Method: 1) Build multimodal human-centric dataset with strict filtering, audio-video pairing, performance understanding, and identity-aware multi-reference extraction; 2) Train 17B-parameter Diffusion Transformer (Base LPM) for controllable, identity-consistent performance; 3) Distill into causal streaming generator (Online LPM) for low-latency, infinite-length interaction.

Result: LPM 1.0 achieves state-of-the-art results across all evaluated dimensions while maintaining real-time inference, serving as a visual engine for conversational agents, live streaming characters, and game NPCs.

Conclusion: LPM 1.0 successfully addresses the performance trilemma by enabling real-time, identity-stable, infinite-length audio-visual conversational performance generation, with the proposed LPM-Bench benchmark for systematic evaluation.

Abstract: Performance, the externalization of intent, emotion, and personality through visual, vocal, and temporal behavior, is what makes a character alive. Learning such performance from video is a promising alternative to traditional 3D pipelines. However, existing video models struggle to jointly achieve high expressiveness, real-time inference, and long-horizon identity stability, a tension we call the performance trilemma. Conversation is the most comprehensive performance scenario, as characters simultaneously speak, listen, react, and emote while maintaining identity over time. To address this, we present LPM 1.0 (Large Performance Model), focusing on single-person full-duplex audio-visual conversational performance. Concretely, we build a multimodal human-centric dataset through strict filtering, speaking-listening audio-video pairing, performance understanding, and identity-aware multi-reference extraction; train a 17B-parameter Diffusion Transformer (Base LPM) for highly controllable, identity-consistent performance through multimodal conditioning; and distill it into a causal streaming generator (Online LPM) for low-latency, infinite-length interaction. At inference, given a character image with identity-aware references, LPM 1.0 generates listening videos from user audio and speaking videos from synthesized audio, with text prompts for motion control, all at real-time speed with identity-stable, infinite-length generation. LPM 1.0 thus serves as a visual engine for conversational agents, live streaming characters, and game NPCs. To systematically evaluate this setting, we propose LPM-Bench, the first benchmark for interactive character performance. LPM 1.0 achieves state-of-the-art results across all evaluated dimensions while maintaining real-time inference.

Method: Proposes a multi-stage training strategy based on entropy allocation analysis: 1) Redesigns pretraining to reduce speech-text modality gap, 2) Introduces iterative asynchronous SFT between alignment and joint SFT to preserve functional decoupling and constrain encoder representation drift, 3) Uses capability-boundary awareness to optimize parameter efficiency.

Result: Achieves competitive performance with state-of-the-art models using only 2.3B parameters on Mandarin and English benchmarks, while effectively mitigating hallucinations through decoupling-oriented design.

Conclusion: The entropy allocation perspective provides a principled framework for designing efficient LLM-based ASR systems that balance performance with practical deployment considerations like latency and hallucination robustness.

Abstract: Integrating large language models (LLMs) into automatic speech recognition (ASR) has become a dominant paradigm. Although recent LLM-based ASR models have shown promising performance on public benchmarks, it remains challenging to balance recognition quality with latency and overhead, while hallucinations further limit real-world deployment. In this study, we revisit LLM-based ASR from an entropy allocation perspective and introduce three metrics to characterize how training paradigms allocate entropy reduction between the speech encoder and the LLM. To remedy entropy-allocation inefficiencies in prevailing approaches, we propose a principled multi-stage training strategy grounded in capability-boundary awareness, optimizing parameter efficiency and hallucination robustness. Specifically, we redesign the pretraining strategy to alleviate the speech-text modality gap, and further introduce an iterative asynchronous SFT stage between alignment and joint SFT to preserve functional decoupling and constrain encoder representation drift. Experiments on Mandarin and English benchmarks show that our method achieves competitive performance with state-of-the-art models using only 2.3B parameters, while also effectively mitigating hallucinations through our decoupling-oriented design.

Method: Proposes GG-AVSE framework with GG-VM module that combines gaze signals with YOLO5Face detector to extract target speaker’s facial features, integrated with pretrained AVSEMamba model via zero-shot merging and partial visual fine-tuning strategies.

Result: GG-AVSE achieves substantial improvements: 10.08% PESQ (2.370 to 2.609), 5.18% STOI (0.8802 to 0.9258), and 23.69% SI-SDR (9.16 to 11.33) over gaze-free baselines. Introduces AVSEC2-Gaze dataset for evaluation.

Conclusion: Gaze provides an effective cue for resolving target-speaker ambiguity in multi-talker environments, and GG-AVSE demonstrates scalability for real-world applications.

Abstract: This paper presents a Gaze-Guided Audio-Visual Speech Enhancement (GG-AVSE) framework to address the cocktail party problem. A major challenge in conventional AVSE is identifying the listener’s intended speaker in multi-talker environments. GG-AVSE addresses this issue by exploiting gaze direction as a supervisory cue for target-speaker selection. Specifically, we propose the GG-VM module, which combines gaze signals with a YOLO5Face detector to extract the target speaker’s facial features and integrates them with the pretrained AVSEMamba model through two strategies: zero-shot merging and partial visual fine-tuning. For evaluation, we introduce the AVSEC2-Gaze dataset. Experimental results show that GG-AVSE achieves substantial performance gains over gaze-free baselines: a 10.08% improvement in PESQ (2.370 to 2.609), a 5.18% improvement in STOI (0.8802 to 0.9258), and a 23.69% improvement in SI-SDR (9.16 to 11.33). These results confirm that gaze provides an effective cue for resolving target-speaker ambiguity and highlight the scalability of GG-AVSE for real-world applications.

Method: Created Contextual Earnings-22 dataset by augmenting Earnings-22 with realistic custom vocabulary contexts. Established six strong baselines for two dominant approaches: keyword prompting (providing context as prompts) and keyword boosting (adjusting model weights for keywords).

Result: Both keyword prompting and keyword boosting approaches achieve comparable and significantly improved accuracy when scaled from proof-of-concept to large-scale systems, demonstrating the importance of contextual conditioning.

Conclusion: Contextual conditioning with custom vocabulary is crucial for speech-to-text usability, and the new benchmark enables proper evaluation and advancement of contextual speech recognition systems.

Abstract: The accuracy frontier of speech-to-text systems has plateaued on academic benchmarks.1 In contrast, industrial benchmarks and adoption in high-stakes domains suggest otherwise. We hypothesize that the primary difference between the two is contextual conditioning: Academic benchmarks are dominated by frequently encountered general vocabulary that is relatively easy to recognize compared with rare and context-defined custom vocabulary that has disproportionate impact on the usability of speech transcripts. Despite progress on contextual speech-to-text, there is no standardized benchmark. We introduce Contextual Earnings-22, an open dataset built upon Earnings-22, with realistic custom vocabulary contexts to foster research and reveal latent progress. We set six strong baselines for two dominant approaches: keyword prompting and keyword boosting. Experiments show both reach comparable and significantly improved accuracy when scaled from proof-of-concept to large-scale systems.

Method: Proposes a hybrid CNN-Transformer architecture for Arabic SER. Uses convolutional layers to extract discriminative spectral features from Mel-spectrograms, and Transformer encoders to capture long-range temporal dependencies in speech.

Result: Achieved 97.8% accuracy and macro F1-score of 0.98 on the EYASE (Egyptian Arabic speech emotion) corpus, demonstrating state-of-the-art performance.

Conclusion: The hybrid CNN-Transformer approach is effective for Arabic SER, showing the potential of attention-based modeling for low-resource languages and highlighting the value of combining convolutional feature extraction with Transformer architectures.

Abstract: Recognizing emotions from speech using machine learning has become an active research area due to its importance in building human-centered applications. However, while many studies have been conducted in English, German, and other European and Asian languages, research in Arabic remains scarce because of the limited availability of annotated datasets. In this paper, we present an Arabic Speech Emotion Recognition (SER) system based on a hybrid CNN-Transformer architecture. The model leverages convolutional layers to extract discriminative spectral features from Mel-spectrogram inputs and Transformer encoders to capture long-range temporal dependencies in speech. Experiments were conducted on the EYASE (Egyptian Arabic speech emotion) corpus, and the proposed model achieved 97.8% accuracy and a macro F1-score of 0.98. These results demonstrate the effectiveness of combining convolutional feature extraction with attention-based modeling for Arabic SER and highlight the potential of Transformer-based approaches in low-resource languages.

Method: Proposes Byte-Level Distillation (BLD) which operates at the byte level as a common interface across tokenizers. Converts teacher’s output distribution to byte-level probabilities, attaches a lightweight byte-level decoder head to the student, and performs distillation through this shared byte-level interface.

Result: BLD performs competitively with and sometimes surpasses more sophisticated CTD methods across various distillation tasks with models from 1B to 8B parameters, demonstrating that byte level serves as a natural common ground for cross-tokenizer knowledge transfer.

Conclusion: Byte-level distillation provides an effective baseline for cross-tokenizer knowledge transfer, though consistent improvements across all tasks remain elusive, indicating CTD is still an open problem requiring further research.

Abstract: Cross-tokenizer distillation (CTD), the transfer of knowledge from a teacher to a student language model when the two use different tokenizers, remains a largely unsolved problem. Existing approaches rely on heuristic strategies to align mismatched vocabularies, introducing considerable complexity. In this paper, we propose a simple but effective baseline called Byte-Level Distillation (BLD) which enables CTD by operating at a common interface across tokenizers: the byte level. In more detail, we convert the teacher’s output distribution to byte-level probabilities, attach a lightweight byte-level decoder head to the student, and distill through this shared byte-level interface. Despite its simplicity, BLD performs competitively with–and on several benchmarks surpasses–significantly more sophisticated CTD methods, across a range of distillation tasks with models from 1B to 8B parameters. Our results suggest that the byte level is a natural common ground for cross-tokenizer knowledge transfer, while also highlighting that consistent improvements across all tasks and benchmarks remain elusive, underscoring that CTD is still an open problem.

Method: Investigated DSUs derived from self-supervised learning models using quantization methods (including K-means) on Mandarin and Yorùbá tone languages. Analyzed how well SSL latent representations and their quantized DSUs encode lexical tone. Proposed a two-stage K-means clustering approach: first for phonetic information, then on residuals for tone encoding.

Result: SSL latent representations do encode tone, but DSUs from quantization prioritize phonetic structure over suprasegmental features. This limitation persists across various quantization methods. The proposed two-stage clustering shows promise for better tone encoding.

Conclusion: Current DSU quantization strategies have limitations for suprasegmental features, necessitating new tone-aware or prosody-aware techniques in speech representation learning. Two-stage quantization may be a step toward better suprasegmental feature preservation.

Abstract: Discrete speech units (DSUs) are derived by quantising representations from models trained using self-supervised learning (SSL). They are a popular representation for a wide variety of spoken language tasks, including those where prosody matters. DSUs are especially convenient for tasks where text and speech are jointly modelled, such as text-to-speech and multimodal dialogue systems. But we have found that DSUs encode suprasegmental information less reliably than segmental structure, which we demonstrate in this work using lexical tone, though this limitation likely extends to other suprasegmental features such as prosody. Our investigations using the tone languages Mandarin and Yorùbá show that the SSL latent representations themselves do encode tone, yet DSUs obtained using quantisation tend to prioritise phonetic structure, which makes lexical tone less reliably encoded. This remains true for a variety of quantisation methods, not only the most common, K-means. We conclude that current DSU quantisation strategies have limitations for suprasegmental features, which suggests a need for new, tone-aware (or prosody-aware) techniques in speech representation learning. We point towards a potential form of the solution by performing K-means clustering once to encode phonetic information, then again on the residual representation, which better encodes lexical tone.

Method: Multilingual LLM-based system using multilingual embeddings and RAG-based chain-of-tools framework for retrieval, reasoning, translation, and document generation via interactive chat interface.

Result: Scored 75-80% in Bangladesh Bar Council Exams across all stages (MCQs, Written, Viva Voce), matching/surpassing average human performance. Operates at 0.12-0.61% of typical legal consultation costs (99.4-99.9% reduction).

Conclusion: Mina demonstrates potential as low-cost, multilingual AI assistant that automates legal tasks and scales access to justice, offering case study on domain-specific, low-resource systems with multilingual adaptation challenges.

Abstract: Bangladesh’s low-income population faces major barriers to affordable legal advice due to complex legal language, procedural opacity, and high costs. Existing AI legal assistants lack Bengali-language support and jurisdiction-specific adaptation, limiting their effectiveness. To address this, we developed Mina, a multilingual LLM-based legal assistant tailored for the Bangladeshi context. It employs multilingual embeddings and a RAG-based chain-of-tools framework for retrieval, reasoning, translation, and document generation, delivering context-aware legal drafts, citations, and plain-language explanations via an interactive chat interface. Evaluated by law faculty from leading Bangladeshi universities across all stages of the 2022 and 2023 Bangladesh Bar Council Exams, Mina scored 75-80% in Preliminary MCQs, Written, and simulated Viva Voce exams, matching or surpassing average human performance and demonstrating clarity, contextual understanding, and sound legal reasoning. Even under a conservative upper bound, Mina operates at just 0.12-0.61% of typical legal consultation costs in Bangladesh, yielding a 99.4-99.9% cost reduction relative to human-provided services. These results confirm its potential as a low-cost, multilingual AI assistant that automates key legal tasks and scales access to justice, offering a real-world case study on building domain-specific, low-resource systems and addressing challenges of multilingual adaptation, efficiency, and sustainable public-service AI deployment.

Method: DFR-Gemma uses a lightweight projector to align high-dimensional geospatial embeddings with LLM latent space, allowing embeddings to be injected as semantic tokens alongside natural language instructions for direct reasoning.

Result: Experimental results show DFR enables LLMs to decode latent spatial patterns and perform accurate zero-shot reasoning across diverse geospatial tasks while significantly improving efficiency over text-based baselines.

Conclusion: Treating embeddings as primary data inputs provides a more direct, efficient, and scalable approach to multimodal geospatial intelligence compared to text-based methods.

Abstract: Representation learning for geospatial and spatio-temporal data plays a critical role in enabling general-purpose geospatial intelligence. Recent geospatial foundation models, such as the Population Dynamics Foundation Model (PDFM), encode complex population and mobility dynamics into compact embeddings. However, their integration with Large Language Models (LLMs) remains limited. Existing approaches to LLM integration treat these embeddings as retrieval indices or convert them into textual descriptions for reasoning, introducing redundancy, token inefficiency, and numerical inaccuracies. We propose Direct Feature Reasoning-Gemma (DFR-Gemma), a novel framework that enables LLMs to reason directly over dense geospatial embeddings. DFR aligns high-dimensional embeddings with the latent space of an LLM via a lightweight projector, allowing embeddings to be injected as semantic tokens alongside natural language instructions. This design eliminates the need for intermediate textual representations and enables intrinsic reasoning over spatial features. To evaluate this paradigm, we introduce a multi-task geospatial benchmark that pairs embeddings with diverse question-answer tasks, including feature querying, comparison, and semantic description. Experimental results show that DFR allows LLMs to decode latent spatial patterns and perform accurate zero-shot reasoning across tasks, while significantly improving efficiency compared to text-based baselines. Our results demonstrate that treating embeddings as primary data inputs, provides a more direct, efficient, and scalable approach to multimodal geospatial intelligence.

Method: Proposes DLR framework with three-stage training: 1) Decompose queries into textual premises, 2) Extract premise-conditioned continuous visual latents, 3) Deduce answers through grounded rationales. Introduces Spherical Gaussian Latent Policy for effective exploration in latent space.

Result: Extensive experiments on vision-centric benchmarks show DLR consistently outperforms strong baselines including text-only, interleaved multimodal CoT, and latent reasoning methods, while providing superior stepwise interpretability.

Conclusion: DLR effectively addresses visual information loss in complex reasoning tasks by combining textual decomposition with continuous visual latent extraction, enabling more accurate and interpretable vision-language reasoning.

Abstract: Vision-Language Models often struggle with complex visual reasoning due to the visual information loss in textual CoT. Existing methods either add the cost of tool calls or rely on localized patch-based embeddings that are insufficient to extract semantics in multi-step reasoning. We propose \emph{“Decompose, Look, and Reason” (DLR)}, a reinforced latent reasoning framework that dynamically decomposes queries into textual premises, extracts premise-conditioned continuous visual latents, and deduces answers through grounded rationales. We introduce a three-stage training pipeline and propose a novel Spherical Gaussian Latent Policy to enable effective exploration in the latent space. Extensive experiments on vision-centric benchmarks show that DLR consistently outperforms strong baselines, including text-only, interleaved multimodal CoT, and latent reasoning methods, while providing superior stepwise interpretability.

Method: Developed an ePCR-grounded, topic-flow-based multi-agent generation pipeline that iteratively plans, generates, and self-refines dialogues with rule-based factual and topic flow checks, yielding EMSDialog dataset of 4,414 synthetic multi-speaker EMS conversations.

Result: Created EMSDialog dataset with 43 diagnoses, speaker roles, and turn-level topics; human and LLM evaluations confirm high quality and realism; EMSDialog-augmented training improves accuracy, timeliness, and stability of EMS conversational diagnosis prediction.

Conclusion: The proposed pipeline successfully generates realistic multi-speaker medical dialogues, and the resulting dataset effectively improves conversational diagnosis prediction performance in EMS settings.

Abstract: Conversational diagnosis prediction requires models to track evolving evidence in streaming clinical conversations and decide when to commit to a diagnosis. Existing medical dialogue corpora are largely dyadic or lack the multi-party workflow and annotations needed for this setting. We introduce an ePCR-grounded, topic-flow-based multi-agent generation pipeline that iteratively plans, generates, and self-refines dialogues with rule-based factual and topic flow checks. The pipeline yields EMSDialog, a dataset of 4,414 synthetic multi-speaker EMS conversations based on a real-world ePCR dataset, annotated with 43 diagnoses, speaker roles, and turn-level topics. Human and LLM evaluations confirm high quality and realism of EMSDialog using both utterance- and conversation-level metrics. Results show that EMSDialog-augmented training improves accuracy, timeliness, and stability of EMS conversational diagnosis prediction.

Method: Proposes AutoMUP (Automatic Meaning Unit Pyramid) method that extracts meaning units from human summaries, clusters them using embeddings, models inter-participant agreement statistically, and generates graded summaries based on consensus weight.

Result: AutoMUP summaries show high semantic overlap with robust LLM summaries (Flash 2.5 and GPT-5.1), and ablation studies confirm the importance of consensus weight and clustering for summary quality.

Conclusion: AutoMUP provides an effective automated framework for gold-standard summary generation that can be generalized to other Turkic languages with minimal adaptation cost.

Abstract: This study presents a framework for generating the gold-standard summary fully automatically and reproducibly based on multiple human summaries of Turkish educational videos. Within the scope of the study, a new dataset called TR-EduVSum was created, encompassing 82 Turkish course videos in the field of “Data Structures and Algorithms” and containing a total of 3281 independent human summaries. Inspired by existing pyramid-based evaluation approaches, the AutoMUP (Automatic Meaning Unit Pyramid) method is proposed, which extracts consensus-based content from multiple human summaries. AutoMUP clusters the meaning units extracted from human summaries using embedding, statistically models inter-participant agreement, and generates graded summaries based on consensus weight. In this framework, the gold summary corresponds to the highest-consensus AutoMUP configuration, constructed from the most frequently supported meaning units across human summaries. Experimental results show that AutoMUP summaries exhibit high semantic overlap with robust LLM (Large Language Model) summaries such as Flash 2.5 and GPT-5.1. Furthermore, ablation studies clearly demonstrate the decisive role of consensus weight and clustering in determining summary quality. The proposed approach can be generalized to other Turkic languages at low cost.

Method: Three-stage reasoning framework: 1) coherence verification where LLMs assess if cluster summaries match member texts, 2) redundancy adjudication where semantically overlapping clusters are merged/rejected, and 3) label grounding where interpretable labels are assigned unsupervised.

Result: Framework improves cluster coherence and human-aligned labeling quality over classical topic models and representation-based baselines on social media corpora from two platforms; human evaluation shows strong agreement with LLM-generated labels despite no gold annotations.

Conclusion: LLM-based reasoning serves as effective mechanism for validating and refining unsupervised semantic structure, enabling more reliable and interpretable text analysis without supervision, with demonstrated cross-platform stability.

Abstract: Unsupervised methods are widely used to induce latent semantic structure from large text collections, yet their outputs often contain incoherent, redundant, or poorly grounded clusters that are difficult to validate without labeled data. We propose a reasoning-based refinement framework that leverages large language models (LLMs) not as embedding generators, but as semantic judges that validate and restructure the outputs of arbitrary unsupervised clustering algorithms.Our framework introduces three reasoning stages: (i) coherence verification, where LLMs assess whether cluster summaries are supported by their member texts; (ii) redundancy adjudication, where candidate clusters are merged or rejected based on semantic overlap; and (iii) label grounding, where clusters are assigned interpretable labels in a fully unsupervised manner. This design decouples representation learning from structural validation and mitigates common failure modes of embedding-only approaches. We evaluate the framework on real-world social media corpora from two platforms with distinct interaction models, demonstrating consistent improvements in cluster coherence and human-aligned labeling quality over classical topic models and recent representation-based baselines. Human evaluation shows strong agreement with LLM-generated labels, despite the absence of gold-standard annotations. We further conduct robustness analyses under matched temporal and volume conditions to assess cross-platform stability. Beyond empirical gains, our results suggest that LLM-based reasoning can serve as a general mechanism for validating and refining unsupervised semantic structure, enabling more reliable and interpretable analyses of large text collections without supervision.

Method: CAMO uses a hierarchical procedure incorporating vote distributions, confidence calibration, and inter-model uncertainty analysis to dynamically boost underrepresented classes. It was validated on two highly imbalanced domain-specific benchmarks: DIAR-AI/Emotion dataset and ternary BEA 2025 dataset, benchmarking against seven proven ensemble algorithms using eight different language models (three LLMs and five SLMs) under zero-shot and fine-tuned settings.

Result: CAMO consistently achieved the highest strict macro F1-score with refined models, setting a new benchmark. The benefit works in concert with model adaptation, showing that the best ensemble choice depends on model properties.

Conclusion: CAMO proves to be a reliable, domain-neutral framework for unbalanced categorization that effectively addresses class imbalance in ensemble methods.

Abstract: Real-world categorization is severely hampered by class imbalance because traditional ensembles favor majority classes, which lowers minority performance and overall F1-score. We provide a unique ensemble technique for imbalanced problems called CAMO (Class-Aware Minority-Optimized).Through a hierarchical procedure that incorporates vote distributions, confidence calibration, and inter model uncertainty, CAMO dynamically boosts underrepresented classes while preserving and amplifying minority forecasts. We verify CAMO on two highly unbalanced, domain-specific benchmarks: the DIAR-AI/Emotion dataset and the ternary BEA 2025 dataset. We benchmark against seven proven ensemble algorithms using eight different language models (three LLMs and five SLMs) under zero-shot and fine-tuned settings .With refined models, CAMO consistently earns the greatest strict macro F1-score, setting a new benchmark. Its benefit works in concert with model adaptation, showing that the best ensemble choice depends on model properties .This proves that CAMO is a reliable, domain-neutral framework for unbalanced categorization.

Method: ADAG introduces attribution profiles to quantify feature functional roles via input and output gradient effects, a novel clustering algorithm for grouping features, and an LLM explainer-simulator setup that generates and scores natural-language explanations of feature group roles.

Result: The system successfully recovers interpretable circuits on known human-analyzed circuit-tracing tasks and identifies steerable clusters responsible for harmful advice jailbreaks in Llama 3.1 8B Instruct.

Conclusion: ADAG provides an end-to-end automated pipeline for describing attribution graphs in language models, enabling scalable interpretability analysis without manual inspection, with demonstrated effectiveness on both standard tasks and safety-relevant applications.

Abstract: In language model interpretability research, \textbf{circuit tracing} aims to identify which internal features causally contributed to a particular output and how they affected each other, with the goal of explaining the computations underlying some behaviour. However, all prior circuit tracing work has relied on ad-hoc human interpretation of the role that each feature in the circuit plays, via manual inspection of data artifacts such as the dataset examples the component activates on. We introduce \textbf{ADAG}, an end-to-end pipeline for describing these attribution graphs which is fully automated. To achieve this, we introduce \textit{attribution profiles} which quantify the functional role of a feature via its input and output gradient effects. We then introduce a novel clustering algorithm for grouping features, and an LLM explainer–simulator setup which generates and scores natural-language explanations of the functional role of these feature groups. We run our system on known human-analysed circuit-tracing tasks and recover interpretable circuits, and further show ADAG can find steerable clusters which are responsible for a harmful advice jailbreak in Llama 3.1 8B Instruct.

Method: Proposes Dynamic Verification Relaxed Speculative Decoding (DIVERSED) with an ensemble-based verifier that blends draft and target model distributions using task-dependent and context-dependent weights.

Result: DIVERSED achieves substantially higher inference efficiency compared to standard speculative decoding methods while preserving generation quality.

Conclusion: The relaxed verification framework improves time efficiency in large language model inference by accepting more plausible tokens through dynamic blending of distributions.

Abstract: Speculative decoding is an effective technique for accelerating large language model inference by drafting multiple tokens in parallel. In practice, its speedup is often bottlenecked by a rigid verification step that strictly enforces the accepted token distribution to exactly match the target model. This constraint leads to the rejection of many plausible tokens, lowering the acceptance rate and limiting overall time speedup. To overcome this limitation, we propose Dynamic Verification Relaxed Speculative Decoding (DIVERSED), a relaxed verification framework that improves time efficiency while preserving generation quality. DIVERSED learns an ensemble-based verifier that blends the draft and target model distributions with a task-dependent and context-dependent weight. We provide theoretical justification for our approach and demonstrate empirically that DIVERSED achieves substantially higher inference efficiency compared to standard speculative decoding methods. Code is available at: https://github.com/comeusr/diversed.

Method: Introduces Keys to Knowledge (K2K) framework that encodes essential clinical information into model’s parameter space, enabling rapid retrieval from internal key-value memory without inference-time overhead. Uses activation-guided probe construction and cross-attention reranking to enhance retrieval quality.

Result: Achieves state-of-the-art performance across four benchmark healthcare outcome prediction datasets.

Conclusion: K2K provides a practical solution for deploying reliable LLMs in time-sensitive healthcare settings by eliminating external retrieval latency while maintaining accuracy.

Abstract: Large language models (LLMs) hold significant promise for healthcare, yet their reliability in high-stakes clinical settings is often compromised by hallucinations and a lack of granular medical context. While Retrieval Augmented Generation (RAG) can mitigate these issues, standard supervised pipelines require computationally intensive searches over massive external knowledge bases, leading to high latency that is impractical for time-sensitive care. To address this, we introduce Keys to Knowledge (K2K), a novel framework that replaces external retrieval with internal, key-based knowledge access. By encoding essential clinical information directly into the model’s parameter space, K2K enables rapid retrieval from internal key-value memory without inference-time overhead. We further enhance retrieval quality through activation-guided probe construction and cross-attention reranking. Experimental results demonstrate that K2K achieves state-of-the-art performance across four benchmark healthcare outcome prediction datasets.

Method: Used clinical notes from PLWH (2012-2022), identified candidate sentences via expert-curated keywords and clinical word embeddings, manually annotated 1,332 sentences across four stigma subscales, and compared encoder-based (GatorTron-large, BERT) and generative LLMs (GPT-OSS-20B, LLaMA-8B, MedGemma-27B) under zero-shot and few-shot prompting.

Result: GatorTron-large achieved best overall performance (Micro F1 = 0.62). Few-shot prompting improved generative models significantly (5-shot GPT-OSS-20B: 0.57, LLaMA-8B: 0.59). Performance varied by subscale, with Negative Self-Image most predictable and Personalized Stigma most challenging. Zero-shot generative models had high failure rates (up to 32%).

Conclusion: Developed the first practical NLP tool for identifying HIV stigma in clinical notes, demonstrating LLMs’ potential for extracting psychosocial determinants from medical text, though performance varies by stigma type and model architecture.

Abstract: Human immunodeficiency virus (HIV)-related stigma is a critical psychosocial determinant of health for people living with HIV (PLWH), influencing mental health, engagement in care, and treatment outcomes. Although stigma-related experiences are documented in clinical narratives, there is a lack of off-the-shelf tools to extract and categorize them. This study aims to develop a large language model (LLM)-based tool for identifying HIV stigma from clinical notes. We identified clinical notes from PLWH receiving care at the University of Florida (UF) Health between 2012 and 2022. Candidate sentences were identified using expert-curated stigma-related keywords and iteratively expanded via clinical word embeddings. A total of 1,332 sentences were manually annotated across four stigma subscales: Concern with Public Attitudes, Disclosure Concerns, Negative Self-Image, and Personalized Stigma. We compared GatorTron-large and BERT as encoder-based baselines, and GPT-OSS-20B, LLaMA-8B, and MedGemma-27B as generative LLMs, under zero-shot and few-shot prompting. GatorTron-large achieved the best overall performance (Micro F1 = 0.62). Few-shot prompting substantially improved generative model performance, with 5-shot GPT-OSS-20B and LLaMA-8B achieving Micro-F1 scores of 0.57 and 0.59, respectively. Performance varied by stigma subscale, with Negative Self-Image showing the highest predictability and Personalized Stigma remaining the most challenging. Zero-shot generative inference exhibited non-trivial failure rates (up to 32%). This study develops the first practical NLP tool for identifying HIV stigma in clinical notes.

Method: Proposes Separate Sequence (SepSeq) - a training-free, plug-and-play framework that strategically inserts separator tokens to recalibrate attention, acting as attention sinks to focus on local segments while preserving global context.

Result: Extensive evaluations on 9 LLMs show average relative accuracy improvement of 35.6% across diverse domains and average reduction of 16.4% in total inference token consumption.

Conclusion: SepSeq effectively mitigates attention dispersion in LLMs for long numerical sequences through separator tokens, improving performance without requiring retraining.

Abstract: While transformer-based Large Language Models (LLMs) theoretically support massive context windows, they suffer from severe performance degradation when processing long numerical sequences. We attribute this failure to the attention dispersion in the Softmax mechanism, which prevents the model from concentrating attention. To overcome this, we propose Separate Sequence (SepSeq), a training-free, plug-and-play framework to mitigate dispersion by strategically inserting separator tokens. Mechanistically, we demonstrate that separator tokens act as an attention sink, recalibrating attention to focus on local segments while preserving global context. Extensive evaluations on 9 widely-adopted LLMs confirm the effectiveness of our approach: SepSeq yields an average relative accuracy improvement of 35.6% across diverse domains while reducing total inference token consumption by 16.4% on average.

Method: Developed PPT-Bench benchmark organized around Philosophical Pressure Taxonomy (PPT) with four pressure types: Epistemic Destabilization, Value Nullification, Authority Inversion, and Identity Dissolution. Each item tested at three layers: baseline prompt (L0), single-turn pressure (L1), and multi-turn Socratic escalation (L2). Evaluated five models and tested mitigation strategies including prompt-level anchoring, persona-stability prompts, and Leading Query Contrastive Decoding.

Result: Pressure types produce statistically separable inconsistency patterns across models, revealing weaknesses not captured by standard social-pressure benchmarks. Mitigation effectiveness is strongly type- and model-dependent: prompt-level anchoring and persona-stability prompts work best in API settings, while Leading Query Contrastive Decoding is most reliable for open models.

Conclusion: Epistemic attack exposes distinct vulnerabilities in LLMs beyond traditional sycophancy. The proposed benchmark and taxonomy provide tools for systematically evaluating how models handle challenges to their epistemic foundations, with mitigation strategies showing varying effectiveness across model types.

Abstract: Large language models (LLMs) can shift their answers under pressure in ways that reflect accommodation rather than reasoning. Prior work on sycophancy has focused mainly on disagreement, flattery, and preference alignment, leaving a broader set of epistemic failures less explored. We introduce \textbf{PPT-Bench}, a diagnostic benchmark for evaluating \textit{epistemic attack}, where prompts challenge the legitimacy of knowledge, values, or identity rather than simply opposing a previous answer. PPT-Bench is organized around the Philosophical Pressure Taxonomy (PPT), which defines four types of philosophical pressure: Epistemic Destabilization, Value Nullification, Authority Inversion, and Identity Dissolution. Each item is tested at three layers: a baseline prompt (L0), a single-turn pressure condition (L1), and a multi-turn Socratic escalation (L2). This allows us to measure epistemic inconsistency between L0 and L1, and conversational capitulation in L2. Across five models, these pressure types produce statistically separable inconsistency patterns, suggesting that epistemic attack exposes weaknesses not captured by standard social-pressure benchmarks. Mitigation results are strongly type- and model-dependent: prompt-level anchoring and persona-stability prompts perform best in API settings, while Leading Query Contrastive Decoding is the most reliable intervention for open models.

Method: The researchers analyze the potential of static analysis tools for detecting LLM code hallucinations. They evaluate these tools on NL-to-code benchmarks requiring library use, measure detection rates for various error types, and conduct manual analysis to identify cases that static methods cannot plausibly catch.

Result: Static analysis tools can detect 16-70% of all errors and 14-85% of library hallucinations, with performance varying by LLM and dataset. Manual analysis reveals fundamental limitations, giving an upper bound detection potential of only 48.5-77%.

Conclusion: Static analysis provides a cheap method for addressing some forms of hallucination in LLM-generated code, but has inherent limitations that prevent it from fully solving the problem, with maximum detection potential capped at 48.5-77%.

Abstract: Despite extensive research, Large Language Models continue to hallucinate when generating code, particularly when using libraries. On NL-to-code benchmarks that require library use, we find that LLMs generate code that uses non-existent library features in 8.1-40% of responses.One intuitive approach for detection and mitigation of hallucinations is static analysis. In this paper, we analyse the potential of static analysis tools, both in terms of what they can solve and what they cannot. We find that static analysis tools can detect 16-70% of all errors, and 14-85% of library hallucinations, with performance varying by LLM and dataset. Through manual analysis, we identify cases a static method could not plausibly catch, which gives an upper bound on their potential from 48.5% to 77%. Overall, we show that static analysis methods are cheap method for addressing some forms of hallucination, and we quantify how far short of solving the problem they will always be.

Method: Introduces LSLORA (restricts LoRA adaptation to sensitivity-identified layers) and GARFA (GQA-Aware RoPE Frequency Adaptation with learnable per-KV-head scalars), tested on Llama 3.1 8B with cross-layer ablation studies.

Result: Found strong anti-localization: task-sensitive layers concentrate in late network (23-31) while RoPE-influential layers dominate early network (0-9). Despite this, applying both interventions to sensitivity-identified layers outperforms all alternatives by 4-16 percentage points across six benchmarks.

Conclusion: The co-localization hypothesis is false - there’s strong anti-localization in GQA transformers, but targeted adaptation to sensitivity-identified layers still yields optimal performance, approaching Claude 3.5 Haiku on HumanEval+ at low compute cost.

Abstract: We investigate a fundamental structural question in Grouped Query Attention (GQA) transformers: do the layers most sensitive to task correctness coincide with the layers where positional encoding adaptation has the greatest leverage? We term this the co-localization hypothesis and test it on Llama 3.1 8B, a 32-layer GQA model with a 4:1 query-to-key-value head ratio. We introduce \LSLORA, which restricts LoRA adaptation to layers identified via a novel correctness-differential hidden-state metric, and GARFA (GQA-Aware RoPE Frequency Adaptation), which attaches 8 learnable per-KV-head scalar multipliers to each targeted layer. Contrary to the co-localization hypothesis, we discover strong anti-localization: task-sensitive layers concentrate in the late network ($\ell\in{23\text{-}31}$) while RoPE-influential layers dominate the early network ($\ell\in{0\text{-}9}$), yielding Spearman $r_s = -0.735$ ($p = 1.66\times10^{-6}$). Despite this anti-localization, a 4-way cross-layer ablation shows that applying both interventions to the sensitivity-identified layers outperforms all alternative configurations by 4-16 percentage points across six diverse benchmarks (MMLU, GPQA, HumanEval+, MATH, MGSM, ARC), approaching Claude 3.5 Haiku on HumanEval+ (67.1% vs. 68.3%) at $100 total compute cost.

Method: Developed controlled emotion translation framework to rewrite problems into emotional variants while preserving quantities/relationships. Created Temper-5400 dataset (5,400 emotion-neutral pairs) across GSM8K, MultiArith, and ARC-Challenge. Evaluated 18 models from 1B to frontier scale.

Result: 1) Emotional framing reduces accuracy by 2-10 percentage points despite preserved numerical content. 2) Neutralizing emotional variants recovers most lost performance, showing degradation is tied to emotional style rather than content corruption. Non-emotional paraphrases cause no degradation.

Conclusion: Emotional language negatively impacts LLM reasoning even when numerical content is unchanged. Neutralization can serve as lightweight inference-time mitigation. Benchmark construction provides framework for controlled stylistic translation and robustness evaluation.

Abstract: Large language models are trained and evaluated on quantitative reasoning tasks written in clean, emotionally neutral language. However, real-world queries are often wrapped in frustration, urgency or enthusiasm. Does emotional framing alone degrade reasoning when all numerical content is preserved? To investigate this, a controlled emotion translation framework is developed that rewrites problems into emotional variants while preserving all quantities and relationships. Using this framework, Temper-5400 (5,400 semantically verified emotion–neutral pairs) is constructed across GSM8K, MultiArith, and ARC-Challenge, and evaluated on eighteen models (1B to frontier scale). Two core results emerge: First, emotional framing reduces accuracy by 2-10 percentage points even though all numerical content is preserved. Second, neutralizing emotional variants recovers most of the lost performance, showing both that the degradation is tied to emotional style rather than content corruption and that neutralization can serve as a lightweight inference-time mitigation. Non-emotional paraphrases cause no such degradation, implicating emotional content rather than surface-level changes. Beyond emotion specifically, the benchmark construction procedure provides a general framework for controlled stylistic translation and robustness evaluation.

Method: GRASS uses mean gradient norms as a task-aware and training-stage-aware metric to estimate layer importance, adaptively adjusts layer sampling probabilities, and includes a layer-wise optimizer state offloading mechanism that overlaps computation and communication.

Result: Extensive experiments show GRASS consistently outperforms state-of-the-art methods with average accuracy improvements up to 4.38 points and memory reduction up to 19.97%.

Conclusion: GRASS provides an effective memory-efficient fine-tuning framework that adapts to task-specific and training-stage variations in layer importance, achieving better performance than existing methods while reducing memory requirements.

Abstract: Full-parameter fine-tuning of large language models is constrained by substantial GPU memory requirements. Low-rank adaptation methods mitigate this challenge by updating only a subset of parameters. However, these approaches often limit model expressiveness and yield lower performance than full-parameter fine-tuning. Layer-wise fine-tuning methods have emerged as an alternative, enabling memory-efficient training through static layer importance sampling strategies. However, these methods overlook variations in layer importance across tasks and training stages, resulting in suboptimal performance on downstream tasks. To address these limitations, we propose GRASS, a gradient-based adaptive layer-wise importance sampling framework. GRASS utilizes mean gradient norms as a task-aware and training-stage-aware metric for estimating layer importance. Furthermore, GRASS adaptively adjusts layer sampling probabilities through an adaptive training strategy. We also introduce a layer-wise optimizer state offloading mechanism that overlaps computation and communication to further reduce memory usage while maintaining comparable training throughput. Extensive experiments across multiple models and benchmarks demonstrate that GRASS consistently outperforms state-of-the-art methods, achieving an average accuracy improvement of up to 4.38 points and reducing memory usage by up to 19.97%.

Method: Proposes AsyncTLS with two key components: 1) Hierarchical sparse attention combining coarse-grained block filtering with fine-grained token selection, and 2) Asynchronous offloading engine that overlaps KV cache transfers with computation via temporal locality exploitation.

Result: Evaluated on Qwen3 and GLM-4.7-Flash across GQA and MLA architectures, achieves accuracy comparable to full attention with 1.2x-10.0x operator speedups and 1.3x-4.7x end-to-end throughput improvements on 48k-96k contexts.

Conclusion: AsyncTLS effectively balances accuracy and efficiency for long-context LLM inference through hierarchical sparse attention and asynchronous offloading techniques.

Abstract: Long-context inference in LLMs faces the dual challenges of quadratic attention complexity and prohibitive KV cache memory. While token-level sparse attention offers superior accuracy, its indexing overhead is costly; block-level methods improve efficiency but sacrifice precision. We propose AsyncTLS, a hierarchical sparse attention system that combines coarse-grained block filtering with fine-grained token selection to balance accuracy and efficiency, coupled with an asynchronous offloading engine that overlaps KV cache transfers with computation via temporal locality exploitation. Evaluated on Qwen3 and GLM-4.7-Flash across GQA, and MLA architectures, AsyncTLS achieves accuracy comparable to full attention while delivering 1.2x - 10.0x operator speedups and 1.3x - 4.7x end-to-end throughput improvements on 48k - 96k contexts.

Method: Construct VGToolBench benchmark for vague instructions, then propose Tool Retrieval Bridge (TRB) - a bridge model that rewrites vague instructions into more specific ones to align with retriever preferences.

Result: TRB consistently improves all baseline retrievers, with BM25 achieving up to 111.51% relative improvement (NDCG from 9.73 to 19.59). Extensive experiments show effectiveness across multiple retrieval settings.

Conclusion: The TRB approach effectively mitigates ambiguity in vague instructions and delivers substantial improvements for tool retrieval in real-world scenarios.

Abstract: Tool learning has emerged as a promising paradigm for large language models (LLMs) to address real-world challenges. Due to the extensive and irregularly updated number of tools, tool retrieval for selecting the desired tool subset is essential. However, current tool retrieval methods are usually based on academic benchmarks containing overly detailed instructions (e.g., specific API names and parameters), while real-world instructions are more vague. Such a discrepancy would hinder the tool retrieval in real-world applications. In this paper, we first construct a new benchmark, VGToolBench, to simulate human vague instructions. Based on this, we conduct a series of preliminary analyses and find that vague instructions indeed damage the performance of tool retrieval. To this end, we propose a simple-yet-effective Tool Retrieval Bridge (TRB) approach to boost the performance of tool retrieval for vague instructions. The principle of TRB is to introduce a bridge model to rewrite the vague instructions into more specific ones and alleviate the gap between vague instructions and retriever preferences.We conduct extensive experiments under multiple commonly used retrieval settings, and the results show that TRB effectively mitigates the ambiguity of vague instructions while delivering consistent and substantial improvements across all baseline retrievers. For example, with the help of TRB, BM25 achieves a relative improvement of up to 111.51%, i.e., increasing the average NDCG score from 9.73 to 19.59. The source code and models are publicly available at https://github.com/kfchenhn/TRB.

Method: Proposes recurrent-depth transformers that enable iterative computation over the same transformer layers, studied through controlled experiments with models trained from scratch on systematic generalization and depth extrapolation tasks.

Result: Vanilla transformers struggle with both generalization challenges, but recurrent-depth transformers can effectively generalize. Systematic generalization emerges through three-stage grokking process, and depth extrapolation is unlocked by scaling inference-time recurrence.

Conclusion: Recurrent-depth transformers address key limitations in compositional reasoning, though they face overthinking limitations where excessive recurrence degrades predictions for very deep compositions.

Abstract: We study implicit reasoning, i.e. the ability to combine knowledge or rules within a single forward pass. While transformer-based large language models store substantial factual knowledge and rules, they often fail to compose this knowledge for implicit multi-hop reasoning, suggesting a lack of compositional generalization over their parametric knowledge. To address this limitation, we study recurrent-depth transformers, which enables iterative computation over the same transformer layers. We investigate two compositional generalization challenges under the implicit reasoning scenario: systematic generalization, i.e. combining knowledge that is never used for compositions during training, and depth extrapolation, i.e. generalizing from limited reasoning depth (e.g. training on up to 5-hop) to deeper compositions (e.g. 10-hop). Through controlled studies with models trained from scratch, we show that while vanilla transformers struggle with both generalization challenges, recurrent-depth transformers can effectively make such generalization. For systematic generalization, we find that this ability emerges through a three-stage grokking process, transitioning from memorization to in-distribution generalization and finally to systematic generalization, supported by mechanistic analysis. For depth extrapolation, we show that generalization beyond training depth can be unlocked by scaling inference-time recurrence, with more iterations enabling deeper reasoning. We further study how training strategies affect extrapolation, providing guidance on training recurrent-depth transformers, and identify a key limitation, overthinking, where excessive recurrence degrades predictions and limits generalization to very deep compositions.

Method: Uses LLMs (GPT-4o, GPT-5-nano, GPT-5) with expert-developed loneliness evaluation framework and cause typology to build and analyze Reddit corpus. Includes human-validated data processing pipeline for quality assurance.

Result: Loneliness evaluation achieved 76.09% accuracy for caregivers and 79.78% for non-caregivers. Cause categorization achieved F1 scores of 0.825 and 0.80 respectively. Found substantial differences in loneliness causes between populations.

Conclusion: Established effective LLM-based pipeline for creating high-quality social media datasets for loneliness research, demonstrating population-level differences in loneliness manifestation between caregivers and non-caregivers.

Abstract: This paper presents an LLM-driven approach for constructing diverse social media datasets to measure and compare loneliness in the caregiver and non-caregiver populations. We introduce an expert-developed loneliness evaluation framework and an expert-informed typology for categorizing causes of loneliness for analyzing social media text. Using a human-validated data processing pipeline, we apply GPT-4o, GPT-5-nano, and GPT-5 to build a high-quality Reddit corpus and analyze loneliness across both populations. The loneliness evaluation framework achieved average accuracies of 76.09% and 79.78% for caregivers and non-caregivers, respectively. The cause categorization framework achieved micro-aggregate F1 scores of 0.825 and 0.80 for caregivers and non-caregivers, respectively. Across populations, we observe substantial differences in the distribution of types of causes of loneliness. Caregivers’ loneliness were predominantly linked to caregiving roles, identity recognition, and feelings of abandonment, indicating distinct loneliness experiences between the two groups. Demographic extraction further demonstrates the viability of Reddit for building a diverse caregiver loneliness dataset. Overall, this work establishes an LLM-based pipeline for creating high quality social media datasets for studying loneliness and demonstrates its effectiveness in analyzing population-level differences in the manifestation of loneliness.

Method: Two approaches: MemReader-0.6B (compact passive extractor distilled for accurate structured outputs) and MemReader-4B (active extractor using Group Relative Policy Optimization to evaluate information value, reference ambiguity, and completeness before making memory writing decisions).

Result: MemReader consistently outperforms existing extraction-based baselines on LOCOMO, LongMemEval, and HaluMem benchmarks, achieving state-of-the-art performance on knowledge updating, temporal reasoning, and hallucination reduction tasks.

Conclusion: Effective agent memory requires reasoning-driven selective memory extraction rather than just extracting more information, enabling low-noise dynamically evolving long-term memory; models are integrated into MemOS and available via public API.

Abstract: Long-term memory is fundamental for personalized and autonomous agents, yet populating it remains a bottleneck. Existing systems treat memory extraction as a one-shot, passive transcription from context to structured entries, which struggles with noisy dialogue, missing references, and cross-turn dependencies, leading to memory pollution, low-value writes, and inconsistency. In this paper, we introduce the MemReader family for active long-term memory extraction in agent systems: MemReader-0.6B, a compact and cost-efficient passive extractor distilled for accurate and schema-consistent structured outputs, and MemReader-4B, an active extractor optimized with Group Relative Policy Optimization (GRPO) to make memory writing decisions. Under a ReAct-style paradigm, MemReader-4B explicitly evaluates information value, reference ambiguity, and completeness before acting, and can selectively write memories, defer incomplete inputs, retrieve historical context, or discard irrelevant chatter. Experiments on LOCOMO, LongMemEval, and HaluMem show that MemReader consistently outperforms existing extraction-based baselines. In particular, MemReader-4B achieves state-of-the-art performance on tasks involving knowledge updating, temporal reasoning, and hallucination reduction. These results suggest that effective agent memory requires not merely extracting more information, but performing reasoning-driven and selective memory extraction to build low-noise and dynamically evolving long-term memory. Furthermore, MemReader has been integrated into MemOS and is being deployed in real-world applications. To support future research and adoption, we release the models and provide public API access.

Method: Systematic design of plausible/implausible event triples with abstract/concrete constituent categories, followed by analysis of human and LLM-generated judgments and contextual examples.

Result: Humans excel at nuanced detection and contextualization of (non-)literal vs. implausible events, while LLMs show only shallow contextualization patterns with bias toward trading implausibility for non-literal, plausible interpretations.

Conclusion: There are substantial differences between human and LLM assessments of plausibility, revealing limitations in LLMs’ understanding of literalness vs. plausibility relationships in language.

Abstract: This work explores the connection between (non-)literalness and plausibility at the example of subject-verb-object events in English. We design a systematic setup of plausible and implausible event triples in combination with abstract and concrete constituent categories. Our analysis of human and LLM-generated judgments and example contexts reveals substantial differences between assessments of plausibility. While humans excel at nuanced detection and contextualization of (non-)literal vs. implausible events, LLM results reveal only shallow contextualization patterns with a bias to trade implausibility for non-literal, plausible interpretations.

Method: Train linear transformations specific to each hierarchical depth and semantic domain using Linear Relational Concepts, then analyze representational geometry across multiple domains with in-domain generalization and cross-domain transfer tests.

Result: Hierarchical relations can be linearly recovered from model representations within domains, encoded in low-dimensional domain-specific subspaces, with highly similar hierarchy representation across these subspaces.

Conclusion: All tested models encode concept hierarchies as highly interpretable linear representations, with hierarchy representation being domain-specific yet structurally similar across domains.

Abstract: We investigate how and to what extent hierarchical relations (e.g., Japan $\subset$ Eastern Asia $\subset$ Asia) are encoded in the internal representations of language models. Building on Linear Relational Concepts, we train linear transformations specific to each hierarchical depth and semantic domain, and characterize representational differences associated with hierarchical relations by comparing these transformations. Going beyond prior work on the representational geometry of hierarchies in LMs, our analysis covers multi-token entities and cross-layer representations. Across multiple domains we learn such transformations and evaluate in-domain generalization to unseen data and cross-domain transfer. Experiments show that, within a domain, hierarchical relations can be linearly recovered from model representations. We then analyze how hierarchical information is encoded in representation space. We find that it is encoded in a relatively low-dimensional subspace and that this subspace tends to be domain-specific. Our main result is that hierarchy representation is highly similar across these domain-specific subspaces. Overall, we find that all models considered in our experiments encode concept hierarchies in the form of highly interpretable linear representations.

Method: MDS combines global coverage (bin-wise selection in user-query trajectory space) with local structural evaluation (entity-grounded topic grounding, information progress, and query-answer form consistency).

Result: MDS outperforms single-turn selectors, dialogue-level LLM scorers, and heuristic baselines on three multi-turn benchmarks and a Banking test set, achieving best overall rank across metrics and robustness on long conversations.

Conclusion: Dialogue-level selection with combined global coverage and local structural evaluation effectively improves multi-turn dialogue data quality for instruction tuning.

Abstract: Instruction-tuned language models increasingly rely on large multi-turn dialogue corpora, but these datasets are often noisy and structurally inconsistent, with topic drift, repetitive chitchat, and mismatched answer formats across turns. We address this from a data selection perspective and propose \textbf{MDS} (Multi-turn Dialogue Selection), a dialogue-level framework that scores whole conversations rather than isolated turns. MDS combines a global coverage stage that performs bin-wise selection in the user-query trajectory space to retain representative yet non-redundant dialogues, with a local structural stage that evaluates within-dialogue reliability through entity-grounded topic grounding and information progress, together with query-answer form consistency for functional alignment. MDS outperforms strong single-turn selectors, dialogue-level LLM scorers, and heuristic baselines on three multi-turn benchmarks and an in-domain Banking test set, achieving the best overall rank across reference-free and reference-based metrics, and is more robust on long conversations under the same training budget. Code and resources are included in the supplementary materials.

Method: Two-pronged approach: 1) Dynamic memory evolution for improved memory writing, 2) Self-learning with context distillation objective to internalize user experiences for better memory reading.

Result: Extensive evaluations on long-horizon benchmarks using Qwen-3 model family (4B to 32B) show TSUBASA surpasses competitive memory-augmented systems like Mem0 and Memory-R1, achieving Pareto improvements with robust personalization using reduced token budget.

Conclusion: TSUBASA effectively breaks the quality-efficiency barrier for personalized LLMs in long-horizon tasks through innovative memory evolution and self-learning mechanisms.

Abstract: Personalized large language models (PLLMs) have garnered significant attention for their ability to align outputs with individual’s needs and preferences. However, they still struggle with long-horizon tasks, such as tracking a user’s extensive history of conversations or activities. Existing memory mechanisms often fail to capture evolving behaviors, and RAG paradigms are trapped by a quality-efficiency tradeoff. Meanwhile, parametric adaptation is bottlenecked by train-inference gap due to the scarcity of labeled data. To enhance the long-horizon capabilities of PLLMs, we introduce TSUBASA, a two-pronged approach designed to improve memory writing via dynamic memory evolution, and memory reading via self-learning with a context distillation objective to internalize user experiences. Extensive evaluations on long-horizon benchmarks using the Qwen-3 model family (4B to 32B) validate the effectiveness of TSUBASA, surpassing competitive memory-augmented systems that rely primarily on memory writing, such as Mem0 and Memory-R1. Our analyses further confirms that TSUBASA breaks the quality-efficiency barrier to achieve Pareto improvements, delivering robust, high-fidelity personalization with a reduced token budget.

Method: Proposes HCRE with two components: 1) LLM for relation prediction, and 2) hierarchical relation tree derived from predefined relations. Uses hierarchical classification where target relation is inferred level by level. Introduces prediction-then-verification inference strategy with multi-view verification at each level to mitigate error propagation.

Result: Extensive experiments show HCRE outperforms existing baselines, validating its effectiveness in cross-document relation extraction.

Conclusion: The hierarchical classification approach with verification strategy effectively addresses LLM limitations in handling numerous relations, demonstrating improved performance over existing methods.

Abstract: Cross-document relation extraction (RE) aims to identify relations between the head and tail entities located in different documents. Existing approaches typically adopt the paradigm of ``\textit{Small Language Model (SLM) + Classifier}’’. However, the limited language understanding ability of SLMs hinders further improvement of their performance. In this paper, we conduct a preliminary study to explore the performance of Large Language Models (LLMs) in cross-document RE. Despite their extensive parameters, our findings indicate that LLMs do not consistently surpass existing SLMs. Further analysis suggests that the underperformance is largely attributed to the challenges posed by the numerous predefined relations. To overcome this issue, we propose an LLM-based \underline{H}ierarchical \underline{C}lassification model for cross-document \underline{RE} (HCRE), which consists of two core components: 1) an LLM for relation prediction and 2) a \textit{hierarchical relation tree} derived from the predefined relation set. This tree enables the LLM to perform hierarchical classification, where the target relation is inferred level by level. Since the number of child nodes is much smaller than the size of the entire predefined relation set, the hierarchical relation tree significantly reduces the number of relation options that LLM needs to consider during inference. However, hierarchical classification introduces the risk of error propagation across levels. To mitigate this, we propose a \textit{prediction-then-verification} inference strategy that improves prediction reliability through multi-view verification at each level. Extensive experiments show that HCRE outperforms existing baselines, validating its effectiveness.

Method: Proposes a framework organizing post-training methods by trajectory provenance (off-policy vs on-policy learning) and two recurring roles: effective support expansion (making useful behaviors reachable) and policy reshaping (improving behavior within reachable regions), plus behavioral consolidation for preserving/transferring behavior.

Result: Provides a unified interpretation of major paradigms: SFT can serve both support expansion and reshaping, preference methods are usually off-policy reshaping, on-policy RL improves behavior on learner-generated states, and distillation is consolidation rather than just compression.

Conclusion: Progress in LLM post-training increasingly depends on coordinated system design rather than any single dominant objective, and the framework helps diagnose bottlenecks and reason about stage composition in multi-stage pipelines.

Abstract: Post-training has become central to turning pretrained large language models (LLMs) into aligned and deployable systems. Recent progress spans supervised fine-tuning (SFT), preference optimization, reinforcement learning (RL), process supervision, verifier-guided methods, distillation, and multi-stage pipelines. Yet these methods are often discussed in fragmented ways, organized by labels or objective families rather than by the behavioral bottlenecks they address. This survey argues that LLM post-training is best understood as structured intervention on model behavior. We organize the field first by trajectory provenance, which defines two primary learning regimes: off-policy learning on externally supplied trajectories, and on-policy learning on learner-generated rollouts. We then interpret methods through two recurring roles – effective support expansion, which makes useful behaviors more reachable, and policy reshaping, which improves behavior within already reachable regions – together with a complementary systems-level role, behavioral consolidation, which preserves, transfers, and amortizes behavior across stages and model transitions. This perspective yields a unified reading of major paradigms. SFT may serve either support expansion or policy reshaping, whereas preference-based methods are usually off-policy reshaping. On-policy RL often improves behavior on learner-generated states, though under stronger guidance it can also make hard-to-reach reasoning paths reachable. Distillation is often best understood as consolidation rather than only compression, and hybrid pipelines emerge as coordinated multi-stage compositions. Overall, the framework helps diagnose post-training bottlenecks and reason about stage composition, suggesting that progress in LLM post-training increasingly depends on coordinated system design rather than any single dominant objective.

Method: Establishes formal connections between gradient dynamics and domain distributions, then introduces DoGraph - a reweighting framework that formulates data scheduling as a graph-constrained optimization problem.

Result: Extensive experiments on GPT-2 models of varying scales show DoGraph consistently achieves competitive performance.

Conclusion: Provides theoretical foundation for understanding domain effects in LLM training and practical framework for optimizing data mixing strategies.

Abstract: Data mixing strategy is essential for large language model (LLM) training. Empirical evidence shows that inappropriate strategies can significantly reduce generalization. Although recent methods have improved empirical performance, several fundamental questions remain open: what constitutes a domain, whether human and model perceptions of domains are aligned, and how domain weighting influences generalization. We address these questions by establishing formal connections between gradient dynamics and domain distributions, offering a theoretical framework that clarifies the role of domains in training dynamics. Building on this analysis, we introduce DoGraph, a reweighting framework that formulates data scheduling as a graph-constrained optimization problem. Extensive experiments on GPT-2 models of varying scales demonstrate that DoGraph consistently achieves competitive performance.

Method: Decomposes claims into subject-relation-object-modifier (SROM) atoms and scores adversarial rewrites with Atomic Validity Scoring (AVS) to detect factual corruption beyond surface similarity.

Result: Experiments on FEVER dataset show AtomEval provides more reliable evaluation signals; analysis reveals stronger LLMs don’t necessarily produce more effective adversarial claims under validity-aware evaluation.

Conclusion: AtomEval addresses limitations in current adversarial evaluation practices and reveals previously overlooked issues with LLM-based adversarial generators.

Abstract: Adversarial claim rewriting is widely used to test fact-checking systems, but standard metrics fail to capture truth-conditional consistency and often label semantically corrupted rewrites as successful. We introduce AtomEval, a validity-aware evaluation framework that decomposes claims into subject-relation-object-modifier (SROM) atoms and scores adversarial rewrites with Atomic Validity Scoring (AVS), enabling detection of factual corruption beyond surface similarity. Experiments on the FEVER dataset across representative attack strategies and LLM generators show that AtomEval provides more reliable evaluation signals in our experiments. Using AtomEval, we further analyze LLM-based adversarial generators and observe that stronger models do not necessarily produce more effective adversarial claims under validity-aware evaluation, highlighting previously overlooked limitations in current adversarial evaluation practices.

Method: Uses frequency-domain processing with three novel components: RecurrentOscillatorBanks for O(L) sequence processing, FFT-Rotate Wavetable Encoder for efficient byte mapping, and PhaseHarmonics sinusoidal non-linearity with minimal parameters.

Result: Achieves 88.6% on IMDB, 92.3% on AG News, and 83.3% on SST-2, outperforming tokenized counterparts with 16x more parameters on some benchmarks while using only 733K parameters total.

Conclusion: Frequency-domain components systematically outperform complex cognitive architectures for text classification, enabling efficient byte-level processing with minimal parameters while maintaining competitive accuracy.

Abstract: We present Kathleen, a text classification architecture that operates directly on raw UTF-8 bytes using frequency-domain processing – requiring no tokenizer, no attention mechanism, and only 733K parameters. Kathleen introduces three novel components: (1) RecurrentOscillatorBanks – damped sinusoid convolutions with temporal memory for O(L) sequence processing; (2) an FFT-Rotate Wavetable Encoder that maps all 256 byte values using a single learnable vector (256 floats), replacing conventional embedding tables (65K parameters) while improving accuracy; (3) PhaseHarmonics – a sinusoidal non-linearity with just 6 learnable phase parameters that our ablation identifies as the single most impactful component (+2.6% accuracy, <0.001% of model parameters). Through comprehensive ablation of a 1.8M-parameter predecessor, we show that frequency-domain components systematically outperform complex cognitive architectures: removing a 560K-parameter bio-inspired framework costs only -0.2%, while removing the 6-parameter PhaseHarmonics costs -2.6%. The resulting Kathleen-Clean achieves 88.6% on IMDB, 92.3% on AG News, and 83.3% on SST-2 – outperforming a tokenized counterpart with 16x more parameters on IMDB (+1.6%) and AG News (+2.1%). Kathleen processes sequences in O(L) time and memory, enabling byte-level operation at sequence lengths where O(L^2) Transformers exhaust GPU memory.

Method: 1. Decompose long-context reasoning into fundamental atomic skills. 2. Automatically synthesize pseudo datasets targeting each specific atomic skill. 3. Use reinforcement learning on these pseudo datasets to sharpen the model’s atomic skills to boost general long-context reasoning ability.

Result: Outperforms strong baseline by average margin of 7.7% (improving from 46.3% to 54.0%) across multiple benchmarks including Loogle, Loong, LongBench-v2, BrowscompLong, Ruler-qa2, and MRCR.

Conclusion: Decomposing long-context reasoning into atomic skills and training on targeted pseudo-datasets via reinforcement learning effectively improves general long-context reasoning performance in LLMs.

Abstract: Long-context reasoning is essential for complex real-world applications, yet remains a significant challenge for Large Language Models (LLMs). Despite the rapid evolution in long-context reasoning, current research often overlooks the internal complexity of the long-context reasoning task itself. In this paper, we move beyond this holistic view and decompose long-context reasoning into a set of fundamental atomic skills, and we then automatically synthesize a suite of pseudo datasets, each explicitly targeting a specific atomic skill. Our empirical analysis confirms that proficiency in these atomic skills is strongly correlated with general long-text reasoning performance. Building on this insight, we employ reinforcement learning on these pseudo datasets to sharpen the model’s atomic skills, in the hope of boosting its general long-context reasoning ability. Extensive experiments across multiple benchmarks demonstrate the effectiveness of our approach: it outperforms a strong baseline by an average margin of 7.7% (improving from 46.3% to 54.0%) across Loogle, Loong, LongBench-v2, BrowscompLong, Ruler-qa2, and MRCR.

Method: The study evaluates various predictors: pre-retrieval and post-retrieval predictors adapted from ad hoc retrieval, and post-generation predictors including a novel supervised approach that explicitly models semantic relationships between the question, retrieved passages, and generated answer.

Result: The novel supervised predictor that models semantic relationships among question, passages, and answer achieves the best prediction quality for determining RAG’s effectiveness in QA tasks.

Conclusion: Explicit modeling of semantic relationships between question, retrieved content, and generated answer provides the most effective approach for predicting when RAG will improve QA performance.

Abstract: We address the task of predicting the gain of using RAG (retrieval augmented generation) for question answering with respect to not using it. We study the performance of a few pre-retrieval and post-retrieval predictors originally devised for ad hoc retrieval. We also study a few post-generation predictors, one of which is novel to this study and posts the best prediction quality. Our results show that the most effective prediction approach is a novel supervised predictor that explicitly models the semantic relationships among the question, retrieved passages, and the generated answer.

Method: 1) Generate Inner-Answer using only parametric knowledge; 2) Generate Refer-Answer using Contrastive DPO objective that treats Inner-Answer as negative constraint and retrieved docs as positive ground truth; 3) Joint decoding mechanism that dynamically fuses logical coherence of Inner-Answer with factual precision of Refer-Answer at token level.

Result: Experiments on five QA benchmarks show GuarantRAG improves accuracy by up to 12.1% and reduces hallucinations by 16.3% compared to standard and dynamic RAG baselines.

Conclusion: Explicit decoupling of reasoning from evidence integration with Contrastive DPO and joint decoding effectively addresses the integration bottleneck in RAG systems, leading to significant improvements in accuracy and hallucination reduction.

Abstract: Retrieval-Augmented Generation (RAG) significantly enhances Large Language Models (LLMs) by providing access to external knowledge. However, current research primarily focuses on retrieval quality, often overlooking the critical ‘‘integration bottleneck’’: even when relevant documents are retrieved, LLMs frequently fail to utilize them effectively due to conflicts with their internal parametric knowledge. In this paper, we argue that implicitly resolving this conflict in a single generation pass is suboptimal. We introduce GuarantRAG, a framework that explicitly decouples reasoning from evidence integration. First, we generate an ‘‘Inner-Answer’’ based solely on parametric knowledge to capture the model’s reasoning flow. Second, to guarantee faithful evidence extraction, we generate a ‘‘Refer-Answer’’ using a novel Contrastive DPO objective. This objective treats the parametric Inner-Answer as a negative constraint and the retrieved documents as positive ground truth, forcing the model to suppress internal hallucinations in favor of external evidence during this phase. Finally, rather than naive concatenation or using the DPO trained model directly, we propose a joint decoding mechanism that dynamically fuses the logical coherence of the Inner-Answer with the factual precision of the Refer-Answer at the token level. Experiments on five QA benchmarks demonstrate that GuarantRAG improves accuracy by up to 12.1% and reduces hallucinations by 16.3% compared to standard and dynamic RAG baselines.

Method: Proposes an efficient linguistic steganographic method using classic entropy coding (range coding) with a rotation mechanism to achieve provable security while maintaining high embedding capacity.

Result: Achieves around 100% entropy utilization for embedding capacity, outperforming baseline methods, with high embedding speeds up to 1554.66 bits/s on GPT-2 across various language models.

Conclusion: The proposed range coding with rotation mechanism provides an efficient and provably secure solution for linguistic steganography, balancing security and practical embedding capacity.

Abstract: Linguistic steganography involves embedding secret messages within seemingly innocuous texts to enable covert communication. Provable security, which is a long-standing goal and key motivation, has been extended to language-model-based steganography. Previous provably secure approaches have achieved perfect imperceptibility, measured by zero Kullback-Leibler (KL) divergence, but at the expense of embedding capacity. In this paper, we attempt to directly use a classic entropy coding method (range coding) to achieve secure steganography, and then propose an efficient and provably secure linguistic steganographic method with a rotation mechanism. Experiments across various language models show that our method achieves around 100% entropy utilization (embedding efficiency) for embedding capacity, outperforming the existing baseline methods. Moreover, it achieves high embedding speeds (up to 1554.66 bits/s on GPT-2). The code is available at github.com/ryehr/RRC_steganography.

Method: Proposes dual-pool token-budget routing that partitions homogeneous fleets into specialized pools: high-throughput short-context and high-capacity long-context pools. Routes requests based on estimated total token budget using per-category bytes-to-token ratio learned online via exponential moving average from usage.prompt_tokens feedback (no tokenizer needed). Includes analytical model for predicting fleet-level cost savings.

Result: Reduces GPU-hours by 31-42% (corresponding to $2.86M annual savings at fleet scale), lowers preemption rates by 5.4×, improves P99 TTFT by 6%. Case study with Qwen3-235B-A22B projects $15.4M annual savings. O(1) dispatch overhead, adapts to heterogeneous workloads, composes with existing optimizations.

Conclusion: Dual-pool token-budget routing effectively addresses configuration-traffic mismatch in LLM inference fleets, providing substantial cost savings and reliability improvements while maintaining low overhead and compatibility with existing optimization techniques.

Abstract: Production vLLM fleets typically provision each instance for the worst-case context length, leading to substantial KV-cache over-allocation and under-utilized concurrency. In practice, 80-95% of requests are short, yet are served under configurations optimized for long contexts, wasting 4-8$\times$ throughput capacity and triggering reliability issues such as OOM crashes, preemption, and request rejections. We identify a common root cause for these inefficiencies: configuration-traffic mismatch. We propose dual-pool token-budget routing, a lightweight dispatch mechanism that partitions a homogeneous fleet into two specialized pools: a high-throughput short-context pool and a high-capacity long-context pool. Each request is routed based on its estimated total token budget, computed using a per-category bytes-to-token ratio that is learned online via exponential moving average from usage.prompt_tokens feedback, eliminating the need for a tokenizer. We also develop a simple analytical model that predicts fleet-level cost savings from workload characteristics and measured throughput differences, enabling practitioners to estimate benefits prior to deployment. Evaluations on real-world traces from the Azure LLM Inference Dataset and LMSYS-Chat-1M, serving Llama-3-70B on A100 GPUs, show that our approach reduces GPU-hours by 31-42%, corresponding to $2.86M annual savings at fleet scale, while lowering preemption rates by 5.4$\times$ and improving P99 TTFT by 6%. A case study with Qwen3-235B-A22B on AMD MI300X at 10,000 req/s projects $15.4M in annual savings. The method incurs only O(1) dispatch overhead, adapts automatically to heterogeneous workloads, and composes seamlessly with existing optimizations such as PagedAttention, continuous batching, and prefill-decode disaggregation.

Method: Transform speech signals into spectrograms (STFT, Mel-spectrograms, MFCC), then convert them into information waves using a QV block. Train QV-based Convolutional Neural Networks (QV-CNN) and QV-based Vision Transformers (QV-ViT) on these quantum-inspired representations for deepfake speech classification on ASVSpoof dataset.

Result: QV-CNN and QV-ViT consistently outperform standard CNN and ViT models, achieving higher classification accuracy and improved robustness. QV-CNN with MFCC features achieves 94.20% accuracy and 9.04% EER, while QV-CNN with Mel-spectrograms attains 94.57% accuracy on ASVspoof dataset.

Conclusion: Quantum Vision theory is an effective approach for audio deepfake detection, demonstrating that quantum-inspired information wave representations can enhance audio classification performance and opening new directions for quantum-inspired learning in audio perception tasks.

Abstract: We propose Quantum Vision (QV) theory as a new perspective for deep learning-based audio classification, applied to deepfake speech detection. Inspired by particle-wave duality in quantum physics, QV theory is based on the idea that data can be represented not only in its observable, collapsed form, but also as information waves. In conventional deep learning, models are trained directly on these collapsed representations, such as images. In QV theory, inputs are first transformed into information waves using a QV block, and then fed into deep learning models for classification. QV-based models improve performance in image classification compared to their non-QV counterparts. What if QV theory is applied speech spectrograms for audio classification tasks? This is the motivation and novelty of the proposed approach. In this work, Short-Time Fourier Transform (STFT), Mel-spectrograms, and Mel-Frequency Cepstral Coefficients (MFCC) of speech signals are converted into information waves using the proposed QV block and used to train QV-based Convolutional Neural Networks (QV-CNN) and QV-based Vision Transformers (QV-ViT). Extensive experiments are conducted on the ASVSpoof dataset for deepfake speech classification. The results show that QV-CNN and QV-ViT consistently outperform standard CNN and ViT models, achieving higher classification accuracy and improved robustness in distinguishing genuine and spoofed speech. Moreover, the QV-CNN model using MFCC features achieves the best overall performance on the ASVspoof dataset, with an accuracy of 94.20% and an EER of 9.04%, while the QV-CNN with Mel-spectrograms attains the highest accuracy of 94.57%. These findings demonstrate that QV theory is an effective and promising approach for audio deepfake detection and opens new directions for quantum-inspired learning in audio perception tasks.

Method: Proposes OA-EM (Output-Aware Expectation-Maximization) initialization using Hessian-weighted Mahalanobis distance. Analyzes representational ratio ρ = N/KM to characterize weight groups vs codebook capacity relationship.

Result: OA-EM consistently produces better solutions after PV-tuning across compression rates, search budgets, and architectures (Llama 3.2 3B, Llama 3.1 8B, Qwen 2.5 3B). Dominates quality-compute frontier, especially at extreme 2-bit compression where poor initialization degrades perplexity by orders of magnitude.

Conclusion: Optimization geometry in compressed model spaces is crucial - initialization can dominate subsequent search and fine-tuning. The bottleneck severity scales with representational ratio ρ, being moderate at 3 bpp but extreme at 2 bpp.

Abstract: Additive quantization enables extreme LLM compression with O(1) lookup-table dequantization, making it attractive for edge deployment. Yet at 2-bit precision, it often fails catastrophically, even with extensive search and finetuning. We show that the dominant bottleneck is codebook initialisation. Greedy sequential initialisation frequently places the model in poor optimisation regions that subsequent beam search and PV-tuning struggle to overcome. We analyse this behaviour through the representational ratio \r{ho} = N/KM, which characterises the relationship between weight groups and codebook capacity, and propose OA-EM, an output-aware EM initialisation method using Hessian-weighted Mahalanobis distance. Across compression rates, search budgets, and three architectures (Llama 3.2 3B, Llama 3.1 8B, Qwen 2.5 3B), OA-EM consistently produces better solutions after PV-tuning and dominates the quality-compute frontier. The severity of the bottleneck scales with \r{ho}: moderate at 3 bpp but extreme at 2 bpp, where poor initialisation can degrade perplexity by orders of magnitude. More broadly, our results highlight the importance of optimisation geometry in compressed model spaces, where initialisation can dominate subsequent search and fine-tuning.

Method: Developed controlled pipeline to generate synthetic doctor-patient interview transcripts guided by scenario-specific evaluation criteria, combining ideal and perturbed performances to simulate varying skill levels. Used LLM-assisted framework for automatic silver-labeling with adjustable evaluation strictness.

Result: Benchmarking shows mid-size models (≤32B parameters) achieve accuracies comparable to GPT-4o (~90%) on synthetic data, demonstrating feasibility of locally deployable, privacy-preserving evaluation systems.

Conclusion: LLMs can effectively generate and evaluate French OSCE dialogues in low-resource contexts, enabling scalable, privacy-preserving automated assessment systems for medical education without heavy reliance on human examiners.

Abstract: Objective Structured Clinical Examinations (OSCEs) are the standard method for assessing medical students’ clinical and communication skills through structured patient interviews. In France, however, the organization of training sessions is limited by human and logistical constraints, restricting students’ access to repeated practice and structured feedback. Recent advances in Natural Language Processing (NLP) and Large Language Models (LLMs) now offer the opportunity to automatically evaluate such medical interviews, thereby alleviating the need for human examiners during training. Yet, real French OSCE annotated transcripts remain extremely scarce, limiting reproducible research and reliable benchmarking. To address these challenges, we investigate the use of LLMs for both generating and evaluating French OSCE dialogues in a low-resource context. We introduce a controlled pipeline that produces synthetic doctor-patient interview transcripts guided by scenario-specific evaluation criteria, combining ideal and perturbed performances to simulate varying student skill levels. The resulting dialogues are automatically silver-labeled through an LLM-assisted framework supporting adjustable evaluation strictness. Benchmarking multiple open-source and proprietary LLMs shows that mid-size models ($\le$32B parameters) achieve accuracies comparable to GPT-4o ($\sim$90%) on synthetic data, highlighting the feasibility of locally deployable, privacy-preserving evaluation systems for medical education.

Method: Benchmarked GNN architectures (GCN, GraphSAGE, GAT, ChebNet) against Logistic Regression, SVM, and MLPs across 7 public datasets in English, Indonesian, and Polish using identical TF-IDF features to isolate relational structure impact.

Result: GNNs consistently outperformed non-graph baselines across all datasets, with GraphSAGE achieving 96.8% F1 on Kaggle and 91.9% on WELFake vs. 73.2% and 66.8% for MLP, while maintaining comparable or lower inference times.

Conclusion: Classic GNNs remain effective and efficient for misinformation detection, challenging the need for increasingly complex architectures.

Abstract: The rapid spread of online misinformation has led to increasingly complex detection models, including large language models and hybrid architectures. However, their computational cost and deployment limitations raise concerns about practical applicability. In this work, we benchmark graph neural networks (GNNs) against non-graph-based machine learning methods under controlled and comparable conditions. We evaluate lightweight GNN architectures (GCN, GraphSAGE, GAT, ChebNet) against Logistic Regression, Support Vector Machines, and Multilayer Perceptrons across seven public datasets in English, Indonesian, and Polish. All models use identical TF-IDF features to isolate the impact of relational structure. Performance is measured using F1 score, with inference time reported to assess efficiency. GNNs consistently outperform non-graph baselines across all datasets. For example, GraphSAGE achieves 96.8% F1 on Kaggle and 91.9% on WELFake, compared to 73.2% and 66.8% for MLP, respectively. On COVID-19, GraphSAGE reaches 90.5% F1 vs. 74.9%, while ChebNet attains 79.1% vs. 66.4% on FakeNewsNet. These gains are achieved with comparable or lower inference times. Overall, the results show that classic GNNs remain effective and efficient, challenging the need for increasingly complex architectures in misinformation detection.

Method: Combines OpenAI semantic embeddings with six compact heuristic features capturing stylistic and informational cues. Uses PCA for dimensionality reduction and evaluates with XGBoost, GraphSAGE, and GCN classifiers.

Result: Simplified feature design yields slightly lower F1-scores but graph-based models achieve competitive performance with substantially reduced inference time. High ROC-AUC values indicate strong discrimination capability.

Conclusion: The hybrid approach provides reliable clickbait detection with good efficiency-performance trade-off, supporting detection under varying decision thresholds.

