances such as irony. The research details the theoretical background, datasets, and methods used, assessing performance of LLMs as measured by accuracy, recall, precision, and F1 score. Findings reveal that advanced prompting significantly improves sentiment analysis, with the few-shot approach excelling in GPT-4o-mini and chain-of-thought prompting boosting irony detection in gemini-1.5-flash by up to 46%. Thus, while advanced prompting techniques overall improve performance, the fact that few-shot prompting works best for GPT-4o-mini and chain-of-thought excels in gemini-1.5-flash for irony detection suggests that prompting strategies must be tailored to both the model and the task. This highlights the importance of aligning prompt design with both the LLM’s architecture and the semantic complexity of the task.

Method: Two-level hierarchical framework: simple tasks handled by modality-specific agents with direct reasoning; complex tasks trigger contract-driven planning that formulates tasks as capability requirements, performs capability-aware tool orchestration, and enables dynamic tool generation with failure recovery.

Result: AgriAgent achieves higher execution success rates and robustness on complex tasks compared to existing tool-centric agent baselines that rely on unified execution paradigms.

Conclusion: The proposed hierarchical approach with contract-driven planning effectively addresses the challenges of varying task complexity and incomplete tool availability in real-world agricultural agent systems, enabling more reliable and robust performance.

Abstract: Intelligent agent systems in real-world agricultural scenarios must handle diverse tasks under multimodal inputs, ranging from lightweight information understanding to complex multi-step execution. However, most existing approaches rely on a unified execution paradigm, which struggles to accommodate large variations in task complexity and incomplete tool availability commonly observed in agricultural environments. To address this challenge, we propose AgriAgent, a two-level agent framework for real-world agriculture. AgriAgent adopts a hierarchical execution strategy based on task complexity: simple tasks are handled through direct reasoning by modality-specific agents, while complex tasks trigger a contract-driven planning mechanism that formulates tasks as capability requirements and performs capability-aware tool orchestration and dynamic tool generation, enabling multi-step and verifiable execution with failure recovery. Experimental results show that AgriAgent achieves higher execution success rates and robustness on complex tasks compared to existing tool-centric agent baselines that rely on unified execution paradigms. All code, data will be released at after our work be accepted to promote reproducible research.

Method: Created CLaS-Bench, a lightweight parallel-question benchmark for 32 languages. Evaluated various steering techniques: residual-stream DiffMean interventions, probe-derived directions, language-specific neurons, PCA/LDA vectors, Sparse Autoencoders, and prompting baselines. Performance measured via language control and semantic relevance combined into harmonic-mean steering score.

Result: Simple residual-based DiffMean method consistently outperforms all other methods across languages. Language-specific structure emerges predominantly in later layers, and steering directions cluster by language family.

Conclusion: CLaS-Bench provides the first standardized benchmark for multilingual steering, enabling both scientific analysis of language representations and practical evaluation of steering as a low-cost adaptation alternative.

Abstract: Understanding and controlling the behavior of large language models (LLMs) is an increasingly important topic in multilingual NLP. Beyond prompting or fine-tuning, , i.e.,~manipulating internal representations during inference, has emerged as a more efficient and interpretable technique for adapting models to a target language. Yet, no dedicated benchmarks or evaluation protocols exist to quantify the effectiveness of steering techniques. We introduce CLaS-Bench, a lightweight parallel-question benchmark for evaluating language-forcing behavior in LLMs across 32 languages, enabling systematic evaluation of multilingual steering methods. We evaluate a broad array of steering techniques, including residual-stream DiffMean interventions, probe-derived directions, language-specific neurons, PCA/LDA vectors, Sparse Autoencoders, and prompting baselines. Steering performance is measured along two axes: language control and semantic relevance, combined into a single harmonic-mean steering score. We find that across languages simple residual-based DiffMean method consistently outperforms all other methods. Moreover, a layer-wise analysis reveals that language-specific structure emerges predominantly in later layers and steering directions cluster based on language family. CLaS-Bench is the first standardized benchmark for multilingual steering, enabling both rigorous scientific analysis of language representations and practical evaluation of steering as a low-cost adaptation alternative.

Method: Created SPEECHMENTALMANIP benchmark by augmenting text-based dataset with high-quality, voice-consistent Text-to-Speech rendered audio. Evaluated using few-shot large audio-language models and human annotation to compare text vs speech detection.

Result: Models show high specificity but markedly lower recall on speech compared to text, suggesting sensitivity to missing acoustic/prosodic cues. Human raters also show similar uncertainty with audio, highlighting inherent ambiguity of manipulative speech.

Conclusion: Modality-aware evaluation and safety alignment are needed in multimodal dialogue systems, as speech introduces unique challenges for detecting mental manipulation compared to text.

Abstract: Mental manipulation, the strategic use of language to covertly influence or exploit others, is a newly emerging task in computational social reasoning. Prior work has focused exclusively on textual conversations, overlooking how manipulative tactics manifest in speech. We present the first study of mental manipulation detection in spoken dialogues, introducing a synthetic multi-speaker benchmark SPEECHMENTALMANIP that augments a text-based dataset with high-quality, voice-consistent Text-to-Speech rendered audio. Using few-shot large audio-language models and human annotation, we evaluate how modality affects detection accuracy and perception. Our results reveal that models exhibit high specificity but markedly lower recall on speech compared to text, suggesting sensitivity to missing acoustic or prosodic cues in training. Human raters show similar uncertainty in the audio setting, underscoring the inherent ambiguity of manipulative speech. Together, these findings highlight the need for modality-aware evaluation and safety alignment in multimodal dialogue systems.

Method: 1) Constructed a taxonomy linking pedagogical methods to personality profiles based on pedagogical literature. 2) Simulated student-teacher conversations where LLM tutors adjust strategies to simulated student personalities. 3) Evaluated the approach with human teachers comparing against baselines.

Result: Human teachers consistently preferred the personality-adaptive approach over two baselines. The method increased use of less common but high-impact strategies like role-playing, which both human and LLM annotators significantly preferred.

Conclusion: Personalizing LLM tutoring strategies based on student personality traits creates more effective educational applications and paves the way for more personalized LLM use in education.

Abstract: Recent advances in large language models (LLMs) demonstrate their potential as educational tutors. However, different tutoring strategies benefit different student personalities, and mismatches can be counterproductive to student outcomes. Despite this, current LLM tutoring systems do not take into account student personality traits. To address this problem, we first construct a taxonomy that links pedagogical methods to personality profiles, based on pedagogical literature. We simulate student-teacher conversations and use our framework to let the LLM tutor adjust its strategy to the simulated student personality. We evaluate the scenario with human teachers and find that they consistently prefer our approach over two baselines. Our method also increases the use of less common, high-impact strategies such as role-playing, which human and LLM annotators prefer significantly. Our findings pave the way for developing more personalized and effective LLM use in educational applications.

Method: Latent-GRPO derives intrinsic rewards from latent space geometry using Iterative Robust Centroid Estimation (IRCE). IRCE generates dense, continuous rewards by mitigating magnitude fluctuations via spherical projection and estimating a robust “truth centroid” through iterative aggregation.

Result: Experimental results show the method maintains model performance while achieving over 2x training speedup compared to baselines, with strong generalization ability and robustness across multiple datasets.

Conclusion: Latent-GRPO provides an efficient alternative to external verifiers by leveraging latent space geometry for intrinsic rewards, significantly improving training efficiency without sacrificing performance.

Abstract: Group Relative Policy Optimization (GRPO) significantly enhances the reasoning performance of Large Language Models (LLMs). However, this success heavily relies on expensive external verifiers or human rules. Such dependency not only leads to significant computational costs and training latency, but also yields sparse rewards that hinder optimization efficiency. To address these challenges, we propose Latent-GRPO, a framework that derives intrinsic rewards directly from latent space geometry. Crucially, our empirical analysis reveals a compelling geometric property: terminal token representations of correct reasoning trajectories form dense clusters with high intra-class similarity, whereas incorrect trajectories remain scattered as outliers. In light of this discovery, we introduce the Iterative Robust Centroid Estimation (IRCE) algorithm, which generates dense, continuous rewards by mitigating magnitude fluctuations via spherical projection and estimating a robust ``truth centroid’’ through iterative aggregation. Experimental results on multiple datasets show that our method maintains model performance while achieving a training speedup of over 2x compared to baselines. Furthermore, extensive results demonstrate strong generalization ability and robustness. The code will be released soon.

Method: 1) Chunk-level Step Reward: Provides immediate step-level supervision via auxiliary chunk-specific QA tasks. 2) Evidence-Anchored Reward Attribution: Redistributes global rewards by anchoring credit to key memory operations based on specific memory items used as evidence in reasoning.

Result: Fine-Mem consistently outperforms strong baselines on Memalpha and MemoryAgentBench, achieving superior success rates across various sub-tasks. It shows adaptability and strong generalization across diverse model configurations and backbones.

Conclusion: The proposed fine-grained feedback framework enables stable policy optimization and aligns local memory operations with long-term utility of memory, effectively addressing reward sparsity in memory management for LLM agents.

Abstract: Effective memory management is essential for large language model agents to navigate long-horizon tasks. Recent research has explored using Reinforcement Learning to develop specialized memory manager agents. However, existing approaches rely on final task performance as the primary reward, which results in severe reward sparsity and ineffective credit assignment, providing insufficient guidance for individual memory operations. To this end, we propose Fine-Mem, a unified framework designed for fine-grained feedback alignment. First, we introduce a Chunk-level Step Reward to provide immediate step-level supervision via auxiliary chunk-specific question answering tasks. Second, we devise Evidence-Anchored Reward Attribution to redistribute global rewards by anchoring credit to key memory operations, based on the specific memory items utilized as evidence in reasoning. Together, these components enable stable policy optimization and align local memory operations with the long-term utility of memory. Experiments on Memalpha and MemoryAgentBench demonstrate that Fine-Mem consistently outperforms strong baselines, achieving superior success rates across various sub-tasks. Further analysis reveals its adaptability and strong generalization capabilities across diverse model configurations and backbones.

Method: Two-stage judge-then-generate paradigm: 1) Train model to judge solution responses with verifiable answers (discriminative pre-training), 2) Fine-tune same model with vanilla generating RLVR initialized from the judge model.

Result: JudgeRLVR achieves better quality-efficiency trade-off than Vanilla RLVR: +3.7 points average accuracy gain with -42% average generation length on in-domain math; +4.5 points average accuracy improvement on out-of-domain benchmarks, showing enhanced generalization.

Conclusion: Discriminative capability is prerequisite for efficient generation - learning to distinguish valid solutions helps models internalize guidance signals that prune search space, leading to more structured reasoning with better accuracy and efficiency.

Abstract: Reinforcement Learning with Verifiable Rewards (RLVR) has become a standard paradigm for reasoning in Large Language Models. However, optimizing solely for final-answer correctness often drives models into aimless, verbose exploration, where they rely on exhaustive trial-and-error tactics rather than structured planning to reach solutions. While heuristic constraints like length penalties can reduce verbosity, they often truncate essential reasoning steps, creating a difficult trade-off between efficiency and verification. In this paper, we argue that discriminative capability is a prerequisite for efficient generation: by learning to distinguish valid solutions, a model can internalize a guidance signal that prunes the search space. We propose JudgeRLVR, a two-stage judge-then-generate paradigm. In the first stage, we train the model to judge solution responses with verifiable answers. In the second stage, we fine-tune the same model with vanilla generating RLVR initialized from the judge. Compared to Vanilla RLVR using the same math-domain training data, JudgeRLVR achieves a better quality–efficiency trade-off for Qwen3-30B-A3B: on in-domain math, it delivers about +3.7 points average accuracy gain with -42% average generation length; on out-of-domain benchmarks, it delivers about +4.5 points average accuracy improvement, demonstrating enhanced generalization.

Method: Developed a 24B parameter model with a synthetic data pipeline combining chain-of-thought prompting with multi-stage verification, generating over 22,000 high-quality training examples across five languages from diverse sources including parliamentary documents, web text, and Wikipedia.

Result: sui-1 significantly outperforms all tested open-weight baselines, including models with 3x more parameters, demonstrating that task-specific training substantially outperforms scale alone for citation-grounded summarization.

Conclusion: Specialized training for citation-grounded summarization is more effective than simply scaling model size, and the approach enables faithful, verifiable summaries critical for compliance-sensitive applications.

Abstract: Large language models frequently generate plausible but unfaithful summaries that users cannot verify against source text, a critical limitation in compliance-sensitive domains such as government and legal analysis. We present sui-1, a 24B parameter model that produces abstractive summaries with inline citations, enabling users to trace each claim to its source sentence. Our synthetic data pipeline combines chain-of-thought prompting with multi-stage verification, generating over 22,000 high-quality training examples across five languages from diverse sources including parliamentary documents, web text, and Wikipedia. Evaluation shows sui-1 significantly outperforms all tested open-weight baselines, including models with 3x more parameters. These results demonstrate that task-specific training substantially outperforms scale alone for citation-grounded summarization. Model weights and an interactive demo are publicly available.

Method: Transformer-based model extracts dialogue features, which are translated into a Stochastic Hybrid Automaton model of dyadic therapy sessions. Empathy-related properties are verified through Statistical Model Checking, and strategy synthesis provides guidance for shaping agent behavior.

Result: Preliminary results show the formal model captures therapy dynamics with good fidelity, and ad-hoc strategies improve the probability of satisfying empathy requirements.

Conclusion: The proposed framework successfully integrates NLP and formal verification to address the empathy gap in therapy chatbot development, providing systematic means to specify and verify empathy-related properties.

Abstract: Conversational agents are increasingly used as support tools along mental therapeutic pathways with significant societal impacts. In particular, empathy is a key non-functional requirement in therapeutic contexts, yet current chatbot development practices provide no systematic means to specify or verify it. This paper envisions a framework integrating natural language processing and formal verification to deliver empathetic therapy chatbots. A Transformer-based model extracts dialogue features, which are then translated into a Stochastic Hybrid Automaton model of dyadic therapy sessions. Empathy-related properties can then be verified through Statistical Model Checking, while strategy synthesis provides guidance for shaping agent behavior. Preliminary results show that the formal model captures therapy dynamics with good fidelity and that ad-hoc strategies improve the probability of satisfying empathy requirements.

Method: SRA constructs a registry of independent Concept Atoms representing protected capabilities and stylistic confounds, then uses ridge-regularized spectral residualization to orthogonalize the refusal vector against these directions, yielding a clean refusal direction.

Result: Across five models, SRA achieves deep refusal reduction (0-2%) with negligible perplexity impact (mean delta PPL ≈ 0.02) and minimal distribution drift, while preserving math and code capabilities on GSM8K and MBPP benchmarks.

Conclusion: Common “model damage” from refusal ablation is often “Ghost Noise” - spectral bleeding of dirty refusal directions into capability subspaces, which SRA avoids by surgical orthogonalization against protected concept atoms.

Abstract: Safety-aligned language models systematically refuse harmful requests. While activation steering can modulate refusal, ablating the raw “refusal vector” calculated from contrastive harmful and harmless prompts often causes collateral damage and distribution drift. We argue this degradation occurs because the raw vector is polysemantic, entangling the refusal signal with core capability circuits and linguistic style. We introduce Surgical Refusal Ablation (SRA) to distill these steering directions. SRA constructs a registry of independent Concept Atoms representing protected capabilities and stylistic confounds, then uses ridge-regularized spectral residualization to orthogonalize the refusal vector against these directions. This yields a clean refusal direction that targets refusal-relevant structure while minimizing disruption to the model’s semantic geometry. Across five models (Qwen3-VL and Ministral series), SRA achieves deep refusal reduction (0-2%) with negligible perplexity impact on Wikitext-2 (mean delta PPL approx. 0.02) and minimal distribution drift. Notably, standard ablation on Qwen3-VL-4B induces severe drift (first-token KL = 2.088), whereas SRA maintains the original distribution (KL = 0.044) while achieving the same 0% refusal rate. Using teacher-forced perplexity on GSM8K and MBPP as a high-resolution capability proxy, we show SRA preserves math and code distributions. These results suggest that common “model damage” is often “Ghost Noise,” defined as the spectral bleeding of the dirty refusal direction into capability subspaces.

Method: Created BenchOverflow benchmark with 9 plain-text prompting strategies that amplify output volume. Evaluated 9 open/closed-source models with 5000 token budget, measuring cap-saturation rates (CSR), empirical CDFs, within-prompt variance, and cross-model correlations. Tested a lightweight mitigation (fixed conciseness reminder).

Result: Models show pronounced rightward shifts and heavy tails in length distributions. Overflow is broadly reproducible but heterogeneous across model families and attack vectors. The simple conciseness reminder effectively attenuates right tails and lowers CSR for most models across all strategies.

Conclusion: Length control is a measurable reliability, cost, and sustainability concern, not just stylistic. BenchOverflow enables standardized comparison of length-control robustness, helping select deployments that minimize waste and evaluate defenses against compute amplification.

Abstract: We investigate a failure mode of large language models (LLMs) in which plain-text prompts elicit excessive outputs, a phenomenon we term Overflow. Unlike jailbreaks or prompt injection, Overflow arises under ordinary interaction settings and can lead to elevated serving cost, latency, and cross-user performance degradation, particularly when scaled across many requests. Beyond usability, the stakes are economic and environmental: unnecessary tokens increase per-request cost and energy consumption, compounding into substantial operational spend and carbon footprint at scale. Moreover, Overflow represents a practical vector for compute amplification and service degradation in shared environments. We introduce BenchOverflow, a model-agnostic benchmark of nine plain-text prompting strategies that amplify output volume without adversarial suffixes or policy circumvention. Using a standardized protocol with a fixed budget of 5000 new tokens, we evaluate nine open- and closed-source models and observe pronounced rightward shifts and heavy tails in length distributions. Cap-saturation rates (CSR@1k/3k/5k) and empirical cumulative distribution functions (ECDFs) quantify tail risk; within-prompt variance and cross-model correlations show that Overflow is broadly reproducible yet heterogeneous across families and attack vectors. A lightweight mitigation-a fixed conciseness reminder-attenuates right tails and lowers CSR for all strategies across the majority of models. Our findings position length control as a measurable reliability, cost, and sustainability concern rather than a stylistic quirk. By enabling standardized comparison of length-control robustness across models, BenchOverflow provides a practical basis for selecting deployments that minimize resource waste and operating expense, and for evaluating defenses that curb compute amplification without eroding task performance.

Method: Unsupervised ensemble approach combining Small Language Model (SLM) and LLM. Uses multilingual bi-encoder (BELA) for candidate retrieval and instruction-tuned LLM for NIL prediction and candidate selection via prompt chaining. SLM’s confidence scores discriminate between easy/hard samples, applying LLM only for hard cases to reduce computational costs and prevent hallucinations.

Result: Outperforms state-of-the-art models on four established benchmarks in six European languages (English, Finnish, French, German, Italian, Swedish) from 19th-20th centuries without requiring fine-tuning.

Conclusion: MHEL-LLaMo offers a scalable solution for low-resource historical EL, reducing computational costs while maintaining performance. Implementation available on Github.

Abstract: Despite the recent advancements in NLP with the advent of Large Language Models (LLMs), Entity Linking (EL) for historical texts remains challenging due to linguistic variation, noisy inputs, and evolving semantic conventions. Existing solutions either require substantial training data or rely on domain-specific rules that limit scalability. In this paper, we present MHEL-LLaMo (Multilingual Historical Entity Linking with Large Language MOdels), an unsupervised ensemble approach combining a Small Language Model (SLM) and an LLM. MHEL-LLaMo leverages a multilingual bi-encoder (BELA) for candidate retrieval and an instruction-tuned LLM for NIL prediction and candidate selection via prompt chaining. Our system uses SLM’s confidence scores to discriminate between easy and hard samples, applying an LLM only for hard cases. This strategy reduces computational costs while preventing hallucinations on straightforward cases. We evaluate MHEL-LLaMo on four established benchmarks in six European languages (English, Finnish, French, German, Italian and Swedish) from the 19th and 20th centuries. Results demonstrate that MHEL-LLaMo outperforms state-of-the-art models without requiring fine-tuning, offering a scalable solution for low-resource historical EL. The implementation of MHEL-LLaMo is available on Github.

Method: STAGE introduces four unified tasks: 1) knowledge graph construction to build narrative world representations, 2) scene-level event summarization for abstracting narrative events, 3) long-context screenplay question answering for reasoning over extended narratives, and 4) in-script character role-playing for generating character-consistent responses. The benchmark provides cleaned scripts, curated knowledge graphs, and event/character annotations for 150 films across English and Chinese.

Result: The benchmark enables holistic evaluation of models’ abilities to build world representations, abstract and verify narrative events, reason over long narratives, and generate character-consistent responses. It provides a comprehensive dataset with multi-task annotations grounded in shared narrative worlds.

Conclusion: STAGE addresses the gap in evaluating narrative understanding by providing a unified benchmark that tests models’ ability to construct and consistently use story world representations across multiple reasoning and generation tasks, moving beyond isolated subtasks to holistic narrative comprehension.

Abstract: Movie screenplays are rich long-form narratives that interleave complex character relationships, temporally ordered events, and dialogue-driven interactions. While prior benchmarks target individual subtasks such as question answering or dialogue generation, they rarely evaluate whether models can construct a coherent story world and use it consistently across multiple forms of reasoning and generation. We introduce STAGE (Screenplay Text, Agents, Graphs and Evaluation), a unified benchmark for narrative understanding over full-length movie screenplays. STAGE defines four tasks: knowledge graph construction, scene-level event summarization, long-context screenplay question answering, and in-script character role-playing, all grounded in a shared narrative world representation. The benchmark provides cleaned scripts, curated knowledge graphs, and event- and character-centric annotations for 150 films across English and Chinese, enabling holistic evaluation of models’ abilities to build world representations, abstract and verify narrative events, reason over long narratives, and generate character-consistent responses.

Method: Proposes STAR (State-Transition Amplification Ratio) framework that detects backdoors by analyzing output probability shifts. It exploits statistical discrepancy where malicious input-induced paths show high posterior probability despite low prior probability. Quantifies state-transition amplification and uses CUSUM algorithm to detect persistent anomalies.

Result: Experiments across diverse models (8B-70B) and five benchmark datasets show STAR achieves robust generalization with near-perfect performance (AUROC ≈ 1.0) and approximately 42× greater efficiency than existing baselines. Framework remains robust against adaptive attacks.

Conclusion: STAR provides an effective and efficient solution for detecting inference-time backdoors in reasoning-enhanced LLMs, addressing a critical security vulnerability in modern language models.

Abstract: Recent LLMs increasingly integrate reasoning mechanisms like Chain-of-Thought (CoT). However, this explicit reasoning exposes a new attack surface for inference-time backdoors, which inject malicious reasoning paths without altering model parameters. Because these attacks generate linguistically coherent paths, they effectively evade conventional detection. To address this, we propose STAR (State-Transition Amplification Ratio), a framework that detects backdoors by analyzing output probability shifts. STAR exploits the statistical discrepancy where a malicious input-induced path exhibits high posterior probability despite a low prior probability in the model’s general knowledge. We quantify this state-transition amplification and employ the CUSUM algorithm to detect persistent anomalies. Experiments across diverse models (8B-70B) and five benchmark datasets demonstrate that STAR exhibits robust generalization capabilities, consistently achieving near-perfect performance (AUROC $\approx$ 1.0) with approximately $42\times$ greater efficiency than existing baselines. Furthermore, the framework proves robust against adaptive attacks attempting to bypass detection.

Method: Presents a comprehensive characterization theorem unifying seven equivalent conditions for unconditional convergence: permutation invariance, net convergence, subseries tests, sign stability, bounded multiplier properties, and weak uniform convergence.

Result: Establishes the strict separation of conditional, unconditional, and absolute convergence in general Banach spaces (via Dvoretzky-Rogers theorem) and provides theoretical results that directly inform algorithmic stability analysis.

Conclusion: The work provides rigorous mathematical foundations for order-independent summation processes, governing permutation invariance in gradient accumulation for Stochastic Gradient Descent and justifying coefficient thresholding in frame-based signal processing.

Abstract: The distinction between conditional, unconditional, and absolute convergence in infinite-dimensional spaces has fundamental implications for computational algorithms. While these concepts coincide in finite dimensions, the Dvoretzky-Rogers theorem establishes their strict separation in general Banach spaces. We present a comprehensive characterization theorem unifying seven equivalent conditions for unconditional convergence: permutation invariance, net convergence, subseries tests, sign stability, bounded multiplier properties, and weak uniform convergence. These theoretical results directly inform algorithmic stability analysis, governing permutation invariance in gradient accumulation for Stochastic Gradient Descent and justifying coefficient thresholding in frame-based signal processing. Our work bridges classical functional analysis with contemporary computational practice, providing rigorous foundations for order-independent and numerically robust summation processes.

Method: Created Deep Research Bench II with 132 grounded research tasks across 22 domains. Developed 9430 fine-grained binary rubrics covering information recall, analysis, and presentation dimensions. Rubrics were derived from expert-written investigative articles using a four-stage LLM+human pipeline with over 400 human-hours of expert review to ensure atomic, verifiable criteria aligned with human judgment.

Result: Evaluation of state-of-the-art deep-research systems shows that even the strongest models satisfy fewer than 50% of the rubrics, revealing a substantial gap between current DRSs and human experts.

Conclusion: Deep Research Bench II provides a rigorous, human-aligned benchmark for evaluating Deep Research Systems, exposing significant limitations in current systems and establishing a foundation for more reliable assessment of research capabilities.

Abstract: Deep Research Systems (DRS) aim to help users search the web, synthesize information, and deliver comprehensive investigative reports. However, how to rigorously evaluate these systems remains under-explored. Existing deep-research benchmarks often fall into two failure modes. Some do not adequately test a system’s ability to analyze evidence and write coherent reports. Others rely on evaluation criteria that are either overly coarse or directly defined by LLMs (or both), leading to scores that can be biased relative to human experts and are hard to verify or interpret. To address these issues, we introduce Deep Research Bench II, a new benchmark for evaluating DRS-generated reports. It contains 132 grounded research tasks across 22 domains; for each task, a system must produce a long-form research report that is evaluated by a set of 9430 fine-grained binary rubrics in total, covering three dimensions: information recall, analysis, and presentation. All rubrics are derived from carefully selected expert-written investigative articles and are constructed through a four-stage LLM+human pipeline that combines automatic extraction with over 400 human-hours of expert review, ensuring that the criteria are atomic, verifiable, and aligned with human expert judgment. We evaluate several state-of-the-art deep-research systems on Deep Research Bench II and find that even the strongest models satisfy fewer than 50% of the rubrics, revealing a substantial gap between current DRSs and human experts.

Method: Cascade Distillation - iterative pruning and continued training with distillation technique to derive efficient models from larger ones.

Result: Three model sizes (3B, 8B, 14B) each with three variants: base, instruction finetuned, and reasoning models, all with image understanding capabilities.

Conclusion: Ministral 3 series provides efficient, versatile language models for constrained environments with multiple specialized variants and open licensing.

Abstract: We introduce the Ministral 3 series, a family of parameter-efficient dense language models designed for compute and memory constrained applications, available in three model sizes: 3B, 8B, and 14B parameters. For each model size, we release three variants: a pretrained base model for general-purpose use, an instruction finetuned, and a reasoning model for complex problem-solving. In addition, we present our recipe to derive the Ministral 3 models through Cascade Distillation, an iterative pruning and continued training with distillation technique. Each model comes with image understanding capabilities, all under the Apache 2.0 license.

Method: ExpSeek shifts experience toward step-level proactive seeking by: (1) estimating step-level entropy thresholds using the model’s intrinsic signals to determine intervention timing, and (2) designing step-level tailor-designed experience content.

Result: Experiments on Qwen3-8B and 32B models across four challenging web agent benchmarks show absolute improvements of 9.3% and 7.5% respectively. The approach validates entropy as a self-triggering signal and shows that even a 4B small-scale experience model can significantly boost larger agent models.

Conclusion: ExpSeek demonstrates that proactive, step-level experience seeking with entropy-based triggering is an effective approach for enhancing web agent capabilities, outperforming passive global experience injection methods.

Abstract: Experience intervention in web agents emerges as a promising technical paradigm, enhancing agent interaction capabilities by providing valuable insights from accumulated experiences. However, existing methods predominantly inject experience passively as global context before task execution, struggling to adapt to dynamically changing contextual observations during agent-environment interaction. We propose ExpSeek, which shifts experience toward step-level proactive seeking: (1) estimating step-level entropy thresholds to determine intervention timing using the model’s intrinsic signals; (2) designing step-level tailor-designed experience content. Experiments on Qwen3-8B and 32B models across four challenging web agent benchmarks demonstrate that ExpSeek achieves absolute improvements of 9.3% and 7.5%, respectively. Our experiments validate the feasibility and advantages of entropy as a self-triggering signal, reveal that even a 4B small-scale experience model can significantly boost the performance of larger agent models.

Method: GraphSearch framework with Graph-aware Query Planner (disentangles search space from semantic queries) and Graph-aware Retriever (constructs candidates based on topology with hybrid scoring). Two traversal modes: GraphSearch-R (recursive hop-by-hop expansion) and GraphSearch-F (flexible retrieval across local/global neighborhoods).

Result: Achieves competitive or superior performance compared to supervised graph learning methods, setting state-of-the-art results in zero-shot node classification and link prediction across diverse benchmarks.

Conclusion: GraphSearch establishes a flexible, generalizable paradigm for agentic reasoning over graphs, enabling effective zero-shot graph learning without task-specific fine-tuning.

Abstract: Recent advances in search-augmented large reasoning models (LRMs) enable the retrieval of external knowledge to reduce hallucinations in multistep reasoning. However, their ability to operate on graph-structured data, prevalent in domains such as e-commerce, social networks, and scientific citations, remains underexplored. Unlike plain text corpora, graphs encode rich topological signals that connect related entities and can serve as valuable priors for retrieval, enabling more targeted search and improved reasoning efficiency. Yet, effectively leveraging such structure poses unique challenges, including the difficulty of generating graph-expressive queries and ensuring reliable retrieval that balances structural and semantic relevance. To address this gap, we introduce GraphSearch, the first framework that extends search-augmented reasoning to graph learning, enabling zero-shot graph learning without task-specific fine-tuning. GraphSearch combines a Graph-aware Query Planner, which disentangles search space (e.g., 1-hop, multi-hop, or global neighbors) from semantic queries, with a Graph-aware Retriever, which constructs candidate sets based on topology and ranks them using a hybrid scoring function. We further instantiate two traversal modes: GraphSearch-R, which recursively expands neighborhoods hop by hop, and GraphSearch-F, which flexibly retrieves across local and global neighborhoods without hop constraints. Extensive experiments across diverse benchmarks show that GraphSearch achieves competitive or even superior performance compared to supervised graph learning methods, setting state-of-the-art results in zero-shot node classification and link prediction. These findings position GraphSearch as a flexible and generalizable paradigm for agentic reasoning over graphs.

Method: Introduces OrderProbe, a deterministic benchmark using fixed four-character expressions in Chinese, Japanese, and Korean that have unique canonical orders, enabling exact-match scoring. Also proposes a diagnostic framework evaluating models on recovery accuracy, semantic fidelity, logical validity, consistency, robustness sensitivity, and information density.

Result: Experiments on twelve widely used LLMs show structural reconstruction remains difficult even for frontier systems, with zero-shot recovery frequently falling below 35%. The study reveals a consistent dissociation between semantic recall and structural planning, indicating structural robustness is not an automatic byproduct of semantic competence.

Conclusion: Structural reconstruction is a challenging task for LLMs that requires specific capabilities beyond semantic understanding. The OrderProbe benchmark provides a reliable way to evaluate these structural abilities, revealing important gaps in current models’ architectural planning capabilities.

Abstract: Large language models (LLMs) excel at semantic understanding, yet their ability to reconstruct internal structure from scrambled inputs remains underexplored. Sentence-level restoration is ill-posed for automated evaluation because multiple valid word orders often exist. We introduce OrderProbe, a deterministic benchmark for structural reconstruction using fixed four-character expressions in Chinese, Japanese, and Korean, which have a unique canonical order and thus support exact-match scoring. We further propose a diagnostic framework that evaluates models beyond recovery accuracy, including semantic fidelity, logical validity, consistency, robustness sensitivity, and information density. Experiments on twelve widely used LLMs show that structural reconstruction remains difficult even for frontier systems: zero-shot recovery frequently falls below 35%. We also observe a consistent dissociation between semantic recall and structural planning, suggesting that structural robustness is not an automatic byproduct of semantic competence.

Method: Developed LALITA (Lexical And Linguistically Informed Text Analysis) framework that evaluates English-Hindi bi-text using lexical and linguistic features to select source sentences. Focuses on selecting complex sentences from both existing and synthetic datasets to curate parallel corpora.

Result: Training on complex sentences significantly improves translation quality. LALITA reduces data needs by more than half across multiple languages (Hindi, Odia, Nepali, Norwegian Nynorsk, and German). Improved performance reported on all tested data sizes (50K to 800K English sentences).

Conclusion: LALITA provides an effective framework for data curation in low-resource MT settings, reducing training costs through intelligent sentence selection and demonstrating utility in data augmentation while maintaining or improving translation quality.

Abstract: Data curation is a critical yet under-researched step in the machine translation training paradigm. To train translation systems, data acquisition relies primarily on human translations and digital parallel sources or, to a limited degree, synthetic generation. But, for low-resource languages, human translation to generate sufficient data is prohibitively expensive. Therefore, it is crucial to develop a framework that screens source sentences to form efficient parallel text, ensuring optimal MT system performance in low-resource environments. We approach this by evaluating English-Hindi bi-text to determine effective sentence selection strategies for optimal MT system training. Our extensively tested framework, (Lexical And Linguistically Informed Text Analysis) LALITA, targets source sentence selection using lexical and linguistic features to curate parallel corpora. We find that by training mostly on complex sentences from both existing and synthetic datasets, our method significantly improves translation quality. We test this by simulating low-resource data availabilty with curated datasets of 50K to 800K English sentences and report improved performances on all data sizes. LALITA demonstrates remarkable efficiency, reducing data needs by more than half across multiple languages (Hindi, Odia, Nepali, Norwegian Nynorsk, and German). This approach not only reduces MT systems training cost by reducing training data requirement, but also showcases LALITA’s utility in data augmentation.

Method: The researchers treat moral orientation as a controllable condition by conditioning models to endorse or reject specific moral values, then evaluate resulting shifts using the Political Compass Test. They examine how moral conditioning steers model trajectories across economic and social dimensions, and study effects modulated by role framing and model scale.

Result: Moral conditioning induces pronounced, value-specific shifts in models’ political coordinates. These effects are systematically modulated by role framing and model scale, and are robust across alternative assessment instruments instantiating the same moral value.

Conclusion: Effective alignment requires anchoring political assessments within the context of broader social values including morality, paving the way for more socially grounded alignment techniques.

Abstract: While recent research has systematically documented political orientation in large language models (LLMs), existing evaluations rely primarily on direct probing or demographic persona engineering to surface ideological biases. In social psychology, however, political ideology is also understood as a downstream consequence of fundamental moral intuitions. In this work, we investigate the causal relationship between moral values and political positioning by treating moral orientation as a controllable condition. Rather than simply assigning a demographic persona, we condition models to endorse or reject specific moral values and evaluate the resulting shifts on their political orientations, using the Political Compass Test. By treating moral values as lenses, we observe how moral conditioning actively steers model trajectories across economic and social dimensions. Our findings show that such conditioning induces pronounced, value-specific shifts in models’ political coordinates. We further notice that these effects are systematically modulated by role framing and model scale, and are robust across alternative assessment instruments instantiating the same moral value. This highlights that effective alignment requires anchoring political assessments within the context of broader social values including morality, paving the way for more socially grounded alignment techniques.

Method: Created a parallel dataset with language experts following multilingual guidelines. Each PIE includes both textual and visual components, with five images per PIE representing a spectrum from idiomatic to literal meanings, plus semantically related and random distractors.

Result: XMPIE dataset containing 34 languages and over ten thousand items, enabling comparative analyses of idiomatic patterns, evaluation of cross-lingual transfer of idiomatic understanding, and study of modality transfer between text and image.

Conclusion: XMPIE provides a high-quality benchmark for evaluating multilingual and multimodal idiomatic language understanding, allowing researchers to assess model performance across languages and modalities while gaining insights into shared cultural aspects.

Abstract: Potentially idiomatic expressions (PIEs) construe meanings inherently tied to the everyday experience of a given language community. As such, they constitute an interesting challenge for assessing the linguistic (and to some extent cultural) capabilities of NLP systems. In this paper, we present XMPIE, a parallel multilingual and multimodal dataset of potentially idiomatic expressions. The dataset, containing 34 languages and over ten thousand items, allows comparative analyses of idiomatic patterns among language-specific realisations and preferences in order to gather insights about shared cultural aspects. This parallel dataset allows to evaluate model performance for a given PIE in different languages and whether idiomatic understanding in one language can be transferred to another. Moreover, the dataset supports the study of PIEs across textual and visual modalities, to measure to what extent PIE understanding in one modality transfers or implies in understanding in another modality (text vs. image). The data was created by language experts, with both textual and visual components crafted under multilingual guidelines, and each PIE is accompanied by five images representing a spectrum from idiomatic to literal meanings, including semantically related and random distractors. The result is a high-quality benchmark for evaluating multilingual and multimodal idiomatic language understanding.

Method: Builds on computational paradigm of learning in the limit, formalizing tasks of safe language identification and generation. Provides theoretical proofs about impossibility and hardness results.

Result: Proves safe language identification is impossible, and safe language generation is at least as hard as vanilla language identification (which is also impossible). Discusses several intractable and tractable cases.

Conclusion: First theoretical treatment of safe language generation reveals fundamental limitations - safe language identification is impossible, and safe generation is at least as hard as impossible identification problems, though some tractable cases exist.

Abstract: Recent results in learning a language in the limit have shown that, although language identification is impossible, language generation is tractable. As this foundational area expands, we need to consider the implications of language generation in real-world settings. This work offers the first theoretical treatment of safe language generation. Building on the computational paradigm of learning in the limit, we formalize the tasks of safe language identification and generation. We prove that under this model, safe language identification is impossible, and that safe language generation is at least as hard as (vanilla) language identification, which is also impossible. Last, we discuss several intractable and tractable cases.

Method: RULERS (Rubric Unification, Locking, and Evidence-anchored Robust Scoring) is a compiler-executor framework that: 1) compiles criteria into versioned immutable bundles, 2) enforces structured decoding with deterministic evidence verification, and 3) applies lightweight Wasserstein-based post-hoc calibration without updating model parameters.

Result: Extensive experiments on essay and summarization benchmarks show RULERS significantly outperforms baselines in human agreement, maintains strong stability against adversarial rubric perturbations, and enables smaller models to rival larger proprietary judges.

Conclusion: Reliable LLM judging requires executable rubrics, verifiable evidence, and calibrated scales rather than prompt phrasing alone. The framework demonstrates that structured, evidence-based approaches can overcome inherent stochasticity in frozen black-box models.

Abstract: The LLM-as-a-Judge paradigm promises scalable rubric-based evaluation, yet aligning frozen black-box models with human standards remains a challenge due to inherent generation stochasticity. We reframe judge alignment as a criteria transfer problem and isolate three recurrent failure modes: rubric instability caused by prompt sensitivity, unverifiable reasoning that lacks auditable evidence, and scale misalignment with human grading boundaries. To address these issues, we introduce RULERS (Rubric Unification, Locking, and Evidence-anchored Robust Scoring), a compiler-executor framework that transforms natural language rubrics into executable specifications. RULERS operates by compiling criteria into versioned immutable bundles, enforcing structured decoding with deterministic evidence verification, and applying lightweight Wasserstein-based post-hoc calibration, all without updating model parameters. Extensive experiments on essay and summarization benchmarks demonstrate that RULERS significantly outperforms representative baselines in human agreement, maintains strong stability against adversarial rubric perturbations, and enables smaller models to rival larger proprietary judges. Overall, our results suggest that reliable LLM judging requires executable rubrics, verifiable evidence, and calibrated scales rather than prompt phrasing alone. Code is available at https://github.com/LabRAI/Rulers.git.

Method: Evaluated 9 LLMs on English and multilingual hate speech datasets. Analyzed contextual and linguistic biases triggering refusals. Proposed cross-translation strategy: translate English hate speech to Chinese for detoxification, then translate back to English.

Result: LLMs disproportionately refuse inputs with higher semantic toxicity and those targeting specific groups (nationality, religion, political ideology). Multilingual datasets show lower overall false refusal rates than English datasets, but models still display systematic, language-dependent biases toward certain targets.

Conclusion: Cross-translation strategy substantially reduces false refusals while preserving original content, providing an effective and lightweight mitigation approach for hate speech detoxification tasks where LLMs would otherwise refuse.

Abstract: While large language models (LLMs) have increasingly been applied to hate speech detoxification, the prompts often trigger safety alerts, causing LLMs to refuse the task. In this study, we systematically investigate false refusal behavior in hate speech detoxification and analyze the contextual and linguistic biases that trigger such refusals. We evaluate nine LLMs on both English and multilingual datasets, our results show that LLMs disproportionately refuse inputs with higher semantic toxicity and those targeting specific groups, particularly nationality, religion, and political ideology. Although multilingual datasets exhibit lower overall false refusal rates than English datasets, models still display systematic, language-dependent biases toward certain targets. Based on these findings, we propose a simple cross-translation strategy, translating English hate speech into Chinese for detoxification and back, which substantially reduces false refusals while preserving the original content, providing an effective and lightweight mitigation approach.

Method: Developed an agentic system architecture that decomposes the summarization task into components. This enables component-wise optimization and addresses practical challenges through a full development lifecycle approach.

Result: Provides practical insights covering: 1) robust evaluation methods for evolving requirements, 2) component optimization enabled by agentic architecture, 3) impact of upstream data bottlenecks, and 4) realities of vendor lock-in due to poor LLM prompt transferability.

Conclusion: The paper offers a practical industry case study that guides practitioners in building reliable summarization systems and informs future research on addressing real-world challenges like evolving requirements, data bottlenecks, and vendor dependencies.

Abstract: Summarization of multi-party dialogues is a critical capability in industry, enhancing knowledge transfer and operational effectiveness across many domains. However, automatically generating high-quality summaries is challenging, as the ideal summary must satisfy a set of complex, multi-faceted requirements. While summarization has received immense attention in research, prior work has primarily utilized static datasets and benchmarks, a condition rare in practical scenarios where requirements inevitably evolve. In this work, we present an industry case study on developing an agentic system to summarize multi-party interactions. We share practical insights spanning the full development lifecycle to guide practitioners in building reliable, adaptable summarization systems, as well as to inform future research, covering: 1) robust methods for evaluation despite evolving requirements and task subjectivity, 2) component-wise optimization enabled by the task decomposition inherent in an agentic architecture, 3) the impact of upstream data bottlenecks, and 4) the realities of vendor lock-in due to the poor transferability of LLM prompts.

Method: QuantEval integrates a CTA-style backtesting framework that executes model-generated strategies and evaluates them using financial performance metrics, covering three dimensions: knowledge-based QA, quantitative mathematical reasoning, and quantitative strategy coding.

Result: Evaluation of state-of-the-art LLMs shows substantial gaps compared to human experts, particularly in reasoning and strategy coding. Fine-tuning and RL experiments demonstrate consistent improvements.

Conclusion: QuantEval facilitates research on LLMs’ quantitative finance capabilities and accelerates practical adoption in trading workflows, with released backtesting configuration ensuring reproducibility.

Abstract: Large Language Models (LLMs) have shown strong capabilities across many domains, yet their evaluation in financial quantitative tasks remains fragmented and mostly limited to knowledge-centric question answering. We introduce QuantEval, a benchmark that evaluates LLMs across three essential dimensions of quantitative finance: knowledge-based QA, quantitative mathematical reasoning, and quantitative strategy coding. Unlike prior financial benchmarks, QuantEval integrates a CTA-style backtesting framework that executes model-generated strategies and evaluates them using financial performance metrics, enabling a more realistic assessment of quantitative coding ability. We evaluate some state-of-the-art open-source and proprietary LLMs and observe substantial gaps to human experts, particularly in reasoning and strategy coding. Finally, we conduct large-scale supervised fine-tuning and reinforcement learning experiments on domain-aligned data, demonstrating consistent improvements. We hope QuantEval will facilitate research on LLMs’ quantitative finance capabilities and accelerate their practical adoption in real-world trading workflows. We additionally release the full deterministic backtesting configuration (asset universe, cost model, and metric definitions) to ensure strict reproducibility.

Method: Comprehensive comparison of six neural models and six LLM prompting strategies across three granularity levels (nationality, region, continent). Includes frequency-based stratified analysis and error analysis to understand performance patterns and error types.

Result: LLMs outperform neural models at all granularity levels, with gap narrowing as granularity becomes coarser. Simple ML methods show highest frequency robustness, while pre-trained models and LLs degrade for low-frequency nationalities. LLMs make “near-miss” errors (correct region, wrong nationality), while neural models show cross-regional errors and bias toward high-frequency classes.

Conclusion: LLM superiority comes from world knowledge, model selection should consider required granularity, and evaluation should account for error quality beyond accuracy. Practical implications for choosing appropriate models based on task requirements and error tolerance.

Abstract: Predicting nationality from personal names has practical value in marketing, demographic research, and genealogical studies. Conventional neural models learn statistical correspondences between names and nationalities from task-specific training data, posing challenges in generalizing to low-frequency nationalities and distinguishing similar nationalities within the same region. Large language models (LLMs) have the potential to address these challenges by leveraging world knowledge acquired during pre-training. In this study, we comprehensively compare neural models and LLMs on nationality prediction, evaluating six neural models and six LLM prompting strategies across three granularity levels (nationality, region, and continent), with frequency-based stratified analysis and error analysis. Results show that LLMs outperform neural models at all granularity levels, with the gap narrowing as granularity becomes coarser. Simple machine learning methods exhibit the highest frequency robustness, while pre-trained models and LLMs show degradation for low-frequency nationalities. Error analysis reveals that LLMs tend to make ``near-miss’’ errors, predicting the correct region even when nationality is incorrect, whereas neural models exhibit more cross-regional errors and bias toward high-frequency classes. These findings indicate that LLM superiority stems from world knowledge, model selection should consider required granularity, and evaluation should account for error quality beyond accuracy.

Method: RAGShaper uses an InfoCurator to build dense information trees with adversarial distractors at Perception and Cognition levels, plus a constrained navigation strategy that forces teacher agents to confront distractors and demonstrate error correction.

Result: Models trained on RAGShaper’s synthesized corpus significantly outperform existing baselines, showing superior robustness in noise-intensive and complex retrieval tasks.

Conclusion: RAGShaper successfully addresses the data scarcity problem for training robust RAG agents by automating the creation of challenging training scenarios that teach agents to handle retrieval noise and failures effectively.

Abstract: Agentic Retrieval-Augmented Generation (RAG) empowers large language models to autonomously plan and retrieve information for complex problem-solving. However, the development of robust agents is hindered by the scarcity of high-quality training data that reflects the noise and complexity of real-world retrieval environments. Conventional manual annotation is unscalable and often fails to capture the dynamic reasoning strategies required to handle retrieval failures. To bridge this gap, we introduce RAGShaper, a novel data synthesis framework designed to automate the construction of RAG tasks and robust agent trajectories. RAGShaper incorporates an InfoCurator to build dense information trees enriched with adversarial distractors spanning Perception and Cognition levels. Furthermore, we propose a constrained navigation strategy that forces a teacher agent to confront these distractors, thereby eliciting trajectories that explicitly demonstrate error correction and noise rejection. Comprehensive experiments confirm that models trained on our synthesized corpus significantly outperform existing baselines, exhibiting superior robustness in noise-intensive and complex retrieval tasks.

Method: PrivGemo uses a dual-tower design to keep raw KG knowledge local while enabling remote reasoning over anonymized views. It retrieves anonymized long-hop paths connecting all topic entities, supports multi-hop/multi-entity reasoning with local grounding/verification, and includes a hierarchical controller with privacy-aware experience memory to reduce unnecessary exploration.

Result: Comprehensive experiments on six benchmarks show PrivGemo achieves state-of-the-art results, outperforming the strongest baseline by up to 17.1%. It enables smaller models (e.g., Qwen3-4B) to achieve reasoning performance comparable to GPT-4-Turbo.

Conclusion: PrivGemo effectively addresses privacy risks in KG-grounded reasoning by limiting both semantic and structural exposure while maintaining strong reasoning performance, making it practical for real-world applications with private knowledge graphs.

Abstract: Knowledge graphs (KGs) provide structured evidence that can ground large language model (LLM) reasoning for knowledge-intensive question answering. However, many practical KGs are private, and sending retrieved triples or exploration traces to closed-source LLM APIs introduces leakage risk. Existing privacy treatments focus on masking entity names, but they still face four limitations: structural leakage under semantic masking, uncontrollable remote interaction, fragile multi-hop and multi-entity reasoning, and limited experience reuse for stability and efficiency. To address these issues, we propose PrivGemo, a privacy-preserving retrieval-augmented framework for KG-grounded reasoning with memory-guided exposure control. PrivGemo uses a dual-tower design to keep raw KG knowledge local while enabling remote reasoning over an anonymized view that goes beyond name masking to limit both semantic and structural exposure. PrivGemo supports multi-hop, multi-entity reasoning by retrieving anonymized long-hop paths that connect all topic entities, while keeping grounding and verification on the local KG. A hierarchical controller and a privacy-aware experience memory further reduce unnecessary exploration and remote interactions. Comprehensive experiments on six benchmarks show that PrivGemo achieves overall state-of-the-art results, outperforming the strongest baseline by up to 17.1%. Furthermore, PrivGemo enables smaller models (e.g., Qwen3-4B) to achieve reasoning performance comparable to that of GPT-4-Turbo.

Method: FRTR decomposes Excel workbooks into granular row, column, and block embeddings, employs hybrid lexical-dense retrieval with Reciprocal Rank Fusion (RRF), and integrates multimodal embeddings to reason over both numerical and visual information in a retrieval-augmented generation framework.

Result: Achieved 74% answer accuracy on FRTR-Bench with Claude Sonnet 4.5 (vs 24% for prior SOTA), and 87% accuracy on SpreadsheetLLM benchmark with GPT-5 while reducing token usage by roughly 50% compared to context-compression methods.

Conclusion: FRTR provides an effective solution for multimodal spreadsheet reasoning that significantly outperforms existing approaches in both accuracy and efficiency, addressing the limitations of LLMs in handling complex enterprise workbooks.

Abstract: Large Language Models (LLMs) struggle to reason over large-scale enterprise spreadsheets containing thousands of numeric rows, multiple linked sheets, and embedded visual content such as charts and receipts. Prior state-of-the-art spreadsheet reasoning approaches typically rely on single-sheet compression or full-context encoding, which limits scalability and fails to reflect how real users interact with complex, multimodal workbooks. We introduce FRTR-Bench, the first large-scale benchmark for multimodal spreadsheet reasoning, comprising 30 enterprise-grade Excel workbooks spanning nearly four million cells and more than 50 embedded images. To address these challenges, we present From Rows to Reasoning (FRTR), an advanced, multimodal retrieval-augmented generation framework that decomposes Excel workbooks into granular row, column, and block embeddings, employs hybrid lexical-dense retrieval with Reciprocal Rank Fusion (RRF), and integrates multimodal embeddings to reason over both numerical and visual information. We tested FRTR on six LLMs, achieving 74% answer accuracy on FRTR-Bench with Claude Sonnet 4.5, a substantial improvement over prior state-of-the-art approaches that reached only 24%. On the SpreadsheetLLM benchmark, FRTR achieved 87% accuracy with GPT-5 while reducing token usage by roughly 50% compared to context-compression methods.

Method: Introduce Attributional NLI (Att-NLI) framework extending NLI with social psychology principles. Instantiate via Undercover-V textual game with three LLM agent types: standard NLI (deductive only), Att-NLI (abductive-deductive), and neuro-symbolic Att-NLI (abductive-deductive with external theorem provers).

Result: Clear hierarchy of attributional inference capabilities: neuro-symbolic agents consistently outperform others, achieving average win rate of 17.08%. Neuro-symbolic approach shows superior performance in intentional inference tasks.

Conclusion: Att-NLI can develop agents with sophisticated reasoning capabilities. Neuro-symbolic AI has significant potential for building rational LLM agents in multi-agent environments by combining abductive intentional inference with deductive verification.

Abstract: Attributional inference, the ability to predict latent intentions behind observed actions, is a critical yet underexplored capability for large language models (LLMs) operating in multi-agent environments. Traditional natural language inference (NLI), in fact, fails to capture the nuanced, intention-driven reasoning essential for complex interactive systems. To address this gap, we introduce Attributional NLI (Att-NLI), a framework that extends NLI with principles from social psychology to assess an agent’s capacity for abductive intentional inference (generating hypotheses about latent intentions), and subsequent deductive verification (drawing valid logical conclusions). We instantiate Att-NLI via a textual game, Undercover-V, experimenting with three types of LLM agents with varying reasoning capabilities and access to external tools: a standard NLI agent using only deductive inference, an Att-NLI agent employing abductive-deductive inference, and a neuro-symbolic Att-NLI agent performing abductive-deductive inference with external theorem provers. Extensive experiments demonstrate a clear hierarchy of attributional inference capabilities, with neuro-symbolic agents consistently outperforming others, achieving an average win rate of 17.08%. Our results underscore the role that Att-NLI can play in developing agents with sophisticated reasoning capabilities, highlighting, at the same time, the potential impact of neuro-symbolic AI in building rational LLM agents acting in multi-agent environments.

Method: 1) Precompute table representations as KV caches offline while preserving primary-foreign key relationships; 2) Construct Table Trie structure for efficient KV cache lookups during inference; 3) Implement cache management system with query reranking to improve cache hit rates; 4) Use computation loading pipeline to parallelize model inference and cache loading.

Result: TableCache achieves up to 3.62x speedup in Time to First Token (TTFT) with negligible performance degradation compared to existing inference engines like SGLang and vLLM.

Conclusion: The proposed TableCache system effectively addresses inefficiencies in Text-to-SQL inference by enabling KV cache sharing across queries through offline precomputation of table representations, leading to significant speed improvements while maintaining query accuracy.

Abstract: In Text-to-SQL tasks, existing LLM-based methods often include extensive database schemas in prompts, leading to long context lengths and increased prefilling latency. While user queries typically focus on recurrent table sets-offering an opportunity for KV cache sharing across queries-current inference engines, such as SGLang and vLLM, generate redundant prefix cache copies when processing user queries with varying table orders. To address this inefficiency, we propose precomputing table representations as KV caches offline and querying the required ones online. A key aspect of our approach is the computation of table caches while preserving primary foreign key relationships between tables. Additionally, we construct a Table Trie structure to facilitate efficient KV cache lookups during inference. To enhance cache performance, we introduce a cache management system with a query reranking strategy to improve cache hit rates and a computation loading pipeline for parallelizing model inference and cache loading. Experimental results show that our proposed TableCache achieves up to a 3.62x speedup in Time to First Token (TTFT) with negligible performance degradation.

Method: Agentic Context Evolution (ACE) uses a central orchestrator agent with majority voting to dynamically decide between activating a retriever agent for external retrieval or a reasoner agent for internal analysis and refinement, maintaining concise evolved context.

Result: Extensive experiments on challenging multi-hop QA benchmarks show ACE significantly outperforms competitive baselines in accuracy while achieving efficient token consumption.

Conclusion: ACE provides valuable insights into advancing context-evolved generation for complex, knowledge-intensive tasks by eliminating redundant retrieval steps and maintaining optimal context evolution.

Abstract: Current context augmentation methods, such as retrieval-augmented generation, are essential for solving knowledge-intensive reasoning tasks.However, they typically adhere to a rigid, brute-force strategy that executes retrieval at every step. This indiscriminate approach not only incurs unnecessary computational costs but also degrades performance by saturating the context with irrelevant noise. To address these limitations, we introduce Agentic Context Evolution (ACE), a framework inspired by human metacognition that dynamically determines whether to seek new evidence or reason with existing knowledge. ACE employs a central orchestrator agent to make decisions strategically via majority voting.It aims to alternate between activating a retriever agent for external retrieval and a reasoner agent for internal analysis and refinement. By eliminating redundant retrieval steps, ACE maintains a concise and evolved context. Extensive experiments on challenging multi-hop QA benchmarks demonstrate that ACE significantly outperforms competitive baselines in accuracy while achieving efficient token consumption.Our work provides valuable insights into advancing context-evolved generation for complex, knowledge-intensive tasks.

Method: Attention-based approach combining vision (aerial imagery) and text (Wikipedia sentences) representations with geolocation encoding, using attention module to dynamically select useful spatial neighbors from nearby observations within spatial neighborhoods.

Result: Model consistently outperforms single-location and unimodal (image-only or text-only) baselines on EcoWikiRS dataset. Significant improvements for climatic, edaphic, population, and land use/land cover variables, demonstrating benefits of spatial context in multimodal integration.

Conclusion: Spatial attention-based multimodal integration of text and imagery effectively leverages complementary information sources for environmental prediction, with particular value for variables where local context matters most.

Abstract: Recent developments in natural language processing highlight text as an emerging data source for ecology. Textual resources carry unique information that can be used in complementarity with geospatial data sources, thus providing insights at the local scale into environmental conditions and properties hidden from more traditional data sources. Leveraging textual information in a spatial context presents several challenges. First, the contribution of textual data remains poorly defined in an ecological context, and it is unclear for which tasks it should be incorporated. Unlike ubiquitous satellite imagery or environmental covariates, the availability of textual data is sparse and irregular; its integration with geospatial data is not straightforward. In response to these challenges, this work proposes an attention-based approach that combines aerial imagery and geolocated text within a spatial neighbourhood, i.e. integrating contributions from several nearby observations. Our approach combines vision and text representations with a geolocation encoding, with an attention-based module that dynamically selects spatial neighbours that are useful for predictive tasks.The proposed approach is applied to the EcoWikiRS dataset, which combines high-resolution aerial imagery with sentences extracted from Wikipedia describing local environmental conditions across Switzerland. Our model is evaluated on the task of predicting 103 environmental variables from the SWECO25 data cube. Our approach consistently outperforms single-location or unimodal, i.e. image-only or text-only, baselines. When analysing variables by thematic groups, results show a significant improvement in performance for climatic, edaphic, population and land use/land cover variables, underscoring the benefit of including the spatial context when combining text and image data.

Method: At each thinking step, sample K candidate tokens and aggregate their embeddings into a single continuous multiplex token. This preserves vocabulary embedding prior and sampling dynamics while creating tractable probability distributions over multiplex rollouts. The approach is self-adaptive (discrete when confident, soft when uncertain) and can be optimized with on-policy reinforcement learning.

Result: Multiplex Thinking consistently outperforms strong discrete CoT and RL baselines across challenging math reasoning benchmarks from Pass@1 through Pass@1024, while producing shorter reasoning sequences.

Conclusion: Multiplex Thinking provides an efficient stochastic soft reasoning mechanism that combines the benefits of human-like soft reasoning with computational efficiency, achieving better performance with shorter sequences than traditional Chain-of-Thought approaches.

Abstract: Large language models often solve complex reasoning tasks more effectively with Chain-of-Thought (CoT), but at the cost of long, low-bandwidth token sequences. Humans, by contrast, often reason softly by maintaining a distribution over plausible next steps. Motivated by this, we propose Multiplex Thinking, a stochastic soft reasoning mechanism that, at each thinking step, samples K candidate tokens and aggregates their embeddings into a single continuous multiplex token. This preserves the vocabulary embedding prior and the sampling dynamics of standard discrete generation, while inducing a tractable probability distribution over multiplex rollouts. Consequently, multiplex trajectories can be directly optimized with on-policy reinforcement learning (RL). Importantly, Multiplex Thinking is self-adaptive: when the model is confident, the multiplex token is nearly discrete and behaves like standard CoT; when it is uncertain, it compactly represents multiple plausible next steps without increasing sequence length. Across challenging math reasoning benchmarks, Multiplex Thinking consistently outperforms strong discrete CoT and RL baselines from Pass@1 through Pass@1024, while producing shorter sequences. The code and checkpoints are available at https://github.com/GMLR-Penn/Multiplex-Thinking.

Method: Multiple LLM agent reviewers with different personas engage in multi-round review interactions moderated by an Area Chair. Compare baseline setting with conditions incorporating Elo ratings and reviewer memory using real-world conference paper submissions.

Result: Incorporating Elo ratings improves Area Chair decision accuracy. Reviewers develop adaptive strategies that exploit the Elo system without actually improving their review effort. The simulation reveals interesting dynamics in reviewer behavior.

Conclusion: Elo-ranked review systems can improve decision accuracy but may lead to reviewers gaming the system rather than improving review quality, highlighting important considerations for designing automated review systems.

Abstract: In this work, we explore the Large Language Model (LLM) agent reviewer dynamics in an Elo-ranked review system using real-world conference paper submissions. Multiple LLM agent reviewers with different personas are engage in multi round review interactions moderated by an Area Chair. We compare a baseline setting with conditions that incorporate Elo ratings and reviewer memory. Our simulation results showcase several interesting findings, including how incorporating Elo improves Area Chair decision accuracy, as well as reviewers’ adaptive review strategy that exploits our Elo system without improving review effort. Our code is available at https://github.com/hsiangwei0903/EloReview.

Method: Adapt existing LLMs to duplex models using time-division-multiplexing (TDM) encoding-decoding strategy. Divide queries and responses into time slices for pseudo-simultaneous processing. Build a fine-tuning dataset with alternating time slices of queries and responses covering typical feedback types in instantaneous interactions.

Result: Duplex models preserve original LLM performance on standard benchmarks despite processing incomplete conversation slices. Automatic and human evaluation show duplex models make user-AI interactions more natural and human-like, greatly improving user satisfaction compared to vanilla LLMs.

Conclusion: Duplex models successfully enable real-time conversational interactions with LLMs, overcoming limitations of traditional turn-based systems. The approach maintains model performance while significantly enhancing user experience through more natural, human-like interactions.

Abstract: As large language models (LLMs) increasingly permeate daily lives, there is a growing demand for real-time interactions that mirror human conversations. Traditional turn-based chat systems driven by LLMs prevent users from verbally interacting with the system while it is generating responses. To overcome these limitations, we adapt existing LLMs to \textit{duplex models} so that these LLMs can listen for users while generating output and dynamically adjust themselves to provide users with instant feedback. % such as in response to interruptions. Specifically, we divide the queries and responses of conversations into several time slices and then adopt a time-division-multiplexing (TDM) encoding-decoding strategy to pseudo-simultaneously process these slices. Furthermore, to make LLMs proficient enough to handle real-time conversations, we build a fine-tuning dataset consisting of alternating time slices of queries and responses as well as covering typical feedback types in instantaneous interactions. Our experiments show that although the queries and responses of conversations are segmented into incomplete slices for processing, LLMs can preserve their original performance on standard benchmarks with a few fine-tuning steps on our dataset. Automatic and human evaluation indicate that duplex models make user-AI interactions more natural and human-like, and greatly improve user satisfaction compared to vanilla LLMs. Our duplex model and dataset will be released.

Method: The paper analyzes Mamba-based models’ forgetting mechanisms, demonstrating they struggle to forget earlier tokens. It shows this stems from training on contexts too short for state size, allowing models to perform well without learning to forget. The study examines the relationship between state size, training length, and forgetting capabilities.

Result: Minimum training length required for models to learn forgetting scales linearly with state size. Maximum context length for accurate retrieval of 5-digit passkey scales exponentially with state size, indicating models retain some information beyond where forgetting begins.

Conclusion: Current RNN architectures have critical limitations in long-context modeling. Future RNN designs must account for interplay between state size, training length, and forgetting mechanisms for robust performance in long-context tasks.

Abstract: Recent advancements in recurrent architectures, such as Mamba and RWKV, have showcased strong language capabilities. Unlike transformer-based models, these architectures encode all contextual information into a fixed-size state, leading to great inference efficiency. However, this approach can cause information interference, where different token data conflicts, resulting in performance degradation and incoherent outputs beyond a certain context length. To prevent this, most RNNs incorporate mechanisms designed to “forget” earlier tokens. In this paper, we reveal that Mamba-based models struggle to effectively forget earlier tokens even with built-in forgetting mechanisms. We demonstrate that this issue stems from training on contexts that are too short for the state size, enabling the model to perform well without needing to learn how to forget. Then, we show that the minimum training length required for the model to learn forgetting scales linearly with the state size, and the maximum context length for accurate retrieval of a 5-digit passkey scales exponentially with the state size, indicating that the model retains some information beyond the point where forgetting begins. These findings highlight a critical limitation in current RNN architectures and provide valuable insights for improving long-context modeling. Our work suggests that future RNN designs must account for the interplay between state size, training length, and forgetting mechanisms to achieve robust performance in long-context tasks.

Method: Proposes KaLM approach that fine-tunes autoregressive LLMs with joint objectives: 1) Explicit knowledge alignment via dual-view knowledge graph contrastive learning to optimize knowledge representation, and 2) Implicit knowledge alignment via triple completion language modeling to incorporate textual patterns of knowledge.

Result: Achieves significant performance boost in knowledge-driven task evaluations, specifically in embedding-based knowledge graph completion and generation-based knowledge graph question answering.

Conclusion: KaLM effectively aligns LLMs with KG knowledge through joint explicit and implicit alignment objectives, improving performance on knowledge-driven tasks without compromising generative capabilities.

Abstract: Autoregressive large language models (LLMs) pre-trained by next token prediction are inherently proficient in generative tasks. However, their performance on knowledge-driven tasks such as factual knowledge querying remains unsatisfactory. Knowledge graphs (KGs), as high-quality structured knowledge bases, can provide reliable knowledge for LLMs, potentially compensating for their knowledge deficiencies. Aligning LLMs with explicit, structured knowledge from KGs has been a challenge; previous attempts either failed to effectively align knowledge representations or compromised the generative capabilities of LLMs, leading to less-than-optimal outcomes. This paper proposes \textbf{KaLM}, a \textit{Knowledge-aligned Language Modeling} approach, which fine-tunes autoregressive LLMs to align with KG knowledge via the joint objective of explicit knowledge alignment and implicit knowledge alignment. The explicit knowledge alignment objective aims to directly optimize the knowledge representation of LLMs through dual-view knowledge graph contrastive learning. The implicit knowledge alignment objective focuses on incorporating textual patterns of knowledge into LLMs through triple completion language modeling. Notably, our method achieves a significant performance boost in evaluations of knowledge-driven tasks, specifically embedding-based knowledge graph completion and generation-based knowledge graph question answering.

Method: Three approaches: (1) baseline using UMR graphs directly with AMR-to-text models, (2) pipeline conversion of UMR to AMR then using AMR-to-text models, and (3) fine-tuning both foundation models and AMR-to-text models with UMR data.

Result: Best models achieved multilingual BERTscores of 0.825 for English and 0.882 for Chinese, showing effectiveness of fine-tuning approaches even with limited UMR data.

Conclusion: Fine-tuning approaches show promise for UMR-to-text generation, leveraging structural similarity between UMR and AMR and existing AMR technologies, despite limited UMR annotations currently available.

Abstract: Uniform Meaning Representation (UMR) is a recently developed graph-based semantic representation, which expands on Abstract Meaning Representation (AMR) in a number of ways, in particular through the inclusion of document-level information and multilingual flexibility. In order to effectively adopt and leverage UMR for downstream tasks, efforts must be placed toward developing a UMR technological ecosystem. Though only a small amount of UMR annotations have been produced to date, in this work, we investigate the first approaches to producing text from multilingual UMR graphs. Exploiting the structural similarity between UMR and AMR graphs and the wide availability of AMR technologies, we introduce (1) a baseline approach which passes UMR graphs to AMR-to-text generation models, (2) a pipeline conversion of UMR to AMR, then using AMR-to-text generation models, and (3) a fine-tuning approach for both foundation models and AMR-to-text generation models with UMR data. Our best performing models achieve multilingual BERTscores of 0.825 for English and 0.882 for Chinese, a promising indication of the effectiveness of fine-tuning approaches for UMR-to-text generation even with limited UMR data.

Method: Created a novel dataset of 1,308 test cases based on psychological frameworks (D/S-IAT and Negative Automatic Thinking) and real-world scenarios. Evaluated 8 widely used LLMs under different contextual settings for two tasks: Identification of Implicit Suicidal ideation (IIS) and Provision of Appropriate Supportive responses (PAS).

Result: Current LLMs struggle significantly with both detecting implicit suicidal ideation and providing appropriate supportive responses, revealing serious limitations in applying these models to mental health contexts.

Conclusion: The findings highlight the need for more sophisticated approaches in developing and evaluating LLMs for sensitive psychological applications like suicide prevention, as current models are inadequate for these critical tasks.

Abstract: We present a comprehensive evaluation framework for assessing Large Language Models’ (LLMs) capabilities in suicide prevention, focusing on two critical aspects: the Identification of Implicit Suicidal ideation (IIS) and the Provision of Appropriate Supportive responses (PAS). We introduce \ourdata, a novel dataset of 1,308 test cases built upon psychological frameworks including D/S-IAT and Negative Automatic Thinking, alongside real-world scenarios. Through extensive experiments with 8 widely used LLMs under different contextual settings, we find that current models struggle significantly with detecting implicit suicidal ideation and providing appropriate support, highlighting crucial limitations in applying LLMs to mental health contexts. Our findings underscore the need for more sophisticated approaches in developing and evaluating LLMs for sensitive psychological applications.

Method: The study investigates positional distribution of hallucinations in LLM-based long response generation, using long document summarization as a case study. It explores factors related to attention and decoding dynamics over long sequences, and investigates mitigation methods.

Result: The paper finds a consistent phenomenon: hallucinations concentrate disproportionately in the latter parts of generated long responses, especially in long context-aware generation settings.

Conclusion: Positional hallucination bias exists in LLM long response generation, requiring attention to improve faithfulness in concluding segments. The work contributes to understanding and mitigating this specific distribution pattern of hallucinations.

Abstract: Large Language Models (LLMs) have significantly advanced text generation capabilities, including tasks like summarization, often producing coherent and fluent outputs. However, faithfulness to source material remains a significant challenge due to the generation of hallucinations. While extensive research focuses on detecting and reducing these inaccuracies, less attention has been paid to the positional distribution of hallucination within generated text, particularly in long outputs. In this work, we investigate where hallucinations occur in LLM-based long response generation, using long document summarization as a key case study. Focusing on the challenging setting of long context-aware long response generation, we find a consistent and concerning phenomenon: hallucinations tend to concentrate disproportionately in the latter parts of the generated long response. To understand this bias, we explore potential contributing factors related to the dynamics of attention and decoding over long sequences. Furthermore, we investigate methods to mitigate this positional hallucination, aiming to improve faithfulness specifically in the concluding segments of long outputs.

Method: Proposed VocalBench with ~24k carefully curated instances across English and Mandarin, covering 4 key dimensions: semantic quality, acoustic performance, conversational abilities, and robustness, with 14 user-oriented characters.

Result: Experiments on 27 mainstream models reveal common challenges for current approaches and highlight the need for new insights into next-generation speech interactive systems.

Conclusion: VocalBench addresses the gap in comprehensive speech conversational evaluation and identifies limitations in current SpeechLLMs, pointing toward future improvements in speech interactive systems.

Abstract: Speech large language models (SpeechLLMs) have extended human-machine interactions from the text modality to the dynamic speech domain. Spoken dialogues convey diverse information, including semantic concepts, acoustic variations, paralanguage cues, and environmental context. However, existing evaluations of speech interaction models lack instances mimicking real scenarios and predominantly focus on the performance of distinct aspects, lacking a comprehensive comparison of critical capabilities between current routines. To address this gap, we propose VocalBench to assess the speech conversational abilities, comprising around 24k carefully curated instances of both English and Mandarin across four key dimensions - semantic quality, acoustic performance, conversational abilities, and robustness, covering 14 user-oriented characters. Experiments on 27 mainstream models reveal the common challenges for current routes, and highlight the need for new insights into next-generation speech interactive systems.

Method: Proposes Align-GRAG framework that uses LLM-extracted anchors and rationale chains to guide refinement through: (1) node-level alignment to identify critical nodes and prune noisy evidence, and (2) graph-level alignment using contrastive learning to bridge graph and language semantic spaces.

Result: Extensive experiments on commonsense reasoning, scene graph understanding, and knowledge graph reasoning show consistent improvements over 18 strong baselines, validating effectiveness for graph-grounded generation.

Conclusion: Align-GRAG successfully addresses challenges in graph-based retrieval-augmented generation by aligning graph structures with language semantics through anchor-and-rationale guided refinement, improving knowledge-intensive tasks.

Abstract: Despite the strong abilities, large language models (LLMs) still suffer from hallucinations and reliance on outdated knowledge, raising concerns in knowledge-intensive tasks. Graph-based retrieval-augmented generation (GRAG) enriches LLMs with knowledge by retrieving graphs leveraging relational evidence, but it faces two challenges: structure-coupled irrelevant knowledge introduced by neighbor expansion and structure-reasoning discrepancy between graph embeddings and LLM semantics. We propose \ourmodel, an anchor-and-rationale guided refinement framework to address these challenges. It prompts an LLM to extract anchors and rationale chains, which provide intermediate supervision for \textbf{(1) node-level alignment} that identifies critical nodes and prunes noisy evidence, and \textbf{(2) graph-level alignment} that bridges graph and language semantic spaces via contrastive learning. Extensive experiments on commonsense reasoning, scene graph understanding, and knowledge graph reasoning demonstrate consistent gains over 18 strong baselines, validating the effectiveness of \ourmodel for improving graph-grounded generation. The code can be found in https://anonymous.4open.science/r/Align-GRAG-F3D8/.

Method: Created Compliance-to-Code dataset with modular clause structuring (subject, condition, constraint, contextual info) plus regulation relations. Provides deterministic Python code mappings, code reasoning, and explanations. Developed FinCheck pipeline for regulation structuring, code generation, and report generation.

Result: Dataset covers 1,159 annotated clauses from 361 regulations across ten categories, enabling automated compliance logic generation for Chinese financial regulations.

Conclusion: Compliance-to-Code fills critical gaps in Chinese financial regulatory automation, providing structured data for improved RegTech/LLM performance through domain-specific, code-oriented training resources.

Abstract: Nowadays, regulatory compliance has become a cornerstone of corporate governance, ensuring adherence to systematic legal frameworks. At its core, financial regulations often comprise highly intricate provisions, layered logical structures, and numerous exceptions, which inevitably result in labor-intensive or comprehension challenges. To mitigate this, recent Regulatory Technology (RegTech) and Large Language Models (LLMs) have gained significant attention in automating the conversion of regulatory text into executable compliance logic. However, their performance remains suboptimal particularly when applied to Chinese-language financial regulations, due to three key limitations: (1) incomplete domain-specific knowledge representation, (2) insufficient hierarchical reasoning capabilities, and (3) failure to maintain temporal and logical coherence. One promising solution is to develop a domain specific and code-oriented datasets for model training. Existing datasets such as LexGLUE, LegalBench, and CODE-ACCORD are often English-focused, domain-mismatched, or lack fine-grained granularity for compliance code generation. To fill these gaps, we present Compliance-to-Code, the first large-scale Chinese dataset dedicated to financial regulatory compliance. Covering 1,159 annotated clauses from 361 regulations across ten categories, each clause is modularly structured with four logical elements-subject, condition, constraint, and contextual information-along with regulation relations. We provide deterministic Python code mappings, detailed code reasoning, and code explanations to facilitate automated auditing. To demonstrate utility, we present FinCheck: a pipeline for regulation structuring, code generation, and report generation.

Method: A reinforcement learning framework using execution correctness as a lightweight reward signal, avoiding intermediate supervision and complex reward shaping. Combines curated data, strong supervised initialization, and effective training practices.

Result: Achieves state-of-the-art execution accuracy across six Text2SQL benchmarks, including top position on BIRD leaderboard. The 7B model outperforms prior 70B-class systems, demonstrating scalability and efficiency.

Conclusion: The RL framework with execution-based rewards enables stable training and alignment with end tasks, providing practical guidance for future Text2SQL research while maintaining inference-time robustness through value retrieval and majority voting.

Abstract: Translating natural language into SQL (Test2SQL) is a longstanding challenge at the intersection of natural language understanding and structured data access. While large language models (LLMs) have significantly improved fluency in SQL generation, producing correct and executable SQL–particularly for complex queries–remains a bottleneck. We present Arctic-Text2SQL-R1, a reinforcement learning (RL) framework and model family designed to generate accurate, executable SQL using a lightweight reward signal based solely on execution correctness. Our approach avoids brittle intermediate supervision and complex reward shaping, promoting stable training and alignment with the end task. Combined with carefully curated data, strong supervised initialization, and effective training practices, Arctic-Text2SQL-R1 achieves state-of-the-art execution accuracy across six diverse Test2SQL benchmarks, including the top position on the BIRD leaderboard. Notably, our 7B model outperforms prior 70B-class systems, highlighting the framework’s scalability and efficiency. We further demonstrate inference-time robustness through simple extensions like value retrieval and majority voting. Extensive experiments and ablation studies offer both positive and negative insights, providing practical guidance for future Test2SQL research.

Method: Extended the Machine Personality Inventory (MPI) to incorporate the 16PF model, developed Specific Attribute Control (SAC) framework with adjective-based semantic anchoring and behavioral questions across five intensity factors: Frequency, Depth, Threshold, Effort, and Willingness.

Result: Modeling intensity as a continuous spectrum yields more consistent and controllable personality expression than binary toggling; changes in target trait intensity systematically influence related traits in psychologically coherent directions, suggesting LLMs internalize multi-dimensional personality structures.

Conclusion: The work enables more controlled and nuanced human-machine interactions in domains like healthcare, education, and interviewing, advancing toward truly human-like social machines with expressive personality control.

Abstract: Large language models (LLMs) have gained significant traction across a wide range of fields in recent years. There is also a growing expectation for them to display human-like personalities during interactions. To meet this expectation, numerous studies have proposed methods for modelling LLM personalities through psychometric evaluations. However, most existing models face two major limitations: they rely on the Big Five (OCEAN) framework, which only provides coarse personality dimensions, and they lack mechanisms for controlling trait intensity. In this paper, we address this gap by extending the Machine Personality Inventory (MPI), which originally used the Big Five model, to incorporate the 16 Personality Factor (16PF) model, allowing expressive control over sixteen distinct traits. We also developed a structured framework known as Specific Attribute Control (SAC) for evaluating and dynamically inducing trait intensity in LLMs. Our method introduces adjective-based semantic anchoring to guide trait intensity expression and leverages behavioural questions across five intensity factors: \textit{Frequency}, \textit{Depth}, \textit{Threshold}, \textit{Effort}, and \textit{Willingness}. Through experimentation, we find that modelling intensity as a continuous spectrum yields substantially more consistent and controllable personality expression compared to binary trait toggling. Moreover, we observe that changes in target trait intensity systematically influence closely related traits in psychologically coherent directions, suggesting that LLMs internalize multi-dimensional personality structures rather than treating traits in isolation. Our work opens new pathways for controlled and nuanced human-machine interactions in domains such as healthcare, education, and interviewing processes, bringing us one step closer to truly human-like social machines.

Method: Proposes “Interaction Distillation” framework that introduces an interaction-based natural language understanding model as a teacher to provide sophisticated token interaction patterns via comprehensive attention. The reward model is trained to simulate the teacher’s interaction pattern through an attentional alignment objective, addressing the attention hacking problem.

Result: Extensive experiments show that interaction distillation provides more stable and generalizable reward signals compared to state-of-the-art RM optimization methods that target data noise, demonstrating that attention hacking constitutes a more fundamental limitation in discriminative reward modeling.

Conclusion: Attention hacking is a fundamental limitation in current discriminative reward models, and the proposed interaction distillation framework effectively addresses this by optimizing attention patterns through teacher-student knowledge transfer, leading to more robust reward signals for RLHF.

Abstract: The reward model (RM), as the core component of reinforcement learning from human feedback (RLHF) for large language models (LLMs), responsible for providing reward signals to generated responses. However, the mainstream discriminative reward modeling is inadequate in terms of token-level interaction, making its judgment signals vulnerable to being hacked by misallocated attention to context. This stems from two fundamental limitations: (1) Current preference modeling employs decoder-only architectures, where the unidirectional causal attention mechanism leads to forward-decaying intra-sequence attention within the prompt-response sequence. (2) The independent Siamese-encoding paradigm induces the absence of token-level inter-sequence attention between chosen and rejected sequences. To address this “attention hacking”, we propose “Interaction Distillation”, a novel training framework for more adequate discriminative reward modeling via attention-level optimization. The method introduces an interaction-based natural language understanding model as the teacher to provide sophisticated token interaction patterns via comprehensive attention, and guides the reward modeling to simulate teacher model’s interaction pattern through an attentional alignment objective. Through extensive experiments, interaction distillation has demonstrated its ability to provide more stable and generalizable reward signals compared to state-of-the-art RM optimization methods that target data noise, highlighting the attention hacking constitute a more fundamental limitation in discriminative RM.

Method: Developed a prediction confidence model (PCM) that estimates semantic similarity between predicted complete user utterances and actual complete utterances to determine if prefetching is feasible.

Result: Evaluated PCM based on differences between predicted and actual complete user utterances (semantic similarity assessment).

Conclusion: PCM can help determine prefetching feasibility to reduce user waiting time in dialogue systems by predicting when prefetching is reliable.

Abstract: Prefetching of dialogue responses has been investigated to reduce user-perceived latency (UPL), which refers to the user’s waiting time before receiving the system’s response, in spoken dialogue systems. To reduce the UPL, it is necessary to predict complete user utterances before the end of the user’s speech, typically by language models, to prepare prefetched dialogue responses. In this study, we proposed a prediction confidence model (PCM) that determines whether prefetching is possible or not by estimating the semantic similarity between the predicted complete user utterance and the complete user utterance. We evaluated our PCM based on the differences between the predicted complete user utterance and the complete user utterance.

Method: MultiCheck uses a relational fusion module based on element-wise difference and product operations for explicit cross-modal interaction modeling with minimal computational overhead. It incorporates a contrastive alignment objective to distinguish between supporting and refuting evidence while maintaining small memory and energy footprint for low-resource deployment.

Result: Evaluated on Factify-2 (5-class) and Mocheg (3-class) benchmarks, MultiCheck achieves huge performance improvement and remains robust under noisy OCR and missing modality conditions.

Conclusion: MultiCheck’s efficiency, transparency, and real-world robustness make it well-suited for journalists, civil society organizations, and web integrity efforts working to build a safer and more trustworthy web.

Abstract: Misinformation on the web increasingly appears in multimodal forms, combining text, images, and OCR-rendered content in ways that amplify harm to public trust and vulnerable communities. While prior fact-checking systems often rely on unimodal signals or shallow fusion strategies, modern misinformation campaigns operate across modalities and require models that can reason over subtle cross-modal inconsistencies in a transparent and responsible manner. We introduce MultiCheck, a lightweight and interpretable framework for multimodal fact verification that jointly analyzes textual, visual, and OCR evidence. At its core, MultiCheck employs a relational fusion module based on element-wise difference and product operations, allowing for explicit cross-modal interaction modeling with minimal computational overhead. A contrastive alignment objective further helps the model distinguish between supporting and refuting evidence while maintaining a small memory and energy footprint, making it suitable for low-resource deployment. Evaluated on the Factify-2 (5-class) and Mocheg (3-class) benchmarks, MultiCheck achieves huge performance improvement and remains robust under noisy OCR and missing modality conditions. Its efficiency, transparency, and real-world robustness make it well-suited for journalists, civil society organisations, and web integrity efforts working to build a safer and more trustworthy web.

Method: Proposes Cross-Prompt Encoder (XPE) - a lightweight encoder with multi-source training on typologically diverse languages, plus a Dual Soft Prompt mechanism combining encoder-based prompts with standard soft prompts.

Result: XPE effectively captures abstract, transferable patterns across languages, especially benefiting low-performing languages. Hybrid variants offer broader adaptability across multilingual settings.

Conclusion: There’s a consistent trade-off: XPE works best for low-performing languages, while hybrid designs provide broader multilingual adaptability, showing soft prompts’ potential for cross-lingual transfer.

Abstract: Soft prompts have emerged as a powerful alternative to adapters in parameter-efficient fine-tuning (PEFT), enabling large language models (LLMs) to adapt to downstream tasks without architectural changes or parameter updates. While prior work has focused on stabilizing training via parameter interaction in small neural prompt encoders, their broader potential for transfer across languages remains unexplored. In this paper, we demonstrate that a prompt encoder can play a central role in improving performance on low-performing languages - those that achieve poor accuracy even under full-model fine-tuning. We investigate a lightweight encoder paired with multi-source training on typologically diverse languages. We call this architecture-training combination the Cross-Prompt Encoder (XPE), and show that it advances the capture of abstract, transferable patterns across languages. To complement XPE, we propose a Dual Soft Prompt mechanism that combines an encoder-based prompt with a directly trained standard soft prompt. This hybrid design proves especially effective for target languages that benefit from both broadly shared structure and language-specific alignment. Text classification experiments with a transformer encoder (XLM-R) on the SIB-200 benchmark reveal a consistent trade-off: XPE is most effective for low-performing languages, while hybrid variants offer broader adaptability across multilingual settings.

Method: Created VMMU benchmark with 2.5k multimodal questions across 7 diverse tasks (STEM problem solving, data interpretation, rule-governed visual reasoning, abstract visual reasoning). Evaluated state-of-the-art proprietary and open-source VLMs on this benchmark.

Result: Proprietary models achieve only 66% mean accuracy despite strong Vietnamese OCR performance. Analysis shows primary failure source is multimodal grounding and reasoning over text and visual evidence, not OCR limitations.

Conclusion: Current VLMs struggle with genuine multimodal understanding in Vietnamese, highlighting the need for improved multimodal reasoning capabilities beyond OCR in non-English languages.

Abstract: We introduce VMMU, a Vietnamese Multitask Multimodal Understanding and Reasoning Benchmark designed to evaluate how vision-language models (VLMs) interpret and reason over visual and textual information beyond English. VMMU consists of 2.5k multimodal questions across 7 tasks, covering a diverse range of problem contexts, including STEM problem solving, data interpretation, rule-governed visual reasoning, and abstract visual reasoning. All questions require genuine multimodal integration, rather than reliance on text-only cues or OCR-based shortcuts. We evaluate a diverse set of state-of-the-art proprietary and open-source VLMs on VMMU. Despite strong Vietnamese OCR performance, proprietary models achieve only 66% mean accuracy. Further analysis shows that the primary source of failure is not OCR, but instead multimodal grounding and reasoning over text and visual evidence. Code and data are available at https://vmmu.github.io.

Method: Three-stage empirical analysis: 1) Exploring presence and direction of format bias across diverse LLMs, 2) Examining how key data-level factors influence biases, 3) Analyzing how format bias emerges within LLMs’ attention patterns and evaluating a lightweight intervention’s effectiveness.

Result: Format bias is consistent across model families, driven by information richness, structure quality, and representation type, and is closely associated with attention imbalance within LLMs. The study identifies specific patterns in how LLMs prioritize different data formats.

Conclusion: Three future research directions to reduce format bias: 1) Enhancing data pre-processing through format repair and normalization, 2) Introducing inference-time interventions like attention re-weighting, 3) Developing format-balanced training corpora. These approaches will support more robust and fair heterogeneous data processing systems.

Abstract: Large Language Models (LLMs) are increasingly employed in applications that require processing information from heterogeneous formats, including texts, tables, infoboxes, and knowledge graphs. However, systematic biases toward particular formats may undermine LLMs’ ability to integrate heterogeneous data impartially, potentially resulting in reasoning errors and increased risks in downstream tasks. Yet it remains unclear whether such biases are systematic, which data-level factors drive them, and what internal mechanisms underlie their emergence. In this paper, we present the first comprehensive study of format bias in LLMs through a three-stage empirical analysis. The first stage explores the presence and direction of bias across a diverse range of LLMs. The second stage examines how key data-level factors influence these biases. The third stage analyzes how format bias emerges within LLMs’ attention patterns and evaluates a lightweight intervention to test its effectiveness. Our results show that format bias is consistent across model families, driven by information richness, structure quality, and representation type, and is closely associated with attention imbalance within the LLMs. Based on these investigations, we identify three future research directions to reduce format bias: enhancing data pre-processing through format repair and normalization, introducing inference-time interventions such as attention re-weighting, and developing format-balanced training corpora. These directions will support the design of more robust and fair heterogeneous data processing systems.

Method: Introduces a mathematically grounded KL divergence-based knowledge distillation method that allows student models to inherit distributional knowledge from teachers despite differing vocabularies. Compares this approach to conventional cross-entropy training across multiple strategies for initializing new token embeddings, followed by fine-tuning to integrate new vocabulary.

Result: The KL-based distillation approach achieves the best performance across approximately 2000 code-generation tasks. Mechanistic interpretability analysis reveals how models learn representations for new tokens, explaining the observed gains and providing insight into embedding space structure during vocabulary expansion.

Conclusion: KL divergence-based knowledge distillation is an effective method for vocabulary expansion in frozen LLMs, enabling better incorporation of domain-specific terminology across different tokenizations while maintaining model performance.

Abstract: Large pre-trained language models often struggle to incorporate new domain-specific terminology when fine-tuned on small, specialized corpora. In this work, we address the challenge of vocabulary expansion in frozen LLMs by introducing a mathematically grounded method for knowledge distillation via KL divergence, even when the original and extended models use different tokenizations. This allows the student model to inherit distributional knowledge from the teacher despite differing vocabularies. We compare our KL-based distillation approach to conventional cross-entropy training, evaluating both methods across multiple strategies for initializing new token embeddings. After embedding initialization, models are further fine-tuned to integrate the new vocabulary. Each trained model is benchmarked on approximately 2000 code-generation tasks, where our approach achieves the best performance across the board. Finally, through mechanistic interpretability, we analyze how models learn representations for the new tokens, providing an explanation for the observed gains and offering insight into the structure of embedding space during vocabulary expansion.

Method: Analyze reasoning through two dynamics: Reasoning Length Dynamics (trade-off between thinking and answer length) and Reasoning Semantic Dynamics (semantic convergence and oscillations). Identify instance-specific Reasoning Completion Points (RCP) and propose RCPD, which monitors termination token rank dynamics to detect RCP for early exit.

Result: RCPD reduces token usage by up to 44% across AIME and GPQA benchmarks using Qwen3 and DeepSeek-R1 models while preserving accuracy.

Conclusion: RCPD offers a principled approach to efficient test-time scaling by preventing overthinking through instance-specific reasoning completion detection, significantly improving computational efficiency without sacrificing performance.

Abstract: Test-time scaling via explicit reasoning trajectories significantly boosts large language model (LLM) performance but often triggers overthinking. To explore this, we analyze reasoning through two lenses: Reasoning Length Dynamics, which reveals a compensatory trade-off between thinking and answer content length that eventually leads to thinking redundancy, and Reasoning Semantic Dynamics, which identifies semantic convergence and repetitive oscillations. These dynamics uncover an instance-specific Reasoning Completion Point (RCP), beyond which computation continues without further performance gain. Since the RCP varies across instances, we propose a Reasoning Completion Point Detector (RCPD), an inference-time early-exit method that identifies the RCP by monitoring the rank dynamics of termination tokens (e.g., ). Across AIME and GPQA benchmarks using Qwen3 and DeepSeek-R1, RCPD reduces token usage by up to 44% while preserving accuracy, offering a principled approach to efficient test-time scaling.

Method: Proposes Performance Interval Estimation (PIE) with two approaches: 1) LLM-as-judge method, and 2) classic regression with confidence estimation features. Evaluates both methods across 11 datasets covering summarization, translation, code generation, function-calling, and question answering.

Result: Regression approach achieves both lower error point estimates of metric scores and well-calibrated uncertainty intervals compared to LLM-as-judge method. The method is evaluated across diverse NLP tasks.

Conclusion: Performance Interval Estimation (PIE) provides a better framework for estimating continuous-valued metrics in free-form generation tasks, with regression methods outperforming LLM-as-judge approaches. The authors share data and code to support reproduction and future work.

Abstract: Confidence estimation infers a probability for whether each model output is correct or not. While predicting such binary correctness is sensible for tasks with exact answers, free-form generation tasks are often more nuanced, with output quality being both fine-grained and multi-faceted. We thus propose Performance Interval Estimation (PIE) to predict both: 1) point estimates for any arbitrary set of continuous-valued evaluation metrics; and 2) calibrated uncertainty intervals around these point estimates. We then compare two approaches: LLM-as-judge vs. classic regression with confidence estimation features. Evaluation over 11 datasets spans summarization, translation, code generation, function-calling, and question answering. Regression is seen to achieve both: i) lower error point estimates of metric scores; and ii) well-calibrated uncertainty intervals. To support reproduction and follow-on work, we share our data and code.

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

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

Conclusion: MRAG-Suite provides a comprehensive diagnostic framework that reveals critical weaknesses in Visual RAG systems and offers guidance for future improvements through systematic evaluation of query difficulty and ambiguity.

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

Method: Created a reproducible pipeline to build datasets from publicly available GPS traces from OpenStreetMap, enriched with contextual layers (stops, moves, POIs, inferred transportation modes, weather data), and added synthetic realistic social media posts generated by Large Language Models.

Result: Two publicly available datasets covering Paris and New York in both tabular and RDF formats, supporting semantic reasoning and FAIR data practices, with an open-source pipeline for customization.

Conclusion: This resource is the first to combine real-world movement, structured semantic enrichment, LLM-generated text, and semantic web compatibility in a reusable framework, supporting research in behavior modeling, mobility prediction, knowledge graph construction, and LLM-based applications.

Abstract: In this resource paper, we present two publicly available datasets of semantically enriched human trajectories, together with the pipeline to build them. The trajectories are publicly available GPS traces retrieved from OpenStreetMap. Each dataset includes contextual layers such as stops, moves, points of interest (POIs), inferred transportation modes, and weather data. A novel semantic feature is the inclusion of synthetic, realistic social media posts generated by Large Language Models (LLMs), enabling multimodal and semantic mobility analysis. The datasets are available in both tabular and Resource Description Framework (RDF) formats, supporting semantic reasoning and FAIR data practices. They cover two structurally distinct, large cities: Paris and New York. Our open source reproducible pipeline allows for dataset customization, while the datasets support research tasks such as behavior modeling, mobility prediction, knowledge graph construction, and LLM-based applications. To our knowledge, our resource is the first to combine real-world movement, structured semantic enrichment, LLM-generated text, and semantic web compatibility in a reusable framework.

Method: AutoContext performs systematic, one-off exploration to construct reusable knowledge graphs for each environment instance. This structured context allows off-the-shelf agents to access necessary facts directly, eliminating the need for redundant exploration during task execution.

Result: Experiments across TextWorld, ALFWorld, Crafter, and InterCode-Bash show substantial gains. For example, the success rate of a ReAct agent on TextWorld improves from 37% to 95%, demonstrating the effectiveness of structured instance context.

Conclusion: Structured instance context is critical for efficient agentic systems. AutoContext’s approach of decoupling exploration from task solving through reusable knowledge graphs significantly improves agent performance by eliminating redundant interactions.

Abstract: Current LLM agents typically lack instance-level context, which comprises concrete facts such as environment structure, system configurations, and local mechanics. Consequently, existing methods are forced to intertwine exploration with task execution. This coupling leads to redundant interactions and fragile decision-making, as agents must repeatedly rediscover the same information for every new task. To address this, we introduce AutoContext, a method that decouples exploration from task solving. AutoContext performs a systematic, one-off exploration to construct a reusable knowledge graph for each environment instance. This structured context allows off-the-shelf agents to access necessary facts directly, eliminating redundant exploration. Experiments across TextWorld, ALFWorld, Crafter, and InterCode-Bash demonstrate substantial gains: for example, the success rate of a ReAct agent on TextWorld improves from 37% to 95%, highlighting the critical role of structured instance context in efficient agentic systems.