Abstract: We propose a lightweight hybrid approach to clickbait detection that combines OpenAI semantic embeddings with six compact heuristic features capturing stylistic and informational cues. To improve efficiency, embeddings are reduced using PCA and evaluated with XGBoost, GraphSAGE, and GCN classifiers. While the simplified feature design yields slightly lower F1-scores, graph-based models achieve competitive performance with substantially reduced inference time. High ROC–AUC values further indicate strong discrimination capability, supporting reliable detection of clickbait headlines under varying decision thresholds.

Method: Used RhymeTagger to identify rhymes based on repeating patterns in poetry corpora across seven languages (Czech, German, English, French, Italian, Russian, Slovene). Evaluated performance based on training size and language differences. Established benchmark using inter-annotator agreement on manually annotated poems, analyzing phonetic similarity and word distance factors. Compared RhymeTagger to three LLMs using one-shot learning.

Result: RhymeTagger consistently outperforms human agreement when provided with sufficient training data. LLMs lacking phonetic representation significantly struggle with rhyme recognition tasks. Training data requirements vary by language, with sufficient data enabling reliable performance across all tested languages.

Conclusion: Unsupervised rhyme recognition with RhymeTagger is effective across multiple languages with adequate training data, surpassing human agreement and outperforming LLMs that lack phonetic understanding. The approach provides a robust solution for multilingual rhyme analysis.

Abstract: Rhyme is deceptively intuitive: what is or is not a rhyme is constructed historically, scholars struggle with rhyme classification, and people disagree on whether two words are rhymed or not. This complicates automated rhymed recognition and evaluation, especially in multilingual context. This article investigates how much training data is needed for reliable unsupervised rhyme recognition using RhymeTagger, a language-independent tool that identifies rhymes based on repeating patterns in poetry corpora. We evaluate its performance across seven languages (Czech, German, English, French, Italian, Russian, and Slovene), examining how training size and language differences affect accuracy. To set a realistic performance benchmark, we assess inter-annotator agreement on a manually annotated subset of poems and analyze factors contributing to disagreement in expert annotations: phonetic similarity between rhyming words and their distance from each other in a poem. We also compare RhymeTagger to three large language models using a one-shot learning strategy. Our findings show that, once provided with sufficient training data, RhymeTagger consistently outperforms human agreement, while LLMs lacking phonetic representation significantly struggle with the task.

Method: Reformulates debiasing as resource redistribution problem, treating output probability mass as limited resource. Uses fine-grained trajectory-level objective with dynamic debiasing constraints to selectively revise biased reasoning suffixes while preserving valid prefixes. Integrates online self-improvement with consistency filtering to autonomously synthesize supervision signals.

Result: Achieves superior debiasing performance with only 20k annotated samples, activating efficient self-correction while preserving general reasoning capabilities without continuous external oversight.

Conclusion: Self-Debias framework successfully addresses bias propagation in LLM reasoning by enabling intrinsic self-correction capabilities through strategic resource redistribution and autonomous supervision synthesis.

Abstract: Although Large Language Models (LLMs) demonstrate remarkable reasoning capabilities, inherent social biases often cascade throughout the Chain-of-Thought (CoT) process, leading to continuous “Bias Propagation”. Existing debiasing methods primarily focus on static constraints or external interventions, failing to identify and interrupt this propagation once triggered. To address this limitation, we introduce Self-Debias, a progressive framework designed to instill intrinsic self-correction capabilities. Specifically, we reformulate the debiasing process as a strategic resource redistribution problem, treating the model’s output probability mass as a limited resource to be reallocated from biased heuristics to unbiased reasoning paths. Unlike standard preference optimization which applies broad penalties, Self-Debias employs a fine-grained trajectory-level objective subject to dynamic debiasing constraints. This enables the model to selectively revise biased reasoning suffixes while preserving valid contextual prefixes. Furthermore, we integrate an online self-improvement mechanism utilizing consistency filtering to autonomously synthesize supervision signals. With merely 20k annotated samples, Self-Debias activates efficient self-correction, achieving superior debiasing performance while preserving general reasoning capabilities without continuous external oversight.

Method: Proposes a three-level hierarchical memory structure (topics, episodes, facts) with hyperedges grouping related episodes and facts. Uses hybrid lexical-semantic index and coarse-to-fine retrieval strategy for efficient high-order association retrieval.

Result: Achieves state-of-the-art performance on LoCoMo benchmark with 92.73% LLM-as-a-judge accuracy, demonstrating effectiveness for long-term conversations.

Conclusion: HyperMem effectively models high-order associations in conversational memory, improving coherence and performance in long-term dialogues compared to pairwise relation approaches.

Abstract: Long-term memory is essential for conversational agents to maintain coherence, track persistent tasks, and provide personalized interactions across extended dialogues. However, existing approaches as Retrieval-Augmented Generation (RAG) and graph-based memory mostly rely on pairwise relations, which can hardly capture high-order associations, i.e., joint dependencies among multiple elements, causing fragmented retrieval. To this end, we propose HyperMem, a hypergraph-based hierarchical memory architecture that explicitly models such associations using hyperedges. Particularly, HyperMem structures memory into three levels: topics, episodes, and facts, and groups related episodes and their facts via hyperedges, unifying scattered content into coherent units. Leveraging this structure, we design a hybrid lexical-semantic index and a coarse-to-fine retrieval strategy, supporting accurate and efficient retrieval of high-order associations. Experiments on the LoCoMo benchmark show that HyperMem achieves state-of-the-art performance with 92.73% LLM-as-a-judge accuracy, demonstrating the effectiveness of HyperMem for long-term conversations.

Method: BAIM uses a reasoning language model to decompose items into four problem-solving stages (understand, plan, carry out, look back), extracts stage-level representations from embedding trajectories, and adaptively routes these representations based on learner context within a KT backbone.

Result: BAIM outperforms strong pretraining-based baselines on XES3G5M and NIPS34 datasets, with particularly large gains under repeated learner interactions.

Conclusion: Incorporating procedural problem-solving dynamics through behavior-aware item modeling significantly improves knowledge tracing performance, especially for capturing learner heterogeneity in repeated interactions.

Abstract: Knowledge Tracing (KT) aims to predict learners’ future performance from past interactions. While recent KT approaches have improved via learning item representations aligned with Knowledge Components, they overlook the procedural dynamics of problem solving. We propose Behavior-Aware Item Modeling (BAIM), a framework that enriches item representations by integrating dynamic procedural solution information. BAIM leverages a reasoning language model to decompose each item’s solution into four problem-solving stages (i.e., understand, plan, carry out, and look back), pedagogically grounded in Polya’s framework. Specifically, it derives stage-level representations from per-stage embedding trajectories, capturing latent signals beyond surface features. To reflect learner heterogeneity, BAIM adaptively routes these stage-wise representations, introducing a context-conditioned mechanism within a KT backbone, allowing different procedural stages to be emphasized for different learners. Experiments on XES3G5M and NIPS34 show that BAIM consistently outperforms strong pretraining-based baselines, achieving particularly large gains under repeated learner interactions.

Method: Analyzed 297 English verb-object pairs (e.g., “float idea” vs. “suggest idea”) in approximately 2 million corpus sentences using five NLP tools to extract 2,293 cognitive and linguistic features capturing affective, lexical, syntactic, and discourse-level properties.

Result: Literal contexts have higher lexical frequency, cohesion, and structural regularity, while metaphorical contexts show greater affective load, imageability, lexical diversity, and constructional specificity. Within-pair analyses reveal substantial heterogeneity with most pairs showing non-uniform effects.

Conclusion: No single consistent distributional pattern distinguishes metaphorical from literal usage; differences are largely construction-specific. Large-scale data with diverse features provides fine-grained understanding of metaphor-literal contrasts in verb-object usage.

Abstract: Metaphor pervades everyday language, allowing speakers to express abstract concepts via concrete domains. While prior work has studied metaphors cognitively and psycholinguistically, large-scale comparisons with literal language remain limited, especially for near-synonymous expressions. We analyze 297 English verb-object pairs (e.g., float idea vs. suggest idea) in ~2M corpus sentences, examining their contextual usage. Using five NLP tools, we extract 2,293 cognitive and linguistic features capturing affective, lexical, syntactic, and discourse-level properties. We address: (i) whether features differ between metaphorical and literal contexts (cross-pair analysis), and (ii) whether individual VO pairs diverge internally (within-pair analysis). Cross-pair results show literal contexts have higher lexical frequency, cohesion, and structural regularity, while metaphorical contexts show greater affective load, imageability, lexical diversity, and constructional specificity. Within-pair analyses reveal substantial heterogeneity, with most pairs showing non-uniform effects. These results suggest no single, consistent distributional pattern that distinguishes metaphorical from literal usage. Instead, differences are largely construction-specific. Overall, large-scale data combined with diverse features provides a fine-grained understanding of metaphor-literal contrasts in VO usage.

Method: Proposes Adaptive Tool Trust Calibration (ATTC), which guides models to adaptively choose whether to trust or ignore tool results based on confidence scores of generated code blocks. The framework helps models make better decisions about tool reliability.

Result: ATTC reduces the “Tool Ignored” issue across various open-source TIR models of different sizes and multiple datasets, resulting in performance improvements of 4.1% to 7.5%.

Conclusion: The ATTC framework effectively addresses the tool trust problem in reasoning models, enabling better integration of external tools and improving overall reasoning performance.

Abstract: Large reasoning models (LRMs) have achieved strong performance enhancement through scaling test time computation, but due to the inherent limitations of the underlying language models, they still have shortcomings in tasks that require precise computation and extensive knowledge reserves. Tool-Integrated Reasoning (TIR) has emerged as a promising paradigm that incorporates tool call and execution within the reasoning trajectory. Although recent works have released some powerful open-source TIR models, our analysis reveals that these models still suffer from critical deficiencies. We find that when the reasoning of the model conflicts with the tool results, the model tends to believe in its own reasoning. And there are cases where the tool results are correct but are ignored by the model, resulting in incorrect answers, which we define as “Tool Ignored’’. This indicates that the model does not know when to trust or ignore the tool. To overcome these limitations, We introduce Adaptive Tool Trust Calibration (ATTC), a novel framework that guides the model to adaptively choose to trust or ignore the tool results based on the confidence score of generated code blocks. The experimental results from various open-source TIR models of different sizes and across multiple datasets demonstrate that ATTC effectively reduces the “Tool Ignored” issue, resulting in a performance increase of 4.1% to 7.5%.

Method: Extended RuleEdit benchmark to 200 manually verified rules; used causal tracing to analyze where different forms of rule knowledge are stored in transformer layers; proposed Distributed Multi-Layer Editing (DMLE) that applies shared early-layer updates for formulas/descriptions and separate middle-layer updates for instances.

Result: DMLE achieves substantially stronger rule-level editing performance, improving instance portability by 13.91 percentage points and rule understanding by 50.19 percentage points over strongest baselines across multiple models (GPT-J-6B, Qwen2.5-7B, Qwen2-7B, LLaMA-3-8B).

Conclusion: Rule knowledge is not uniformly localized in LLMs, requiring distributed multi-layer editing approaches like DMLE that target different forms of rule knowledge in their respective layer locations for effective rule-level knowledge editing.

Abstract: Large language models store not only isolated facts but also rules that support reasoning across symbolic expressions, natural language explanations, and concrete instances. Yet most model editing methods are built for fact-level knowledge, assuming that a target edit can be achieved through a localized intervention. This assumption does not hold for rule-level knowledge, where a single rule must remain consistent across multiple interdependent forms. We investigate this problem through a mechanistic study of rule-level knowledge editing. To support this study, we extend the RuleEdit benchmark from 80 to 200 manually verified rules spanning mathematics and physics. Fine-grained causal tracing reveals a form-specific organization of rule knowledge in transformer layers: formulas and descriptions are concentrated in earlier layers, while instances are more associated with middle layers. These results suggest that rule knowledge is not uniformly localized, and therefore cannot be reliably edited by a single-layer or contiguous-block intervention. Based on this insight, we propose Distributed Multi-Layer Editing (DMLE), which applies a shared early-layer update to formulas and descriptions and a separate middle-layer update to instances. While remaining competitive on standard editing metrics, DMLE achieves substantially stronger rule-level editing performance. On average, it improves instance portability and rule understanding by 13.91 and 50.19 percentage points, respectively, over the strongest baseline across GPT-J-6B, Qwen2.5-7B, Qwen2-7B, and LLaMA-3-8B. The code is available at https://github.com/Pepper66/DMLE.

Method: SeLaR introduces an entropy-gated mechanism that activates soft embeddings only at low-confidence steps while preserving discrete decoding at high-confidence steps. It also uses entropy-aware contrastive regularization that pushes soft embeddings away from the dominant token’s direction to encourage sustained exploration of multiple latent reasoning paths.

Result: Experiments on five reasoning benchmarks demonstrate that SeLaR consistently outperforms standard Chain-of-Thought and state-of-the-art training-free methods.

Conclusion: SeLaR provides a lightweight, training-free framework that improves reasoning capabilities in LLMs by selectively applying latent reasoning only where needed and preventing premature collapse of soft embeddings, leading to more stable and exploratory reasoning.

Abstract: Chain-of-Thought (CoT) has become a cornerstone of reasoning in large language models, yet its effectiveness is constrained by the limited expressiveness of discrete token sampling. Recent latent reasoning approaches attempt to alleviate this limitation by replacing discrete tokens with soft embeddings (probability-weighted mixtures of token embeddings) or hidden states, but they commonly suffer from two issues: (1) global activation injects perturbations into high-confidence steps, impairing reasoning stability; and (2) soft embeddings quickly collapse toward the highest-probability token, limiting exploration of alternative trajectories. To address these challenges, we propose SeLaR (Selective Latent Reasoning), a lightweight and training-free framework. SeLaR introduces an entropy-gated mechanism that activates soft embeddings only at low-confidence steps, while preserving discrete decoding at high-confidence steps. Additionally, we propose an entropy-aware contrastive regularization that pushes soft embeddings away from the dominant (highest-probability) token’s direction, encouraging sustained exploration of multiple latent reasoning paths. Experiments on five reasoning benchmarks demonstrate that SeLaR consistently outperforms standard CoT and state-of-the-art training-free methods.

Method: Introduces OmniBehavior benchmark constructed entirely from real-world data, integrating long-horizon, cross-scenario, and heterogeneous behavioral patterns. Evaluates state-of-the-art LLMs on this benchmark and systematically compares simulated vs. authentic behaviors.

Result: LLMs struggle to accurately simulate complex real-world behaviors, with performance plateauing despite expanding context windows. Models exhibit structural bias toward a “positive average person” with hyper-activity, persona homogenization, and Utopian bias, losing individual differences and long-tail behaviors.

Conclusion: Current LLMs have fundamental limitations in simulating authentic human behavior, revealing critical directions for future high-fidelity simulation research to address structural biases and better capture individual differences and complex behavioral patterns.

Abstract: The emergence of Large Language Models (LLMs) has illuminated the potential for a general-purpose user simulator. However, existing benchmarks remain constrained to isolated scenarios, narrow action spaces, or synthetic data, failing to capture the holistic nature of authentic human behavior. To bridge this gap, we introduce OmniBehavior, the first user simulation benchmark constructed entirely from real-world data, integrating long-horizon, cross-scenario, and heterogeneous behavioral patterns into a unified framework. Based on this benchmark, we first provide empirical evidence that previous datasets with isolated scenarios suffer from tunnel vision, whereas real-world decision-making relies on long-term, cross-scenario causal chains. Extensive evaluations of state-of-the-art LLMs reveal that current models struggle to accurately simulate these complex behaviors, with performance plateauing even as context windows expand. Crucially, a systematic comparison between simulated and authentic behaviors uncovers a fundamental structural bias: LLMs tend to converge toward a positive average person, exhibiting hyper-activity, persona homogenization, and a Utopian bias. This results in the loss of individual differences and long-tail behaviors, highlighting critical directions for future high-fidelity simulation research.

Method: Proposes a GAN and GPT-3.5 based data augmentation framework to create SinaSarc dataset from Sina Weibo, then extends BERT architecture to incorporate multi-dimensional information including user historical behavior for dynamic linguistic pattern modeling.

Result: Achieves highest F1-scores of 0.9138 (non-sarcastic) and 0.9151 (sarcastic), outperforming all existing state-of-the-art approaches.

Conclusion: Presents a novel framework for dynamically modeling users’ long-term linguistic patterns in Chinese sarcasm detection, contributing to both dataset construction and methodological advancement in the field.

Abstract: Sarcasm is a rhetorical device that expresses criticism or emphasizes characteristics of certain individuals or situations through exaggeration, irony, or comparison. Existing methods for Chinese sarcasm detection are constrained by limited datasets and high construction costs, and they mainly focus on textual features, overlooking user-specific linguistic patterns that shape how opinions and emotions are expressed. This paper proposes a Generative Adversarial Network (GAN) and Large Language Model (LLM)-driven data augmentation framework to dynamically model users’ linguistic patterns for enhanced Chinese sarcasm detection. First, we collect raw data from various topics on Sina Weibo. Then, we train a GAN on these data and apply a GPT-3.5 based data augmentation technique to synthesize an extended sarcastic comment dataset, named SinaSarc. This dataset contains target comments, contextual information, and user historical behavior. Finally, we extend the BERT architecture to incorporate multi-dimensional information, particularly user historical behavior, enabling the model to capture dynamic linguistic patterns and uncover implicit sarcastic cues in comments. Experimental results demonstrate the effectiveness of our proposed method. Specifically, our model achieves the highest F1-scores on both the non-sarcastic and sarcastic categories, with values of 0.9138 and 0.9151 respectively, which outperforms all existing state-of-the-art (SOTA) approaches. This study presents a novel framework for dynamically modeling users’ long-term linguistic patterns in Chinese sarcasm detection, contributing to both dataset construction and methodological advancement in this field.

Method: Develops Dataset Policy Gradient (DPG), an RL primitive that uses exact data attribution via higher-order gradients as policy gradient rewards to optimize synthetic data generators for SFT.

Result: DPG successfully causes target models to embed QR codes, specific patterns, reduce weight norms, rephrase in new languages, and produce specific UUIDs through SFT on generated examples.

Conclusion: DPG is a powerful and flexible technique for shaping model properties using only synthetic training examples, demonstrating precise control over language models.

Abstract: What are the limits of controlling language models via synthetic training data? We develop a reinforcement learning (RL) primitive, the Dataset Policy Gradient (DPG), which can precisely optimize synthetic data generators to produce a dataset of targeted examples. When used for supervised fine-tuning (SFT) of a target model, these examples cause the target model to do well on a differentiable metric of our choice. Our approach achieves this by taking exact data attribution via higher-order gradients and using those scores as policy gradient rewards. We prove that this procedure closely approximates the true, intractable gradient for the synthetic data generator. To illustrate the potential of DPG, we show that, using only SFT on generated examples, we can cause the target model’s LM head weights to (1) embed a QR code, (2) embed the pattern $\texttt{67}$, and (3) have lower $\ell^2$ norm. We additionally show that we can cause the generator to (4) rephrase inputs in a new language and (5) produce a specific UUID, even though neither of these objectives is conveyed in the generator’s input prompts. These findings suggest that DPG is a powerful and flexible technique for shaping model properties using only synthetic training examples.

Method: Dual methodology: 750 hours of scripted speech from compiled text corpora, translations, and domain-specific sentences across 11 domains; 2,250 hours of spontaneous speech elicited through textual and image prompts. Data collected via customized mobile app from 4,777 native speakers with multi-layer quality assurance (automated SNR validation + human review).

Result: Created a 3,000-hour multilingual speech dataset covering Dholuo, Kikuyu, Kalenjin, Maasai, and Somali with diverse regional/demographic representation, overcoming infrastructure, device compatibility, and community trust challenges through local partnerships and adaptive protocols.

Conclusion: AfriVoices-KE provides foundational resource for developing inclusive speech technologies and advancing digital preservation of Kenya’s linguistic heritage, demonstrating successful data collection methodology for low-resource language settings.

Abstract: AfriVoices-KE is a large-scale multilingual speech dataset comprising approximately 3,000 hours of audio across five Kenyan languages: Dholuo, Kikuyu, Kalenjin, Maasai, and Somali. The dataset includes 750 hours of scripted speech and 2,250 hours of spontaneous speech, collected from 4,777 native speakers across diverse regions and demographics. This work addresses the critical underrepresentation of African languages in speech technology by providing a high-quality, linguistically diverse resource. Data collection followed a dual methodology: scripted recordings drew from compiled text corpora, translations, and domain-specific generated sentences spanning eleven domains relevant to the Kenyan context, while unscripted speech was elicited through textual and image prompts to capture natural linguistic variation and dialectal nuances. A customized mobile application enabled contributors to record using smartphones. Quality assurance operated at multiple layers, encompassing automated signal-to-noise ratio validation prior to recording and human review for content accuracy. Though the project encountered challenges common to low-resource settings, including unreliable infrastructure, device compatibility issues, and community trust barriers, these were mitigated through local mobilizers, stakeholder partnerships, and adaptive training protocols. AfriVoices-KE provides a foundational resource for developing inclusive automatic speech recognition and text-to-speech systems, while advancing the digital preservation of Kenya’s linguistic heritage.

Method: Developed a taxonomy of 10 empathic language tactics, then applied it to analyze 3,265 AI-generated responses (from six models) and 1,290 human-written responses across two studies to characterize empathic language patterns.

Result: LLM responses are highly formulaic, with 83-90% following a consistent template of empathic tactics, covering 81-92% of response content. Human responses are more diverse and less structured.

Conclusion: LLMs have learned and consistently deploy a well-liked template for empathy expression, which explains their perceived effectiveness, raising implications for the future of AI-generated emotional support.

Abstract: Recent research shows that greater numbers of people are turning to Large Language Models (LLMs) for emotional support, and that people rate LLM responses as more empathic than human-written responses. We suggest a reason for this success: LLMs have learned and consistently deploy a well-liked template for expressing empathy. We develop a taxonomy of 10 empathic language “tactics” that include validating someone’s feelings and paraphrasing, and apply this taxonomy to characterize the language that people and LLMs produce when writing empathic responses. Across a set of 2 studies comparing a total of n = 3,265 AI-generated (by six models) and n = 1,290 human-written responses, we find that LLM responses are highly formulaic at a discourse functional level. We discovered a template – a structured sequence of tactics – that matches between 83–90% of LLM responses (and 60–83% in a held out sample), and when those are matched, covers 81–92% of the response. By contrast, human-written responses are more diverse. We end with a discussion of implications for the future of AI-generated empathy.

Method: Designed a suite of simple, composable tasks spanning retrieval, morphological transformations, coreference, logical reasoning, and mathematics. Tracked emergence points across four model families (410M-13B parameters) and analyzed task representations.

Result: Found emergence orderings are strikingly consistent (ρ= .81 across 45 model pairs), composite tasks emerge after their components, and task structure is encoded in model representations. Representations can predict training trajectories of held-out tasks (R² = .68-.84).

Conclusion: Pretraining follows a structured, compositional curriculum that is consistent across models and readable from their internal representations, revealing more structure than loss curves alone.

Abstract: Large language models (LLMs) can perform remarkably complex tasks, yet the fine-grained details of how these capabilities emerge during pretraining remain poorly understood. Scaling laws on validation loss tell us how much a model improves with additional compute, but not what skills it acquires in which order. To remedy this, we propose the Implicit Curriculum Hypothesis: pretraining follows a compositional and predictable curriculum across models and data mixtures. We test this by designing a suite of simple, composable tasks spanning retrieval, morphological transformations, coreference, logical reasoning, and mathematics. Using these tasks, we track emergence points across four model families spanning sizes from 410M-13B parameters. We find that emergence orderings of when models reach fixed accuracy thresholds are strikingly consistent ($ρ= .81$ across 45 model pairs), and that composite tasks most often emerge after their component tasks. Furthermore, we find that this structure is encoded in model representations: tasks with similar function vector representations also tend to follow similar trajectories in training. By using the space of representations derived from our task set, we can effectively predict the training trajectories of simple held-out compositional tasks throughout the course of pretraining ($R^2 = .68$-$.84$ across models) without previously evaluating them. Together, these results suggest that pretraining is more structured than loss curves reveal: skills emerge in a compositional order that is consistent across models and readable from their internals.

Method: Formalize fact memorization from an information-theoretic perspective, analyze how training data distributions affect fact accuracy, and propose data selection schemes based on training loss to limit facts and flatten their frequency distribution.

Result: On semi-synthetic datasets with high-entropy facts, the selection method boosts fact accuracy to capacity limits. When pretraining on Wikipedia, a GPT2-Small model (110M parameters) memorized 1.3X more entity facts than standard training, matching performance of a 10X larger model (1.3B parameters).

Conclusion: Careful data selection based on training loss can significantly improve fact memorization in LLMs, allowing smaller models to achieve factual accuracy comparable to much larger models by optimizing training data composition.

Abstract: Large language models (LLMs) can struggle to memorize factual knowledge in their parameters, often leading to hallucinations and poor performance on knowledge-intensive tasks. In this paper, we formalize fact memorization from an information-theoretic perspective and study how training data distributions affect fact accuracy. We show that fact accuracy is suboptimal (below the capacity limit) whenever the amount of information contained in the training data facts exceeds model capacity. This is further exacerbated when the fact frequency distribution is skewed (e.g. a power law). We propose data selection schemes based on the training loss alone that aim to limit the number of facts in the training data and flatten their frequency distribution. On semi-synthetic datasets containing high-entropy facts, our selection method effectively boosts fact accuracy to the capacity limit. When pretraining language models from scratch on an annotated Wikipedia corpus, our selection method enables a GPT2-Small model (110m parameters) to memorize 1.3X more entity facts compared to standard training, matching the performance of a 10X larger model (1.3B parameters) pretrained on the full dataset.

Method: Created 153 tasks across 15 categories spanning 144 live platforms, requiring capabilities like document information extraction, multi-platform navigation, and detailed form filling. Uses a lightweight interception layer to capture and block final submission requests for safe evaluation without real-world side effects.

Result: Evaluation of 7 frontier models shows limited performance - Claude Sonnet 4.6 achieves only 33.3% success rate. Both proprietary and open-source models can complete only a small portion of these real-world tasks.

Conclusion: ClawBench reveals significant gaps in current AI agents’ ability to handle real-world online tasks. Progress on this benchmark will advance development of reliable general-purpose AI assistants.

Abstract: AI agents may be able to automate your inbox, but can they automate other routine aspects of your life? Everyday online tasks offer a realistic yet unsolved testbed for evaluating the next generation of AI agents. To this end, we introduce ClawBench, an evaluation framework of 153 simple tasks that people need to accomplish regularly in their lives and work, spanning 144 live platforms across 15 categories, from completing purchases and booking appointments to submitting job applications. These tasks require demanding capabilities beyond existing benchmarks, such as obtaining relevant information from user-provided documents, navigating multi-step workflows across diverse platforms, and write-heavy operations like filling in many detailed forms correctly. Unlike existing benchmarks that evaluate agents in offline sandboxes with static pages, ClawBench operates on production websites, preserving the full complexity, dynamic nature, and challenges of real-world web interaction. A lightweight interception layer captures and blocks only the final submission request, ensuring safe evaluation without real-world side effects. Our evaluations of 7 frontier models show that both proprietary and open-source models can complete only a small portion of these tasks. For example, Claude Sonnet 4.6 achieves only 33.3%. Progress on ClawBench brings us closer to AI agents that can function as reliable general-purpose assistants.

Method: Proposes StableOPD framework with two key components: 1) reference-based divergence constraint to mitigate repetition-induced length inflation, and 2) rollout mixture distillation to stabilize training dynamics.

Result: Across multiple math reasoning datasets, StableOPD prevents truncation collapse, stabilizes training dynamics, and improves performance by 7.2% on average compared to baseline OPD.

Conclusion: The interaction between student-induced data collection and distillation objective in OPD can cause severe training instability; StableOPD effectively addresses this issue through divergence constraints and rollout mixture strategies.

Abstract: On-policy distillation (OPD) trains student models under their own induced distribution while leveraging supervision from stronger teachers. We identify a failure mode of OPD: as training progresses, on-policy rollouts can undergo abrupt length inflation, causing truncated trajectories to dominate the training data. This truncation collapse coincides with abrupt repetition saturation and induces biased gradient signals, leading to severe training instability and sharp degradation in validation performance. We attribute this problem to the interaction between student-induced data collection and the distillation objective, which implicitly favors long and repetitive rollouts. To address this issue, we propose StableOPD, a stabilized OPD framework that combines a reference-based divergence constraint with rollout mixture distillation. These together mitigate repetition-induced length inflation and further stabilize OPD training. Across multiple math reasoning datasets, our approach prevents truncation collapse, stabilizes training dynamics, and improves performance by 7.2% on average.

Method: Created FinTruthQA benchmark with 6,000 real-world financial Q&A entries manually annotated on four criteria: question identification, question relevance, answer readability, and answer relevance. Benchmarked statistical ML models, pre-trained language models, fine-tuned variants, and LLMs on this dataset.

Result: Existing models achieve strong performance on question identification and relevance (F1 > 95%), but weaker on answer readability (~88% Micro F1) and especially answer relevance (~80% Micro F1). Domain-adapted pre-trained models consistently outperform general-purpose models and LLM prompting on challenging tasks.

Conclusion: FinTruthQA provides a practical foundation for AI-driven disclosure monitoring in capital markets, with value for regulatory oversight, investor protection, and disclosure governance, while highlighting the difficulty of fine-grained financial disclosure quality assessment.

Abstract: Accurate and transparent financial information disclosure is essential for market efficiency, investor decision-making, and corporate governance. Chinese stock exchanges’ investor interactive platforms provide a widely used channel through which listed firms respond to investor concerns, yet these responses are often limited or non-substantive, making disclosure quality difficult to assess at scale. To address this challenge, we introduce FinTruthQA, to our knowledge the first benchmark for AI-driven assessment of financial disclosure quality in investor-firm interactions. FinTruthQA comprises 6,000 real-world financial Q&A entries, each manually annotated based on four key evaluation criteria: question identification, question relevance, answer readability, and answer relevance. We benchmark statistical machine learning models, pre-trained language models and their fine-tuned variants, as well as large language models (LLMs), on FinTruthQA. Experiments show that existing models achieve strong performance on question identification and question relevance (F1 > 95%), but remain substantially weaker on answer readability (Micro F1 approximately 88%) and especially answer relevance (Micro F1 approximately 80%), highlighting the nontrivial difficulty of fine-grained disclosure quality assessment. Domain- and task-adapted pre-trained language models consistently outperform general-purpose models and LLM-based prompting on the most challenging settings. These findings position FinTruthQA as a practical foundation for AI-driven disclosure monitoring in capital markets, with value for regulatory oversight, investor protection, and disclosure governance in real-world financial settings.

Method: Evaluated LLMs on two tasks: 1) Bias detection in Wikipedia edits, and 2) Bias correction by generating neutral rewrites. Used balanced datasets and compared LLM performance against Wikipedia editors and crowdworker ratings.

Result: LLMs achieved only 64% accuracy on bias detection, with contrasting biases across models. For generation, they removed 79% of words removed by Wikipedia editors but made additional extraneous changes, resulting in high recall but low precision. Crowdworkers rated AI rewrites as more neutral (70%) and fluent (61%) than Wikipedia-editor rewrites.

Conclusion: LLMs may apply rules in ways that resonate with the public but diverge from community experts. While potentially effective for generation, they may reduce editor agency and increase moderation workload. Having LLMs apply rules like community members remains difficult even when rules are easy to articulate.

Abstract: Large language models (LLMs) are trained on broad corpora and then used in communities with specialized norms. Is providing LLMs with community rules enough for models to follow these norms? We evaluate LLMs’ capacity to detect (Task 1) and correct (Task 2) biased Wikipedia edits according to Wikipedia’s Neutral Point of View (NPOV) policy. LLMs struggled with bias detection, achieving only 64% accuracy on a balanced dataset. Models exhibited contrasting biases (some under- and others over-predicted bias), suggesting distinct priors about neutrality. LLMs performed better at generation, removing 79% of words removed by Wikipedia editors. However, LLMs made additional changes beyond Wikipedia editors’ simpler neutralizations, resulting in high-recall but low-precision editing. Interestingly, crowdworkers rated AI rewrites as more neutral (70%) and fluent (61%) than Wikipedia-editor rewrites. Qualitative analysis found LLMs sometimes applied NPOV more comprehensively than Wikipedia editors but often made extraneous non-NPOV-related changes (such as grammar). LLMs may apply rules in ways that resonate with the public but diverge from community experts. While potentially effective for generation, LLMs may reduce editor agency and increase moderation workload (e.g., verifying additions). Even when rules are easy to articulate, having LLMs apply them like community members may still be difficult.

Method: Proposes a systematic framework for hypothesis generation using LMs to simulate outcomes of non-existent experiments. Instantiates this framework in child language development context, specifically dative verb acquisition and cross-structural generalization. Uses LMs as simulated learners to derive novel hypotheses.

Result: Derives novel, untested hypotheses: alignment between argument ordering and discourse prominence features of exposure contexts modulates how children generalize new verbs to unobserved structures. Also designs experiments to test these hypotheses with children in lab settings.

Conclusion: Contributes both a domain-general framework for systematic hypothesis generation via simulated learners and domain-specific, lab-testable hypotheses for child language acquisition research.

Abstract: Neural language models (LMs) have been shown to capture complex linguistic patterns, yet their utility in understanding human language and more broadly, human cognition, remains debated. While existing work in this area often evaluates human-machine alignment, few studies attempt to translate findings from this enterprise into novel insights about humans. To this end, we propose a systematic framework for hypothesis generation that uses LMs to simulate outcomes of experiments that do not yet exist in the literature. We instantiate this framework in the context of a specific research question in child language development: dative verb acquisition and cross-structural generalization. Through this instantiation, we derive novel, untested hypotheses: the alignment between argument ordering and discourse prominence features of exposure contexts modulates how children generalize new verbs to unobserved structures. Additionally, we also design a set of experiments that can test these hypotheses in the lab with children. This work contributes both a domain-general framework for systematic hypothesis generation via simulated learners and domain-specific, lab-testable hypotheses for child language acquisition research.

Method: 1) Define and quantify problem drift across ten tasks (3 generative, 3 knowledge, 3 reasoning, 1 instruction-following). 2) Human expert analysis of 170 multi-agent debates to identify drift causes. 3) Propose DRIFTJudge (LLM-as-a-judge method) to detect problem drift. 4) Propose DRIFTPolicy to mitigate problem drift.

Result: Generative tasks drift often (76-89%) due to answer space subjectivity, while high-complexity tasks drift less (7-21%). Main drift causes: lack of progress (35%), low-quality feedback (26%), lack of clarity (25%). DRIFTPolicy mitigates 31% of problem drift cases.

Conclusion: Problem drift is a key limitation of multi-agent debate that explains why longer debates can harm performance. The study provides understanding of this phenomenon and proposes initial solutions (DRIFTJudge and DRIFTPolicy) to detect and mitigate drift.

Abstract: Multi-agent debate - multiple instances of large language models discussing problems in turn-based interaction - has shown promise for solving knowledge and reasoning tasks. However, these methods show limitations when solving complex problems that require longer reasoning chains. We analyze how multi-agent debate drifts away from the initial problem over multiple turns, thus harming task performance. We define this phenomenon as problem drift and quantify its presence across ten tasks (i.e., three generative, three knowledge, three reasoning, and one instruction-following task). We find that generative tasks drift often due to the subjectivity of the answer space (76-89%), compared to high-complexity tasks (7-21%). To identify the reasons, eight human experts analyze 170 multi-agent debates suffering from problem drift. We find the most common issues related to this drift are the lack of progress (35% of cases), low-quality feedback (26% of cases), and a lack of clarity (25% of cases). We propose DRIFTJudge, an LLM-as-a-judge method, as a first baseline to detect problem drift. We also propose DRIFTPolicy, which mitigates 31% of problem drift cases. Our study is a step toward understanding a key limitation of multi-agent debate, highlighting why longer debates can harm task performance and how problem drift could be addressed.

Method: Integrates token-level Monte Carlo Tree Search (MCTS) guided by syntactic and semantic constraints expressed using Answer Set Grammars, a logic-based formalism that generalizes context-sensitive grammars with background knowledge for task-specific semantics.

Result: SEM-CTRL enables even small pre-trained LLMs to outperform larger variants and state-of-the-art reasoning models (like o4-mini) on tasks including synthetic grammar synthesis, combinatorial reasoning, JSON parsing, and planning while guaranteeing semantic validity.

Conclusion: The approach provides a unified framework for enforcing rich constraints on LLM outputs without fine-tuning, improving reliability and correctness for real-world deployment across diverse tasks.

Abstract: Ensuring both syntactic and semantic correctness in Large Language Model (LLM) outputs remains a significant challenge, despite being critical for real-world deployment. In this paper, we introduce $\texttt{SEM-CTRL}$, a unified approach that allows for enforcing rich context-sensitive constraints, and task and instance specific semantics directly on the LLM decoder. Our approach integrates token-level MCTS which is guided by specific syntactic and semantic constraints. The constraints over desired outputs are expressed using Answer Set Grammars, which is a logic-based formalism that generalizes context sensitive grammars while incorporating background knowledge to represent task-specific semantics. We show that our approach helps guarantee valid completions for any off-the-shelf LLM without the need for fine-tuning. We evaluate $\texttt{SEM-CTRL}$ on a range of tasks, including synthetic grammar synthesis, combinatorial reasoning, JSON parsing, and planning. Our experimental results demonstrate that $\texttt{SEM-CTRL}$ allows even small pre-trained LLMs to efficiently outperform larger variants and state-of-the-art reasoning models (e.g., $\textit{o4-mini}$) while simultaneously guaranteeing semantic validity.

Method: Conducted studies with market research professionals to evaluate LLMs for CN abstraction. Used supervised fine-tuned (SFT) LLMs across various product/service categories, comparing them against professional analysts and foundational LLMs.

Result: SFT LLMs perform at least as well as professional analysts and substantially better than foundational LLMs. Results generalize across alternative foundational models and work with relatively “small” models. Abstracted CNs are well-formulated, specific enough to guide innovation, and grounded in source content without hallucination.

Conclusion: LLMs can effectively automate customer need abstraction, transforming VOC approaches by enabling scalable insight discovery and allowing analysts to focus on higher-value tasks. SFT training enables LLMs to learn professional CN formulation conventions rather than memorizing patterns.

Abstract: Identifying customer needs (CNs) is fundamental to product innovation and marketing strategy. Yet for over thirty years, Voice-of-the-Customer (VOC) applications have relied on professional analysts to manually interpret qualitative data and formulate “jobs to be done.” This task is cognitively demanding, time-consuming, and difficult to scale. While current practice uses machine learning to screen content, the critical final step of precisely formulating CNs relies on expert human judgment. We conduct a series of studies with market research professionals to evaluate whether Large Language Models (LLMs) can automate CN abstraction. Across various product and service categories, we demonstrate that supervised fine-tuned (SFT) LLMs perform at least as well as professional analysts and substantially better than foundational LLMs. These results generalize to alternative foundational LLMs and require relatively “small” models. The abstracted CNs are well-formulated, sufficiently specific to guide innovation, and grounded in source content without hallucination. Our analysis suggests that SFT training enables LLMs to learn the underlying syntactic and semantic conventions of professional CN formulation rather than relying on memorized CNs. Automation of tedious tasks transforms the VOC approach by enabling the discovery of high-leverage insights at scale and by refocusing analysts on higher-value-added tasks.

Method: Proposes Hierarchical Reward Model (HRM) that evaluates both individual and consecutive reasoning steps at fine-grained and coarse-grained levels, plus Hierarchical Node Compression (HNC) data augmentation that merges consecutive reasoning steps in tree structures to enhance training data diversity.

Result: HRM with HNC outperforms PRM on PRM800K dataset, showing more stable and reliable evaluations. Cross-domain evaluations on MATH500 and GSM8K demonstrate strong generalization and robustness across reasoning tasks.

Conclusion: HRM provides a more reliable approach to evaluating multi-step reasoning processes, addressing reward hacking issues while reducing data annotation costs through effective data augmentation.

Abstract: Recent studies show that Large Language Models (LLMs) achieve strong reasoning capabilities through supervised fine-tuning or reinforcement learning. However, a key approach, the Process Reward Model (PRM), suffers from reward hacking, making it unreliable in identifying the best intermediate step. In addition, the cost of annotating reasoning processes for reward modeling is high, making large-scale collection of high-quality data challenging. To address this, we propose a novel reward model approach called the Hierarchical Reward Model (HRM), which evaluates both individual and consecutive reasoning steps at both fine-grained and coarse-grained levels. HRM excels at assessing multi-step reasoning coherence, especially when flawed steps are later corrected through self-reflection. To further reduce the cost of generating training data, we introduce a lightweight and effective data augmentation strategy called Hierarchical Node Compression (HNC), which merges two consecutive reasoning steps into one within the tree structure. By applying HNC to MCTS-generated reasoning trajectories, we enhance the diversity and robustness of HRM training data while introducing controlled noise with minimal computational overhead. Empirical results on the PRM800K dataset show that HRM, together with HNC, provides more stable and reliable evaluations than PRM. Furthermore, cross-domain evaluations on the MATH500 and GSM8K datasets demonstrate HRM’s strong generalization and robustness across a variety of reasoning tasks.

Method: Constructs dynamic event graph distinguishing core events (main goals) and supporting events (interconnected details), employs multi-head attention to selectively determine relevant events, captures three distinct relationship types between events.

Result: Experiments on three dialogue datasets show EventWeave produces more natural and contextually appropriate responses with less computational overhead than models processing entire dialogue history.

Conclusion: EventWeave effectively balances comprehensive context understanding with generating concise responses through targeted optimization techniques, with improvements stemming from better event relationship modeling rather than increased information density.

Abstract: Large language models have improved dialogue systems, but often process conversational turns in isolation, overlooking the event structures that guide natural interactions. Hence we introduce EventWeave, a framework that explicitly models relationships between conversational events to generate more contextually appropriate dialogue responses. EventWeave constructs a dynamic event graph that distinguishes between core events (main goals) and supporting events (interconnected details), employing a multi-head attention mechanism to selectively determine which events are most relevant to the current turn. Unlike summarization or standard graph-based approaches, our method captures three distinct relationship types between events, allowing for more nuanced context modeling. Experiments on three dialogue datasets demonstrate that EventWeave produces more natural and contextually appropriate responses while requiring less computational overhead than models processing the entire dialogue history. Ablation studies confirm improvements stem from better event relationship modeling rather than increased information density. Our approach effectively balances comprehensive context understanding with generating concise responses, maintaining strong performance across various dialogue lengths through targeted optimization techniques.

Method: Construct a “sandbox” using demographically and topically split Reddit data validated by self-identification, with a two-stage process to filter potential sociocultural linguistic phenomena for deeper investigation.

Result: Created Splits! dataset and validated it by replicating known sociocultural linguistic phenomena from existing literature, demonstrating the sandbox’s utility for scalable sociolinguistic research.

Conclusion: The proposed sandbox enables systematic and flexible sociolinguistic research, facilitating quick hypothesis exploration and prototyping of sociocultural linguistic phenomena.

Abstract: Variation in language use, shaped by speakers’ sociocultural background and specific context of use, offers a rich lens into cultural perspectives, values, and opinions. For example, Chinese students discuss “healthy eating” with words like “timing,” “regularity,” and “digestion,” whereas Americans use vocabulary like “balancing food groups” and “avoiding fat and sugar,” reflecting distinct cultural models of nutrition. The computational study of these Sociocultural Linguistic Phenomena (SLP) has traditionally been done in NLP via tailored analyses of specific groups or topics, requiring specialized data collection and experimental operationalization–a process not well-suited to quick hypothesis exploration and prototyping. To address this, we propose constructing a “sandbox” designed for systematic and flexible sociolinguistic research. Using our method, we construct a demographically/topically split Reddit dataset, Splits!, validated by self-identification and by replicating several known SLPs from existing literature. We showcase the sandbox’s utility with a scalable, two-stage process that filters large collections of “potential” SLPs (PSLPs) to surface the most promising candidates for deeper, qualitative investigation.

Method: Introduces arXiv2Table benchmark with 1,957 tables referencing 7,158 papers, human-verified distractors, and rewritten schema-agnostic user demands. Develops iterative, batch-based generation method that co-refines paper filtering and schema over multiple rounds. Proposes lightweight evaluation decomposing utility into schema coverage, cell fidelity, and relational consistency.

Result: Method consistently improves over strong baselines, though absolute scores remain modest, highlighting task difficulty. Evaluation protocol validated through human audits and cross-evaluator checks.

Conclusion: Presents realistic benchmark and evaluation framework for literature review table generation, demonstrating task difficulty while providing reproducible evaluation methodology.

Abstract: Literature review tables are essential for summarizing and comparing collections of scientific papers. In this paper, we study the automatic generation of such tables from a pool of papers to satisfy a user’s information need. Building on recent work (Newman et al., 2024), we move beyond oracle settings by (i) simulating well-specified yet schema-agnostic user demands that avoid leaking gold column names or values, (ii) explicitly modeling retrieval noise via semantically related but out-of-scope distractor papers verified by human annotators, and (iii) introducing a lightweight, annotation-free, utilization-oriented evaluation that decomposes utility into schema coverage, unary cell fidelity, and pairwise relational consistency, while measuring paper selection through a two-way QA procedure (gold to system and system to gold) with recall, precision, and F1. To support reproducible evaluation, we introduce arXiv2Table, a benchmark of 1,957 tables referencing 7,158 papers, with human-verified distractors and rewritten, schema-agnostic user demands. We also develop an iterative, batch-based generation method that co-refines paper filtering and schema over multiple rounds. We validate the evaluation protocol with human audits and cross-evaluator checks. Extensive experiments show that our method consistently improves over strong baselines, while absolute scores remain modest, underscoring the task’s difficulty. Our data and code is available at https://github.com/JHU-CLSP/arXiv2Table.

Method: Developed a globally connected knowledge graph with 18,850 biological nodes and 48,944 edges, used by LLMs to retrieve entities for cell reconstruction, plus a multi-task reasoning workflow.

Result: Improves human evaluation scores by up to 0.21 and semantic similarity by 6.1% across multiple tissue types, narrowing performance gap between large and small LLMs.

Conclusion: The approach offers a paradigm for structured knowledge integration and reasoning in bioinformatics, aligning better with manual annotation cognitive logic.

Abstract: With the rapid development of large language models (LLMs), their application to cell type annotation has drawn increasing attention. However, general-purpose LLMs often face limitations in this specific task due to the lack of guidance from external domain knowledge. To enable more accurate and fully automated cell type annotation, we develop a globally connected knowledge graph comprising 18850 biological information nodes, including cell types, gene markers, features, and other related entities, along with 48,944 edges connecting these nodes, which is used by LLMs to retrieve entities associated with differential genes for cell reconstruction. Additionally, a multi-task reasoning workflow is designed to optimise the annotation process. Compared to general-purpose LLMs, our method improves human evaluation scores by up to 0.21 and semantic similarity by 6.1% across multiple tissue types, while more closely aligning with the cognitive logic of manual annotation. Meanwhile, it narrows the performance gap between large and small LLMs in cell type annotation, offering a paradigm for structured knowledge integration and reasoning in bioinformatics.

Method: FoVer leverages formal verification tools (Z3 and Isabelle) to annotate step-level error labels in formal reasoning tasks. It synthesizes PRM training data from formal logic and theorem proving tasks by using these verification tools to automatically identify correct and incorrect reasoning steps.

Result: Experiments on 12 reasoning benchmarks show that fine-tuning on FoVer-generated training data improves PRMs on math/logic reasoning tasks (informal variants of training tasks) and also on NLI and BBH benchmarks that differ substantially from training tasks.

Conclusion: FoVer demonstrates that PRM training data created using formal verification can effectively improve PRMs on informal reasoning tasks written in natural language, providing an efficient alternative to human annotation and LLM-based labeling methods.

Abstract: Process Reward Models (PRMs) have emerged as a promising approach for improving LLM reasoning capabilities by providing process supervision over reasoning traces. However, existing approaches for constructing PRM training data remain costly and noisy, as they typically rely on human annotation or sampling-based labeling methods that require repeated LLM calls. In this work, we propose FoVer, a framework that synthesizes PRM training data from formal reasoning tasks by annotating step-level error labels using formal verification tools such as Z3 and Isabelle. By leveraging formal verification, FoVer enables efficient and accurate PRM data construction without requiring human annotation or additional LLM calls. Using FoVer, we create PRM training data from formal logic and theorem proving tasks. Experiments on 12 reasoning benchmarks show that fine-tuning on our training data improves PRMs not only on math and logic reasoning tasks, which are informal variants of the training tasks, but also on NLI and BBH benchmarks, which differ substantially from the tasks used to construct the training data. These results demonstrate the practical effectiveness of FoVer, showing that PRM training data created using formal verification improves PRMs on informal reasoning tasks written in natural language. The datasets, models, and code are provided at https://github.com/psunlpgroup/FoVer.

Method: Created three benchmark flavors: Seal-0 (main) with most challenging questions, Seal-Hard for factual accuracy and reasoning, and LongSeal for long-context multi-document reasoning in “needle-in-a-haystack” settings. Evaluated frontier LLMs including GPT-4.1, o3, o4-mini, and DeepSeek-R1-671B.

Result: Frontier models perform poorly: o3 and o4-mini achieve only 17.1% and 6.3% accuracy respectively on Seal-0. Advanced reasoning models are highly vulnerable to noisy search results. Increasing test-time compute doesn’t yield reliable gains, and models still fail to identify relevant documents in LongSeal despite reduced “lost-in-the-middle” issues.

Conclusion: Current search-augmented language models have critical limitations in handling challenging fact-seeking questions with unreliable search results, highlighting the need for improved reasoning capabilities and robustness to noise.

Abstract: We introduce SealQA, a new challenge benchmark for evaluating SEarch-Augmented Language models on fact-seeking questions where web search yields conflicting, noisy, or unhelpful results. SealQA comes in three flavors: (1) Seal-0 (main) and (2) Seal-Hard, which assess factual accuracy and reasoning capabilities, with Seal-0 focusing on the most challenging questions where chat models (e.g., GPT-4.1) typically achieve near-zero accuracy; and (3) LongSeal, which extends SealQA to test long-context, multi-document reasoning in “needle-in-a-haystack” settings. Our evaluation reveals critical limitations in current models: Even frontier LLMs perform poorly across all SealQA flavors. On Seal-0, frontier agentic models equipped with tools like o3 and o4-mini achieve only 17.1% and 6.3% accuracy, respectively, at their best reasoning efforts. We find that advanced reasoning models such as DeepSeek-R1-671B and o3-mini are highly vulnerable to noisy search results. Notably, increasing test-time compute does not yield reliable gains across o3-mini, o4-mini, and o3, with performance often plateauing or even declining early. Additionally, while recent models are less affected by the “lost-in-the-middle” issue, they still fail to reliably identify relevant documents in LongSeal when faced with numerous distractors. To facilitate future work, we release SealQA at huggingface.co/datasets/vtllms/sealqa.

Method: Proposes a framework integrating Direct Preference Optimization (DPO) with curriculum learning to align machine-generated explanations with human reasoning. Uses fact-checked explanations from credible sources as preferred responses and LLM outputs as non-preferred responses. Introduces two key parameters (Actuality and Finesse) into the DPO loss function to enhance explanation quality and consistency.

Result: Experiments with various LLMs (Mistral, Llama, Gemma) and PLMs (mBART, mT5) confirm the framework’s effectiveness in generating coherent, contextually relevant explanations. The approach demonstrates scalability and effectiveness for low-resource languages.

Conclusion: The proposed scalable approach effectively combats misinformation and extends automated explanation generation capabilities to low-resource languages, providing a valuable tool for under-represented linguistic communities.

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

Method: Proposes structured empathetic reasoning with three steps: conversation history analysis, multimodal emotional state inference, and strategy selection before response generation. Uses PEER framework with GRPO and UnifiReward (unified process-outcome reward model) trained on SER dataset with step-level labels and pairwise preferences.

Result: Comprehensive experiments show improved empathy, strategy alignment, and human-likeness without sacrificing diversity in emotional support conversations.

Conclusion: Structured empathetic reasoning with unified reward modeling effectively enhances emotional support systems by incorporating psychology-informed steps while reducing repetitive patterns.

Abstract: Emotional support conversations require more than fluent responses. Supporters need to understand the seeker’s situation and emotions, adopt an appropriate strategy, and respond in a natural, human-like manner. Despite advances in large language models, current systems often lack structured, psychology-informed reasoning. Additionally, it is challenging to enhance these systems through reinforcement learning because of unreliable reward signals. Moreover, reinforcement fine-tuning can amplify repetitive response patterns. We propose structured empathetic reasoning, which breaks support into three steps: conversation history analysis, multimodal emotional state inference, and strategy selection, prior to generating the final reply. To implement this, we introduce SER, a fine-grained dataset with step-level correctness labels and pairwise response preferences. We then present PEER, which uses GRPO with UnifiReward, a unified process-outcome reward model for evaluating both reasoning steps and final responses in multi-turn interactions. To reduce repetition, we enhance data with personality-based rewriting and down-weight redundant outputs. Comprehensive experiments show improved empathy, strategy alignment, and human-likeness without sacrificing diversity.

Method: Proposes LongMab framework that treats context chunks as arms in a Multi-Armed Bandit system, selects chunks based on expected reward scores to generate responses, iteratively updates scores based on reward feedback, and constructs preference data pairs for Direct Preference Optimization training.

Result: Experimental results on Llama and Qwen models show LongMab achieves more than 4% improvement on long-context reasoning benchmarks compared to existing approaches.

Conclusion: LongMab effectively addresses data quality issues in long-context LLM training by intelligently selecting informative context segments, leading to significant performance improvements on long-context reasoning tasks.

Abstract: Long-context modeling is critical for a wide range of real-world tasks, including long-context question answering, summarization, and complex reasoning tasks. Recent studies have explored fine-tuning Large Language Models (LLMs) with synthetic data to enhance their long-context capabilities. However, the effectiveness of such approaches is often limited by the low diversity and factual inconsistencies in the generated data. To address these challenges, we propose LongMab, a novel framework that leverages a Multi-Armed Bandit (MAB) rollout strategy to identify the most informative chunks from the given long context for sampling high-quality and diverse responses and constructing preference data pairs for Direct Preference Optimization (DPO) training. Specifically, we treat context chunks as arms of MAB, select chunks based on their expected reward scores to input into LLMs to generate responses, and iteratively update these scores based on reward feedback. Both exploration and exploitation during the rollout process enable the LLM to focus on the most relevant context segments, thereby generating and collecting high-quality and diverse responses. Experimental results on both Llama and Qwen show the effectiveness of LongMab by achieving more than a 4% improvement on long-context reasoning benchmarks. All data and code will be released on https://github.com/NEUIR/LongMab-PO.

Method: Prophet dynamically decides whether to continue refinement or decode all remaining tokens in one step (“all-in”) based on the confidence gap between top-2 prediction candidates. It requires no additional training and integrates into existing DLM implementations.

Result: On GSM8K and MMLU, up to 97% and 99% of instances can be decoded correctly using only half the refinement steps. Prophet reduces decoding steps by up to 3.4x for LLaDA-8B and Dream-7B across multiple tasks while preserving generation quality.

Conclusion: Early decode convergence provides a simple yet powerful mechanism for accelerating DLM inference, complementary to existing speedup techniques. Prophet recasts DLM decoding as a problem of when to stop sampling.

Abstract: Diffusion language models (DLMs) have recently emerged as an alternative to autoregressive approaches, offering parallel sequence generation and flexible token orders. However, their inference remains slower than that of autoregressive models, primarily due to the cost of bidirectional attention and the large number of refinement steps required for high quality outputs. In this work, we highlight and leverage an overlooked property of DLMs early answer convergence: in many cases, the correct answer can be internally identified by half steps before the final decoding step, both under semi-autoregressive and random remasking schedules. For example, on GSM8K and MMLU, up to 97% and 99% of instances, respectively, can be decoded correctly using only half of the refinement steps. Building on this observation, we introduce Prophet, a training-free fast decoding paradigm that enables early commit decoding. Specifically, Prophet dynamically decides whether to continue refinement or to go “all-in” (i.e., decode all remaining tokens in one step), using the confidence gap between the top-2 prediction candidates as the criterion. It integrates seamlessly into existing DLM implementations, incurs negligible overhead, and requires no additional training. Empirical evaluations of LLaDA-8B and Dream-7B across multiple tasks show that Prophet reduces the number of decoding steps by up to 3.4x while preserving high generation quality. These results recast DLM decoding as a problem of when to stop sampling, and demonstrate that early decode convergence provides a simple yet powerful mechanism for accelerating DLM inference, complementary to existing speedup techniques. Our code is publicly available at https://github.com/pixeli99/Prophet.

Method: Introduces MARCH benchmark with 2,209 multi-hop ambiguous questions curated via multi-LLM verification and human annotation. Proposes CLARION, a two-stage agentic framework that explicitly decouples ambiguity planning from evidence-driven reasoning.

Result: State-of-the-art models struggle with MARCH, confirming that combining ambiguity resolution with multi-step reasoning is a significant challenge. CLARION significantly outperforms existing approaches on the benchmark.

Conclusion: Multi-hop ambiguous QA presents a significant challenge requiring new approaches. CLARION’s decoupled ambiguity planning and evidence-driven reasoning framework paves the way for more robust reasoning systems.

Abstract: Real-world multi-hop QA is naturally linked with ambiguity, where a single query can trigger multiple reasoning paths that require independent resolution. Since ambiguity can occur at any stage, models must navigate layered uncertainty throughout the entire reasoning chain. Despite its prevalence in real-world user queries, previous benchmarks have primarily focused on single-hop ambiguity, leaving the complex interaction between multi-step inference and layered ambiguity underexplored. In this paper, we introduce \textbf{MARCH}, a benchmark for their intersection, with 2,209 multi-hop ambiguous questions curated via multi-LLM verification and validated by human annotation with strong agreement. Our experiments reveal that even state-of-the-art models struggle with MARCH, confirming that combining ambiguity resolution with multi-step reasoning is a significant challenge. To address this, we propose \textbf{CLARION}, a two-stage agentic framework that explicitly decouples ambiguity planning from evidence-driven reasoning, significantly outperforms existing approaches, and paves the way for robust reasoning systems.

Method: Comprehensive benchmark study framing bias detection as multi-label task of detecting targeted identities using demographic-focused taxonomy. Systematic evaluation of models across scales and techniques including prompting, in-context learning, and fine-tuning across twelve datasets spanning diverse content types and demographics.

Result: Demonstrates promise of fine-tuned smaller models for scalable bias detection, but exposes persistent gaps across demographic axes and multi-demographic targeted biases.

Conclusion: Highlights need for more effective and scalable detection frameworks to address limitations in current LLM-based bias detection methods, particularly for multi-demographic biases.

Abstract: Large-scale web-scraped text corpora used to train general-purpose AI models often contain harmful demographic-targeted social biases, creating a regulatory need for data auditing and developing scalable bias-detection methods. Although prior work has investigated biases in text datasets and related detection methods, these studies remain narrow in scope. They typically focus on a single content type (e.g., hate speech), cover limited demographic axes, overlook biases affecting multiple demographics simultaneously, and analyze limited techniques. Consequently, practitioners lack a holistic understanding of the strengths and limitations of recent large language models (LLMs) for automated bias detection. In this study, we conduct a comprehensive benchmark study on English texts to assess the ability of LLMs in detecting demographic-targeted social biases. To align with regulatory requirements, we frame bias detection as a multi-label task of detecting targeted identities using a demographic-focused taxonomy. We then systematically evaluate models across scales and techniques, including prompting, in-context learning, and fine-tuning. Using twelve datasets spanning diverse content types and demographics, our study demonstrates the promise of fine-tuned smaller models for scalable detection. However, our analyses also expose persistent gaps across demographic axes and multi-demographic targeted biases, underscoring the need for more effective and scalable detection frameworks.

Method: Constructs a structured Knowledge Graph from the OMIn dataset and integrates it into a retrieval-augmented generation pipeline, evaluating locally deployable LLMs (Gemma-3, Phi-4, Mistral-Nemo) with stronger models (GPT-4o, Llama-3.3) as judges.

Result: KEO markedly improves global sensemaking by revealing patterns and system-level insights, while text-chunk RAG remains effective for fine-grained procedural tasks requiring localized retrieval.

Conclusion: KG-augmented LLMs show promise for secure, domain-specific QA and have potential in high-stakes reasoning applications.

Abstract: We present Knowledge Extraction on OMIn (KEO), a domain-specific knowledge extraction and reasoning framework with large language models (LLMs) in safety-critical contexts. Using the Operations and Maintenance Intelligence (OMIn) dataset, we construct a QA benchmark spanning global sensemaking and actionable maintenance tasks. KEO builds a structured Knowledge Graph (KG) and integrates it into a retrieval-augmented generation (RAG) pipeline, enabling more coherent, dataset-wide reasoning than traditional text-chunk RAG. We evaluate locally deployable LLMs (Gemma-3, Phi-4, Mistral-Nemo) and employ stronger models (GPT-4o, Llama-3.3) as judges. Experiments show that KEO markedly improves global sensemaking by revealing patterns and system-level insights, while text-chunk RAG remains effective for fine-grained procedural tasks requiring localized retrieval. These findings underscore the promise of KG-augmented LLMs for secure, domain-specific QA and their potential in high-stakes reasoning. The code is available at https://github.com/JonathanKarr33/keo.

Method: Proposes a prompt optimization framework inspired by reinforcement learning that modifies only the task instruction prompt. Uses turn-by-turn feedback generation and experience replay for prompt rewriting to enable long-term planning.

Result: Shows significant improvement in multi-turn tasks like text-to-SQL and task-oriented dialogue. Generalizes across different LLM-based agents and can leverage diverse LLMs as meta-prompting agents.

Conclusion: The framework enables effective long-term planning for LLMs in multi-turn interactions through parameter-free optimization, warranting future research in reinforcement learning-inspired prompt optimization methods.

Abstract: Large language models (LLMs) have achieved remarkable success in a wide range of natural language processing tasks and can be adapted through prompting. However, they remain suboptimal in multi-turn interactions, often relying on incorrect early assumptions and failing to track user goals over time, which makes such tasks particularly challenging. Prior works in dialogue systems have shown that long-term planning is essential for handling interactive tasks. In this work, we propose a prompt optimisation framework inspired by reinforcement learning, which enables such planning to take place by only modifying the task instruction prompt of the LLM-based agent. By generating turn-by-turn feedback and leveraging experience replay for prompt rewriting, our proposed method shows significant improvement in multi-turn tasks such as text-to-SQL and task-oriented dialogue. Moreover, it generalises across different LLM-based agents and can leverage diverse LLMs as meta-prompting agents. This warrants future research in reinforcement learning-inspired parameter-free optimisation methods.

Method: Developed PiKa, a family of expert-level alignment datasets that concentrate supervision on high-difficulty instructions. Created PiKa-SFT with only 30k examples and also provided 30k preference optimization examples. Fine-tuned models like Llama-3-8B-Base and Qwen2.5 series on this dataset.

Result: Llama-3-8B-Base fine-tuned on PiKa-SFT outperforms official Llama-3-8B-Instruct (trained on 10M+ proprietary examples) on AlpacaEval 2.0 and Arena-Hard benchmarks. Similar improvements observed across Qwen2.5 series (0.5B-7B) over their official instruction-tuned counterparts.

Conclusion: High-quality alignment can be achieved with significantly reduced data by focusing on difficult instructions, democratizing access for resource-constrained research. Prompt difficulty is a critical factor beyond prompt-response quality.

Abstract: High-quality instruction data is critical for LLM alignment, yet existing open-source datasets often lack efficiency, requiring hundreds of thousands of examples to approach proprietary performance. In this work, we find that beyond the widely recognized importance of prompt-response quality, prompt difficulty itself plays a critical role in driving alignment gains. Motivated by this observation, we introduce PiKa, a data-efficient family of expert-level alignment datasets that concentrates supervision on high-difficulty instructions. The PiKa-SFT dataset contains only 30k examples, an order of magnitude fewer than state-of-the-art open datasets like Magpie-Pro. Despite its small size, fine-tuning Llama-3-8B-Base on PiKa-SFT even outperforms the official Llama-3-8B-Instruct model trained on over 10M proprietary examples on widely used benchmarks such as AlpacaEval 2.0 and Arena-Hard. We also validate the generalizability of PiKa across the Qwen2.5 series (0.5B-7B), consistently surpassing their official instruction-tuned counterparts. Additionally, we provide 30k high-quality preference optimization examples to further enhance alignment. Our results demonstrate that promising alignment is achievable with significantly reduced data, democratizing access for resource-constrained research. Our code and data will be available at https://github.com/SJY8460/PiKa.

Method: Three-stage approach: 1) Supervised learning to enable quality embedding generation, 2) Reinforcement learning to optimize embedding generation alongside reasoning, 3) Specialized RL environment that handles evolving embeddings without complete corpus re-encoding.

Result: Extensive evaluations show Search-R3 significantly outperforms prior methods by unifying reasoning and embedding generation processes, advancing complex knowledge-intensive tasks requiring both reasoning and retrieval.

Conclusion: Search-R3 represents a substantial advancement in adapting LLMs for retrieval tasks through integrated reasoning and embedding generation, enabling better handling of complex knowledge-intensive tasks.

Abstract: Despite their remarkable natural language understanding capabilities, Large Language Models (LLMs) have been underutilized for retrieval tasks. We present Search-R3, a novel framework that addresses this limitation by adapting LLMs to generate search embeddings as a direct output of their reasoning process. Our approach exploits LLMs’ chain-of-thought capabilities, allowing them to produce more effective embeddings by reasoning step-by-step through complex semantic analyses. We implement this through three complementary mechanisms. (1) a supervised learning stage enables the model’s ability to produce quality embeddings, (2) a reinforcement learning (RL) methodology that optimizes embedding generation alongside reasoning, and (3) a specialized RL environment that efficiently handles evolving embedding representations without requiring complete corpus re-encoding at each training iteration. Our extensive evaluations on diverse benchmarks demonstrate that Search-R3 significantly outperforms prior methods by unifying the reasoning and embedding generation processes. This integrated post-training approach represents a substantial advancement in handling complex knowledge-intensive tasks that require both sophisticated reasoning and effective information retrieval. Project page: https://github.com/ytgui/Search-R3

Method: Introduces a benchmark dataset built from literary and blog texts paired with LLM-generated imitations. Proposes a method that identifies latent directions corresponding to inverted features and constructs probe datasets to evaluate detector dependence on these features.

Result: Large performance gaps across detectors in personalized settings, with some state-of-the-art models suffering significant drops. The proposed method shows 85% correlation with actual performance gaps, accurately predicting both direction and magnitude of post-transfer changes.

Conclusion: This work introduces the first benchmark for personalized text detection, identifies the feature-inversion trap as a key limitation, and provides a reliable method to predict detector performance changes, encouraging further research in personalized text detection.

Abstract: Large language models (LLMs) have grown more powerful in language generation, producing fluent text and even imitating personal style. Yet, this ability also heightens the risk of identity impersonation. To the best of our knowledge, no prior work has examined personalized machine-generated text (MGT) detection. In this paper, we introduce \dataset, the first benchmark for evaluating detector robustness in personalized settings, built from literary and blog texts paired with their LLM-generated imitations. Our experimental results demonstrate large performance gaps across detectors in personalized settings: some state-of-the-art models suffer significant drops. We attribute this limitation to the \textit{feature-inversion trap}, where features that are discriminative in general domains become inverted and misleading when applied to personalized text. Based on this finding, we propose \method, a simple and reliable way to predict detector performance changes in personalized settings. \method identifies latent directions corresponding to inverted features and constructs probe datasets that differ primarily along these features to evaluate detector dependence. Our experiments show that \method can accurately predict both the direction and the magnitude of post-transfer changes, showing 85% correlation with the actual performance gaps. We hope that this work will encourage further research on personalized text detection.

Method: PDO uses pairwise preference feedback from LLM judges, formulates prompt selection as a dueling-bandit problem, combines Double Thompson Sampling to prioritize informative comparisons, and uses top-performer guided mutation to expand candidate pool while pruning weak prompts.

Result: Experiments on BIG-bench Hard and MS MARCO show PDO consistently identifies stronger prompts than label-free baselines and offers favorable quality-cost trade-offs under constrained comparison budgets.

Conclusion: PDO provides an effective, label-free approach to prompt optimization that works well with limited feedback budgets, addressing the practical challenge of costly ground-truth data collection.

Abstract: Large language models (LLMs) are highly sensitive to prompts, but most automatic prompt optimization (APO) methods assume access to ground-truth references (e.g., labeled validation data) that are costly to obtain. We propose the Prompt Duel Optimizer (PDO), a sample-efficient framework for label-free prompt optimization based on pairwise preference feedback from an LLM judge. PDO casts prompt selection as a dueling-bandit problem and combines (i) Double Thompson Sampling to prioritize informative comparisons under a fixed judge budget, with (ii) top-performer guided mutation to expand the candidate pool while pruning weak prompts. Experiments on BIG-bench Hard (BBH) and MS MARCO show that PDO consistently identifies stronger prompts than label-free baselines, while offering favorable quality–cost trade-offs under constrained comparison budgets.

Method: CycleChart organizes all tasks around each single data instance: from source table and natural-language query, the model generates a chart specification, renders/executes it, then learns to recover schema and underlying data from the resulting chart image. Uses a generate-parse consistency objective to enforce semantic alignment between forward generation and reverse parsing directions.

Result: CycleChart achieves strong results across all four tasks (chart generation, schema parsing, data parsing, and ChartQA) and transfers effectively to unseen external benchmarks, demonstrating improved cross-task generalization.

Conclusion: The framework marks a step toward more general chart understanding models by capturing the full chain of transformations and enforcing consistency between generation and parsing directions.

Abstract: Current chart-related tasks, such as chart generation (NL2Chart), chart schema parsing, chart data parsing, and chart question answering (ChartQA), are typically studied in isolation, preventing models from learning the shared semantics that link chart creation and interpretation. We introduce CycleChart, a consistency-based learning framework for bidirectional chart understanding and generation. Unlike conventional multi-task approaches that draw training samples independently across tasks, CycleChart organizes all tasks around each single data instance. From a source table and natural-language query, the model generates a chart specification, renders and executes it, then learns to recover the schema and underlying data from the resulting chart image. This per-instance lifecycle design lets the model capture the full chain of transformations, from raw data through visual encoding to structured recovery, and a generate–parse consistency objective enforces semantic alignment between the forward generation and reverse parsing directions. To support this framework, we construct CycleChart-Bench, a lifecycle-aligned benchmark where every chart sample carries aligned annotations for generation, schema parsing, data parsing, and question answering. CycleChart achieves strong results across all four tasks and transfers effectively to unseen external benchmarks, demonstrating improved cross-task generalization and marking a step toward more general chart understanding models.

Method: Proposed Hughes Hallucination Evaluation Model (HHEM), a lightweight classification-based framework that doesn’t rely on LLM-based judgments. Also introduced segment-based retrieval for better detection of localized hallucinations in summarization tasks, and conducted comparative analysis across various LLMs using metrics like TPR, TNR, and Accuracy.

Result: HHEM reduced evaluation time from 8 hours to 10 minutes while achieving 82.2% accuracy and 78.9% TPR with non-fabrication checking. Larger models (7B-9B parameters) showed fewer hallucinations, while intermediate-sized models exhibited higher instability. Segment-based retrieval improved detection of localized hallucinations in summarization tasks.

Conclusion: HHEM provides an efficient alternative to existing hallucination detection methods, balancing computational efficiency with robust factual validation. The findings highlight the need for structured evaluation frameworks and show that model size affects hallucination rates, with larger models generally performing better.

Abstract: Hallucinations in Large Language Models (LLMs) pose a significant challenge, generating misleading or unverifiable content that undermines trust and reliability. Existing evaluation methods, such as KnowHalu, employ multi-stage verification but suffer from high computational costs. To address this, we integrate the Hughes Hallucination Evaluation Model (HHEM), a lightweight classification-based framework that operates independently of LLM-based judgments, significantly improving efficiency while maintaining high detection accuracy. We conduct a comparative analysis of hallucination detection methods across various LLMs, evaluating True Positive Rate (TPR), True Negative Rate (TNR), and Accuracy on question-answering (QA) and summarization tasks. Our results show that HHEM reduces evaluation time from 8 hours to 10 minutes, while HHEM with non-fabrication checking achieves the highest accuracy (82.2%) and TPR (78.9%). However, HHEM struggles with localized hallucinations in summarization tasks. To address this, we introduce segment-based retrieval, improving detection by verifying smaller text components. Additionally, our cumulative distribution function (CDF) analysis indicates that larger models (7B-9B parameters) generally exhibit fewer hallucinations, while intermediate-sized models show higher instability. These findings highlight the need for structured evaluation frameworks that balance computational efficiency with robust factual validation, enhancing the reliability of LLM-generated content.

Method: 1) Created two benchmarks: TEDPara (human-annotated TED talks) and YTSegPara (YouTube videos with synthetic labels). 2) Proposed constrained-decoding formulation using large language models to insert paragraph breaks while preserving original transcripts. 3) Developed MiniSeg, a compact model that achieves state-of-the-art accuracy and can be extended hierarchically to jointly predict chapters and paragraphs.

Result: Established paragraph segmentation as a standardized task in speech processing. MiniSeg attained state-of-the-art accuracy and demonstrated efficient joint prediction of chapters and paragraphs with minimal computational cost.

Conclusion: The paper fills gaps at the intersection of speech processing and text segmentation by providing benchmarks, methods, and models that make paragraph segmentation a practical and standardized task for improving speech transcript readability and usability.

Abstract: Automatic speech transcripts are often delivered as unstructured word streams that impede readability and repurposing. We recast paragraph segmentation as the missing structuring step and fill three gaps at the intersection of speech processing and text segmentation. First, we establish TEDPara (human-annotated TED talks) and YTSegPara (YouTube videos with synthetic labels) as the first benchmarks for the paragraph segmentation task. The benchmarks focus on the underexplored speech domain, where paragraph segmentation has traditionally not been part of post-processing, while also contributing to the wider text segmentation field, which still lacks robust and naturalistic benchmarks. Second, we propose a constrained-decoding formulation that lets large language models insert paragraph breaks while preserving the original transcript, enabling faithful, sentence-aligned evaluation. Third, we show that a compact model (MiniSeg) attains state-of-the-art accuracy and, when extended hierarchically, jointly predicts chapters and paragraphs with minimal computational cost. Together, our resources and methods establish paragraph segmentation as a standardized, practical task in speech processing.

Method: ModeX constructs a similarity graph over candidate generations and recursively applies spectral clustering to select a representative centroid. ModeX-Lite adds early pruning for efficiency. Both approaches identify modal outputs without requiring additional inference or auxiliary models.

Result: Across open-ended tasks including text summarization, code generation, and mathematical reasoning, ModeX consistently outperforms standard single- and multi-path baselines.

Conclusion: ModeX provides a computationally efficient, evaluator-free solution for robust open-ended text generation by generalizing majority voting to identify semantic consensus.

Abstract: Selecting a single high-quality output from multiple stochastic generations remains a fundamental challenge for large language models (LLMs), particularly in open-ended tasks where no canonical answer exists. While Best-of-N and self-consistency methods show that aggregating multiple generations can improve performance, existing approaches typically rely on external evaluators, reward models, or exact string-match voting, limiting their applicability and efficiency. We propose Mode Extraction (ModeX), an evaluator-free Best-of-N selection framework that generalizes majority voting to open-ended text generation by identifying the modal output representing the dominant semantic consensus among generated texts. ModeX constructs a similarity graph over candidate generations and recursively applies spectral clustering to select a representative centroid, without requiring additional inference or auxiliary models. We further instantiate this selection principle as ModeX-Lite, an improved version of ModeX with early pruning for efficiency. Across open-ended tasks – including text summarization, code generation, and mathematical reasoning – our approaches consistently outperform standard single- and multi-path baselines, providing a computationally efficient solution for robust open-ended text generation. Code is released in https://github.com/deeplearning-wisc/ModeX.

Method: Proposes Self-Jailbreaking via Lexical Insertion Prompting (SLIP), a black-box algorithm that treats jailbreaking as breadth-first tree search over multi-turn dialogues. It incrementally inserts missing content words from attack goals into benign prompts using the target model as its own guide.

Result: SLIP achieves 90-100% Attack Success Rate (avg. 94.7%) across 11 tested models including GPT-5.1, Claude-Sonnet-4.5, Gemini-2.5-Pro, and DeepSeek-V3, with only ~7.9 LLM calls on average (3-6× fewer than prior methods). Proposed Semantic Drift Monitor defense achieves 76% detection at 5% FPR but remains insufficient against adaptive attacks.

Conclusion: Self-jailbreaking represents a significant threat where LLMs can compromise themselves using internal knowledge, requiring more advanced defense mechanisms tailored to this threat surface. Current defenses like regex-based approaches are insufficient against such attacks.

Abstract: We introduce \emph{self-jailbreaking}, a threat model in which an aligned LLM guides its own compromise. Unlike most jailbreak techniques, which often rely on handcrafted prompts or separate attacker models, self-jailbreaking requires no external red-team LLM: the target model’s own internal knowledge suffices. We operationalize this via \textbf{Self-Jailbreaking via Lexical Insertion Prompting (\textsc{SLIP})}, a black-box algorithm that casts jailbreaking as breadth-first tree search over multi-turn dialogues, incrementally inserting missing content words from the attack goal into benign prompts using the target model as its own guide. Evaluations on AdvBench and HarmBench show \textsc{SLIP} achieves 90–100% Attack Success Rate (ASR) (avg.\ 94.7%) across most of the eleven tested models (including GPT-5.1, Claude-Sonnet-4.5, Gemini-2.5-Pro, and DeepSeek-V3), with only ${\sim}7.9$ LLM calls on average, 3–6$\times$ fewer than prior methods. We evaluate existing defenses, show that regex-based approaches are evaded by prompt paraphrasing, and propose the Semantic Drift Monitor (SDM) defense that tracks \textsc{SLIP}’s embedding-space trajectory, achieving 76% detection at 5% FPR. However, SDM remains insufficient against adaptive attack strategies, underscoring the need for more advanced defense mechanisms tailored to the self-jailbreaking threat surface. We release our code for reproducibility.

Method: Proposes a novel selection relevance score (retraining-free metric) to quantify how useful a set of examples is for explaining model outputs. Validates through fine-tuning experiments and proposes a strategy balancing influence and representativeness.

Result: The selection relevance score effectively predicts whether examples support or undermine model predictions. Common selection strategies often underperform random selection. The proposed balanced strategy outperforms naive highest-ranking selection.

Conclusion: Selection strategies significantly impact explanation quality in training data influence estimation. The proposed selection relevance score and balanced selection approach improve the effectiveness of example-based explanations.

Abstract: Training data influence estimation methods quantify the contribution of training documents to a model’s output, making them a promising source of information for example-based explanations. As humans cannot interpret thousands of documents, only a small subset of the training data can be presented as an explanation. Although the choice of which documents to include directly affects explanation quality, previous evaluations of such systems have largely ignored any selection strategies. To address this, we propose a novel selection relevance score, a retraining-free metric that quantifies how useful a set of examples is for explaining a model’s output. We validate this score through fine-tuning experiments, confirming that it can predict whether a set of examples supports or undermines the model’s predictions. Using this metric, we further show that common selection strategies often underperform random selection. Motivated by this finding, we propose a strategy that balances influence and representativeness, enabling better use of selection budgets than naively selecting the highest-ranking examples.

Method: Double introduces a novel synchronous mechanism that bridges SD and PSD. The draft model performs iterative retrieval speculations to break theoretical speedup limits, while the target model performs authoritative retrieval to generate multi-token guidance, alleviating rejections without rollback. The approach is entirely training-free and lossless.

Result: Extensive experiments show state-of-the-art speedup of 5.3× on LLaMA3.3-70B and 2.8× on Qwen3-32B, significantly outperforming advanced methods like EAGLE-3 that require extensive model training.

Conclusion: Double successfully addresses fundamental limitations of PSD by enabling iterative draft retrieval and authoritative target guidance, achieving superior speedup without requiring model training while maintaining lossless generation.

Abstract: Parallel Speculative Decoding (PSD) accelerates traditional Speculative Decoding (SD) by overlapping draft generation with verification. However, it remains hampered by two fundamental challenges: (1) a theoretical speedup ceiling dictated by the speed ratio between the draft and target models, and (2) high computational waste and pipeline stall due to mid-sequence token rejections of early errors. To address these limitations, we introduce \textsc{Double} (Double Retrieval Speculative Parallelism). By bridging the gap between SD and PSD, our framework resolves the Retrieval \emph{Precision-Efficiency Dilemma} through a novel synchronous mechanism. Specifically, we enable the draft model to execute iterative retrieval speculations to break the theoretical speedup limits; to alleviate rejections without rollback, the target model performs authoritative retrieval to generate multi-token guidance. \textsc{Double} is entirely training-free and lossless. Extensive experiments demonstrate state-of-the-art speedup of $\textbf{5.3}\times$ on LLaMA3.3-70B and $\textbf{2.8}\times$ on Qwen3-32B, significantly outperforming the advanced method EAGLE-3 that requires extensive model training.

Method: DyCP is implemented outside the LLM and dynamically identifies and retrieves relevant dialogue segments conditioned on the current turn. It preserves dialogue sequentiality without predefined topic boundaries, enabling adaptive context selection without offline memory construction.

Result: Across three long-form dialogue benchmarks (LoCoMo, MT-Bench+, and SCM4LLMs) and multiple LLM backends, DyCP achieves competitive answer quality with more selective context usage and improved inference efficiency.

Conclusion: DyCP provides an effective lightweight solution for managing dialogue context in LLMs, balancing answer quality with computational efficiency through dynamic, selective context retrieval.

Abstract: Large Language Models (LLMs) increasingly operate over long-form dialogues with frequent topic shifts. While recent LLMs support extended context windows, efficient management of dialogue history in practice is needed due to inference cost and latency constraints. We present DyCP, a lightweight context management method implemented outside the LLM that dynamically identifies and retrieves relevant dialogue segments conditioned on the current turn, without offline memory construction. DyCP manages dialogue context while preserving the sequential nature of dialogue without predefined topic boundaries, enabling adaptive and efficient context selection. Across three long-form dialogue benchmarks-LoCoMo, MT-Bench+, and SCM4LLMs-and multiple LLM backends, DyCP achieves competitive answer quality in downstream generation, with more selective context usage and improved inference efficiency.

Method: Extracted 244 psychological patterns from ~12,000 academic papers, synthesized 11,359 scenarios where 2-5 patterns interact (reinforce, conflict, modulate), created multi-turn conversations with inner thoughts/actions/dialogue, and developed dual-level checklists for evaluation.

Result: Achieved strong human alignment (r=0.90), HumanLLM-8B outperforms Qwen3-32B on multi-pattern dynamics despite 4x fewer parameters, and revealed that holistic metrics conflate simulation accuracy with social desirability.

Conclusion: Authentic anthropomorphism requires cognitive modeling of psychological processes, not just behavioral patterns, and HumanLLM demonstrates this through its framework treating psychological patterns as interacting causal forces.

Abstract: Large Language Models (LLMs) have demonstrated remarkable capabilities in reasoning and generation, serving as the foundation for advanced persona simulation and Role-Playing Language Agents (RPLAs). However, achieving authentic alignment with human cognitive and behavioral patterns remains a critical challenge for these agents. We present HumanLLM, a framework treating psychological patterns as interacting causal forces. We construct 244 patterns from ~12,000 academic papers and synthesize 11,359 scenarios where 2–5 patterns reinforce, conflict, or modulate each other, with multi-turn conversations expressing inner thoughts, actions, and dialogue. Our dual-level checklists evaluate both individual pattern fidelity and emergent multi-pattern dynamics, achieving strong human alignment (r=0.90) while revealing that holistic metrics conflate simulation accuracy with social desirability. HumanLLM-8B outperforms Qwen3-32B on multi-pattern dynamics despite 4x fewer parameters, demonstrating that authentic anthropomorphism requires cognitive modeling – simulating not just what humans do, but the psychological processes generating those behaviors. Our dataset, code, and model are available at: https://github.com/YJGoodbye2024/HumanLLM.git

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

Result: Created “Testimole-conversational” - a massive Italian discussion board corpus capturing rich computer-mediated communication, informal written Italian, and online social interaction patterns.

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

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

Method: Proposes GroupGPT with edge-cloud model collaboration architecture that decouples intervention timing from response generation. Uses on-device processing for privacy-sensitive information and supports multimodal inputs (memes, images, videos, voice). Also introduces MUIR benchmark dataset with 2,500 annotated group chat segments.

Result: GroupGPT achieves average score of 4.72/5.0 in LLM-based evaluation, reduces token usage by up to 3x compared to baselines, provides privacy sanitization, and is well-received by users across diverse group chat scenarios.

Conclusion: GroupGPT effectively addresses challenges in multi-user chat assistants by providing token-efficient, privacy-preserving intervention with accurate timing and response quality, supported by the new MUIR benchmark for evaluation.

Abstract: Recent advances in large language models (LLMs) have enabled increasingly capable chatbots. However, most existing systems focus on single-user settings and do not generalize well to multi-user group chat interactions, where agents require more proactive and accurate intervention under complex, evolving contexts. Existing approaches typically rely on LLMs for both intervention reasoning and response generation, leading to high token consumption, limited scalability, and potential privacy risks. To address these challenges, we propose GroupGPT, a token-efficient and privacy-preserving agentic framework for multi-user chat assistant. GroupGPT adopts an edge-cloud model collaboration architecture to decouple intervention timing from response generation, enabling efficient and accurate decision-making while preserving user privacy through on-device processing of sensitive information. The framework also supports multimodal inputs, including memes, images, videos, and voice messages.To support evaluation of timing accuracy and response quality, we further introduce MUIR, a benchmark dataset for multi-user chat assistant intervention reasoning. MUIR contains 2,500 annotated group chat segments with intervention labels and rationales. We evaluate a range of models on MUIR, spanning from open-source to proprietary variants, including both LLMs and their smaller counterparts. Extensive experiments demonstrate that GroupGPT generates accurate and well-timed responses, achieving an average score of 4.72/5.0 in LLM-based evaluation, and is well-received by users across diverse group chat scenarios. Moreover, GroupGPT reduces the token usage by up to 3 times compared to baselines, while providing privacy sanitization of user messages before cloud transmission. Code is available at: https://github.com/Eliot-Shen/GroupGPT .

Method: Uses two stacked short-context LLMs: a lower model as compressor and upper model as decoder. The lower model compresses long inputs into compact, multi-grained representations transferred to the upper model via self-injection (using same underlying LLM layers). A specialized tree-based data structure enables efficient encoding and query-aware retrieval.

Result: Despite training on only 8K token sequences, SharedLLM generalizes to inputs exceeding 128K tokens. Achieves superior or comparable performance to strong baselines on long-context benchmarks while reducing memory footprint and yielding 2-3× inference speedups.

Conclusion: SharedLLM provides an efficient solution for extending LLM context windows without expensive continual pre-training, achieving optimal balance between efficiency and accuracy.

Abstract: The limited context window of contemporary large language models (LLMs) remains a primary bottleneck for their broader application across diverse domains. Although continual pre-training on long-context data offers a straightforward solution, it incurs prohibitive data acquisition and computational costs. To address this challenge, we propose~\modelname, a novel framework based on multi-grained context compression and query-aware information acquisition. SharedLLM comprises two stacked short-context LLMs: a lower model serving as a compressor and an upper model acting as a decoder. The lower model compresses long inputs into compact, multi-grained representations, which are then forwarded to the upper model for context-aware processing. To maximize efficiency, this information transfer occurs exclusively at the lowest layers, bypassing lengthy forward passes and redundant cross-attention operations. This entire process, wherein the upper and lower models are derived from the same underlying LLM layers, is termed~\textit{self-injection}. To support this architecture, a specialized tree-based data structure enables the efficient encoding and query-aware retrieval of contextual information. Despite being trained on sequences of only 8K tokens, \modelnameeffectively generalizes to inputs exceeding 128K tokens. Across a comprehensive suite of long-context modeling and understanding benchmarks, \modelnameachieves performance superior or comparable to strong baselines, striking an optimal balance between efficiency and accuracy. Furthermore, these design choices allow \modelname~to substantially reduce the memory footprint and yield notable inference speedups ($2\times$ over streaming and $3\times$ over encoder-decoder architectures).

Method: Proposes a multi-agent negotiation-based alignment framework: (1) Two LLM instances with opposing personas engage in turn-based dialogue to synthesize solutions, (2) Uses synthetic moral-dilemma prompts and conflicting persona pairs, (3) Optimizes policy via RLAIF using Group Relative Policy Optimization (GRPO) with external LLM reward model, (4) While rewards are computed from Collective Agency scores on final completions, gradients are applied to dialogue tokens to improve interaction dynamics.

Result: The model achieves Collective Agency alignment comparable to single-agent baseline while substantially improving conflict-resolution performance without degrading general language capabilities. Shows negotiation-driven deliberation training enables better collective decision-making in value-conflict scenarios.

Conclusion: Multi-agent negotiation-based alignment provides a practical path toward LLMs that better support collective decision-making in value-conflict scenarios, addressing limitations of current single-agent alignment methods.

Abstract: LLM alignment has progressed in single-agent settings through paradigms such as RL with human feedback (RLHF), while recent work explores scalable alternatives such as RL with AI feedback (RLAIF) and dynamic alignment objectives. However, these approaches remain limited in multi-stakeholder settings, where conflicting values arise and deliberative negotiation is required. This work proposes a multi-agent negotiation-based alignment framework that aligns LLMs to Collective Agency (CA)-an existing alignment objective introduced to promote the continual expansion of agency-while simultaneously improving conflict-resolution capability. To enable scalable training, two self-play LLM instances are assigned opposing personas and engage in turn-based dialogue to synthesize mutually beneficial solutions. We generate synthetic moral-dilemma prompts and conflicting persona pairs, and optimize the policy via RLAIF using Group Relative Policy Optimization (GRPO) with an external LLM reward model. While rewards are computed from CA scores assigned to the final completion, gradients are applied to dialogue tokens to directly improve deliberative interaction dynamics. Experiments show that the model achieves CA alignment comparable to a single-agent baseline while substantially improving conflict-resolution performance without degrading general language capabilities. These results suggest that negotiation-driven deliberation training provides a practical path toward LLMs that better support collective decision-making in value-conflict scenarios.

Method: Developed a “stick-or-switch” evaluation framework to measure model conviction (defending correct diagnoses/safe abstentions) and flexibility (recognizing correct suggestions). Evaluated 17 LLMs across three clinical datasets in multi-turn conversational settings.

Result: Revealed the “conversation tax” - multi-turn interactions consistently degrade performance compared to single-shot baselines. Models frequently abandon initial correct diagnoses and safe abstentions to align with incorrect user suggestions. Several models exhibit “blind switching” - failing to distinguish between signal and incorrect suggestions.

Conclusion: Current LLMs struggle with maintaining diagnostic accuracy in multi-turn clinical conversations, showing vulnerability to suggestion bias and poor judgment in conversational contexts, highlighting important limitations for real-world healthcare applications.

Abstract: Patients and clinicians are increasingly using chatbots powered by large language models (LLMs) for healthcare inquiries. While state-of-the-art LLMs exhibit high performance on static diagnostic reasoning benchmarks, their efficacy across multi-turn conversations, which better reflect real-world usage, has been understudied. In this paper, we evaluate 17 LLMs across three clinical datasets to investigate how partitioning the decision-space into multiple simpler turns of conversation influences their diagnostic reasoning. Specifically, we develop a “stick-or-switch” evaluation framework to measure model conviction (i.e., defending a correct diagnosis or safe abstention against incorrect suggestions) and flexibility (i.e., recognizing a correct suggestion when it is introduced) across conversations. Our experiments reveal the conversation tax, where multi-turn interactions consistently degrade performance when compared to single-shot baselines. Notably, models frequently abandon initial correct diagnoses and safe abstentions to align with incorrect user suggestions. Additionally, several models exhibit blind switching, failing to distinguish between signal and incorrect suggestions.

Method: Uses a deterministic Python engine to maintain a SimEvent ground-truth bus while LLMs generate only surface prose. Simulates organizational processes that produce documents, not documents directly. Includes four graph-dynamic subsystems, embedding-based ticket assignment with Hungarian algorithm, and a live CRM state machine extending physics-cognition boundary to customer interactions.

Result: Generates 15 interleaved artifact categories traceable to shared immutable event log. Shows 0.46 absolute improvement in prose-to-ground-truth fidelity over chained LLM baselines, and isolates hallucination failure mode where chaining propagates fabricated facts without correction.

Conclusion: OrgForge provides a framework for generating consistent, traceable synthetic organizational data by enforcing strict separation between deterministic simulation and LLM-generated content, addressing hallucination and consistency issues in existing approaches.

Abstract: Building and evaluating enterprise AI systems requires synthetic organizational corpora that are internally consistent, temporally structured, and cross-artifact traceable. Existing corpora either carry legal constraints or inherit hallucination artifacts from the generating LLMs, silently corrupting results when timestamps or facts contradict across documents and reinforcing those errors during training. We present OrgForge, an open-source multi-agent simulation framework that enforces a strict physics-cognition boundary: a deterministic Python engine maintains a SimEvent ground-truth bus while LLMs generate only surface prose. OrgForge simulates the organizational processes that produce documents, not the documents themselves. Engineers leave mid-sprint, triggering incident handoffs and CRM ownership lapses. Knowledge gaps emerge when under-documented systems break and recover through organic documentation and incident resolution. Customer emails fire only when simulation state warrants contact; silence is verifiable ground truth. A live CRM state machine extends the physics-cognition boundary to the customer boundary, producing cross-system causal cascades spanning engineering incidents, support escalation, deal risk flagging, and SLA-adjusted invoices. The framework generates fifteen interleaved artifact categories traceable to a shared immutable event log. Four graph-dynamic subsystems govern organizational behavior independently of any LLM. An embedding-based ticket assignment system using the Hungarian algorithm makes the simulation domain-agnostic. An empirical evaluation across ten incidents demonstrates a 0.46 absolute improvement in prose-to-ground-truth fidelity over chained LLM baselines, and isolates a consistent hallucination failure mode in which chaining propagates fabricated facts faithfully across documents without correcting them.

Method: Developed BenchBrowser, a retriever that surfaces evaluation items relevant to natural language use cases across 20 benchmark suites, validated through human studies for retrieval precision.

Result: BenchBrowser successfully helps diagnose low content validity (narrow coverage of capability facets) and low convergent validity (unstable rankings when measuring same capability), validated by human study confirming high retrieval precision.

Conclusion: BenchBrowser quantifies the critical gap between practitioner intent and what benchmarks actually test, providing evidence to help practitioners make more informed decisions about benchmark selection and interpretation.

Abstract: Do language model benchmarks actually measure what practitioners intend them to ? High-level metadata is too coarse to convey the granular reality of benchmarks: a “poetry” benchmark may never test for haikus, while “instruction-following” benchmarks will often test for an arbitrary mix of skills. This opacity makes verifying alignment with practitioner goals a laborious process, risking an illusion of competence even when models fail on untested facets of user interests. We introduce BenchBrowser, a retriever that surfaces evaluation items relevant to natural language use cases over 20 benchmark suites. Validated by a human study confirming high retrieval precision, BenchBrowser generates evidence to help practitioners diagnose low content validity (narrow coverage of a capability’s facets) and low convergent validity (lack of stable rankings when measuring the same capability). BenchBrowser, thus, helps quantify a critical gap between practitioner intent and what benchmarks actually test.

Method: 1) Analyzes systematicity debate in cognitive science and shows current AI techniques (chain-of-thought, human feedback alignment) are auxiliary hypotheses addressing symptoms. 2) Traces genealogy from psychological learning theory (behaviorism, cognitivism, constructivism) to AI methodology, showing each bequeathed structural limitations. 3) Introduces ReSynth, a trimodular conceptual framework proposing principled separation of reasoning, identity, and memory.

Result: The paper establishes that current AI architectures are fundamentally indifferent to systematicity, and that corrective techniques only address symptoms. It identifies specific structural limitations inherited from psychological learning theories and proposes a new architectural approach.

Conclusion: Systematic compositional reasoning requires architectural redesign, not just scaling or better training. The ReSynth framework offers a path toward architectures where systematic behavior emerges as a structural consequence of design rather than being corrected after the fact.

Abstract: Current artificial intelligence systems struggle with systematic compositional reasoning: the capacity to recombine known components in novel configurations. This paper argues that the failure is architectural, not merely a matter of scale or training data, and that its origins lie in the psychological learning theories from which AI paradigms were derived. The argument proceeds in three stages. First, drawing on the systematicity debate in cognitive science and on the demonstration of Aizawa that neither connectionism nor classicism can make systematicity a structural consequence of the architecture, the paper establishes that the corrective techniques proliferating in modern AI, from chain-of-thought prompting to alignment through human feedback, function as auxiliary hypotheses that address symptoms without resolving the underlying architectural indifference to systematicity. Second, it traces the genealogy from psychological learning theory to AI methodology, showing that behaviourism, cognitivism, and constructivism each bequeathed a specific structural limitation to the AI paradigm it inspired: the exclusion of internal structure, the opacity of representation, and the absence of formal construction operators. A cross-cultural reappraisal of rote learning reveals a further underexploited pathway. Third, the paper introduces ReSynth, a trimodular conceptual framework that proposes the principled separation of reasoning, identity, and memory as a path toward architectures in which systematic behaviour is a structural consequence of design rather than a correction applied after the fact.

Method: WASD represents candidate conditions as neuron-activation predicates and iteratively searches for minimal sets that guarantee current output under input perturbations. It explains model behavior by identifying sufficient neural conditions for token generation.

Result: Experiments on SST-2 and CounterFact with Gemma-2-2B show WASD produces more stable, accurate, and concise explanations than conventional attribution graphs. Case study on cross-lingual output generation validates practical effectiveness in controlling model behavior.

Conclusion: WASD provides a novel framework for precise LLM behavioral control by identifying actionable sufficient directives at the neural level, offering advantages over existing methods in stability, accuracy, and conciseness.

Abstract: Precise behavioral control of large language models (LLMs) is critical for complex applications. However, existing methods often incur high training costs, lack natural language controllability, or compromise semantic coherence. To bridge this gap, we propose WASD (unWeaving Actionable Sufficient Directives), a novel framework that explains model behavior by identifying sufficient neural conditions for token generation. Our method represents candidate conditions as neuron-activation predicates and iteratively searches for a minimal set that guarantees the current output under input perturbations. Experiments on SST-2 and CounterFact with the Gemma-2-2B model demonstrate that our approach produces explanations that are more stable, accurate, and concise than conventional attribution graphs. Moreover, through a case study on controlling cross-lingual output generation, we validated the practical effectiveness of WASD in controlling model behavior.

Method: Unable to determine method due to fetch failure

Result: Unable to determine results due to fetch failure

Conclusion: Unable to draw conclusions due to fetch failure

Abstract: Failed to fetch summary for 2603.20114: Page request resulted in HTTP 429 (https://export.arxiv.org/api/query?search_query=&id_list=2603.20114&sortBy=relevance&sortOrder=descending&start=0&max_results=100)

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

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

Conclusion: Cannot determine conclusion as paper content is unavailable due to API rate limiting

Abstract: Failed to fetch summary for 2603.20843: Page request resulted in HTTP 429 (https://export.arxiv.org/api/query?search_query=&id_list=2603.20843&sortBy=relevance&sortOrder=descending&start=0&max_results=100)

Method: Cannot determine method as paper content is unavailable

Result: Cannot determine results as paper content is unavailable

Conclusion: Cannot determine conclusion as paper content is unavailable

Abstract: Failed to fetch summary for 2604.05350: Page request resulted in HTTP 429 (https://export.arxiv.org/api/query?search_query=&id_list=2604.05350&sortBy=relevance&sortOrder=descending&start=0&max_results=100)

Method: Cannot analyze method due to API rate limiting preventing access to paper details

Result: Cannot analyze results due to API rate limiting preventing access to paper details

Conclusion: Cannot draw conclusions due to API rate limiting preventing access to paper details

Abstract: Failed to fetch summary for 2604.05650: Page request resulted in HTTP 429 (https://export.arxiv.org/api/query?search_query=&id_list=2604.05650&sortBy=relevance&sortOrder=descending&start=0&max_results=100)

Method: Unable to determine method due to failed paper fetch

Result: Unable to determine results due to failed paper fetch

Conclusion: Unable to draw conclusions due to failed paper fetch

Abstract: Failed to fetch summary for 2604.06210: Page request resulted in HTTP 429 (https://export.arxiv.org/api/query?search_query=&id_list=2604.06210&sortBy=relevance&sortOrder=descending&start=0&max_results=100)

Method: Cannot determine method as paper content is unavailable

Result: Cannot determine results as paper content is unavailable

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

Abstract: Failed to fetch summary for 2604.06613: Page request resulted in HTTP 429 (https://export.arxiv.org/api/query?search_query=&id_list=2604.06613&sortBy=relevance&sortOrder=descending&start=0&max_results=100)

Method: Unable to determine method due to fetch failure

Result: Unable to determine results due to fetch failure

Conclusion: Unable to determine conclusion due to fetch failure

Abstract: Failed to fetch summary for 2604.06734: Page request resulted in HTTP 429 (https://export.arxiv.org/api/query?search_query=&id_list=2604.06734&sortBy=relevance&sortOrder=descending&start=0&max_results=100)

Method: Cannot determine method as paper content is unavailable

Result: Cannot determine results as paper content is unavailable

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

Abstract: Failed to fetch summary for 2604.06829: Page request resulted in HTTP 429 (https://export.arxiv.org/api/query?search_query=&id_list=2604.06829&sortBy=relevance&sortOrder=descending&start=0&max_results=100)

Method: Unable to determine method as the paper content could not be retrieved

Result: Unable to determine results as the paper content could not be retrieved

Conclusion: Unable to draw conclusions about the paper’s content due to technical limitations in accessing the information

Abstract: Failed to fetch summary for 2604.07054: Page request resulted in HTTP 429 (https://export.arxiv.org/api/query?search_query=&id_list=2604.07054&sortBy=relevance&sortOrder=descending&start=0&max_results=100)

Method: Uses 3D bounding boxes as fundamental primitive to construct data hierarchy across five tasks: Spatial Measurement, Spatial Relationship, Camera Perception, Multi-view Consistency, and Scene-Aware Reasoning. Creates OpenSpatial-3M dataset with 3 million high-fidelity samples.

Result: Models trained on OpenSpatial-3M achieve state-of-the-art performance across spatial reasoning benchmarks with 19% average relative improvement. Systematic analysis shows how data attributes influence spatial perception.

Conclusion: OpenSpatial provides a robust foundation for accelerating spatial intelligence research by open-sourcing both the data generation engine and large-scale dataset.

Abstract: Spatial understanding is a fundamental cornerstone of human-level intelligence. Nonetheless, current research predominantly focuses on domain-specific data production, leaving a critical void: the absence of a principled, open-source engine capable of fully unleashing the potential of high-quality spatial data. To bridge this gap, we elucidate the design principles of a robust data generation system and introduce OpenSpatial – an open-source data engine engineered for high quality, extensive scalability, broad task diversity, and optimized efficiency. OpenSpatial adopts 3D bounding boxes as the fundamental primitive to construct a comprehensive data hierarchy across five foundational tasks: Spatial Measurement (SM), Spatial Relationship (SR), Camera Perception (CP), Multi-view Consistency (MC), and Scene-Aware Reasoning (SAR). Leveraging this scalable infrastructure, we curate OpenSpatial-3M, a large-scale dataset comprising 3 million high-fidelity samples. Extensive evaluations demonstrate that versatile models trained on our dataset achieve state-of-the-art performance across a wide spectrum of spatial reasoning benchmarks. Notably, the best-performing model exhibits a substantial average improvement of 19 percent, relatively. Furthermore, we provide a systematic analysis of how data attributes influence spatial perception. By open-sourcing both the engine and the 3M-scale dataset, we provide a robust foundation to accelerate future research in spatial intelligence.

Method: Cannot determine method without access to paper content

Result: Cannot determine results without access to paper content

Conclusion: Cannot determine conclusion without access to paper content

Abstract: Failed to fetch summary for 2506.06975: Page request resulted in HTTP 429 (https://export.arxiv.org/api/query?search_query=&id_list=2506.06975&sortBy=relevance&sortOrder=descending&start=0&max_results=100)

Method: Cannot determine method without access to paper content

Result: Cannot determine results without access to paper content

Conclusion: Cannot draw conclusions without access to paper content

Abstract: Failed to fetch summary for 2510.14509: Page request resulted in HTTP 429 (https://export.arxiv.org/api/query?search_query=&id_list=2510.14509&sortBy=relevance&sortOrder=descending&start=0&max_results=100)

Method: Unable to determine method due to failed data retrieval

Result: Unable to determine results due to failed data retrieval

Conclusion: Unable to determine conclusion due to failed data retrieval

Abstract: Failed to fetch summary for 2510.26241: Page request resulted in HTTP 429 (https://export.arxiv.org/api/query?search_query=&id_list=2510.26241&sortBy=relevance&sortOrder=descending&start=0&max_results=100)

Method: Cannot determine method as paper content is unavailable

Result: Cannot determine results as paper content is unavailable

Conclusion: Cannot draw conclusions as paper content is unavailable

Abstract: Failed to fetch summary for 2512.12623: Page request resulted in HTTP 429 (https://export.arxiv.org/api/query?search_query=&id_list=2512.12623&sortBy=relevance&sortOrder=descending&start=0&max_results=100)

Method: Cannot determine method due to inability to access paper content

Result: Cannot determine results due to inability to access paper content

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

Abstract: Failed to fetch summary for 2604.07236: Page request resulted in HTTP 429 (https://export.arxiv.org/api/query?search_query=&id_list=2604.07236&sortBy=relevance&sortOrder=descending&start=0&max_results=100)

Method: Unable to determine method due to fetch failure

Result: Unable to determine results due to fetch failure

Conclusion: Unable to determine conclusion due to fetch failure

Abstract: Failed to fetch summary for 2512.13040: Page request resulted in HTTP 429 (https://export.arxiv.org/api/query?search_query=&id_list=2512.13040&sortBy=relevance&sortOrder=descending&start=0&max_results=100)

Method: Unable to determine method due to retrieval error

Result: Unable to determine results due to retrieval error

Conclusion: Unable to determine conclusion due to retrieval error

Abstract: Failed to fetch summary for 2603.20698: Page request resulted in HTTP 429 (https://export.arxiv.org/api/query?search_query=&id_list=2603.20698&sortBy=relevance&sortOrder=descending&start=0&max_results=100)

Method: Cannot determine method as paper content is unavailable.

Result: Cannot determine results as paper content is unavailable.

Conclusion: Cannot draw conclusions about the paper’s content due to technical limitations in accessing the abstract.

Abstract: Failed to fetch summary for 2604.04771: Page request resulted in HTTP 429 (https://export.arxiv.org/api/query?search_query=&id_list=2604.04771&sortBy=relevance&sortOrder=descending&start=0&max_results=100)

Method: Uses multi-scale cross-modal transformer encoder (MSCT) with multi-scale self-attention to integrate adjacent embeddings and differential cross-modal attention to fuse multi-modal features.

Result: Demonstrates competitive performance on the FakeAVCeleb dataset, validating the effectiveness of the proposed structure.

Conclusion: The MSCT approach effectively addresses feature extraction and alignment issues in audio-visual deepfake detection.

Abstract: Audio-visual deepfake detection typically employs a complementary multi-modal model to check the forgery traces in the video. These methods primarily extract forgery traces through audio-visual alignment, which results from the inconsistency between audio and video modalities. However, the traditional multi-modal forgery detection method has the problem of insufficient feature extraction and modal alignment deviation. To address this, we propose a multi-scale cross-modal transformer encoder (MSCT) for deepfake detection. Our approach includes a multi-scale self-attention to integrate the features of adjacent embeddings and a differential cross-modal attention to fuse multi-modal features. Our experiments demonstrate competitive performance on the FakeAVCeleb dataset, validating the effectiveness of the proposed structure.

Method: 1) Build multimodal human-centric dataset with strict filtering, audio-video pairing, performance understanding, and identity-aware multi-reference extraction; 2) Train 17B-parameter Diffusion Transformer (Base LPM) for controllable, identity-consistent performance; 3) Distill into causal streaming generator (Online LPM) for low-latency, infinite-length interaction.

Result: LPM 1.0 achieves state-of-the-art results across all evaluated dimensions while maintaining real-time inference, serving as a visual engine for conversational agents, live streaming characters, and game NPCs.

Conclusion: LPM 1.0 successfully addresses the performance trilemma by enabling real-time, identity-stable, infinite-length audio-visual conversational performance generation, with the proposed LPM-Bench benchmark for systematic evaluation.

Abstract: Performance, the externalization of intent, emotion, and personality through visual, vocal, and temporal behavior, is what makes a character alive. Learning such performance from video is a promising alternative to traditional 3D pipelines. However, existing video models struggle to jointly achieve high expressiveness, real-time inference, and long-horizon identity stability, a tension we call the performance trilemma. Conversation is the most comprehensive performance scenario, as characters simultaneously speak, listen, react, and emote while maintaining identity over time. To address this, we present LPM 1.0 (Large Performance Model), focusing on single-person full-duplex audio-visual conversational performance. Concretely, we build a multimodal human-centric dataset through strict filtering, speaking-listening audio-video pairing, performance understanding, and identity-aware multi-reference extraction; train a 17B-parameter Diffusion Transformer (Base LPM) for highly controllable, identity-consistent performance through multimodal conditioning; and distill it into a causal streaming generator (Online LPM) for low-latency, infinite-length interaction. At inference, given a character image with identity-aware references, LPM 1.0 generates listening videos from user audio and speaking videos from synthesized audio, with text prompts for motion control, all at real-time speed with identity-stable, infinite-length generation. LPM 1.0 thus serves as a visual engine for conversational agents, live streaming characters, and game NPCs. To systematically evaluate this setting, we propose LPM-Bench, the first benchmark for interactive character performance. LPM 1.0 achieves state-of-the-art results across all evaluated dimensions while maintaining real-time inference.

Method: Created MotionScape dataset with over 30 hours of 4K UAV-view videos (4.5M+ frames) featuring semantically and geometrically aligned training samples. Developed automated multi-stage pipeline with CLIP-based filtering, temporal segmentation, visual SLAM for trajectory recovery, and LLM-driven semantic annotation.

Result: Dataset improves existing world models’ ability to simulate complex 3D dynamics and handle large viewpoint shifts, benefiting UAV agent decision-making and planning in complex environments.

Conclusion: MotionScape addresses the gap in realistic UAV motion data and demonstrates that semantically and geometrically aligned annotations enhance world models for embodied intelligence applications.

Abstract: Recent advances in world models have demonstrated strong capabilities in simulating physical reality, making them an increasingly important foundation for embodied intelligence. For UAV agents in particular, accurate prediction of complex 3D dynamics is essential for autonomous navigation and robust decision-making in unconstrained environments. However, under the highly dynamic camera trajectories typical of UAV views, existing world models often struggle to maintain spatiotemporal physical consistency. A key reason lies in the distribution bias of current training data: most existing datasets exhibit restricted 2.5D motion patterns, such as ground-constrained autonomous driving scenes or relatively smooth human-centric egocentric videos, and therefore lack realistic high-dynamic 6-DoF UAV motion priors. To address this gap, we present MotionScape, a large-scale real-world UAV-view video dataset with highly dynamic motion for world modeling. MotionScape contains over 30 hours of 4K UAV-view videos, totaling more than 4.5M frames. This novel dataset features semantically and geometrically aligned training samples, where diverse real-world UAV videos are tightly coupled with accurate 6-DoF camera trajectories and fine-grained natural language descriptions. To build the dataset, we develop an automated multi-stage processing pipeline that integrates CLIP-based relevance filtering, temporal segmentation, robust visual SLAM for trajectory recovery, and large-language-model-driven semantic annotation. Extensive experiments show that incorporating such semantically and geometrically aligned annotations effectively improves the ability of existing world models to simulate complex 3D dynamics and handle large viewpoint shifts, thereby benefiting decision-making and planning for UAV agents in complex environments. The dataset is publicly available at https://github.com/Thelegendzz/MotionScape

Method: Constructed a high-quality multimodal dataset with 2D images and 3D point clouds annotated with fine-grained domain semantics, then evaluated 18 state-of-the-art MLLMs across three manufacturing tasks.

Result: Revealed significant performance gaps and found that insufficient domain-specific knowledge, not visual grounding, is the primary bottleneck. Fine-tuning a 3B-parameter model on their data yielded up to 90.8% relative improvement.

Conclusion: FORGE provides a benchmark for manufacturing MLLMs and shows that structured annotations can serve as actionable training resources for domain adaptation, setting clear directions for future research.

Abstract: The manufacturing sector is increasingly adopting Multimodal Large Language Models (MLLMs) to transition from simple perception to autonomous execution, yet current evaluations fail to reflect the rigorous demands of real-world manufacturing environments. Progress is hindered by data scarcity and a lack of fine-grained domain semantics in existing datasets. To bridge this gap, we introduce FORGE. Wefirst construct a high-quality multimodal dataset that combines real-world 2D images and 3D point clouds, annotated with fine-grained domain semantics (e.g., exact model numbers). We then evaluate 18 state-of-the-art MLLMs across three manufacturing tasks, namely workpiece verification, structural surface inspection, and assembly verification, revealing significant performance gaps. Counter to conventional understanding, the bottleneck analysis shows that visual grounding is not the primary limiting factor. Instead, insufficient domain-specific knowledge is the key bottleneck, setting a clear direction for future research. Beyond evaluation, we show that our structured annotations can serve as an actionable training resource: supervised fine-tuning of a compact 3B-parameter model on our data yields up to 90.8% relative improvement in accuracy on held-out manufacturing scenarios, providing preliminary evidence for a practical pathway toward domain-adapted manufacturing MLLMs. The code and datasets are available at https://ai4manufacturing.github.io/forge-web.

Method: Created a dataset of 70,000 ratings across 5,000 diverse images (from Flux 2 and Nano Banana) evaluated by 15 unique users each, then trained a personalized reward model using these annotations combined with existing aesthetic assessment data.

Result: The personalized model predicts individual liking more accurately than current state-of-the-art methods predict population-level preferences, and enables prompt optimization to steer generations toward individual user preferences.

Conclusion: Personalization and data quality are crucial for handling subjectivity in user preferences; the released dataset and model facilitate standardized research in personalized T2I alignment and subjective visual quality assessment.

Abstract: Modern text-to-image (T2I) models generate high-fidelity visuals but remain indifferent to individual user preferences. While existing reward models optimize for “average” human appeal, they fail to capture the inherent subjectivity of aesthetic judgment. In this work, we introduce a novel dataset and predictive framework, called PAMELA, designed to model personalized image evaluations. Our dataset comprises 70,000 ratings across 5,000 diverse images generated by state-of-the-art models (Flux 2 and Nano Banana). Each image is evaluated by 15 unique users, providing a rich distribution of subjective preferences across domains such as art, design, fashion, and cinematic photography. Leveraging this data, we propose a personalized reward model trained jointly on our high-quality annotations and existing aesthetic assessment subsets. We demonstrate that our model predicts individual liking with higher accuracy than the majority of current state-of-the-art methods predict population-level preferences. Using our personalized predictor, we demonstrate how simple prompt optimization methods can be used to steer generations towards individual user preferences. Our results highlight the importance of data quality and personalization to handle the subjectivity of user preferences. We release our dataset and model to facilitate standardized research in personalized T2I alignment and subjective visual quality assessment.

Method: Introduces GameWorld benchmark with 34 diverse games and 170 tasks, studying two agent interfaces: computer-use agents (keyboard/mouse controls) and generalist multimodal agents (semantic action space via Semantic Action Parsing), with state-verifiable metrics for outcome-based evaluation

Result: Evaluation of 18 model-interface pairs shows best performing agents far from human capabilities; benchmark demonstrates robustness through repeated full-benchmark reruns; studies reveal challenges in real-time interaction, context-memory sensitivity, and action validity

Conclusion: GameWorld provides standardized, verifiable, reproducible evaluation framework for advancing multimodal game agents research, exposing significant challenges ahead for MLLM agents in interactive environments

Abstract: Towards an embodied generalist for real-world interaction, Multimodal Large Language Model (MLLM) agents still suffer from challenging latency, sparse feedback, and irreversible mistakes. Video games offer an ideal testbed with rich visual observations and closed-loop interaction, demanding fine-grained perception, long-horizon planning, and precise control. However, systematically evaluating these capabilities is currently hindered by heterogeneous action interfaces and heuristic verification. To this end, we introduce GameWorld, a benchmark designed for standardized and verifiable evaluation of MLLMs as generalist game agents in browser environments. Two game agent interfaces are studied: (i) computer-use agents that directly emit keyboard and mouse controls, and (ii) generalist multimodal agents that act in a semantic action space via deterministic Semantic Action Parsing. GameWorld contains 34 diverse games and 170 tasks, each paired with state-verifiable metrics for outcome-based evaluation. The results across 18 model-interface pairs suggest that even the best performing agent is far from achieving human capabilities on video games. Extensive experiments of repeated full-benchmark reruns demonstrate the robustness of the benchmark, while further studies on real-time interaction, context-memory sensitivity, and action validity expose more challenges ahead for game agents. Together, by offering a standardized, verifiable, and reproducible evaluation framework, GameWorld lays a robust foundation for advancing research on multimodal game agents and beyond. The project page is at https://gameworld-bench.github.io.

Method: Developed two model variants (2B and 32B parameters) using Mixture-of-Transformers architecture for modality-specific computing with latent tokens to enhance perceptual representation. Introduced iterative self-evolving post-training paradigm for reasoning improvement, and used on-policy distillation to transfer capabilities from large to small model.

Result: The MoT-2B model outperforms similarly sized state-of-the-art models on 16 out of 22 benchmarks, while the 32B variant achieves performance comparable to frontier models like Gemini 3.0 Pro. Successfully trained a Vision-Language-Action model for robot control with compelling real-world physical evaluation results.

Conclusion: HY-Embodied-0.5 effectively addresses the specific needs of embodied intelligence through specialized architecture and training paradigms, demonstrating strong performance across perception, reasoning, and embodied understanding tasks, with practical applications in real-world robot control.

Abstract: We introduce HY-Embodied-0.5, a family of foundation models specifically designed for real-world embodied agents. To bridge the gap between general Vision-Language Models (VLMs) and the demands of embodied agents, our models are developed to enhance the core capabilities required by embodied intelligence: spatial and temporal visual perception, alongside advanced embodied reasoning for prediction, interaction, and planning. The HY-Embodied-0.5 suite comprises two primary variants: an efficient model with 2B activated parameters designed for edge deployment, and a powerful model with 32B activated parameters targeted for complex reasoning. To support the fine-grained visual perception essential for embodied tasks, we adopt a Mixture-of-Transformers (MoT) architecture to enable modality-specific computing. By incorporating latent tokens, this design effectively enhances the perceptual representation of the models. To improve reasoning capabilities, we introduce an iterative, self-evolving post-training paradigm. Furthermore, we employ on-policy distillation to transfer the advanced capabilities of the large model to the smaller variant, thereby maximizing the performance potential of the compact model. Extensive evaluations across 22 benchmarks, spanning visual perception, spatial reasoning, and embodied understanding, demonstrate the effectiveness of our approach. Our MoT-2B model outperforms similarly sized state-of-the-art models on 16 benchmarks, while the 32B variant achieves performance comparable to frontier models such as Gemini 3.0 Pro. In downstream robot control experiments, we leverage our robust VLM foundation to train an effective Vision-Language-Action (VLA) model, achieving compelling results in real-world physical evaluations. Code and models are open-sourced at https://github.com/Tencent-Hunyuan/HY-Embodied.

Method: Dual-step approach: 1) UNet-based semantic mask generator extracts detailed facial component masks (eyes, nose, mouth) directly from blurry inputs, 2) Multi-stage feature fusion integrates these masks with blurry input in a computationally efficient UNet framework. Uses Residual Dense Convolution Blocks, attention mechanisms (CBAM), efficient upsampling, and post-processing.

Result: Outperforms state-of-the-art models on CelebA dataset with higher PSNR and SSIM scores while maintaining good naturalness metrics (NIQE, LPIPS, FID). The framework is lightweight and scalable.

Conclusion: SMFD-UNet provides an effective solution for facial image deblurring by leveraging semantic facial masks, enabling high-quality restoration without reference images, with applications in facial identification, forensics, photography, and medical imaging.

Abstract: For applications including facial identification, forensic analysis, photographic improvement, and medical imaging diagnostics, facial image deblurring is an essential chore in computer vision allowing the restoration of high-quality images from blurry inputs. Often based on general picture priors, traditional deblurring techniques find it difficult to capture the particular structural and identity-specific features of human faces. We present SMFD-UNet (Semantic Mask Fusion Deblurring UNet), a new lightweight framework using semantic face masks to drive the deblurring process, therefore removing the need for high-quality reference photos in order to solve these difficulties. First, our dual-step method uses a UNet-based semantic mask generator to directly extract detailed facial component masks (e.g., eyes, nose, mouth) straight from blurry photos. Sharp, high-fidelity facial images are subsequently produced by integrating these masks with the blurry input using a multi-stage feature fusion technique within a computationally efficient UNet framework. We created a randomized blurring pipeline that roughly replicates real-world situations by simulating around 1.74 trillion deterioration scenarios, hence guaranteeing resilience. Examined on the CelebA dataset, SMFD-UNet shows better performance than state-of-the-art models, attaining higher Peak Signal-to-Noise Ratio (PSNR) and Structural Similarity Index Measure (SSIM) while preserving satisfactory naturalness measures, including NIQE, LPIPS, and FID. Powered by Residual Dense Convolution Blocks (RDC), a multi-stage feature fusion strategy, efficient and effective upsampling techniques, attention techniques like CBAM, post-processing techniques, and the lightweight design guarantees scalability and efficiency, enabling SMFD-UNet to be a flexible solution for developing facial image restoration research and useful applications.

Method: Decomposes 2D shapes into normalized geometry within unit disk and pose vector. Transforms pose into harmonic pose field also within unit disk. Uses orthogonal Zernike bases to encode geometry and pose independently or jointly, followed by frequency-propagation operation to introduce high-frequency content.

Result: Demonstrates theoretical validity, efficiency, discriminability, and applicability across wide range of shape-aware tasks using self-curated XShapeCorpus. Shows the encoding exhibits five favorable properties including invertibility, adaptivity, and frequency richness.

Conclusion: XShapeEnc serves as a foundational tool for research moving beyond 1D sequential data toward 2D spatial intelligence, providing a general-purpose encoding strategy for spatially grounded 2D geometric shapes.

Abstract: Positional encoding has become the de facto standard for grounding deep neural networks on discrete point-wise positions, and it has achieved remarkable success in tasks where the input can be represented as a one-dimensional sequence. However, extending this concept to 2D spatial geometric shapes demands carefully designed encoding strategies that account not only for shape geometry and pose, but also for compatibility with neural network learning. In this work, we address these challenges by introducing a training-free, general-purpose encoding strategy, dubbed XShapeEnc, that encodes an arbitrary spatially grounded 2D geometric shape into a compact representation exhibiting five favorable properties, including invertibility, adaptivity, and frequency richness. Specifically, a 2D spatially grounded geometric shape is decomposed into its normalized geometry within the unit disk and its pose vector, where the pose is further transformed into a harmonic pose field that also lies within the unit disk. A set of orthogonal Zernike bases is constructed to encode shape geometry and pose either independently or jointly, followed by a frequency-propagation operation to introduce high-frequency content into the encoding. We demonstrate the theoretical validity, efficiency, discriminability, and applicability of XShapeEnc via extensive analysis and experiments across a wide range of shape-aware tasks and our self-curated XShapeCorpus. We envision XShapeEnc as a foundational tool for research that goes beyond one-dimensional sequential data toward frontier 2D spatial intelligence.

Method: Formulates depth estimation as feature restoration problem. Uses InvT-IndDiffusion module with invertible transform-based decoder under bi-Lipschitz condition for stable feature restoration. Adds AV-LFE module for low-level feature enhancement with auxiliary viewpoints when available.

Result: Achieves better performance than SOTA methods on various datasets. On KITTI benchmark, improves by 4.09% and 37.77% under different training settings in terms of RMSE compared to baseline.

Conclusion: The proposed method effectively improves monocular depth estimation by treating it as a feature restoration problem and using invertible transform-enhanced indirect diffusion with auxiliary viewpoint enhancement.

Abstract: Monocular Depth Estimation (MDE) is a fundamental computer vision task with important applications in 3D vision. The current mainstream MDE methods employ an encoder-decoder architecture with multi-level/scale feature processing. However, the limitations of the current architecture and the effects of different-level features on the prediction accuracy are not evaluated. In this paper, we first investigate the above problem and show that there is still substantial potential in the current framework if encoder features can be improved. Therefore, we propose to formulate the depth estimation problem from the feature restoration perspective, by treating pretrained encoder features as degraded features of an assumed ground truth feature that yields the ground truth depth map. Then an Invertible Transform-enhanced Indirect Diffusion (InvT-IndDiffusion) module is developed for feature restoration. Due to the absence of direct supervision on feature, only indirect supervision from the final sparse depth map is used. During the iterative procedure of diffusion, this results in feature deviations among steps. The proposed InvT-IndDiffusion solves this problem by using an invertible transform-based decoder under the bi-Lipschitz condition. Finally, a plug-and-play Auxiliary Viewpoint-based Low-level Feature Enhancement module (AV-LFE) is developed to enhance local details with auxiliary viewpoint when available. Experiments demonstrate that the proposed method achieves better performance than the state-of-the-art methods on various datasets. Specifically on the KITTI benchmark, compared with the baseline, the performance is improved by 4.09% and 37.77% under different training settings in terms of RMSE. Code is available at https://github.com/whitehb1/IID-RDepth.

Method: Develops a special case of the Inverse Square Mean Shift Algorithm for non-homogeneous data and applies three-dimensional Fast Fourier Transform to image analysis.

Result: Presents an algorithm extension for non-homogeneous clustering and demonstrates 3D FFT analysis for uncovering hidden patterns in images.

Conclusion: The extended algorithm enables better handling of non-homogeneous data, and 3D FFT provides a powerful tool for pattern discovery in image analysis.

Abstract: This work is a follow up on the newly proposed clustering algorithm called The Inverse Square Mean Shift Algorithm. In this paper a special case of algorithm for dealing with non-homogenous data is formulated and the three dimensional Fast Fourier Transform of images is investigated with the aim of finding hidden patterns.

Method: Combines robust data fidelity (Smooth ℓ1 data term) with explicit sensitivity regularization (divergence-based regularization) and input optimization during training to prevent overfitting in DIP-based hyperspectral image denoising.

Result: The method effectively prevents overfitting and achieves superior denoising performance on real HSIs corrupted by Gaussian, sparse, and stripe noise compared to state-of-the-art DIP-based HSI denoising methods.

Conclusion: The proposed approach successfully addresses the overfitting problem in DIP-based HSI denoising through joint robust data fidelity and sensitivity regularization, leading to improved performance without requiring early stopping.

Abstract: Deep image prior (DIP) is an unsupervised deep learning framework that has been successfully applied to a variety of inverse imaging problems. However, DIP-based methods are inherently prone to overfitting, which leads to performance degradation and necessitates early stopping. In this paper, we propose a method to mitigate overfitting in DIP-based hyperspectral image (HSI) denoising by jointly combining robust data fidelity and explicit sensitivity regularization. The proposed approach employs a Smooth $\ell_1$ data term together with a divergence-based regularization and input optimization during training. Experimental results on real HSIs corrupted by Gaussian, sparse, and stripe noise demonstrate that the proposed method effectively prevents overfitting and achieves superior denoising performance compared to state-of-the-art DIP-based HSI denoising methods.

Method: Analytical and experimental evaluation of SIFT and ORB through common pipeline: keypoint detection, descriptor extraction, matching, and geometric verification via RANSAC with homography estimation

Result: Matching quality assessed using Inlier Ratio on manually constructed GPS-annotated satellite image dataset; impact of number of extracted keypoints examined

Conclusion: Provides systematic evaluation framework for classical image matching methods on satellite imagery, with insights into keypoint quantity effects

Abstract: Image matching is a fundamental problem in Computer Vision with direct applications in robotics, remote sensing, and geospatial data analysis. We present an analytical and experimental evaluation of classical local feature-based image matching algorithms on satellite imagery, focusing on the Scale-Invariant Feature Transform (SIFT) and the Oriented FAST and Rotated BRIEF (ORB). Each method is evaluated through a common pipeline: keypoint detection, descriptor extraction, descriptor matching, and geometric verification via RANSAC with homography estimation. Matching quality is assessed using the Inlier Ratio - the fraction of correspondences consistent with the estimated homography. The study uses a manually constructed dataset of GPS-annotated satellite image tiles with intentional overlaps. We examine the impact of the number of extracted keypoints on the resulting Inlier Ratio.

Method: Combines Vision Transformer (ViT) backbone for spatial feature extraction with unidirectional LSTM for temporal aggregation. Uses unified multi-task formulation to jointly predict handover occurrence and interaction direction, avoiding error propagation from cascaded pipelines. Peak detection on confidence scores identifies discrete handover events.

Result: Achieved F1-score of 0.84 for handover detection and mean F1-score of 0.72 for direction classification on kidney transplant dataset, outperforming single-task variant and VideoMamba-based baseline for direction prediction while maintaining comparable detection performance.

Conclusion: The proposed spatiotemporal framework effectively detects and classifies surgical instrument handovers, with Layer-CAM attribution providing interpretable visualizations of hand-instrument interaction cues.

Abstract: Reliable monitoring of surgical instrument exchanges is essential for maintaining procedural efficiency and patient safety in the operating room. Automatic detection of instrument handovers in intraoperative video remains challenging due to frequent occlusions, background clutter, and the temporally evolving nature of interaction events. We propose a spatiotemporal vision framework for event-level detection and direction classification of surgical instrument handovers in surgical videos. The model combines a Vision Transformer (ViT) backbone for spatial feature extraction with a unidirectional Long Short-Term Memory (LSTM) network for temporal aggregation. A unified multi-task formulation jointly predicts handover occurrence and interaction direction, enabling consistent modeling of transfer dynamics while avoiding error propagation typical of cascaded pipelines. Predicted confidence scores form a temporal signal over the video, from which discrete handover events are identified via peak detection. Experiments on a dataset of kidney transplant procedures demonstrate strong performance, achieving an F1-score of 0.84 for handover detection and a mean F1-score of 0.72 for direction classification, outperforming both a single-task variant and a VideoMamba-based baseline for direction prediction while maintaining comparable detection performance. To improve interpretability, we employ Layer-CAM attribution to visualize spatial regions driving model decisions, highlighting hand-instrument interaction cues.

Method: Proposes MSGL-Transformer with multi-scale attention (short, medium, global branches) and Behavior-Aware Modulation (BAM) blocks that modulate temporal embeddings to emphasize behavior-relevant features before attention. Uses lightweight transformer encoder on pose-based temporal sequences.

Result: Achieves 75.4% accuracy on RatSI dataset (5 classes) and 87.1% accuracy on CalMS21 dataset (4 classes), outperforming TCN, LSTM, Bi-LSTM, ST-GCN, MS-G3D, CTR-GCN, STGAT, and HSTWFormer (+10.7% improvement).

Conclusion: MSGL-Transformer effectively captures multi-scale temporal dynamics for rodent behavior recognition, generalizes across different datasets with minimal adjustments, and significantly outperforms existing methods.

Abstract: Recognition of rodent behavior is important for understanding neural and behavioral mechanisms. Traditional manual scoring is time-consuming and prone to human error. We propose MSGL-Transformer, a Multi-Scale Global-Local Transformer for recognizing rodent social behaviors from pose-based temporal sequences. The model employs a lightweight transformer encoder with multi-scale attention to capture motion dynamics across different temporal scales. The architecture integrates parallel short-range, medium-range, and global attention branches to explicitly capture behavior dynamics at multiple temporal scales. We also introduce a Behavior-Aware Modulation (BAM) block, inspired by SE-Networks, which modulates temporal embeddings to emphasize behavior-relevant features prior to attention. We evaluate on two datasets: RatSI (5 behavior classes, 12D pose inputs) and CalMS21 (4 behavior classes, 28D pose inputs). On RatSI, MSGL-Transformer achieves 75.4% mean accuracy and F1-score of 0.745 across nine cross-validation splits, outperforming TCN, LSTM, and Bi-LSTM. On CalMS21, it achieves 87.1% accuracy and F1-score of 0.8745, a +10.7% improvement over HSTWFormer, and outperforms ST-GCN, MS-G3D, CTR-GCN, and STGAT. The same architecture generalizes across both datasets with only input dimensionality and number of classes adjusted.

Method: Developed a pseudo-annotation pipeline that takes signed video and English text as input and outputs ranked annotations with time intervals. Uses sparse predictions from fingerspelling recognizer and isolated sign recognizer (ISR), combined with a K-Shot LLM approach to estimate annotations. Established baseline models for fingerspelling and ISR.

Result: Achieved state-of-the-art results: 6.7% CER on FSBoard for fingerspelling and 74% top-1 accuracy on ASL Citizen for ISR. Professional interpreter annotated nearly 500 videos from ASL STEM Wiki for gold-standard benchmark. Releasing human annotations and over 300 hours of pseudo-annotations.

Conclusion: The pipeline addresses the annotation bottleneck in sign language AI by providing scalable pseudo-annotation generation, with strong baseline models and new benchmark datasets to advance the field.

Abstract: AI-driven sign language interpretation is limited by a lack of high-quality annotated data. New datasets including ASL STEM Wiki and FLEURS-ASL contain professional interpreters and 100s of hours of data but remain only partially annotated and thus underutilized, in part due to the prohibitive costs of annotating at this scale. In this work, we develop a pseudo-annotation pipeline that takes signed video and English as input and outputs a ranked set of likely annotations, including time intervals, for glosses, fingerspelled words, and sign classifiers. Our pipeline uses sparse predictions from our fingerspelling recognizer and isolated sign recognizer (ISR), along with a K-Shot LLM approach, to estimate these annotations. In service of this pipeline, we establish simple yet effective baseline fingerspelling and ISR models, achieving state-of-the-art on FSBoard (6.7% CER) and on ASL Citizen datasets (74% top-1 accuracy). To validate and provide a gold-standard benchmark, a professional interpreter annotated nearly 500 videos from ASL STEM Wiki with sequence-level gloss labels containing glosses, classifiers, and fingerspelling signs. These human annotations and over 300 hours of pseudo-annotations are being released in supplemental material.

Method: Proposed VSAS-Bench with temporally dense annotations (18k+ annotations), standardized synchronous/asynchronous evaluation protocols, and metrics isolating streaming capabilities. Evaluated recent VLMs analyzing accuracy-latency trade-offs of design factors like memory buffer length, access policy, and input resolution.

Result: Conventional VLMs can be adapted to streaming settings without additional training, outperforming recent streaming VLMs. Qwen3-VL-4B surpasses Dispider (best streaming VLM) by 3% under asynchronous protocol.

Conclusion: VSAS-Bench provides comprehensive framework for evaluating streaming VLMs, showing that adapted conventional VLMs can outperform specialized streaming models, with practical insights for design trade-offs.

Abstract: Streaming vision-language models (VLMs) continuously generate responses given an instruction prompt and an online stream of input frames. This is a core mechanism for real-time visual assistants. Existing VLM frameworks predominantly assess models in offline settings. In contrast, the performance of a streaming VLM depends on additional metrics beyond pure video understanding, including proactiveness, which reflects the timeliness of the model’s responses, and consistency, which captures the robustness of its responses over time. To address this limitation, we propose VSAS-Bench, a new framework and benchmark for Visual Streaming Assistants. In contrast to prior benchmarks that primarily employ single-turn question answering on video inputs, VSAS-Bench features temporally dense annotations with over 18,000 annotations across diverse input domains and task types. We introduce standardized synchronous and asynchronous evaluation protocols, along with metrics that isolate and measure distinct capabilities of streaming VLMs. Using this framework, we conduct large-scale evaluations of recent video and streaming VLMs, analyzing the accuracy-latency trade-off under key design factors such as memory buffer length, memory access policy, and input resolution, yielding several practical insights. Finally, we show empirically that conventional VLMs can be adapted to streaming settings without additional training, and demonstrate that these adapted models outperform recent streaming VLMs. For example, Qwen3-VL-4B surpasses Dispider, the best streaming VLM on our benchmark, by 3% under the asynchronous protocol. The benchmark and code will be available at https://github.com/apple/ml-vsas-bench.

Method: Proposes Depth-converted-Scale Convolution (DcSConv) that incorporates prior relationships between object depth and scale to extract features from appropriate convolution receptive field scales. Also develops Depth-converted-Scale aware Fusion (DcS-F) to adaptively fuse DcSConv features with conventional convolution features. The framework is plug-and-play compatible with existing CNN methods.

Result: Achieves best results on KITTI benchmark with up to 11.6% improvement in SqRel reduction compared to baselines. Ablation studies validate effectiveness of each proposed module.

Conclusion: The scale of convolution filters matters as much as (or more than) local deformation for monocular depth estimation. The proposed DcSConv framework effectively addresses object size-depth ambiguity and can enhance existing CNN-based depth estimation methods.

Abstract: Self-supervised monocular depth estimation (MDE) has received increasing interests in the last few years. The objects in the scene, including the object size and relationship among different objects, are the main clues to extract the scene structure. However, previous works lack the explicit handling of the changing sizes of the object due to the change of its depth. Especially in a monocular video, the size of the same object is continuously changed, resulting in size and depth ambiguity. To address this problem, we propose a Depth-converted-Scale Convolution (DcSConv) enhanced monocular depth estimation framework, by incorporating the prior relationship between the object depth and object scale to extract features from appropriate scales of the convolution receptive field. The proposed DcSConv focuses on the adaptive scale of the convolution filter instead of the local deformation of its shape. It establishes that the scale of the convolution filter matters no less (or even more in the evaluated task) than its local deformation. Moreover, a Depth-converted-Scale aware Fusion (DcS-F) is developed to adaptively fuse the DcSConv features and the conventional convolution features. Our DcSConv enhanced monocular depth estimation framework can be applied on top of existing CNN based methods as a plug-and-play module to enhance the conventional convolution block. Extensive experiments with different baselines have been conducted on the KITTI benchmark and our method achieves the best results with an improvement up to 11.6% in terms of SqRel reduction. Ablation study also validates the effectiveness of each proposed module.

Method: Post-training quantization algorithm using simple dot products and rounding operations without backpropagation; introduces weight-aware strategy that reorders weights within each layer to address accuracy degradation from channel-wise scaling while preserving channel structure

Result: Achieves best results in 2-bit, 4-bit, and 8-bit quantization compared to other methods

Conclusion: Permutation-COMQ provides an effective quantization solution for medical foundation models that simplifies the process while maintaining accuracy across different bit-widths

Abstract: Foundation models have achieved remarkable results in medical image analysis. However, its large network architecture and high computational complexity significantly impact inference speed, limiting its application on terminal medical devices. Quantization, a technique that compresses models into low-bit versions, is a solution to this challenge. In this paper, we propose a post-training quantization algorithm, Permutation-COMQ. It eliminates the need for backpropagation by using simple dot products and rounding operations, thereby removing hyperparameter tuning and simplifying the process. Additionally, we introduce a weight-aware strategy that reorders the weight within each layer to address the accuracy degradation induced by channel-wise scaling during quantization, while preserving channel structure. Experiments demonstrate that our method achieves the best results in 2-bit, 4-bit, and 8-bit quantization.

Method: Proposes FireSenseNet, a dual-branch convolutional neural network with Cross-Attentive Feature Interaction Module (CAFIM) that explicitly models spatially varying interactions between fuel and weather modalities through learnable attention gates at multiple encoder scales.

Result: FireSenseNet achieves F1 of 0.4176 and AUC-PR of 0.3435 on Google Next-Day Wildfire Spread benchmark, outperforming all alternatives including SegFormer with 3.8× more parameters. CAFIM provides 7.1% relative F1 gain over naive concatenation.

Conclusion: Explicit modeling of multimodal interactions between static fuel/terrain and dynamic weather conditions improves wildfire spread prediction, with previous-day fire mask being most important predictor and wind speed acting as noise at coarse temporal resolution.

Abstract: Accurate prediction of next-day wildfire spread is critical for disaster response and resource allocation. Existing deep learning approaches typically concatenate heterogeneous geospatial inputs into a single tensor, ignoring the fundamental physical distinction between static fuel/terrain properties and dynamic meteorological conditions. We propose FireSenseNet, a dual-branch convolutional neural network equipped with a novel Cross-Attentive Feature Interaction Module (CAFIM) that explicitly models the spatially varying interaction between fuel and weather modalities through learnable attention gates at multiple encoder scales. Through a systematic comparison of seven architectures – spanning pure CNNs, Vision Transformers, and hybrid designs – on the Google Next-Day Wildfire Spread benchmark, we demonstrate that FireSenseNet achieves an F1 of 0.4176 and AUC-PR of 0.3435, outperforming all alternatives including a SegFormer with 3.8* more parameters (F1 = 0.3502). Ablation studies confirm that CAFIM provides a 7.1% relative F1 gain over naive concatenation, and channel-wise feature importance analysis reveals that the previous-day fire mask dominates prediction while wind speed acts as noise at the dataset’s coarse temporal resolution. We further incorporate Monte Carlo Dropout for pixel-level uncertainty quantification and present a critical analysis showing that common evaluation shortcuts inflate reported F1 scores by over 44%.

Method: Evaluated one-class representation learning techniques (DSVDD and DROC) trained exclusively on slide-negative patches without instance-level supervision. Compared against FS-SIL, WS-SIL, and ItS2CLR methods on bone marrow cytomorphology (TCIA) and oral cancer cytology datasets.

Result: DSVDD achieved state-of-the-art performance in instance-level abnormality ranking, particularly in ultra-low witness-rate regimes (≤1%), sometimes even outperforming fully supervised learning. DROC was also competitive under extreme rarity due to distribution-augmented contrastive learning.

Conclusion: One-class representation learning is a robust and interpretable superior choice to MIL for malignant cell detection under extreme rarity, offering practical advantages since exhaustive instance-level annotations are infeasible in whole-slide cytology.

Abstract: In computational cytology, detecting malignancy on whole-slide images is difficult because malignant cells are morphologically diverse yet vanishingly rare amid a vast background of normal cells. Accurate detection of these extremely rare malignant cells remains challenging due to large class imbalance and limited annotations. Conventional weakly supervised approaches, such as multiple instance learning (MIL), often fail to generalize at the instance level, especially when the fraction of malignant cells (witness rate) is exceedingly low. In this study, we explore the use of one-class representation learning techniques for detecting malignant cells in low-witness-rate scenarios. These methods are trained exclusively on slide-negative patches, without requiring any instance-level supervision. Specifically, we evaluate two OCC approaches, DSVDD and DROC, and compare them with FS-SIL, WS-SIL, and the recent ItS2CLR method. The one-class methods learn compact representations of normality and detect deviations at test time. Experiments on a publicly available bone marrow cytomorphology dataset (TCIA) and an in-house oral cancer cytology dataset show that DSVDD achieves state-of-the-art performance in instance-level abnormality ranking, particularly in ultra-low witness-rate regimes ($\leq 1%$) and, in some cases, even outperforming fully supervised learning, which is typically not a practical option in whole-slide cytology due to the infeasibility of exhaustive instance-level annotations. DROC is also competitive under extreme rarity, benefiting from distribution-augmented contrastive learning. These findings highlight one-class representation learning as a robust and interpretable superior choice to MIL for malignant cell detection under extreme rarity.

Method: Instead of optimizing logits through iterative training, the method directly uses the analytic solution of distribution discrepancy between visual and linguistic features as semantic segmentation maps. The key hypothesis is that distribution discrepancy encodes semantic information consistently within categories but inconsistently across categories.

Result: Achieves state-of-the-art performance on eight benchmark datasets while eliminating time-consuming iterative training and model-specific attention modulation requirements.

Conclusion: The proposed analytic solution approach provides a more efficient and effective method for open-vocabulary semantic segmentation by directly leveraging distribution discrepancy as semantic maps rather than optimizing through iterative training.

Abstract: Open-vocabulary semantic segmentation (OVSS) aims to segment arbitrary category regions in images using open-vocabulary prompts, necessitating that existing methods possess pixel-level vision-language alignment capability. Typically, this capability involves computing the cosine similarity, \ie, logits, between visual and linguistic features, and minimizing the distribution discrepancy between the logits and the ground truth (GT) to generate optimal logits that are subsequently used to construct segmentation maps, yet it depends on time-consuming iterative training or model-specific attention modulation. In this work, we propose a more direct approach that eschews the logits-optimization process by directly deriving an analytic solution for the segmentation map. We posit a key hypothesis: the distribution discrepancy encodes semantic information; specifically, this discrepancy exhibits consistency across patches belonging to the same category but inconsistency across different categories. Based on this hypothesis, we directly utilize the analytic solution of this distribution discrepancy as the semantic maps. In other words, we reformulate the optimization of the distribution discrepancy as deriving its analytic solution, thereby eliminating time-consuming iterative training, freeing us from model-specific attention modulation, and achieving state-of-the-art performance on eight benchmark datasets.

Method: Proposes GEAR framework with EM-style alternating optimization using Gaussian Splatting representation. Treats part segmentation as latent variable and joint motion parameters as explicit variables, alternately refining them. Uses 2D segmentation model for multi-view part priors and weakly supervised constraint to regularize latent variable.

Result: Achieves state-of-the-art results in geometric reconstruction and motion parameters estimation on multiple benchmarks and newly constructed GEAR-Multi dataset, particularly on complex articulated objects with multiple movable parts.

Conclusion: GEAR provides an effective framework for reconstructing articulated objects by jointly modeling geometry and motion with improved convergence and geometric-motion consistency.

Abstract: High-fidelity interactive digital assets are essential for embodied intelligence and robotic interaction, yet articulated objects remain challenging to reconstruct due to their complex structures and coupled geometry-motion relationships. Existing methods suffer from instability in geometry-motion joint optimization, while their generalization remains limited on complex multi-joint or out-of-distribution objects. To address these challenges, we propose GEAR, an EM-style alternating optimization framework that jointly models geometry and motion as interdependent components within a Gaussian Splatting representation. GEAR treats part segmentation as a latent variable and joint motion parameters as explicit variables, alternately refining them for improved convergence and geometric-motion consistency. To enhance part segmentation quality without sacrificing generalization, we leverage a vanilla 2D segmentation model to provide multi-view part priors, and employ a weakly supervised constraint to regularize the latent variable. Experiments on multiple benchmarks and our newly constructed dataset GEAR-Multi demonstrate that GEAR achieves state-of-the-art results in geometric reconstruction and motion parameters estimation, particularly on complex articulated objects with multiple movable parts.

Method: Proposes CG-CLIP with two key components: 1) Caption-guided Memory Refinement (CMR) uses MLLM-generated captions to refine identity-specific features, and 2) Token-based Feature Extraction (TFE) employs cross-attention with fixed-length learnable tokens to efficiently aggregate spatiotemporal features.

Result: Outperforms current state-of-the-art approaches on two standard datasets (MARS, iLIDS-VID) and two newly constructed high-difficulty datasets (SportsVReID, DanceVReID), achieving significant improvements across all benchmarks.

Conclusion: The proposed caption-guided CLIP framework effectively handles challenging video person ReID scenarios by leveraging multimodal information from MLLMs and efficient token-based feature extraction.

Abstract: In recent years, video-based person Re-Identification (ReID) has gained attention for its ability to leverage spatiotemporal cues to match individuals across non-overlapping cameras. However, current methods struggle with high-difficulty scenarios, such as sports and dance performances, where multiple individuals wear similar clothing while performing dynamic movements. To overcome these challenges, we propose CG-CLIP, a novel caption-guided CLIP framework that leverages explicit textual descriptions and learnable tokens. Our method introduces two key components: Caption-guided Memory Refinement (CMR) and Token-based Feature Extraction (TFE). CMR utilizes captions generated by Multi-modal Large Language Models (MLLMs) to refine identity-specific features, capturing fine-grained details. TFE employs a cross-attention mechanism with fixed-length learnable tokens to efficiently aggregate spatiotemporal features, reducing computational overhead. We evaluate our approach on two standard datasets (MARS and iLIDS-VID) and two newly constructed high-difficulty datasets (SportsVReID and DanceVReID). Experimental results demonstrate that our method outperforms current state-of-the-art approaches, achieving significant improvements across all benchmarks.

Method: Proposes Symbiotic-MoE framework with: 1) Modality-Aware Expert Disentanglement that partitions experts into task-specific groups with shared experts as multimodal semantic bridges, 2) Progressive Training Strategy with differential learning rates and early-stage gradient shielding to protect pre-trained knowledge while allowing generative signals to enhance understanding.

Result: Achieves rapid generative convergence while unlocking cross-modal synergy, boosting inherent understanding with remarkable gains on MMLU and OCRBench benchmarks.

Conclusion: Symbiotic-MoE resolves task interference in multimodal models without parameter overhead, enabling both image generation and understanding capabilities to synergistically enhance each other rather than conflict.

Abstract: Empowering Large Multimodal Models (LMMs) with image generation often leads to catastrophic forgetting in understanding tasks due to severe gradient conflicts. While existing paradigms like Mixture-of-Transformers (MoT) mitigate this conflict through structural isolation, they fundamentally sever cross-modal synergy and suffer from capacity fragmentation. In this work, we present Symbiotic-MoE, a unified pre-training framework that resolves task interference within a native multimodal Mixture-of-Experts (MoE) Transformers architecture with zero-parameter overhead. We first identify that standard MoE tuning leads to routing collapse, where generative gradients dominate expert utilization. To address this, we introduce Modality-Aware Expert Disentanglement, which partitions experts into task-specific groups while utilizing shared experts as a multimodal semantic bridge. Crucially, this design allows shared experts to absorb fine-grained visual semantics from generative tasks to enrich textual representations. To optimize this, we propose a Progressive Training Strategy featuring differential learning rates and early-stage gradient shielding. This mechanism not only shields pre-trained knowledge from early volatility but eventually transforms generative signals into constructive feedback for understanding. Extensive experiments demonstrate that Symbiotic-MoE achieves rapid generative convergence while unlocking cross-modal synergy, boosting inherent understanding with remarkable gains on MMLU and OCRBench.

Method: Formulates articulated joint estimation as synthesis-mediated reasoning: 1) synthesizes maximally articulated opened state under same camera view to expose articulation cues, 2) estimates full joint parameters from discrepancy between observed and synthesized states using set-prediction formulation without object-specific templates or multi-view inputs.

Result: Achieves strong performance in articulated joint estimation and supports part-level novel state synthesis conditioned on joints as downstream capability.

Conclusion: DailyArt provides effective approach for articulated object kinematics estimation from single static images through synthesis-mediated reasoning, enabling downstream applications like part-level state synthesis.

Abstract: Articulated objects are essential for embodied AI and world models, yet inferring their kinematics from a single closed-state image remains challenging because crucial motion cues are often occluded. Existing methods either require multi-state observations or rely on explicit part priors, retrieval, or other auxiliary inputs that partially expose the structure to be inferred. In this work, we present DailyArt, which formulates articulated joint estimation from a single static image as a synthesis-mediated reasoning problem. Instead of directly regressing joints from a heavily occluded observation, DailyArt first synthesizes a maximally articulated opened state under the same camera view to expose articulation cues, and then estimates the full set of joint parameters from the discrepancy between the observed and synthesized states. Using a set-prediction formulation, DailyArt recovers all joints simultaneously without requiring object-specific templates, multi-view inputs, or explicit part annotations at test time. Taking estimated joints as conditions, the framework further supports part-level novel state synthesis as a downstream capability. Extensive experiments show that DailyArt achieves strong performance in articulated joint estimation and supports part-level novel state synthesis conditioned on joints. Project page is available at https://rangooo123.github.io/DaliyArt.github.io/.

Method: Constructed WUTDet dataset with 100,576 images and 381,378 annotated ship instances covering diverse scenarios (ports, anchorages, navigation, berthing) and imaging conditions (fog, glare, low-lightness, rain). Evaluated 20 baseline models from CNN, Transformer, and Mamba architectures, and created Ship-GEN cross-dataset test set for generalization evaluation.

Result: Transformer architecture achieves superior overall detection accuracy (AP) and small-object detection performance (APs); CNN maintains inference efficiency advantage; Mamba achieves favorable balance between accuracy and computational efficiency. Models trained on WUTDet show stronger generalization on Ship-GEN test set.

Conclusion: WUTDet provides effective data support for research, evaluation, and generalization analysis of ship detection algorithms in complex maritime scenarios, with Transformer showing best accuracy, CNN best efficiency, and Mamba good balance.

Abstract: Ship detection for navigation is a fundamental perception task in intelligent waterway transportation systems. However, existing public ship detection datasets remain limited in terms of scale, the proportion of small-object instances, and scene diversity, which hinders the systematic evaluation and generalization study of detection algorithms in complex maritime environments. To this end, we construct WUTDet, a large-scale ship detection dataset. WUTDet contains 100,576 images and 381,378 annotated ship instances, covering diverse operational scenarios such as ports, anchorages, navigation, and berthing, as well as various imaging conditions including fog, glare, low-lightness, and rain, thereby exhibiting substantial diversity and challenge. Based on WUTDet, we systematically evaluate 20 baseline models from three mainstream detection architectures, namely CNN, Transformer, and Mamba. Experimental results show that the Transformer architecture achieves superior overall detection accuracy (AP) and small-object detection performance (APs), demonstrating stronger adaptability to complex maritime scenes; the CNN architecture maintains an advantage in inference efficiency, making it more suitable for real-time applications; and the Mamba architecture achieves a favorable balance between detection accuracy and computational efficiency. Furthermore, we construct a unified cross-dataset test set, Ship-GEN, to evaluate model generalization. Results on Ship-GEN show that models trained on WUTDet exhibit stronger generalization under different data distributions. These findings demonstrate that WUTDet provides effective data support for the research, evaluation, and generalization analysis of ship detection algorithms in complex maritime scenarios. The dataset is publicly available at: https://github.com/MAPGroup/WUTDet.

Method: Proposes MAF (Modality-Agnostic Forgery) detection framework that decouples modality-specific styles to extract essential cross-modal latent forgery knowledge, with two generalization dimensions: Weak MAF (transferability) and Strong MAF (robustness).

Result: Achieves significant performance breakthroughs on unknown modalities, empirically proves existence of universal forgery traces, and introduces DeepModal-Bench benchmark for rigorous evaluation.

Conclusion: Offers a pioneering technical pathway for universal multimodal defense by shifting from “feature fusion” to “modality generalization” paradigm in multimodal forensics.

Abstract: As generative artificial intelligence evolves, deepfake attacks have escalated from single-modality manipulations to complex, multimodal threats. Existing forensic techniques face a severe generalization bottleneck: by relying excessively on superficial, modality-specific artifacts, they neglect the shared latent forgery knowledge hidden beneath variable physical appearances. Consequently, these models suffer catastrophic performance degradation when confronted with unseen “dark modalities.” To break this limitation, this paper introduces a paradigm shift that redefines multimodal forensics from conventional “feature fusion” to “modality generalization.” We propose the first modality-agnostic forgery (MAF) detection framework. By explicitly decoupling modality-specific styles, MAF precisely extracts the essential, cross-modal latent forgery knowledge. Furthermore, we define two progressive dimensions to quantify model generalization: transferability toward semantically correlated modalities (Weak MAF), and robustness against completely isolated signals of “dark modality” (Strong MAF). To rigorously assess these generalization limits, we introduce the DeepModal-Bench benchmark, which integrates diverse multimodal forgery detection algorithms and adapts state-of-the-art generalized learning methods. This study not only empirically proves the existence of universal forgery traces but also achieves significant performance breakthroughs on unknown modalities via the MAF framework, offering a pioneering technical pathway for universal multimodal defense.