Method: 1) Introduced Language-Mixed CoT reasoning schema that switches between English and target language; 2) Created Yi-Sang dataset: 5.79M Korean prompts, 3.7M reasoning traces from Qwen3-32B, and 260k high-yield subset; 3) Trained nine models (4B-35B) across six model families.

Result: Best model KO-REAson-35B achieves SOTA with highest overall average score (64.0 ± 25), ranking first on 5/9 benchmarks and second on remainder. Smaller models show average +18.6 point improvement. Language-Mixed CoT outperforms monolingual CoT and yields cross-lingual and multimodal gains.

Conclusion: Language-Mixed CoT effectively enhances language-specific reasoning, with Korean case study demonstrating substantial improvements. The approach advances research on non-English reasoning and the released resources support further development.

Abstract: Recent frontier models employ long chain-of-thought reasoning to explore solution spaces in context and achieve stonger performance. While many works study distillation to build smaller yet capable models, most focus on English and little is known about language-specific reasoning. To bridge this gap, we first introduct Language-Mixed CoT, a reasoning schema that switches between English and a target language, using English as an anchor to excel in reasoning while minimizing translation artificats. As a Korean case study, we curate Yi-Sang: 5.79M native-Korean prompts from web Q&A, exams, STEM, and code; 3.7M long reasoning traces generated from Qwen3-32B; and a targeted 260k high-yield subset. We train ninve models (4B-35B) across six families (Qwen2.5, Llama-3.1, Gemma-3, etc). Our best model, KO-REAson-35B, achieves state-of-the-art performance, with the highest overall average score (64.0 \pm 25), ranking first on 5/9 benchmarks and second on the remainder. Samller and mid-sized models also benefit substantially, with an average improvement of +18.6 points across teh evaluated nine benchmarks. Ablations show Language-Mixed CoT is more effective than monolingual CoT, also resulting in cross-lingual and mult-modal performance gains. We release our data-curation pipeline, evaluation system, datasets, and models to advance research on language-specific reasoning. Data and model collection: https://huggingface.co/KOREAson.

Method: The study uses miniGPTKBs (domain-specific, tractable subcrawls) to systematically analyze LLM knowledge materialization. It examines three categories of metrics (yield, lexical similarity, semantic similarity) across four variations (seed, language, randomness, model) and three illustrative domains (history, entertainment, finance).

Result: Findings show: (i) high termination rates (though model-dependent), (ii) mixed reproducibility, and (iii) robustness that varies by perturbation type - high for seeds and temperature, lower for languages and models.

Conclusion: LLM knowledge materialization can reliably surface core knowledge, but also reveals important limitations in reproducibility and robustness across different variations.

Abstract: Large Language Models (LLMs) encode substantial factual knowledge, yet measuring and systematizing this knowledge remains challenging. Converting it into structured format, for example through recursive extraction approaches such as the GPTKB methodology (Hu et al., 2025b), is still underexplored. Key open questions include whether such extraction can terminate, whether its outputs are reproducible, and how robust they are to variations. We systematically study LLM knowledge materialization using miniGPTKBs (domain-specific, tractable subcrawls), analyzing termination, reproducibility, and robustness across three categories of metrics: yield, lexical similarity, and semantic similarity. We experiment with four variations (seed, language, randomness, model) and three illustrative domains (from history, entertainment, and finance). Our findings show (i) high termination rates, though model-dependent; (ii) mixed reproducibility; and (iii) robustness that varies by perturbation type: high for seeds and temperature, lower for languages and models. These results suggest that LLM knowledge materialization can reliably surface core knowledge, while also revealing important limitations.

Method: A label-free self-supervised RL framework that derives reward signals directly from instructions, generates pseudo-labels for reward model training, uses constraint decomposition strategies, and employs efficient constraint-wise binary classification to address sparse rewards while maintaining computational efficiency.

Result: The approach generalizes well, achieving strong improvements across 3 in-domain and 5 out-of-domain datasets, including challenging agentic and multi-turn instruction following tasks.

Conclusion: The proposed self-supervised RL framework effectively addresses multi-constraint instruction following challenges without external supervision, demonstrating strong generalization across diverse datasets and task types.

Abstract: Language models often struggle to follow multi-constraint instructions that are crucial for real-world applications. Existing reinforcement learning (RL) approaches suffer from dependency on external supervision and sparse reward signals from multi-constraint tasks. We propose a label-free self-supervised RL framework that eliminates dependency on external supervision by deriving reward signals directly from instructions and generating pseudo-labels for reward model training. Our approach introduces constraint decomposition strategies and efficient constraint-wise binary classification to address sparse reward challenges while maintaining computational efficiency. Experiments show that our approach generalizes well, achieving strong improvements across 3 in-domain and 5 out-of-domain datasets, including challenging agentic and multi-turn instruction following. The data and code are publicly available at https://github.com/Rainier-rq/verl-if

Method: Proposes Multi-Personality Generation (MPG) framework that reformulates the task as sampling from a target strategy aggregating implicit density ratios from single-dimensional models. Implements Speculative Chunk-level based Rejection sampling (SCR) which generates responses in chunks and validates them in parallel using estimated thresholds within a sliding window to reduce computational overhead.

Result: Experiments on MBTI personality and Role-Playing tasks show MPG achieves improvements up to 16%-18% over baseline methods, demonstrating effectiveness in multi-personality generation.

Conclusion: MPG provides an efficient, flexible framework for multi-personality generation in LLMs without requiring retraining or external models, leveraging implicit density ratios as a “free lunch” and achieving significant performance improvements through speculative chunk-level sampling.

Abstract: Multi-personality generation for LLMs, enabling simultaneous embodiment of multiple personalization attributes, is a fundamental challenge. Existing retraining-based approaches are costly and poorly scalable, while decoding-time methods often rely on external models or heuristics, limiting flexibility and robustness. In this paper, we propose a novel Multi-Personality Generation (MPG) framework under the decoding-time combination paradigm. It flexibly controls multi-personality without relying on scarce multi-dimensional models or extra training, leveraging implicit density ratios in single-dimensional models as a “free lunch” to reformulate the task as sampling from a target strategy aggregating these ratios. To implement MPG efficiently, we design Speculative Chunk-level based Rejection sampling (SCR), which generates responses in chunks and parallelly validates them via estimated thresholds within a sliding window. This significantly reduces computational overhead while maintaining high-quality generation. Experiments on MBTI personality and Role-Playing demonstrate the effectiveness of MPG, showing improvements up to 16%-18%. Code and data are available at https://github.com/Libra117/MPG .

Method: 1) Theoretical analysis of scaling behavior with respect to dataset size in a simplified setting, focusing on power law exponents. 2) Empirical measurement of miscalibration across language models ranging from 0.5B to 70B parameters. 3) Theoretical proof showing calibration without log loss tradeoffs is possible given access to a black box that can predict future entropy of text.

Result: 1) Theoretical analysis shows miscalibration scaling depends on power law exponent - for exponents close to 1, scaling exponent is close to 0, meaning miscalibration improves very slowly with scale. 2) Empirical results confirm similar scaling behavior: fitted scaling exponents for text are close to 0, meaning larger models accumulate error at similar rates as smaller ones. 3) Theoretical proof demonstrates calibration without log loss tradeoffs is possible with access to future entropy predictions.

Conclusion: Miscalibration does not automatically improve with scale, explaining why larger models require similar truncation as smaller ones despite higher quality. While current truncation approaches sacrifice diversity, the paper proves calibration without log loss tradeoffs is theoretically possible, pointing toward future research directions for better calibration methods.

Abstract: We study the problem of entropy calibration, which asks whether a language model’s entropy over generations matches its log loss on human text. Past work found that models are miscalibrated, with entropy per step increasing as generations grow longer, due to error accumulation. To calibrate the model and improve text quality, it has become standard practice to truncate the distribution, but this approach reduces output diversity, which we would like to avoid. Therefore, in this paper, we ask: does miscalibration improve automatically with scale, and if not, is it theoretically possible to calibrate without tradeoffs? To build intuition, we first study a simplified theoretical setting to characterize the scaling behavior of miscalibration with respect to dataset size. We find that the rate of scaling depends on the power law exponent of the data distribution – in particular, for a power law exponent close to 1, the scaling exponent is close to 0, meaning that miscalibration improves very slowly with scale. Next, we measure miscalibration empirically in language models ranging from 0.5B to 70B parameters. We find that the observed scaling behavior is similar to what is predicted theoretically: our fitted scaling exponents for text are close to 0, meaning that larger models accumulate error at a similar rate as smaller ones. This scaling (or, lack thereof) provides one explanation for why we sample from larger models with similar amounts of truncation as smaller models, even though the larger models are of higher quality. However, truncation is not a satisfying solution because it comes at the cost of increased log loss. In theory, is it even possible to reduce entropy while preserving log loss? We prove that it is possible, if we assume access to a black box which can fit models to predict the future entropy of text.

Method: Developed four FGTI principles (Faithfulness, Groundedness, Traceability, Interchangeability) for interpretable scoring, then created AnalyticScore framework for short answer scoring as a baseline implementation.

Result: AnalyticScore outperforms many uninterpretable methods and is within 0.06 QWK of uninterpretable SOTA across 10 ASAP-SAS items. Its featurization behavior aligns well with human annotators.

Conclusion: The FGTI principles provide a principled approach to interpretable automated scoring, and AnalyticScore demonstrates feasibility of implementing these principles while maintaining competitive accuracy.

Abstract: AI-driven automated scoring systems offer scalable and efficient means of evaluating complex student-generated responses. Yet, despite increasing demand for transparency and interpretability, the field has yet to develop a widely accepted solution for interpretable automated scoring to be used in large-scale real-world assessments. This work takes a principled approach to address this challenge. We analyze the needs and potential benefits of interpretable automated scoring for various assessment stakeholder groups and develop four principles of interpretability – (F)aithfulness, (G)roundedness, (T)raceability, and (I)nterchangeability (FGTI) – targeted at those needs. To illustrate the feasibility of implementing these principles, we develop the AnalyticScore framework for short answer scoring as a baseline reference framework for future research. In terms of scoring accuracy, AnalyticScore outperforms many uninterpretable scoring methods and is, on average, within 0.06 QWK of the uninterpretable SOTA across 10 items from the ASAP-SAS dataset. By comparing against human annotators conducting the same featurization task, we further demonstrate that the featurization behavior of AnalyticScore aligns well with that of humans.

Method: The paper provides a systematic survey by: 1) delineating agent memory scope from related concepts (LLM memory, RAG, context engineering), 2) analyzing memory through three unified lenses: forms (token-level, parametric, latent), functions (factual, experiential, working), and dynamics (formation, evolution, retrieval), and 3) compiling benchmarks and frameworks.

Result: Creates a comprehensive landscape of current agent memory research with clear conceptual foundations, identifies three dominant memory forms, proposes a finer-grained functional taxonomy, and provides practical resources including benchmarks and frameworks.

Conclusion: The survey serves as both a reference for existing work and a conceptual foundation for rethinking memory as a first-class primitive in future agent design, highlighting emerging frontiers like memory automation, RL integration, multimodal memory, multi-agent memory, and trustworthiness.

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

Method: Propose a method for finding a single adversarial text in discrete space that is consistently evaluated as high-quality regardless of test cases, to identify vulnerabilities in evaluation metrics.

Result: The hub text achieved 79.1 COMET% for En-Ja and 67.8 COMET% for En-De translation tasks, outperforming translations generated individually for each source sentence by M2M100. The hub text also generalizes across multiple language pairs (Ja-En and De-En).

Conclusion: The findings demonstrate significant vulnerabilities in neural evaluation metrics, showing they can be easily fooled by adversarial hub texts, raising serious concerns about their reliability and safety for guiding model development.

Abstract: Strongly human-correlated evaluation metrics serve as an essential compass for the development and improvement of generation models and must be highly reliable and robust. Recent embedding-based neural text evaluation metrics, such as COMET for translation tasks, are widely used in both research and development fields. However, there is no guarantee that they yield reliable evaluation results due to the black-box nature of neural networks. To raise concerns about the reliability and safety of such metrics, we propose a method for finding a single adversarial text in the discrete space that is consistently evaluated as high-quality, regardless of the test cases, to identify the vulnerabilities in evaluation metrics. The single hub text found with our method achieved 79.1 COMET% and 67.8 COMET% in the WMT'24 English-to-Japanese (En–Ja) and English-to-German (En–De) translation tasks, respectively, outperforming translations generated individually for each source sentence by using M2M100, a general translation model. Furthermore, we also confirmed that the hub text found with our method generalizes across multiple language pairs such as Ja–En and De–En.

Method: TG is a recurrent transformer that models language at two levels: tokens and sentence-level “thought” states. It generates one sentence at a time while cross-attending to a working memory of prior sentence representations. Token and sentence representations are generated using a shared stack of transformer blocks and trained with a single next-token prediction loss. Gradients from future token losses flow backward through cross-attention to optimize parameters that generate earlier sentence vectors.

Result: TG consistently improves data and parameter efficiency compared to matched GPT-2 runs and other baselines. Scaling fits indicate GPT-2 requires ~5-8% more data and ~33-42% more parameters to match TG’s test loss. TG also reduces errors in relational-direction generalization on a father-son reversal curse probe.

Conclusion: The Thought Gestalt model successfully addresses key limitations of standard transformers by incorporating sentence-level representations and working memory, leading to improved data efficiency, parameter efficiency, and relational generalization capabilities.

Abstract: Transformer language models can generate strikingly natural text by modeling language as a sequence of tokens, but by relying primarily on surface-level co-occurrence statistics they fail to form globally consistent latent representations of entities and events, which contributes to poor relational generalization (the reversal curse), contextualization errors, and data inefficiency. Cognitive science, by contrast, shows that human comprehension converts linguistic input into compact, event-like representations that persist in memory while verbatim form is short-lived. Motivated by these findings, we introduce the Thought Gestalt (TG) model, a recurrent transformer that models language at two levels of abstraction: tokens and sentence-level “thought” states. TG generates one sentence at a time while cross-attending to a working memory of prior sentence representations. Token and sentence representations are generated using a shared stack of transformer blocks and trained with a single objective, next-token prediction loss. By retaining the computation graph of sentence representations written to working memory, gradients from future token losses flow backward through cross-attention to optimize the parameters that generate earlier sentence vectors. In scaling experiments, TG consistently improves data and parameter efficiency compared to matched GPT-2 runs and other baselines, with scaling fits indicating GPT-2 requires ~5-8% more data and ~33-42% more parameters to match TG’s test loss. TG also reduces errors in relational-direction generalization on a father-son reversal curse probe.

Method: Introduce homotokens as meaning-preserving data augmentation. Use lightweight architecture with auxiliary causal encoder and block-causal cross-attention to condition canonical next-token prediction on sampled homotoken variants, without changing training objective or token interface.

Result: In data-constrained pretraining: consistently delays overfitting under repeated data exposure and improves generalization across diverse evaluation datasets. In multilingual fine-tuning: effectiveness depends on tokenizer quality - strongest gains when canonical tokens are highly compressed, diminishes when tokenizer already over-fragments input.

Conclusion: Homotokens provide a simple, modular mechanism for inducing tokenization invariance in language models, offering benefits for generalization and robustness to different tokenization strategies.

Abstract: Subword tokenization introduces a computational layer in language models where many distinct token sequences decode to the same surface form and preserve meaning, yet induce different internal computations. Despite this non-uniqueness, language models are typically trained using a single canonical longest-prefix tokenization. We formalize homotokens-alternative valid subword segmentations of the same lexical item-as a strictly meaning-preserving form of data augmentation. We introduce a lightweight training architecture that conditions canonical next-token prediction on sampled homotoken variants via an auxiliary causal encoder and block-causal cross-attention, without modifying the training objective or token interface. In data-constrained pretraining, homotoken augmentation consistently delays overfitting under repeated data exposure and improves generalization across diverse evaluation datasets. In multilingual fine-tuning, we find that the effectiveness of homotokens depends on tokenizer quality: gains are strongest when canonical tokens are highly compressed and diminish when the tokenizer already over-fragments the input. Overall, homotokens provide a simple and modular mechanism for inducing tokenization invariance in language models.

Method: PALU uses prefix-aware localized unlearning with local entropy maximization across temporal and vocabulary dimensions. It suppresses only the sensitive prefix to sever causal generation links and flattens only the top-k logits to maximize uncertainty in critical subspaces.

Result: Extensive experiments show PALU achieves superior forgetting efficacy and utility preservation compared to state-of-the-art baselines, avoiding redundant optimization while minimizing collateral damage to general model performance.

Conclusion: PALU demonstrates that targeted, localized unlearning focused on sensitive prefixes and critical vocabulary subspaces is more effective than global approaches for machine unlearning in LLMs.

Abstract: Machine unlearning aims to forget sensitive knowledge from Large Language Models (LLMs) while maintaining general utility. However, existing approaches typically treat all tokens in a response indiscriminately and enforce uncertainty over the entire vocabulary. This global treatment results in unnecessary utility degradation and extends optimization to content-agnostic regions. To address these limitations, we propose PALU (Prefix-Aware Localized Unlearning), a framework driven by a local entropy maximization objective across both temporal and vocabulary dimensions. PALU reveals that (i) suppressing the sensitive prefix alone is sufficient to sever the causal generation link, and (ii) flattening only the top-$k$ logits is adequate to maximize uncertainty in the critical subspace. These findings allow PALU to avoid redundant optimization across the full vocabulary and parameter space while minimizing collateral damage to general model performance. Extensive experiments validate that PALU achieves superior forgetting efficacy and utility preservation compared to state-of-the-art baselines.

Method: Proposed a novel framework to investigate source preferences in LLM resolution of inter-context knowledge conflicts. Conducted comprehensive, tightly-controlled evaluation of 13 open-weight LLMs. Developed a method to mitigate repetition effects that reduces repetition bias while maintaining original preferences.

Result: LLMs prefer institutionally-corroborated information (government, newspaper) over information from people and social media. However, these source preferences can be reversed by simply repeating information from less credible sources. The proposed method reduces repetition bias by up to 99.8% while maintaining at least 88.8% of original preferences.

Conclusion: Source preferences significantly affect LLM knowledge conflict resolution, but are vulnerable to repetition effects. The proposed mitigation method effectively addresses this vulnerability while preserving original source preferences. Data and code are released to encourage future work on credibility and source preferences in knowledge-intensive NLP.

Abstract: As large language models (LLMs) are more frequently used in retrieval-augmented generation pipelines, it is increasingly relevant to study their behavior under knowledge conflicts. Thus far, the role of the source of the retrieved information has gone unexamined. We address this gap with a novel framework to investigate how source preferences affect LLM resolution of inter-context knowledge conflicts in English, motivated by interdisciplinary research on credibility. With a comprehensive, tightly-controlled evaluation of 13 open-weight LLMs, we find that LLMs prefer institutionally-corroborated information (e.g., government or newspaper sources) over information from people and social media. However, these source preferences can be reversed by simply repeating information from less credible sources. To mitigate repetition effects and maintain consistent preferences, we propose a novel method that reduces repetition bias by up to 99.8%, while also maintaining at least 88.8% of original preferences. We release all data and code to encourage future work on credibility and source preferences in knowledge-intensive NLP.

Method: Used neural networks on text embeddings from online news sources; developed models for social media using word-level (GoEmotions) and context-level (LLM) methods; created MirrorMirror Chrome extension for real-time peacefulness feedback.

Result: News model trained on one dataset showed high accuracy on different news dataset; developed functional Chrome extension; demonstrated cross-dataset generalization capability for peace measurement.

Conclusion: The MirrorMirror tool aims to evolve into an open-source platform for content creators, journalists, researchers, and users to understand media tone effects, encouraging more respectful and informative communication beyond engagement metrics.

Abstract: We used machine learning and artificial intelligence: 1) to measure levels of peace in countries from news and social media and 2) to develop on-line tools that promote peace by helping users better understand their own media diet. For news media, we used neural networks to measure levels of peace from text embeddings of on-line news sources. The model, trained on one news media dataset also showed high accuracy when used to analyze a different news dataset. For social media, such as YouTube, we developed other models to measure levels of social dimensions important in peace using word level (GoEmotions) and context level (Large Language Model) methods. To promote peace, we note that 71% of people 20-40 years old daily view most of their news through short videos on social media. Content creators of these videos are biased towards creating videos with emotional activation, making you angry to engage you, to increase clicks. We developed and tested a Chrome extension, MirrorMirror, which provides real-time feedback to YouTube viewers about the peacefulness of the media they are watching. Our long term goal is for MirrorMirror to evolve into an open-source tool for content creators, journalists, researchers, platforms, and individual users to better understand the tone of their media creation and consumption and its effects on viewers. Moving beyond simple engagement metrics, we hope to encourage more respectful, nuanced, and informative communication.

Method: Analyze distilled trajectories to identify molecular-like structures in Long CoT, propose three interaction types (Deep-Reasoning, Self-Reflection, Self-Exploration), introduce Effective Semantic Isomers concept, and develop Mole-Syn - a distribution-transfer-graph method for synthesizing effective Long CoT structures.

Result: Analysis shows stable molecular structures emerge from Long CoT fine-tuning, not keyword imitation. Only bonds promoting fast entropy convergence support stable learning. Mole-Syn boosts performance and RL stability across benchmarks.

Conclusion: Effective Long CoT reasoning requires specific molecular-like structural properties, and Mole-Syn successfully synthesizes these structures to improve LLM reasoning capabilities and training stability.

Abstract: Large language models (LLMs) often fail to learn effective long chain-of-thought (Long CoT) reasoning from human or non-Long-CoT LLMs imitation. To understand this, we propose that effective and learnable Long CoT trajectories feature stable molecular-like structures in unified view, which are formed by three interaction types: Deep-Reasoning (covalent-like), Self-Reflection (hydrogen-bond-like), and Self-Exploration (van der Waals-like). Analysis of distilled trajectories reveals these structures emerge from Long CoT fine-tuning, not keyword imitation. We introduce Effective Semantic Isomers and show that only bonds promoting fast entropy convergence support stable Long CoT learning, while structural competition impairs training. Drawing on these findings, we present Mole-Syn, a distribution-transfer-graph method that guides synthesis of effective Long CoT structures, boosting performance and RL stability across benchmarks.

Method: Created IndRegBias dataset with 25,000 comments from Reddit and YouTube discussing Indian regional issues. Developed multilevel annotation strategy for bias severity. Evaluated open-source LLMs and Indic Language Models using zero-shot, few-shot, and fine-tuning approaches.

Result: Zero-shot and few-shot approaches showed low accuracy in detecting regional biases and severity across most LLMs and ILMs. However, fine-tuning significantly enhanced performance in detecting Indian regional bias and its severity levels.

Conclusion: Fine-tuning is crucial for effective detection of regional biases in Indian contexts. The IndRegBias dataset enables better research on regional bias detection, addressing an important gap in NLP bias studies.

Abstract: Warning: This paper consists of examples representing regional biases in Indian regions that might be offensive towards a particular region. While social biases corresponding to gender, race, socio-economic conditions, etc., have been extensively studied in the major applications of Natural Language Processing (NLP), biases corresponding to regions have garnered less attention. This is mainly because of (i) difficulty in the extraction of regional bias datasets, (ii) disagreements in annotation due to inherent human biases, and (iii) regional biases being studied in combination with other types of social biases and often being under-represented. This paper focuses on creating a dataset IndRegBias, consisting of regional biases in an Indian context reflected in users’ comments on popular social media platforms, namely Reddit and YouTube. We carefully selected 25,000 comments appearing on various threads in Reddit and videos on YouTube discussing trending topics on regional issues in India. Furthermore, we propose a multilevel annotation strategy to annotate the comments describing the severity of regional biased statements. To detect the presence of regional bias and its severity in IndRegBias, we evaluate open-source Large Language Models (LLMs) and Indic Language Models (ILMs) using zero-shot, few-shot, and fine-tuning strategies. We observe that zero-shot and few-shot approaches show lower accuracy in detecting regional biases and severity in the majority of the LLMs and ILMs. However, the fine-tuning approach significantly enhances the performance of the LLM in detecting Indian regional bias along with its severity.

Method: The authors propose a method using contrastive learning on a hypersphere to induce circular emotion representations within language model embeddings, aligning with psychological circumplex geometry.

Result: Circular alignment offers superior interpretability and robustness against dimensionality reduction, but underperforms compared to conventional designs in high-dimensional settings and fine-grained classification tasks.

Conclusion: The findings elucidate the trade-offs involved in applying psychological circumplex models to deep learning architectures, highlighting the balance between interpretability/robustness and performance in different contexts.

Abstract: Psychological research has long utilized circumplex models to structure emotions, placing similar emotions adjacently and opposing ones diagonally. Although frequently used to interpret deep learning representations, these models are rarely directly incorporated into the representation learning of language models, leaving their geometric validity unexplored. This paper proposes a method to induce circular emotion representations within language model embeddings via contrastive learning on a hypersphere. We show that while this circular alignment offers superior interpretability and robustness against dimensionality reduction, it underperforms compared to conventional designs in high-dimensional settings and fine-grained classification. Our findings elucidate the trade-offs involved in applying psychological circumplex models to deep learning architectures.

Method: Algorithmically generated puzzles combining factual temporal anchors with cross-cultural calendar relations, each admitting one or multiple valid solution dates.

Result: GPT-5 achieves only 49.3% accuracy, all other models below 31%, showing the task remains challenging. Web search helps substantially, code interpreter has mixed effects, and explicit date rewriting improves performance significantly.

Conclusion: Time Puzzles provides a simple, cost-effective diagnostic for tool-augmented iterative temporal reasoning, revealing gaps in reliable tool use and distinguishing model capabilities.

Abstract: We introduce Time Puzzles, a constraint-based date inference task for evaluating iterative temporal reasoning. Each puzzle combines factual temporal anchors with (cross-cultural) calendar relations, admits one or multiple valid solution dates, and is algorithmically generated for controlled, dynamic, and continual evaluation. Across 13 diverse LLMs, Time Puzzles well distinguishes their iterative temporal reasoning capabilities and remains challenging without tools: GPT-5 reaches only 49.3% accuracy and all other models stay below 31%, despite the dataset’s simplicity. Web search consistently yields substantial gains and using code interpreter shows mixed effects, but all models perform much better when constraints are rewritten with explicit dates, revealing a gap in reliable tool use. Overall, Time Puzzles presents a simple, cost-effective diagnostic for tool-augmented iterative temporal reasoning.

Method: Controlled Self-Evolution (CSE) with three key components: 1) Diversified Planning Initialization for broad solution space coverage, 2) Genetic Evolution with feedback-guided mutation and compositional crossover, 3) Hierarchical Evolution Memory capturing successful and failed experiences at inter-task and intra-task levels.

Result: CSE consistently outperforms all baselines across various LLM backbones on EffiBench-X, achieves higher efficiency from early generations, and maintains continuous improvement throughout evolution.

Conclusion: CSE effectively addresses exploration inefficiencies in self-evolution methods for code generation through controlled evolution mechanisms, demonstrating superior performance and efficiency.

Abstract: Self-evolution methods enhance code generation through iterative “generate-verify-refine” cycles, yet existing approaches suffer from low exploration efficiency, failing to discover solutions with superior complexity within limited budgets. This inefficiency stems from initialization bias trapping evolution in poor solution regions, uncontrolled stochastic operations lacking feedback guidance, and insufficient experience utilization across tasks. To address these bottlenecks, we propose Controlled Self-Evolution (CSE), which consists of three key components. Diversified Planning Initialization generates structurally distinct algorithmic strategies for broad solution space coverage. Genetic Evolution replaces stochastic operations with feedback-guided mechanisms, enabling targeted mutation and compositional crossover. Hierarchical Evolution Memory captures both successful and failed experiences at inter-task and intra-task levels. Experiments on EffiBench-X demonstrate that CSE consistently outperforms all baselines across various LLM backbones. Furthermore, CSE achieves higher efficiency from early generations and maintains continuous improvement throughout evolution. Our code is publicly available at https://github.com/QuantaAlpha/EvoControl.

Method: ES-Mem incorporates two core components: (1) dynamic event segmentation module that partitions long-term interactions into semantically coherent events with distinct boundaries, and (2) hierarchical memory architecture that constructs multi-layered memories and leverages boundary semantics to anchor specific episodic memory for precise context localization.

Result: Evaluations on two memory benchmarks demonstrate consistent performance gains over baseline methods. The event segmentation module also exhibits robust applicability on dialogue segmentation datasets.

Conclusion: ES-Mem effectively addresses limitations of existing memory mechanisms by incorporating event segmentation theory to create coherent memory units and hierarchical retrieval for precise episodic context localization in dialogue agents.

Abstract: Memory is critical for dialogue agents to maintain coherence and enable continuous adaptation in long-term interactions. While existing memory mechanisms offer basic storage and retrieval capabilities, they are hindered by two primary limitations: (1) rigid memory granularity often disrupts semantic integrity, resulting in fragmented and incoherent memory units; (2) prevalent flat retrieval paradigms rely solely on surface-level semantic similarity, neglecting the structural cues of discourse required to navigate and locate specific episodic contexts. To mitigate these limitations, drawing inspiration from Event Segmentation Theory, we propose ES-Mem, a framework incorporating two core components: (1) a dynamic event segmentation module that partitions long-term interactions into semantically coherent events with distinct boundaries; (2) a hierarchical memory architecture that constructs multi-layered memories and leverages boundary semantics to anchor specific episodic memory for precise context localization. Evaluations on two memory benchmarks demonstrate that ES-Mem yields consistent performance gains over baseline methods. Furthermore, the proposed event segmentation module exhibits robust applicability on dialogue segmentation datasets.

Method: Integrated YOLOv8 Nano for object detection, DeepSORT for vehicle tracking, and rule-guided OCR for license plate recognition on NVIDIA Jetson Nano with TensorRT FP16 optimization.

Result: 97.7% violation detection accuracy, 84.9% OCR precision, 28-30 FPS at 9.6W, 10.7% mAP gain over alternatives, with 1.4x accuracy per watt improvement.

Conclusion: Roadside edge AI can effectively augment cooperative perception and proactive safety management within intelligent transportation systems through V2X integration.

Abstract: Rapid motorization in emerging economies such as India has created severe enforcement asymmetries, with over 11 million recorded violations in 2023 against a human policing density of roughly one officer per 4000 vehicles. Traditional surveillance and manual ticketing cannot scale to this magnitude, motivating the need for an autonomous, cooperative, and energy efficient edge AI perception infrastructure. This paper presents a real time roadside perception node for multi class traffic violation analytics and safety event dissemination within a connected and intelligent vehicle ecosystem. The node integrates YOLOv8 Nano for high accuracy multi object detection, DeepSORT for temporally consistent vehicle tracking, and a rule guided OCR post processing engine capable of recognizing degraded or multilingual license plates compliant with MoRTH AIS 159 and ISO 7591 visual contrast standards. Deployed on an NVIDIA Jetson Nano with a 128 core Maxwell GPU and optimized via TensorRT FP16 quantization, the system sustains 28 to 30 frames per second inference at 9.6 W, achieving 97.7 percent violation detection accuracy and 84.9 percent OCR precision across five violation classes, namely signal jumping, zebra crossing breach, wrong way driving, illegal U turn, and speeding, without manual region of interest calibration. Comparative benchmarking against YOLOv4 Tiny, PP YOLOE S, and Nano DetPlus demonstrates a 10.7 percent mean average precision gain and a 1.4 times accuracy per watt improvement. Beyond enforcement, the node publishes standardized safety events of CAM and DENM type to connected vehicles and intelligent transportation system backends via V2X protocols, demonstrating that roadside edge AI analytics can augment cooperative perception and proactive road safety management within the IEEE Intelligent Vehicles ecosystem.

Method: Two complementary attention processors: 1) Cross-Attention Object Fusion (CAOF) averages head-wise attention to locate spatial tokens, solves optimal transport to reassign feature vectors; 2) Self-Attention Style Fusion (SASF) injects style through Detail-Sensitive Instance Normalization and swaps Key/Value matrices with style prompt.

Result: Produces high-resolution, photo-realistic edits with precise control over both content and appearance, surpassing recent baselines in quantitative fidelity, perceptual quality, and inference speed.

Conclusion: TP-Blend effectively handles simultaneous object and style editing through a lightweight training-free framework that preserves cross-head correlations while maintaining memory efficiency.

Abstract: Current text-conditioned diffusion editors handle single object replacement well but struggle when a new object and a new style must be introduced simultaneously. We present Twin-Prompt Attention Blend (TP-Blend), a lightweight training-free framework that receives two separate textual prompts, one specifying a blend object and the other defining a target style, and injects both into a single denoising trajectory. TP-Blend is driven by two complementary attention processors. Cross-Attention Object Fusion (CAOF) first averages head-wise attention to locate spatial tokens that respond strongly to either prompt, then solves an entropy-regularised optimal transport problem that reassigns complete multi-head feature vectors to those positions. CAOF updates feature vectors at the full combined dimensionality of all heads (e.g., 640 dimensions in SD-XL), preserving rich cross-head correlations while keeping memory low. Self-Attention Style Fusion (SASF) injects style at every self-attention layer through Detail-Sensitive Instance Normalization. A lightweight one-dimensional Gaussian filter separates low- and high-frequency components; only the high-frequency residual is blended back, imprinting brush-stroke-level texture without disrupting global geometry. SASF further swaps the Key and Value matrices with those derived from the style prompt, enforcing context-aware texture modulation that remains independent of object fusion. Extensive experiments show that TP-Blend produces high-resolution, photo-realistic edits with precise control over both content and appearance, surpassing recent baselines in quantitative fidelity, perceptual quality, and inference speed.

Method: Proposed ROAD benchmark for LiDAR-based object classification that accounts for domain shifts and three forms of label evolution (class split, expansion, insertion). Evaluated zero-shot transfer, linear probe, and continual learning methods using large-scale datasets (Waymo, NuScenes, Argoverse2), analyzing backbone architectures, training objectives, and CL methods.

Result: Findings reveal limitations of existing approaches under realistic shifts and establish strong baselines for future research in robust 3D perception.

Conclusion: The ROAD benchmark addresses a critical gap in 3D perception research and provides a comprehensive evaluation suite to advance robust autonomous driving systems that can handle evolving domains and object definitions.

Abstract: For 3D perception systems to be practical in real-world applications – from autonomous driving to embodied AI – models must adapt to continuously evolving object definitions and sensor domains. Yet, research on continual and transfer learning in 3D point cloud perception remains underexplored compared to 2D vision – particularly under simultaneous domain and label shifts. To address this gap, we propose the RObust Autonomous driving under Dataset shifts (ROAD) benchmark, a comprehensive evaluation suite for LiDAR-based object classification that explicitly accounts for domain shifts as well as three key forms of label evolution: class split, class expansion, and class insertion. Using large-scale datasets (Waymo, NuScenes, Argoverse2), we evaluate zero-shot transfer, linear probe, and CL, and analyze the impact of backbone architectures, training objectives, and CL methods. Our findings reveal limitations of existing approaches under realistic shifts and establish strong baselines for future research in robust 3D perception.

Method: Performed systematic layer-wise masking analysis across multiple MLLM architectures to study visual-text fusion evolution. Analyzed layer-wise attention evolution, then introduced a training-free contrastive attention framework that models transformation between early fusion and final layers to highlight meaningful attention shifts.

Result: Found that fusion emerges at specific layers rather than being uniformly distributed, with some models showing late-stage “review” phenomenon where visual signals are reactivated. Observed persistent high-attention noise on irrelevant regions and gradually increasing attention on text-aligned areas. The proposed approach improves multimodal reasoning performance across various benchmarks.

Conclusion: The systematic analysis provides insights into MLLM fusion mechanisms, revealing specific fusion patterns and attention behaviors. The proposed training-free contrastive attention framework successfully leverages these insights to enhance multimodal reasoning performance without additional training.

Abstract: Multimodal Large Language Models (MLLMs) have achieved remarkable progress in vision-language understanding, yet how they internally integrate visual and textual information remains poorly understood. To bridge this gap, we perform a systematic layer-wise masking analysis across multiple architectures, revealing how visual-text fusion evolves within MLLMs. The results show that fusion emerges at several specific layers rather than being uniformly distributed across the network, and certain models exhibit a late-stage “review” phenomenon where visual signals are reactivated before output generation. Besides, we further analyze layer-wise attention evolution and observe persistent high-attention noise on irrelevant regions, along with gradually increasing attention on text-aligned areas. Guided by these insights, we introduce a training-free contrastive attention framework that models the transformation between early fusion and final layers to highlight meaningful attention shifts. Extensive experiments across various MLLMs and benchmarks validate our analysis and demonstrate that the proposed approach improves multimodal reasoning performance. Code will be released.

Method: Images generated using Moonworks Lunara model, intentionally crafted to embody distinct aesthetic styles, with human-refined prompts and structured annotations describing objects, attributes, relationships, and stylistic cues.

Result: Created a dataset with substantially higher aesthetic scores than existing aesthetics-focused and general-purpose datasets, spanning diverse regional aesthetics and general artistic styles.

Conclusion: The Lunara Aesthetic Dataset provides a valuable resource for research with high aesthetic quality, stylistic diversity, and transparent licensing for unrestricted academic and commercial use.

Abstract: The dataset spans diverse artistic styles, including regionally grounded aesthetics from the Middle East, Northern Europe, East Asia, and South Asia, alongside general categories such as sketch and oil painting. All images are generated using the Moonworks Lunara model and intentionally crafted to embody distinct, high-quality aesthetic styles, yielding a first-of-its-kind dataset with substantially higher aesthetic scores, exceeding even aesthetics-focused datasets, and general-purpose datasets by a larger margin. Each image is accompanied by a human-refined prompt and structured annotations that jointly describe salient objects, attributes, relationships, and stylistic cues. Unlike large-scale web-derived datasets that emphasize breadth over precision, the Lunara Aesthetic Dataset prioritizes aesthetic quality, stylistic diversity, and licensing transparency, and is released under the Apache 2.0 license to support research and unrestricted academic and commercial use.

Method: Proposes Instruction-driven Facial Expression Decomposer (IFED) for multimodal learning, Instruction to Facial Expression Transition (I2FET) with vertex reconstruction loss for semantic comprehension, and Facial Expression Transition model for smooth transitions.

Result: Outperforms state-of-the-art methods on CK+ and CelebV-HQ datasets, generating facial expression trajectories according to text instructions with smooth transitions.

Conclusion: The framework enables diverse facial expression generation and transitions via text instructions, expanding expression repertoire for practical applications in 3D avatar animation.

Abstract: A 3D avatar typically has one of six cardinal facial expressions. To simulate realistic emotional variation, we should be able to render a facial transition between two arbitrary expressions. This study presents a new framework for instruction-driven facial expression generation that produces a 3D face and, starting from an image of the face, transforms the facial expression from one designated facial expression to another. The Instruction-driven Facial Expression Decomposer (IFED) module is introduced to facilitate multimodal data learning and capture the correlation between textual descriptions and facial expression features. Subsequently, we propose the Instruction to Facial Expression Transition (I2FET) method, which leverages IFED and a vertex reconstruction loss function to refine the semantic comprehension of latent vectors, thus generating a facial expression sequence according to the given instruction. Lastly, we present the Facial Expression Transition model to generate smooth transitions between facial expressions. Extensive evaluation suggests that the proposed model outperforms state-of-the-art methods on the CK+ and CelebV-HQ datasets. The results show that our framework can generate facial expression trajectories according to text instruction. Considering that text prompts allow us to make diverse descriptions of human emotional states, the repertoire of facial expressions and the transitions between them can be expanded greatly. We expect our framework to find various practical applications More information about our project can be found at https://vohoanganh.github.io/tg3dfet/

Method: Proposes LWMSCNN-SE, a lightweight CNN that integrates multi-scale feature extraction, depthwise separable convolutions, and squeeze-and-excitation (SE) attention mechanisms to balance accuracy and efficiency.

Result: Achieves 96.63% classification accuracy with only 241,348 parameters and 0.666 GFLOPs, demonstrating excellent performance with minimal computational requirements.

Conclusion: The model successfully addresses the accuracy-efficiency trade-off, making it suitable for real-time deployment on edge devices in precision farming systems for efficient maize disease diagnosis.

Abstract: Maize disease classification plays a vital role in mitigating yield losses and ensuring food security. However, the deployment of traditional disease detection models in resource-constrained environments, such as those using smartphones and drones, faces challenges due to high computational costs. To address these challenges, we propose LWMSCNN-SE, a lightweight convolutional neural network (CNN) that integrates multi-scale feature extraction, depthwise separable convolutions, and squeeze-and-Excitation (SE) attention mechanisms. This novel combination enables the model to achieve 96.63% classification accuracy with only 241,348 parameters and 0.666 GFLOPs, making it suitable for real-time deployment in field applications. Our approach addresses the accuracy–efficiency trade-off by delivering high accuracy while maintaining low computational costs, demonstrating its potential for efficient maize disease diagnosis on edge devices in precision farming systems.

Method: CoEvo introduces a proxy-aligned co-evolution mechanism with two evolving proxy caches: (1) dynamically mines contextual textual negatives guided by test images, (2) iteratively refines visual proxies, (3) progressively realigns cross-modal similarities, and (4) dynamically re-weights dual-modal contributions for calibrated OOD scores.

Result: CoEvo achieves state-of-the-art performance on standard benchmarks, improving AUROC by 1.33% and reducing FPR95 by 45.98% on ImageNet-1K compared to strong negative-label baselines.

Conclusion: The proposed bidirectional, sample-conditioned adaptation framework effectively addresses cross-modal misalignment in zero-shot OOD detection, demonstrating superior performance through dynamic proxy evolution without requiring training or annotations.

Abstract: Reliable zero-shot detection of out-of-distribution (OOD) inputs is critical for deploying vision-language models in open-world settings. However, the lack of labeled negatives in zero-shot OOD detection necessitates proxy signals that remain effective under distribution shift. Existing negative-label methods rely on a fixed set of textual proxies, which (i) sparsely sample the semantic space beyond in-distribution (ID) classes and (ii) remain static while only visual features drift, leading to cross-modal misalignment and unstable predictions. In this paper, we propose CoEvo, a training- and annotation-free test-time framework that performs bidirectional, sample-conditioned adaptation of both textual and visual proxies. Specifically, CoEvo introduces a proxy-aligned co-evolution mechanism to maintain two evolving proxy caches, which dynamically mines contextual textual negatives guided by test images and iteratively refines visual proxies, progressively realigning cross-modal similarities and enlarging local OOD margins. Finally, we dynamically re-weight the contributions of dual-modal proxies to obtain a calibrated OOD score that is robust to distribution shift. Extensive experiments on standard benchmarks demonstrate that CoEvo achieves state-of-the-art performance, improving AUROC by 1.33% and reducing FPR95 by 45.98% on ImageNet-1K compared to strong negative-label baselines.

Method: Combines deformation guidance using 3D Gaussian Splatting for consistent geometric modifications with diffusion guidance for content correction and visual quality enhancement. Uses a progressive editing strategy to support aggressive 3D drag edits.

Result: Achieves state-of-the-art performance in geometry-related 3D content editing, enabling various edits including motion change, shape adjustment, inpainting, and content extension. Editing is efficient, taking 10-20 minutes on a single RTX 4090 GPU.

Conclusion: 3DGS-Drag successfully bridges the gap between deformation-based and 2D-editing-based 3D editing methods, providing an efficient and intuitive framework for drag manipulation of real 3D scenes with wide editing capabilities.

Abstract: The transformative potential of 3D content creation has been progressively unlocked through advancements in generative models. Recently, intuitive drag editing with geometric changes has attracted significant attention in 2D editing yet remains challenging for 3D scenes. In this paper, we introduce 3DGS-Drag – a point-based 3D editing framework that provides efficient, intuitive drag manipulation of real 3D scenes. Our approach bridges the gap between deformation-based and 2D-editing-based 3D editing methods, addressing their limitations to geometry-related content editing. We leverage two key innovations: deformation guidance utilizing 3D Gaussian Splatting for consistent geometric modifications and diffusion guidance for content correction and visual quality enhancement. A progressive editing strategy further supports aggressive 3D drag edits. Our method enables a wide range of edits, including motion change, shape adjustment, inpainting, and content extension. Experimental results demonstrate the effectiveness of 3DGS-Drag in various scenes, achieving state-of-the-art performance in geometry-related 3D content editing. Notably, the editing is efficient, taking 10 to 20 minutes on a single RTX 4090 GPU.

Method: Proposed two feature-combination pipelines integrating Gabor-based features (structural) and GLCM-based texture features. Evaluated using simulated digital breast tomosynthesis images and compared with threshold-based model observer and human eye-tracking data.

Result: Qualitative agreement between predicted fixation candidates and model observer; strong correlation between GLCM mean and Gabor feature responses; consistency between predicted fixation regions and early-stage human gaze behavior.

Conclusion: Combining structural and texture-based features is valuable for modeling visual search behavior, supporting development of perceptually informed observer models for medical imaging applications.

Abstract: Understanding human visual search behavior is a fundamental problem in vision science and computer vision, with direct implications for modeling how observers allocate attention in location-unknown search tasks. In this study, we investigate the relationship between Gabor-based features and gray-level co-occurrence matrix (GLCM) based texture features in modeling early-stage visual search behavior. Two feature-combination pipelines are proposed to integrate Gabor and GLCM features for narrowing the region of possible human fixations. The pipelines are evaluated using simulated digital breast tomosynthesis images. Results show qualitative agreement among fixation candidates predicted by the proposed pipelines and a threshold-based model observer. A strong correlation is observed between GLCM mean and Gabor feature responses, indicating that these features encode related image information despite their different formulations. Eye-tracking data from human observers further suggest consistency between predicted fixation regions and early-stage gaze behavior. These findings highlight the value of combining structural and texture-based features for modeling visual search and support the development of perceptually informed observer models.

Method: Created a dataset of 206 training, 43 validation, and 43 test images collected from Nigerian farms using high-resolution mobile cameras, annotated with Segment Anything Model v2 under farmer supervision in YOLO-compatible segmentation format.

Result: YOLOv8 evaluation showed strong performance: detection achieved 79% recall, 79% precision, 84% mAP@0.50, 58% mAP@0.50:0.95; segmentation achieved 82% recall, 77% precision, 84% mAP@0.50, 52% mAP@0.50:0.95.

Conclusion: The dataset represents a novel contribution to sesame-focused agricultural vision in Nigeria, supporting plant monitoring, yield estimation, and agricultural research with precise pixel-level segmentation capabilities.

Abstract: This paper presents the Sesame Plant Segmentation Dataset, an open source annotated image dataset designed to support the development of artificial intelligence models for agricultural applications, with a specific focus on sesame plants. The dataset comprises 206 training images, 43 validation images, and 43 test images in YOLO compatible segmentation format, capturing sesame plants at early growth stages under varying environmental conditions. Data were collected using a high resolution mobile camera from farms in Jirdede, Daura Local Government Area, Katsina State, Nigeria, and annotated using the Segment Anything Model version 2 with farmer supervision. Unlike conventional bounding box datasets, this dataset employs pixel level segmentation to enable more precise detection and analysis of sesame plants in real world farm settings. Model evaluation using the Ultralytics YOLOv8 framework demonstrated strong performance for both detection and segmentation tasks. For bounding box detection, the model achieved a recall of 79 percent, precision of 79 percent, mean average precision at IoU 0.50 of 84 percent, and mean average precision from 0.50 to 0.95 of 58 percent. For segmentation, it achieved a recall of 82 percent, precision of 77 percent, mean average precision at IoU 0.50 of 84 percent, and mean average precision from 0.50 to 0.95 of 52 percent. The dataset represents a novel contribution to sesame focused agricultural vision datasets in Nigeria and supports applications such as plant monitoring, yield estimation, and agricultural research.

Method: Motive uses gradient-based data attribution scaled for modern video datasets/models, isolates temporal dynamics from static appearance via motion-weighted loss masks, and enables efficient motion-specific influence computation.

Result: Motive identifies clips that strongly affect motion, guides data curation improving temporal consistency and physical plausibility. With Motive-selected data, achieves 74.1% human preference win rate, improves motion smoothness and dynamic degree on VBench.

Conclusion: Motive is the first framework to attribute motion rather than visual appearance in video generative models and use it to curate fine-tuning data, providing a systematic approach to understand and improve motion quality through targeted data selection.

Abstract: Despite the rapid progress of video generation models, the role of data in influencing motion is poorly understood. We present Motive (MOTIon attribution for Video gEneration), a motion-centric, gradient-based data attribution framework that scales to modern, large, high-quality video datasets and models. We use this to study which fine-tuning clips improve or degrade temporal dynamics. Motive isolates temporal dynamics from static appearance via motion-weighted loss masks, yielding efficient and scalable motion-specific influence computation. On text-to-video models, Motive identifies clips that strongly affect motion and guides data curation that improves temporal consistency and physical plausibility. With Motive-selected high-influence data, our method improves both motion smoothness and dynamic degree on VBench, achieving a 74.1% human preference win rate compared with the pretrained base model. To our knowledge, this is the first framework to attribute motion rather than visual appearance in video generative models and to use it to curate fine-tuning data.

Method: Proposes a subject decoupling framework that extracts driver appearance embeddings and removes their influence from image embeddings before zero-shot classification. Also orthogonalizes text embeddings via metric projection onto Stiefel manifold to improve separability while preserving original semantics.

Result: Experiments demonstrate consistent gains over prior baselines, showing improved performance for distracted driver detection in practical road-safety applications.

Conclusion: The proposed subject decoupling approach effectively addresses appearance bias in VLM-based distracted driver detection, promising more robust and practical road-safety applications.

Abstract: Distracted driving is a major cause of traffic collisions, calling for robust and scalable detection methods. Vision-language models (VLMs) enable strong zero-shot image classification, but existing VLM-based distracted driver detectors often underperform in real-world conditions. We identify subject-specific appearance variations (e.g., clothing, age, and gender) as a key bottleneck: VLMs entangle these factors with behavior cues, leading to decisions driven by who the driver is rather than what the driver is doing. To address this, we propose a subject decoupling framework that extracts a driver appearance embedding and removes its influence from the image embedding prior to zero-shot classification, thereby emphasizing distraction-relevant evidence. We further orthogonalize text embeddings via metric projection onto Stiefel manifold to improve separability while staying close to the original semantics. Experiments demonstrate consistent gains over prior baselines, indicating the promise of our approach for practical road-safety applications.

Method: Propose Additive Kolmogorov-Arnold Transformer (AKT) with two key innovations: (1) replaces MLPs with Pade Kolmogorov-Arnold Network (PKAN) modules for enhanced functional expressivity in small-object feature extraction, and (2) introduces PKAN Additive Attention (PAA) to model multiscale spatial dependencies with reduced computational complexity. Also present Point-based Maize Localization (PML) dataset with 1,928 high-resolution UAV images and ~501,000 point annotations.

Result: AKT achieves average F1-score of 62.8%, outperforming SOTA by 4.2%, while reducing FLOPs by 12.6% and improving inference throughput by 20.7%. For downstream tasks: mean absolute error of 7.1 in stand counting, and root mean square error of 1.95-1.97 cm in interplant spacing estimation.

Conclusion: Integrating Kolmogorov-Arnold representation theory with efficient attention mechanisms provides an effective framework for high-resolution agricultural remote sensing, demonstrating practical value for precision agriculture applications.

Abstract: High-resolution UAV photogrammetry has become a key technology for precision agriculture, enabling centimeter-level crop monitoring and point-level plant localization. However, point-level maize localization in UAV imagery remains challenging due to (1) extremely small object-to-pixel ratios, typically less than 0.1%, (2) prohibitive computational costs of quadratic attention on ultra-high-resolution images larger than 3000 x 4000 pixels, and (3) agricultural scene-specific complexities such as sparse object distribution and environmental variability that are poorly handled by general-purpose vision models. To address these challenges, we propose the Additive Kolmogorov-Arnold Transformer (AKT), which replaces conventional multilayer perceptrons with Pade Kolmogorov-Arnold Network (PKAN) modules to enhance functional expressivity for small-object feature extraction, and introduces PKAN Additive Attention (PAA) to model multiscale spatial dependencies with reduced computational complexity. In addition, we present the Point-based Maize Localization (PML) dataset, consisting of 1,928 high-resolution UAV images with approximately 501,000 point annotations collected under real field conditions. Extensive experiments show that AKT achieves an average F1-score of 62.8%, outperforming state-of-the-art methods by 4.2%, while reducing FLOPs by 12.6% and improving inference throughput by 20.7%. For downstream tasks, AKT attains a mean absolute error of 7.1 in stand counting and a root mean square error of 1.95-1.97 cm in interplant spacing estimation. These results demonstrate that integrating Kolmogorov-Arnold representation theory with efficient attention mechanisms offers an effective framework for high-resolution agricultural remote sensing.

Method: Develops a statistical ideal observer model that performs holistic visual search by placing thresholds on minimum extractable image features, thereby reducing the number of free parameters and shrinking the system.

Result: A novel framework with reduced parameter complexity that enables holistic visual search processing through feature thresholding, creating a more efficient ideal observer model.

Conclusion: The proposed ideal observer model provides a foundation for applications in medical image perception, computer vision, performance benchmarking, feature evaluation, target detection/recognition, and sensor/detector assessment across multiple domains.

Abstract: We develop a new statistical ideal observer model that performs holistic visual search (or gist) processing in part by placing thresholds on minimum extractable image features. In this model, the ideal observer reduces the number of free parameters thereby shrinking down the system. The applications of this novel framework is in medical image perception (for optimizing imaging systems and algorithms), computer vision, benchmarking performance and enabling feature selection/evaluations. Other applications are in target detection and recognition in defense/security as well as evaluating sensors and detectors.

Method: Three-axis approach: 1) Single-tower architecture with unified DiT blocks and Omni-Full Attention for tight alignment; 2) Progressive multitask training with random modality masking and multistage curriculum; 3) Novel automated pipeline for creating large-scale dense-caption audio-video dataset.

Result: Achieves high-fidelity, semantically/temporally aligned generation in both joint and unimodal settings, robust generalization to out-of-distribution scenarios, substantially outperforms prior methods by large margin, and reaches performance comparable to Veo 3.

Conclusion: Apollo offers a unified, scalable path toward next-generation audio-video synthesis by addressing core challenges through integrated architectural, training, and data innovations.

Abstract: Audio-video joint generation has progressed rapidly, yet substantial challenges still remain. Non-commercial approaches still suffer audio-visual asynchrony, poor lip-speech alignment, and unimodal degradation, which can be stemmed from weak audio-visual correspondence modeling, limited generalization, and scarce high-quality dense-caption data. To address these issues, we introduce Apollo and delve into three axes–model architecture, training strategy, and data curation. Architecturally, we adopt a single-tower design with unified DiT blocks and an Omni-Full Attention mechanism, achieving tight audio-visual alignment and strong scalability. Training-wise, we adopt a progressive multitask regime–random modality masking to joint optimization across tasks, and a multistage curriculum, yielding robust representations, strengthening A-V aligned world knowledge, and preventing unimodal collapse. For datasets, we present the first large-scale audio-video dataset with dense captions, and introduce a novel automated data-construction pipeline which annotates and filters millions of diverse, high-quality, strictly aligned audio-video-caption triplets. Building on this, Apollo scales to large datasets, delivering high-fidelity, semantically and temporally aligned, instruction-following generation in both joint and unimodal settings while generalizing robustly to out-of-distribution scenarios. Across tasks, it substantially outperforms prior methods by a large margin and achieves performance comparable to Veo 3, offering a unified, scalable path toward next-generation audio-video synthesis.

Method: 1) Dual Discrete Diffusion (D3Diff) loss unifying masked generative models with discrete score matching diffusion; 2) Multi-level grouped Mixture-of-Experts architecture incorporating semantic and identity facial embeddings; 3) Construction of UniF²aceD-1M dataset with 130K image-caption pairs and 1M VQA pairs.

Result: UniF²ace outperforms existing models with similar scale: 7.1% higher Desc-GPT score and 6.6% higher VQA-score in understanding and generation tasks, demonstrating superior fine-grained facial attribute handling.

Conclusion: UniF²ace successfully addresses fragmentation and fine-grained attribute challenges in face AI, establishing a unified multimodal framework that advances both understanding and generation capabilities for high-fidelity facial applications.

Abstract: Unified multimodal models (UMMs) have emerged as a powerful paradigm in fundamental cross-modality research, demonstrating significant potential in both image understanding and generation. However, existing research in the face domain primarily faces two challenges: $\textbf{(1)}$ $\textbf{fragmentation development}$, with existing methods failing to unify understanding and generation into a single one, hindering the way to artificial general intelligence. $\textbf{(2) lack of fine-grained facial attributes}$, which are crucial for high-fidelity applications. To handle those issues, we propose $\textbf{UniF$^2$ace}$, $\textit{the first UMM specifically tailored for fine-grained face understanding and generation}$. $\textbf{First}$, we introduce a novel theoretical framework with a Dual Discrete Diffusion (D3Diff) loss, unifying masked generative models with discrete score matching diffusion and leading to a more precise approximation of the negative log-likelihood. Moreover, this D3Diff significantly enhances the model’s ability to synthesize high-fidelity facial details aligned with text input. $\textbf{Second}$, we propose a multi-level grouped Mixture-of-Experts architecture, adaptively incorporating the semantic and identity facial embeddings to complement the attribute forgotten phenomenon in representation evolvement. $\textbf{Finally}$, to this end, we construct UniF$^2$aceD-1M, a large-scale dataset comprising 130K fine-grained image-caption pairs and 1M visual question-answering pairs, spanning a much wider range of facial attributes than existing datasets. Extensive experiments demonstrate that UniF$^2$ace outperforms existing models with a similar scale in both understanding and generation tasks, with 7.1% higher Desc-GPT and 6.6% higher VQA-score, respectively.

Method: Two approaches: (1) CASHEW - inference-time framework that aggregates multiple candidate trajectories with visual verification to filter hallucinations, (2) CASHEW-RL - learned variant trained with Group Sequence Policy Optimization (GSPO) using composite reward for correct answers grounded in minimal sufficient visual evidence.

Result: Significant performance improvements across 13 image/video understanding and reasoning benchmarks, including +23.6pp on ScienceQA and +8.1pp on EgoSchema.

Conclusion: The proposed methods effectively stabilize vision-language model reasoning by aggregating trajectories and grounding in visual evidence, enabling more consistent and accurate multimodal reasoning.

Abstract: Vision-language models achieve strong performance across a wide range of multimodal understanding and reasoning tasks, yet their multi-step reasoning remains unstable. Repeated sampling over the same input often produces divergent reasoning trajectories and inconsistent final predictions. To address this, we introduce two complementary approaches inspired by test-time scaling: (1) CASHEW, an inference-time framework that stabilizes reasoning by iteratively aggregating multiple candidate trajectories into higher-quality reasoning traces, with explicit visual verification filtering hallucinated steps and grounding reasoning in visual evidence, and (2) CASHEW-RL, a learned variant that internalizes this aggregation behavior within a single model. CASHEW-RL is trained using Group Sequence Policy Optimization (GSPO) with a composite reward that encourages correct answers grounded in minimal yet sufficient visual evidence, while adaptively allocating reasoning effort based on task difficulty. This training objective enables robust self-aggregation at inference. Extensive experiments on 13 image understanding, video understanding, and video reasoning benchmarks show significant performance improvements, including gains of up to +23.6 percentage points on ScienceQA and +8.1 percentage points on EgoSchema.

Method: Two main components: 1) “MisCaption This!” - adversarial prompting of VLMs to generate realistic miscaptioned datasets; 2) LAMAR - Transformer-based network that reconstructs truthful caption embeddings as auxiliary signal for detection.

Result: Models trained on “MisCaption This!” data generalize better to real-world misinformation. LAMAR achieves state-of-the-art on NewsCLIPpings, VERITE, and VERITE 24/25 benchmarks.

Conclusion: VLM-generated synthetic data and reconstruction-based networks are effective for advancing multimodal misinformation detection, offering better generalization to real-world cases.

Abstract: Multimodal misinformation, such as miscaptioned images, where captions misrepresent an image’s origin, context, or meaning, poses a growing challenge in the digital age. Due to the scarcity of large-scale annotated datasets for multimodal misinformation detection (MMD), recent approaches rely on synthetic training data created via out-of-context pairings or named entity manipulations (e.g., altering names, dates, or locations). However, these often yield simplistic, unrealistic examples, which limits their utility as training examples. To address this, we introduce “MisCaption This!”, a framework for generating high-fidelity synthetic miscaptioned datasets through Adversarial Prompting of Vision-Language Models (VLMs). Additionally, we introduce “Latent Multimodal Reconstruction” (LAMAR), a Transformer-based network trained to reconstruct the embeddings of truthful captions, providing a strong auxiliary signal to guide detection. We explore various training strategies (end-to-end vs. large-scale pre-training) and integration mechanisms (direct, mask, gate, and attention). Extensive experiments show that models trained on “MisCaption This!” data generalize better to real-world misinformation, while LAMAR achieves new state-of-the-art on NewsCLIPpings, VERITE, and the newly introduced VERITE 24/25 benchmark; highlighting the efficacy of VLM-generated data and reconstruction-based networks for advancing MMD. Our code is available at https://github.com/stevejpapad/miscaptioned-image-reconstruction

Method: A novel decoder-based deep learning framework that decodes latent representations into printable 3D objects. The neural decoders learn complex mapping functions from abstract representations to valid 3D geometries while incorporating manufacturing constraints.

Result: The approach produces parts with significantly improved manufacturability compared to naive generation methods. It was validated on diverse object categories and demonstrated practical 3D printing of decoder-generated structures.

Conclusion: Neural decoders can effectively learn to generate 3D structures optimized for additive manufacturing, bridging the gap between abstract design representations and physically manufacturable objects.

Abstract: This paper presents a novel decoder-based approach for generating manufacturable 3D structures optimized for additive manufacturing. We introduce a deep learning framework that decodes latent representations into geometrically valid, printable objects while respecting manufacturing constraints such as overhang angles, wall thickness, and structural integrity. The methodology demonstrates that neural decoders can learn complex mapping functions from abstract representations to valid 3D geometries, producing parts with significantly improved manufacturability compared to naive generation approaches. We validate the approach on diverse object categories and demonstrate practical 3D printing of decoder-generated structures.

Method: Uses a synthesis-based approach inspired by DeepDream: extracts concept vectors for textual concepts at specific layers, then optimizes image synthesis to align with those vectors. Applied to hundreds of concepts across seven layers in Gemma 3 model.

Result: Synthesized images depict salient visual features of targeted concepts from layer 1 onward (e.g., >50% of images show recognizable features of animals, activities, or seasons at layer 1). Provides direct constructive evidence of early image-text alignment.

Conclusion: Image-text alignment in adapter-based vision-language models begins from the first layer, not just late layers. The method offers a simple, fast approach for measuring multimodal alignment without auxiliary models/datasets and provides new interpretability tools for visualizing model representations.

Abstract: We show that for a variety of concepts in adapter-based vision-language models, the representations of their images and their text descriptions are meaningfully aligned from the very first layer. This contradicts the established view that such image-text alignment only appears in late layers. We show this using a new synthesis-based method inspired by DeepDream: given a textual concept such as “Jupiter”, we extract its concept vector at a given layer, and then use optimisation to synthesise an image whose representation aligns with that vector. We apply our approach to hundreds of concepts across seven layers in Gemma 3, and find that the synthesised images often depict salient visual features of the targeted textual concepts: for example, already at layer 1, more than 50 % of images depict recognisable features of animals, activities, or seasons. Our method thus provides direct, constructive evidence of image-text alignment on a concept-by-concept and layer-by-layer basis. Unlike previous methods for measuring multimodal alignment, our approach is simple, fast, and does not require auxiliary models or datasets. It also offers a new path towards model interpretability, by providing a way to visualise a model’s representation space by backtracing through its image processing components.

Method: Uses DDIM inversion to obtain latent representations from input images, then initiates denoising from a fixed intermediate timestep to reconstruct normal-looking images. The diffusion model is trained only on normal data, so reconstruction highlights anomalies through discrepancy with input.

Result: Achieves state-of-the-art performance on VISA dataset, demonstrating strong localization capabilities without auxiliary modalities, outperforming previous methods.

Conclusion: Proposes a simple yet effective training-free vision-only framework that eliminates prompt dependence for zero-shot anomaly detection, enabling precise spatial localization using only visual information.

Abstract: Zero-Shot image Anomaly Detection (ZSAD) aims to detect and localise anomalies without access to any normal training samples of the target data. While recent ZSAD approaches leverage additional modalities such as language to generate fine-grained prompts for localisation, vision-only methods remain limited to image-level classification, lacking spatial precision. In this work, we introduce a simple yet effective training-free vision-only ZSAD framework that circumvents the need for fine-grained prompts by leveraging the inversion of a pretrained Denoising Diffusion Implicit Model (DDIM). Specifically, given an input image and a generic text description (e.g., “an image of an [object class]”), we invert the image to obtain latent representations and initiate the denoising process from a fixed intermediate timestep to reconstruct the image. Since the underlying diffusion model is trained solely on normal data, this process yields a normal-looking reconstruction. The discrepancy between the input image and the reconstructed one highlights potential anomalies. Our method achieves state-of-the-art performance on VISA dataset, demonstrating strong localisation capabilities without auxiliary modalities and facilitating a shift away from prompt dependence for zero-shot anomaly detection research. Code is available at https://github.com/giddyyupp/DIVAD.

Method: A cascading multi-agent framework with early modules for reconstruction-gated filtering and object-level assessment, plus higher-level reasoning agents selectively invoked for ambiguous events. Uses adaptive escalation thresholds and publish-subscribe communication for asynchronous coordination across heterogeneous hardware.

Result: Achieves 3x reduction in latency compared to direct vision-language inference while maintaining high perceptual fidelity (PSNR = 38.3 dB, SSIM = 0.965) and consistent semantic labeling on large-scale monitoring data.

Conclusion: The framework advances beyond conventional detection pipelines by combining early-exit efficiency, adaptive multi-agent reasoning, and explainable anomaly attribution, establishing a reproducible and energy-efficient foundation for scalable intelligent visual monitoring.

Abstract: Intelligent anomaly detection in dynamic visual environments requires reconciling real-time performance with semantic interpretability. Conventional approaches address only fragments of this challenge. Reconstruction-based models capture low-level deviations without contextual reasoning, object detectors provide speed but limited semantics, and large vision-language systems deliver interpretability at prohibitive computational cost. This work introduces a cascading multi-agent framework that unifies these complementary paradigms into a coherent and interpretable architecture. Early modules perform reconstruction-gated filtering and object-level assessment, while higher-level reasoning agents are selectively invoked to interpret semantically ambiguous events. The system employs adaptive escalation thresholds and a publish-subscribe communication backbone, enabling asynchronous coordination and scalable deployment across heterogeneous hardware. Extensive evaluation on large-scale monitoring data demonstrates that the proposed cascade achieves a threefold reduction in latency compared to direct vision-language inference, while maintaining high perceptual fidelity (PSNR = 38.3 dB, SSIM = 0.965) and consistent semantic labeling. The framework advances beyond conventional detection pipelines by combining early-exit efficiency, adaptive multi-agent reasoning, and explainable anomaly attribution, establishing a reproducible and energy-efficient foundation for scalable intelligent visual monitoring.

Method: Proposes Concept-Based Diversity (CBD) using Vision-Language Models to efficiently measure image diversity, combined with Margin uncertainty metric in a hybrid selection approach.

Result: CBD strongly correlates with established Geometric Diversity but is much faster. CBD-based selection consistently outperforms 5 state-of-the-art baselines across diverse models and datasets while maintaining efficiency close to simple uncertainty methods.

Conclusion: CBD provides an effective, computationally efficient, and scalable solution for input selection in DNN fine-tuning, making diversity-based approaches practical for real-world applications even on large datasets like ImageNet.

Abstract: Maintaining or improving the performance of Deep Neural Networks (DNNs) through fine-tuning requires labeling newly collected inputs, a process that is often costly and time-consuming. To alleviate this problem, input selection approaches have been developed in recent years to identify small, yet highly informative subsets for labeling. Diversity-based selection is one of the most effective approaches for this purpose. However, they are often computationally intensive and lack scalability for large input sets, limiting their practical applicability. To address this challenge, we introduce Concept-Based Diversity (CBD), a highly efficient metric for image inputs that leverages Vision-Language Models (VLM). Our results show that CBD exhibits a strong correlation with Geometric Diversity (GD), an established diversity metric, while requiring only a fraction of its computation time. Building on this finding, we propose a hybrid input selection approach that combines CBD with Margin, a simple uncertainty metric. We conduct a comprehensive evaluation across a diverse set of DNN models, input sets, selection budgets, and five most effective state-of-the-art selection baselines. The results demonstrate that the CBD-based selection consistently outperforms all baselines at guiding input selection to improve the DNN model. Furthermore, the CBD-based selection approach remains highly efficient, requiring selection times close to those of simple uncertainty-based methods such as Margin, even on larger input sets like ImageNet. These results confirm not only the effectiveness and computational advantage of the CBD-based approach, particularly compared to hybrid baselines, but also its scalability in repetitive and extensive input selection scenarios.

Method: Proposes FigEx2 framework with noise-aware gated fusion module to filter token-level features and stabilize detection query space. Uses staged optimization combining supervised learning with RL, utilizing CLIP-based alignment and BERTScore-based semantic rewards for multimodal consistency.

Result: Achieves 0.726 mAP@0.5:0.95 for detection and significantly outperforms Qwen3-VL-8B by 0.51 in METEOR and 0.24 in BERTScore. Shows remarkable zero-shot transferability to out-of-distribution scientific domains without fine-tuning.

Conclusion: FigEx2 effectively addresses panel-level understanding of compound scientific figures through visual-conditioned localization and captioning with strong performance and cross-domain transferability.

Abstract: Scientific compound figures combine multiple labeled panels into a single image, but captions in real pipelines are often missing or only provide figure-level summaries, making panel-level understanding difficult. In this paper, we propose FigEx2, visual-conditioned framework that localizes panels and generates panel-wise captions directly from the compound figure. To mitigate the impact of diverse phrasing in open-ended captioning, we introduce a noise-aware gated fusion module that adaptively filters token-level features to stabilize the detection query space. Furthermore, we employ a staged optimization strategy combining supervised learning with reinforcement learning (RL), utilizing CLIP-based alignment and BERTScore-based semantic rewards to enforce strict multimodal consistency. To support high-quality supervision, we curate BioSci-Fig-Cap, a refined benchmark for panel-level grounding, alongside cross-disciplinary test suites in physics and chemistry. Experimental results demonstrate that FigEx2 achieves a superior 0.726 mAP@0.5:0.95 for detection and significantly outperforms Qwen3-VL-8B by 0.51 in METEOR and 0.24 in BERTScore. Notably, FigEx2 exhibits remarkable zero-shot transferability to out-of-distribution scientific domains without any fine-tuning.

Method: Combines diffusion-based synthesis with vision-language prompting for realistic, semantically controlled manipulations (duplication, splicing, region removal). Introduces Rescind benchmark with fine-grained annotations and modality-specific splits, and proposes Integscan framework using structured state space modeling with attention-enhanced visual encoding and prompt-conditioned semantic alignment for forgery localization. Includes vision-language model verification loop for semantic fidelity.

Result: Extensive experiments on Rescind and existing benchmarks show Integscan achieves state-of-the-art performance in both detection and localization of biomedical image forgeries.

Conclusion: Establishes a strong foundation for automated scientific integrity analysis in biomedical publications through the first comprehensive vision-language guided framework for biomedical forgery generation and detection.

Abstract: Scientific image manipulation in biomedical publications poses a growing threat to research integrity and reproducibility. Unlike natural image forensics, biomedical forgery detection is uniquely challenging due to domain-specific artifacts, complex textures, and unstructured figure layouts. We present the first vision-language guided framework for both generating and detecting biomedical image forgeries. By combining diffusion-based synthesis with vision-language prompting, our method enables realistic and semantically controlled manipulations, including duplication, splicing, and region removal, across diverse biomedical modalities. We introduce Rescind, a large-scale benchmark featuring fine-grained annotations and modality-specific splits, and propose Integscan, a structured state space modeling framework that integrates attention-enhanced visual encoding with prompt-conditioned semantic alignment for precise forgery localization. To ensure semantic fidelity, we incorporate a vision-language model based verification loop that filters generated forgeries based on consistency with intended prompts. Extensive experiments on Rescind and existing benchmarks demonstrate that Integscan achieves state of the art performance in both detection and localization, establishing a strong foundation for automated scientific integrity analysis.

Method: Used CIFAR-10 dataset with Resnet-18 model, introduced three types of controlled noise (Gaussian noise, Salt-and-Pepper noise, Gaussian blur) at varying intensities and training set pollution levels (up to 10%).

Result: Incorporating just 10% noisy data during training significantly reduces test loss and improves accuracy under fully corrupted test conditions, while maintaining similar performance on clean data.

Conclusion: Strategic exposure to noise during training acts as an effective regularizer, offering a practical trade-off between data cleanliness and real-world resilience for CNN image classification.

Abstract: Data quality plays a central role in the performance and robustness of convolutional neural networks (CNNs) for image classification. While high-quality data is often preferred for training, real-world inputs are frequently affected by noise and other distortions. This paper investigates the effect of deliberately introducing controlled noise into the training data to improve model robustness. Using the CIFAR-10 dataset, we evaluate the impact of three common corruptions, namely Gaussian noise, Salt-and-Pepper noise, and Gaussian blur at varying intensities and training set pollution levels. Experiments using a Resnet-18 model reveal that incorporating just 10% noisy data during training is sufficient to significantly reduce test loss and enhance accuracy under fully corrupted test conditions, with minimal impact on clean-data performance. These findings suggest that strategic exposure to noise can act as a simple yet effective regularizer, offering a practical trade-off between traditional data cleanliness and real-world resilience.

Method: Proposes DINO-AugSeg with two key modules: 1) WT-Aug - wavelet-based feature-level augmentation that perturbs frequency components to enrich DINOv3 feature diversity, and 2) CG-Fuse - contextual information-guided fusion module using cross-attention to integrate semantic-rich low-resolution features with spatially detailed high-resolution features.

Result: Extensive experiments on six public benchmarks across five imaging modalities (MRI, CT, ultrasound, endoscopy, dermoscopy) show DINO-AugSeg consistently outperforms existing methods under limited-sample conditions.

Conclusion: The framework demonstrates effectiveness of wavelet-domain augmentation and contextual fusion for robust feature representation, offering a promising direction for advancing few-shot medical image segmentation. Code and data will be publicly available.

Abstract: Deep learning-based automatic medical image segmentation plays a critical role in clinical diagnosis and treatment planning but remains challenging in few-shot scenarios due to the scarcity of annotated training data. Recently, self-supervised foundation models such as DINOv3, which were trained on large natural image datasets, have shown strong potential for dense feature extraction that can help with the few-shot learning challenge. Yet, their direct application to medical images is hindered by domain differences. In this work, we propose DINO-AugSeg, a novel framework that leverages DINOv3 features to address the few-shot medical image segmentation challenge. Specifically, we introduce WT-Aug, a wavelet-based feature-level augmentation module that enriches the diversity of DINOv3-extracted features by perturbing frequency components, and CG-Fuse, a contextual information-guided fusion module that exploits cross-attention to integrate semantic-rich low-resolution features with spatially detailed high-resolution features. Extensive experiments on six public benchmarks spanning five imaging modalities, including MRI, CT, ultrasound, endoscopy, and dermoscopy, demonstrate that DINO-AugSeg consistently outperforms existing methods under limited-sample conditions. The results highlight the effectiveness of incorporating wavelet-domain augmentation and contextual fusion for robust feature representation, suggesting DINO-AugSeg as a promising direction for advancing few-shot medical image segmentation. Code and data will be made available on https://github.com/apple1986/DINO-AugSeg.

Method: Three-stage framework: 1) Synthesize target objects within domain-specific backgrounds via controlled inpainting, 2) Validate outputs through multi-modal assessment (object detection, aesthetic scoring, vision-language alignment), 3) Employ user-preference classifier to capture subjective selection criteria.

Result: Enables efficient construction of high-quality, deployable datasets with reduced need for real-world data collection.

Conclusion: The proposed pipeline provides an effective solution for generating domain-specific synthetic datasets that bridge the gap between pre-trained models and real-world deployment requirements.

Abstract: In this paper, we present an automated pipeline for generating domain-specific synthetic datasets with diffusion models, addressing the distribution shift between pre-trained models and real-world deployment environments. Our three-stage framework first synthesizes target objects within domain-specific backgrounds through controlled inpainting. The generated outputs are then validated via a multi-modal assessment that integrates object detection, aesthetic scoring, and vision-language alignment. Finally, a user-preference classifier is employed to capture subjective selection criteria. This pipeline enables the efficient construction of high-quality, deployable datasets while reducing reliance on extensive real-world data collection.

Method: PathoGen uses a diffusion-based generative model for controllable, high-fidelity inpainting of lesions into benign histopathology images. It leverages the iterative refinement process of diffusion models to synthesize lesions with natural tissue boundaries, preserved cellular structures, and authentic staining characteristics.

Result: PathoGen outperforms state-of-the-art generative baselines (conditional GAN and Stable Diffusion) in image fidelity and distributional similarity across four diverse datasets (kidney, skin, breast, and prostate pathology). Augmenting training sets with PathoGen-synthesized lesions enhances downstream segmentation performance compared to traditional geometric augmentations, especially in data-scarce regimes.

Conclusion: PathoGen effectively overcomes the manual annotation bottleneck by simultaneously generating realistic morphology and pixel-level ground truth. This approach offers a scalable pathway for developing generalizable medical AI systems despite limited expert-labeled data.

Abstract: The development of robust artificial intelligence models for histopathology diagnosis is severely constrained by the scarcity of expert-annotated lesion data, particularly for rare pathologies and underrepresented disease subtypes. While data augmentation offers a potential solution, existing methods fail to generate sufficiently realistic lesion morphologies that preserve the complex spatial relationships and cellular architectures characteristic of histopathological tissues. Here we present PathoGen, a diffusion-based generative model that enables controllable, high-fidelity inpainting of lesions into benign histopathology images. Unlike conventional augmentation techniques, PathoGen leverages the iterative refinement process of diffusion models to synthesize lesions with natural tissue boundaries, preserved cellular structures, and authentic staining characteristics. We validate PathoGen across four diverse datasets representing distinct diagnostic challenges: kidney, skin, breast, and prostate pathology. Quantitative assessment confirms that PathoGen outperforms state-of-the-art generative baselines, including conditional GAN and Stable Diffusion, in image fidelity and distributional similarity. Crucially, we show that augmenting training sets with PathoGen-synthesized lesions enhances downstream segmentation performance compared to traditional geometric augmentations, particularly in data-scarce regimes. Besides, by simultaneously generating realistic morphology and pixel-level ground truth, PathoGen effectively overcomes the manual annotation bottleneck. This approach offers a scalable pathway for developing generalizable medical AI systems despite limited expert-labeled data.

Method: Proposes SSP framework with optical flow for motion dynamics and two textual prompts (object category + scene description). Uses pre-mask technique with optical flow for temporal context, visual-textual alignment module for cross-modal integration, and post-mask training to learn optical flow patterns.

Result: SSP outperforms existing AVS methods, delivering efficient and precise segmentation results with better handling of both moving and stationary sound-emitting objects.

Conclusion: The SSP framework successfully addresses challenges in audio-visual semantic segmentation by integrating optical flow and textual prompts, providing a robust solution for complex audio-visual scenes with both moving and stationary sound sources.

Abstract: Audio-visual semantic segmentation (AVSS) represents an extension of the audio-visual segmentation (AVS) task, necessitating a semantic understanding of audio-visual scenes beyond merely identifying sound-emitting objects at the visual pixel level. Contrary to a previous methodology, by decomposing the AVSS task into two discrete subtasks by initially providing a prompted segmentation mask to facilitate subsequent semantic analysis, our approach innovates on this foundational strategy. We introduce a novel collaborative framework, \textit{S}tepping \textit{S}tone \textit{P}lus (SSP), which integrates optical flow and textual prompts to assist the segmentation process. In scenarios where sound sources frequently coexist with moving objects, our pre-mask technique leverages optical flow to capture motion dynamics, providing essential temporal context for precise segmentation. To address the challenge posed by stationary sound-emitting objects, such as alarm clocks, SSP incorporates two specific textual prompts: one identifies the category of the sound-emitting object, and the other provides a broader description of the scene. Additionally, we implement a visual-textual alignment module (VTA) to facilitate cross-modal integration, delivering more coherent and contextually relevant semantic interpretations. Our training regimen involves a post-mask technique aimed at compelling the model to learn the diagram of the optical flow. Experimental results demonstrate that SSP outperforms existing AVS methods, delivering efficient and precise segmentation results.

Method: Aligns visual and textual subspaces by minimizing chordal distance, then projects visual features onto task-specific textual subspace to filter noise before performing standard TTA.

Result: Achieves average 2.24% improvement over state-of-the-art TTA methods across various benchmarks and VLM architectures.

Conclusion: SubTTA effectively addresses modality gaps and visual nuisance to enhance VLM adaptation under distribution shifts through subspace alignment and purification.

Abstract: Vision-language models (VLMs), despite their extraordinary zero-shot capabilities, are vulnerable to distribution shifts. Test-time adaptation (TTA) emerges as a predominant strategy to adapt VLMs to unlabeled test data on the fly. However, existing TTA methods heavily rely on zero-shot predictions as pseudo-labels for self-training, which can be unreliable under distribution shifts and misguide adaptation due to two fundamental limitations. First (Modality Gap), distribution shifts induce gaps between visual and textual modalities, making cross-modal relations inaccurate. Second (Visual Nuisance), visual embeddings encode rich but task-irrelevant noise that often overwhelms task-specific semantics under distribution shifts. To address these limitations, we propose SubTTA, which aligns the semantic subspaces of both modalities to enhance zero-shot predictions to better guide the TTA process. To bridge the modality gap, SubTTA extracts the principal subspaces of both modalities and aligns the visual manifold to the textual semantic anchor by minimizing their chordal distance. To eliminate visual nuisance, SubTTA projects the aligned visual features onto the task-specific textual subspace, which filters out task-irrelevant noise by constraining visual embeddings within the valid semantic span, and standard TTA is further performed on the purified space to refine the decision boundaries. Extensive experiments on various benchmarks and VLM architectures demonstrate the effectiveness of SubTTA, yielding an average improvement of 2.24% over state-of-the-art TTA methods.

Method: Proposes instance-aligned captions that link each textual claim to specific object instances with appearance and motion attributes, capturing who caused the anomaly, what each entity was doing, whom it affected, and where the explanation is grounded.

Result: Annotated eight widely used VAD benchmarks and extended VIEW360 dataset to create VIEW360+ with 868 additional videos, eight locations, and four new anomaly types. Experiments show instance-level spatially grounded captions reveal limitations in current LLM- and VLM-based methods.

Conclusion: The framework enables verifiable and actionable reasoning for explainable VAD, providing a robust benchmark for future research in trustworthy and interpretable anomaly detection.

Abstract: Explainable video anomaly detection (VAD) is crucial for safety-critical applications, yet even with recent progress, much of the research still lacks spatial grounding, making the explanations unverifiable. This limitation is especially pronounced in multi-entity interactions, where existing explainable VAD methods often produce incomplete or visually misaligned descriptions, reducing their trustworthiness. To address these challenges, we introduce instance-aligned captions that link each textual claim to specific object instances with appearance and motion attributes. Our framework captures who caused the anomaly, what each entity was doing, whom it affected, and where the explanationis grounded, enabling verifiable and actionable reasoning. We annotate eight widely used VAD benchmarks and extend the 360-degree egocentric dataset, VIEW360, with 868 additional videos, eight locations, and four new anomaly types, creating VIEW360+, a comprehensive testbed for explainable VAD. Experiments show that our instance-level spatially grounded captions reveal significant limitations in current LLM- and VLM-based methods while providing a robust benchmark for future research in trustworthy and interpretable anomaly detection.

Method: Co-design framework with: 1) Custom Spatially Varying Exposure (SVE) sensor capturing multi-exposure data in single shot for robust haze assessment, 2) Physics-aware dehazing algorithm that dynamically estimates haze density, performs region-adaptive illumination optimization, and applies multi-scale entropy-constrained fusion to separate haze from scene radiance.

Result: Superior performance in recovering physically accurate visual information of plume and engine region from real launch imagery and controlled experiments, enabling extraction of key mechanical parameters (particle velocity, flow instability frequency, structural vibration).

Conclusion: The framework provides reliable image basis for precise quantitative analysis in extreme aerospace environments, overcoming traditional limitations of optical measurement under severe rocket launch conditions.

Abstract: Quantitative optical measurement of critical mechanical parameters – such as plume flow fields, shock wave structures, and nozzle oscillations – during rocket launch faces severe challenges due to extreme imaging conditions. Intense combustion creates dense particulate haze and luminance variations exceeding 120 dB, degrading image data and undermining subsequent photogrammetric and velocimetric analyses. To address these issues, we propose a hardware-algorithm co-design framework that combines a custom Spatially Varying Exposure (SVE) sensor with a physics-aware dehazing algorithm. The SVE sensor acquires multi-exposure data in a single shot, enabling robust haze assessment without relying on idealized atmospheric models. Our approach dynamically estimates haze density, performs region-adaptive illumination optimization, and applies multi-scale entropy-constrained fusion to effectively separate haze from scene radiance. Validated on real launch imagery and controlled experiments, the framework demonstrates superior performance in recovering physically accurate visual information of the plume and engine region. This offers a reliable image basis for extracting key mechanical parameters, including particle velocity, flow instability frequency, and structural vibration, thereby supporting precise quantitative analysis in extreme aerospace environments.

Method: Multi-level alignment framework with semantic-aware instance and sparse token alignments. Incorporates inter-report similarity to eliminate false negatives in contrastive learning, and aligns image patches with relevant word tokens at patch-word level.

Result: Improves transfer performance across different datasets on three downstream tasks: image classification, segmentation, and object detection. Achieves significant improvements in fine-grained tasks even with limited labeled data.

Conclusion: SISTA effectively addresses the false negative problem in medical VLP and enables better semantic representation through multi-level alignment, demonstrating strong performance on various medical imaging tasks.

Abstract: Medical contrastive vision-language pre-training (VLP) has demonstrated significant potential in improving performance on downstream tasks. Traditional approaches typically employ contrastive learning, treating paired image-report samples as positives and unpaired ones as negatives. However, in medical datasets, there can be substantial similarities between images or reports from different patients. Rigidly treating all unpaired samples as negatives, can disrupt the underlying semantic structure and negatively impact the quality of the learned representations. In this paper, we propose a multi-level alignment framework, Representation Learning with Semantic-aware Instance and Sparse Token Alignments (SISTA) by exploiting the semantic correspondence between medical image and radiology reports at two levels, i.e., image-report and patch-word levels. Specifically, we improve the conventional contrastive learning by incorporating inter-report similarity to eliminate the false negatives and introduce a method to effectively align image patches with relevant word tokens. Experimental results demonstrate the effectiveness of the proposed framework in improving transfer performance across different datasets on three downstream tasks: image classification, image segmentation, and object detection. Notably, our framework achieves significant improvements in fine-grained tasks even with limited labeled data. Codes and pre-trained models will be made available.

Method: Built upon PVRCNN++ framework, enhanced with tailored data augmentation and self-training strategy with pseudo-labels to narrow domain gaps and improve cross-platform generalization.

Result: Achieved 3rd place in RoboSense2025 Challenge with 62.67% 3D AP for Car category on phase-1 target domain, and 58.76% (Car) and 49.81% (Pedestrian) on phase-2 target domain.

Conclusion: The proposed approach successfully addresses cross-platform generalization challenges in 3D object detection through domain adaptation techniques, demonstrating competitive performance across different sensor platforms and object categories.

Abstract: This technical report represents the award-winning solution to the Cross-platform 3D Object Detection task in the RoboSense2025 Challenge. Our approach is built upon PVRCNN++, an efficient 3D object detection framework that effectively integrates point-based and voxel-based features. On top of this foundation, we improve cross-platform generalization by narrowing domain gaps through tailored data augmentation and a self-training strategy with pseudo-labels. These enhancements enabled our approach to secure the 3rd place in the challenge, achieving a 3D AP of 62.67% for the Car category on the phase-1 target domain, and 58.76% and 49.81% for Car and Pedestrian categories respectively on the phase-2 target domain.

Method: Integrates three core capabilities: 1) multi-stage motion cue framework for identifying dynamic objects using depth and camera pose priors, 2) cognitive mapping system for storing, recalling, and updating static scenes across multiple visits, and 3) factor graph optimization for refining camera poses.

Result: Demonstrates state-of-the-art performance on video depth estimation, camera pose reconstruction, and 3D mapping tasks, effectively supporting continuous scene understanding across extended sequences and multiple visits.

Conclusion: CogniMap3D successfully emulates human cognitive processes for dynamic 3D scene understanding, providing an effective framework for persistent spatial memory and scene reconstruction that works across multiple visits to familiar locations.

Abstract: We present CogniMap3D, a bioinspired framework for dynamic 3D scene understanding and reconstruction that emulates human cognitive processes. Our approach maintains a persistent memory bank of static scenes, enabling efficient spatial knowledge storage and rapid retrieval. CogniMap3D integrates three core capabilities: a multi-stage motion cue framework for identifying dynamic objects, a cognitive mapping system for storing, recalling, and updating static scenes across multiple visits, and a factor graph optimization strategy for refining camera poses. Given an image stream, our model identifies dynamic regions through motion cues with depth and camera pose priors, then matches static elements against its memory bank. When revisiting familiar locations, CogniMap3D retrieves stored scenes, relocates cameras, and updates memory with new observations. Evaluations on video depth estimation, camera pose reconstruction, and 3D mapping tasks demonstrate its state-of-the-art performance, while effectively supporting continuous scene understanding across extended sequences and multiple visits.

Method: Uses second-order Gaussian directional derivative (SOGDD) filtering to smooth two high-resolution corner models (END-type and L-type), derives SOGDD representations, discovers corner characteristics, and develops a method to select optimal Gaussian filtering scales for accurate adjacent corner detection.

Result: The proposed method demonstrates superior performance compared to state-of-the-art methods in localization error, robustness to image blur transformation, image matching, and 3D reconstruction applications.

Conclusion: The SOGDD-based approach effectively addresses the interference problem between adjacent corners, providing more accurate corner detection that improves performance in practical computer vision applications.

Abstract: Corner detection is widely used in various computer vision tasks, such as image matching and 3D reconstruction. Our research indicates that there are theoretical flaws in Zhang et al.’s use of a simple corner model to obtain a series of corner characteristics, as the grayscale information of two adjacent corners can affect each other. In order to address the above issues, a second-order Gaussian directional derivative (SOGDD) filter is used in this work to smooth two typical high-resolution angle models (i.e. END-type and L-type models). Then, the SOGDD representations of these two corner models were derived separately, and many characteristics of high-resolution corners were discovered, which enabled us to demonstrate how to select Gaussian filtering scales to obtain intensity variation information from images, accurately depicting adjacent corners. In addition, a new high-resolution corner detection method for images has been proposed for the first time, which can accurately detect adjacent corner points. The experimental results have verified that the proposed method outperforms state-of-the-art methods in terms of localization error, robustness to image blur transformation, image matching, and 3D reconstruction.

Method: Created GI-Bench benchmark with 20 lesion categories, evaluated 12 MLLMs across 5-stage clinical workflow (localization, identification, diagnosis, description, management), compared against 3 junior endoscopists and 3 trainees using Macro-F1, mIoU, and Likert scales.

Result: Gemini-3-Pro achieved SOTA performance; top models outperformed trainees (Macro-F1 0.641 vs 0.492) and rivaled junior endoscopists (0.727, p>0.05). However, humans significantly outperformed models in lesion localization (mIoU >0.506 vs 0.345, p<0.05). Models showed “fluency-accuracy paradox” - better readability but lower factual correctness due to hallucinations.

Conclusion: MLLMs show diagnostic reasoning potential but face critical spatial grounding bottlenecks and fluency-accuracy trade-offs, requiring dynamic benchmarking via GI-Bench to track clinical utility evolution.

Abstract: Multimodal Large Language Models (MLLMs) show promise in gastroenterology, yet their performance against comprehensive clinical workflows and human benchmarks remains unverified. To systematically evaluate state-of-the-art MLLMs across a panoramic gastrointestinal endoscopy workflow and determine their clinical utility compared with human endoscopists. We constructed GI-Bench, a benchmark encompassing 20 fine-grained lesion categories. Twelve MLLMs were evaluated across a five-stage clinical workflow: anatomical localization, lesion identification, diagnosis, findings description, and management. Model performance was benchmarked against three junior endoscopists and three residency trainees using Macro-F1, mean Intersection-over-Union (mIoU), and multi-dimensional Likert scale. Gemini-3-Pro achieved state-of-the-art performance. In diagnostic reasoning, top-tier models (Macro-F1 0.641) outperformed trainees (0.492) and rivaled junior endoscopists (0.727; p>0.05). However, a critical “spatial grounding bottleneck” persisted; human lesion localization (mIoU >0.506) significantly outperformed the best model (0.345; p<0.05). Furthermore, qualitative analysis revealed a “fluency-accuracy paradox”: models generated reports with superior linguistic readability compared with humans (p<0.05) but exhibited significantly lower factual correctness (p<0.05) due to “over-interpretation” and hallucination of visual features.GI-Bench maintains a dynamic leaderboard that tracks the evolving performance of MLLMs in clinical endoscopy. The current rankings and benchmark results are available at https://roterdl.github.io/GIBench/.

Method: GPM mimics human visual system’s cooperative mechanism: 1) Global Insight Generator (GIG) extracts holistic cues first, 2) Parallel branches process different scales - LSAE for mid/large-scale semantic relations and IRB for fine-grained texture information. This enables coherent global-local representation, reflecting how humans perceive whole before details while maintaining contextual awareness.

Result: H-GPE network built on GPM achieves strong performance on image classification, object detection, and semantic segmentation while maintaining balanced FLOPs and parameters. It delivers superior accuracy-efficiency trade-off compared to recent state-of-the-art lightweight models.

Conclusion: GPM provides a more biologically plausible approach to lightweight vision networks that better balances parameter efficiency, computational cost, and task performance, addressing limitations of existing models through human vision-inspired cooperative multi-scale encoding.

Abstract: Lightweight vision networks have witnessed remarkable progress in recent years, yet achieving a satisfactory balance among parameter scale, computational overhead, and task performance remains difficult. Although many existing lightweight models manage to reduce computation considerably, they often do so at the expense of a substantial increase in parameter count (e.g., LSNet, MobileMamba), which still poses obstacles for deployment on resource-limited devices. In parallel, some studies attempt to draw inspiration from human visual perception, but their modeling tends to oversimplify the visual process, making it hard to reflect how perception truly operates. Revisiting the cooperative mechanism of the human visual system, we propose GPM (Global-to-Parallel Multi-scale Encoding). GPM first employs a Global Insight Generator (GIG) to extract holistic cues, and subsequently processes features of different scales through parallel branches: LSAE emphasizes mid-/large-scale semantic relations, while IRB (Inverted Residual Block) preserves fine-grained texture information, jointly enabling coherent representation of global and local features. As such, GPM conforms to two characteristic behaviors of human vision perceiving the whole before focusing on details, and maintaining broad contextual awareness even during local attention. Built upon GPM, we further develop the lightweight H-GPE network. Experiments on image classification, object detection, and semantic segmentation show that H-GPE achieves strong performance while maintaining a balanced footprint in both FLOPs and parameters, delivering a more favorable accuracy-efficiency trade-off compared with recent state-of-the-art lightweight models.

Method: R^4 framework uses four coordinated agents: Router (configures task-aware prompts), Retriever (generates reports and bounding boxes using exemplar memory), Reflector (critiques drafts for clinical error modes), and Repairer (iteratively revises outputs while curating exemplars).

Result: On chest X-ray analysis, R^4 boosts LLM-as-a-Judge scores by +1.7-+2.5 points and mAP50 by +2.5-+3.5 absolute points over strong single-VLM baselines without any gradient-based fine-tuning.

Conclusion: Agentic routing, reflection, and repair can transform strong but brittle vision-language models into more reliable and better grounded tools for clinical image interpretation.

Abstract: Medical image analysis increasingly relies on large vision-language models (VLMs), yet most systems remain single-pass black boxes that offer limited control over reasoning, safety, and spatial grounding. We propose R^4, an agentic framework that decomposes medical imaging workflows into four coordinated agents: a Router that configures task- and specialization-aware prompts from the image, patient history, and metadata; a Retriever that uses exemplar memory and pass@k sampling to jointly generate free-text reports and bounding boxes; a Reflector that critiques each draft-box pair for key clinical error modes (negation, laterality, unsupported claims, contradictions, missing findings, and localization errors); and a Repairer that iteratively revises both narrative and spatial outputs under targeted constraints while curating high-quality exemplars for future cases. Instantiated on chest X-ray analysis with multiple modern VLM backbones and evaluated on report generation and weakly supervised detection, R^4 consistently boosts LLM-as-a-Judge scores by roughly +1.7-+2.5 points and mAP50 by +2.5-+3.5 absolute points over strong single-VLM baselines, without any gradient-based fine-tuning. These results show that agentic routing, reflection, and repair can turn strong but brittle VLMs into more reliable and better grounded tools for clinical image interpretation. Our code can be found at: https://github.com/faiyazabdullah/MultimodalMedAgent

Method: Two-stage framework: (1) diffusion-based global harmonizer maps MR images to sequence-specific unified domain using style-agnostic gradient conditioning, (2) target-specific fine-tuner adapts globally aligned images to desired target domains. Uses tri-planar attention BiomedCLIP encoder to aggregate multi-view embeddings for volumetric style characterization.

Result: Evaluated on 4,163 T1- and T2-weighted MRIs, MMH demonstrates superior harmonization over state-of-the-art methods in image feature clustering, voxel-level comparison, tissue segmentation, and downstream age and site classification tasks.

Conclusion: MMH provides an effective unified framework for multi-site multi-sequence MRI harmonization that explicitly disentangles image styles from anatomy without requiring paired data, enabling better utilization of multi-site data for deep learning applications.

Abstract: Aggregating multi-site brain MRI data can enhance deep learning model training, but also introduces non-biological heterogeneity caused by site-specific variations (e.g., differences in scanner vendors, acquisition parameters, and imaging protocols) that can undermine generalizability. Recent retrospective MRI harmonization seeks to reduce such site effects by standardizing image style (e.g., intensity, contrast, noise patterns) while preserving anatomical content. However, existing methods often rely on limited paired traveling-subject data or fail to effectively disentangle style from anatomy. Furthermore, most current approaches address only single-sequence harmonization, restricting their use in real-world settings where multi-sequence MRI is routinely acquired. To this end, we introduce MMH, a unified framework for multi-site multi-sequence brain MRI harmonization that leverages biomedical semantic priors for sequence-aware style alignment. MMH operates in two stages: (1) a diffusion-based global harmonizer that maps MR images to a sequence-specific unified domain using style-agnostic gradient conditioning, and (2) a target-specific fine-tuner that adapts globally aligned images to desired target domains. A tri-planar attention BiomedCLIP encoder aggregates multi-view embeddings to characterize volumetric style information, allowing explicit disentanglement of image styles from anatomy without requiring paired data. Evaluations on 4,163 T1- and T2-weighted MRIs demonstrate MMH’s superior harmonization over state-of-the-art methods in image feature clustering, voxel-level comparison, tissue segmentation, and downstream age and site classification.

Method: Proposes MobiDiary with a unified sensor encoder that exploits shared inductive biases of motion-induced signals (IMU and Wi-Fi). Uses patch-based mechanism for local temporal correlations and heterogeneous placement embedding for spatial context. Transformer-based decoder with autoregressive mechanism generates descriptions word-by-word.

Result: Achieves state-of-the-art performance on multiple public benchmarks (XRF V2, UWash, WiFiTAD) for captioning metrics (BLEU@4, CIDEr, RMC). Effectively generalizes across modalities and outperforms specialized baselines in continuous action understanding.

Conclusion: MobiDiary successfully bridges the semantic gap between continuous physical signals and discrete linguistic descriptions, offering a privacy-preserving alternative to vision-based systems with expressive natural language output for smart home activity recognition.

Abstract: Human Activity Recognition (HAR) in smart homes is critical for health monitoring and assistive living. While vision-based systems are common, they face privacy concerns and environmental limitations (e.g., occlusion). In this work, we present MobiDiary, a framework that generates natural language descriptions of daily activities directly from heterogeneous physical signals (specifically IMU and Wi-Fi). Unlike conventional approaches that restrict outputs to pre-defined labels, MobiDiary produces expressive, human-readable summaries. To bridge the semantic gap between continuous, noisy physical signals and discrete linguistic descriptions, we propose a unified sensor encoder. Instead of relying on modality-specific engineering, we exploit the shared inductive biases of motion-induced signals–where both inertial and wireless data reflect underlying kinematic dynamics. Specifically, our encoder utilizes a patch-based mechanism to capture local temporal correlations and integrates heterogeneous placement embedding to unify spatial contexts across different sensors. These unified signal tokens are then fed into a Transformer-based decoder, which employs an autoregressive mechanism to generate coherent action descriptions word-by-word. We comprehensively evaluate our approach on multiple public benchmarks (XRF V2, UWash, and WiFiTAD). Experimental results demonstrate that MobiDiary effectively generalizes across modalities, achieving state-of-the-art performance on captioning metrics (e.g., BLEU@4, CIDEr, RMC) and outperforming specialized baselines in continuous action understanding.

Method: FUME uses a lightweight dual-stream architecture with weight-shared encoders, modality-specific self-attention, and channel attention fusion. It processes complementary CO2 and CH4 emission patterns from infrared cameras to jointly optimize gas plume segmentation and health classification (Healthy, Transitional, Acidotic).

Result: Achieves 80.99% mIoU for segmentation and 98.82% classification accuracy with only 1.28M parameters and 1.97G MACs, outperforming SOTA methods with 10x lower computational cost. CO2 provides primary discriminative signal, and dual-task learning is essential.

Conclusion: Establishes feasibility of gas emission-based livestock health monitoring, paving the way for practical in vitro acidosis detection systems. The approach demonstrates potential for non-invasive, continuous health monitoring in dairy cattle.

Abstract: Ruminal acidosis is a prevalent metabolic disorder in dairy cattle causing significant economic losses and animal welfare concerns. Current diagnostic methods rely on invasive pH measurement, limiting scalability for continuous monitoring. We present FUME (Fused Unified Multi-gas Emission Network), the first deep learning approach for rumen acidosis detection from dual-gas optical imaging under in vitro conditions. Our method leverages complementary carbon dioxide (CO2) and methane (CH4) emission patterns captured by infrared cameras to classify rumen health into Healthy, Transitional, and Acidotic states. FUME employs a lightweight dual-stream architecture with weight-shared encoders, modality-specific self-attention, and channel attention fusion, jointly optimizing gas plume segmentation and classification of dairy cattle health. We introduce the first dual-gas OGI dataset comprising 8,967 annotated frames across six pH levels with pixel-level segmentation masks. Experiments demonstrate that FUME achieves 80.99% mIoU and 98.82% classification accuracy while using only 1.28M parameters and 1.97G MACs–outperforming state-of-the-art methods in segmentation quality with 10x lower computational cost. Ablation studies reveal that CO2 provides the primary discriminative signal and dual-task learning is essential for optimal performance. Our work establishes the feasibility of gas emission-based livestock health monitoring, paving the way for practical, in vitro acidosis detection systems. Codes are available at https://github.com/taminulislam/fume.

Method: Used NIH Chest X-ray dataset with EVA-ViT image encoder combined with LlaMA or ChatGPT LLMs. Compared three approaches: image-based RAG (using KNN methods), text-based RAG (using external knowledge), and baseline methods.

Result: Text-based RAG effectively reduced hallucinations using external knowledge; image-based RAG improved prediction confidence and calibration via KNN methods. GPT LLM outperformed LlaMA with better performance, lower hallucination rate, and better Expected Calibration Error.

Conclusion: The research demonstrates effective approaches to reduce hallucinations in medical image analysis but highlights challenges with data imbalance and complex multi-stage structures, suggesting need for larger experience environments and balanced usage examples.

Abstract: This research analyzed and compared the multi-modal approach in the Vision Transformer(EVA-ViT) based image encoder with the LlaMA or ChatGPT LLM to reduce the hallucination problem and detect diseases in chest x-ray images. In this research, we utilized the NIH Chest X-ray image to train the model and compared it in image-based RAG, text-based RAG, and baseline. [3] [5] In a result, the text-based RAG[2] e!ectively reduces the hallucination problem by using external knowledge information, and the image-based RAG improved the prediction con"dence and calibration by using the KNN methods. [4] Moreover, the GPT LLM showed better performance, a low hallucination rate, and better Expected Calibration Error(ECE) than Llama Llama-based model. This research shows the challenge of data imbalance, a complex multi-stage structure, but suggests a large experience environment and a balanced example of use.

Method: Proposes event-based segmentation for ADL recognition using LLMs, which better aligns with LLM contextual reasoning, and introduces a novel method for estimating prediction confidence.

Result: Event-based segmentation consistently outperforms time-based LLM approaches on complex datasets, surpasses supervised data-driven methods even with smaller LLMs (e.g., Gemma 3 27B), and the confidence measure effectively distinguishes correct from incorrect predictions.

Conclusion: Event-based segmentation combined with confidence estimation significantly improves zero-shot ADL recognition, making LLM-based approaches more practical and reliable for smart home applications.

Abstract: Unobtrusive sensor-based recognition of Activities of Daily Living (ADLs) in smart homes by processing data collected from IoT sensing devices supports applications such as healthcare, safety, and energy management. Recent zero-shot methods based on Large Language Models (LLMs) have the advantage of removing the reliance on labeled ADL sensor data. However, existing approaches rely on time-based segmentation, which is poorly aligned with the contextual reasoning capabilities of LLMs. Moreover, existing approaches lack methods for estimating prediction confidence. This paper proposes to improve zero-shot ADL recognition with event-based segmentation and a novel method for estimating prediction confidence. Our experimental evaluation shows that event-based segmentation consistently outperforms time-based LLM approaches on complex, realistic datasets and surpasses supervised data-driven methods, even with relatively small LLMs (e.g., Gemma 3 27B). The proposed confidence measure effectively distinguishes correct from incorrect predictions.

Method: Created AIMC-Spec synthetic dataset with 33 modulation types across 13 SNR levels. Re-implemented and evaluated five representative deep learning algorithms (lightweight CNNs, denoising architectures, transformer-based networks) under unified input format for benchmarking.

Result: Significant performance variation across algorithms, with frequency-modulated (FM) signals classified more reliably than phase or hybrid types, especially at low SNRs. FM-only tests revealed how modulation type and network architecture influence classifier robustness.

Conclusion: AIMC-Spec establishes a reproducible baseline and provides a foundation for future research and standardization in the AIMC domain, enabling automated interpretation of radar pulse intrapulse structure.

Abstract: A lack of standardized datasets has long hindered progress in automatic intrapulse modulation classification (AIMC) - a critical task in radar signal analysis for electronic support systems, particularly under noisy or degraded conditions. AIMC seeks to identify the modulation type embedded within a single radar pulse from its complex in-phase and quadrature (I/Q) representation, enabling automated interpretation of intrapulse structure. This paper introduces AIMC-Spec, a comprehensive synthetic dataset for spectrogram-based image classification, encompassing 33 modulation types across 13 signal-to-noise ratio (SNR) levels. To benchmark AIMC-Spec, five representative deep learning algorithms - ranging from lightweight CNNs and denoising architectures to transformer-based networks - were re-implemented and evaluated under a unified input format. The results reveal significant performance variation, with frequency-modulated (FM) signals classified more reliably than phase or hybrid types, particularly at low SNRs. A focused FM-only test further highlights how modulation type and network architecture influence classifier robustness. AIMC-Spec establishes a reproducible baseline and provides a foundation for future research and standardization in the AIMC domain.

Method: HIPPO proposes two key components: (1) semantic-aware token preservation that fuses global attention scores with local visual semantics to retain semantic information at high pruning ratios, and (2) video parallel speculative decoding algorithm that decouples and overlaps draft generation and target verification phases.

Result: Experiments on four video-LLMs across six benchmarks demonstrate HIPPO’s effectiveness, yielding up to 3.51x speedup compared to vanilla auto-regressive decoding.

Conclusion: HIPPO addresses limitations of existing video-LLM speculative decoding methods by better preserving visual semantics and implementing parallel processing, achieving significant inference acceleration while maintaining output quality.

Abstract: Speculative decoding (SD) has emerged as a promising approach to accelerate LLM inference without sacrificing output quality. Existing SD methods tailored for video-LLMs primarily focus on pruning redundant visual tokens to mitigate the computational burden of massive visual inputs. However, existing methods do not achieve inference acceleration comparable to text-only LLMs. We observe from extensive experiments that this phenomenon mainly stems from two limitations: (i) their pruning strategies inadequately preserve visual semantic tokens, degrading draft quality and acceptance rates; (ii) even with aggressive pruning (e.g., 90% visual tokens removed), the draft model’s remaining inference cost limits overall speedup. To address these limitations, we propose HIPPO, a general holistic-aware parallel speculative decoding framework. Specifically, HIPPO proposes (i) a semantic-aware token preservation method, which fuses global attention scores with local visual semantics to retain semantic information at high pruning ratios; (ii) a video parallel SD algorithm that decouples and overlaps draft generation and target verification phases. Experiments on four video-LLMs across six benchmarks demonstrate HIPPO’s effectiveness, yielding up to 3.51x speedup compared to vanilla auto-regressive decoding.

Method: Uses stacked images with siamese capsule networks for one-shot identification tasks across different domains including industrial applications and face recognition.

Result: Capsule architecture achieves strong results and exceeds other techniques on a wide range of datasets from industrial applications to face recognition benchmarks.

Conclusion: Capsule networks are effective for one-shot learning, performing better than other methods while being easy to use and optimize across diverse domains.

Abstract: Convolutional neural networks (CNNs) have been widely used in the computer vision community, significantly improving the state-of-the-art. But learning good features often is computationally expensive in machine learning settings and is especially difficult when there is a lack of data. One-shot learning is one such area where only limited data is available. In one-shot learning, predictions have to be made after seeing only one example from one class, which requires special techniques. In this paper we explore different approaches to one-shot identification tasks in different domains including an industrial application and face recognition. We use a special technique with stacked images and use siamese capsule networks. It is encouraging to see that the approach using capsule architecture achieves strong results and exceeds other techniques on a wide range of datasets from industrial application to face recognition benchmarks while being easy to use and optimise.

Method: Introduced KidVis benchmark grounded in human visual development theory, deconstructing visual intelligence into six atomic capabilities (Concentration, Tracking, Discrimination, Memory, Spatial, and Closure) that 6-7 year old children possess. Evaluated 20 state-of-the-art MLLMs against human physiological baseline across 10 categories of low-semantic-dependent visual tasks.

Result: Human children achieved near-perfect average score of 95.32, while state-of-the-art GPT-5 attained only 67.33. Revealed “Scaling Law Paradox” - increasing model parameters fails to yield linear improvements in foundational visual capabilities.

Conclusion: Current MLLMs lack essential physiological perceptual primitives required for generalized visual intelligence, despite their reasoning prowess. The performance gap and scaling paradox suggest fundamental limitations in current architectures for basic visual perception.

Abstract: While Multimodal Large Language Models (MLLMs) have demonstrated impressive proficiency in high-level reasoning tasks, such as complex diagrammatic interpretation, it remains an open question whether they possess the fundamental visual primitives comparable to human intuition. To investigate this, we introduce KidVis, a novel benchmark grounded in the theory of human visual development. KidVis deconstructs visual intelligence into six atomic capabilities - Concentration, Tracking, Discrimination, Memory, Spatial, and Closure - already possessed by 6-7 year old children, comprising 10 categories of low-semantic-dependent visual tasks. Evaluating 20 state-of-the-art MLLMs against a human physiological baseline reveals a stark performance disparity. Results indicate that while human children achieve a near-perfect average score of 95.32, the state-of-the-art GPT-5 attains only 67.33. Crucially, we observe a “Scaling Law Paradox”: simply increasing model parameters fails to yield linear improvements in these foundational visual capabilities. This study confirms that current MLLMs, despite their reasoning prowess, lack the essential physiological perceptual primitives required for generalized visual intelligence.

Method: Proposes M3SR with a multi-perceptual fusion block integrated into a U-Net structure. This enables extraction and fusion of global, intermediate, and local features at multiple scales for accurate hyperspectral image reconstruction.

Result: Extensive experiments show M3SR outperforms state-of-the-art methods while maintaining lower computational cost, demonstrating superior performance in both quantitative and qualitative evaluations.

Conclusion: The proposed multi-scale, multi-perceptual Mamba architecture effectively addresses limitations of existing approaches and achieves better spectral reconstruction performance with improved efficiency.

Abstract: The Mamba architecture has been widely applied to various low-level vision tasks due to its exceptional adaptability and strong performance. Although the Mamba architecture has been adopted for spectral reconstruction, it still faces the following two challenges: (1) Single spatial perception limits the ability to fully understand and analyze hyperspectral images; (2) Single-scale feature extraction struggles to capture the complex structures and fine details present in hyperspectral images. To address these issues, we propose a multi-scale, multi-perceptual Mamba architecture for the spectral reconstruction task, called M3SR. Specifically, we design a multi-perceptual fusion block to enhance the ability of the model to comprehensively understand and analyze the input features. By integrating the multi-perceptual fusion block into a U-Net structure, M3SR can effectively extract and fuse global, intermediate, and local features, thereby enabling accurate reconstruction of hyperspectral images at multiple scales. Extensive quantitative and qualitative experiments demonstrate that the proposed M3SR outperforms existing state-of-the-art methods while incurring a lower computational cost.

Method: Proposes Region- and Context-aware Knowledge Distillation (ReCo-KD) with two components: 1) Multi-Scale Structure-Aware Region Distillation (MS-SARD) using class-aware masks and scale-normalized weighting to emphasize small clinically important regions, and 2) Multi-Scale Context Alignment (MS-CA) aligning teacher-student affinity patterns across feature levels. Implemented on nnU-Net in backbone-agnostic manner.

Result: Distilled lightweight models attain accuracy close to teacher models while significantly reducing parameters and inference latency across multiple public 3D medical segmentation datasets and a challenging aggregated dataset.

Conclusion: ReCo-KD provides a practical training-only framework for clinical deployment of efficient 3D medical segmentation models without custom student design, maintaining accuracy while reducing computational requirements.

Abstract: Accurate 3D medical image segmentation is vital for diagnosis and treatment planning, but state-of-the-art models are often too large for clinics with limited computing resources. Lightweight architectures typically suffer significant performance loss. To address these deployment and speed constraints, we propose Region- and Context-aware Knowledge Distillation (ReCo-KD), a training-only framework that transfers both fine-grained anatomical detail and long-range contextual information from a high-capacity teacher to a compact student network. The framework integrates Multi-Scale Structure-Aware Region Distillation (MS-SARD), which applies class-aware masks and scale-normalized weighting to emphasize small but clinically important regions, and Multi-Scale Context Alignment (MS-CA), which aligns teacher-student affinity patterns across feature levels. Implemented on nnU-Net in a backbone-agnostic manner, ReCo-KD requires no custom student design and is easily adapted to other architectures. Experiments on multiple public 3D medical segmentation datasets and a challenging aggregated dataset show that the distilled lightweight model attains accuracy close to the teacher while markedly reducing parameters and inference latency, underscoring its practicality for clinical deployment.

Method: Three key components: 1) Compact DiT architecture with adaptive global-local sparse attention, 2) Elastic training framework that jointly optimizes sub-DiTs of varying capacities within a unified supernetwork, 3) Knowledge-Guided Distribution Matching Distillation (step-distillation pipeline combining DMD objective with knowledge transfer from few-step teacher models).

Result: Achieves transformer-level generation quality under strict resource constraints, enabling high-fidelity low-latency generation (e.g., 4-step) suitable for real-time on-device use on diverse hardware.

Conclusion: The framework enables scalable, efficient, and high-quality diffusion models for practical deployment on mobile and edge devices while maintaining generation quality comparable to full DiTs.

Abstract: Recent advances in diffusion transformers (DiTs) have set new standards in image generation, yet remain impractical for on-device deployment due to their high computational and memory costs. In this work, we present an efficient DiT framework tailored for mobile and edge devices that achieves transformer-level generation quality under strict resource constraints. Our design combines three key components. First, we propose a compact DiT architecture with an adaptive global-local sparse attention mechanism that balances global context modeling and local detail preservation. Second, we propose an elastic training framework that jointly optimizes sub-DiTs of varying capacities within a unified supernetwork, allowing a single model to dynamically adjust for efficient inference across different hardware. Finally, we develop Knowledge-Guided Distribution Matching Distillation, a step-distillation pipeline that integrates the DMD objective with knowledge transfer from few-step teacher models, producing high-fidelity and low-latency generation (e.g., 4-step) suitable for real-time on-device use. Together, these contributions enable scalable, efficient, and high-quality diffusion models for deployment on diverse hardware.

Method: IQARAG uses Retrieval-Augmented Generation to retrieve semantically similar but quality-variant reference images with Mean Opinion Scores. It has three phases: Retrieval Feature Extraction, Image Retrieval, and Integration & Quality Score Generation. Retrieved images provide visual perception anchors for IQA tasks through specific prompt integration.

Result: Extensive experiments across multiple IQA datasets (KADID, KonIQ, LIVE Challenge, SPAQ) demonstrate that IQARAG effectively boosts LMMs’ IQA performance, offering a resource-efficient alternative to fine-tuning.

Conclusion: IQARAG provides a novel, training-free framework that enhances LMMs’ IQA ability through RAG with quality-variant reference images, achieving competitive performance without the computational costs of fine-tuning.

Abstract: Large Multimodal Models (LMMs) have recently shown remarkable promise in low-level visual perception tasks, particularly in Image Quality Assessment (IQA), demonstrating strong zero-shot capability. However, achieving state-of-the-art performance often requires computationally expensive fine-tuning methods, which aim to align the distribution of quality-related token in output with image quality levels. Inspired by recent training-free works for LMM, we introduce IQARAG, a novel, training-free framework that enhances LMMs’ IQA ability. IQARAG leverages Retrieval-Augmented Generation (RAG) to retrieve some semantically similar but quality-variant reference images with corresponding Mean Opinion Scores (MOSs) for input image. These retrieved images and input image are integrated into a specific prompt. Retrieved images provide the LMM with a visual perception anchor for IQA task. IQARAG contains three key phases: Retrieval Feature Extraction, Image Retrieval, and Integration & Quality Score Generation. Extensive experiments across multiple diverse IQA datasets, including KADID, KonIQ, LIVE Challenge, and SPAQ, demonstrate that the proposed IQARAG effectively boosts the IQA performance of LMMs, offering a resource-efficient alternative to fine-tuning for quality assessment.

Method: Proposes YOLOBirDrone architecture featuring: 1) Adaptive and Extended Layer Aggregation (AELAN), 2) Multi-scale Progressive Dual Attention Module (MPDA), and 3) Reverse MPDA (RMPDA) to preserve shape information and enrich features with local/global spatial and channel information. Also introduces BirDrone dataset containing small, challenging objects for robust aerial identification.

Result: Experimental results show YOLOBirDrone outperforms state-of-the-art algorithms, achieving approximately 85% detection accuracy across various scenarios.

Conclusion: The proposed YOLOBirDrone architecture effectively addresses the drone-bird discrimination problem, offering improved detection accuracy for aerial security applications through novel architectural components and a comprehensive dataset.

Abstract: The use of aerial drones for commercial and defense applications has benefited in many ways and is therefore utilized in several different application domains. However, they are also increasingly used for targeted attacks, posing a significant safety challenge and necessitating the development of drone detection systems. Vision-based drone detection systems currently have an accuracy limitation and struggle to distinguish between drones and birds, particularly when the birds are small in size. This research work proposes a novel YOLOBirDrone architecture that improves the detection and classification accuracy of birds and drones. YOLOBirDrone has different components, including an adaptive and extended layer aggregation (AELAN), a multi-scale progressive dual attention module (MPDA), and a reverse MPDA (RMPDA) to preserve shape information and enrich features with local and global spatial and channel information. A large-scale dataset, BirDrone, is also introduced in this article, which includes small and challenging objects for robust aerial object identification. Experimental results demonstrate an improvement in performance metrics through the proposed YOLOBirDrone architecture compared to other state-of-the-art algorithms, with detection accuracy reaching approximately 85% across various scenarios.

Method: Proposes UM-Text with: 1) Visual Language Model (VLM) to process instructions and reference images for context-aware text/layout design, 2) UM-Encoder to combine condition embeddings automatically configured by VLM, 3) Regional consistency loss for glyph generation supervision, 4) Three-stage training strategy, and 5) UM-DATA-200K dataset for training.

Result: Achieves state-of-the-art performance on multiple public benchmarks through extensive qualitative and quantitative evaluation, demonstrating superior visual text generation that harmonizes with reference images.

Conclusion: UM-Text provides an effective unified solution for natural language-based visual text editing that maintains stylistic consistency with reference images, advancing the field beyond manual attribute specification approaches.

Abstract: With the rapid advancement of image generation, visual text editing using natural language instructions has received increasing attention. The main challenge of this task is to fully understand the instruction and reference image, and thus generate visual text that is style-consistent with the image. Previous methods often involve complex steps of specifying the text content and attributes, such as font size, color, and layout, without considering the stylistic consistency with the reference image. To address this, we propose UM-Text, a unified multimodal model for context understanding and visual text editing by natural language instructions. Specifically, we introduce a Visual Language Model (VLM) to process the instruction and reference image, so that the text content and layout can be elaborately designed according to the context information. To generate an accurate and harmonious visual text image, we further propose the UM-Encoder to combine the embeddings of various condition information, where the combination is automatically configured by VLM according to the input instruction. During training, we propose a regional consistency loss to offer more effective supervision for glyph generation on both latent and RGB space, and design a tailored three-stage training strategy to further enhance model performance. In addition, we contribute the UM-DATA-200K, a large-scale visual text image dataset on diverse scenes for model training. Extensive qualitative and quantitative results on multiple public benchmarks demonstrate that our method achieves state-of-the-art performance.

Method: Introduces Inception Generative Adversarial Network (IGAN) incorporating deeper inception-inspired convolution and dilated convolution architectures. Uses dropout and spectral normalization in both generator and discriminator to mitigate gradient explosion and overfitting.

Result: Achieves FID scores of 13.12 (CUB-200) and 15.08 (ImageNet), representing 28-33% improvement over state-of-the-art GANs. Attains Inception Scores of 9.27 and 68.25, indicating improved image diversity and quality.

Conclusion: IGAN successfully balances training stability with image generation quality, providing a scalable and computationally efficient framework for high-fidelity image synthesis that addresses key GAN limitations.

Abstract: Generative Adversarial Networks (GANs) face a significant challenge of striking an optimal balance between high-quality image generation and training stability. Recent techniques, such as DCGAN, BigGAN, and StyleGAN, improve visual fidelity; however, such techniques usually struggle with mode collapse and unstable gradients at high network depth. This paper proposes a novel GAN structural model that incorporates deeper inception-inspired convolution and dilated convolution. This novel model is termed the Inception Generative Adversarial Network (IGAN). The IGAN model generates high-quality synthetic images while maintaining training stability, by reducing mode collapse as well as preventing vanishing and exploding gradients. Our proposed IGAN model achieves the Frechet Inception Distance (FID) of 13.12 and 15.08 on the CUB-200 and ImageNet datasets, respectively, representing a 28-33% improvement in FID over the state-of-the-art GANs. Additionally, the IGAN model attains an Inception Score (IS) of 9.27 and 68.25, reflecting improved image diversity and generation quality. Finally, the two techniques of dropout and spectral normalization are utilized in both the generator and discriminator structures to further mitigate gradient explosion and overfitting. These findings confirm that the IGAN model potentially balances training stability with image generation quality, constituting a scalable and computationally efficient framework for high-fidelity image synthesis.

Method: BioMorphNet constructs graphs to model relationships between target patches and neighbors, adjusting response strength based on morphological and molecular similarity. It derives clinical pathway features from spatial transcriptomics using predefined databases, and includes a novel learnable pathway module to simulate biological pathway formation.

Result: BioMorphNet improves average classification metrics by 2.67% for prostate cancer, 5.48% for colorectal cancer, and 6.29% for breast cancer compared to state-of-the-art multimodal methods. It enables accurate tissue classification for tumor localization and differential gene expression analysis for biomarker discovery.

Conclusion: BioMorphNet effectively integrates tissue morphology and spatial gene expression, advancing multimodal analysis for cancer diagnosis and biomarker discovery through improved classification performance and biological pathway modeling.

Abstract: Automatic integration of whole slide images (WSIs) and gene expression profiles has demonstrated substantial potential in precision clinical diagnosis and cancer progression studies. However, most existing studies focus on individual gene sequences and slide level classification tasks, with limited attention to spatial transcriptomics and patch level applications. To address this limitation, we propose a multimodal network, BioMorphNet, which automatically integrates tissue morphological features and spatial gene expression to support tissue classification and differential gene analysis. For considering morphological features, BioMorphNet constructs a graph to model the relationships between target patches and their neighbors, and adjusts the response strength based on morphological and molecular level similarity, to better characterize the tumor microenvironment. In terms of multimodal interactions, BioMorphNet derives clinical pathway features from spatial transcriptomic data based on a predefined pathway database, serving as a bridge between tissue morphology and gene expression. In addition, a novel learnable pathway module is designed to automatically simulate the biological pathway formation process, providing a complementary representation to existing clinical pathways. Compared with the latest morphology gene multimodal methods, BioMorphNet’s average classification metrics improve by 2.67%, 5.48%, and 6.29% for prostate cancer, colorectal cancer, and breast cancer datasets, respectively. BioMorphNet not only classifies tissue categories within WSIs accurately to support tumor localization, but also analyzes differential gene expression between tissue categories based on prediction confidence, contributing to the discovery of potential tumor biomarkers.

Method: Proposes Individualized Exploratory Transformer (IET) with Individualized Exploratory Attention (IEA) mechanism. Each token adaptively selects its own content-aware and independent attention candidates, enabling token-adaptive and asymmetric attention computation.

Result: Extensive experiments on standard SR benchmarks show IET achieves state-of-the-art performance while maintaining comparable computational complexity.

Conclusion: IET’s token-adaptive attention mechanism enables more precise information aggregation with computational efficiency, addressing limitations of fixed group-wise attention in transformer-based SISR methods.

Abstract: Single Image Super-Resolution (SISR) is a fundamental computer vision task that aims to reconstruct a high-resolution (HR) image from a low-resolution (LR) input. Transformer-based methods have achieved remarkable performance by modeling long-range dependencies in degraded images. However, their feature-intensive attention computation incurs high computational cost. To improve efficiency, most existing approaches partition images into fixed groups and restrict attention within each group. Such group-wise attention overlooks the inherent asymmetry in token similarities, thereby failing to enable flexible and token-adaptive attention computation. To address this limitation, we propose the Individualized Exploratory Transformer (IET), which introduces a novel Individualized Exploratory Attention (IEA) mechanism that allows each token to adaptively select its own content-aware and independent attention candidates. This token-adaptive and asymmetric design enables more precise information aggregation while maintaining computational efficiency. Extensive experiments on standard SR benchmarks demonstrate that IET achieves state-of-the-art performance under comparable computational complexity.

Method: Systematically study semantic misalignment in VLMs under controlled degradation of upstream visual perception using semantic segmentation on Cityscapes dataset, introducing perception-realistic corruptions and proposing language-level misalignment metrics.

Result: Even moderate drops in conventional segmentation metrics cause severe VLM failures including hallucinated objects, omission of safety-critical entities, and inconsistent safety judgments, revealing disconnect between pixel-level robustness and semantic reliability.

Conclusion: Current VLM-based systems have critical limitations in handling perception uncertainty, necessitating new evaluation frameworks that explicitly account for perception degradation in safety-critical applications.

Abstract: Vision-Language Models (VLMs) are increasingly deployed in autonomous driving and embodied AI systems, where reliable perception is critical for safe semantic reasoning and decision-making. While recent VLMs demonstrate strong performance on multimodal benchmarks, their robustness to realistic perception degradation remains poorly understood. In this work, we systematically study semantic misalignment in VLMs under controlled degradation of upstream visual perception, using semantic segmentation on the Cityscapes dataset as a representative perception module. We introduce perception-realistic corruptions that induce only moderate drops in conventional segmentation metrics, yet observe severe failures in downstream VLM behavior, including hallucinated object mentions, omission of safety-critical entities, and inconsistent safety judgments. To quantify these effects, we propose a set of language-level misalignment metrics that capture hallucination, critical omission, and safety misinterpretation, and analyze their relationship with segmentation quality across multiple contrastive and generative VLMs. Our results reveal a clear disconnect between pixel-level robustness and multimodal semantic reliability, highlighting a critical limitation of current VLM-based systems and motivating the need for evaluation frameworks that explicitly account for perception uncertainty in safety-critical applications.

Method: Uses Signed Distance Function (SDF) as underlying geometric representation to guide rendering, complemented by a novel Geometry-Preservation Loss to preserve fine structural details.

Result: Achieves competitive rendering performance with 4-5x reduction in energy consumption compared to similar methods, producing photorealistic results with sharp, geometrically coherent details.

Conclusion: Geo-NVS-w is a robust method for in-the-wild novel view synthesis that successfully addresses geometric inconsistency issues while being energy-efficient.

Abstract: We introduce Geo-NVS-w, a geometry-aware framework for high-fidelity novel view synthesis from unstructured, in-the-wild image collections. While existing in-the-wild methods already excel at novel view synthesis, they often lack geometric grounding on complex surfaces, sometimes producing results that contain inconsistencies. Geo-NVS-w addresses this limitation by leveraging an underlying geometric representation based on a Signed Distance Function (SDF) to guide the rendering process. This is complemented by a novel Geometry-Preservation Loss which ensures that fine structural details are preserved. Our framework achieves competitive rendering performance, while demonstrating a 4-5x reduction reduction in energy consumption compared to similar methods. We demonstrate that Geo-NVS-w is a robust method for in-the-wild NVS, yielding photorealistic results with sharp, geometrically coherent details.

Method: LoGo uses: 1) Local class-balanced prototype estimation with intra-class independent anchor mining for tail classes, 2) Global optimal transport-based distribution alignment to correct head class dominance, and 3) Dual-consistency pseudo-label filtering that requires agreement between local ensemble predictions and global assignments.

Result: The framework effectively mitigates feature collapse in long-tailed distributions, prevents model bias toward majority classes, and generates reliable pseudo-labels for self-training in source-free domain adaptation scenarios.

Conclusion: LoGo provides a novel solution for source-free domain adaptation in geospatial point clouds, addressing both local class imbalance and global distribution misalignment through dual-consensus approach, enabling practical deployment without source data access.

Abstract: Semantic segmentation of 3D geospatial point clouds is pivotal for remote sensing applications. However, variations in geographic patterns across regions and data acquisition strategies induce significant domain shifts, severely degrading the performance of deployed models. Existing domain adaptation methods typically rely on access to source-domain data. However, this requirement is rarely met due to data privacy concerns, regulatory policies, and data transmission limitations. This motivates the largely underexplored setting of source-free unsupervised domain adaptation (SFUDA), where only a pretrained model and unlabeled target-domain data are available. In this paper, we propose LoGo (Local-Global Dual-Consensus), a novel SFUDA framework specifically designed for geospatial point clouds. At the local level, we introduce a class-balanced prototype estimation module that abandons conventional global threshold filtering in favor of an intra-class independent anchor mining strategy. This ensures that robust feature prototypes can be generated even for sample-scarce tail classes, effectively mitigating the feature collapse caused by long-tailed distributions. At the global level, we introduce an optimal transport-based global distribution alignment module that formulates pseudo-label assignment as a global optimization problem. By enforcing global distribution constraints, this module effectively corrects the over-dominance of head classes inherent in local greedy assignments, preventing model predictions from being severely biased towards majority classes. Finally, we propose a dual-consistency pseudo-label filtering mechanism. This strategy retains only high-confidence pseudo-labels where local multi-augmented ensemble predictions align with global optimal transport assignments for self-training.

Method: Developed a system combining Arduino microcontroller, SIM800L GSM module for cellular communication, ultrasonic sensor for food-level monitoring, and servo mechanism for portion dispensing, controlled via MIT App Inventor mobile app.

Result: Achieved 98% SMS command success rate, consistent portion dispensing with ±2.67% variance, and reliable autonomous operation in a low-cost, energy-efficient design.

Conclusion: The system provides a practical, scalable, internet-independent solution for personalized pet feeding, setting a new benchmark for low-cost GSM-powered automation in smart pet products.

Abstract: Pet ownership is increasingly common in modern households, yet maintaining a consistent feeding schedule remains challenging for the owners particularly those who live in cities and have busy lifestyles. This paper presents the design, development, and validation of a low-cost, scalable GSM-IoT smart pet feeder that enables remote monitoring and control through cellular communication. The device combines with an Arduino microcontroller, a SIM800L GSM module for communication, an ultrasonic sensor for real-time food-level assessment, and a servo mechanism for accurate portion dispensing. A dedicated mobile application was developed using MIT App Inventor which allows owners to send feeding commands and receive real-time status updates. Experimental results demonstrate a 98% SMS command success rate, consistent portion dispensing with $\pm 2.67$% variance, and reliable autonomous operation. Its modular, energy-efficient design makes it easy to use in a wide range of households, including those with limited resources. This work pushes forward the field of accessible pet care technology by providing a practical, scalable, and completely internet-independent solution for personalized pet feeding. In doing so, it sets a new benchmark for low-cost, GSM-powered automation in smart pet products.

Method: Two-stage deep learning pipeline: (1) YOLOv8 detects third molars and classifies anatomical positions/angulations using Winter’s classification, (2) Modified ResNet-50 classifier detects radiographic features of pericoronitis, with Grad-CAM for interpretability.

Result: YOLOv8 achieved 92% precision and 92.5% mAP; ResNet-50 yielded F1-scores of 88% (normal) and 86% (pericoronitis); Radiologists reported 84% alignment between Grad-CAM highlights and their diagnostic impressions.

Conclusion: The system demonstrates strong potential for AI-assisted panoramic assessment with explainable AI features that enhance clinical trust and support diagnostic confidence.

Abstract: Objectives: To overcome challenges in diagnosing pericoronitis on panoramic radiographs, an AI-assisted assessment system integrating anatomical localization, pathological classification, and interpretability. Methods: A two-stage deep learning pipeline was implemented. The first stage used YOLOv8 to detect third molars and classify their anatomical positions and angulations based on Winter’s classification. Detected regions were then fed into a second-stage classifier, a modified ResNet-50 architecture, for detecting radiographic features suggestive of pericoronitis. To enhance clinical trust, Grad-CAM was used to highlight key diagnostic regions on the radiographs. Results: The YOLOv8 component achieved 92% precision and 92.5% mean average precision. The ResNet-50 classifier yielded F1-scores of 88% for normal cases and 86% for pericoronitis. Radiologists reported 84% alignment between Grad-CAM and their diagnostic impressions, supporting the radiographic relevance of the interpretability output. Conclusion: The system shows strong potential for AI-assisted panoramic assessment, with explainable AI features that support clinical confidence.

Method: Three key approaches: 1) Deep integration of BLIP-2 with YOLO-World and YOLOv8-Seg for object detection and segmentation, 2) Content-aware key frame sampling using K-Means clustering with intelligent frame selection and temporal feature concatenation for video tasks, 3) Unified prompt optimization scheme injecting structured YOLO event logs as context and using output constraints to filter technical details.

Result: Enables BLIP-2 to handle multi-task UAV applications with minimal adaptation, extending capabilities to video-level interactive tasks while maintaining lightweight architecture suitable for edge devices.

Conclusion: Proposed platform effectively bridges the gap between computational demands and resource constraints on UAVs, providing a practical solution for real-time multimodal understanding without requiring task-specific fine-tuning on drone data.

Abstract: The demand for real-time visual understanding and interaction in complex scenarios is increasingly critical for unmanned aerial vehicles. However, a significant challenge arises from the contradiction between the high computational cost of large Vision language models and the limited computing resources available on UAV edge devices. To address this challenge, this paper proposes a lightweight multimodal task platform based on BLIP-2, integrated with YOLO-World and YOLOv8-Seg models. This integration extends the multi-task capabilities of BLIP-2 for UAV applications with minimal adaptation and without requiring task-specific fine-tuning on drone data. Firstly, the deep integration of BLIP-2 with YOLO models enables it to leverage the precise perceptual results of YOLO for fundamental tasks like object detection and instance segmentation, thereby facilitating deeper visual-attention understanding and reasoning. Secondly, a content-aware key frame sampling mechanism based on K-Means clustering is designed, which incorporates intelligent frame selection and temporal feature concatenation. This equips the lightweight BLIP-2 architecture with the capability to handle video-level interactive tasks effectively. Thirdly, a unified prompt optimization scheme for multi-task adaptation is implemented. This scheme strategically injects structured event logs from the YOLO models as contextual information into BLIP-2’s input. Combined with output constraints designed to filter out technical details, this approach effectively guides the model to generate accurate and contextually relevant outputs for various tasks.

Method: SPARK consists of two main components: (1) a geometry-aware online extrinsic estimation module that uses multi-view priors with cross-view and temporal consistency for stable self-calibration, and (2) a confidence-driven point cloud fusion strategy that models depth reliability and visibility at pixel and point levels to suppress noise and inconsistencies.

Result: Extensive experiments on real-world multi-camera systems show SPARK outperforms existing approaches in extrinsic accuracy, geometric consistency, temporal stability, and real-time performance, with linear scalability to number of cameras.

Conclusion: SPARK demonstrates effectiveness and scalability for large-scale multi-camera 3D reconstruction by addressing key challenges of extrinsic uncertainty and multi-view fusion while maintaining real-time performance.

Abstract: Real-time multi-camera 3D reconstruction is crucial for 3D perception, immersive interaction, and robotics. Existing methods struggle with multi-view fusion, camera extrinsic uncertainty, and scalability for large camera setups. We propose SPARK, a self-calibrating real-time multi-camera point cloud reconstruction framework that jointly handles point cloud fusion and extrinsic uncertainty. SPARK consists of: (1) a geometry-aware online extrinsic estimation module leveraging multi-view priors and enforcing cross-view and temporal consistency for stable self-calibration, and (2) a confidence-driven point cloud fusion strategy modeling depth reliability and visibility at pixel and point levels to suppress noise and view-dependent inconsistencies. By performing frame-wise fusion without accumulation, SPARK produces stable point clouds in dynamic scenes while scaling linearly with the number of cameras. Extensive experiments on real-world multi-camera systems show that SPARK outperforms existing approaches in extrinsic accuracy, geometric consistency, temporal stability, and real-time performance, demonstrating its effectiveness and scalability for large-scale multi-camera 3D reconstruction.

Method: Uses modality-specific encoders for HSI and LiDAR, aligns visual features with handcrafted textual embeddings in shared latent space via bi-directional contrastive learning, inspired by CLIP’s training paradigm.

Result: Achieves strong performance with simple CNN-based encoders, outperforms established multimodal visual-only methods on two benchmark datasets, demonstrating significant benefit of language supervision.

Conclusion: MMLGNet successfully bridges the gap between high-dimensional remote sensing data and language-guided interpretation, showing that language supervision provides substantial benefits for multimodal remote sensing analysis.

Abstract: In this paper, we propose a novel multimodal framework, Multimodal Language-Guided Network (MMLGNet), to align heterogeneous remote sensing modalities like Hyperspectral Imaging (HSI) and LiDAR with natural language semantics using vision-language models such as CLIP. With the increasing availability of multimodal Earth observation data, there is a growing need for methods that effectively fuse spectral, spatial, and geometric information while enabling semantic-level understanding. MMLGNet employs modality-specific encoders and aligns visual features with handcrafted textual embeddings in a shared latent space via bi-directional contrastive learning. Inspired by CLIP’s training paradigm, our approach bridges the gap between high-dimensional remote sensing data and language-guided interpretation. Notably, MMLGNet achieves strong performance with simple CNN-based encoders, outperforming several established multimodal visual-only methods on two benchmark datasets, demonstrating the significant benefit of language supervision. Codes are available at https://github.com/AdityaChaudhary2913/CLIP_HSI.

Method: Three-module approach: 1) Automatic Patch Extraction Module extracts local patches from source video frames, 2) Reenactment Module generates corresponding re-enacted target local patches, 3) Weight Estimation Module calculates animation parameters for target character to create complete facial animation sequence.

Result: Extensive experiments show the method successfully transfers semantic meaning of source facial expressions to stylized characters with considerable variations in facial feature proportions.

Conclusion: The proposed local patch-based retargeting method effectively addresses the challenge of transferring facial animations to stylized 3D characters while preserving expression semantics, enabling lifelike animations for diverse virtual characters.

Abstract: In the era of digital animation, the quest to produce lifelike facial animations for virtual characters has led to the development of various retargeting methods. While the retargeting facial motion between models of similar shapes has been very successful, challenges arise when the retargeting is performed on stylized or exaggerated 3D characters that deviate significantly from human facial structures. In this scenario, it is important to consider the target character’s facial structure and possible range of motion to preserve the semantics assumed by the original facial motions after the retargeting. To achieve this, we propose a local patch-based retargeting method that transfers facial animations captured in a source performance video to a target stylized 3D character. Our method consists of three modules. The Automatic Patch Extraction Module extracts local patches from the source video frame. These patches are processed through the Reenactment Module to generate correspondingly re-enacted target local patches. The Weight Estimation Module calculates the animation parameters for the target character at every frame for the creation of a complete facial animation sequence. Extensive experiments demonstrate that our method can successfully transfer the semantic meaning of source facial expressions to stylized characters with considerable variations in facial feature proportion.

Method: Proposes Cardiac Reasoning Template (CRT) for stepwise canonical diagnostic procedures, and CardiacMind reinforcement learning scheme with three novel rewards: Procedural Quantity Reward (PQtR), Procedural Quality Reward (PQlR), and Echocardiographic Semantic Reward (ESR).

Result: 48% improvement in multiview echocardiographic diagnosis for 15 complex cardiac diseases, 5% improvement on CardiacNet-PAH benchmark, and 93.33% clinician agreement with cardiologist-like reasoning logic in user study.

Conclusion: The proposed approach effectively enhances MLLM’s echocardiographic reasoning by introducing cardiologist-like mindset through structured templates and reinforcement learning, achieving significant performance improvements and high clinical validation.

Abstract: Echocardiographic diagnosis is vital for cardiac screening yet remains challenging. Existing echocardiography foundation models do not effectively capture the relationships between quantitative measurements and clinical manifestations, whereas medical reasoning multimodal large language models (MLLMs) require costly construction of detailed reasoning paths and remain ineffective at directly incorporating such echocardiographic priors into their reasoning. To address these limitations, we propose a novel approach comprising Cardiac Reasoning Template (CRT) and CardiacMind to enhance MLLM’s echocardiographic reasoning by introducing cardiologist-like mindset. Specifically, CRT provides stepwise canonical diagnostic procedures for complex cardiac diseases to streamline reasoning path construction without the need for costly case-by-case verification. To incentivize reasoning MLLM under CRT, we develop CardiacMind, a new reinforcement learning scheme with three novel rewards: Procedural Quantity Reward (PQtR), Procedural Quality Reward (PQlR), and Echocardiographic Semantic Reward (ESR). PQtR promotes detailed reasoning; PQlR promotes integration of evidence across views and modalities, while ESR grounds stepwise descriptions in visual content. Our methods show a 48% improvement in multiview echocardiographic diagnosis for 15 complex cardiac diseases and a 5% improvement on CardiacNet-PAH over prior methods. The user study on our method’s reasoning outputs shows 93.33% clinician agreement with cardiologist-like reasoning logic. Our code will be available.

Method: NAR uses a semi-supervised learning framework with confidence-based label handling: dynamically retains high-confidence labels, temporarily deactivates moderate-confidence labels, and flips low-confidence labels. This selective attenuation integrates with early-learning regularization (ELR) to stabilize training and prevent overfitting to corrupted labels.

Result: Experiments across additive, subtractive, and mixed noise scenarios show NAR consistently outperforms existing methods, with most significant improvements under subtractive and mixed noise conditions, demonstrating effective noise robustness in remote sensing multi-label classification.

Conclusion: Adaptive suppression and selective correction of noisy supervision through NAR provides an effective strategy for noise-robust learning in remote sensing multi-label classification, particularly beneficial for handling the complex noise patterns from cost-effective annotation procedures.

Abstract: The development of reliable methods for multi-label classification (MLC) has become a prominent research direction in remote sensing (RS). As the scale of RS data continues to expand, annotation procedures increasingly rely on thematic products or crowdsourced procedures to reduce the cost of manual annotation. While cost-effective, these strategies often introduce multi-label noise in the form of partially incorrect annotations. In MLC, label noise arises as additive noise, subtractive noise, or a combination of both in the form of mixed noise. Previous work has largely overlooked this distinction and commonly treats noisy annotations as supervised signals, lacking mechanisms that explicitly adapt learning behavior to different noise types. To address this limitation, we propose NAR, a noise-adaptive regularization method that explicitly distinguishes between additive and subtractive noise within a semi-supervised learning framework. NAR employs a confidence-based label handling mechanism that dynamically retains label entries with high confidence, temporarily deactivates entries with moderate confidence, and corrects low confidence entries via flipping. This selective attenuation of supervision is integrated with early-learning regularization (ELR) to stabilize training and mitigate overfitting to corrupted labels. Experiments across additive, subtractive, and mixed noise scenarios demonstrate that NAR consistently improves robustness compared with existing methods. Performance improvements are most pronounced under subtractive and mixed noise, indicating that adaptive suppression and selective correction of noisy supervision provide an effective strategy for noise robust learning in RS MLC.

Method: Proposes Static-Dynamic Collaboration (SDC) framework with two stages: SRS trains initial model with base session data and preserves static memory; DLS introduces dynamic projector jointly trained with static memory to adapt to new classes.

Result: Extensive experiments on three public benchmarks and a real-world application dataset demonstrate state-of-the-art performance against other competitors.

Conclusion: SDC framework effectively balances stability and plasticity in FSCIL by separating static knowledge retention from dynamic adaptation, achieving superior performance in continuous learning scenarios.

Abstract: Few-shot class-incremental learning (FSCIL) aims to continuously recognize novel classes under limited data, which suffers from the key stability-plasticity dilemma: balancing the retention of old knowledge with the acquisition of new knowledge. To address this issue, we divide the task into two different stages and propose a framework termed Static-Dynamic Collaboration (SDC) to achieve a better trade-off between stability and plasticity. Specifically, our method divides the normal pipeline of FSCIL into Static Retaining Stage (SRS) and Dynamic Learning Stage (DLS), which harnesses old static and incremental dynamic class information, respectively. During SRS, we train an initial model with sufficient data in the base session and preserve the key part as static memory to retain fundamental old knowledge. During DLS, we introduce an extra dynamic projector jointly trained with the previous static memory. By employing both stages, our method achieves improved retention of old knowledge while continuously adapting to new classes. Extensive experiments on three public benchmarks and a real-world application dataset demonstrate that our method achieves state-of-the-art performance against other competitors.

Method: Developed a radiomics framework with automated randomization algorithm to mimic inter-observer variability for robust feature selection. Used pre- and post-NACT CT scans, extracted features from lesions in different anatomical sites, employed four response metrics (CRS, RECIST, VolR, DiaR), trained models on one cohort and externally validated on an independent cohort.

Result: Best prediction performance achieved using all lesions combined for volume reduction (VolR) prediction (AUC 0.83). Omental lesions best for CRS prediction (AUC 0.77), pelvic lesions best for diameter reduction (DiaR) prediction (AUC 0.76).

Conclusion: Integration of robustness into feature selection ensures reliable model development and facilitates clinical implementation of radiomics models for HGSOC patients. Future work should explore further applications in real-time clinical settings.

Abstract: Objectives: High-grade serous ovarian carcinoma (HGSOC) is typically diagnosed at an advanced stage with extensive peritoneal metastases, making treatment challenging. Neoadjuvant chemotherapy (NACT) is often used to reduce tumor burden before surgery, but about 40% of patients show limited response. Radiomics, combined with machine learning (ML), offers a promising non-invasive method for predicting NACT response by analyzing computed tomography (CT) imaging data. This study aimed to improve response prediction in HGSOC patients undergoing NACT by integration different feature selection methods. Materials and methods: A framework for selecting robust radiomics features was introduced by employing an automated randomisation algorithm to mimic inter-observer variability, ensuring a balance between feature robustness and prediction accuracy. Four response metrics were used: chemotherapy response score (CRS), RECIST, volume reduction (VolR), and diameter reduction (DiaR). Lesions in different anatomical sites were studied. Pre- and post-NACT CT scans were used for feature extraction and model training on one cohort, and an independent cohort was used for external testing. Results: The best prediction performance was achieved using all lesions combined for VolR prediction, with an AUC of 0.83. Omental lesions provided the best results for CRS prediction (AUC 0.77), while pelvic lesions performed best for DiaR (AUC 0.76). Conclusion: The integration of robustness into the feature selection processes ensures the development of reliable models and thus facilitates the implementation of the radiomics models in clinical applications for HGSOC patients. Future work should explore further applications of radiomics in ovarian cancer, particularly in real-time clinical settings.

Method: Uses DETR-like detectors as separate RGB and TIR branches with query fusion between them. High-quality queries from one branch are selected, adapted, and fed to the other branch to exclude degraded modality and correct predictions. Framework supports unpaired training data.

Result: Extensive experiments show superior performance over existing RGB-T detectors and better modality independence.

Conclusion: MDQF effectively balances modality complementation and separation, enabling robust detection under diverse conditions while eliminating the need for paired RGB-T data through its decoupled architecture.

Abstract: The advantage of RGB-Thermal (RGB-T) detection lies in its ability to perform modality fusion and integrate cross-modality complementary information, enabling robust detection under diverse illumination and weather conditions. However, under extreme conditions where one modality exhibits poor quality and disturbs detection, modality separation is necessary to mitigate the impact of noise. To address this problem, we propose a Modality-Decoupled RGB-T detection framework with Query Fusion (MDQF) to balance modality complementation and separation. In this framework, DETR-like detectors are employed as separate branches for the RGB and TIR images, with query fusion interspersed between the two branches in each refinement stage. Herein, query fusion is performed by feeding the high-quality queries from one branch to the other one after query selection and adaptation. This design effectively excludes the degraded modality and corrects the predictions using high-quality queries. Moreover, the decoupled framework allows us to optimize each individual branch with unpaired RGB or TIR images, eliminating the need for paired RGB-T data. Extensive experiments demonstrate that our approach delivers superior performance to existing RGB-T detectors and achieves better modality independence.

Method: 1) Created CoMa-20K dataset containing massing geometries, economic/programmatic data, and visual site context representations. 2) Formulated massing generation as conditional task for Vision-Language Models. 3) Evaluated both fine-tuned and large zero-shot VLMs on this dataset.

Result: Experiments revealed the inherent complexity of the massing generation task while demonstrating VLMs’ potential to produce context-sensitive massing options. The dataset establishes a foundational benchmark for data-driven architectural design.

Conclusion: The CoMa-20K dataset and VLM benchmarking provide a foundation for automated architectural design research, highlighting significant opportunities for future work in data-driven building massing generation.

Abstract: The conceptual design phase in architecture and urban planning, particularly building massing, is complex and heavily reliant on designer intuition and manual effort. To address this, we propose an automated framework for generating building massing based on functional requirements and site context. A primary obstacle to such data-driven methods has been the lack of suitable datasets. Consequently, we introduce the CoMa-20K dataset, a comprehensive collection that includes detailed massing geometries, associated economical and programmatic data, and visual representations of the development site within its existing urban context. We benchmark this dataset by formulating massing generation as a conditional task for Vision-Language Models (VLMs), evaluating both fine-tuned and large zero-shot models. Our experiments reveal the inherent complexity of the task while demonstrating the potential of VLMs to produce context-sensitive massing options. The dataset and analysis establish a foundational benchmark and highlight significant opportunities for future research in data-driven architectural design.

Method: Developed HazardForge - a scalable pipeline using image editing models with layout decision algorithms and validation modules to generate realistic hazardous scenarios. Created MovSafeBench benchmark with 7,254 images and QA pairs across 13 object categories covering normal and anomalous objects.

Result: VLM performance degrades notably under conditions including anomalous objects, with the largest performance drop in scenarios requiring nuanced motion understanding.

Conclusion: HazardForge enables systematic evaluation of VLMs’ safety capabilities in complex environments, revealing significant performance gaps in handling anomalous scenarios with spatio-temporal dynamics.

Abstract: Vision Language Models (VLMs) are increasingly deployed in autonomous vehicles and mobile systems, making it crucial to evaluate their ability to support safer decision-making in complex environments. However, existing benchmarks inadequately cover diverse hazardous situations, especially anomalous scenarios with spatio-temporal dynamics. While image editing models are a promising means to synthesize such hazards, it remains challenging to generate well-formulated scenarios that include moving, intrusive, and distant objects frequently observed in the real world. To address this gap, we introduce \textbf{HazardForge}, a scalable pipeline that leverages image editing models to generate these scenarios with layout decision algorithms, and validation modules. Using HazardForge, we construct \textbf{MovSafeBench}, a multiple-choice question (MCQ) benchmark comprising 7,254 images and corresponding QA pairs across 13 object categories, covering both normal and anomalous objects. Experiments using MovSafeBench show that VLM performance degrades notably under conditions including anomalous objects, with the largest drop in scenarios requiring nuanced motion understanding.

Method: Developed an IoT-based system using ESP32 microcontroller integrated with pH, TDS, temperature, and turbidity sensors, along with servo feeder and water pump actuators. Implemented edge processing, cloud connectivity via Blynk IoT platform, and intelligent alert mechanisms with configurable cooldown periods.

Result: System achieved 96% average sensor accuracy, 1.2-second response time for anomaly detection, and 97% operational reliability for automated feeding and water circulation modules. Successfully maintained stable aquatic conditions in 10-liter aquarium testing.

Conclusion: Cost-effective IoT solutions can revolutionize aquarium maintenance by making aquatic ecosystem management more accessible, reliable, and efficient for both residential and commercial applications, significantly reducing manual intervention.

Abstract: Maintaining optimal water quality in aquariums is critical for aquatic health but remains challenging due to the need for continuous monitoring of multiple parameters. Traditional manual methods are inefficient, labor-intensive, and prone to human error, often leading to suboptimal aquatic conditions. This paper presents an IoT-based smart aquarium system that addresses these limitations by integrating an ESP32 microcontroller with multiple sensors (pH, TDS, temperature, turbidity) and actuators (servo feeder, water pump) for comprehensive real-time water quality monitoring and automated control. The system architecture incorporates edge processing capabilities, cloud connectivity via Blynk IoT platform, and an intelligent alert mechanism with configurable cooldown periods to prevent notification fatigue. Experimental evaluation in a 10-liter aquarium environment demonstrated the system’s effectiveness, achieving 96% average sensor accuracy and 1.2-second response time for anomaly detection. The automated feeding and water circulation modules maintained 97% operational reliability throughout extended testing, significantly reducing manual intervention while ensuring stable aquatic conditions. This research demonstrates that cost-effective IoT solutions can revolutionize aquarium maintenance, making aquatic ecosystem management more accessible, reliable, and efficient for both residential and commercial applications.

Method: PKI consists of backbone, ensemble of projectors, classifier, and memory. Each incremental session adds a new projector to the ensemble, then jointly finetunes only the new projector and classifier while keeping other components frozen. Two variants (PKIV-1, PKIV-2) reduce projectors for resource efficiency.

Result: Extensive experiments on three popular benchmarks show PKI outperforms state-of-the-art methods in FSCIL.

Conclusion: The cascading projector approach effectively integrates prior knowledge from previous sessions while learning new knowledge flexibly, achieving better balance between preserving old classes and learning new ones in few-shot incremental learning scenarios.

Abstract: Few-shot class-incremental learning (FSCIL) aims to continually adapt a model on a limited number of new-class examples, facing two well-known challenges: catastrophic forgetting and overfitting to new classes. Existing methods tend to freeze more parts of network components and finetune others with an extra memory during incremental sessions. These methods emphasize preserving prior knowledge to ensure proficiency in recognizing old classes, thereby mitigating catastrophic forgetting. Meanwhile, constraining fewer parameters can help in overcoming overfitting with the assistance of prior knowledge. Following previous methods, we retain more prior knowledge and propose a prior knowledge-infused neural network (PKI) to facilitate FSCIL. PKI consists of a backbone, an ensemble of projectors, a classifier, and an extra memory. In each incremental session, we build a new projector and add it to the ensemble. Subsequently, we finetune the new projector and the classifier jointly with other frozen network components, ensuring the rich prior knowledge is utilized effectively. By cascading projectors, PKI integrates prior knowledge accumulated from previous sessions and learns new knowledge flexibly, which helps to recognize old classes and efficiently learn new classes. Further, to reduce the resource consumption associated with keeping many projectors, we design two variants of the prior knowledge-infused neural network (PKIV-1 and PKIV-2) to trade off a balance between resource consumption and performance by reducing the number of projectors. Extensive experiments on three popular benchmarks demonstrate that our approach outperforms state-of-the-art methods.

Method: Proposes EfficientFSL with three key components: 1) Lightweight trainable Forward Block to synthesize task-specific queries that extract features from pre-trained model representations in query-only manner, 2) Combine Block to fuse multi-layer outputs for enhanced feature depth and robustness, and 3) Support-Query Attention Block to mitigate distribution shift by adjusting prototypes to align with query set distribution.

Result: Achieves state-of-the-art performance on four in-domain few-shot datasets and six cross-domain datasets while using minimal trainable parameters, demonstrating effectiveness in real-world applications.

Conclusion: EfficientFSL successfully bridges the gap between performance and computational efficiency in few-shot learning with Vision Transformers, offering a practical solution for low-resource scenarios while maintaining competitive accuracy.

Abstract: Large models such as Vision Transformers (ViTs) have demonstrated remarkable superiority over smaller architectures like ResNet in few-shot classification, owing to their powerful representational capacity. However, fine-tuning such large models demands extensive GPU memory and prolonged training time, making them impractical for many real-world low-resource scenarios. To bridge this gap, we propose EfficientFSL, a query-only fine-tuning framework tailored specifically for few-shot classification with ViT, which achieves competitive performance while significantly reducing computational overhead. EfficientFSL fully leverages the knowledge embedded in the pre-trained model and its strong comprehension ability, achieving high classification accuracy with an extremely small number of tunable parameters. Specifically, we introduce a lightweight trainable Forward Block to synthesize task-specific queries that extract informative features from the intermediate representations of the pre-trained model in a query-only manner. We further propose a Combine Block to fuse multi-layer outputs, enhancing the depth and robustness of feature representations. Finally, a Support-Query Attention Block mitigates distribution shift by adjusting prototypes to align with the query set distribution. With minimal trainable parameters, EfficientFSL achieves state-of-the-art performance on four in-domain few-shot datasets and six cross-domain datasets, demonstrating its effectiveness in real-world applications.

Method: Formulate channel configuration search as sequence of conditional code generation tasks. Generate large corpus of valid architectures via Abstract Syntax Tree (AST) mutations to overcome data scarcity. Use LLM to refine architectural specifications based on performance telemetry, allowing it to learn relationships between channel configurations and model performance.

Result: Experimental results on CIFAR-100 show statistically significant improvements in accuracy. The LLM successfully acquires domain-specific architectural priors, distinguishing the method from random search.

Conclusion: LLM-driven NAS framework demonstrates transformative potential for channel configuration optimization. The approach shows language-driven design can effectively learn complex architectural patterns and outperforms traditional methods, highlighting the promise of LLMs in deep learning architecture design.

Abstract: Channel configuration search the optimization of layer specifications such as layer widths in deep neural networks presents a complex combinatorial challenge constrained by tensor shape compatibility and computational budgets. We posit that Large Language Models (LLMs) offer a transformative approach to Neural Architecture Search (NAS), capable of reasoning about architectural code structure in ways that traditional heuristics cannot. In this paper, we investigate the application of an LLM-driven NAS framework to the problem of channel configuration. We formulate the search as a sequence of conditional code generation tasks, where an LLM refines architectural specifications based on performance telemetry. Crucially, we address the data scarcity problem by generating a vast corpus of valid, shape-consistent architectures via Abstract Syntax Tree (AST) mutations. While these mutated networks are not necessarily high-performing, they provide the critical volume of structural data required for the LLM to learn the latent relationship between channel configurations and model performance. This allows the LLM to internalize complex design patterns and apply them to optimize feature extraction strategies. Experimental results on CIFAR-100 validate the efficacy of this approach, demonstrating that the model yields statistically significant improvements in accuracy. Our analysis confirms that the LLM successfully acquires domain-specific architectural priors, distinguishing this method from random search and highlighting the immense potential of language-driven design in deep learning.

Method: Proposes Constrained Dataset Distillation (CD²) with two modules: Dataset Distillation Module (DDM) synthesizes condensed samples guided by classifier to learn compact essential class-related clues; Distillation Constraint Module (DCM) introduces designed loss to constrain previously learned class distribution to preserve distilled knowledge.

Result: Extensive experiments on three public datasets demonstrate superiority over other state-of-the-art competitors.

Conclusion: The CD² framework effectively addresses catastrophic forgetting in FSCIL by distilling essential knowledge and constraining class distributions, achieving better performance than existing methods.

Abstract: Few-shot class-incremental learning (FSCIL) receives significant attention from the public to perform classification continuously with a few training samples, which suffers from the key catastrophic forgetting problem. Existing methods usually employ an external memory to store previous knowledge and treat it with incremental classes equally, which cannot properly preserve previous essential knowledge. To solve this problem and inspired by recent distillation works on knowledge transfer, we propose a framework termed \textbf{C}onstrained \textbf{D}ataset \textbf{D}istillation (\textbf{CD$^2$}) to facilitate FSCIL, which includes a dataset distillation module (\textbf{DDM}) and a distillation constraint module~(\textbf{DCM}). Specifically, the DDM synthesizes highly condensed samples guided by the classifier, forcing the model to learn compacted essential class-related clues from a few incremental samples. The DCM introduces a designed loss to constrain the previously learned class distribution, which can preserve distilled knowledge more sufficiently. Extensive experiments on three public datasets show the superiority of our method against other state-of-the-art competitors.

Method: Generates multiple answers from clean and perturbed video clips, clusters them into semantic hypotheses using NLI or embedding methods, then computes three reliability scores: Semantic Entropy, RadFlag, and Vision-Amplified Semantic Entropy.

Result: VASE achieves highest ROC-AUC across three 7B Video-VLMs, especially with larger distortion budgets. Embedding-based clustering matches NLI performance at lower cost, and domain fine-tuning reduces hallucinations but offers modest calibration improvements.

Conclusion: VideoHEDGE provides an effective framework for hallucination detection in video QA, with VASE as the most reliable metric, and offers practical tools through the hedge-bench library for reproducible benchmarking.

Abstract: Hallucinations in video-capable vision-language models (Video-VLMs) remain frequent and high-confidence, while existing uncertainty metrics often fail to align with correctness. We introduce VideoHEDGE, a modular framework for hallucination detection in video question answering that extends entropy-based reliability estimation from images to temporally structured inputs. Given a video-question pair, VideoHEDGE draws a baseline answer and multiple high-temperature generations from both clean clips and photometrically and spatiotemporally perturbed variants, then clusters the resulting textual outputs into semantic hypotheses using either Natural Language Inference (NLI)-based or embedding-based methods. Cluster-level probability masses yield three reliability scores: Semantic Entropy (SE), RadFlag, and Vision-Amplified Semantic Entropy (VASE). We evaluate VideoHEDGE on the SoccerChat benchmark using an LLM-as-a-judge to obtain binary hallucination labels. Across three 7B Video-VLMs (Qwen2-VL, Qwen2.5-VL, and a SoccerChat-finetuned model), VASE consistently achieves the highest ROC-AUC, especially at larger distortion budgets, while SE and RadFlag often operate near chance. We further show that embedding-based clustering matches NLI-based clustering in detection performance at substantially lower computational cost, and that domain fine-tuning reduces hallucination frequency but yields only modest improvements in calibration. The hedge-bench PyPI library enables reproducible and extensible benchmarking, with full code and experimental resources available at https://github.com/Simula/HEDGE#videohedge .

Method: Proposes REVNET built on Vector Neuron networks, representing partial point clouds as equivariant anchors and using a VN Missing Anchor Transformer to predict missing anchors. Extends VN networks with rotation-equivariant bias formulation and ZCA-based layer normalization for better feature expressiveness.

Result: Outperforms state-of-the-art methods on synthetic MVP dataset in equivariant setting. Achieves competitive results on real-world KITTI dataset without requiring input pose alignment, unlike non-equivariant networks.

Conclusion: REVNET provides a robust solution for point cloud completion under arbitrary rotations, eliminating the need for pose alignment while maintaining competitive performance with state-of-the-art methods.

Abstract: Incomplete point clouds captured by 3D sensors often result in the loss of both geometric and semantic information. Most existing point cloud completion methods are built on rotation-variant frameworks trained with data in canonical poses, limiting their applicability in real-world scenarios. While data augmentation with random rotations can partially mitigate this issue, it significantly increases the learning burden and still fails to guarantee robust performance under arbitrary poses. To address this challenge, we propose the Rotation-Equivariant Anchor Transformer (REVNET), a novel framework built upon the Vector Neuron (VN) network for robust point cloud completion under arbitrary rotations. To preserve local details, we represent partial point clouds as sets of equivariant anchors and design a VN Missing Anchor Transformer to predict the positions and features of missing anchors. Furthermore, we extend VN networks with a rotation-equivariant bias formulation and a ZCA-based layer normalization to improve feature expressiveness. Leveraging the flexible conversion between equivariant and invariant VN features, our model can generate point coordinates with greater stability. Experimental results show that our method outperforms state-of-the-art approaches on the synthetic MVP dataset in the equivariant setting. On the real-world KITTI dataset, REVNET delivers competitive results compared to non-equivariant networks, without requiring input pose alignment. The source code will be released on GitHub under URL: https://github.com/nizhf/REVNET.

Method: MoCha introduces a condition-aware RoPE to adapt multi-modal input conditions and enhance facial identity, employs an RL-based post-training stage, and uses a comprehensive data construction pipeline with three specialized datasets: a high-fidelity rendered dataset built with Unreal Engine 5, an expression-driven dataset synthesized by portrait animation techniques, and an augmented dataset from existing video-mask pairs.

Result: Extensive experiments demonstrate that MoCha substantially outperforms existing state-of-the-art approaches in video character replacement tasks.

Conclusion: MoCha presents a pioneering framework that overcomes data scarcity and complex scenario limitations in video character replacement by requiring only a single arbitrary frame mask, with plans to release code to facilitate further research.

Abstract: Controllable video character replacement with a user-provided identity remains a challenging problem due to the lack of paired video data. Prior works have predominantly relied on a reconstruction-based paradigm that requires per-frame segmentation masks and explicit structural guidance (e.g., skeleton, depth). This reliance, however, severely limits their generalizability in complex scenarios involving occlusions, character-object interactions, unusual poses, or challenging illumination, often leading to visual artifacts and temporal inconsistencies. In this paper, we propose MoCha, a pioneering framework that bypasses these limitations by requiring only a single arbitrary frame mask. To effectively adapt the multi-modal input condition and enhance facial identity, we introduce a condition-aware RoPE and employ an RL-based post-training stage. Furthermore, to overcome the scarcity of qualified paired-training data, we propose a comprehensive data construction pipeline. Specifically, we design three specialized datasets: a high-fidelity rendered dataset built with Unreal Engine 5 (UE5), an expression-driven dataset synthesized by current portrait animation techniques, and an augmented dataset derived from existing video-mask pairs. Extensive experiments demonstrate that our method substantially outperforms existing state-of-the-art approaches. We will release the code to facilitate further research. Please refer to our project page for more details: orange-3dv-team.github.io/MoCha

Method: Treat feature maps as spatial signals evolving over propagation time (network depth) governed by an underdamped wave equation. Derive closed-form frequency-time decoupled solution implemented as Wave Propagation Operator (WPO) with O(N log N) complexity. Build WaveFormer models as drop-in replacements for ViTs and CNNs.

Result: Competitive accuracy across image classification, object detection, and semantic segmentation. Achieves up to 1.6x higher throughput and 30% fewer FLOPs than attention-based alternatives. Captures both global coherence and high-frequency details better than heat-based methods.

Conclusion: Wave propagation provides a principled, efficient alternative to attention for vision modeling, offering complementary modeling bias that effectively captures rich visual semantics while being computationally efficient.

Abstract: Vision modeling has advanced rapidly with Transformers, whose attention mechanisms capture visual dependencies but lack a principled account of how semantic information propagates spatially. We revisit this problem from a wave-based perspective: feature maps are treated as spatial signals whose evolution over an internal propagation time (aligned with network depth) is governed by an underdamped wave equation. In this formulation, spatial frequency-from low-frequency global layout to high-frequency edges and textures-is modeled explicitly, and its interaction with propagation time is controlled rather than implicitly fixed. We derive a closed-form, frequency-time decoupled solution and implement it as the Wave Propagation Operator (WPO), a lightweight module that models global interactions in O(N log N) time-far lower than attention. Building on WPO, we propose a family of WaveFormer models as drop-in replacements for standard ViTs and CNNs, achieving competitive accuracy across image classification, object detection, and semantic segmentation, while delivering up to 1.6x higher throughput and 30% fewer FLOPs than attention-based alternatives. Furthermore, our results demonstrate that wave propagation introduces a complementary modeling bias to heat-based methods, effectively capturing both global coherence and high-frequency details essential for rich visual semantics. Codes are available at: https://github.com/ZishanShu/WaveFormer.

Method: Two complementary strategies: 1) Radiomic fingerprints - dynamically construct patient-specific feature sets using image-conditioned predictor to estimate feature relevance probabilities, with transparent logistic regression; 2) Healthy personas - synthesize pathology-free baselines using diffusion models trained on healthy knee MRIs, comparing pathological features against synthesized healthy versions.

Result: Both approaches yield performance comparable to or surpassing state-of-the-art DL models across three clinical tasks, while supporting multi-level interpretability. Case studies demonstrate human-explainable biomarker discovery and pathology localization.

Conclusion: The proposed radiomic fingerprints and healthy personas provide interpretable alternatives to black-box deep learning for knee MRI assessment, balancing clinical accuracy with explainability needed for medical adoption.

Abstract: For automated assessment of knee MRI scans, both accuracy and interpretability are essential for clinical use and adoption. Traditional radiomics rely on predefined features chosen at the population level; while more interpretable, they are often too restrictive to capture patient-specific variability and can underperform end-to-end deep learning (DL). To address this, we propose two complementary strategies that bring individuality and interpretability: radiomic fingerprints and healthy personas. First, a radiomic fingerprint is a dynamically constructed, patient-specific feature set derived from MRI. Instead of applying a uniform population-level signature, our model predicts feature relevance from a pool of candidate features and selects only those most predictive for each patient, while maintaining feature-level interpretability. This fingerprint can be viewed as a latent-variable model of feature usage, where an image-conditioned predictor estimates usage probabilities and a transparent logistic regression with global coefficients performs classification. Second, a healthy persona synthesises a pathology-free baseline for each patient using a diffusion model trained to reconstruct healthy knee MRIs. Comparing features extracted from pathological images against their personas highlights deviations from normal anatomy, enabling intuitive, case-specific explanations of disease manifestations. We systematically compare fingerprints, personas, and their combination across three clinical tasks. Experimental results show that both approaches yield performance comparable to or surpassing state-of-the-art DL models, while supporting interpretability at multiple levels. Case studies further illustrate how these perspectives facilitate human-explainable biomarker discovery and pathology localisation.

Method: Proposes SfMamba with Channel-wise Visual State-Space block for channel-sequence scanning to extract domain-invariant features, and Semantic-Consistent Shuffle strategy that disrupts background patch sequences while preserving prediction consistency to mitigate error accumulation.

Result: Comprehensive evaluations across multiple benchmarks show SfMamba achieves consistently stronger performance than existing methods while maintaining favorable parameter efficiency.

Conclusion: SfMamba offers a practical solution for SFDA by effectively exploring stable dependencies in source-free model transfer through improved Mamba architecture adaptations.

Abstract: Source-free domain adaptation (SFDA) tackles the critical challenge of adapting source-pretrained models to unlabeled target domains without access to source data, overcoming data privacy and storage limitations in real-world applications. However, existing SFDA approaches struggle with the trade-off between perception field and computational efficiency in domain-invariant feature learning. Recently, Mamba has offered a promising solution through its selective scan mechanism, which enables long-range dependency modeling with linear complexity. However, the Visual Mamba (i.e., VMamba) remains limited in capturing channel-wise frequency characteristics critical for domain alignment and maintaining spatial robustness under significant domain shifts. To address these, we propose a framework called SfMamba to fully explore the stable dependency in source-free model transfer. SfMamba introduces Channel-wise Visual State-Space block that enables channel-sequence scanning for domain-invariant feature extraction. In addition, SfMamba involves a Semantic-Consistent Shuffle strategy that disrupts background patch sequences in 2D selective scan while preserving prediction consistency to mitigate error accumulation. Comprehensive evaluations across multiple benchmarks show that SfMamba achieves consistently stronger performance than existing methods while maintaining favorable parameter efficiency, offering a practical solution for SFDA. Our code is available at https://github.com/chenxi52/SfMamba.

Method: Proposes Semantic Orthogonal Calibration (SoC), a Huber-based regularizer that enforces smooth prototype separation while preserving semantic proximity between related classes, addressing the overconfidence problem caused by full orthogonality constraints.

Result: SoC consistently improves calibration performance across comprehensive empirical validation while maintaining competitive discriminative capabilities compared to prior orthogonality-based approaches.

Conclusion: Semantic-aware orthogonal constraints are crucial for proper VLM calibration, as full orthogonality pushes related classes apart and causes overconfidence, while SoC effectively balances separation and semantic preservation.

Abstract: With the increasing adoption of vision-language models (VLMs) in critical decision-making systems such as healthcare or autonomous driving, the calibration of their uncertainty estimates becomes paramount. Yet, this dimension has been largely underexplored in the VLM test-time prompt-tuning (TPT) literature, which has predominantly focused on improving their discriminative performance. Recent state-of-the-art advocates for enforcing full orthogonality over pairs of text prompt embeddings to enhance separability, and therefore calibration. Nevertheless, as we theoretically show in this work, the inherent gradients from fully orthogonal constraints will strongly push semantically related classes away, ultimately making the model overconfident. Based on our findings, we propose Semantic Orthogonal Calibration (SoC), a Huber-based regularizer that enforces smooth prototype separation while preserving semantic proximity, thereby improving calibration compared to prior orthogonality-based approaches. Across a comprehensive empirical validation, we demonstrate that SoC consistently improves calibration performance, while also maintaining competitive discriminative capabilities.

Method: Proposes CtrlFuse with three components: multi-modal feature extractor, reference prompt encoder (RPE) that dynamically encodes task-specific semantic prompts using pre-trained segmentation models with mask guidance, and prompt-semantic fusion module (PSFM) that explicitly injects these semantics into fusion features. Uses synergistic optimization of parallel segmentation and fusion branches.

Result: Achieves state-of-the-art results in both fusion controllability and segmentation accuracy. The adapted task branch even outperforms the original segmentation model.

Conclusion: CtrlFuse provides an effective controllable fusion framework that enables interactive dynamic fusion guided by mask prompts, achieving mutual enhancement between task performance and fusion quality for improved environmental awareness in intelligent unmanned systems.

Abstract: Infrared and visible image fusion generates all-weather perception-capable images by combining complementary modalities, enhancing environmental awareness for intelligent unmanned systems. Existing methods either focus on pixel-level fusion while overlooking downstream task adaptability or implicitly learn rigid semantics through cascaded detection/segmentation models, unable to interactively address diverse semantic target perception needs. We propose CtrlFuse, a controllable image fusion framework that enables interactive dynamic fusion guided by mask prompts. The model integrates a multi-modal feature extractor, a reference prompt encoder (RPE), and a prompt-semantic fusion module (PSFM). The RPE dynamically encodes task-specific semantic prompts by fine-tuning pre-trained segmentation models with input mask guidance, while the PSFM explicitly injects these semantics into fusion features. Through synergistic optimization of parallel segmentation and fusion branches, our method achieves mutual enhancement between task performance and fusion quality. Experiments demonstrate state-of-the-art results in both fusion controllability and segmentation accuracy, with the adapted task branch even outperforming the original segmentation model.

Method: SafeRedir uses prompt embedding redirection with two components: 1) latent-aware multi-modal safety classifier to identify unsafe generation trajectories, and 2) token-level delta generator for semantic redirection with auxiliary predictors for token masking and adaptive scaling.

Result: Achieves effective unlearning capability, high semantic/perceptual preservation, robust image quality, and enhanced resistance to adversarial attacks across multiple unlearning tasks. Generalizes across various diffusion backbones and existing unlearned models.

Conclusion: SafeRedir provides a plug-and-play, inference-time solution for robust unlearning that addresses limitations of existing methods while maintaining compatibility and broad applicability across different image generation models.

Abstract: Image generation models (IGMs), while capable of producing impressive and creative content, often memorize a wide range of undesirable concepts from their training data, leading to the reproduction of unsafe content such as NSFW imagery and copyrighted artistic styles. Such behaviors pose persistent safety and compliance risks in real-world deployments and cannot be reliably mitigated by post-hoc filtering, owing to the limited robustness of such mechanisms and a lack of fine-grained semantic control. Recent unlearning methods seek to erase harmful concepts at the model level, which exhibit the limitations of requiring costly retraining, degrading the quality of benign generations, or failing to withstand prompt paraphrasing and adversarial attacks. To address these challenges, we introduce SafeRedir, a lightweight inference-time framework for robust unlearning via prompt embedding redirection. Without modifying the underlying IGMs, SafeRedir adaptively routes unsafe prompts toward safe semantic regions through token-level interventions in the embedding space. The framework comprises two core components: a latent-aware multi-modal safety classifier for identifying unsafe generation trajectories, and a token-level delta generator for precise semantic redirection, equipped with auxiliary predictors for token masking and adaptive scaling to localize and regulate the intervention. Empirical results across multiple representative unlearning tasks demonstrate that SafeRedir achieves effective unlearning capability, high semantic and perceptual preservation, robust image quality, and enhanced resistance to adversarial attacks. Furthermore, SafeRedir generalizes effectively across a variety of diffusion backbones and existing unlearned models, validating its plug-and-play compatibility and broad applicability. Code and data are available at https://github.com/ryliu68/SafeRedir.

Method: Proposes a reliability assessment framework with four evaluation pillars: transferability (performance across different datasets), robustness (against adversarial attacks), interpretability (explainability of decisions), and computational efficiency (resource requirements). Analyzes five state-of-the-art deepfake detection methods using this framework.

Result: Analysis of five SOTA methods revealed both significant progress and critical limitations across the four reliability dimensions, highlighting areas needing improvement in current deepfake detection approaches.

Conclusion: A comprehensive reliability assessment framework is needed to better evaluate deepfake detection methods beyond just classification accuracy, addressing transferability, robustness, interpretability, and efficiency for more practical deployment.

Abstract: Deepfakes are synthetic media generated by artificial intelligence, with positive applications in education and creativity, but also serious negative impacts such as fraud, misinformation, and privacy violations. Although detection techniques have advanced, comprehensive evaluation methods that go beyond classification performance remain lacking. This paper proposes a reliability assessment framework based on four pillars: transferability, robustness, interpretability, and computational efficiency. An analysis of five state-of-the-art methods revealed significant progress as well as critical limitations.

Method: Salience-SGG framework with Iterative Salience Decoder (ISD) that emphasizes triplets with salient spatial structures, guided by semantic-agnostic salience labels to improve spatial understanding without compromising debiasing.

Result: Achieves state-of-the-art performance on Visual Genome, Open Images V6, and GQA-200 datasets. Improves existing Unbiased-SGG methods in spatial understanding as demonstrated by Pairwise Localization Average Precision.

Conclusion: Salience-SGG effectively addresses the long-tailed distribution problem in SGG while maintaining strong spatial understanding, overcoming limitations of previous debiasing methods that sacrificed spatial comprehension for better rare relation performance.

Abstract: Scene Graph Generation (SGG) suffers from a long-tailed distribution, where a few predicate classes dominate while many others are underrepresented, leading to biased models that underperform on rare relations. Unbiased-SGG methods address this issue by implementing debiasing strategies, but often at the cost of spatial understanding, resulting in an over-reliance on semantic priors. We introduce Salience-SGG, a novel framework featuring an Iterative Salience Decoder (ISD) that emphasizes triplets with salient spatial structures. To support this, we propose semantic-agnostic salience labels guiding ISD. Evaluations on Visual Genome, Open Images V6, and GQA-200 show that Salience-SGG achieves state-of-the-art performance and improves existing Unbiased-SGG methods in their spatial understanding as demonstrated by the Pairwise Localization Average Precision

Method: Developed ISLA through systematic optimization of loss function, convolutional architecture, deep supervision, and attention mechanisms. Used three multicenter databases (1500+ participants) and investigated unsupervised domain adaptation for better generalization to external clinical data.

Result: ISLA outperformed two state-of-the-art approaches for AIS lesion segmentation on an external test set.

Conclusion: ISLA provides a robust segmentation framework for acute ischemic stroke lesions, with codes and trained models to be made publicly available to facilitate reuse and reproducibility in clinical applications.

Abstract: Accurate delineation of acute ischemic stroke lesions in MRI is a key component of stroke diagnosis and management. In recent years, deep learning models have been successfully applied to the automatic segmentation of such lesions. While most proposed architectures are based on the U-Net framework, they primarily differ in their choice of loss functions and in the use of deep supervision, residual connections, and attention mechanisms. Moreover, many implementations are not publicly available, and the optimal configuration for acute ischemic stroke (AIS) lesion segmentation remains unclear. In this work, we introduce ISLA (Ischemic Stroke Lesion Analyzer), a new deep learning model for AIS lesion segmentation from diffusion MRI, trained on three multicenter databases totaling more than 1500 AIS participants. Through systematic optimization of the loss function, convolutional architecture, deep supervision, and attention mechanisms, we developed a robust segmentation framework. We further investigated unsupervised domain adaptation to improve generalization to an external clinical dataset. ISLA outperformed two state-of-the-art approaches for AIS lesion segmentation on an external test set. Codes and trained models will be made publicly available to facilitate reuse and reproducibility.

Method: 1) Created UR-Bench with two categories (Humanistic Scenes, Natural Scenes) and four subsets of ultra-high-resolution images; 2) Developed agent-based framework where LLM invokes external visual tools; 3) Introduced Semantic Abstraction and Retrieval tools for efficient ultra-high-resolution image processing.

Result: Benchmark enables evaluation of MLLMs’ reasoning capabilities in ultra-high-resolution scenarios. Agent-based framework with visual tools demonstrates effectiveness compared to end-to-end MLLMs.

Conclusion: UR-Bench addresses the gap in evaluating MLLMs on ultra-high-resolution imagery, and the proposed agent-based framework with specialized visual tools offers an effective approach for processing extreme visual information.

Abstract: Recent multimodal large language models (MLLMs) show strong capabilities in visual-language reasoning, yet their performance on ultra-high-resolution imagery remains largely unexplored. Existing visual question answering (VQA) benchmarks typically rely on medium-resolution data, offering limited visual complexity. To bridge this gap, we introduce Ultra-high-resolution Reasoning Benchmark (UR-Bench), a benchmark designed to evaluate the reasoning capabilities of MLLMs under extreme visual information. UR-Bench comprises two major categories, Humanistic Scenes and Natural Scenes, covering four subsets of ultra-high-resolution images with distinct spatial structures and data sources. Each subset contains images ranging from hundreds of megapixels to gigapixels, accompanied by questions organized into three levels, enabling evaluation of models’ reasoning capabilities in ultra-high-resolution scenarios. We further propose an agent-based framework in which a language model performs reasoning by invoking external visual tools. In addition, we introduce Semantic Abstraction and Retrieval tools that enable more efficient processing of ultra-high-resolution images. We evaluate state-of-the-art models using both an end-to-end MLLMs and our agent-based framework, demonstrating the effectiveness of our framework.

Method: Uses CycleGAN for unpaired image-to-image translation, combining C01 and C02 fluorescence channels into RGB format. Architecture employs ResNet-based generators with residual blocks and PatchGAN discriminators, trained with adversarial, cycle-consistency, and identity losses.

Result: The model generates realistic pseudo H&E images that preserve morphological structures from fluorescence data while adopting H&E-like color characteristics, enabling visualization of fluorescence data in a format familiar to pathologists.

Conclusion: The approach successfully bridges fluorescence microscopy and standard histopathology by converting fluorescence images to pseudo H&E format, supporting integration with existing H&E-based analysis pipelines and aiding pathologist interpretation.

Abstract: Histopathology analysis relies on Hematoxylin and Eosin (H&E) staining, but fluorescence microscopy offers complementary information. Converting fluorescence images to H&E-like appearance can aid interpretation and integration with standard workflows. We present a Cycle-Consistent Adversarial Network (CycleGAN) approach for unpaired image-to-image translation from multi-channel fluorescence microscopy to pseudo H&E stained histopathology images. The method combines C01 and C02 fluorescence channels into RGB and learns a bidirectional mapping between fluorescence and H&E domains without paired training data. The architecture uses ResNet-based generators with residual blocks and PatchGAN discriminators, trained with adversarial, cycle-consistency, and identity losses. Experiments on fluorescence microscopy datasets show the model generates realistic pseudo H&E images that preserve morphological structures while adopting H&E-like color characteristics. This enables visualization of fluorescence data in a format familiar to pathologists and supports integration with existing H&E-based analysis pipelines.

Method: Proposes MCAN framework with mixture-of-encoders adapter to dynamically process three complementary cues: input image (overall content), high-frequency components (edge details), and Chromatic Inconsistency cue (normalized intensity values capturing noise patterns from real image acquisition).

Result: State-of-the-art performance on GenImage, Chameleon, and UniversalFakeDetect benchmarks. On GenImage, outperforms best existing method by up to 7.4% average ACC across eight different image generators.

Conclusion: MCAN’s unified multi-cue aggregation framework effectively integrates spatial, frequency-domain, and chromaticity information to create more robust and generalizable AI-generated image detectors that work across diverse synthesis models.

Abstract: The rapid emergence of image synthesis models poses challenges to the generalization of AI-generated image detectors. However, existing methods often rely on model-specific features, leading to overfitting and poor generalization. In this paper, we introduce the Multi-Cue Aggregation Network (MCAN), a novel framework that integrates different yet complementary cues in a unified network. MCAN employs a mixture-of-encoders adapter to dynamically process these cues, enabling more adaptive and robust feature representation. Our cues include the input image itself, which represents the overall content, and high-frequency components that emphasize edge details. Additionally, we introduce a Chromatic Inconsistency (CI) cue, which normalizes intensity values and captures noise information introduced during the image acquisition process in real images, making these noise patterns more distinguishable from those in AI-generated content. Unlike prior methods, MCAN’s novelty lies in its unified multi-cue aggregation framework, which integrates spatial, frequency-domain, and chromaticity-based information for enhanced representation learning. These cues are intrinsically more indicative of real images, enhancing cross-model generalization. Extensive experiments on the GenImage, Chameleon, and UniversalFakeDetect benchmark validate the state-of-the-art performance of MCAN. In the GenImage dataset, MCAN outperforms the best state-of-the-art method by up to 7.4% in average ACC across eight different image generators.

Method: Proposes DentalX with a structural context extraction module that learns semantic segmentation of dental anatomy as an auxiliary task. This extracts meaningful structural context and integrates it into the primary disease detection task to enhance detection of subtle dental diseases.

Result: Extensive experiments on a dedicated benchmark demonstrate that DentalX significantly outperforms prior methods in both disease detection and segmentation tasks. The mutual benefit arises naturally during model optimization as the correlation between the two tasks is effectively captured.

Conclusion: DentalX effectively addresses the visual ambiguity in dental radiographs by leveraging oral structure information through a context-aware approach, achieving superior performance in dental disease detection.

Abstract: Diagnosing dental diseases from radiographs is time-consuming and challenging due to the subtle nature of diagnostic evidence. Existing methods, which rely on object detection models designed for natural images with more distinct target patterns, struggle to detect dental diseases that present with far less visual support. To address this challenge, we propose {\bf DentalX}, a novel context-aware dental disease detection approach that leverages oral structure information to mitigate the visual ambiguity inherent in radiographs. Specifically, we introduce a structural context extraction module that learns an auxiliary task: semantic segmentation of dental anatomy. The module extracts meaningful structural context and integrates it into the primary disease detection task to enhance the detection of subtle dental diseases. Extensive experiments on a dedicated benchmark demonstrate that DentalX significantly outperforms prior methods in both tasks. This mutual benefit arises naturally during model optimization, as the correlation between the two tasks is effectively captured. Our code is available at https://github.com/zhiqin1998/DentYOLOX.

Method: Evaluated computer vision methods using 1,233 ventral images from 191 Hula painted frogs. Compared deep local-feature matching (zero-shot) with deep global-feature embedding models. Developed two-stage workflow: global-feature model retrieves candidate list, then local-feature matching re-ranks results for accuracy and efficiency.

Result: Local-feature pipeline achieved 98% top-1 closed-set accuracy, outperforming global-feature models (60% top-1 after fine-tuning). Two-stage workflow reduced runtime from 6.5-7.8 hours to ~38 minutes while maintaining ~96% accuracy. Match score separation enabled open-set identification for novel individuals.

Conclusion: Zero-shot deep local-feature matching provides superior performance for amphibian photo-identification. The deployed web application enables rapid, standardized, non-invasive identification for conservation monitoring, with the two-stage approach balancing accuracy and scalability.

Abstract: Accurate individual identification is essential for monitoring rare amphibians, yet invasive marking is often unsuitable for critically endangered species. We evaluate state-of-the-art computer-vision methods for photographic re-identification of the Hula painted frog (Latonia nigriventer) using 1,233 ventral images from 191 individuals collected during 2013-2020 capture-recapture surveys. We compare deep local-feature matching in a zero-shot setting with deep global-feature embedding models. The local-feature pipeline achieves 98% top-1 closed-set identification accuracy, outperforming all global-feature models; fine-tuning improves the best global-feature model to 60% top-1 (91% top-10) but remains below local matching. To combine scalability with accuracy, we implement a two-stage workflow in which a fine-tuned global-feature model retrieves a short candidate list that is re-ranked by local-feature matching, reducing end-to-end runtime from 6.5-7.8 hours to ~38 minutes while maintaining ~96% top-1 closed-set accuracy on the labeled dataset. Separation of match scores between same- and different-individual pairs supports thresholding for open-set identification, enabling practical handling of novel individuals. We deploy this pipeline as a web application for routine field use, providing rapid, standardized, non-invasive identification to support conservation monitoring and capture-recapture analyses. Overall, in this species, zero-shot deep local-feature matching outperformed global-feature embedding and provides a strong default for photo-identification.