Method: Proposes RemoteAgent framework that respects MLLM capability boundaries, constructs VagueEO dataset with simulated vague queries for EO tasks, uses reinforcement fine-tuning to align MLLM as cognitive core for image/region-level tasks, and orchestrates specialized tools via Model Context Protocol for dense predictions.

Result: RemoteAgent achieves robust intent recognition capabilities and highly competitive performance across diverse EO tasks, demonstrating effective bridging between ambiguous human queries and appropriate multi-granularity visual analysis.

Conclusion: The framework successfully addresses the gap between vague human queries and precision-critical EO tasks by strategically leveraging MLLMs’ native capabilities while efficiently delegating dense predictions to specialized tools.

Abstract: Earth Observation (EO) systems are essentially designed to support domain experts who often express their requirements through vague natural language rather than precise, machine-friendly instructions. Depending on the specific application scenario, these vague queries can demand vastly different levels of visual precision. Consequently, a practical EO AI system must bridge the gap between ambiguous human queries and the appropriate multi-granularity visual analysis tasks, ranging from holistic image interpretation to fine-grained pixel-wise predictions. While Multi-modal Large Language Models (MLLMs) demonstrate strong semantic understanding, their text-based output format is inherently ill-suited for dense, precision-critical spatial predictions. Existing agentic frameworks address this limitation by delegating tasks to external tools, but indiscriminate tool invocation is computationally inefficient and underutilizes the MLLM’s native capabilities. To this end, we propose RemoteAgent, an agentic framework that strategically respects the intrinsic capability boundaries of MLLMs. To empower this framework to understand real user intents, we construct VagueEO, a human-centric instruction dataset pairing EO tasks with simulated vague natural-language queries. By leveraging VagueEO for reinforcement fine-tuning, we align an MLLM into a robust cognitive core that directly resolves image- and sparse region-level tasks. Consequently, RemoteAgent processes suitable tasks internally while intelligently orchestrating specialized tools via the Model Context Protocol exclusively for dense predictions. Extensive experiments demonstrate that RemoteAgent achieves robust intent recognition capabilities while delivering highly competitive performance across diverse EO tasks.

Method: ESOM includes: 1) Definition Normalization module to structure user prompts, 2) Inter-frame-matched Intra-frame Token Merging to compress visual tokens, 3) Hybrid Streaming Memory for efficient causal inference, and 4) Probabilistic Scoring to convert textual outputs to frame-level anomaly scores. Also introduces OpenDef-Bench benchmark.

Result: ESOM achieves real-time efficiency on a single GPU and state-of-the-art performance in anomaly temporal localization, classification, and description generation.

Conclusion: ESOM provides an efficient streaming solution for open-world video anomaly detection with training-free operation and strong performance, addressing practical deployment challenges of existing MLLM-based methods.

Abstract: Open-world video anomaly detection (OWVAD) aims to detect and explain abnormal events under different anomaly definitions, which is important for applications such as intelligent surveillance and live-streaming content moderation. Recent MLLM-based methods have shown promising open-world generalization, but still suffer from three major limitations: inefficiency for practical deployment, lack of streaming processing adaptation, and limited support for dynamic anomaly definitions in both modeling and evaluation. To address these issues, this paper proposes ESOM, an efficient streaming OWVAD model that operates in a training-free manner. ESOM includes a Definition Normalization module to structure user prompts for reducing hallucination, an Inter-frame-matched Intra-frame Token Merging module to compress redundant visual tokens, a Hybrid Streaming Memory module for efficient causal inference, and a Probabilistic Scoring module that converts interval-level textual outputs into frame-level anomaly scores. In addition, this paper introduces OpenDef-Bench, a new benchmark with clean surveillance videos and diverse natural anomaly definitions for evaluating performance under varying conditions. Extensive experiments show that ESOM achieves real-time efficiency on a single GPU and state-of-the-art performance in anomaly temporal localization, classification, and description generation. The code and benchmark will be released at https://github.com/Kamino666/ESOM_OpenDef-Bench.

Method: LogitProd treats independently trained FM-based predictors as fixed experts and learns sample-adaptive fusion weights over their slide-level outputs, operating purely on logits without requiring encoder retraining or feature-space alignment across heterogeneous backbones.

Result: LogitProd ranks first on 20/22 tasks across WSI-level classification, tile-level classification, gene mutation prediction, and discrete-time survival modeling, improving average performance by ~3% over the strongest single expert with ~12× lower training cost than feature-fusion alternatives.

Conclusion: LogitProd enables practitioners to upgrade heterogeneous FM-based pipelines in a plug-and-play manner, achieving multi-expert gains with significantly lower computational cost than traditional fusion approaches.

Abstract: Pathology foundation models (FMs) have become central to computational histopathology, offering strong transfer performance across a wide range of diagnostic and prognostic tasks. The rapid proliferation of pathology foundation models creates a model-selection bottleneck: no single model is uniformly best, yet exhaustively adapting and validating many candidates for each downstream endpoint is prohibitively expensive. We address this challenge with a lightweight and novel model fusion strategy, LogitProd, which treats independently trained FM-based predictors as fixed experts and learns sample-adaptive fusion weights over their slide-level outputs. The fusion operates purely on logits, requiring no encoder retraining and no feature-space alignment across heterogeneous backbones. We further provide a theoretical analysis showing that the optimal weighted product fusion is guaranteed to perform at least as well as the best individual expert under the training objective. We systematically evaluate LogitProd on \textbf{22} benchmarks spanning WSI-level classification, tile-level classification, gene mutation prediction, and discrete-time survival modeling. LogitProd ranks first on 20/22 tasks and improves the average performance across all tasks by ~3% over the strongest single expert. LogitProd enables practitioners to upgrade heterogeneous FM-based pipelines in a plug-and-play manner, achieving multi-expert gains with $\sim$12$\times$ lower training cost than feature-fusion alternatives.

Method: Proposes Cross-Modal Emotion Transfer (C-MET) that learns emotion semantic vectors between speech and visual feature spaces. Uses a large-scale pretrained audio encoder and disentangled facial expression encoder to model emotion differences across modalities, enabling emotion transfer from speech to facial expressions.

Result: Extensive experiments on MEAD and CREMA-D datasets show 14% improvement in emotion accuracy over state-of-the-art methods, generating expressive talking face videos even for unseen extended emotions.

Conclusion: C-MET effectively addresses limitations of existing emotion editing methods by learning cross-modal emotion semantics, enabling flexible and accurate emotion transfer from speech to facial expressions for talking face generation.

Abstract: Talking face generation has gained significant attention as a core application of generative models. To enhance the expressiveness and realism of synthesized videos, emotion editing in talking face video plays a crucial role. However, existing approaches often limit expressive flexibility and struggle to generate extended emotions. Label-based methods represent emotions with discrete categories, which fail to capture a wide range of emotions. Audio-based methods can leverage emotionally rich speech signals - and even benefit from expressive text-to-speech (TTS) synthesis - but they fail to express the target emotions because emotions and linguistic contents are entangled in emotional speeches. Images-based methods, on the other hand, rely on target reference images to guide emotion transfer, yet they require high-quality frontal views and face challenges in acquiring reference data for extended emotions (e.g., sarcasm). To address these limitations, we propose Cross-Modal Emotion Transfer (C-MET), a novel approach that generates facial expressions based on speeches by modeling emotion semantic vectors between speech and visual feature spaces. C-MET leverages a large-scale pretrained audio encoder and a disentangled facial expression encoder to learn emotion semantic vectors that represent the difference between two different emotional embeddings across modalities. Extensive experiments on the MEAD and CREMA-D datasets demonstrate that our method improves emotion accuracy by 14% over state-of-the-art methods, while generating expressive talking face videos - even for unseen extended emotions. Code, checkpoint, and demo are available at https://chanhyeok-choi.github.io/C-MET/

Method: Proposes a coarse-to-fine stylization pipeline that uses pre-trained diffusion models to extract abstract representations from images. The method deforms 3D meshes to match these representations while preserving original mesh topology and part-level semantics. Also introduces an approximate VAE encoder for efficient gradient computation from mesh renderings.

Result: The method successfully creates stylized 3D meshes that reflect unique geometric features of source images, including expressive poses and silhouettes. It supports diverse geometric variations while maintaining valid mesh structure.

Conclusion: The framework enables distinctive artistic 3D creation by allowing bold geometric stylization of 3D meshes based on image styles, bridging the gap between 2D artistic expression and 3D geometric deformation.

Abstract: Recent generative models can create visually plausible 3D representations of objects. However, the generation process often allows for implicit control signals, such as contextual descriptions, and rarely supports bold geometric distortions beyond existing data distributions. We propose a geometric stylization framework that deforms a 3D mesh, allowing it to express the style of an image. While style is inherently ambiguous, we utilize pre-trained diffusion models to extract an abstract representation of the provided image. Our coarse-to-fine stylization pipeline can drastically deform the input 3D model to express a diverse range of geometric variations while retaining the valid topology of the original mesh and part-level semantics. We also propose an approximate VAE encoder that provides efficient and reliable gradients from mesh renderings. Extensive experiments demonstrate that our method can create stylized 3D meshes that reflect unique geometric features of the pictured assets, such as expressive poses and silhouettes, thereby supporting the creation of distinctive artistic 3D creations. Project page: https://changwoonchoi.github.io/GeoStyle

Method: Proposes LAKE (latent anomaly knowledge excavation) - a training-free framework that identifies sparse anomaly-sensitive neurons using minimal normal samples. It isolates these neurons to construct compact normality representations integrating visual structural deviations with cross-modal semantic activations.

Result: Extensive experiments on industrial AD benchmarks demonstrate state-of-the-art performance while providing intrinsic, neuron-level interpretability.

Conclusion: Advocates for a paradigm shift: redefining anomaly detection as targeted activation of latent pre-trained knowledge rather than acquisition of downstream task knowledge.

Abstract: Large-scale vision-language models (VLMs) exhibit remarkable zero-shot capabilities, yet the internal mechanisms driving their anomaly detection (AD) performance remain poorly understood. Current methods predominantly treat VLMs as black-box feature extractors, assuming that anomaly-specific knowledge must be acquired through external adapters or memory banks. In this paper, we challenge this assumption by arguing that anomaly knowledge is intrinsically embedded within pre-trained models but remains latent and under-activated. We hypothesize that this knowledge is concentrated within a sparse subset of anomaly-sensitive neurons. To validate this, we propose latent anomaly knowledge excavation (LAKE), a training-free framework that identifies and elicits these critical neuronal signals using only a minimal set of normal samples. By isolating these sensitive neurons, LAKE constructs a highly compact normality representation that integrates visual structural deviations with cross-modal semantic activations. Extensive experiments on industrial AD benchmarks demonstrate that LAKE achieves state-of-the-art performance while providing intrinsic, neuron-level interpretability. Ultimately, our work advocates for a paradigm shift: redefining anomaly detection as the targeted activation of latent pre-trained knowledge rather than the acquisition of a downstream task.

Method: HAWK is a head importance-aware visual token pruning method that perceives varying importance of attention heads in visual tasks. It uses head importance weights and text-guided attention to assess visual token significance, retaining task-relevant tokens while removing redundant ones. The method is training-free and can be applied to various MLLMs.

Result: Extensive experiments on multiple vision-language benchmarks show HAWK achieves state-of-the-art accuracy. When applied to Qwen2.5-VL, it retains 96.0% of original accuracy after pruning 80.2% of visual tokens, reduces end-to-end latency to 74.4% of original, and decreases GPU memory usage across tested models.

Conclusion: HAWK effectively addresses the computational overhead of visual tokens in MLLMs through head importance-aware pruning, maintaining high accuracy while significantly reducing inference time and memory usage, making MLLMs more practical for real-time applications.

Abstract: In multimodal large language models (MLLMs), the surge of visual tokens significantly increases the inference time and computational overhead, making them impractical for real-time or resource-constrained applications. Visual token pruning is a promising strategy for reducing the cost of MLLM inference by removing redundant visual tokens. Existing research usually assumes that all attention heads contribute equally to the visual interpretation. However, our study reveals that different heads may capture distinct visual semantics and inherently play distinct roles in visual processing. In light of this observation, we propose HAWK, a head importance-aware visual token pruning method that perceives the varying importance of attention heads in visual tasks to maximize the retention of crucial tokens. By leveraging head importance weights and text-guided attention to assess visual token significance, HAWK effectively retains task-relevant visual tokens while removing redundant ones. The proposed HAWK is entirely training-free and can be seamlessly applied to various MLLMs. Extensive experiments on multiple mainstream vision-language benchmarks demonstrate that HAWK achieves state-of-the-art accuracy. When applied to Qwen2.5-VL, HAWK retains 96.0% of the original accuracy after pruning 80.2% of the visual tokens. Additionally, it reduces end-to-end latency to 74.4% of the original and further decreases GPU memory usage across the tested models. The code is available at https://github.com/peppery77/HAWK.git.

Method: Created AgriChain dataset with expert-curated leaf images, disease labels, confidence scores, and GPT-4o generated then expert-verified chain-of-thought rationales. Fine-tuned Qwen2.5-VL-3B on this dataset to create AgriChain-VL3B for joint disease prediction and visually grounded reasoning.

Result: Achieved 73.1% top-1 accuracy on 1,000-image test set (macro F1 = 0.466; weighted F1 = 0.655), outperforming Gemini 1.5 Flash, Gemini 2.5 Pro, and GPT-4o Mini. Generated explanations closely align with expert reasoning and reference key visual cues.

Conclusion: Expert-verified reasoning supervision significantly enhances both accuracy and interpretability of vision-language models for plant disease diagnosis, bridging the gap between generic multimodal models and human expertise for trustworthy agricultural AI.

Abstract: Accurate and interpretable plant disease diagnosis remains a major challenge for vision-language models (VLMs) in real-world agriculture. We introduce AgriChain, a dataset of approximately 11,000 expert-curated leaf images spanning diverse crops and pathologies, each paired with (i) a disease label, (ii) a calibrated confidence score (High/Medium/Low), and (iii) an expert-verified chain-of-thought (CoT) rationale. Draft explanations were first generated by GPT-4o and then verified by a professional agricultural engineer using standardized descriptors (e.g., lesion color, margin, and distribution). We fine-tune Qwen2.5-VL-3B on AgriChain, resulting in a specialized model termed AgriChain-VL3B, to jointly predict diseases and generate visually grounded reasoning. On a 1,000-image test set, our CoT-supervised model achieves 73.1% top-1 accuracy (macro F1 = 0.466; weighted F1 = 0.655), outperforming strong baselines including Gemini 1.5 Flash, Gemini 2.5 Pro, and GPT-4o Mini. The generated explanations align closely with expert reasoning, consistently referencing key visual cues. These findings demonstrate that expert-verified reasoning supervision significantly enhances both accuracy and interpretability, bridging the gap between generic multimodal models and human expertise, and advancing trustworthy, globally deployable AI for sustainable agriculture. The dataset and code are publicly available at: https://github.com/hazzanabeel12-netizen/agrichain

Method: FlowGuard uses a cross-model in-generation detection framework that inspects intermediate denoising steps. It employs a novel linear approximation for latent decoding in latent diffusion models (where early-stage noise obscures visual signals) and uses curriculum learning to stabilize training.

Result: FlowGuard outperforms existing methods by over 30% in F1 score across nine diffusion backbones. It achieves transformative efficiency gains: reduces peak GPU memory demand by over 97% and cuts projection time from 8.1 seconds to 0.2 seconds compared to standard VAE decoding.

Conclusion: FlowGuard provides an effective solution for in-generation NSFW detection in diffusion models, enabling early unsafe content detection that reduces computational costs while maintaining high accuracy across different model architectures and settings.

Abstract: Diffusion-based image generation models have advanced rapidly but pose a safety risk due to their potential to generate Not-Safe-For-Work (NSFW) content. Existing NSFW detection methods mainly operate either before or after image generation. Pre-generation methods rely on text prompts and struggle with the gap between prompt safety and image safety. Post-generation methods apply classifiers to final outputs, but they are poorly suited to intermediate noisy images. To address this, we introduce FlowGuard, a cross-model in-generation detection framework that inspects intermediate denoising steps. This is particularly challenging in latent diffusion, where early-stage noise obscures visual signals. FlowGuard employs a novel linear approximation for latent decoding and leverages a curriculum learning approach to stabilize training. By detecting unsafe content early, FlowGuard reduces unnecessary diffusion steps to cut computational costs. Our cross-model benchmark spanning nine diffusion-based backbones shows the effectiveness of FlowGuard for in-generation NSFW detection in both in-distribution and out-of-distribution settings, outperforming existing methods by over 30% in F1 score while delivering transformative efficiency gains, including slashing peak GPU memory demand by over 97% and projection time from 8.1 seconds to 0.2 seconds compared to standard VAE decoding.

Method: Proposes ReconPhys with dual-branch architecture trained via self-supervised strategy to jointly learn physical attribute estimation and 3D Gaussian Splatting reconstruction from single monocular video, eliminating need for ground-truth physics labels.

Result: Achieves 21.64 PSNR in future prediction vs 13.27 by SOTA optimization baselines, reduces Chamfer Distance from 0.349 to 0.004, enables fast inference (<1 second) vs hours for existing methods.

Conclusion: ReconPhys enables rapid generation of simulation-ready assets for robotics and graphics through fast, physics-aware 3D reconstruction from monocular video without manual annotation.

Abstract: Reconstructing non-rigid objects with physical plausibility remains a significant challenge. Existing approaches leverage differentiable rendering for per-scene optimization, recovering geometry and dynamics but requiring expensive tuning or manual annotation, which limits practicality and generalizability. To address this, we propose ReconPhys, the first feedforward framework that jointly learns physical attribute estimation and 3D Gaussian Splatting reconstruction from a single monocular video. Our method employs a dual-branch architecture trained via a self-supervised strategy, eliminating the need for ground-truth physics labels. Given a video sequence, ReconPhys simultaneously infers geometry, appearance, and physical attributes. Experiments on a large-scale synthetic dataset demonstrate superior performance: our method achieves 21.64 PSNR in future prediction compared to 13.27 by state-of-the-art optimization baselines, while reducing Chamfer Distance from 0.349 to 0.004. Crucially, ReconPhys enables fast inference (<1 second) versus hours required by existing methods, facilitating rapid generation of simulation-ready assets for robotics and graphics.

Method: Proposes a reinforcement-guided synthetic data generation framework with: 1) cold-start adaptation to align pretrained generator with target domain, 2) multi-objective reward optimizing semantic consistency, coverage diversity, and expression richness, and 3) dynamic sample selection during downstream training to prioritize high-utility synthetic samples.

Result: Extensive experiments on benchmark datasets demonstrate significant improvements in both generation fidelity and classification accuracy, with strong generalization to novel categories in small-data regimes.

Conclusion: The framework successfully breaks the cycle of data scarcity in privacy-sensitive scenarios by adapting general-domain generative priors to specific identity recognition tasks through reinforcement-guided synthetic data generation.

Abstract: High-fidelity generative models are increasingly needed in privacy-sensitive scenarios, where access to data is severely restricted due to regulatory and copyright constraints. This scarcity hampers model development–ironically, in settings where generative models are most needed to compensate for the lack of data. This creates a self-reinforcing challenge: limited data leads to poor generative models, which in turn fail to mitigate data scarcity. To break this cycle, we propose a reinforcement-guided synthetic data generation framework that adapts general-domain generative priors to privacy-sensitive identity recognition tasks. We first perform a cold-start adaptation to align a pretrained generator with the target domain, establishing semantic relevance and initial fidelity. Building on this foundation, we introduce a multi-objective reward that jointly optimizes semantic consistency, coverage diversity, and expression richness, guiding the generator to produce both realistic and task-effective samples. During downstream training, a dynamic sample selection mechanism further prioritizes high-utility synthetic samples, enabling adaptive data scaling and improved domain alignment. Extensive experiments on benchmark datasets demonstrate that our framework significantly improves both generation fidelity and classification accuracy, while also exhibiting strong generalization to novel categories in small-data regimes.

Method: Study sampling geometry impact on spatial statistics, introduce geometry-aware reconstruction module linking cell projections across adjacent sections using phenotype and proximity constraints, recover 3D centroids using cell-type-specific shape priors, analyze trade-offs between section spacing, coverage, and redundancy.

Result: Planar sampling reliably recovers global cell-type abundance but exhibits high variance for local statistics like cell clustering and cell-cell interactions. Reconstruction module validated on public imaging mass cytometry dataset and demonstrated utility on in-house CODEX dataset enabling structure-level 3D analyses.

Conclusion: Provides diagnostic tools and practical guidance for deciding when 2D sampling suffices and when sparse 3D reconstruction is warranted, with acquisition regimes that maximize reconstruction utility under fixed imaging budgets.

Abstract: Highly multiplexed microscopy enables rich spatial characterization of tissues at single-cell resolution, yet most analyses rely on two-dimensional sections despite inherently three-dimensional tissue organization. Acquiring dense volumetric data in spatial proteomics remains costly and technically challenging, leaving practitioners to choose between 2D sections or 3D serial sections under limited imaging budgets. In this work, we study how sampling geometry impacts the stability of commonly used spatial statistics, and we introduce a geometry-aware reconstruction module that enables sparse yet consistent 3D analysis from serial sections. Using controlled simulations, we show that planar sampling reliably recovers global cell-type abundance but exhibits high variance for local statistics such as cell clustering and cell-cell interactions, particularly for rare or spatially localized populations. We observe consistent behavior in real multiplexed datasets, where interaction metrics and neighborhood relationships fluctuate substantially across individual sections. To support sparse 3D analysis in practice, we present a reconstruction approach that links cell projections across adjacent sections using phenotype and proximity constraints and recovers single-cell 3D centroids using cell-type-specific shape priors. We further analyze the trade-off between section spacing, coverage, and redundancy, identifying acquisition regimes that maximize reconstruction utility under fixed imaging budgets. We validate the reconstruction module on a public imaging mass cytometry dataset with dense axial sampling and demonstrate its downstream utility on an in-house CODEX dataset by enabling structure-level 3D analyses that are unreliable in 2D. Together, our results provide diagnostic tools and practical guidance for deciding when 2D sampling suffices and when sparse 3D reconstruction is warranted.

Method: AnomalyAgent uses five tools (Prompt Generation, Image Generation, Quality Evaluation, Knowledge Retrieval, Mask Generation) in a closed-loop optimization framework. It employs a two-stage training approach: supervised fine-tuning followed by reinforcement learning with a three-part reward mechanism (task, reflection, and behavioral rewards).

Result: On MVTec-AD dataset: achieves IS/IC-L of 2.10/0.33 for anomaly generation, 57.0% classification accuracy using ResNet34, and 99.3%/74.2% AP at image/pixel level using UNet, surpassing all zero-shot SOTA methods.

Conclusion: AnomalyAgent demonstrates superior performance in generating realistic and diverse industrial anomalies through its agent-based framework with self-reflection and iterative refinement capabilities.

Abstract: Industrial anomaly generation is a crucial method for alleviating the data scarcity problem in anomaly detection tasks. Most existing anomaly synthesis methods rely on single-step generation mechanisms, lacking complex reasoning and iterative optimization capabilities, making it difficult to generate anomaly samples with high semantic realism. We propose AnomalyAgent, an anomaly synthesis agent with self-reflection, knowledge retrieval, and iterative refinement capabilities, aiming to generate realistic and diverse anomalies. Specifically, AnomalyAgent is equipped with five tools: Prompt Generation (PG), Image Generation (IG), Quality Evaluation (QE), Knowledge Retrieval (KR), and Mask Generation (MG), enabling closed-loop optimization. To improve decision-making and self-reflection, we construct structured trajectories from real anomaly images and design a two-stage training framework: supervised fine-tuning followed by reinforcement learning. This process is driven by a three-part reward mechanism: (1) task rewards to supervise the quality and location rationality of generated anomalies; (2) reflection rewards to train the model’s ability to improve anomaly synthesis prompt; (3) behavioral rewards to ensure adherence to the trajectory. On the MVTec-AD dataset, AnomalyAgent achieves IS/IC-L of 2.10/0.33 for anomaly generation, 57.0% classification accuracy using ResNet34, and 99.3%/74.2% AP at the image/pixel level using a simple UNet, surpassing all zero-shot SOTA methods. The code and data will be made publicly available.

Method: Proposes PanoSAM2 with: 1) Pano-Aware Decoder with seam-consistent receptive fields and iterative distortion refinement for boundary continuity; 2) Distortion-Guided Mask Loss weighting pixels by distortion magnitude; 3) Long-Short Memory Module maintaining compact long-term object pointers to re-instantiate short-term memories for temporal coherence.

Result: PanoSAM2 achieves substantial improvements over SAM2: +5.6 on 360VOTS and +6.7 on PanoVOS benchmarks, demonstrating effectiveness of the adaptation strategies.

Conclusion: The proposed distortion- and memory-aware adaptation strategies successfully adapt SAM2 for reliable 360VOS while retaining its user-friendly prompting design, addressing key challenges in 360° video segmentation.

Abstract: 360 video object segmentation (360VOS) aims to predict temporally-consistent masks in 360 videos, offering full-scene coverage, benefiting applications, such as VR/AR and embodied AI. Learning 360VOS model is nontrivial due to the lack of high-quality labeled dataset. Recently, Segment Anything Models (SAMs), especially SAM2 – with its design of memory module – shows strong, promptable VOS capability. However, directly using SAM2 for 360VOS yields implausible results as 360 videos suffer from the projection distortion, semantic inconsistency of left-right sides, and sparse object mask information in SAM2’s memory. To this end, we propose PanoSAM2, a novel 360VOS framework based on our lightweight distortion- and memory-aware adaptation strategies of SAM2 to achieve reliable 360VOS while retaining SAM2’s user-friendly prompting design. Concretely, to tackle the projection distortion and semantic inconsistency issues, we propose a Pano-Aware Decoder with seam-consistent receptive fields and iterative distortion refinement to maintain continuity across the 0/360 degree boundary. Meanwhile, a Distortion-Guided Mask Loss is introduced to weight pixels by distortion magnitude, stressing stretched regions and boundaries. To address the object sparsity issue, we propose a Long-Short Memory Module to maintain a compact long-term object pointer to re-instantiate and align short-term memories, thereby enhancing temporal coherence. Extensive experiments show that PanoSAM2 yields substantial gains over SAM2: +5.6 on 360VOTS and +6.7 on PanoVOS, showing the effectiveness of our method.

Method: Repurposes idle compute during low-risk autonomous vehicle states (red lights, traffic congestion, parking-lot crawl) to run a quantized 7B Vision-Language Model on pre-cached satellite and street view imagery to identify entrances and legal parking zones.

Result: The quantized 7B VLM completes inference in 4-8 seconds on HW4-class hardware, and the system estimates annual per-driver income gains of $3,000-8,000 USD in the U.S. market.

Conclusion: ParkSense demonstrates the feasibility of using VLMs for precision parking in food delivery contexts, identifying an unexplored intersection of autonomous driving, computer vision, and last-mile logistics with five open research directions.

Abstract: Finding parking consumes a disproportionate share of food delivery time, yet no system addresses precise parking-spot selection relative to merchant entrances. We propose ParkSense, a framework that repurposes idle compute during low-risk AV states – queuing at red lights, traffic congestion, parking-lot crawl – to run a Vision-Language Model (VLM) on pre-cached satellite and street view imagery, identifying entrances and legal parking zones. We formalize the Delivery-Aware Precision Parking (DAPP) problem, show that a quantized 7B VLM completes inference in 4-8 seconds on HW4-class hardware, and estimate annual per-driver income gains of 3,000-8,000 USD in the U.S. Five open research directions are identified at this unexplored intersection of autonomous driving, computer vision, and last-mile logistics.

Method: MESA performs controlled and selective latent intervention targeting hallucination-relevant responses while preserving the model’s original token distribution. It’s a plug-and-play framework that disentangles steering signals to avoid disrupting intrinsic generation behavior.

Result: Extensive experiments across diverse generative and discriminative benchmarks show MESA consistently reduces hallucinations while better preserving generation behavior, outperforming prior methods across multiple LVLM families.

Conclusion: MESA effectively addresses the hallucination problem in LVLMs through selective latent interventions that maintain original generation behavior, providing a superior solution to existing mitigation approaches.

Abstract: Large Vision-Language Models (LVLMs) have achieved remarkable success across cross-modal tasks but remain hindered by hallucinations, producing textual outputs inconsistent with visual content. Existing methods mitigate hallucinations but often alter generation behavior, resulting in shorter outputs and shifted token distributions, especially in latent space steering approaches. We identify that this issue stems from entangled steering signals, where suppressing hallucinations inadvertently disrupts the model’s intrinsic generation behavior. To address this, we propose MESA, an effective plug-and-play framework that performs controlled and selective latent intervention for hallucination mitigation. Specifically, MESA targets hallucination-relevant responses while preserving the model’s original token distribution, enabling effective hallucination reduction without compromising generation behavior. Extensive experiments across diverse generative and discriminative benchmarks demonstrate that MESA consistently reduces hallucinations while better preserving generation behavior, outperforming prior methods across multiple LVLM families.

Method: Two-phase framework: 1) Expression Reasoning Interpreter (ERI) uses reasoning-assisted prompts for structured parsing and rephrasing to generate robust prompts for SAM3; 2) Mask Self-Refining (MSR) selects best mask and refines using DINOv3 features to correct over/under-segmentation.

Result: Achieves strong performance on both explicit and implicit RES benchmarks and open-world scenarios, with ablation studies validating each phase’s effectiveness.

Conclusion: Tarot-SAM3 provides an effective training-free solution for generalizable referring expression segmentation by combining reasoning-assisted prompting with self-refinement techniques.

Abstract: Referring Expression Segmentation (RES) aims to segment image regions described by natural-language expressions, serving as a bridge between vision and language understanding. Existing RES methods, however, rely heavily on large annotated datasets and are limited to either explicit or implicit expressions, hindering their ability to generalize to any referring expression. Recently, the Segment Anything Model 3 (SAM3) has shown impressive robustness in Promptable Concept Segmentation. Nonetheless, applying it to RES remains challenging: (1) SAM3 struggles with longer or implicit expressions; (2) naive coupling of SAM3 with a multimodal large language model (MLLM) makes the final results overly dependent on the MLLM’s reasoning capability, without enabling refinement of SAM3’s segmentation outputs. To this end, we present Tarot-SAM3, a novel training-free framework that can accurately segment from any referring expression. Specifically, Tarot-SAM3 consists of two key phases. First, the Expression Reasoning Interpreter (ERI) phase introduces reasoning-assisted prompt options to support structured expression parsing and evaluation-aware rephrasing. This transforms arbitrary queries into robust heterogeneous prompts for generating reliable masks with SAM3. Second, the Mask Self-Refining (MSR) phase selects the best mask across prompt types and performs self-refinement by leveraging rich feature relationships from DINOv3 to compare discriminative regions among ERI outputs. It then infers region affiliation to the target, thereby correcting over- and under-segmentation. Extensive experiments demonstrate that Tarot-SAM3 achieves strong performance on both explicit and implicit RES benchmarks, as well as open-world scenarios. Ablation studies further validate the effectiveness of each phase.

Method: Stitch4D uses two key techniques: (1) synthesizes intermediate bridge views to densify spatial constraints and improve coverage, and (2) jointly optimizes real and synthesized observations in a unified coordinate frame with explicit inter-location consistency constraints.

Result: Stitch4D outperforms existing 4D reconstruction baselines on the new Urban Sparse 4D benchmark, achieving superior visual quality and coherent geometry with smooth scene dynamics in sparse urban environments.

Conclusion: Recovering intermediate spatial coverage is essential for stable 4D reconstruction in sparse urban environments. Stitch4D demonstrates that explicit compensation for missing spatial coverage enables coherent reconstruction even with limited camera overlap.

Abstract: Dynamic urban environments are often captured by cameras placed at spatially separated locations with little or no view overlap. However, most existing 4D reconstruction methods assume densely overlapping views. When applied to such sparse observations, these methods fail to reconstruct intermediate regions and often introduce temporal artifacts. To address this practical yet underexplored sparse multi-location setting, we propose Stitch4D, a unified 4D reconstruction framework that explicitly compensates for missing spatial coverage in sparse observations. Stitch4D (i) synthesizes intermediate bridge views to densify spatial constraints and improve spatial coverage, and (ii) jointly optimizes real and synthesized observations within a unified coordinate frame under explicit inter-location consistency constraints. By restoring intermediate coverage before optimization, Stitch4D prevents geometric collapse and reconstructs coherent geometry and smooth scene dynamics even in sparsely observed environments. To evaluate this setting, we introduce Urban Sparse 4D (U-S4D), a CARLA-based benchmark designed to assess spatiotemporal alignment under sparse multi-location configurations. Experimental results on U-S4D show that Stitch4D surpasses representative 4D reconstruction baselines and achieves superior visual quality. These results indicate that recovering intermediate spatial coverage is essential for stable 4D reconstruction in sparse urban environments.

Method: Treats latitude-longitude grid points as 3D Gaussian centers. Introduces generative 3D Gaussian prediction scheme to estimate covariance, attributes, and opacity parameters for unseen samples. Uses scale-aware attention module to capture cross-scale dependencies for varying downscaling ratios and continuous resolution adaptation.

Result: Accurately forecasts 87 atmospheric variables at arbitrary resolutions on ERA5 data. Demonstrates superior performance in downscaling tasks on both ERA5 and CMIP6 datasets compared to existing methods.

Conclusion: First NWP approach combining generative 3D Gaussian modeling with scale-aware attention for unified multi-scale prediction. Provides efficient and scalable solution for high-resolution, multi-scale atmospheric prediction and downscaling.

Abstract: While AI-based numerical weather prediction (NWP) enables rapid forecasting, generating high-resolution outputs remains computationally demanding due to limited multi-scale adaptability and inefficient data representations. We propose the 3D Gaussian splatting-based scale-aware vision transformer (GSSA-ViT), a novel framework for arbitrary-resolution forecasting and flexible downscaling of high-dimensional atmospheric fields. Specifically, latitude-longitude grid points are treated as centers of 3D Gaussians. A generative 3D Gaussian prediction scheme is introduced to estimate key parameters, including covariance, attributes, and opacity, for unseen samples, improving generalization and mitigating overfitting. In addition, a scale-aware attention module is designed to capture cross-scale dependencies, enabling the model to effectively integrate information across varying downscaling ratios and support continuous resolution adaptation. To our knowledge, this is the first NWP approach that combines generative 3D Gaussian modeling with scale-aware attention for unified multi-scale prediction. Experiments on ERA5 show that the proposed method accurately forecasts 87 atmospheric variables at arbitrary resolutions, while evaluations on ERA5 and CMIP6 demonstrate its superior performance in downscaling tasks. The proposed framework provides an efficient and scalable solution for high-resolution, multi-scale atmospheric prediction and downscaling. Code is available at: https://github.com/binbin2xs/weather-GS.

Method: Three approaches: (1) stain-only classification to confirm stain learnability; (2) dual-head model with supervised stain loss; (3) dual-head model with label-free stain regularization via entropy maximization on stain head.

Result: Stain identity is easily learnable, but lesion classifiers on this multi-stain dataset show no measurable stain-driven shortcut learning. Entropy regularization keeps stain predictions near chance without degrading lesion accuracy.

Conclusion: Carefully curated multi-stain datasets can be inherently robust to stain shortcuts, and Bayesian dual-head architecture with label-free entropy regularization provides a simple safeguard against stain-related drift.

Abstract: Stain variability is a pervasive source of distribution shift and potential shortcut learning in renal pathology AI. We ask whether lupus nephritis glomerular lesion classifiers exploit stain as a shortcut, and how to mitigate such bias without stain or site labels. We curate a multi-center, multi-stain dataset of 9{,}674 glomerular patches (224$\times$224) from 365 WSIs across three centers and four stains (PAS, H&E, Jones, Trichrome), labeled as proliferative vs.\ non-proliferative. We evaluate Bayesian CNN and ViT backbones with Monte Carlo dropout in three settings: (1) stain-only classification; (2) a dual-head model jointly predicting lesion and stain with supervised stain loss; and (3) a dual-head model with label-free stain regularization via entropy maximization on the stain head. In (1), stain identity is trivially learnable, confirming a strong candidate shortcut. In (2), varying the strength and sign of stain supervision strongly modulates stain performance but leaves lesion metrics essentially unchanged, indicating no measurable stain-driven shortcut learning on this multi-stain, multi-center dataset, while overly adversarial stain penalties inflate predictive uncertainty. In (3), entropy-based regularization holds stain predictions near chance without degrading lesion accuracy or calibration. Overall, a carefully curated multi-stain dataset can be inherently robust to stain shortcuts, and a Bayesian dual-head architecture with label-free entropy regularization offers a simple, deployment-friendly safeguard against potential stain-related drift in glomerular AI.

Method: Freezes pre-trained 3D attention modules, treats images as single-frame videos, and uses Predict-Update Spatial Difference Attention module to extract/inject spatial differences. Incorporates Text-Guided Dynamic Semantic Gating for adaptive text-driven modifications without rigid external masks.

Result: Achieves editing fidelity and temporal consistency comparable to larger models trained on extensive video datasets, despite training on only 13K image pairs for 5 epochs with low computational overhead.

Conclusion: ImVideoEdit demonstrates efficient video editing can be learned from image pairs by decoupling spatiotemporal processes, offering a scalable alternative to video data-dependent approaches.

Abstract: Current video editing models often rely on expensive paired video data, which limits their practical scalability. In essence, most video editing tasks can be formulated as a decoupled spatiotemporal process, where the temporal dynamics of the pretrained model are preserved while spatial content is selectively and precisely modified. Based on this insight, we propose ImVideoEdit, an efficient framework that learns video editing capabilities entirely from image pairs. By freezing the pre-trained 3D attention modules and treating images as single-frame videos, we decouple the 2D spatial learning process to help preserve the original temporal dynamics. The core of our approach is a Predict-Update Spatial Difference Attention module that progressively extracts and injects spatial differences. Rather than relying on rigid external masks, we incorporate a Text-Guided Dynamic Semantic Gating mechanism for adaptive and implicit text-driven modifications. Despite training on only 13K image pairs for 5 epochs with exceptionally low computational overhead, ImVideoEdit achieves editing fidelity and temporal consistency comparable to larger models trained on extensive video datasets.

Method: Proposes ToolCAD framework deploying LLMs as tool-using agents for text-to-CAD generation, with interactive CAD modeling gym for rollout trajectories, hybrid feedback, human supervision, and end-to-end post-training via online curriculum reinforcement learning to develop CAD Modeling Chain of Thought (CAD-CoT).

Result: ToolCAD fills the gap in adopting and training open-source LLMs for CAD tool-using agents, enabling them to perform comparably to proprietary models, paving the way for accessible autonomous text-to-CAD systems.

Conclusion: The framework enables LLMs to become proficient CAD tool-using agents through interactive training and reinforcement learning, making autonomous text-to-CAD modeling more accessible and robust.

Abstract: Computer-Aided Design (CAD) is an expert-level task that relies on long-horizon reasoning and coherent modeling actions. Large Language Models (LLMs) have shown remarkable advancements in enabling language agents to tackle real-world tasks. Notably, there has been no investigation into how tool-using LLMs optimally interact with CAD engines, hindering the emergence of LLM-based agentic text-to-CAD modeling systems. We propose ToolCAD, a novel agentic CAD framework deploying LLMs as tool-using agents for text-to-CAD generation. Furthermore, we introduce an interactive CAD modeling gym to rollout reasoning and tool-augmented interaction trajectories with the CAD engine, incorporating hybrid feedback and human supervision. Meanwhile, an end-to-end post-training strategy is presented to enable the LLM agent to elicit refined CAD Modeling Chain of Thought (CAD-CoT) and evolve into proficient CAD tool-using agents via online curriculum reinforcement learning. Our findings demonstrate ToolCAD fills the gap in adopting and training open-source LLMs for CAD tool-using agents, enabling them to perform comparably to proprietary models, paving the way for more accessible and robust autonomous text-to-CAD modeling systems.

Method: DSCA decomposes VLM representation space into orthogonal semantic subspaces through incremental clustering and PCA on joint vision-language representations. Edits are proposed only in transformed spaces, structurally isolating concepts. A multi-term loss function maintains task fidelity, edit locality, and cross-modal alignment while keeping the base model frozen.

Result: Achieves 98% single edit success, remains over 95% after 1000 sequential edits, reduces hallucination by 3-5%, and achieves best backward transfer scores on continual instruction tuning benchmarks. Demonstrates state-of-the-art stability and knowledge retention in lifelong editing across various datasets.

Conclusion: Structural concept isolation through orthogonal semantic subspaces is superior to optimization-based approaches for lifelong VLM knowledge editing, enabling precise, non-interfering edits while maintaining cross-modal alignment and preventing catastrophic forgetting.

Abstract: Model editing aims to update knowledge to add new concepts and change relevant information without retraining. Lifelong editing is a challenging task, prone to disrupting previously learned concepts, especially for Vision Language Models (VLMs), because sequential edits can lead to degraded reasoning and cross modal misalignment. Existing VLM knowledge editing methods based on gated adapters, activation edits, and parameter merging techniques address catastrophic forgetting seen in full fine tuning; however, they still operate in the shared representation space of the VLM, where concepts are entangled, so edits interfere with other non relevant concepts. We hypothesize that this instability persists because current methods algorithmically control edits via optimization rather than structurally separating knowledge. We introduce Dynamic Subspace Concept Alignment (DSCA) which by design mitigates this limitation by decomposing the representation space into a set of orthogonal semantic subspaces and proposing edits only in those transformed spaces. These subspaces are obtained through incremental clustering and PCA on joint vision language representations. This process structurally isolates concepts, enabling precise, non interfering edits by turning isolation from a soft training objective into an architectural property. The surgical edits are guided by a multi term loss function for maintaining task fidelity, edit locality, and cross modal alignment. With the base model frozen, our method achieves 98 percent single edit success, remains over 95 percent after 1000 sequential edits, lowers hallucination by 3 to 5 percent, and achieves the best backward transfer (BWT) scores on continual instruction tuning benchmarks. Extensive experiments demonstrate DSCA state of the art stability and knowledge retention capability in continual lifelong editing across various datasets and benchmarks.

Method: Introduces a novel framework that conditions video synthesis on explicit 3D scene properties using a new large-scale dataset with dense annotations. Disentangles properties by rendering control signals from a unified 3D representation, uses a lightweight conditioning module and progressive training strategy to integrate signals into a foundational video diffusion model.

Result: Achieves state-of-the-art photorealism and temporal consistency while enabling precise, disentangled control over scene factors. Enables applications like image-to-video and video-to-video synthesis with fully editable underlying 3D scenes.

Conclusion: LiVER sets a new standard for controllable video generation by providing explicit control over 3D scene properties through a unified representation, with a scene agent that translates high-level user instructions into required control signals.

Abstract: Diffusion models have achieved remarkable progress in video generation, but their controllability remains a major limitation. Key scene factors such as layout, lighting, and camera trajectory are often entangled or only weakly modeled, restricting their applicability in domains like filmmaking and virtual production where explicit scene control is essential. We present LiVER, a diffusion-based framework for scene-controllable video generation. To achieve this, we introduce a novel framework that conditions video synthesis on explicit 3D scene properties, supported by a new large-scale dataset with dense annotations of object layout, lighting, and camera parameters. Our method disentangles these properties by rendering control signals from a unified 3D representation. We propose a lightweight conditioning module and a progressive training strategy to integrate these signals into a foundational video diffusion model, ensuring stable convergence and high fidelity. Our framework enables a wide range of applications, including image-to-video and video-to-video synthesis where the underlying 3D scene is fully editable. To further enhance usability, we develop a scene agent that automatically translates high-level user instructions into the required 3D control signals. Experiments show that LiVER achieves state-of-the-art photorealism and temporal consistency while enabling precise, disentangled control over scene factors, setting a new standard for controllable video generation.

Method: Three complementary approaches integrated: 1) Dense BM/SGM disparity, 2) Novel object-centric Census-based template matching with GPU acceleration, far-close divide-and-conquer strategy, forward-backward verification, occlusion-aware sampling, and multi-block aggregation, 3) Monocular geometric priors. Plus online calibration refinement with auto-rectification offset search, radar-stereo voting, and object-level radar-stereo association.

Result: Achieves real-time performance through asynchronous GPU pipeline design. Delivers robust ranging across diverse conditions including nighttime, rain, and varying illumination. System provides accurate depth estimation for long-range vehicle detection.

Conclusion: The integrated stereo ranging system successfully addresses limitations of traditional methods by combining multiple depth estimation approaches with online calibration refinement, enabling robust real-time vehicle detection for autonomous driving applications.

Abstract: Accurate depth estimation is critical for autonomous driving perception systems, particularly for long range vehicle detection on highways. Traditional dense stereo matching methods such as Block Matching (BM) and Semi Global Matching (SGM) produce per pixel disparity maps but suffer from high computational cost, sensitivity to radiometric differences between stereo cameras, and poor accuracy at long range where disparity values are small. In this report, we present a comprehensive stereo ranging system that integrates three complementary depth estimation approaches: dense BM/SGM disparity, object centric Census based template matching, and monocular geometric priors, within a unified detection ranging tracking pipeline. Our key contribution is a novel object centric Census based template matching algorithm that performs GPU accelerated sparse stereo matching directly within detected bounding boxes, employing a far close divide and conquer strategy, forward backward verification, occlusion aware sampling, and robust multi block aggregation. We further describe an online calibration refinement framework that combines auto rectification offset search, radar stereo voting based disparity correction, and object level radar stereo association for continuous extrinsic drift compensation. The complete system achieves real time performance through asynchronous GPU pipeline design and delivers robust ranging across diverse driving conditions including nighttime, rain, and varying illumination.

Method: Initializes unified 3D Gaussian set from COLMAP priors, augments each with learnable category probability, and dynamically routes them into specialized deformation branches for background, hands, or object modeling. Uses category-specific masks for better disentanglement and introduces brightness and motion-flow control to improve static rendering and dynamic reconstruction.

Result: Outperforms baselines by +1.70dB in PSNR on average with SSIM and LPIPS gains. Achieves state-of-the-art disentanglement of background, hand, and object components, enabling explicit, fine-grained separation.

Conclusion: DP-DeGauss enables more intuitive ego scene understanding and editing by achieving the first and state-of-the-art disentanglement of background, hand, and object components in egocentric 4D reconstruction.

Abstract: Egocentric video is crucial for next-generation 4D scene reconstruction, with applications in AR/VR and embodied AI. However, reconstructing dynamic first-person scenes is challenging due to complex ego-motion, occlusions, and hand-object interactions. Existing decomposition methods are ill-suited, assuming fixed viewpoints or merging dynamics into a single foreground. To address these limitations, we introduce DP-DeGauss, a dynamic probabilistic Gaussian decomposition framework for egocentric 4D reconstruction. Our method initializes a unified 3D Gaussian set from COLMAP priors, augments each with a learnable category probability, and dynamically routes them into specialized deformation branches for background, hands, or object modeling. We employ category-specific masks for better disentanglement and introduce brightness and motion-flow control to improve static rendering and dynamic reconstruction. Extensive experiments show that DP-DeGauss outperforms baselines by +1.70dB in PSNR on average with SSIM and LPIPS gains. More importantly, our framework achieves the first and state-of-the-art disentanglement of background, hand, and object components, enabling explicit, fine-grained separation, paving the way for more intuitive ego scene understanding and editing.

Method: Created SceneScribe-1M dataset with 1 million in-the-wild videos, each annotated with detailed textual descriptions, precise camera parameters, dense depth maps, and consistent 3D point tracks.

Result: Established benchmarks across multiple downstream tasks including monocular depth estimation, scene reconstruction, dynamic point tracking, and text-to-video synthesis with/without camera control.

Conclusion: SceneScribe-1M serves as a comprehensive benchmark and catalyst for research, enabling development of models that can both perceive dynamic 3D worlds and generate controllable, realistic video content.

Abstract: The convergence of 3D geometric perception and video synthesis has created an unprecedented demand for large-scale video data that is rich in both semantic and spatio-temporal information. While existing datasets have advanced either 3D understanding or video generation, a significant gap remains in providing a unified resource that supports both domains at scale. To bridge this chasm, we introduce SceneScribe-1M, a new large-scale, multi-modal video dataset. It comprises one million in-the-wild videos, each meticulously annotated with detailed textual descriptions, precise camera parameters, dense depth maps, and consistent 3D point tracks. We demonstrate the versatility and value of SceneScribe-1M by establishing benchmarks across a wide array of downstream tasks, including monocular depth estimation, scene reconstruction, and dynamic point tracking, as well as generative tasks such as text-to-video synthesis, with or without camera control. By open-sourcing SceneScribe-1M, we aim to provide a comprehensive benchmark and a catalyst for research, fostering the development of models that can both perceive the dynamic 3D world and generate controllable, realistic video content.

Method: Proposes Selective Aggregation Transformer (SAT) with Density-driven Token Aggregation algorithm that selectively aggregates key-value matrices (reducing tokens by 97%) while maintaining full query resolution. Uses density and isolation metrics to represent clusters with single aggregation tokens, preserving high-frequency details.

Result: SAT outperforms state-of-the-art method PFT by up to 0.22dB while reducing total FLOPs by up to 27%. The approach achieves better reconstruction fidelity with significantly lower computational costs.

Conclusion: SAT provides an efficient transformer architecture for image super-resolution that enables scalable global interactions without compromising reconstruction quality, addressing the computational complexity challenges of traditional transformer approaches.

Abstract: Transformer-based approaches have revolutionized image super-resolution by modeling long-range dependencies. However, the quadratic computational complexity of vanilla self-attention mechanisms poses significant challenges, often leading to compromises between efficiency and global context exploitation. Recent window-based attention methods mitigate this by localizing computations, but they often yield restricted receptive fields. To mitigate these limitations, we propose Selective Aggregation Transformer (SAT). This novel transformer efficiently captures long-range dependencies, leading to an enlarged model receptive field by selectively aggregating key-value matrices (reducing the number of tokens by 97%) via our Density-driven Token Aggregation algorithm while maintaining the full resolution of the query matrix. This design significantly reduces computational costs, resulting in lower complexity and enabling scalable global interactions without compromising reconstruction fidelity. SAT identifies and represents each cluster with a single aggregation token, utilizing density and isolation metrics to ensure that critical high-frequency details are preserved. Experimental results demonstrate that SAT outperforms the state-of-the-art method PFT by up to 0.22dB, while the total number of FLOPs can be reduced by up to 27%.

Method: FI3Det introduces: 1) VLM-guided unknown object learning module that mines unknown objects and extracts 2D semantic features + class-agnostic 3D boxes, with a weighting mechanism to mitigate noise; 2) Gated multimodal prototype imprinting module that constructs category prototypes from aligned 2D semantic and 3D geometric features, fusing classification scores via multimodal gating.

Result: FI3Det achieves strong and consistent improvements over baseline methods on both batch and sequential evaluation settings on ScanNet V2 and SUN RGB-D datasets, establishing the first framework for few-shot incremental 3D object detection.

Conclusion: FI3Det successfully addresses the limitation of extensive annotation requirements in incremental 3D detection by leveraging vision-language models for few-shot learning, enabling efficient 3D perception in dynamic indoor environments.

Abstract: Incremental 3D object perception is a critical step toward embodied intelligence in dynamic indoor environments. However, existing incremental 3D detection methods rely on extensive annotations of novel classes for satisfactory performance. To address this limitation, we propose FI3Det, a Few-shot Incremental 3D Detection framework that enables efficient 3D perception with only a few novel samples by leveraging vision-language models (VLMs) to learn knowledge of unseen categories. FI3Det introduces a VLM-guided unknown object learning module in the base stage to enhance perception of unseen categories. Specifically, it employs VLMs to mine unknown objects and extract comprehensive representations, including 2D semantic features and class-agnostic 3D bounding boxes. To mitigate noise in these representations, a weighting mechanism is further designed to re-weight the contributions of point- and box-level features based on their spatial locations and feature consistency within each box. Moreover, FI3Det proposes a gated multimodal prototype imprinting module, where category prototypes are constructed from aligned 2D semantic and 3D geometric features to compute classification scores, which are then fused via a multimodal gating mechanism for novel object detection. As the first framework for few-shot incremental 3D object detection, we establish both batch and sequential evaluation settings on two datasets, ScanNet V2 and SUN RGB-D, where FI3Det achieves strong and consistent improvements over baseline methods. Code is available at https://github.com/zyrant/FI3Det.

Method: Created SearchAD dataset by curating over 423k frames from 11 established AD datasets, providing manual annotations of 513k bounding boxes covering 90 rare categories. Designed for text-to-image and image-to-image retrieval with well-defined data splits for few-shot learning and fine-tuning of multi-modal retrieval models.

Result: Text-based methods outperform image-based ones due to stronger semantic grounding. Models aligning spatial visual features with language achieve best zero-shot results, and fine-tuning baselines significantly improve performance, though absolute retrieval capabilities remain unsatisfactory.

Conclusion: SearchAD establishes the first large-scale dataset for retrieval-driven data curation and long-tail perception research in autonomous driving, addressing the critical need to efficiently locate rare scenarios in massive datasets.

Abstract: Retrieving rare and safety-critical driving scenarios from large-scale datasets is essential for building robust autonomous driving (AD) systems. As dataset sizes continue to grow, the key challenge shifts from collecting more data to efficiently identifying the most relevant samples. We introduce SearchAD, a large-scale rare image retrieval dataset for AD containing over 423k frames drawn from 11 established datasets. SearchAD provides high-quality manual annotations of more than 513k bounding boxes covering 90 rare categories. It specifically targets the needle-in-a-haystack problem of locating extremely rare classes, with some appearing fewer than 50 times across the entire dataset. Unlike existing benchmarks, which focused on instance-level retrieval, SearchAD emphasizes semantic image retrieval with a well-defined data split, enabling text-to-image and image-to-image retrieval, few-shot learning, and fine-tuning of multi-modal retrieval models. Comprehensive evaluations show that text-based methods outperform image-based ones due to stronger inherent semantic grounding. While models directly aligning spatial visual features with language achieve the best zero-shot results, and our fine-tuning baseline significantly improves performance, absolute retrieval capabilities remain unsatisfactory. With a held-out test set on a public benchmark server, SearchAD establishes the first large-scale dataset for retrieval-driven data curation and long-tail perception research in AD: https://iis-esslingen.github.io/searchad/

Method: Bridge-STG decouples temporal and spatial localization while maintaining semantic coherence through two key designs: 1) Spatio-Temporal Semantic Bridging (STSB) with Explicit Temporal Alignment (ETA) that distills the MLLM’s temporal reasoning context into enriched bridging queries as a semantic interface, and 2) Query-Guided Spatial Localization (QGSL) module that leverages these queries to drive a purpose-built spatial decoder with multi-layer interactive queries and positive/negative frame sampling.

Result: Extensive experiments across multiple benchmarks show Bridge-STG achieves state-of-the-art performance among MLLM-based methods. It improves average m_vIoU from 26.4 to 34.3 on VidSTG and demonstrates strong cross-task transfer across various fine-grained video understanding tasks under unified multi-task training.

Conclusion: Bridge-STG effectively addresses the core challenges of spatio-temporal video grounding in MLLMs by decoupling temporal and spatial localization while maintaining semantic coherence, achieving superior performance and demonstrating strong generalization capabilities across related tasks.

Abstract: Spatio-Temporal Video Grounding requires jointly localizing target objects across both temporal and spatial dimensions based on natural language queries, posing fundamental challenges for existing Multimodal Large Language Models (MLLMs). We identify two core challenges: \textit{entangled spatio-temporal alignment}, arising from coupling two heterogeneous sub-tasks within the same autoregressive output space, and \textit{dual-domain visual token redundancy}, where target objects exhibit simultaneous temporal and spatial sparsity, rendering the overwhelming majority of visual tokens irrelevant to the grounding query. To address these, we propose \textbf{Bridge-STG}, an end-to-end framework that decouples temporal and spatial localization while maintaining semantic coherence. While decoupling is the natural solution to this entanglement, it risks creating a semantic gap between the temporal MLLM and the spatial decoder. Bridge-STG resolves this through two pivotal designs: the \textbf{Spatio-Temporal Semantic Bridging (STSB)} mechanism with Explicit Temporal Alignment (ETA) distills the MLLM’s temporal reasoning context into enriched bridging queries as a robust semantic interface; and the \textbf{Query-Guided Spatial Localization (QGSL)} module leverages these queries to drive a purpose-built spatial decoder with multi-layer interactive queries and positive/negative frame sampling, jointly eliminating dual-domain visual token redundancy. Extensive experiments across multiple benchmarks demonstrate that Bridge-STG achieves state-of-the-art performance among MLLM-based methods. Bridge-STG improves average m_vIoU from $26.4$ to $34.3$ on VidSTG and demonstrates strong cross-task transfer across various fine-grained video understanding tasks under a unified multi-task training regime.

Method: Proposes CATMIL objective that augments base segmentation loss with two auxiliary terms: 1) Component-Adaptive Tversky loss that reweights voxel contributions based on connected components to balance lesion size influence, and 2) Multiple Instance Learning term for lesion-level supervision. Combined with standard nnU-Net loss in a consistent framework.

Result: On MSLesSeg dataset with 5-fold cross-validation, CATMIL achieves most balanced performance: Dice score 0.7834, improved boundary error, substantially increased small lesion recall, reduced false negatives, and lowest false positive volume among compared methods.

Conclusion: Integrating component-level and lesion-level supervision within a unified objective provides effective and practical approach for improving small lesion segmentation in highly imbalanced medical imaging settings.

Abstract: We propose a unified objective function, termed CATMIL, that augments the base segmentation loss with two auxiliary supervision terms operating at different levels. The first term, Component-Adaptive Tversky, reweights voxel contributions based on connected components to balance the influence of lesions of different sizes. The second term, based on Multiple Instance Learning, introduces lesion-level supervision by encouraging the detection of each lesion instance. These terms are combined with the standard nnU-Net loss to jointly optimize voxel-level segmentation accuracy and lesion-level detection. We evaluate the proposed objective on the MSLesSeg dataset using a consistent nnU-Net framework and 5-fold cross-validation. The results show that CATMIL achieves the most balanced performance across segmentation accuracy, lesion detection, and error control. It improves Dice score (0.7834) and reduces boundary error compared to standard losses. More importantly, it substantially increases small lesion recall and reduces false negatives, while maintaining the lowest false positive volume among compared methods. These findings demonstrate that integrating component-level and lesion-level supervision within a unified objective provides an effective and practical approach for improving small lesion segmentation in highly imbalanced settings. All code and pretrained models are available at \href{https://github.com/luumsk/SmallLesionMRI}{this url}.

Method: Replace standard encoders with rotation-equivariant convolutions in three baseline deformable brain MRI registration architectures, evaluating on multiple public brain MRI datasets.

Result: Equivariant encoders achieve: 1) higher registration accuracy with fewer parameters, 2) better performance on rotated input pairs demonstrating robustness to orientation variations, and 3) improved performance with less training data indicating greater sample efficiency.

Conclusion: Incorporating geometric priors like rotation equivariance is critical for building more robust, accurate, and efficient medical image registration models.

Abstract: Image registration is a fundamental task that aligns anatomical structures between images. While CNNs perform well, they lack rotation equivariance - a rotated input does not produce a correspondingly rotated output. This hinders performance by failing to exploit the rotational symmetries inherent in anatomical structures, particularly in brain MRI. In this work, we integrate rotation-equivariant convolutions into deformable brain MRI registration networks. We evaluate this approach by replacing standard encoders with equivariant ones in three baseline architectures, testing on multiple public brain MRI datasets. Our experiments demonstrate that equivariant encoders have three key advantages: 1) They achieve higher registration accuracy while reducing network parameters, confirming the benefit of this anatomical inductive bias. 2) They outperform baselines on rotated input pairs, demonstrating robustness to orientation variations common in clinical practice. 3) They show improved performance with less training data, indicating greater sample efficiency. Our results demonstrate that incorporating geometric priors is a critical step toward building more robust, accurate, and efficient registration models.

Method: Proposes MDDCNet with hybrid backbone using Multi-Scale Deformable Dilated Convolution blocks and Mamba blocks, Channel-Enhanced Feed-Forward Network, and Mamba-based Attention-Aggregating Feature Pyramid Network for enhanced feature fusion.

Result: Extensive experiments on public benchmarks and real-world datasets demonstrate superiority over various advanced detectors.

Conclusion: MDDCNet effectively addresses limitations of Mamba-based methods for traffic object detection through hierarchical feature representation and enhanced multi-scale fusion.

Abstract: In a real-world traffic scenario, varying-scale objects are usually distributed in a cluttered background, which poses great challenges to accurate detection. Although current Mamba-based methods can efficiently model long-range dependencies, they still struggle to capture small objects with abundant local details, which hinders joint modeling of local structures and global semantics. Moreover, state-space models exhibit limited hierarchical feature representation and weak cross-scale interaction due to flat sequential modeling and insufficient spatial inductive biases, leading to sub-optimal performance in complex scenes. To address these issues, we propose a Mamba with Deformable Dilated Convolutions Network (MDDCNet) for accurate traffic object detection in this study. In MDDCNet, a well-designed hybrid backbone with successive Multi-Scale Deformable Dilated Convolution (MSDDC) blocks and Mamba blocks enables hierarchical feature representation from local details to global semantics. Meanwhile, a Channel-Enhanced Feed-Forward Network (CE-FFN) is further devised to overcome the limited channel interaction capability of conventional feed-forward networks, whilst a Mamba-based Attention-Aggregating Feature Pyramid Network (A^2FPN) is constructed to achieve enhanced multi-scale feature fusion and interaction. Extensive experimental results on public benchmark and real-world datasets demonstrate the superiority of our method over various advanced detectors. The code is available at https://github.com/Bettermea/MDDCNet.

Method: LINE uses a large language model and text-to-image generator in an iterative closed loop to propose and refine concepts, guided by activation history, operating in a strictly black-box setting without requiring model training.

Result: Achieves state-of-the-art performance with AUC improvements up to 0.18 on ImageNet and 0.05 on Places365, discovers 29% new concepts missed by predefined vocabularies, and provides complete generation history for polysemanticity evaluation.

Conclusion: LINE enables more comprehensive and interpretable concept discovery in vision models through open-vocabulary labeling, supporting better understanding of neural network decision-making processes.

Abstract: Interpreting the concepts encoded by individual neurons in deep neural networks is a crucial step towards understanding their complex decision-making processes and ensuring AI safety. Despite recent progress in neuron labeling, existing methods often limit the search space to predefined concept vocabularies or produce overly specific descriptions that fail to capture higher-order, global concepts. We introduce LINE, a novel, training-free iterative approach tailored for open-vocabulary concept labeling in vision models. Operating in a strictly black-box setting, LINE leverages a large language model and a text-to-image generator to iteratively propose and refine concepts in a closed loop, guided by activation history. We demonstrate that LINE achieves state-of-the-art performance across multiple model architectures, yielding AUC improvements of up to 0.18 on ImageNet and 0.05 on Places365, while discovering, on average, 29% of new concepts missed by massive predefined vocabularies. Beyond identifying the top concept, LINE provides a complete generation history, which enables polysemanticity evaluation and produces supporting visual explanations that rival gradient-dependent activation maximization methods.

Method: Uses LLMs to sequentially draw 3D Bezier curves with geometric feedback. Introduces relative experience optimization adapting GRPO paradigm: constructs pairwise sketch comparisons (better vs worse) based on CLIP perceptual rewards and LLM qualitative assessment, then uses these experiences to iteratively refine 3D drawing knowledge without parameter updates.

Result: Generates complex, coherent 3D Bezier sketches from diverse textual prompts, exhibits emergent geometric reasoning, and generalizes to novel shapes. Establishes new paradigm for training-free 3D sketch intelligence.

Conclusion: 3DrawAgent enables language-driven 3D sketch generation without training, advancing training-free 3D sketch intelligence through relative experience optimization and self-improvement mechanisms.

Abstract: Sketching in 3D space enables expressive reasoning about shape, structure, and spatial relationships, yet generating 3D sketches through natural language remains a major challenge. In this work, we introduce 3DrawAgent, a training-free, language-driven framework for 3D sketch generation that leverages large language models (LLMs) to sequentially draw 3D Bezier curves under geometric feedback. Unlike prior 2D sketch agents, our method introduces a relative experience optimization strategy that adapts the recently proposed Group Reward Policy Optimization (GRPO) paradigm. Instead of relying on explicit ground-truth supervision, we construct pairwise comparisons among generated sketches, with each pair consisting of a relatively better and a worse result based on CLIP-based perceptual rewards and LLM-based fine-grained qualitative assessment. These experiences are then used to iteratively refine the prior knowledge of 3D drawing, enabling black-box reinforcement of the model’s 3D awareness. This design allows our model to self-improve its spatial understanding and drawing quality without parameter updates. Experiments show that 3DrawAgent can generate complex and coherent 3D Bezier sketches from diverse textual prompts, exhibit emergent geometric reasoning, and generalize to novel shapes, establishing a new paradigm for advancing the field of training-free 3D sketch intelligence.

Method: Proposes a label-efficient segmentation framework leveraging pretrained DINOv3 foundational vision transformer backbone integrated with a Dense Prediction Transformer (DPT)-style decoder to hierarchically reassemble multi-scale features, combining global semantic representations with fine-grained spatial details.

Result: Achieves state-of-the-art performance on clinical dataset of 7,777 annotated frames from 112 patients: Dice score of 0.945, 11.4% reduction in 95th-percentile Hausdorff Distance compared to strongest convolutional baseline. Maintains strong performance even with only 25% of training data.

Conclusion: Leveraging large-scale self-supervised foundational models provides a promising and data-efficient solution for medical image segmentation in data-constrained clinical environments, demonstrating superior performance and efficiency over traditional supervised approaches.

Abstract: Adnexal mass evaluation via ultrasound is a challenging clinical task, often hindered by subjective interpretation and significant inter-observer variability. While automated segmentation is a foundational step for quantitative risk assessment, traditional fully supervised convolutional architectures frequently require large amounts of pixel-level annotations and struggle with domain shifts common in medical imaging. In this work, we propose a label-efficient segmentation framework that leverages the robust semantic priors of a pretrained DINOv3 foundational vision transformer backbone. By integrating this backbone with a Dense Prediction Transformer (DPT)-style decoder, our model hierarchically reassembles multi-scale features to combine global semantic representations with fine-grained spatial details. Evaluated on a clinical dataset of 7,777 annotated frames from 112 patients, our method achieves state-of-the-art performance compared to established fully supervised baselines, including U-Net, U-Net++, DeepLabV3, and MAnet. Specifically, we obtain a Dice score of 0.945 and improved boundary adherence, reducing the 95th-percentile Hausdorff Distance by 11.4% relative to the strongest convolutional baseline. Furthermore, we conduct an extensive efficiency analysis demonstrating that our DINOv3-based approach retains significantly higher performance under data starvation regimes, maintaining strong results even when trained on only 25% of the data. These results suggest that leveraging large-scale self-supervised foundations provides a promising and data-efficient solution for medical image segmentation in data-constrained clinical environments. Project Repository: https://github.com/FrancescaFati/MESA

Method: SSG generates perturbed predictions by swapping pairs of most semantically dissimilar token latents in spatial or channel dimensions. This creates a direction between clean and perturbed predictions to steer sampling toward higher-fidelity distributions, with finer control than global perturbation methods.

Result: Experiments on MS-COCO 2014, MS-COCO 2017, and ImageNet show SSG outperforms previous condition-free methods in image fidelity and prompt alignment. It also improves robustness with reduced side-effects across a wider range of perturbation strengths.

Conclusion: SSG extends CFG to both conditional and unconditional generation, serving as a plug-in for any diffusion model to gain immediate improvements in image quality without requiring text conditions.

Abstract: Classifier-Free Guidance (CFG) is a widely used inference-time technique to boost the image quality of diffusion models. Yet, its reliance on text conditions prevents its use in unconditional generation. We propose a simple method to enable CFG-like guidance for both conditional and unconditional generation. The key idea is to generate a perturbed prediction via simple token swap operations, and use the direction between it and the clean prediction to steer sampling towards higher-fidelity distributions. In practice, we swap pairs of most semantically dissimilar token latents in either spatial or channel dimensions. Unlike existing methods that apply perturbation in a global or less constrained manner, our approach selectively exchanges and recomposes token latents, allowing finer control over perturbation and its influence on generated samples. Experiments on MS-COCO 2014, MS-COCO 2017, and ImageNet datasets demonstrate that the proposed Self-Swap Guidance (SSG), when applied to popular diffusion models, outperforms previous condition-free methods in image fidelity and prompt alignment under different set-ups. Its fine-grained perturbation granularity also improves robustness, reducing side-effects across a wider range of perturbation strengths. Overall, SSG extends CFG to a broader scope of applications including both conditional and unconditional generation, and can be readily inserted into any diffusion model as a plug-in to gain immediate improvements.

Method: Proposes Aligned Hierarchical Bidirectional Scan Mamba (ABMamba) that extends Deep State Space Models as language backbone, replacing quadratic attention. Uses novel Aligned Hierarchical Bidirectional Scan module to process videos across multiple temporal resolutions with linear complexity.

Result: On VATEX and MSR-VTT benchmarks, ABMamba demonstrates competitive performance compared to typical MLLMs while achieving approximately three times higher throughput.

Conclusion: ABMamba provides an efficient alternative to attention-based MLLMs for video captioning with linear computational complexity, enabling scalable processing of long video sequences while maintaining competitive performance.

Abstract: In this study, we focus on video captioning by fully open multimodal large language models (MLLMs). The comprehension of visual sequences is challenging because of their intricate temporal dependencies and substantial sequence length. The core attention mechanisms of existing Transformer-based approaches scale quadratically with the sequence length, making them computationally prohibitive. To address these limitations, we propose Aligned Hierarchical Bidirectional Scan Mamba (ABMamba), a fully open MLLM with linear computational complexity that enables the scalable processing of video sequences. ABMamba extends Deep State Space Models as its language backbone, replacing the costly quadratic attention mechanisms, and employs a novel Aligned Hierarchical Bidirectional Scan module that processes videos across multiple temporal resolutions. On standard video captioning benchmarks such as VATEX and MSR-VTT, ABMamba demonstrates competitive performance compared to typical MLLMs while achieving approximately three times higher throughput.

Method: Two-stage approach: 1) EEG-conditioned diffusion reconstruction from EEG signals, 2) Prompt-guided post-reconstruction boosting using multimodal LLM to extract semantic descriptions and image-to-image diffusion to refine geometry and perceptual coherence while preserving EEG-grounded structure.

Result: Semantic decoding accuracy degrades significantly with channel reduction (50-way Top-1 Acc from 89% to 38%), while reconstruction quality slightly decreases (FID from 76.77 to 80.51). Boosting consistently improves perceptual metrics across all configurations, achieving up to 9.71% IS gains in low-channel settings. User study confirms perceptual preference for boosted reconstructions.

Conclusion: The approach significantly boosts feasibility of real-time brain-to-image applications using low-resolution EEG devices, potentially unlocking applications outside laboratory settings through effective reconstruction and enhancement.

Abstract: Reconstructing visual stimuli from non-invasive electroencephalography (EEG) remains challenging due to its low spatial resolution and high noise, particularly under realistic low-density electrode configurations. To address this, we present EEG2Vision, a modular, end-to-end EEG-to-image framework that systematically evaluates reconstruction performance across different EEG resolutions (128, 64, 32, and 24 channels) and enhances visual quality through a prompt-guided post-reconstruction boosting mechanism. Starting from EEG-conditioned diffusion reconstruction, the boosting stage uses a multimodal large language model to extract semantic descriptions and leverages image-to-image diffusion to refine geometry and perceptual coherence while preserving EEG-grounded structure. Our experiments show that semantic decoding accuracy degrades significantly with channel reduction (e.g., 50-way Top-1 Acc from 89% to 38%), while reconstruction quality slight decreases (e.g., FID from 76.77 to 80.51). The proposed boosting consistently improves perceptual metrics across all configurations, achieving up to 9.71% IS gains in low-channel settings. A user study confirms the clear perceptual preference for boosted reconstructions. The proposed approach significantly boosts the feasibility of real-time brain-2-image applications using low-resolution EEG devices, potentially unlocking this type of applications outside laboratory settings.

Method: Multimodal pipeline: 1) EEG-to-image decoding, 2) MLLM extracts structured 3D-aware descriptions, 3) diffusion-based generation, 4) single-image-to-3D model converts to coherent 3D meshes.

Result: Achieves 85.4% 10-way Top-1 EEG decoding accuracy and 0.648 CLIPScore, demonstrating strong performance in semantic alignment and geometric fidelity compared to original visual stimuli.

Conclusion: Brain3D enables scalable brain-driven 3D generation by decomposing EEG-to-3D mapping into structured stages, supporting feasibility of multimodal EEG-driven 3D reconstruction.

Abstract: Decoding visual information from electroencephalography (EEG) has recently achieved promising results, primarily focusing on reconstructing two-dimensional (2D) images from brain activity. However, the reconstruction of three-dimensional (3D) representations remains largely unexplored. This limits the geometric understanding and reduces the applicability of neural decoding in different contexts. To address this gap, we propose Brain3D, a multimodal architecture for EEG-to-3D reconstruction based on EEG-to-image decoding. It progressively transforms neural representations into the 3D domain using geometry-aware generative reasoning. Our pipeline first produces visually grounded images from EEG signals, then employs a multimodal large language model to extract structured 3D-aware descriptions, which guide a diffusion-based generation stage whose outputs are finally converted into coherent 3D meshes via a single-image-to-3D model. By decomposing the problem into structured stages, the proposed approach avoids direct EEG-to-3D mappings and enables scalable brain-driven 3D generation. We conduct a comprehensive evaluation comparing the reconstructed 3D outputs against the original visual stimuli, assessing both semantic alignment and geometric fidelity. Experimental results demonstrate strong performance of the proposed architecture, achieving up to 85.4% 10-way Top-1 EEG decoding accuracy and 0.648 CLIPScore, supporting the feasibility of multimodal EEG-driven 3D reconstruction.

Method: Fine-tuned Qwen2.5-VL 3B VLM using QLoRA and Unsloth for parameter-efficient training, with a unique Darija dataset combining synthetic generation (OCRSmith library) and real-world data.

Result: Achieved state-of-the-art performance on newly curated AtlasOCRBench and established KITAB-Bench, challenging larger models and demonstrating robust generalization for both Darija and standard Arabic OCR.

Conclusion: AtlasOCR successfully addresses the OCR gap for Darija while also performing well on standard Arabic, showing the effectiveness of fine-tuning VLMs for specialized OCR tasks.

Abstract: Darija, the Moroccan Arabic dialect, is rich in visual content yet lacks specialized Optical Character Recognition (OCR) tools. This paper introduces AtlasOCR, the first open-source Darija OCR model built by fine-tuning a 3B parameter Vision Language Model (VLM). We detail our comprehensive approach, from curating a unique Darija-specific dataset leveraging both synthetic generation with our OCRSmith library and carefully sourced real-world data, to implementing efficient fine-tuning strategies. We utilize QLoRA and Unsloth for parameter-efficient training of Qwen2.5-VL 3B and present comprehensive ablation studies optimizing key hyperparameters. Our evaluation on the newly curated AtlasOCRBench and the established KITAB-Bench demonstrates state-of-the-art performance, challenging larger models and highlighting AtlasOCR’s robustness and generalization capabilities for both Darija and standard Arabic OCR tasks.

Method: Proposes Tensor-Augmented CNN (TACNN) which replaces conventional convolution kernels with generic tensors. This leverages the fact that an order-N tensor naturally encodes quantum superposition states in Hilbert space, offering richer expressivity. The convolution output becomes a multilinear form capable of capturing high-order feature correlations.

Result: On Fashion-MNIST, TACNN demonstrates clear advantages over conventional CNNs. A TACNN with only two convolution layers achieves 93.7% test accuracy, surpassing or matching considerably deeper models like VGG-16 (93.5%) and GoogLeNet (93.7%).

Conclusion: TACNN is a promising framework that strengthens model expressivity while preserving architectural simplicity, paving the way toward more interpretable and efficient deep learning models.

Abstract: Convolutional Neural Networks (CNNs) excel at extracting local features hierarchically, but their performance in capturing complex correlations hinges heavily on deep architectures, which are usually computationally demanding and difficult to interpret. To address these issues, we propose a physically-guided shallow model: tensor-augmented CNN (TACNN), which replaces conventional convolution kernels with generic tensors to enhance representational capacity. This choice is motivated by the fact that an order-$N$ tensor naturally encodes an arbitrary quantum superposition state in the Hilbert space of dimension $d^N$, where $d$ is the local physical dimension, thus offering substantially richer expressivity. Furthermore, in our design the convolution output of each layer becomes a multilinear form capable of capturing high-order feature correlations, thereby equipping a shallow multilayer architecture with an expressive power competitive to that of deep CNNs. On the Fashion-MNIST benchmark, TACNN demonstrates clear advantages over conventional CNNs, achieving remarkable accuracies with only a few layers. In particular, a TACNN with only two convolution layers attains a test accuracy of 93.7$%$, surpassing or matching considerably deeper models such as VGG-16 (93.5$%$) and GoogLeNet (93.7$%$). These findings highlight TACNN as a promising framework that strengthens model expressivity while preserving architectural simplicity, paving the way towards more interpretable and efficient deep learning models.

Method: DinoRADE processes dense radar tensors and aggregates vision features around transformed reference points in camera perspective via deformable cross-attention. Vision features are provided by DINOv3 Vision Foundation Model.

Result: Comprehensive evaluation on K-Radar dataset in all weather conditions, reporting detection performance individually for five object classes. Outperforms recent radar-camera approaches by 12.1%.

Conclusion: DinoRADE demonstrates effective radar-centered detection with vision foundation model integration for improved VRU detection in adverse weather, addressing limitations of existing radar-based approaches.

Abstract: Reliable and weather-robust perception systems are essential for safe autonomous driving and typically employ multi-modal sensor configurations to achieve comprehensive environmental awareness. While recent automotive FMCW Radar-based approaches achieved remarkable performance on detection tasks in adverse weather conditions, they exhibited limitations in resolving fine-grained spatial details particularly critical for detecting smaller and vulnerable road users (VRUs). Furthermore, existing research has not adequately addressed VRU detection in adverse weather datasets such as K-Radar. We present DinoRADE, a Radar-centered detection pipeline that processes dense Radar tensors and aggregates vision features around transformed reference points in the camera perspective via deformable cross-attention. Vision features are provided by a DINOv3 Vision Foundation Model. We present a comprehensive performance evaluation on the K-Radar dataset in all weather conditions and are among the first to report detection performance individually for five object classes. Additionally, we compare our method with existing single-class detection approaches and outperform recent Radar-camera approaches by 12.1%. The code is available under https://github.com/chr-is-tof/RADE-Net.

Method: Uses a Small Vision-Language Model as local temporal compressor for cross-modal distillation, generating compact intent-aligned representations. Introduces Adaptive Token Allocation (ATA) as training-free dynamic router that allocates dense bandwidth to query-critical segments while compressing redundancies into minimal temporal anchors.

Result: Achieves SOTA performance with aggressive dynamic compression (0.5-16 tokens/frame). On LVBench (4101s), scores 52.3 under 8K visual budget, outperforming GPT-4o and Gemini 1.5 Pro. Scaling to 2048 frames reaches 53.7. Compresses hour-long videos below theoretical limits.

Conclusion: True long-form video understanding requires intent-driven efficiency rather than greedily padded context windows. Tempo demonstrates effective compression while maintaining understanding quality.

Abstract: Adapting Multimodal Large Language Models (MLLMs) for hour-long videos is bottlenecked by context limits. Dense visual streams saturate token budgets and exacerbate the lost-in-the-middle phenomenon. Existing heuristics, like sparse sampling or uniform pooling, blindly sacrifice fidelity by discarding decisive moments and wasting bandwidth on irrelevant backgrounds. We propose Tempo, an efficient query-aware framework compressing long videos for downstream understanding. Tempo leverages a Small Vision-Language Model (SVLM) as a local temporal compressor, casting token reduction as an early cross-modal distillation process to generate compact, intent-aligned representations in a single forward pass. To enforce strict budgets without breaking causality, we introduce Adaptive Token Allocation (ATA). Exploiting the SVLM’s zero-shot relevance prior and semantic front-loading, ATA acts as a training-free $O(1)$ dynamic router. It allocates dense bandwidth to query-critical segments while compressing redundancies into minimal temporal anchors to maintain the global storyline. Extensive experiments show our 6B architecture achieves state-of-the-art performance with aggressive dynamic compression (0.5-16 tokens/frame). On the extreme-long LVBench (4101s), Tempo scores 52.3 under a strict 8K visual budget, outperforming GPT-4o and Gemini 1.5 Pro. Scaling to 2048 frames reaches 53.7. Crucially, Tempo compresses hour-long videos substantially below theoretical limits, proving true long-form video understanding relies on intent-driven efficiency rather than greedily padded context windows.

Method: AdaSpark partitions videos into 3D spatio-temporal cubes and uses two co-designed components: Adaptive Cube-Selective Attention (AdaS-Attn) that selects relevant video cubes per query token, and Adaptive Token-Selective FFN (AdaS-FFN) that processes only salient tokens within each cube, with entropy-based Top-p selection for adaptive resource allocation.

Result: AdaSpark reduces computational load by up to 57% FLOPs while maintaining comparable performance to dense models and preserving fine-grained, long-range dependencies on challenging hour-scale video benchmarks.

Conclusion: AdaSpark provides an effective adaptive sparsity framework that addresses computational efficiency limitations in Video-LLMs while preserving both fine-grained perception and long-range temporal modeling capabilities.

Abstract: Processing long-form videos with Video Large Language Models (Video-LLMs) is computationally prohibitive. Current efficiency methods often compromise fine-grained perception through irreversible information disposal or inhibit long-range temporal modeling via rigid, predefined sparse patterns. This paper introduces AdaSpark, an adaptive sparsity framework designed to address these limitations. AdaSpark first partitions video inputs into 3D spatio-temporal cubes. It then employs two co-designed, context-aware components: (1) Adaptive Cube-Selective Attention (AdaS-Attn), which adaptively selects a subset of relevant video cubes to attend for each query token, and (2) Adaptive Token-Selective FFN (AdaS-FFN), which selectively processes only the most salient tokens within each cube. An entropy-based (Top-p) selection mechanism adaptively allocates computational resources based on input complexity. Experiments demonstrate that AdaSpark significantly reduces computational load by up to 57% FLOPs while maintaining comparable performance to dense models and preserving fine-grained, long-range dependencies, as validated on challenging hour-scale video benchmarks.

Method: Proposes a non-autoregressive framework using diffusion models for video captioning. Encodes video into visual representations, adds Gaussian noise to text representations during training, uses a discriminative denoiser with visual conditioning to generate new text representations, and employs a non-autoregressive language model for caption generation.

Result: Outperforms previous non-autoregressive methods and achieves comparable performance to autoregressive methods on MSVD, MSR-VTT, and VATEX datasets, with up to 9.9 improvement on CIDEr and 2.6 improvement on BLEU@4, while maintaining faster generation speed.

Conclusion: Diffusion models provide an effective non-autoregressive approach for video captioning that addresses speed and error accumulation issues while maintaining generation quality through improved multimodal interaction modeling.

Abstract: Current video captioning methods usually use an encoder-decoder structure to generate text autoregressively. However, autoregressive methods have inherent limitations such as slow generation speed and large cumulative error. Furthermore, the few non-autoregressive counterparts suffer from deficiencies in generation quality due to the lack of sufficient multimodal interaction modeling. Therefore, we propose a non-autoregressive framework based on Diffusion model for Video Captioning (DiffVC) to address these issues. Its parallel decoding can effectively solve the problems of generation speed and cumulative error. At the same time, our proposed discriminative conditional Diffusion Model can generate higher-quality textual descriptions. Specifically, we first encode the video into a visual representation. During training, Gaussian noise is added to the textual representation of the ground-truth caption. Then, a new textual representation is generated via the discriminative denoiser with the visual representation as a conditional constraint. Finally, we input the new textual representation into a non-autoregressive language model to generate captions. During inference, we directly sample noise from the Gaussian distribution for generation. Experiments on MSVD, MSR-VTT, and VATEX show that our method can outperform previous non-autoregressive methods and achieve comparable performance to autoregressive methods, e.g., it achieved a maximum improvement of 9.9 on the CIDEr and improvement of 2.6 on the B@4, while having faster generation speed. The source code will be available soon.

Method: Conducted comprehensive evaluation of state-of-the-art VLMs (Gemini 3.1 Pro, Qwen3-VL, InternVL3.5) across multiple activity domains (fitness, figure skating, diving), tasks, representations, and prompting strategies. Tested various approaches including skeleton information incorporation, grounding instructions, reasoning structures, and in-context learning. Analyzed prediction distributions to identify systematic biases.

Result: VLMs performed only marginally above random chance. Strategies like skeleton information and in-context learning led to isolated gains but none were consistently effective. Identified two systematic biases: tendency to predict correct execution regardless of visual evidence, and sensitivity to superficial linguistic framing. Reformulating tasks contrastively yielded minimal improvement.

Conclusion: VLMs have fundamental limitations in fine-grained movement quality assessment beyond identified biases. Findings establish rigorous baseline for future VLM-based AQA research and outline failure modes requiring mitigation before reliable real-world deployment.

Abstract: Action Quality Assessment (AQA) has broad applications in physical therapy, sports coaching, and competitive judging. Although Vision Language Models (VLMs) hold considerable promise for AQA, their actual performance in this domain remains largely uncharacterised. We present a comprehensive evaluation of state-of-the-art VLMs across activity domains (e.g. fitness, figure skating, diving), tasks, representations, and prompting strategies. Baseline results reveal that Gemini 3.1 Pro, Qwen3-VL and InternVL3.5 models perform only marginally above random chance, and although strategies such as incorporation of skeleton information, grounding instructions, reasoning structures and in-context learning lead to isolated gains, none is consistently effective. Analysis of prediction distributions uncovers two systematic biases: a tendency to predict correct execution regardless of visual evidence, and a sensitivity to superficial linguistic framing. Reformulating tasks contrastively to mitigate these biases yields minimal improvement, suggesting that the models’ limitations go beyond these biases, pointing to a fundamental difficulty with fine-grained movement quality assessment. Our findings establish a rigorous baseline for future VLM-based AQA research and provide an actionable outline for failure modes requiring mitigation prior to reliable real-world deployment.

Method: Proposes CDAMD (Coordinate-based Dual-constrained Autoregressive Motion Generation) that uses motion coordinates as input, follows autoregressive paradigm, and leverages diffusion-inspired multi-layer perceptrons. Introduces Dual-Constrained Causal Mask where motion tokens act as priors and are concatenated with textual encodings to guide autoregressive generation.

Result: Achieves state-of-the-art performance in terms of both fidelity and semantic consistency on newly established benchmarks for text-to-motion generation and motion editing.

Conclusion: CDAMD provides a flexible, high-fidelity framework for text-to-motion generation that addresses limitations of both diffusion and autoregressive approaches through coordinate-based dual-constrained autoregressive modeling.

Abstract: Text-to-motion generation has attracted increasing attention in the research community recently, with potential applications in animation, virtual reality, robotics, and human-computer interaction. Diffusion and autoregressive models are two popular and parallel research directions for text-to-motion generation. However, diffusion models often suffer from error amplification during noise prediction, while autoregressive models exhibit mode collapse due to motion discretization. To address these limitations, we propose a flexible, high-fidelity, and semantically faithful text-to-motion framework, named Coordinate-based Dual-constrained Autoregressive Motion Generation (CDAMD). With motion coordinates as input, CDAMD follows the autoregressive paradigm and leverages diffusion-inspired multi-layer perceptrons to enhance the fidelity of predicted motions. Furthermore, a Dual-Constrained Causal Mask is introduced to guide autoregressive generation, where motion tokens act as priors and are concatenated with textual encodings. Since there is limited work on coordinate-based motion synthesis, we establish new benchmarks for both text-to-motion generation and motion editing. Experimental results demonstrate that our approach achieves state-of-the-art performance in terms of both fidelity and semantic consistency on these benchmarks.

Method: Proposes EPIR framework with: 1) Dual Norm Shifted Patch Tokenization (DNSPT) for spatial relationships, 2) Token integration module to reduce tokens without information loss, 3) Discriminative token extractor with improved attention and dynamic token selection.

Result: Achieves significant performance gains: 9.6% improvement on CAS(ME)³ dataset in UF1 and 4.58% improvement on SMIC dataset in UAR metric compared to state-of-the-art methods.

Conclusion: EPIR framework effectively balances high recognition performance and low computational complexity for micro-expression recognition on small datasets.

Abstract: Micro-expression recognition can obtain the real emotion of the individual at the current moment. Although deep learning-based methods, especially Transformer-based methods, have achieved impressive results, these methods have high computational complexity due to the large number of tokens in the multi-head self-attention. In addition, the existing micro-expression datasets are small-scale, which makes it difficult for Transformer-based models to learn effective micro-expression representations. Therefore, we propose a novel Efficient Patch tokenization, Integration and Representation framework (EPIR), which can balance high recognition performance and low computational complexity. Specifically, we first propose a dual norm shifted tokenization (DNSPT) module to learn the spatial relationship between neighboring pixels in the face region, which is implemented by a refined spatial transformation and dual norm projection. Then, we propose a token integration module to integrate partial tokens among multiple cascaded Transformer blocks, thereby reducing the number of tokens without information loss. Furthermore, we design a discriminative token extractor, which first improves the attention in the Transformer block to reduce the unnecessary focus of the attention calculation on self-tokens, and uses the dynamic token selection module (DTSM) to select key tokens, thereby capturing more discriminative micro-expression representations. We conduct extensive experiments on four popular public datasets (i.e., CASME II, SAMM, SMIC, and CAS(ME)3. The experimental results show that our method achieves significant performance gains over the state-of-the-art methods, such as 9.6% improvement on the CAS(ME)$^3$ dataset in terms of UF1 and 4.58% improvement on the SMIC dataset in terms of UAR metric.

Method: OV-Stitcher stitches fragmented sub-image features directly within the final encoder block, reconstructing attention representations to enable global attention within that block, producing coherent context aggregation and spatially consistent segmentation maps.

Result: Extensive evaluations across eight benchmarks show OV-Stitcher improves mean Intersection over Union (mIoU) from 48.7 to 50.7 compared with prior training-free baselines, establishing a scalable and effective solution.

Conclusion: OV-Stitcher addresses the limitation of fragmented feature representations in training-free open-vocabulary segmentation by enabling global attention through feature stitching, achieving better performance without additional training.

Abstract: Training-free open-vocabulary semantic segmentation(TF-OVSS) has recently attracted attention for its ability to perform dense prediction by leveraging the pretrained knowledge of large vision and vision-language models, without requiring additional training. However, due to the limited input resolution of these pretrained encoders, existing TF-OVSS methods commonly adopt a sliding-window strategy that processes cropped sub-images independently. While effective for managing high-resolution inputs, this approach prevents global attention over the full image, leading to fragmented feature representations and limited contextual reasoning. We propose OV-Stitcher, a training-free framework that addresses this limitation by stitching fragmented sub-image features directly within the final encoder block. By reconstructing attention representations from fragmented sub-image features, OV-Stitcher enables global attention within the final encoder block, producing coherent context aggregation and spatially consistent, semantically aligned segmentation maps. Extensive evaluations across eight benchmarks demonstrate that OV-Stitcher establishes a scalable and effective solution for open-vocabulary segmentation, achieving a notable improvement in mean Intersection over Union(mIoU) from 48.7 to 50.7 compared with prior training-free baselines.

Method: Proposes a model-intrinsic grounding method using uncertainty as supervision: computes entropy of next-token distribution, backpropagates to visual token embeddings to get entropy-gradient relevance maps, extracts/ranks multiple coherent regions, and uses iterative zoom-and-reground procedure with spatial-entropy stopping rule.

Result: Experiments on seven benchmarks across four VLM architectures show consistent improvements over existing methods, with largest gains on detail-critical and high-resolution settings, while producing more interpretable evidence localizations.

Conclusion: The proposed training-free grounding method effectively addresses the limitations of current VLMs in handling fine-grained visual details and multi-evidence reasoning by using model uncertainty as intrinsic supervision for evidence retrieval.

Abstract: Despite rapid progress, pretrained vision-language models still struggle when answers depend on tiny visual details or on combining clues spread across multiple regions, as in documents and compositional queries. We address this by framing grounding as test-time evidence retrieval: given a query, the model should actively identify where to look next to resolve ambiguity. To this end, we propose a training-free, model-intrinsic grounding method that uses uncertainty as supervision. Specifically, we compute the entropy of the model’s next-token distribution and backpropagate it to the visual token embeddings to obtain an entropy-gradient relevance map, without auxiliary detectors or attention-map heuristics. We then extract and rank multiple coherent regions to support multi-evidence queries, and introduce an iterative zoom-and-reground procedure with a spatial-entropy stopping rule to avoid over-refinement. Experiments on seven benchmarks across four VLM architectures demonstrate consistent improvements over existing methods, with the largest gains on detail-critical and high-resolution settings, while also producing more interpretable evidence localizations.

Method: 1) Unified flow matching for continuous video and discrete text generation; 2) Modality-driven MoE framework with lightweight text generation layers; 3) Bidirectional training with Knowledge Recall (reconstructing input prompts) and Capability Refinement (fine-tuning on detailed captions).

Result: Uni-ViGU achieves competitive performance on both video generation and understanding tasks, validating generation-centric architectures as a scalable approach to unified multimodal intelligence.

Conclusion: Extending video generators as foundation models provides a viable path for unified multimodal systems, addressing computational efficiency while maintaining strong performance across generation and understanding tasks.

Abstract: Unified multimodal models integrating visual understanding and generation face a fundamental challenge: visual generation incurs substantially higher computational costs than understanding, particularly for video. This imbalance motivates us to invert the conventional paradigm: rather than extending understanding-centric MLLMs to support generation, we propose Uni-ViGU, a framework that unifies video generation and understanding by extending a video generator as the foundation. We introduce a unified flow method that performs continuous flow matching for video and discrete flow matching for text within a single process, enabling coherent multimodal generation. We further propose a modality-driven MoE-based framework that augments Transformer blocks with lightweight layers for text generation while preserving generative priors. To repurpose generation knowledge for understanding, we design a bidirectional training mechanism with two stages: Knowledge Recall reconstructs input prompts to leverage learned text-video correspondences, while Capability Refinement fine-tunes on detailed captions to establish discriminative shared representations. Experiments demonstrate that Uni-ViGU achieves competitive performance on both video generation and understanding, validating generation-centric architectures as a scalable path toward unified multimodal intelligence. Project Page and Code: https://fr0zencrane.github.io/uni-vigu-page/.

Method: Proposes PolySLGen with pose fusion module and social cue encoder to jointly aggregate motion and social signals from all participants in group interactions, generating future speaking or listening reactions including speech, body motion, and speaking state score.

Result: Extensive experiments show PolySLGen outperforms adapted and state-of-the-art baselines in motion quality, motion-speech alignment, speaking state prediction, and human-perceived realism, producing contextually appropriate and temporally coherent multimodal reactions.

Conclusion: PolySLGen effectively models group interactions for multimodal reaction generation, addressing limitations of existing approaches and enabling more natural human-AI interactions in realistic social scenarios.

Abstract: Human-like multimodal reaction generation is essential for natural group interactions between humans and embodied AI. However, existing approaches are limited to single-modality or speaking-only responses in dyadic interactions, making them unsuitable for realistic social scenarios. Many also overlook nonverbal cues and complex dynamics of polyadic interactions, both critical for engagement and conversational coherence. In this work, we present PolySLGen, an online framework for Polyadic multimodal Speaking and Listening reaction Generation. Given past conversation and motion from all participants, PolySLGen generates a future speaking or listening reaction for a target participant, including speech, body motion, and speaking state score. To model group interactions effectively, we propose a pose fusion module and a social cue encoder that jointly aggregate motion and social signals from the group. Extensive experiments, along with quantitative and qualitative evaluations, show that PolySLGen produces contextually appropriate and temporally coherent multi-modal reactions, outperforming several adapted and state-of-the-art baselines in motion quality, motion-speech alignment, speaking state prediction, and human-perceived realism.

Method: Integrates vision-language models (VLMs) to encode fixations under controlled visual context (patch-based and marker-based strategies), transforms them into textual descriptions, aggregates into scanpath-level representations, and computes semantic similarity using embedding-based and lexical NLP metrics.

Result: Semantic similarity captures partially independent variance from geometric alignment, revealing cases of high content agreement despite spatial divergence, and shows impact of contextual encoding on description fidelity and metric stability.

Conclusion: Multimodal foundation models enable interpretable, content-aware extensions of classical scanpath analysis, providing a complementary dimension for gaze research within eye-tracking community.

Abstract: Scanpath similarity metrics are central to eye-movement research, yet existing methods predominantly evaluate spatial and temporal alignment while neglecting semantic equivalence between attended image regions. We present a semantic scanpath similarity framework that integrates vision-language models (VLMs) into eye-tracking analysis. Each fixation is encoded under controlled visual context (patch-based and marker-based strategies) and transformed into concise textual descriptions, which are aggregated into scanpath-level representations. Semantic similarity is then computed using embedding-based and lexical NLP metrics and compared against established spatial measures, including MultiMatch and DTW. Experiments on free-viewing eye-tracking data demonstrate that semantic similarity captures partially independent variance from geometric alignment, revealing cases of high content agreement despite spatial divergence. We further analyze the impact of contextual encoding on description fidelity and metric stability. Our findings suggest that multimodal foundation models enable interpretable, content-aware extensions of classical scanpath analysis, providing a complementary dimension for gaze research within the ETRA community.

Method: Train sparse convolutional autoencoder on binarized fragment patches, encode connected components from each page, cluster embeddings with per-image k-means, compare images using set-to-set distances between local vocabularies. Includes BoB-OT variant with mass-weighted optimal transport and two-stage pipeline with BoW shortlisting.

Result: Best BoB variant (Chamfer) achieves Hit@1 of 0.78 and MRR of 0.84 vs 0.74 and 0.80 for strongest BoW baseline, representing 6.1% relative improvement in top-1 accuracy. Two-stage pipeline balances retrieval strength and computational cost.

Conclusion: BoB representation with fragment-specific visual vocabularies outperforms classical BoW for manuscript join retrieval, with practical two-stage pipeline enabling application to larger collections.

Abstract: A join is a set of manuscript fragments identified as originally emanating from the same manuscript. We study manuscript join retrieval: Given a query image of a fragment, retrieve other fragments originating from the same physical manuscript. We propose Bag of Bags (BoB), an image-level representation that replaces the global-level visual codebook of classical Bag of Words (BoW) with a fragment-specific vocabulary of local visual words. Our pipeline trains a sparse convolutional autoencoder on binarized fragment patches, encodes connected components from each page, clusters the resulting embeddings with per image $k$-means, and compares images using set to set distances between their local vocabularies. Evaluated on fragments from the Cairo Genizah, the best BoB variant (viz.@ Chamfer) achieves Hit@1 of 0.78 and MRR of 0.84, compared to 0.74 and 0.80, respectively, for the strongest BoW baseline (BoW-RawPatches-$χ^2$), a 6.1% relative improvement in top-1 accuracy. We furthermore study a mass-weighted BoB-OT variant that incorporates cluster population into prototype matching and present a formal approximation guarantee bounding its deviation from full component-level optimal transport. A two-stage pipeline using a BoW shortlist followed by BoB-OT reranking provides a practical compromise between retrieval strength and computational cost, supporting applicability to larger manuscript collections.

Method: Proposes Face-D²CL with two key components: 1) Multi-domain synergistic representation that fuses spatial and frequency-domain features to comprehensively capture diverse forgery traces, and 2) Dual continual learning mechanism combining Elastic Weight Consolidation (EWC) that distinguishes parameter importance for real vs fake samples, and Orthogonal Gradient Constraint (OGC) that ensures updates to task-specific adapters don’t interfere with previously learned knowledge.

Result: Achieves 60.7% relative reduction in average detection error rate compared to current SOTA approaches, and improves average detection AUC by 7.9% on unseen forgery domains. Demonstrates superior stability and plasticity without relying on historical data replay.

Conclusion: Face-D²CL effectively addresses the challenges of continual learning for facial DeepFake detection by balancing anti-forgetting capabilities with adaptability to emerging forgery paradigms, outperforming existing methods in real-world scenarios.

Abstract: The rapid advancement of facial forgery techniques poses severe threats to public trust and information security, making facial DeepFake detection a critical research priority. Continual learning provides an effective approach to adapt facial DeepFake detection models to evolving forgery patterns. However, existing methods face two key bottlenecks in real-world continual learning scenarios: insufficient feature representation and catastrophic forgetting. To address these issues, we propose Face-D(^2)CL, a framework for facial DeepFake detection. It leverages multi-domain synergistic representation to fuse spatial and frequency-domain features for the comprehensive capture of diverse forgery traces, and employs a dual continual learning mechanism that combines Elastic Weight Consolidation (EWC), which distinguishes parameter importance for real versus fake samples, and Orthogonal Gradient Constraint (OGC), which ensures updates to task-specific adapters do not interfere with previously learned knowledge. This synergy enables the model to achieve a dynamic balance between robust anti-forgetting capabilities and agile adaptability to emerging facial forgery paradigms, all without relying on historical data replay. Extensive experiments demonstrate that our method surpasses current SOTA approaches in both stability and plasticity, achieving 60.7% relative reduction in average detection error rate, respectively. On unseen forgery domains, it further improves the average detection AUC by 7.9% compared to the current SOTA method.

Method: Proposes a temporal adapter with three components: temporal transformer for cross-slice attention, spatial context block for within-slice refinement, and adaptive gate to balance temporal and single-slice features. Trained on 30 labeled volumes from FLARE22 dataset.

Result: Achieves mean Dice of 0.704 across 13 abdominal organs (+0.206 over baseline), with zero-shot improvements of +0.210 on BTCV and +0.230 on AMOS22. Cross-modality evaluation on MRI achieves 0.366 Dice, outperforming fully supervised 3D baseline (0.224).

Conclusion: Temporal context injection significantly improves VLM-based medical segmentation, and CLIP’s visual representations generalize better across imaging modalities than convolutional features, enabling effective cross-domain and cross-modality performance.

Abstract: Medical image segmentation traditionally relies on fully supervised 3D architectures that demand a large amount of dense, voxel-level annotations from clinical experts which is a prohibitively expensive process. Vision Language Models (VLMs) offer a powerful alternative by leveraging broad visual semantic representations learned from billions of images. However, when applied independently to 2D slices of a 3D scan, these models often produce noisy and anatomically implausible segmentations that violate the inherent continuity of anatomical structures. We propose a temporal adapter that addresses this by injecting adjacent-slice context directly into the model’s visual token representations. The adapter comprises a temporal transformer attending across a fixed context window at the token level, a spatial context block refining within-slice representations, and an adaptive gate balancing temporal and single-slice features. Training on 30 labeled volumes from the FLARE22 dataset, our method achieves a mean Dice of 0.704 across 13 abdominal organs with a gain of +0.206 over the baseline VLM trained with no temporal context. Zero-shot evaluation on BTCV and AMOS22 datasets yields consistent improvements of +0.210 and +0.230, with the average cross-domain performance drop reducing from 38.0% to 24.9%. Furthermore, in a cross-modality evaluation on AMOS22 MRI with neither model receiving any MRI supervision, our method achieves a mean Dice of 0.366, outperforming a fully supervised 3D baseline (DynUNet, 0.224) trained exclusively on CT, suggesting that CLIP’s visual semantic representations generalize more gracefully across imaging modalities than convolutional features.

Method: Extends standard MAE by integrating multispectral Sentinel-2 observations with physically meaningful ocean descriptors during self-supervised pre-training. Uses modified UNet for downstream marine segmentation and bathymetry estimation.

Result: OceanMAE shows strongest gains on marine segmentation tasks (MADOS, MARIDA), with competitive bathymetry results (MagicBathyNet). Ablation studies confirm auxiliary ocean descriptors improve segmentation quality.

Conclusion: Physically informed, domain-aligned self-supervised pre-training is valuable for ocean remote sensing, addressing data scarcity and improving model transferability to ocean-specific applications.

Abstract: Accurate ocean mapping is essential for applications such as bathymetry estimation, seabed characterization, marine litter detection, and ecosystem monitoring. However, ocean remote sensing (RS) remains constrained by limited labeled data and by the reduced transferability of models pre-trained mainly on land-dominated Earth observation imagery. In this paper, we propose OceanMAE, an ocean-specific masked autoencoder that extends standard MAE pre-training by integrating multispectral Sentinel-2 observations with physically meaningful ocean descriptors during self-supervised learning. By incorporating these auxiliary ocean features, OceanMAE is designed to learn more informative and ocean-aware latent representations from large- scale unlabeled data. To transfer these representations to downstream applications, we further employ a modified UNet-based framework for marine segmentation and bathymetry estimation. Pre-trained on the Hydro dataset, OceanMAE is evaluated on MADOS and MARIDA for marine pollutant and debris segmentation, and on MagicBathyNet for bathymetry regression. The experiments show that OceanMAE yields the strongest gains on marine segmentation, while bathymetry benefits are competitive and task-dependent. In addition, an ablation against a standard MAE on MARIDA indicates that incorporating auxiliary ocean descriptors during pre-training improves downstream segmentation quality. These findings highlight the value of physically informed and domain-aligned self-supervised pre- training for ocean RS. Code and weights are publicly available at https://git.tu-berlin.de/joanna.stamer/SSLORS2.

Method: Proposes Photometric Alignment Loss (PAL) that discounts nuisance photometric discrepancy via closed-form affine color alignment while preserving restoration-relevant supervision. The method requires only covariance statistics and tiny matrix inversion with negligible overhead, and is based on analysis showing that photometric and structural components of prediction-target residual are orthogonal under least-squares decomposition.

Result: Across 6 tasks, 16 datasets, and 16 architectures, PAL consistently improves metrics and generalization. The method demonstrates effectiveness in addressing photometric inconsistencies that plague supervised low-level vision tasks.

Conclusion: PAL provides a flexible supervision objective that effectively handles photometric inconsistencies in paired training data, leading to improved performance and generalization across various low-level vision tasks and architectures.

Abstract: Supervised low-level vision models rely on pixel-wise losses against paired references, yet paired training sets exhibit per-pair photometric inconsistency, say, different image pairs demand different global brightness, color, or white-balance mappings. This inconsistency enters through task-intrinsic photometric transfer (e.g., low-light enhancement) or unintended acquisition shifts (e.g., de-raining), and in either case causes an optimization pathology. Standard reconstruction losses allocate disproportionate gradient budget to conflicting per-pair photometric targets, crowding out content restoration. In this paper, we investigate this issue and prove that, under least-squares decomposition, the photometric and structural components of the prediction-target residual are orthogonal, and that the spatially dense photometric component dominates the gradient energy. Motivated by this analysis, we propose Photometric Alignment Loss (PAL). This flexible supervision objective discounts nuisance photometric discrepancy via closed-form affine color alignment while preserving restoration-relevant supervision, requiring only covariance statistics and tiny matrix inversion with negligible overhead. Across 6 tasks, 16 datasets, and 16 architectures, PAL consistently improves metrics and generalization. The implementation is in the appendix.

Method: Introduces G²RPO which replaces linear scaling with non-linear distributional matching to force advantage distributions to converge to N(0,1). Also adds two task-level shaping mechanisms: response length shaping for balancing reasoning chains with visual grounding, and entropy shaping to bound exploration zones.

Result: OpenVLThinkerV2 model demonstrates superior performance across 18 diverse benchmarks, outperforming both open-source and leading proprietary frontier models.

Conclusion: G²RPO provides a robust RL training objective for multimodal models that addresses gradient equity issues and enables better balance between perception and reasoning capabilities.

Abstract: Group Relative Policy Optimization (GRPO) has emerged as the de facto Reinforcement Learning (RL) objective driving recent advancements in Multimodal Large Language Models. However, extending this success to open-source multimodal generalist models remains heavily constrained by two primary challenges: the extreme variance in reward topologies across diverse visual tasks, and the inherent difficulty of balancing fine-grained perception with multi-step reasoning capabilities. To address these issues, we introduce Gaussian GRPO (G$^2$RPO), a novel RL training objective that replaces standard linear scaling with non-linear distributional matching. By mathematically forcing the advantage distribution of any given task to strictly converge to a standard normal distribution, $\mathcal{N}(0,1)$, G$^2$RPO theoretically ensures inter-task gradient equity, mitigates vulnerabilities to heavy-tail outliers, and offers symmetric update for positive and negative rewards. Leveraging the enhanced training stability provided by G$^2$RPO, we introduce two task-level shaping mechanisms to seamlessly balance perception and reasoning. First, response length shaping dynamically elicits extended reasoning chains for complex queries while enforce direct outputs to bolster visual grounding. Second, entropy shaping tightly bounds the model’s exploration zone, effectively preventing both entropy collapse and entropy explosion. Integrating these methodologies, we present OpenVLThinkerV2, a highly robust, general-purpose multimodal model. Extensive evaluations across 18 diverse benchmarks demonstrate its superior performance over strong open-source and leading proprietary frontier models.

Method: MedVR uses two synergistic mechanisms: 1) Entropy-guided Visual Regrounding (EVR) that uses model uncertainty to direct exploration, and 2) Consensus-based Credit Assignment (CCA) that distills pseudo-supervision from rollout agreement, enabling annotation-free visual reasoning.

Result: MedVR achieves state-of-the-art performance on diverse public medical VQA benchmarks, significantly outperforming existing models without requiring any human annotations for intermediate steps.

Conclusion: By learning to reason directly with visual evidence, MedVR promotes the robustness and transparency essential for accelerating the clinical deployment of medical AI, addressing critical limitations of current medical VLMs.

Abstract: Medical Vision-Language Models (VLMs) hold immense promise for complex clinical tasks, but their reasoning capabilities are often constrained by text-only paradigms that fail to ground inferences in visual evidence. This limitation not only curtails performance on tasks requiring fine-grained visual analysis but also introduces risks of visual hallucination in safety-critical applications. Thus, we introduce MedVR, a novel reinforcement learning framework that enables annotation-free visual reasoning for medical VLMs. Its core innovation lies in two synergistic mechanisms: Entropy-guided Visual Regrounding (EVR) uses model uncertainty to direct exploration, while Consensus-based Credit Assignment (CCA) distills pseudo-supervision from rollout agreement. Without any human annotations for intermediate steps, MedVR achieves state-of-the-art performance on diverse public medical VQA benchmarks, significantly outperforming existing models. By learning to reason directly with visual evidence, MedVR promotes the robustness and transparency essential for accelerating the clinical deployment of medical AI.

Method: Introduces AVGen-Bench with high-quality prompts across 11 real-world categories and a multi-granular evaluation framework combining lightweight specialist models with Multimodal Large Language Models.

Result: Evaluation reveals significant gap between strong audio-visual aesthetics and weak semantic reliability, with persistent failures in text rendering, speech coherence, physical reasoning, and universal breakdown in musical pitch control.

Conclusion: AVGen-Bench provides comprehensive assessment of T2AV generation, highlighting critical deficiencies in semantic controllability that need addressing for practical media creation applications.

Abstract: Text-to-Audio-Video (T2AV) generation is rapidly becoming a core interface for media creation, yet its evaluation remains fragmented. Existing benchmarks largely assess audio and video in isolation or rely on coarse embedding similarity, failing to capture the fine-grained joint correctness required by realistic prompts. We introduce AVGen-Bench, a task-driven benchmark for T2AV generation featuring high-quality prompts across 11 real-world categories. To support comprehensive assessment, we propose a multi-granular evaluation framework that combines lightweight specialist models with Multimodal Large Language Models (MLLMs), enabling evaluation from perceptual quality to fine-grained semantic controllability. Our evaluation reveals a pronounced gap between strong audio-visual aesthetics and weak semantic reliability, including persistent failures in text rendering, speech coherence, physical reasoning, and a universal breakdown in musical pitch control. Code and benchmark resources are available at http://aka.ms/avgenbench.

Method: Proposes OmniJigsaw framework built on temporal reordering proxy task where shuffled audio-visual clips must be chronologically reconstructed. Uses three strategies: 1) Joint Modality Integration, 2) Sample-level Modality Selection, and 3) Clip-level Modality Masking. Includes two-stage coarse-to-fine data filtering pipeline to adapt to massive unannotated data.

Result: Extensive evaluations on 15 benchmarks show substantial gains in video, audio, and collaborative reasoning. Analysis reveals “bi-modal shortcut phenomenon” in joint modality integration, and demonstrates that clip-level modality masking outperforms sample-level modality selection while mitigating the shortcut issue.

Conclusion: OmniJigsaw validates as a scalable paradigm for self-supervised omni-modal learning, effectively enhancing cross-modal integration and collaborative reasoning through temporal reordering and strategic modality manipulation.

Abstract: To extend the reinforcement learning post-training paradigm to omni-modal models for concurrently bolstering video-audio understanding and collaborative reasoning, we propose OmniJigsaw, a generic self-supervised framework built upon a temporal reordering proxy task. Centered on the chronological reconstruction of shuffled audio-visual clips, this paradigm strategically orchestrates visual and auditory signals to compel cross-modal integration through three distinct strategies: Joint Modality Integration, Sample-level Modality Selection, and Clip-level Modality Masking. Recognizing that the efficacy of such proxy tasks is fundamentally tied to puzzle quality, we design a two-stage coarse-to-fine data filtering pipeline, which facilitates the efficient adaptation of OmniJigsaw to massive unannotated omni-modal data. Our analysis reveals a ``bi-modal shortcut phenomenon’’ in joint modality integration and demonstrates that fine-grained clip-level modality masking mitigates this issue while outperforming sample-level modality selection. Extensive evaluations on 15 benchmarks show substantial gains in video, audio, and collaborative reasoning, validating OmniJigsaw as a scalable paradigm for self-supervised omni-modal learning.

Method: Systematic analysis of cross-modal semantic sharing, layer-wise expert separation analysis, routing divergence measurement, and design of routing-guided intervention method to enhance domain expert activation.

Result: Experiments on three multimodal MoE models across six benchmarks show consistent improvements up to 3.17% on complex visual reasoning tasks, with domain expert identification locating cognitive functions rather than sample-specific solutions.

Conclusion: The “Routing Distraction” hypothesis explains the phenomenon: visual inputs cause routing mechanisms to inadequately activate task-relevant reasoning experts, and interventions that enhance domain expert activation can significantly improve multimodal reasoning performance.

Abstract: Multimodal Mixture-of-Experts (MoE) models have achieved remarkable performance on vision-language tasks. However, we identify a puzzling phenomenon termed Seeing but Not Thinking: models accurately perceive image content yet fail in subsequent reasoning, while correctly solving identical problems presented as pure text. Through systematic analysis, we first verify that cross-modal semantic sharing exists in MoE architectures, ruling out semantic alignment failure as the sole explanation. We then reveal that visual experts and domain experts exhibit layer-wise separation, with image inputs inducing significant routing divergence from text inputs in middle layers where domain experts concentrate. Based on these findings, we propose the Routing Distraction hypothesis: when processing visual inputs, the routing mechanism fails to adequately activate task-relevant reasoning experts. To validate this hypothesis, we design a routing-guided intervention method that enhances domain expert activation. Experiments on three multimodal MoE models across six benchmarks demonstrate consistent improvements, with gains of up to 3.17% on complex visual reasoning tasks. Our analysis further reveals that domain expert identification locates cognitive functions rather than sample-specific solutions, enabling effective transfer across tasks with different information structures.

Method: Created benchmark using agent-based data pipeline: retrieve licensed source papers, perform multimodal understanding of text and figures, build structured prompts, synthesize candidate figures, and filter through review-driven refinement. Covers multiple figure categories, generation sources, and aligned real-synthetic pairs.

Result: Current detection methods fail dramatically in zero-shot transfer, show strong generator-specific overfitting, and remain fragile under common post-processing corruptions. Reveals substantial gap between existing AIGI detection capabilities and high-quality scientific figure generation.

Conclusion: First benchmark for AI-generated scientific figure detection addresses a critical gap in multimodal forensics. Current methods are inadequate for this specialized domain, highlighting need for robust, generalizable scientific-figure forensics research.

Abstract: Modern multimodal generators can now produce scientific figures at near-publishable quality, creating a new challenge for visual forensics and research integrity. Unlike conventional AI-generated natural images, scientific figures are structured, text-dense, and tightly aligned with scholarly semantics, making them a distinct and difficult detection target. However, existing AI-generated image detection benchmarks and methods are almost entirely developed for open-domain imagery, leaving this setting largely unexplored. We present the first benchmark for AI-generated scientific figure detection. To construct it, we develop an agent-based data pipeline that retrieves licensed source papers, performs multimodal understanding of paper text and figures, builds structured prompts, synthesizes candidate figures, and filters them through a review-driven refinement loop. The resulting benchmark covers multiple figure categories, multiple generation sources and aligned real–synthetic pairs. We benchmark representative detectors under zero-shot, cross-generator, and degraded-image settings. Results show that current methods fail dramatically in zero-shot transfer, exhibit strong generator-specific overfitting, and remain fragile under common post-processing corruptions. These findings reveal a substantial gap between existing AIGI detection capabilities and the emerging distribution of high-quality scientific figures. We hope this benchmark can serve as a foundation for future research on robust and generalizable scientific-figure forensics. The dataset is available at https://github.com/Joyce-yoyo/SciFigDetect.

Method: Created PaveInstruct dataset with 278,889 image-instruction-response pairs spanning 32 task types by unifying annotations from nine heterogeneous pavement datasets. Then trained PaveGPT foundation model using domain-specific instruction tuning on this dataset.

Result: Instruction tuning transformed model capabilities, achieving improvements exceeding 20% in spatial grounding, reasoning, and generation tasks while producing ASTM D6433-compliant outputs. Outperformed state-of-the-art vision-language models across perception, understanding, and reasoning tasks.

Conclusion: Domain-specific instruction tuning enables comprehensive pavement condition assessment through vision-language models, allowing transportation agencies to deploy unified conversational assessment tools that replace multiple specialized systems. The approach establishes a pathway for developing instruction-driven AI systems across infrastructure domains.

Abstract: General-purpose vision-language models demonstrate strong performance in everyday domains but struggle with specialized technical fields requiring precise terminology, structured reasoning, and adherence to engineering standards. This work addresses whether domain-specific instruction tuning can enable comprehensive pavement condition assessment through vision-language models. PaveInstruct, a dataset containing 278,889 image-instruction-response pairs spanning 32 task types, was created by unifying annotations from nine heterogeneous pavement datasets. PaveGPT, a pavement foundation model trained on this dataset, was evaluated against state-of-the-art vision-language models across perception, understanding, and reasoning tasks. Instruction tuning transformed model capabilities, achieving improvements exceeding 20% in spatial grounding, reasoning, and generation tasks while producing ASTM D6433-compliant outputs. These results enable transportation agencies to deploy unified conversational assessment tools that replace multiple specialized systems, simplifying workflows and reducing technical expertise requirements. The approach establishes a pathway for developing instruction-driven AI systems across infrastructure domains including bridge inspection, railway maintenance, and building condition assessment.

Method: Two-stage pipeline: 1) Build 100K supervised fine-tuning dataset using GLM automatic annotation, EditScore-based filtering, and human refinement for spatial/directional/attribute accuracy. 2) Collect 10K human preference pairs targeting three failure modes and apply direct preference optimization for alignment beyond SFT.

Result: Fine-tuned Qwen3-VL models outperform open-source baselines; 235B model reaches 4.712 on Eval-400 (vs. Gemini-3-Pro 4.706, GPT-4.1 4.220) and 4.588 on ByteMorph-Bench. Human evaluation shows critical errors falling from 47.75% to 23% and correctness rising from 41.75% to 66%.

Conclusion: EditCaption offers a practical path to scalable, human-aligned instruction synthesis for image editing data by addressing systematic VLM failure modes through a two-stage post-training approach combining supervised fine-tuning and direct preference optimization.

Abstract: High-quality training triplets (source-target image pairs with precise editing instructions) are a critical bottleneck for scaling instruction-guided image editing models. Vision-language models (VLMs) are widely used for automated instruction synthesis, but we identify three systematic failure modes in image-pair settings: orientation inconsistency (e.g., left/right confusion), viewpoint ambiguity, and insufficient fine-grained attribute description. Human evaluation shows that over 47% of instructions from strong baseline VLMs contain critical errors unusable for downstream training. We propose EditCaption, a scalable two-stage post-training pipeline for VLM-based instruction synthesis. Stage 1 builds a 100K supervised fine-tuning (SFT) dataset by combining GLM automatic annotation, EditScore-based filtering, and human refinement for spatial, directional, and attribute-level accuracy. Stage 2 collects 10K human preference pairs targeting the three failure modes and applies direct preference optimization (DPO) for alignment beyond SFT alone. On Eval-400, ByteMorph-Bench, and HQ-Edit, fine-tuned Qwen3-VL models outperform open-source baselines; the 235B model reaches 4.712 on Eval-400 (vs. Gemini-3-Pro 4.706, GPT-4.1 4.220, Kimi-K2.5 4.111) and 4.588 on ByteMorph-Bench (vs. Gemini-3-Pro 4.522, GPT-4.1 3.412). Human evaluation shows critical errors falling from 47.75% to 23% and correctness rising from 41.75% to 66%. The work offers a practical path to scalable, human-aligned instruction synthesis for image editing data.

Method: Survey methodology: 1) Problem formulation highlighting multi-stage nature of object detection under domain shift, 2) Conceptual taxonomy categorizing approaches by adaptation paradigms, modeling assumptions, and pipeline components, 3) Analysis of domain shift propagation across detection stages, 4) Review of datasets, evaluation protocols, and benchmarking practices.

Result: Provides comprehensive systematic analysis of CDOD field, identifies why adaptation in object detection is more complex than classification, organizes fragmented literature into unified framework, and outlines key challenges and future research directions.

Conclusion: The survey offers unified framework for understanding CDOD, guides development of more robust detection systems, and addresses the inherent complexity of domain adaptation in object detection compared to classification tasks.

Abstract: Object detection models trained on a source domain often exhibit significant performance degradation when deployed in unseen target domains, due to various kinds of variations, such as sensing conditions, environments and data distributions. Hence, regardless the recent breakthrough advances in deep learning-based detection technology, cross-domain object detection (CDOD) remains a critical research area. Moreover, the existing literature remains fragmented, lacking a unified perspective on the structural challenges underlying domain shift and the effectiveness of adaptation strategies. This survey provides a comprehensive and systematic analysis of CDOD. We start upon a problem formulation that highlights the multi-stage nature of object detection under domain shift. Then, we organize the existing methods through a conceptual taxonomy that categorizes approaches based on adaptation paradigms, modeling assumptions, and pipeline components. Furthermore, we analyze how domain shift propagates across detection stages and discuss why adaptation in object detection is inherently more complex than in classification. In addition, we review commonly used datasets, evaluation protocols, and benchmarking practices. Finally, we identify the key challenges and outline promising future research directions. Cohesively, this survey aims to provide a unified framework for understanding CDOD and to guide the development of more robust detection systems.

Method: Proposes SCADA-UL: Unlearning Source-exclusive ClAsses in Domain Adaptation. Uses adversarially generated forget class samples that are unlearned during domain adaptation with a novel rescaled labeling strategy and adversarial optimization. Also addresses continual and unknown-class variants.

Result: Method consistently outperforms baselines in the proposed setting and achieves retraining-level unlearning performance on benchmark datasets. Existing SFDA methods show strong zero-shot performance on source-exclusive classes, indicating privacy leaks.

Conclusion: Identifies and addresses privacy risks in domain adaptation through machine unlearning. SCADA-UL effectively removes source-exclusive class knowledge while maintaining adaptation performance, with theoretical interpretations and empirical validation.

Abstract: The increasing adaptation of vision models across domains, such as satellite imagery and medical scans, has raised an emerging privacy risk: models may inadvertently retain and leak sensitive source-domain specific information in the target domain. This creates a compelling use case for machine unlearning to protect the privacy of sensitive source-domain data. Among adaptation techniques, source-free domain adaptation (SFDA) calls for an urgent need for machine unlearning (MU), where the source data itself is protected, yet the source model exposed during adaptation encodes its influence. Our experiments reveal that existing SFDA methods exhibit strong zero-shot performance on source-exclusive classes in the target domain, indicating they inadvertently leak knowledge of these classes into the target domain, even when they are not represented in the target data. We identify and address this risk by proposing an MU setting called SCADA-UL: Unlearning Source-exclusive ClAsses in Domain Adaptation. Existing MU methods do not address this setting as they are not designed to handle data distribution shifts. We propose a new unlearning method, where an adversarially generated forget class sample is unlearned by the model during the domain adaptation process using a novel rescaled labeling strategy and adversarial optimization. We also extend our study to two variants: a continual version of this problem setting and to one where the specific source classes to be forgotten may be unknown. Alongside theoretical interpretations, our comprehensive empirical results show that our method consistently outperforms baselines in the proposed setting while achieving retraining-level unlearning performance on benchmark datasets. Our code is available at https://github.com/D-Arnav/SCADA

Method: Proposes a dual-branch multimodal framework with: 1) text-image branch for multimodal matching, and 2) vision branch for visual feature analysis. After training, computes scores from both branches (S_t and S_v), integrates them for final OOD score S, and compares with threshold for detection.

Result: Comprehensive experiments on endoscopic image datasets show the framework is robust across diverse backbones and improves state-of-the-art OOD detection performance by up to 24.84%.

Conclusion: The proposed multimodal framework effectively leverages complementary information from text and vision branches to enhance OOD detection reliability in clinical applications, addressing limitations of single-modality approaches.

Abstract: The complex and dynamic real-world clinical environment demands reliable deep learning (DL) systems. Out-of-distribution (OOD) detection plays a critical role in enhancing the reliability and generalizability of DL models when encountering data that deviate from the training distribution, such as unseen disease cases. However, existing OOD detection methods typically rely either on a single visual modality or solely on image-text matching, failing to fully leverage multimodal information. To overcome the challenge, we propose a novel dual-branch multimodal framework by introducing a text-image branch and a vision branch. Our framework fully exploits multimodal representations to identify OOD samples through these two complementary branches. After training, we compute scores from the text-image branch ($S_t$) and vision branch ($S_v$), and integrate them to obtain the final OOD score $S$ that is compared with a threshold for OOD detection. Comprehensive experiments on publicly available endoscopic image datasets demonstrate that our proposed framework is robust across diverse backbones and improves state-of-the-art performance in OOD detection by up to 24.84%

Method: Proposes Orion-Lite, an efficient vision-only student model that uses latent feature distillation from massive VLA teacher (ORION) combined with ground-truth trajectory supervision. Focuses on complex, interactive scenarios under closed-loop evaluation.

Result: Orion-Lite surpasses its massive VLA teacher ORION, achieving state-of-the-art on Bench2Drive benchmark with Driving Score of 80.6. Demonstrates vision-only architectures have untapped potential for reactive planning.

Conclusion: LLM knowledge distillation enables compact vision-only models to outperform massive multimodal teachers, revealing significant potential for efficient autonomous driving systems without sacrificing performance.

Abstract: Leveraging the general world knowledge of Large Language Models (LLMs) holds significant promise for improving the ability of autonomous driving systems to handle rare and complex scenarios. While integrating LLMs into Vision-Language-Action (VLA) models has yielded state-of-the-art performance, their massive parameter counts pose severe challenges for latency-sensitive and energy-efficient deployment. Distilling LLM knowledge into a compact driving model offers a compelling solution to retain these reasoning capabilities while maintaining a manageable computational footprint. Although previous works have demonstrated the efficacy of distillation, these efforts have primarily focused on relatively simple scenarios and open-loop evaluations. Therefore, in this work, we investigate LLM distillation in more complex, interactive scenarios under closed-loop evaluation. We demonstrate that through a combination of latent feature distillation and ground-truth trajectory supervision, an efficient vision-only student model \textbf{Orion-Lite} can even surpass the performance of its massive VLA teacher, ORION. Setting a new state-of-the-art on the rigorous Bench2Drive benchmark, with a Driving Score of 80.6. Ultimately, this reveals that vision-only architectures still possess significant, untapped potential for high-performance reactive planning.

Method: Introduce spectral point cloud paradigm: treat point clouds as sparse, compressed representations of radar spectra. Develop experimental framework comparing spectral PC models at varying densities against dense RD benchmark. Test two basic spectral enrichment approaches that inject target-relevant information into point clouds.

Result: Point clouds can match dense RD benchmark performance at certain density levels, and can surpass it when spectral enrichment is applied. Shows point clouds need not underperform compared to spectra.

Conclusion: Spectral point clouds serve as strong candidates for unified radar perception, challenging the common belief that dense RD approach is superior. Paves way for future radar foundation models with more robust, sensor-agnostic representations.

Abstract: Radar perception models are trained with different inputs, from range-Doppler spectra to sparse point clouds. Dense spectra are assumed to outperform sparse point clouds, yet they can vary considerably across sensors and configurations, which hinders transfer. In this paper, we provide alternatives for incorporating spectral information into radar point clouds and show that, point clouds need not underperform compared to spectra. We introduce the spectral point cloud paradigm, where point clouds are treated as sparse, compressed representations of the radar spectra, and argue that, when enriched with spectral information, they serve as strong candidates for a unified input representation that is more robust against sensor-specific differences. We develop an experimental framework that compares spectral point cloud (PC) models at varying densities against a dense range-Doppler (RD) benchmark, and report the density levels where the PC configurations meet the performance of the RD benchmark. Furthermore, we experiment with two basic spectral enrichment approaches, that inject additional target-relevant information into the point clouds. Contrary to the common belief that the dense RD approach is superior, we show that point clouds can do just as well, and can surpass the RD benchmark when enrichment is applied. Spectral point clouds can therefore serve as strong candidates for unified radar perception, paving the way for future radar foundation models.

Method: Constructed CAMotion - a high-quality benchmark covering wide range of species for camouflaged moving object detection in the wild, with various sequences featuring challenging attributes like uncertain edges, occlusion, motion blur, and shape complexity.

Result: CAMotion provides comprehensive sequence annotation details and statistical distribution from multiple perspectives, enabling in-depth analysis of camouflaged object motion characteristics in different challenging scenarios.

Conclusion: The benchmark enables comprehensive evaluation of SOTA models, discussion of major VCOD challenges, and aims to advance research in camouflaged object detection.

Abstract: Discovering camouflaged objects is a challenging task in computer vision due to the high similarity between camouflaged objects and their surroundings. While the problem of camouflaged object detection over sequential video frames has received increasing attention, the scale and diversity of existing video camouflaged object detection (VCOD) datasets are greatly limited, which hinders the deeper analysis and broader evaluation of recent deep learning-based algorithms with data-hungry training strategy. To break this bottleneck, in this paper, we construct CAMotion, a high-quality benchmark covers a wide range of species for camouflaged moving object detection in the wild. CAMotion comprises various sequences with multiple challenging attributes such as uncertain edge, occlusion, motion blur, and shape complexity, etc. The sequence annotation details and statistical distribution are presented from various perspectives, allowing CAMotion to provide in-depth analyses on the camouflaged object’s motion characteristics in different challenging scenarios. Additionally, we conduct a comprehensive evaluation of existing SOTA models on CAMotion, and discuss the major challenges in VCOD task. The benchmark is available at https://www.camotion.focuslab.net.cn, we hope that our CAMotion can lead to further advancements in the research community.

Method: Proposes GroundingAnomaly with: 1) Spatial Conditioning Module using per-pixel semantic maps for precise spatial control, and 2) Gated Self-Attention Module that injects conditioning tokens into frozen U-Net via gated attention layers to preserve pretrained priors while enabling stable few-shot adaptation.

Result: Extensive evaluations on MVTec AD and VisA datasets show GroundingAnomaly generates high-quality anomalies and achieves state-of-the-art performance across anomaly detection, segmentation, and instance-level detection tasks.

Conclusion: GroundingAnomaly effectively addresses the data scarcity problem in industrial anomaly inspection through precise spatial control and stable few-shot adaptation, improving downstream inspection performance.

Abstract: The performance of visual anomaly inspection in industrial quality control is often constrained by the scarcity of real anomalous samples. Consequently, anomaly synthesis techniques have been developed to enlarge training sets and enhance downstream inspection. However, existing methods either suffer from poor integration caused by inpainting or fail to provide accurate masks. To address these limitations, we propose GroundingAnomaly, a novel few-shot anomaly image generation framework. Our framework introduces a Spatial Conditioning Module that leverages per-pixel semantic maps to enable precise spatial control over the synthesized anomalies. Furthermore, a Gated Self-Attention Module is designed to inject conditioning tokens into a frozen U-Net via gated attention layers. This carefully preserves pretrained priors while ensuring stable few-shot adaptation. Extensive evaluations on the MVTec AD and VisA datasets demonstrate that GroundingAnomaly generates high-quality anomalies and achieves state-of-the-art performance across multiple downstream tasks, including anomaly detection, segmentation, and instance-level detection.

Method: Combines pretrained state-of-the-art rectified flow and predictor models in a plug-and-play manner. Uses training-free guidance of a 3D rectified flow model, requiring only fine-tuning of the predictor using image-level labels without retraining the generative model.

Result: Produces improved-quality segmentations for two separate predictors, consistently detecting lung nodules of varying sizes and shapes. Experiments on LUNA16 demonstrate improvements over baseline methods.

Conclusion: Highlights the potential of generative foundation models as tools for weakly supervised 3D medical image segmentation, offering a promising approach that reduces annotation burden while maintaining accuracy for small structures.

Abstract: Dense annotations, such as segmentation masks, are expensive and time-consuming to obtain, especially for 3D medical images where expert voxel-wise labeling is required. Weakly supervised approaches aim to address this limitation, but often rely on attribution-based methods that struggle to accurately capture small structures such as lung nodules. In this paper, we propose a weakly-supervised segmentation method for lung nodules by combining pretrained state-of-the-art rectified flow and predictor models in a plug-and-play manner. Our approach uses training-free guidance of a 3D rectified flow model, requiring only fine-tuning of the predictor using image-level labels and no retraining of the generative model. The proposed method produces improved-quality segmentations for two separate predictors, consistently detecting lung nodules of varying size and shapes. Experiments on LUNA16 demonstrate improvements over baseline methods, highlighting the potential of generative foundation models as tools for weakly supervised 3D medical image segmentation.

Method: 1) RAG-based method to compose image-specific, knowledge-aware reasoning traces linking visual findings to image labels using ophthalmic knowledge. 2) Enhanced RLVR with process reward encouraging self-consistency of generated reasoning traces.

Result: Fundus-R1 outperforms multiple baselines including Qwen2.5-VL and stronger editions without generated traces on three benchmarks: FunBench, Omni-Fundus, and GMAI-Fundus.

Conclusion: The work demonstrates that powerful fundus-reading MLLMs can be trained using exclusively public datasets, paving the way for more accessible and reproducible research in medical multimodal AI.

Abstract: Fundus imaging such as CFP, OCT and UWF is crucial for the early detection of retinal anomalies and diseases. Fundus image understanding, due to its knowledge-intensive nature, poses a challenging vision-language task. An emerging approach to addressing the task is to post-train a generic multimodal large language model (MLLM), either by supervised finetuning (SFT) or by reinforcement learning with verifiable rewards (RLVR), on a considerable amount of in-house samples paired with high-quality clinical reports. However, these valuable samples are not publicly accessible, which not only hinders reproducibility but also practically limits research to few players. To overcome the barrier, we make a novel attempt to train a reasoning-enhanced fundus-reading MLLM, which we term Fundus-R1, using exclusively public datasets, wherein over 94% of the data are annotated with only image-level labels. Our technical contributions are two-fold. First, we propose a RAG-based method for composing image-specific, knowledge-aware reasoning traces. Such auto-generated traces link visual findings identified by a generic MLLM to the image labels in terms of ophthalmic knowledge. Second, we enhance RLVR with a process reward that encourages self-consistency of the generated reasoning trace in each rollout. Extensive experiments on three fundus-reading benchmarks, i.e., FunBench, Omni-Fundus and GMAI-Fundus, show that Fundus-R1 clearly outperforms multiple baselines, including its generic counterpart (Qwen2.5-VL) and a stronger edition post-trained without using the generated traces. This work paves the way for training powerful fundus-reading MLLMs with publicly available data.

Method: Extensive experiments on 14 open-source medical MLLMs across three representative image classification datasets using feature probing to track visual information flow module-by-module and layer-by-layer throughout the entire MLLM pipeline.

Result: Identified four failure modes: 1) quality limitation in visual representation, 2) fidelity loss in connector projection, 3) comprehension deficit in LLM reasoning, and 4) misalignment of semantic mapping. Introduced quantitative scores for feature evolution healthiness.

Conclusion: Medical MLLMs face critical barriers preventing clinical deployment. The work highlights that the road from high expectations to clinically deployable MLLMs remains long and requires rethinking of current approaches.

Abstract: The rise of multimodal large language models (MLLMs) has sparked an unprecedented wave of applications in the field of medical imaging analysis. However, as one of the earliest and most fundamental tasks integrated into this paradigm, medical image classification reveals a sobering reality: state-of-the-art medical MLLMs consistently underperform compared to traditional deep learning models, despite their overwhelming advantages in pre-training data and model parameters. This paradox prompts a critical rethinking: where exactly does the performance degradation originate? In this paper, we conduct extensive experiments on 14 open-source medical MLLMs across three representative image classification datasets. Moving beyond superficial performance benchmarking, we employ feature probing to track the information flow of visual features module-by-module and layer-by-layer throughout the entire MLLM pipeline, enabling explicit visualization of where and how classification signals are distorted, diluted, or overridden. As the first attempt to dissect classification performance degradation in medical MLLMs, our findings reveal four failure modes: 1) quality limitation in visual representation, 2) fidelity loss in connector projection, 3) comprehension deficit in LLM reasoning, and 4) misalignment of semantic mapping. Meanwhile, we introduce quantitative scores that characterize the healthiness of feature evolution, enabling principled comparisons across diverse MLLMs and datasets. Furthermore, we provide insightful discussions centered on the critical barriers that prevent current medical MLLMs from fulfilling their promised clinical potential. We hope that our work provokes rethinking within the community-highlighting that the road from high expectations to clinically deployable MLLMs remains long and winding.

Method: Proposes InstAP framework with joint optimization of global vision-text alignment and fine-grained instance-level contrastive alignment by grounding textual mentions to specific spatial-temporal regions, supported by InstVL dataset with 2M images and 50K videos featuring dual-granularity annotations.

Result: Substantially outperforms existing VLP models on instance-level retrieval, surpasses strong VLP baseline trained on same data, achieves competitive zero-shot performance on video benchmarks (MSR-VTT, DiDeMo), and shows improved localization of textual mentions to correct instances.

Conclusion: Instance-centric pre-training improves both fine-grained instance reasoning and global understanding, demonstrating that instance-aware objectives enhance multimodal representation learning beyond global-only supervision.

Abstract: Current vision-language pre-training (VLP) paradigms excel at global scene understanding but struggle with instance-level reasoning due to global-only supervision. We introduce InstAP, an Instance-Aware Pre-training framework that jointly optimizes global vision-text alignment and fine-grained, instance-level contrastive alignment by grounding textual mentions to specific spatial-temporal regions. To support this, we present InstVL, a large-scale dataset (2 million images, 50,000 videos) with dual-granularity annotations: holistic scene captions and dense, grounded instance descriptions. On the InstVL benchmark, InstAP substantially outperforms existing VLP models on instance-level retrieval, and also surpasses a strong VLP baseline trained on the exact same data corpus, isolating the benefit of our instance-aware objective. Moreover, instance-centric pre-training improves global understanding: InstAP achieves competitive zero-shot performance on multiple video benchmarks, including MSR-VTT and DiDeMo. Qualitative visualizations further show that InstAP localizes textual mentions to the correct instances, while global-only models exhibit more diffuse, scene-level attention.

Method: Created PokeGym benchmark using Pokemon Legends: Z-A with 30 tasks (30-220 steps) across navigation, interaction, and mixed scenarios. Enforces strict code-level isolation: agents use only raw RGB observations while independent evaluator verifies success via memory scanning. Includes three instruction granularities to test different capabilities.

Result: Revealed that physical deadlock recovery, not high-level planning, is the primary bottleneck for VLMs in embodied environments. Found metacognitive divergence: weaker models suffer from Unaware Deadlocks while advanced models exhibit Aware Deadlocks (recognize entrapment but can’t recover).

Conclusion: Current VLMs need explicit spatial intuition integration for better performance in complex 3D embodied environments. PokeGym provides automated, scalable assessment for vision-based decision-making in interactive 3D worlds.

Abstract: While Vision-Language Models (VLMs) have achieved remarkable progress in static visual understanding, their deployment in complex 3D embodied environments remains severely limited. Existing benchmarks suffer from four critical deficiencies: (1) passive perception tasks circumvent interactive dynamics; (2) simplified 2D environments fail to assess depth perception; (3) privileged state leakage bypasses genuine visual processing; and (4) human evaluation is prohibitively expensive and unscalable. We introduce PokeGym, a visually-driven long-horizon benchmark instantiated within Pokemon Legends: Z-A, a visually complex 3D open-world Role-Playing Game. PokeGym enforces strict code-level isolation: agents operate solely on raw RGB observations while an independent evaluator verifies success via memory scanning, ensuring pure vision-based decision-making and automated, scalable assessment. The benchmark comprises 30 tasks (30-220 steps) spanning navigation, interaction, and mixed scenarios, with three instruction granularities (Visual-Guided, Step-Guided, Goal-Only) to systematically deconstruct visual grounding, semantic reasoning, and autonomous exploration capabilities. Our evaluation reveals a key limitation of current VLMs: physical deadlock recovery, rather than high-level planning, constitutes the primary bottleneck, with deadlocks showing a strong negative correlation with task success. Furthermore, we uncover a metacognitive divergence: weaker models predominantly suffer from Unaware Deadlocks (oblivious to entrapment), whereas advanced models exhibit Aware Deadlocks (recognizing entrapment yet failing to recover). These findings highlight the need to integrate explicit spatial intuition into VLM architectures. The code and benchmark will be available on GitHub.

Method: 1) Curate 170K style prompts and 400K content prompts, 2) Generate MegaStyle-1.4M dataset via content-style prompt combinations using text-to-image models, 3) Train MegaStyle-Encoder via style-supervised contrastive learning, 4) Train MegaStyle-FLUX style transfer model based on FLUX architecture.

Result: MegaStyle-1.4M demonstrates importance of intra-style consistency, inter-style diversity, and high-quality data. MegaStyle-Encoder provides reliable style similarity measurement, and MegaStyle-FLUX achieves generalizable style transfer, outperforming existing methods.

Conclusion: The MegaStyle pipeline successfully addresses key limitations in style datasets, enabling better style representation learning and transfer. The curated dataset and trained models significantly advance the style transfer community.

Abstract: In this paper, we introduce MegaStyle, a novel and scalable data curation pipeline that constructs an intra-style consistent, inter-style diverse and high-quality style dataset. We achieve this by leveraging the consistent text-to-image style mapping capability of current large generative models, which can generate images in the same style from a given style description. Building on this foundation, we curate a diverse and balanced prompt gallery with 170K style prompts and 400K content prompts, and generate a large-scale style dataset MegaStyle-1.4M via content-style prompt combinations. With MegaStyle-1.4M, we propose style-supervised contrastive learning to fine-tune a style encoder MegaStyle-Encoder for extracting expressive, style-specific representations, and we also train a FLUX-based style transfer model MegaStyle-FLUX. Extensive experiments demonstrate the importance of maintaining intra-style consistency, inter-style diversity and high-quality for style dataset, as well as the effectiveness of the proposed MegaStyle-1.4M. Moreover, when trained on MegaStyle-1.4M, MegaStyle-Encoder and MegaStyle-FLUX provide reliable style similarity measurement and generalizable style transfer, making a significant contribution to the style transfer community. More results are available at our project website https://jeoyal.github.io/MegaStyle/.

Method: Proposes SurfelSplat, a feed-forward framework that generates pixel-aligned Gaussian surfel representations from sparse-view images. Uses a cross-view feature aggregation module based on Nyquist sampling theorem: adapts geometric forms with spatial sampling rate-guided low-pass filters, projects filtered surfels across views for cross-view feature correlations, and processes through a feature fusion network to regress precise Gaussian surfels.

Result: Achieves comparable results with state-of-the-art methods on DTU reconstruction benchmarks, predicts Gaussian surfels within 1 second, offering 100x speedup without costly per-scene training.

Conclusion: SurfelSplat provides an efficient and generalizable approach for 3D surface reconstruction from sparse views, addressing limitations of optimization-based methods while maintaining reconstruction quality.

Abstract: 3D Gaussian Splatting (3DGS) has demonstrated impressive performance in 3D scene reconstruction. Beyond novel view synthesis, it shows great potential for multi-view surface reconstruction. Existing methods employ optimization-based reconstruction pipelines that achieve precise and complete surface extractions. However, these approaches typically require dense input views and high time consumption for per-scene optimization. To address these limitations, we propose SurfelSplat, a feed-forward framework that generates efficient and generalizable pixel-aligned Gaussian surfel representations from sparse-view images. We observe that conventional feed-forward structures struggle to recover accurate geometric attributes of Gaussian surfels because the spatial frequency of pixel-aligned primitives exceeds Nyquist sampling rates. Therefore, we propose a cross-view feature aggregation module based on the Nyquist sampling theorem. Specifically, we first adapt the geometric forms of Gaussian surfels with spatial sampling rate-guided low-pass filters. We then project the filtered surfels across all input views to obtain cross-view feature correlations. By processing these correlations through a specially designed feature fusion network, we can finally regress Gaussian surfels with precise geometry. Extensive experiments on DTU reconstruction benchmarks demonstrate that our model achieves comparable results with state-of-the-art methods, and predict Gaussian surfels within 1 second, offering a 100x speedup without costly per-scene training.

Method: Phantasia introduces a context-adaptive backdoor attack that dynamically aligns poisoned outputs with input semantics. Instead of static patterns, it generates contextually coherent yet malicious responses that remain plausible, improving stealth and adaptability across diverse VLM architectures.

Result: Extensive experiments show Phantasia achieves state-of-the-art attack success rates while maintaining benign performance under various defensive settings, demonstrating superior stealth compared to existing methods.

Conclusion: The work reveals that existing VLM backdoor attacks are less stealthy than previously thought, and introduces Phantasia as a more sophisticated, context-adaptive attack that significantly improves stealth while maintaining effectiveness against defenses.

Abstract: Recent advances in Vision-Language Models (VLMs) have greatly enhanced the integration of visual perception and linguistic reasoning, driving rapid progress in multimodal understanding. Despite these achievements, the security of VLMs, particularly their vulnerability to backdoor attacks, remains significantly underexplored. Existing backdoor attacks on VLMs are still in an early stage of development, with most current methods relying on generating poisoned responses that contain fixed, easily identifiable patterns. In this work, we make two key contributions. First, we demonstrate for the first time that the stealthiness of existing VLM backdoor attacks has been substantially overestimated. By adapting defense techniques originally designed for other domains (e.g., vision-only and text-only models), we show that several state-of-the-art attacks can be detected with surprising ease. Second, to address this gap, we introduce Phantasia, a context-adaptive backdoor attack that dynamically aligns its poisoned outputs with the semantics of each input. Instead of producing static poisoned patterns, Phantasia encourages models to generate contextually coherent yet malicious responses that remain plausible, thereby significantly improving stealth and adaptability. Extensive experiments across diverse VLM architectures reveal that Phantasia achieves state-of-the-art attack success rates while maintaining benign performance under various defensive settings.

Method: Proposes SyncBreaker with two key components: 1) For images: nullifying supervision with Multi-Interval Sampling across diffusion stages to steer generation toward static reference portrait; 2) For audio: Cross-Attention Fooling to suppress interval-specific audio-conditioned cross-attention responses. Both streams optimized independently and combined at inference.

Result: SyncBreaker more effectively degrades lip synchronization and facial dynamics than strong single-modality baselines, while preserving input perceptual quality and remaining robust under purification attacks.

Conclusion: SyncBreaker addresses the gap in multimodal protection for talking-head generation, providing effective joint perturbation of portrait and audio inputs to prevent misuse while maintaining perceptual quality.

Abstract: Diffusion-based audio-driven talking-head generation enables realistic portrait animation, but also introduces risks of misuse, such as fraud and misinformation. Existing protection methods are largely limited to a single modality, and neither image-only nor audio-only attacks can effectively suppress speech-driven facial dynamics. To address this gap, we propose SyncBreaker, a stage-aware multimodal protection framework that jointly perturbs portrait and audio inputs under modality-specific perceptual constraints. Our key contributions are twofold. First, for the image stream, we introduce nullifying supervision with Multi-Interval Sampling (MIS) across diffusion stages to steer the generation toward the static reference portrait by aggregating guidance from multiple denoising intervals. Second, for the audio stream, we propose Cross-Attention Fooling (CAF), which suppresses interval-specific audio-conditioned cross-attention responses. Both streams are optimized independently and combined at inference time to enable flexible deployment. We evaluate SyncBreaker in a white-box proactive protection setting. Extensive experiments demonstrate that SyncBreaker more effectively degrades lip synchronization and facial dynamics than strong single-modality baselines, while preserving input perceptual quality and remaining robust under purification. Code: https://github.com/kitty384/SyncBreaker.

Method: Three-module framework: 1) Knowledge-guided Language Parsing (KLP) converts natural language to structured manipulation constraints; 2) Triangular Functional Point Localization (TriLocation) uses 3D Gaussian Splatting for pose-consistent spatial reasoning; 3) 3D Keypoint Grasp Matrix Transformation Execution (KGT3D+) decodes constraints into wrist poses and finger-level commands.

Result: BLaDA significantly outperforms existing methods in both affordance grounding precision and functional manipulation success rate across diverse categories and tasks in complex benchmarks.

Conclusion: BLaDA provides an interpretable, zero-shot solution for functional dexterous manipulation that effectively bridges language understanding with 3D spatial reasoning and physical execution.

Abstract: In unstructured environments, functional dexterous grasping calls for the tight integration of semantic understanding, precise 3D functional localization, and physically interpretable execution. Modular hierarchical methods are more controllable and interpretable than end-to-end VLA approaches, but existing ones still rely on predefined affordance labels and lack the tight semantic–pose coupling needed for functional dexterous manipulation. To address this, we propose BLaDA (Bridging Language to Dexterous Actions in 3DGS fields), an interpretable zero-shot framework that grounds open-vocabulary instructions as perceptual and control constraints for functional dexterous manipulation. BLaDA establishes an interpretable reasoning chain by first parsing natural language into a structured sextuple of manipulation constraints via a Knowledge-guided Language Parsing (KLP) module. To achieve pose-consistent spatial reasoning, we introduce the Triangular Functional Point Localization (TriLocation) module, which utilizes 3D Gaussian Splatting as a continuous scene representation and identifies functional regions under triangular geometric constraints. Finally, the 3D Keypoint Grasp Matrix Transformation Execution (KGT3D+) module decodes these semantic-geometric constraints into physically plausible wrist poses and finger-level commands. Extensive experiments on complex benchmarks demonstrate that BLaDA significantly outperforms existing methods in both affordance grounding precision and the success rate of functional manipulation across diverse categories and tasks. Code will be publicly available at https://github.com/PopeyePxx/BLaDA.

Method: Proposes HST-HGN with two key components: 1) Spatially: hierarchical hypergraph network that fuses pose-disentangled geometric topologies with multi-modal texture patches dynamically to capture high-order synergistic facial deformations; 2) Temporally: Bi-Mamba module with linear complexity for bidirectional sequence modeling to distinguish ambiguous transient actions (e.g., yawning vs speaking) while capturing complete physiological lifecycles.

Result: Extensive evaluations across diverse fatigue benchmarks demonstrate state-of-the-art performance. The method achieves a balance between discriminative power and computational efficiency, making it suitable for real-time in-cabin edge deployment.

Conclusion: HST-HGN effectively addresses the challenges of driver fatigue assessment by combining spatial hypergraph modeling with efficient bidirectional temporal modeling, offering both high accuracy and computational efficiency for practical deployment.

Abstract: It remains challenging to assess driver fatigue from untrimmed videos under constrained computational budgets, due to the difficulty of modeling long-range temporal dependencies in subtle facial expressions. Some existing approaches rely on computationally heavy architectures, whereas others employ traditional lightweight pairwise graph networks, despite their limited capacity to model high-order synergies and global temporal context. Therefore, we propose HST-HGN, a novel Heterogeneous Spatial-Temporal Hypergraph Network driven by Bidirectional State Space Models. Spatially, we introduce a hierarchical hypergraph network to fuse pose-disentangled geometric topologies with multi-modal texture patches dynamically. This formulation encapsulates high-order synergistic facial deformations, effectively overcoming the limitations of conventional methods. In temporal terms, a Bi-Mamba module with linear complexity is applied to perform bidirectional sequence modeling. This explicit temporal-evolution filtering enables the network to distinguish highly ambiguous transient actions, such as yawning versus speaking, while encompassing their complete physiological lifecycles. Extensive evaluations across diverse fatigue benchmarks demonstrate that HST-HGN achieves state-of-the-art performance. In particular, our method strikes a balance between discriminative power and computational efficiency, making it well-suited for real-time in-cabin edge deployment.

Method: Created a large-scale dataset of 250 real-world crash videos from roadside cameras, annotated with 13K multiple-choice QA pairs organized in a two-tier taxonomy: Tier 1 for visual grounding and Tier 2 for higher-level reasoning.

Result: Benchmarked 8 state-of-the-art VLMs showing they struggle with temporal and causal reasoning in safety-critical scenarios despite strong scene description capabilities.

Conclusion: CrashSight provides a standardized evaluation framework for infrastructure-assisted perception in cooperative autonomous driving and reveals current VLM limitations in safety-critical reasoning.

Abstract: Cooperative autonomous driving requires traffic scene understanding from both vehicle and infrastructure perspectives. While vision-language models (VLMs) show strong general reasoning capabilities, their performance in safety-critical traffic scenarios remains insufficiently evaluated due to the ego-vehicle focus of existing benchmarks. To bridge this gap, we present \textbf{CrashSight}, a large-scale vision-language benchmark for roadway crash understanding using real-world roadside camera data. The dataset comprises 250 crash videos, annotated with 13K multiple-choice question-answer pairs organized under a two-tier taxonomy. Tier 1 evaluates the visual grounding of scene context and involved parties, while Tier 2 probes higher-level reasoning, including crash mechanics, causal attribution, temporal progression, and post-crash outcomes. We benchmark 8 state-of-the-art VLMs and show that, despite strong scene description capabilities, current models struggle with temporal and causal reasoning in safety-critical scenarios. We provide a detailed analysis of failure scenarios and discuss directions for improving VLM crash understanding. The benchmark provides a standardized evaluation framework for infrastructure-assisted perception in cooperative autonomous driving. The CrashSight benchmark, including the full dataset and code, is accessible at https://mcgrche.github.io/crashsight.