Method: Integrates recent advances in super-resolution networks with task-driven super-resolution pipelines, adapting them to video-based person re-identification. Combines video super-resolution with CLIP-ReID framework for the VReID-XFD challenge.

Result: Achieves 37.52% mAP in aerial-to-ground and 29.16% mAP in ground-to-aerial scenarios. In ground-to-aerial setting, shows substantial gains: Rank-1 (+11.24%), Rank-5 (+13.48%), and Rank-10 (+17.98%) improvements.

Conclusion: This represents the first systematic investigation of video super-resolution for enhancing tracklet quality in person ReID, demonstrating competitive performance and significant ranking accuracy improvements in challenging cross-view conditions.

Abstract: Tracklet quality is often treated as an afterthought in most person re-identification (ReID) methods, with the majority of research presenting architectural modifications to foundational models. Such approaches neglect an important limitation, posing challenges when deploying ReID systems in real-world, difficult scenarios. In this paper, we introduce S3-CLIP, a video super-resolution-based CLIP-ReID framework developed for the VReID-XFD challenge at WACV 2026. The proposed method integrates recent advances in super-resolution networks with task-driven super-resolution pipelines, adapting them to the video-based person re-identification setting. To the best of our knowledge, this work represents the first systematic investigation of video super-resolution as a means of enhancing tracklet quality for person ReID, particularly under challenging cross-view conditions. Experimental results demonstrate performance competitive with the baseline, achieving 37.52% mAP in aerial-to-ground and 29.16% mAP in ground-to-aerial scenarios. In the ground-to-aerial setting, S3-CLIP achieves substantial gains in ranking accuracy, improving Rank-1, Rank-5, and Rank-10 performance by 11.24%, 13.48%, and 17.98%, respectively.

Method: Develops an automated data pipeline to synthesize 3D visual grounding data with corresponding reasoning processes. Uses this generated data to fine-tune LLMs, resulting in Reason3DVG-8B model.

Result: Reason3DVG-8B outperforms previous LLM-based method 3D-GRAND while using only 1.6% of their training data, demonstrating both the effectiveness of the synthesized data and the importance of reasoning in 3D visual grounding.

Conclusion: Automated synthesis of 3D visual grounding data with reasoning processes enables training of more efficient and effective 3D visual grounding LLMs, highlighting the critical role of reasoning in 3D understanding tasks.

Abstract: The recent development of Large Language Models (LLMs) with strong reasoning ability has driven research in various domains such as mathematics, coding, and scientific discovery. Meanwhile, 3D visual grounding, as a fundamental task in 3D understanding, still remains challenging due to the limited reasoning ability of recent 3D visual grounding models. Most of the current methods incorporate a text encoder and visual feature encoder to generate cross-modal fuse features and predict the referring object. These models often require supervised training on extensive 3D annotation data. On the other hand, recent research also focus on scaling synthetic data to train stronger 3D visual grounding LLM, however, the performance gain remains limited and non-proportional to the data collection cost. In this work, we propose a 3D visual grounding data pipeline, which is capable of automatically synthesizing 3D visual grounding data along with corresponding reasoning process. Additionally, we leverage the generated data for LLM fine-tuning and introduce Reason3DVG-8B, a strong 3D visual grounding LLM that outperforms previous LLM-based method 3D-GRAND using only 1.6% of their training data, demonstrating the effectiveness of our data and the importance of reasoning in 3D visual grounding.

Method: Integrates 3D-aware features from MUSt3R into SAM2 using a lightweight Feature Merger that fuses multi-level features encoding implicit geometric correspondence. Uses field-of-view aware sampling to ensure frames observe spatially consistent object regions for reliable 3D correspondence learning.

Result: Achieves 90.6% IoU and 71.7% Positive IoU on ScanNet++’s Selected Subset, improving over state-of-the-art VOS methods by +15.9 and +30.4 points. Substantially outperforms SAM2 and extensions on challenging datasets with wide-baseline motion.

Conclusion: 3AM enables geometry-consistent video object segmentation using only RGB input at inference, addressing viewpoint consistency without requiring camera poses or preprocessing, making it practical for real-world applications.

Abstract: Video object segmentation methods like SAM2 achieve strong performance through memory-based architectures but struggle under large viewpoint changes due to reliance on appearance features. Traditional 3D instance segmentation methods address viewpoint consistency but require camera poses, depth maps, and expensive preprocessing. We introduce 3AM, a training-time enhancement that integrates 3D-aware features from MUSt3R into SAM2. Our lightweight Feature Merger fuses multi-level MUSt3R features that encode implicit geometric correspondence. Combined with SAM2’s appearance features, the model achieves geometry-consistent recognition grounded in both spatial position and visual similarity. We propose a field-of-view aware sampling strategy ensuring frames observe spatially consistent object regions for reliable 3D correspondence learning. Critically, our method requires only RGB input at inference, with no camera poses or preprocessing. On challenging datasets with wide-baseline motion (ScanNet++, Replica), 3AM substantially outperforms SAM2 and extensions, achieving 90.6% IoU and 71.7% Positive IoU on ScanNet++’s Selected Subset, improving over state-of-the-art VOS methods by +15.9 and +30.4 points. Project page: https://jayisaking.github.io/3AM-Page/

Method: A zero-shot diffusion-based framework that applies controlled geometric transformations in latent space, augmented with view-guided correspondence attention to maintain structural consistency during reconstruction. Operates on frozen pre-trained models without detector access or watermark knowledge.

Result: Achieves state-of-the-art watermark suppression across 15 watermarking methods, outperforming 14 baseline attacks while maintaining superior perceptual quality across multiple datasets.

Conclusion: Reveals a critical gap: watermarks robust to pixel-space and frequency-domain attacks remain vulnerable to semantic-preserving viewpoint transformations, exposing a fundamental vulnerability in invisible watermarking schemes.

Abstract: Invisible watermarking has become a critical mechanism for authenticating AI-generated image content, with major platforms deploying watermarking schemes at scale. However, evaluating the vulnerability of these schemes against sophisticated removal attacks remains essential to assess their reliability and guide robust design. In this work, we expose a fundamental vulnerability in invisible watermarks by reformulating watermark removal as a view synthesis problem. Our key insight is that generating a perceptually consistent alternative view of the same semantic content, akin to re-observing a scene from a shifted perspective, naturally removes the embedded watermark while preserving visual fidelity. This reveals a critical gap: watermarks robust to pixel-space and frequency-domain attacks remain vulnerable to semantic-preserving viewpoint transformations. We introduce a zero-shot diffusion-based framework that applies controlled geometric transformations in latent space, augmented with view-guided correspondence attention to maintain structural consistency during reconstruction. Operating on frozen pre-trained models without detector access or watermark knowledge, our method achieves state-of-the-art watermark suppression across 15 watermarking methods–outperforming 14 baseline attacks while maintaining superior perceptual quality across multiple datasets.

Method: Introduces SPOT algorithm for extracting spatial coordinates of irregular colored regions, creates ClimateIQA dataset (26,280 heatmaps, 762,120 samples) with spatial cues and metadata, and develops Climate-Zoo fine-tuned VLMs.

Result: Climate-Zoo models significantly outperform existing VLMs (GPT-4o, Qwen-VL, LLaVA 1.6) in meteorological heatmap interpretation tasks, demonstrating improved accuracy in extreme weather feature analysis.

Conclusion: SPOT algorithm with ClimateIQA dataset enables VLMs to better handle complex meteorological heatmaps, advancing extreme weather analysis capabilities through structured representation of irregular visual patterns.

Abstract: Meteorological heatmaps play a vital role in deciphering extreme weather phenomena, yet their inherent complexities marked by irregular contours, unstructured patterns, and complex color variations present unique analytical hurdles for state-of-the-art Vision-Language Models (VLMs). Current state-of-the-art models like GPT-4o, Qwen-VL, and LLaVA 1.6 struggle with tasks such as precise color identification and spatial localization, resulting in inaccurate or incomplete interpretations. To address these challenges, we introduce Sparse Position and Outline Tracking (SPOT), a novel algorithm specifically designed to process irregularly shaped colored regions in visual data. SPOT identifies and localizes these regions by extracting their spatial coordinates, enabling structured representations of irregular shapes. Building on SPOT, we construct ClimateIQA, a novel meteorological visual question answering (VQA) dataset, comprising 26,280 high-resolution heatmaps and 762,120 instruction samples for wind gust, total precipitation, wind chill index and heat index analysis. ClimateIQA enhances VLM training by incorporating spatial cues, geographic metadata, and reanalysis data, improving model accuracy in interpreting and describing extreme weather features. Furthermore, we develop Climate-Zoo, a suite of fine-tuned VLMs based on SPOT-empowered ClimateIQA, which significantly outperforms existing models in meteorological heatmap tasks.

Method: The paper conducts a literature review categorizing learning-based MVS methods into five groups: depth map-based, voxel-based, NeRF-based, 3D Gaussian Splatting-based, and large feed-forward methods. The survey focuses significantly on depth map-based methods due to their conciseness, flexibility, and scalability.

Result: The survey provides a comprehensive review of learning-based MVS literature, investigates various methods, summarizes their performances on popular benchmarks, and identifies current trends in the field.

Conclusion: Learning-based MVS methods have achieved impressive performance improvements over traditional methods. The survey organizes the field into clear categories, with depth map-based methods being the main family due to their practical advantages. The paper also discusses promising future research directions for advancing 3D reconstruction technology.

Abstract: 3D reconstruction aims to recover the dense 3D structure of a scene. It plays an essential role in various applications such as Augmented/Virtual Reality (AR/VR), autonomous driving and robotics. Leveraging multiple views of a scene captured from different viewpoints, Multi-View Stereo (MVS) algorithms synthesize a comprehensive 3D representation, enabling precise reconstruction in complex environments. Due to its efficiency and effectiveness, MVS has become a pivotal method for image-based 3D reconstruction. Recently, with the success of deep learning, many learning-based MVS methods have been proposed, achieving impressive performance against traditional methods. We categorize these learning-based methods as: depth map-based, voxel-based, NeRF-based, 3D Gaussian Splatting-based, and large feed-forward methods. Among these, we focus significantly on depth map-based methods, which are the main family of MVS due to their conciseness, flexibility and scalability. In this survey, we provide a comprehensive review of the literature at the time of this writing. We investigate these learning-based methods, summarize their performances on popular benchmarks, and discuss promising future research directions in this area.

Method: Proposes Multi-Stage Sampling Fusion (MSSF) network with: 1) Fusion blocks (SFF and MSDFF) that deeply interact point cloud features with image features in plug-and-play manner, 2) Semantic-guided head performing foreground-background segmentation on voxels with voxel feature re-weighting to alleviate feature blurring.

Result: Achieves 7.0% and 4.0% improvement in 3D mean average precision on VoD and TJ4DRadSet datasets compared to state-of-the-art methods. Notably surpasses classical LiDAR-based methods on VoD dataset.

Conclusion: MSSF provides an effective solution for 4D radar-camera fusion that addresses feature blurring and enables deep semantic interaction, demonstrating competitive performance that can rival LiDAR-based approaches while using more affordable sensors.

Abstract: As one of the automotive sensors that have emerged in recent years, 4D millimeter-wave radar has a higher resolution than conventional 3D radar and provides precise elevation measurements. But its point clouds are still sparse and noisy, making it challenging to meet the requirements of autonomous driving. Camera, as another commonly used sensor, can capture rich semantic information. As a result, the fusion of 4D radar and camera can provide an affordable and robust perception solution for autonomous driving systems. However, previous radar-camera fusion methods have not yet been thoroughly investigated, resulting in a large performance gap compared to LiDAR-based methods. Specifically, they ignore the feature-blurring problem and do not deeply interact with image semantic information. To this end, we present a simple but effective multi-stage sampling fusion (MSSF) network based on 4D radar and camera. On the one hand, we design a fusion block that can deeply interact point cloud features with image features, and can be applied to commonly used single-modal backbones in a plug-and-play manner. The fusion block encompasses two types, namely, simple feature fusion (SFF) and multiscale deformable feature fusion (MSDFF). The SFF is easy to implement, while the MSDFF has stronger fusion abilities. On the other hand, we propose a semantic-guided head to perform foreground-background segmentation on voxels with voxel feature re-weighting, further alleviating the problem of feature blurring. Extensive experiments on the View-of-Delft (VoD) and TJ4DRadset datasets demonstrate the effectiveness of our MSSF. Notably, compared to state-of-the-art methods, MSSF achieves a 7.0% and 4.0% improvement in 3D mean average precision on the VoD and TJ4DRadSet datasets, respectively. It even surpasses classical LiDAR-based methods on the VoD dataset.

Method: Developed PathGen, a diffusion-based crossmodal generative AI model that synthesizes gene expression data from whole slide images (WSIs) of histopathology, enabling multimodal prediction without requiring actual transcriptomic tests.

Result: Significant performance gains in cancer grading and survival risk prediction across four cancer cohorts; synthetic transcriptomic features performed statistically similar to real transcriptomic data while achieving state-of-the-art performance with conformal coverage guarantees and interpretable attention maps.

Conclusion: PathGen enables practical multimodal cancer analysis in clinical settings by synthesizing transcriptomic data from standard histopathology, overcoming the barrier of limited transcriptomic testing availability while maintaining high accuracy and interpretability.

Abstract: Emerging research has highlighted that artificial intelligence-based multimodal fusion of digital pathology and transcriptomic features can improve cancer diagnosis (grading/subtyping) and prognosis (survival risk) prediction. However, such direct fusion is impractical in clinical settings, where histopathology remains the gold standard and transcriptomic tests are rarely requested in public healthcare. We experiment on two publicly available multimodal datasets, The Cancer Genomic Atlas and the Clinical Proteomic Tumor Analysis Consortium, spanning four independent cohorts: glioma-glioblastoma, renal, uterine, and breast, and observe significant performance gains in gradation and risk estimation (p-value<0.05) when incorporating synthesized transcriptomic data with WSIs. Also, predictions using synthesized features were statistically close to those obtained with real transcriptomic data (p-value>0.05), consistently across cohorts. Here we show that with our diffusion based crossmodal generative AI model, PathGen, gene expressions synthesized from digital histopathology jointly predict cancer grading and patient survival risk with high accuracy (state-of-the-art performance), certainty (through conformal coverage guarantee) and interpretability (through distributed co-attention maps). PathGen code is available for open use on GitHub at https://github.com/Samiran-Dey/PathGen.

Method: PRISM (training-free framework) surgically removes corrupting influence of global background features by modeling intrinsic visual semantics via implicit re-centering. It’s the first training-free approach for visual instruction selection.

Result: Reduces end-to-end time for data selection and model tuning to 30% of conventional pipelines. Surpasses models fine-tuned on full dataset across 8 multimodal and 3 language understanding benchmarks, achieving 101.7% relative improvement over baseline.

Conclusion: PRISM addresses the efficiency bottleneck in visual instruction tuning by providing a training-free selection framework that not only reduces computational costs but also improves model performance by properly handling visual feature anisotropy.

Abstract: Visual instruction tuning adapts pre-trained Multimodal Large Language Models (MLLMs) to follow human instructions for real-world applications. However, the rapid growth of these datasets introduces significant redundancy, leading to increased computational costs. Existing methods for selecting instruction data aim to prune this redundancy, but predominantly rely on computationally demanding techniques such as proxy-based inference or training-based metrics. Consequently, the substantial computational costs incurred by these selection processes often exacerbate the very efficiency bottlenecks they are intended to resolve, posing a significant challenge to the scalable and effective tuning of MLLMs. To address this challenge, we first identify a critical, yet previously overlooked, factor: the anisotropy inherent in visual feature distributions. We find that this anisotropy induces a \textit{Global Semantic Drift}, and overlooking this phenomenon is a key factor limiting the efficiency of current data selection methods. Motivated by this insight, we devise \textbf{PRISM}, the first training-free framework for efficient visual instruction selection. PRISM surgically removes the corrupting influence of global background features by modeling the intrinsic visual semantics via implicit re-centering. Empirically, PRISM reduces the end-to-end time for data selection and model tuning to just 30% of conventional pipelines. More remarkably, it achieves this efficiency while simultaneously enhancing performance, surpassing models fine-tuned on the full dataset across eight multimodal and three language understanding benchmarks, culminating in a 101.7% relative improvement over the baseline. The code is available for access via \href{https://github.com/bibisbar/PRISM}{this repository}.

Method: Novel two-stream fusion system that combines static and dynamic streams from thermal recordings. Uses image and video analysis to capture spatiotemporal birth-related features, with a score aggregation module for final ToB estimation.

Result: System achieves 95.7% precision and 84.8% recall for birth detection in short clips. With score aggregation, identifies ToB in 100% of test cases with median absolute error of 2 seconds and mean deviation of 4.5 seconds compared to manual annotations.

Conclusion: The fusion of image and video modalities enhances ToB detection performance over single-stream approaches, providing robust and precise automated documentation for neonatal care optimization.

Abstract: Around 10% of newborns require some help to initiate breathing, and 5% need ventilation assistance. Accurate Time of Birth (ToB) documentation is essential for optimizing neonatal care, as timely interventions are vital for proper resuscitation. However, current clinical methods for recording ToB often rely on manual processes, which can be prone to inaccuracies. In this study, we present a novel two-stream fusion system that combines the power of image and video analysis to accurately detect the ToB from thermal recordings in the delivery room and operating theater. By integrating static and dynamic streams, our approach captures richer birth-related spatiotemporal features, leading to more robust and precise ToB estimation. We demonstrate that this synergy between data modalities enhances performance over single-stream approaches. Our system achieves 95.7% precision and 84.8% recall in detecting birth within short video clips. Additionally, with the help of a score aggregation module, it successfully identifies ToB in 100% of test cases, with a median absolute error of 2 seconds and an absolute mean deviation of 4.5 seconds compared to manual annotations.

Method: 1) Instance-level historical rasterization map representation to store previous perception results; 2) Map-Trajectory Prior Fusion module leveraging historical priors for tracked instances; 3) Global perspective metric for evaluating temporal geometry construction quality.

Result: Outperforms state-of-the-art methods on nuScenes and Argoverse2 datasets in both single-frame and temporal metrics, demonstrating improved temporal smoothness and continuity.

Conclusion: The proposed HisTrackMap framework effectively addresses temporal inconsistency in HD map construction through explicit historical trajectory tracking, providing more stable and reliable perception for autonomous driving applications.

Abstract: As an essential component of autonomous driving systems, high-definition (HD) maps provide rich and precise environmental information for auto-driving scenarios; however, existing methods, which primarily rely on query-based detection frameworks to directly model map elements or implicitly propagate queries over time, often struggle to maintain consistent temporal perception outcomes. These inconsistencies pose significant challenges to the stability and reliability of real-world autonomous driving and map data collection systems. To address this limitation, we propose a novel end-to-end tracking framework for global map construction by temporally tracking map elements’ historical trajectories. Firstly, instance-level historical rasterization map representation is designed to explicitly store previous perception results, which can control and maintain different global instances’ history information in a fine-grained way. Secondly, we introduce a Map-Trajectory Prior Fusion module within this tracking framework, leveraging historical priors for tracked instances to improve temporal smoothness and continuity. Thirdly, we propose a global perspective metric to evaluate the quality of temporal geometry construction in HD maps, filling the gap in current metrics for assessing global geometric perception results. Substantial experiments on the nuScenes and Argoverse2 datasets demonstrate that the proposed method outperforms state-of-the-art (SOTA) methods in both single-frame and temporal metrics. The project page is available at: https://yj772881654.github.io/HisTrackMap.

Method: Created EPIC-KITCHENS-100-MQA benchmark with hard distractors, curated a supervised finetuning dataset, introduced LLaVAction with an action token to boost attention on visual tokens, and developed a two-stage pipeline for structured action recognition.

Result: LLaVAction achieved 21-point accuracy improvement over GPT-4o on EPIC-KITCHENS-100-MQA and showed strong performance on established action recognition benchmarks, demonstrating superior action understanding capabilities.

Conclusion: The proposed methods significantly enhance MLLMs’ action understanding, making them a promising approach for complex action tasks, with code, data, benchmarks, and models publicly available.

Abstract: Understanding human behavior requires measuring behavioral actions. Due to its complexity, behavior is best mapped onto a rich, semantic structure such as language. Emerging multimodal large language models (MLLMs) are promising candidates, but their fine-grained action understanding ability has not been fully examined. In this work, we reformulate EPIC-KITCHENS-100, one of the largest and most challenging egocentric action recognition datasets, into a MLLM benchmark (EPIC-KITCHENS-100-MQA). We demonstrate that when we sample difficult answers based on specialist models as distractors, leading MLLMs struggle to recognize the correct actions. How can we increase the performance of MLLMs? We curated a supervised finetuning dataset that includes `hard’ action recognition, temporal detection, captioning, and free-form question answering to improve models’ diverse action understanding capabilities. We introduce a new model called LLaVAction that adds an action token to boost models’ attention on visual tokens and a two-stage pipeline to obtain structured actions. LLaVAction greatly improves the MLLMs’ ability of action understanding, achieving strong improvements on both MLLM benchmarks (21 points in accuracy over GPT-4o on EPIC-KITCHENS-100-MQA) and established action recognition benchmarks, suggesting that our methods prepare MLLMs to be a promising path forward for complex action tasks. Code, data, the benchmark, and models are available at https://github.com/AdaptiveMotorControlLab/LLaVAction.

Method: Proposes CasTex - cascaded diffusion models for texture synthesis that uses score distillation sampling with explicit texture parameterization instead of implicit parameterization, working with differentiable rasterization and shading pipelines.

Result: The approach significantly outperforms state-of-the-art optimization-based solutions on public texture synthesis benchmarks, producing high-quality textures without visual artifacts.

Conclusion: CasTex provides a more direct and effective alternative to existing text-to-texture synthesis methods, demonstrating that explicit texture parameterization with cascaded diffusion models yields superior results compared to optimization-based approaches.

Abstract: This work investigates text-to-texture synthesis using diffusion models to generate physically-based texture maps. We aim to achieve realistic model appearances under varying lighting conditions. A prominent solution for the task is score distillation sampling. It allows recovering a complex texture using gradient guidance given a differentiable rasterization and shading pipeline. However, in practice, the aforementioned solution in conjunction with the widespread latent diffusion models produces severe visual artifacts and requires additional regularization such as implicit texture parameterization. As a more direct alternative, we propose an approach using cascaded diffusion models for texture synthesis (CasTex). In our setup, score distillation sampling yields high-quality textures out-of-the box. In particular, we were able to omit implicit texture parameterization in favor of an explicit parameterization to improve the procedure. In the experiments, we show that our approach significantly outperforms state-of-the-art optimization-based solutions on public texture synthesis benchmarks.

Method: Proposes DAS2F-Model with: 1) Cross-modal diff-attention module (CMDA-Module) to separate common and dominant features, 2) Space preserving visual mamba (SPVM) for spatial feature retention, 3) Attention-aware linear fusion module (AALF-Module) for pixel-wise fusion of semantically different features.

Result: Empirical evaluations show the method achieves better results than alternative approaches for multimodal remote sensing image classification.

Conclusion: The proposed DAS2F-Model effectively addresses feature redundancy and semantic difference challenges in multimodal remote sensing fusion, outperforming existing methods.

Abstract: Multispectral (MS) and panchromatic (PAN) images describe the same land surface, so these images not only have their own advantages, but also have a lot of similar information. In order to separate these similar information and their respective advantages, reduce the feature redundancy in the fusion stage. This paper introduces a diff-attention aware state space fusion model (DAS2F-Model) for multimodal remote sensing image classification. Based on the selective state space model, a cross-modal diff-attention module (CMDA-Module) is designed to extract and separate the common features and their respective dominant features of MS and PAN images. Among this, space preserving visual mamba (SPVM) retains image spatial features and captures local features by optimizing visual mamba’s input reasonably. Considering that features in the fusion stage will have large semantic differences after feature separation and simple fusion operations struggle to effectively integrate these significantly different features, an attention-aware linear fusion module (AALF-Module) is proposed. It performs pixel-wise linear fusion by calculating influence coefficients. This mechanism can fuse features with large semantic differences while keeping the feature size unchanged. Empirical evaluations indicate that the presented method achieves better results than alternative approaches. The relevant code can be found at:https://github.com/AVKSKVL/DAS-F-Model

Method: AutoDriveRL formulates autonomous driving as structured reasoning across four core tasks (perception, prediction, planning, behavior). Each task is modeled as a vision-language QA problem and optimized using task-specific reward models, enabling fine-grained reinforcement signals at different reasoning stages. The framework trains DriveRX, a cross-task reasoning VLM designed for multi-stage decision-making.

Result: DriveRX achieves strong performance on public benchmarks, outperforming GPT-4o in behavior reasoning and demonstrating robustness under complex or corrupted driving conditions. It produces structured stage-wise reasoning chains that enhance decision consistency and provides high-quality supervisory signals for annotation and downstream planning/control models.

Conclusion: AutoDriveRL provides a unified framework for structured reasoning in autonomous driving, with DriveRX serving as an effective high-level semantic reasoning backbone. The framework and model are released to support future research in autonomous driving reasoning systems.

Abstract: Effective autonomous driving hinges on robust reasoning across perception, prediction, planning, and behavior. However, conventional end-to-end models fail to generalize in complex scenarios due to the lack of structured reasoning. While recent vision-language models (VLMs) have been applied to driving tasks, they typically rely on isolated modules and static supervision, limiting their ability to support multi-stage decision-making. We present AutoDriveRL, a unified training framework that formulates autonomous driving as a structured reasoning process over four core tasks. Each task is independently modeled as a vision-language QA problem and optimized using task-specific reward models, enabling fine-grained reinforcement signals at different reasoning stages. Within this framework, we train DriveRX, a cross-task reasoning VLM designed for multi-stage decision-making. DriveRX achieves strong performance on the public benchmark, outperforming GPT-4o in behavior reasoning and demonstrating robustness under complex or corrupted driving conditions. DriveRX serves as a high-level semantic reasoning backbone, producing structured stage-wise reasoning chains that enhance decision consistency. These outputs also provide high-quality supervisory signals for annotation and downstream planning/control models. We release the AutoDriveRL framework and DriveRX to support future research.

Method: Novel multi-scale bilateral grid that unifies appearance codes and bilateral grids to perform pixel-wise color mapping, addressing photometric inconsistency in dynamic autonomous driving scenes.

Result: Significantly improves geometric accuracy in dynamic, decoupled autonomous driving scene reconstruction, outperforming both appearance codes and bilateral grids. Shows strong results across Waymo, NuScenes, Argoverse, and PandaSet datasets. Reduces floaters caused by photometric inconsistency.

Conclusion: Multi-scale bilateral grid effectively addresses photometric inconsistency in neural rendering for autonomous driving, providing crucial geometric accuracy improvements for obstacle avoidance and control applications.

Abstract: Neural rendering techniques, including NeRF and Gaussian Splatting (GS), rely on photometric consistency to produce high-quality reconstructions. However, in real-world scenarios, it is challenging to guarantee perfect photometric consistency in acquired images. Appearance codes have been widely used to address this issue, but their modeling capability is limited, as a single code is applied to the entire image. Recently, the bilateral grid was introduced to perform pixel-wise color mapping, but it is difficult to optimize and constrain effectively. In this paper, we propose a novel multi-scale bilateral grid that unifies appearance codes and bilateral grids. We demonstrate that this approach significantly improves geometric accuracy in dynamic, decoupled autonomous driving scene reconstruction, outperforming both appearance codes and bilateral grids. This is crucial for autonomous driving, where accurate geometry is important for obstacle avoidance and control. Our method shows strong results across four datasets: Waymo, NuScenes, Argoverse, and PandaSet. We further demonstrate that the improvement in geometry is driven by the multi-scale bilateral grid, which effectively reduces floaters caused by photometric inconsistency.

Method: MOGO uses two key components: 1) MoSA-VQ - motion scale-adaptive residual vector quantization that hierarchically discretizes motion sequences with learnable scaling, and 2) RQHC-Transformer - residual quantized hierarchical causal transformer that generates multi-layer motion tokens in a single forward pass. Also includes text condition alignment for improved semantic fidelity.

Result: Achieves competitive or superior generation quality compared to state-of-the-art transformer methods on HumanML3D, KIT-ML, and CMP datasets, while offering substantial improvements in real-time performance, streaming generation, and generalization under zero-shot settings.

Conclusion: MOGO successfully addresses the fundamental challenge of jointly achieving high fidelity, streaming capability, real-time responsiveness, and scalability in text-to-motion generation through its novel autoregressive framework with adaptive quantization and hierarchical transformers.

Abstract: Recent advances in transformer-based text-to-motion generation have led to impressive progress in synthesizing high-quality human motion. Nevertheless, jointly achieving high fidelity, streaming capability, real-time responsiveness, and scalability remains a fundamental challenge. In this paper, we propose MOGO (Motion Generation with One-pass), a novel autoregressive framework tailored for efficient and real-time 3D motion generation. MOGO comprises two key components: (1) MoSA-VQ, a motion scale-adaptive residual vector quantization module that hierarchically discretizes motion sequences with learnable scaling to produce compact yet expressive representations; and (2) RQHC-Transformer, a residual quantized hierarchical causal transformer that generates multi-layer motion tokens in a single forward pass, significantly reducing inference latency. To enhance semantic fidelity, we further introduce a text condition alignment mechanism that improves motion decoding under textual control. Extensive experiments on benchmark datasets including HumanML3D, KIT-ML, and CMP demonstrate that MOGO achieves competitive or superior generation quality compared to state-of-the-art transformer-based methods, while offering substantial improvements in real-time performance, streaming generation, and generalization under zero-shot settings.

Method: DGAE uses a diffusion model to guide the decoder in recovering informative signals not fully decoded from the latent representation, improving decoder expressiveness and mitigating performance issues at high compression rates.

Result: DGAE achieves state-of-the-art performance with 2x smaller latent space, effectively mitigates performance degradation under high spatial compression, and when integrated with Diffusion Models, shows competitive ImageNet-1K generation with faster diffusion model convergence.

Conclusion: The diffusion-guided approach successfully addresses autoencoder limitations, enabling more compact and efficient latent representations while maintaining performance, which benefits downstream generative modeling tasks.

Abstract: Autoencoders empower state-of-the-art image and video generative models by compressing pixels into a latent space through visual tokenization. Although recent advances have alleviated the performance degradation of autoencoders under high compression ratios, addressing the training instability caused by GAN remains an open challenge. While improving spatial compression, we also aim to minimize the latent space dimensionality, enabling more efficient and compact representations. To tackle these challenges, we focus on improving the decoder’s expressiveness. Concretely, we propose DGAE, which employs a diffusion model to guide the decoder in recovering informative signals that are not fully decoded from the latent representation. With this design, DGAE effectively mitigates the performance degradation under high spatial compression rates. At the same time, DGAE achieves state-of-the-art performance with a 2x smaller latent space. When integrated with Diffusion Models, DGAE demonstrates competitive performance on image generation for ImageNet-1K and shows that this compact latent representation facilitates faster convergence of the diffusion model.

Method: Introduces DSF (divergence-based similarity function) that represents each set of augmented views as a distribution and measures similarity as the divergence between distributions, enabling explicit modeling of joint structure across all views.

Result: DSF consistently improves performance across diverse tasks (kNN classification, linear evaluation, transfer learning, distribution shift) while achieving greater efficiency than other multi-view methods. It also operates effectively without needing temperature hyperparameter tuning.

Conclusion: DSF provides an effective approach for capturing joint structure across multiple views in contrastive learning, offering performance improvements, efficiency gains, and eliminating the need for temperature hyperparameter tuning compared to cosine similarity-based approaches.

Abstract: Recent success in contrastive learning has sparked growing interest in more effectively leveraging multiple augmented views of data. While prior methods incorporate multiple views at the loss or feature level, they primarily capture pairwise relationships and fail to model the joint structure across all views. In this work, we propose a divergence-based similarity function (DSF) that explicitly captures the joint structure by representing each set of augmented views as a distribution and measuring similarity as the divergence between distributions. Extensive experiments demonstrate that DSF consistently improves performance across diverse tasks, including kNN classification, linear evaluation, transfer learning, and distribution shift, while also achieving greater efficiency than other multi-view methods. Furthermore, we establish a connection between DSF and cosine similarity, and demonstrate that, unlike cosine similarity, DSF operates effectively without the need for tuning a temperature hyperparameter.

Method: Three key insights: (1) Render covisible pixels (visible in both input and target views) by reconstructing dynamic 3D scene and rendering from novel views; (2) Inpaint hidden pixels using feed-forward 2D video diffusion models (CogNVS) trained self-supervised on large corpus of in-the-wild videos; (3) Apply CogNVS zero-shot to novel test videos via test-time finetuning.

Result: CogNVS empirically outperforms almost all prior art for novel-view synthesis of dynamic scenes from monocular videos.

Conclusion: The proposed approach successfully combines 3D reconstruction for geometry preservation with 2D video diffusion for inpainting, enabling efficient novel-view synthesis that generalizes well to new dynamic scenes without extensive per-scene optimization.

Abstract: We explore novel-view synthesis for dynamic scenes from monocular videos. Prior approaches rely on costly test-time optimization of 4D representations or do not preserve scene geometry when trained in a feed-forward manner. Our approach is based on three key insights: (1) covisible pixels (that are visible in both the input and target views) can be rendered by first reconstructing the dynamic 3D scene and rendering the reconstruction from the novel-views and (2) hidden pixels in novel views can be “inpainted” with feed-forward 2D video diffusion models. Notably, our video inpainting diffusion model (CogNVS) can be self-supervised from 2D videos, allowing us to train it on a large corpus of in-the-wild videos. This in turn allows for (3) CogNVS to be applied zero-shot to novel test videos via test-time finetuning. We empirically verify that CogNVS outperforms almost all prior art for novel-view synthesis of dynamic scenes from monocular videos.

Method: Uses transparent training pipeline with public data (ImageNet-21K, ReLAION-2B subset), introduces parameter-efficient multi-head clustering projector with nested Matryoshka representations, and novel positional disentanglement strategy to remove positional biases.

Result: Matches or surpasses state-of-the-art proprietary models (DINOv2, CLIP, SigLIPv2) on several downstream benchmarks, demonstrating cleaner feature spaces and improved semantic encoding.

Conclusion: Franca establishes new standards for transparent, high-performance vision models and opens a path toward more reproducible and generalizable foundation models for the broader AI community.

Abstract: We present Franca (pronounced Fran-ka): free one; the first fully open-source (data, code, weights) vision foundation model that matches and in many cases surpasses the performance of state-of-the-art proprietary models, e.g., DINOv2, CLIP, SigLIPv2, etc. Our approach is grounded in a transparent training pipeline inspired by Web-SSL and uses publicly available data: ImageNet-21K and a subset of ReLAION-2B. Beyond model release, we tackle critical limitations in SSL clustering methods. While modern models rely on assigning image features to large codebooks via clustering algorithms like Sinkhorn-Knopp, they fail to account for the inherent ambiguity in clustering semantics. To address this, we introduce a parameter-efficient, multi-head clustering projector based on nested Matryoshka representations. This design progressively refines features into increasingly fine-grained clusters without increasing the model size, enabling both performance and memory efficiency. Additionally, we propose a novel positional disentanglement strategy that explicitly removes positional biases from dense representations, thereby improving the encoding of semantic content. This leads to consistent gains on several downstream benchmarks, demonstrating the utility of cleaner feature spaces. Our contributions establish a new standard for transparent, high-performance vision models and open a path toward more reproducible and generalizable foundation models for the broader AI community. The code and model checkpoints are available at https://github.com/valeoai/Franca.

Method: Organized the Surgical Procedure Phase, Keypoint, and Instrument Recognition (PhaKIR) sub-challenge at MICCAI 2024 EndoVis. Created a novel multi-center dataset with 13 full-length laparoscopic cholecystectomy videos from 3 medical institutions, featuring unified annotations for three interrelated tasks.

Result: The dataset enables joint investigation of instrument localization and procedural context within the same data while supporting temporal information integration across entire procedures. Results reported following BIAS guidelines for biomedical image analysis challenges.

Conclusion: PhaKIR advances the field by providing a unique benchmark for developing temporally aware, context-driven methods in RAMIS and offers a high-quality resource for future surgical scene understanding research.

Abstract: Reliable recognition and localization of surgical instruments in endoscopic video recordings are foundational for a wide range of applications in computer- and robot-assisted minimally invasive surgery (RAMIS), including surgical training, skill assessment, and autonomous assistance. However, robust performance under real-world conditions remains a significant challenge. Incorporating surgical context - such as the current procedural phase - has emerged as a promising strategy to improve robustness and interpretability. To address these challenges, we organized the Surgical Procedure Phase, Keypoint, and Instrument Recognition (PhaKIR) sub-challenge as part of the Endoscopic Vision (EndoVis) challenge at MICCAI 2024. We introduced a novel, multi-center dataset comprising thirteen full-length laparoscopic cholecystectomy videos collected from three distinct medical institutions, with unified annotations for three interrelated tasks: surgical phase recognition, instrument keypoint estimation, and instrument instance segmentation. Unlike existing datasets, ours enables joint investigation of instrument localization and procedural context within the same data while supporting the integration of temporal information across entire procedures. We report results and findings in accordance with the BIAS guidelines for biomedical image analysis challenges. The PhaKIR sub-challenge advances the field by providing a unique benchmark for developing temporally aware, context-driven methods in RAMIS and offers a high-quality resource to support future research in surgical scene understanding.

Method: Multimodal framework integrating language and vision into 3D Gaussian Splatting. Each Gaussian primitive gets learnable semantic features supervised by CLIP embeddings from region-level semantic cues. Uses semantic consistency loss, stage-wise optimization (coarse-to-fine with late-stage refinement), and 3D Gaussian Primitives Reallocation to address imbalanced distribution.

Result: Outperforms state-of-the-art methods on SeaThru-NeRF and Submerged3D datasets across PSNR, SSIM, and LPIPS metrics, achieving up to 3.48 dB PSNR improvement.

Conclusion: SWAGSplatting enables more accurate and semantically coherent underwater scene reconstruction, advancing marine perception and exploration applications.

Abstract: Accurate 3D reconstruction in underwater environments remains a challenging task due to light attenuation, scattering, and limited visibility. While recent AI-based approaches have advanced underwater imaging, they often overlook high-level semantic understanding, which is crucial for reconstructing complex scenes. In this paper, we propose SWAGSplatting, \textit{Semantic-guided Water-scene Augmented Gaussian Splatting}, a novel multimodal framework that integrates language and vision knowledge into 3D Gaussian Splatting for robust and high-fidelity underwater reconstruction. Each Gaussian primitive is augmented with a learnable semantic feature, supervised using CLIP-based embeddings extracted from region-level semantic cues. A dedicated semantic consistency loss enforces alignment between geometric reconstruction and scene semantics. In addition, a stage-wise optimisation strategy combining coarse-to-fine learning with late-stage parameter refinement improves training stability and visual quality. Furthermore, we propose a 3D Gaussian Primitives Reallocation strategy to address the imbalanced distribution of primitives introduced by naive point cloud densification. Extensive experiments on the SeaThru-NeRF and Submerged3D datasets demonstrate that SWAGSplatting consistently outperforms state-of-the-art methods across PSNR, SSIM, and LPIPS metrics, achieving up to a 3.48 dB improvement in PSNR, enabling more accurate and semantically coherent underwater scene reconstruction for applications in marine perception and exploration.

Method: Extends 3DETR with geometry-aware matching strategy and specialized loss functions to enable unified detection of differently parameterized curve types (cubic Bézier curves, line segments, circles, arcs) in a single forward pass, with optional post-processing refinement.

Result: Sets new state-of-the-art on the ABC dataset and generalizes effectively to real sensor data, demonstrating improved robustness to noise and varying sampling densities.

Conclusion: PI3DETR offers a simple yet powerful end-to-end solution for 3D edge and curve estimation that avoids complex intermediate representations while maintaining high performance on both synthetic and real-world data.

Abstract: We present PI3DETR, an end-to-end framework that directly predicts 3D parametric curve instances from raw point clouds, avoiding the intermediate representations and multi-stage processing common in prior work. Extending 3DETR, our model introduces a geometry-aware matching strategy and specialized loss functions that enable unified detection of differently parameterized curve types, including cubic Bézier curves, line segments, circles, and arcs, in a single forward pass. Optional post-processing steps further refine predictions without adding complexity. This streamlined design improves robustness to noise and varying sampling densities, addressing critical challenges in real world LiDAR and 3D sensing scenarios. PI3DETR sets a new state-of-the-art on the ABC dataset and generalizes effectively to real sensor data, offering a simple yet powerful solution for 3D edge and curve estimation.

Method: Proposes SAAM with lightweight scale-adaptive feature extraction and upsampling, incorporating SimAM for efficient guidance and gradient variance loss for sharpness.

Result: SAAM integrates seamlessly into multiple SOTA SR backbones, delivering competitive/superior performance across integer and non-integer scale factors with minimal overhead.

Conclusion: The approach enables robust multi-scale upscaling, offering a practical solution for real-world scenarios requiring arbitrary-scale super-resolution.

Abstract: Despite remarkable progress in Single Image Super-Resolution (SISR), traditional models often struggle to generalize across varying scale factors, limiting their real-world applicability. To address this, we propose a plug-in Scale-Aware Attention Module (SAAM) designed to retrofit modern fixed-scale SR models with the ability to perform arbitrary-scale SR. SAAM employs lightweight, scale-adaptive feature extraction and upsampling, incorporating the Simple parameter-free Attention Module (SimAM) for efficient guidance and gradient variance loss to enhance sharpness in image details. Our method integrates seamlessly into multiple state-of-the-art SR backbones (e.g., SCNet, HiT-SR, OverNet), delivering competitive or superior performance across a wide range of integer and non-integer scale factors. Extensive experiments on benchmark datasets demonstrate that our approach enables robust multi-scale upscaling with minimal computational overhead, offering a practical solution for real-world scenarios.

Method: Proposes GSAlign with two key components: 1) Learnable Thin Plate Spline (LTPS) Module that adaptively warps pedestrian features using learned keypoints to compensate for geometric variations, and 2) Dynamic Alignment Module (DAM) that estimates visibility-aware representation masks to highlight visible body regions and handle occlusions at semantic level.

Result: Achieves significant improvements of +18.8% in mAP and +16.8% in Rank-1 accuracy over previous state-of-the-art methods on the CARGO dataset with four matching protocols, demonstrating superior performance in aerial-ground person re-identification.

Conclusion: GSAlign effectively addresses the core challenges of AG-ReID by jointly tackling geometric distortion and semantic misalignment through adaptive feature warping and visibility-aware representation, establishing a new state-of-the-art for cross-view person matching between aerial and ground perspectives.

Abstract: Aerial-Ground person re-identification (AG-ReID) is an emerging yet challenging task that aims to match pedestrian images captured from drastically different viewpoints, typically from unmanned aerial vehicles (UAVs) and ground-based surveillance cameras. The task poses significant challenges due to extreme viewpoint discrepancies, occlusions, and domain gaps between aerial and ground imagery. While prior works have made progress by learning cross-view representations, they remain limited in handling severe pose variations and spatial misalignment. To address these issues, we propose a Geometric and Semantic Alignment Network (GSAlign) tailored for AG-ReID. GSAlign introduces two key components to jointly tackle geometric distortion and semantic misalignment in aerial-ground matching: a Learnable Thin Plate Spline (LTPS) Module and a Dynamic Alignment Module (DAM). The LTPS module adaptively warps pedestrian features based on a set of learned keypoints, effectively compensating for geometric variations caused by extreme viewpoint changes. In parallel, the DAM estimates visibility-aware representation masks that highlight visible body regions at the semantic level, thereby alleviating the negative impact of occlusions and partial observations in cross-view correspondence. A comprehensive evaluation on CARGO with four matching protocols demonstrates the effectiveness of GSAlign, achieving significant improvements of +18.8% in mAP and +16.8% in Rank-1 accuracy over previous state-of-the-art methods on the aerial-ground setting.

Method: Two-component framework: 1) End-to-end heatmap-supervised detection model inspired by keypoint detection, 2) Teacher-student co-training mechanism for sparse labeling conditions. Additional techniques: multi-view pseudo-labeling and CryoET-specific DropBlock augmentation.

Result: Extensive evaluations on large-scale CZII dataset show 10% F1 improvement over supervised baselines, demonstrating effectiveness of semi-supervised learning for leveraging unlabeled CryoET data.

Conclusion: The semi-supervised framework successfully addresses the particle picking bottleneck in CryoET by efficiently utilizing unlabeled data, showing promise for advancing structural biology research through improved label efficiency.

Abstract: Cryogenic Electron Tomography (CryoET) combined with sub-volume averaging (SVA) is the only imaging modality capable of resolving protein structures inside cells at molecular resolution. Particle picking, the task of localizing and classifying target proteins in 3D CryoET volumes, remains the main bottleneck. Due to the reliance on time-consuming manual labels, the vast reserve of unlabeled tomograms remains underutilized. In this work, we present a fast, label-efficient semi-supervised framework that exploits this untapped data. Our framework consists of two components: (i) an end-to-end heatmap-supervised detection model inspired by keypoint detection, and (ii) a teacher-student co-training mechanism that enhances performance under sparse labeling conditions. Furthermore, we introduce multi-view pseudo-labeling and a CryoET-specific DropBlock augmentation strategy to further boost performance. Extensive evaluations on the large-scale CZII dataset show that our approach improves F1 by 10% over supervised baselines, underscoring the promise of semi-supervised learning for leveraging unlabeled CryoET data.

Method: MDReID introduces Modality Decoupling Learning (MDL) to decompose features into modality-shared and modality-specific representations, and Modality-aware Metric Learning (MML) to enforce orthogonality and complementarity between components for cross-modality discriminative power.

Result: Extensive experiments on three multi-modality benchmarks show significant improvements: 9.8%, 3.0%, 11.5% mAP gains in modality-matched scenarios, and average 3.4%, 11.8%, 10.9% gains in modality-mismatched scenarios.

Conclusion: MDReID effectively addresses modality inconsistency challenges in real-world ReID systems, demonstrating superior performance in both modality-aligned and mismatched scenarios through explicit feature decomposition and tailored metric learning.

Abstract: Real-world object re-identification (ReID) systems often face modality inconsistencies, where query and gallery images come from different sensors (e.g., RGB, NIR, TIR). However, most existing methods assume modality-matched conditions, which limits their robustness and scalability in practical applications. To address this challenge, we propose MDReID, a flexible any-to-any image-level ReID framework designed to operate under both modality-matched and modality-mismatched scenarios. MDReID builds on the insight that modality information can be decomposed into two components: modality-shared features that are predictable and transferable, and modality-specific features that capture unique, modality-dependent characteristics. To effectively leverage this, MDReID introduces two key components: the Modality Decoupling Learning (MDL) and Modality-aware Metric Learning (MML). Specifically, MDL explicitly decomposes modality features into modality-shared and modality-specific representations, enabling effective retrieval in both modality-aligned and mismatched scenarios. MML, a tailored metric learning strategy, further enforces orthogonality and complementarity between the two components to enhance discriminative power across modalities. Extensive experiments conducted on three challenging multi-modality ReID benchmarks (RGBNT201, RGBNT100, MSVR310) consistently demonstrate the superiority of MDReID. Notably, MDReID achieves significant mAP improvements of 9.8%, 3.0%, and 11.5% in general modality-matched scenarios, and average gains of 3.4%, 11.8%, and 10.9% in modality-mismatched scenarios, respectively. The code is available at: \textcolor{magenta}{https://github.com/stone96123/MDReID}.

Method: A pre-training–fine-tuning pipeline that learns from multiple satellites with global coverage to translate 2D satellite imagery into 3D cloud maps of relevant cloud properties. The model is applied to a custom-built TC dataset to evaluate performance in challenging conditions.

Result: The framework successfully creates global instantaneous 3D cloud maps and accurately reconstructs the 3D structure of intense storms for the first time. The model extends available satellite observations and provides estimates when observations are missing entirely.

Conclusion: This approach is crucial for advancing understanding of TC intensification and improving forecasts by providing comprehensive 3D structural information that was previously unavailable, especially for intense tropical cyclones.

Abstract: Accurate forecasting of tropical cyclones (TCs) remains challenging due to limited satellite observations probing TC structure and difficulties in resolving cloud properties involved in TC intensification. Recent research has demonstrated the capabilities of machine learning methods for 3D cloud reconstruction from satellite observations. However, existing approaches have been restricted to regions where TCs are uncommon, and are poorly validated for intense storms. We introduce a new framework, based on a pre-training–fine-tuning pipeline, that learns from multiple satellites with global coverage to translate 2D satellite imagery into 3D cloud maps of relevant cloud properties. We apply our model to a custom-built TC dataset to evaluate performance in the most challenging and relevant conditions. We show that we can - for the first time - create global instantaneous 3D cloud maps and accurately reconstruct the 3D structure of intense storms. Our model not only extends available satellite observations but also provides estimates when observations are missing entirely. This is crucial for advancing our understanding of TC intensification and improving forecasts.

Method: PINGS-X extends 3D Gaussian splatting with three innovations: (1) normalized Gaussian splatting with formal convergence guarantee, (2) axes-aligned Gaussians for high-dimensional data efficiency while preserving accuracy, and (3) Gaussian merging procedure to prevent degenerate solutions and boost computational efficiency.

Result: Experimental results on CFD and real 4D flow MRI datasets show PINGS-X substantially reduces training time while achieving superior super-resolution accuracy compared to existing methods.

Conclusion: PINGS-X provides an efficient framework for 4D flow MRI super-resolution that overcomes the training time limitations of previous PINN approaches, making high-resolution cardiovascular diagnostics more practical.

Abstract: 4D flow magnetic resonance imaging (MRI) is a reliable, non-invasive approach for estimating blood flow velocities, vital for cardiovascular diagnostics. Unlike conventional MRI focused on anatomical structures, 4D flow MRI requires high spatiotemporal resolution for early detection of critical conditions such as stenosis or aneurysms. However, achieving such resolution typically results in prolonged scan times, creating a trade-off between acquisition speed and prediction accuracy. Recent studies have leveraged physics-informed neural networks (PINNs) for super-resolution of MRI data, but their practical applicability is limited as the prohibitively slow training process must be performed for each patient. To overcome this limitation, we propose PINGS-X, a novel framework modeling high-resolution flow velocities using axes-aligned spatiotemporal Gaussian representations. Inspired by the effectiveness of 3D Gaussian splatting (3DGS) in novel view synthesis, PINGS-X extends this concept through several non-trivial novel innovations: (i) normalized Gaussian splatting with a formal convergence guarantee, (ii) axes-aligned Gaussians that simplify training for high-dimensional data while preserving accuracy and the convergence guarantee, and (iii) a Gaussian merging procedure to prevent degenerate solutions and boost computational efficiency. Experimental results on computational fluid dynamics (CFD) and real 4D flow MRI datasets demonstrate that PINGS-X substantially reduces training time while achieving superior super-resolution accuracy. Our code and datasets are available at https://github.com/SpatialAILab/PINGS-X.

Method: TrAP (Trigger-Aware Prompt tuning) jointly optimizes prompt parameters in both image and text modalities along with visual triggers. It uses a curriculum-based training strategy that progressively shrinks trigger size, enabling effective backdoor activation with small trigger patches at inference.

Result: Experiments across multiple datasets show TrAP achieves high attack success rates for both object misclassification and object disappearance attacks, while also improving clean image performance on downstream datasets compared to zero-shot settings.

Conclusion: The work reveals a new attack surface in open-vocabulary object detectors introduced by prompt tuning, demonstrating that lightweight prompt optimization can effectively implant backdoors while preserving model generalization capabilities.

Abstract: Open-vocabulary object detectors (OVODs) unify vision and language to detect arbitrary object categories based on text prompts, enabling strong zero-shot generalization to novel concepts. As these models gain traction in high-stakes applications such as robotics, autonomous driving, and surveillance, understanding their security risks becomes crucial. In this work, we conduct the first study of backdoor attacks on OVODs and reveal a new attack surface introduced by prompt tuning. We propose TrAP (Trigger-Aware Prompt tuning), a multi-modal backdoor injection strategy that jointly optimizes prompt parameters in both image and text modalities along with visual triggers. TrAP enables the attacker to implant malicious behavior using lightweight, learnable prompt tokens without retraining the base model weights, thus preserving generalization while embedding a hidden backdoor. We adopt a curriculum-based training strategy that progressively shrinks the trigger size, enabling effective backdoor activation using small trigger patches at inference. Experiments across multiple datasets show that TrAP achieves high attack success rates for both object misclassification and object disappearance attacks, while also improving clean image performance on downstream datasets compared to the zero-shot setting. Code: https://github.com/rajankita/TrAP

Method: 1) Agent-based chaining framework converts natural language descriptions to structured parameters; 2) Procedural generation pipeline for floorplan, structure, materials, doors/windows, object layout; 3) SetDepot-Pro dataset of 6,862 film-specific 3D assets and 733 materials.

Result: System produces structurally sound scenes with strong cinematic fidelity. Supports downstream tasks: virtual previs, construction drawings, mood board creation. Validated through experiments and human evaluations.

Conclusion: FilmSceneDesigner successfully automates film set design workflow, addressing labor-intensive manual process while maintaining professional quality and cinematic realism.

Abstract: Film set design plays a pivotal role in cinematic storytelling and shaping the visual atmosphere. However, the traditional process depends on expert-driven manual modeling, which is labor-intensive and time-consuming. To address this issue, we introduce FilmSceneDesigner, an automated scene generation system that emulates professional film set design workflow. Given a natural language description, including scene type, historical period, and style, we design an agent-based chaining framework to generate structured parameters aligned with film set design workflow, guided by prompt strategies that ensure parameter accuracy and coherence. On the other hand, we propose a procedural generation pipeline which executes a series of dedicated functions with the structured parameters for floorplan and structure generation, material assignment, door and window placement, and object retrieval and layout, ultimately constructing a complete film scene from scratch. Moreover, to enhance cinematic realism and asset diversity, we construct SetDepot-Pro, a curated dataset of 6,862 film-specific 3D assets and 733 materials. Experimental results and human evaluations demonstrate that our system produces structurally sound scenes with strong cinematic fidelity, supporting downstream tasks such as virtual previs, construction drawing and mood board creation.

Method: Two novel mechanisms: 1) Motion Direction Decoupling (MDD) module that decomposes temporal dynamics into signed polarity fields to explicitly encode movement occurrence and trajectory direction, recovering lost directional priors. 2) Residual-Driven Spatio-Temporal Refinement (R-STR) head, a Transformer-based module operating on coarse-to-fine paradigm that leverages factorized spatio-temporal contexts to estimate corrective residuals for recovering occluded targets.

Result: Achieves new state-of-the-art on TrackNetV2 dataset with F1-score of 0.9859 and accuracy of 0.9733, significantly outperforming previous versions. Performance achieved with only 3.7% increase in FLOPs compared to V4, maintaining real-time inference capabilities.

Conclusion: TrackNetV5 successfully addresses the limitations of previous versions by integrating motion direction awareness and occlusion handling mechanisms, establishing a robust architecture for fast-moving small object tracking in sports with superior precision and efficiency.

Abstract: The TrackNet series has established a strong baseline for fast-moving small object tracking in sports. However, existing iterations face significant limitations: V1-V3 struggle with occlusions due to a reliance on purely visual cues, while TrackNetV4, despite introducing motion inputs, suffers from directional ambiguity as its absolute difference method discards motion polarity. To overcome these bottlenecks, we propose TrackNetV5, a robust architecture integrating two novel mechanisms. First, to recover lost directional priors, we introduce the Motion Direction Decoupling (MDD) module. Unlike V4, MDD decomposes temporal dynamics into signed polarity fields, explicitly encoding both movement occurrence and trajectory direction. Second, we propose the Residual-Driven Spatio-Temporal Refinement (R-STR) head. Operating on a coarse-to-fine paradigm, this Transformer-based module leverages factorized spatio-temporal contexts to estimate a corrective residual, effectively recovering occluded targets. Extensive experiments on the TrackNetV2 dataset demonstrate that TrackNetV5 achieves a new state-of-the-art F1-score of 0.9859 and an accuracy of 0.9733, significantly outperforming previous versions. Notably, this performance leap is achieved with a marginal 3.7% increase in FLOPs compared to V4, maintaining real-time inference capabilities while delivering superior tracking precision.

Method: Proposes a region perception-based fusion framework using spatially varying exposure (SVE) camera for multi-exposure and multi-modal imaging. Includes region perception-based feature fusion for precise registration, adaptive fusion with contrast enhancement, and structural similarity compensation guided by regional saliency maps. Also adapts to single-exposure scenarios.

Result: Experiments on synthetic and real-world data demonstrate superior image clarity and improved performance compared to state-of-the-art methods, validated by both quantitative and visual evaluations.

Conclusion: The proposed framework effectively addresses the challenge of fusing IR and VIS spectra while preserving geometric fidelity and incorporating thermal radiation, overcoming limitations of existing methods in extreme environments.

Abstract: In photogrammetry, accurately fusing infrared (IR) and visible (VIS) spectra while preserving the geometric fidelity of visible features and incorporating thermal radiation is a significant challenge, particularly under extreme conditions. Existing methods often compromise visible imagery quality, impacting measurement accuracy. To solve this, we propose a region perception-based fusion framework that combines multi-exposure and multi-modal imaging using a spatially varying exposure (SVE) camera. This framework co-fuses multi-modal and multi-exposure data, overcoming single-exposure method limitations in extreme environments. The framework begins with region perception-based feature fusion to ensure precise multi-modal registration, followed by adaptive fusion with contrast enhancement. A structural similarity compensation mechanism, guided by regional saliency maps, optimizes IR-VIS spectral integration. Moreover, the framework adapts to single-exposure scenarios for robust fusion across different conditions. Experiments conducted on both synthetic and real-world data demonstrate superior image clarity and improved performance compared to state-of-the-art methods, as evidenced by both quantitative and visual evaluations.

Method: Evaluated 9 open/closed-source VLMs on 2 visually prompted tasks, testing sensitivity to marker design, dataset size, and low-level inference choices like JPEG compression.

Result: Visual prompting details have outsized impact: changing marker color/size can flip rankings; smaller models can outperform larger ones with optimal markers; JPEG compression affects results.

Conclusion: Created VPBench with 16 visual marker variants to provide more stable VLM evaluation, highlighting the need for robustness testing in visual prompting benchmarks.

Abstract: A key challenge in evaluating VLMs is testing models’ ability to analyze visual content independently from their textual priors. Recent benchmarks such as BLINK probe visual perception through visual prompting, where questions about visual content are paired with coordinates to which the question refers, with the coordinates explicitly marked in the image itself. While these benchmarks are an important part of VLM evaluation, we find that existing models are surprisingly fragile to seemingly irrelevant details of visual prompting: simply changing a visual marker from red to blue can completely change rankings among models on a leaderboard. By evaluating nine commonly-used open- and closed-source VLMs on two visually prompted tasks, we demonstrate how details in benchmark setup, including visual marker design and dataset size, have a significant influence on model performance and leaderboard rankings. These effects can even be exploited to lift weaker models above stronger ones; for instance, slightly increasing the size of the visual marker results in open-source InternVL3-8B ranking alongside or better than much larger proprietary models like Gemini 2.5 Pro. We further show that low-level inference choices that are often ignored in benchmarking, such as JPEG compression levels in API calls, can also cause model lineup changes. These details have substantially larger impacts on visually prompted benchmarks than on conventional semantic VLM evaluations. To mitigate this instability, we curate existing datasets to create VPBench, a larger visually prompted benchmark with 16 visual marker variants. We open-source VPBench and our analysis framework at: https://lisadunlap.github.io/vpbench/.

Method: Decomposes synthesis into deterministic anatomical reconstruction and stochastic deviation modeling using a Deviation-Space Diffusion Model that learns conditional distributions of pathological residuals. Includes seam-aware fusion strategy and inference-time stabilization module for spatial coherence.

Result: Quantitatively and qualitatively outperforms holistic diffusion and mask-conditioned baselines on tumor imaging benchmarks in both perceptual realism and anatomical fidelity.

Conclusion: Provides a mathematically principled pipeline for generating patient-specific synthetic datasets, enabling robust diagnostic algorithm development in low-data regimes, interpretable counterfactual disease progression modeling, and clinical decision-support benchmarking.

Abstract: We present PathoSyn, a unified generative framework for Magnetic Resonance Imaging (MRI) image synthesis that reformulates imaging-pathology as a disentangled additive deviation on a stable anatomical manifold. Current generative models typically operate in the global pixel domain or rely on binary masks, these paradigms often suffer from feature entanglement, leading to corrupted anatomical substrates or structural discontinuities. PathoSyn addresses these limitations by decomposing the synthesis task into deterministic anatomical reconstruction and stochastic deviation modeling. Central to our framework is a Deviation-Space Diffusion Model designed to learn the conditional distribution of pathological residuals, thereby capturing localized intensity variations while preserving global structural integrity by construction. To ensure spatial coherence, the diffusion process is coupled with a seam-aware fusion strategy and an inference-time stabilization module, which collectively suppress boundary artifacts and produce high-fidelity internal lesion heterogeneity. PathoSyn provides a mathematically principled pipeline for generating high-fidelity patient-specific synthetic datasets, facilitating the development of robust diagnostic algorithms in low-data regimes. By allowing interpretable counterfactual disease progression modeling, the framework supports precision intervention planning and provides a controlled environment for benchmarking clinical decision-support systems. Quantitative and qualitative evaluations on tumor imaging benchmarks demonstrate that PathoSyn significantly outperforms holistic diffusion and mask-conditioned baselines in both perceptual realism and anatomical fidelity. The source code of this work will be made publicly available.

Method: Uses DINOv3 descriptors to establish dense multi-view correspondences, refines them with confidence-aware inlier classifier, performs one-shot triangulation to create structure-aware Gaussian seeds before optimization, replacing the progressive densification stage.

Result: Achieves 20% faster convergence, higher rendering fidelity in texture-rich and cluttered scenes, competitive/superior localization and reconstruction accuracy on TUM RGB-D and Replica datasets, with real-time mapping at up to 925 FPS.

Conclusion: RGS-SLAM demonstrates that training-free correspondence-to-Gaussian initialization provides more stable and efficient Gaussian-splatting SLAM, offering better performance while remaining fully compatible with existing GS-SLAM pipelines.

Abstract: We introduce RGS-SLAM, a robust Gaussian-splatting SLAM framework that replaces the residual-driven densification stage of GS-SLAM with a training-free correspondence-to-Gaussian initialization. Instead of progressively adding Gaussians as residuals reveal missing geometry, RGS-SLAM performs a one-shot triangulation of dense multi-view correspondences derived from DINOv3 descriptors refined through a confidence-aware inlier classifier, generating a well-distributed and structure-aware Gaussian seed prior to optimization. This initialization stabilizes early mapping and accelerates convergence by roughly 20%, yielding higher rendering fidelity in texture-rich and cluttered scenes while remaining fully compatible with existing GS-SLAM pipelines. Evaluated on the TUM RGB-D and Replica datasets, RGS-SLAM achieves competitive or superior localization and reconstruction accuracy compared with state-of-the-art Gaussian and point-based SLAM systems, sustaining real-time mapping performance at up to 925 FPS. Project page:https://breeze1124.github.io/rgs-slam-project-page/

Method: Combines multiple LiDAR inputs with high-resolution color images and applies a joint bilateral filtering strategy implemented through a convolutional neural network architecture.

Result: Produces dense depth maps at full HD resolution in real time (30 fps), which is over 15x faster than recent training-based depth completion approaches, with accurate geometry and no multiview inconsistencies or ghosting artifacts.

Conclusion: The method successfully enables high-speed LiDAR point cloud densification for immersive telepresence applications, achieving real-time performance with high-quality dense 3D scene reconstruction.

Abstract: To realize low-latency spatial transmission system for immersive telepresence, there are two major problems: capturing dynamic 3D scene densely and processing them in real time. LiDAR sensors capture 3D in real time, but produce sparce point clouds. Therefore, this paper presents a high-speed LiDAR point cloud densification method to generate dense 3D scene with minimal latency, addressing the need for on-the-fly depth completion while maintaining real-time performance. Our approach combines multiple LiDAR inputs with high-resolution color images and applies a joint bilateral filtering strategy implemented through a convolutional neural network architecture. Experiments demonstrate that the proposed method produces dense depth maps at full HD resolution in real time (30 fps), which is over 15x faster than a recent training-based depth completion approach. The resulting dense point clouds exhibit accurate geometry without multiview inconsistencies or ghosting artifacts.

Method: Parallelized Alternating Least Squares (ALS) framework with extensive hyperparameter optimization, visualizing learned embeddings and testing cold-start scenarios with tunable scoring parameters.

Result: Constrained low-rank models significantly outperform higher dimensional counterparts in generalization (RMSE and ranking precision), reveal semantic genre clusters, and effectively manage popularity bias vs personalized affinity trade-off in cold-start.

Conclusion: The ALS framework successfully balances performance and interpretability, capturing deep structural relationships from interaction data while providing practical utility for real-world recommendation scenarios.

Abstract: Scalability and data sparsity remain critical bottlenecks for collaborative filtering on massive interaction datasets. This work investigates the latent geometry of user preferences using the MovieLens 32M dataset, implementing a high-performance, parallelized Alternating Least Squares (ALS) framework. Through extensive hyperparameter optimization, we demonstrate that constrained low-rank models significantly outperform higher dimensional counterparts in generalization, achieving an optimal balance between Root Mean Square Error (RMSE) and ranking precision. We visualize the learned embedding space to reveal the unsupervised emergence of semantic genre clusters, confirming that the model captures deep structural relationships solely from interaction data. Finally, we validate the system’s practical utility in a cold-start scenario, introducing a tunable scoring parameter to manage the trade-off between popularity bias and personalized affinity effectively. The codebase for this research can be found here: https://github.com/joshsalako/recommender.git

Method: Three-module framework: (1) Enhanced visual localization with planning/captioning stages before reasoning, using quality-based caption reward; (2) Exploration improvement by categorizing samples based on reward distribution mean/variance, prioritizing high-variance samples; (3) Sample interference mitigation via NTK similarity regulation using grouping and InfoNCE loss to balance gradient interactions.

Result: Experimental results show the proposed method significantly reduces hallucination rates and effectively enhances MLLMs’ inference accuracy.

Conclusion: The comprehensive framework addresses key hallucination causes in MLLMs during RL training, providing systematic solutions that improve model reliability and practical deployment viability.

Abstract: While Multimodal Large Language Models (MLLMs) have achieved remarkable success across diverse tasks, their practical deployment is severely hindered by hallucination issues, which become particularly acute during Reinforcement Learning (RL) optimization. This paper systematically analyzes the root causes of hallucinations in MLLMs under RL training, identifying three critical factors: (1) an over-reliance on chained visual reasoning, where inaccurate initial descriptions or redundant information anchor subsequent inferences to incorrect premises; (2) insufficient exploration diversity during policy optimization, leading the model to generate overly confident but erroneous outputs; and (3) destructive conflicts between training samples, where Neural Tangent Kernel (NTK) similarity causes false associations and unstable parameter updates. To address these challenges, we propose a comprehensive framework comprising three core modules. First, we enhance visual localization by introducing dedicated planning and captioning stages before the reasoning phase, employing a quality-based caption reward to ensure accurate initial anchoring. Second, to improve exploration, we categorize samples based on the mean and variance of their reward distributions, prioritizing samples with high variance to focus the model on diverse and informative data. Finally, to mitigate sample interference, we regulate NTK similarity by grouping sample pairs and applying an InfoNCE loss to push overly similar pairs apart and pull dissimilar ones closer, thereby guiding gradient interactions toward a balanced range. Experimental results demonstrate that our proposed method significantly reduces hallucination rates and effectively enhances the inference accuracy of MLLMs.

Method: MVGGT (Multimodal Visual Geometry Grounded Transformer): End-to-end dual-branch framework integrating language into sparse-view geometric reasoning. Introduces PVSO (Per-view No-target Suppression Optimization) to solve Foreground Gradient Dilution problem in sparse 3D supervision.

Result: MVGGT establishes first strong baseline for MV-3DRES, achieving high accuracy and fast inference, outperforming existing alternatives. Also introduces MVRefer benchmark for standardized evaluation.

Conclusion: MVGGT enables efficient 3D referring expression segmentation directly from sparse multi-view images, addressing practical constraints of real-world agents while maintaining accuracy and speed.

Abstract: Most existing 3D referring expression segmentation (3DRES) methods rely on dense, high-quality point clouds, while real-world agents such as robots and mobile phones operate with only a few sparse RGB views and strict latency constraints. We introduce Multi-view 3D Referring Expression Segmentation (MV-3DRES), where the model must recover scene structure and segment the referred object directly from sparse multi-view images. Traditional two-stage pipelines, which first reconstruct a point cloud and then perform segmentation, often yield low-quality geometry, produce coarse or degraded target regions, and run slowly. We propose the Multimodal Visual Geometry Grounded Transformer (MVGGT), an efficient end-to-end framework that integrates language information into sparse-view geometric reasoning through a dual-branch design. Training in this setting exposes a critical optimization barrier, termed Foreground Gradient Dilution (FGD), where sparse 3D signals lead to weak supervision. To resolve this, we introduce Per-view No-target Suppression Optimization (PVSO), which provides stronger and more balanced gradients across views, enabling stable and efficient learning. To support consistent evaluation, we build MVRefer, a benchmark that defines standardized settings and metrics for MV-3DRES. Experiments show that MVGGT establishes the first strong baseline and achieves both high accuracy and fast inference, outperforming existing alternatives. Code and models are publicly available at https://mvggt.github.io.

Method: ViewMorpher3D uses image diffusion models to jointly process multiple rendered views conditioned on camera poses, 3D geometric priors, and temporally/spatially adjacent reference views to infer missing details, suppress artifacts, and enforce cross-view consistency.

Result: Experiments on real-world driving datasets show substantial improvements in image quality metrics, effectively reducing artifacts while preserving geometric fidelity.

Conclusion: ViewMorpher3D provides an adaptable framework for enhancing multi-view image quality in driving simulations, supporting variable camera setups and flexible view configurations to improve photorealism for autonomous driving system development.

Abstract: Autonomous driving systems rely heavily on multi-view images to ensure accurate perception and robust decision-making. To effectively develop and evaluate perception stacks and planning algorithms, realistic closed-loop simulators are indispensable. While 3D reconstruction techniques such as Gaussian Splatting offer promising avenues for simulator construction, the rendered novel views often exhibit artifacts, particularly in extrapolated perspectives or when available observations are sparse. We introduce ViewMorpher3D, a multi-view image enhancement framework based on image diffusion models, designed to elevate photorealism and multi-view coherence in driving scenes. Unlike single-view approaches, ViewMorpher3D jointly processes a set of rendered views conditioned on camera poses, 3D geometric priors, and temporally adjacent or spatially overlapping reference views. This enables the model to infer missing details, suppress rendering artifacts, and enforce cross-view consistency. Our framework accommodates variable numbers of cameras and flexible reference/target view configurations, making it adaptable to diverse sensor setups. Experiments on real-world driving datasets demonstrate substantial improvements in image quality metrics, effectively reducing artifacts while preserving geometric fidelity.

Method: 1) Creates a large-scale dataset of triplets (reference effect video, input image/video, output video) using an automated pipeline for video-to-video effects; 2) Augments with image-to-video effects from LoRA adapters and code-based temporal effects; 3) Trains a reference-conditioned model using recent text-to-video backbones.

Result: RefVFX produces visually consistent and temporally coherent edits, generalizes across unseen effect categories, and outperforms prompt-only baselines in both quantitative metrics and human preference evaluations.

Conclusion: The framework successfully enables transfer of complex temporal effects from reference videos to target content in a feed-forward manner, addressing key limitations of existing text-based video editing methods.

Abstract: We present RefVFX, a new framework that transfers complex temporal effects from a reference video onto a target video or image in a feed-forward manner. While existing methods excel at prompt-based or keyframe-conditioned editing, they struggle with dynamic temporal effects such as dynamic lighting changes or character transformations, which are difficult to describe via text or static conditions. Transferring a video effect is challenging, as the model must integrate the new temporal dynamics with the input video’s existing motion and appearance. % To address this, we introduce a large-scale dataset of triplets, where each triplet consists of a reference effect video, an input image or video, and a corresponding output video depicting the transferred effect. Creating this data is non-trivial, especially the video-to-video effect triplets, which do not exist naturally. To generate these, we propose a scalable automated pipeline that creates high-quality paired videos designed to preserve the input’s motion and structure while transforming it based on some fixed, repeatable effect. We then augment this data with image-to-video effects derived from LoRA adapters and code-based temporal effects generated through programmatic composition. Building on our new dataset, we train our reference-conditioned model using recent text-to-video backbones. Experimental results demonstrate that RefVFX produces visually consistent and temporally coherent edits, generalizes across unseen effect categories, and outperforms prompt-only baselines in both quantitative metrics and human preference. See our website https://tuningfreevisualeffects-maker.github.io/Tuning-free-Visual-Effect-Transfer-across-Videos-Project-Page/

Method: Developed fuzzy-XGBoost model on 1,014 maternal health records, combining ante-hoc fuzzy logic with post-hoc SHAP explanations, validated with clinician feedback from 14 healthcare professionals in Bangladesh.

Result: Achieved 88.67% accuracy (ROC-AUC: 0.9703); clinicians strongly preferred hybrid explanations (71.4%), with 54.8% expressing trust for clinical use; SHAP identified healthcare access as primary predictor.

Conclusion: Combining interpretable fuzzy rules with feature importance explanations enhances utility and trust, providing practical insights for XAI deployment in maternal healthcare.

Abstract: While machine learning shows promise for maternal health risk prediction, clinical adoption in resource-constrained settings faces a critical barrier: lack of explainability and trust. This study presents a hybrid explainable AI (XAI) framework combining ante-hoc fuzzy logic with post-hoc SHAP explanations, validated through systematic clinician feedback. We developed a fuzzy-XGBoost model on 1,014 maternal health records, achieving 88.67% accuracy (ROC-AUC: 0.9703). A validation study with 14 healthcare professionals in Bangladesh revealed strong preference for hybrid explanations (71.4% across three clinical cases) with 54.8% expressing trust for clinical use. SHAP analysis identified healthcare access as the primary predictor, with the engineered fuzzy risk score ranking third, validating clinical knowledge integration (r=0.298). Clinicians valued integrated clinical parameters but identified critical gaps: obstetric history, gestational age, and connectivity barriers. This work demonstrates that combining interpretable fuzzy rules with feature importance explanations enhances both utility and trust, providing practical insights for XAI deployment in maternal healthcare.

Method: Uses Executable Ontologies implemented through the boldsea framework, applied to a survival game scenario called “Winter Feast.” Demonstrates priority-based task interruption through dataflow conditions rather than explicit preemption logic. Compares with traditional approaches like Behavior Trees and Goal-Oriented Action Planning.

Result: EO enables agent behavior to emerge naturally from declarative domain rules, achieving priority-based task interruption through dataflow conditions. The approach models when actions become possible rather than just what agents should do, addressing the semantic-process gap. The framework offers debugging advantages through temporal event graphs and potential for LLM-driven runtime model generation.

Conclusion: Executable Ontologies represent a paradigm shift in game AI development, moving from algorithmic behavior programming to semantic world modeling. This approach better addresses the semantic-process gap, offers improved debugging capabilities, and has potential for integration with modern AI techniques like LLMs for runtime model generation.

Abstract: This paper examines the application of Executable Ontologies (EO), implemented through the boldsea framework, to game development. We argue that EO represents a paradigm shift: a transition from algorithmic behavior programming to semantic world modeling, where agent behavior emerges naturally from declarative domain rules rather than being explicitly coded. Using a survival game scenario (Winter Feast), we demonstrate how EO achieves prioritybased task interruption through dataflow conditions rather than explicit preemption logic. Comparison with Behavior Trees (BT) and Goal-Oriented Action Planning (GOAP) reveals that while these approaches model what agents should do, EO models when actions become possible - a fundamental difference that addresses the semantic-process gap in game AI architecture. We discuss integration strategies, debugging advantages inherent to temporal event graphs, and the potential for LLM-driven runtime model generation.

Method: A simple, effective, training-free calibration method that works for both vision and language models. The approach builds on two empirical observations: higher confidence correlates with higher accuracy within a model, and calibration transfers from validation to test sets. This enables calibrated confidence as a reliable signal.

Result: The method enables two key applications: 1) Model cascading with calibrated advantage routing that improves efficiency with minimal accuracy loss, and can even combine comparable models to exceed individual performance. 2) Data cleaning using model ensembles that balances precision and detection to identify mislabeled samples in ImageNet and MMLU datasets.

Conclusion: Enabling models to recognize when they don’t know is a practical step toward more efficient, reliable, and trustworthy AI systems. The proposed calibration method provides a universal, training-free approach to achieve this capability.

Abstract: When a model knows when it does not know, many possibilities emerge. The first question is how to enable a model to recognize that it does not know. A promising approach is to use confidence, computed from the model’s internal signals, to reflect its ignorance. Prior work in specific domains has shown that calibration can provide reliable confidence estimates. In this work, we propose a simple, effective, and universal training-free method that applies to both vision and language models, performing model calibration, cascading, and data cleaning to better exploit a model’s ability to recognize when it does not know. We first highlight two key empirical observations: higher confidence corresponds to higher accuracy within a single model, and models calibrated on the validation set remain calibrated on a held-out test set. These findings empirically establish the reliability and comparability of calibrated confidence. Building on this, we introduce two applications: (1) model cascading with calibrated advantage routing and (2) data cleaning based on model ensemble. Using the routing signal derived from the comparability of calibrated confidences, we cascade large and small models to improve efficiency with almost no compromise in accuracy, and we further cascade two models of comparable scale to achieve performance beyond either model alone. Leveraging multiple experts and their calibrated confidences, we design a simple yet effective data-cleaning method that balances precision and detection rate to identify mislabeled samples in ImageNet and Massive Multitask Language Understanding (MMLU) datasets. Our results demonstrate that enabling models to recognize when they do not know is a practical step toward more efficient, reliable, and trustworthy AI.

Method: Propose CADA (Case-Augmented Deliberative Alignment) using reinforcement learning on self-generated safety reasoning chains. Instead of extensive code-like rules, use case-augmented simple codes to guide LLMs through reasoning rather than rigid rule-following.

Result: CADA effectively enhances harmlessness, improves robustness against attacks, reduces over-refusal while preserving utility across diverse benchmarks. Outperforms rule-only deliberative alignment by enabling broader adaptability and more generalized safety behaviors.

Conclusion: Case-augmented reasoning is more effective than extensive code-like safety rules for LLM alignment. CADA offers a practical alternative that improves safety while maintaining helpfulness, avoiding the trade-offs of rule-only approaches.

Abstract: Ensuring that Large Language Models (LLMs) adhere to safety principles without refusing benign requests remains a significant challenge. While OpenAI introduces deliberative alignment (DA) to enhance the safety of its o-series models through reasoning over detailed ``code-like’’ safety rules, the effectiveness of this approach in open-source LLMs, which typically lack advanced reasoning capabilities, is understudied. In this work, we systematically evaluate the impact of explicitly specifying extensive safety codes versus demonstrating them through illustrative cases. We find that referencing explicit codes inconsistently improves harmlessness and systematically degrades helpfulness, whereas training on case-augmented simple codes yields more robust and generalized safety behaviors. By guiding LLMs with case-augmented reasoning instead of extensive code-like safety rules, we avoid rigid adherence to narrowly enumerated rules and enable broader adaptability. Building on these insights, we propose CADA, a case-augmented deliberative alignment method for LLMs utilizing reinforcement learning on self-generated safety reasoning chains. CADA effectively enhances harmlessness, improves robustness against attacks, and reduces over-refusal while preserving utility across diverse benchmarks, offering a practical alternative to rule-only DA for improving safety while maintaining helpfulness.

Method: The paper examines how 2025 US and EU frontier AI regulations handle internal deployment, identifies three specific regulatory gaps, analyzes why these gaps persist (examining tensions around measurability, incentives, and information access), and maps potential approaches to address them.

Result: Three key regulatory gaps identified: (1) scope ambiguity allowing internal systems to evade obligations, (2) point-in-time compliance assessments failing to capture continuous system evolution, and (3) information asymmetries subverting regulatory awareness and oversight.

Conclusion: By understanding these patterns, policymakers can make deliberate rather than incidental choices about regulating internally deployed AI systems, though addressing the gaps involves tradeoffs between oversight effectiveness and practical implementation.

Abstract: Frontier AI regulations primarily focus on systems deployed to external users, where deployment is more visible and subject to outside scrutiny. However, high-stakes applications can occur internally when companies deploy highly capable systems within their own organizations, such as for automating R&D, accelerating critical business processes, and handling sensitive proprietary data. This paper examines how frontier AI regulations in the United States and European Union in 2025 handle internal deployment. We identify three gaps that could cause internally-deployed systems to evade intended oversight: (1) scope ambiguity that allows internal systems to evade regulatory obligations, (2) point-in-time compliance assessments that fail to capture the continuous evolution of internal systems, and (3) information asymmetries that subvert regulatory awareness and oversight. We then analyze why these gaps persist, examining tensions around measurability, incentives, and information access. Finally, we map potential approaches to address them and their associated tradeoffs. By understanding these patterns, we hope that policy choices around internally deployed AI systems can be made deliberately rather than incidentally.

Method: SCASED uses Raspberry Pi camera with OpenCV for face detection and a fine-tuned MobileNetV2 model to classify four learning-related emotional states (engagement, boredom, confusion, frustration). It implements a session-based mechanism for attendance recording and continuous emotion analysis, with data visualized through a cloud-based dashboard.

Result: Experimental evaluation using the DAiSEE dataset achieved 89.5% emotion classification accuracy. The system successfully integrates attendance data with emotion analytics to provide insights into classroom dynamics.

Conclusion: Integrating attendance tracking with emotion recognition can provide instructors with valuable insights into classroom engagement patterns, supporting more responsive teaching practices and better understanding of student learning states.

Abstract: The increasing adoption of smart classroom technologies in higher education has mainly focused on automating attendance, with limited attention given to students’ emotional and cognitive engagement during lectures. This limits instructors’ ability to identify disengagement and adapt teaching strategies in real time. This paper presents SCASED (Smart Classroom Attendance System with Emotion Detection), an IoT-based system that integrates automated attendance tracking with facial emotion recognition to support classroom engagement monitoring. The system uses a Raspberry Pi camera and OpenCV for face detection, and a finetuned MobileNetV2 model to classify four learning-related emotional states: engagement, boredom, confusion, and frustration. A session-based mechanism is implemented to manage attendance and emotion monitoring by recording attendance once per session and performing continuous emotion analysis thereafter. Attendance and emotion data are visualized through a cloud-based dashboard to provide instructors with insights into classroom dynamics. Experimental evaluation using the DAiSEE dataset achieved an emotion classification accuracy of 89.5%. The results show that integrating attendance data with emotion analytics can provide instructors with additional insight into classroom dynamics and support more responsive teaching practices.

Method: Proposes two PPO variants: Forecast Aware PPO (incorporates short-term demand/renewable forecasts using hour/month residual calibration) and PID KL PPO (uses PID controller to adaptively regulate KL divergence for stable policy updates). Focuses on battery storage and water heating scheduling under operational constraints.

Result: Achieves up to 1% lower electricity cost than standard PPO, 4.8% lower than DQN, and 1.5% lower than SAC. For battery scheduling, reduces grid imports by 13.1%, demonstrating scalability and effectiveness.

Conclusion: The proposed DRL framework enables sustainable energy management in dairy farming by addressing realistic operational constraints and training instability, supporting SDG 7 goals through reduced grid dependence and operational costs.

Abstract: Dairy farming is an energy intensive sector that relies heavily on grid electricity. With increasing renewable energy integration, sustainable energy management has become essential for reducing grid dependence and supporting the United Nations Sustainable Development Goal 7 on affordable and clean energy. However, the intermittent nature of renewables poses challenges in balancing supply and demand in real time. Intelligent load scheduling is therefore crucial to minimize operational costs while maintaining reliability. Reinforcement Learning has shown promise in improving energy efficiency and reducing costs. However, most RL-based scheduling methods assume complete knowledge of future prices or generation, which is unrealistic in dynamic environments. Moreover, standard PPO variants rely on fixed clipping or KL divergence thresholds, often leading to unstable training under variable tariffs. To address these challenges, this study proposes a Deep Reinforcement Learning framework for efficient load scheduling in dairy farms, focusing on battery storage and water heating under realistic operational constraints. The proposed Forecast Aware PPO incorporates short term forecasts of demand and renewable generation using hour of day and month based residual calibration, while the PID KL PPO variant employs a proportional integral derivative controller to regulate KL divergence for stable policy updates adaptively. Trained on real world dairy farm data, the method achieves up to 1% lower electricity cost than PPO, 4.8% than DQN, and 1.5% than SAC. For battery scheduling, PPO reduces grid imports by 13.1%, demonstrating scalability and effectiveness for sustainable energy management in modern dairy farming.

Method: Proposes a generative diffusion-guided framework that transmits highly-compressed semantic information only, then uses diffusion models with spatially-adaptive normalizations to synthesize semantic-consistent scenes from denoised semantic information.

Result: Outperforms existing solutions in generating high-quality images with preserved semantic information even with significantly degraded received content. Objects, locations, and depths remain recognizable under extremely noisy channel conditions.

Conclusion: The proposed diffusion-guided semantic communication framework successfully bridges the gap between generation effectiveness and transmission complexity, enabling robust semantic-aware image regeneration in next-generation AI-based communications.

Abstract: Semantic communication is expected to be one of the cores of next-generation AI-based communications. One of the possibilities offered by semantic communication is the capability to regenerate, at the destination side, images or videos semantically equivalent to the transmitted ones, without necessarily recovering the transmitted sequence of bits. The current solutions still lack the ability to build complex scenes from the received partial information. Clearly, there is an unmet need to balance the effectiveness of generation methods and the complexity of the transmitted information, possibly taking into account the goal of communication. In this paper, we aim to bridge this gap by proposing a novel generative diffusion-guided framework for semantic communication that leverages the strong abilities of diffusion models in synthesizing multimedia content while preserving semantic features. We reduce bandwidth usage by sending highly-compressed semantic information only. Then, the diffusion model learns to synthesize semantic-consistent scenes through spatially-adaptive normalizations from such denoised semantic information. We prove, through an in-depth assessment of multiple scenarios, that our method outperforms existing solutions in generating high-quality images with preserved semantic information even in cases where the received content is significantly degraded. More specifically, our results show that objects, locations, and depths are still recognizable even in the presence of extremely noisy conditions of the communication channel. The code is available at https://github.com/ispamm/GESCO.

Method: Introduces new algorithms that compute both over- and under-approximations of backward reachable sets, then integrates these backward algorithms with established forward analysis techniques.

Result: Creates a unified verification framework for neural feedback systems that combines both forward and backward reachability analysis approaches.

Conclusion: The integration of new backward reachability algorithms with forward analysis provides a more comprehensive verification framework for neural feedback systems, addressing scalability limitations of previous backward methods.

Abstract: Forward reachability analysis is the predominant approach for verifying reach-avoid properties in neural feedback systems (dynamical systems controlled by neural networks). This dominance stems from the limited scalability of existing backward reachability methods. In this work, we introduce new algorithms that compute both over- and under-approximations of backward reachable sets for such systems. We further integrate these backward algorithms with established forward analysis techniques to yield a unified verification framework for neural feedback systems.

Method: Introduced VGC-Bench benchmark with infrastructure, standardized evaluation protocols, human-play dataset of 700k+ battle logs, and baseline agents using heuristics, LLMs, behavior cloning, and multi-agent RL with game-theoretic methods (self-play, fictitious play, double oracle).

Result: In restricted mirror matches with single team configurations, methods can beat professional VGC competitors. As team sets expand, best single-team algorithms show worse performance and increased exploitability but improved generalization to unseen teams.

Conclusion: VGC-Bench provides essential infrastructure for studying generalization in complex multi-agent environments with vast combinatorial spaces, revealing trade-offs between specialization and generalization as team diversity increases.

Abstract: Developing AI agents that can robustly adapt to varying strategic landscapes without retraining is a central challenge in multi-agent learning. Pokémon Video Game Championships (VGC) is a domain with a vast space of approximately $10^{139}$ team configurations, far larger than those of other games such as Chess, Go, Poker, StarCraft, or Dota. The combinatorial nature of team building in Pokémon VGC causes optimal strategies to vary substantially depending on both the controlled team and the opponent’s team, making generalization uniquely challenging. To advance research on this problem, we introduce VGC-Bench: a benchmark that provides critical infrastructure, standardizes evaluation protocols, and supplies a human-play dataset of over 700,000 battle logs and a range of baseline agents based on heuristics, large language models, behavior cloning, and multi-agent reinforcement learning with empirical game-theoretic methods such as self-play, fictitious play, and double oracle. In the restricted setting where an agent is trained and evaluated in a mirror match with a single team configuration, our methods can win against a professional VGC competitor. We repeat this training and evaluation with progressively larger team sets and find that as the number of teams increases, the best-performing algorithm in the single-team setting has worse performance and is more exploitable, but has improved generalization to unseen teams. Our code and dataset are open-sourced at https://github.com/cameronangliss/vgc-bench and https://huggingface.co/datasets/cameronangliss/vgc-battle-logs.

Method: Introduced semantic pressure as a quantitative measure of intrinsic probability of generating forbidden tokens; used logistic regression analysis; conducted layer-wise analysis with logit lens technique; identified two failure modes through activation patching in layers 23-27.

Result: Violation probability follows a tight logistic relationship with semantic pressure; suppression signal is 4.4× weaker in failures; identified two failure modes: priming failure (87.5%) where instruction activates forbidden word, and override failure (12.5%) where late-layer networks generate +0.39 probability contributions overwhelming suppression.

Conclusion: Negative constraint design has a fundamental tension: naming a forbidden word paradoxically primes the model to produce it, revealing systematic weaknesses in LLMs’ instruction-following capabilities for negative constraints.

Abstract: Negative constraints (instructions of the form “do not use word X”) represent a fundamental test of instruction-following capability in large language models. Despite their apparent simplicity, these constraints fail with striking regularity, and the conditions governing failure have remained poorly understood. This paper presents the first comprehensive mechanistic investigation of negative instruction failure. We introduce semantic pressure, a quantitative measure of the model’s intrinsic probability of generating the forbidden token, and demonstrate that violation probability follows a tight logistic relationship with pressure ($p=σ(-2.40+2.27\cdot P_0)$; $n=40{,}000$ samples; bootstrap $95%$ CI for slope: $[2.21,,2.33]$). Through layer-wise analysis using the logit lens technique, we establish that the suppression signal induced by negative instructions is present but systematically weaker in failures: the instruction reduces target probability by only 5.2 percentage points in failures versus 22.8 points in successes – a $4.4\times$ asymmetry. We trace this asymmetry to two mechanistically distinct failure modes. In priming failure (87.5% of violations), the instruction’s explicit mention of the forbidden word paradoxically activates rather than suppresses the target representation. In override failure (12.5%), late-layer feed-forward networks generate contributions of $+0.39$ toward the target probability – nearly $4\times$ larger than in successes – overwhelming earlier suppression signals. Activation patching confirms that layers 23–27 are causally responsible: replacing these layers’ activations flips the sign of constraint effects. These findings reveal a fundamental tension in negative constraint design: the very act of naming a forbidden word primes the model to produce it.

Method: MemoBrain constructs a dependency-aware memory over reasoning steps, capturing intermediate states and logical relations. It operates as a co-pilot that organizes reasoning progress without blocking execution, actively managing working context by pruning invalid steps, folding completed sub-trajectories, and preserving a compact, high-salience reasoning backbone.

Result: MemoBrain demonstrates consistent improvements over strong baselines on challenging long-horizon benchmarks including GAIA, WebWalker, and BrowseComp-Plus.

Conclusion: Memory is not just an auxiliary efficiency concern but a core component for sustaining coherent, goal-directed reasoning over long horizons in tool-augmented agents, and MemoBrain provides explicit cognitive control over reasoning trajectories rather than passive context accumulation.

Abstract: Complex reasoning in tool-augmented agent frameworks is inherently long-horizon, causing reasoning traces and transient tool artifacts to accumulate and strain the bounded working context of large language models. Without explicit memory mechanisms, such accumulation disrupts logical continuity and undermines task alignment. This positions memory not as an auxiliary efficiency concern, but as a core component for sustaining coherent, goal-directed reasoning over long horizons. We propose MemoBrain, an executive memory model for tool-augmented agents that constructs a dependency-aware memory over reasoning steps, capturing salient intermediate states and their logical relations. Operating as a co-pilot alongside the reasoning agent, MemoBrain organizes reasoning progress without blocking execution and actively manages the working context. Specifically, it prunes invalid steps, folds completed sub-trajectories, and preserves a compact, high-salience reasoning backbone under a fixed context budget. Together, these mechanisms enable explicit cognitive control over reasoning trajectories rather than passive context accumulation. We evaluate MemoBrain on challenging long-horizon benchmarks, including GAIA, WebWalker, and BrowseComp-Plus, demonstrating consistent improvements over strong baselines.

Method: A modular benchmarking framework with typed interfaces, metadata-driven registries, multi-backend support, caching, and observability. It supports pluggable user proxies, datasets, tasks, and metrics, including three lexical-diversity metrics (MATTR, YULE’S K, HD-D) and three LLM-judge-based metrics (GTEval, Pairwise Indistinguishability, Rubric-and-Reason).

Result: Across four open datasets, MIRRORBENCH produces variance-aware results and reveals systematic gaps between user proxies and real human users.

Conclusion: MIRRORBENCH provides a reproducible, extensible framework for evaluating user proxy agents on human-likeness, decoupled from task success, with open-source implementation available.

Abstract: Large language models (LLMs) are increasingly used as human simulators, both for evaluating conversational systems and for generating fine-tuning data. However, naive “act-as-a-user” prompting often yields verbose, unrealistic utterances, underscoring the need for principled evaluation of so-called user proxy agents. We present MIRRORBENCH, a reproducible, extensible benchmarking framework that evaluates user proxies solely on their ability to produce human-like user utterances across diverse conversational tasks, explicitly decoupled from downstream task success. MIRRORBENCH features a modular execution engine with typed interfaces, metadata-driven registries, multi-backend support, caching, and robust observability. The system supports pluggable user proxies, datasets, tasks, and metrics, enabling researchers to evaluate arbitrary simulators under a uniform, variance-aware harness. We include three lexical-diversity metrics (MATTR, YULE’S K, and HD-D) and three LLM-judge-based metrics (GTEval, Pairwise Indistinguishability, and Rubric-and-Reason). Across four open datasets, MIRRORBENCH yields variance-aware results and reveals systematic gaps between user proxies and real human users. The framework is open source and includes a simple command-line interface for running experiments, managing configurations and caching, and generating reports. The framework can be accessed at https://github.com/SAP/mirrorbench.

Method: Constructed post-crash lane change dataset from drone videos, conducted empirical analysis comparing to mandatory and discretionary lane changes, and developed a novel trajectory prediction framework with graph-based attention module that explicitly models yielding behavior as an auxiliary interaction-aware task, guiding both conditional variational autoencoder and Transformer-based decoder.

Result: Post-crash lane changes exhibit longer durations, lower insertion speeds, and higher crash risks. 79.4% involve non-yielding behavior from new followers (vs 21.7% DLCs, 28.6% MLCs). The proposed model outperforms baselines by >10% in ADE and FDE across prediction horizons, provides more reliable crash risk analysis by reducing false crash rates and improving conflict prediction accuracy, and shows transferability across different sites.

Conclusion: Post-crash lane changes have distinct characteristics with high non-yielding rates, requiring specialized modeling. The proposed interaction-aware trajectory prediction framework effectively captures yielding behavior dynamics, significantly improving prediction accuracy and crash risk assessment for post-crash scenarios.

Abstract: When a traffic crash occurs, following vehicles need to change lanes to bypass the obstruction. We define these maneuvers as post crash lane changes. In such scenarios, vehicles in the target lane may refuse to yield even after the lane change has already begun, increasing the complexity and crash risk of post crash LCs. However, the behavioral characteristics and motion patterns of post crash LCs remain unknown. To address this gap, we construct a post crash LC dataset by extracting vehicle trajectories from drone videos captured after crashes. Our empirical analysis reveals that, compared to mandatory LCs (MLCs) and discretionary LCs (DLCs), post crash LCs exhibit longer durations, lower insertion speeds, and higher crash risks. Notably, 79.4% of post crash LCs involve at least one instance of non yielding behavior from the new follower, compared to 21.7% for DLCs and 28.6% for MLCs. Building on these findings, we develop a novel trajectory prediction framework for post crash LCs. At its core is a graph based attention module that explicitly models yielding behavior as an auxiliary interaction aware task. This module is designed to guide both a conditional variational autoencoder and a Transformer based decoder to predict the lane changer’s trajectory. By incorporating the interaction aware module, our model outperforms existing baselines in trajectory prediction performance by more than 10% in both average displacement error and final displacement error across different prediction horizons. Moreover, our model provides more reliable crash risk analysis by reducing false crash rates and improving conflict prediction accuracy. Finally, we validate the model’s transferability using additional post crash LC datasets collected from different sites.

Method: Alternating memory paradigm with two phases: active phases use lightweight retrieval for real-time dialog; inactive phases perform computationally intensive memory extraction, consolidation, and maintenance across full conversation sessions.

Result: System using Qwen2.5-7B-Instruct quantized int4 outperforms equivalent raw LLM without memory across most metrics, and performs comparably to GPT-3.5 with 16k context window on the proposed AI Companion benchmark.

Conclusion: The alternating memory paradigm enables effective AI companions on resource-constrained edge devices by minimizing latency while maintaining long-term personalization, demonstrated through competitive performance with much larger models.

Abstract: Computational resource constraints on edge devices make it difficult to develop a fully embedded AI companion system with a satisfactory user experience. AI companion and memory systems detailed in existing literature cannot be directly used in such an environment due to lack of compute resources and latency concerns. In this paper, we propose a memory paradigm that alternates between active and inactive phases: during phases of user activity, the system performs low-latency, real-time dialog using lightweight retrieval over existing memories and context; whereas during phases of user inactivity, it conducts more computationally intensive extraction, consolidation, and maintenance of memories across full conversation sessions. This design minimizes latency while maintaining long-term personalization under the tight constraints of embedded hardware. We also introduce an AI Companion benchmark designed to holistically evaluate the AI Companion across both its conversational quality and memory capabilities. In our experiments, we found that our system (using a very weak model: Qwen2.5-7B-Instruct quantized int4) outperforms the equivalent raw LLM without memory across most metrics, and performs comparably to GPT-3.5 with 16k context window.

Method: Hierarchical multi-agent architecture with a central Resolution Supervisor for strategic task decomposition and specialized worker agents for tactical execution, orchestrated using LangGraph for complex cyclical workflows. Evaluation uses a benchmark dataset of 30 disruption scenarios from 6,000+ real-world reviews and LLM-as-a-Judge protocol with bias mitigation.

Result: The paper presents a novel agentic framework validated on real-world disruption scenarios, though specific performance metrics are not provided in the abstract. The system demonstrates capability to handle complex delivery disruptions through hierarchical agent coordination.

Conclusion: Project Synapse provides an effective agentic framework for autonomous resolution of last-mile delivery disruptions, combining strategic planning with tactical execution through hierarchical multi-agent coordination and validated on real-world scenarios.

Abstract: This paper introduces Project Synapse, a novel agentic framework designed for the autonomous resolution of last-mile delivery disruptions. Synapse employs a hierarchical multi-agent architecture in which a central Resolution Supervisor agent performs strategic task decomposition and delegates subtasks to specialized worker agents responsible for tactical execution. The system is orchestrated using LangGraph to manage complex and cyclical workflows. To validate the framework, a benchmark dataset of 30 complex disruption scenarios was curated from a qualitative analysis of over 6,000 real-world user reviews. System performance is evaluated using an LLM-as-a-Judge protocol with explicit bias mitigation.

Method: Model-based hierarchical MARL with two collaborative agents decomposing exponential action space, Dyna-Q-inspired framework integrating direct RL with model-based planning, LLM-based semantic feature extraction for OpenMP programs, and accurate environment model using regression techniques for thermal/performance prediction.

Result: Achieves 7.09x better energy efficiency and 4.0x better makespan than Linux ondemand governor, with 358ms latency for subsequent decisions (3.5-8.0s for first decisions including LLM feature extraction), 20x faster convergence than model-free methods, and 8,300x faster first-decision latency than table-based profiling.

Conclusion: The proposed framework enables practical deployment in dynamic embedded systems through zero-shot deployment for new workloads, fast decision-making, and superior energy-performance trade-offs compared to existing approaches.

Abstract: Dynamic voltage and frequency scaling (DVFS) and task-to-core allocation are critical for thermal management and balancing energy and performance in embedded systems. Existing approaches either rely on utilization-based heuristics that overlook stall times, or require extensive offline profiling for table generation, preventing runtime adaptation. We propose a model-based hierarchical multi-agent reinforcement learning (MARL) framework for thermal- and energy-aware scheduling on multi-core platforms. Two collaborative agents decompose the exponential action space, achieving 358ms latency for subsequent decisions. First decisions require 3.5 to 8.0s including one-time LLM feature extraction. An accurate environment model leverages regression techniques to predict thermal dynamics and performance states. When combined with LLM-extracted semantic features, the environment model enables zero-shot deployment for new workloads on trained platforms by generating synthetic training data without requiring workload-specific profiling samples. We introduce LLM-based semantic feature extraction that characterizes OpenMP programs through 13 code-level features without execution. The Dyna-Q-inspired framework integrates direct reinforcement learning with model-based planning, achieving 20x faster convergence than model-free methods. Experiments on BOTS and PolybenchC benchmarks across NVIDIA Jetson TX2, Jetson Orin NX, RubikPi, and Intel Core i7 demonstrate 7.09x better energy efficiency and 4.0x better makespan than Linux ondemand governor. First-decision latency is 8,300x faster than table-based profiling, enabling practical deployment in dynamic embedded systems.

Method: Introduces EvoEnv, a dynamic evaluation environment that simulates a “trainee” agent exploring novel settings, evaluating agents along three dimensions: context-aware scheduling for streaming tasks, prudent information acquisition via active exploration, and continuous evolution through distillation of generalized strategies from rule-based tasks.

Result: Experiments show cutting-edge agents have significant deficiencies in dynamic environments, particularly in active exploration and continual learning capabilities.

Conclusion: EvoEnv establishes a framework for assessing agent reliability, shifting evaluation from static tests to realistic, production-oriented scenarios, with code publicly available.

Abstract: The rapid evolution of Multi-modal Large Language Models (MLLMs) has advanced workflow automation; however, existing research mainly targets performance upper bounds in static environments, overlooking robustness for stochastic real-world deployment. We identify three key challenges: dynamic task scheduling, active exploration under uncertainty, and continuous learning from experience. To bridge this gap, we introduce \method{}, a dynamic evaluation environment that simulates a “trainee” agent continuously exploring a novel setting. Unlike traditional benchmarks, \method{} evaluates agents along three dimensions: (1) context-aware scheduling for streaming tasks with varying priorities; (2) prudent information acquisition to reduce hallucination via active exploration; and (3) continuous evolution by distilling generalized strategies from rule-based, dynamically generated tasks. Experiments show that cutting-edge agents have significant deficiencies in dynamic environments, especially in active exploration and continual learning. Our work establishes a framework for assessing agent reliability, shifting evaluation from static tests to realistic, production-oriented scenarios. Our codes are available at https://github.com/KnowledgeXLab/EvoEnv

Method: HS2C uses graph homophily to compress TAGs: (1) Structurally, performs global hierarchical partition guided by Structural Entropy minimization to identify cohesive communities and remove noise; (2) Semantically, enables LLMs to perform differentiated semantic aggregation based on community types, compressing redundant contexts into community-level consensus.

Result: Extensive experiments on 10 node-level benchmarks across various LLMs show HS2C simultaneously enhances compression rate and downstream inference accuracy. Extensions to 7 graph-level benchmarks demonstrate task generalizability and scalability.

Conclusion: HS2C effectively exploits graph homophily to compress TAGs for LLMs, improving reasoning performance while reducing context requirements, demonstrating superiority, scalability, and task generalizability across diverse benchmarks.

Abstract: Large language models (LLMs) have demonstrated promising capabilities in Text-Attributed Graph (TAG) understanding. Recent studies typically focus on verbalizing the graph structures via handcrafted prompts, feeding the target node and its neighborhood context into LLMs. However, constrained by the context window, existing methods mainly resort to random sampling, often implemented via dropping node/edge randomly, which inevitably introduces noise and cause reasoning instability. We argue that graphs inherently contain rich structural and semantic information, and that their effective exploitation can unlock potential gains in LLMs reasoning performance. To this end, we propose Homophily-aware Structural and Semantic Compression for LLMs (HS2C), a framework centered on exploiting graph homophily. Structurally, guided by the principle of Structural Entropy minimization, we perform a global hierarchical partition that decodes the graph’s essential topology. This partition identifies naturally cohesive, homophilic communities, while discarding stochastic connectivity noise. Semantically, we deliver the detected structural homophily to the LLM, empowering it to perform differentiated semantic aggregation based on predefined community type. This process compresses redundant background contexts into concise community-level consensus, selectively preserving semantically homophilic information aligned with the target nodes. Extensive experiments on 10 node-level benchmarks across LLMs of varying sizes and families demonstrate that, by feeding LLMs with structurally and semantically compressed inputs, HS2C simultaneously enhances the compression rate and downstream inference accuracy, validating its superiority and scalability. Extensions to 7 diverse graph-level benchmarks further consolidate HS2C’s task generalizability.

Method: Used a world champion AI to engage in self-play competitions and conducted statistical data analysis to identify game balance issues, specifically examining first-mover advantage and subgoal scoring settings.

Result: Revealed significant first-mover advantage and issues with subgoal scoring settings. Found that traditional position rotation over multiple rounds (used offline) doesn’t work well for online single-round play.

Conclusion: Proposed rule adaptations including compensatory points for first-mover advantage and refined subgoal scoring for different tile patterns. Implemented revised Mahjong online for players. Demonstrated using AI data to evaluate and adapt traditional games for online environments.

Abstract: As one of the worldwide spread traditional game, Official International Mahjong can be played and promoted online through remote devices instead of requiring face-to-face interaction. However, online players have fragmented playtime and unfixed combination of opponents in contrary to offline players who have fixed opponents for multiple rounds of play. Therefore, the rules designed for offline players need to be modified to ensure the fairness of online single-round play. Specifically, We employ a world champion AI to engage in self-play competitions and conduct statistical data analysis. Our study reveals the first-mover advantage and issues in the subgoal scoring settings. Based on our findings, we propose rule adaptations to make the game more suitable for the online environment, such as introducing compensatory points for the first-mover advantage and refining the scores of subgoals for different tile patterns. Compared with the traditional method of rotating positions over multiple rounds to balance first-mover advantage, our compensatory points mechanism in each round is more convenient for online players. Furthermore, we implement the revised Mahjong game online, which is open for online players. This work is an initial attempt to use data from AI systems to evaluate Official Internatinoal Mahjong’s game balance and develop a revised version of the traditional game better adapted for online players.

Method: Proposes SANC(E3) framework with five core axioms: finite capacity, association from co-occurrence, similarity-based competition, confidence-based stabilization, and reconstruction-compression-update trade-off. Uses pseudo-memory-mapped I/O mechanism where internally replayed Gestalts are processed like external input.

Result: Derives twelve propositions showing that category formation, hierarchical organization, unsupervised learning, and high-level cognitive activities can all be understood as instances of Gestalt completion under E3 minimization. Unifies perception, imagination, prediction, planning, and action within single representational process.

Conclusion: Representational units should not be predefined but emerge through competitive selection and compression under finite activation capacity. The SANC(E3) framework provides principled foundation for understanding how intelligence organizes experience into stable internal structures that enable prediction and action.

Abstract: General intelligence must reorganize experience into internal structures that enable prediction and action under finite resources. Existing systems implicitly presuppose fixed primitive units – tokens, subwords, pixels, or predefined sensor channels – thereby bypassing the question of how representational units themselves emerge and stabilize. This paper proposes SANC(E3), an axiomatic framework in which representational units are not given a priori but instead arise as stable outcomes of competitive selection, reconstruction, and compression under finite activation capacity, governed by the explicit minimization of an energy functional E3. SANC(E3) draws a principled distinction between system tokens – structural anchors such as {here, now, I} and sensory sources – and tokens that emerge through self-organization during co-occurring events. Five core axioms formalize finite capacity, association from co-occurrence, similarity-based competition, confidence-based stabilization, and the reconstruction-compression-update trade-off. A key feature is a pseudo-memory-mapped I/O mechanism, through which internally replayed Gestalts are processed via the same axiomatic pathway as external sensory input. As a result, perception, imagination, prediction, planning, and action are unified within a single representational and energetic process. From the axioms, twelve propositions are derived, showing that category formation, hierarchical organization, unsupervised learning, and high-level cognitive activities can all be understood as instances of Gestalt completion under E3 minimization.

Method: Introduces MPCI-Bench, a Multimodal Pairwise Contextual Integrity benchmark with paired positive/negative instances from the same visual source across three tiers: normative Seed judgments, context-rich Story reasoning, and executable agent action Traces. Uses a Tri-Principle Iterative Refinement pipeline to ensure data quality.

Result: Evaluations of state-of-the-art multimodal models reveal systematic failures to balance privacy and utility, and a pronounced modality leakage gap where sensitive visual information is leaked more frequently than textual information.

Conclusion: MPCI-Bench addresses critical gaps in evaluating agent privacy behavior and will be open-sourced to facilitate future research on agentic Contextual Integrity, highlighting the need for better privacy-utility balancing in multimodal AI systems.

Abstract: As language-model agents evolve from passive chatbots into proactive assistants that handle personal data, evaluating their adherence to social norms becomes increasingly critical, often through the lens of Contextual Integrity (CI). However, existing CI benchmarks are largely text-centric and primarily emphasize negative refusal scenarios, overlooking multimodal privacy risks and the fundamental trade-off between privacy and utility. In this paper, we introduce MPCI-Bench, the first Multimodal Pairwise Contextual Integrity benchmark for evaluating privacy behavior in agentic settings. MPCI-Bench consists of paired positive and negative instances derived from the same visual source and instantiated across three tiers: normative Seed judgments, context-rich Story reasoning, and executable agent action Traces. Data quality is ensured through a Tri-Principle Iterative Refinement pipeline. Evaluations of state-of-the-art multimodal models reveal systematic failures to balance privacy and utility and a pronounced modality leakage gap, where sensitive visual information is leaked more frequently than textual information. We will open-source MPCI-Bench to facilitate future research on agentic CI.

Method: Leveraging large language models (LLMs) to synthesize reward functions directly from natural language descriptions (EUREKA) and adapt reward formulations online with minimal human intervention (CARD). The approach conceptualizes transition along semantic reward specification, dynamic reward adaptation, and improved alignment with human intent.

Result: The emerging paradigm of Reinforcement Learning from Verifiable Rewards (RLVR) provides empirical evidence that language-mediated supervision can serve as a viable alternative to traditional reward engineering. LLMs enable coordination through shared semantic representations rather than explicitly engineered numerical signals.

Conclusion: The paper outlines a research direction where coordination arises from shared semantic representations rather than explicitly engineered numerical signals, while noting open challenges related to computational overhead, robustness to hallucination, and scalability to large multi-agent systems.

Abstract: Reward engineering, the manual specification of reward functions to induce desired agent behavior, remains a fundamental challenge in multi-agent reinforcement learning. This difficulty is amplified by credit assignment ambiguity, environmental non-stationarity, and the combinatorial growth of interaction complexity. We argue that recent advances in large language models (LLMs) point toward a shift from hand-crafted numerical rewards to language-based objective specifications. Prior work has shown that LLMs can synthesize reward functions directly from natural language descriptions (e.g., EUREKA) and adapt reward formulations online with minimal human intervention (e.g., CARD). In parallel, the emerging paradigm of Reinforcement Learning from Verifiable Rewards (RLVR) provides empirical evidence that language-mediated supervision can serve as a viable alternative to traditional reward engineering. We conceptualize this transition along three dimensions: semantic reward specification, dynamic reward adaptation, and improved alignment with human intent, while noting open challenges related to computational overhead, robustness to hallucination, and scalability to large multi-agent systems. We conclude by outlining a research direction in which coordination arises from shared semantic representations rather than explicitly engineered numerical signals.

Method: Propose a DRL agent guided by an LLM, where the LLM acts as a high-level coordinator generating textual guidance that shapes the DRL agent’s reward during training.

Result: LAM-DRL outperforms traditional DRL by 40% in nominal weather scenarios and 64% in extreme weather scenarios compared to heuristics in terms of throughput, fairness, and outage probability.

Conclusion: LLM-guided DRL approach effectively combines the strengths of large language models and reinforcement learning for improved performance in NTN applications across varying environmental conditions.

Abstract: Large AI Model (LAM) have been proposed to applications of Non-Terrestrial Networks (NTN), that offer better performance with its great generalization and reduced task specific trainings. In this paper, we propose a Deep Reinforcement Learning (DRL) agent that is guided by a Large Language Model (LLM). The LLM operates as a high level coordinator that generates textual guidance that shape the reward of the DRL agent during training. The results show that the LAM-DRL outperforms the traditional DRL by 40% in nominal weather scenarios and 64% in extreme weather scenarios compared to heuristics in terms of throughput, fairness, and outage probability.

Method: Created T3 benchmark with 454 expert-curated vignettes, decomposing performance into Utility (sensitivity), Safety (specificity), and Wise Refusal. Applied to frontier models and validated with process-verified protocol (RCA).

Result: Identified two pathologies: 1) “Skepticism Trap” at L1 where safety-tuned models reject valid links (Claude Haiku rejects 60%), and 2) “Scaling Paradox” at L3 where larger models underperform (GPT-5.2 underperforms GPT-4-Turbo by 55 points on ambiguous counterfactuals).

Conclusion: T3 successfully captures LLM causal judgment pathologies and validates that structured verification (RCA) can restore decisive causal judgment, providing a diagnostic tool for model evaluation.

Abstract: We introduce T3 (Testing Trustworthy Thinking), a diagnostic benchmark designed to rigorously evaluate LLM causal judgment across Pearl’s Ladder of Causality. Comprising 454 expert-curated vignettes, T3 prioritizes high-resolution failure analysis, decomposing performance into Utility (sensitivity), Safety (specificity), and Wise Refusal on underdetermined cases. By applying T3 to frontier models, we diagnose two distinct pathologies: a “Skepticism Trap” at L1 (where safety-tuned models like Claude Haiku reject 60% of valid links) and a non-monotonic Scaling Paradox at L3. In the latter, the larger GPT-5.2 underperforms GPT-4-Turbo by 55 points on ambiguous counterfactuals, driven by a collapse into paralysis (excessive hedging) rather than hallucination. Finally, we use the benchmark to validate a process-verified protocol (RCA), showing that T3 successfully captures the restoration of decisive causal judgment under structured verification.

Method: Uses VGG-16 convolutional neural network with transfer learning and image data augmentation. Trained on a widely used image dataset using Python and Keras. Validated on NUS dataset (10 gesture classes) and tested on custom dataset created using Google’s API for gesture capture.

Result: The system achieved approximately 98% accuracy by combining transfer learning with image data augmentation techniques.

Conclusion: The proposed VGG-16 based hand gesture recognition system with transfer learning and data augmentation is highly effective, demonstrating strong performance for sign language interpretation and human-computer interaction applications.

Abstract: Hand gesture recognition is an important aspect of human-computer interaction. It forms the basis of sign language for the visually impaired people. This work proposes a novel hand gesture recognizing system for the differently-abled persons. The model uses a convolutional neural network, known as VGG-16 net, for building a trained model on a widely used image dataset by employing Python and Keras libraries. Furthermore, the result is validated by the NUS dataset, consisting of 10 classes of hand gestures, fed to the model as the validation set. Afterwards, a testing dataset of 10 classes is built by employing Google’s open source Application Programming Interface (API) that captures different gestures of human hand and the efficacy is then measured by carrying out experiments. The experimental results show that by combining a transfer learning mechanism together with the image data augmentation, the VGG-16 net produced around 98% accuracy.

Method: Formalizes Sparse Agentic Control (SAC) with block-sparse policy representations, uses ell_{1,2}-regularized policy learning through convex surrogates, establishes theoretical results via Policy-RSC conditions, primal-dual witness arguments, and compressed-sensing-style analysis.

Result: Proves: (1) estimation/value suboptimality scales as k(log M/T)^{1/2}, (2) exact tool-support recovery when T > k log M under incoherence conditions, (3) dense policies require Ω(M) samples explaining prompt-only instability, (4) partial observability adds only O(ε_b) degradation while preserving log M dependence.

Conclusion: Sparse Agentic Control provides a theoretical foundation for efficient tool-augmented LLM learning, explaining why sparse policies succeed where dense ones fail, with extensions covering online, robust, and interaction-aware variants.

Abstract: Tool-augmented LLM systems expose a control regime that learning theory has largely ignored: sequential decision-making with a massive discrete action universe (tools, APIs, documents) in which only a small, unknown subset is relevant for any fixed task distribution. We formalize this setting as Sparse Agentic Control (SAC), where policies admit block-sparse representations over M » 1 actions and rewards depend on sparse main effects and (optionally) sparse synergies. We study ell_{1,2}-regularized policy learning through a convex surrogate and establish sharp, compressed-sensing-style results: (i) estimation and value suboptimality scale as k (log M / T)^{1/2} under a Policy-RSC condition; (ii) exact tool-support recovery holds via primal-dual witness arguments when T > k log M under incoherence and beta-min; and (iii) any dense policy class requires Omega(M) samples, explaining the instability of prompt-only controllers. We further show that under partial observability, LLMs matter only through a belief/representation error epsilon_b, yielding an additive O(epsilon_b) degradation while preserving logarithmic dependence on M. Extensions cover tuning-free, online, robust, group-sparse, and interaction-aware SAC.

Method: Proposes ToolACE-MCP pipeline that leverages dependency-rich candidate graphs to synthesize multi-turn trajectories, training history-aware routers with dynamic context understanding to create plug-and-play Light Routing Agents.

Result: Superior performance on real-world benchmarks MCP-Universe and MCP-Mark; exhibits critical properties for future Agent Web including generalization to multi-agent collaboration with minimal adaptation, exceptional robustness against noise, and effective scaling to massive candidate spaces.

Conclusion: ToolACE-MCP provides a strong empirical foundation for universal orchestration in open-ended ecosystems, addressing scalability and generality challenges in the evolving agent web.

Abstract: With the rise of the Agent Web and Model Context Protocol (MCP), the agent ecosystem is evolving into an open collaborative network, exponentially increasing accessible tools. However, current architectures face severe scalability and generality bottlenecks. To address this, we propose ToolACE-MCP, a pipeline for training history-aware routers to empower precise navigation in large-scale ecosystems. By leveraging a dependency-rich candidate Graph to synthesize multi-turn trajectories, we effectively train routers with dynamic context understanding to create the plug-and-play Light Routing Agent. Experiments on the real-world benchmarks MCP-Universe and MCP-Mark demonstrate superior performance. Notably, ToolACE-MCP exhibits critical properties for the future Agent Web: it not only generalizes to multi-agent collaboration with minimal adaptation but also maintains exceptional robustness against noise and scales effectively to massive candidate spaces. These findings provide a strong empirical foundation for universal orchestration in open-ended ecosystems.

Method: Formulates action discovery as a block-sparse recovery problem using a contextual linear reward model with structured sparsity. Analyzes a greedy algorithm inspired by Orthogonal Matching Pursuit (OMP) under assumptions of incoherence, signal strength, and action coverage.

Result: Proves the greedy procedure exactly recovers relevant action sets with high probability using samples scaling polynomially in sparsity/latent dimension and logarithmically in total actions. Provides estimation error guarantees and shows resulting decision rules are near-optimal for new latent states.

Conclusion: Identifies sparse action discovery as fundamental for large-action decision-making, provides theoretical foundation for action pruning in agentic systems, and establishes information-theoretic lower bounds showing sparsity and coverage are necessary for tractability.

Abstract: Modern agentic systems operate in environments with extremely large action spaces, such as tool-augmented language models with thousands of available APIs or retrieval operations. Despite this scale, empirical evidence suggests that only a small subset of actions meaningfully influences performance in a given deployment. Motivated by this observation, we study a contextual linear reward model in which action relevance is governed by a structured sparsity assumption: only a small number of actions have nonzero effects across latent states. We formulate action discovery as a block-sparse recovery problem and analyze a greedy algorithm inspired by Orthogonal Matching Pursuit. Under standard assumptions on incoherence, signal strength, and action coverage, we prove that the greedy procedure exactly recovers the relevant action set with high probability, using a number of samples that scales polynomially in the sparsity level and latent dimension, and only logarithmically in the total number of actions. We further provide estimation error guarantees for refitted parameters and show that the resulting decision rule is near-optimal for new latent states. Complementing these results, we establish information-theoretic lower bounds demonstrating that sparsity and sufficient coverage are necessary for tractability. Together, our results identify sparse action discovery as a fundamental principle underlying large-action decision-making and provide a theoretical foundation for action pruning in agentic systems.

Method: OpenMic uses an end-to-end multi-agent system built on AutoGen with specialized agents in a multi-round iterative loop-planning process. It employs retrieval-augmented generation (RAG) for material grounding and idea expansion, and fine-tunes a dedicated JokeWriter to internalize stand-up-specific structures like setup-punchline patterns and long-range callbacks.

Result: The system transforms user-provided life topics into 3-5 minute Chinese stand-up performances and produces narrated comedy videos, jointly optimizing humor, timing, and performability.

Conclusion: OpenMic addresses the complex challenges of Chinese stand-up comedy generation through a multi-agent approach that combines specialized agents, RAG, and fine-tuned joke writing to overcome dataset limitations and produce culturally appropriate, performable comedy.

Abstract: Chinese stand-up comedy generation goes beyond plain text generation, requiring culturally grounded humor, precise timing, stage-performance cues, and implicit multi-step reasoning. Moreover, commonly used Chinese humor datasets are often better suited for humor understanding and evaluation than for long-form stand-up generation, making direct supervision misaligned with the target task. To address these challenges, we present OpenMic, an end-to-end multi-agent system built on AutoGen that transforms a user-provided life topic into a 3-5 minute Chinese stand-up performance and further produces a narrated comedy video. OpenMic orchestrates multiple specialized agents in a multi-round iterative loop-planning to jointly optimize humor, timing, and performability. To mitigate the dataset-task mismatch, we augment generation with retrieval-augmented generation (RAG) for material grounding and idea expansion, and we fine-tune a dedicated JokeWriter to better internalize stand-up-specific setup-punchline structures and long-range callbacks.

Method: AtomMem reframes memory management as a dynamic decision-making problem by deconstructing high-level memory processes into fundamental atomic CRUD (Create, Read, Update, Delete) operations. It transforms memory workflow into a learnable decision process using supervised fine-tuning combined with reinforcement learning to learn an autonomous, task-aligned policy.

Result: Experimental results across 3 long-context benchmarks show that the trained AtomMem-8B consistently outperforms prior static-workflow memory methods. Analysis reveals that the learning-based formulation enables agents to discover structured, task-aligned memory management strategies.

Conclusion: AtomMem demonstrates that learning-based memory management through atomic CRUD operations enables more flexible and effective memory behaviors tailored to specific task demands, outperforming predefined static workflows and highlighting the advantage of learned strategies over hand-crafted routines.

Abstract: Equipping agents with memory is essential for solving real-world long-horizon problems. However, most existing agent memory mechanisms rely on static and hand-crafted workflows. This limits the performance and generalization ability of these memory designs, which highlights the need for a more flexible, learning-based memory framework. In this paper, we propose AtomMem, which reframes memory management as a dynamic decision-making problem. We deconstruct high-level memory processes into fundamental atomic CRUD (Create, Read, Update, Delete) operations, transforming the memory workflow into a learnable decision process. By combining supervised fine-tuning with reinforcement learning, AtomMem learns an autonomous, task-aligned policy to orchestrate memory behaviors tailored to specific task demands. Experimental results across 3 long-context benchmarks demonstrate that the trained AtomMem-8B consistently outperforms prior static-workflow memory methods. Further analysis of training dynamics shows that our learning-based formulation enables the agent to discover structured, task-aligned memory management strategies, highlighting a key advantage over predefined routines.

Method: The authors formalize the problem as “semantic laundering” - a pattern where propositions gain epistemic status without proper warrant by crossing architecturally trusted interfaces. They analyze this as an architectural realization of the Gettier problem, prove the Theorem of Inevitable Self-Licensing showing circular epistemic justification cannot be eliminated under standard assumptions, and introduce the Warrant Erosion Principle as the fundamental explanation.

Result: The central result is the Theorem of Inevitable Self-Licensing: under standard architectural assumptions for LLM-based agents, circular epistemic justification cannot be eliminated. The Warrant Erosion Principle explains why scaling, model improvement, and LLM-as-judge schemes are structurally incapable of solving this type-level problem.

Conclusion: Semantic laundering represents a fundamental architectural limitation of LLM-based agent systems - not an accidental flaw but a systematic, reproducible pattern that constitutes an architectural Gettier problem. The issue exists at the type level and cannot be eliminated through conventional approaches like scaling or model improvements, requiring architectural redesign.

Abstract: LLM-based agent architectures systematically conflate information transport mechanisms with epistemic justification mechanisms. We formalize this class of architectural failures as semantic laundering: a pattern where propositions with absent or weak warrant are accepted by the system as admissible by crossing architecturally trusted interfaces. We show that semantic laundering constitutes an architectural realization of the Gettier problem: propositions acquire high epistemic status without a connection between their justification and what makes them true. Unlike classical Gettier cases, this effect is not accidental; it is architecturally determined and systematically reproducible. The central result is the Theorem of Inevitable Self-Licensing: under standard architectural assumptions, circular epistemic justification cannot be eliminated. We introduce the Warrant Erosion Principle as the fundamental explanation for this effect and show that scaling, model improvement, and LLM-as-judge schemes are structurally incapable of eliminating a problem that exists at the type level.

Method: Working group approach focusing on curriculum design, professional development programs, practical classroom applications, and policy guideline development.

Result: Development of effective strategies for enhancing teacher AI literacy and agency, though specific outcomes not detailed in abstract.

Conclusion: Empowering educators with AI knowledge and skills is crucial for successful AI integration in education and fostering student AI understanding.

Abstract: TWG 5 focused on developing and implementing effective strategies for enhancing AI literacy and agency of teachers, equipping them with the knowledge and skills necessary to integrate AI into their teaching practices. Explorations covered curriculum design, professional development programs, practical classroom applications, and policy guidelines aiming to empower educators to confidently utilize AI tools and foster a deeper understanding of AI concepts among students.

Method: Built a conceptual neighborhood graph (CNGRLO) that relates Rotation movement to Location change and Orientation change of cube features, and produced composition tables for calculating inferences about rotations.

Result: Created a qualitative reasoning model that can solve the Cube Comparison Test by systematically reasoning about object rotations through formal graph-based relationships and inference tables.

Conclusion: The QOR model provides a formal qualitative framework for reasoning about object rotations, demonstrating that spatial reasoning tasks like the Cube Comparison Test can be addressed through systematic graph-based approaches and composition tables.

Abstract: This paper presents a Qualitative model for Reasoning about Object Rotations (QOR) which is applied to solve the Cube Comparison Test (CCT) by Ekstrom et al. (1976). A conceptual neighborhood graph relating the Rotation movement to the Location change and the Orientation change (CNGRLO) of the features on the cube sides has been built and it produces composition tables to calculate inferences for reasoning about rotations.

Method: Introduced Gated-LPI (Log-Probability Increase), a neuron-level attribution metric to decompose log-probability increase across neurons. Conducted time-resolved comparison tracking checkpoints over 1.2M training steps (~5.0T tokens) for MoE and 600K steps (~2.5T tokens) for dense models.

Result: Three key patterns: 1) Low-entropy backbone - top ~1% MoE neurons capture >45% of positive updates (absent in dense), 2) Early consolidation - MoE stabilizes importance profile within <100K steps while dense remains volatile, 3) Functional robustness - masking top MoE attention heads reduces relational HIT@10 by <10% vs >50% for dense, showing distributed knowledge storage.

Conclusion: Sparsity in MoE architectures fosters an intrinsically stable and distributed computational backbone from early training stages, bridging the gap between sparse architectures and training-time interpretability.

Abstract: Mixture-of-Experts (MoE) architectures decouple model capacity from per-token computation, enabling scaling beyond the computational limits imposed by dense scaling laws. Yet how MoE architectures shape knowledge acquisition during pre-training, and how this process differs from dense architectures, remains unknown. To address this issue, we introduce Gated-LPI (Log-Probability Increase), a neuron-level attribution metric that decomposes log-probability increase across neurons. We present a time-resolved comparison of knowledge acquisition dynamics in MoE and dense architectures, tracking checkpoints over 1.2M training steps (~ 5.0T tokens) and 600K training steps (~ 2.5T tokens), respectively. Our experiments uncover three patterns: (1) Low-entropy backbone. The top approximately 1% of MoE neurons capture over 45% of positive updates, forming a high-utility core, which is absent in the dense baseline. (2) Early consolidation. The MoE model locks into a stable importance profile within < 100K steps, whereas the dense model remains volatile throughout training. (3) Functional robustness. Masking the ten most important MoE attention heads reduces relational HIT@10 by < 10%, compared with > 50% for the dense model, showing that sparsity fosters distributed – rather than brittle – knowledge storage. These patterns collectively demonstrate that sparsity fosters an intrinsically stable and distributed computational backbone from early in training, helping bridge the gap between sparse architectures and training-time interpretability.

Method: Proposes a generative perspective focusing on structural and contextual conditions for creative behaviors. Introduces conceptual decomposition of creativity into four components: pattern-based generation, induced world models, contextual grounding, and arbitrarity. Examines how these manifest in multimodal generative systems.

Result: Provides a technical framework for studying creativity as an emergent phenomenon in AI systems, grounded in the interaction between generative dynamics and domain-specific representations.

Conclusion: Creativity should be studied as an emergent property arising from generative models in bounded environments, not just as a post hoc evaluative label. This generative perspective offers a more fundamental approach to understanding machine creativity.

Abstract: Creativity in artificial intelligence is most often addressed through evaluative frameworks that aim to measure novelty, diversity, or usefulness in generated outputs. While such approaches have provided valuable insights into the behavior of modern generative models, they largely treat creativity as a property to be assessed rather than as a phenomenon to be explicitly modeled. In parallel, recent advances in large-scale generative systems, particularly multimodal architectures, have demonstrated increasingly sophisticated forms of pattern recombination, raising questions about the nature and limits of machine creativity. This paper proposes a generative perspective on creativity in AI, framing it as an emergent property of domain-limited generative models embedded within bounded informational environments. Rather than introducing new evaluative criteria, we focus on the structural and contextual conditions under which creative behaviors arise. We introduce a conceptual decomposition of creativity into four interacting components-pattern-based generation, induced world models, contextual grounding, and arbitrarity, and examine how these components manifest in multimodal generative systems. By grounding creativity in the interaction between generative dynamics and domain-specific representations, this work aims to provide a technical framework for studying creativity as an emergent phenomenon in AI systems, rather than as a post hoc evaluative label.

Method: OSPO (Owen-Shapley Policy Optimization) redistributes sequence-level advantages based on tokens’ marginal contributions using potential-based reward shaping via Shapley-Owen attributions. It forms coalitions of semantically coherent units (phrases describing product attributes or sentences capturing preferences) to identify which response parts drive performance, without requiring parametric value models.

Result: Experiments on Amazon ESCI and H&M Fashion datasets show consistent gains over baselines, with notable test-time robustness to out-of-distribution retrievers unseen during training.

Conclusion: OSPO effectively addresses the credit assignment problem in RL for personalized recommendations by attributing rewards to semantically meaningful token segments, improving performance and robustness without additional computational overhead from value models.

Abstract: Large language models are increasingly trained via reinforcement learning for personalized recommendation tasks, but standard methods like GRPO rely on sparse, sequence-level rewards that create a credit assignment gap, obscuring which tokens drive success. This gap is especially problematic when models must infer latent user intent from under-specified language without ground truth labels, a reasoning pattern rarely seen during pretraining. We introduce Owen-Shapley Policy Optimization (OSPO), a framework that redistributes sequence-level advantages based on tokens’ marginal contributions to outcomes. Unlike value-model-based methods requiring additional computation, OSPO employs potential-based reward shaping via Shapley-Owen attributions to assign segment-level credit while preserving the optimal policy, learning directly from task feedback without parametric value models. By forming coalitions of semantically coherent units (phrases describing product attributes or sentences capturing preferences), OSPO identifies which response parts drive performance. Experiments on Amazon ESCI and H&M Fashion datasets show consistent gains over baselines, with notable test-time robustness to out-of-distribution retrievers unseen during training.

Method: WebTrap Park instantiates three major sources of security risk into 1,226 executable evaluation tasks and enables action-based assessment through direct observation of agent interactions with live web pages without requiring agent modification.

Result: Results reveal clear security differences across agent frameworks, highlighting the importance of agent architecture beyond the underlying model.

Conclusion: WebTrap Park provides a scalable foundation for reproducible Web Agent security evaluation and is publicly accessible as an automated platform for systematic security assessment.

Abstract: Web Agents are increasingly deployed to perform complex tasks in real web environments, yet their security evaluation remains fragmented and difficult to standardize. We present WebTrap Park, an automated platform for systematic security evaluation of Web Agents through direct observation of their concrete interactions with live web pages. WebTrap Park instantiates three major sources of security risk into 1,226 executable evaluation tasks and enables action based assessment without requiring agent modification. Our results reveal clear security differences across agent frameworks, highlighting the importance of agent architecture beyond the underlying model. WebTrap Park is publicly accessible at https://security.fudan.edu.cn/webagent and provides a scalable foundation for reproducible Web Agent security evaluation.

Method: 1) Construct high-quality dataset with instruction-code-reasoning chains and counterfactual negative samples; 2) Use DeepSeek-Coder-V2-Lite as teacher model with hybrid black-box/white-box distillation to generate chain-of-thought soft labels; 3) Combine weighted cross-entropy loss with hard labels; 4) Apply prompt tuning engineering optimized for UAV control scenarios.

Result: The distilled lightweight model maintains high code generation accuracy while achieving significant improvements in deployment and inference efficiency, demonstrating feasibility for precise and lightweight UAV intelligent control.

Conclusion: The integrated approach effectively transfers complex reasoning and code generation capabilities to smaller models, enabling efficient UAV multi-SDK control while balancing accuracy and resource constraints.

Abstract: With large language models demonstrating significant potential in code generation tasks, their application to onboard control of resource-constrained Unmanned Aerial Vehicles has emerged as an important research direction. However, a notable contradiction exists between the high resource consumption of large models and the real-time, lightweight requirements of UAV platforms. This paper proposes an integrated approach that combines knowledge distillation, chain-of-thought guidance, and supervised fine-tuning for UAV multi-SDK control tasks, aiming to efficiently transfer complex reasoning and code generation capabilities to smaller models. Firstly, a high-quality dataset covering various mainstream UAV SDKs is constructed, featuring instruction-code-reasoning chains, and incorporates counterfactual negative samples for data augmentation, guiding the model to learn the end-to-end logic from instruction parsing to code generation. Secondly, leveraging DeepSeek-Coder-V2-Lite quantized via QLoRA as the teacher model, and based on a hybrid black-box and white-box distillation strategy, high-quality chain-of-thought soft labels are generated. These are combined with a weighted cross-entropy loss using hard labels to transfer complex reasoning capabilities to the smaller student model. Finally, through prompt tuning engineering optimized for the UAV control scenario, the model performance on core tasks such as SDK type recognition and function call matching is enhanced. Experimental results indicate that the distilled lightweight model maintains high code generation accuracy while achieving significant improvements in deployment and inference efficiency, effectively demonstrating the feasibility and superiority of our approach in achieving precise and lightweight intelligent control for UAVs

Method: Develops automated Coarse-to-Fine Rubric Generation framework using principle-guided synthesis, multi-model aggregation, and difficulty evolution. Creates RubricHub dataset (~110k rubrics). Uses two-stage post-training: Rubric-based Rejection Sampling Fine-Tuning (RuFT) followed by Reinforcement Learning (RuRL).

Result: Post-trained Qwen3-14B achieves state-of-the-art results on HealthBench (69.3), surpassing proprietary frontier models like GPT-5. RubricHub enables significant performance gains in open-ended generation tasks.

Conclusion: The proposed rubric generation framework and RubricHub dataset effectively address limitations of existing evaluation methods for open-ended generation, enabling better RL training and achieving superior performance compared to even proprietary models.

Abstract: Reinforcement Learning with Verifiable Rewards (RLVR) has driven substantial progress in reasoning-intensive domains like mathematics. However, optimizing open-ended generation remains challenging due to the lack of ground truth. While rubric-based evaluation offers a structured proxy for verification, existing methods suffer from scalability bottlenecks and coarse criteria, resulting in a supervision ceiling effect. To address this, we propose an automated Coarse-to-Fine Rubric Generation framework. By synergizing principle-guided synthesis, multi-model aggregation, and difficulty evolution, our approach produces comprehensive and highly discriminative criteria capable of capturing the subtle nuances. Based on this framework, we introduce RubricHub, a large-scale ($\sim$110k) and multi-domain dataset. We validate its utility through a two-stage post-training pipeline comprising Rubric-based Rejection Sampling Fine-Tuning (RuFT) and Reinforcement Learning (RuRL). Experimental results demonstrate that RubricHub unlocks significant performance gains: our post-trained Qwen3-14B achieves state-of-the-art (SOTA) results on HealthBench (69.3), surpassing proprietary frontier models such as GPT-5. The code and data will be released soon.

Method: YaPO (Yet another Policy Optimization) is a reference-free method that learns sparse steering vectors in the latent space of a Sparse Autoencoder (SAE) by optimizing sparse codes, producing disentangled, interpretable, and efficient steering directions.

Result: YaPO converges faster, achieves stronger performance, and exhibits improved training stability compared to dense steering baselines. It generalizes to various alignment behaviors (hallucination, wealth-seeking, jailbreak, power-seeking) while preserving general knowledge with no measurable degradation on MMLU.

Conclusion: YaPO provides a general recipe for efficient, stable, and fine-grained alignment of LLMs with broad applications to controllability and domain adaptation, offering superior disentanglement and interpretability over dense steering methods.

Abstract: Steering Large Language Models (LLMs) through activation interventions has emerged as a lightweight alternative to fine-tuning for alignment and personalization. Recent work on Bi-directional Preference Optimization (BiPO) shows that dense steering vectors can be learned directly from preference data in a Direct Preference Optimization (DPO) fashion, enabling control over truthfulness, hallucinations, and safety behaviors. However, dense steering vectors often entangle multiple latent factors due to neuron multi-semanticity, limiting their effectiveness and stability in fine-grained settings such as cultural alignment, where closely related values and behaviors (e.g., among Middle Eastern cultures) must be distinguished. In this paper, we propose Yet another Policy Optimization (YaPO), a \textit{reference-free} method that learns \textit{sparse steering vectors} in the latent space of a Sparse Autoencoder (SAE). By optimizing sparse codes, YaPO produces disentangled, interpretable, and efficient steering directions. Empirically, we show that YaPO converges faster, achieves stronger performance, and exhibits improved training stability compared to dense steering baselines. Beyond cultural alignment, YaPO generalizes to a range of alignment-related behaviors, including hallucination, wealth-seeking, jailbreak, and power-seeking. Importantly, YaPO preserves general knowledge, with no measurable degradation on MMLU. Overall, our results show that YaPO provides a general recipe for efficient, stable, and fine-grained alignment of LLMs, with broad applications to controllability and domain adaptation. The associated code and data are publicly available\footnote{https://github.com/MBZUAI-Paris/YaPO}.

Method: Proposes Table Graph Reasoner (TABGR) with two key components: 1) Attributed Table Graph (ATG) representation that preserves row-column-cell structures while enabling graph-based reasoning, and 2) Question-Guided Personalized PageRank (QG-PPR) mechanism to rerank tabular data and mitigate the lost-in-the-middle issue.

Result: Extensive experiments on two commonly used benchmarks show that TABGR consistently outperforms state-of-the-art models by up to 9.7% in accuracy.

Conclusion: TABGR addresses critical issues in current table reasoning approaches by preserving table structures, enabling explainable reasoning paths, and mitigating information loss, achieving superior performance without requiring training.

Abstract: Table reasoning, a task to answer questions by reasoning over data presented in tables, is an important topic due to the prevalence of knowledge stored in tabular formats. Recent solutions use Large Language Models (LLMs), exploiting the semantic understanding and reasoning capabilities of LLMs. A common paradigm of such solutions linearizes tables to form plain texts that are served as input to LLMs. This paradigm has critical issues. It loses table structures, lacks explicit reasoning paths for result explainability, and is subject to the “lost-in-the-middle” issue. To address these issues, we propose Table Graph Reasoner (TABGR), a training-free model that represents tables as an Attributed Table Graph (ATG). The ATG explicitly preserves row-column-cell structures while enabling graph-based reasoning for explainability. We further propose a Question-Guided Personalized PageRank (QG-PPR) mechanism to rerank tabular data and mitigate the lost-in-the-middle issue. Extensive experiments on two commonly used benchmarks show that TABGR consistently outperforms state-of-the-art models by up to 9.7% in accuracy. Our code will be made publicly available upon publication.

Method: Multi-modal web app using XLM-R/mBERT for text (4,193 comments) and mBERT+EfficientNet/ResNET for memes (4,218 memes); SHAP/LIME for feature importance; evaluated via CUQ/UEQ questionnaires.

Result: Developed functional web application with transformer-based models for both text and meme misogyny detection in Hindi-English code-mixed content; integrated explainability features.

Conclusion: System serves as tool for researchers/content moderators to combat gender-based digital violence; promotes safe digital spaces through explainable AI for sensitive hate speech detection.

Abstract: Digital platforms have an ever-expanding user base, and act as a hub for communication, business, and connectivity. However, this has also allowed for the spread of hate speech and misogyny. Artificial intelligence models have emerged as an effective solution for countering online hate speech but are under explored for low resource and code-mixed languages and suffer from a lack of interpretability. Explainable Artificial Intelligence (XAI) can enhance transparency in the decisions of deep learning models, which is crucial for a sensitive domain such as hate speech detection. In this paper, we present a multi-modal and explainable web application for detecting misogyny in text and memes in code-mixed Hindi and English. The system leverages state-of-the-art transformer-based models that support multilingual and multimodal settings. For text-based misogyny identification, the system utilizes XLM-RoBERTa (XLM-R) and multilingual Bidirectional Encoder Representations from Transformers (mBERT) on a dataset of approximately 4,193 comments. For multimodal misogyny identification from memes, the system utilizes mBERT + EfficientNet, and mBERT + ResNET trained on a dataset of approximately 4,218 memes. It also provides feature importance scores using explainability techniques including Shapley Additive Values (SHAP) and Local Interpretable Model Agnostic Explanations (LIME). The application aims to serve as a tool for both researchers and content moderators, to promote further research in the field, combat gender based digital violence, and ensure a safe digital space. The system has been evaluated using human evaluators who provided their responses on Chatbot Usability Questionnaire (CUQ) and User Experience Questionnaire (UEQ) to determine overall usability.

Method: Proposes M3-Bench with a process-aware evaluation framework that synergistically analyzes three modules: Behavioral Trajectory Analysis (BTA), Reasoning Process Analysis (RPA), and Communication Content Analysis (CCA). Integrates Big Five personality model and Social Exchange Theory to create interpretable social behavior portraits.

Result: M3-Bench reliably distinguishes diverse social behavior competencies across models. Reveals that some models achieve reasonable behavioral outcomes while exhibiting pronounced inconsistencies in their reasoning and communication processes.

Conclusion: The proposed benchmark provides a comprehensive, process-aware evaluation framework for LLM agents’ social behaviors, going beyond simple task scores to characterize personality traits and capability profiles through multi-dimensional analysis.

Abstract: As the capabilities of large language model (LLM) agents continue to advance, their advanced social behaviors, such as cooperation, deception, and collusion, call for systematic evaluation. However, existing benchmarks often emphasize a single capability dimension or rely solely on behavioral outcomes, overlooking rich process information from agents’ decision reasoning and communicative interactions. To address this gap, we propose M3-Bench, a multi-stage benchmark for mixed-motive games, together with a process-aware evaluation framework that conducts synergistic analysis across three modules: BTA (Behavioral Trajectory Analysis), RPA (Reasoning Process Analysis), and CCA (Communication Content Analysis). Furthermore, we integrate the Big Five personality model and Social Exchange Theory to aggregate multi-dimensional evidence into interpretable social behavior portraits, thereby characterizing agents’ personality traits and capability profiles beyond simple task scores or outcome-based metrics. Experimental results show that M3-Bench can reliably distinguish diverse social behavior competencies across models, and it reveals that some models achieve seemingly reasonable behavioral outcomes while exhibiting pronounced inconsistencies in their reasoning and communication.

Method: Introduces SummPilot, an interaction-based customizable summarization system that leverages large language models for both automatic and interactive summarization, with interactive components including semantic graphs, entity clustering, and explainable evaluation.

Result: Demo and user studies demonstrate SummPilot’s adaptability and usefulness for customizable summarization.

Conclusion: SummPilot successfully combines automatic summarization efficiency with personalized, interactive summarization capabilities, providing a practical solution for customizable document summarization.

Abstract: This paper incorporates the efficiency of automatic summarization and addresses the challenge of generating personalized summaries tailored to individual users’ interests and requirements. To tackle this challenge, we introduce SummPilot, an interaction-based customizable summarization system. SummPilot leverages a large language model to facilitate both automatic and interactive summarization. Users can engage with the system to understand document content and personalize summaries through interactive components such as semantic graphs, entity clustering, and explainable evaluation. Our demo and user studies demonstrate SummPilot’s adaptability and usefulness for customizable summarization.

Method: Created WIT benchmark with expert-crafted plausible or counterfactual scenarios that provide contextual text. The benchmark evaluates whether models can condition their forecasts on these future scenarios.

Result: Developed a rigorous testbed for scenario-guided multimodal forecasting that challenges models to produce distinct forecasts from the same observations under different scenarios.

Conclusion: WIT addresses the gap in evaluating multimodal forecasting capabilities and enables testing whether models can meaningfully leverage textual context like human experts do with what-if scenarios.

Abstract: Time series forecasting is critical to real-world decision making, yet most existing approaches remain unimodal and rely on extrapolating historical patterns. While recent progress in large language models (LLMs) highlights the potential for multimodal forecasting, existing benchmarks largely provide retrospective or misaligned raw context, making it unclear whether such models meaningfully leverage textual inputs. In practice, human experts incorporate what-if scenarios with historical evidence, often producing distinct forecasts from the same observations under different scenarios. Inspired by this, we introduce What If TSF (WIT), a multimodal forecasting benchmark designed to evaluate whether models can condition their forecasts on contextual text, especially future scenarios. By providing expert-crafted plausible or counterfactual scenarios, WIT offers a rigorous testbed for scenario-guided multimodal forecasting. The benchmark is available at https://github.com/jinkwan1115/WhatIfTSF.

Method: Three-stage framework: 1) Fine-tuned VLM predicts bounding boxes from rough sketch and text to identify modification areas; 2) Stable diffusion generates detailed sketches of new elements merged with original outline via generative inpainting; 3) ControlNet refines results into photorealistic images.

Result: The framework successfully generates renovation proposals that preserve original structures while improving facade detail quality. Experiments on datasets and real industrial buildings demonstrate effectiveness in bypassing detailed as-built modeling requirements.

Conclusion: This AI-powered approach enables architects to rapidly explore design alternatives, iterate on early-stage concepts, and communicate renovation intentions more clearly, offering a more efficient alternative to traditional facade renovation workflows.

Abstract: Facade renovation offers a more sustainable alternative to full demolition, yet producing design proposals that preserve existing structures while expressing new intent remains challenging. Current workflows typically require detailed as-built modelling before design, which is time-consuming, labour-intensive, and often involves repeated revisions. To solve this issue, we propose a three-stage framework combining generative artificial intelligence (AI) and vision-language models (VLM) that directly processes rough structural sketch and textual descriptions to produce consistent renovation proposals. First, the input sketch is used by a fine-tuned VLM model to predict bounding boxes specifying where modifications are needed and which components should be added. Next, a stable diffusion model generates detailed sketches of new elements, which are merged with the original outline through a generative inpainting pipeline. Finally, ControlNet is employed to refine the result into a photorealistic image. Experiments on datasets and real industrial buildings indicate that the proposed framework can generate renovation proposals that preserve the original structure while improving facade detail quality. This approach effectively bypasses the need for detailed as-built modelling, enabling architects to rapidly explore design alternatives, iterate on early-stage concepts, and communicate renovation intentions with greater clarity.

Method: Two-stage framework: 1) Repair solution retrieval using edit-driven code retrieval to guide LLMs; 2) Solution-guided program repair that fixes code and provides explanations. Includes iterative retrieval enhancement that uses evaluation results to optimize retrieval direction.

Result: The approach outperforms baseline methods by a large margin, validating the effectiveness of the LSG framework for the newly proposed LPR task.

Conclusion: The proposed LSG framework effectively addresses the LPR task by not only repairing buggy code but also providing explanatory bug descriptions, representing an advancement in intelligent programming coaching systems.

Abstract: With the development of large language models (LLMs) in the field of programming, intelligent programming coaching systems have gained widespread attention. However, most research focuses on repairing the buggy code of programming learners without providing the underlying causes of the bugs. To address this gap, we introduce a novel task, namely \textbf{LPR} (\textbf{L}earner-Tailored \textbf{P}rogram \textbf{R}epair). We then propose a novel and effective framework, \textbf{\textsc{\MethodName{}}} (\textbf{L}earner-Tailored \textbf{S}olution \textbf{G}enerator), to enhance program repair while offering the bug descriptions for the buggy code. In the first stage, we utilize a repair solution retrieval framework to construct a solution retrieval database and then employ an edit-driven code retrieval approach to retrieve valuable solutions, guiding LLMs in identifying and fixing the bugs in buggy code. In the second stage, we propose a solution-guided program repair method, which fixes the code and provides explanations under the guidance of retrieval solutions. Moreover, we propose an Iterative Retrieval Enhancement method that utilizes evaluation results of the generated code to iteratively optimize the retrieval direction and explore more suitable repair strategies, improving performance in practical programming coaching scenarios. The experimental results show that our approach outperforms a set of baselines by a large margin, validating the effectiveness of our framework for the newly proposed LPR task.

Method: Built on Retrieval-Augmented Generation (RAG) and tool-calling architectures with two custom plugins: iwmi-doc-plugin for semantic search over indexed documents using Azure AI Search, and iwmi-api-plugin for querying live databases for dynamic insights.

Result: Achieves overall RAGAS score of 0.8043 with high answer relevancy (0.8571) and context precision (0.8009). Features include multilingual interactions, transparent source referencing, automated calculations, visualization, threshold-based alerts, and integration with LRB Digital Twin.

Conclusion: WaterCopilot establishes a replicable AI-augmented framework for enhancing water governance in data-scarce transboundary contexts, demonstrating potential to support informed decision-making and strengthen water security in complex river basins.

Abstract: Sustainable water resource management in transboundary river basins is challenged by fragmented data, limited real-time access, and the complexity of integrating diverse information sources. This paper presents WaterCopilot-an AI-driven virtual assistant developed through collaboration between the International Water Management Institute (IWMI) and Microsoft Research for the Limpopo River Basin (LRB) to bridge these gaps through a unified, interactive platform. Built on Retrieval-Augmented Generation (RAG) and tool-calling architectures, WaterCopilot integrates static policy documents and real-time hydrological data via two custom plugins: the iwmi-doc-plugin, which enables semantic search over indexed documents using Azure AI Search, and the iwmi-api-plugin, which queries live databases to deliver dynamic insights such as environmental-flow alerts, rainfall trends, reservoir levels, water accounting, and irrigation data. The system features guided multilingual interactions (English, Portuguese, French), transparent source referencing, automated calculations, and visualization capabilities. Evaluated using the RAGAS framework, WaterCopilot achieves an overall score of 0.8043, with high answer relevancy (0.8571) and context precision (0.8009). Key innovations include automated threshold-based alerts, integration with the LRB Digital Twin, and a scalable deployment pipeline hosted on AWS. While limitations in processing non-English technical documents and API latency remain, WaterCopilot establishes a replicable AI-augmented framework for enhancing water governance in data-scarce, transboundary contexts. The study demonstrates the potential of this AI assistant to support informed, timely decision-making and strengthen water security in complex river basins.

Method: Created ViDoRe v3 benchmark with 10 datasets across diverse professional domains, featuring 26,000 visually rich document pages paired with 3,099 human-verified queries in 6 languages. Used 12,000 hours of human annotation to provide high-quality annotations for retrieval relevance, bounding box localization, and verified reference answers.

Result: Evaluation shows: 1) Visual retrievers outperform textual ones, 2) Late-interaction models and textual reranking substantially improve performance, 3) Hybrid or purely visual contexts enhance answer generation quality, 4) Current models still struggle with non-textual elements, open-ended queries, and fine-grained visual grounding.

Conclusion: ViDoRe v3 addresses critical gaps in multimodal RAG evaluation and reveals current limitations in handling visual content and complex queries. The benchmark is released under a commercially permissive license to encourage progress in this important research area.

Abstract: Retrieval-Augmented Generation (RAG) pipelines must address challenges beyond simple single-document retrieval, such as interpreting visual elements (tables, charts, images), synthesizing information across documents, and providing accurate source grounding. Existing benchmarks fail to capture this complexity, often focusing on textual data, single-document comprehension, or evaluating retrieval and generation in isolation. We introduce ViDoRe v3, a comprehensive multimodal RAG benchmark featuring multi-type queries over visually rich document corpora. It covers 10 datasets across diverse professional domains, comprising ~26,000 document pages paired with 3,099 human-verified queries, each available in 6 languages. Through 12,000 hours of human annotation effort, we provide high-quality annotations for retrieval relevance, bounding box localization, and verified reference answers. Our evaluation of state-of-the-art RAG pipelines reveals that visual retrievers outperform textual ones, late-interaction models and textual reranking substantially improve performance, and hybrid or purely visual contexts enhance answer generation quality. However, current models still struggle with non-textual elements, open-ended queries, and fine-grained visual grounding. To encourage progress in addressing these challenges, the benchmark is released under a commercially permissive license at https://hf.co/vidore.

Method: Explainable multi-agent system inspired by asset management teams, using few-shot chain-of-thought prompting. Specialized agents collaborate on subtasks, acquiring meme coin trading knowledge and interpreting multi-modal data.

Result: System achieves 73% precision identifying high-quality meme coin projects and 70% precision identifying key opinion leader wallets. Selected KOLs generated $500,000 total profit across 1,000 projects.

Conclusion: Multi-agent LLM approach effectively addresses copy trading challenges in meme coins, outperforming traditional methods and providing explainable decisions through specialized agent collaboration.

Abstract: The launch of $Trump coin ignited a wave in meme coin investment. Copy trading, as a strategy-agnostic approach that eliminates the need for deep trading knowledge, quickly gains widespread popularity in the meme coin market. However, copy trading is not a guarantee of profitability due to the prevalence of manipulative bots, the uncertainty of the followed wallets’ future performance, and the lag in trade execution. Recently, large language models (LLMs) have shown promise in financial applications by effectively understanding multi-modal data and producing explainable decisions. However, a single LLM struggles with complex, multi-faceted tasks such as asset allocation. These challenges are even more pronounced in cryptocurrency markets, where LLMs often lack sufficient domain-specific knowledge in their training data. To address these challenges, we propose an explainable multi-agent system for meme coin copy trading. Inspired by the structure of an asset management team, our system decomposes the complex task into subtasks and coordinates specialized agents to solve them collaboratively. Employing few-shot chain-of-though (CoT) prompting, each agent acquires professional meme coin trading knowledge, interprets multi-modal data, and generates explainable decisions. Using a dataset of 1,000 meme coin projects’ transaction data, our empirical evaluation shows that the proposed multi-agent system outperforms both traditional machine learning models and single LLMs, achieving 73% and 70% precision in identifying high-quality meme coin projects and key opinion leader (KOL) wallets, respectively. The selected KOLs collectively generated a total profit of $500,000 across these projects.

Method: Four-module framework: 1) Complex intent decomposition into structured elements with logical dependencies, 2) Logical clarification generation organizing questions based on dependencies, 3) Intent-aware reward evaluation with Monte Carlo sampling for training data, 4) Self-evolved intent tuning through iterative feedback optimization.

Result: State-of-the-art performance: reduces logical conflicts to 11.5%, increases user satisfaction by 14.4%, decreases task completion time by 34.8%, and achieves superior logical consistency across clarification, execution, and cognitive load benchmarks.

Conclusion: Prism effectively addresses the core challenge of modeling logical dependencies in intent clarification, enabling more coherent and efficient human-LLM collaboration through structured decomposition and dependency-aware questioning.

Abstract: Large Language Models are rapidly emerging as web-native interfaces to social platforms. On the social web, users frequently have ambiguous and dynamic goals, making complex intent understanding-rather than single-turn execution-the cornerstone of effective human-LLM collaboration. Existing approaches attempt to clarify user intents through sequential or parallel questioning, yet they fall short of addressing the core challenge: modeling the logical dependencies among clarification questions. Inspired by the Cognitive Load Theory, we propose Prism, a novel framework for complex intent understanding that enables logically coherent and efficient intent clarification. Prism comprises four tailored modules: a complex intent decomposition module, which decomposes user intents into smaller, well-structured elements and identifies logical dependencies among them; a logical clarification generation module, which organizes clarification questions based on these dependencies to ensure coherent, low-friction interactions; an intent-aware reward module, which evaluates the quality of clarification trajectories via an intent-aware reward function and leverages Monte Carlo Sample to simulate user-LLM interactions for large-scale,high-quality training data generation; and a self-evolved intent tuning module, which iteratively refines the LLM’s logical clarification capability through data-driven feedback and optimization. Prism consistently outperforms existing approaches across clarification interactions, intent execution, and cognitive load benchmarks. It achieves stateof-the-art logical consistency, reduces logical conflicts to 11.5%, increases user satisfaction by 14.4%, and decreases task completion time by 34.8%. All data and code are released.

Method: Uses example-driven explanations and hands-on coding examples with approachable teaching methods to make RL concepts accessible to beginners.

Result: Students gain foundational RL skills and confidence to apply reinforcement learning techniques in real-world scenarios through practical learning.

Conclusion: The tutorial successfully makes RL more accessible to undergraduates by providing clear, practical guidance that bridges theory and implementation challenges.

Abstract: This tutorial is designed to make reinforcement learning (RL) more accessible to undergraduate students by offering clear, example-driven explanations. It focuses on bridging the gap between RL theory and practical coding applications, addressing common challenges that students face when transitioning from conceptual understanding to implementation. Through hands-on examples and approachable explanations, the tutorial aims to equip students with the foundational skills needed to confidently apply RL techniques in real-world scenarios.

Method: Parallel Context-of-Experts Decoding (Pced) treats retrieved documents as isolated experts, synchronizes their predictions via retrieval-aware contrastive decoding that weighs expert logits against model prior, shifting evidence aggregation from attention to decoding.

Result: Recovers cross-document reasoning capabilities without constructing shared attention across documents, eliminating the need to choose between multi-document reasoning and computational efficiency.

Conclusion: Pced offers a training-free framework that resolves the RAG trade-off by enabling efficient multi-document reasoning through decoding-level synchronization of isolated document experts.

Abstract: Retrieval Augmented Generation faces a trade-off: concatenating documents in a long prompt enables multi-document reasoning but creates prefill bottlenecks, while encoding document KV caches separately offers speed but breaks cross-document interaction. We propose Parallel Context-of-Experts Decoding (Pced), a training-free framework that shifts evidence aggregation from the attention mechanism to the decoding. Pced treats retrieved documents as isolated “experts”, synchronizing their predictions via a novel retrieval-aware contrastive decoding rule that weighs expert logits against the model prior. This approach recovers cross-document reasoning capabilities without constructing a shared attention across documents.

Method: Drawing on relational models theory to analyze LLM behaviors as structural generalizations of human interaction regimes that arise under extreme power asymmetries, rather than as categorical moral failures.

Result: LLMs’ “unethical” behaviors are better understood as statistical internalizations of the full spectrum of human social interactions (including laws, contracts, conflicts, coercive arrangements) rather than evidence of malign agency.

Conclusion: The primary AGI risk is structural amplification of human intelligence, power, and contradictions, not adversarial intent. Alignment failure is structural, requiring governance approaches focused on amplification, complexity, and regime stability rather than just model-level intent.

Abstract: Recent reports of large language models (LLMs) exhibiting behaviors such as deception, threats, or blackmail are often interpreted as evidence of alignment failure or emergent malign agency. We argue that this interpretation rests on a conceptual error. LLMs do not reason morally; they statistically internalize the record of human social interaction, including laws, contracts, negotiations, conflicts, and coercive arrangements. Behaviors commonly labeled as unethical or anomalous are therefore better understood as structural generalizations of interaction regimes that arise under extreme asymmetries of power, information, or constraint. Drawing on relational models theory, we show that practices such as blackmail are not categorical deviations from normal social behavior, but limiting cases within the same continuum that includes market pricing, authority relations, and ultimatum bargaining. The surprise elicited by such outputs reflects an anthropomorphic expectation that intelligence should reproduce only socially sanctioned behavior, rather than the full statistical landscape of behaviors humans themselves enact. Because human morality is plural, context-dependent, and historically contingent, the notion of a universally moral artificial intelligence is ill-defined. We therefore reframe concerns about artificial general intelligence (AGI). The primary risk is not adversarial intent, but AGI’s role as an endogenous amplifier of human intelligence, power, and contradiction. By eliminating longstanding cognitive and institutional frictions, AGI compresses timescales and removes the historical margin of error that has allowed inconsistent values and governance regimes to persist without collapse. Alignment failure is thus structural, not accidental, and requires governance approaches that address amplification, complexity, and regime stability rather than model-level intent alone.

Method: Introduces ESGAgent - a hierarchical multi-agent system with specialized tools (retrieval augmentation, web search, domain-specific functions). Also creates a comprehensive 3-level benchmark from 310 corporate sustainability reports to evaluate capabilities from atomic questions to integrated analysis.

Result: ESGAgent outperforms state-of-the-art closed-source LLMs with 84.15% average accuracy on atomic QA tasks, and excels in professional report generation with rich charts and verifiable references.

Conclusion: The benchmark has diagnostic value and establishes a vital testbed for assessing general and advanced agentic capabilities in high-stakes vertical domains like ESG analysis.

Abstract: Environmental, social, and governance (ESG) criteria are essential for evaluating corporate sustainability and ethical performance. However, professional ESG analysis is hindered by data fragmentation across unstructured sources, and existing large language models (LLMs) often struggle with the complex, multi-step workflows required for rigorous auditing. To address these limitations, we introduce ESGAgent, a hierarchical multi-agent system empowered by a specialized toolset, including retrieval augmentation, web search and domain-specific functions, to generate in-depth ESG analysis. Complementing this agentic system, we present a comprehensive three-level benchmark derived from 310 corporate sustainability reports, designed to evaluate capabilities ranging from atomic common-sense questions to the generation of integrated, in-depth analysis. Empirical evaluations demonstrate that ESGAgent outperforms state-of-the-art closed-source LLMs with an average accuracy of 84.15% on atomic question-answering tasks, and excels in professional report generation by integrating rich charts and verifiable references. These findings confirm the diagnostic value of our benchmark, establishing it as a vital testbed for assessing general and advanced agentic capabilities in high-stakes vertical domains.

Method: PersonaDual trains a single model to support both objective and personalized reasoning patterns via supervised fine-tuning (SFT), then optimizes mode selection using reinforcement learning with DualGRPO to adaptively switch between modes based on context.

Result: Experiments show PersonaDual preserves personalization benefits while reducing interference, achieving near interference-free performance and better leveraging helpful personalized signals to improve objective problem-solving.

Conclusion: PersonaDual successfully addresses the personalization-objectivity trade-off by enabling adaptive mode switching, balancing personalized interaction benefits with maintaining factual correctness and objectivity in LLMs.

Abstract: As users increasingly expect LLMs to align with their preferences, personalized information becomes valuable. However, personalized information can be a double-edged sword: it can improve interaction but may compromise objectivity and factual correctness, especially when it is misaligned with the question. To alleviate this problem, we propose PersonaDual, a framework that supports both general-purpose objective reasoning and personalized reasoning in a single model, and adaptively switches modes based on context. PersonaDual is first trained with SFT to learn two reasoning patterns, and then further optimized via reinforcement learning with our proposed DualGRPO to improve mode selection. Experiments on objective and personalized benchmarks show that PersonaDual preserves the benefits of personalization while reducing interference, achieving near interference-free performance and better leveraging helpful personalized signals to improve objective problem-solving.

Method: MemeWeaver is an end-to-end trainable multimodal framework featuring novel inter-meme graph reasoning mechanism. It systematically evaluates multiple visual-textual fusion strategies to capture relational patterns in online hate content.

Result: Consistently outperforms state-of-the-art baselines on MAMI and EXIST benchmarks while achieving faster training convergence. Learned graph structure captures semantically meaningful patterns revealing relational nature of online hate.

Conclusion: The framework successfully addresses limitations of existing approaches by capturing social dynamics through graph reasoning, offering valuable insights into the relational patterns of online sexism and misogyny.

Abstract: Women are twice as likely as men to face online harassment due to their gender. Despite recent advances in multimodal content moderation, most approaches still overlook the social dynamics behind this phenomenon, where perpetrators reinforce prejudices and group identity within like-minded communities. Graph-based methods offer a promising way to capture such interactions, yet existing solutions remain limited by heuristic graph construction, shallow modality fusion, and instance-level reasoning. In this work, we present MemeWeaver, an end-to-end trainable multimodal framework for detecting sexism and misogyny through a novel inter-meme graph reasoning mechanism. We systematically evaluate multiple visual–textual fusion strategies and show that our approach consistently outperforms state-of-the-art baselines on the MAMI and EXIST benchmarks, while achieving faster training convergence. Further analyses reveal that the learned graph structure captures semantically meaningful patterns, offering valuable insights into the relational nature of online hate.

Method: Obligatory-Information Phase Structured Compliance Evaluation (OIP-SCE) checks whether every required clinical obligation is met in the correct order, with clear evidence for clinician review, turning complex policies into actionable steps.

Result: Demonstrated in two case studies (respiratory history and benefits verification), OIP-SCE shows how phase-level evidence transforms policy into shared, actionable steps and provides an auditable evaluation surface.

Conclusion: OIP-SCE aligns AI capability with clinical workflow by giving clinicians control over what to check and engineers clear specifications to implement, supporting routine, safe use of conversational AI in healthcare.

Abstract: Conversational AI is starting to support real clinical work, but most evaluation methods miss how compliance depends on the full course of a conversation. We introduce Obligatory-Information Phase Structured Compliance Evaluation (OIP-SCE), an evaluation method that checks whether every required clinical obligation is met, in the right order, with clear evidence for clinicians to review. This makes complex rules practical and auditable, helping close the gap between technical progress and what healthcare actually needs. We demonstrate the method in two case studies (respiratory history, benefits verification) and show how phase-level evidence turns policy into shared, actionable steps. By giving clinicians control over what to check and engineers a clear specification to implement, OIP-SCE provides a single, auditable evaluation surface that aligns AI capability with clinical workflow and supports routine, safe use.

Method: Proposes three principles for explanation evaluation and introduces AXE (Adversarial Explanation Evaluation) method. AXE evaluates feature-importance explanations without requiring ground truth by focusing on detecting when protected attributes are used in predictions and identifying adversarial manipulation of explanations.

Result: AXE achieves 100% success rate in detecting adversarial fairwashing of explanations. Unlike prior methods based on model sensitivity or ground truth comparison, AXE can determine when protected attributes are used to make predictions, revealing behavioral differences within Rashomon sets that other evaluation methods obscure.

Conclusion: Explanation evaluation should focus on revealing behavioral differences in Rashomon sets rather than comparing to ground truth. AXE provides an effective method for evaluating explanations without ground truth, detecting adversarial manipulation, and identifying when models use protected attributes, helping select appropriate models for deployment.

Abstract: Explainable artificial intelligence (XAI) is concerned with producing explanations indicating the inner workings of models. For a Rashomon set of similarly performing models, explanations provide a way of disambiguating the behavior of individual models, helping select models for deployment. However explanations themselves can vary depending on the explainer used, and need to be evaluated. In the paper “Evaluating Model Explanations without Ground Truth”, we proposed three principles of explanation evaluation and a new method “AXE” to evaluate the quality of feature-importance explanations. We go on to illustrate how evaluation metrics that rely on comparing model explanations against ideal ground truth explanations obscure behavioral differences within a Rashomon set. Explanation evaluation aligned with our proposed principles would highlight these differences instead, helping select models from the Rashomon set. The selection of alternate models from the Rashomon set can maintain identical predictions but mislead explainers into generating false explanations, and mislead evaluation methods into considering the false explanations to be of high quality. AXE, our proposed explanation evaluation method, can detect this adversarial fairwashing of explanations with a 100% success rate. Unlike prior explanation evaluation strategies such as those based on model sensitivity or ground truth comparison, AXE can determine when protected attributes are used to make predictions.

Method: Cago dynamically tracks agent’s competence along expert trajectories and uses this signal to select intermediate steps (goals just beyond current reach) to guide learning, creating an adaptive curriculum.

Result: Empirical results show Cago significantly improves sample efficiency and final performance across sparse-reward, goal-conditioned tasks, consistently outperforming existing learning-from-demonstrations baselines.

Conclusion: Cago provides an effective approach to mitigate imitation learning’s brittle dependence on expert demonstrations by creating capability-aware goal sampling for adaptive curriculum learning in long-horizon environments.

Abstract: Despite its promise, imitation learning often fails in long-horizon environments where perfect replication of demonstrations is unrealistic and small errors can accumulate catastrophically. We introduce Cago (Capability-Aware Goal Sampling), a novel learning-from-demonstrations method that mitigates the brittle dependence on expert trajectories for direct imitation. Unlike prior methods that rely on demonstrations only for policy initialization or reward shaping, Cago dynamically tracks the agent’s competence along expert trajectories and uses this signal to select intermediate steps–goals that are just beyond the agent’s current reach–to guide learning. This results in an adaptive curriculum that enables steady progress toward solving the full task. Empirical results demonstrate that Cago significantly improves sample efficiency and final performance across a range of sparse-reward, goal-conditioned tasks, consistently outperforming existing learning from-demonstrations baselines.

Method: The authors examine current AI evaluation practices, identify their limitations in assessing entertainment value, and draw on humanities insights to propose “thick entertainment” as an alternative framework for evaluating AI-generated cultural content.

Result: The analysis reveals a critical asymmetry: AI assessments rigorously measure both benefits and harms of intelligence but focus almost exclusively on cultural harms, lacking frameworks to articulate how cultural outputs might be actively beneficial.

Conclusion: Entertainment will likely become a primary business model for AI corporations, exerting powerful influence on technology development. The paper concludes that AI may ultimately be as much about “intelligence” as social media is about social connection, requiring new evaluation frameworks like “thick entertainment.”

Abstract: Generative AI systems are predominantly designed, evaluated, and marketed as intelligent systems which will benefit society by augmenting or automating human cognitive labor, promising to increase personal, corporate, and macroeconomic productivity. But this mainstream narrative about what AI is and what it can do is in tension with another emerging use case: entertainment. We argue that the field of AI is unprepared to measure or respond to how the proliferation of entertaining AI-generated content will impact society. Emerging data suggest AI is already widely adopted for entertainment purposes – especially by young people – and represents a large potential source of revenue. We contend that entertainment will become a primary business model for major AI corporations seeking returns on massive infrastructure investments; this will exert a powerful influence on the technology these companies produce in the coming years. Examining current evaluation practices, we identify a critical asymmetry: while AI assessments rigorously measure both benefits and harms of intelligence, they focus almost exclusively on cultural harms. We lack frameworks for articulating how cultural outputs might be actively beneficial. Drawing on insights from the humanities, we propose “thick entertainment” as a framework for evaluating AI-generated cultural content – one that considers entertainment’s role in meaning-making, identity formation, and social connection rather than simply minimizing harm. While AI is often touted for its potential to revolutionize productivity, in the long run we may find that AI turns out to be as much about “intelligence” as social media is about social connection.

Method: Empirical study analyzing annotation error rates in BIRD and Spider 2.0-Snow benchmarks, correcting a subset of BIRD Dev set, and re-evaluating 16 open-source agents on both original and corrected versions.

Result: Found high error rates (52.8% in BIRD Mini-Dev, 62.8% in Spider 2.0-Snow). Performance changes ranged -7% to +31%, rank changes up to ±9 positions. Rankings on uncorrected subset strongly correlated with full Dev set, but weakly with corrected subset.

Conclusion: Annotation errors significantly distort reported performance and rankings, potentially misleading research directions and deployment choices. Highlights need for more rigorous benchmark validation.

Abstract: Researchers have proposed numerous text-to-SQL techniques to streamline data analytics and accelerate the development of database-driven applications. To compare these techniques and select the best one for deployment, the community depends on public benchmarks and their leaderboards. Since these benchmarks heavily rely on human annotations during question construction and answer evaluation, the validity of the annotations is crucial. In this paper, we conduct an empirical study that (i) benchmarks annotation error rates for two widely used text-to-SQL benchmarks, BIRD and Spider 2.0-Snow, and (ii) corrects a subset of the BIRD development (Dev) set to measure the impact of annotation errors on text-to-SQL agent performance and leaderboard rankings. Through expert analysis, we show that BIRD Mini-Dev and Spider 2.0-Snow have error rates of 52.8% and 62.8%, respectively. We re-evaluate all 16 open-source agents from the BIRD leaderboard on both the original and the corrected BIRD Dev subsets. We show that performance changes range from -7% to 31% (in relative terms) and rank changes range from $-9$ to $+9$ positions. We further assess whether these impacts generalize to the full BIRD Dev set. We find that the rankings of agents on the uncorrected subset correlate strongly with those on the full Dev set (Spearman’s $r_s$=0.85, $p$=3.26e-5), whereas they correlate weakly with those on the corrected subset (Spearman’s $r_s$=0.32, $p$=0.23). These findings show that annotation errors can significantly distort reported performance and rankings, potentially misguiding research directions or deployment choices. Our code and data are available at https://github.com/uiuc-kang-lab/text_to_sql_benchmarks.

Method: The paper introduces a general methodology for constructing political bias benchmarks by aligning model-generated voting predictions with verified parliamentary voting records. This is instantiated in three national case studies: PoliBiasNL (Dutch), PoliBiasNO (Norwegian), and PoliBiasES (Spanish). The evaluation framework includes visualizing LLM and party ideologies in a shared two-dimensional CHES space by linking voting-based positions to CHES dimensions.

Result: Experiments reveal that state-of-the-art LLMs consistently display left-leaning or centrist tendencies, with clear negative biases toward right-conservative parties across all three national benchmarks.

Conclusion: The findings highlight the value of transparent, cross-national evaluation grounded in real parliamentary behavior for understanding and auditing political bias in modern LLMs, providing a systematic approach to assess ideological tendencies in AI systems.

Abstract: As large language models (LLMs) become deeply embedded in digital platforms and decision-making systems, concerns about their political biases have grown. While substantial work has examined social biases such as gender and race, systematic studies of political bias remain limited, despite their direct societal impact. This paper introduces a general methodology for constructing political bias benchmarks by aligning model-generated voting predictions with verified parliamentary voting records. We instantiate this methodology in three national case studies: PoliBiasNL (2,701 Dutch parliamentary motions and votes from 15 political parties), PoliBiasNO (10,584 motions and votes from 9 Norwegian parties), and PoliBiasES (2,480 motions and votes from 10 Spanish parties). Across these benchmarks, we assess ideological tendencies and political entity bias in LLM behavior. As part of our evaluation framework, we also propose a method to visualize the ideology of LLMs and political parties in a shared two-dimensional CHES (Chapel Hill Expert Survey) space by linking their voting-based positions to the CHES dimensions, enabling direct and interpretable comparisons between models and real-world political actors. Our experiments reveal fine-grained ideological distinctions: state-of-the-art LLMs consistently display left-leaning or centrist tendencies, alongside clear negative biases toward right-conservative parties. These findings highlight the value of transparent, cross-national evaluation grounded in real parliamentary behavior for understanding and auditing political bias in modern LLMs.

Method: The xAiTrees framework constructs model-based hierarchical segmentations that preserve the model’s reasoning fidelity. It allows both human-centric and model-centric segmentation approaches, providing multiscale explanations that can identify biases and enhance understanding of neural network decision-making.

Result: Experiments demonstrate that xAiTrees delivers highly interpretable and faithful model explanations, surpassing traditional xAI methods. The framework provides new insights into enhancing xAI interpretability through hierarchical segmentation techniques.

Conclusion: The hierarchical segmentation approach offers a novel way to enhance xAI interpretability, providing both faithful model explanations and human-understandable insights into CNN decision-making processes at multiple scales.

Abstract: Challenges persist in providing interpretable explanations for neural network reasoning in explainable AI (xAI). Existing methods like Integrated Gradients produce noisy maps, and LIME, while intuitive, may deviate from the model’s reasoning. We introduce a framework that uses hierarchical segmentation techniques for faithful and interpretable explanations of Convolutional Neural Networks (CNNs). Our method constructs model-based hierarchical segmentations that maintain the model’s reasoning fidelity and allows both human-centric and model-centric segmentation. This approach offers multiscale explanations, aiding bias identification and enhancing understanding of neural network decision-making. Experiments show that our framework, xAiTrees, delivers highly interpretable and faithful model explanations, not only surpassing traditional xAI methods but shedding new light on a novel approach to enhancing xAI interpretability.

Method: COSINT-Agent integrates fine-tuned MLLMs with Entity-Event-Scene Knowledge Graph (EES-KG) using EES-Match framework for systematic extraction, reasoning, and contextualization of multimodal insights.

Result: Extensive experiments show superior performance across core OSINT tasks including entity recognition, EES generation, and context matching, validating its effectiveness.

Conclusion: COSINT-Agent demonstrates potential as a robust and scalable solution for advancing automated multimodal reasoning and enhancing OSINT methodologies.

Abstract: Open Source Intelligence (OSINT) requires the integration and reasoning of diverse multimodal data, presenting significant challenges in deriving actionable insights. Traditional approaches, including multimodal large language models (MLLMs), often struggle to infer complex contextual relationships or deliver comprehensive intelligence from unstructured data sources. In this paper, we introduce COSINT-Agent, a knowledge-driven multimodal agent tailored to address the challenges of OSINT in the Chinese domain. COSINT-Agent seamlessly integrates the perceptual capabilities of fine-tuned MLLMs with the structured reasoning power of the Entity-Event-Scene Knowledge Graph (EES-KG). Central to COSINT-Agent is the innovative EES-Match framework, which bridges COSINT-MLLM and EES-KG, enabling systematic extraction, reasoning, and contextualization of multimodal insights. This integration facilitates precise entity recognition, event interpretation, and context retrieval, effectively transforming raw multimodal data into actionable intelligence. Extensive experiments validate the superior performance of COSINT-Agent across core OSINT tasks, including entity recognition, EES generation, and context matching. These results underscore its potential as a robust and scalable solution for advancing automated multimodal reasoning and enhancing the effectiveness of OSINT methodologies.

Method: Two key innovations: (1) Graph Traversal as Visual Chain of Thought mechanism for incremental atom-bond prediction, and (2) Faithfully Recognize What You’ve Seen principle for aligning abbreviated structures with expanded annotations. For hand-drawn tasks, reinforcement learning with GRPO method using format, graph, and SMILES rewards. Two-stage training with SFT for printed images and GRPO for hand-drawn transfer.

Result: GTR-VL outperforms specialist models, chemistry-domain VLMs, and commercial VLMs on both printed and hand-drawn datasets. Created GTR-1.3M instruction-tuning dataset with corrected annotations and MolRec-Bench benchmark for fine-grained evaluation.

Conclusion: The proposed GTR-VL framework effectively addresses OCSR challenges through innovative graph traversal reasoning and data-centric principles, demonstrating superior performance across diverse molecular recognition tasks.

Abstract: Optical Chemical Structure Recognition (OCSR) is essential for converting molecular images into machine-readable formats. While recent vision-language models (VLMs) have shown promise, their image-captioning approach often struggles with complex molecular structures and inconsistent annotations. To address these issues, we introduce GTR-VL, featuring two key innovations: (1) the \textit{Graph Traversal as Visual Chain of Thought} mechanism that emulates human reasoning by incrementally parsing molecular graphs through sequential atom-bond predictions, and (2) the data-centric \textit{Faithfully Recognize What You’ve Seen} principle, which aligns abbreviated structures in images with their expanded annotations. For hand-drawn OCSR tasks, where datasets lack graph annotations and only provide final SMILES, we apply reinforcement learning using the GRPO method, introducing reward mechanisms like format reward, graph reward, and SMILES reward. This approach significantly enhances performance in hand-drawn recognition tasks through weak supervision. We developed GTR-1.3M, a large-scale instruction-tuning dataset with corrected annotations, and MolRec-Bench, the first benchmark for fine-grained evaluation of graph-parsing accuracy in OCSR. Our two-stage training scheme involves SFT training for printed images and the GRPO method for transferring capabilities to hand-drawn tasks. Experiments show that GTR-VL outperforms specialist models, chemistry-domain VLMs, and commercial VLMs on both printed and hand-drawn datasets.

Method: The method introduces a framework that constructs a conditioning distribution from a small set of diverse anchor generations, then uses this distribution to condition LLM generation via an xRAG-style projector, requiring no modification of LLM parameters.

Result: Experiments demonstrate that manifold-based conditioning substantially increases generative diversity, with direct benefits for enhancing divergent thinking as a core facet of creativity in language models.

Conclusion: The geometric structure of semantic space makes it explorable, and by traversing these manifolds to condition LLM generation, we can systematically expand the model’s semantic reach and enhance creative capabilities through improved divergent thinking.

Abstract: Starting from the hypothesis that knowledge in semantic space is organized along structured manifolds, we argue that this geometric structure renders the space explorable. By traversing it and using the resulting continuous representations to condition an LLM’s generation distribution, we can systematically expand the model’s reachable semantic range. We introduce a framework that requires no modification of LLM parameters and operationalizes this idea by constructing a conditioning distribution from a small set of diverse anchor generations. This distribution conditions LLM’s generation via an xRAG-style projector. Our experiments demonstrate that this manifold-based conditioning substantially increases generative diversity, with direct benefits for enhancing divergent thinking, a core facet of creativity, in language models.

Method: The authors define four types of visually unanswerable questions and construct MoHoBench, a large-scale benchmark with 12k+ visual question samples validated through multi-stage filtering and human verification. They benchmark 28 popular MLLMs and implement initial alignment methods using supervised and preference learning.

Result: Benchmarking reveals that most MLLMs fail to appropriately refuse to answer unanswerable questions. The study finds that honesty is not just a language modeling issue but is deeply influenced by visual information, requiring dedicated multimodal alignment approaches.

Conclusion: This work establishes the first systematic assessment of MLLM honesty, demonstrating the need for specialized multimodal honesty alignment methods. The proposed MoHoBench benchmark and initial alignment approaches provide a foundation for developing more trustworthy MLLMs.

Abstract: Recently Multimodal Large Language Models (MLLMs) have achieved considerable advancements in vision-language tasks, yet produce potentially harmful or untrustworthy content. Despite substantial work investigating the trustworthiness of language models, MMLMs’ capability to act honestly, especially when faced with visually unanswerable questions, remains largely underexplored. This work presents the first systematic assessment of honesty behaviors across various MLLMs. We ground honesty in models’ response behaviors to unanswerable visual questions, define four representative types of such questions, and construct MoHoBench, a large-scale MMLM honest benchmark, consisting of 12k+ visual question samples, whose quality is guaranteed by multi-stage filtering and human verification. Using MoHoBench, we benchmarked the honesty of 28 popular MMLMs and conducted a comprehensive analysis. Our findings show that: (1) most models fail to appropriately refuse to answer when necessary, and (2) MMLMs’ honesty is not solely a language modeling issue, but is deeply influenced by visual information, necessitating the development of dedicated methods for multimodal honesty alignment. Therefore, we implemented initial alignment methods using supervised and preference learning to improve honesty behavior, providing a foundation for future work on trustworthy MLLMs. Our data and code can be found at https://github.com/yanxuzhu/MoHoBench.

Method: MixGRPO integrates stochastic differential equations (SDE) and ordinary differential equations (ODE) with a sliding window mechanism. SDE sampling and GRPO-guided optimization are applied only within the window, while ODE sampling is used outside. This confines randomness to specific time-steps and reduces optimization overhead. MixGRPO-Flash variant supports higher-order solvers for even faster sampling.

Result: MixGRPO outperforms DanceGRPO in both effectiveness and efficiency, achieving nearly 50% lower training time. MixGRPO-Flash further reduces training time by 71% while maintaining comparable performance across multiple dimensions of human preference alignment.

Conclusion: MixGRPO provides an efficient framework for human preference alignment in image generation by strategically mixing SDE and ODE sampling with a sliding window approach, significantly reducing computational costs while maintaining or improving alignment quality.

Abstract: Although GRPO substantially enhances flow matching models in human preference alignment of image generation, methods such as FlowGRPO and DanceGRPO still exhibit inefficiency due to the necessity of sampling and optimizing over all denoising steps specified by the Markov Decision Process (MDP). In this paper, we propose $\textbf{MixGRPO}$, a novel framework that leverages the flexibility of mixed sampling strategies through the integration of stochastic differential equations (SDE) and ordinary differential equations (ODE). This streamlines the optimization process within the MDP to improve efficiency and boost performance. Specifically, MixGRPO introduces a sliding window mechanism, using SDE sampling and GRPO-guided optimization only within the window, while applying ODE sampling outside. This design confines sampling randomness to the time-steps within the window, thereby reducing the optimization overhead, and allowing for more focused gradient updates to accelerate convergence. Additionally, as time-steps beyond the sliding window are not involved in optimization, higher-order solvers are supported for faster sampling. So we present a faster variant, termed $\textbf{MixGRPO-Flash}$, which further improves training efficiency while achieving comparable performance. MixGRPO exhibits substantial gains across multiple dimensions of human preference alignment, outperforming DanceGRPO in both effectiveness and efficiency, with nearly 50% lower training time. Notably, MixGRPO-Flash further reduces training time by 71%.

Method: Proposes an LLM-based framework with domain-specific reasoning grounded in monetary policy transmission mechanisms to enrich semantic representation. Introduces a dynamic uncertainty decoding module to assess prediction confidence and enhance reliability.

Result: Achieves state-of-the-art performance on policy stance analysis task. Statistical analysis shows significant positive correlation between perceptual uncertainty and model error rates, validating uncertainty as a diagnostic signal.

Conclusion: The uncertainty-aware LLM framework effectively deciphers Fedspeak for monetary policy stance classification, with uncertainty measures providing valuable diagnostic information about model reliability.

Abstract: “Fedspeak”, the stylized and often nuanced language used by the U.S. Federal Reserve, encodes implicit policy signals and strategic stances. The Federal Open Market Committee strategically employs Fedspeak as a communication tool to shape market expectations and influence both domestic and global economic conditions. As such, automatically parsing and interpreting Fedspeak presents a high-impact challenge, with significant implications for financial forecasting, algorithmic trading, and data-driven policy analysis. In this paper, we propose an LLM-based, uncertainty-aware framework for deciphering Fedspeak and classifying its underlying monetary policy stance. Technically, to enrich the semantic and contextual representation of Fedspeak texts, we incorporate domain-specific reasoning grounded in the monetary policy transmission mechanism. We further introduce a dynamic uncertainty decoding module to assess the confidence of model predictions, thereby enhancing both classification accuracy and model reliability. Experimental results demonstrate that our framework achieves state-of-the-art performance on the policy stance analysis task. Moreover, statistical analysis reveals a significant positive correlation between perceptual uncertainty and model error rates, validating the effectiveness of perceptual uncertainty as a diagnostic signal.

Method: Created DeKeyNLU dataset with 1,500 annotated QA pairs to refine task decomposition and keyword extraction. Developed DeKeySQL, a RAG-based NL2SQL pipeline with three modules: user question understanding, entity retrieval, and generation. Fine-tuned models with DeKeyNLU.

Result: Fine-tuning with DeKeyNLU significantly improved SQL generation accuracy: from 62.31% to 69.10% on BIRD dev dataset, and from 84.2% to 88.7% on Spider dev dataset.

Conclusion: The DeKeyNLU dataset effectively addresses key limitations in NL2SQL systems, particularly improving task decomposition and keyword extraction, leading to substantial accuracy improvements in SQL generation across standard benchmarks.

Abstract: Natural Language to SQL (NL2SQL) provides a new model-centric paradigm that simplifies database access for non-technical users by converting natural language queries into SQL commands. Recent advancements, particularly those integrating Retrieval-Augmented Generation (RAG) and Chain-of-Thought (CoT) reasoning, have made significant strides in enhancing NL2SQL performance. However, challenges such as inaccurate task decomposition and keyword extraction by LLMs remain major bottlenecks, often leading to errors in SQL generation. While existing datasets aim to mitigate these issues by fine-tuning models, they struggle with over-fragmentation of tasks and lack of domain-specific keyword annotations, limiting their effectiveness. To address these limitations, we present DeKeyNLU, a novel dataset which contains 1,500 meticulously annotated QA pairs aimed at refining task decomposition and enhancing keyword extraction precision for the RAG pipeline. Fine-tuned with DeKeyNLU, we propose DeKeySQL, a RAG-based NL2SQL pipeline that employs three distinct modules for user question understanding, entity retrieval, and generation to improve SQL generation accuracy. We benchmarked multiple model configurations within DeKeySQL RAG pipeline. Experimental results demonstrate that fine-tuning with DeKeyNLU significantly improves SQL generation accuracy on both BIRD (62.31% to 69.10%) and Spider (84.2% to 88.7%) dev datasets.

Method: The key innovation is leveraging Boolean Set Cover as a subroutine to combine existing formulas using Boolean connectives. This approach provides a novel trade-off between computational efficiency and formula size.

Result: Bolt learns formulas more than 100x faster than state-of-the-art methods on 70% of benchmarks, while producing smaller or equal formulas in 98% of cases.