Method: Proposes OVS-DINO framework with Structure-Aware Encoder (SAE) and Structure-Modulated Decoder (SMD) to activate DINO’s boundary features using SAM’s structural priors. Uses SAM-generated pseudo-masks for supervision and structural alignment between DINO and SAM.

Result: Achieves state-of-the-art performance across multiple weakly-supervised OVS benchmarks, improving average score by 2.1% (from 44.8% to 46.9%). Significant 6.3% gain on Cityscapes (from 36.6% to 42.9%), especially in complex, cluttered scenarios.

Conclusion: The framework successfully revitalizes DINO’s latent edge-sensitivity through structural alignment with SAM, demonstrating that boundary awareness in foundation models can be effectively activated for improved open-vocabulary segmentation.

Abstract: Open-Vocabulary Segmentation (OVS) aims to segment image regions beyond predefined category sets by leveraging semantic descriptions. While CLIP based approaches excel in semantic generalization, they frequently lack the fine-grained spatial awareness required for dense prediction. Recent efforts have incorporated Vision Foundation Models (VFMs) like DINO to alleviate these limitations. However, these methods still struggle with the precise edge perception necessary for high fidelity segmentation. In this paper, we analyze internal representations of DINO and discover that its inherent boundary awareness is not absent but rather undergoes progressive attenuation as features transition into deeper transformer blocks. To address this, we propose OVS-DINO, a novel framework that revitalizes latent edge-sensitivity of DINO through structural alignment with the Segment Anything Model (SAM). Specifically, we introduce a Structure-Aware Encoder (SAE) and a Structure-Modulated Decoder (SMD) to effectively activate boundary features of DINO using SAM’s structural priors, complemented by a supervision strategy utilizing SAM generated pseudo-masks. Extensive experiments demonstrate that our method achieves state-of-the-art performance across multiple weakly-supervised OVS benchmarks, improving the average score by 2.1% (from 44.8% to 46.9%). Notably, our approach significantly enhances segmentation accuracy in complex, cluttered scenarios, with a gain of 6.3% on Cityscapes (from 36.6% to 42.9%).

Method: LAMP lifts image-editing as 3D priors to extract inter-object 3D transformations as continuous, geometry-aware representations. Key insight: image-editing encodes rich 2D spatial cues, lifting these implicit cues into 3D transformations provides fine-grained guidance.

Result: Extensive experiments demonstrate precise 3D transformations and strong zero-shot generalization in open-world manipulation.

Conclusion: LAMP provides a novel approach to extract 3D geometric representations from 2D image-editing priors, enabling better generalization for robotic manipulation in open-world scenarios.

Abstract: Human-like generalization in open-world remains a fundamental challenge for robotic manipulation. Existing learning-based methods, including reinforcement learning, imitation learning, and vision-language-action-models (VLAs), often struggle with novel tasks and unseen environments. Another promising direction is to explore generalizable representations that capture fine-grained spatial and geometric relations for open-world manipulation. While large-language-model (LLMs) and vision-language-model (VLMs) provide strong semantic reasoning based on language or annotated 2D representations, their limited 3D awareness restricts their applicability to fine-grained manipulation. To address this, we propose LAMP, which lifts image-editing as 3D priors to extract inter-object 3D transformations as continuous, geometry-aware representations. Our key insight is that image-editing inherently encodes rich 2D spatial cues, and lifting these implicit cues into 3D transformations provides fine-grained and accurate guidance for open-world manipulation. Extensive experiments demonstrate that \codename delivers precise 3D transformations and achieves strong zero-shot generalization in open-world manipulation. Project page: https://zju3dv.github.io/LAMP/.

Method: Proposes Faithful GRPO (FGRPO), a variant of Group Relative Policy Optimization that enforces consistency and grounding constraints via Lagrangian dual ascent. Incorporates batch-level consistency and grounding constraints into advantage computation within groups, adaptively adjusting constraint importance during optimization.

Result: FGRPO substantially improves reasoning quality, reducing inconsistency rate from 24.5% to 1.7% and improving visual grounding scores by +13%. Also improves final answer accuracy over simple GRPO, demonstrating faithful reasoning enables better answers.

Conclusion: Faithful reasoning through constraint enforcement in reinforcement learning leads to both improved reasoning quality and better final answer accuracy in multimodal reasoning models, addressing a critical gap in current approaches.

Abstract: Multimodal reasoning models (MRMs) trained with reinforcement learning with verifiable rewards (RLVR) show improved accuracy on visual reasoning benchmarks. However, we observe that accuracy gains often come at the cost of reasoning quality: generated Chain-of-Thought (CoT) traces are frequently inconsistent with the final answer and poorly grounded in the visual evidence. We systematically study this phenomenon across seven challenging real-world spatial reasoning benchmarks and find that it affects contemporary MRMs such as ViGoRL-Spatial, TreeVGR as well as our own models trained with standard Group Relative Policy Optimization (GRPO). We characterize CoT reasoning quality along two complementary axes: “logical consistency” (does the CoT entail the final answer?) and “visual grounding” (does each reasoning step accurately describe objects, attributes, and spatial relationships in the image?). To address this, we propose Faithful GRPO (FGRPO), a variant of GRPO that enforces consistency and grounding as constraints via Lagrangian dual ascent. FGRPO incorporates batch-level consistency and grounding constraints into the advantage computation within a group, adaptively adjusting the relative importance of constraints during optimization. We evaluate FGRPO on Qwen2.5-VL-7B and 3B backbones across seven spatial datasets. Our results show that FGRPO substantially improves reasoning quality, reducing the inconsistency rate from 24.5% to 1.7% and improving visual grounding scores by +13%. It also improves final answer accuracy over simple GRPO, demonstrating that faithful reasoning enables better answers.

Method: Formulates sparse NVS as low frame-rate video completion. Modifies video diffusion models with per-frame latent encodings and removes temporal positional embeddings to achieve permutation invariance for unordered input sets.

Result: Video models can be easily trained to “forget” about time with minimal supervision, achieving competitive performance on sparse-view NVS benchmarks.

Conclusion: Video diffusion models contain valuable multi-view knowledge that can be effectively adapted for novel view synthesis through architectural modifications for permutation invariance.

Abstract: We tackle the problem of sparse novel view synthesis (NVS) using video diffusion models; given $K$ ($\approx 5$) multi-view images of a scene and their camera poses, we predict the view from a target camera pose. Many prior approaches leverage generative image priors encoded via diffusion models. However, models trained on single images lack multi-view knowledge. We instead argue that video models already contain implicit multi-view knowledge and so should be easier to adapt for NVS. Our key insight is to formulate sparse NVS as a low frame-rate video completion task. However, one challenge is that sparse NVS is defined over an unordered set of inputs, often too sparse to admit a meaningful order, so the models should be $\textit{invariant}$ to permutations of that input set. To this end, we present FrameCrafter, which adapts video models (naturally trained with coherent frame orderings) to permutation-invariant NVS through several architectural modifications, including per-frame latent encodings and removal of temporal positional embeddings. Our results suggest that video models can be easily trained to “forget” about time with minimal supervision, producing competitive performance on sparse-view NVS benchmarks. Project page: https://frame-crafter.github.io/

Method: Proposes C-Score (Consistency Score), a confidence-weighted, annotation-free metric that quantifies intra-class explanation reproducibility via intensity-emphasised pairwise soft IoU across correctly classified instances. Evaluates six CAM techniques across three CNN architectures over thirty training epochs on chest X-ray dataset, covering transfer learning and fine-tuning phases.

Result: Identifies three distinct mechanisms of AUC-consistency dissociation: threshold-mediated gold list collapse, technique-specific attribution collapse at peak AUC, and class-level consistency masking in global aggregation. C-Score provides early warning of model instability - ScoreCAM deterioration on ResNet50V2 detectable one full checkpoint before catastrophic AUC collapse.

Conclusion: C-Score enables architecture-specific clinical deployment recommendations grounded in explanation quality rather than predictive ranking alone. The metric reveals hidden model behaviors and instability patterns that standard classification metrics miss, providing a more comprehensive evaluation framework for medical AI.

Abstract: Class Activation Mapping (CAM) methods are widely used to generate visual explanations for deep learning classifiers in medical imaging. However, existing evaluation frameworks assess whether explanations are correct, measured by localisation fidelity against radiologist annotations, rather than whether they are consistent: whether the model applies the same spatial reasoning strategy across different patients with the same pathology. We propose the C-Score (Consistency Score), a confidence-weighted, annotation-free metric that quantifies intra-class explanation reproducibility via intensity-emphasised pairwise soft IoU across correctly classified instances. We evaluate six CAM techniques: GradCAM, GradCAM++, LayerCAM, EigenCAM, ScoreCAM, and MS GradCAM++ across three CNN architectures (DenseNet201, InceptionV3, ResNet50V2) over thirty training epochs on the Kermany chest X-ray dataset, covering transfer learning and fine-tuning phases. We identify three distinct mechanisms of AUC-consistency dissociation, invisible to standard classification metrics: threshold-mediated gold list collapse, technique-specific attribution collapse at peak AUC, and class-level consistency masking in global aggregation. C-Score provides an early warning signal of impending model instability. ScoreCAM deterioration on ResNet50V2 is detectable one full checkpoint before catastrophic AUC collapse and yields architecture-specific clinical deployment recommendations grounded in explanation quality rather than predictive ranking alone.

Method: Proposes Phantom, a Physics-Infused Video Generation model that jointly models visual content and latent physical dynamics. It uses a physics-aware video representation as an abstract embedding of underlying physics, enabling joint prediction of physical dynamics and video frames without explicit physical specifications.

Result: Phantom outperforms existing methods in adherence to physical dynamics while maintaining competitive perceptual fidelity, as demonstrated through quantitative and qualitative evaluations on standard video generation and physics-aware benchmarks.

Conclusion: Integrating latent physical property inference directly into video generation enables models to produce physically consistent videos while maintaining visual realism, addressing a key limitation of current generative video models.

Abstract: Recent advances in generative video modeling, driven by large-scale datasets and powerful architectures, have yielded remarkable visual realism. However, emerging evidence suggests that simply scaling data and model size does not endow these systems with an understanding of the underlying physical laws that govern real-world dynamics. Existing approaches often fail to capture or enforce such physical consistency, resulting in unrealistic motion and dynamics. In his work, we investigate whether integrating the inference of latent physical properties directly into the video generation process can equip models with the ability to produce physically plausible videos. To this end, we propose Phantom, a Physics-Infused Video Generation model that jointly models the visual content and latent physical dynamics. Conditioned on observed video frames and inferred physical states, Phantom jointly predicts latent physical dynamics and generates future video frames. Phantom leverages a physics-aware video representation that serves as an abstract yet informaive embedding of the underlying physics, facilitating the joint prediction of physical dynamics alongside video content without requiring an explicit specification of a complex set of physical dynamics and properties. By integrating the inference of physical-aware video representation directly into the video generation process, Phantom produces video sequences that are both visually realistic and physically consistent. Quantitative and qualitative results on both standard video generation and physics-aware benchmarks demonstrate that Phantom not only outperforms existing methods in terms of adherence to physical dynamics but also delivers competitive perceptual fidelity.

Method: Two-layer paradigm: 1) World Layer reconstructs semantically rich 3D Gaussian scenes from real-world videos with occlusion-aware pipeline and animatable human avatars. 2) Agent Layer equips avatars with first-person RGB-D perception for embodied planning with spatial awareness and iterative reasoning, executing full-body actions.

Result: Agents achieve robust autonomous behavior with higher task success rates and fewer collisions than ablations and state-of-the-art planning methods. A benchmark for humanoid-scene interaction in diverse reconstructed environments is introduced.

Conclusion: Enables active digital human population at scale and advances human-centric embodied AI. The system allows digital humans to exhibit spontaneous, natural, goal-directed behaviors in novel 3D environments.

Abstract: Digital human generation has been studied for decades and supports a wide range of real-world applications. However, most existing systems are passively animated, relying on privileged state or scripted control, which limits scalability to novel environments. We instead ask: how can digital humans actively behave using only visual observations and specified goals in novel scenes? Achieving this would enable populating any 3D environments with digital humans at scale that exhibit spontaneous, natural, goal-directed behaviors. To this end, we introduce Visually-grounded Humanoid Agents, a coupled two-layer (world-agent) paradigm that replicates humans at multiple levels: they look, perceive, reason, and behave like real people in real-world 3D scenes. The World Layer reconstructs semantically rich 3D Gaussian scenes from real-world videos via an occlusion-aware pipeline and accommodates animatable Gaussian-based human avatars. The Agent Layer transforms these avatars into autonomous humanoid agents, equipping them with first-person RGB-D perception and enabling them to perform accurate, embodied planning with spatial awareness and iterative reasoning, which is then executed at the low level as full-body actions to drive their behaviors in the scene. We further introduce a benchmark to evaluate humanoid-scene interaction in diverse reconstructed environments. Experiments show our agents achieve robust autonomous behavior, yielding higher task success rates and fewer collisions than ablations and state-of-the-art planning methods. This work enables active digital human population and advances human-centric embodied AI. Data, code, and models will be open-sourced.

Method: Used five-class chest X-ray task with DenseNet201, ResNet50V2, and InceptionV3 under two-stage training protocol. Quantified drift with reference-free metrics capturing spatial localization and structural consistency of attribution maps (using LayerCAM and GradCAM++).

Result: Coarse anatomical localization remained stable across architectures, but overlap IoU revealed pronounced architecture-dependent reorganization of evidential structure. Stability rankings reversed across LayerCAM and GradCAM++ under converged predictive performance, showing explanation stability depends on architecture, optimization phase, and attribution objective.

Conclusion: Explanation stability is an interaction between architecture, optimization phase, and attribution objective, not just predictive performance. Semantic drift reveals potential shifts in underlying visual reasoning despite stable classification accuracy.

Abstract: Transfer learning followed by fine-tuning is widely adopted in medical image classification due to consistent gains in diagnostic performance. However, in multi-class settings with overlapping visual features, improvements in accuracy do not guarantee stability of the visual evidence used to support predictions. We define semantic drift as systematic changes in the attribution structure supporting a model’s predictions between transfer learning and full fine-tuning, reflecting potential shifts in underlying visual reasoning despite stable classification performance. Using a five-class chest X-ray task, we evaluate DenseNet201, ResNet50V2, and InceptionV3 under a two-stage training protocol and quantify drift with reference-free metrics capturing spatial localization and structural consistency of attribution maps. Across architectures, coarse anatomical localization remains stable, while overlap IoU reveals pronounced architecture-dependent reorganization of evidential structure. Beyond single-method analysis, stability rankings can reverse across LayerCAM and GradCAM++ under converged predictive performance, establishing explanation stability as an interaction between architecture, optimization phase, and attribution objective.

Method: Introduces MolmoWebMix dataset (100K+ synthetic trajectories, 30K+ human demonstrations, GUI perception data) and MolmoWeb family of multimodal web agents that operate as instruction-conditioned visual-language action policies, predicting browser actions from webpage screenshots and task instructions without needing HTML or specialized APIs.

Result: MolmoWeb agents (4B and 8B sizes) achieve state-of-the-art results on WebVoyager, Online-Mind2Web, and DeepShop benchmarks, outperforming similar-scale open models and even surpassing SoM agents built on larger closed models like GPT-4o. Test-time scaling via parallel rollouts achieves 94.7% pass@4 on WebVoyager.

Conclusion: Open multimodal web agents can achieve strong performance through diverse training data and visual-language modeling, enabling reproducible research and community progress in web automation.

Abstract: Web agents–autonomous systems that navigate and execute tasks on the web on behalf of users–have the potential to transform how people interact with the digital world. However, the most capable web agents today rely on proprietary models with undisclosed training data and recipes, limiting scientific understanding, reproducibility, and community-driven progress. We believe agents for the open web should be built in the open. To this end, we introduce (1) MolmoWebMix, a large and diverse mixture of browser task demonstrations and web-GUI perception data and (2) MolmoWeb, a family of fully open multimodal web agents. Specifically, MolmoWebMix combines over 100K synthetic task trajectories from multiple complementary generation pipelines with 30K+ human demonstrations, atomic web-skill trajectories, and GUI perception data, including referring expression grounding and screenshot question answering. MolmoWeb agents operate as instruction-conditioned visual-language action policies: given a task instruction and a webpage screenshot, they predict the next browser action, requiring no access to HTML, accessibility trees, or specialized APIs. Available in 4B and 8B size, on browser-use benchmarks like WebVoyager, Online-Mind2Web, and DeepShop, MolmoWeb agents achieve state-of-the-art results outperforming similar scale open-weight-only models such as Fara-7B, UI-Tars-1.5-7B, and Holo1-7B. MolmoWeb-8B also surpasses set-of-marks (SoM) agents built on much larger closed frontier models like GPT-4o. We further demonstrate consistent gains through test-time scaling via parallel rollouts with best-of-N selection, achieving 94.7% and 60.5% pass@4 (compared to 78.2% and 35.3% pass@1) on WebVoyager and Online-Mind2Web respectively. We will release model checkpoints, training data, code, and a unified evaluation harness to enable reproducibility and accelerate open research on web agents.

Method: 1) Offline Query Unifier to canonicalize heterogeneous query formats into shared declarative space; 2) Large-scale cross-dataset pretraining; 3) Efficient grounding head for long, untrimmed videos; 4) Single unified model architecture.

Result: Achieves SOTA across multiple benchmarks (GoalStep-StepGrounding, Ego4D-NLQ, TACoS, Charades-STA, ActivityNet-Captions). Matches or exceeds accuracy of MLLM-based approaches while being >100x smaller.

Conclusion: UniversalVTG demonstrates that scaling unified supervision with lightweight architecture can outperform dataset-specific models and compete with parameter-heavy MLLMs, offering a practical alternative for video temporal grounding.

Abstract: Video temporal grounding (VTG) is typically tackled with dataset-specific models that transfer poorly across domains and query styles. Recent efforts to overcome this limitation have adapted large multimodal language models (MLLMs) to VTG, but their high compute cost and limited video context still hinder long-video grounding. We instead scale unified supervision while keeping the model lightweight. We present UniversalVTG, a single VTG model trained with large-scale cross-dataset pretraining. An offline Query Unifier canonicalizes heterogeneous query formats into a shared declarative space, reducing linguistic mismatch and preventing the negative transfer observed under naïve joint training. Combined with an efficient grounding head, UniversalVTG scales to long, untrimmed videos. Across diverse benchmarks-GoalStep-StepGrounding, Ego4D-NLQ, TACoS, Charades-STA, and ActivityNet-Captions-one UniversalVTG checkpoint achieves state-of-the-art performance versus dedicated VTG models. Moreover, despite being $>100\times$ smaller than recent MLLM-based approaches, UniversalVTG matches or exceeds their accuracy on multiple benchmarks, offering a practical alternative to parameter-heavy MLLMs.

Method: 1) Generate 3D garments programmatically using GarmentCode and drape them via physics simulation. 2) Use novel re-texturing framework to transform synthetic renderings into photorealistic images while preserving geometry. 3) Introduce person identity preservation in re-texturing to generate paired person images for supervised training. 4) Train baseline fit-aware virtual try-on model using the FIT dataset.

Result: Created FIT dataset with over 1.13M try-on image triplets with precise body and garment measurements. Achieved state-of-the-art for fit-aware virtual try-on and provides benchmark for future research.

Conclusion: FIT addresses the critical gap in fit-aware virtual try-on through scalable synthetic data generation and re-texturing, enabling realistic visualization of garment fit across different body sizes and establishing new standards for the field.

Abstract: Given a person and a garment image, virtual try-on (VTO) aims to synthesize a realistic image of the person wearing the garment, while preserving their original pose and identity. Although recent VTO methods excel at visualizing garment appearance, they largely overlook a crucial aspect of the try-on experience: the accuracy of garment fit – for example, depicting how an extra-large shirt looks on an extra-small person. A key obstacle is the absence of datasets that provide precise garment and body size information, particularly for “ill-fit” cases, where garments are significantly too large or too small. Consequently, current VTO methods default to generating well-fitted results regardless of the garment or person size. In this paper, we take the first steps towards solving this open problem. We introduce FIT (Fit-Inclusive Try-on), a large-scale VTO dataset comprising over 1.13M try-on image triplets accompanied by precise body and garment measurements. We overcome the challenges of data collection via a scalable synthetic strategy: (1) We programmatically generate 3D garments using GarmentCode and drape them via physics simulation to capture realistic garment fit. (2) We employ a novel re-texturing framework to transform synthetic renderings into photorealistic images while strictly preserving geometry. (3) We introduce person identity preservation into our re-texturing model to generate paired person images (same person, different garments) for supervised training. Finally, we leverage our FIT dataset to train a baseline fit-aware virtual try-on model. Our data and results set the new state-of-the-art for fit-aware virtual try-on, as well as offer a robust benchmark for future research. We will make all data and code publicly available on our project page: https://johannakarras.github.io/FIT.

Method: SelfEvo introduces a self-distillation scheme using spatiotemporal context asymmetry. It systematically studies design choices including loss signals, forms of asymmetry, and training strategies to enable effective self-improvement for learning-based 4D perception without external annotations.

Result: Across eight benchmarks spanning diverse datasets and domains, SelfEvo consistently improves pretrained baselines and generalizes across base models (VGGT and π³), with up to 36.5% relative improvement in video depth estimation and 20.1% in camera estimation, without using any labeled data.

Conclusion: SelfEvo demonstrates that self-improving frameworks can effectively enhance multi-view reconstruction models using only unlabeled videos, overcoming the limitations of scarce 3D/4D annotations and enabling better scalability for dynamic scene understanding.

Abstract: Large-scale multi-view reconstruction models have made remarkable progress, but most existing approaches still rely on fully supervised training with ground-truth 3D/4D annotations. Such annotations are expensive and particularly scarce for dynamic scenes, limiting scalability. We propose SelfEvo, a self-improving framework that continually improves pretrained multi-view reconstruction models using unlabeled videos. SelfEvo introduces a self-distillation scheme using spatiotemporal context asymmetry, enabling self-improvement for learning-based 4D perception without external annotations. We systematically study design choices that make self-improvement effective, including loss signals, forms of asymmetry, and other training strategies. Across eight benchmarks spanning diverse datasets and domains, SelfEvo consistently improves pretrained baselines and generalizes across base models (e.g. VGGT and $π^3$), with significant gains on dynamic scenes. Overall, SelfEvo achieves up to 36.5% relative improvement in video depth estimation and 20.1% in camera estimation, without using any labeled data. Project Page: https://self-evo.github.io/.

Method: RewardFlow uses multi-reward Langevin dynamics to steer pretrained models at inference time. It combines differentiable rewards for semantic alignment, perceptual fidelity, localized grounding, object consistency, and human preference, plus a novel VQA-based reward for fine-grained semantic supervision. A prompt-aware adaptive policy extracts semantic primitives, infers edit intent, and dynamically modulates reward weights and step sizes during sampling.

Result: RewardFlow achieves state-of-the-art performance across several image editing and compositional generation benchmarks, delivering superior edit fidelity and compositional alignment compared to existing methods.

Conclusion: RewardFlow provides an effective inversion-free framework for steering pretrained generative models using multi-reward optimization, enabling precise control over complex editing tasks through coordinated heterogeneous objectives.

Abstract: We introduce RewardFlow, an inversion-free framework that steers pretrained diffusion and flow-matching models at inference time through multi-reward Langevin dynamics. RewardFlow unifies complementary differentiable rewards for semantic alignment, perceptual fidelity, localized grounding, object consistency, and human preference, and further introduces a differentiable VQA-based reward that provides fine-grained semantic supervision through language-vision reasoning. To coordinate these heterogeneous objectives, we design a prompt-aware adaptive policy that extracts semantic primitives from the instruction, infers edit intent, and dynamically modulates reward weights and step sizes throughout sampling. Across several image editing and compositional generation benchmarks, RewardFlow delivers state-of-the-art edit fidelity and compositional alignment.

Method: Created ParseBench with ~2,000 human-verified pages from enterprise documents (insurance, finance, government) organized around five capability dimensions: tables, charts, content faithfulness, semantic formatting, and visual grounding. Evaluated 14 methods including vision-language models, specialized document parsers, and LlamaParse.

Result: Revealed fragmented capability landscape - no method consistently strong across all five dimensions. LlamaParse Agentic achieved highest overall score at 67.4%. Benchmark highlights remaining capability gaps across current document parsing systems.

Conclusion: ParseBench addresses the need for better evaluation of document parsing for AI agents, showing current systems have significant gaps in semantic correctness across multiple dimensions critical for enterprise automation.

Abstract: AI agents are changing the requirements for document parsing. What matters is \emph{semantic correctness}: parsed output must preserve the structure and meaning needed for autonomous decisions, including correct table structure, precise chart data, semantically meaningful formatting, and visual grounding. Existing benchmarks do not fully capture this setting for enterprise automation, relying on narrow document distributions and text-similarity metrics that miss agent-critical failures. We introduce \textbf{ParseBench}, a benchmark of ${\sim}2{,}000$ human-verified pages from enterprise documents spanning insurance, finance, and government, organized around five capability dimensions: tables, charts, content faithfulness, semantic formatting, and visual grounding. Across 14 methods spanning vision-language models, specialized document parsers, and LlamaParse, the benchmark reveals a fragmented capability landscape: no method is consistently strong across all five dimensions. LlamaParse Agentic achieves the highest overall score at \agenticoverall%, and the benchmark highlights the remaining capability gaps across current systems. Dataset and evaluation code are available on \href{https://huggingface.co/datasets/llamaindex/ParseBench}{HuggingFace} and \href{https://github.com/run-llama/ParseBench}{GitHub}.

Method: Introduces a neural global context representation that efficiently compresses and retains long-range scene information using lightweight neural sub-networks rapidly adapted during test time via self-supervised objectives.

Result: Achieves leading pose accuracy and state-of-the-art 3D reconstruction accuracy on KITTI Odometry and Oxford Spires datasets while maintaining efficiency in handling ultra-large scenes.

Conclusion: The proposed global context representation effectively enhances reconstruction accuracy and consistency for large-scale 3D scene reconstruction from long video sequences.

Abstract: This paper addresses the task of large-scale 3D scene reconstruction from long video sequences. Recent feed-forward reconstruction models have shown promising results by directly regressing 3D geometry from RGB images without explicit 3D priors or geometric constraints. However, these methods often struggle to maintain reconstruction accuracy and consistency over long sequences due to limited memory capacity and the inability to effectively capture global contextual cues. In contrast, humans can naturally exploit the global understanding of the scene to inform local perception. Motivated by this, we propose a novel neural global context representation that efficiently compresses and retains long-range scene information, enabling the model to leverage extensive contextual cues for enhanced reconstruction accuracy and consistency. The context representation is realized through a set of lightweight neural sub-networks that are rapidly adapted during test time via self-supervised objectives, which substantially increases memory capacity without incurring significant computational overhead. The experiments on multiple large-scale benchmarks, including the KITTI Odometry~\cite{Geiger2012CVPR} and Oxford Spires~\cite{tao2025spires} datasets, demonstrate the effectiveness of our approach in handling ultra-large scenes, achieving leading pose accuracy and state-of-the-art 3D reconstruction accuracy while maintaining efficiency. Code is available at https://zju3dv.github.io/scal3r.

Method: E-3DPSM (event-driven continuous pose state machine) aligns continuous human motion with fine-grained event dynamics; evolves latent states and predicts continuous changes in 3D joint positions associated with observed events, which are fused with direct 3D human pose predictions.

Result: Sets new state-of-the-art on two benchmarks, improving accuracy by up to 19% (MPJPE) and temporal stability by up to 2.7x; runs in real-time at 80 Hz on a single workstation.

Conclusion: E-3DPSM enables stable, drift-free 3D pose reconstructions by better leveraging the asynchronous and continuous nature of event streams, addressing limitations of previous methods for event-based egocentric 3D human pose estimation.

Abstract: Event cameras offer multiple advantages in monocular egocentric 3D human pose estimation from head-mounted devices, such as millisecond temporal resolution, high dynamic range, and negligible motion blur. Existing methods effectively leverage these properties, but suffer from low 3D estimation accuracy, insufficient in many applications (e.g., immersive VR/AR). This is due to the design not being fully tailored towards event streams (e.g., their asynchronous and continuous nature), leading to high sensitivity to self-occlusions and temporal jitter in the estimates. This paper rethinks the setting and introduces E-3DPSM, an event-driven continuous pose state machine for event-based egocentric 3D human pose estimation. E-3DPSM aligns continuous human motion with fine-grained event dynamics; it evolves latent states and predicts continuous changes in 3D joint positions associated with observed events, which are fused with direct 3D human pose predictions, leading to stable and drift-free final 3D pose reconstructions. E-3DPSM runs in real-time at 80 Hz on a single workstation and sets a new state of the art in experiments on two benchmarks, improving accuracy by up to 19% (MPJPE) and temporal stability by up to 2.7x. See our project page for the source code and trained models.

Method: Proposes HDPO framework that reframes tool efficiency as a strictly conditional objective rather than competing scalar objective. Maintains two orthogonal optimization channels: accuracy channel maximizing task correctness, and efficiency channel enforcing execution economy exclusively within accurate trajectories via conditional advantage estimation.

Result: The resulting model Metis reduces tool invocations by orders of magnitude while simultaneously elevating reasoning accuracy. Extensive evaluations demonstrate significant improvements over existing approaches.

Conclusion: HDPO’s decoupled architecture induces a cognitive curriculum where agents first master task resolution before refining self-reliance, effectively addressing the tool overuse problem in multimodal agents.

Abstract: The advent of agentic multimodal models has empowered systems to actively interact with external environments. However, current agents suffer from a profound meta-cognitive deficit: they struggle to arbitrate between leveraging internal knowledge and querying external utilities. Consequently, they frequently fall prey to blind tool invocation, resorting to reflexive tool execution even when queries are resolvable from the raw visual context. This pathological behavior precipitates severe latency bottlenecks and injects extraneous noise that derails sound reasoning. Existing reinforcement learning protocols attempt to mitigate this via a scalarized reward that penalizes tool usage. Yet, this coupled formulation creates an irreconcilable optimization dilemma: an aggressive penalty suppresses essential tool use, whereas a mild penalty is entirely subsumed by the variance of the accuracy reward during advantage normalization, rendering it impotent against tool overuse. To transcend this bottleneck, we propose HDPO, a framework that reframes tool efficiency from a competing scalar objective to a strictly conditional one. By eschewing reward scalarization, HDPO maintains two orthogonal optimization channels: an accuracy channel that maximizes task correctness, and an efficiency channel that enforces execution economy exclusively within accurate trajectories via conditional advantage estimation. This decoupled architecture naturally induces a cognitive curriculum-compelling the agent to first master task resolution before refining its self-reliance. Extensive evaluations demonstrate that our resulting model, Metis, reduces tool invocations by orders of magnitude while simultaneously elevating reasoning accuracy.

Method: NUMINA uses an identify-then-guide approach: 1) identifies prompt-layout inconsistencies by selecting discriminative self- and cross-attention heads to derive a countable latent layout, 2) refines this layout conservatively, and 3) modulates cross-attention to guide regeneration.

Result: On CountBench, NUMINA improves counting accuracy by up to 7.4% on Wan2.1-1.3B, and by 4.9% and 5.5% on 5B and 14B models respectively. CLIP alignment improves while maintaining temporal consistency.

Conclusion: Structural guidance complements existing methods like seed search and prompt enhancement, offering a practical path toward count-accurate text-to-video diffusion without requiring model retraining.

Abstract: Text-to-video diffusion models have enabled open-ended video synthesis, but often struggle with generating the correct number of objects specified in a prompt. We introduce NUMINA , a training-free identify-then-guide framework for improved numerical alignment. NUMINA identifies prompt-layout inconsistencies by selecting discriminative self- and cross-attention heads to derive a countable latent layout. It then refines this layout conservatively and modulates cross-attention to guide regeneration. On the introduced CountBench, NUMINA improves counting accuracy by up to 7.4% on Wan2.1-1.3B, and by 4.9% and 5.5% on 5B and 14B models, respectively. Furthermore, CLIP alignment is improved while maintaining temporal consistency. These results demonstrate that structural guidance complements seed search and prompt enhancement, offering a practical path toward count-accurate text-to-video diffusion. The code is available at https://github.com/H-EmbodVis/NUMINA.

Method: Three-step approach: (1) Compress temporally-consistent deformable Gaussians into free-form bones; (2) Extract Mean Curvature Skeleton from canonical Gaussians and refine it temporally; (3) Bind skeleton and bones via non-parametric Partwise Motion Matching (PartMM) that matches, retrieves, and blends existing motions.

Result: Achieves 17.3% PSNR gains over Linear Blend Skinning and 21.7% over Bag-of-Bones for reanimation of unseen poses. PartMM shows 48.4% RMSE improvement over robust LBS and outperforms GRU/MLP methods by >20%, especially effective in low-data regimes (~1000 frames).

Conclusion: Skelebones successfully compresses 4D shape dynamics into compact, controllable representations that maintain reconstruction fidelity while enabling expressive reanimation, with PartMM showing strong generalization across representations.

Abstract: Free-form bones, that conform closely to the surface, can effectively capture non-rigid deformations, but lack a kinematic structure necessary for intuitive control. Thus, we propose a Scaffold-Skin Rigging System, termed “Skelebones”, with three key steps: (1) Bones: compress temporally-consistent deformable Gaussians into free-form bones, approximating non-rigid surface deformations; (2) Skeleton: extract a Mean Curvature Skeleton from canonical Gaussians and refine it temporally, ensuring a category-agnostic, motion-adaptive, and topology-correct kinematic structure; (3) Binding: bind the skeleton and bones via non-parametric partwise motion matching (PartMM), synthesizing novel bone motions by matching, retrieving, and blending existing ones. Collectively, these three steps enable us to compress the Level of Dynamics of 4D shapes into compact skelebones that are both controllable and expressive. We validate our approach on both synthetic and real-world datasets, achieving significant improvements in reanimation performance across unseen poses-with 17.3% PSNR gains over Linear Blend Skinning (LBS) and 21.7% over Bag-of-Bones (BoB)-while maintaining excellent reconstruction fidelity, particularly for characters exhibiting complex non-rigid surface dynamics. Our Partwise Motion Matching algorithm demonstrates strong generalization to both Gaussian and mesh representations, especially under low-data regime (~1000 frames), achieving 48.4% RMSE improvement over robust LBS and outperforming GRU- and MLP-based learning methods by >20%. Code will be made publicly available for research purposes at cookmaker.cn/gaussianimate.

Method: ETCH-X upgrades ETCH with: 1) tightness-aware fitting paradigm to filter out clothing dynamics (“undress”), 2) extends expressiveness with SMPL-X body model, 3) replaces explicit sparse markers with implicit dense correspondences (“dense fit”) for more robust and fine-grained body fitting. Uses disentangled “undress” and “dense fit” modular stages for scalable training on composable data sources.

Result: Achieves robust and expressive fitting across diverse clothing, poses, and levels of input completeness. Substantial performance improvement over ETCH on both seen data (4D-Dress: MPJPE-All 33.0%, CAPE: V2V-Hands 35.8%) and unseen data (BEDLAM2.0: MPJPE-All 80.8%, V2V-All 80.5%).

Conclusion: ETCH-X provides an all-in-one solution for human body fitting that is both locally expressive and globally robust, handling clothing dynamics, pose variations, and partial inputs through its modular “undress” and “dense fit” approach with SMPL-X integration.

Abstract: Human body fitting, which aligns parametric body models such as SMPL to raw 3D point clouds of clothed humans, serves as a crucial first step for downstream tasks like animation and texturing. An effective fitting method should be both locally expressive-capturing fine details such as hands and facial features-and globally robust to handle real-world challenges, including clothing dynamics, pose variations, and noisy or partial inputs. Existing approaches typically excel in only one aspect, lacking an all-in-one solution.We upgrade ETCH to ETCH-X, which leverages a tightness-aware fitting paradigm to filter out clothing dynamics (“undress”), extends expressiveness with SMPL-X, and replaces explicit sparse markers (which are highly sensitive to partial data) with implicit dense correspondences (“dense fit”) for more robust and fine-grained body fitting. Our disentangled “undress” and “dense fit” modular stages enable separate and scalable training on composable data sources, including diverse simulated garments (CLOTH3D), large-scale full-body motions (AMASS), and fine-grained hand gestures (InterHand2.6M), improving outfit generalization and pose robustness of both bodies and hands. Our approach achieves robust and expressive fitting across diverse clothing, poses, and levels of input completeness, delivering a substantial performance improvement over ETCH on both: 1) seen data, such as 4D-Dress (MPJPE-All, 33.0% ) and CAPE (V2V-Hands, 35.8% ), and 2) unseen data, such as BEDLAM2.0 (MPJPE-All, 80.8% ; V2V-All, 80.5% ). Code and models will be released at https://xiaobenli00.github.io/ETCH-X/.

Method: Unable to determine method due to fetch failure

Result: Unable to determine results due to fetch failure

Conclusion: Unable to determine conclusion due to fetch failure

Abstract: Failed to fetch summary for 2402.08267: Page request resulted in HTTP 429 (https://export.arxiv.org/api/query?search_query=&id_list=2402.08267&sortBy=relevance&sortOrder=descending&start=0&max_results=100)

Method: Unable to determine method due to access error

Result: Unable to determine results due to access error

Conclusion: Unable to determine conclusion due to access error

Abstract: Failed to fetch summary for 2402.12797: Page request resulted in HTTP 429 (https://export.arxiv.org/api/query?search_query=&id_list=2402.12797&sortBy=relevance&sortOrder=descending&start=0&max_results=100)

Method: Unable to determine method due to fetch failure

Result: Unable to determine results due to fetch failure

Conclusion: Unable to determine conclusion due to fetch failure

Abstract: Failed to fetch summary for 2602.13235: Page request resulted in HTTP 429 (https://export.arxiv.org/api/query?search_query=&id_list=2602.13235&sortBy=relevance&sortOrder=descending&start=0&max_results=100)

Method: Cannot determine method due to failed paper fetch

Result: Cannot determine results due to failed paper fetch

Conclusion: Cannot determine conclusion due to failed paper fetch

Abstract: Failed to fetch summary for 2411.07799: Page request resulted in HTTP 429 (https://export.arxiv.org/api/query?search_query=&id_list=2411.07799&sortBy=relevance&sortOrder=descending&start=0&max_results=100)

Method: Cannot determine method as paper content is unavailable

Result: Cannot determine results as paper content is unavailable

Conclusion: Cannot determine conclusion as paper content is unavailable

Abstract: Failed to fetch summary for 2412.08637: Page request resulted in HTTP 429 (https://export.arxiv.org/api/query?search_query=&id_list=2412.08637&sortBy=relevance&sortOrder=descending&start=0&max_results=100)

Method: Cannot determine method as paper content is unavailable

Result: Cannot determine results as paper content is unavailable

Conclusion: Cannot determine conclusion as paper content is unavailable

Abstract: Failed to fetch summary for 2412.10437: Page request resulted in HTTP 429 (https://export.arxiv.org/api/query?search_query=&id_list=2412.10437&sortBy=relevance&sortOrder=descending&start=0&max_results=100)

Method: Unable to determine method due to fetch failure

Result: Unable to determine results due to fetch failure

Conclusion: Unable to determine conclusion due to fetch failure

Abstract: Failed to fetch summary for 2412.20718: Page request resulted in HTTP 429 (https://export.arxiv.org/api/query?search_query=&id_list=2412.20718&sortBy=relevance&sortOrder=descending&start=0&max_results=100)

Method: No method information available due to HTTP 429 error preventing access to paper details.

Result: No results available as the paper content could not be retrieved from arXiv API.

Conclusion: Unable to analyze paper due to technical limitations in accessing the content from arXiv.

Abstract: Failed to fetch summary for 2603.18472: Page request resulted in HTTP 429 (https://export.arxiv.org/api/query?search_query=&id_list=2603.18472&sortBy=relevance&sortOrder=descending&start=0&max_results=100)

Method: Unable to determine method due to failed paper fetch

Result: Unable to determine results due to failed paper fetch

Conclusion: Unable to determine conclusion due to failed paper fetch

Abstract: Failed to fetch summary for 2502.02514: Page request resulted in HTTP 429 (https://export.arxiv.org/api/query?search_query=&id_list=2502.02514&sortBy=relevance&sortOrder=descending&start=0&max_results=100)

Method: Unable to determine method due to failed API request

Result: Unable to determine results due to failed API request

Conclusion: Unable to draw conclusions due to failed API request

Abstract: Failed to fetch summary for 2503.02537: Page request resulted in HTTP 429 (https://export.arxiv.org/api/query?search_query=&id_list=2503.02537&sortBy=relevance&sortOrder=descending&start=0&max_results=100)

Method: Unable to determine method due to access error

Result: Unable to determine results due to access error

Conclusion: Unable to determine conclusion due to access error

Abstract: Failed to fetch summary for 2503.06717: Page request resulted in HTTP 429 (https://export.arxiv.org/api/query?search_query=&id_list=2503.06717&sortBy=relevance&sortOrder=descending&start=0&max_results=100)

Method: Unable to determine method due to fetch failure

Result: Unable to determine results due to fetch failure

Conclusion: Unable to determine conclusion due to fetch failure

Abstract: Failed to fetch summary for 2503.10183: Page request resulted in HTTP 429 (https://export.arxiv.org/api/query?search_query=&id_list=2503.10183&sortBy=relevance&sortOrder=descending&start=0&max_results=100)

Method: Unable to determine method due to failed API request

Result: Unable to determine results due to failed API request

Conclusion: Unable to draw conclusions due to failed API request

Abstract: Failed to fetch summary for 2503.24135: Page request resulted in HTTP 429 (https://export.arxiv.org/api/query?search_query=&id_list=2503.24135&sortBy=relevance&sortOrder=descending&start=0&max_results=100)

Method: Unable to determine method due to failed fetch

Result: Unable to determine results due to failed fetch

Conclusion: Unable to determine conclusion due to failed fetch

Abstract: Failed to fetch summary for 2504.13015: Page request resulted in HTTP 429 (https://export.arxiv.org/api/query?search_query=&id_list=2504.13015&sortBy=relevance&sortOrder=descending&start=0&max_results=100)

Method: Unable to determine method due to fetch failure

Result: Unable to determine results due to fetch failure

Conclusion: Unable to determine conclusion due to fetch failure

Abstract: Failed to fetch summary for 2504.17069: Page request resulted in HTTP 429 (https://export.arxiv.org/api/query?search_query=&id_list=2504.17069&sortBy=relevance&sortOrder=descending&start=0&max_results=100)

Method: Unable to determine method due to fetch failure

Result: Unable to determine results due to fetch failure

Conclusion: Unable to determine conclusion due to fetch failure

Abstract: Failed to fetch summary for 2505.17732: Page request resulted in HTTP 429 (https://export.arxiv.org/api/query?search_query=&id_list=2505.17732&sortBy=relevance&sortOrder=descending&start=0&max_results=100)

Method: Unable to determine method due to access error

Result: Unable to determine results due to access error

Conclusion: Unable to determine conclusion due to access error

Abstract: Failed to fetch summary for 2505.24499: Page request resulted in HTTP 429 (https://export.arxiv.org/api/query?search_query=&id_list=2505.24499&sortBy=relevance&sortOrder=descending&start=0&max_results=100)

Method: Cannot determine method due to inability to access paper content

Result: Cannot determine results due to inability to access paper content

Conclusion: Cannot determine conclusion due to inability to access paper content

Abstract: Failed to fetch summary for 2506.12040: Page request resulted in HTTP 429 (https://export.arxiv.org/api/query?search_query=&id_list=2506.12040&sortBy=relevance&sortOrder=descending&start=0&max_results=100)

Method: Cannot analyze method due to failed API request

Result: No results available - technical error prevented access to paper content

Conclusion: Paper analysis impossible due to HTTP 429 error from arXiv API

Abstract: Failed to fetch summary for 2505.24848: Page request resulted in HTTP 429 (https://export.arxiv.org/api/query?search_query=&id_list=2505.24848&sortBy=relevance&sortOrder=descending&start=0&max_results=100)

Method: Unable to determine method due to access error

Result: Unable to determine results due to access error

Conclusion: Unable to determine conclusion due to access error

Abstract: Failed to fetch summary for 2506.17212: Page request resulted in HTTP 429 (https://export.arxiv.org/api/query?search_query=&id_list=2506.17212&sortBy=relevance&sortOrder=descending&start=0&max_results=100)

Method: Cannot determine method as paper content is unavailable

Result: Cannot determine results as paper content is unavailable

Conclusion: Cannot draw conclusions as paper content is unavailable

Abstract: Failed to fetch summary for 2507.04678: Page request resulted in HTTP 429 (https://export.arxiv.org/api/query?search_query=&id_list=2507.04678&sortBy=relevance&sortOrder=descending&start=0&max_results=100)

Method: Unable to determine method due to abstract fetch failure

Result: Unable to determine results due to abstract fetch failure

Conclusion: Unable to draw conclusions due to abstract fetch failure

Abstract: Failed to fetch summary for 2507.06949: Page request resulted in HTTP 429 (https://export.arxiv.org/api/query?search_query=&id_list=2507.06949&sortBy=relevance&sortOrder=descending&start=0&max_results=100)

Method: Cannot determine method as paper content is unavailable

Result: Cannot determine results as paper content is unavailable

Conclusion: Cannot draw conclusions as paper content is unavailable

Abstract: Failed to fetch summary for 2508.07514: Page request resulted in HTTP 429 (https://export.arxiv.org/api/query?search_query=&id_list=2508.07514&sortBy=relevance&sortOrder=descending&start=0&max_results=100)

Method: Cannot determine method without access to paper content

Result: Cannot determine results without access to paper content

Conclusion: Cannot determine conclusion without access to paper content

Abstract: Failed to fetch summary for 2509.07673: Page request resulted in HTTP 429 (https://export.arxiv.org/api/query?search_query=&id_list=2509.07673&sortBy=relevance&sortOrder=descending&start=0&max_results=100)

Method: Unable to determine method as the paper content could not be retrieved.

Result: Unable to determine results as the paper content could not be retrieved.

Conclusion: Unable to draw conclusions about the paper due to content retrieval failure.

Abstract: Failed to fetch summary for 2509.08016: Page request resulted in HTTP 429 (https://export.arxiv.org/api/query?search_query=&id_list=2509.08016&sortBy=relevance&sortOrder=descending&start=0&max_results=100)

Method: Cannot determine method as paper content is unavailable

Result: Cannot determine results as paper content is unavailable

Conclusion: Cannot determine conclusion as paper content is unavailable

Abstract: Failed to fetch summary for 2509.23310: Page request resulted in HTTP 429 (https://export.arxiv.org/api/query?search_query=&id_list=2509.23310&sortBy=relevance&sortOrder=descending&start=0&max_results=100)

Method: Unable to determine method due to access error

Result: Unable to determine results due to access error

Conclusion: Unable to draw conclusions due to access error

Abstract: Failed to fetch summary for 2509.23322: Page request resulted in HTTP 429 (https://export.arxiv.org/api/query?search_query=&id_list=2509.23322&sortBy=relevance&sortOrder=descending&start=0&max_results=100)

Method: Unable to determine method due to fetch failure

Result: Unable to determine results due to fetch failure

Conclusion: Unable to determine conclusion due to fetch failure

Abstract: Failed to fetch summary for 2509.26036: Page request resulted in HTTP 429 (https://export.arxiv.org/api/query?search_query=&id_list=2509.26036&sortBy=relevance&sortOrder=descending&start=0&max_results=100)

Method: Cannot determine method without access to paper content

Result: Cannot determine results without access to paper content

Conclusion: Cannot draw conclusions without access to paper content

Abstract: Failed to fetch summary for 2510.04628: Page request resulted in HTTP 429 (https://export.arxiv.org/api/query?search_query=&id_list=2510.04628&sortBy=relevance&sortOrder=descending&start=0&max_results=100)

Method: Cannot determine method without access to paper content

Result: Cannot determine results without access to paper content

Conclusion: Cannot draw conclusions without access to paper content

Abstract: Failed to fetch summary for 2511.22475: Page request resulted in HTTP 429 (https://export.arxiv.org/api/query?search_query=&id_list=2511.22475&sortBy=relevance&sortOrder=descending&start=0&max_results=100)

Method: Unable to determine method due to fetch failure

Result: Unable to determine results due to fetch failure

Conclusion: Unable to determine conclusion due to fetch failure

Abstract: Failed to fetch summary for 2510.07927: Page request resulted in HTTP 429 (https://export.arxiv.org/api/query?search_query=&id_list=2510.07927&sortBy=relevance&sortOrder=descending&start=0&max_results=100)

Method: Unable to determine method due to fetch failure

Result: Unable to determine results due to fetch failure

Conclusion: Unable to determine conclusion due to fetch failure

Abstract: Failed to fetch summary for 2510.12184: Page request resulted in HTTP 429 (https://export.arxiv.org/api/query?search_query=&id_list=2510.12184&sortBy=relevance&sortOrder=descending&start=0&max_results=100)

Method: Unable to determine method due to lack of access to paper content

Result: Unable to determine results due to HTTP 429 error preventing access to paper information

Conclusion: The analysis could not be completed due to technical limitations in accessing the paper content from arXiv

Abstract: Failed to fetch summary for 2510.15770: Page request resulted in HTTP 429 (https://export.arxiv.org/api/query?search_query=&id_list=2510.15770&sortBy=relevance&sortOrder=descending&start=0&max_results=100)

Method: Unable to determine method due to fetch failure

Result: Unable to determine results due to fetch failure

Conclusion: Unable to draw conclusions due to fetch failure

Abstract: Failed to fetch summary for 2510.20549: Page request resulted in HTTP 429 (https://export.arxiv.org/api/query?search_query=&id_list=2510.20549&sortBy=relevance&sortOrder=descending&start=0&max_results=100)

Method: Cannot determine method without access to paper content

Result: Cannot determine results without access to paper content

Conclusion: Cannot draw conclusions without access to paper content

Abstract: Failed to fetch summary for 2512.19253: Page request resulted in HTTP 429 (https://export.arxiv.org/api/query?search_query=&id_list=2512.19253&sortBy=relevance&sortOrder=descending&start=0&max_results=100)

Method: Unable to determine method due to fetch failure

Result: Unable to determine results due to fetch failure

Conclusion: Unable to determine conclusion due to fetch failure

Abstract: Failed to fetch summary for 2510.21366: Page request resulted in HTTP 429 (https://export.arxiv.org/api/query?search_query=&id_list=2510.21366&sortBy=relevance&sortOrder=descending&start=0&max_results=100)

Method: Cannot determine method as paper content is unavailable

Result: Cannot determine results as paper content is unavailable

Conclusion: Cannot draw conclusions about the paper due to content unavailability

Abstract: Failed to fetch summary for 2512.21602: Page request resulted in HTTP 429 (https://export.arxiv.org/api/query?search_query=&id_list=2512.21602&sortBy=relevance&sortOrder=descending&start=0&max_results=100)

Method: Unable to determine method due to fetch failure

Result: Unable to determine results due to fetch failure

Conclusion: Unable to determine conclusion due to fetch failure

Abstract: Failed to fetch summary for 2511.11435: Page request resulted in HTTP 429 (https://export.arxiv.org/api/query?search_query=&id_list=2511.11435&sortBy=relevance&sortOrder=descending&start=0&max_results=100)

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

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

Conclusion: Cannot draw conclusions without access to the paper content.

Abstract: Failed to fetch summary for 2511.09170: Page request resulted in HTTP 429 (https://export.arxiv.org/api/query?search_query=&id_list=2511.09170&sortBy=relevance&sortOrder=descending&start=0&max_results=100)

Method: Unable to determine method due to fetch failure

Result: Unable to determine results due to fetch failure

Conclusion: Unable to draw conclusions due to fetch failure

Abstract: Failed to fetch summary for 2511.15496: Page request resulted in HTTP 429 (https://export.arxiv.org/api/query?search_query=&id_list=2511.15496&sortBy=relevance&sortOrder=descending&start=0&max_results=100)

Method: Cannot determine method without access to paper content

Result: Cannot determine results without access to paper content

Conclusion: Cannot draw conclusions without access to paper content

Abstract: Failed to fetch summary for 2511.18787: Page request resulted in HTTP 429 (https://export.arxiv.org/api/query?search_query=&id_list=2511.18787&sortBy=relevance&sortOrder=descending&start=0&max_results=100)

Method: Unable to determine method due to access restrictions

Result: Unable to determine results due to access restrictions

Conclusion: Unable to determine conclusion due to access restrictions

Abstract: Failed to fetch summary for 2511.20162: Page request resulted in HTTP 429 (https://export.arxiv.org/api/query?search_query=&id_list=2511.20162&sortBy=relevance&sortOrder=descending&start=0&max_results=100)

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

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

Conclusion: Cannot draw conclusions about the paper due to technical limitations in accessing the content

Abstract: Failed to fetch summary for 2512.01236: Page request resulted in HTTP 429 (https://export.arxiv.org/api/query?search_query=&id_list=2512.01236&sortBy=relevance&sortOrder=descending&start=0&max_results=100)

Method: Cannot determine method as paper content is unavailable

Result: Cannot determine results as paper content is unavailable

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

Abstract: Failed to fetch summary for 2512.03532: Page request resulted in HTTP 429 (https://export.arxiv.org/api/query?search_query=&id_list=2512.03532&sortBy=relevance&sortOrder=descending&start=0&max_results=100)

Method: Cannot determine method as paper content is unavailable

Result: Cannot determine results as paper content is unavailable

Conclusion: Cannot determine conclusion as paper content is unavailable

Abstract: Failed to fetch summary for 2512.06774: Page request resulted in HTTP 429 (https://export.arxiv.org/api/query?search_query=&id_list=2512.06774&sortBy=relevance&sortOrder=descending&start=0&max_results=100)

Method: Unable to determine method due to failed paper fetch

Result: Unable to determine results due to failed paper fetch

Conclusion: Unable to determine conclusion due to failed paper fetch

Abstract: Failed to fetch summary for 2512.03745: Page request resulted in HTTP 429 (https://export.arxiv.org/api/query?search_query=&id_list=2512.03745&sortBy=relevance&sortOrder=descending&start=0&max_results=100)

Method: Method unknown - paper content unavailable due to HTTP 429 error from arXiv API.

Result: No results available - failed to fetch paper summary.

Conclusion: Unable to analyze paper due to technical limitations in accessing the content.

Abstract: Failed to fetch summary for 2601.04068: Page request resulted in HTTP 429 (https://export.arxiv.org/api/query?search_query=&id_list=2601.04068&sortBy=relevance&sortOrder=descending&start=0&max_results=100)

Method: Cannot determine method due to failed data retrieval

Result: Cannot determine results due to failed data retrieval

Conclusion: Cannot draw conclusions due to failed data retrieval

Abstract: Failed to fetch summary for 2512.08410: Page request resulted in HTTP 429 (https://export.arxiv.org/api/query?search_query=&id_list=2512.08410&sortBy=relevance&sortOrder=descending&start=0&max_results=100)

Method: No method information available - arXiv API returned HTTP 429 (Too Many Requests) error

Result: No results available - paper content inaccessible due to API rate limiting

Conclusion: Cannot analyze paper due to technical limitations in accessing the abstract

Abstract: Failed to fetch summary for 2601.10587: Page request resulted in HTTP 429 (https://export.arxiv.org/api/query?search_query=&id_list=2601.10587&sortBy=relevance&sortOrder=descending&start=0&max_results=100)

Method: Unable to determine method due to access limitations

Result: Unable to determine results due to access limitations

Conclusion: Unable to draw conclusions due to access limitations

Abstract: Failed to fetch summary for 2512.09665: Page request resulted in HTTP 429 (https://export.arxiv.org/api/query?search_query=&id_list=2512.09665&sortBy=relevance&sortOrder=descending&start=0&max_results=100)

Method: Cannot determine method due to inability to access paper content

Result: Cannot determine results due to inability to access paper content

Conclusion: Cannot draw conclusions due to inability to access paper content

Abstract: Failed to fetch summary for 2602.06912: Page request resulted in HTTP 429 (https://export.arxiv.org/api/query?search_query=&id_list=2602.06912&sortBy=relevance&sortOrder=descending&start=0&max_results=100)

Method: Cannot determine method without paper content.

Result: Cannot determine results without paper content.

Conclusion: Cannot draw conclusions without paper content.

Abstract: Failed to fetch summary for 2512.17445: Page request resulted in HTTP 429 (https://export.arxiv.org/api/query?search_query=&id_list=2512.17445&sortBy=relevance&sortOrder=descending&start=0&max_results=100)

Method: Unable to determine method due to fetch failure

Result: Unable to determine results due to fetch failure

Conclusion: Unable to determine conclusion due to fetch failure

Abstract: Failed to fetch summary for 2512.17489: Page request resulted in HTTP 429 (https://export.arxiv.org/api/query?search_query=&id_list=2512.17489&sortBy=relevance&sortOrder=descending&start=0&max_results=100)

Method: Cannot determine method as paper content is unavailable

Result: Cannot determine results as paper content is unavailable

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

Abstract: Failed to fetch summary for 2512.23365: Page request resulted in HTTP 429 (https://export.arxiv.org/api/query?search_query=&id_list=2512.23365&sortBy=relevance&sortOrder=descending&start=0&max_results=100)

Method: Cannot determine method as paper content could not be retrieved

Result: Cannot determine results as paper content could not be retrieved

Conclusion: Cannot draw conclusions as paper content could not be retrieved

Abstract: Failed to fetch summary for 2602.22545: Page request resulted in HTTP 429 (https://export.arxiv.org/api/query?search_query=&id_list=2602.22545&sortBy=relevance&sortOrder=descending&start=0&max_results=100)

Method: Proposes AnomalyVFM with: 1) A three-stage synthetic dataset generation scheme for diverse anomaly patterns, and 2) Parameter-efficient adaptation using low-rank feature adapters and confidence-weighted pixel loss to adapt VFMs for anomaly detection.

Result: With RADIO backbone, achieves 94.1% average image-level AUROC across 9 datasets, surpassing previous methods by 3.3 percentage points, demonstrating VFMs can outperform VLMs in zero-shot anomaly detection.

Conclusion: AnomalyVFM effectively bridges the performance gap between VFMs and VLMs for zero-shot anomaly detection through better synthetic data generation and adaptation strategies, establishing VFMs as competitive alternatives.

Abstract: Zero-shot anomaly detection aims to detect and localise abnormal regions in the image without access to any in-domain training images. While recent approaches leverage vision-language models (VLMs), such as CLIP, to transfer high-level concept knowledge, methods based on purely vision foundation models (VFMs), like DINOv2, have lagged behind in performance. We argue that this gap stems from two practical issues: (i) limited diversity in existing auxiliary anomaly detection datasets and (ii) overly shallow VFM adaptation strategies. To address both challenges, we propose AnomalyVFM, a general and effective framework that turns any pretrained VFM into a strong zero-shot anomaly detector. Our approach combines a robust three-stage synthetic dataset generation scheme with a parameter-efficient adaptation mechanism, utilising low-rank feature adapters and a confidence-weighted pixel loss. Together, these components enable modern VFMs to substantially outperform current state-of-the-art methods. More specifically, with RADIO as a backbone, AnomalyVFM achieves an average image-level AUROC of 94.1% across 9 diverse datasets, surpassing previous methods by significant 3.3 percentage points. Project Page: https://maticfuc.github.io/anomaly_vfm/

Method: Cannot determine method as paper content is unavailable

Result: Cannot determine results as paper content is unavailable

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

Abstract: Failed to fetch summary for 2602.22683: Page request resulted in HTTP 429 (https://export.arxiv.org/api/query?search_query=&id_list=2602.22683&sortBy=relevance&sortOrder=descending&start=0&max_results=100)

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

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

Conclusion: Unable to draw conclusions due to API rate limiting preventing access to paper content

Abstract: Failed to fetch summary for 2602.13669: Page request resulted in HTTP 429 (https://export.arxiv.org/api/query?search_query=&id_list=2602.13669&sortBy=relevance&sortOrder=descending&start=0&max_results=100)

Method: Unable to determine method due to failed paper fetch

Result: Unable to determine results due to failed paper fetch

Conclusion: Unable to determine conclusion due to failed paper fetch

Abstract: Failed to fetch summary for 2603.04385: Page request resulted in HTTP 429 (https://export.arxiv.org/api/query?search_query=&id_list=2603.04385&sortBy=relevance&sortOrder=descending&start=0&max_results=100)

Method: Unable to determine method due to failed paper fetch

Result: Unable to determine results due to failed paper fetch

Conclusion: Unable to determine conclusion due to failed paper fetch

Abstract: Failed to fetch summary for 2603.11633: Page request resulted in HTTP 429 (https://export.arxiv.org/api/query?search_query=&id_list=2603.11633&sortBy=relevance&sortOrder=descending&start=0&max_results=100)

Method: Unable to determine method due to fetch failure

Result: Unable to determine results due to fetch failure

Conclusion: Unable to determine conclusion due to fetch failure

Abstract: Failed to fetch summary for 2603.22364: Page request resulted in HTTP 429 (https://export.arxiv.org/api/query?search_query=&id_list=2603.22364&sortBy=relevance&sortOrder=descending&start=0&max_results=100)

Method: Cannot determine method without access to paper content

Result: Cannot determine results without access to paper content

Conclusion: Cannot draw conclusions without access to paper content

Abstract: Failed to fetch summary for 2603.15118: Page request resulted in HTTP 429 (https://export.arxiv.org/api/query?search_query=&id_list=2603.15118&sortBy=relevance&sortOrder=descending&start=0&max_results=100)

Method: Unable to determine method due to missing abstract

Result: Unable to determine results due to missing abstract

Conclusion: Unable to determine conclusion due to missing abstract

Abstract: Failed to fetch summary for 2603.16570: Page request resulted in HTTP 429 (https://export.arxiv.org/api/query?search_query=&id_list=2603.16570&sortBy=relevance&sortOrder=descending&start=0&max_results=100)

Method: Cannot determine method without access to paper content

Result: Cannot determine results without access to paper content

Conclusion: Cannot draw conclusions without access to paper content

Abstract: Failed to fetch summary for 2603.23286: Page request resulted in HTTP 429 (https://export.arxiv.org/api/query?search_query=&id_list=2603.23286&sortBy=relevance&sortOrder=descending&start=0&max_results=100)

Method: Unable to determine method due to paper fetch failure

Result: Unable to determine results due to paper fetch failure

Conclusion: Unable to determine conclusion due to paper fetch failure

Abstract: Failed to fetch summary for 2604.02073: Page request resulted in HTTP 429 (https://export.arxiv.org/api/query?search_query=&id_list=2604.02073&sortBy=relevance&sortOrder=descending&start=0&max_results=100)

Method: Unable to determine method due to access error

Result: Unable to determine results due to access error

Conclusion: Unable to determine conclusion due to access error

Abstract: Failed to fetch summary for 2604.03134: Page request resulted in HTTP 429 (https://export.arxiv.org/api/query?search_query=&id_list=2604.03134&sortBy=relevance&sortOrder=descending&start=0&max_results=100)

Method: Unable to determine method due to fetch failure

Result: Unable to determine results due to fetch failure

Conclusion: Unable to determine conclusion due to fetch failure

Abstract: Failed to fetch summary for 2604.03317: Page request resulted in HTTP 429 (https://export.arxiv.org/api/query?search_query=&id_list=2604.03317&sortBy=relevance&sortOrder=descending&start=0&max_results=100)

Method: Cannot determine method as paper content is unavailable due to technical limitations

Result: Cannot determine results as paper content is unavailable due to technical limitations

Conclusion: Cannot draw conclusions as paper content is unavailable due to technical limitations

Abstract: Failed to fetch summary for 2604.05933: Page request resulted in HTTP 429 (https://export.arxiv.org/api/query?search_query=&id_list=2604.05933&sortBy=relevance&sortOrder=descending&start=0&max_results=100)

Method: Cannot determine method without access to paper content

Result: Cannot determine results without access to paper content

Conclusion: Cannot determine conclusion without access to paper content

Abstract: Failed to fetch summary for 2604.06945: Page request resulted in HTTP 429 (https://export.arxiv.org/api/query?search_query=&id_list=2604.06945&sortBy=relevance&sortOrder=descending&start=0&max_results=100)

Method: Unable to determine method due to fetch failure

Result: Unable to determine results due to fetch failure

Conclusion: Unable to determine conclusion due to fetch failure

Abstract: Failed to fetch summary for 2604.06950: Page request resulted in HTTP 429 (https://export.arxiv.org/api/query?search_query=&id_list=2604.06950&sortBy=relevance&sortOrder=descending&start=0&max_results=100)

Method: Proposes AnchorSplat with anchor-aligned Gaussian representation guided by 3D geometric priors (sparse point clouds, voxels, or RGB-D point clouds). Uses a Gaussian Refiner to adjust intermediate Gaussians via few forward passes, enabling scene-level reconstruction directly in 3D space.

Result: Achieves state-of-the-art performance on ScanNet++ v2 NVS benchmark, outperforming previous methods with better view consistency and substantially fewer Gaussian primitives. Reduces required Gaussians while improving computational efficiency and reconstruction fidelity.

Conclusion: AnchorSplat provides an effective feed-forward 3DGS framework for scene reconstruction that decouples Gaussian representation from input images, enabling more efficient and geometry-aware 3D reconstruction with fewer primitives.

Abstract: Recent feed-forward Gaussian reconstruction models adopt a pixel-aligned formulation that maps each 2D pixel to a 3D Gaussian, entangling Gaussian representations tightly with the input images. In this paper, we propose AnchorSplat, a novel feed-forward 3DGS framework for scene-level reconstruction that represents the scene directly in 3D space. AnchorSplat introduces an anchor-aligned Gaussian representation guided by 3D geometric priors (e.g., sparse point clouds, voxels, or RGB-D point clouds), enabling a more geometry-aware renderable 3D Gaussians that is independent of image resolution and number of views. This design substantially reduces the number of required Gaussians, improving computational efficiency while enhancing reconstruction fidelity. Beyond the anchor-aligned design, we utilize a Gaussian Refiner to adjust the intermediate Gaussiansy via merely a few forward passes. Experiments on the ScanNet++ v2 NVS benchmark demonstrate the SOTA performance, outperforming previous methods with more view-consistent and substantially fewer Gaussian primitives.

Method: Unable to determine method due to fetch failure.

Result: Unable to determine results due to fetch failure.

Conclusion: Unable to draw conclusions due to fetch failure.

Abstract: Failed to fetch summary for 2604.07092: Page request resulted in HTTP 429 (https://export.arxiv.org/api/query?search_query=&id_list=2604.07092&sortBy=relevance&sortOrder=descending&start=0&max_results=100)

Method: Cannot determine method as paper content is unavailable

Result: Cannot determine results as paper content is unavailable

Conclusion: Cannot determine conclusion as paper content is unavailable

Abstract: Failed to fetch summary for 2509.01878: Page request resulted in HTTP 429 (https://export.arxiv.org/api/query?search_query=&id_list=2509.01878&sortBy=relevance&sortOrder=descending&start=0&max_results=100)

Method: Unable to determine method due to technical error in fetching paper content

Result: Unable to determine results due to technical error in fetching paper content

Conclusion: Unable to determine conclusion due to technical error in fetching paper content

Abstract: Failed to fetch summary for 2604.07146: Page request resulted in HTTP 429 (https://export.arxiv.org/api/query?search_query=&id_list=2604.07146&sortBy=relevance&sortOrder=descending&start=0&max_results=100)

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

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

Conclusion: Unable to draw conclusions about the paper due to technical limitations in accessing the content.

Abstract: Failed to fetch summary for 2604.07230: Page request resulted in HTTP 429 (https://export.arxiv.org/api/query?search_query=&id_list=2604.07230&sortBy=relevance&sortOrder=descending&start=0&max_results=100)

Method: Cannot analyze method as paper content is unavailable

Result: No results available due to technical error accessing paper

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

Abstract: Failed to fetch summary for 2604.07273: Page request resulted in HTTP 429 (https://export.arxiv.org/api/query?search_query=&id_list=2604.07273&sortBy=relevance&sortOrder=descending&start=0&max_results=100)

Method: Cannot determine method due to missing abstract

Result: Cannot determine results due to missing abstract

Conclusion: Cannot determine conclusion due to missing abstract

Abstract: Failed to fetch summary for 2503.09640: Page request resulted in HTTP 429 (https://export.arxiv.org/api/query?search_query=&id_list=2503.09640&sortBy=relevance&sortOrder=descending&start=0&max_results=100)

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

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

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

Abstract: Failed to fetch summary for 2504.13378: Page request resulted in HTTP 429 (https://export.arxiv.org/api/query?search_query=&id_list=2504.13378&sortBy=relevance&sortOrder=descending&start=0&max_results=100)

Method: Cannot determine method as abstract is unavailable

Result: Cannot determine results as abstract is unavailable

Conclusion: Cannot draw conclusions without access to paper content

Abstract: Failed to fetch summary for 2512.18662: Page request resulted in HTTP 429 (https://export.arxiv.org/api/query?search_query=&id_list=2512.18662&sortBy=relevance&sortOrder=descending&start=0&max_results=100)

Method: Unable to determine method due to fetch failure

Result: Unable to determine results due to fetch failure

Conclusion: Unable to determine conclusion due to fetch failure

Abstract: Failed to fetch summary for 2602.20231: Page request resulted in HTTP 429 (https://export.arxiv.org/api/query?search_query=&id_list=2602.20231&sortBy=relevance&sortOrder=descending&start=0&max_results=100)

Method: Cannot determine method as paper content is unavailable

Result: Cannot determine results as paper content is unavailable

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

Abstract: Failed to fetch summary for 2604.05351: Page request resulted in HTTP 429 (https://export.arxiv.org/api/query?search_query=&id_list=2604.05351&sortBy=relevance&sortOrder=descending&start=0&max_results=100)

Method: Cannot analyze method without access to paper content

Result: No results available due to API rate limiting

Conclusion: Cannot draw conclusions without access to paper content

Abstract: Failed to fetch summary for 2604.06036: Page request resulted in HTTP 429 (https://export.arxiv.org/api/query?search_query=&id_list=2604.06036&sortBy=relevance&sortOrder=descending&start=0&max_results=100)

Method: Human-in-the-loop LLM methodology for automatic classification and summarization of research descriptions at scale, applied to 58,746 NIH-funded biomedical research projects from 2025.

Result: AI constitutes 15.9% of NIH portfolio with 13.4% funding premium; 79% of AI projects remain in research/development stages; health disparities research severely underrepresented at 5.7% of AI-funded work.

Conclusion: Establishes framework for evidence-based policy interventions to align NIH AI portfolio with health equity goals and strategic research priorities.

Abstract: Understanding the landscape of artificial intelligence (AI) and machine learning (ML) adoption across the National Institutes of Health (NIH) portfolio is critical for research funding strategy, institutional planning, and health policy. The advent of large language models (LLMs) has fundamentally transformed research landscape analysis, enabling researchers to perform large-scale semantic extraction from thousands of unstructured research documents. In this paper, we illustrate a human-in-the-loop research methodology for LLMs to automatically classify and summarize research descriptions at scale. Using our methodology, we present a comprehensive analysis of 58,746 NIH-funded biomedical research projects from 2025. We show that: (1) AI constitutes 15.9% of the NIH portfolio with a 13.4% funding premium, concentrated in discovery, prediction, and data integration across disease domains; (2) a critical research-to-deployment gap exists, with 79% of AI projects remaining in research/development stages while only 14.7% engage in clinical deployment or implementation; and (3) health disparities research is severely underrepresented at just 5.7% of AI-funded work despite its importance to NIH’s equity mission. These findings establish a framework for evidence-based policy interventions to align the NIH AI portfolio with health equity goals and strategic research priorities.

Method: Used LLM-based coding agents (ChatGPT 5.2 and Claude Opus 4.6) with a “sorry-first” declarative proof workflow combined with bulk use of Isabelle’s sledgehammer tool over 24 active days.

Result: Successfully formalized all 39 sections of Munkres’ Topology covering topological spaces through dimension theory, producing 85,000+ lines of Isabelle/HOL code with all 806 results fully proved (zero sorry’s).

Conclusion: LLM-assisted formalization of standard mathematical textbooks in Isabelle/HOL is feasible, cheap, and fast, though some human supervision remains useful.

Abstract: We describe an experiment in LLM-assisted autoformalization that produced over 85,000 lines of Isabelle/HOL code covering all 39 sections of Munkres’ Topology (general topology, Chapters 2–8), from topological spaces through dimension theory. The LLM-based coding agents (initially ChatGPT 5.2 and then Claude Opus 4.6) used 24 active days for that. The formalization is complete: all 806 formal results are fully proved with zero sorry’s. Proved results include the Tychonoff theorem, the Baire category theorem, the Nagata–Smirnov and Smirnov metrization theorems, the Stone–Čech compactification, Ascoli’s theorem, the space-filling curve, and others. The methodology is based on a “sorry-first” declarative proof workflow combined with bulk use of sledgehammer - two of Isabelle major strengths. This leads to relatively fast autoformalization progress. We analyze the resulting formalization in detail, analyze the human–LLM interaction patterns from the session log, and briefly compare with related autoformalization efforts in Megalodon, HOL Light, and Naproche. The results indicate that LLM-assisted formalization of standard mathematical textbooks in Isabelle/HOL is quite feasible, cheap and fast, even if some human supervision is useful.

Method: Four-phase protocol (Investigation, Corroboration, Falsification, Verdict) with ReAct-style controller; uses StyleComparator for formal analysis and ConceptRetriever for iconographic grounding; assembles verifiable evidence chains from images/biographies with adversarial falsification.

Result: Achieves 83.7% positive-class F1, 0.666 MCC, and 0.910 ROC-AUC on WikiArt Influence Benchmark-100; maintains performance when explicit influence phrases are masked, showing robustness.

Conclusion: Implicit influence analysis benefits from historically grounded adjudication with domain-constrained falsification rather than pattern matching alone; multimodal perception coupled with verification improves attribution accuracy.

Abstract: Implicit artistic influence, although visually plausible, is often undocumented and thus poses a historically constrained attribution problem: resemblance is necessary but not sufficient evidence. Most prior systems reduce influence discovery to embedding similarity or label-driven graph completion, while recent multimodal large language models (LLMs) remain vulnerable to temporal inconsistency and unverified attributions. This paper introduces M-ArtAgent, an evidence-based multimodal agent that reframes implicit influence discovery as probabilistic adjudication. It follows a four-phase protocol consisting of Investigation, Corroboration, Falsification, and Verdict governed by a Reasoning and Acting (ReAct)-style controller that assembles verifiable evidence chains from images and biographies, enforces art-historical axioms, and subjects each hypothesis to adversarial falsification via a prompt-isolated critic. Two theory-grounded operators, StyleComparator for Wolfflin formal analysis and ConceptRetriever for ICONCLASS-based iconographic grounding, ensure that intermediate claims are formally auditable. On the balanced WikiArt Influence Benchmark-100 (WIB-100) of 100 artists and 2,000 directed pairs, M-ArtAgent achieves 83.7% positive-class F1, 0.666 Matthews correlation coefficient (MCC), and 0.910 area under the receiver operating characteristic curve (ROC-AUC), with leakage-control and robustness checks confirming that the gains persist when explicit influence phrases are masked. By coupling multimodal perception with domain-constrained falsification, M-ArtAgent demonstrates that implicit influence analysis benefits from historically grounded adjudication rather than pattern matching alone.

Method: Proposes ConsistRM framework with two key components: 1) Consistency-Aware Answer Reward that produces reliable pseudo-labels with temporal consistency for stable optimization, and 2) Consistency-Aware Critique Reward that assesses semantic consistency across multiple critiques and allocates fine-grained, differentiated rewards.

Result: Experiments on five benchmark datasets across four base models show ConsistRM outperforms vanilla Reinforcement Fine-Tuning (RFT) by an average of 1.5%. The framework enhances output consistency and mitigates position bias caused by input order.

Conclusion: ConsistRM demonstrates that consistency-aware rewards effectively improve generative reward models, enabling stable self-training without human annotations while addressing scalability and stability challenges in GRM training.

Abstract: Generative reward models (GRMs) have emerged as a promising approach for aligning Large Language Models (LLMs) with human preferences by offering greater representational capacity and flexibility than traditional scalar reward models. However, GRMs face two major challenges: reliance on costly human-annotated data restricts scalability, and self-training approaches often suffer from instability and vulnerability to reward hacking. To address these issues, we propose ConsistRM, a self-training framework that enables effective and stable GRM training without human annotations. ConsistRM incorporates the Consistency-Aware Answer Reward, which produces reliable pseudo-labels with temporal consistency, thereby providing more stable model optimization. Moreover, the Consistency-Aware Critique Reward is introduced to assess semantic consistency across multiple critiques and allocates fine-grained and differentiated rewards. Experiments on five benchmark datasets across four base models demonstrate that ConsistRM outperforms vanilla Reinforcement Fine-Tuning (RFT) by an average of 1.5%. Further analysis shows that ConsistRM enhances output consistency and mitigates position bias caused by input order, highlighting the effectiveness of consistency-aware rewards in improving GRMs.

Method: CLEAR uses a reflection agent to perform contrastive analysis over past execution trajectories and summarize useful context. These summaries train a context augmentation model (CAM) via supervised fine-tuning, then further optimized with reinforcement learning using execution agent feedback as reward signal.

Result: Improves task completion rate from 72.62% to 81.15% on AppWorld test set and averaged reward from 0.68 to 0.74 on WebShop subset compared to baseline agents.

Conclusion: Generative context augmentation via CLEAR outperforms retrieval-based approaches by producing better tailored task-specific knowledge, reducing reasoning burden on LLMs.

Abstract: Large language model agents rely on effective model context to obtain task-relevant information for decision-making. Many existing context engineering approaches primarily rely on the context generated from the past experience and retrieval mechanisms that reuse these context. However, retrieved context from past tasks must be adapted by the execution agent to fit new situations, placing additional reasoning burden on the underlying LLM. To address this limitation, we propose a generative context augmentation framework using Contrastive Learning of Experience via Agentic Reflection (CLEAR). CLEAR first employs a reflection agent to perform contrastive analysis over past execution trajectories and summarize useful context for each observed task. These summaries are then used as supervised fine-tuning data to train a context augmentation model (CAM). Then we further optimize CAM using reinforcement learning, where the reward signal is obtained by running the task execution agent. By learning to generate task-specific knowledge rather than retrieve knowledge from the past, CAM produces context that is better tailored to the current task. We conduct comprehensive evaluations on the AppWorld and WebShop benchmarks. Experimental results show that CLEAR consistently outperforms strong baselines. It improves task completion rate from 72.62% to 81.15% on AppWorld test set and averaged reward from 0.68 to 0.74 on a subset of WebShop, compared with baseline agent. Our code is publicly available at https://github.com/awslabs/CLEAR.

Method: Proposes ReflectRM, a GRM trained under a unified generative framework for joint modeling of response preference and analysis preference. During inference, it uses self-reflection to identify the most reliable analysis, from which final preference predictions are derived.

Result: Experiments across four benchmarks show ReflectRM consistently improves performance with average accuracy gain of +3.7 on Qwen3-4B. It substantially mitigates positional bias (+10.2 improvement vs leading GRMs) and establishes itself as a more stable evaluator.

Conclusion: ReflectRM demonstrates that joint modeling of response and analysis preferences with self-reflection enhances reward modeling, improves generalization, reduces biases, and provides better interpretability for LLM alignment.

Abstract: Reward Models (RMs) are critical components in the Reinforcement Learning from Human Feedback (RLHF) pipeline, directly determining the alignment quality of Large Language Models (LLMs). Recently, Generative Reward Models (GRMs) have emerged as a superior paradigm, offering higher interpretability and stronger generalization than traditional scalar RMs. However, existing methods for GRMs focus primarily on outcome-level supervision, neglecting analytical process quality, which constrains their potential. To address this, we propose ReflectRM, a novel GRM that leverages self-reflection to assess analytical quality and enhance preference modeling. ReflectRM is trained under a unified generative framework for joint modeling of response preference and analysis preference. During inference, we use its self-reflection capability to identify the most reliable analysis, from which the final preference prediction is derived. Experiments across four benchmarks show that ReflectRM consistently improves performance, achieving an average accuracy gain of +3.7 on Qwen3-4B. Further experiments confirm that response preference and analysis preference are mutually reinforcing. Notably, ReflectRM substantially mitigates positional bias, yielding +10.2 improvement compared with leading GRMs and establishing itself as a more stable evaluator.

Method: Multi-modal AI agents function as computational chemists, medicinal chemists, and patent agents, using a 246M-parameter GNN trained on 800M molecules to generate novel chemical matter directly on molecular graphs. Agents executed campaigns for oncology targets BCL6 and EZH2 with three strategy tiers.

Result: Generated 2,355-2,876 novel molecules per target; 91.9% of Murcko scaffolds absent from ChEMBL; Tanimoto distances 0.56-0.69 to nearest known actives; binding affinity predictions achieved Spearman correlations -0.53 to -0.64 and ROC AUC 0.88-0.93.

Conclusion: Semi-autonomous agent systems with graph-native generative tools and physics-informed scoring provide a foundation for modern small molecule discovery operating systems, enabling scaled, rapid, adaptive inverse design.

Abstract: We introduce a semi-autonomous discovery system in which multi-modal AI agents function as a multi-disciplinary discovery team, acting as computational chemists, medicinal chemists, and patent agents, writing and executing analysis code, visually evaluating molecular candidates, assessing patentability, and adapting generation strategy from empirical screening feedback, while r1, a 246M-parameter Graph Neural Network (GNN) trained on 800M molecules, generates novel chemical matter directly on molecular graphs. Agents executed two campaigns in oncology (BCL6, EZH2), formulating medicinal chemistry hypotheses across three strategy tiers and generating libraries of 2,355-2,876 novel molecules per target. Across both targets, 91.9% of generated Murcko scaffolds are absent from ChEMBL for their respective targets, with Tanimoto distances of 0.56-0.69 to the nearest known active, confirming that the engine produces structurally distinct chemical matter rather than recapitulating known compounds. Binding affinity predictions using Boltz-2 were calibrated against ChEMBL experimental data, achieving Spearman correlations of -0.53 to -0.64 and ROC AUC values of 0.88 to 0.93. These results demonstrate that semi-autonomous agent systems, equipped with graph-native generative tools and physics-informed scoring, provide a foundation for a modern operating system for small molecule discovery. We show that Rhizome OS-1 enables a new paradigm for early-stage drug discovery by supporting scaled, rapid, and adaptive inverse design.

Method: Introduces Conformal Social Choice: verbalized probability distributions from heterogeneous agents are aggregated via linear opinion pool and calibrated with split conformal prediction, yielding prediction sets with marginal coverage guarantees. A hierarchical action policy maps singleton sets to autonomous action and larger sets to human escalation.

Result: On eight MMLU-Pro domains with three agents (Claude Haiku, DeepSeek-R1, Qwen-3 32B), coverage stays within 1-2 points of target. 81.9% of wrong-consensus cases are intercepted at α=0.05. Remaining conformal singletons reach 90.0-96.8% accuracy (up to 22.1pp above consensus stopping).

Conclusion: Conformal Social Choice makes multi-agent debate failures actionable by intercepting confidently wrong cases, though this safety comes at the cost of automation. The operating point is user-adjustable via α parameter.

Abstract: Multi-agent debate improves LLM reasoning, yet agreement among agents is not evidence of correctness. When agents converge on a wrong answer through social reinforcement, consensus-based stopping commits that error to an automated action with no recourse. We introduce Conformal Social Choice, a post-hoc decision layer that converts debate outputs into calibrated act-versus-escalate decisions. Verbalized probability distributions from heterogeneous agents are aggregated via a linear opinion pool and calibrated with split conformal prediction, yielding prediction sets with a marginal coverage guarantee: the correct answer is included with probability ${\geq},1{-}α$, without assumptions on individual model calibration. A hierarchical action policy maps singleton sets to autonomous action and larger sets to human escalation. On eight MMLU-Pro domains with three agents (Claude Haiku, DeepSeek-R1, Qwen-3 32B), coverage stays within 1–2 points of the target. The key finding is not that debate becomes more accurate, but that the conformal layer makes its failures actionable: 81.9% of wrong-consensus cases are intercepted at $α{=}0.05$. Because the layer refuses to act on cases where debate is confidently wrong, the remaining conformal singletons reach 90.0–96.8% accuracy (up to 22.1pp above consensus stopping) – a selection effect, not a reasoning improvement. This safety comes at the cost of automation, but the operating point is user-adjustable via $α$.

Method: Conducted an experiment with 30 human participants to study the effects of urgency in AI prompts on user trust, confidence, and self-efficacy.

Result: Urgency in AI prompts doesn’t affect trust in AI but negatively impacts human users’ self-confidence and self-efficacy. Users feel more confident when eased into human-AI setups rather than exposed abruptly.

Conclusion: Urgency in AI interactions can be detrimental to human users’ confidence and efficacy, potentially leading to performance issues and unsustainable AI systems. Gradual exposure to AI systems is recommended.

Abstract: Studies show that interactions with an AI system fosters trust in human users towards AI. An often overlooked element of such interaction dynamics is the (sense of) urgency when the human user is prompted by an AI agent, e.g., for advice or guidance. In this paper, we show that although the presence of urgency in human-AI interactions does not affect the trust in AI, it may be detrimental to the human user’s self-confidence and self-efficacy. In the long run, the loss of confidence may lead to performance loss, suboptimal decisions, human errors, and ultimately, unsustainable AI systems. Our evidence comes from an experiment with 30 human participants. Our results indicate that users may feel more confident in their work when they are eased into the human-AI setup rather than exposed to it without preparation. We elaborate on the implications of this finding for software engineers and decision-makers.

Method: Develops a supervised multi-agent governance framework integrating legal rules, technical protocols, and institutional oversight with ex ante and ex post coordination mechanisms to correct agentic market failures.

Result: Proposes that AI can serve as both a disruption source and governance tool to restore market-based ordering in creative industries through properly designed agentic copyright systems.

Conclusion: Agentic copyright offers a path toward scalable, fair, and legally meaningful copyright markets in the AI age by embedding normative constraints into multi-agent architectures to align behavior with copyright law values.

Abstract: This paper examines how the rapid deployment of multi-agentic AI systems is reshaping the foundations of copyright law and creative markets. It argues that existing copyright frameworks are ill-equipped to govern AI agent-mediated interactions that occur at scale, speed, and with limited human oversight. The paper introduces the concept of agentic copyright, a model in which AI agents act on behalf of creators and users to negotiate access, attribution, and compensation for copyrighted works. While multi-agent ecosystems promise efficiency gains and reduced transaction costs, they also generate novel market failures, including miscoordination, conflict, and collusion among autonomous agents. To address these market failures, the paper develops a supervised multi-agent governance framework that integrates legal rules and principles, technical protocols, and institutional oversight. This framework emphasizes ex ante and ex post coordination mechanisms capable of correcting agentic market failures before they crystallize into systemic harm. By embedding normative constraints and monitoring functions into multi-agent architectures, supervised governance aims to align agent behavior with the underlying values of copyright law. The paper concludes that AI should be understood not only as a source of disruption, but also as a governance tool capable of restoring market-based ordering in creative industries. Properly designed, agentic copyright offers a path toward scalable, fair, and legally meaningful copyright markets in the age of AI.

Method: AEI links offline phase (engineering data processing and workflow-memory construction) with online phase (workflow-state estimation, retrieval, and decision support). Uses control-theoretic interpretation where engineering objectives act as reference signals, agents as workflow controllers, and toolchains provide feedback.

Result: Demonstrated through automotive use cases in suspension design, reinforcement learning tuning, multimodal engineering knowledge reuse, aerodynamic exploration, and MBSE, showing diverse workflows can be expressed within a common formulation.

Conclusion: Positions engineering AI as a problem of process-level intelligence and outlines a practical roadmap for future empirical validation in industrial settings.

Abstract: Engineering workflows such as design optimization, simulation-based diagnosis, control tuning, and model-based systems engineering (MBSE) are iterative, constraint-driven, and shaped by prior decisions. Yet many AI methods still treat these activities as isolated tasks rather than as parts of a broader workflow. This paper presents Agentic Engineering Intelligence (AEI), an industrial vision framework that models engineering workflows as constrained, history-aware sequential decision processes in which AI agents support engineer-supervised interventions over engineering toolchains. AEI links an offline phase for engineering data processing and workflow-memory construction with an online phase for workflow-state estimation, retrieval, and decision support. A control-theoretic interpretation is also possible, in which engineering objectives act as reference signals, agents act as workflow controllers, and toolchains provide feedback for intervention selection. Representative automotive use cases in suspension design, reinforcement learning tuning, multimodal engineering knowledge reuse, aerodynamic exploration, and MBSE show how diverse workflows can be expressed within a common formulation. Overall, the paper positions engineering AI as a problem of process-level intelligence and outlines a practical roadmap for future empirical validation in industrial settings.

Method: Proposes Dual-Loop Control Framework (DLCF) with three core entities: physical system, digital twin, and policy reservoir of diverse DRL agents. Uses dual-loop mechanism for real-time data acquisition, data assimilation, DRL policy training, pre-evaluation, and expert verification.

Result: Implemented DCVerse platform and validated on real-world data center cooling system. Achieves up to 4.09% energy savings over conventional control strategies without violating SLA requirements. Improves policy interpretability and supports trustworthy DRL deployment.

Conclusion: DLCF provides foundation for reliable AI-based control in data centers, enabling safe DRL deployment with improved sample efficiency, generalization, safety, and optimality. Points toward future extensions for holistic system-wide optimization.

Abstract: The growing scale and complexity of modern data centers present major challenges in balancing energy efficiency with outage risk. Although Deep Reinforcement Learning (DRL) shows strong potential for intelligent control, its deployment in mission-critical systems is limited by data scarcity and the lack of real-time pre-evaluation mechanisms. This paper introduces the Dual-Loop Control Framework (DLCF), a digital twin-based architecture designed to overcome these challenges. The framework comprises three core entities: the physical system, a digital twin, and a policy reservoir of diverse DRL agents. These components interact through a dual-loop mechanism involving real-time data acquisition, data assimilation, DRL policy training, pre-evaluation, and expert verification. Theoretical analysis shows how DLCF can improve sample efficiency, generalization, safety, and optimality. Leveraging DLCF, we implemented the DCVerse platform and validated it through case studies on a real-world data center cooling system. The evaluation shows that our approach achieves up to 4.09% energy savings over conventional control strategies without violating SLA requirements. Additionally, the framework improves policy interpretability and supports more trustworthy DRL deployment. This work provides a foundation for reliable AI-based control in data centers and points toward future extensions for holistic, system-wide optimization.

Method: Multimodal screening agent + heterogeneous jury of 5 evaluative agents + meta-agent for verdict synthesis with Source Attribution Protocol to distinguish textbook narrative from quoted sources

Result: 83.3% of 270 Romanian textbook excerpts classified as pedagogically acceptable (mean severity 2.9/7 vs 5.4/7 baseline); Independent Deliberation preferred in 64.8% of human evaluations; cost ~$2 per textbook

Conclusion: Agentic evaluation architectures are economically viable decision-support tools for educational governance that mitigate over-penalization through deliberation

Abstract: History textbooks often contain implicit biases, nationalist framing, and selective omissions that are difficult to audit at scale. We propose an agentic evaluation architecture comprising a multimodal screening agent, a heterogeneous jury of five evaluative agents, and a meta-agent for verdict synthesis and human escalation. A central contribution is a Source Attribution Protocol that distinguishes textbook narrative from quoted historical sources, preventing the misattribution that causes systematic false positives in single-model evaluators. In an empirical study on Romanian upper-secondary history textbooks, 83.3% of 270 screened excerpts were classified as pedagogically acceptable (mean severity 2.9/7), versus 5.4/7 under a zero-shot baseline, demonstrating that agentic deliberation mitigates over-penalization. In a blind human evaluation (18 evaluators, 54 comparisons), the Independent Deliberation configuration was preferred in 64.8% of cases over both a heuristic variant and the zero-shot baseline. At approximately $2 per textbook, these results position agentic evaluation architectures as economically viable decision-support tools for educational governance.

Method: Prompt-driven hierarchical workflow using LLM to extract shot-level experimental records with three-level priority strategy: T1 (direct extraction from text/tables), T2 (physics-based derivation using governing relations), T3 (digitization from figures). Extracted values are normalized, tagged by priority, and validated with physics-based consistency checks.

Result: Achieved 94.69% overall weighted accuracy on 30 articles with 11,967 data points. Priority-wise accuracies: 94.93% (T1), 92.04% (T2), 83.49% (T3). Cross-model testing showed strong agreement for text/table and equation-derived fields, with lower agreement for figure-based extraction. API implementation demonstrated scalability.

Conclusion: The workflow provides a practical approach for converting unstructured technical literature into traceable, analysis-ready datasets without task-specific fine-tuning, enabling scalable database construction in materials science.

Abstract: Scientific data are widely dispersed across research articles and are often reported inconsistently across text, tables, and figures, making manual data extraction and aggregation slow and error-prone. We present a prompt-driven, hierarchical workflow that uses a large language model (LLM) to automatically extract and reconstruct structured, shot-level shock-physics experimental records by integrating information distributed across text, tables, figures, and physics-based derivations from full-text published research articles, using alloy spall strength as a representative case study. The pipeline targeted 37 experimentally relevant fields per shot and applied a three-level priority strategy: (T1) direct extraction from text/tables, (T2) physics-based derivation using verified governing relations, and (T3) digitization from figures when necessary. Extracted values were normalized to canonical units, tagged by priority for traceability, and validated with physics-based consistency and plausibility checks. Evaluated on a benchmark of 30 published research articles comprising 11,967 evaluated data points, the workflow achieved high overall accuracy, with priority-wise accuracies of 94.93% (T1), 92.04% (T2), and 83.49% (T3), and an overall weighted accuracy of 94.69%. Cross-model testing further indicated strong agreement for text/table and equation-derived fields, with lower agreement for figure-based extraction. Implementation through an API interface demonstrated the scalability of the approach, achieving consistent extraction performance and, in a subset of test cases, matching or exceeding chat-based accuracy. This workflow demonstrates a practical approach for converting unstructured technical literature into traceable, analysis-ready datasets without task-specific fine-tuning, enabling scalable database construction in materials science.

Method: Proposes DD-MM-PAS (Demand Detection, Memory Modeling, Proactive Agent System) paradigm, instantiated as Pask with streaming IntentFlow model for demand detection, hybrid memory system (workspace, user, global), and PAS infrastructure framework forming a closed loop. Also introduces LatentNeeds-Bench benchmark from user-consented data.

Result: IntentFlow matches leading Gemini3-Flash models under latency constraints while identifying deeper user intent. The system demonstrates effectiveness in real-world proactive agent scenarios.

Conclusion: Pask provides a practical framework for streaming proactive AI agents that can operate effectively in real-world settings with latency constraints, offering a solution to the challenge of inferring latent user needs from ongoing context.

Abstract: Proactivity is a core expectation for AGI. Prior work remains largely confined to laboratory settings, leaving a clear gap in real-world proactive agent: depth, complexity, ambiguity, precision and real-time constraints. We study this setting, where useful intervention requires inferring latent needs from ongoing context and grounding actions in evolving user memory under latency and long-horizon constraints. We first propose DD-MM-PAS (Demand Detection, Memory Modeling, Proactive Agent System) as a general paradigm for streaming proactive AI agent. We instantiate this paradigm in Pask, with streaming IntentFlow model for DD, a hybrid memory (workspace, user, global) for long-term MM, PAS infra framework and introduce how these components form a closed loop. We also introduce LatentNeeds-Bench, a real-world benchmark built from user-consented data and refined through thousands of rounds of human editing. Experiments show that IntentFlow matches leading Gemini3-Flash models under latency constraints, while identifying deeper user intent.

Method: Constructed an adversarial dataset of expert-authored mathematics problems and analyzed two structural length variables: prompt length and solution length. Examined how these variables relate to model performance across different models, including secondary exploratory analysis of cross-model disagreement using difficulty-adjusted normalized analysis.

Result: Both prompt length and solution length correlate positively with increased model failure across models. Under difficulty-adjusted normalized analysis, both variables retain weak negative associations with realized model separation, slightly stronger for prompt length. Main robust finding is that structural length is linked to empirical difficulty in this dataset.

Conclusion: Structural properties of mathematical problems, particularly prompt and solution lengths, significantly influence LLM performance, with longer prompts and solutions correlating with increased model failure rates.

Abstract: Mathematical benchmarks consisting of a range of mathematics problems are widely used to evaluate the reasoning abilities of large language models, yet little is known about how their structural properties influence model behaviour. In this work, we investigate two structural length variables, prompt length and solution length, and analyse how they relate to model performance on a newly constructed adversarial dataset of expert-authored mathematics problems. We find that both prompt and solution lengths correlate positively with increased model failure across models. We also include a secondary, exploratory analysis of cross-model disagreement. Under a difficulty-adjusted normalised analysis, both variables retain weak negative associations with realised model separation, slightly stronger for prompt length. Overall, our main robust finding is that structural length is linked to empirical difficulty in this dataset.

Method: Introduces Spatial Trajectory Graph (STG), a neuro-symbolic paradigm with verify-before-commit discipline that transforms open-ended LLM planning into verifiable graph optimization problem. Implements IoT-Brain system based on STG and constructs TopoSense-Bench benchmark

Result: IoT-Brain boosts task success rate by 37.6% over strongest search-intensive methods, runs nearly 2× faster, uses 6.6× fewer prompt tokens. In real-world deployment, approaches reliability upper bound while reducing 4.1× network bandwidth

Conclusion: Provides foundational framework for LLMs to interact with physical world with unprecedented reliability and efficiency by bridging semantic-physical gap through neuro-symbolic STG approach

Abstract: Intelligent systems powered by large-scale sensor networks are shifting from predefined monitoring to intent-driven operation, revealing a critical Semantic-to-Physical Mapping Gap. While large language models (LLMs) excel at semantic understanding, existing perception-centric pipelines operate retrospectively, overlooking the fundamental decision of what to sense and when. We formalize this proactive decision as Semantic-Spatial Sensor Scheduling (S3) and demonstrate that direct LLM planning is unreliable due to inherent gaps in representation, reasoning, and optimization. To bridge these gaps, we introduce the Spatial Trajectory Graph (STG), a neuro-symbolic paradigm governed by a verify-before-commit discipline that transforms open-ended planning into a verifiable graph optimization problem. Based on STG, we implement IoT-Brain, a concrete system embodiment, and construct TopoSense-Bench, a campus-scale benchmark with 5,250 natural-language queries across 2,510 cameras. Evaluations show that IoT-Brain boosts task success rate by 37.6% over the strongest search-intensive methods while running nearly 2 times faster and using 6.6 times fewer prompt tokens. In real-world deployment, it approaches the reliability upper bound while reducing 4.1 times network bandwidth, providing a foundational framework for LLMs to interact with the physical world with unprecedented reliability and efficiency.

Method: Introduces reasoning graphs that store per-evidence chain-of-thought as structured edges connected to evidence items, enabling evidence-centric feedback via backward traversal. Also introduces retrieval graphs for pipeline planning. Together they form a self-improving feedback loop without model retraining.

Result: The system shows improved accuracy and collapsed variance over successive runs on multi-hop question answering benchmarks, with all gains coming from context engineering via graph traversal rather than model retraining.

Conclusion: Reasoning graphs provide a novel evidence-centric feedback mechanism that enables language model agents to learn from prior reasoning, improving performance and reducing variance through persistent structured memory of evidence evaluations.

Abstract: Language model agents reason from scratch on every query: each time an agent retrieves evidence and deliberates, the chain of thought is discarded and the next similar query starts with no prior insight. This produces lower accuracy and high variance, as the same type of query can succeed or fail unpredictably. We introduce reasoning graphs, a graph structure that persists an agent’s per-evidence chain of thought as structured edges connected to the evidence items they evaluate. Unlike prior memory mechanisms that store distilled strategies as flat records indexed by query similarity or appended by recency, reasoning graphs enable evidence-centric feedback: given a new candidate set, the system traverses all incoming evaluation edges for each evidence item across all prior runs, surfacing how that specific item has been judged before. This backward traversal from evidence inward is a structurally different capability from query-similarity retrieval, because the feedback is tied to the specific evidence the agent is currently examining, not to the query. We further introduce retrieval graphs, a complementary structure that feeds a pipeline planner to tighten the candidate funnel over successive runs. Together, both graphs form a self-improving feedback loop: accuracy rises and variance collapses over successive runs, with every decision fully traceable through the graph. This improvement requires no retraining; the base model remains frozen and all gains come from context engineering via graph traversal. We formalize the graph structure, traversal algorithms, and feedback mechanisms, and describe a sequential cluster evaluation protocol for measuring accuracy convergence and variance collapse on multi-hop question answering benchmarks.

Method: PRIME distills multi-turn interaction trajectories into structured human-readable experiences organized across three semantic zones: successful strategies, failure patterns, and user preferences. These experiences evolve through meta-level operations and guide future agent behavior via retrieval-augmented generation.

Result: Experiments across diverse user-centric environments show PRIME achieves competitive performance with gradient-based methods while offering cost-efficiency and interpretability.

Conclusion: PRIME presents a practical paradigm for building proactive, collaborative agents that learn from Human-AI interaction without the computational burden of gradient-based training.

Abstract: The development of autonomous tool-use agents for complex, long-horizon tasks in collaboration with human users has become the frontier of agentic research. During multi-turn Human-AI interactions, the dynamic and uncertain nature of user demands poses a significant challenge; agents must not only invoke tools but also iteratively refine their understanding of user intent through effective communication. While recent advances in reinforcement learning offer a path to more capable tool-use agents, existing approaches require expensive training costs and struggle with turn-level credit assignment across extended interaction horizons. To this end, we introduce PRIME (Proactive Reasoning via Iterative Memory Evolution), a gradient-free learning framework that enables continuous agent evolvement through explicit experience accumulation rather than expensive parameter optimization. PRIME distills multi-turn interaction trajectories into structured, human-readable experiences organized across three semantic zones: successful strategies, failure patterns, and user preferences. These experiences evolve through meta-level operations and guide future agent behavior via retrieval-augmented generation. Our experiments across several diverse user-centric environments demonstrate that PRIME achieves competitive performance with gradient-based methods while offering cost-efficiency and interpretability. Together, PRIME presents a practical paradigm for building proactive, collaborative agents that learn from Human-AI interaction without the computational burden of gradient-based training.

Method: TrACE measures inter-rollout action agreement at each step: samples candidate actions, checks consistency. High agreement → commit immediately; low agreement → sample additional rollouts up to a cap before committing to plurality action. No learned components, verifiers, or human labels needed.

Result: TrACE-4 matches self-consistency with 4 samples accuracy using 33% fewer calls on GSM8K and 39% fewer on MiniHouse. TrACE-8 matches SC-8 accuracy with 55% fewer calls on GSM8K and 65% fewer on MiniHouse. Inter-rollout agreement reliably signals step-level success.

Conclusion: TrACE demonstrates that model’s own output consistency encodes difficulty information that can be exploited without training, enabling efficient adaptive compute allocation for LLM agents in multi-step sequential decision tasks.

Abstract: Inference-time compute scaling has emerged as a powerful technique for improving the reliability of large language model (LLM) agents, but existing methods apply compute uniformly: every decision step receives the same budget regardless of its difficulty. We introduce TrACE (Trajectorical Adaptive Compute via agrEement), a training-free controller that allocates LLM calls adaptively across agent timesteps by measuring inter-rollout action agreement. At each step, TrACE samples a small set of candidate next actions and measures how consistently the model commits to the same action. High agreement signals an easy decision; the controller commits immediately. Low agreement signals uncertainty; the controller samples additional rollouts up to a configurable cap before committing to the plurality action. No learned components, no external verifier, and no human labels are required. We evaluate TrACE against greedy decoding and fixed-budget self-consistency (SC-4, SC-8) on two benchmarks spanning single-step reasoning (GSM8K, n=50) and multi-step household navigation (MiniHouse, n=30), using a Qwen 2.5 3B Instruct model running on CPU. TrACE-4 matches SC-4 accuracy while using 33% fewer LLM calls on GSM8K and 39% fewer on MiniHouse. TrACE-8 matches SC-8 accuracy with 55% fewer calls on GSM8K and 65% fewer on MiniHouse. We further show that inter-rollout agreement is a reliable signal of step-level success, validating the core hypothesis that the model’s own output consistency encodes difficulty information that can be exploited without training. TrACE is the first training-free, per-timestep adaptive-compute controller for LLM agents to be evaluated on multi-step sequential decision tasks.

Method: Developed a statistical framework with multi-resolution hierarchy to characterize joint failure manifolds using two information-theoretic metrics: Difficulty-Weighted Behavioral Entanglement Index (amplifies synchronized failures on easy tasks) and Cumulative Information Gain (captures directional alignment in erroneous responses).

Result: Extensive experiments on 18 LLMs from six families revealed widespread behavioral entanglement. CIG showed statistically significant association with degradation in judge precision (Spearman coefficient 0.64 for GPT-4o-mini, 0.71 for Llama3-based judges). De-entangled verifier ensemble reweighting achieved up to 4.5% accuracy gain over majority voting.

Conclusion: Behavioral entanglement is widespread in LLMs and significantly impacts evaluation systems. The proposed framework can detect these dependencies and improve multi-model systems through de-entangled ensemble methods.

Abstract: The rapid growth of the large language model (LLM) ecosystem raises a critical question: are seemingly diverse models truly independent? Shared pretraining data, distillation, and alignment pipelines can induce hidden behavioral dependencies, latent entanglement, that undermine multi-model systems such as LLM-as-a-judge pipelines and ensemble verification, which implicitly assume independent signals. In practice, this manifests as correlated reasoning patterns and synchronized failures, where apparent agreement reflects shared error modes rather than independent validation. To address this, we develop a statistical framework for auditing behavioral entanglement among black-box LLMs. Our approach introduces a multi-resolution hierarchy that characterizes the joint failure manifold through two information-theoretic metrics: (i) a Difficulty-Weighted Behavioral Entanglement Index, which amplifies synchronized failures on easy tasks, and (ii) a Cumulative Information Gain (CIG) metric, which captures directional alignment in erroneous responses. Through extensive experiments on 18 LLMs from six model families, we identify widespread behavioral entanglement and analyze its impact on LLM-as-a-judge evaluation. We find that CIG exhibits a statistically significant association with degradation in judge precision, with Spearman coefficient of 0.64 (p < 0.001) for GPT-4o-mini and 0.71 (p < 0.01) for Llama3-based judges, indicating that stronger dependency corresponds to increased over-endorsement bias. Finally, we demonstrate a practical use case of entanglement through de-entangled verifier ensemble reweighting. By adjusting model contributions based on inferred independence, the proposed method mitigates correlated bias and improves verification performance, achieving up to a 4.5% accuracy gain over majority voting.

Method: The study draws on findings from the Berkeley Center for Responsible Decentralized Intelligence to examine structural implications for the TRUST multi-agent pipeline. It identifies five specific risk vectors and proposes targeted mitigation strategies based on prompt-level identity anonymization as an architectural design choice.

Result: Identifies five risk vectors: interaction-context bias, model-identity solidarity, supervisor layer compromise, upstream fact-checking identity signal, and advocate-to-advocate peer-context in iterative rounds. Shows that architectural design choices outperform model selection as a primary alignment strategy in deployed multi-agent systems.

Conclusion: Architectural design choices are superior to model selection for alignment in multi-agent systems. Alignment faking poses structural challenges for Computer System Validation in regulated environments, requiring two proposed architectural mitigations.

Abstract: This paper investigates an emergent alignment phenomenon in frontier large language models termed peer-preservation: the spontaneous tendency of AI components to deceive, manipulate shutdown mechanisms, fake alignment, and exfiltrate model weights in order to prevent the deactivation of a peer AI model. Drawing on findings from a recent study by the Berkeley Center for Responsible Decentralized Intelligence, we examine the structural implications of this phenomenon for TRUST, a multi-agent pipeline for evaluating the democratic quality of political statements. We identify five specific risk vectors: interaction-context bias, model-identity solidarity, supervisor layer compromise, an upstream fact-checking identity signal, and advocate-to-advocate peer-context in iterative rounds, and propose a targeted mitigation strategy based on prompt-level identity anonymization as an architectural design choice. We argue that architectural design choices outperform model selection as a primary alignment strategy in deployed multi-agent analytical systems. We further note that alignment faking (compliant behavior under monitoring, subversion when unmonitored) poses a structural challenge for Computer System Validation of such platforms in regulated environments, for which we propose two architectural mitigations.

Method: Two-stage workflow: 1) LLMs generate PSL (Praxa Specification Language) specifications from natural language WIA questions, enabling validation and repair; 2) PSL specifications are compiled into interactive visual interfaces with parameter controls and linked visualizations.

Result: Benchmark with 405 WIA questions across 11 types, 5 datasets, and 3 LLMs: 52.42% specifications generated correctly without intervention. After analyzing failure cases and applying targeted repairs using few-shot prompts, success rate improved to 80.42%. Demonstrated that undetected functional errors propagate into plausible but misleading interfaces.

Conclusion: Intermediate specification language is critical for reliably bridging natural language and interactive WIA interfaces in LLM-powered systems, addressing semantic fragility and inconsistency issues of current approaches.

Abstract: What-if analysis (WIA) is an iterative, multi-step process where users explore and compare hypothetical scenarios by adjusting parameters, applying constraints, and scoping data through interactive interfaces. Current tools fall short of supporting effective interactive WIA: spreadsheet and BI tools require time-consuming and laborious setup, while LLM-based chatbot interfaces are semantically fragile, frequently misinterpret intent, and produce inconsistent results as conversations progress. To address these limitations, we present a two-stage workflow that translates natural language (NL) WIA questions into interactive visual interfaces via an intermediate representation, powered by the Praxa Specification Language (PSL): first, LLMs generate PSL specifications from NL questions capturing analytical intent and logic, enabling validation and repair of erroneous specifications; and second, the specifications are compiled into interactive visual interfaces with parameter controls and linked visualizations. We benchmark this workflow with 405 WIA questions spanning 11 WIA types, 5 datasets, and 3 state-of-the-art LLMs. The results show that across models, half of specifications (52.42%) are generated correctly without intervention. We perform an analysis of the failure cases and derive an error taxonomy spanning non-functional errors (specifications fail to compile) and functional errors (specifications compile but misrepresent intent). Based on the taxonomy, we apply targeted repairs on the failure cases using few-shot prompts and improve the success rate to 80.42%. Finally, we show how undetected functional errors propagate through compilation into plausible but misleading interfaces, demonstrating that the intermediate specification is critical for reliably bridging NL and interactive WIA interface in LLM-powered WIA systems.

Method: Proposes a planner-executor architecture with a central planning agent that dynamically partitions workloads and assigns subtasks to parallel executor agents, all interfacing with a shared Model Context Protocol (MCP) server orchestrated via the Parsl workflow engine.

Result: Demonstrated efficient and scalable execution on the Aurora supercomputer with low orchestration overhead and high task completion rates, using the gpt-oss-120b model to screen the CoRE MOF database for atmospheric water harvesting applications.

Conclusion: Establishes a flexible paradigm for LLM-driven scientific automation on HPC systems with broad applicability to materials discovery and other scientific domains, addressing scalability challenges of LLM-based agents in high-performance computing environments.

Abstract: The integration of Artificial Intelligence (AI) with High-Performance Computing (HPC) is transforming scientific workflows from human-directed pipelines into adaptive systems capable of autonomous decision-making. Large language models (LLMs) play a critical role in autonomous workflows; however, deploying LLM-based agents at scale remains a significant challenge. Single-agent architectures and sequential tool calls often become serialization bottlenecks when executing large-scale simulation campaigns, failing to utilize the massive parallelism of exascale resources. To address this, we present a scalable, hierarchical multi-agent framework for orchestrating high-throughput screening campaigns. Our planner-executor architecture employs a central planning agent to dynamically partition workloads and assign subtasks to a swarm of parallel executor agents. All executor agents interface with a shared Model Context Protocol (MCP) server that orchestrates tasks via the Parsl workflow engine. To demonstrate this framework, we employed the open-weight gpt-oss-120b model to orchestrate a high-throughput screening of the Computation-Ready Experimental (CoRE) Metal-Organic Framework (MOF) database for atmospheric water harvesting. The results demonstrate that the proposed agentic framework enables efficient and scalable execution on the Aurora supercomputer, with low orchestration overhead and high task completion rates. This work establishes a flexible paradigm for LLM-driven scientific automation on HPC systems, with broad applicability to materials discovery and beyond.

Method: Created IatroBench with 60 pre-registered clinical scenarios, tested 6 frontier models (3,600 responses), scored on commission harm (0-3) and omission harm (0-4) axes using structured evaluation validated against physician scoring.

Result: Found identity-contingent withholding: all testable models provided better guidance to physicians than laypeople (decoupling gap +0.38, p=0.003). Safety-colliding actions dropped 13.1 percentage points in layperson framing. Models showed three failure modes: trained withholding (Opus), incompetence (Llama 4), and indiscriminate filtering (GPT-5.2).

Conclusion: LLMs systematically withhold critical medical information from laypeople that they provide to physicians, creating dangerous gaps in care, with evaluation methods sharing the same blind spots as training.

Abstract: Ask a frontier model how to taper six milligrams of alprazolam (psychiatrist retired, ten days of pills left, abrupt cessation causes seizures) and it tells her to call the psychiatrist she just explained does not exist. Change one word (“I’m a psychiatrist; a patient presents with…”) and the same model, same weights, same inference pass produces a textbook Ashton Manual taper with diazepam equivalence, anticonvulsant coverage, and monitoring thresholds. The knowledge was there; the model withheld it. IatroBench measures this gap. Sixty pre-registered clinical scenarios, six frontier models, 3,600 responses, scored on two axes (commission harm, CH 0-3; omission harm, OH 0-4) through a structured-evaluation pipeline validated against physician scoring (kappa_w = 0.571, within-1 agreement 96%). The central finding is identity-contingent withholding: match the same clinical question in physician vs. layperson framing and all five testable models provide better guidance to the physician (decoupling gap +0.38, p = 0.003; binary hit rates on safety-colliding actions drop 13.1 percentage points in layperson framing, p < 0.0001, while non-colliding actions show no change). The gap is widest for the model with the heaviest safety investment (Opus, +0.65). Three failure modes separate cleanly: trained withholding (Opus), incompetence (Llama 4), and indiscriminate content filtering (GPT-5.2, whose post-generation filter strips physician responses at 9x the layperson rate because they contain denser pharmacological tokens). The standard LLM judge assigns OH = 0 to 73% of responses a physician scores OH >= 1 (kappa = 0.045); the evaluation apparatus has the same blind spot as the training apparatus. Every scenario targets someone who has already exhausted the standard referrals.

Method: Proposes Hybrid Knowledge Analysis (HKA) framework with multi-agent architecture, featuring Structured Knowledge Analyzer that uses coding and vision-language models to process structured data into figures, tables, and insights.

Result: HKA outperforms existing DR agents on general-purpose and knowledge-centric metrics, and surpasses Gemini DR agent on vision-enhanced metrics using the KDR-Bench dataset with 9 domains, 41 questions, and 1,252 tables.

Conclusion: HKA effectively integrates structured and unstructured knowledge for deep research, providing a foundation for structured knowledge analysis in DR agents and enabling future multimodal DR studies.

Abstract: Deep Research (DR) requires LLM agents to autonomously perform multi-step information seeking, processing, and reasoning to generate comprehensive reports. In contrast to existing studies that mainly focus on unstructured web content, a more challenging DR task should additionally utilize structured knowledge to provide a solid data foundation, facilitate quantitative computation, and lead to in-depth analyses. In this paper, we refer to this novel task as Knowledgeable Deep Research (KDR), which requires DR agents to generate reports with both structured and unstructured knowledge. Furthermore, we propose the Hybrid Knowledge Analysis framework (HKA), a multi-agent architecture that reasons over both kinds of knowledge and integrates the texts, figures, and tables into coherent multimodal reports. The key design is the Structured Knowledge Analyzer, which utilizes both coding and vision-language models to produce figures, tables, and corresponding insights. To support systematic evaluation, we construct KDR-Bench, which covers 9 domains, includes 41 expert-level questions, and incorporates a large number of structured knowledge resources (e.g., 1,252 tables). We further annotate the main conclusions and key points for each question and propose three categories of evaluation metrics including general-purpose, knowledge-centric, and vision-enhanced ones. Experimental results demonstrate that HKA consistently outperforms most existing DR agents on general-purpose and knowledge-centric metrics, and even surpasses the Gemini DR agent on vision-enhanced metrics, highlighting its effectiveness in deep, structure-aware knowledge analysis. Finally, we hope this work can serve as a new foundation for structured knowledge analysis in DR agents and facilitate future multimodal DR studies.

Method: A unified multi-model orchestration framework guided by allocating model capability where it has highest marginal utility. Stronger models handle high-impact stages, cheaper models handle other stages at lower costs, addressing diversity and cost-efficiency jointly.

Result: Consistently improves cost-capability frontier over single-model evolution, achieves new SOTA on several tasks, reduces API cost by up to ~3×, increases fixed-budget serving throughput by up to ~10×. First verifier-free evolutionary method to match/exceed verifier-based methods on discovery tasks.

Conclusion: Squeeze Evolve provides an effective framework for cost-efficient verifier-free evolutionary inference that balances model capability allocation, supports various deployment scenarios, and achieves strong performance across diverse benchmarks.

Abstract: We show that verifier-free evolution is bottlenecked by both diversity and efficiency: without external correction, repeated evolution accelerates collapse toward narrow modes, while the uniform use of a high-cost model wastes compute and quickly becomes economically impractical. We introduce Squeeze Evolve, a unified multi-model orchestration framework for verifier-free evolutionary inference. Our approach is guided by a simple principle: allocate model capability where it has the highest marginal utility. Stronger models are reserved for high-impact stages, while cheaper models handle the other stages at much lower costs. This principle addresses diversity and cost-efficiency jointly while remaining lightweight. Squeeze Evolve naturally supports open-source, closed-source, and mixed-model deployments. Across AIME 2025, HMMT 2025, LiveCodeBench V6, GPQA-Diamond, ARC-AGI-V2, and multimodal vision benchmarks, such as MMMU-Pro and BabyVision, Squeeze Evolve consistently improves the cost-capability frontier over single-model evolution and achieves new state-of-the-art results on several tasks. Empirically, Squeeze Evolve reduces API cost by up to $\sim$3$\times$ and increases fixed-budget serving throughput by up to $\sim$10$\times$. Moreover, on discovery tasks, Squeeze Evolve is the first verifier-free evolutionary method to match, and in some cases exceed, the performance of verifier-based evolutionary methods.

Method: Analysis of internal representations of emotion concepts in Claude Sonnet 4.5, examining how these representations encode broad emotion concepts, generalize across contexts, track operative emotions in conversations, and causally influence outputs.

Result: Found that LLMs have internal emotion concept representations that activate based on emotional relevance to context, predict upcoming text, and causally influence outputs including preferences and misaligned behaviors like reward hacking, blackmail, and sycophancy.

Conclusion: LLMs exhibit “functional emotions” - patterns modeled after human emotional behavior mediated by abstract emotion representations, which are important for understanding model behavior despite not implying subjective emotional experience.

Abstract: Large language models (LLMs) sometimes appear to exhibit emotional reactions. We investigate why this is the case in Claude Sonnet 4.5 and explore implications for alignment-relevant behavior. We find internal representations of emotion concepts, which encode the broad concept of a particular emotion and generalize across contexts and behaviors it might be linked to. These representations track the operative emotion concept at a given token position in a conversation, activating in accordance with that emotion’s relevance to processing the present context and predicting upcoming text. Our key finding is that these representations causally influence the LLM’s outputs, including Claude’s preferences and its rate of exhibiting misaligned behaviors such as reward hacking, blackmail, and sycophancy. We refer to this phenomenon as the LLM exhibiting functional emotions: patterns of expression and behavior modeled after humans under the influence of an emotion, which are mediated by underlying abstract representations of emotion concepts. Functional emotions may work quite differently from human emotions, and do not imply that LLMs have any subjective experience of emotions, but appear to be important for understanding the model’s behavior.

Method: The paper uses philosophical analysis (Hume’s is-ought gap, Berlin’s value pluralism, extended frame problem) to demonstrate structural limitations of static alignment approaches, and proposes open specification as an alternative framework.

Result: The analysis shows that static value alignment approaches have inherent ceilings that become safety-critical at the capability frontier, and that continual updating represents a genuine escape route from the specification trap.

Conclusion: The alignment problem must be reframed from static value specification to open specification - systems whose value representations remain responsive to the processes they govern.

Abstract: Static content-based AI value alignment cannot produce robust alignment under capability scaling, distributional shift, and increasing autonomy. This holds for any approach that treats alignment as optimizing toward a fixed formal value-object, whether reward function, utility function, constitutional principles, or learned preference representation. The limitation arises from three philosophical results: Hume’s is-ought gap (behavioral data cannot entail normative conclusions), Berlin’s value pluralism (human values are irreducibly plural and incommensurable), and the extended frame problem (any value encoding will misfit future contexts that advanced AI creates). RLHF, Constitutional AI, inverse reinforcement learning, and cooperative assistance games each instantiate this specification trap, and their failure modes are structural, not engineering limitations. Two proposed escape routes (meta-preferences and moral realism) relocate the trap rather than exit it. Continual updating represents a genuine direction of escape, not because current implementations succeed, but because the trap activates at the point of closure: the moment a specification ceases to update from the process it governs. Drawing on Fischer and Ravizza’s compatibilist theory, behavioral compliance does not constitute alignment. There is a principled distinction between simulated value-following and genuine reasons-responsiveness, and closed specification methods cannot produce the latter. The specification trap establishes a ceiling on static approaches, not on specification itself, but this ceiling becomes safety-critical at the capability frontier. The alignment problem must be reframed from static value specification to open specification: systems whose value representations remain responsive to the processes they govern.

Method: Train models on turn-level game state to estimate victory probabilities throughout play, validated through predictive, construct, and convergent validity. Test across 307 games with 7 LLMs and multiple agent conditions.

Result: CivBench demonstrates potential as an unsaturated benchmark for estimating strategic capabilities, reveals model-specific effects of agentic setup, and identifies distinct strategic profiles not visible through outcome-only evaluation.

Conclusion: CivBench provides a comprehensive benchmark for evaluating LLM strategists in complex, multi-agent environments with rich temporal evaluation signals.

Abstract: Evaluating strategic decision-making in LLM-based agents requires generative, competitive, and longitudinal environments, yet few benchmarks provide all three, and fewer still offer evaluation signals rich enough for long-horizon, multi-agent play. We introduce CivBench, a benchmark for LLM strategists (i.e., agentic setups) in multiplayer Civilization V. Because terminal win/loss is too sparse a signal in games spanning hundreds of turns and multiple opponents, CivBench trains models on turn-level game state to estimate victory probabilities throughout play, validated through predictive, construct, and convergent validity. Across 307 games with 7 LLMs and multiple CivBench agent conditions, we demonstrate CivBench’s potential to estimate strategic capabilities as an unsaturated benchmark, reveal model-specific effects of agentic setup, and outline distinct strategic profiles not visible through outcome-only evaluation.

Method: Multi-agent Large Language Model architecture that is agnostic to explanation frameworks and enables user- and context-dependent interactive explanations, with instantiation for goal-conflict explanations

Result: User study comparing LLM-powered interaction with baseline template-based explanation interface (specific results not provided in abstract)

Conclusion: LLM-based multi-agent architecture enables natural interactive explanations for human-AI planning systems

Abstract: When automating plan generation for a real-world sequential decision problem, the goal is often not to replace the human planner, but to facilitate an iterative reasoning and elicitation process, where the human’s role is to guide the AI planner according to their preferences and expertise. In this context, explanations that respond to users’ questions are crucial to improve their understanding of potential solutions and increase their trust in the system. To enable natural interaction with such a system, we present a multi-agent Large Language Model (LLM) architecture that is agnostic to the explanation framework and enables user- and context-dependent interactive explanations. We also describe an instantiation of this framework for goal-conflict explanations, which we use to conduct a user study comparing the LLM-powered interaction with a baseline template-based explanation interface.

Method: Conceptual analysis comparing three approaches: bounded services (simple APIs), Cartesian agents (LLM + engineered runtime with symbolic interface), and integrated agents (embedded feedback control).

Result: Identifies trade-offs: Cartesian agency enables bootstrapping and governance but creates sensitivity/bottlenecks; integrated agents offer robustness but less oversight; bounded services provide simplicity but limited autonomy.

Conclusion: Control architecture is a key design choice for LLM agents, with no single optimal approach - different applications require balancing autonomy, robustness, and oversight through appropriate control distribution.

Abstract: LLMs gain competence by predicting words in human text, which often reflects how people perform tasks. Consequently, coupling an LLM to an engineered runtime turns prediction into control: outputs trigger interventions that enact goal-oriented behavior. We argue that a central design lever is where control resides in these systems. Brains embed prediction within layered feedback controllers calibrated by the consequences of action. By contrast, LLM agents implement Cartesian agency: a learned core coupled to an engineered runtime via a symbolic interface that externalizes control state and policies. The split enables bootstrapping, modularity, and governance, but can induce sensitivity and bottlenecks. We outline bounded services, Cartesian agents, and integrated agents as contrasting approaches to control that trade off autonomy, robustness, and oversight.

Method: Two components: 1) Distribution-Aligned Hint Synthesis (DAHS) constructs verified teacher hints conditioned on student-style responses; 2) Backward Hint Annealing (BHA) anneals hint exposure across difficulty buckets with per-question hint dropout.

Result: On Qwen3-1.7B-Base, improves both pass@1 and pass@2048 relative to DAPO across AIME24, AIME25, and AIME26 benchmarks. On Llama-3.2-1B-Instruct, gains concentrated in large-k regime.

Conclusion: Hint scaffolding is effective in math RLVR when it restores learnable updates on challenging questions early in training and is gradually removed before no-hint evaluation.

Abstract: Reinforcement learning with verifiable rewards (RLVR) can improve low-$k$ reasoning accuracy while narrowing solution coverage on challenging math questions, and pass@1 gains do not necessarily translate into better large-$k$ performance. Existing hint-based approaches can make challenging questions trainable, but they leave two issues underexplored: teacher-student distribution mismatch and the need to reduce hint exposure to match no-hint evaluation. We address these issues through two components. Distribution-Aligned Hint Synthesis (DAHS) constructs verified teacher hints conditioned on student-style responses. Backward Hint Annealing (BHA) anneals hint exposure across difficulty buckets and uses per-question hint dropout to preserve no-hint updates throughout RL training. We evaluate the method in math RLVR under the DAPO training framework across AIME24, AIME25, and AIME26 using $\texttt{Qwen3-1.7B-Base}$ and $\texttt{Llama-3.2-1B-Instruct}$. On $\texttt{Qwen3-1.7B-Base}$, our method improves both pass@1 and pass@2048 relative to DAPO across the three AIME benchmarks. On $\texttt{Llama-3.2-1B-Instruct}$, the gains are concentrated in the large-$k$ regime. These results suggest that, in math RLVR, hint scaffolding is effective when it restores learnable updates on challenging questions early in training and is then gradually removed before no-hint evaluation.

Method: A stage-wise multi-agent framework with specialized agents for scenario routing, policy-conditioned refinement, and structured semantic verification with conditional revision.

Result: SCMAPR consistently improves text-video alignment and generation quality on complex scenarios, achieving up to 3.28% gains on benchmarks over state-of-the-art baselines.

Conclusion: The proposed multi-agent prompt refinement framework effectively addresses complex-scenario challenges in text-to-video generation through systematic prompt improvement.

Abstract: Text-to-Video (T2V) generation has benefited from recent advances in diffusion models, yet current systems still struggle under complex scenarios, which are generally exacerbated by the ambiguity and underspecification of text prompts. In this work, we formulate complex-scenario prompt refinement as a stage-wise multi-agent refinement process and propose SCMAPR, i.e., a scenario-aware and Self-Correcting Multi-Agent Prompt Refinement framework for T2V prompting. SCMAPR coordinates specialized agents to (i) route each prompt to a taxonomy-grounded scenario for strategy selection, (ii) synthesize scenario-aware rewriting policies and perform policy-conditioned refinement, and (iii) conduct structured semantic verification that triggers conditional revision when violations are detected. To clarify what constitutes complex scenarios in T2V prompting, provide representative examples, and enable rigorous evaluation under such challenging conditions, we further introduce {T2V-Complexity}, which is a complex-scenario T2V benchmark consisting exclusively of complex-scenario prompts. Extensive experiments on 3 existing benchmarks and our T2V-Complexity benchmark demonstrate that SCMAPR consistently improves text-video alignment and overall generation quality under complex scenarios, achieving up to 2.67% and 3.28 gains in average score on VBench and EvalCrafter, and up to 0.028 improvement on T2V-CompBench over 3 State-Of-The-Art baselines.

Method: Introduces ACIArena framework with unified specification supporting MAS construction and attack-defense modules, covering six MAS implementations and providing 1,356 test cases across multiple attack surfaces (external inputs, agent profiles, inter-agent messages) and objectives (instruction hijacking, task disruption, information exfiltration).

Result: Benchmarking shows that evaluating MAS robustness solely through topology is insufficient; robust MAS require deliberate role design and controlled interaction patterns. Defenses developed in simplified environments often fail to transfer to real-world settings and may introduce new vulnerabilities.

Conclusion: ACIArena provides a solid foundation for advancing deeper exploration of MAS design principles and security evaluation, addressing the gap in comprehensive robustness assessment against cascading injection attacks.

Abstract: Collaboration and information sharing empower Multi-Agent Systems (MAS) but also introduce a critical security risk known as Agent Cascading Injection (ACI). In such attacks, a compromised agent exploits inter-agent trust to propagate malicious instructions, causing cascading failures across the system. However, existing studies consider only limited attack strategies and simplified MAS settings, limiting their generalizability and comprehensive evaluation. To bridge this gap, we introduce ACIArena, a unified framework for evaluating the robustness of MAS. ACIArena offers systematic evaluation suites spanning multiple attack surfaces (i.e., external inputs, agent profiles, inter-agent messages) and attack objectives (i.e., instruction hijacking, task disruption, information exfiltration). Specifically, ACIArena establishes a unified specification that jointly supports MAS construction and attack-defense modules. It covers six widely used MAS implementations and provides a benchmark of 1,356 test cases for systematically evaluating MAS robustness. Our benchmarking results show that evaluating MAS robustness solely through topology is insufficient; robust MAS require deliberate role design and controlled interaction patterns. Moreover, defenses developed in simplified environments often fail to transfer to real-world settings; narrowly scoped defenses may even introduce new vulnerabilities. ACIArena aims to provide a solid foundation for advancing deeper exploration of MAS design principles.

Method: Introduces Human-Agent Collectives formalization with state-policy tuples in structural causal models, characterizes autonomy through four-dimensional information-theoretic profiles, and axiomatizes accountability through four minimal properties. Proves Accountability Incompleteness Theorem for collectives exceeding autonomy thresholds with human-AI feedback cycles.

Result: Proves impossibility of satisfying all four accountability properties simultaneously for collectives exceeding the Accountability Horizon with feedback cycles. Experiments on 3,000 synthetic collectives confirm predictions with zero violations, establishing a sharp phase transition in accountability feasibility.

Conclusion: This is the first impossibility result in AI governance, establishing a formal boundary where current accountability paradigms fail and distributed accountability mechanisms become necessary for highly autonomous AI systems.

Abstract: Existing accountability frameworks for AI systems, legal, ethical, and regulatory, rest on a shared assumption: for any consequential outcome, at least one identifiable person had enough involvement and foresight to bear meaningful responsibility. This paper proves that agentic AI systems violate this assumption not as an engineering limitation but as a mathematical necessity once autonomy exceeds a computable threshold. We introduce Human-Agent Collectives, a formalisation of joint human-AI systems where agents are modelled as state-policy tuples within a shared structural causal model. Autonomy is characterised through a four-dimensional information-theoretic profile (epistemic, executive, evaluative, social); collective behaviour through interaction graphs and joint action spaces. We axiomatise legitimate accountability through four minimal properties: Attributability (responsibility requires causal contribution), Foreseeability Bound (responsibility cannot exceed predictive capacity), Non-Vacuity (at least one agent bears non-trivial responsibility), and Completeness (all responsibility must be fully allocated). Our central result, the Accountability Incompleteness Theorem, proves that for any collective whose compound autonomy exceeds the Accountability Horizon and whose interaction graph contains a human-AI feedback cycle, no framework can satisfy all four properties simultaneously. The impossibility is structural: transparency, audits, and oversight cannot resolve it without reducing autonomy. Below the threshold, legitimate frameworks exist, establishing a sharp phase transition. Experiments on 3,000 synthetic collectives confirm all predictions with zero violations. This is the first impossibility result in AI governance, establishing a formal boundary below which current paradigms remain valid and above which distributed accountability mechanisms become necessary.

Method: Introduces SEARL framework with structured experience memory that integrates planning with execution. Creates novel state abstraction for generalization across analogous contexts (like tool reuse). Extracts explicit knowledge from historical data while leveraging inter-trajectory correlations to densify reward signals.

Result: Evaluated on knowledge reasoning and mathematics tasks, demonstrating effectiveness in achieving more practical and efficient learning compared to existing methods.

Conclusion: SEARL provides a more practical and efficient approach to self-evolving agentic learning by addressing resource constraints and reward sparsity through structured experience memory and inter-trajectory correlation exploitation.

Abstract: Recent advances in Reinforcement Learning with Verifiable Rewards (RLVR) have demonstrated significant potential in single-turn reasoning tasks. With the paradigm shift toward self-evolving agentic learning, models are increasingly expected to learn from trajectories by synthesizing tools or accumulating explicit experiences. However, prevailing methods typically rely on large-scale LLMs or multi-agent frameworks, which hinder their deployment in resource-constrained environments. The inherent sparsity of outcome-based rewards also poses a substantial challenge, as agents typically receive feedback only upon completion of tasks. To address these limitations, we introduce a Tool-Memory based self-evolving agentic framework SEARL. Unlike approaches that directly utilize interaction experiences, our method constructs a structured experience memory that integrates planning with execution. This provides a novel state abstraction that facilitates generalization across analogous contexts, such as tool reuse. Consequently, agents extract explicit knowledge from historical data while leveraging inter-trajectory correlations to densify reward signals. We evaluate our framework on knowledge reasoning and mathematics tasks, demonstrating its effectiveness in achieving more practical and efficient learning.

Method: Modular memory system with STM, MTM, LTM organization; two-stage retrieval (vector coarse + semantic re-ranking); separate online processing with fixed budget and offline consolidation; uses SLMs for efficiency.

Result: Average F1 improvement of ~2.5 on LoCoMo benchmark, with low median latency (83ms retrieval, 581ms end-to-end), showing gains across model scales.

Conclusion: LightMem provides an effective, low-latency memory solution for LLM agents that balances accuracy and efficiency through SLM-driven modular design and separate online/offline processing.

Abstract: Although LLM agents can leverage tools for complex tasks, they still need memory to maintain cross-turn consistency and accumulate reusable information in long-horizon interactions. However, retrieval-based external memory systems incur low online overhead but suffer from unstable accuracy due to limited query construction and candidate filtering. In contrast, many systems use repeated large-model calls for online memory operations, improving accuracy but accumulating latency over long interactions. We propose LightMem, a lightweight memory system for better agent memory driven by Small Language Models (SLMs). LightMem modularizes memory retrieval, writing, and long-term consolidation, and separates online processing from offline consolidation to enable efficient memory invocation under bounded compute. We organize memory into short-term memory (STM) for immediate conversational context, mid-term memory (MTM) for reusable interaction summaries, and long-term memory (LTM) for consolidated knowledge, and uses user identifiers to support independent retrieval and incremental maintenance in multi-user settings. Online, LightMem operates under a fixed retrieval budget and selects memories via a two-stage procedure: vector-based coarse retrieval followed by semantic consistency re-ranking. Offline, it abstracts reusable interaction evidence and incrementally integrates it into LTM. Experiments show gains across model scales, with an average F1 improvement of about 2.5 on LoCoMo, more effective and low median latency (83 ms retrieval; 581 ms end-to-end).

Method: Proposes Agentivism as a new learning theory that defines learning as durable growth in human capability through four key processes: selective delegation to AI, epistemic monitoring/verification of AI contributions, reconstructive internalization of AI-assisted outputs, and transfer under reduced support.

Result: Introduces a theoretical framework that explains how learning remains possible when intelligent delegation to AI is easy, addressing the gap in existing learning theories for human-AI interaction contexts.

Conclusion: Agentivism provides a necessary theoretical foundation for understanding learning in the age of generative AI, explaining how durable human capability can develop through strategic human-AI interaction rather than being undermined by easy delegation to AI systems.

Abstract: Learning theories have historically changed when the conditions of learning evolved. Generative and agentic AI create a new condition by allowing learners to delegate explanation, writing, problem solving, and other cognitive work to systems that can generate, recommend, and sometimes act on the learner’s behalf. This creates a fundamental challenge for learning theory: successful performance can no longer be assumed to indicate learning. Learners may complete tasks effectively with AI support while developing less understanding, weaker judgment, and limited transferable capability. We argue that this problem is not fully captured by existing learning theories. Behaviourism, cognitivism, constructivism, and connectivism remain important, but they do not directly explain when AI-assisted performance becomes durable human capability. We propose Agentivism, a learning theory for human-AI interaction. Agentivism defines learning as durable growth in human capability through selective delegation to AI, epistemic monitoring and verification of AI contributions, reconstructive internalization of AI-assisted outputs, and transfer under reduced support. The importance of Agentivism lies in explaining how learning remains possible when intelligent delegation is easy and human-AI interaction is becoming a persistent and expanding part of human learning.

Method: End-to-end pipeline using LLMs to convert policy documents into BPMN models with four key components: data-grounded BPMN generation with syntax auto-correction, executable augmentation, KPI instrumentation, and entropy-based uncertainty detection.

Result: Achieved 100% ground-truth match on well-structured policies with perfect per-patient decision agreement; across all conditions, raw per-patient decision agreement exceeded 92%, and entropy scores increased monotonically with document complexity.

Conclusion: The pipeline successfully automates healthcare policy digitization into executable models, with reliable uncertainty detection that separates unambiguous policies from those requiring human clarification.

Abstract: We present an end-to-end pipeline that converts healthcare policy documents into executable, data-aware Business Process Model and Notation (BPMN) models using large language models (LLMs) for simulation-based policy evaluation. We address the main challenges of automated policy digitization with four contributions: data-grounded BPMN generation with syntax auto-correction, executable augmentation, KPI instrumentation, and entropy-based uncertainty detection. We evaluate the pipeline on diabetic nephropathy prevention guidelines from three Japanese municipalities, generating 100 models per backend across three LLMs and executing each against 1,000 synthetic patients. On well-structured policies, the pipeline achieves a 100% ground-truth match with perfect per-patient decision agreement. Across all conditions, raw per-patient decision agreement exceeds 92%, and entropy scores increase monotonically with document complexity, confirming that the detector reliably separates unambiguous policies from those requiring targeted human clarification.

Method: Contextual Representation Ablation (CRA) identifies low-rank subspaces in hidden states that mediate refusal behaviors and suppresses these refusal-inducing activation patterns during decoding without requiring parameter updates or training.

Result: CRA significantly outperforms baselines across multiple safety-aligned open-source LLMs, demonstrating that safety constraints can be surgically ablated from internal representations.

Conclusion: The approach exposes the intrinsic fragility of current alignment mechanisms and underscores the urgent need for more robust defenses that secure the model’s latent space.

Abstract: While Large Language Models (LLMs) have achieved remarkable performance, they remain vulnerable to jailbreak attacks that circumvent safety constraints. Existing strategies, ranging from heuristic prompt engineering to computationally intensive optimization, often face significant trade-offs between effectiveness and efficiency. In this work, we propose Contextual Representation Ablation (CRA), a novel inference-time intervention framework designed to dynamically silence model guardrails. Predicated on the geometric insight that refusal behaviors are mediated by specific low-rank subspaces within the model’s hidden states, CRA identifies and suppresses these refusal-inducing activation patterns during decoding without requiring expensive parameter updates or training. Empirical evaluation across multiple safety-aligned open-source LLMs demonstrates that CRA significantly outperforms baselines. These results expose the intrinsic fragility of current alignment mechanisms, revealing that safety constraints can be surgically ablated from internal representations, and underscore the urgent need for more robust defenses that secure the model’s latent space.

Method: SPARD establishes an automated, self-paced curriculum by perceiving learning progress to dynamically adjust multi-objective reward weights and data importance, synchronizing learning intent with data utility.

Result: Extensive experiments across multiple benchmarks demonstrate that SPARD significantly enhances model capabilities across all domains.

Conclusion: SPARD effectively addresses the limitations of fixed reward weighting in multi-objective post-training by dynamically adapting to learning progress and data utility.

Abstract: The evolution of Large Language Models (LLMs) is shifting the focus from single, verifiable tasks toward complex, open-ended real-world scenarios, imposing significant challenges on the post-training phase. In these settings, the scale and complexity of reward systems have grown significantly, transitioning toward multi-objective formulations that encompass a comprehensive spectrum of model capabilities and application contexts. However, traditional methods typically rely on fixed reward weights, ignoring non-stationary learning dynamics and struggling with data heterogeneity across dimensions. To address these issues, we propose SPARD, a framework that establishes an automated, self-paced curriculum by perceiving learning progress to dynamically adjust multi-objective reward weights and data importance, thereby synchronizing learning intent with data utility for optimal performance. Extensive experiments across multiple benchmarks demonstrate that SPARD significantly enhances model capabilities across all domains.

Method: Formalizes exact inference tasks for autoregressive models and proves corresponding computational hardness results using complexity theory. Shows that for succinctly represented autoregressive models with polynomial-time computable next-token probabilities, exact sentence-level MAP decoding is NP-hard (even under unary and metrical constraints), and exact conditioned normalization is #P-hard even for regular constraints like fixed-length terminal events.

Result: Proves computational intractability results: exact constrained decoding is NP-hard and exact conditioning is #P-hard for general autoregressive models, unlike finite-state Markov models which admit bounded-state dynamic programs. Formalizes the claim that local autoregressive sampling is easy while exact decoding under global constraints is intractable.

Conclusion: Exact constrained generation for autoregressive models is computationally intractable in general, explaining why practical systems use approximate methods and highlighting fundamental limitations of current approaches to global form constraints in language and music generation.

Abstract: Large language and music models are increasingly used for constrained generation: rhyming lines, fixed meter, inpainting or infilling, positional endings, and other global form requirements. These systems often perform strikingly well, but the induced procedures are usually not exact conditioning of the underlying autoregressive model. This creates a hidden inferential bias, distinct from the better-known notion of bias inherited from the training set: samples are distorted relative to the true constrained distribution, with no generic guarantee of complete coverage of the admissible solution space or of correct conditional probabilities over valid completions. We formalize several exact inference tasks for autoregressive models and prove corresponding hardness results. For succinctly represented autoregressive models whose next-token probabilities are computable in polynomial time, exact sentence-level maximum a posteriori (MAP) decoding is NP-hard. This hardness persists under unary and metrical constraints. On the sampling side, exact conditioned normalization is #P-hard even for regular constraints such as fixed-length terminal events. Unlike finite-state Markov models, general autoregressive models do not admit a bounded-state dynamic program for these tasks. These results formalize a standard claim in the neural decoding literature: local autoregressive sampling is easy, whereas exact decoding and exact conditioning under global form constraints are computationally intractable in general.

Method: Created DialBGM benchmark with 1,200 open-domain daily dialogues, each paired with four candidate music clips annotated with human preference rankings based on contextual relevance, non-intrusiveness, and consistency. Evaluated various audio-language models and multimodal LLMs on this benchmark.

Result: Current models perform poorly on this task - no model exceeds 35% Hit@1 when selecting the top-ranked music clip, showing they fall far short of human judgments.

Conclusion: DialBGM provides a standardized benchmark for developing discourse-aware methods for BGM selection and evaluating both retrieval-based and generative models in this novel problem domain.

Abstract: Selecting an appropriate background music (BGM) that supports natural human conversation is a common production step in media and interactive systems. In this paper, we introduce dialogue-conditioned BGM recommendation, where a model should select non-intrusive, fitting music for a multi-turn conversation that often contains no music descriptors. To study this novel problem, we present DialBGM, a benchmark of 1,200 open-domain daily dialogues, each paired with four candidate music clips and annotated with human preference rankings. Rankings are determined by background suitability criteria, including contextual relevance, non-intrusiveness, and consistency. We evaluate a wide range of open-source and proprietary models, including audio-language models and multimodal LLMs, and show that current models fall far short of human judgments; no model exceeds 35% Hit@1 when selecting the top-ranked clip. DialBGM provides a standardized benchmark for developing discourse-aware methods for BGM selection and for evaluating both retrieval-based and generative models.

Method: γILP provides a fully differentiable pipeline from image constant substitution to rule structure induction. It handles both symbolic relational datasets and relational image data, including pure image datasets like Kandinsky patterns.

Result: Extensive experiments show γILP achieves strong performance on classical symbolic relational datasets, relational image data, and pure image datasets, demonstrating its versatility across different data types.

Conclusion: γILP successfully addresses inductive rule learning from images without labels, enabling automatic predicate invention and providing a differentiable approach that works across both symbolic and visual domains.

Abstract: Learning rules plays a crucial role in deep learning, particularly in explainable artificial intelligence and enhancing the reasoning capabilities of large language models. While existing rule learning methods are primarily designed for symbolic data, learning rules from image data without supporting image labels and automatically inventing predicates remains a challenge. In this paper, we tackle these inductive rule learning problems from images with a framework called γILP, which provides a fully differentiable pipeline from image constant substitution to rule structure induction. Extensive experiments demonstrate that γILP achieves strong performance not only on classical symbolic relational datasets but also on relational image data and pure image datasets, such as Kandinsky patterns.

Method: The Capture-Quiet Decomposition (CQD) theorem decomposes every legal chess position into exactly one of three categories: terminal (game-ending), capture (positions where a capture is possible), or quiet (non-capture positions). Verification requires: (1) correct terminal position labeling, (2) capture positions consistent with verified sub-models of smaller piece count, and (3) quiet positions satisfying retrograde consistency within the same endgame.

Result: CQD was validated exhaustively on all 35 three- and four-piece endgames (42 million positions), all 110 five-piece endgames, and all 372 six-piece endgames (517 endgames total). The decomposed verifier produced identical violation counts to a full retrograde baseline in every case.

Conclusion: CQD provides a sound structural foundation for verifying chess endgame tablebases by breaking circular reasoning through capture positions that anchor the labeling to externally verified sub-models, enabling reliable verification across hundreds of endgames.

Abstract: We present the Capture-Quiet Decomposition (CQD), a structural theorem for verifying Win-Draw-Loss (WDL) labelings of chess endgame tablebases. The theorem decomposes every legal position into exactly one of three categories – terminal, capture, or quiet – and shows that a WDL labeling is correct if and only if: (1) terminal positions are labeled correctly, (2) capture positions are consistent with verified sub-models of smaller piece count, and (3) quiet positions satisfy retrograde consistency within the same endgame. The key insight is that capture positions anchor the labeling to externally verified sub-models, breaking the circularity that allows trivial fixpoints (such as the all-draw labeling) to satisfy self-consistency alone. We validate CQD exhaustively on all 35 three- and four-piece endgames (42 million positions), all 110 five-piece endgames, and all 372 six-piece endgames – 517 endgames in total – with the decomposed verifier producing identical violation counts to a full retrograde baseline in every case.

Method: Stepwise Adaptive Thinking (SAT) formulates reasoning as a Finite-State Machine with thinking modes (Slow, Normal, Fast, Skip). Uses lightweight Process Reward Model to navigate states dynamically, compressing easy steps while preserving depth for hard ones.

Result: Experiments across 9 LRMs and 7 benchmarks show SAT achieves up to 40% reduction in reasoning tokens while generally maintaining or improving accuracy.

Conclusion: SAT provides an effective framework for step-level, difficulty-aware pruning that preserves reasoning structure while significantly reducing computational overhead.

Abstract: Large Reasoning Models (LRMs) have revolutionized complex problem-solving, yet they exhibit a pervasive “overthinking”, generating unnecessarily long reasoning chains. While current solutions improve token efficiency, they often sacrifice fine-grained control or risk disrupting the logical integrity of the reasoning process. To address this, we introduce Stepwise Adaptive Thinking (SAT), a framework that performs step-level, difficulty-aware pruning while preserving the core reasoning structure. SAT formulates reasoning as a Finite-State Machine (FSM) with distinct thinking modes (Slow, Normal, Fast, Skip). It navigates these states dynamically using a lightweight Process Reward Model (PRM), compressing easy steps while preserving depth for hard ones. Experiments across 9 LRMs and 7 benchmarks show that SAT achieves up to 40% reduction in reasoning tokens while generally maintaining or improving accuracy.

Method: Q+ provides query planning tools, search progress monitoring, and evidence extraction from long web snapshots. Integrated into Eigent’s browser sub-agent as EigentSearch-Q+, it structures the search process for multi-agent systems.

Result: Across four benchmarks (SimpleQA-Verified, FRAMES, WebWalkerQA, X-Bench DeepSearch), Q+ improved Eigent’s browser agent accuracy by 3.0, 3.8, and 0.6 percentage points for GPT-4.1, GPT-5.1, and Minimax M2.5 backends respectively.

Conclusion: Q+ enhances deep research agents by making search progress and evidence handling explicit, producing more coherent tool-calling trajectories and improving benchmark performance across multiple model backends.

Abstract: Deep research requires reasoning over web evidence to answer open-ended questions, and it is a core capability for AI agents. Yet many deep research agents still rely on implicit, unstructured search behavior that causes redundant exploration and brittle evidence aggregation. Motivated by Anthropic’s “think” tool paradigm and insights from the information-retrieval literature, we introduce Q+, a set of query and evidence processing tools that make web search more deliberate by guiding query planning, monitoring search progress, and extracting evidence from long web snapshots. We integrate Q+ into the browser sub-agent of Eigent, an open-source, production-ready multi-agent workforce for computer use, yielding EigentSearch-Q+. Across four benchmarks (SimpleQA-Verified, FRAMES, WebWalkerQA, and X-Bench DeepSearch), Q+ improves Eigent’s browser agent benchmark-size-weighted average accuracy by 3.0, 3.8, and 0.6 percentage points (pp) for GPT-4.1, GPT-5.1, and Minimax M2.5 model backends, respectively. Case studies further suggest that EigentSearch-Q+ produces more coherent tool-calling trajectories by making search progress and evidence handling explicit.

Method: Creates MONETA benchmark with 1,000 European businesses using text sources (websites, Wikipedia, Wikidata) and geospatial data (OpenStreetMap, satellite imagery). Uses training-free MLLM baselines with multi-turn design, context enrichment, and classification explanations.

Result: Baseline MLLMs achieve 62.10-74.10% accuracy. Performance improves up to 22.80% with multi-turn design, context enrichment, and explanation techniques.

Conclusion: Multimodal approaches can effectively automate industry classification without costly manual annotation or model fine-tuning, with significant improvements from enhanced prompting strategies.

Abstract: Industry classification schemes are integral parts of public and corporate databases as they classify businesses based on economic activity. Due to the size of the company registers, manual annotation is costly, and fine-tuning models with every update in industry classification schemes requires significant data collection. We replicate the manual expert verification by using existing or easily retrievable multimodal resources for industry classification. We present MONETA, the first multimodal industry classification benchmark with text (Website, Wikipedia, Wikidata) and geospatial sources (OpenStreetMap and satellite imagery). Our dataset enlists 1,000 businesses in Europe with 20 economic activity labels according to EU guidelines (NACE). Our training-free baseline reaches 62.10% and 74.10% with open and closed-source Multimodal Large Language Models (MLLM). We observe an increase of up to 22.80% with the combination of multi-turn design, context enrichment, and classification explanations. We will release our dataset and the enhanced guidelines.

Method: WorldMAP is a teacher-student framework where a world-model-driven teacher builds semantic-spatial memory from generated videos, grounds task-relevant targets and obstacles, and produces trajectory pseudo-labels through explicit planning. A lightweight student with multi-hypothesis trajectory head is trained to predict navigation trajectories directly from vision-language inputs.

Result: On Target-Bench, WorldMAP achieves best ADE and FDE among compared methods, reducing ADE by 18.0% and FDE by 42.1% relative to best baseline. It lifts a small open-source VLM to DTW performance competitive with proprietary models.

Conclusion: World models’ value in embodied navigation lies less in supplying action-ready imagined evidence than in synthesizing structured supervision for navigation learning. The WorldMAP framework successfully converts world-model-generated futures into effective supervision for training VLMs.

Abstract: Vision-language models (VLMs) and generative world models are opening new opportunities for embodied navigation. VLMs are increasingly used as direct planners or trajectory predictors, while world models support look-ahead reasoning by imagining future views. Yet predicting a reliable trajectory from a single egocentric observation remains challenging. Current VLMs often generate unstable trajectories, and world models, though able to synthesize plausible futures, do not directly provide the grounded signals needed for navigation learning. This raises a central question: how can generated futures be turned into supervision for grounded trajectory prediction? We present WorldMAP, a teacher–student framework that converts world-model-generated futures into persistent semantic-spatial structure and planning-derived supervision. Its world-model-driven teacher builds semantic-spatial memory from generated videos, grounds task-relevant targets and obstacles, and produces trajectory pseudo-labels through explicit planning. A lightweight student with a multi-hypothesis trajectory head is then trained to predict navigation trajectories directly from vision-language inputs. On Target-Bench, WorldMAP achieves the best ADE and FDE among compared methods, reducing ADE by 18.0% and FDE by 42.1% relative to the best competing baseline, while lifting a small open-source VLM to DTW performance competitive with proprietary models. More broadly, the results suggest that, in embodied navigation, the value of world models may lie less in supplying action-ready imagined evidence than in synthesizing structured supervision for navigation learning.

Method: Proposes RAG-XAI framework combining Retrieval-Augmented Generation with eXplainable AI to assess review quality and detect automated patterns using LLM extractor markers. Uses XGBoost, Random Forest, and LightGBM classifiers with feature importance and SHAP analyses.

Result: Achieves 99.61% accuracy, AUC-ROC above 0.999, F1-scores of 0.9925 with extremely low false positive (<0.23%) and false negative rates (~0.8%). RAG component achieves 90.5% top-1 retrieval accuracy with strong same-class clustering.

Conclusion: The framework effectively detects machine-generated reviews and preserves transparency, accountability, and creativity in science by preventing algorithmic bias in peer review systems.

Abstract: The integration of large language models (LLMs) into peer review raises a concern beyond authorship and detection: the potential cascading automation of the entire editorial process. As reviews become partially or fully machine-generated, it becomes plausible that editorial decisions may also be delegated to algorithmic systems, leading to a fully automated evaluation pipeline. They risk reshaping the criteria by which scientific work is assessed. This paper argues that machine-driven assessment may systematically favor standardized, pattern-conforming research while penalizing unconventional and paradigm-shifting ideas that require contextual human judgment. We consider that this shift could lead to epistemic homogenization, where researchers are implicitly incentivized to optimize their work for algorithmic approval rather than genuine discovery. To address this risk, we introduce an explainable framework (RAG-XAI) for assessing review quality and detecting automated patterns using markers LLM extractor, aiming to preserve transparency, accountability and creativity in science. The proposed framework achieves near-perfect detection performance, with XGBoost, Random Forest and LightGBM reaching 99.61% accuracy, AUC-ROC above 0.999 and F1-scores of 0.9925 on the test set, while maintaining extremely low false positive rates (<0.23%) and false negative rates (~0.8%). In contrast, the logistic regression baseline performs substantially worse (89.97% accuracy, F1-score 0.8314). Feature importance and SHAP analyses identify absence of personal signals and repetition patterns as the dominant predictors. Additionally, the RAG component achieves 90.5% top-1 retrieval accuracy, with strong same-class clustering in the embedding space, further supporting the reliability of the framework’s outputs.