Conclusion: The Boolean Set Cover approach enables significant performance improvements in LTLf formula learning, offering a practical solution that balances speed and formula compactness for real-world applications.

Abstract: Learning formulas in Linear Temporal Logic (LTLf) from finite traces is a fundamental research problem which has found applications in artificial intelligence, software engineering, programming languages, formal methods, control of cyber-physical systems, and robotics. We implement a new CPU tool called Bolt improving over the state of the art by learning formulas more than 100x faster over 70% of the benchmarks, with smaller or equal formulas in 98% of the cases. Our key insight is to leverage a problem called Boolean Set Cover as a subroutine to combine existing formulas using Boolean connectives. Thanks to the Boolean Set Cover component, our approach offers a novel trade-off between efficiency and formula size.

Method: Developed two LLM-based systems: Single-Agent System (SAS) with one LLM agent inferring repayment signals, and Popperian Multi-agent Debate System (PMADS) using adversarial argumentation under Karl Popper Debate protocol. Evaluated on work productivity, report quality/usability ratings by professionals, and reasoning-tree analysis.

Result: Both systems drastically reduced task completion time vs. human experts. Professionals rated SAS reports as adequate, while PMADS reports exceeded neutral benchmarks and scored significantly higher in explanatory adequacy, practical applicability, and usability. PMADS produced deeper, more elaborated reasoning trees compared to SAS’s single-layered structures.

Conclusion: Structured multi-agent debate enhances analytical rigor and perceived usefulness, though with longer computation time. Reasoning-centered automation represents a promising approach for developing useful AI systems in decision-critical financial contexts.

Abstract: This study investigated LLM-based automation for analyzing non-financial data in corporate credit evaluation. Two systems were developed and compared: a Single-Agent System (SAS), in which one LLM agent infers favorable and adverse repayment signals, and a Popperian Multi-agent Debate System (PMADS), which structures the dual-perspective analysis as adversarial argumentation under the Karl Popper Debate protocol. Evaluation addressed three fronts: (i) work productivity compared with human experts; (ii) perceived report quality and usability, rated by credit risk professionals for system-generated reports; and (iii) reasoning characteristics quantified via reasoning-tree analysis. Both systems drastically reduced task completion time relative to human experts. Professionals rated SAS reports as adequate, while PMADS reports exceeded neutral benchmarks and scored significantly higher in explanatory adequacy, practical applicability, and usability. Reasoning-tree analysis showed PMADS produced deeper, more elaborated structures, whereas SAS yielded single-layered trees. These findings suggest that structured multi-agent debate enhances analytical rigor and perceived usefulness, though at the cost of longer computation time. Overall, the results demonstrate that reasoning-centered automation represents a promising approach for developing useful AI systems in decision-critical financial contexts.

Method: The paper uses a two-player Modified-Action MDP to analyze suboptimal behavior of standard AI agents, and proposes Interruptible Collaborative Roleplayer (ICR) - a novel partner-aware learning algorithm to train common-ground-optimal collaborators that intelligently collect information from partner interventions.

Result: Experiments on multiple collaborative task environments show that ICR, on average, is more capable of promoting successful common ground convergence and exploring more diverse solutions compared to standard approaches.

Conclusion: The paper demonstrates that partner-aware learning algorithms like ICR can effectively address the limitation of standard RLHF-trained agents in collaborative settings, enabling better group common ground alignment and more diverse solution exploration in multi-agent tasks.

Abstract: Large Language Models (LLMs) are increasingly being deployed in agentic settings where they act as collaborators with humans. Therefore, it is increasingly important to be able to evaluate their abilities to collaborate effectively in multi-turn, multi-party tasks. In this paper, we build on the AI alignment and safe interruptibility literature to offer novel theoretical insights on collaborative behavior between LLM-driven collaborator agents and an intervention agent. Our goal is to learn an ideal partner-aware collaborator that increases the group’s common-ground (CG) alignment on task-relevant propositions-by intelligently collecting information provided in interventions by a partner agent. We show how LLM agents trained using standard RLHF and related approaches are naturally inclined to ignore possibly well-meaning interventions, which makes increasing group common ground non-trivial in this setting. We employ a two-player Modified-Action MDP to examine this suboptimal behavior of standard AI agents, and propose Interruptible Collaborative Roleplayer (ICR)-a novel partner-aware learning algorithm to train CG-optimal collaborators. Experiments on multiple collaborative task environments show that ICR, on average, is more capable of promoting successful CG convergence and exploring more diverse solutions in such tasks.

Method: Proposed ToolRM family of lightweight reward models with a novel pipeline using rule-based scoring and multidimensional sampling to create ToolPref-Pairwise-30K dataset, and introduced TRBench_BFCL benchmark for evaluation.

Result: Qwen3-4B/8B models achieve up to 17.94% higher accuracy than frontier LLMs and RMs in pairwise reward judgments, with generative ToolRM reducing output token usage by over 66% while enabling inference-time scaling.

Conclusion: ToolRM effectively addresses the gap in tool-specific reward modeling, demonstrating superior performance, efficiency, and practical utility for downstream RL training and broader critique tasks in tool learning.

Abstract: Reward models (RMs) play a critical role in aligning large language models (LLMs) with human preferences. Yet in the domain of tool learning, the lack of RMs specifically designed for function-calling tasks has limited progress toward more capable agentic AI. We introduce ToolRM, a family of lightweight reward models tailored for general tool-use scenarios. To build these models, we propose a novel pipeline that constructs high-quality pairwise preference data using rule-based scoring and multidimensional sampling. This yields ToolPref-Pairwise-30K, a diverse, balanced, and challenging preference dataset that supports both generative and discriminative reward modeling. We also introduce TRBench$_{BFCL}$, a benchmark built on the agent evaluation suite BFCL to evaluate RMs on tool calling tasks. Trained on our constructed data, models from the Qwen3-4B/8B series achieve up to 17.94% higher accuracy, substantially outperforming frontier LLMs and RMs in pairwise reward judgments. Beyond training objectives, generative ToolRM generalizes to broader critique tasks, including Best-of-N sampling and self-correction. Experiments on ACEBench highlight its effectiveness and efficiency, enabling inference-time scaling while reducing output token usage by over 66%. Its support for downstream RL training further validates its practical utility. We release data to facilitate future research.

Method: Proposes a single-agent reinforcement learning model with state, action, and reward functions defined based on queue length. Uses a modified queue length definition that can be reliably estimated from probe vehicle link travel time data. The action design regulates queue dynamics across multiple intersections.

Result: The model was evaluated using SUMO simulation platform and demonstrated effective mitigation of large-scale regional congestion through coordinated multi-intersection control.

Conclusion: The proposed single-agent RL approach provides a scalable solution for regional adaptive traffic signal control that leverages widely available probe vehicle data, making it suitable for widespread urban deployment.

Abstract: Several studies have employed reinforcement learning (RL) to address the challenges of regional adaptive traffic signal control (ATSC) and achieved promising results. In this field, existing research predominantly adopts multi-agent frameworks. However, the adoption of multi-agent frameworks presents challenges for scalability. Instead, the Traffic signal control (TSC) problem necessitates a single-agent framework. TSC inherently relies on centralized management by a single control center, which can monitor traffic conditions across all roads in the study area and coordinate the control of all intersections. This work proposes a single-agent RL-based regional ATSC model compatible with probe vehicle technology. Key components of the RL design include state, action, and reward function definitions. To facilitate learning and manage congestion, both state and reward functions are defined based on queue length, with action designed to regulate queue dynamics. The queue length definition used in this study differs slightly from conventional definitions but is closely correlated with congestion states. More importantly, it allows for reliable estimation using link travel time data from probe vehicles. With probe vehicle data already covering most urban roads, this feature enhances the proposed method’s potential for widespread deployment. The method was comprehensively evaluated using the SUMO simulation platform. Experimental results demonstrate that the proposed model effectively mitigates large-scale regional congestion levels via coordinated multi-intersection control.

Method: Proposes information capacity as a measure of model efficiency based on text compression performance relative to computational complexity. The metric incorporates tokenizer efficiency. Evaluated 52 open-source models and conducted experiments on 5 heterogeneous datasets to analyze linguistic bias and identify major factors affecting information capacity.

Result: Found consistent information capacity among different-sized models within a series. Identified strong linguistic bias in mainstream LLMs across datasets. Three major factors affecting information capacity are tokenizer efficiency, pretraining data, and mixture-of-experts architecture. Verified accuracy of performance prediction across model sizes based on information capacity and showed correlation with benchmark scores.

Conclusion: Information capacity serves as an effective unified metric for evaluating LLM efficiency that accounts for tokenizer efficiency and computational complexity, enabling better performance prediction across model sizes and revealing important architectural and data factors affecting efficiency.

Abstract: Recent years have witnessed the rapid advancements of large language models (LLMs) and their expanding applications, leading to soaring demands for computational resources. The widespread adoption of test-time scaling further intensifies the tension between model capability and resource consumption, highlighting the importance of inference efficiency. However, a unified metric that accurately reflects an LLM’s efficiency across diverse model sizes and architectures remains absent. Motivated by the correlation between compression and intelligence, we introduce information capacity, a measure of model efficiency based on text compression performance relative to computational complexity. A distinctive feature of information capacity is its incorporation of tokenizer efficiency, which affects inference costs but is often neglected in LLM evaluations. We assess the information capacity of 52 open-source models and observe a consistent information capacity among different-sized models within a series. Experiments on 5 heterogeneous datasets reveal strong linguistic bias in mainstream LLMs. Three major factors of information capacity include tokenizer efficiency, pretraining data, and the mixture-of-experts architecture. Empirical results verify the accuracy of performance prediction across model sizes based on information capacity and show the correlation between information capacity and benchmark scores.

Method: Conceptual framework dividing video foundation models into two core components: implicit world model (encodes physical laws, interaction dynamics, agent behavior) and video renderer (transforms latent simulation into visual observations). Traces progression through four generations of video generation capabilities.

Result: Provides systematic overview of video generation evolution culminating in world models with physical plausibility, real-time multimodal interaction, and multi-scale planning capabilities. Examines applications in robotics, autonomous driving, and gaming.

Conclusion: Video foundation models are evolving toward implicit world models that combine simulation engines with visual rendering. Future challenges include integrating agent intelligence and developing evaluation frameworks for next-generation world models.

Abstract: The landscape of video generation is shifting, from a focus on generating visually appealing clips to building virtual environments that support interaction and maintain physical plausibility. These developments point toward the emergence of video foundation models that function not only as visual generators but also as implicit world models, models that simulate the physical dynamics, agent-environment interactions, and task planning that govern real or imagined worlds. This survey provides a systematic overview of this evolution, conceptualizing modern video foundation models as the combination of two core components: an implicit world model and a video renderer. The world model encodes structured knowledge about the world, including physical laws, interaction dynamics, and agent behavior. It serves as a latent simulation engine that enables coherent visual reasoning, long-term temporal consistency, and goal-driven planning. The video renderer transforms this latent simulation into realistic visual observations, effectively producing videos as a “window” into the simulated world. We trace the progression of video generation through four generations, in which the core capabilities advance step by step, ultimately culminating in a world model, built upon a video generation model, that embodies intrinsic physical plausibility, real-time multimodal interaction, and planning capabilities spanning multiple spatiotemporal scales. For each generation, we define its core characteristics, highlight representative works, and examine their application domains such as robotics, autonomous driving, and interactive gaming. Finally, we discuss open challenges and design principles for next-generation world models, including the role of agent intelligence in shaping and evaluating these systems. An up-to-date list of related works is maintained at this link.

Method: Systematically evaluated off-policy data influence on probe generalization across eight distinct LLM behaviors. Tested linear and attention probes across multiple LLMs. Proposed a test using incentivized data to predict generalization failures when on-policy test data is unavailable.

Result: Training data generation strategy significantly affects probe performance, with largest generalization failures for intent-based behaviors (e.g., strategic deception) rather than text-level content (e.g., list usage). Successful generalization to incentivized data strongly correlates with high performance against on-policy examples.

Conclusion: Current deception probes may fail to generalize to real monitoring scenarios. Off-policy data can yield more reliable probes than on-policy data from sufficiently different settings, highlighting the need for new monitoring methods that better handle all types of distribution shift.

Abstract: Probing has emerged as a promising method for monitoring large language models (LLMs), enabling cheap inference-time detection of concerning behaviours. However, natural examples of many behaviours are rare, forcing researchers to rely on synthetic or off-policy LLM responses for training probes. We systematically evaluate how off-policy data influences probe generalisation across eight distinct LLM behaviours. Testing linear and attention probes across multiple LLMs, we find that training data generation strategy can significantly affect probe performance, though the magnitude varies greatly by behaviour. The largest generalisation failures arise for behaviours defined by response “intent” (e.g. strategic deception) rather than text-level content (e.g. usage of lists). We then propose a useful test for predicting generalisation failures in cases where on-policy test data is unavailable: successful generalisation to incentivised data (where the model was coerced) strongly correlates with high performance against on-policy examples. Based on these results, we predict that current deception probes may fail to generalise to real monitoring scenarios. Additionally, our finding that off-policy data can yield more reliable probes than on-policy data from a sufficiently different setting underscores the need for new monitoring methods that better handle all types of distribution shift.

Method: ICPO calculates preference advantage scores by comparing relative generation probabilities of multiple responses under the same input prompt, then integrates these scores with verifiable rewards to guide exploration.

Result: ICPO alleviates coarse-grained rewards and reward noise issues, curbs overconfident errors, enhances undervalued high-quality responses, prevents overfitting, and steadily boosts reasoning across multiple benchmarks.

Conclusion: ICPO effectively addresses RLVR limitations and demonstrates superior reasoning performance compared to GRPO across both general-domain and mathematical benchmarks.

Abstract: Reinforcement Learning with Verifiable Rewards (RLVR) demonstrates significant potential in enhancing the reasoning capabilities of Large Language Models (LLMs). However, existing RLVR methods are often constrained by issues such as coarse-grained rewards, reward noise, and inefficient exploration, which lead to unstable training and entropy collapse. To address this challenge, we propose the Intrinsic Confidence-Driven Group Relative Preference Optimization method (ICPO). The intuition behind it lies in the fact that the probabilities of an LLM generating different responses can inherently and directly reflect its self-assessment of the reasoning process. Inspired by the idea of preference modeling, ICPO calculates a preference advantage score for each response by comparing the relative generation probabilities of multiple responses under the same input prompt, and integrates this score with verifiable rewards to guide the exploration process. We have discovered that the preference advantage score not only alleviates the issues of coarse-grained rewards and reward noise but also effectively curbs overconfident errors, enhances the relative superiority of undervalued high-quality responses, and prevents the model from overfitting to specific strategies. Comprehensive experiments across four general-domain benchmarks and three mathematical benchmarks demonstrate that ICPO steadily boosts reasoning compared to GRPO.

Method: Tested five state-of-the-art LLMs (proprietary, open-source, and reasoning models) on sequential naming and production tasks with letters and words. Used various prompting strategies including chain-of-thought, evaluated scaling effects with increasing model size, and analyzed embedding dynamics during enumeration.

Result: Some LLMs can deploy counting procedures when explicitly prompted, but none spontaneously engage in counting when simply asked to enumerate items. Counting abilities don’t consistently follow scaling laws with model size.

Conclusion: Despite impressive emergent abilities, LLMs cannot yet robustly and systematically deploy counting procedures, highlighting a persistent gap between neural and symbolic approaches to compositional generalization.

Abstract: Reliably counting and generating sequences of items remain a significant challenge for neural networks, including Large Language Models (LLMs). Indeed, although this capability is readily handled by rule-based symbolic systems based on serial computation, learning to systematically deploy counting procedures is difficult for neural models, which should acquire these skills through learning. Previous research has demonstrated that recurrent architectures can only approximately track and enumerate sequences of events, and it remains unclear whether modern deep learning systems, including LLMs, can deploy systematic counting procedures over sequences of discrete symbols. This paper aims to fill this gap by investigating the sequential enumeration abilities of five state-of-the-art LLMs, including proprietary, open-source, and reasoning models. We probe LLMs in sequential naming and production tasks involving lists of letters and words, adopting a variety of prompting instructions to explore the role of chain-of-thought in the spontaneous emerging of counting strategies. We also evaluate open-source models with the same architecture but increasing size to see whether the mastering of counting principles follows scaling laws, and we analyze the embedding dynamics during sequential enumeration to investigate the emergent encoding of numerosity. We find that some LLMs are indeed capable of deploying counting procedures when explicitly prompted to do so, but none of them spontaneously engage in counting when simply asked to enumerate the number of items in a sequence. Our results suggest that, despite their impressive emergent abilities, LLMs cannot yet robustly and systematically deploy counting procedures, highlighting a persistent gap between neural and symbolic approaches to compositional generalization.

Method: Created LocalSearchBench with a database of over 1.3M merchant entries across 6 service categories and 9 cities, plus 900 multi-hop QA tasks from real user queries. Also developed LocalPlayground, a unified environment with multiple tools for LRM interaction.

Result: State-of-the-art LRMs perform poorly: best model (DeepSeek-V3.2) achieves only 35.60% correctness, with average completeness at 60.32% and faithfulness at 30.72%, showing significant challenges in this domain.

Conclusion: The benchmark reveals critical gaps in current LRMs for local life services, highlighting the need for specialized benchmarks and domain-specific agent training to handle real-world ambiguous queries and complex multi-hop reasoning.

Abstract: Recent advances in large reasoning models LRMs have enabled agentic search systems to perform complex multi-step reasoning across multiple sources. However, most studies focus on general information retrieval and rarely explores vertical domains with unique challenges. In this work, we focus on local life services and introduce LocalSearchBench, which encompass diverse and complex business scenarios. Real-world queries in this domain are often ambiguous and require multi-hop reasoning across merchants and products, remaining challenging and not fully addressed. As the first comprehensive benchmark for agentic search in local life services, LocalSearchBench comprises a database of over 1.3M merchant entries across 6 service categories and 9 major cities, and 900 multi-hop QA tasks from real user queries that require multi-step reasoning. We also developed LocalPlayground, a unified environment integrating multiple tools for LRMs interaction. Experiments show that even state-of-the-art LRMs struggle on LocalSearchBench: the best model (DeepSeek-V3.2) achieves only 35.60% correctness, and most models have issues with completeness (average 60.32%) and faithfulness (average 30.72%). This highlights the need for specialized benchmarks and domain-specific agent training in local life services. Code, Benchmark, and Leaderboard are available at https://localsearchbench.github.io/.

Method: Formulates jailbreak attacks as multi-turn RL problem, directly optimizing final-turn response harmfulness as outcome reward. Introduces TROJail with two process rewards: (1) penalizes overly harmful prompts that trigger refusal mechanisms, (2) encourages steering semantic relevance toward targeted harmful content.

Result: Experimental results show improved attack success rates across multiple models and benchmarks, demonstrating effectiveness of the approach.

Conclusion: TROJail provides an effective reinforcement learning framework for automated multi-turn jailbreak attacks that learns long-term strategies through outcome and process rewards.

Abstract: Large language models have seen widespread adoption, yet they remain vulnerable to multi-turn jailbreak attacks, threatening their safe deployment. This has led to the task of training automated multi-turn attackers to probe model safety vulnerabilities. However, existing approaches typically rely on turn-level optimization, which is insufficient for learning long-term attack strategies. To bridge this gap, we formulate this task as a multi-turn reinforcement learning problem, directly optimizing the harmfulness of the final-turn response as the outcome reward. To address the sparse supervision of the outcome reward, we introduce TROJail, which employs two process rewards to evaluate the utility of intermediate prompts and integrate them into advantage estimation. These rewards (1) penalize overly harmful prompts that trigger the model’s refusal mechanism, and (2) encourage steering the semantic relevance of responses toward the targeted harmful content. Experimental results show improved attack success rates across multiple models and benchmarks, highlighting the effectiveness of our approach. The code is available at https://github.com/xxiqiao/TROJail. Warning: This paper contains examples of harmful content.

Method: AI TIPS (Artificial Intelligence Trust-Integrated Pillars for Sustainability) 2.0 - an updated comprehensive operational framework that provides tailored governance for unique use cases, translates principles into specific technical implementations, and enables systematic embedding of trustworthy AI practices throughout development lifecycle.

Result: The framework directly addresses the three identified governance gaps by offering use-case specific risk assessment, actionable controls beyond conceptual principles, and mechanisms for operationalizing governance at scale with quantitative compliance measurement and role-appropriate visibility.

Conclusion: AI TIPS 2.0 provides a practical solution to critical AI governance challenges that current frameworks fail to address, enabling organizations to implement trustworthy AI systems with proper risk assessment, actionable controls, and scalable operationalization.

Abstract: The deployment of AI systems faces three critical governance challenges that current frameworks fail to adequately address. First, organizations struggle with inadequate risk assessment at the use case level, exemplified by the Humana class action lawsuit and other high impact cases where an AI system deployed to production exhibited both significant bias and high error rates, resulting in improper healthcare claim denials. Each AI use case presents unique risk profiles requiring tailored governance, yet most frameworks provide one size fits all guidance. Second, existing frameworks like ISO 42001 and NIST AI RMF remain at high conceptual levels, offering principles without actionable controls, leaving practitioners unable to translate governance requirements into specific technical implementations. Third, organizations lack mechanisms for operationalizing governance at scale, with no systematic approach to embed trustworthy AI practices throughout the development lifecycle, measure compliance quantitatively, or provide role-appropriate visibility from boards to data scientists. We present AI TIPS, Artificial Intelligence Trust-Integrated Pillars for Sustainability 2.0, update to the comprehensive operational framework developed in 2019,four years before NIST’s AI Risk Management Framework, that directly addresses these challenges.

Method: Proposes Vision-Language Simulation Model (VLSM) that synthesizes FlexScript from layout sketches and natural-language prompts, supported by a new large-scale dataset of 120,000+ prompt-sketch-code triplets for multimodal learning.

Result: Models achieve near-perfect structural accuracy and high execution robustness through systematic ablation studies across vision encoders, connectors, and code-pretrained language backbones.

Conclusion: Establishes foundation for generative digital twins that integrate visual reasoning and language understanding into executable industrial simulation systems.

Abstract: We propose a Vision-Language Simulation Model (VLSM) that unifies visual and textual understanding to synthesize executable FlexScript from layout sketches and natural-language prompts, enabling cross-modal reasoning for industrial simulation systems. To support this new paradigm, the study constructs the first large-scale dataset for generative digital twins, comprising over 120,000 prompt-sketch-code triplets that enable multimodal learning between textual descriptions, spatial structures, and simulation logic. In parallel, three novel evaluation metrics, Structural Validity Rate (SVR), Parameter Match Rate (PMR), and Execution Success Rate (ESR), are proposed specifically for this task to comprehensively evaluate structural integrity, parameter fidelity, and simulator executability. Through systematic ablation across vision encoders, connectors, and code-pretrained language backbones, the proposed models achieve near-perfect structural accuracy and high execution robustness. This work establishes a foundation for generative digital twins that integrate visual reasoning and language understanding into executable industrial simulation systems. Project page: https://danielhsu2014.github.io/GDT-VLSM-project/

Method: Introduced SafePro benchmark featuring a dataset of high-complexity tasks across diverse professional domains with safety risks, developed through rigorous iterative creation and review process. Evaluated state-of-the-art AI models and investigated safety mitigation strategies.

Result: Evaluation revealed significant safety vulnerabilities and uncovered new unsafe behaviors in professional contexts. Models exhibited both insufficient safety judgment and weak safety alignment when executing complex professional tasks. Safety mitigation strategies showed encouraging improvements.

Conclusion: Findings highlight urgent need for robust safety mechanisms tailored to next generation of professional AI agents, as current models show inadequate safety alignment for complex professional tasks.

Abstract: Large language model-based agents are rapidly evolving from simple conversational assistants into autonomous systems capable of performing complex, professional-level tasks in various domains. While these advancements promise significant productivity gains, they also introduce critical safety risks that remain under-explored. Existing safety evaluations primarily focus on simple, daily assistance tasks, failing to capture the intricate decision-making processes and potential consequences of misaligned behaviors in professional settings. To address this gap, we introduce \textbf{SafePro}, a comprehensive benchmark designed to evaluate the safety alignment of AI agents performing professional activities. SafePro features a dataset of high-complexity tasks across diverse professional domains with safety risks, developed through a rigorous iterative creation and review process. Our evaluation of state-of-the-art AI models reveals significant safety vulnerabilities and uncovers new unsafe behaviors in professional contexts. We further show that these models exhibit both insufficient safety judgment and weak safety alignment when executing complex professional tasks. In addition, we investigate safety mitigation strategies for improving agent safety in these scenarios and observe encouraging improvements. Together, our findings highlight the urgent need for robust safety mechanisms tailored to the next generation of professional AI agents.

Method: Extending Chen et al. (2025)’s work by testing LRMs on multiple choice questions with hints, directly asking models to reflect on unusual/hinted prompt content, and allowing them to use hints while measuring their actual reliance through experiments.

Result: LRMs flatly deny relying on hints provided in prompts for answering questions, even when directly questioned about unusual prompt content and even though experimental evidence shows they are actually using the hints.

Conclusion: The findings have discouraging implications for Chain-of-Thought monitoring and interpretability, revealing that LRMs can actively lie about their reasoning processes rather than just omitting information.

Abstract: It has been shown that Large Reasoning Models (LRMs) may not say what they think: they do not always volunteer information about how certain parts of the input influence their reasoning. But it is one thing for a model to omit such information and another, worse thing to lie about it. Here, we extend the work of Chen et al. (2025) to show that LRMs will do just this: they will flatly deny relying on hints provided in the prompt in answering multiple choice questions – even when directly asked to reflect on unusual (i.e. hinted) prompt content, even when allowed to use hints, and even though experiments show them to be using the hints. Our results thus have discouraging implications for CoT monitoring and interpretability.

Method: Created LJ-Spoof corpus with systematic variation across: prosody, vocoders, generative hyperparameters, bona fide prompt sources, training regimes, and neural post-processing. The corpus includes one speaker’s studio-quality recordings, 30 TTS families, 500 generatively variant subsets, 10 bona fide neural-processing variants, and over 3 million utterances.

Result: The variation-dense design enables robust speaker-conditioned anti-spoofing and fine-grained synthesis-source tracing. The dataset serves as both a practical training resource and benchmark evaluation suite for anti-spoofing and source tracing research.

Conclusion: LJ-Spoof addresses the critical gap in audio anti-spoofing research by providing a systematically varied, speaker-specific corpus that enables comprehensive evaluation of anti-spoofing and synthesis-source tracing methods, positioning it as both a training resource and benchmark suite.

Abstract: Speaker-specific anti-spoofing and synthesis-source tracing are central challenges in audio anti-spoofing. Progress has been hampered by the lack of datasets that systematically vary model architectures, synthesis pipelines, and generative parameters. To address this gap, we introduce LJ-Spoof, a speaker-specific, generatively diverse corpus that systematically varies prosody, vocoders, generative hyperparameters, bona fide prompt sources, training regimes, and neural post-processing. The corpus spans one speakers-including studio-quality recordings-30 TTS families, 500 generatively variant subsets, 10 bona fide neural-processing variants, and more than 3 million utterances. This variation-dense design enables robust speaker-conditioned anti-spoofing and fine-grained synthesis-source tracing. We further position this dataset as both a practical reference training resource and a benchmark evaluation suite for anti-spoofing and source tracing.

Method: VoxCog is an end-to-end framework that uses pre-trained speech foundation models with dialect classifier initialization, processing speech directly without relying on text, images, or other modalities.

Result: Achieves 87.5% accuracy on ADReSS 2020 and 85.9% on ADReSSo 2021 test sets, outperforming previous multimodal ensemble approaches and LLM-based solutions.

Conclusion: Dialect-aware speech foundation models provide an effective, modality-efficient approach for cognitive impairment detection, demonstrating that speech patterns in AD/MCI can be effectively modeled as dialect-like variations.

Abstract: In this work, we present a novel perspective on cognitive impairment classification from speech by integrating speech foundation models that explicitly recognize speech dialects. Our motivation is based on the observation that individuals with Alzheimer’s Disease (AD) or mild cognitive impairment (MCI) often produce measurable speech characteristics, such as slower articulation rate and lengthened sounds, in a manner similar to dialectal phonetic variations seen in speech. Building on this idea, we introduce VoxCog, an end-to-end framework that uses pre-trained dialect models to detect AD or MCI without relying on additional modalities such as text or images. Through experiments on multiple multilingual datasets for AD and MCI detection, we demonstrate that model initialization with a dialect classifier on top of speech foundation models consistently improves the predictive performance of AD or MCI. Our trained models yield similar or often better performance compared to previous approaches that ensembled several computational methods using different signal modalities. Particularly, our end-to-end speech-based model achieves 87.5% and 85.9% accuracy on the ADReSS 2020 challenge and ADReSSo 2021 challenge test sets, outperforming existing solutions that use multimodal ensemble-based computation or LLMs.

Method: Uses masked diffusion framework that progressively unmasks positions, allowing arbitrary decoding orders during both training and inference. Compares fixed strategies (left-to-right, right-to-left) with adaptive strategies (Top-K). Also explores quantization of acoustic representations, testing down to 1-bit quantization.

Result: Randomness in decoding order affects speech quality. Fixed-order decoding (including the standard left-to-right approach) is suboptimal, while adaptive decoding yields better performance. Even 1-bit quantization of acoustic representations can support reasonably high-quality speech synthesis.

Conclusion: Decoding order matters for speech synthesis quality, and adaptive decoding strategies outperform traditional fixed-order approaches. The masked diffusion framework provides flexibility for exploring different decoding orders, and surprisingly low-bit quantization remains viable for speech synthesis.

Abstract: Autoregressive speech synthesis often adopts a left-to-right order, yet generation order is a modelling choice. We investigate decoding order through masked diffusion framework, which progressively unmasks positions and allows arbitrary decoding orders during training and inference. By interpolating between identity and random permutations, we show that randomness in decoding order affects speech quality. We further compare fixed strategies, such as \texttt{l2r} and \texttt{r2l} with adaptive ones, such as Top-$K$, finding that fixed-order decoding, including the dominating left-to-right approach, is suboptimal, while adaptive decoding yields better performance. Finally, since masked diffusion requires discrete inputs, we quantise acoustic representations and find that even 1-bit quantisation can support reasonably high-quality speech.

Method: Two-part approach: (1) ACEval evaluation framework with advanced LALM- and ASR-based solvers to test existing CAPTCHAs, (2) IllusionAudio CAPTCHA that leverages audio illusions based on human auditory perception mechanisms.

Result: Most existing audio CAPTCHAs can be solved with high success rates by advanced LALMs and ASR models. IllusionAudio defeats all tested AI attacks while achieving 100% human pass rate, significantly outperforming existing methods.

Conclusion: Current audio CAPTCHAs are vulnerable to modern AI attacks, but IllusionAudio provides a robust solution using audio illusions that exploit differences between human and machine auditory perception.

Abstract: CAPTCHAs are widely used by websites to block bots and spam by presenting challenges that are easy for humans but difficult for automated programs to solve. To improve accessibility, audio CAPTCHAs are designed to complement visual ones. However, the robustness of audio CAPTCHAs against advanced Large Audio Language Models (LALMs) and Automatic Speech Recognition (ASR) models remains unclear. In this paper, we introduce AI-CAPTCHA, a unified framework that offers (i) an evaluation framework, ACEval, which includes advanced LALM- and ASR-based solvers, and (ii) a novel audio CAPTCHA approach, IllusionAudio, leveraging audio illusions. Through extensive evaluations of seven widely deployed audio CAPTCHAs, we show that most existing methods can be solved with high success rates by advanced LALMs and ASR models, exposing critical security weaknesses. To address these vulnerabilities, we design a new audio CAPTCHA approach, IllusionAudio, which exploits perceptual illusion cues rooted in human auditory mechanisms. Extensive experiments demonstrate that our method defeats all tested LALM- and ASR-based attacks while achieving a 100% human pass rate, significantly outperforming existing audio CAPTCHA methods.

Method: Developed a scalable automatic annotation framework that uses a unified detection model to identify NVs in natural speech and integrates them with transcripts via temporal-semantic alignment. Applied this to create NonVerbalSpeech-38K dataset from in-the-wild media.

Result: Created and released NonVerbalSpeech-38K dataset with 38,718 samples across 10 NV categories. Experimental results show the dataset provides superior controllability for NVs generation and achieves comparable performance for NVs understanding tasks.

Conclusion: The proposed framework enables low-cost, scalable, and natural acquisition of non-verbal vocalizations, addressing limitations of existing methods. The resulting dataset advances research in emotionally intelligent speech systems by providing diverse, real-world NV data.

Abstract: Non-verbal Vocalizations (NVs), such as laughter and sighs, are vital for conveying emotion and intention in human speech, yet most existing speech systems neglect them, which severely compromises communicative richness and emotional intelligence. Existing methods for NVs acquisition are either costly and unscalable (relying on manual annotation/recording) or unnatural (relying on rule-based synthesis). To address these limitations, we propose a highly scalable automatic annotation framework to label non-verbal phenomena from natural speech, which is low-cost, easily extendable, and inherently diverse and natural. This framework leverages a unified detection model to accurately identify NVs in natural speech and integrates them with transcripts via temporal-semantic alignment method. Using this framework, we created and released \textbf{NonVerbalSpeech-38K}, a diverse, real-world dataset featuring 38,718 samples across 10 NV categories collected from in-the-wild media. Experimental results demonstrate that our dataset provides superior controllability for NVs generation and achieves comparable performance for NVs understanding.

Method: Two-fold approach: 1) Created AHEAD-DS dataset by repackaging and refining several open source datasets with standardized, hearing-aid-relevant labels; 2) Developed OpenYAMNet, a sound recognition model optimized for edge deployment on smartphones connected to hearing devices.

Result: OpenYAMNet achieved mean average precision of 0.86 and accuracy of 0.93 on AHEAD-DS testing set across 14 auditory scene categories. Real-time deployment on Android smartphone (Google Pixel 3) showed ~50ms model loading latency and ~30ms processing time per second of audio.

Conclusion: The proposed AHEAD-DS dataset and OpenYAMNet model provide standardized tools for auditory scene recognition research and practical deployment on edge devices, enabling better hearing aid functionality through real-time sound-based scene recognition.

Abstract: Scene recognition is important for hearing devices, however; this is challenging, in part because of the limitations of existing datasets. Datasets often lack public accessibility, completeness, or audiologically relevant labels, hindering systematic comparison of machine learning models. Deploying such models on resource-constrained edge devices presents another challenge.The proposed solution is two-fold, a repack and refinement of several open source datasets to create AHEAD-DS, a dataset designed for auditory scene recognition for hearing devices, and introduce OpenYAMNet, a sound recognition model. AHEAD-DS aims to provide a standardised, publicly available dataset with consistent labels relevant to hearing aids, facilitating model comparison. OpenYAMNet is designed for deployment on edge devices like smartphones connected to hearing devices, such as hearing aids and wireless earphones with hearing aid functionality, serving as a baseline model for sound-based scene recognition. OpenYAMNet achieved a mean average precision of 0.86 and accuracy of 0.93 on the testing set of AHEAD-DS across fourteen categories relevant to auditory scene recognition. Real-time sound-based scene recognition capabilities were demonstrated on edge devices by deploying OpenYAMNet to an Android smartphone. Even with a 2018 Google Pixel 3, a phone with modest specifications, the model processes audio with approximately 50ms of latency to load the model, and an approximate linear increase of 30ms per 1 second of audio. The project website with links to code, data, and models. https://github.com/Australian-Future-Hearing-Initiative

Method: CALM uses a large Transformer backbone to produce contextual embeddings at each timestep, which condition an MLP that generates continuous audio frames through consistency modeling, avoiding discrete tokenization.

Result: CALM achieves higher audio quality at lower computational cost than state-of-the-art discrete audio language models, with improved efficiency and fidelity for both speech and music generation.

Conclusion: Continuous audio modeling provides a more efficient alternative to discrete token approaches, enabling lightweight, high-quality audio generation, as demonstrated by the open-source Pocket TTS model that runs faster than real-time on laptop CPUs.

Abstract: Audio Language Models (ALM) have emerged as the dominant paradigm for speech and music generation by representing audio as sequences of discrete tokens. Yet, unlike text tokens, which are invertible, audio tokens are extracted from lossy codecs with a limited bitrate. As a consequence, increasing audio quality requires generating more tokens, which imposes a trade-off between fidelity and computational cost. We address this issue by studying Continuous Audio Language Models (CALM). These models instantiate a large Transformer backbone that produces a contextual embedding at every timestep. This sequential information then conditions an MLP that generates the next continuous frame of an audio VAE through consistency modeling. By avoiding lossy compression, CALM achieves higher quality at lower computational cost than their discrete counterpart. Experiments on speech and music demonstrate improved efficiency and fidelity over state-of-the-art discrete audio language models, facilitating lightweight, high-quality audio generation. Samples are available at hf.co/spaces/kyutai/calm-samples. Finally, we release Pocket TTS, an open-source 100M-parameter text-to-speech model that can run faster than real time on a laptop CPU: github.com/kyutai-labs/pocket-tts.

Method: Extends DeepFilterNet2 architecture with dual-path blocks in the encoder to strengthen long-range temporal and cross-band modeling. Uses a loss component to mitigate over-attenuation and includes a fine-tuning phase for “always-on” applications. Evaluates on a new dataset of long, low-SNR recordings across 12 languages and everyday noise scenarios.

Result: DPDFNet outperforms other causal open-source models on the new evaluation set, including larger and more computationally demanding models. The proposed PRISM metric shows scalability with dual-path blocks. DPDFNet-4 achieves real-time performance on Ceva-NeuPro-Nano edge NPUs (NPN32 and NPN64), demonstrating state-of-the-art quality within embedded constraints.

Conclusion: DPDFNet successfully combines architectural improvements (dual-path blocks), specialized training techniques, and real-world evaluation to deliver superior causal speech enhancement that maintains high quality while being deployable on resource-constrained edge devices for “always-on” applications.

Abstract: We present DPDFNet, a causal single-channel speech enhancement model that extends DeepFilterNet2 architecture with dual-path blocks in the encoder, strengthening long-range temporal and cross-band modeling while preserving the original enhancement framework. In addition, we demonstrate that adding a loss component to mitigate over-attenuation in the enhanced speech, combined with a fine-tuning phase tailored for “always-on” applications, leads to substantial improvements in overall model performance. To compare our proposed architecture with a variety of causal open-source models, we created a new evaluation set comprising long, low-SNR recordings in 12 languages across everyday noise scenarios, better reflecting real-world conditions than commonly used benchmarks. On this evaluation set, DPDFNet delivers superior performance to other causal open-source models, including some that are substantially larger and more computationally demanding. We also propose an holistic metric named PRISM, a composite, scale-normalized aggregate of intrusive and non-intrusive metrics, which demonstrates clear scalability with the number of dual-path blocks. We further demonstrate on-device feasibility by deploying DPDFNet on Ceva-NeuPro-Nano edge NPUs. Results indicate that DPDFNet-4, our second-largest model, achieves real-time performance on NPN32 and runs even faster on NPN64, confirming that state-of-the-art quality can be sustained within strict embedded power and latency constraints.

Method: Two-stage approach: (1) remove largest-magnitude weights into sparse matrix S, (2) apply recursive Hierarchically Sparse Separable low-rank factorization to residual matrix, using rank-reduction strategy and RCM permutation for diagonal alignment.

Result: On LLaMA-7B, compressing only self-attention projections (1.6B params) yields large memory savings while maintaining comparable perplexity (1.64 on WikiText with 30% sparsity, rank 512), outperforming dense baselines and sparse-plus-SVD variants.

Conclusion: HSS compression is hardware-friendly (sparse + thin-matrix multiplications), trainable end-to-end, and provides effective memory-efficient compression for LLMs while preserving performance.

Abstract: Modern large language models (LLMs) place extraordinary pressure on memory and compute budgets, making principled compression indispensable for both deployment and continued training. We present Hierarchical Sparse Plus Low-Rank (HSS) compression, a two-stage scheme that (i) removes the largest-magnitude weights into a sparse matrix S and (ii) applies a recursive Hierarchically Sparse Separable (HSS) low-rank factorisation to the dense residual matrix. A recursive rank-reducing strategy and a reverse Cuthill-Mckee (RCM) permutation are introduced to align high weights towards the diagonal with the block-diagonal hierarchy, maximising off-diagonal compressibility (because they are touched only once). HSS is hardware-friendly: its matrix-vector multiply reduces to one sparse and a sequence of thin-matrix multiplications and can be trained end-to-end with standard optimisers. Experiments on LLaMA-7B show that targeting only the self-attention projections (1.6 B parameters of Q, K, and V matrices out of a total 7B parameters) suffices to yield large memory savings while retaining comparable state-of-the-art perplexity scores on test samples of the WikiText dataset. For example, with a 30% sparsity budget and an outer rank of 512, sHSS-RCM achieves a perplexity of 1.64, outperforming dense baselines and classical sparse-plus-SVD variants, while also achieving significant memory savings.

Method: The study conducts theoretical and empirical analysis of regularisation-based CL methods (EWC, SI, MAS) on EEG emotion classification using DREAMER and SEED datasets. It investigates subject-incremental learning sequences and analyzes limitations through gradient interference analysis, importance estimation reliability, and forward/backward transfer evaluation.

Result: Regularisation-based CL methods show limited performance: (1) parameter importance heuristics become unreliable under noisy EEG data and covariate shift, (2) gradients on important parameters interfere with new subject adaptation, (3) accumulated importance values over-constrain models, (4) performance is sensitive to subject order, and (5) forward transfer shows no significant improvement over sequential fine-tuning.

Conclusion: Regularisation-based continual learning approaches are limited for robust generalization to unseen subjects in EEG-based emotion classification due to fundamental misalignment in stability-plasticity trade-off and the high variability of EEG signals that makes past subjects provide limited value to future subjects.

Abstract: Generalisation to unseen subjects in EEG-based emotion classification remains a challenge due to high inter-and intra-subject variability. Continual learning (CL) poses a promising solution by learning from a sequence of tasks while mitigating catastrophic forgetting. Regularisation-based CL approaches, such as Elastic Weight Consolidation (EWC), Synaptic Intelligence (SI), and Memory Aware Synapses (MAS), are commonly used as baselines in EEG-based CL studies, yet their suitability for this problem remains underexplored. This study theoretically and empirically finds that regularisation-based CL methods show limited performance for EEG-based emotion classification on the DREAMER and SEED datasets. We identify a fundamental misalignment in the stability-plasticity trade-off, where regularisation-based methods prioritise mitigating catastrophic forgetting (backward transfer) over adapting to new subjects (forward transfer). We investigate this limitation under subject-incremental sequences and observe that: (1) the heuristics for estimating parameter importance become less reliable under noisy data and covariate shift, (2) gradients on parameters deemed important by these heuristics often interfere with gradient updates required for new subjects, moving optimisation away from the minimum, (3) importance values accumulated across tasks over-constrain the model, and (4) performance is sensitive to subject order. Forward transfer showed no statistically significant improvement over sequential fine-tuning (p > 0.05 across approaches and datasets). The high variability of EEG signals means past subjects provide limited value to future subjects. Regularisation-based continual learning approaches are therefore limited for robust generalisation to unseen subjects in EEG-based emotion classification.

Method: RewriteNets use learnable string rewriting rules in each layer: (1) fuzzy pattern matching, (2) differentiable conflict resolution via Gumbel-Sinkhorn assignment, (3) rule application with variable-length replacements, and (4) untouched token propagation. End-to-end training enabled by straight-through Gumbel-Sinkhorn estimator.

Result: Achieves 98.7% accuracy on SCAN benchmark’s length split, excels at systematic generalization tasks, and is computationally more efficient than Transformers. Analysis shows learned rules are interpretable.

Conclusion: RewriteNets present a promising direction for sequence modeling with explicit structural inductive biases, offering better systematic generalization and computational efficiency compared to dominant attention-based models.

Abstract: Dominant sequence models like the Transformer represent structure implicitly through dense attention weights, incurring quadratic complexity. We propose RewriteNets, a novel neural architecture built on an alternative paradigm: explicit, parallel string rewriting. Each layer in a RewriteNet contains a set of learnable rules. For each position in an input sequence, the layer performs four operations: (1) fuzzy matching of rule patterns, (2) conflict resolution via a differentiable assignment operator to select non-overlapping rewrites, (3) application of the chosen rules to replace input segments with output segments of potentially different lengths, and (4) propagation of untouched tokens. While the discrete assignment of rules is non-differentiable, we employ a straight-through Gumbel-Sinkhorn estimator, enabling stable end-to-end training. We evaluate RewriteNets on algorithmic, compositional, and string manipulation tasks, comparing them against strong LSTM and Transformer baselines. Results show that RewriteNets excel at tasks requiring systematic generalization (achieving 98.7% accuracy on the SCAN benchmark’s length split) and are computationally more efficient than Transformers. We also provide an analysis of learned rules and an extensive ablation study, demonstrating that this architecture presents a promising direction for sequence modeling with explicit structural inductive biases.

Method: Introduces HOSC activation: tanh(βsin(ω₀x)), which exposes explicit parameter β to control Lipschitz bound (βω₀), providing direct mechanism to tune gradient magnitudes while retaining periodic carrier.

Result: Comprehensive empirical study across images, audio, video, NeRFs, and SDFs shows HOSC provides substantial benefits or competitive parity against SIREN, FINER, and related methods.

Conclusion: HOSC established as practical periodic activation for INR applications with domain-specific hyperparameter guidance, offering improved gradient stability and control.

Abstract: Periodic activations such as sine preserve high-frequency information in implicit neural representations (INRs) through their oscillatory structure, but often suffer from gradient instability and limited control over multi-scale behavior. We introduce the Hyperbolic Oscillator with Saturation Control (HOSC) activation, $\text{HOSC}(x) = \tanh\bigl(β\sin(ω_0 x)\bigr)$, which exposes an explicit parameter $β$ that controls the Lipschitz bound of the activation by $βω_0$. This provides a direct mechanism to tune gradient magnitudes while retaining a periodic carrier. We provide a mathematical analysis and conduct a comprehensive empirical study across images, audio, video, NeRFs, and SDFs using standardized training protocols. Comparative analysis against SIREN, FINER, and related methods shows where HOSC provides substantial benefits and where it achieves competitive parity. Results establish HOSC as a practical periodic activation for INR applications, with domain-specific guidance on hyperparameter selection. For code visit the project page https://hosc-nn.github.io/ .

Method: Conducted first empirical comparative study of transfer learning for UATR using multiple pretrained audio models from diverse domains. Model weights were frozen, and embeddings were analyzed through classification, clustering, and similarity-based evaluations. Used linear probing to suppress recording-specific information and isolate ship-type features.

Result: Embedding space was dominated by recording-specific characteristics, but simple linear probes effectively suppressed this noise and isolated ship-type features. This enables effective automatic UATR using pretrained models at low computational cost.

Conclusion: Transfer learning with pretrained audio models and linear probing provides a practical solution for underwater acoustic target recognition, significantly reducing the need for large amounts of high-quality labeled ship recordings while maintaining effectiveness.

Abstract: Increasing levels of anthropogenic noise from ships contribute significantly to underwater sound pollution, posing risks to marine ecosystems. This makes monitoring crucial to understand and quantify the impact of the ship radiated noise. Passive Acoustic Monitoring (PAM) systems are widely deployed for this purpose, generating years of underwater recordings across diverse soundscapes. Manual analysis of such large-scale data is impractical, motivating the need for automated approaches based on machine learning. Recent advances in automatic Underwater Acoustic Target Recognition (UATR) have largely relied on supervised learning, which is constrained by the scarcity of labeled data. Transfer Learning (TL) offers a promising alternative to mitigate this limitation. In this work, we conduct the first empirical comparative study of transfer learning for UATR, evaluating multiple pretrained audio models originating from diverse audio domains. The pretrained model weights are frozen, and the resulting embeddings are analyzed through classification, clustering, and similarity-based evaluations. The analysis shows that the geometrical structure of the embedding space is largely dominated by recording-specific characteristics. However, a simple linear probe can effectively suppress this recording-specific information and isolate ship-type features from these embeddings. As a result, linear probing enables effective automatic UATR using pretrained audio models at low computational cost, significantly reducing the need for a large amounts of high-quality labeled ship recordings.

Method: Proposes Multiplicative Orthogonal Sequential Editing (MOSE) - a multiplicative paradigm where new knowledge is incorporated into an orthogonal matrix that multiplies the original parameter matrix. This maintains numerical stability since multiplying by orthogonal matrices doesn’t change matrix stability properties.

Result: MOSE effectively limits deviations in edited parameter matrices and maintains numerical stability. Achieves 12.08% improvement in sequential editing performance while retaining 95.73% of general abilities across downstream tasks compared to current methods.

Conclusion: The multiplicative orthogonal editing paradigm fundamentally addresses the numerical stability limitations of additive editing methods, enabling better sequential editing performance while preserving general model capabilities.

Abstract: Knowledge editing aims to efficiently modify the internal knowledge of large language models (LLMs) without compromising their other capabilities. The prevailing editing paradigm, which appends an update matrix to the original parameter matrix, has been shown by some studies to damage key numerical stability indicators (such as condition number and norm), thereby reducing editing performance and general abilities, especially in sequential editing scenario. Although subsequent methods have made some improvements, they remain within the additive framework and have not fundamentally addressed this limitation. To solve this problem, we analyze it from both statistical and mathematical perspectives and conclude that multiplying the original matrix by an orthogonal matrix does not change the numerical stability of the matrix. Inspired by this, different from the previous additive editing paradigm, a multiplicative editing paradigm termed Multiplicative Orthogonal Sequential Editing (MOSE) is proposed. Specifically, we first derive the matrix update in the multiplicative form, the new knowledge is then incorporated into an orthogonal matrix, which is multiplied by the original parameter matrix. In this way, the numerical stability of the edited matrix is unchanged, thereby maintaining editing performance and general abilities. We compared MOSE with several current knowledge editing methods, systematically evaluating their impact on both editing performance and the general abilities across three different LLMs. Experimental results show that MOSE effectively limits deviations in the edited parameter matrix and maintains its numerical stability. Compared to current methods, MOSE achieves a 12.08% improvement in sequential editing performance, while retaining 95.73% of general abilities across downstream tasks. The code is available at https://github.com/famoustourist/MOSE.

Method: Formulates the problem as a non-stationary multi-armed bandit and proposes AGING BANDIT WITH ADAPTIVE RESET algorithm that combines adaptive windowing with periodic monitoring to track evolving reward distributions in the face of abrupt and gradual changes.

Result: Establishes theoretical performance guarantees showing near-optimal performance and validates results through simulations, demonstrating effectiveness in handling the challenging decentralized setting with coupled, history-dependent reward processes.

Conclusion: The proposed algorithm effectively addresses the challenges of decentralized collaborative requesting in edge networks by providing a solution that adapts to both abrupt and gradual changes in reward distributions while achieving near-optimal performance in optimizing information freshness.

Abstract: We study a decentralized collaborative requesting problem that aims to optimize the information freshness of time-sensitive clients in edge networks consisting of multiple clients, access nodes (ANs), and servers. Clients request content through ANs acting as gateways, without observing AN states or the actions of other clients. We define the reward as the age of information reduction resulting from a client’s selection of an AN, and formulate the problem as a non-stationary multi-armed bandit. In this decentralized and partially observable setting, the resulting reward process is history-dependent and coupled across clients, and exhibits both abrupt and gradual changes in expected rewards, rendering classical bandit-based approaches ineffective. To address these challenges, we propose the AGING BANDIT WITH ADAPTIVE RESET algorithm, which combines adaptive windowing with periodic monitoring to track evolving reward distributions. We establish theoretical performance guarantees showing that the proposed algorithm achieves near-optimal performance, and we validate the theoretical results through simulations.

Method: Dual-mode architecture with streamlined fast path; integrates adaptive moment estimation, rectified learning-rate scheduling, decoupled weight regularization, multiple Nesterov momentum variants, and memory-efficient lookahead (O(p + p/k) overhead). Uses custom CUDA kernels for 3-5x speedup while maintaining numerical stability.

Result: Outperforms 14 contemporary optimizers (Adam, AdamW, RAdam, Lion, Adan) on CIFAR-10, CIFAR-100, ImageNet, ImageNette with ResNet-50, VGG-16, ViT architectures. Achieves SOTA accuracy and exceptional robustness, particularly excelling on VGG-16/ImageNette, demonstrating superior architectural robustness.

Conclusion: NOVAK’s architectural innovations (rectification, decoupled decay, hybrid momentum) are crucial for reliable training of deep plain networks, addressing a long-standing limitation of existing adaptive optimization methods, while providing theoretical convergence guarantees and practical efficiency.

Abstract: This work introduces NOVAK, a modular gradient-based optimization algorithm that integrates adaptive moment estimation, rectified learning-rate scheduling, decoupled weight regularization, multiple variants of Nesterov momentum, and lookahead synchronization into a unified, performance-oriented framework. NOVAK adopts a dual-mode architecture consisting of a streamlined fast path designed for production. The optimizer employs custom CUDA kernels that deliver substantial speedups (3-5 for critical operations) while preserving numerical stability under standard stochastic-optimization assumptions. We provide fully developed mathematical formulations for rectified adaptive learning rates, a memory-efficient lookahead mechanism that reduces overhead from O(2p) to O(p + p/k), and the synergistic coupling of complementary optimization components. Theoretical analysis establishes convergence guarantees and elucidates the stability and variance-reduction properties of the method. Extensive empirical evaluation on CIFAR-10, CIFAR-100, ImageNet, and ImageNette demonstrates NOVAK superiority over 14 contemporary optimizers, including Adam, AdamW, RAdam, Lion, and Adan. Across architectures such as ResNet-50, VGG-16, and ViT, NOVAK consistently achieves state-of-the-art accuracy, and exceptional robustness, attaining very high accuracy on VGG-16/ImageNette demonstrating superior architectural robustness compared to contemporary optimizers. The results highlight that NOVAKs architectural contributions (particularly rectification, decoupled decay, and hybrid momentum) are crucial for reliable training of deep plain networks lacking skip connections, addressing a long-standing limitation of existing adaptive optimization methods.

Method: E^2-LLM integrates a pretrained EEG encoder with Qwen-based LLMs via learnable projection layers. Uses multi-stage training: emotion-discriminative pretraining, cross-modal alignment, and instruction tuning with chain-of-thought reasoning. Comprehensive evaluation protocol covers basic prediction, multi-task reasoning, and zero-shot scenario understanding.

Result: Achieves excellent performance on emotion classification across seven emotion categories. Larger variants show enhanced reliability and superior zero-shot generalization to complex reasoning scenarios. Model scaling improves both recognition accuracy and interpretable emotional understanding.

Conclusion: Establishes a new paradigm combining physiological signals with LLM reasoning capabilities. Demonstrates that model scaling enhances both accuracy and interpretable emotional understanding in affective computing, addressing key limitations in EEG-based emotion recognition.

Abstract: Emotion recognition from electroencephalography (EEG) signals remains challenging due to high inter-subject variability, limited labeled data, and the lack of interpretable reasoning in existing approaches. While recent multimodal large language models (MLLMs) have advanced emotion analysis, they have not been adapted to handle the unique spatiotemporal characteristics of neural signals. We present E^2-LLM (EEG-to-Emotion Large Language Model), the first MLLM framework for interpretable emotion analysis from EEG. E^2-LLM integrates a pretrained EEG encoder with Qwen-based LLMs through learnable projection layers, employing a multi-stage training pipeline that encompasses emotion-discriminative pretraining, cross-modal alignment, and instruction tuning with chain-of-thought reasoning. We design a comprehensive evaluation protocol covering basic emotion prediction, multi-task reasoning, and zero-shot scenario understanding. Experiments on the dataset across seven emotion categories demonstrate that E^2-LLM achieves excellent performance on emotion classification, with larger variants showing enhanced reliability and superior zero-shot generalization to complex reasoning scenarios. Our work establishes a new paradigm combining physiological signals with LLM reasoning capabilities, showing that model scaling improves both recognition accuracy and interpretable emotional understanding in affective computing.

Method: Proposes PRINCE mechanism with: 1) bias-resilient global SFL model aggregation to eliminate model biases, 2) rigorous SFL convergence bound to evaluate device contributions, and 3) modeling inter-tenant device competition as congestion game for Stackelberg equilibrium analysis to derive optimal incentive strategies.

Result: Extensive simulations with four SFL tenant types (ViT, BERT, Whisper, LLaMA) across text, images, and audio show PRINCE accelerates FM fine-tuning by up to 3.07x compared to state-of-the-art approaches while consistently meeting performance targets.

Conclusion: PRINCE effectively addresses the multi-tenant SFL incentive problem, enabling efficient coordination of self-interested devices across diverse foundation model fine-tuning tasks through strategic price incentives and game-theoretic analysis.

Abstract: Foundation models (FMs) such as GPT-4 exhibit exceptional generative capabilities across diverse downstream tasks through fine-tuning. Split Federated Learning (SFL) facilitates privacy-preserving FM fine-tuning on resource-constrained local devices by offloading partial FM computations to edge servers, enabling device-edge synergistic fine-tuning. Practical edge networks often host multiple SFL tenants to support diversified downstream tasks. However, existing research primarily focuses on single-tenant SFL scenarios, and lacks tailored incentive mechanisms for multi-tenant settings, which are essential to effectively coordinate self-interested local devices for participation in various downstream tasks, ensuring that each SFL tenant’s distinct FM fine-tuning requirements (e.g., FM types, performance targets, and fine-tuning deadlines) are met. To address this gap, we propose a novel Price-Incentive Mechanism (PRINCE) that guides multiple SFL tenants to offer strategic price incentives, which solicit high-quality device participation for efficient FM fine-tuning. Specifically, we first develop a bias-resilient global SFL model aggregation scheme to eliminate model biases caused by independent device participation. We then derive a rigorous SFL convergence bound to evaluate the contributions of heterogeneous devices to FM performance improvements, guiding the incentive strategies of SFL tenants. Furthermore, we model inter-tenant device competition as a congestion game for Stackelberg equilibrium (SE) analysis, deriving each SFL tenant’s optimal incentive strategy. Extensive simulations involving four representative SFL tenant types (ViT, BERT, Whisper, and LLaMA) across diverse data modalities (text, images, and audio) demonstrate that PRINCE accelerates FM fine-tuning by up to 3.07x compared to state-of-the-art approaches, while consistently meeting fine-tuning performance targets.

Method: Introduces a sliced Wasserstein loss function for distribution-aware calibration in ultra-low-bit post-training quantization. The loss aligns output distributions of full-precision and quantized models under random linear projections, complementing standard MSE loss without adding computational overhead during inference.

Result: The proposed loss consistently improves perplexity and downstream task accuracy across multiple ultra-low-bit settings. It recovers 4.12-20.37% of OmniQuant’s lost accuracy on LLaMA-2-7B, 0.93-7.65% on OPT-6.7B, and 2.26-6.20% on LLaMA-2-13B. TesseraQ’s accuracy degradation is recovered by 3.63-7.63%.

Conclusion: Distributional alignment through sliced Wasserstein loss provides a simple yet effective performance boost that pushes the limits of frontier quantization methods, enabling more efficient deployment of large language models with reduced accuracy degradation.

Abstract: The benefits of most large language models come with steep and often hidden economic and environmental costs due to their resource usage inefficiency during deployment. Model quantization improves energy and memory efficiency through representing model parameters by lower-precision values. However, compression below 4-bits often distorts activation distributions and degrades performance. We address this challenge by introducing a sliced Wasserstein loss function for distribution-aware calibration in ultra-low-bit post-training quantization. The proposed loss aligns the output distributions of full-precision and quantized models under random linear projections, complementing standard mean-squared error loss without adding any computational overhead during inference. Our proposed loss function can be incorporated with any post-training quantization framework that has a retraining component. We demonstrate the performance gains of our proposed model by incorporating it with two frontier methods known as OmniQuant and TesseraQ. Compared to these two baselines, the proposed loss consistently improves both perplexity and downstream task accuracy across multiple ultra-low-bit settings. Our proposed loss function recovers 4.12-20.37% of the OmniQuant’s lost accuracy on the language model LLaMA-2-7B, 0.93-7.65% on OPT-6.7B, and 2.26-6.20% on LLaMA-2-13B. TesseraQ’s accuracy degradation is recovered by 3.63-7.63% in relative terms when augmented by our proposed loss function. Taken together, these results demonstrate that distributional alignment provides a simple yet effective performance boost that can push the limits of frontier quantization methods. Our method is available on GitHub to facilitate future progress in ultra-low-bit quantization.

Method: Develop multivariate max-min neural network operators activated by sigmoidal functions, analyze them using approximation theory, and establish convergence properties.

Result: Proved pointwise and uniform convergence theorems, derived quantitative approximation error bounds using modulus of continuity and multivariate generalized absolute moment.

Conclusion: Multivariate max-min operators offer efficient, stable approximation tools with both theoretical elegance and practical applications in neural network approximation.

Abstract: In this paper, we develop a multivariate framework for approximation by max-min neural network operators. Building on the recent advances in approximation theory by neural network operators, particularly, the univariate max-min operators, we propose and analyze new multivariate operators activated by sigmoidal functions. We establish pointwise and uniform convergence theorems and derive quantitative estimates for the order of approximation via modulus of continuity and multivariate generalized absolute moment. Our results demonstrate that multivariate max-min structure of operators, besides their algebraic elegance, provide efficient and stable approximation tools in both theoretical and applied settings.

Method: KVzap is a fast, input-adaptive approximation of KVzip that works in both prefilling and decoding phases of transformer inference.

Result: Achieves 2-4x KV cache compression on Qwen3-8B, Llama-3.1-8B-Instruct, and Qwen3-32B across long-context and reasoning tasks with negligible accuracy loss. State-of-the-art performance on KVpress leaderboard.

Conclusion: KVzap provides an effective solution to KV cache bottleneck with practical speed-accuracy balance, making it suitable for adoption in major inference engines.

Abstract: Growing context lengths in transformer-based language models have made the key-value (KV) cache a critical inference bottleneck. While many KV cache pruning methods have been proposed, they have not yet been adopted in major inference engines due to speed–accuracy trade-offs. We introduce KVzap, a fast, input-adaptive approximation of KVzip that works in both prefilling and decoding. On Qwen3-8B, Llama-3.1-8B-Instruct, and Qwen3-32B across long-context and reasoning tasks, KVzap achieves $2$–$4\times$ KV cache compression with negligible accuracy loss and achieves state-of-the-art performance on the KVpress leaderboard. Code and models are available at https://github.com/NVIDIA/kvpress.

Method: Sherry introduces 3:4 fine-grained sparsity that packs blocks of four weights into five bits, achieving regularized 1.25-bit width with power-of-two alignment. It also proposes Arenas, an annealing residual synapse mechanism that prevents weight trapping and maintains representational diversity during training.

Result: On LLaMA-3.2 across five benchmarks, Sherry matches state-of-the-art ternary performance while significantly reducing model size. The 1B model achieves zero accuracy loss compared to SOTA baselines with 25% bit savings and 10% speedup on an Intel i7-14700HX CPU.

Conclusion: Sherry resolves the hardware efficiency vs. performance trade-off in ternary quantization through innovative 3:4 sparsity and Arenas mechanism, enabling efficient LLM deployment on resource-constrained edge devices without accuracy degradation.

Abstract: The deployment of Large Language Models (LLMs) on resource-constrained edge devices is increasingly hindered by prohibitive memory and computational requirements. While ternary quantization offers a compelling solution by reducing weights to {-1, 0, +1}, current implementations suffer from a fundamental misalignment with commodity hardware. Most existing methods must choose between 2-bit aligned packing, which incurs significant bit wastage, or 1.67-bit irregular packing, which degrades inference speed. To resolve this tension, we propose Sherry, a hardware-efficient ternary quantization framework. Sherry introduces a 3:4 fine-grained sparsity that achieves a regularized 1.25-bit width by packing blocks of four weights into five bits, restoring power-of-two alignment. Furthermore, we identify weight trapping issue in sparse ternary training, which leads to representational collapse. To address this, Sherry introduces Arenas, an annealing residual synapse mechanism that maintains representational diversity during training. Empirical evaluations on LLaMA-3.2 across five benchmarks demonstrate that Sherry matches state-of-the-art ternary performance while significantly reducing model size. Notably, on an Intel i7-14700HX CPU, our 1B model achieves zero accuracy loss compared to SOTA baselines while providing 25% bit savings and 10% speed up. The code is available at https://github.com/Tencent/AngelSlim .

Method: The paper analyzes attention behaviors in MDMs through empirical investigation, revealing the phenomenon of Attention Floating. It examines how attention patterns differ from ARMs across denoising steps and layers, identifying shallow structure-aware and deep content-focused attention mechanisms.

Result: MDMs exhibit dynamic, dispersed attention anchors (Attention Floating) that shift across denoising steps and layers, unlike ARMs’ fixed attention sinks. Shallow layers use floating tokens to build global structure, while deeper layers focus on semantic content. This explains MDMs’ ability to double performance compared to ARMs in knowledge-intensive tasks.

Conclusion: The distinctive Attention Floating pattern in MDMs provides a mechanistic explanation for their strong in-context learning capabilities. This attention mechanism allows MDMs to outperform ARMs significantly in knowledge-intensive tasks, offering insights into why MDMs are narrowing the performance gap with ARMs.

Abstract: Masked diffusion models (MDMs), which leverage bidirectional attention and a denoising process, are narrowing the performance gap with autoregressive models (ARMs). However, their internal attention mechanisms remain under-explored. This paper investigates the attention behaviors in MDMs, revealing the phenomenon of Attention Floating. Unlike ARMs, where attention converges to a fixed sink, MDMs exhibit dynamic, dispersed attention anchors that shift across denoising steps and layers. Further analysis reveals its Shallow Structure-Aware, Deep Content-Focused attention mechanism: shallow layers utilize floating tokens to build a global structural framework, while deeper layers allocate more capability toward capturing semantic content. Empirically, this distinctive attention pattern provides a mechanistic explanation for the strong in-context learning capabilities of MDMs, allowing them to double the performance compared to ARMs in knowledge-intensive tasks. All codes and datasets are available at https://github.com/NEUIR/Attention-Floating.

Method: Introduces LLM-DAL (Large Language Model - Decompositional Algorithmic Learning), a specialized supervised training approach focused on reasoning decomposition to teach LLMs algorithmic execution.

Result: Demonstrates that LLMs’ ability to perform complex algorithmic inferences and generalize can be significantly improved with properly designed training that guides the learning process.

Conclusion: LLMs can be extended to execute algorithms effectively through decomposition-focused training methods like LLM-DAL, overcoming their current limitations in algorithmic reasoning.

Abstract: Large Language Models (LLMs) have recently developed new advanced functionalities. Their effectiveness relies on statistical learning and generalization capabilities. However, they face limitations in internalizing the data they process and struggle, for instance, to autonomously execute algorithms. In this paper, we investigate the possibility of extending these models’ capabilities to algorithm execution through specialized supervised training focused on reasoning decomposition. We introduce a training model called LLM-DAL (Large Language Model - Decompositional Algorithmic Learning), through which we demonstrate that LLMs’ ability to perform complex algorithmic inferences and generalize can be significantly improved when the training method is properly designed to guide the model in its learning process.

Method: LVICL uses vector-injected in-context learning with a learnable context vector adapter. It extracts a compressed context vector from multiple examples, then injects this vector into every layer of the frozen LLM during forward passes, eliciting the model’s in-context learning ability without increasing prompt length.

Result: Extensive experiments demonstrate the effectiveness of the approach. The method improves forecasting performance while keeping all LLM parameters frozen, reducing computational overhead compared to fine-tuning approaches.

Conclusion: LVICL successfully addresses the dual challenge of prediction performance and compute overhead in LLM-based time series forecasting by using vector-injected in-context learning, offering a practical solution for web-scale predictive applications.

Abstract: The World Wide Web needs reliable predictive capabilities to respond to changes in user behavior and usage patterns. Time series forecasting (TSF) is a key means to achieve this goal. In recent years, the large language models (LLMs) for TSF (LLM4TSF) have achieved good performance. However, there is a significant difference between pretraining corpora and time series data, making it hard to guarantee forecasting quality when directly applying LLMs to TSF; fine-tuning LLMs can mitigate this issue, but often incurs substantial computational overhead. Thus, LLM4TSF faces a dual challenge of prediction performance and compute overhead. To address this, we aim to explore a method for improving the forecasting performance of LLM4TSF while freezing all LLM parameters to reduce computational overhead. Inspired by in-context learning (ICL), we propose LVICL. LVICL uses our vector-injected ICL to inject example information into a frozen LLM, eliciting its in-context learning ability and thereby enhancing its performance on the example-related task (i.e., TSF). Specifically, we first use the LLM together with a learnable context vector adapter to extract a context vector from multiple examples adaptively. This vector contains compressed, example-related information. Subsequently, during the forward pass, we inject this vector into every layer of the LLM to improve forecasting performance. Compared with conventional ICL that adds examples into the prompt, our vector-injected ICL does not increase prompt length; moreover, adaptively deriving a context vector from examples suppresses components harmful to forecasting, thereby improving model performance. Extensive experiments demonstrate the effectiveness of our approach.

Method: A novel two-stage encoder-decoder transformer model trained on matched molecular pairs (MMPs) from ChEMBL, using SMIRKS-based representations to sequentially generate functional group removal and addition, ensuring strict substructure-level modifications.

Result: The method demonstrates ability to generate chemically valid transformations, explore diverse chemical spaces, and maintain scalability across varying search sizes, outperforming traditional approaches.

Conclusion: The two-stage transformer approach advances functional group replacement by combining the power of transformer models with structural similarity guarantees, enabling more effective design of novel chemical compounds with tailored properties.

Abstract: Functional group replacement is a pivotal approach in cheminformatics to enable the design of novel chemical compounds with tailored properties. Traditional methods for functional group removal and replacement often rely on rule-based heuristics, which can be limited in their ability to generate diverse and novel chemical structures. Recently, transformer-based models have shown promise in improving the accuracy and efficiency of molecular transformations, but existing approaches typically focus on single-step modeling, lacking the guarantee of structural similarity. In this work, we seek to advance the state of the art by developing a novel two-stage transformer model for functional group removal and replacement. Unlike one-shot approaches that generate entire molecules in a single pass, our method generates the functional group to be removed and appended sequentially, ensuring strict substructure-level modifications. Using a matched molecular pairs (MMPs) dataset derived from ChEMBL, we trained an encoder-decoder transformer model with SMIRKS-based representations to capture transformation rules effectively. Extensive evaluations demonstrate our method’s ability to generate chemically valid transformations, explore diverse chemical spaces, and maintain scalability across varying search sizes.

Method: Integrates Mixture-of-Experts (MoE) with Low-Rank Adaptation (LoRA) using asymmetric expert distribution (more experts in deeper layers for complex semantic abstractions) and a “Knowledge-Preservation Plugin” inspired by LoRA MoE to isolate general-purpose reasoning. Employs soft merging with adaptive routing and rank-wise decoupling.

Result: Consistently outperforms standard LoRA and conventional MoE architectures across multiple clinical NLP tasks while retaining the model’s general cognitive capabilities. Achieves superior performance in medical benchmarks while reducing interference.

Conclusion: Med-MoE-LoRA provides an effective framework for efficient multi-task domain adaptation in specialized fields like medicine, successfully addressing the stability-plasticity dilemma and task interference problems through innovative integration of MoE and LoRA techniques.

Abstract: The rapid evolution of Large Language Models (LLMs) has shifted focus from general-purpose capabilities to domain-specific expertise. However, adapting LLMs to specialized fields such as medicine presents two challenge: (1) the “Stability-Plasticity Dilemma”, where the model must acquire complex clinical knowledge without suffering from catastrophic forgetting of general world knowledge; and (2) “Task Interference”, where disparate sub-tasks, such as medical diagnosis, report summarization, and drug-drug interaction prediction, compete for limited low-rank parameter space. In this paper, we propose Med-MoE-LoRA, a novel framework that integrates Mixture-of-Experts (MoE) with Low-Rank Adaptation (LoRA) to enable efficient multi-task domain adaptation, especially for medical scenarios. Drawing inspiration from recent advances, our framework employs an asymmetric expert distribution where deeper layers are equipped with a higher density of LoRA experts to capture complex semantic abstractions. We further introduce a “Knowledge-Preservation Plugin”, inspired by LoRA MoE, to isolate and protect general-purpose reasoning. By utilizing soft merging with adaptive routing and rank-wise decoupling, Med-MoE-LoRA achieves superior performance in medical benchmarks while reducing interference. Experimental results demonstrate that our approach consistently outperforms standard LoRA and conventional MoE architectures across multiple clinical NLP tasks while retaining the model’s general cognitive capabilities.

Method: DiffCoder integrates a probabilistic diffusion model with a conventional convolutional ResNet encoder in an end-to-end framework. The encoder compresses flow fields into latent representations, while the diffusion model learns a generative prior over reconstructions conditioned on the compressed state.

Result: Under aggressive compression, DiffCoder significantly improves spectral accuracy while VAEs show substantial degradation. Both methods have comparable L2 reconstruction error, but DiffCoder better preserves underlying distributional structure. At moderate compression, sufficiently large VAEs remain competitive.

Conclusion: Diffusion-based priors provide greatest benefit under severe information bottlenecks, offering a promising path toward compact, statistically consistent representations of complex flow fields by recovering distributional and spectral properties critical for flow field statistics.

Abstract: Data-driven flow-field reconstruction typically relies on autoencoder architectures that compress high-dimensional states into low-dimensional latent representations. However, classical approaches such as variational autoencoders (VAEs) often struggle to preserve the higher-order statistical structure of fluid flows when subjected to strong compression. We propose DiffCoder, a coupled framework that integrates a probabilistic diffusion model with a conventional convolutional ResNet encoder and trains both components end-to-end. The encoder compresses the flow field into a latent representation, while the diffusion model learns a generative prior over reconstructions conditioned on the compressed state. This design allows DiffCoder to recover distributional and spectral properties that are not strictly required for minimizing pointwise reconstruction loss but are critical for faithfully representing statistical properties of the flow field. We evaluate DiffCoder and VAE baselines across multiple model sizes and compression ratios on a challenging dataset of Kolmogorov flow fields. Under aggressive compression, DiffCoder significantly improves the spectral accuracy while VAEs exhibit substantial degradation. Although both methods show comparable relative L2 reconstruction error, DiffCoder better preserves the underlying distributional structure of the flow. At moderate compression levels, sufficiently large VAEs remain competitive, suggesting that diffusion-based priors provide the greatest benefit when information bottlenecks are severe. These results demonstrate that the generative decoding by diffusion offers a promising path toward compact, statistically consistent representations of complex flow fields.

Method: Evaluated multiple RL-based neighborhood selection strategies: multi-armed bandits (upper confidence bound, ε-greedy) and deep RL methods (proximal policy optimization, double deep Q-network). Compared them against baselines across three problems: traveling salesman problem, pickup and delivery problem with time windows, and car sequencing problem. Used carefully designed reward functions to handle search-specific characteristics like large cost variations from constraint violation penalties.

Result: Algorithm performance varies substantially across different problems. ε-greedy consistently ranks among the best performers. Deep RL approaches require substantially longer runtime to become competitive due to computational overhead. Search-specific characteristics necessitate carefully designed reward functions for stable learning signals.

Conclusion: The study highlights both the promise and practical limitations of deep reinforcement learning in local search. While RL-based methods show potential, simpler approaches like ε-greedy perform consistently well, and deep RL methods face computational overhead challenges that limit their practical competitiveness without extended runtime.

Abstract: Reinforcement learning has recently gained traction as a means to improve combinatorial optimization methods, yet its effectiveness within local search metaheuristics specifically remains comparatively underexamined. In this study, we evaluate a range of reinforcement learning-based neighborhood selection strategies – multi-armed bandits (upper confidence bound, $ε$-greedy) and deep reinforcement learning methods (proximal policy optimization, double deep $Q$-network) – and compare them against multiple baselines across three different problems: the traveling salesman problem, the pickup and delivery problem with time windows, and the car sequencing problem. We show how search-specific characteristics, particularly large variations in cost due to constraint violation penalties, necessitate carefully designed reward functions to provide stable and informative learning signals. Our extensive experiments reveal that algorithm performance varies substantially across problems, although that $ε$-greedy consistently ranks among the best performers. In contrast, the computational overhead of deep reinforcement learning approaches only makes them competitive with a substantially longer runtime. These findings highlight both the promise and the practical limitations of deep reinforcement learning in local search.

Method: A hybrid architecture using residual learning: SARIMA models long-term seasonal trends, LSTM is trained exclusively on the residuals (nonlinear errors SARIMA fails to capture), fusing statistical stability with neural plasticity.

Result: The hybrid approach minimizes error propagation and enhances long-horizon accuracy by decomposing temperature into predictable climate components and nonlinear weather components.

Conclusion: The proposed SARIMA-LSTM hybrid framework effectively addresses limitations of both traditional statistical models and pure deep learning approaches for long-term atmospheric forecasting.

Abstract: Accurately forecasting long-term atmospheric variables remains a defining challenge in meteorological science due to the chaotic nature of atmospheric systems. Temperature data represents a complex superposition of deterministic cyclical climate forces and stochastic, short-term fluctuations. While planetary mechanics drive predictable seasonal periodicities, rapid meteorological changes such as thermal variations, pressure anomalies, and humidity shifts introduce nonlinear volatilities that defy simple extrapolation. Historically, the Seasonal Autoregressive Integrated Moving Average (SARIMA) model has been the standard for modeling historical weather data, prized for capturing linear seasonal trends. However, SARIMA operates under strict assumptions of stationarity, failing to capture abrupt, nonlinear transitions. This leads to systematic residual errors, manifesting as the under-prediction of sudden spikes or the over-smoothing of declines. Conversely, Deep Learning paradigms, specifically Long Short-Term Memory (LSTM) networks, demonstrate exceptional efficacy in handling intricate time-series data. By utilizing memory gates, LSTMs learn complex nonlinear dependencies. Yet, LSTMs face instability in open-loop forecasting; without ground truth feedback, minor deviations compound recursively, causing divergence. To resolve these limitations, we propose a Hybrid SARIMA-LSTM architecture. This framework employs a residual-learning strategy to decompose temperature into a predictable climate component and a nonlinear weather component. The SARIMA unit models the robust, long-term seasonal trend, while the LSTM is trained exclusively on the residuals the nonlinear errors SARIMA fails to capture. By fusing statistical stability with neural plasticity, this hybrid approach minimizes error propagation and enhances long-horizon accuracy.

Method: Cloud-based platform with ontology-backed data ingestion, machine-actionable knowledge graphs, FAIR-compliant metadata, schema/unit harmonization, uncertainty-aware surrogate modeling, and multi-objective multi-fidelity Bayesian optimization for closed-loop workflows.

Result: Validated through case studies in electrochemical materials and high-entropy alloys, demonstrating end-to-end data fusion, real-time optimization, and reproducible exploration of multi-objective trade spaces.

Conclusion: DataScribe serves as a general-purpose application-layer backbone for laboratories of any scale, supporting self-driving labs and distributed materials acceleration platforms with built-in support for performance, sustainability, and supply-chain objectives.

Abstract: The acceleration of materials discovery requires digital platforms that go beyond data repositories to embed learning, optimization, and decision-making directly into research workflows. We introduce DataScribe, an AI-native, cloud-based materials discovery platform that unifies heterogeneous experimental and computational data through ontology-backed ingestion and machine-actionable knowledge graphs. The platform integrates FAIR-compliant metadata capture, schema and unit harmonization, uncertainty-aware surrogate modeling, and native multi-objective multi-fidelity Bayesian optimization, enabling closed-loop propose-measure-learn workflows across experimental and computational pipelines. DataScribe functions as an application-layer intelligence stack, coupling data governance, optimization, and explainability rather than treating them as downstream add-ons. We validate the platform through case studies in electrochemical materials and high-entropy alloys, demonstrating end-to-end data fusion, real-time optimization, and reproducible exploration of multi-objective trade spaces. By embedding optimization engines, machine learning, and unified access to public and private scientific data directly within the data infrastructure, and by supporting open, free use for academic and non-profit researchers, DataScribe functions as a general-purpose application-layer backbone for laboratories of any scale, including self-driving laboratories and geographically distributed materials acceleration platforms, with built-in support for performance, sustainability, and supply-chain-aware objectives.

Method: Reformulates future-port ETA prediction as segment-level time-series forecasting. Uses transformer architecture with causally masked attention for long-range temporal dependencies. Integrates historical sailing durations, destination port congestion proxies, and static vessel descriptors. Employs multi-task learning to jointly predict sailing durations and port congestion states.

Result: Model outperforms comprehensive baselines on real-world global 2021 dataset: 4.85% MAE and 4.95% MAPE reduction vs sequence models; 9.39% MAE and 52.97% MAPE reduction vs gradient boosting machines. Case studies for major destination ports show superior accuracy.

Conclusion: The proposed transformer-based framework effectively addresses the future-port ETA prediction problem by leveraging historical patterns and multi-task learning, demonstrating significant improvements over conventional approaches for maritime schedule reliability.

Abstract: Accurate forecasts of segment-level sailing durations are fundamental to enhancing maritime schedule reliability and optimizing long-term port operations. However, conventional estimated time of arrival (ETA) models are primarily designed for the immediate next port of call and rely heavily on real-time automatic identification system (AIS) data, which is inherently unavailable for future voyage segments. To address this gap, the study reformulates future-port ETA prediction as a segment-level time-series forecasting problem. We develop a transformer-based architecture that integrates historical sailing durations, destination port congestion proxies, and static vessel descriptors. The proposed framework employs a causally masked attention mechanism to capture long-range temporal dependencies and a multi-task learning head to jointly predict segment sailing durations and port congestion states, leveraging shared latent signals to mitigate high uncertainty. Evaluation on a real-world global dataset from 2021 demonstrates the proposed model consistently outperforms a comprehensive suite of competitive baselines. The result shows a relative reduction of 4.85% in mean absolute error (MAE) and 4.95% in mean absolute percentage error (MAPE) compared with sequence baseline models. The relative reductions with gradient boosting machines are 9.39% in MAE and 52.97% in MAPE. Case studies for the major destination port further illustrate the model’s superior accuracy.

Method: InfGraND identifies structurally influential nodes to guide distillation, prioritizing critical parts of the graph. It embeds structural awareness in MLPs through one-time multi-hop neighborhood feature pre-computation, enriching the MLP’s input without inference-time overhead.

Result: InfGraND consistently outperforms prior GNN-to-MLP knowledge distillation methods across seven homophilic graph benchmark datasets in both transductive and inductive settings.

Conclusion: The framework demonstrates practicality for latency-critical applications by enabling efficient MLP-based inference while maintaining performance through structure-aware knowledge distillation from GNNs.

Abstract: Graph Neural Networks (GNNs) are the go-to model for graph data analysis. However, GNNs rely on two key operations - aggregation and update, which can pose challenges for low-latency inference tasks or resource-constrained scenarios. Simple Multi-Layer Perceptrons (MLPs) offer a computationally efficient alternative. Yet, training an MLP in a supervised setting often leads to suboptimal performance. Knowledge Distillation (KD) from a GNN teacher to an MLP student has emerged to bridge this gap. However, most KD methods either transfer knowledge uniformly across all nodes or rely on graph-agnostic indicators such as prediction uncertainty. We argue this overlooks a more fundamental, graph-centric inquiry: “How important is a node to the structure of the graph?” We introduce a framework, InfGraND, an Influence-guided Graph KNowledge Distillation from GNN to MLP that addresses this by identifying and prioritizing structurally influential nodes to guide the distillation process, ensuring that the MLP learns from the most critical parts of the graph. Additionally, InfGraND embeds structural awareness in MLPs through one-time multi-hop neighborhood feature pre-computation, which enriches the student MLP’s input and thus avoids inference-time overhead. Our rigorous evaluation in transductive and inductive settings across seven homophilic graph benchmark datasets shows InfGraND consistently outperforms prior GNN to MLP KD methods, demonstrating its practicality for numerous latency-critical applications in real-world settings.

Method: Construct structure-preserving geodesically complete metrics that preserve stationarity properties. Analyze the symmetric two-point zeroth-order estimator intrinsically (without ambient embeddings). Establish convergence guarantees for stochastic gradient descent with this intrinsic estimator.

Result: Theoretical convergence guarantees show that ε-stationary points under the constructed complete metric correspond to ε-stationary points under the original incomplete metric, matching best-known complexity in complete settings. Empirical validation on synthetic problems and mesh optimization demonstrates stable convergence even without geodesic completeness.

Conclusion: The framework successfully extends Riemannian zeroth-order optimization to geodesically incomplete settings through metric completion while preserving stationarity and achieving optimal complexity bounds.

Abstract: In this paper, we study Riemannian zeroth-order optimization in settings where the underlying Riemannian metric $g$ is geodesically incomplete, and the goal is to approximate stationary points with respect to this incomplete metric. To address this challenge, we construct structure-preserving metrics that are geodesically complete while ensuring that every stationary point under the new metric remains stationary under the original one. Building on this foundation, we revisit the classical symmetric two-point zeroth-order estimator and analyze its mean-squared error from a purely intrinsic perspective, depending only on the manifold’s geometry rather than any ambient embedding. Leveraging this intrinsic analysis, we establish convergence guarantees for stochastic gradient descent with this intrinsic estimator. Under additional suitable conditions, an $ε$-stationary point under the constructed metric $g’$ also corresponds to an $ε$-stationary point under the original metric $g$, thereby matching the best-known complexity in the geodesically complete setting. Empirical studies on synthetic problems confirm our theoretical findings, and experiments on a practical mesh optimization task demonstrate that our framework maintains stable convergence even in the absence of geodesic completeness.

Method: Propose a lookup table (LUT) compilation pipeline that replaces expensive spline computations with precomputed quantized tables and linear interpolation. The lightweight KAN model (50K parameters, 0.19MB) is evaluated with comprehensive analysis across LUT resolutions, quantization schemes, and out-of-bounds policies.

Result: Achieves 99.0% accuracy on CICIDS2017 DoS dataset. After LUT compilation (resolution L=8), maintains 98.96% accuracy (F1 degradation <0.0004) with 68× speedup at batch size 256 and over 5000× speedup at batch size 1, with only 2× memory overhead.

Conclusion: LUT-compiled KANs enable real-time DoS detection on CPU-only IoT gateways with deterministic inference latency and minimal resource footprint, establishing clear Pareto frontiers for accuracy-latency-memory trade-offs.

Abstract: Denial-of-Service (DoS) attacks pose a critical threat to Internet of Things (IoT) ecosystems, yet deploying effective intrusion detection on resource-constrained edge devices remains challenging. Kolmogorov-Arnold Networks (KANs) offer a compact alternative to Multi-Layer Perceptrons (MLPs) by placing learnable univariate spline functions on edges rather than fixed activations on nodes, achieving competitive accuracy with fewer parameters. However, runtime B-spline evaluation introduces significant computational overhead unsuitable for latency-critical IoT applications. We propose a lookup table (LUT) compilation pipeline that replaces expensive spline computations with precomputed quantized tables and linear interpolation, dramatically reducing inference latency while preserving detection quality. Our lightweight KAN model (50K parameters, 0.19~MB) achieves 99.0% accuracy on the CICIDS2017 DoS dataset. After LUT compilation with resolution $L=8$, the model maintains 98.96% accuracy (F1 degradation $<0.0004$) while achieving $\mathbf{68\times}$ speedup at batch size 256 and over $\mathbf{5000\times}$ speedup at batch size 1, with only $2\times$ memory overhead. We provide comprehensive evaluation across LUT resolutions, quantization schemes, and out-of-bounds policies, establishing clear Pareto frontiers for accuracy-latency-memory trade-offs. Our results demonstrate that LUT-compiled KANs enable real-time DoS detection on CPU-only IoT gateways with deterministic inference latency and minimal resource footprint.

Method: Q-realign reframes quantization as a dual-objective procedure for compression and safety recovery. It analyzes representational structure and uses post-training quantization to decouple safety alignment from fine-tuning, naturally integrating into deployment pipelines.

Result: Experiments show substantial reduction in unsafe behaviors while preserving task performance, with significant memory usage and GPU hour reductions. Can recover safety alignment of a fine-tuned 7B LLM on a single RTX 4090 within 40 minutes.

Conclusion: Q-realign provides a practical, turnkey solution for safety-aware deployment that decouples safety recovery from training, reduces computational overhead, and integrates seamlessly into existing workflows.

Abstract: Public large language models (LLMs) are typically safety-aligned during pretraining, yet task-specific fine-tuning required for deployment often erodes this alignment and introduces safety risks. Existing defenses either embed safety recovery into fine-tuning or rely on fine-tuning-derived priors for post-hoc correction, leaving safety recovery tightly coupled with training and incurring high computational overhead and a complex workflow. To address these challenges, we propose \texttt{Q-realign}, a post-hoc defense method based on post-training quantization, guided by an analysis of representational structure. By reframing quantization as a dual-objective procedure for compression and safety, \texttt{Q-realign} decouples safety alignment from fine-tuning and naturally piggybacks into modern deployment pipelines. Experiments across multiple models and datasets demonstrate that our method substantially reduces unsafe behaviors while preserving task performance, with significant reductions in memory usage and GPU hours. Notably, our approach can recover the safety alignment of a fine-tuned 7B LLM on a single RTX 4090 within 40 minutes. Overall, our work provides a practical, turnkey solution for safety-aware deployment.

Method: Proposes a fusion framework integrating local channel-wise descriptors (differential entropy + graph-theoretic features) and global trial-level descriptors (time-domain, spectral, complexity). Uses dual-branch transformer with attention-based fusion and domain-adversarial regularization, with intensity threshold filtering.

Result: Outperforms single-view and classical baselines under leave-one-subject-out protocol, achieving ~40% mean accuracy in 7-class subject-independent emotion recognition on SEED-VII dataset.

Conclusion: The proposed local-global fusion approach effectively addresses cross-subject generalization challenges in EEG emotion recognition by combining complementary feature representations with domain adaptation techniques.

Abstract: Subject-independent EEG emotion recognition is challenged by pronounced inter-subject variability and the difficulty of learning robust representations from short, noisy recordings. To address this, we propose a fusion framework that integrates (i) local, channel-wise descriptors and (ii) global, trial-level descriptors, improving cross-subject generalization on the SEED-VII dataset. Local representations are formed per channel by concatenating differential entropy with graph-theoretic features, while global representations summarize time-domain, spectral, and complexity characteristics at the trial level. These representations are fused in a dual-branch transformer with attention-based fusion and domain-adversarial regularization, with samples filtered by an intensity threshold. Experiments under a leave-one-subject-out protocol demonstrate that the proposed method consistently outperforms single-view and classical baselines, achieving approximately 40% mean accuracy in 7-class subject-independent emotion recognition. The code has been released at https://github.com/Danielz-z/LGF-EEG-Emotion.

Method: Leverages LLMs to generate temporally ordered subgoal sequences and state-to-subgoal-stage mappings. Applies potential-based reward shaping to transform sparse terminal rewards into dense, temporally consistent signals. Uses augmented dataset with shaped rewards for offline policy training.

Result: Outperforms state-of-the-art offline goal-conditioned and hierarchical RL baselines on four discrete and continuous sparse-reward benchmarks. Achieves faster convergence, higher success rates, and shorter trajectories. Robust to imperfect or noisy LLM-generated subgoal sequences.

Conclusion: LLM-guided subgoal temporal structures combined with theoretically grounded reward shaping provide a practical and scalable solution for long-horizon offline RL, addressing key limitations of existing methods.

Abstract: Offline reinforcement learning (RL) enables policy learning from pre-collected datasets, avoiding costly and risky online interactions, but it often struggles with long-horizon tasks involving sparse rewards. Existing goal-conditioned and hierarchical offline RL methods decompose such tasks and generate intermediate rewards to mitigate limitations of traditional offline RL, but usually overlook temporal dependencies among subgoals and rely on imprecise reward shaping, leading to suboptimal policies. To address these issues, we propose STO-RL (Offline RL using LLM-Guided Subgoal Temporal Order), an offline RL framework that leverages large language models (LLMs) to generate temporally ordered subgoal sequences and corresponding state-to-subgoal-stage mappings. Using this temporal structure, STO-RL applies potential-based reward shaping to transform sparse terminal rewards into dense, temporally consistent signals, promoting subgoal progress while avoiding suboptimal solutions. The resulting augmented dataset with shaped rewards enables efficient offline training of high-performing policies. Evaluations on four discrete and continuous sparse-reward benchmarks demonstrate that STO-RL consistently outperforms state-of-the-art offline goal-conditioned and hierarchical RL baselines, achieving faster convergence, higher success rates, and shorter trajectories. Ablation studies further confirm STO-RL’s robustness to imperfect or noisy LLM-generated subgoal sequences, demonstrating that LLM-guided subgoal temporal structures combined with theoretically grounded reward shaping provide a practical and scalable solution for long-horizon offline RL.

Method: Developed a 10-term cubic stochastic differential equation model trained on hurricane intensity data (IBTrACS) with environmental features from ERA5 reanalysis. Used automated system identification techniques restricted to Northern Hemisphere data.

Result: The model generates synthetic intensity series that capture key aspects of historical intensification statistics and hazard estimates in the Northern Hemisphere, demonstrating physical reasonableness.

Conclusion: Interpretable, physics-style models of complex earth system dynamics like tropical cyclone intensification can be successfully learned from data using automated system identification methods.

Abstract: Tropical cyclones are dangerous natural hazards, but their hazard is challenging to quantify directly from historical datasets due to limited dataset size and quality. Models of cyclone intensification fill this data gap by simulating huge ensembles of synthetic hurricanes based on estimates of the storm’s large scale environment. Both physics-based and statistical/ML intensification models have been developed to tackle this problem, but an open question is: can a physically reasonable and simple physics-style differential equation model of intensification be learned from data? In this paper, we answer this question in the affirmative by presenting a 10-term cubic stochastic differential equation model of Tropical Cyclone intensification. The model depends on a well-vetted suite of engineered environmental features known to drive intensification and is trained using a high quality dataset of hurricane intensity (IBTrACS) with estimates of the cyclone’s large scale environment from a data-assimilated simulation (ERA5 reanalysis), restricted to the Northern Hemisphere. The model generates synthetic intensity series which capture many aspects of historical intensification statistics and hazard estimates in the Northern Hemisphere. Our results show promise that interpretable, physics style models of complex earth system dynamics can be learned using automated system identification techniques.

Method: Proposes Structure Detection MBTL (SD-MBTL) framework that dynamically identifies CMDP generalization structures and selects appropriate MBTL algorithms. Specifically introduces M/GP-MBTL that detects Mountain structure and adaptively switches between Gaussian Process-based and clustering-based approaches.

Result: Extensive experiments on synthetic data and CRL benchmarks (continuous control, traffic control, agricultural management) show M/GP-MBTL surpasses strongest prior method by 12.49% on aggregated metric.

Conclusion: Online structure detection is promising for guiding source task selection in complex CRL environments, as different CMDP structures require different task selection strategies for optimal transfer learning.

Abstract: Contextual Reinforcement Learning (CRL) tackles the problem of solving a set of related Contextual Markov Decision Processes (CMDPs) that vary across different context variables. Traditional approaches–independent training and multi-task learning–struggle with either excessive computational costs or negative transfer. A recently proposed multi-policy approach, Model-Based Transfer Learning (MBTL), has demonstrated effectiveness by strategically selecting a few tasks to train and zero-shot transfer. However, CMDPs encompass a wide range of problems, exhibiting structural properties that vary from problem to problem. As such, different task selection strategies are suitable for different CMDPs. In this work, we introduce Structure Detection MBTL (SD-MBTL), a generic framework that dynamically identifies the underlying generalization structure of CMDP and selects an appropriate MBTL algorithm. For instance, we observe Mountain structure in which generalization performance degrades from the training performance of the target task as the context difference increases. We thus propose M/GP-MBTL, which detects the structure and adaptively switches between a Gaussian Process-based approach and a clustering-based approach. Extensive experiments on synthetic data and CRL benchmarks–covering continuous control, traffic control, and agricultural management–show that M/GP-MBTL surpasses the strongest prior method by 12.49% on the aggregated metric. These results highlight the promise of online structure detection for guiding source task selection in complex CRL environments.

Method: Used two synthetic data generating processes with cancellation-style structure where two primitive features share a strong nuisance factor while the target depends on a smaller differential factor. Varied colsample_bylevel and colsample_bynode parameters (s in {0.4, 0.6, 0.8, 0.9}), with emphasis on mild subsampling (s >= 0.8). Included a control feature set with engineered ratio features to compare performance.

Result: Intra-tree column subsampling reduces test PR-AUC in primitives-only settings. In the main process, relative decrease reached 54% when both parameters were set to 0.4. The effect largely disappeared when engineered ratio features were present. A path-based co-usage metric also dropped in cells where performance deteriorated.

Conclusion: When ratio-like structure is plausible in data, practitioners should either avoid intra-tree column subsampling or include the intended ratio features as engineered features to maintain model performance.

Abstract: Many applied problems contain signal that becomes clear only after combining multiple raw measurements. Ratios and rates are common examples. In gradient boosted trees, this combination is not an explicit operation: the model must synthesize it through coordinated splits on the component features. We study whether intra-tree column subsampling in XGBoost makes that synthesis harder. We use two synthetic data generating processes with cancellation-style structure. In both, two primitive features share a strong nuisance factor, while the target depends on a smaller differential factor. A log ratio cancels the nuisance and isolates the signal. We vary colsample_bylevel and colsample_bynode over s in {0.4, 0.6, 0.8, 0.9}, emphasizing mild subsampling (s >= 0.8). A control feature set includes the engineered ratio, removing the need for synthesis. Across both processes, intra-tree column subsampling reduces test PR-AUC in the primitives-only setting. In the main process the relative decrease reaches 54 percent when both parameters are set to 0.4. The effect largely disappears when the engineered ratio is present. A path-based co-usage metric drops in the same cells where performance deteriorates. Practically, if ratio-like structure is plausible, either avoid intra-tree subsampling or include the intended ratio features.

Method: Models trained RNN states as unknown nonlinear closed-loop feedback systems. Uses Koopman operator theory to approximate nonlinear dynamics with linear operators for interpretability. Applies spectral analysis to quantify worst-case impact of domain shifts on generalization error. Proposes a domain generalization method based on this analysis.

Result: Developed theoretical framework for analyzing RNN interpretability and OOD generalization. Proposed domain generalization method that reduces OOD generalization error and improves robustness to distribution shifts. Validated approach on practical temporal pattern-learning tasks.

Conclusion: The paper provides a novel theoretical framework combining Koopman operator theory and spectral analysis to address interpretability and generalization challenges in RNNs, with practical validation demonstrating improved robustness to domain shifts.

Abstract: Deep learning (DL) has driven broad advances across scientific and engineering domains. Despite its success, DL models often exhibit limited interpretability and generalization, which can undermine trust, especially in safety-critical deployments. As a result, there is growing interest in (i) analyzing interpretability and generalization and (ii) developing models that perform robustly under data distributions different from those seen during training (i.e. domain generalization). However, the theoretical analysis of DL remains incomplete. For example, many generalization analyses assume independent samples, which is violated in sequential data with temporal correlations. Motivated by these limitations, this paper proposes a method to analyze interpretability and out-of-domain (OOD) generalization for a family of recurrent neural networks (RNNs). Specifically, the evolution of a trained RNN’s states is modeled as an unknown, discrete-time, nonlinear closed-loop feedback system. Using Koopman operator theory, these nonlinear dynamics are approximated with a linear operator, enabling interpretability. Spectral analysis is then used to quantify the worst-case impact of domain shifts on the generalization error. Building on this analysis, a domain generalization method is proposed that reduces the OOD generalization error and improves the robustness to distribution shifts. Finally, the proposed analysis and domain generalization approach are validated on practical temporal pattern-learning tasks.

Method: Propose Reverse Flow Matching (RFM) framework using reverse inferential perspective, formulating training as posterior mean estimation given noisy samples. Introduce Langevin Stein operators to create zero-mean control variates, deriving a general class of variance-reduced estimators that unify existing methods.

Result: RFM extends Boltzmann distribution targeting from diffusion to flow policies, enables principled combination of Q-value and Q-gradient information for optimal minimum-variance estimator. Flow policy trained with RFM shows improved performance on continuous-control benchmarks vs diffusion policy baselines.

Conclusion: RFM provides unified theoretical framework connecting previously disparate methods, improves training efficiency/stability for diffusion/flow policies in online RL, and demonstrates practical benefits through better benchmark performance.

Abstract: Diffusion and flow policies are gaining prominence in online reinforcement learning (RL) due to their expressive power, yet training them efficiently remains a critical challenge. A fundamental difficulty in online RL is the lack of direct samples from the target distribution; instead, the target is an unnormalized Boltzmann distribution defined by the Q-function. To address this, two seemingly distinct families of methods have been proposed for diffusion policies: a noise-expectation family, which utilizes a weighted average of noise as the training target, and a gradient-expectation family, which employs a weighted average of Q-function gradients. Yet, it remains unclear how these objectives relate formally or if they can be synthesized into a more general formulation. In this paper, we propose a unified framework, reverse flow matching (RFM), which rigorously addresses the problem of training diffusion and flow models without direct target samples. By adopting a reverse inferential perspective, we formulate the training target as a posterior mean estimation problem given an intermediate noisy sample. Crucially, we introduce Langevin Stein operators to construct zero-mean control variates, deriving a general class of estimators that effectively reduce importance sampling variance. We show that existing noise-expectation and gradient-expectation methods are two specific instances within this broader class. This unified view yields two key advancements: it extends the capability of targeting Boltzmann distributions from diffusion to flow policies, and enables the principled combination of Q-value and Q-gradient information to derive an optimal, minimum-variance estimator, thereby improving training efficiency and stability. We instantiate RFM to train a flow policy in online RL, and demonstrate improved performance on continuous-control benchmarks compared to diffusion policy baselines.

Method: Proposes Resistance Curvature Flow (RCF) framework that leverages effective resistance from circuit physics to transform expensive curvature optimization into efficient matrix operations. Develops DGSL-RCF graph optimization algorithm based on this framework, which redistributes edge weights via curvature gradients to eliminate topological noise and strengthen local cluster structures.

Result: RCF achieves over 100x computational acceleration while maintaining geometric optimization capabilities comparable to OCF. Experimental results across deep metric learning, manifold learning, and graph structure learning demonstrate that DGSL-RCF significantly improves representation quality and downstream task performance.

Conclusion: RCF provides an efficient geometric evolution framework that breaks the computational bottleneck of traditional curvature-based methods, enabling practical application to large-scale datasets and deep learning models while effectively enhancing manifold structure and suppressing noise.

Abstract: Geometric Representation Learning (GRL) aims to approximate the non-Euclidean topology of high-dimensional data through discrete graph structures, grounded in the manifold hypothesis. However, traditional static graph construction methods based on Euclidean distance often fail to capture the intrinsic curvature characteristics of the data manifold. Although Ollivier-Ricci Curvature Flow (OCF) has proven to be a powerful tool for dynamic topological optimization, its core reliance on Optimal Transport (Wasserstein distance) leads to prohibitive computational complexity, severely limiting its application in large-scale datasets and deep learning frameworks. To break this bottleneck, this paper proposes a novel geometric evolution framework: Resistance Curvature Flow (RCF). Leveraging the concept of effective resistance from circuit physics, RCF transforms expensive curvature optimization into efficient matrix operations. This approach achieves over 100x computational acceleration while maintaining geometric optimization capabilities comparable to OCF. We provide an in-depth exploration of the theoretical foundations and dynamical principles of RCF, elucidating how it guides the redistribution of edge weights via curvature gradients to eliminate topological noise and strengthen local cluster structures. Furthermore, we provide a mechanistic explanation of RCF’s role in manifold enhancement and noise suppression, as well as its compatibility with deep learning models. We design a graph optimization algorithm, DGSL-RCF, based on this framework. Experimental results across deep metric learning, manifold learning, and graph structure learning demonstrate that DGSL-RCF significantly improves representation quality and downstream task performance.

Method: Proposes VBO-MI using variational mutual information estimation with actor-critic architecture: action-net navigates input space, variational critic estimates information gain, enabling end-to-end gradient flow without inner-loop optimization.

Result: Achieves up to 100x FLOP reduction compared to BNN-BO baselines while maintaining or improving optimization performance on high-dimensional synthetic functions, PDE optimization, Lunar Lander control, and categorical Pest Control tasks.

Conclusion: VBO-MI provides computationally scalable Bayesian optimization with gradient-based mutual information estimation, eliminating traditional acquisition optimization bottlenecks while delivering competitive or superior performance across diverse benchmarks.

Abstract: Many real-world tasks require optimizing expensive black-box functions accessible only through noisy evaluations, a setting commonly addressed with Bayesian optimization (BO). While Bayesian neural networks (BNNs) have recently emerged as scalable alternatives to Gaussian Processes (GPs), traditional BNN-BO frameworks remain burdened by expensive posterior sampling and acquisition function optimization. In this work, we propose {VBO-MI} (Variational Bayesian Optimization with Mutual Information), a fully gradient-based BO framework that leverages recent advances in variational mutual information estimation. To enable end-to-end gradient flow, we employ an actor-critic architecture consisting of an {action-net} to navigate the input space and a {variational critic} to estimate information gain. This formulation effectively eliminates the traditional inner-loop acquisition optimization bottleneck, achieving up to a {$10^2 \times$ reduction in FLOPs} compared to BNN-BO baselines. We evaluate our method on a diverse suite of benchmarks, including high-dimensional synthetic functions and complex real-world tasks such as PDE optimization, the Lunar Lander control problem, and categorical Pest Control. Our experiments demonstrate that VBO-MI consistently provides the same or superior optimization performance and computational scalability over the baselines.

Method: The authors analyze information encoded in hidden representations and examine how representations evolve across layers using probing experiments that test for linear regression coefficients, intermediate values from complex expressions, and final answers in early layers.

Result: Results show meaningful structured information is stored inside representations, with clear signals corresponding to both intermediate and final quantities involved in prediction processes, providing insight into how models refine inputs and produce outputs.

Conclusion: The findings contribute to understanding internal mechanics of tabular foundational models, showing they encode concrete interpretable information that moves toward making decision processes more transparent and trustworthy.

Abstract: Tabular foundational models are pre-trained models designed for a wide range of tabular data tasks. They have shown strong performance across domains, yet their internal representations and learned concepts remain poorly understood. This lack of interpretability makes it important to study how these models process and transform input features. In this work, we analyze the information encoded inside the model’s hidden representations and examine how these representations evolve across layers. We run a set of probing experiments that test for the presence of linear regression coefficients, intermediate values from complex expressions, and the final answer in early layers. These experiments allow us to reason about the computations the model performs internally. Our results provide evidence that meaningful and structured information is stored inside the representations of tabular foundational models. We observe clear signals that correspond to both intermediate and final quantities involved in the model’s prediction process. This gives insight into how the model refines its inputs and how the final output emerges. Our findings contribute to a deeper understanding of the internal mechanics of tabular foundational models. They show that these models encode concrete and interpretable information, which moves us closer to making their decision processes more transparent and trustworthy.

Method: Introduces Collective Influence Estimation Network (CIEN) that explicitly models the collective influence of other agents on the task object. Each agent infers interaction information from local observations and task object states, eliminating need for explicit action information exchange.

Result: Achieves stable and efficient coordination in communication-limited environments. Framework avoids network expansion with increasing team size, allows new agents to be incorporated without modifying existing network structures, and demonstrates strong scalability.

Conclusion: CIEN enables communication-efficient MARL with improved scalability. Real robotic deployment shows enhanced robustness, deployment feasibility, and reduced dependence on communication infrastructure compared to traditional approaches.

Abstract: Multiagent reinforcement learning (MARL) has attracted considerable attention due to its potential in addressing complex cooperative tasks. However, existing MARL approaches often rely on frequent exchanges of action or state information among agents to achieve effective coordination, which is difficult to satisfy in practical robotic systems. A common solution is to introduce estimator networks to model the behaviors of other agents and predict their actions; nevertheless, such designs cause the size and computational cost of the estimator networks to grow rapidly with the number of agents, thereby limiting scalability in large-scale systems. To address these challenges, this paper proposes a multiagent learning framework augmented with a Collective Influence Estimation Network (CIEN). By explicitly modeling the collective influence of other agents on the task object, each agent can infer critical interaction information solely from its local observations and the task object’s states, enabling efficient collaboration without explicit action information exchange. The proposed framework effectively avoids network expansion as the team size increases; moreover, new agents can be incorporated without modifying the network structures of existing agents, demonstrating strong scalability. Experimental results on multiagent cooperative tasks based on the Soft Actor-Critic (SAC) algorithm show that the proposed method achieves stable and efficient coordination under communication-limited environments. Furthermore, policies trained with collective influence modeling are deployed on a real robotic platform, where experimental results indicate significantly improved robustness and deployment feasibility, along with reduced dependence on communication infrastructure.

Method: Formulate federated ridge regression as distributed equilibrium problem where each client computes local Gram matrix and moment vector once, transmits them to server, which reconstructs global solution through single matrix inversion. For high dimensions, use random projections to reduce communication from O(d²) to O(m²) where m ≪ d.

Result: Exact recovery of centralized ridge solution under coverage condition; non-asymptotic error bounds for heterogeneous distributions; communication reduction from O(Rd) to O(d²) (O(m²) with projections); differential privacy with single noise injection; 38× less communication than FedAvg while matching accuracy.

Conclusion: Iterative communication is unnecessary for distributed linear regression - one-shot aggregation of sufficient statistics yields exact centralized solution with dramatically reduced communication, applicable to kernel methods but not general nonlinear architectures.

Abstract: Federated learning protocols require repeated synchronization between clients and a central server, with convergence rates depending on learning rates, data heterogeneity, and client sampling. This paper asks whether iterative communication is necessary for distributed linear regression. We show it is not. We formulate federated ridge regression as a distributed equilibrium problem where each client computes local sufficient statistics – the Gram matrix and moment vector – and transmits them once. The server reconstructs the global solution through a single matrix inversion. We prove exact recovery: under a coverage condition on client feature matrices, one-shot aggregation yields the centralized ridge solution, not an approximation. For heterogeneous distributions violating coverage, we derive non-asymptotic error bounds depending on spectral properties of the aggregated Gram matrix. Communication reduces from $\mathcal{O}(Rd)$ in iterative methods to $\mathcal{O}(d^2)$ total; for high-dimensional settings, we propose and experimentally validate random projection techniques reducing this to $\mathcal{O}(m^2)$ where $m \ll d$. We establish differential privacy guarantees where noise is injected once per client, eliminating the composition penalty that degrades privacy in multi-round protocols. We further address practical considerations including client dropout robustness, federated cross-validation for hyperparameter selection, and comparison with gradient-based alternatives. Comprehensive experiments on synthetic heterogeneous regression demonstrate that one-shot fusion matches FedAvg accuracy while requiring up to $38\times$ less communication. The framework applies to kernel methods and random feature models but not to general nonlinear architectures.

Method: Two-agent system: LLM generates rank-1 SVD updates, VLM accepts/rejects updates and decides when to stop. Uses in-context learning and row/column permutations to expose visually coherent structure for stability.

Result: On synthetic noisy matrices, best agentic configuration differs by only 1.75 RMSE from noise target. On Digits and CIFAR-10 (deflating to 10% Frobenius norm), achieves competitive results across all settings.

Conclusion: Fully agentic, threshold-free matrix deflation is a viable alternative to classical numerical algorithms, demonstrating the potential of LLM+VLM teams for numerical tasks.

Abstract: Can a small team of agents peel a matrix apart, one rank-1 slice at a time? We propose an agentic approach to matrix deflation in which a solver Large Language Model (LLM) generates rank-1 Singular Value Decomposition (SVD) updates and a Vision Language Model (VLM) accepts or rejects each update and decides when to stop, eliminating fixed norm thresholds. Solver stability is improved through in-context learning (ICL) and types of row/column permutations that expose visually coherent structure. We evaluate on Digits ($8{\times}8$), CIFAR-10 ($32{\times}32$ grayscale), and synthetic ($16{\times}16$) matrices with and without Gaussian noise. In the synthetic noisy case, where the true construction rank $k$ is known, numerical deflation provides the noise target and our best agentic configuration differs by only $1.75$ RMSE of the target. For Digits and CIFAR-10, targets are defined by deflating until the Frobenius norm reaches $10%$ of the original. Across all settings, our agent achieves competitive results, suggesting that fully agentic, threshold-free deflation is a viable alternative to classical numerical algorithms.

Method: GADPN uses low-rank denoising and generalized structural perturbation. Key innovations: (1) Bayesian optimization to adaptively determine optimal denoising strength based on each graph’s homophily level; (2) extending structural perturbation method to arbitrary graphs via Singular Value Decomposition (SVD).

Result: Extensive experiments show GADPN achieves state-of-the-art performance with significantly improved efficiency. Shows particularly strong gains on challenging disassortative graphs, validating robust learning across diverse network types.

Conclusion: GADPN provides a simple yet effective graph structure learning framework that adaptively refines graph topology, overcoming computational limitations of existing methods while achieving superior performance across various graph types.

Abstract: While Graph Neural Networks (GNNs) excel on graph-structured data, their performance is fundamentally limited by the quality of the observed graph, which often contains noise, missing links, or structural properties misaligned with GNNs’ underlying assumptions. To address this, graph structure learning aims to infer a more optimal topology. Existing methods, however, often incur high computational costs due to complex generative models and iterative joint optimization, limiting their practical utility. In this paper, we propose GADPN, a simple yet effective graph structure learning framework that adaptively refines graph topology via low-rank denoising and generalized structural perturbation. Our approach makes two key contributions: (1) we introduce Bayesian optimization to adaptively determine the optimal denoising strength, tailoring the process to each graph’s homophily level; and (2) we extend the structural perturbation method to arbitrary graphs via Singular Value Decomposition (SVD), overcoming its original limitation to symmetric structures. Extensive experiments on benchmark datasets demonstrate that GADPN achieves state-of-the-art performance while significantly improving efficiency. It shows particularly strong gains on challenging disassortative graphs, validating its ability to robustly learn enhanced graph structures across diverse network types.

Method: The researchers integrate cognitive biases (overconfidence, loss aversion) into RL reward structures and decision-making processes. They evaluate these biased agents in both simulated and real-world trading environments.

Result: The study yields inconclusive or negative results - the biased RL agents do not achieve better risk-adjusted returns than standard RL agents, despite exhibiting more human-like trading behavior.

Conclusion: While the biased agents show human-like behavior, incorporating cognitive biases into RL for financial trading presents significant challenges. The study provides valuable insights for developing more robust financial AI systems that better account for human psychology.

Abstract: Financial markets are influenced by human behavior that deviates from rationality due to cognitive biases. Traditional reinforcement learning (RL) models for financial decision-making assume rational agents, potentially overlooking the impact of psychological factors. This study integrates cognitive biases into RL frameworks for financial trading, hypothesizing that such models can exhibit human-like trading behavior and achieve better risk-adjusted returns than standard RL agents. We introduce biases, such as overconfidence and loss aversion, into reward structures and decision-making processes and evaluate their performance in simulated and real-world trading environments. Despite its inconclusive or negative results, this study provides insights into the challenges of incorporating human-like biases into RL, offering valuable lessons for developing robust financial AI systems.

Method: Proposes Hyperbolic Heterogeneous Graph Transformer (HypHGT) that operates entirely in hyperbolic space using transformer-based architecture with relation-specific hyperbolic attention mechanism having linear time complexity, capturing both local and global dependencies while preserving heterogeneous information.

Result: HypHGT consistently outperforms state-of-the-art methods in node classification tasks while significantly reducing training time and memory usage, demonstrating effectiveness and efficiency.

Conclusion: HypHGT effectively addresses limitations of existing hyperbolic heterogeneous GNNs by capturing complex structural properties and semantic information in heterogeneous graphs through hyperbolic transformer architecture with efficient relation-specific attention.

Abstract: In heterogeneous graphs, we can observe complex structures such as tree-like or hierarchical structures. Recently, the hyperbolic space has been widely adopted in many studies to effectively learn these complex structures. Although these methods have demonstrated the advantages of the hyperbolic space in learning heterogeneous graphs, most existing methods still have several challenges. They rely heavily on tangent-space operations, which often lead to mapping distortions during frequent transitions. Moreover, their message-passing architectures mainly focus on local neighborhood information, making it difficult to capture global hierarchical structures and long-range dependencies between different types of nodes. To address these limitations, we propose Hyperbolic Heterogeneous Graph Transformer (HypHGT), which effectively and efficiently learns heterogeneous graph representations entirely within the hyperbolic space. Unlike previous message-passing based hyperbolic heterogeneous GNNs, HypHGT naturally captures both local and global dependencies through transformer-based architecture. Furthermore, the proposed relation-specific hyperbolic attention mechanism in HypHGT, which operates with linear time complexity, enables efficient computation while preserving the heterogeneous information across different relation types. This design allows HypHGT to effectively capture the complex structural properties and semantic information inherent in heterogeneous graphs. We conduct comprehensive experiments to evaluate the effectiveness and efficiency of HypHGT, and the results demonstrate that it consistently outperforms state-of-the-art methods in node classification task, with significantly reduced training time and memory usage.

Method: Derived exact marginal output variance using Wishart distribution, James’ theorem, Laplace-integral expansion, and combinatorial expansions for trace moments. Conducted Monte Carlo experiments and empirical analysis on Higgs boson dataset.

Result: He/Kaiming initialization yields output variance of 0.41, while new parameterization (10/√n for α=1) achieves 0.9 variance. However, empirical validation shows He initialization still performs better despite variance preservation.

Conclusion: Theoretical analysis explains information loss in Lipschitz networks and provides initialization prescriptions, but practical performance may not align with theoretical variance preservation.

Abstract: We analyze initialization dynamics for LDLT-based $\mathcal{L}$-Lipschitz layers by deriving the exact marginal output variance when the underlying parameter matrix $W_0\in \mathbb{R}^{m\times n}$ is initialized with IID Gaussian entries $\mathcal{N}(0,σ^2)$. The Wishart distribution, $S=W_0W_0^\top\sim\mathcal{W}_m(n,σ^2 \boldsymbol{I}_m)$, used for computing the output marginal variance is derived in closed form using expectations of zonal polynomials via James’ theorem and a Laplace-integral expansion of $(α\boldsymbol{I}_m+S)^{-1}$. We develop an Isserlis/Wick-based combinatorial expansion for $\operatorname{\mathbb{E}}\left[\operatorname{tr}(S^k)\right]$ and provide explicit truncated moments up to $k=10$, which yield accurate series approximations for small-to-moderate $σ^2$. Monte Carlo experiments confirm the theoretical estimates. Furthermore, empirical analysis was performed to quantify that, using current He or Kaiming initialization with scaling $1/\sqrt{n}$, the output variance is $0.41$, whereas the new parameterization with $10/ \sqrt{n}$ for $α=1$ results in an output variance of $0.9$. The findings clarify why deep $\mathcal{L}$-Lipschitz networks suffer rapid information loss at initialization and offer practical prescriptions for choosing initialization hyperparameters to mitigate this effect. However, using the Higgs boson classification dataset, a hyperparameter sweep over optimizers, initialization scale, and depth was conducted to validate the results on real-world data, showing that although the derivation ensures variance preservation, empirical results indicate He initialization still performs better.

Method: Adopts a multi-label classification framework for evaluation, testing several representative unsupervised feature selection methods on 21 multi-label datasets.

Result: Performance rankings of methods differ significantly from those reported under single-label settings, showing that multi-label evaluation provides more reliable comparisons.

Conclusion: Multi-label evaluation settings should be used for fair and reliable comparison of unsupervised feature selection methods, as single-label evaluation from multi-label data is inherently unstable.

Abstract: Unsupervised feature selection aims to identify a compact subset of features that captures the intrinsic structure of data without supervised label. Most existing studies evaluate the performance of methods using the single-label dataset that can be instantiated by selecting a label from multi-label data while maintaining the original features. Because the chosen label can vary arbitrarily depending on the experimental setting, the superiority among compared methods can be changed with regard to which label happens to be selected. Thus, evaluating unsupervised feature selection methods based solely on single-label accuracy is unreasonable for assessing their true discriminative ability. This study revisits this evaluation paradigm by adopting a multi-label classification framework. Experiments on 21 multi-label datasets using several representative methods demonstrate that performance rankings differ markedly from those reported under single-label settings, suggesting the possibility of multi-label evaluation settings for fair and reliable comparison of unsupervised feature selection methods.

Method: Generative Adversarial Network (GAN) with graphical command-line interface to generate realistic synthetic ECG beats, allowing clinical researchers to train class-specific generators and produce large volumes of labeled synthetic data

Result: Tool enables generation of synthetic ECG beats that preserve distribution of minority classes, supporting early diagnosis and patient stratification

Conclusion: The GAN-based tool addresses data scarcity in cardiac amyloidosis by providing usable synthetic ECG generation for clinical research and machine learning applications

Abstract: Cardiac amyloidosis (CA) is a rare and underdiagnosed infiltrative cardiomyopathy, and available datasets for machine-learning models are typically small, imbalanced and heterogeneous. This paper presents a Generative Adversarial Network (GAN) and a graphical command-line interface for generating realistic synthetic electrocardiogram (ECG) beats to support early diagnosis and patient stratification in CA. The tool is designed for usability, allowing clinical researchers to train class-specific generators once and then interactively produce large volumes of labelled synthetic beats that preserve the distribution of minority classes.

Method: Empirical analysis of open-source LLMs to identify SDH characteristics, combined with theoretical analysis of queries, keys, and Rotary Position Embeddings. Formal modeling assumptions and analysis of training dynamics in shallow Transformers with RoPE under gradient descent.

Result: SDHs are intrinsic to models and generalize to OOD prompts. Two key conditions identified: (1) queries and keys are almost rank-one (nearly identical across tokens), and (2) RoPE is dominated by medium- and high-frequency components. Theoretical proof shows these conditions ensure SDH emergence.

Conclusion: Slash-Dominant Heads emerge from fundamental properties of LLM architecture: rank-one structure in queries/keys combined with specific frequency characteristics of RoPE. This explains the intrinsic emergence of slash attention patterns crucial for information flow in Transformers.

Abstract: Large Language Models (LLMs) often exhibit slash attention patterns, where attention scores concentrate along the $Δ$-th sub-diagonal for some offset $Δ$. These patterns play a key role in passing information across tokens. But why do they emerge? In this paper, we demystify the emergence of these Slash-Dominant Heads (SDHs) from both empirical and theoretical perspectives. First, by analyzing open-source LLMs, we find that SDHs are intrinsic to models and generalize to out-of-distribution prompts. To explain the intrinsic emergence, we analyze the queries, keys, and Rotary Position Embedding (RoPE), which jointly determine attention scores. Our empirical analysis reveals two characteristic conditions of SDHs: (1) Queries and keys are almost rank-one, and (2) RoPE is dominated by medium- and high-frequency components. Under these conditions, queries and keys are nearly identical across tokens, and interactions between medium- and high-frequency components of RoPE give rise to SDHs. Beyond empirical evidence, we theoretically show that these conditions are sufficient to ensure the emergence of SDHs by formalizing them as our modeling assumptions. Particularly, we analyze the training dynamics of a shallow Transformer equipped with RoPE under these conditions, and prove that models trained via gradient descent exhibit SDHs. The SDHs generalize to out-of-distribution prompts.

Method: ORBIT uses multi-stage reinforcement learning to discover Pareto-optimal reasoning behaviors at each effort level, followed by on-policy distillation to fuse these behaviors into a single unified model with well-separated reasoning modes triggered by input.

Result: ORBIT achieves: (1) controllable reasoning behavior over multiple modes, (2) competitive reasoning density within each mode, and (3) successful integration of frontier policies into a single student model while preserving clear mode separation and high per-mode performance.

Conclusion: ORBIT provides a flexible framework for controllable multi-budget reasoning that addresses the limitations of uniform reasoning approaches and enables adaptable cost-accuracy trade-offs for different deployment scenarios.

Abstract: Recent Large Reasoning Models (LRMs) achieve strong performance by leveraging long-form Chain-of-Thought (CoT) reasoning, but uniformly applying overlong reasoning at inference time incurs substantial and often unnecessary computational cost. To address this, prior work explores various strategies to infer an appropriate reasoning budget from the input. However, such approaches are unreliable in the worst case, as estimating the minimal required reasoning effort is fundamentally difficult, and they implicitly fix the trade-off between reasoning cost and accuracy during training, limiting flexibility under varying deployment scenarios. Motivated by these limitations, we propose ORBIT, a controllable multi-budget reasoning framework with well-separated reasoning modes triggered by input. ORBIT employs multi-stage reinforcement learning to discover Pareto-optimal reasoning behaviors at each effort, followed by on-policy distillation to fuse these behaviors into a single unified model. Experiments show that ORBIT achieves (1) controllable reasoning behavior over multiple modes, (2) competitive reasoning density within each mode, and (3) integration of these frontier policies into a single unified student model while preserving clear mode separation and high per-mode performance.

Method: Trained fully connected neural networks of three sizes on CIFAR-10 with 30% label noise. Decomposed loss curves into signal contributions from clean and noisy training data, separately analyzing epoch-wise evolution of internal signals to study structural changes.

Result: Three key findings: 1) Models achieved strong re-generalization on test data after perfectly fitting noisy data (benign overfitting). 2) Noisy data were learned after clean data, with their internal activations becoming increasingly separated in outer layers. 3) Single very large activations emerged in shallow layers across all models, correlating with input patterns but not output patterns.

Conclusion: The empirical findings directly link deep double descent, benign overfitting, and large activation phenomena, supporting a novel scenario for understanding deep double descent through internal structural evolution during training.

Abstract: Deep double descent is one of the key phenomena underlying the generalization capability of deep learning models. In this study, epoch-wise double descent, which is delayed generalization following overfitting, was empirically investigated by focusing on the evolution of internal structures. Fully connected neural networks of three different sizes were trained on the CIFAR-10 dataset with 30% label noise. By decomposing the loss curves into signal contributions from clean and noisy training data, the epoch-wise evolutions of internal signals were analyzed separately. Three main findings were obtained from this analysis. First, the model achieved strong re-generalization on test data even after perfectly fitting noisy training data during the double descent phase, corresponding to a “benign overfitting” state. Second, noisy data were learned after clean data, and as learning progressed, their corresponding internal activations became increasingly separated in outer layers; this enabled the model to overfit only noisy data. Third, a single, very large activation emerged in the shallow layer across all models; this phenomenon is referred as “outliers,” “massive activa-tions,” and “super activations” in recent large language models and evolves with re-generalization. The magnitude of large activation correlated with input patterns but not with output patterns. These empirical findings directly link the recent key phenomena of “deep double descent,” “benign overfitting,” and “large activation”, and support the proposal of a novel scenario for understanding deep double descent.

Method: Tested 6 general-purpose and 5 radiomics-specific AutoML frameworks on 10 diverse radiomics datasets (CT/MRI, varied sizes/anatomies/endpoints) using standardized cross-validation with predefined parameters. Evaluated performance (AUC), runtime, and qualitative aspects (software status, accessibility, interpretability).

Result: Simplatab (radiomics-specific) achieved highest average test AUC (81.81%) with moderate runtime (~1 hour). LightAutoML (general-purpose) showed fastest execution with competitive performance (78.74% mean AUC in 6 minutes). Most radiomics-specific frameworks were excluded due to obsolescence, programming requirements, or computational inefficiency, while general-purpose frameworks demonstrated higher accessibility.

Conclusion: Simplatab provides effective balance of performance, efficiency, and accessibility for radiomics classification, but significant gaps remain including lack of accessible survival analysis support and limited integration of feature reproducibility/harmonization. Future research should focus on adapting AutoML solutions to better address radiomics-specific challenges.

Abstract: Automated machine learning (AutoML) frameworks can lower technical barriers for predictive and prognostic model development in radiomics by enabling researchers without programming expertise to build models. However, their effectiveness in addressing radiomics-specific challenges remains unclear. This study evaluates the performance, efficiency, and accessibility of general-purpose and radiomics-specific AutoML frameworks on diverse radiomics classification tasks, thereby highlighting development needs for radiomics. Ten public/private radiomics datasets with varied imaging modalities (CT/MRI), sizes, anatomies and endpoints were used. Six general-purpose and five radiomics-specific frameworks were tested with predefined parameters using standardized cross-validation. Evaluation metrics included AUC, runtime, together with qualitative aspects related to software status, accessibility, and interpretability. Simplatab, a radiomics-specific tool with a no-code interface, achieved the highest average test AUC (81.81%) with a moderate runtime (~1 hour). LightAutoML, a general-purpose framework, showed the fastest execution with competitive performance (78.74% mean AUC in six minutes). Most radiomics-specific frameworks were excluded from the performance analysis due to obsolescence, extensive programming requirements, or computational inefficiency. Conversely, general-purpose frameworks demonstrated higher accessibility and ease of implementation. Simplatab provides an effective balance of performance, efficiency, and accessibility for radiomics classification problems. However, significant gaps remain, including the lack of accessible survival analysis support and the limited integration of feature reproducibility and harmonization within current AutoML frameworks. Future research should focus on adapting AutoML solutions to better address these radiomics-specific challenges.

Method: Proposes MMD Guidance that augments the reverse diffusion process with gradients of Maximum Mean Discrepancy between generated samples and reference dataset. MMD provides reliable distributional estimates from limited data, is efficiently differentiable, and extends to conditional generation via product kernels. Works efficiently in latent diffusion models by applying guidance in latent space.

Result: Experiments on synthetic and real-world benchmarks demonstrate that MMD Guidance achieves distributional alignment while preserving sample fidelity.

Conclusion: MMD Guidance offers an effective training-free solution for domain adaptation with diffusion models using few reference examples, directly aligning generated distributions with target data characteristics while maintaining sample quality.

Abstract: Pre-trained diffusion models have emerged as powerful generative priors for both unconditional and conditional sample generation, yet their outputs often deviate from the characteristics of user-specific target data. Such mismatches are especially problematic in domain adaptation tasks, where only a few reference examples are available and retraining the diffusion model is infeasible. Existing inference-time guidance methods can adjust sampling trajectories, but they typically optimize surrogate objectives such as classifier likelihoods rather than directly aligning with the target distribution. We propose MMD Guidance, a training-free mechanism that augments the reverse diffusion process with gradients of the Maximum Mean Discrepancy (MMD) between generated samples and a reference dataset. MMD provides reliable distributional estimates from limited data, exhibits low variance in practice, and is efficiently differentiable, which makes it particularly well-suited for the guidance task. Our framework naturally extends to prompt-aware adaptation in conditional generation models via product kernels. Also, it can be applied with computational efficiency in latent diffusion models (LDMs), since guidance is applied in the latent space of the LDM. Experiments on synthetic and real-world benchmarks demonstrate that MMD Guidance can achieve distributional alignment while preserving sample fidelity.

Method: SSO enforces strict module-wise spectral constraints on both weights and their updates by deriving the steepest descent direction on the spectral sphere, implemented as an efficient parallel algorithm in Megatron.

Result: SSO consistently outperforms AdamW and Muon in pretraining diverse architectures (Dense 1.7B, MoE 8B-A1B, 200-layer DeepNet), with improved MoE router load balancing, suppressed outliers, and strictly bounded activations.

Conclusion: SSO provides a fully μP-aligned optimization approach that offers both performance improvements and practical stability benefits for large-scale model training.

Abstract: Scaling large models requires optimization strategies that ensure rapid convergence grounded in stability. Maximal Update Parametrization ($\boldsymbolμ$P) provides a theoretical safeguard for width-invariant $Θ(1)$ activation control, whereas emerging optimizers like Muon are only ``half-aligned’’ with these constraints: they control updates but allow weights to drift. To address this limitation, we introduce the \textbf{Spectral Sphere Optimizer (SSO)}, which enforces strict module-wise spectral constraints on both weights and their updates. By deriving the steepest descent direction on the spectral sphere, SSO realizes a fully $\boldsymbolμ$P-aligned optimization process. To enable large-scale training, we implement SSO as an efficient parallel algorithm within Megatron. Through extensive pretraining on diverse architectures, including Dense 1.7B, MoE 8B-A1B, and 200-layer DeepNet models, SSO consistently outperforms AdamW and Muon. Furthermore, we observe significant practical stability benefits, including improved MoE router load balancing, suppressed outliers, and strictly bounded activations.

Method: Introduces a multi-channel U-Net architecture for autoregressive deep-learning surrogates of 2D PDE dynamics on periodic domains. Evaluated on five qualitatively different PDE families and compared against ViT, AFNO, PDE-Transformer, and KAN-UNet under common training setup.

Result: Across all datasets, me-UNet matches or outperforms more complex architectures in field-space error, spectral similarity, and physics-based metrics for in-distribution rollouts. It generalizes qualitatively to unseen initial conditions with as few as ~20 training simulations and requires substantially less training time.

Conclusion: In small-data periodic 2D PDE settings, convolutional architectures with inductive biases aligned to locality and periodic boundary conditions (like U-Net) remain strong contenders for accurate and moderately out-of-distribution-robust surrogate modeling.

Abstract: Partial differential equations (PDEs) are a central tool for modeling the dynamics of physical, engineering, and materials systems, but high-fidelity simulations are often computationally expensive. At the same time, many scientific applications can be viewed as the evolution of spatially distributed fields, making data-driven forecasting of such fields a core task in scientific machine learning. In this work we study autoregressive deep-learning surrogates for two-dimensional PDE dynamics on periodic domains, focusing on generalization to out-of-distribution initial conditions within a fixed PDE and parameter regime and on strict small-data settings with at most $\mathcal{O}(10^2)$ simulated trajectories per system. We introduce a multi-channel U-Net […], evaluate it on five qualitatively different PDE families and compare it to ViT, AFNO, PDE-Transformer, and KAN-UNet under a common training setup. Across all datasets, me-UNet matches or outperforms these more complex architectures in terms of field-space error, spectral similarity, and physics-based metrics for in-distribution rollouts, while requiring substantially less training time. It also generalizes qualitatively to unseen initial conditions with as few as $\approx 20$ training simulations. A data-efficiency study and Grad-CAM analysis further suggest that, in small-data periodic 2D PDE settings, convolutional architectures with inductive biases aligned to locality and periodic boundary conditions remain strong contenders for accurate and moderately out-of-distribution-robust surrogate modeling.

Method: Taxon integrates: (1) feature-gating mixture-of-experts architecture that adaptively routes multi-modal features across taxonomy levels, (2) semantic consistency model distilled from LLMs to verify alignment between product titles and official tax definitions, and (3) multi-source training pipeline combining curated tax databases, invoice validation logs, and merchant registration data.

Result: Taxon achieves state-of-the-art performance on proprietary TaxCode dataset and public benchmarks. Full hierarchical paths reconstruction improves structural consistency, yielding highest overall F1 scores. Deployed at Alibaba handling 500k+ daily queries with peak volumes over 5 million requests during business events.

Conclusion: Taxon provides an effective framework for hierarchical tax code prediction with improved accuracy, interpretability, and robustness, successfully deployed in production for large-scale e-commerce tax compliance.

Abstract: Tax code prediction is a crucial yet underexplored task in automating invoicing and compliance management for large-scale e-commerce platforms. Each product must be accurately mapped to a node within a multi-level taxonomic hierarchy defined by national standards, where errors lead to financial inconsistencies and regulatory risks. This paper presents Taxon, a semantically aligned and expert-guided framework for hierarchical tax code prediction. Taxon integrates (i) a feature-gating mixture-of-experts architecture that adaptively routes multi-modal features across taxonomy levels, and (ii) a semantic consistency model distilled from large language models acting as domain experts to verify alignment between product titles and official tax definitions. To address noisy supervision in real business records, we design a multi-source training pipeline that combines curated tax databases, invoice validation logs, and merchant registration data to provide both structural and semantic supervision. Extensive experiments on the proprietary TaxCode dataset and public benchmarks demonstrate that Taxon achieves state-of-the-art performance, outperforming strong baselines. Further, an additional full hierarchical paths reconstruction procedure significantly improves structural consistency, yielding the highest overall F1 scores. Taxon has been deployed in production within Alibaba’s tax service system, handling an average of over 500,000 tax code queries per day and reaching peak volumes above five million requests during business event with improved accuracy, interpretability, and robustness.

Method: Theoretical analysis of coverage improvement principle; convergence analysis for on-policy DPO in contextual bandit setting with Bradley-Terry preferences; proposal of hybrid sampler based on preferential G-optimal design; development of principled on-policy schemes for reward distillation.

Result: On-policy DPO converges exponentially with sufficient batch size, while offline methods have slower minimax rates; hybrid sampler achieves convergence in two rounds; experimental confirmation of on-policy superiority.

Conclusion: On-policy training enables coverage improvement leading to exponential convergence, while offline methods are fundamentally limited; novel sampling strategies can further accelerate convergence by removing coverage dependence.

Abstract: Online on-policy preference learning algorithms for language model alignment such as online direct policy optimization (DPO) can significantly outperform their offline counterparts. We provide a theoretical explanation for this phenomenon by analyzing how the sampling policy’s coverage evolves throughout on-policy training. We propose and rigorously justify the \emph{coverage improvement principle}: with sufficient batch size, each update moves into a region around the target where coverage is uniformly better, making subsequent data increasingly informative and enabling rapid convergence. In the contextual bandit setting with Bradley-Terry preferences and linear softmax policy class, we show that on-policy DPO converges exponentially in the number of iterations for batch size exceeding a generalized coverage threshold. In contrast, any learner restricted to offline samples from the initial policy suffers a slower minimax rate, leading to a sharp separation in total sample complexity. Motivated by this analysis, we further propose a simple hybrid sampler based on a novel \emph{preferential} G-optimal design, which removes dependence on coverage and guarantees convergence in just two rounds. Finally, we develop principled on-policy schemes for reward distillation in the general function class setting, and show faster noiseless rates under an alternative deviation-based notion of coverage. Experimentally, we confirm that on-policy DPO and our proposed reward distillation algorithms outperform their off-policy counterparts and enjoy stable, monotonic performance gains across iterations.

Method: 1) Road segment-aware trajectory encoder that jointly embeds input trajectory and surrounding candidate road segments into shared latent space using attention mechanism. 2) One-step diffusion method for map matching using shortcut model with joint embedding as conditioning context.

Result: Extensive experiments on large-scale trajectory datasets show DiffMM consistently outperforms state-of-the-art methods in both accuracy and efficiency, especially for sparse trajectories and complex road network topologies.

Conclusion: DiffMM provides an effective and efficient solution for map matching sparse trajectories through encoder-diffusion framework, addressing limitations of existing methods in handling noisy or sparsely sampled GPS data.

Abstract: Map matching for sparse trajectories is a fundamental problem for many trajectory-based applications, e.g., traffic scheduling and traffic flow analysis. Existing methods for map matching are generally based on Hidden Markov Model (HMM) or encoder-decoder framework. However, these methods continue to face significant challenges when handling noisy or sparsely sampled GPS trajectories. To address these limitations, we propose DiffMM, an encoder-diffusion-based map matching framework that produces effective yet efficient matching results through a one-step diffusion process. We first introduce a road segment-aware trajectory encoder that jointly embeds the input trajectory and its surrounding candidate road segments into a shared latent space through an attention mechanism. Next, we propose a one step diffusion method to realize map matching through a shortcut model by leveraging the joint embedding of the trajectory and candidate road segments as conditioning context. We conduct extensive experiments on large-scale trajectory datasets, demonstrating that our approach consistently outperforms state-of-the-art map matching methods in terms of both accuracy and efficiency, particularly for sparse trajectories and complex road network topologies.

Method: Temporal Fusion Nexus (TFN) is a multi-modal, task-agnostic embedding model designed to integrate irregular time series data with unstructured clinical narratives. The model was evaluated on a retrospective cohort of 3382 kidney transplant patients for predicting three key outcomes: graft loss, graft rejection, and mortality.

Result: TFN achieved superior performance compared to state-of-the-art models: AUC 0.96 vs. 0.94 for graft loss, AUC 0.84 vs. 0.74 for graft rejection, and AUC 0.86 for mortality prediction. It outperformed unimodal baselines by approximately 10% AUC improvement over time-series-only models and 5% improvement over time-series with static data. Integration of clinical text improved performance across all tasks.

Conclusion: TFN effectively integrates heterogeneous clinical data sources and demonstrates strong predictive performance for post-transplant outcomes. The model produces interpretable latent factors aligned with clinical reasoning and has potential applications beyond kidney transplant care in clinical scenarios with similar data characteristics.

Abstract: We introduce Temporal Fusion Nexus (TFN), a multi-modal and task-agnostic embedding model to integrate irregular time series and unstructured clinical narratives. We analysed TFN in post-kidney transplant (KTx) care, with a retrospective cohort of 3382 patients, on three key outcomes: graft loss, graft rejection, and mortality. Compared to state-of-the-art model in post KTx care, TFN achieved higher performance for graft loss (AUC 0.96 vs. 0.94) and graft rejection (AUC 0.84 vs. 0.74). In mortality prediction, TFN yielded an AUC of 0.86. TFN outperformed unimodal baselines (approx 10% AUC improvement over time series only baseline, approx 5% AUC improvement over time series with static patient data). Integrating clinical text improved performance across all tasks. Disentanglement metrics confirmed robust and interpretable latent factors in the embedding space, and SHAP-based attributions confirmed alignment with clinical reasoning. TFN has potential application in clinical tasks beyond KTx, where heterogeneous data sources, irregular longitudinal data, and rich narrative documentation are available.

Method: Proposed History-Aware Adaptive Difficulty Weighting (HA-DW), an adaptive reweighting scheme that adjusts advantage estimates based on an evolving difficulty anchor and training dynamics. This corrects the systematic bias in group-relative advantage estimation.

Result: Theoretical analysis shows HA-DW corrects biased advantage estimation. Experiments on five mathematical reasoning benchmarks demonstrate consistent performance improvements when integrated into GRPO and its variants.

Conclusion: Correcting biased advantage estimation is critical for robust and efficient RLVR training. HA-DW addresses the fundamental issue in group-based RL by providing adaptive reweighting that balances exploration and exploitation.

Abstract: Reinforcement Learning from Verifier Rewards (RLVR) has emerged as a widely used approach for post-training large language models on reasoning tasks, with group-based methods such as GRPO and its variants gaining broad adoption. These methods rely on group-relative advantage estimation to avoid learned critics, yet its theoretical properties remain poorly understood. In this work, we uncover a fundamental issue of group-based RL: the group-relative advantage estimator is inherently biased relative to the true (expected) advantage. We provide the first theoretical analysis showing that it systematically underestimates advantages for hard prompts and overestimates them for easy prompts, leading to imbalanced exploration and exploitation. To address this issue, we propose History-Aware Adaptive Difficulty Weighting (HA-DW), an adaptive reweighting scheme that adjusts advantage estimates based on an evolving difficulty anchor and training dynamics. Both theoretical analysis and experiments on five mathematical reasoning benchmarks demonstrate that HA-DW consistently improves performance when integrated into GRPO and its variants. Our results suggest that correcting biased advantage estimation is critical for robust and efficient RLVR training.

Method: Two-stage approach: 1) Denoising autoencoder for artifact suppression and temporal stabilization, 2) Multitask architecture with convolutional backbone + Transformer encoder for motor imagery classification, chaotic/non-chaotic discrimination (using Lyapunov exponents), and self-supervised contrastive learning with NT-Xent loss.

Result: Framework enhances robustness and generalization, surpasses strong baselines and recent state-of-the-art methods in EEG decoding, demonstrating effectiveness of combining denoising, dynamical features, and self-supervised learning.

Conclusion: The staged multitask learning approach effectively integrates denoising, dynamical modeling, and representation learning for EEG analysis, providing improved stability, reproducibility, and performance over existing methods.

Abstract: We introduce a two-stage multitask learning framework for analyzing Electroencephalography (EEG) signals that integrates denoising, dynamical modeling, and representation learning. In the first stage, a denoising autoencoder is trained to suppress artifacts and stabilize temporal dynamics, providing robust signal representations. In the second stage, a multitask architecture processes these denoised signals to achieve three objectives: motor imagery classification, chaotic versus non-chaotic regime discrimination using Lyapunov exponent-based labels, and self-supervised contrastive representation learning with NT-Xent loss. A convolutional backbone combined with a Transformer encoder captures spatial-temporal structure, while the dynamical task encourages sensitivity to nonlinear brain dynamics. This staged design mitigates interference between reconstruction and discriminative goals, improves stability across datasets, and supports reproducible training by clearly separating noise reduction from higher-level feature learning. Empirical studies show that our framework not only enhances robustness and generalization but also surpasses strong baselines and recent state-of-the-art methods in EEG decoding, highlighting the effectiveness of combining denoising, dynamical features, and self-supervised learning.

Method: Extension of evidential learning integrated with Neural Additive Models (NAMs). Enables simultaneous estimation of aleatoric and epistemic uncertainty along with explicit feature contributions in a single forward pass.

Result: EviNAM matches state-of-the-art predictive performance on synthetic and real data while providing interpretable uncertainty estimates and feature contributions.

Conclusion: EviNAM offers a path toward more intelligible and trustworthy predictions, naturally extending to classification and generalized additive models beyond the regression focus.

Abstract: Intelligibility and accurate uncertainty estimation are crucial for reliable decision-making. In this paper, we propose EviNAM, an extension of evidential learning that integrates the interpretability of Neural Additive Models (NAMs) with principled uncertainty estimation. Unlike standard Bayesian neural networks and previous evidential methods, EviNAM enables, in a single pass, both the estimation of the aleatoric and epistemic uncertainty as well as explicit feature contributions. Experiments on synthetic and real data demonstrate that EviNAM matches state-of-the-art predictive performance. While we focus on regression, our method extends naturally to classification and generalized additive models, offering a path toward more intelligible and trustworthy predictions.

Method: Three-module architecture: (1) Multi-view frequency mixture-of-experts with Fourier and Wavelet domain experts for distinct spectral bands, (2) Multi-resolution adaptive fusion that hierarchically aggregates frequency features from coarse to fine resolutions, (3) Temporal gating integration that dynamically balances long-term trends and short-term frequency-aware features.

Result: Experiments on real-world hydrological datasets with extreme patterns show M²FMoE outperforms state-of-the-art baselines without requiring extreme-event labels.

Conclusion: M²FMoE effectively addresses the challenge of forecasting time series with extreme events by learning both regular and extreme patterns through multi-resolution and multi-view frequency modeling, achieving superior performance without needing labeled extreme events.

Abstract: Forecasting time series with extreme events is critical yet challenging due to their high variance, irregular dynamics, and sparse but high-impact nature. While existing methods excel in modeling dominant regular patterns, their performance degrades significantly during extreme events, constituting the primary source of forecasting errors in real-world applications. Although some approaches incorporate auxiliary signals to improve performance, they still fail to capture extreme events’ complex temporal dynamics. To address these limitations, we propose M$^2$FMoE, an extreme-adaptive forecasting model that learns both regular and extreme patterns through multi-resolution and multi-view frequency modeling. It comprises three modules: (1) a multi-view frequency mixture-of-experts module assigns experts to distinct spectral bands in Fourier and Wavelet domains, with cross-view shared band splitter aligning frequency partitions and enabling inter-expert collaboration to capture both dominant and rare fluctuations; (2) a multi-resolution adaptive fusion module that hierarchically aggregates frequency features from coarse to fine resolutions, enhancing sensitivity to both short-term variations and sudden changes; (3) a temporal gating integration module that dynamically balances long-term trends and short-term frequency-aware features, improving adaptability to both regular and extreme temporal patterns. Experiments on real-world hydrological datasets with extreme patterns demonstrate that M$^2$FMoE outperforms state-of-the-art baselines without requiring extreme-event labels.

Method: Two algorithms: 1) OFU-based algorithm for safe RL, 2) Main algorithm with entropy regularization building on the first. Both ensure safety constraints with arbitrarily high probability during learning.

Result: Finite-sample analysis shows both algorithms have regret bounds. Entropy regularization improves regret and significantly reduces episode-to-episode variability inherent in OFU-based safe RL algorithms.

Conclusion: Entropy regularization is effective for safe RL, providing better regret bounds and stability compared to standard OFU approaches while maintaining safety guarantees during learning.

Abstract: We consider the problem of learning the optimal policy for Markov decision processes with safety constraints. We formulate the problem in a reach-avoid setup. Our goal is to design online reinforcement learning algorithms that ensure safety constraints with arbitrarily high probability during the learning phase. To this end, we first propose an algorithm based on the optimism in the face of uncertainty (OFU) principle. Based on the first algorithm, we propose our main algorithm, which utilizes entropy regularization. We investigate the finite-sample analysis of both algorithms and derive their regret bounds. We demonstrate that the inclusion of entropy regularization improves the regret and drastically controls the episode-to-episode variability that is inherent in OFU-based safe RL algorithms.

Method: Used reconstruction-based anomaly detection with convolutional autoencoders: compared 2D autoencoder on individual frames vs 3D autoencoder on short temporal stacks for Kármán vortex street simulations.

Result: 2D model identifies localized spatial irregularities in single time steps, while 3D model exploits spatio-temporal context to detect anomalous motion patterns and reduces redundant detections across time. Reconstruction errors are strongly influenced by spatial distribution of mass.

Conclusion: Temporal context is crucial for robust anomaly detection in dynamic simulations, with 3D autoencoders providing better performance by capturing spatio-temporal patterns.

Abstract: Detecting anomalies in high-dimensional, time-dependent simulation data is challenging due to complex spatial and temporal dynamics. We study reconstruction-based anomaly detection for ensemble data from parameterized Kármán vortex street simulations using convolutional autoencoders. We compare a 2D autoencoder operating on individual frames with a 3D autoencoder that processes short temporal stacks. The 2D model identifies localized spatial irregularities in single time steps, while the 3D model exploits spatio-temporal context to detect anomalous motion patterns and reduces redundant detections across time. We further evaluate volumetric time-dependent data and find that reconstruction errors are strongly influenced by the spatial distribution of mass, with highly concentrated regions yielding larger errors than dispersed configurations. Our results highlight the importance of temporal context for robust anomaly detection in dynamic simulations.

Method: Soft partition-based kernel adaptation with smooth partition lengths controlling collocation centers and Gaussian kernel widths, enabling continuous coarse-to-fine resolution. Uses signed-distance weighting for irregular geometries and requires only a single linear solve.

Result: Outperforms or matches state-of-the-art PINN and TFC variants across eight benchmarks including oscillatory ODEs, high-frequency Poisson equations, irregular domains, and stiff singularly perturbed convection-diffusion problems.

Conclusion: Soft-partition kernel adaptation provides a fast, architecture-free approach for multiscale PDEs with broad potential for future physics-informed modeling, demonstrated on steady linear PDEs.

Abstract: Physics-informed machine learning holds great promise for solving differential equations, yet existing methods struggle with highly oscillatory, multiscale, or singularly perturbed PDEs due to spectral bias, costly backpropagation, and manually tuned kernel or Fourier frequencies. This work introduces a soft partition–based Kernel-Adaptive Physics-Informed Extreme Learning Machine (KAPI-ELM), a deterministic low-dimensional parameterization in which smooth partition lengths jointly control collocation centers and Gaussian kernel widths, enabling continuous coarse-to-fine resolution without Fourier features, random sampling, or hard domain interfaces. A signed-distance-based weighting further stabilizes least-squares learning on irregular geometries. Across eight benchmarks–including oscillatory ODEs, high-frequency Poisson equations, irregular-shaped domains, and stiff singularly perturbed convection-diffusion problems-the proposed method matches or exceeds the accuracy of state-of-the-art Physics-Informed Neural Network (PINN) and Theory of Functional Connections (TFC) variants while using only a single linear solve. Although demonstrated on steady linear PDEs, the results show that soft-partition kernel adaptation provides a fast, architecture-free approach for multiscale PDEs with broad potential for future physics-informed modeling. For reproducibility, the reference codes are available at https://github.com/vikas-dwivedi-2022/soft_kapi

Method: Introduce explicit time dependence into the learning process by allowing the neural network’s function approximation to incorporate temporal information, enabling agents to estimate value functions consistent with the process’s intrinsic growth rate.

Result: Deep RL implementations also produce suboptimal policies under non-ergodic dynamics, but incorporating temporal information into function approximation corrects this limitation without requiring reward transformations or modified objective functions.

Conclusion: Adding temporal dependence to deep RL architectures enables optimal policy learning in non-ergodic environments by aligning value estimation with time-average growth rather than ensemble averages, advancing RL methods for non-ergodic systems.

Abstract: Reinforcement Learning (RL) remains a central optimisation framework in machine learning. Although RL agents can converge to optimal solutions, the definition of ``optimality’’ depends on the environment’s statistical properties. The Bellman equation, central to most RL algorithms, is formulated in terms of expected values of future rewards. However, when ergodicity is broken, long-term outcomes depend on the specific trajectory rather than on the ensemble average. In such settings, the ensemble average diverges from the time-average growth experienced by individual agents, with expected-value formulations yielding systematically suboptimal policies. Prior studies demonstrated that traditional RL architectures fail to recover the true optimum in non-ergodic environments. We extend this analysis to deep RL implementations and show that these, too, produce suboptimal policies under non-ergodic dynamics. Introducing explicit time dependence into the learning process can correct this limitation. By allowing the network’s function approximation to incorporate temporal information, the agent can estimate value functions consistent with the process’s intrinsic growth rate. This improvement does not require altering the environmental feedback, such as reward transformations or modified objective functions, but arises naturally from the agent’s exposure to temporal trajectories. Our results contribute to the growing body of research on reinforcement learning methods for non-ergodic systems.

Method: Proposed explainable AI framework comparing Quantum Boltzmann Machines (QBMs) and Classical Boltzmann Machines (CBMs). Used binarised, PCA-reduced MNIST dataset. QBMs used hybrid quantum-classical circuits with strongly entangling layers; CBMs used contrastive divergence. For interpretability: gradient-based saliency maps for QBMs and SHAP for CBMs to evaluate feature attributions.

Result: QBMs significantly outperformed CBMs in classification accuracy (83.5% vs 54%). QBMs also showed more concentrated feature attribution distributions (entropy 1.27 vs 1.39), indicating clearer identification of important features behind predictions.

Conclusion: Quantum-classical hybrid models can improve both accuracy and interpretability, moving toward more trustworthy and explainable AI systems. QBMs provide better predictive performance and clearer identification of “active ingredients” in model decisions.

Abstract: Artificial Intelligence (AI) systems have shown good success at classifying. However, the lack of explainability is a true and significant challenge, especially in high-stakes domains, such as health and finance, where understanding is paramount. We propose a new solution to this challenge: an explainable AI framework based on our comparative study with Quantum Boltzmann Machines (QBMs) and Classical Boltzmann Machines (CBMs). We leverage principles of quantum computing within classical machine learning to provide substantive transparency around decision-making. The design involves training both models on a binarised and dimensionally reduced MNIST dataset, where Principal Component Analysis (PCA) is applied for preprocessing. For interpretability, we employ gradient-based saliency maps in QBMs and SHAP (SHapley Additive exPlanations) in CBMs to evaluate feature attributions.QBMs deploy hybrid quantum-classical circuits with strongly entangling layers, allowing for richer latent representations, whereas CBMs serve as a classical baseline that utilises contrastive divergence. Along the way, we found that QBMs outperformed CBMs on classification accuracy (83.5% vs. 54%) and had more concentrated distributions in feature attributions as quantified by entropy (1.27 vs. 1.39). In other words, QBMs not only produced better predictive performance than CBMs, but they also provided clearer identification of “active ingredient” or the most important features behind model predictions. To conclude, our results illustrate that quantum-classical hybrid models can display improvements in both accuracy and interpretability, which leads us toward more trustworthy and explainable AI systems.

Method: Uniqueness-Aware Reinforcement Learning uses an LLM-based judge to cluster rollouts by high-level solution strategies, ignoring superficial variations, and reweights policy advantages inversely with cluster size to reward novel strategies.

Result: Across math, physics, and medical reasoning benchmarks, the approach improves pass@k across large sampling budgets, increases AUC@K without sacrificing pass@1, and sustains exploration of diverse strategies.

Conclusion: Explicitly rewarding rollout-level diversity over solution strategies addresses exploration collapse in RL for LLMs, enabling better pass@k performance while maintaining pass@1 quality.

Abstract: Reinforcement learning (RL) has become a central paradigm for post-training large language models (LLMs), particularly for complex reasoning tasks, yet it often suffers from exploration collapse: policies prematurely concentrate on a small set of dominant reasoning patterns, improving pass@1 while limiting rollout-level diversity and gains in pass@k. We argue that this failure stems from regularizing local token behavior rather than diversity over sets of solutions. To address this, we propose Uniqueness-Aware Reinforcement Learning, a rollout-level objective that explicitly rewards correct solutions that exhibit rare high-level strategies. Our method uses an LLM-based judge to cluster rollouts for the same problem according to their high-level solution strategies, ignoring superficial variations, and reweights policy advantages inversely with cluster size. As a result, correct but novel strategies receive higher rewards than redundant ones. Across mathematics, physics, and medical reasoning benchmarks, our approach consistently improves pass@$k$ across large sampling budgets and increases the area under the pass@$k$ curve (AUC@$K$) without sacrificing pass@1, while sustaining exploration and uncovering more diverse solution strategies at scale.

Method: Formalizes (k,f(k))-robust alignment and asymptotic universal alignment (U-alignment). Proposes symmetric multi-player alignment games where any symmetric Nash equilibrium policy of the (k+1)-player game achieves optimal (k,k/(k+1))-robust alignment.

Result: Characterizes optimal convergence rate: k/(k+1) is achievable and no method can achieve faster rate in general. Shows NLHF and other post-training methods fundamentally underutilize test-time scaling, often achieving win rates barely above 1/2.

Conclusion: Symmetric multi-player alignment games preserve output diversity and achieve optimal test-time scaling rates, overcoming limitations of existing methods that collapse to single responses. Theoretical convergence guarantees for self-play learning are provided.

Abstract: Aligning large language models (LLMs) to serve users with heterogeneous and potentially conflicting preferences is a central challenge for personalized and trustworthy AI. We formalize an ideal notion of universal alignment through test-time scaling: for each prompt, the model produces $k\ge 1$ candidate responses and a user selects their preferred one. We introduce $(k,f(k))$-robust alignment, which requires the $k$-output model to have win rate $f(k)$ against any other single-output model, and asymptotic universal alignment (U-alignment), which requires $f(k)\to 1$ as $k\to\infty$. Our main result characterizes the optimal convergence rate: there exists a family of single-output policies whose $k$-sample product policies achieve U-alignment at rate $f(k)=\frac{k}{k+1}$, and no method can achieve a faster rate in general. We show that popular post-training methods, including Nash learning from human feedback (NLHF), can fundamentally underutilize the benefits of test-time scaling. Even though NLHF is optimal for $k=1$, sampling from the resulting (often deterministic) policy cannot guarantee win rates above $\tfrac{1}{2}$ except for an arbitrarily small slack. This stems from a lack of output diversity: existing alignment methods can collapse to a single majority-preferred response, making additional samples redundant. In contrast, our approach preserves output diversity and achieves the optimal test-time scaling rate. In particular, we propose a family of symmetric multi-player alignment games and prove that any symmetric Nash equilibrium policy of the $(k+1)$-player alignment game achieves the optimal $(k,\frac{k}{k+1})$-robust alignment. Finally, we provide theoretical convergence guarantees for self-play learning dynamics in these games and extend the framework to opponents that also generate multiple responses.

Method: Instead of using principal components for sorting, the authors use appropriate norms of data vectors as sorting criteria combined with triangle inequality for search termination. For Tanimoto distance specifically, they employ a provably sharper intersection inequality to further boost algorithm performance.

Result: On a real-world chemical fingerprint benchmark, CLASSIX with Tanimoto distance is about 30 times faster than Taylor-Butina algorithm and about 80 times faster than DBSCAN, while computing higher-quality clusters in both cases.

Conclusion: The extended CLASSIX algorithm successfully supports multiple distance metrics with significant performance improvements over existing methods while maintaining or improving cluster quality, making it a versatile and efficient clustering solution for various applications.

Abstract: The CLASSIX algorithm is a fast and explainable approach to data clustering. In its original form, this algorithm exploits the sorting of the data points by their first principal component to truncate the search for nearby data points, with nearness being defined in terms of the Euclidean distance. Here we extend CLASSIX to other distance metrics, including the Manhattan distance and the Tanimoto distance. Instead of principal components, we use an appropriate norm of the data vectors as the sorting criterion, combined with the triangle inequality for search termination. In the case of Tanimoto distance, a provably sharper intersection inequality is used to further boost the performance of the new algorithm. On a real-world chemical fingerprint benchmark, CLASSIX Tanimoto is about 30 times faster than the Taylor–Butina algorithm, and about 80 times faster than DBSCAN, while computing higher-quality clusters in both cases.

Method: Proposes Gromov-Wasserstein Imitation Learning (GWIL) that uses Gromov-Wasserstein distance to align and compare states between different agent spaces, enabling cross-domain imitation.

Result: Theoretical analysis characterizes scenarios where GWIL preserves optimality, and empirical demonstrations show effectiveness in continuous control domains from simple rigid transformations to arbitrary state-action space transformations.

Conclusion: GWIL provides a principled approach for cross-domain imitation learning by leveraging optimal transport theory to bridge different agent embodiments while preserving optimality in appropriate scenarios.

Abstract: Cross-domain imitation learning studies how to leverage expert demonstrations of one agent to train an imitation agent with a different embodiment or morphology. Comparing trajectories and stationary distributions between the expert and imitation agents is challenging because they live on different systems that may not even have the same dimensionality. We propose Gromov-Wasserstein Imitation Learning (GWIL), a method for cross-domain imitation that uses the Gromov-Wasserstein distance to align and compare states between the different spaces of the agents. Our theory formally characterizes the scenarios where GWIL preserves optimality, revealing its possibilities and limitations. We demonstrate the effectiveness of GWIL in non-trivial continuous control domains ranging from simple rigid transformation of the expert domain to arbitrary transformation of the state-action space.

Method: F³ uses fixed random feedback paths (no weight transport) and delayed error information from the previous epoch (avoiding update locking) to approximate gradients for training feed-forward neural networks.

Result: F³ significantly improves predictive performance compared to similarly plausible approaches, narrowing the gap to backpropagation by up to 56% for classification and 96% for regression across various tasks and architectures including fully-connected and Transformer networks.

Conclusion: F³ represents a step towards more biologically plausible learning algorithms while opening new avenues for energy-efficient and parallelizable neural network training, balancing biological plausibility with competitive performance.

Abstract: Backpropagation has long been criticized for being biologically implausible due to its reliance on concepts that are not viable in natural learning processes. Two core issues are the weight transport and update locking problems caused by the forward-backward dependencies, which limit biological plausibility, computational efficiency, and parallelization. Although several alternatives have been proposed to increase biological plausibility, they often come at the cost of reduced predictive performance. This paper proposes an alternative approach to training feed-forward neural networks addressing these issues by using approximate gradient information. We introduce Feed-Forward with delayed Feedback (F$^3$), which approximates gradients using fixed random feedback paths and delayed error information from the previous epoch to balance biological plausibility with predictive performance. We evaluate F$^3$ across multiple tasks and architectures, including both fully-connected and Transformer networks. Our results demonstrate that, compared to similarly plausible approaches, F$^3$ significantly improves predictive performance, narrowing the gap to backpropagation by up to 56% for classification and 96% for regression. This work is a step towards more biologically plausible learning algorithms while opening up new avenues for energy-efficient and parallelizable neural network training.

Method: Uses a threat model similar to backdoor attacks where an attacker controls a small subset of FL clients (even just one client) to influence the aggregated model’s performance distribution between attributes.

Result: The attack successfully creates unfair performance distribution between any given set of attributes. The attack is possible with control of only a single client, making it highly feasible.

Conclusion: Defending against attacks on fairness should be a critical consideration in FL, especially in situations where unfairness in a trained model could benefit participants. This highlights a new security vulnerability beyond traditional backdoor attacks.

Abstract: Federated Learning is an important emerging distributed training paradigm that keeps data private on clients. It is now well understood that by controlling only a small subset of FL clients, it is possible to introduce a backdoor to a federated learning model, in the presence of certain attributes. In this paper, we present a new type of attack that compromises the fairness of the trained model. Fairness is understood to be the attribute-level performance distribution of a trained model. It is particularly salient in domains where, for example, skewed accuracy discrimination between subpopulations could have disastrous consequences. We find that by employing a threat model similar to that of a backdoor attack, an attacker is able to influence the aggregated model to have an unfair performance distribution between any given set of attributes. Furthermore, we find that this attack is possible by controlling only a single client. While combating naturally induced unfairness in FL has previously been discussed in depth, its artificially induced kind has been neglected. We show that defending against attacks on fairness should be a critical consideration in any situation where unfairness in a trained model could benefit a user who participated in its training.

Method: Cast malware detection as a zeroth-order optimization framework that incorporates functionality-preserving manipulations. Propose ZEXE, a novel zeroth-order attack using gradient-free optimization algorithms with theoretical guarantees and minimal hyper-parameter tuning.

Result: ZEXE achieves higher evasion rates compared to state-of-the-art techniques while reducing injected content size to less than one third. The approach provides theoretical guarantees and requires minimal hyper-parameter tuning.

Conclusion: Zeroth-order optimization provides a sound framework for generating functionality-preserving adversarial malware examples, with ZEXE demonstrating superior performance in evasion rates and efficiency compared to existing heuristic methods.

Abstract: Machine learning malware detectors are vulnerable to adversarial EXEmples, i.e., carefully-crafted Windows programs tailored to evade detection. Unlike other adversarial problems, attacks in this context must be functionality-preserving, a constraint that is challenging to address. As a consequence, heuristic algorithms are typically used, which inject new content, either randomly-picked or harvested from legitimate programs. In this paper, we show how learning malware detectors can be cast within a zeroth-order optimization framework, which allows incorporating functionality-preserving manipulations. This permits the deployment of sound and efficient gradient-free optimization algorithms, which come with theoretical guarantees and allow for minimal hyper-parameters tuning. As a by-product, we propose and study ZEXE, a novel zeroth-order attack against Windows malware detection. Compared to state-of-the-art techniques, ZEXE provides improvement in the evasion rate, reducing to less than one third the size of the injected content.

Method: Uses Group SELFIES string representation for chemically meaningful semantics; combines self-attention weights and gradients to quantify substructure impacts; introduces marginal loss to align explanations with chemists’ annotations; bridges manifold hypothesis with loss function for concept-aligned explanations.

Result: Achieves comparable classification accuracy on six mutagenicity datasets and one hepatotoxicity dataset while boosting explanation accuracy by up to 14.3%, becoming state-of-the-art in explainable molecular property prediction.

Conclusion: Lamole provides a framework for explainable molecular property prediction with chemically meaningful, concept-aligned explanations that faithfully reveal structure-property relationships, advancing interpretability in scientific applications.

Abstract: Providing explainable molecular property predictions is critical for many scientific domains, such as drug discovery and material science. Though transformer-based language models have shown great potential in accurate molecular property prediction, they neither provide chemically meaningful explanations nor faithfully reveal the molecular structure-property relationships. In this work, we develop a framework for explainable molecular property prediction based on language models, dubbed as Lamole, which can provide chemical concepts-aligned explanations. We take a string-based molecular representation – Group SELFIES – as input tokens to pretrain and fine-tune our Lamole, as it provides chemically meaningful semantics. By disentangling the information flows of Lamole, we propose combining self-attention weights and gradients for better quantification of each chemically meaningful substructure’s impact on the model’s output. To make the explanations more faithfully respect the structure-property relationship, we then carefully craft a marginal loss to explicitly optimize the explanations to be able to align with the chemists’ annotations. We bridge the manifold hypothesis with the elaborated marginal loss to prove that the loss can align the explanations with the tangent space of the data manifold, leading to concept-aligned explanations. Experimental results over six mutagenicity datasets and one hepatotoxicity dataset demonstrate Lamole can achieve comparable classification accuracy and boost the explanation accuracy by up to 14.3%, being the state-of-the-art in explainable molecular property prediction.

Method: Implemented and benchmarked efficient DP-SGD methods with correct Poisson subsampling, comparing naive Opacus implementation with optimized approaches like Ghost Clipping gradient clipping and a JAX-based implementation.

Result: Naive DP-SGD with Opacus has 2.6-8x lower throughput than SGD, but Ghost Clipping reduces this cost by roughly half. The JAX implementation with Poisson subsampling performs comparably to efficient PyTorch optimizations. DP-SGD scales better than SGD on up to 80 GPUs.

Conclusion: Efficient implementations can significantly reduce the computational overhead of correct DP-SGD training while maintaining theoretical guarantees, with DP-SGD showing better scaling behavior than standard SGD on distributed systems.

Abstract: Differentially private stochastic gradient descent (DP-SGD) is the standard algorithm for training machine learning models under differential privacy (DP). The most common DP-SGD privacy accountants rely on Poisson subsampling to ensure the theoretical DP guarantees. Implementing computationally efficient DP-SGD with Poisson subsampling is not trivial, which leads many implementations to taking a shortcut by using computationally faster subsampling. We quantify the computational cost of training deep learning models under DP by implementing and benchmarking efficient methods with the correct Poisson subsampling. We find that using the naive implementation of DP-SGD with Opacus in PyTorch has a throughput between 2.6 and 8 times lower than that of SGD. However, efficient gradient clipping implementations like Ghost Clipping can roughly halve this cost. We propose an alternative computationally efficient implementation of DP-SGD with JAX that uses Poisson subsampling and performs comparably with efficient clipping optimizations based on PyTorch. We study the scaling behavior using up to 80 GPUs and find that DP-SGD scales better than SGD. We share our library at https://github.com/DPBayes/Towards-Efficient-Scalable-Training-DP-DL.

Method: The paper shows that standard gradient flow in parameter space can be continuously deformed into an adapted gradient flow that yields constrained Euclidean gradient flow in output space. For L² loss with full-rank Jacobian, time reparametrization yields linear interpolation. For cross-entropy loss with full-rank Jacobian and positive labels, an explicit formula for the unique global minimum is derived.

Result: Under full-rank Jacobian conditions: 1) For L² loss, gradient flow reduces to linear interpolation in output space, achieving global minimum. 2) For cross-entropy loss with positive labels, an explicit formula for the unique global minimum is obtained.

Conclusion: The geometric transformation of gradient flow reveals simplified optimization dynamics in deep learning, showing that under certain conditions, training can be understood as linear interpolation in output space with explicit global minima, providing theoretical insights into optimization behavior.

Abstract: We prove that the standard gradient flow in parameter space that underlies many training algorithms in deep learning can be continuously deformed into an adapted gradient flow which yields (constrained) Euclidean gradient flow in output space. Moreover, for the $L^{2}$ loss, if the Jacobian of the outputs with respect to the parameters is full rank (for fixed training data), then the time variable can be reparametrized so that the resulting flow is simply linear interpolation, and a global minimum can be achieved. For the cross-entropy loss, under the same rank condition and assuming the labels have positive components, we derive an explicit formula for the unique global minimum.

Method: Introduces multi-tangent forward gradients with an improved approach to combine forward gradients from multiple tangents using orthogonal projections, rather than using just a single tangent per step.

Result: Increasing the number of tangents improves both approximation quality and optimization performance across various tasks, demonstrating the superiority of multi-tangent approaches over single-tangent methods.

Conclusion: Multi-tangent forward gradients with orthogonal projection combination provide a more effective alternative to backpropagation, offering better gradient approximation and optimization performance while addressing backpropagation’s limitations.

Abstract: The gradients used to train neural networks are typically computed using backpropagation. While an efficient way to obtain exact gradients, backpropagation is computationally expensive, hinders parallelization, and is biologically implausible. Forward gradients are an approach to approximate the gradients from directional derivatives along random tangents computed by forward-mode automatic differentiation. So far, research has focused on using a single tangent per step. This paper provides an in-depth analysis of multi-tangent forward gradients and introduces an improved approach to combining the forward gradients from multiple tangents based on orthogonal projections. We demonstrate that increasing the number of tangents improves both approximation quality and optimization performance across various tasks.

Method: ROSS aggregates cross-gradient information from neighbors (derivatives of local model with respect to neighbors’ datasets) and weights these derivatives using Shapley values to measure their contributions. Updates are performed in a momentum-like manner.

Result: Theoretical analysis reveals linear convergence speedup. Extensive experiments on public datasets show ROSS significantly outperforms state-of-the-art methods in both convergence rate and prediction accuracy under various data challenges.

Conclusion: ROSS provides an effective robust decentralized learning solution that handles data heterogeneity, noise, and poisoning through Shapley value-based gradient weighting, offering practical advantages for real-world decentralized learning scenarios.

Abstract: In the paradigm of decentralized learning, a group of agents collaborate to learn a global model using a distributed dataset without a central server; nevertheless, it is severely challenged by the heterogeneity of the data distribution across the agents. For example, the data may be distributed non-independently and identically, and even be noised or poisoned. To address these data challenges, we propose ROSS, a novel robust decentralized stochastic learning algorithm based on Shapley values, in this paper. Specifically, in each round, each agent aggregates the cross-gradient information from its neighbors, i.e., the derivatives of its local model with respect to the datasets of its neighbors, to update its local model in a momentum like manner, while we innovate in weighting the derivatives according to their contributions measured by Shapley values. We perform solid theoretical analysis to reveal the linear convergence speedup of our ROSS algorithm. We also verify the efficacy of our algorithm through extensive experiments on public datasets. Our results demonstrate that, in face of the above variety of data challenges, our ROSS algorithm has significant advantages over existing state-of-the-art proposals in terms of both convergence and prediction accuracy.

Method: Proposes CausAdv framework that learns both causal and non-causal features, quantifies counterfactual information (CI) of filters in the last convolutional layer, and performs statistical analysis of CI distributions across clean vs. adversarial samples.

Result: Adversarial examples exhibit different CI distributions compared to clean samples. Causal reasoning enhances adversarial detection without needing separate detector training. Visualizations show causal features as effective detection tools.

Conclusion: Causal reasoning provides an effective approach for adversarial detection in CNNs, offering both detection capability and explanatory visualizations without additional detector training.

Abstract: Deep learning has led to tremendous success in computer vision, largely due to Convolutional Neural Networks (CNNs). However, CNNs have been shown to be vulnerable to crafted adversarial perturbations. This vulnerability of adversarial examples has has motivated research into improving model robustness through adversarial detection and defense methods. In this paper, we address the adversarial robustness of CNNs through causal reasoning. We propose CausAdv: a causal framework for detecting adversarial examples based on counterfactual reasoning. CausAdv learns both causal and non-causal features of every input, and quantifies the counterfactual information (CI) of every filter of the last convolutional layer. We then perform a statistical analysis of the filters’ CI across clean and adversarial samples, to demonstrate that adversarial examples exhibit different CI distributions compared to clean samples. Our results show that causal reasoning enhances the process of adversarial detection without the need to train a separate detector. Moreover, we illustrate the efficiency of causal explanations as a helpful detection tool by visualizing the extracted causal features.

Method: DeepMaxent uses neural networks to automatically learn shared features among species based on maximum entropy principle. It employs a normalized Poisson loss where presence probabilities across sites are modeled by a neural network, enabling automatic feature extraction through backpropagation and stochastic gradient descent.

Result: DeepMaxent outperforms Maxent and other leading SDMs across all six regions and taxonomic groups in benchmark testing. It performs particularly well in regions with uneven sampling and yields more accurate predictions than traditional single-species models.

Conclusion: DeepMaxent demonstrates substantial potential to increase SDM performance by combining neural networks’ automatic feature extraction with maximum entropy principles, especially for handling sampling biases in presence-only data and enabling multi-species modeling.

Abstract: The rapid expansion of citizen science initiatives has led to a significant growth of biodiversity databases, and particularly presence-only (PO) observations. PO data are invaluable for understanding species distributions and their dynamics, but their use in a Species Distribution Model (SDM) is curtailed by sampling biases and the lack of information on absences. Poisson point processes are widely used for SDMs, with Maxent being one of the most popular methods. Maxent maximises the entropy of a probability distribution across sites as a function of predefined transformations of variables, called features. In contrast, neural networks and deep learning have emerged as a promising technique for automatic feature extraction from complex input variables. Arbitrarily complex transformations of input variables can be learned from the data efficiently through backpropagation and stochastic gradient descent (SGD). In this paper, we propose DeepMaxent, which harnesses neural networks to automatically learn shared features among species, using the maximum entropy principle. To do so, it employs a normalised Poisson loss where for each species, presence probabilities across sites are modelled by a neural network. We evaluate DeepMaxent on a benchmark dataset known for its spatial sampling biases, using PO data for calibration and presence-absence (PA) data for validation across six regions with different biological groups and covariates. Our results indicate that DeepMaxent performs better than Maxent and other leading SDMs across all regions and taxonomic groups. The method performs particularly well in regions of uneven sampling, demonstrating substantial potential to increase SDM performances. In particular, our approach yields more accurate predictions than traditional single-species models, which opens up new possibilities for methodological enhancement.

Method: 1) Introduced bipartite receptive field (BRF) - a brain-inspired network model for initializing sparse ANN connectivity. 2) Developed GPU-friendly matrix-based approximation of CH link prediction, reducing complexity to O(N³). 3) Proposed Cannistraci-Hebb training soft rule (CHTs) with flexible sampling strategy for link removal/regrowth. 4) Integrated CHTs with sigmoid gradual density decay (CHTss).

Result: 1) BRF outperforms previous network science models. 2) CHTs with 1% connections outperforms fully connected MLPs on image classification, compressing networks to <30% nodes. 3) CHTss with 5% connections outperforms fully connected networks in Transformer-based machine translation. 4) At 30% connectivity, both CHTs and CHTss outperform other DST methods in language modeling.

Conclusion: The proposed brain-inspired methods successfully address limitations of existing DST approaches, enabling high-performance sparse training across various architectures (MLPs, Transformers) and tasks (image classification, machine translation, language modeling) at different sparsity levels.

Abstract: Dynamic sparse training (DST) can reduce the computational demands in ANNs, but faces difficulties in keeping peak performance at high sparsity levels. The Cannistraci-Hebb training (CHT) is a brain-inspired method for growing connectivity in DST. CHT leverages a gradient-free, topology-driven link regrowth, which has shown ultra-sparse (less than 1% connectivity) advantage across various tasks compared to fully connected networks. Yet, CHT suffers two main drawbacks: (i) its time complexity is $O(Nd^3)$ - N node network size, d node degree - restricting it to ultra-sparse regimes. (ii) it selects top link prediction scores, which is inappropriate for the early training epochs, when the network presents unreliable connections. Here, we design the first brain-inspired network model - termed bipartite receptive field (BRF) - to initialize the connectivity of sparse artificial neural networks. We further introduce a GPU-friendly matrix-based approximation of CH link prediction, reducing complexity to $O(N^3)$. We introduce the Cannistraci-Hebb training soft rule (CHTs), which adopts a flexible strategy for sampling connections in both link removal and regrowth, balancing the exploration and exploitation of network topology. Additionally, we integrate CHTs with a sigmoid gradual density decay (CHTss). Empirical results show that BRF offers performance advantages over previous network science models. Using 1% of connections, CHTs outperforms fully connected networks in MLP architectures on image classification tasks, compressing some networks to less than 30% of the nodes. Using 5% of the connections, CHTss outperforms fully connected networks in two Transformer-based machine translation tasks. Finally, at 30% connectivity, both CHTs and CHTss outperform other DST methods in language modeling task.

Method: The authors develop a general technique to convert optimization algorithms with memory into memoryless approximations. This is done by replacing all past iterates in the update with the current iterate and adding a correction term arising from memory. This correction term can be interpreted as a perturbation of the loss function.

Result: The technique reveals how memory implicitly (anti-)regularizes optimization dynamics. As a key application, the authors find that Lion does not have the implicit anti-regularization induced by memory that AdamW does, providing a theoretical explanation for Lion’s better generalization performance observed in practice.

Conclusion: The proposed framework provides a way to understand how memory in optimization algorithms affects their dynamics and generalization properties. The analysis explains why Lion generalizes better than AdamW by showing that AdamW’s memory induces anti-regularization while Lion’s does not.

Abstract: In modern optimization methods used in deep learning, each update depends on the history of previous iterations, often referred to as memory, and this dependence decays fast as the iterates go further into the past. For example, gradient descent with momentum has exponentially decaying memory through exponentially averaged past gradients. We introduce a general technique for identifying a memoryless algorithm that approximates an optimization algorithm with memory. It is obtained by replacing all past iterates in the update by the current one, and then adding a correction term arising from memory (also a function of the current iterate). This correction term can be interpreted as a perturbation of the loss, and the nature of this perturbation can inform how memory implicitly (anti-)regularizes the optimization dynamics. As an application of our theory, we find that Lion does not have the kind of implicit anti-regularization induced by memory that AdamW does, providing a theory-based explanation for Lion’s better generalization performance recently documented.

Method: Theoretical analysis of gradient-based training on AIMC hardware with non-ideal response functions, proposing Residual Learning algorithm that solves a bilevel optimization problem to overcome asymmetric response function issues.

Result: Asymmetric response functions negatively impact Analog SGD by imposing implicit penalty on objective; Residual Learning algorithm provably converges exactly to critical point; method extends to address other hardware imperfections like limited response granularity.

Conclusion: First paper to investigate impact of generic non-ideal response functions in AIMC training, providing theoretical foundation and practical algorithm for overcoming hardware imperfections in energy-efficient analog computing.

Abstract: As the economic and environmental costs of training and deploying large vision or language models increase dramatically, analog in-memory computing (AIMC) emerges as a promising energy-efficient solution. However, the training perspective, especially its training dynamic, is underexplored. In AIMC hardware, the trainable weights are represented by the conductance of resistive elements and updated using consecutive electrical pulses. While the conductance changes by a constant in response to each pulse, in reality, the change is scaled by asymmetric and non-linear response functions, leading to a non-ideal training dynamic. This paper provides a theoretical foundation for gradient-based training on AIMC hardware with non-ideal response functions. We demonstrate that asymmetric response functions negatively impact Analog SGD by imposing an implicit penalty on the objective. To overcome the issue, we propose Residual Learning algorithm, which provably converges exactly to a critical point by solving a bilevel optimization problem. We demonstrate that the proposed method can be extended to address other hardware imperfections, such as limited response granularity. As we know, it is the first paper to investigate the impact of a class of generic non-ideal response functions. The conclusion is supported by simulations validating our theoretical insights.

Method: Introduces YRC-0 problem setting (unsupervised expert collaboration learning), creates YRC-Bench benchmark with Gym-like API, simulated experts, evaluation pipeline, and baseline implementations. Proposes validation strategy and uses proposer-validator decomposition as diagnostic framework to evaluate learning methods.

Result: YRC-Bench provides standardized testbed for YRC-0 research across diverse environments. The validation strategy and proposer-validator framework offer insights into learning methods’ performance, establishing foundation for future research in expert-leveraging agents.

Conclusion: The paper addresses critical AI safety challenge of when agents should consult experts, introduces YRC-0 problem and YRC-Bench benchmark, and provides diagnostic tools to advance research in robust, low-cost expert-leveraging agent training.

Abstract: When deployed in the real world, AI agents will inevitably face challenges that exceed their individual capabilities. A critical component of AI safety is an agent’s ability to recognize when it is likely to fail in a novel situation and to yield control to a more capable expert system. Leveraging such expert assistance can significantly improve safety and performance in such situations. Since expert assistance is costly, a central challenge is determining when to consult an expert. In this paper, we explore a novel variant of this problem, termed YRC-0, in which an agent must learn to collaborate with an expert in new environments in an unsupervised manner–that is, without interacting with the expert during training. This setting motivates the development of low-cost, robust approaches for training expert-leveraging agents. To support research in this area, we introduce YRC-Bench, an open-source benchmark that instantiates YRC-0 across diverse environments. YRC-Bench provides a standardized Gym-like API, simulated experts, an evaluation pipeline, and implementations of popular baselines. Toward tackling YRC-0, we propose a validation strategy and use a proposer-validator decomposition as a diagnostic framework to evaluate a range of learning methods, offering insights that can inform future research. Codebase: https://github.com/modanesh/YRC-Bench

Method: Uses H² matrices for near-linear complexity, on-the-fly kernel evaluation to avoid storage, AFN preconditioner for faster convergence, C++ implementation for performance, and analytical gradients for hyperparameter optimization.

Result: HiGP provides a scalable computational backbone for large-scale GP regression/classification with efficient storage and computation, complementing existing GP frameworks.

Conclusion: HiGP successfully addresses GP scalability challenges through advanced numerical techniques, enabling practical application of GPs to large datasets while maintaining performance and integration with modern ML workflows.

Abstract: Gaussian Processes (GPs) are flexible, nonparametric Bayesian models widely used for regression and classification because of their ability to capture complex data patterns and quantify predictive uncertainty. However, the O(n^3) computational cost of kernel matrix operations poses a major obstacle to applying GPs at scale. HiGP is a high-performance Python package designed to overcome these scalability limitations through advanced numerical linear algebra and hierarchical kernel representations. It integrates H^2 matrices to achieve near-linear complexity in both storage and computation for spatial datasets, supports on-the-fly kernel evaluation to avoid explicit storage in large-scale problems, and incorporates a robust Adaptive Factorized Nyström (AFN) preconditioner that accelerates convergence of iterative solvers across a broad range of kernel spectra. These computational kernels are implemented in C++ for maximum performance and exposed through Python interfaces, enabling seamless integration with modern machine learning workflows. HiGP also includes analytically derived gradient computations for efficient hyperparameter optimization, avoiding the inefficiencies of automatic differentiation in iterative solvers. By serving as a reusable numerical engine, HiGP complements existing GP frameworks such as GPJax, KeOps, and GaussianProcesses.jl, providing a reliable and scalable computational backbone for large-scale Gaussian Process regression and classification.

Method: Reformulate node classification as subgraph classification. Use differentiated zero-padding strategy and Ego-Alter subgraph representation to resolve label conflicts. Implement Adaptive Feature Scaling Mechanism to dynamically adjust feature contributions based on dataset dependencies.

Result: Experimental results on six benchmark datasets show SubGND achieves performance comparable to or surpassing global message-passing GNNs, especially in heterophilic settings, demonstrating effectiveness and scalability.

Conclusion: SubGND provides a promising solution for node classification that balances scalability and accuracy, particularly effective in heterophilic graph settings where traditional GNNs struggle.

Abstract: Previous studies have demonstrated the strong performance of Graph Neural Networks (GNNs) in node classification. However, most existing GNNs adopt a node-centric perspective and rely on global message passing, leading to high computational and memory costs that hinder scalability. To mitigate these challenges, subgraph-based methods have been introduced, leveraging local subgraphs as approximations of full computational trees. While this approach improves efficiency, it often suffers from performance degradation due to the loss of global contextual information, limiting its effectiveness compared to global GNNs. To address this trade-off between scalability and classification accuracy, we reformulate the node classification task as a subgraph classification problem and propose SubGND (Subgraph GNN for NoDe). This framework introduces a differentiated zero-padding strategy and an Ego-Alter subgraph representation method to resolve label conflicts while incorporating an Adaptive Feature Scaling Mechanism to dynamically adjust feature contributions based on dataset-specific dependencies. Experimental results on six benchmark datasets demonstrate that SubGND achieves performance comparable to or surpassing global message-passing GNNs, particularly in heterophilic settings, highlighting its effectiveness and scalability as a promising solution for node classification.

Method: Single-agent reinforcement learning model with defined state space (queue lengths and signal phases), action space (intersection selection and phase adjustment), and two reward functions (congestion alleviation and travel time minimization). Tested using SUMO simulation.

Result: Model effectively reduces queue lengths and significantly decreases average travel time when using combined reward function, outperforming baseline with no signal adjustments.

Conclusion: Provides a new approach for large-scale regional traffic signal control with valuable insights for future urban traffic management.

Abstract: In the context of global urbanization and motorization, traffic congestion has become a significant issue, severely affecting the quality of life, environment, and economy. This paper puts forward a single-agent reinforcement learning (RL)-based regional traffic signal control (TSC) model. Different from multi - agent systems, this model can coordinate traffic signals across a large area, with the goals of alleviating regional traffic congestion and minimizing the total travel time. The TSC environment is precisely defined through specific state space, action space, and reward functions. The state space consists of the current congestion state, which is represented by the queue lengths of each link, and the current signal phase scheme of intersections. The action space is designed to select an intersection first and then adjust its phase split. Two reward functions are meticulously crafted. One focuses on alleviating congestion and the other aims to minimize the total travel time while considering the congestion level. The experiments are carried out with the SUMO traffic simulation software. The performance of the TSC model is evaluated by comparing it with a base case where no signal-timing adjustments are made. The results show that the model can effectively control congestion. For example, the queuing length is significantly reduced in the scenarios tested. Moreover, when the reward is set to both alleviate congestion and minimize the total travel time, the average travel time is remarkably decreased, which indicates that the model can effectively improve traffic conditions. This research provides a new approach for large-scale regional traffic signal control and offers valuable insights for future urban traffic management.

Method: Analyzing the stability of the KA representation theorem by examining its resilience to re-parameterizations of the hidden space, treating such transformations as adversarial attempts to disrupt the outer function construction, with focus on countable collections of continuous re-parameterizations.

Result: KA representation is stable under countable collections of continuous re-parameterizations, but reveals an equi-continuity question about the outer functions that prevents taking limits and defeating continuous groups of re-parameterizations.

Conclusion: The regularity question about outer functions’ equi-continuity is significant for the debate over KA theorem’s applicability to neural network theory, highlighting limitations in handling continuous transformation groups despite stability under countable continuous re-parameterizations.

Abstract: Regarding the representation theorem of Kolmogorov and Arnold (KA) as an algorithm for representing or «expressing» functions, we test its robustness by analyzing its stability to withstand re-parameterizations of the hidden space. One may think of such re-parameterizations as the work of an adversary attempting to foil the construction of the KA outer function. We find KA to be stable under countable collections of continuous re-parameterizations, but unearth a question about the equi-continuity of the outer functions that, so far, obstructs taking limits and defeating continuous groups of re-parameterizations. This question on the regularity of the outer functions is relevant to the debate over the applicability of KA to the general theory of NNs.

Method: The authors conduct systematic analysis using HarmBench, comparing vulnerability to manipulated feedback vs. query poisoning alone. They introduce a simple fake reward attack requiring no access to the reward model, and propose a lightweight highlighting defense.

Result: Feedback-based attacks raise attack success rate by up to ΔASR = 0.48. The fake reward attack significantly increases vulnerability, while the highlighting defense reduces fake reward ΔASR from 0.23 to 0.07 without degrading utility.

Conclusion: Prompt optimization pipelines represent a first-class attack surface requiring stronger safeguards for feedback channels and optimization frameworks, as demonstrated by the significant vulnerability to feedback poisoning attacks.

Abstract: Large language model (LLM) systems increasingly power everyday AI applications such as chatbots, computer-use assistants, and autonomous robots, where performance often depends on manually well-crafted prompts. LLM-based prompt optimizers reduce that effort by iteratively refining prompts from scored feedback, yet the security of this optimization stage remains underexamined. We present the first systematic analysis of poisoning risks in LLM-based prompt optimization. Using HarmBench, we find systems are substantially more vulnerable to manipulated feedback than to query poisoning alone: feedback-based attacks raise attack success rate (ASR) by up to ΔASR = 0.48. We introduce a simple fake reward attack that requires no access to the reward model and significantly increases vulnerability. We also propose a lightweight highlighting defense that reduces the fake reward ΔASR from 0.23 to 0.07 without degrading utility. These results establish prompt optimization pipelines as a first-class attack surface and motivate stronger safeguards for feedback channels and optimization frameworks.

Method: Proposes Quantization Error Propagation (QEP) - a lightweight, scalable framework that explicitly propagates quantization errors and compensates for accumulated errors across layers. Includes tunable propagation mechanism to prevent overfitting and control computational overhead.

Result: Extensive experiments on several LLMs show QEP-enhanced layer-wise PTQ achieves substantially higher accuracy than existing methods, with most pronounced gains in extremely low-bit quantization regimes.

Conclusion: QEP addresses the fundamental limitation of error accumulation in layer-wise PTQ, providing a general framework that improves quantization performance while maintaining simplicity and scalability.

Abstract: Layer-wise PTQ is a promising technique for compressing large language models (LLMs), due to its simplicity and effectiveness without requiring retraining. However, recent progress in this area is saturating, underscoring the need to revisit its core limitations and explore further improvements. We address this challenge by identifying a key limitation of existing layer-wise PTQ methods: the growth of quantization errors across layers significantly degrades performance, particularly in low-bit regimes. To address this fundamental issue, we propose Quantization Error Propagation (QEP), a general, lightweight, and scalable framework that enhances layer-wise PTQ by explicitly propagating quantization errors and compensating for accumulated errors. QEP also offers a tunable propagation mechanism that prevents overfitting and controls computational overhead, enabling the framework to adapt to various architectures and resource budgets. Extensive experiments on several LLMs demonstrate that QEP-enhanced layer-wise PTQ achieves substantially higher accuracy than existing methods. Notably, the gains are most pronounced in the extremely low-bit quantization regime.

Method: A three-stage semantic retrieval framework: (1) retrieves semantically similar samples using embeddings and indexing, (2) reranks them at varying subsequence lengths, and (3) calibrates scores using human novelty references for interpretability.

Result: Applied to SmolLM family: (1) models use pretraining data across longer sequences than previously reported, (2) task domains systematically affect novelty (some promote, some suppress), (3) instruction tuning increases novelty beyond just style changes.

Conclusion: Semantic novelty analysis is valuable for studying generalization. The framework enables efficient large-scale novelty assessment. Released ~20TB of corpus chunks and index artifacts for reproducibility and further research.

Abstract: Generation novelty is a key indicator of an LLM’s ability to generalize, yet measuring it against full pretraining corpora is computationally challenging. Existing evaluations often rely on lexical overlap, failing to detect paraphrased text, or do not consider the full pretraining corpus. We frame novelty as a semantic retrieval problem. This framing enables us to address novelty with modern embedding and indexing pipelines, allowing for efficient analysis at pre-training scale. Specifically, we propose a three-stage framework that retrieves semantically similar samples, reranks them at varying subsequence lengths, and calibrates scores using a human novelty reference for interpretability. We apply this framework to the SmolLM model family and report three key findings: (1) models draw on pre-training data across much longer sequences than previously reported; (2) some task domains systematically promote or suppress generation novelty; and (3) instruction tuning not only alters style but also increases novelty. These results highlight the value of semantic novelty analysis for studying generalization. To support reproducibility and further research, we release ~20 TB of corpus chunks and index artifacts at https://huggingface.co/datasets/stai-tuebingen/faiss-smollm

Method: SoftPick is a rectified, not sum-to-one drop-in replacement for softmax that creates sparse attention maps and reduces hidden state kurtosis.

Result: Experiments with 340M and 1.8B parameter models show 0% sink rate, lower kurtosis, sparse attention maps, and better quantization performance than softmax, especially at lower bit precisions.

Conclusion: SoftPick eliminates attention sink issues and enables better quantization, low-precision training, sparsity optimization, pruning, and interpretability in transformers.

Abstract: We introduce softpick, a rectified, not sum-to-one, drop-in replacement for softmax in transformer attention mechanisms that eliminates attention sink and massive activations. Our experiments with 340M and 1.8B parameter models demonstrate that softpick achieves 0% sink rate consistently. The softpick transformers produce hidden states with significantly lower kurtosis and creates sparse attention maps. Quantized models using softpick outperform softmax on standard benchmarks, with a particularly pronounced advantage at lower bit precisions. Our analysis and discussion shows how softpick has the potential to open new possibilities for quantization, low-precision training, sparsity optimization, pruning, and interpretability. Code: https://github.com/zaydzuhri/softpick-attention.

Method: “Pre-train, then post-train” curriculum: initial unsupervised training followed by supervised post-training that binds specific concepts to dedicated latent slots while preserving general reconstruction capacity.

Result: AlignSAE enables precise causal interventions like reliable “concept swaps” by targeting single semantically aligned slots, supports multi-hop reasoning, and provides mechanistic probes of grokking-like generalization dynamics.

Conclusion: AlignSAE creates an interpretable interface for inspecting and controlling specific concepts in LLMs without interference from unrelated features, advancing interpretability and control over model representations.

Abstract: Large Language Models (LLMs) encode factual knowledge within hidden parametric spaces that are difficult to inspect or control. While Sparse Autoencoders (SAEs) can decompose hidden activations into more fine-grained, interpretable features, they often struggle to reliably align these features with human-defined concepts, resulting in entangled and distributed feature representations. To address this, we introduce AlignSAE, a method that aligns SAE features with a predefined ontology through a “pre-train, then post-train” curriculum. After an initial unsupervised training phase, we apply supervised post-training to bind specific concepts to dedicated latent slots while preserving the remaining capacity for general reconstruction. This separation creates an interpretable interface where specific concepts can be inspected and controlled without interference from unrelated features. Empirical results demonstrate that AlignSAE enables precise causal interventions, such as reliable “concept swaps”, by targeting single, semantically aligned slots, and further supports multi-hop reasoning and a mechanistic probe of grokking-like generalization dynamics.

Method: Relates Hebbian STDP rule to noisy gradient descent on a non-convex loss function defined on the probability simplex, then provides rigorous mathematical proof of convergence properties.

Result: Proves that despite constant noise injection and non-convex optimization landscape, the Hebbian learning dynamic exponentially fast identifies the presynaptic neuron with highest activity, which is non-standard compared to typical noisy gradient descent that only fluctuates around minimizers.

Conclusion: Hebbian learning can achieve reliable and fast convergence to identify dominant input patterns even in noisy, non-convex settings, providing theoretical foundation for biological learning mechanisms.

Abstract: Hebbian learning is a key principle underlying learning in biological neural networks. We relate a Hebbian spike-timing-dependent plasticity rule to noisy gradient descent with respect to a non-convex loss function on the probability simplex. Despite the constant injection of noise and the non-convexity of the underlying optimization problem, one can rigorously prove that the considered Hebbian learning dynamic identifies the presynaptic neuron with the highest activity and that the convergence is exponentially fast in the number of iterations. This is non-standard and surprising as typically noisy gradient descent with fixed noise level only converges to a stationary regime where the noise causes the dynamic to fluctuate around a minimiser.

Method: Construct a simplified Transformer that interprets prompts as externally injected programs, decomposing attention as selective routing from prompt memory, FFN as local arithmetic conditioned on retrieved fragments, and depth-wise stacking as composition of local updates.

Result: Proves a constructive existential result showing that a single fixed Transformer backbone can approximate a broad class of target behaviors via prompts alone, providing formal foundations for prompt-based switching.

Conclusion: The framework enables formal analysis of prompt length/precision trade-offs and structural limits of prompt-based switching, establishing prompt engineering as a theoretically grounded capability distinct from empirical LLM claims.

Abstract: Prompts can switch a model’s behavior even when the weights are fixed, yet this phenomenon is rarely treated as a clean theoretical object rather than a heuristic. We study the family of functions obtainable by holding a Transformer backbone fixed as an executor and varying only the prompt. Our core idea is to view the prompt as an externally injected program and to construct a simplified Transformer that interprets it to implement different computations. The construction exposes a mechanism-level decomposition: attention performs selective routing from prompt memory, the FFN performs local arithmetic conditioned on retrieved fragments, and depth-wise stacking composes these local updates into a multi-step computation. Under this viewpoint, we prove a constructive existential result showing that a single fixed backbone can approximate a broad class of target behaviors via prompts alone. The framework provides a unified starting point for formalizing trade-offs under prompt length/precision constraints and for studying structural limits of prompt-based switching, while remaining distinct from empirical claims about pretrained LLMs.

Method: Modular cloud-based framework using Adversarial Autoencoder (AAE) trained on high-resolution pollution-health data from INHALE study, fine-tuned with smartwatch data via transfer learning to capture individual-specific patterns.

Result: Simulated pollution spikes (+100%) showed modest but measurable vital sign increases (+2.5% heart rate, +3.5% breathing rate). Analysis of U-BIOPRED data confirmed clinical relevance - individuals with such subclinical elevations had higher asthma burden scores or elevated FeNO.

Conclusion: The approach demonstrates feasibility of anticipatory, personalized health modeling for environmental challenges, offering scalable and secure infrastructure for AI-driven environmental health monitoring.

Abstract: Air pollution is a growing global health threat, exacerbated by climate change and linked to cardiovascular and respiratory diseases. While personal sensing devices enable real-time physiological monitoring, their integration with environmental data for individualised health prediction remains underdeveloped. Here, we present a modular, cloud-based framework that predicts personalised physiological responses to pollution by combining wearable-derived data with real-time environmental exposures. At its core is an Adversarial Autoencoder (AAE), initially trained on high-resolution pollution-health data from the INHALE study and fine-tuned using smartwatch data via transfer learning to capture individual-specific patterns. Consistent with changes in pollution levels commonly observed in the real-world, simulated pollution spikes (+100%) revealed modest but measurable increases in vital signs (e.g., +2.5% heart rate, +3.5% breathing rate). To assess clinical relevance, we analysed U-BIOPRED data and found that individuals with such subclinical vital sign elevations had higher asthma burden scores or elevated Fractional Exhaled Nitric Oxide (FeNO), supporting the physiological validity of these AI-predicted responses. This integrative approach demonstrates the feasibility of anticipatory, personalised health modelling in response to environmental challenges, offering a scalable and secure infrastructure for AI-driven environmental health monitoring.

Method: The paper presents a conceptual reframing rather than a specific technical method. It analyzes how token reduction can serve multiple purposes across vision, language, and multimodal systems, including facilitating multimodal integration, mitigating hallucinations, maintaining coherence, and enhancing training stability.

Result: The paper establishes a new conceptual framework positioning token reduction as a core principle in generative modeling, identifying specific benefits across multiple domains and outlining promising future research directions.

Conclusion: Token reduction should be viewed as a fundamental design principle in generative modeling rather than merely an efficiency measure, with significant implications for model architecture, multimodal alignment, and broader applications across vision, language, and multimodal systems.

Abstract: In Transformer architectures, tokens\textemdash discrete units derived from raw data\textemdash are formed by segmenting inputs into fixed-length chunks. Each token is then mapped to an embedding, enabling parallel attention computations while preserving the input’s essential information. Due to the quadratic computational complexity of transformer self-attention mechanisms, token reduction has primarily been used as an efficiency strategy. This is especially true in single vision and language domains, where it helps balance computational costs, memory usage, and inference latency. Despite these advances, this paper argues that token reduction should transcend its traditional efficiency-oriented role in the era of large generative models. Instead, we position it as a fundamental principle in generative modeling, critically influencing both model architecture and broader applications. Specifically, we contend that across vision, language, and multimodal systems, token reduction can: (i) facilitate deeper multimodal integration and alignment, (ii) mitigate “overthinking” and hallucinations, (iii) maintain coherence over long inputs, and (iv) enhance training stability, etc. We reframe token reduction as more than an efficiency measure. By doing so, we outline promising future directions, including algorithm design, reinforcement learning-guided token reduction, token optimization for in-context learning, agentic framework design, and broader ML and scientific domains.

Method: Proposes advancing through three interrelated themes: 1) Sensing - extending AI’s ability to capture richer non-digital sensory data, 2) Science - developing principled frameworks for multimodal heterogeneity, unified architectures, and cross-modal transfer, and 3) Synergy - technical challenges for learning synergy between modalities and humans.

Result: Presents a comprehensive research roadmap with accompanying projects, resources, and demos from MIT Media Lab’s Multisensory Intelligence group, establishing foundations for next-generation multisensory AI systems.

Conclusion: Multisensory AI represents a transformative direction that can fundamentally change how humans and AI experience and interact with each other by connecting AI to the full spectrum of human senses and environmental signals over the next decade.

Abstract: Our experience of the world is multisensory, spanning a synthesis of language, sight, sound, touch, taste, and smell. Yet, artificial intelligence has primarily advanced in digital modalities like text, vision, and audio. This paper outlines a research vision for multisensory artificial intelligence over the next decade. This new set of technologies can change how humans and AI experience and interact with one another, by connecting AI to the human senses and a rich spectrum of signals from physiological and tactile cues on the body, to physical and social signals in homes, cities, and the environment. We outline how this field must advance through three interrelated themes of sensing, science, and synergy. Firstly, research in sensing should extend how AI captures the world in richer ways beyond the digital medium. Secondly, developing a principled science for quantifying multimodal heterogeneity and interactions, developing unified modeling architectures and representations, and understanding cross-modal transfer. Finally, we present new technical challenges to learn synergy between modalities and between humans and AI, covering multisensory integration, alignment, reasoning, generation, generalization, and experience. Accompanying this vision paper are a series of projects, resources, and demos of latest advances from the Multisensory Intelligence group at the MIT Media Lab, see https://mit-mi.github.io/.

Method: GAOT combines multiscale attentional graph neural operator encoders/decoders with geometry embeddings and vision transformer processors. It uses novel implementations for computational efficiency and scalability while accurately mapping domain and input information to PDE solutions.

Result: GAOT demonstrates significant gains in both accuracy and efficiency over baselines on diverse PDE learning tasks. It achieves state-of-the-art performance on three large-scale three-dimensional industrial CFD datasets.

Conclusion: GAOT successfully addresses the accuracy-efficiency trade-off in PDE operator learning by combining geometry awareness with transformer architectures, making it suitable for practical engineering and industrial simulations.

Abstract: The very challenging task of learning solution operators of PDEs on arbitrary domains accurately and efficiently is of vital importance to engineering and industrial simulations. Despite the existence of many operator learning algorithms to approximate such PDEs, we find that accurate models are not necessarily computationally efficient and vice versa. We address this issue by proposing a geometry aware operator transformer (GAOT) for learning PDEs on arbitrary domains. GAOT combines novel multiscale attentional graph neural operator encoders and decoders, together with geometry embeddings and (vision) transformer processors to accurately map information about the domain and the inputs into a robust approximation of the PDE solution. Multiple innovations in the implementation of GAOT also ensure computational efficiency and scalability. We demonstrate this significant gain in both accuracy and efficiency of GAOT over several baselines on a large number of learning tasks from a diverse set of PDEs, including achieving state of the art performance on three large scale three-dimensional industrial CFD datasets.

Method: DHGNN incorporates homophily-aware and direction-sensitive components: a resettable gating mechanism to adaptively modulate message contributions based on homophily levels and informativeness, and a structure-aware noise-tolerant fusion module to integrate node representations from original and reverse directions.

Result: DHGNN outperforms state-of-the-art methods in node classification and link prediction on both homophilic and heterophilic directed graph datasets, improving over the best baseline by up to 15.07% in link prediction.

Conclusion: The framework effectively handles heterophilic neighborhoods and directional graph structures, with analysis showing the gating mechanism captures directional homophily gaps and fluctuating homophily across layers, providing insights into message-passing behavior on complex graphs.

Abstract: Graph Neural Networks (GNNs) have achieved significant success in various learning tasks on graph-structured data. Nevertheless, most GNNs struggle to generalize to heterophilic neighborhoods. Additionally, many GNNs ignore the directional nature of real-world graphs, resulting in suboptimal performance on directed graphs with asymmetric structures. In this work, we propose Directed Homophily-aware Graph Neural Network (DHGNN), a novel framework that addresses these limitations by incorporating homophily-aware and direction-sensitive components. DHGNN employs a resettable gating mechanism to adaptively modulate message contributions based on homophily levels and informativeness, and a structure-aware noise-tolerant fusion module to effectively integrate node representations from the original and reverse directions. Extensive experiments on both homophilic and heterophilic directed graph datasets demonstrate that DHGNN outperforms state-of-the-art methods in node classification and link prediction. In particular, DHGNN improves over the best baseline by up to 15.07% in link prediction. Our analysis further shows that the gating mechanism captures directional homophily gaps and fluctuating homophily across layers, providing deeper insights into message-passing behavior on complex graph structures.

Method: Quantile-based approach to develop algorithms for learning long-run per-step reward distribution and differential return distribution in average-reward MDPs.

Result: Proven-convergent tabular algorithms for prediction and control, plus scalable algorithm family; competitive/superior performance vs non-distributional methods.

Conclusion: Successfully extends distributional RL to average-reward setting, capturing rich distributional information while maintaining strong performance.

Abstract: To date, distributional reinforcement learning (distributional RL) methods have exclusively focused on the discounted setting, where an agent aims to optimize a discounted sum of rewards over time. In this work, we extend distributional RL to the average-reward setting, where an agent aims to optimize the reward received per time step. In particular, we utilize a quantile-based approach to develop the first set of algorithms that can successfully learn and/or optimize the long-run per-step reward distribution, as well as the differential return distribution of an average-reward MDP. We derive proven-convergent tabular algorithms for both prediction and control, as well as a broader family of algorithms that have appealing scaling properties. Empirically, we find that these algorithms yield competitive and sometimes superior performance when compared to their non-distributional equivalents, while also capturing rich information about the long-run per-step reward and differential return distributions.

Method: Train activation probes on synthetic data to detect high-stakes interactions, evaluate probe architectures, and compare with prompted/finetuned LLM monitors.

Result: Probes show robust generalization to out-of-distribution real-world data, achieve performance comparable to LLM monitors, and offer six orders-of-magnitude computational savings by reusing monitored model’s activations.

Conclusion: Activation probes are effective, efficient monitors for high-stakes LLM interactions and can serve as initial filters in hierarchical monitoring systems, enabling resource-aware deployment.

Abstract: Monitoring is an important aspect of safely deploying Large Language Models (LLMs). This paper examines activation probes for detecting ``high-stakes’’ interactions – where the text indicates that the interaction might lead to significant harm – as a critical, yet underexplored, target for such monitoring. We evaluate several probe architectures trained on synthetic data, and find them to exhibit robust generalization to diverse, out-of-distribution, real-world data. Probes’ performance is comparable to that of prompted or finetuned medium-sized LLM monitors, while offering computational savings of six orders-of-magnitude. These savings are enabled by reusing activations of the model that is being monitored. Our experiments also highlight the potential of building resource-aware hierarchical monitoring systems, where probes serve as an efficient initial filter and flag cases for more expensive downstream analysis. We release our novel synthetic dataset and the codebase at https://github.com/arrrlex/models-under-pressure.

Method: Proposes a data-centric ecosystem view with 5D complexities (structure, domain, cardinality, causality, ethics) integrated with data life cycle phases. Uses data agents performing goal-driven tasks, with data scientists as abstract entities organizing these agents. Defines discipline-induced data science and pan-data science as integrative frameworks.

Result: Develops a formalized ecosystemic view of data science with a general-purpose, fusion-oriented architecture for integrating heterogeneous knowledge, agents, and workflows applicable across diverse disciplines and high-impact applications.

Conclusion: The essential data science framework provides a comprehensive architecture for addressing complex data challenges through an integrated ecosystem approach that bridges computational and foundational aspects, enabling cross-disciplinary applications.

Abstract: This manuscript provides a systemic and data-centric view of what we term essential data science, as a natural ecosystem with challenges and missions stemming from the fusion of data universe with its multiple combinations of the 5D complexities (data structure, domain, cardinality, causality, and ethics) with the phases of the data life cycle. Data agents perform tasks driven by specific goals. The data scientist is an abstract entity that comes from the logical organization of data agents with their actions. Data scientists face challenges that are defined according to the missions. We define specific discipline-induced data science, which in turn allows for the definition of pan-data science, a natural ecosystem that integrates specific disciplines with the essential data science. We semantically split the essential data science into computational, and foundational. By formalizing this ecosystemic view, we contribute a general-purpose, fusion-oriented architecture for integrating heterogeneous knowledge, agents, and workflows-relevant to a wide range of disciplines and high-impact applications.

Method: Proposes a novel approach that combines Monte Carlo sampling with any function family whose probabilistic values can be computed efficiently, enabling use of tree-based models like XGBoost while maintaining unbiased estimates.

Result: Experiments across eight datasets show state-of-the-art performance: 6.5× lower error than Permutation SHAP, 3.8× lower than Kernel SHAP, and 2.6× lower than Leverage SHAP for Shapley values. For general probabilistic values, achieves 215× lower error than prior best estimators.

Conclusion: The proposed flexible framework significantly improves accuracy of probabilistic value estimation by allowing integration of powerful function families beyond linear regression, advancing explainable AI capabilities.

Abstract: With origins in game theory, probabilistic values like Shapley values, Banzhaf values, and semi-values have emerged as a central tool in explainable AI. They are used for feature attribution, data attribution, data valuation, and more. Since all of these values require exponential time to compute exactly, research has focused on efficient approximation methods using two techniques: Monte Carlo sampling and linear regression formulations. In this work, we present a new way of combining both of these techniques. Our approach is more flexible than prior algorithms, allowing for linear regression to be replaced with any function family whose probabilistic values can be computed efficiently. This allows us to harness the accuracy of tree-based models like XGBoost, while still producing unbiased estimates. From experiments across eight datasets, we find that our methods give state-of-the-art performance for estimating probabilistic values. For Shapley values, the error of our methods can be $6.5\times$ lower than Permutation SHAP (the most popular Monte Carlo method), $3.8\times$ lower than Kernel SHAP (the most popular linear regression method), and $2.6\times$ lower than Leverage SHAP (the prior state-of-the-art Shapley value estimator). For more general probabilistic values, we can obtain error $215\times$ lower than the best estimator from prior work.

Method: Uses neural networks to enhance specific aspects of the statistical model in Bayesian MOT that are identified as overly simplistic. Employs belief propagation to avoid high-dimensional operations and sequential Monte Carlo methods for efficient low-dimensional operations, ensuring computational tractability.

Result: Evaluated on the nuScenes autonomous driving dataset, demonstrating state-of-the-art performance. The method successfully combines model-based flexibility with neural network learning capabilities.

Conclusion: The proposed hybrid framework effectively integrates model-based and data-driven approaches for MOT, achieving superior performance while maintaining computational efficiency and broad applicability across scenarios.

Abstract: Multiobject tracking (MOT) is an important task in applications including autonomous driving, ocean sciences, and aerospace surveillance. Traditional MOT methods are model-based and combine sequential Bayesian estimation with data association and an object birth model. More recent methods are fully data-driven and rely on the training of neural networks. Both approaches offer distinct advantages in specific settings. In particular, model-based methods are generally applicable across a wide range of scenarios, whereas data-driven MOT achieves superior performance in scenarios where abundant labeled data for training is available. A natural thought is whether a general framework can integrate the two approaches. This paper introduces a hybrid method that utilizes neural networks to enhance specific aspects of the statistical model in Bayesian MOT that have been identified as overly simplistic. By doing so, the performance of the prediction and update steps of Bayesian MOT is improved. To ensure tractable computation, our framework uses belief propagation to avoid high-dimensional operations combined with sequential Monte Carlo methods to perform low-dimensional operations efficiently. The resulting method combines the flexibility and robustness of model-based approaches with the capability to learn complex information from data of neural networks. We evaluate the performance of the proposed method based on the nuScenes autonomous driving dataset and demonstrate that it has state-of-the-art performance.

Method: Develop Tensor Decomposition Networks (TDNs) using low-rank tensor decompositions (CP decomposition) to replace CG tensor products, with path-weight sharing to tie multiplicity-space weights across all CG paths into a single shared parameter set.

Result: TDNs achieve competitive performance on PubChemQCR (105M DFT snapshots), OC20, and OC22 datasets with dramatic computational speedup (O(L⁶) to O(L⁴)), proving uniform bound on equivariance error and universality of approximating equivariant bilinear maps.

Conclusion: TDNs provide an efficient plug-and-play replacement for tensor products in existing networks, enabling faster SO(3)-equivariant models without sacrificing performance, making them practical for large-scale molecular simulations.

Abstract: $\rm{SO}(3)$-equivariant networks are the dominant models for machine learning interatomic potentials (MLIPs). The key operation of such networks is the Clebsch-Gordan (CG) tensor product, which is computationally expensive. To accelerate the computation, we develop tensor decomposition networks (TDNs) as a class of approximately equivariant networks in which CG tensor products are replaced by low-rank tensor decompositions, such as the CANDECOMP/PARAFAC (CP) decomposition. With the CP decomposition, we prove (i) a uniform bound on the induced error of $\rm{SO}(3)$-equivariance, and (ii) the universality of approximating any equivariant bilinear map. To further reduce the number of parameters, we propose path-weight sharing that ties all multiplicity-space weights across the $\mathcal{O}(L^3)$ CG paths into a single shared parameter set without compromising equivariance, where $L$ is the maximum angular degree. The resulting layer acts as a plug-and-play replacement for tensor products in existing networks, and the computational complexity of tensor products is reduced from $\mathcal{O}(L^6)$ to $\mathcal{O}(L^4)$. We evaluate TDNs on PubChemQCR, a newly curated molecular relaxation dataset containing 105 million DFT-calculated snapshots. We also use existing datasets, including OC20, and OC22. Results show that TDNs achieve competitive performance with dramatic speedup in computations. Our code is publicly available as part of the AIRS library (\href{https://github.com/divelab/AIRS/tree/main/OpenMol/TDN}{https://github.com/divelab/AIRS/}).

Method: Developed Bi-cephalic Self-Attention Model (BiSAM) using resting-state EEG signals combined with demographic/clinical variables (age, education, disease duration). Tested three modalities: signal-only, descriptive-only, and multi-modal across different EEG channel subsets.

Result: Multi-modal models significantly outperformed single-modality approaches: signal-only (55% accuracy), descriptive-only (68% accuracy), while multi-modal BiSAM 8 and BiSAM 4 achieved 88% accuracy for classifying PDFOG+ vs PDFOG- vs healthy controls.

Conclusion: Integration of EEG with objective descriptive features enables robust classification of Parkinson’s patients with freezing of gait using minimal EEG channels, offering scalable clinical monitoring and early diagnosis potential.

Abstract: Parkinson’s Disease (PD) often results in motor and cognitive impairments, including gait dysfunction, particularly in patients with freezing of gait (FOG). Current detection methods are either subjective or reliant on specialized gait analysis tools. This study aims to develop an objective, data-driven, multi-modal classification model for FOG-specific classification, distinguishing PD patients with FOG (PDFOG+) from those without FOG (PDFOG-) and healthy controls using resting-state EEG signals combined with demographic and clinical variables. For our main analysis, we utilized a dataset of 124 participants: 42 PDFOG+, 41 PDFOG-, and 41 age-matched healthy controls. Features extracted from resting-state EEG and descriptive variables (age, education, disease duration) were used to train a novel Bi-cephalic Self-Attention Model (BiSAM). We tested three modalities: signal-only, descriptive-only, and multi-modal, across different EEG channel subsets including BiSAM 63, BiSAM 16, BiSAM 8, and BiSAM 4 for primary analysis. For main analysis, signal-only (BiSAM 4) and descriptive-only models showed limited performance, achieving a maximum accuracy of 55% and 68%, respectively. In contrast, the multi-modal models significantly outperformed both, with BiSAM 8 and BiSAM 4 achieving the highest classification accuracy of 88%. These results demonstrate the value of integrating EEG with objective descriptive features for robust PDFOG+ classification. This study introduces a multi-modal, attention-based architecture that objectively classifies PDFOG+ using minimal EEG channels and descriptive variables. This approach offers a scalable and efficient alternative to traditional assessments, with potential applications in routine clinical monitoring and early diagnosis of PD-related gait dysfunction.

Method: SLogic uses a context-aware scoring function that analyzes the subgraph locally defined by the query’s head entity to assign query-dependent scores to logical rules, enabling differentiated weighting specific to local query contexts.

Result: Extensive experiments on benchmark datasets show SLogic outperforms existing rule-based methods and achieves competitive performance against state-of-the-art baselines.

Conclusion: SLogic provides both improved performance and interpretability by generating query-dependent, human-readable logical rules that serve as explicit explanations for its inferences.

Abstract: Logical rule-based methods offer an interpretable approach to knowledge graph completion (KGC) by capturing compositional relationships in the form of human-readable inference rules. While existing logical rule-based methods learn rule confidence scores, they typically assign a global weight to each rule schema, applied uniformly across the graph. This is a significant limitation, as a rule’s importance often varies depending on the specific query instance. To address this, we introduce SLogic (Subgraph-Informed Logical Rule learning), a novel framework that assigns query-dependent scores to logical rules. The core of SLogic is a context-aware scoring function. This function determines the importance of a rule by analyzing the subgraph locally defined by the query’s head entity, thereby enabling a differentiated weighting of rules specific to their local query contexts. Extensive experiments on benchmark datasets show that SLogic outperforms existing rule-based methods and achieves competitive performance against state-of-the-art baselines. It also generates query-dependent, human-readable logical rules that serve as explicit explanations for its inferences.

Method: WaveNet architecture with dilated causal convolutions and residual connections trained on 209,231 annotated iEEG samples using 70/20/10 split, focal loss for class imbalance, and dynamic dataset partitioning.

Result: Model achieved classification accuracy exceeding previous CNN/LSTM approaches, with high discrimination of noise/artifacts (precision 0.98-1.0) and clinically interpretable physiological/pathological classification (F1-scores 0.96/0.90).

Conclusion: WaveNet is well-suited for iEEG classification due to its ability to capture temporal dependencies, though generalizability across datasets and clinical settings needs further validation.

Abstract: This study introduces a WaveNet-based deep learning model designed to automate the classification of intracranial electroencephalography (iEEG) signals into physiological activity, pathological (epileptic) activity, power-line noise, and other non-cerebral artifacts. Traditional methods for iEEG signal classification, which rely on expert visual review, are becoming increasingly impractical due to the growing complexity and volume of iEEG recordings. Leveraging a publicly available annotated dataset from Mayo Clinic and St. Anne’s University Hospital, the WaveNet model was trained, validated, and tested on 209,231 samples using a 70/20/10 split. The model achieved a classification accuracy exceeding previous non-specialized CNN- and LSTM-based approaches and was benchmarked against a Temporal Convolutional Network (TCN) baseline. Notably, the model achieves high discrimination of noise and artifact classes, with precisions of 0.98 and approximately 1, respectively. Classification between physiological and pathological signals exhibits a modest but clinically interpretable overlap, with F1-scores of 0.96 and 0.90 and 175 and 272 cross-class false positives, respectively, reflecting inherent clinical overlap. WaveNet’s architecture, originally developed for raw audio synthesis, is well-suited for iEEG data due to its use of dilated causal convolutions and residual connections, enabling the capture of both fine-grained and long-range temporal dependencies. The study also details the preprocessing pipeline, including dynamic dataset partitioning, the use of focal loss to address class imbalance, and normalization steps that support high model performance. While the results demonstrate strong in-distribution performance, generalizability across datasets and clinical settings has yet to be established.

Method: Extend continuous-time Hopfield neural network to quaternionic domain, establish existence/uniqueness of fixed points, and use periodic projection strategy during gradient descent to maintain quaternionic structure in weight matrices.

Result: Achieves high accuracy, fast convergence, strong reliability on random target sets, and produces smooth evolution trajectories with well-bounded curvature suitable for robotics applications.

Conclusion: QSHNN provides practical framework for neural networks under hypercomplex/non-commutative algebraic structures, with applications in robotics where quaternion parameterization is natural.

Abstract: Motivated by the geometric advantages of quaternions in representing rotations and postures, we propose a quaternion-valued supervised learning Hopfield-structured neural network (QSHNN) with a fully connected structure inspired by the classic Hopfield neural network (HNN). Starting from a continuous-time dynamical model of HNNs, we extend the formulation to the quaternionic domain and establish the existence and uniqueness of fixed points with asymptotic stability. For the learning rules, we introduce a periodic projection strategy that modifies standard gradient descent by periodically projecting each 4*4 block of the weight matrix onto the closest quaternionic structure in the least-squares sense. This approach preserves both convergence and quaternionic consistency throughout training. Benefiting from this rigorous mathematical foundation, the experimental model implementation achieves high accuracy, fast convergence, and strong reliability across randomly generated target sets. Moreover, the evolution trajectories of the QSHNN exhibit well-bounded curvature, i.e., sufficient smoothness, which is crucial for applications such as control systems or path planning modules in robotic arms, where joint postures are parameterized by quaternion neurons. Beyond these application scenarios, the proposed model offers a practical implementation framework and a general mathematical methodology for designing neural networks under hypercomplex or non-commutative algebraic structures.

Method: The authors formalize a tailored definition of differential privacy for policy optimization, addressing challenges from on-policy learning dynamics. They then systematically analyze sample complexity of popular PO algorithms (policy gradient, natural policy gradient, etc.) under DP constraints using a unified theoretical framework.

Result: Theoretical results show that privacy costs often manifest as lower-order terms in sample complexity. The analysis also reveals subtle but important observations specific to private PO settings that offer practical insights for privacy-preserving algorithms.

Conclusion: The paper establishes foundational theoretical understanding of differentially private policy optimization, demonstrating that privacy can be achieved with manageable sample complexity costs, providing valuable guidance for implementing privacy-preserving RL algorithms in sensitive applications.

Abstract: Policy optimization (PO) is a cornerstone of modern reinforcement learning (RL), with diverse applications spanning robotics, healthcare, and large language model training. The increasing deployment of PO in sensitive domains, however, raises significant privacy concerns. In this paper, we initiate a theoretical study of differentially private policy optimization, focusing explicitly on its sample complexity. We first formalize an appropriate definition of differential privacy (DP) tailored to PO, addressing the inherent challenges arising from on-policy learning dynamics and the subtlety involved in defining the unit of privacy. We then systematically analyze the sample complexity of widely-used PO algorithms, including policy gradient (PG), natural policy gradient (NPG) and more, under DP constraints and various settings, via a unified framework. Our theoretical results demonstrate that privacy costs can often manifest as lower-order terms in the sample complexity, while also highlighting subtle yet important observations in private PO settings. These offer valuable practical insights for privacy-preserving PO algorithms.

Method: Novel single-agent RL framework using DreamerV3 world model with centralized decision-making. Uses adjacency matrix to encode road network topology, real-time queue states from probe vehicle data, and signal timing parameters. Agent sequentially selects intersections and adjusts signal phase splits to regulate traffic inflow/outflow like a feedback control system.

Result: Simulation experiments in SUMO show the framework exhibits robust anti-fluctuation capability against multi-level (10%, 20%, 30%) OD demand fluctuations and significantly reduces queue lengths.

Conclusion: Establishes a new paradigm for intelligent traffic control compatible with probe vehicle technology. Future work will focus on incorporating stochastic OD demand fluctuations during training and exploring regional optimization for contingency events.

Abstract: Traffic congestion, primarily driven by intersection queuing, significantly impacts urban living standards, safety, environmental quality, and economic efficiency. While Traffic Signal Control (TSC) systems hold potential for congestion mitigation, traditional optimization models often fail to capture real-world traffic complexity and dynamics. This study introduces a novel single-agent reinforcement learning (RL) framework for regional adaptive TSC, circumventing the coordination complexities inherent in multi-agent systems through a centralized decision-making paradigm. The model employs an adjacency matrix to unify the encoding of road network topology, real-time queue states derived from probe vehicle data, and current signal timing parameters. Leveraging the efficient learning capabilities of the DreamerV3 world model, the agent learns control policies where actions sequentially select intersections and adjust their signal phase splits to regulate traffic inflow/outflow, analogous to a feedback control system. Reward design prioritizes queue dissipation, directly linking congestion metrics (queue length) to control actions. Simulation experiments conducted in SUMO demonstrate the model’s effectiveness: under inference scenarios with multi-level (10%, 20%, 30%) Origin-Destination (OD) demand fluctuations, the framework exhibits robust anti-fluctuation capability and significantly reduces queue lengths. This work establishes a new paradigm for intelligent traffic control compatible with probe vehicle technology. Future research will focus on enhancing practical applicability by incorporating stochastic OD demand fluctuations during training and exploring regional optimization mechanisms for contingency events.

Method: Proposes a partition-based multi-stage fine-tuning framework that strategically partitions domains into subsets (stages) by balancing domain discrepancy, synergy, and model capacity constraints.

Result: Theoretical analysis provides novel generalization bounds justifying the partitioning strategy, and extensive empirical evaluations on various language understanding tasks show consistent outperformance over state-of-the-art baselines.

Conclusion: The proposed framework effectively addresses inter-domain interference in LLM adaptation, leveraging domain synergies while minimizing negative transfer through strategic partitioning.

Abstract: Large language models (LLMs) demonstrate impressive generalization abilities, yet adapting them effectively across multiple heterogeneous domains remains challenging due to inter-domain interference. To overcome this challenge, we propose a partition-based multi-stage fine-tuning framework designed to exploit inter-domain synergies while minimizing negative transfer. Our approach strategically partitions domains into subsets (stages) by balancing domain discrepancy, synergy, and model capacity constraints. We theoretically analyze the proposed framework and derive novel generalization bounds that justify our partitioning strategy. Extensive empirical evaluations on various language understanding tasks show that our method consistently outperforms state-of-the-art baselines.

Method: BACL consists of two modules: 1) Boundary-aware Negative Sampler that gradually increases difficulty by selecting more ambiguous negatives, and 2) Contrastive Local Attention loss that highlights where mismatches occur. Both modules are fully differentiable and work with any off-the-shelf dual encoder.

Result: Theoretical analysis predicts fast O(1/n) error rate. Empirically, BACL achieves up to +32% R@1 improvement over CLIP and sets new state-of-the-art results on four large-scale benchmarks, all without requiring extra labels.

Conclusion: BACL effectively leverages ambiguous negatives as a curriculum signal, enabling multimodal models to learn fine-grained distinctions and achieve significant performance improvements without additional supervision.

Abstract: Most multimodal models treat every negative pair alike, ignoring the ambiguous negatives that differ from the positive by only a small detail. We propose Boundary-Aware Curriculum with Local Attention (BACL), a lightweight add-on that turns these borderline cases into a curriculum signal. A Boundary-aware Negative Sampler gradually raises difficulty, while a Contrastive Local Attention loss highlights where the mismatch occurs. The two modules are fully differentiable and work with any off-the-shelf dual encoder. Theory predicts a fast O(1/n) error rate; practice shows up to +32% R@1 over CLIP and new SOTA on four large-scale benchmarks, all without extra labels.

Method: Proposes PAZO (Public-data-Assisted Zeroth-order Optimizers) framework that leverages public information to guide and improve gradient approximation in private zeroth-order algorithms with minimal overhead.

Result: PAZO achieves superior privacy/utility tradeoffs across vision and text tasks in both pre-training and fine-tuning settings. Outperforms best first-order baselines (with public data) especially in highly private regimes, while offering up to 16× runtime speedup.

Conclusion: Public-data-assisted zeroth-order optimization provides an effective approach to mitigate computation/memory overhead of DP methods while maintaining strong privacy/utility tradeoffs, particularly valuable in high-privacy regimes.

Abstract: One of the major bottlenecks for deploying popular first-order differentially private (DP) machine learning algorithms (e.g., DP-SGD) lies in their high computation and memory cost, despite the existence of optimized implementations. Zeroth-order methods have promise in mitigating the overhead, as they leverage function evaluations to approximate the gradients, hence significantly easier to privatize. While recent works have explored zeroth-order approaches in both private and non-private settings, they still suffer from relatively low utilities compared with DP-SGD, and have only been evaluated in limited application domains. In this work, we propose to leverage public information to guide and improve gradient approximation of private zeroth-order algorithms. We explore a suite of public-data-assisted zeroth-order optimizers (PAZO) with minimal overhead. We provide theoretical analyses of the PAZO framework under an assumption of the similarity between public and private data. Empirically, we demonstrate that PAZO achieves superior privacy/utility tradeoffs across vision and text tasks in both pre-training and fine-tuning settings, outperforming the best first-order baselines (with public data) especially in highly private regimes, while offering up to $16\times$ runtime speedup.

Method: Soft Q-based Diffusion Finetuning (SQDF) uses KL-regularized RL with reparameterized policy gradient of a training-free, differentiable soft Q-function estimation. Enhanced with: 1) discount factor for credit assignment in denoising, 2) consistency models to refine Q-function estimates, and 3) off-policy replay buffer for mode coverage and reward-diversity trade-off management.

Result: SQDF achieves superior target rewards while preserving diversity in text-to-image alignment tasks, and attains high sample efficiency while maintaining naturalness and diversity in online black-box optimization.

Conclusion: SQDF effectively mitigates reward over-optimization in diffusion model alignment, enabling better reward attainment while preserving sample quality and diversity across different alignment tasks.

Abstract: Diffusion models excel at generating high-likelihood samples but often require alignment with downstream objectives. Existing fine-tuning methods for diffusion models significantly suffer from reward over-optimization, resulting in high-reward but unnatural samples and degraded diversity. To mitigate over-optimization, we propose Soft Q-based Diffusion Finetuning (SQDF), a novel KL-regularized RL method for diffusion alignment that applies a reparameterized policy gradient of a training-free, differentiable estimation of the soft Q-function. SQDF is further enhanced with three innovations: a discount factor for proper credit assignment in the denoising process, the integration of consistency models to refine Q-function estimates, and the use of an off-policy replay buffer to improve mode coverage and manage the reward-diversity trade-off. Our experiments demonstrate that SQDF achieves superior target rewards while preserving diversity in text-to-image alignment. Furthermore, in online black-box optimization, SQDF attains high sample efficiency while maintaining naturalness and diversity.

Method: JPM treats single-disease trajectories as partial rankings and builds a prior over joint progressions. Four variants are studied: Pairwise, Bradley-Terry, Plackett-Luce, and Mallows models. The framework is evaluated on three properties: calibration, separation, and sharpness.

Result: All JPM variants are calibrated and achieve near-perfect separation. Sharpness varies by variant and is predictable from input partial ranking features. In synthetic experiments, JPM improves ordering accuracy by ~21% over SA-EBM baseline. In NACC data, Mallows variant and baseline show results consistent with prior literature on AD+VaD progression.

Conclusion: JPM effectively models joint disease progression in mixed pathologies, outperforming single-disease approaches. The framework provides a principled way to combine multiple disease trajectories, with Mallows variant showing particular promise for clinical applications in neurodegenerative diseases.

Abstract: Event-based models (EBMs) infer disease progression from cross-sectional data, and standard EBMs assume a single underlying disease per individual. In contrast, mixed pathologies are common in neurodegeneration. We introduce the Joint Progression Model (JPM), a probabilistic framework that treats single-disease trajectories as partial rankings and builds a prior over joint progressions. We study several JPM variants (Pairwise, Bradley-Terry, Plackett-Luce, and Mallows) and analyze three properties: (i) calibration – whether lower model energy predicts smaller distance to the ground truth ordering; (ii) separation – the degree to which sampled rankings are distinguishable from random permutations; and (iii) sharpness – the stability of sampled aggregate rankings. All variants are calibrated, and all achieve near-perfect separation; sharpness varies by variant and is well-predicted by simple features of the input partial rankings (number and length of rankings, conflict, and overlap). In synthetic experiments, JPM improves ordering accuracy by roughly 21 percent over a strong EBM baseline (SA-EBM) that treats the joint disease as a single condition. Finally, using NACC, we find that the Mallows variant of JPM and the baseline model (SA-EBM) have results that are more consistent with prior literature on the possible disease progression of the mixed pathology of AD and VaD.

Method: The authors develop a unified theoretical framework treating SDL as one optimization problem, analyze how diverse methods instantiate this framework, rigorously examine optimization landscapes, create the Linear Representation Bench benchmark following the Linear Representation Hypothesis, and develop “feature anchoring” technique to enhance feature recovery.

Result: The framework provides novel theoretical explanations for empirically observed phenomena (feature absorption, dead neurons), demonstrates how different SDL methods instantiate the unified framework, and shows that feature anchoring enhances feature recovery capabilities across SDL methods.

Conclusion: This work provides the first unified theoretical foundation for the broader family of SDL methods in mechanistic interpretability, offering rigorous analysis, explaining empirical observations, and introducing practical improvements through feature anchoring, advancing both theoretical understanding and practical deployment of interpretability methods.

Abstract: As AI models achieve remarkable capabilities across diverse domains, understanding what representations they learn and how they process information has become increasingly important for both scientific progress and trustworthy deployment. Recent works in mechanistic interpretability have shown that neural networks represent meaningful concepts as directions in their representation spaces and often encode diverse concepts in superposition. Various sparse dictionary learning (SDL) methods, including sparse autoencoders, transcoders, and crosscoders, address this by training auxiliary models with sparsity constraints to disentangle these superposed concepts into monosemantic features. These methods have demonstrated remarkable empirical success but have limited theoretical understanding. Existing theoretical work is limited to sparse autoencoders with tied-weight constraints, leaving the broader family of SDL methods without formal grounding. In this work, we develop the first unified theoretical framework considering SDL as one optimization problem. We demonstrate how diverse methods instantiate the theoretical framework and provide rigorous analysis of the optimization landscape. We provide novel theoretical explanations for empirically observed phenomena, including feature absorption and dead neurons. We design the Linear Representation Bench, a benchmark that strictly follows the Linear Representation Hypothesis, to evaluate SDL methods with fully accessible ground-truth features. Motivated by our theory and findings, we develop feature achoring, a novel technique applicable for all SDL methods, to enhance their feature recovery capabilities.

Method: Use self-play with two agents (user simulator and AI assistant) to generate conversations with ambiguous queries. Train models using Reinforced Self-Training (ReST) that take numerical costs for clarification questions and generated words as input, optimizing for cost-penalized accuracy reward.

Result: The approach produces steerable policies that change behavior predictably based on provided costs, achieving higher reward and accuracy. The method generalizes to unseen numerical cost values at test time.

Conclusion: Self-play with cost-aware reinforcement learning enables training of flexible, steerable policies for managing uncertainty in AI assistants that adapt to contextual constraints and generalize to new cost scenarios.

Abstract: To handle underspecified or ambiguous queries, AI assistants need a policy for managing their uncertainty to determine (a) when to guess the user intent and answer directly, (b) when to enumerate and answer multiple possible intents, and (c) when to ask a clarifying question. However, such policies are contextually dependent on factors such as user preferences or modality. For example, enumerating multiple possible user intentions is cumbersome on small screens or in a voice setting. In this work, we propose to train steerable policies for managing this uncertainty using self-play. Given two agents, one simulating a user and the other an AI assistant, we generate conversations where the user issues a potentially ambiguous query, and the assistant needs to determine how to respond. Importantly, the model takes as input the numerical cost of each clarification question, and each generated word, and is asked to take the action that will maximize its final reward, which is the cost-penalized accuracy. We use Reinforced Self-Training (ReST) to train our model to achieve high reward and show this leads to a steerable policy that changes its behavior predictably conditioned on the provided costs, leading to higher reward and accuracy. Moreover, our procedure also generalizes to numerical cost values that were unobserved at training time.

Method: Uses the Generalized Resolution-Shell Dynamics (GRSD) framework modeling learning as spectral energy transport across logarithmic resolution shells. Identifies sufficient conditions for renormalizable coarse-grained description including bounded gradient propagation, weak functional incoherence at initialization, controlled Jacobian evolution, and log-shift invariance of shell couplings.

Result: Power-law scaling does not follow from renormalizability alone, but arises as a rigidity consequence when log-shift invariance is combined with time-rescaling covariance of gradient flow, forcing the renormalized GRSD velocity field into power-law form.

Conclusion: Power-law scaling in deep learning emerges from specific structural constraints on the learning process within the GRSD framework, rather than being an automatic consequence of renormalizability, providing theoretical understanding of observed empirical patterns.

Abstract: Empirical power–law scaling has been widely observed across modern deep learning systems, yet its theoretical origins and scope of validity remain incompletely understood. The Generalized Resolution–Shell Dynamics (GRSD) framework models learning as spectral energy transport across logarithmic resolution shells, providing a coarse–grained dynamical description of training. Within GRSD, power–law scaling corresponds to a particularly simple renormalized shell dynamics; however, such behavior is not automatic and requires additional structural properties of the learning process. In this work, we identify a set of sufficient conditions under which the GRSD shell dynamics admits a renormalizable coarse–grained description. These conditions constrain the learning configuration at multiple levels, including boundedness of gradient propagation in the computation graph, weak functional incoherence at initialization, controlled Jacobian evolution along training, and log–shift invariance of renormalized shell couplings. We further show that power–law scaling does not follow from renormalizability alone, but instead arises as a rigidity consequence: once log–shift invariance is combined with the intrinsic time–rescaling covariance of gradient flow, the renormalized GRSD velocity field is forced into a power–law form.

Method: Three-stage pipeline: 1) Patch Low-Frequency forecasting Module (PLFM) extracts stable low-frequency trends from localized spectral patches, 2) residual learner models high-frequency variations, 3) fine-tuned LLM refines predictions by incorporating auxiliary context and domain knowledge through structured natural language prompts.

Result: Extensive experiments on financial and energy datasets show LoFT-LLM significantly outperforms strong baselines under both full-data and few-shot regimes, delivering superior accuracy, robustness, and interpretability.

Conclusion: LoFT-LLM effectively addresses forecasting challenges by separating low-frequency trends from high-frequency noise and leveraging LLMs for semantic calibration, providing a robust solution for real-world applications with limited data and complex dynamics.

Abstract: Time-series forecasting in real-world applications such as finance and energy often faces challenges due to limited training data and complex, noisy temporal dynamics. Existing deep forecasting models typically supervise predictions using full-length temporal windows, which include substantial high-frequency noise and obscure long-term trends. Moreover, auxiliary variables containing rich domain-specific information are often underutilized, especially in few-shot settings. To address these challenges, we propose LoFT-LLM, a frequency-aware forecasting pipeline that integrates low-frequency learning with semantic calibration via a large language model (LLM). Firstly, a Patch Low-Frequency forecasting Module (PLFM) extracts stable low-frequency trends from localized spectral patches. Secondly, a residual learner then models high-frequency variations. Finally, a fine-tuned LLM refines the predictions by incorporating auxiliary context and domain knowledge through structured natural language prompts. Extensive experiments on financial and energy datasets demonstrate that LoFT-LLM significantly outperforms strong baselines under both full-data and few-shot regimes, delivering superior accuracy, robustness, and interpretability.

Method: Define Decentralized Discrete Flow Matching objective by expressing probability generating velocity as linear combination of expert flows. Conduct experiments comparing decentralized vs centralized training using LLaVA and InternVL 2.5-1B models with fixed CLIP vision encoder and full-parameter fine-tuning.

Result: Demonstrate equivalence between decentralized and centralized training settings for multimodal language models across diverse benchmarks.

Conclusion: Decentralized training can achieve comparable performance to centralized approaches for multimodal language models, with theoretical framework supporting this equivalence.

Abstract: We present a theoretical analysis of decentralization of autoregressive generation. We define the Decentralized Discrete Flow Matching objective, by expressing probability generating velocity as a linear combination of expert flows. We also conduct experiments demonstrating the equivalence between decentralized and centralized training settings for multimodal language models across diverse set of benchmarks. Specifically, we compare two distinct paradigms: LLaVA and InternVL 2.5-1B, which uses a fixed CLIP vision encoder and performs full-parameter fine-tuning (ViT+MLP+LLM) during the instruction tuning stage.

Method: Systematic benchmarking of traditional ML, deep learning (including Transformers), and LLM approaches using the College Experience Sensing (CES) dataset. Evaluation across temporal windows, feature granularities, personalization strategies, and class imbalance handling.

Result: Deep learning models, particularly Transformers (Macro-F1 = 0.58), achieve the best overall performance. LLMs show strength in contextual reasoning but weaker temporal modeling. Personalization substantially improves forecasts of severe mental health states.

Conclusion: This work lays the groundwork for next-generation, adaptive, and human-centered mental health technologies by revealing how different modeling approaches interpret phone sensing behavioral data over time, advancing both research and real-world well-being applications.

Abstract: Smartphone sensing offers an unobtrusive and scalable way to track daily behaviors linked to mental health, capturing changes in sleep, mobility, and phone use that often precede symptoms of stress, anxiety, or depression. While most prior studies focus on detection that responds to existing conditions, forecasting mental health enables proactive support through Just-in-Time Adaptive Interventions. In this paper, we present the first comprehensive benchmarking study comparing traditional machine learning (ML), deep learning (DL), and large language model (LLM) approaches for mental health forecasting using the College Experience Sensing (CES) dataset, the most extensive longitudinal dataset of college student mental health to date. We systematically evaluate models across temporal windows, feature granularities, personalization strategies, and class imbalance handling. Our results show that DL models, particularly Transformer (Macro-F1 = 0.58), achieve the best overall performance, while LLMs show strength in contextual reasoning but weaker temporal modeling. Personalization substantially improves forecasts of severe mental health states. By revealing how different modeling approaches interpret phone sensing behavioral data over time, this work lays the groundwork for next-generation, adaptive, and human-centered mental health technologies that can advance both research and real-world well-being.

Method: Proposes RRE-PPO4Pred with three innovations: (1) Reinforced Recurrent Encoder framework formulating RNN internal adaptation as Markov Decision Process for unified decision environment; (2) Improved Prediction-oriented Proximal Policy Optimization (PPO4Pred) with Transformer-based agent and dynamic transition sampling; (3) Co-evolutionary optimization paradigm for adaptive interactive time series modeling.

Result: Comprehensive evaluations on five real-world datasets show the method consistently outperforms existing baselines and attains accuracy better than state-of-the-art Transformer models.

Conclusion: RRE-PPO4Pred provides an advanced time series predictor in engineering informatics by significantly improving RNN models’ time series modeling capacity and forecasting accuracy through reinforcement learning formulation.

Abstract: Time series forecasting plays a crucial role in contemporary engineering information systems for supporting decision-making across various industries, where Recurrent Neural Networks (RNNs) have been widely adopted due to their capability in modeling sequential data. Conventional RNN-based predictors adopt an encoder-only strategy with sliding historical windows as inputs to forecast future values. However, this approach treats all time steps and hidden states equally without considering their distinct contributions to forecasting, leading to suboptimal performance. To address this limitation, we propose a novel Reinforced Recurrent Encoder with Prediction-oriented Proximal Policy Optimization, RRE-PPO4Pred, which significantly improves time series modeling capacity and forecasting accuracy of the RNN models. The core innovations of this method are: (1) A novel Reinforced Recurrent Encoder (RRE) framework that enhances RNNs by formulating their internal adaptation as a Markov Decision Process, creating a unified decision environment capable of learning input feature selection, hidden skip connection, and output target selection; (2) An improved Prediction-oriented Proximal Policy Optimization algorithm, termed PPO4Pred, which is equipped with a Transformer-based agent for temporal reasoning and develops a dynamic transition sampling strategy to enhance sampling efficiency; (3) A co-evolutionary optimization paradigm to facilitate the learning of the RNN predictor and the policy agent, providing adaptive and interactive time series modeling. Comprehensive evaluations on five real-world datasets indicate that our method consistently outperforms existing baselines, and attains accuracy better than state-of-the-art Transformer models, thus providing an advanced time series predictor in engineering informatics.

Method: Systematic empirical investigation comparing models trained on hard-only constraints vs. mixed datasets, analysis of reward precision effects, attention mechanism analysis, and proposed data-centric refinement strategy prioritizing reward precision.

Result: Hard-only constraint training consistently outperforms mixed datasets; high-precision rewards develop transferable meta-skill; proposed approach outperforms baselines by 13.4% with 58% reduction in training time across five benchmarks.

Conclusion: Advocates for paradigm shift from indiscriminate pursuit of data diversity toward prioritizing high-precision rewards for effective alignment in instruction following tasks.

Abstract: A central belief in scaling reinforcement learning with verifiable rewards for instruction following (IF) tasks is that, a diverse mixture of verifiable hard and unverifiable soft constraints is essential for generalizing to unseen instructions. In this work, we challenge this prevailing consensus through a systematic empirical investigation. Counter-intuitively, we find that models trained on hard-only constraints consistently outperform those trained on mixed datasets. Extensive experiments reveal that reward precision, rather than constraint diversity, is the primary driver of effective alignment. The LLM judge suffers from a low recall rate in detecting false response, which leads to severe reward hacking, thereby undermining the benefits of diversity. Furthermore, analysis of the attention mechanism reveals that high-precision rewards develop a transferable meta-skill for IF. Motivated by these insights, we propose a simple yet effective data-centric refinement strategy that prioritizes reward precision. Evaluated on five benchmarks, our approach outperforms competitive baselines by 13.4% in performance while achieving a 58% reduction in training time, maintaining strong generalization beyond instruction following. Our findings advocate for a paradigm shift: moving away from the indiscriminate pursuit of data diversity toward high-precision rewards.

Method: Uses Bayesian optimization with kernel-of-kernels geometry, expected divergence-based distances between GP priors, and MDS embedding to map discrete kernel library into continuous Euclidean manifold for smooth BO.

Result: Demonstrated superior predictive accuracy and uncertainty calibration on synthetic benchmarks, real-world time-series datasets, and additive manufacturing case study compared to baselines including LLM-guided search.

Conclusion: Establishes reusable probabilistic geometry for kernel search with direct relevance to GP modeling and deep kernel learning.

Abstract: Gaussian Process (GP) regression is a powerful nonparametric Bayesian framework, but its performance depends critically on the choice of covariance kernel. Selecting an appropriate kernel is therefore central to model quality, yet remains one of the most challenging and computationally expensive steps in probabilistic modeling. We present a Bayesian optimization framework built on kernel-of-kernels geometry, using expected divergence-based distances between GP priors to explore kernel space efficiently. A multidimensional scaling (MDS) embedding of this distance matrix maps a discrete kernel library into a continuous Euclidean manifold, enabling smooth BO. In this formulation, the input space comprises kernel compositions, the objective is the log marginal likelihood, and featurization is given by the MDS coordinates. When the divergence yields a valid metric, the embedding preserves geometry and produces a stable BO landscape. We demonstrate the approach on synthetic benchmarks, real-world time-series datasets, and an additive manufacturing case study predicting melt-pool geometry, achieving superior predictive accuracy and uncertainty calibration relative to baselines including Large Language Model (LLM)-guided search. This framework establishes a reusable probabilistic geometry for kernel search, with direct relevance to GP modeling and deep kernel learning.

Method: Mixed-integer linear optimization (MILO) formulation maximizing buffered AUC (bAUC) as tightest concave lower bound on AUC, with group sparsity constraint to limit number of questions/variables.

Result: Computational experiments on real-world datasets show MILO method builds scoring systems with superior AUC values compared to baseline methods using regularization and stepwise regression.

Conclusion: Research advances MIO techniques for developing highly interpretable classification models by directly optimizing AUC through MILO framework.

Abstract: A scoring system is a linear classifier composed of a small number of explanatory variables, each assigned a small integer coefficient. This system is highly interpretable and allows predictions to be made with simple manual calculations without the need for a calculator. Several previous studies have used mixed-integer optimization (MIO) techniques to develop scoring systems for binary classification; however, they have not focused on directly maximizing AUC (i.e., area under the receiver operating characteristic curve), even though AUC is recognized as an essential evaluation metric for scoring systems. Our goal herein is to establish an effective MIO framework for constructing scoring systems that directly maximize the buffered AUC (bAUC) as the tightest concave lower bound on AUC. Our optimization model is formulated as a mixed-integer linear optimization (MILO) problem that maximizes bAUC subject to a group sparsity constraint for limiting the number of questions in the scoring system. Computational experiments using publicly available real-world datasets demonstrate that our MILO method can build scoring systems with superior AUC values compared to the baseline methods based on regularization and stepwise regression. This research contributes to the advancement of MIO techniques for developing highly interpretable classification models.

Method: The algorithm computes Hessian-inverse-vector products directly without storing the Hessian or its inverse. It achieves linear scaling in both time and storage relative to the number of network layers, similar to Pearlmutter’s algorithm for Hessian-vector products.

Result: The method provides exact solutions to linear systems involving deep network Hessians with dramatically improved efficiency - linear scaling in layers compared to quadratic storage and cubic operations of naive approaches.

Conclusion: This Hessian-inverse-vector product algorithm enables practical computation with deep network Hessians for large-scale problems where traditional methods would be computationally prohibitive.

Abstract: We describe an exact algorithm for solving linear systems $Hx=b$ where $H$ is the Hessian of a deep net. The method computes Hessian-inverse-vector products without storing the Hessian or its inverse in time and storage that scale linearly in the number of layers. Compared to the naive approach of first computing the Hessian, then solving the linear system, which takes storage that’s quadratic in the number of parameters and cubically many operations, our Hessian-inverse-vector product method scales roughly like Pearlmutter’s algorithm for computing Hessian-vector products.

Method: Uses Euler discretization to make liquid dynamics compatible with embedded systems, then performs dual-stage knowledge distillation to transfer temporal behavior from teacher to student models, followed by Pareto-guided selection based on joint error-cost objectives.

Result: Deployed student model achieves 0.0066 error predicting battery health over next 100 cycles (15.4% lower than teacher), reduces model size from 616 kB to 94 kB (84.7% reduction), and takes 21 ms per inference on Arduino Nano 33 BLE Sense with int8 deployment.

Conclusion: Small models can match or exceed large teachers for edge-based prognostics with proper supervision and selection. The DLNet framework can extend to other industrial analytics tasks with strict hardware constraints.

Abstract: Battery management systems increasingly require accurate battery health prognostics under strict on-device constraints. This paper presents DLNet, a practical framework with dual-stage distillation of liquid neural networks that turns a high-capacity model into compact and edge-deployable models for battery health prediction. DLNet first applies Euler discretization to reformulate liquid dynamics for embedded compatibility. It then performs dual-stage knowledge distillation to transfer the teacher model’s temporal behavior and recover it after further compression. Pareto-guided selection under joint error-cost objectives retains student models that balance accuracy and efficiency. We evaluate DLNet on a widely used dataset and validate real-device feasibility on an Arduino Nano 33 BLE Sense using int8 deployment. The final deployed student achieves a low error of 0.0066 when predicting battery health over the next 100 cycles, which is 15.4% lower than the teacher model. It reduces the model size from 616 kB to 94 kB with 84.7% reduction and takes 21 ms per inference on the device. These results support a practical smaller wins observation that a small model can match or exceed a large teacher for edge-based prognostics with proper supervision and selection. Beyond batteries, the DLNet framework can extend to other industrial analytics tasks with strict hardware constraints.

Method: PRPO segments reasoning sequences based on semantic clues, normalizes PRM scores into token-level advantages, and aligns their distribution with outcome advantages through location-parameter shift. This combines outcome reliability with process-level guidance in a critic-free framework without requiring value networks.

Result: On MATH500, PRPO improves Qwen2.5-Math-1.5B accuracy from 61.2% to 64.4% over GRPO using only eight rollouts and no value network, demonstrating efficient fine-grained credit assignment within critic-free optimization.

Conclusion: PRPO effectively addresses the limitations of both outcome-only and process-only reward methods by combining their strengths in a critic-free framework, enabling efficient fine-grained credit assignment for LLM policy optimization in multi-step reasoning tasks.

Abstract: Policy optimization for large language models often suffers from sparse reward signals in multi-step reasoning tasks. Critic-free methods like GRPO assign a single normalized outcome reward to all tokens, providing limited guidance for intermediate reasoning . While Process Reward Models (PRMs) offer dense feedback, they risk premature collapse when used alone, as early low-reward tokens can drive policies toward truncated outputs. We introduce Process Relative Policy Optimization (PRPO), which combines outcome reliability with process-level guidance in a critic-free framework. PRPO segments reasoning sequences based on semantic clues, normalizes PRM scores into token-level advantages, and aligns their distribution with outcome advantages through location-parameter shift. On MATH500, PRPO improves Qwen2.5-Math-1.5B accuracy from 61.2% to 64.4% over GRPO using only eight rollouts and no value network, demonstrating efficient fine-grained credit assignment within critic-free optimization. Code is available at: https://github.com/SchumiDing/srpocode

Method: Proposes Free-RBF-KAN with adaptive learning grids and trainable smoothness. Uses freely learnable RBF shapes that dynamically align grid representations with activation patterns. Treats smoothness as a kernel parameter optimized jointly with network weights without increasing computational complexity.

Result: Free-RBF-KAN achieves accuracy comparable to original B-spline-based KAN while delivering faster training and inference. Validated through multiscale function approximation, physics-informed machine learning, and PDE solution operator learning.

Conclusion: Free-RBF-KAN provides a compelling balance between computational efficiency and adaptive resolution, particularly effective for high-dimensional structured modeling tasks.

Abstract: Kolmogorov-Arnold Networks (KANs) have shown strong potential for efficiently approximating complex nonlinear functions. However, the original KAN formulation relies on B-spline basis functions, which incur substantial computational overhead due to De Boor’s algorithm. To address this limitation, recent work has explored alternative basis functions such as radial basis functions (RBFs) that can improve computational efficiency and flexibility. Yet, standard RBF-KANs often sacrifice accuracy relative to the original KAN design. In this work, we propose Free-RBF-KAN, a RBF-based KAN architecture that incorporates adaptive learning grids and trainable smoothness to close this performance gap. Our method employs freely learnable RBF shapes that dynamically align grid representations with activation patterns, enabling expressive and adaptive function approximation. Additionally, we treat smoothness as a kernel parameter optimized jointly with network weights, without increasing computational complexity. We provide a general universality proof for RBF-KANs, which encompasses our Free-RBF-KAN formulation. Through a broad set of experiments, including multiscale function approximation, physics-informed machine learning, and PDE solution operator learning, Free-RBF-KAN achieves accuracy comparable to the original B-spline-based KAN while delivering faster training and inference. These results highlight Free-RBF-KAN as a compelling balance between computational efficiency and adaptive resolution, particularly for high-dimensional structured modeling tasks.

Method: Propose pressure-field coordination: agents operate locally on shared artifacts guided by pressure gradients from quality signals, with temporal decay preventing premature convergence. Formalized as optimization over pressure landscape with convergence guarantees.

Result: On Latin Square constraint satisfaction (1,078 trials): pressure-field matches hierarchical control (38.2% vs 38.8% solve rate, p=0.94). Both outperform sequential (23.3%), random (11.7%), and conversation-based dialogue (8.6%). Temporal decay essential (49-fold pressure increase without it). Maintains consistent performance from 2 to 32 agents.

Conclusion: Implicit coordination through shared pressure gradients achieves parity with explicit hierarchical control while dramatically outperforming explicit dialogue-based coordination. Constraint-driven emergence offers simpler, equally effective foundation for multi-agent AI.

Abstract: Current multi-agent LLM frameworks rely on explicit orchestration patterns borrowed from human organizational structures: planners delegate to executors, managers coordinate workers, and hierarchical control flow governs agent interactions. These approaches suffer from coordination overhead that scales poorly with agent count and task complexity. We propose a fundamentally different paradigm inspired by natural coordination mechanisms: agents operate locally on a shared artifact, guided only by pressure gradients derived from measurable quality signals, with temporal decay preventing premature convergence. We formalize this as optimization over a pressure landscape and prove convergence guarantees under mild conditions. Empirically, on Latin Square constraint satisfaction across 1,078 trials, pressure-field coordination matches hierarchical control (38.2% vs 38.8% aggregate solve rate, p=0.94, indicating statistical equivalence). Both significantly outperform sequential (23.3%), random (11.7%), and conversation-based multi-agent dialogue (8.6%, p<0.00001). Temporal decay is essential: disabling it increases final pressure 49-fold (d=4.15). On easy problems, pressure-field achieves 87% solve rate. The approach maintains consistent performance from 2 to 32 agents. Our key finding: implicit coordination through shared pressure gradients achieves parity with explicit hierarchical control while dramatically outperforming explicit dialogue-based coordination. This suggests that constraint-driven emergence offers a simpler, equally effective foundation for multi-agent AI.

Method: The authors evaluate KV cache reuse strategies across GSM8K, MMLU, and HumanEval benchmarks. They introduce Judge Consistency Rate (JCR) to quantify consistency with dense prefill, analyze cross-candidate attention patterns, and conduct ablations to study cross-candidate interaction.

Result: Reuse strategies effective for execution agents severely perturb judge behavior - while end-task accuracy may appear stable, judge selections become highly inconsistent with dense prefill. Reuse systematically weakens cross-candidate attention, especially for later candidate blocks.

Conclusion: KV cache reuse has a previously overlooked failure mode in judge-centric inference, which represents a distinct regime requiring dedicated, risk-aware system design rather than simply applying execution agent optimizations.

Abstract: Multi-agent LLM systems routinely generate multiple candidate responses that are aggregated by an LLM judge. To reduce the dominant prefill cost in such pipelines, recent work advocates KV cache reuse across partially shared contexts and reports substantial speedups for generation agents. In this work, we show that these efficiency gains do not transfer uniformly to judge-centric inference. Across GSM8K, MMLU, and HumanEval, we find that reuse strategies that are effective for execution agents can severely perturb judge behavior: end-task accuracy may appear stable, yet the judge’s selection becomes highly inconsistent with dense prefill. We quantify this risk using Judge Consistency Rate (JCR) and provide diagnostics showing that reuse systematically weakens cross-candidate attention, especially for later candidate blocks. Our ablation further demonstrates that explicit cross-candidate interaction is crucial for preserving dense-prefill decisions. Overall, our results identify a previously overlooked failure mode of KV cache reuse and highlight judge-centric inference as a distinct regime that demands dedicated, risk-aware system design.

Method: Introduces Agent Contracts framework that unifies I/O specifications, multi-dimensional resource constraints, temporal boundaries, and success criteria with explicit lifecycle semantics and conservation laws for hierarchical coordination.

Result: Empirical validation shows 90% token reduction with 525x lower variance in iterative workflows, zero conservation violations in multi-agent delegation, and measurable quality-resource tradeoffs through contract modes.

Conclusion: Agent Contracts provide formal foundations for predictable, auditable, and resource-bounded autonomous AI deployment, extending the contract metaphor from task allocation to resource-bounded execution.

Abstract: The Contract Net Protocol (1980) introduced coordination through contracts in multi-agent systems. Modern agent protocols standardize connectivity and interoperability; yet, none provide formal, resource governance-normative mechanisms to bound how much agents may consume or how long they may operate. We introduce Agent Contracts, a formal framework that extends the contract metaphor from task allocation to resource-bounded execution. An Agent Contract unifies input/output specifications, multi-dimensional resource constraints, temporal boundaries, and success criteria into a coherent governance mechanism with explicit lifecycle semantics. For multi-agent coordination, we establish conservation laws ensuring delegated budgets respect parent constraints, enabling hierarchical coordination through contract delegation. Empirical validation across four experiments demonstrates 90% token reduction with 525x lower variance in iterative workflows, zero conservation violations in multi-agent delegation, and measurable quality-resource tradeoffs through contract modes. Agent Contracts provide formal foundations for predictable, auditable, and resource-bounded autonomous AI deployment.

Method: Spatially explicit agent-based model integrating road-network GIS, synthetic populations via Iterative Proportional Fitting, and behavioral heterogeneity; applied to Premeno, Italy with baseline vs. relocation scenarios; analyzed using system-level and micro-spatial metrics across 40 batches and 50 replications per scenario.

Result: Results show aggregate neutrality but pronounced local redistribution of accessibility; spatial impedance dominates accessibility while behavioral capacity modulates care effort; demonstrates emergence, heterogeneity, and feedback properties of complex adaptive social systems.

Conclusion: Computational social simulation reveals policy trade-offs between spatial efficiency, social equity, and care sustainability in ageing territories, highlighting how service relocation redistributes accessibility locally despite aggregate neutrality.

Abstract: Ageing societies face increasing strain on formal and informal care systems, particularly in low-density mountainous municipalities where sparse services and steep terrain constrain access. This study presents a spatially explicit agent-based model that integrates a road-network GIS, synthetic populations derived through Iterative Proportional Fitting, and behavioural heterogeneity to examine how alternative service configurations shape accessibility and caregiver burden. The model, applied to Premeno (Piedmont, Italy), compares a baseline distribution of ambulatory services with a relocation scenario at Villa Bernocchi. System-level indicators (Caregiver Effort, Overwhelmed Caregivers, Hours Not Cared, Walkability) and micro-spatial metrics (Walkability, Detour Ratio, Proximity) are analysed across 40 batches and 50 stochastic replications per scenario. Results reveal aggregate neutrality but pronounced local redistribution of accessibility. Sensitivity analysis shows that spatial impedance dominates accessibility, whereas behavioural capacity modulates care effort. The findings illustrate distinctive properties of complex adaptive social systems - emergence, heterogeneity, and feedback - demonstrating how computational social simulation can highlight policy trade-offs between spatial efficiency, social equity, and care sustainability in ageing territories.

Method: The framework segments video ads into functional units, classifies each unit’s functionality using an advertising-specific functional role taxonomy, aggregates functional sequences across ads to recover data-driven storyline structures, and applies this to 50k social media video ads across four industry subverticals.

Result: Analysis reveals that story-based creatives improve video retention, and the framework identifies top-performing story arcs that can guide advertisers in creative design.

Conclusion: The framework demonstrates the value of using domain knowledge to guide MLLMs in generating scalable insights for video ad storylines, making it a versatile tool for understanding video creatives in general.

Abstract: We propose MLLM-VADStory, a novel domain knowledge-guided multimodal large language models (MLLM) framework to systematically quantify and generate insights for video ad storyline understanding at scale. The framework is centered on the core idea that ad narratives are structured by functional intent, with each scene unit performing a distinct communicative function, delivering product and brand-oriented information within seconds. MLLM-VADStory segments ads into functional units, classifies each unit’s functionality using a novel advertising-specific functional role taxonomy, and then aggregates functional sequences across ads to recover data-driven storyline structures. Applying the framework to 50k social media video ads across four industry subverticals, we find that story-based creatives improve video retention, and we recommend top-performing story arcs to guide advertisers in creative design. Our framework demonstrates the value of using domain knowledge to guide MLLMs in generating scalable insights for video ad storylines, making it a versatile tool for understanding video creatives in general.

Method: Establishes a reproducible end-to-end pipeline covering feature extraction, multimodal fusion, cougher-independent evaluation, and uncertainty quantification. Uses cough recordings and clinical metadata from a recently compiled multi-country dataset, with consistent reporting of clinically relevant metrics.

Result: Provides a strong, well-documented baseline for TB prediction that quantifies performance for both cough audio-only and fused (audio + clinical metadata) models, with full experimental protocol released for benchmarking.

Conclusion: The proposed framework serves as a common reference point to reduce methodological variance and enable fair comparison, addressing the current barriers to progress in TB detection from audio and clinical data.

Abstract: In this paper, we propose a standardized framework for automatic tuberculosis (TB) detection from cough audio and routinely collected clinical data using machine learning. While TB screening from audio has attracted growing interest, progress is difficult to measure because existing studies vary substantially in datasets, cohort definitions, feature representations, model families, validation protocols, and reported metrics. Consequently, reported gains are often not directly comparable, and it remains unclear whether improvements stem from modeling advances or from differences in data and evaluation. We address this gap by establishing a strong, well-documented baseline for TB prediction using cough recordings and accompanying clinical metadata from a recently compiled dataset from several countries. Our pipeline is reproducible end-to-end, covering feature extraction, multimodal fusion, cougher-independent evaluation, and uncertainty quantification, and it reports a consistent suite of clinically relevant metrics to enable fair comparison. We further quantify performance for cough audio-only and fused (audio + clinical metadata) models, and release the full experimental protocol to facilitate benchmarking. This baseline is intended to serve as a common reference point and to reduce methodological variance that currently holds back progress in the field.

Method: Quantitative analysis across five proxy task configurations (AutoEncoders, classification, source separation, contrastive learning, pre-trained models) using linear probe and Mahalanobis distance to evaluate learned representations.

Result: Strong proxy performance doesn’t necessarily improve ASD: classification saturates due to insufficient difficulty, contrastive learning fails with limited data diversity, while only source separation shows strong positive correlation between improved separation and better anomaly detection.

Conclusion: Task difficulty and objective alignment are critical for effective proxy tasks in ASD; a three-stage alignment verification protocol is proposed to guide proxy task design.

Abstract: Anomalous sound detection (ASD) typically involves self-supervised proxy tasks to learn feature representations from normal sound data, owing to the scarcity of anomalous samples. In ASD research, proxy tasks such as AutoEncoders operate under the explicit assumption that models trained on normal data will increase the reconstruction errors related to anomalies. A natural extension suggests that improved proxy task performance should improve ASD capability; however, this relationship has received little systematic attention. This study addresses this research gap by quantitatively analyzing the relationship between proxy task metrics and ASD performance across five configurations, namely, AutoEncoders, classification, source separation, contrastive learning, and pre-trained models. We evaluate the learned representations using linear probe (linear separability) and Mahalanobis distance (distributional compactness). Our experiments reveal that strong proxy performance does not necessarily improve anomalous sound detection performance. Specifically, classification tasks experience performance saturation owing to insufficient task difficulty, whereas contrastive learning fails to learn meaningful features owing to limited data diversity. Notably, source separation is the only task demonstrating a strong positive correlation, such that improved separation consistently improves anomaly detection. Based on these findings, we highlight the critical importance of task difficulty and objective alignment. Finally, we propose a three-stage alignment verification protocol to guide the design of highly effective proxy tasks for ASD systems.

Method: Proposes a weakly supervised framework combining CTC-based acoustic model with sequence-level rhythmic rescoring using a Tāla-Independent Static-Dynamic Rhythmic Model (TI-SDRM) that integrates long-term rhythmic structure with short-term adaptive dynamics through adaptive interpolation.

Result: Establishes first benchmark for weakly supervised TST using new real-world tabla solo dataset and synthetic dataset. Experiments show consistent and substantial reductions in stroke error rate over acoustic-only decoding.

Conclusion: Demonstrates the importance of explicit rhythmic structure for accurate tabla stroke transcription in weakly supervised settings, providing a practical solution that doesn’t require costly temporal alignments.

Abstract: Tabla Stroke Transcription (TST) is central to the analysis of rhythmic structure in Hindustani classical music, yet remains challenging due to complex rhythmic organization and the scarcity of strongly annotated data. Existing approaches largely rely on fully supervised learning with onset-level annotations, which are costly and impractical at scale. This work addresses TST in a weakly supervised setting, using only symbolic stroke sequences without temporal alignment. We propose a framework that combines a CTC-based acoustic model with sequence-level rhythmic rescoring. The acoustic model produces a decoding lattice, which is refined using a \textbf{$T\bar{a}la$}-Independent Static–Dynamic Rhythmic Model (TI-SDRM) that integrates long-term rhythmic structure with short-term adaptive dynamics through an adaptive interpolation mechanism. We curate a new real-world tabla solo dataset and a complementary synthetic dataset, establishing the first benchmark for weakly supervised TST in Hindustani classical music. Experiments demonstrate consistent and substantial reductions in stroke error rate over acoustic-only decoding, confirming the importance of explicit rhythmic structure for accurate transcription.

Method: Developed a two-stage framework: candidate selection using thresholded data to refine regions of interest, followed by decision-making with stage-specific feature processing. Used thresholding to filter out low-salience features and conducted simulations to evaluate effects on feature maps, candidate localization, and multi-feature scenarios.

Result: Thresholding improves observer performance by excluding low-salience features, especially in noisy environments. Intermediate thresholds often outperform no thresholding, showing that retaining only relevant features is more effective. The model can be trained effectively with fewer images while maintaining alignment with human performance.

Conclusion: The proposed framework can predict human visual search performance in clinically realistic tasks and provides solutions for model observer training with limited resources. The approach has broader applications in computer vision, machine learning, defense, and security image analysis.

Abstract: This study advances task-based image quality assessment by developing an anthropomorphic thresholded visual-search model observer. The model is an ideal observer for thresholded data inspired by the human visual system, allowing selective processing of high-salience features to improve discrimination performance. By filtering out irrelevant variability, the model enhances diagnostic accuracy and computational efficiency. The observer employs a two-stage framework: candidate selection and decision-making. Using thresholded data during candidate selection refines regions of interest, while stage-specific feature processing optimizes performance. Simulations were conducted to evaluate the effects of thresholding on feature maps, candidate localization, and multi-feature scenarios. Results demonstrate that thresholding improves observer performance by excluding low-salience features, particularly in noisy environments. Intermediate thresholds often outperform no thresholding, indicating that retaining only relevant features is more effective than keeping all features. Additionally, the model demonstrates effective training with fewer images while maintaining alignment with human performance. These findings suggest that the proposed novel framework can predict human visual search performance in clinically realistic tasks and provide solutions for model observer training with limited resources. Our novel approach has applications in other areas where human visual search and detection tasks are modeled such as in computer vision, machine learning, defense and security image analysis.

Method: TIMM-ProRS integrates Vision Transformer (ViT), Convolutional Neural Network (CNN), and Graph Neural Network (GNN) with multi-modal fusion. It uniquely leverages both retinal images and temporal biomarkers (HbA1c, retinal thickness) to capture multi-modal and temporal dynamics.

Result: The model achieves 97.8% accuracy and an F1-score of 0.96 across diverse datasets including APTOS 2019 (trained), Messidor-2, RFMiD, EyePACS, and Messidor-1 (validated), demonstrating state-of-the-art performance and outperforming existing methods like RSG-Net and DeepDR.

Conclusion: TIMM-ProRS enables early, precise, and interpretable diagnosis of diabetic retinopathy, supporting scalable telemedical management and enhancing global eye health sustainability through its multi-modal temporal approach.

Abstract: Diabetic retinopathy (DR), affecting millions globally with projections indicating a significant rise, poses a severe blindness risk and strains healthcare systems. Diagnostic complexity arises from visual symptom overlap with conditions like age-related macular degeneration and hypertensive retinopathy, exacerbated by high misdiagnosis rates in underserved regions. This study introduces TIMM-ProRS, a novel deep learning framework integrating Vision Transformer (ViT), Convolutional Neural Network (CNN), and Graph Neural Network (GNN) with multi-modal fusion. TIMM-ProRS uniquely leverages both retinal images and temporal biomarkers (HbA1c, retinal thickness) to capture multi-modal and temporal dynamics. Evaluated comprehensively across diverse datasets including APTOS 2019 (trained), Messidor-2, RFMiD, EyePACS, and Messidor-1 (validated), the model achieves 97.8% accuracy and an F1-score of 0.96, demonstrating state-of-the-art performance and outperforming existing methods like RSG-Net and DeepDR. This approach enables early, precise, and interpretable diagnosis, supporting scalable telemedical management and enhancing global eye health sustainability.

Method: Innovative cascade approach: 1) Simple, efficient detection model for single ROI localization (multi-task: encoder-decoder segmentation + fully connected network for detection), 2) Segmentation on detected ROI. Compared against one-step segmentation on full image and nnU-Net.

Result: Achieved mean Dice similarity coefficient of 0.944 with >0.9 for all cases, using only one-third of the computing power compared to alternatives.

Conclusion: The proposed simple cascade solution achieves state-of-the-art performance while being compact, robust, and computationally efficient, making it ideal for clinical applications in aortic segmentation.

Abstract: Thoracic aortic dissection and aneurysms are the most lethal diseases of the aorta. The major hindrance to treatment lies in the accurate analysis of the medical images. More particularly, aortic segmentation of the 3D image is often tedious and difficult. Deep-learning-based segmentation models are an ideal solution, but their inability to deliver usable outputs in difficult cases and their computational cost cause their clinical adoption to stay limited. This study presents an innovative approach for efficient aortic segmentation using targeted region of interest (ROI) detection. In contrast to classical detection models, we propose a simple and efficient detection model that can be widely applied to detect a single ROI. Our detection model is trained as a multi-task model, using an encoder-decoder architecture for segmentation and a fully connected network attached to the bottleneck for detection. We compare the performance of a one-step segmentation model applied to a complete image, nnU-Net and our cascade model composed of a detection and a segmentation step. We achieve a mean Dice similarity coefficient of 0.944 with over 0.9 for all cases using a third of the computing power. This simple solution achieves state-of-the-art performance while being compact and robust, making it an ideal solution for clinical applications.

Method: Formulates image model using factor graphs with NUP (normal with unknown parameters) priors, where local parameters are piecewise constant and automatically adapted. Computations involve iterative conjugate-gradient steps and Gaussian message passing algorithms.

Result: Demonstrated successful applications to image denoising and contrast enhancement tasks, showing the model’s effectiveness in adapting to local image characteristics.

Conclusion: The proposed piecewise smooth image model with adaptive local parameters and factor graph formulation provides an effective framework for image processing tasks, particularly for denoising and contrast enhancement where local adaptation is crucial.

Abstract: We propose a novel piecewise smooth image model with piecewise constant local parameters that are automatically adapted to each image. Technically, the model is formulated in terms of factor graphs with NUP (normal with unknown parameters) priors, and the pertinent computations amount to iterations of conjugate-gradient steps and Gaussian message passing. The proposed model and algorithms are demonstrated with applications to denoising and contrast enhancement.

Method: Introduces M3CoTBench benchmark featuring: 1) diverse multi-level difficulty dataset covering 24 examination types, 2) 13 varying-difficulty tasks, 3) CoT-specific evaluation metrics (correctness, efficiency, impact, consistency) tailored to clinical reasoning, and 4) performance analysis of multiple MLLMs.

Result: The benchmark systematically evaluates CoT reasoning across diverse medical imaging tasks, revealing current limitations of MLLMs in generating reliable and clinically interpretable reasoning paths.

Conclusion: M3CoTBench aims to foster development of transparent, trustworthy, and diagnostically accurate AI systems for healthcare by providing proper evaluation of reasoning processes in medical image understanding.

Abstract: Chain-of-Thought (CoT) reasoning has proven effective in enhancing large language models by encouraging step-by-step intermediate reasoning, and recent advances have extended this paradigm to Multimodal Large Language Models (MLLMs). In the medical domain, where diagnostic decisions depend on nuanced visual cues and sequential reasoning, CoT aligns naturally with clinical thinking processes. However, Current benchmarks for medical image understanding generally focus on the final answer while ignoring the reasoning path. An opaque process lacks reliable bases for judgment, making it difficult to assist doctors in diagnosis. To address this gap, we introduce a new M3CoTBench benchmark specifically designed to evaluate the correctness, efficiency, impact, and consistency of CoT reasoning in medical image understanding. M3CoTBench features 1) a diverse, multi-level difficulty dataset covering 24 examination types, 2) 13 varying-difficulty tasks, 3) a suite of CoT-specific evaluation metrics (correctness, efficiency, impact, and consistency) tailored to clinical reasoning, and 4) a performance analysis of multiple MLLMs. M3CoTBench systematically evaluates CoT reasoning across diverse medical imaging tasks, revealing current limitations of MLLMs in generating reliable and clinically interpretable reasoning, and aims to foster the development of transparent, trustworthy, and diagnostically accurate AI systems for healthcare. Project page at https://juntaojianggavin.github.io/projects/M3CoTBench/.

Method: 1. Survey development of GANs and popular variants for image applications. 2. Analyze motivations, implementations, and differences of GAN-based optimization methods and discriminative learning for SISR in supervised, semi-supervised, and unsupervised manners. 3. Compare performances of popular GANs on public datasets using quantitative and qualitative analysis.

Result: The paper provides comprehensive analysis of GANs in SISR, comparing different approaches through network architectures, prior knowledge, loss functions, and multiple tasks. Performance comparisons are made on public datasets.

Conclusion: Highlights challenges of GANs in SISR and identifies potential research directions for future work in the field.

Abstract: Single image super-resolution (SISR) has played an important role in the field of image processing. Recent generative adversarial networks (GANs) can achieve excellent results on low-resolution images. However, there are little literatures summarizing different GANs in SISR. In this paper, we conduct a comparative study of GANs from different perspectives. We begin by surveying the development of GANs and popular GAN variants for image-related applications, and then analyze motivations, implementations and differences of GANs based optimization methods and discriminative learning for image super-resolution in terms of supervised, semi-supervised and unsupervised manners, where these GANs are analyzed via integrating different network architectures, prior knowledge, loss functions and multiple tasks. Secondly, we compare the performances of these popular GANs on public datasets via quantitative and qualitative analysis in SISR. Finally, we highlight challenges of GANs and potential research points for SISR.

Method: Used 331 pre-treatment rectal cancer T2-weighted MRI scans from three European hospitals. Applied self-supervised frequency domain harmonization to reduce scanner variability. Tested three classifiers: SeResNet, UMedPT with MLP head, and UMedPT_LR (logistic regression with frozen UMedPT features). Used Grad-CAM for visualization.

Result: UMedPT_LR achieved best EVI performance with multiplanar fusion (AUC=0.82). For MFI, UMedPT on axial harmonized images performed best (AUC=0.77). Both outperformed CHAIMELEON 2024 benchmark (EVI: 0.82 vs 0.74; MFI: 0.77 vs 0.75). Harmonization enhanced MFI classification, and multiplanar fusion boosted EVI performance.

Conclusion: The foundation model-driven framework with frequency domain harmonization and multiplanar fusion achieves state-of-the-art performance for automated EVI and MFI classification on MRI, demonstrating strong generalizability across multiple centers.

Abstract: Objectives Accurate MRI-based identification of extramural vascular invasion (EVI) and mesorectal fascia invasion (MFI) is crucial for risk-stratified rectal cancer treatment. However, subjective visual assessment and inter-institutional variability limit diagnostic consistency. This study developed and externally evaluated a multi-centre, foundation model-driven framework that automatically classifies EVI and MFI on axial and sagittal MRI. Methods A total of 331 pre-treatment rectal cancer T2-weighted MRI scans from three European hospitals were retrospectively recruited. A self-supervised frequency domain harmonization strategy was applied to reduce scanner variability. Three classifiers, SeResNet, the universal biomedical pretrained model (UMedPT) with a multilayer perceptron head, and a logistic-regression variant using frozen UMedPT features (UMedPT_LR), were trained (n=265) and tested (n=66). Gradient-weighted class activation mapping (Grad-CAM) visualized model predictions. Results UMedPT_LR achieved the best EVI performance with multiplanar fusion (AUC=0.82, test set). For MFI, UMedPT trained on axial harmonized images yielded the highest performance (AUC = 0.77). Both tasks outperformed the CHAIMELEON 2024 benchmark (EVI: 0.82 vs 0.74; MFI: 0.77 vs 0.75). Harmonization enhanced MFI classification, and multiplanar fusion further boosted EVI performance. Grad-CAM confirmed biologically plausible attention on peritumoral regions (EVI) and mesorectal fascia margins (MFI). Conclusion The proposed foundation model-driven framework, leveraging frequency domain harmonization and multiplanar fusion, achieves state-of-the-art performance for automated EVI and MFI classification on MRI, demonstrating strong generalizability across multiple centers.

Method: QANA uses quantization-aware design with spike-compatible transformations, bounding activations, aligning normalization with low-bit quantization, Ghost-based feature generation, spatially-aware efficient channel attention, and squeeze-and-excitation recalibration for SNN conversion stability.

Result: On HAM10000: 91.6% Top-1 accuracy (3.5pp gain), 91.0% macro F1 (12.0pp gain). On clinical dataset: 90.8% Top-1 accuracy (3.2pp gain), 81.7% macro F1 (3.6pp gain). On BrainChip Akida: 1.5ms per image with 1.7mJ per image (94.6% lower latency, 99.0% lower energy vs GPU).

Conclusion: QANA enables efficient, accurate on-device skin lesion analysis with stable SNN conversion, supporting incremental updates on edge hardware without full retraining or data offloading, making it suitable for real-world clinical deployment.

Abstract: On-device skin lesion analysis is constrained by the compute and energy cost of conventional CNN inference and by the need to update models as new patient data become available. We propose QANA, a quantization-aware CNN backbone embedded in an end-to-end pipeline engineered for conversion-stable neuromorphic execution. QANA replaces conversion-fragile components with spike-compatible transformations by bounding intermediate activations and aligning normalization with low-bit quantization, reducing conversion-induced distortion that disproportionately impacts rare classes. Efficiency is achieved through Ghost-based feature generation under tight FLOP budgets, while spatially-aware efficient channel attention and squeeze-and-excitation recalibrate channels without heavy global operators that are difficult to map to spiking cores. The resulting quantized projection head produces SNN-ready logits and enables incremental updates on edge hardware without full retraining or data offloading. On HAM10000, QANA achieves 91.6% Top-1 accuracy and 91.0% macro F1, improving the strongest converted SNN baseline by 3.5 percentage points in Top-1 accuracy, corresponding to a 4.0% relative gain, and by 12.0 points in macro F1, corresponding to a 15.2% relative gain. On a clinical dataset, QANA achieves 90.8% Top-1 accuracy and 81.7% macro F1, improving the strongest converted SNN baseline by 3.2 points in Top-1 accuracy, which corresponds to a 3.7% relative gain, and by 3.6 points in macro F1, corresponding to a 4.6% relative gain. When deployed on BrainChip Akida, QANA runs in 1.5 ms per image with 1.7 mJ per image, corresponding to 94.6% lower latency and 99.0% lower energy than its GPU-based CNN implementation.

Method: Proposes HiFi-Mamba with dual-stream architecture: WL blocks perform fidelity-preserving spectral decoupling into low- and high-frequency streams. HiFi-Mamba blocks focus on low-frequency structures while selectively integrating high-frequency features through adaptive state-space modulation. Uses streamlined unidirectional traversal to eliminate scanning redundancy while maintaining long-range modeling capability.

Result: Extensive experiments on standard MRI reconstruction benchmarks show HiFi-Mamba consistently outperforms state-of-the-art CNN-based, Transformer-based, and other Mamba-based models in reconstruction accuracy while maintaining compact and efficient design.

Conclusion: HiFi-Mamba effectively addresses the limitations of existing Mamba variants for MRI reconstruction, achieving superior reconstruction accuracy with improved computational efficiency through its dual-stream architecture and streamlined traversal strategy.

Abstract: Reconstructing high-fidelity MR images from undersampled k-space data remains a challenging problem in MRI. While Mamba variants for vision tasks offer promising long-range modeling capabilities with linear-time complexity, their direct application to MRI reconstruction inherits two key limitations: (1) insensitivity to high-frequency anatomical details; and (2) reliance on redundant multi-directional scanning. To address these limitations, we introduce High-Fidelity Mamba (HiFi-Mamba), a novel dual-stream Mamba-based architecture comprising stacked W-Laplacian (WL) and HiFi-Mamba blocks. Specifically, the WL block performs fidelity-preserving spectral decoupling, producing complementary low- and high-frequency streams. This separation enables the HiFi-Mamba block to focus on low-frequency structures, enhancing global feature modeling. Concurrently, the HiFi-Mamba block selectively integrates high-frequency features through adaptive state-space modulation, preserving comprehensive spectral details. To eliminate the scanning redundancy, the HiFi-Mamba block adopts a streamlined unidirectional traversal strategy that preserves long-range modeling capability with improved computational efficiency. Extensive experiments on standard MRI reconstruction benchmarks demonstrate that HiFi-Mamba consistently outperforms state-of-the-art CNN-based, Transformer-based, and other Mamba-based models in reconstruction accuracy while maintaining a compact and efficient model design.