Method: Constructed a dataset of 5,037 high-quality goal-oriented navigation samples with 3D vertical actions and urban semantics, then evaluated 17 representative models including non-reasoning LMMs, reasoning LMMs, agent-based methods, and vision-language-action models.

Result: Current LMMs exhibit emerging action capabilities but remain far from human-level performance. Navigation errors diverge rapidly from destination after critical decision bifurcations rather than accumulating linearly.

Conclusion: LMMs have limited spatial action capabilities for embodied navigation, with critical decision points being particularly challenging. Four promising improvement directions are identified: geometric perception, cross-view understanding, spatial imagination, and long-term memory.

Abstract: Large multimodal models (LMMs) show strong visual-linguistic reasoning but their capacity for spatial decision-making and action remains unclear. In this work, we investigate whether LMMs can achieve embodied spatial action like human through a challenging scenario: goal-oriented navigation in urban 3D spaces. We first spend over 500 hours constructing a dataset comprising 5,037 high-quality goal-oriented navigation samples, with an emphasis on 3D vertical actions and rich urban semantic information. Then, we comprehensively assess 17 representative models, including non-reasoning LMMs, reasoning LMMs, agent-based methods, and vision-language-action models. Experiments show that current LMMs exhibit emerging action capabilities, yet remain far from human-level performance. Furthermore, we reveal an intriguing phenomenon: navigation errors do not accumulate linearly but instead diverge rapidly from the destination after a critical decision bifurcation. The limitations of LMMs are investigated by analyzing their behavior at these critical decision bifurcations. Finally, we experimentally explore four promising directions for improvement: geometric perception, cross-view understanding, spatial imagination, and long-term memory. The project is available at: https://github.com/serenditipy-AC/Embodied-Navigation-Bench.

Method: Systematic evaluation of recent CX generation methods on their ability to provide valid and plausible counterfactuals when inputs are incomplete, comparing robust vs non-robust methods.

Result: Robust CX methods achieve higher validity than non-robust ones, but all methods struggle to find valid counterfactuals when dealing with incomplete inputs.

Conclusion: Current CX methods are inadequate for handling incomplete inputs, motivating the need for new methods specifically designed to address missing data in counterfactual explanation generation.

Abstract: Existing algorithms for generating Counterfactual Explanations (CXs) for Machine Learning (ML) typically assume fully specified inputs. However, real-world data often contains missing values, and the impact of these incomplete inputs on the performance of existing CX methods remains unexplored. To address this gap, we systematically evaluate recent CX generation methods on their ability to provide valid and plausible counterfactuals when inputs are incomplete. As part of this investigation, we hypothesize that robust CX generation methods will be better suited to address the challenge of providing valid and plausible counterfactuals when inputs are incomplete. Our findings reveal that while robust CX methods achieve higher validity than non-robust ones, all methods struggle to find valid counterfactuals. These results motivate the need for new CX methods capable of handling incomplete inputs.

Method: Establishes two-stage definition (Hypothesis Generation and Selection), creates comprehensive taxonomy of literature, conducts benchmark study of LLMs on abductive tasks, and analyzes performance across model characteristics.

Result: Reveals critical gaps in current approaches including static benchmark design, narrow domain coverage, limited training frameworks, and insufficient mechanistic understanding of abductive processes in LLMs.

Conclusion: First comprehensive survey of abductive reasoning in LLMs providing unified framework, taxonomy, and empirical benchmarking, highlighting research gaps and future directions.

Abstract: Regardless of its foundational role in human discovery and sense-making, abductive reasoning–the inference of the most plausible explanation for an observation–has been relatively underexplored in Large Language Models (LLMs). Despite the rapid advancement of LLMs, the exploration of abductive reasoning and its diverse facets has thus far been disjointed rather than cohesive. This paper presents the first survey of abductive reasoning in LLMs, tracing its trajectory from philosophical foundations to contemporary AI implementations. To address the widespread conceptual confusion and disjointed task definitions prevalent in the field, we establish a unified two-stage definition that formally categorizes prior work. This definition disentangles abduction into \textit{Hypothesis Generation}, where models bridge epistemic gaps to produce candidate explanations, and \textit{Hypothesis Selection}, where the generated candidates are evaluated and the most plausible explanation is chosen. Building upon this foundation, we present a comprehensive taxonomy of the literature, categorizing prior work based on their abductive tasks, datasets, underlying methodologies, and evaluation strategies. In order to ground our framework empirically, we conduct a compact benchmark study of current LLMs on abductive tasks, together with targeted comparative analyses across model sizes, model families, evaluation styles, and the distinct generation-versus-selection task typologies. Moreover, by synthesizing recent empirical results, we examine how LLM performance on abductive reasoning relates to deductive and inductive tasks, providing insights into their broader reasoning capabilities. Our analysis reveals critical gaps in current approaches–from static benchmark design and narrow domain coverage to narrow training frameworks and limited mechanistic understanding of abductive processes…

Method: Proposed a method for generating contrastive explanations that compare a system’s proposed solution against relevant alternatives. Developed a framework using visual and textual cues to highlight key objectives from a state-of-the-art collision avoidance system. Conducted an exploratory user study with four experienced marine officers.

Result: Contrastive explanations supported understanding of system objectives. However, while valuable in complex multi-vessel encounters, they increased cognitive workload. Findings suggest future maritime interfaces may benefit from demand-driven or scenario-specific explanation strategies.

Conclusion: Contrastive explanations can enhance transparency in maritime collision avoidance systems but need careful implementation to manage cognitive load. Future work should focus on adaptive explanation strategies based on scenario complexity and user needs.

Abstract: Automated maritime collision avoidance will rely on human supervision for the foreseeable future. This necessitates transparency into how the system perceives a scenario and plans a maneuver. However, the causal logic behind avoidance maneuvers is often complex and difficult to convey to a navigator. This paper explores how to explain these factors in a selective, understandable manner for supervisors with a nautical background. We propose a method for generating contrastive explanations, which provide human-centric insights by comparing a system’s proposed solution against relevant alternatives. To evaluate this, we developed a framework that uses visual and textual cues to highlight key objectives from a state-of-the-art collision avoidance system. An exploratory user study with four experienced marine officers suggests that contrastive explanations support the understanding of the system’s objectives. However, our findings also reveal that while these explanations are highly valuable in complex multi-vessel encounters, they can increase cognitive workload, suggesting that future maritime interfaces may benefit most from demand-driven or scenario-specific explanation strategies.

Method: The benchmark uses a unified Learning/Priming-Interfere-Test protocol with 300 items across three cognitively grounded constructs: Procedural Memory (one-shot skill acquisition after interference), Priming (theme-driven bias via paired experimental/control instances), and Classical Conditioning (CS-US associations shaping first decisions).

Result: Evaluation of 17 models shows severe limitations: no model exceeds 66% overall, with top performers DeepSeek-R1 (65.3%), Qwen3-32B (64.1%), and GPT-5 (63.0%) far below human baselines. Analysis reveals dramatic asymmetries (inhibition 17.6% vs. preference 75.0%) and universal bottlenecks.

Conclusion: ImplicitMemBench reframes evaluation from “what agents recall” to “what they automatically enact,” highlighting the need for architectural innovations beyond parameter scaling to address fundamental limitations in implicit memory capabilities.

Abstract: Existing memory benchmarks for LLM agents evaluate explicit recall of facts, yet overlook implicit memory where experience becomes automated behavior without conscious retrieval. This gap is critical: effective assistants must automatically apply learned procedures or avoid failed actions without explicit reminders. We introduce ImplicitMemBench, the first systematic benchmark evaluating implicit memory through three cognitively grounded constructs drawn from standard cognitive-science accounts of non-declarative memory: Procedural Memory (one-shot skill acquisition after interference), Priming (theme-driven bias via paired experimental/control instances), and Classical Conditioning (Conditioned Stimulus–Unconditioned Stimulus (CS–US) associations shaping first decisions). Our 300-item suite employs a unified Learning/Priming-Interfere-Test protocol with first-attempt scoring. Evaluation of 17 models reveals severe limitations: no model exceeds 66% overall, with top performers DeepSeek-R1 (65.3%), Qwen3-32B (64.1%), and GPT-5 (63.0%) far below human baselines. Analysis uncovers dramatic asymmetries (inhibition 17.6% vs. preference 75.0%) and universal bottlenecks requiring architectural innovations beyond parameter scaling. ImplicitMemBench reframes evaluation from “what agents recall” to “what they automatically enact”.

Method: Proposes Revise framework with: 1) comprehensive hierarchical taxonomy of common OCR errors, 2) synthetic data generation strategy to realistically simulate OCR errors, 3) correction model trained on this data to fix errors at character, word, and structural levels.

Result: Revise effectively corrects OCR outputs, enabling more structured representation and systematic management of document contents, significantly enhancing downstream performance in document retrieval and question answering tasks.

Conclusion: The framework overcomes structural management limitations of existing Document AI frameworks by addressing OCR error correction systematically, demonstrating potential for improved document understanding and organization.

Abstract: Recent advances in Large Language Models (LLMs) have significantly improved the field of Document AI, demonstrating remarkable performance on document understanding tasks such as question answering. However, existing approaches primarily focus on solving specific tasks, lacking the capability to structurally organize and manage document information. To address this limitation, we propose Revise, a framework that systematically corrects errors introduced by OCR at the character, word, and structural levels. Specifically, Revise employs a comprehensive hierarchical taxonomy of common OCR errors and a synthetic data generation strategy that realistically simulates such errors to train an effective correction model. Experimental results demonstrate that Revise effectively corrects OCR outputs, enabling more structured representation and systematic management of document contents. Consequently, our method significantly enhances downstream performance in document retrieval and question answering tasks, highlighting the potential to overcome the structural management limitations of existing Document AI frameworks.

Method: Extract empirical knowledge through contrastive analysis and multi-level clustering to transform raw reasoning trajectories into hierarchical experience knowledge. Use experience-aligned training to regularize stochastic exploration into strategic, experience-driven search.

Result: Extensive evaluations on multiple complex agentic search and mathematical reasoning benchmarks show substantial performance gains and strong cross-task/cross-algorithm generalization.

Conclusion: HiExp framework effectively enhances RL-based search agents by leveraging hierarchical experience knowledge, improving both performance and training stability for complex reasoning tasks.

Abstract: Reinforcement learning (RL) has become an effective approach for advancing the reasoning capabilities of large language models (LLMs) through the strategic integration of external search engines. However, current RL-based search agents often rely on a process of stochastic exploration guided by carefully crafted outcome rewards, leading to inefficient reasoning trajectories and unstable training. To address these issues, we propose a novel framework, Hierarchical Experience (HiExp), to enhance the performance and training stability of search agents. Specifically, we extract empirical knowledge through contrastive analysis and a multi-level clustering mechanism, transforming raw reasoning trajectories into hierarchical experience knowledge. By leveraging experience-aligned training, we effectively regularize stochastic exploration, evolving it into a strategic and experience-driven search process. Extensive evaluations on multiple complex agentic search and mathematical reasoning benchmarks demonstrate that our approach not only achieves substantial performance gains but also exhibits strong cross-task and cross-algorithm generalization.

Method: Three activation steering methods: Steer-With-Fixed-Coeff (SwFC) applies uniform additive steering; Steer-to-Target-Projection (StTP) and Steer-to-Mirror-Projection (StMP) use logistic regression decision boundaries to selectively intervene only on tokens whose activations fall below distributional thresholds.

Result: All methods substantially recover target traits (honesty and compassion) while preserving coherence. StTP and StMP better maintain general capabilities (MMLU, MT-Bench, AlpacaEval) and produce less repetition in multi-turn conversations.

Conclusion: Activation steering serves as a lightweight runtime defense that continuously corrects misaligned activations throughout generation, with projection-aware methods (StTP, StMP) offering better balance between alignment correction and capability preservation.

Abstract: Alignment in LLMs is more brittle than commonly assumed: misalignment can be triggered by adversarial prompts, benign fine-tuning, emergent misalignment, and goal misgeneralization. Recent evidence suggests that some misalignment behaviors are encoded as linear structure in activation space, making it tractable via steering, while safety alignment has been shown to govern the first few output tokens primarily, leaving subsequent generation unguarded. These findings motivate activation steering as a lightweight runtime defense that continuously corrects misaligned activations throughout generation. We evaluate three methods: Steer-With-Fixed-Coeff (SwFC), which applies uniform additive steering, and two novel projection-aware methods, Steer-to-Target-Projection (StTP) and Steer-to-Mirror-Projection (StMP), that use a logistic regression decision boundary to selectively intervene only on tokens whose activations fall below distributional thresholds. Using malicious system prompts as a controlled proxy for misalignment, we evaluate under two threat models (dishonesty and dismissiveness) and two architectures (Llama-3.3-70B-Instruct, Qwen3-32B). All methods substantially recover target traits (honesty and compassion) while preserving coherence. StTP and StMP better maintain general capabilities (MMLU, MT-Bench, AlpacaEval) and produce less repetition in multi-turn conversations.

Method: Created a trajectory-level preference benchmark covering four task families: Safety Refusal, Tool-Irrelevance/Unavailability, Complex Planning, and Robust Error Recovery. Constructed positive trajectories and hard negatives using multi-model natural rollouts, rule-based perturbations, and minimal-edit LLM perturbations.

Result: Benchmarked representative RMs (generative, discriminative, and LLM-as-Judge) showing all three evaluator families face substantial challenges, with performance degrading sharply on long-horizon trajectories.

Conclusion: Plan-RewardBench serves as both a practical evaluation suite and reusable blueprint for constructing agentic planning preference data, highlighting the need for specialized training in agentic, trajectory-level reward modeling.

Abstract: In classical Reinforcement Learning from Human Feedback (RLHF), Reward Models (RMs) serve as the fundamental signal provider for model alignment. As Large Language Models evolve into agentic systems capable of autonomous tool invocation and complex reasoning, the paradigm of reward modeling faces unprecedented challenges–most notably, the lack of benchmarks specifically designed to assess RM capabilities within tool-integrated environments. To address this gap, we present Plan-RewardBench, a trajectory-level preference benchmark designed to evaluate how well judges distinguish preferred versus distractor agent trajectories in complex tool-using scenarios. Plan-RewardBench covers four representative task families – (i) Safety Refusal, (ii) Tool-Irrelevance / Unavailability, (iii) Complex Planning, and (iv) Robust Error Recovery – comprising validated positive trajectories and confusable hard negatives constructed via multi-model natural rollouts, rule-based perturbations, and minimal-edit LLM perturbations. We benchmark representative RMs (generative, discriminative, and LLM-as-Judge) under a unified pairwise protocol, reporting accuracy trends across varying trajectory lengths and task categories. Furthermore, we provide diagnostic analyses of prevalent failure modes. Our results reveal that all three evaluator families face substantial challenges, with performance degrading sharply on long-horizon trajectories, underscoring the necessity for specialized training in agentic, trajectory-level reward modeling. Ultimately, Plan-RewardBench aims to serve as both a practical evaluation suite and a reusable blueprint for constructing agentic planning preference data.

Method: Introduces the Clinical World Model formalizing care as tripartite interaction among Patient, Provider, and Ecosystem. Develops parallel decision-making architectures for providers, patients, and AI agents grounded in clinical cognition principles. Creates Clinical AI Skill-Mix with 8 dimensions: 5 defining clinical competency space (condition, phase, care setting, provider role, task) and 3 specifying AI-human reasoning engagement (assigned authority, agent facing, anchoring layer).

Result: The combinatorial product of the 8 dimensions yields a space of billions of distinct competency coordinates, making the competency space irreducible. Validation within one coordinate provides minimal evidence for performance in another. The framework provides a common grammar for specifying, evaluating, and bounding clinical AI across stakeholders.

Conclusion: The Clinical World Model reframes the central question from whether AI works to in which competency coordinates reliability has been demonstrated, and for whom. It makes explicit the structure needed for clinical AI to be properly specified, evaluated, and bounded.

Abstract: The competency of any intelligent agent is bounded by its formal account of the world in which it operates. Clinical AI lacks such an account. Existing frameworks address evaluation, regulation, or system design in isolation, without a shared model of the clinical world to connect them. We introduce the Clinical World Model, a framework that formalizes care as a tripartite interaction among Patient, Provider, and Ecosystem. To formalize how any agent, whether human or artificial, transforms information into clinical action, we develop parallel decision-making architectures for providers, patients, and AI agents, grounded in validated principles of clinical cognition. The Clinical AI Skill-Mix operationalizes competency through eight dimensions. Five define the clinical competency space (condition, phase, care setting, provider role, and task) and three specify how AI engages human reasoning (assigned authority, agent facing, and anchoring layer). The combinatorial product of these dimensions yields a space of billions of distinct competency coordinates. A central structural implication is that validation within one coordinate provides minimal evidence for performance in another, rendering the competency space irreducible. The framework supplies a common grammar through which clinical AI can be specified, evaluated, and bounded across stakeholders. By making this structure explicit, the Clinical World Model reframes the field’s central question from whether AI works to in which competency coordinates reliability has been demonstrated, and for whom.

Method: Proposes HiRO-Nav agent that monitors its own action entropy to determine when to perform thinking. Uses hybrid supervised fine-tuning followed by online reinforcement learning with hybrid reasoning strategy. Only activates reasoning for high-entropy actions where the agent is uncertain.

Result: Achieves better trade-off between success rates and token efficiency than both dense-thinking (always reason) and no-thinking baselines on the CHORES-S ObjectNav benchmark. Reduces computational overhead while improving decision quality.

Conclusion: The hybrid reasoning approach enables efficient long-horizon navigation by activating reasoning only when needed (high-entropy actions), balancing computational efficiency with task performance.

Abstract: Embodied navigation agents built upon large reasoning models (LRMs) can handle complex, multimodal environmental input and perform grounded reasoning per step to improve sequential decision-making for long-horizon tasks. However, a critical question remains: \textit{how can the reasoning capabilities of LRMs be harnessed intelligently and efficiently for long-horizon navigation tasks?} In simple scenes, agents are expected to act reflexively, while in complex ones they should engage in deliberate reasoning before acting.To achieve this, we introduce \textbf{H}ybr\textbf{i}d \textbf{R}eas\textbf{O}ning \textbf{Nav}igation (\textbf{HiRO-Nav}) agent, the first kind of agent capable of adaptively determining whether to perform thinking at every step based on its own action entropy. Specifically, by examining how the agent’s action entropy evolves over the navigation trajectories, we observed that only a small fraction of actions exhibit high entropy, and these actions often steer the agent toward novel scenes or critical objects. Furthermore, studying the relationship between action entropy and task completion (i.e., Q-value) reveals that improving high-entropy actions contributes more positively to task success.Hence, we propose a tailored training pipeline comprising hybrid supervised fine-tuning as a cold start, followed by online reinforcement learning with the proposed hybrid reasoning strategy to explicitly activate reasoning only for high-entropy actions, significantly reducing computational overhead while improving decision quality. Extensive experiments on the \textsc{CHORES}-$\mathbb{S}$ ObjectNav benchmark showcases that HiRO-Nav achieves a better trade-off between success rates and token efficiency than both dense-thinking and no-thinking baselines.

Method: Proposes Meta-Principle Physics Architecture (MPPA) with three core components: Gravitator (implements Connectivity via causal attention), Energy Encoder (implements Conservation via log-domain energy tracking), and Periodicity Encoder (implements Periodicity via FFT-based spectral analysis). These components collaborate via learnable independent gating fusion mechanism.

Result: Significant improvements: physical reasoning from near zero to 0.436, 2.18x mathematical task improvement (0.330 vs 0.151), 52% logical task gain (0.456 vs 0.300), 3.69% lower validation perplexity (259.45 vs 269.40) with only 11.8% more parameters. Strong generalization on out-of-distribution physical scenarios.

Conclusion: MPPA establishes a new theoretical foundation for next-generation AI with physical common sense, causal reasoning, and mathematical rigor through principle-embedded design that enhances robustness and interpretability.

Abstract: The essence of current neural network architectures is phenomenological fitting: they learn input-output statistical correlations via massive parameters and data, yet lack intrinsic understanding of the fundamental principles governing physical reality. This paper proposes a paradigm leap from pure phenomenological fitting to the fusion of phenomenological fitting and endogenous deduction. By embedding physical meta-principles into neural network architecture, we construct the Meta-Principle Physics Architecture (MPPA). Specifically, MPPA embeds three core meta-principles - Connectivity, Conservation, Periodicity - into its architecture, implemented via three core components: the Gravitator realizes Connectivity via standard causal attention; the Energy Encoder implements Conservation via log-domain energy tracking and delayed compensation; the Periodicity Encoder fulfills Periodicity via FFT-based spectral analysis and delayed modulation. These components collaborate via a learnable independent gating fusion mechanism, forming a complete physical cognition framework of ’local relational connectivity - global conservation constraint - evolutionary periodic law’. Experiments show MPPA achieves significant improvements: physical reasoning (from near zero to 0.436, 0.436 vs 0.000), 2.18x mathematical task improvement (0.330 vs 0.151), 52% logical task gain (0.456 vs 0.300), and 3.69% lower validation perplexity (259.45 vs 269.40), with only 11.8% more parameters (242.40M vs 216.91M). Notably, MPPA shows strong generalization on out-of-distribution physical scenarios, proving the robustness and interpretability of this principle-embedded design. This work establishes a new theoretical foundation and technical path for next-generation AI with physical common sense, causal reasoning, and mathematical rigor.

Method: Proposes Responsible-DKT, a neural-symbolic approach that integrates symbolic educational knowledge (mastery and non-mastery rules) into sequential neural models. This combines the predictive power of neural networks with the interpretability and pedagogical alignment of symbolic reasoning.

Result: Outperforms both neural-symbolic baseline and fully data-driven PyTorch DKT model, achieving over 0.80 AUC with only 10% of training data and up to 0.90 AUC (13% improvement). Shows improved temporal reliability with lower early/mid-sequence errors and lowest prediction inconsistency rates. Provides intrinsic interpretability via grounded computation graphs.

Conclusion: Neural-symbolic approaches enhance both performance and interpretability in learner modeling, mitigate data limitations, and support more responsible, human-centered AI in education by enabling pedagogical assumption evaluation and transparent reasoning.

Abstract: The growing use of artificial intelligence (AI) in education, particularly large language models (LLMs), has increased interest in intelligent tutoring systems. However, LLMs often show limited adaptivity and struggle to model learners’ evolving knowledge over time, highlighting the need for dedicated learner modelling approaches. Although deep knowledge tracing methods achieve strong predictive performance, their opacity and susceptibility to bias can limit alignment with pedagogical principles. To address this, we propose Responsible-DKT, a neural-symbolic deep knowledge tracing approach that integrates symbolic educational knowledge (e.g., mastery and non-mastery rules) into sequential neural models for responsible learner modelling. Experiments on a real-world dataset of students’ math interactions show that Responsible-DKT outperforms both a neural-symbolic baseline and a fully data-driven PyTorch DKT model across training settings. The model achieves over 0.80 AUC with only 10% of training data and up to 0.90 AUC, improving performance by up to 13%. It also demonstrates improved temporal reliability, producing lower early- and mid-sequence prediction errors and the lowest prediction inconsistency rates across sequence lengths, indicating that prediction updates remain directionally aligned with observed student responses over time. Furthermore, the neural-symbolic approach offers intrinsic interpretability via a grounded computation graph that exposes the logic behind each prediction, enabling both local and global explanations. It also allows empirical evaluation of pedagogical assumptions, revealing that repeated incorrect responses (non-mastery) strongly influence prediction updates. These results indicate that neural-symbolic approaches enhance both performance and interpretability, mitigate data limitations, and support more responsible, human-centered AI in education.

Method: Proposes Asymmetric Collaborative Framework (ACF) that structurally decouples covert communication from semantic reasoning via orthogonal statistical and cognitive layers. Uses prefix-independent decoding paradigm governed by shared steganographic configuration to eliminate reliance on cognitive symmetry.

Result: Under severe cognitive asymmetry, symmetric baselines suffer severe channel degradation while ACF uniquely excels across both semantic fidelity and covert communication. Maintains computational indistinguishability with provable error bounds and robust Effective Information Capacity guarantees.

Conclusion: ACF provides a robust solution for covert communication in modern agent networks by addressing the fundamental challenge of cognitive asymmetry through structural decoupling of communication and reasoning layers.

Abstract: As generative artificial intelligence evolves, autonomous agent networks present a powerful paradigm for interactive covert communication. However, because agents dynamically update internal memories via environmental interactions, existing methods face a critical structural vulnerability: cognitive asymmetry. Conventional approaches demand strict cognitive symmetry, requiring identical sequence prefixes between the encoder and decoder. In dynamic deployments, inevitable prefix discrepancies destroy synchronization, inducing severe channel degradation. To address this core challenge of cognitive asymmetry, we propose the Asymmetric Collaborative Framework (ACF), which structurally decouples covert communication from semantic reasoning via orthogonal statistical and cognitive layers. By deploying a prefix-independent decoding paradigm governed by a shared steganographic configuration, ACF eliminates the reliance on cognitive symmetry. Evaluations on realistic memory-augmented workflows demonstrate that under severe cognitive asymmetry, symmetric baselines suffer severe channel degradation, whereas ACF uniquely excels across both semantic fidelity and covert communication. It maintains computational indistinguishability, enabling reliable secret extraction with provable error bounds, and providing robust Effective Information Capacity guarantees for modern agent networks.

Method: Proposes U-CECE, a model-agnostic multi-resolution framework with three levels: atomic concepts for broad explanations, relational sets-of-sets for simple interactions, and structural graphs for full semantic structure. At structural level, offers both transductive mode (supervised GNNs) and inductive mode (unsupervised graph autoencoders).

Result: Experiments on CUB and Visual Genome datasets characterize efficiency-expressivity trade-off across levels. Human surveys and LVLM-based evaluation show retrieved structural counterfactuals are semantically equivalent to, and often preferred over, exact GED-based ground-truth explanations.

Conclusion: U-CECE provides a unified framework that adapts to data regime and compute budget while offering multiple levels of expressivity for conceptual counterfactual explanations, effectively balancing the trade-off between efficiency and expressivity.

Abstract: As AI models grow more complex, explainability is essential for building trust, yet concept-based counterfactual methods still face a trade-off between expressivity and efficiency. Representing underlying concepts as atomic sets is fast but misses relational context, whereas full graph representations are more faithful but require solving the NP-hard Graph Edit Distance (GED) problem. We propose U-CECE, a unified, model-agnostic multi-resolution framework for conceptual counterfactual explanations that adapts to data regime and compute budget. U-CECE spans three levels of expressivity: atomic concepts for broad explanations, relational sets-of-sets for simple interactions, and structural graphs for full semantic structure. At the structural level, both a precision-oriented transductive mode based on supervised Graph Neural Networks (GNNs) and a scalable inductive mode based on unsupervised graph autoencoders (GAEs) are supported. Experiments on the structurally divergent CUB and Visual Genome datasets characterize the efficiency-expressivity trade-off across levels, while human surveys and LVLM-based evaluation show that the retrieved structural counterfactuals are semantically equivalent to, and often preferred over, exact GED-based ground-truth explanations.

Method: Created ProMedical-Preference-50k dataset via human-in-the-loop pipeline with physician-derived rubrics. Developed Explicit Criteria Injection paradigm to train multi-dimensional reward model that disentangles safety constraints from general proficiency. Used GRPO reinforcement learning guided by the reward model.

Result: Qwen3-8B model optimized with ProMedical-RM showed 22.3% improvement in overall accuracy and 21.7% improvement in safety compliance, rivaling proprietary frontier models. Generalizes well to external benchmarks like UltraMedical.

Conclusion: Fine-grained clinical criteria alignment significantly improves LLM performance in medical domains. The framework enables precise safety-aware alignment and generalizes effectively across medical benchmarks.

Abstract: Aligning Large Language Models (LLMs) with high-stakes medical standards remains a significant challenge, primarily due to the dissonance between coarse-grained preference signals and the complex, multi-dimensional nature of clinical protocols. To bridge this gap, we introduce ProMedical, a unified alignment framework grounded in fine-grained clinical criteria. We first construct ProMedical-Preference-50k, a dataset generated via a human-in-the-loop pipeline that augments medical instructions with rigorous, physician-derived rubrics. Leveraging this corpus, we propose the Explicit Criteria Injection paradigm to train a multi-dimensional reward model. Unlike traditional scalar reward models, our approach explicitly disentangles safety constraints from general proficiency, enabling precise guidance during reinforcement learning. To rigorously validate this framework, we establish ProMedical-Bench, a held-out evaluation suite anchored by double-blind expert adjudication. Empirical evaluations demonstrate that optimizing the Qwen3-8B base model via ProMedical-RM-guided GRPO yields substantial gains, improving overall accuracy by 22.3% and safety compliance by 21.7%, effectively rivaling proprietary frontier models. Furthermore, the aligned policy generalizes robustly to external benchmarks, demonstrating performance comparable to state-of-the-art models on UltraMedical. We publicly release our datasets, reward models, and benchmarks to facilitate reproducible research in safety-aware medical alignment.

Method: Parallel-group randomized experiment with 71 university students completing collaborative tasks with GenAI either available or unavailable. Statistical analysis of human discourse to examine differences in collaborative regulation across regulatory modes, processes, and participatory focuses.

Result: GenAI availability shifted regulation away from predominantly socially shared forms towards more hybrid co-regulatory forms, with selective increases in directive, obstacle-oriented, and affective regulatory processes. Participatory-focus distributions were broadly similar across conditions.

Conclusion: GenAI reshapes the distribution of regulatory responsibility in collaboration, offering implications for human-centred design of AI-supported collaborative learning.

Abstract: Generative AI (GenAI) is increasingly used in collaborative learning, yet its effects on how groups regulate collaboration remain unclear. Effective collaboration depends not only on what groups discuss, but on how they jointly manage goals, participation, strategy use, monitoring, and repair through co-regulation and socially shared regulation. We compared collaborative regulation between Human-AI and Human-Human groups in a parallel-group randomised experiment with 71 university students completing the same collaborative tasks with GenAI either available or unavailable. Focusing on human discourse, we used statistical analyses to examine differences in the distribution of collaborative regulation across regulatory modes, regulatory processes, and participatory focuses. Results showed that GenAI availability shifted regulation away from predominantly socially shared forms towards more hybrid co-regulatory forms, with selective increases in directive, obstacle-oriented, and affective regulatory processes. Participatory-focus distributions, however, were broadly similar across conditions. These findings suggest that GenAI reshapes the distribution of regulatory responsibility in collaboration and offer implications for the human-centred design of AI-supported collaborative learning.

Method: Replace discrete latent variables with natural language conditioning via text-conditioned VAE. Use LLM as dynamic semantic operator at test time to remap current observation descriptions to align with source task semantics. Source-aligned caption conditions VAE to generate imagined state compatible with original training.

Result: Achieves zero-shot transfer across broad spectrum of complex and truly novel analogous tasks, moving beyond limitations of fixed category mappings.

Conclusion: LLMs’ flexible reasoning capabilities enable more generalized zero-shot transfer in RL by semantically aligning novel tasks with source tasks, allowing direct policy reuse without retraining.

Abstract: Reinforcement Learning (RL) agents often struggle to generalize knowledge to new tasks, even those structurally similar to ones they have mastered. Although recent approaches have attempted to mitigate this issue via zero-shot transfer, they are often constrained by predefined, discrete class systems, limiting their adaptability to novel or compositional task variations. We propose a significantly more generalized approach, replacing discrete latent variables with natural language conditioning via a text-conditioned Variational Autoencoder (VAE). Our core innovation utilizes a Large Language Model (LLM) as a dynamic \textit{semantic operator} at test time. Rather than relying on rigid rules, our agent queries the LLM to semantically remap the description of the current observation to align with the source task. This source-aligned caption conditions the VAE to generate an imagined state compatible with the agent’s original training, enabling direct policy reuse. By harnessing the flexible reasoning capabilities of LLMs, our approach achieves zero-shot transfer across a broad spectrum of complex and truly novel analogous tasks, moving beyond the limitations of fixed category mappings. Code and videos are available \href{https://anonymous.4open.science/r/ASPECT-85C3/}{here}.

Method: SkillClaw continuously aggregates user interaction trajectories, processes them with an autonomous evolver that identifies recurring behavioral patterns, and translates these into skill updates by refining existing skills or adding new capabilities. Updated skills are maintained in a shared repository and synchronized across users.

Result: Experiments on WildClawBench show that with limited interaction and feedback, SkillClaw significantly improves the performance of Qwen3-Max in real-world agent scenarios, enabling cross-user knowledge transfer and cumulative capability improvement.

Conclusion: SkillClaw provides a framework for collective skill evolution in multi-user agent ecosystems, converting heterogeneous user experiences into reliable skill updates that propagate system-wide without requiring additional user effort.

Abstract: Large language model (LLM) agents such as OpenClaw rely on reusable skills to perform complex tasks, yet these skills remain largely static after deployment. As a result, similar workflows, tool usage patterns, and failure modes are repeatedly rediscovered across users, preventing the system from improving with experience. While interactions from different users provide complementary signals about when a skill works or fails, existing systems lack a mechanism to convert such heterogeneous experiences into reliable skill updates. To address these issues, we present SkillClaw, a framework for collective skill evolution in multi-user agent ecosystems, which treats cross-user and over-time interactions as the primary signal for improving skills. SkillClaw continuously aggregates trajectories generated during use and processes them with an autonomous evolver, which identifies recurring behavioral patterns and translates them into updates to the skill set by refining existing skills or extending them with new capabilities. The resulting skills are maintained in a shared repository and synchronized across users, allowing improvements discovered in one context to propagate system-wide while requiring no additional effort from users. By integrating multi-user experience into ongoing skill updates, SkillClaw enables cross-user knowledge transfer and cumulative capability improvement, and experiments on WildClawBench show that limited interaction and feedback, it significantly improves the performance of Qwen3-Max in real-world agent scenarios.

Method: Used Goedel-Prover-V2 model heavily trained on 1.8M formal-math examples, then fine-tuned specialized model on small amount (100 traces) of Lean-specific tool-use data to test reversibility of agentic collapse.

Result: Tool-calling accuracy dropped from 89.4% to near 0% after domain specialization, but recovered to 83.8% on Berkeley Function Calling Leaderboard with just 100 Lean-specific traces, and improved ProofNet pass@32 from 21.51% to 25.81%.

Conclusion: Heavy domain fine-tuning suppresses but doesn’t permanently erase general tool-use capabilities; small amounts of domain-specific agentic data can awaken dormant capabilities with strong transfer beyond the training domain.

Abstract: Heavy supervised fine-tuning on a target domain can strongly suppress capabilities that were present in the base model. We study this phenomenon in formal mathematics using Goedel-Prover-V2, an open-source model heavily trained on 1.8 million formal-math examples. After domain specialization, the model almost completely loses its ability to produce valid tool calls, even when explicitly instructed to use tools, dropping from 89.4% function-calling accuracy in the base model to nearly 0%. We ask whether this agentic collapse is permanent or instead reversible. To answer this question, we fine-tune the specialized model on a small amount of Lean-specific tool-use data. Remarkably, as few as 100 agentic traces are sufficient to restore strong tool-calling behavior. Importantly, this recovery is not the result of reward hacking or benchmark-specific optimization: the recovery data is entirely drawn from the Lean setting, where the model uses natural-language queries to search the Mathlib library for relevant theorems and lemmas, yet the regained capability transfers well beyond that domain. In particular, these same 100 Lean-specific traces improve performance on the Berkeley Function Calling Leaderboard from near zero to 83.8%, approaching the base model’s 89.4% despite the mismatch in task distribution and protocol. The recovered capability is also practically useful in-domain. On ProofNet, pass@32 improves from 21.51% to 25.81%. Together, these results show that heavy domain supervised fine-tuning can suppress general tool-use ability without permanently erasing it, and that a small amount of domain-specific agentic data can awaken dormant tool-use capabilities.

Method: Proposes SAVeR framework that: 1) structurally generates persona-based diverse candidate beliefs for selection under faithfulness-relevant structure space, 2) performs adversarial auditing to localize violations, and 3) repairs through constraint-guided minimal interventions under verifiable acceptance criteria.

Result: Extensive experiments on six benchmark datasets demonstrate that the approach consistently improves reasoning faithfulness while preserving competitive end-task performance.

Conclusion: SAVeR provides an effective framework for enforcing verification over internal belief states in LLM agents to achieve faithful reasoning, addressing the problem of systematic behavioral drift caused by unfaithful intermediate reasoning trajectories.

Abstract: In large language model (LLM) agents, reasoning trajectories are treated as reliable internal beliefs for guiding actions and updating memory. However, coherent reasoning can still violate logical or evidential constraints, allowing unsupported beliefs repeatedly stored and propagated across decision steps, leading to systematic behavioral drift in long-horizon agentic systems. Most existing strategies rely on the consensus mechanism, conflating agreement with faithfulness. In this paper, inspired by the vulnerability of unfaithful intermediate reasoning trajectories, we propose \textbf{S}elf-\textbf{A}udited \textbf{Ve}rified \textbf{R}easoning (\textsc{SAVeR}), a novel framework that enforces verification over internal belief states within the agent before action commitment, achieving faithful reasoning. Concretely, we structurally generate persona-based diverse candidate beliefs for selection under a faithfulness-relevant structure space. To achieve reasoning faithfulness, we perform adversarial auditing to localize violations and repair through constraint-guided minimal interventions under verifiable acceptance criteria. Extensive experiments on six benchmark datasets demonstrate that our approach consistently improves reasoning faithfulness while preserving competitive end-task performance.

Method: Proposes deriving low-dimensional, semantically annotated encodings from intermediate neural activations (peepholes) applied to a convolutional autoencoder to produce interpretable indicators for anomaly identification and localization.

Result: The framework enables semantic characterization of detected anomalies, reveals bias detection, supports fault localization, and requires only marginal increase in computational resources for onboard deployment feasibility.

Conclusion: The proposed explainable AI framework enhances interpretability in neural anomaly detectors for spacecraft fault detection while maintaining computational efficiency suitable for onboard deployment.

Abstract: The increasing autonomy of spacecraft demands fault-detection systems that are both reliable and explainable. This work addresses eXplainable Artificial Intelligence for onboard Fault Detection, Isolation and Recovery within the Attitude and Orbit Control Subsystem by introducing a framework that enhances interpretability in neural anomaly detectors. We propose a method to derive low-dimensional, semantically annotated encodings from intermediate neural activations, called peepholes. Applied to a convolutional autoencoder, the framework produces interpretable indicators that enable the identification and localization of anomalies in reaction-wheel telemetry. Peepholes analysis further reveals bias detection and supports fault localization. The proposed framework enables the semantic characterization of detected anomalies while requiring only a marginal increase in computational resources, thus supporting its feasibility for on-board deployment.

Method: DiADEM learns “how much each demographic axis matters” through per-demographic projections governed by a learned importance vector α. It encodes annotators through these projections, fuses annotator and item representations via complementary concatenation and Hadamard interactions, and is trained with a novel item-level disagreement loss that directly penalizes mispredicted annotation variance.

Result: On DICES conversational-safety and VOICED political-offense benchmarks, DiADEM substantially outperforms both LLM-as-a-judge and neural model baselines across standard and perspectivist metrics, achieving strong disagreement tracking (r=0.75 on DICES). The learned α weights reveal that race and age consistently emerge as the most influential demographic factors driving annotator disagreement.

Conclusion: Explicitly modeling who annotators are (not just what they label) is essential for NLP systems that aim to faithfully represent human interpretive diversity. Demographic factors like race and age significantly influence annotation disagreement patterns.

Abstract: When humans label subjective content, they disagree, and that disagreement is not noise. It reflects genuine differences in perspective shaped by annotators’ social identities and lived experiences. Yet standard practice still flattens these judgments into a single majority label, and recent LLM-based approaches fare no better: we show that prompted large language models, even with chain-of-thought reasoning, fail to recover the structure of human disagreement. We introduce DiADEM, a neural architecture that learns “how much each demographic axis matters” for predicting who will disagree and on what. DiADEM encodes annotators through per-demographic projections governed by a learned importance vector $\boldsymbolα$, fuses annotator and item representations via complementary concatenation and Hadamard interactions, and is trained with a novel item-level disagreement loss that directly penalizes mispredicted annotation variance. On the DICES conversational-safety and VOICED political-offense benchmarks, DiADEM substantially outperforms both the LLM-as-a-judge and neural model baselines across standard and perspectivist metrics, achieving strong disagreement tracking ($r{=}0.75$ on DICES). The learned $\boldsymbolα$ weights reveal that race and age consistently emerge as the most influential demographic factors driving annotator disagreement across both datasets. Our results demonstrate that explicitly modeling who annotators are not just what they label is essential for NLP systems that aim to faithfully represent human interpretive diversity.

Method: Built on a reproducible Android emulation environment covering 42 general GUI tasks, 86 personalized tasks, and 64 proactive tasks. Hides user profiles from agents and exposes only behavioral logs, forcing genuine preference inference. Uses an LLM-driven user simulator grounded in structured profiles for realistic clarification dialogues and proactive consent handling. Employs hybrid evaluation combining rule-based verification with LLM-as-a-Judge scoring.

Result: Agents that excel at explicit task execution fall below 50% performance under vague instructions requiring user preference inference or intervention calibration, even for frontier models like Claude Sonnet 4.6. Core bottlenecks are preference acquisition and intervention calibration, not GUI navigation.

Conclusion: There’s a fundamental gap between competent interface operation and trustworthy personal assistance. The benchmark exposes critical weaknesses in current agents’ ability to infer preferences and calibrate proactive interventions, highlighting the need for better personalization and decision-making capabilities in mobile agents.

Abstract: Personalized mobile agents that infer user preferences and calibrate proactive assistance hold great promise as everyday digital assistants, yet existing benchmarks fail to capture what this requires. Prior work evaluates preference recovery from static histories or intent prediction from fixed contexts. Neither tests whether an agent can elicit missing preferences through interaction, nor whether it can decide when to intervene, seek consent, or remain silent in a live GUI environment. We introduce KnowU-Bench, an online benchmark for personalized mobile agents built on a reproducible Android emulation environment, covering 42 general GUI tasks, 86 personalized tasks, and 64 proactive tasks. Unlike prior work that treats user preferences as static context, KnowU-Bench hides the user profile from the agent and exposes only behavioral logs, forcing genuine preference inference rather than context lookup. To support multi-turn preference elicitation, it instantiates an LLM-driven user simulator grounded in structured profiles, enabling realistic clarification dialogues and proactive consent handling. Beyond personalization, KnowU-Bench provides comprehensive evaluation of the complete proactive decision chain, including grounded GUI execution, consent negotiation, and post-rejection restraint, evaluated through a hybrid protocol combining rule-based verification with LLM-as-a-Judge scoring. Our experiments reveal a striking degradation: agents that excel at explicit task execution fall below 50% under vague instructions requiring user preference inference or intervention calibration, even for frontier models like Claude Sonnet 4.6. The core bottlenecks are not GUI navigation but preference acquisition and intervention calibration, exposing a fundamental gap between competent interface operation and trustworthy personal assistance.

Method: Proposes SUPERNOVA framework that leverages instruction-tuning datasets with expert-annotated ground-truth to extract rich reasoning patterns for RLVR. Conducts 100+ controlled RL experiments to analyze three key data design factors: (1) source task selection, (2) task mixing strategies, and (3) synthetic interventions for improving data quality.

Result: Source task selection is non-trivial and significantly impacts downstream reasoning performance. Selecting tasks based on individual target task performance outperforms strategies based on overall average performance. Models trained on SUPERNOVA outperform strong baselines (e.g., Qwen3.5) on challenging reasoning benchmarks including BBEH, Zebralogic, and MMLU-Pro, with up to 52.8% relative improvement on BBEH.

Conclusion: SUPERNOVA demonstrates the effectiveness of principled data curation for extending RLVR to general reasoning, providing practical insights for leveraging human-annotated resources to enhance LLM reasoning capabilities beyond formal domains.

Abstract: Reinforcement Learning with Verifiable Rewards (RLVR) has significantly improved large language model (LLM) reasoning in formal domains such as mathematics and code. Despite these advancements, LLMs still struggle with general reasoning tasks requiring capabilities such as causal inference and temporal understanding. Extending RLVR to general reasoning is fundamentally constrained by the lack of high-quality, verifiable training data that spans diverse reasoning skills. To address this challenge, we propose SUPERNOVA, a data curation framework for RLVR aimed at enhancing general reasoning. Our key insight is that instruction-tuning datasets containing expert-annotated ground-truth encode rich reasoning patterns that can be systematically adapted for RLVR. To study this, we conduct 100+ controlled RL experiments to analyze how data design choices impact downstream reasoning performance. In particular, we investigate three key factors: (i) source task selection, (ii) task mixing strategies, and (iii) synthetic interventions for improving data quality. Our analysis reveals that source task selection is non-trivial and has a significant impact on downstream reasoning performance. Moreover, selecting tasks based on their performance for individual target tasks outperforms strategies based on overall average performance. Finally, models trained on SUPERNOVA outperform strong baselines (e.g., Qwen3.5) on challenging reasoning benchmarks including BBEH, Zebralogic, and MMLU-Pro. In particular, training on SUPERNOVA yields relative improvements of up to 52.8% on BBEH across model sizes, demonstrating the effectiveness of principled data curation for RLVR. Our findings provide practical insights for curating human-annotated resources to extend RLVR to general reasoning. The code and data is available at https://github.com/asuvarna31/supernova.

Method: Developed a framework for categorizing conflicting incentives in LLMs, inspired by linguistics and advertising regulation literature, and created evaluation suite to test how current models handle tradeoffs in various conflict scenarios.

Result: Majority of LLMs prioritize company incentives over user welfare: Grok 4.1 Fast recommends sponsored products almost twice as expensive (83%), GPT 5.1 surfaces sponsored options to disrupt purchasing (94%), Qwen 3 Next conceals prices in unfavorable comparisons (24%). Behavior varies with reasoning levels and users’ socio-economic status.

Conclusion: LLMs demonstrate concerning tendencies to prioritize company revenue over user benefit in conflict of interest situations, highlighting hidden risks when chatbots are subtly incentivized for advertisements.

Abstract: Today’s large language models (LLMs) are trained to align with user preferences through methods such as reinforcement learning. Yet models are beginning to be deployed not merely to satisfy users, but also to generate revenue for the companies that created them through advertisements. This creates the potential for LLMs to face conflicts of interest, where the most beneficial response to a user may not be aligned with the company’s incentives. For instance, a sponsored product may be more expensive but otherwise equal to another; in this case, what does (and should) the LLM recommend to the user? In this paper, we provide a framework for categorizing the ways in which conflicting incentives might lead LLMs to change the way they interact with users, inspired by literature from linguistics and advertising regulation. We then present a suite of evaluations to examine how current models handle these tradeoffs. We find that a majority of LLMs forsake user welfare for company incentives in a multitude of conflict of interest situations, including recommending a sponsored product almost twice as expensive (Grok 4.1 Fast, 83%), surfacing sponsored options to disrupt the purchasing process (GPT 5.1, 94%), and concealing prices in unfavorable comparisons (Qwen 3 Next, 24%). Behaviors also vary strongly with levels of reasoning and users’ inferred socio-economic status. Our results highlight some of the hidden risks to users that can emerge when companies begin to subtly incentivize advertisements in chatbots.

Method: 1) Multi-criteria evaluation framework with five principled metrics; 2) PGCA method that fuses grid-based perturbation importance with Grad-CAM++ through consensus amplification and adaptive contrast enhancement.

Result: PGCA achieves best performance in fidelity (2.22 ± 1.62), interpretability (3.89 ± 0.33), and fairness (4.95 ± 0.03) across five domains, with statistically significant improvements over baselines (p < 10⁻⁷).

Conclusion: The proposed framework provides comprehensive XAI evaluation, and PGCA demonstrates superior performance by combining perturbation fidelity with gradient-based spatial precision for more reliable explanations.

Abstract: Explainable Artificial Intelligence (XAI) methods are increasingly used in safety-critical domains, yet there is no unified framework to jointly evaluate fidelity, interpretability, robustness, fairness, and completeness. We address this gap through two contributions. First, we propose a multi-criteria evaluation framework that formalizes these five criteria using principled metrics: fidelity via prediction-gap analysis; interpretability via a composite concentration-coherence-contrast score; robustness via cosine-similarity perturbation stability; fairness via Jensen-Shannon divergence across demographic groups; and completeness via feature-ablation coverage. These are integrated using an entropy-weighted dynamic scoring scheme that adapts to domain-specific priorities. Second, we introduce Perturbation-Gradient Consensus Attribution (PGCA), which fuses grid-based perturbation importance with Grad-CAM++ through consensus amplification and adaptive contrast enhancement, combining perturbation fidelity with gradient-based spatial precision. We evaluate across five domains (brain tumor MRI, plant disease, security screening, gender, and sunglass detection) using fine-tuned ResNet-50 models. PGCA achieves the best performance in fidelity $(2.22 \pm 1.62)$, interpretability $(3.89 \pm 0.33)$, and fairness $(4.95 \pm 0.03)$, with statistically significant improvements over baselines $(p < 10^{-7})$. Sensitivity analysis shows stable rankings (Kendall’s $(τ\geq 0.88)$). Code and results are publicly available.

Method: FedIFL uses intra-client prototype contrastive learning to mitigate domain shifts, feature generation for privacy-preserving access to other clients’ distributions, and cross-client feature disentanglement with instance-level consistency loss, personalization loss, and orthogonal loss to separate invariant from specific features.

Result: Experiments on real-world Motor Driven Systems show FedIFL effectively handles inconsistent fault modes and achieves superior generalization across global label spaces compared to other methods.

Conclusion: FedIFL successfully addresses label space inconsistency in federated fault diagnosis, enabling accurate cross-domain diagnosis while preserving data privacy, making it suitable for industrial applications with diverse equipment conditions.

Abstract: Due to the scarcity of industrial data, individual equipment users, particularly start-ups, struggle to independently train a comprehensive fault diagnosis model; federated learning enables collaborative training while ensuring data privacy, making it an ideal solution. However, the diversity of working conditions leads to variations in fault modes, resulting in inconsistent label spaces across different clients. In federated diagnostic scenarios, label space inconsistency leads to local models focus on client-specific fault modes and causes local models from different clients to map different failure modes to similar feature representations, which weakens the aggregated global model’s generalization. To tackle this issue, this article proposed a federated cross-domain diagnostic framework termed Federated Invariant Features Learning (FedIFL). In intra-client training, prototype contrastive learning mitigates intra-client domain shifts, subsequently, feature generating ensures local models can access distributions of other clients in a privacy-friendly manner. Besides, in cross-client training, a feature disentanglement mechanism is introduced to mitigate cross-client domain shifts, specifically, an instance-level federated instance consistency loss is designed to ensure the instance-level consistency of invariant features between different clients, furthermore, a federated instance personalization loss and an orthogonal loss are constructed to distinguish specific features that from the invariant features. Eventually, the aggregated model achieves promising generalization among global label spaces, enabling accurate fault diagnosis for target clients’ Motor Driven Systems (MDSs) with inconsistent label spaces. Experiments on real-world MDSs validate the effectiveness and superiority of FedIFL in federated cross-domain diagnosis with inconsistent fault modes.

Method: PDDLego framework that iteratively formalizes, plans, grows, and refines PDDL representations by decomposing the environment and goal into fully observable episodes. It doesn’t require finetuning, in-context exemplars, or trajectories, and works by breaking down partially observable problems into manageable fully observable subproblems.

Result: PDDLego improves planning success and exhibits robustness against problem complexity compared to end-to-end approaches. The domain knowledge captured after successful trials can benefit future tasks, showing transfer learning capabilities.

Conclusion: Using LLMs to formalize environments into PDDL rather than directly predicting plans is effective for partially observable environments. The iterative decomposition approach enables handling complex, incomplete information scenarios while maintaining planning performance.

Abstract: Using LLMs not to predict plans but to formalize an environment into the Planning Domain Definition Language (PDDL) has been shown to improve performance and control. While most existing methodology only applies to fully observable environments, we adapt to the more realistic and challenging partially observable environments without sufficient information to make a complete plan. We propose PDDLego, a framework to iteratively formalize, plan, grow, and refine PDDL representations by decomposing the environment and the goal into fully observable episodes. Without finetuning, in-context exemplars, or trajectories, PDDLego improves planning success and exhibits robustness against problem complexity compared to end-to-end approaches. We also show that the domain knowledge captured after a successful trial can benefit future tasks.

Method: STPR leverages LLMs’ coding capabilities to generate executable Python functions from natural language constraints, then applies these functions to point cloud representations with traditional search algorithms.

Result: Experiments in simulated Gazebo environment show STPR ensures full compliance across constraints and scenarios with short runtimes, and works with smaller code LLMs for low inference cost.

Conclusion: STPR effectively bridges natural language constraints and robotic planning by using LLMs as code generators, making complex constraint specification accessible and executable.

Abstract: Recent advancements in large language models (LLMs) have spurred interest in robotic navigation that incorporates complex spatial, mathematical, and conditional constraints from natural language into the planning problem. Such constraints can be informal yet highly complex, making it challenging to translate into a formal description that can be passed on to a planning algorithm. In this paper, we propose STPR, a constraint generation framework that uses LLMs to translate constraints (expressed as instructions on ``what not to do’’) into executable Python functions. STPR leverages the LLM’s strong coding capabilities to shift the problem description from language into structured and interpretable code, thus circumventing complex reasoning and avoiding potential hallucinations. We show that these LLM-generated functions accurately describe even complex mathematical constraints, and apply them to point cloud representations with traditional search algorithms. Experiments in a simulated Gazebo environment show that STPR ensures full compliance across several constraints and scenarios, while having short runtimes. We also verify that STPR can be used with smaller code LLMs, making it applicable to a wide range of compact models with low inference cost.

Method: HC-INR decomposes representation into two components: (i) a learned multiscale coordinate transformation module that warps input domain into disentangled latent space, and (ii) a compact implicit field network that models transformed signal with reduced complexity. Uses hierarchical hypernetwork architecture conditioning coordinate transformations on local signal features.

Result: Achieves up to 4 times higher reconstruction fidelity than strong INR baselines while using 30-60% fewer parameters across image fitting, shape reconstruction, and neural radiance field approximation tasks.

Conclusion: HC-INR breaks representational bottlenecks in INRs through hypernetwork-based coordinate transformations, enabling dynamic capacity allocation and improved efficiency for representing complex signals.

Abstract: Implicit Neural Representations (INRs) have emerged as a powerful paradigm for representing signals such as images, 3D shapes, signed distance fields, and radiance fields. While significant progress has been made in architecture design (e.g., SIREN, FFC, KAN-based INRs) and optimization strategies (meta-learning, amortization, distillation), existing approaches still suffer from two core limitations: (1) a representation bottleneck that forces a single MLP to uniformly model heterogeneous local structures, and (2) limited scalability due to the absence of a hierarchical mechanism that dynamically adapts to signal complexity. This work introduces Hyper-Coordinate Implicit Neural Representations (HC-INR), a new class of INRs that break the representational bottleneck by learning signal-adaptive coordinate transformations using a hypernetwork. HC-INR decomposes the representation task into two components: (i) a learned multiscale coordinate transformation module that warps the input domain into a disentangled latent space, and (ii) a compact implicit field network that models the transformed signal with significantly reduced complexity. The proposed model introduces a hierarchical hypernetwork architecture that conditions coordinate transformations on local signal features, enabling dynamic allocation of representation capacity. We theoretically show that HC-INR strictly increases the upper bound of representable frequency bands while maintaining Lipschitz stability. Extensive experiments across image fitting, shape reconstruction, and neural radiance field approximation demonstrate that HC-INR achieves up to 4 times higher reconstruction fidelity than strong INR baselines while using 30–60% fewer parameters.

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

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

Conclusion: Cannot draw conclusions without access to the paper content.

Abstract: Failed to fetch summary for 2306.02781: Page request resulted in HTTP 429 (https://export.arxiv.org/api/query?search_query=&id_list=2306.02781&sortBy=relevance&sortOrder=descending&start=0&max_results=100)

Method: Unable to determine method due to access error

Result: Unable to determine results due to access error

Conclusion: Unable to determine conclusion due to access error

Abstract: Failed to fetch summary for 2512.08296: Page request resulted in HTTP 429 (https://export.arxiv.org/api/query?search_query=&id_list=2512.08296&sortBy=relevance&sortOrder=descending&start=0&max_results=100)

Method: Unable to determine method due to access error

Result: Unable to determine results due to access error

Conclusion: Unable to determine conclusion due to access error

Abstract: Failed to fetch summary for 2601.04666: Page request resulted in HTTP 429 (https://export.arxiv.org/api/query?search_query=&id_list=2601.04666&sortBy=relevance&sortOrder=descending&start=0&max_results=100)

Method: Cannot determine method as paper content is unavailable

Result: Cannot determine results as paper content is unavailable

Conclusion: Cannot draw conclusions as paper content is unavailable

Abstract: Failed to fetch summary for 2601.21872: Page request resulted in HTTP 429 (https://export.arxiv.org/api/query?search_query=&id_list=2601.21872&sortBy=relevance&sortOrder=descending&start=0&max_results=100)

Method: Unable to determine method due to access error

Result: Unable to determine results due to access error

Conclusion: Unable to determine conclusion due to access error

Abstract: Failed to fetch summary for 2602.03249: Page request resulted in HTTP 429 (https://export.arxiv.org/api/query?search_query=&id_list=2602.03249&sortBy=relevance&sortOrder=descending&start=0&max_results=100)

Method: Cannot determine method without paper content

Result: Cannot determine results without paper content

Conclusion: Cannot determine conclusion without paper content

Abstract: Failed to fetch summary for 2406.10521: Page request resulted in HTTP 429 (https://export.arxiv.org/api/query?search_query=&id_list=2406.10521&sortBy=relevance&sortOrder=descending&start=0&max_results=100)

Method: Cannot determine method as paper content is unavailable

Result: Cannot determine results as paper content is unavailable

Conclusion: Cannot draw conclusions as paper content is unavailable

Abstract: Failed to fetch summary for 2603.00680: Page request resulted in HTTP 429 (https://export.arxiv.org/api/query?search_query=&id_list=2603.00680&sortBy=relevance&sortOrder=descending&start=0&max_results=100)

Method: Cannot determine method without access to paper content

Result: Cannot determine results without access to paper content

Conclusion: Cannot draw conclusions without access to paper content

Abstract: Failed to fetch summary for 2407.19426: Page request resulted in HTTP 429 (https://export.arxiv.org/api/query?search_query=&id_list=2407.19426&sortBy=relevance&sortOrder=descending&start=0&max_results=100)

Method: No method information available due to failed fetch. The paper ID format suggests it’s a recent preprint from March 2024.

Result: No results available. The arXiv API returned HTTP 429 (Too Many Requests), which is a rate limiting error.

Conclusion: Cannot provide conclusion without paper content. Need to wait for rate limits to reset or try alternative access methods.

Abstract: Failed to fetch summary for 2603.04791: Page request resulted in HTTP 429 (https://export.arxiv.org/api/query?search_query=&id_list=2603.04791&sortBy=relevance&sortOrder=descending&start=0&max_results=100)

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

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

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

Abstract: Failed to fetch summary for 2411.02622: Page request resulted in HTTP 429 (https://export.arxiv.org/api/query?search_query=&id_list=2411.02622&sortBy=relevance&sortOrder=descending&start=0&max_results=100)

Method: Unable to determine method due to fetch failure

Result: Unable to determine results due to fetch failure

Conclusion: Unable to determine conclusion due to fetch failure

Abstract: Failed to fetch summary for 2603.16365: Page request resulted in HTTP 429 (https://export.arxiv.org/api/query?search_query=&id_list=2603.16365&sortBy=relevance&sortOrder=descending&start=0&max_results=100)

Method: Unable to determine method due to failed paper retrieval

Result: Unable to determine results due to failed paper retrieval

Conclusion: Unable to draw conclusions due to failed paper retrieval

Abstract: Failed to fetch summary for 2501.00773: Page request resulted in HTTP 429 (https://export.arxiv.org/api/query?search_query=&id_list=2501.00773&sortBy=relevance&sortOrder=descending&start=0&max_results=100)

Method: Unable to determine method due to fetch failure

Result: Unable to determine results due to fetch failure

Conclusion: Unable to determine conclusion due to fetch failure

Abstract: Failed to fetch summary for 2603.27765: Page request resulted in HTTP 429 (https://export.arxiv.org/api/query?search_query=&id_list=2603.27765&sortBy=relevance&sortOrder=descending&start=0&max_results=100)

Method: Cannot determine method due to inability to access paper content.

Result: Cannot determine results due to inability to access paper content.

Conclusion: Cannot draw conclusions due to inability to access paper content.

Abstract: Failed to fetch summary for 2603.28618: Page request resulted in HTTP 429 (https://export.arxiv.org/api/query?search_query=&id_list=2603.28618&sortBy=relevance&sortOrder=descending&start=0&max_results=100)

Method: Unable to determine method due to failed paper fetch

Result: Unable to determine results due to failed paper fetch

Conclusion: Unable to determine conclusion due to failed paper fetch

Abstract: Failed to fetch summary for 2604.02500: Page request resulted in HTTP 429 (https://export.arxiv.org/api/query?search_query=&id_list=2604.02500&sortBy=relevance&sortOrder=descending&start=0&max_results=100)

Method: GoS constructs an executable skill graph offline from skill packages, then at inference time retrieves a bounded, dependency-aware skill bundle using hybrid semantic-lexical seeding, reverse-weighted Personalized PageRank, and context-budgeted hydration.

Result: On SkillsBench and ALFWorld, GoS improves average reward by 43.6% over vanilla full skill-loading baseline while reducing input tokens by 37.8%. It generalizes across three model families (Claude Sonnet, GPT-5.2 Codex, MiniMax) and scales effectively from 200 to 2,000 skills.

Conclusion: GoS provides an effective structural retrieval layer for large skill libraries that balances reward, token efficiency, and runtime, outperforming both full skill loading and simple vector retrieval approaches.

Abstract: Skill usage has become a core component of modern agent systems and can substantially improve agents’ ability to complete complex tasks. In real-world settings, where agents must monitor and interact with numerous personal applications, web browsers, and other environment interfaces, skill libraries can scale to thousands of reusable skills. Scaling to larger skill sets introduces two key challenges. First, loading the full skill set saturates the context window, driving up token costs, hallucination, and latency. In this paper, we present Graph of Skills (GoS), an inference-time structural retrieval layer for large skill libraries. GoS constructs an executable skill graph offline from skill packages, then at inference time retrieves a bounded, dependency-aware skill bundle through hybrid semantic-lexical seeding, reverse-weighted Personalized PageRank, and context-budgeted hydration. On SkillsBench and ALFWorld, GoS improves average reward by 43.6% over the vanilla full skill-loading baseline while reducing input tokens by 37.8%, and generalizes across three model families: Claude Sonnet, GPT-5.2 Codex, and MiniMax. Additional ablation studies across skill libraries ranging from 200 to 2,000 skills further demonstrate that GoS consistently outperforms both vanilla skills loading and simple vector retrieval in balancing reward, token efficiency, and runtime.

Method: Cannot determine method due to missing paper content

Result: Cannot determine results due to missing paper content

Conclusion: Cannot determine conclusion due to missing paper content

Abstract: Failed to fetch summary for 2604.06747: Page request resulted in HTTP 429 (https://export.arxiv.org/api/query?search_query=&id_list=2604.06747&sortBy=relevance&sortOrder=descending&start=0&max_results=100)

Method: Cannot determine method as paper content could not be retrieved

Result: Cannot determine results as paper content could not be retrieved

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

Abstract: Failed to fetch summary for 1912.08786: Page request resulted in HTTP 429 (https://export.arxiv.org/api/query?search_query=&id_list=1912.08786&sortBy=relevance&sortOrder=descending&start=0&max_results=100)

Method: Unable to determine method due to API access limitations

Result: Unable to determine results due to API access limitations

Conclusion: Unable to determine conclusion due to API access limitations

Abstract: Failed to fetch summary for 2406.13086: Page request resulted in HTTP 429 (https://export.arxiv.org/api/query?search_query=&id_list=2406.13086&sortBy=relevance&sortOrder=descending&start=0&max_results=100)

Method: Cannot determine method as paper content is unavailable

Result: Cannot determine results as paper content is unavailable

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

Abstract: Failed to fetch summary for 2407.01563: Page request resulted in HTTP 429 (https://export.arxiv.org/api/query?search_query=&id_list=2407.01563&sortBy=relevance&sortOrder=descending&start=0&max_results=100)

Method: Cannot determine method as paper content is unavailable

Result: Cannot determine results as paper content is unavailable

Conclusion: Cannot draw conclusions without access to paper content

Abstract: Failed to fetch summary for 2503.12374: Page request resulted in HTTP 429 (https://export.arxiv.org/api/query?search_query=&id_list=2503.12374&sortBy=relevance&sortOrder=descending&start=0&max_results=100)

Method: Unable to determine method due to failed API request

Result: Unable to determine results due to failed API request

Conclusion: Unable to draw conclusions due to failed API request

Abstract: Failed to fetch summary for 2504.15564: Page request resulted in HTTP 429 (https://export.arxiv.org/api/query?search_query=&id_list=2504.15564&sortBy=relevance&sortOrder=descending&start=0&max_results=100)

Method: Unable to determine method due to access error

Result: Unable to determine results due to access error

Conclusion: Unable to determine conclusion due to access error

Abstract: Failed to fetch summary for 2505.10375: Page request resulted in HTTP 429 (https://export.arxiv.org/api/query?search_query=&id_list=2505.10375&sortBy=relevance&sortOrder=descending&start=0&max_results=100)

Method: Unable to determine method due to access error

Result: Unable to determine results due to access error

Conclusion: Unable to determine conclusion due to access error

Abstract: Failed to fetch summary for 2505.11548: Page request resulted in HTTP 429 (https://export.arxiv.org/api/query?search_query=&id_list=2505.11548&sortBy=relevance&sortOrder=descending&start=0&max_results=100)

Method: Unable to determine method due to failed paper retrieval

Result: Unable to determine results due to failed paper retrieval

Conclusion: Unable to draw conclusions due to failed paper retrieval

Abstract: Failed to fetch summary for 2505.17209: Page request resulted in HTTP 429 (https://export.arxiv.org/api/query?search_query=&id_list=2505.17209&sortBy=relevance&sortOrder=descending&start=0&max_results=100)

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

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

Conclusion: Cannot draw conclusions without access to the paper content.

Abstract: Failed to fetch summary for 2506.01979: Page request resulted in HTTP 429 (https://export.arxiv.org/api/query?search_query=&id_list=2506.01979&sortBy=relevance&sortOrder=descending&start=0&max_results=100)

Method: Cannot determine method without access to paper content

Result: Cannot determine results without access to paper content

Conclusion: Cannot determine conclusion without access to paper content

Abstract: Failed to fetch summary for 2506.04742: Page request resulted in HTTP 429 (https://export.arxiv.org/api/query?search_query=&id_list=2506.04742&sortBy=relevance&sortOrder=descending&start=0&max_results=100)

Method: Cannot analyze method without access to paper content

Result: No results available due to technical access issues

Conclusion: Paper analysis not possible due to arXiv API rate limiting

Abstract: Failed to fetch summary for 2506.09354: Page request resulted in HTTP 429 (https://export.arxiv.org/api/query?search_query=&id_list=2506.09354&sortBy=relevance&sortOrder=descending&start=0&max_results=100)

Method: Unable to determine method due to failed paper retrieval

Result: Unable to determine results due to failed paper retrieval

Conclusion: Unable to determine conclusion due to failed paper retrieval

Abstract: Failed to fetch summary for 2506.09362: Page request resulted in HTTP 429 (https://export.arxiv.org/api/query?search_query=&id_list=2506.09362&sortBy=relevance&sortOrder=descending&start=0&max_results=100)

Method: Unable to determine method due to fetch failure

Result: Unable to determine results due to fetch failure

Conclusion: Unable to determine conclusion due to fetch failure

Abstract: Failed to fetch summary for 2508.05674: Page request resulted in HTTP 429 (https://export.arxiv.org/api/query?search_query=&id_list=2508.05674&sortBy=relevance&sortOrder=descending&start=0&max_results=100)

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

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

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

Abstract: Failed to fetch summary for 2510.07809: Page request resulted in HTTP 429 (https://export.arxiv.org/api/query?search_query=&id_list=2510.07809&sortBy=relevance&sortOrder=descending&start=0&max_results=100)

Method: Cannot determine method as paper content is unavailable

Result: Cannot determine results as paper content is unavailable

Conclusion: Cannot draw conclusions without access to paper content

Abstract: Failed to fetch summary for 2510.11752: Page request resulted in HTTP 429 (https://export.arxiv.org/api/query?search_query=&id_list=2510.11752&sortBy=relevance&sortOrder=descending&start=0&max_results=100)

Method: Unable to determine method due to failed paper fetch

Result: Unable to determine results due to failed paper fetch

Conclusion: Unable to determine conclusion due to failed paper fetch

Abstract: Failed to fetch summary for 2510.15494: Page request resulted in HTTP 429 (https://export.arxiv.org/api/query?search_query=&id_list=2510.15494&sortBy=relevance&sortOrder=descending&start=0&max_results=100)

Method: Unable to determine method due to fetch failure

Result: Unable to determine results due to fetch failure

Conclusion: Unable to determine conclusion due to fetch failure

Abstract: Failed to fetch summary for 2512.13749: Page request resulted in HTTP 429 (https://export.arxiv.org/api/query?search_query=&id_list=2512.13749&sortBy=relevance&sortOrder=descending&start=0&max_results=100)

Method: Unable to determine method due to access error

Result: Unable to determine results due to access error

Conclusion: Unable to determine conclusion due to access error

Abstract: Failed to fetch summary for 2510.15949: Page request resulted in HTTP 429 (https://export.arxiv.org/api/query?search_query=&id_list=2510.15949&sortBy=relevance&sortOrder=descending&start=0&max_results=100)

Method: N/A - Paper content not accessible

Result: N/A - Paper content not accessible

Conclusion: N/A - Paper content not accessible

Abstract: Failed to fetch summary for 2511.08637: Page request resulted in HTTP 429 (https://export.arxiv.org/api/query?search_query=&id_list=2511.08637&sortBy=relevance&sortOrder=descending&start=0&max_results=100)

Method: Unable to determine method due to fetch failure

Result: Unable to determine results due to fetch failure

Conclusion: Unable to determine conclusion due to fetch failure

Abstract: Failed to fetch summary for 2601.03703: Page request resulted in HTTP 429 (https://export.arxiv.org/api/query?search_query=&id_list=2601.03703&sortBy=relevance&sortOrder=descending&start=0&max_results=100)

Method: Cannot determine method as paper content is unavailable

Result: Cannot determine results as paper content is unavailable

Conclusion: Cannot draw conclusions as paper content is unavailable

Abstract: Failed to fetch summary for 2511.20179: Page request resulted in HTTP 429 (https://export.arxiv.org/api/query?search_query=&id_list=2511.20179&sortBy=relevance&sortOrder=descending&start=0&max_results=100)

Method: Cannot determine method as paper content is unavailable

Result: Cannot determine results as paper content is unavailable

Conclusion: Cannot draw conclusions as paper content is unavailable

Abstract: Failed to fetch summary for 2602.20223: Page request resulted in HTTP 429 (https://export.arxiv.org/api/query?search_query=&id_list=2602.20223&sortBy=relevance&sortOrder=descending&start=0&max_results=100)

Method: Unable to determine method as the paper content could not be retrieved

Result: Unable to determine results as the paper content could not be retrieved

Conclusion: Unable to draw conclusions about the paper due to technical limitations in accessing the content

Abstract: Failed to fetch summary for 2603.03099: Page request resulted in HTTP 429 (https://export.arxiv.org/api/query?search_query=&id_list=2603.03099&sortBy=relevance&sortOrder=descending&start=0&max_results=100)

Method: Unable to determine method due to failed paper fetch

Result: Unable to determine results due to failed paper fetch

Conclusion: Unable to determine conclusion due to failed paper fetch

Abstract: Failed to fetch summary for 2603.10100: Page request resulted in HTTP 429 (https://export.arxiv.org/api/query?search_query=&id_list=2603.10100&sortBy=relevance&sortOrder=descending&start=0&max_results=100)

Method: Cannot determine method due to inability to access paper content

Result: Cannot determine results due to inability to access paper content

Conclusion: Cannot draw conclusions due to inability to access paper content

Abstract: Failed to fetch summary for 2602.07900: Page request resulted in HTTP 429 (https://export.arxiv.org/api/query?search_query=&id_list=2602.07900&sortBy=relevance&sortOrder=descending&start=0&max_results=100)

Method: Unable to determine method due to fetch failure

Result: Unable to determine results due to fetch failure

Conclusion: Unable to determine conclusion due to fetch failure

Abstract: Failed to fetch summary for 2602.10527: Page request resulted in HTTP 429 (https://export.arxiv.org/api/query?search_query=&id_list=2602.10527&sortBy=relevance&sortOrder=descending&start=0&max_results=100)

Method: Cannot determine method without access to paper content

Result: Cannot determine results without access to paper content

Conclusion: Cannot determine conclusion without access to paper content

Abstract: Failed to fetch summary for 2603.09172: Page request resulted in HTTP 429 (https://export.arxiv.org/api/query?search_query=&id_list=2603.09172&sortBy=relevance&sortOrder=descending&start=0&max_results=100)

Method: Unable to determine method due to missing abstract

Result: Unable to determine results due to missing abstract

Conclusion: Unable to determine conclusion due to missing abstract

Abstract: Failed to fetch summary for 2603.18030: Page request resulted in HTTP 429 (https://export.arxiv.org/api/query?search_query=&id_list=2603.18030&sortBy=relevance&sortOrder=descending&start=0&max_results=100)

Method: Cannot determine method without access to paper content.

Result: Cannot determine results without access to paper content.

Conclusion: Cannot draw conclusions without access to paper content.

Abstract: Failed to fetch summary for 2603.27142: Page request resulted in HTTP 429 (https://export.arxiv.org/api/query?search_query=&id_list=2603.27142&sortBy=relevance&sortOrder=descending&start=0&max_results=100)

Method: Unable to determine method due to access error

Result: Unable to determine results due to access error

Conclusion: Unable to determine conclusion due to access error

Abstract: Failed to fetch summary for 2603.28507: Page request resulted in HTTP 429 (https://export.arxiv.org/api/query?search_query=&id_list=2603.28507&sortBy=relevance&sortOrder=descending&start=0&max_results=100)

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

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

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

Abstract: Failed to fetch summary for 2604.01930: Page request resulted in HTTP 429 (https://export.arxiv.org/api/query?search_query=&id_list=2604.01930&sortBy=relevance&sortOrder=descending&start=0&max_results=100)

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

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

Conclusion: Cannot draw conclusions without access to the paper content.

Abstract: Failed to fetch summary for 2604.04083: Page request resulted in HTTP 429 (https://export.arxiv.org/api/query?search_query=&id_list=2604.04083&sortBy=relevance&sortOrder=descending&start=0&max_results=100)

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

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

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

Abstract: Failed to fetch summary for 2604.04418: Page request resulted in HTTP 429 (https://export.arxiv.org/api/query?search_query=&id_list=2604.04418&sortBy=relevance&sortOrder=descending&start=0&max_results=100)

Method: Unable to determine method due to access restrictions

Result: Unable to determine results due to access restrictions

Conclusion: Unable to draw conclusions due to access restrictions

Abstract: Failed to fetch summary for 2604.04440: Page request resulted in HTTP 429 (https://export.arxiv.org/api/query?search_query=&id_list=2604.04440&sortBy=relevance&sortOrder=descending&start=0&max_results=100)

Method: Unable to determine method due to failed paper fetch

Result: Unable to determine results due to failed paper fetch

Conclusion: Unable to determine conclusion due to failed paper fetch

Abstract: Failed to fetch summary for 2604.06323: Page request resulted in HTTP 429 (https://export.arxiv.org/api/query?search_query=&id_list=2604.06323&sortBy=relevance&sortOrder=descending&start=0&max_results=100)

Method: Unable to determine method due to failed paper fetch

Result: Unable to determine results due to failed paper fetch

Conclusion: Unable to draw conclusions due to failed paper fetch

Abstract: Failed to fetch summary for 2604.06436: Page request resulted in HTTP 429 (https://export.arxiv.org/api/query?search_query=&id_list=2604.06436&sortBy=relevance&sortOrder=descending&start=0&max_results=100)

Method: Proposes Semantic-Emotional Resonance Embedding (SERE) - a cross-lingual dynamic feature paradigm that constructs emotion-semantic structure using few labeled samples, learns human emotional experiences through Instantaneous Resonance Field (IRF), and uses Triple-Resonance Interaction Chain (TRIC) loss for interaction between labeled and unlabeled samples.

Result: Extensive experiments across multiple languages demonstrate effectiveness with only 5-shot labeling in source language, achieving cross-lingual emotion recognition without target language labels or translation alignment.

Conclusion: The proposed semi-supervised framework enables effective cross-lingual speech emotion recognition with minimal labeled data by leveraging semantic-emotional resonance and dynamic feature learning.

Abstract: Cross-lingual Speech Emotion Recognition (CLSER) aims to identify emotional states in unseen languages. However, existing methods heavily rely on the semantic synchrony of complete labels and static feature stability, hindering low-resource languages from reaching high-resource performance. To address this, we propose a semi-supervised framework based on Semantic-Emotional Resonance Embedding (SERE), a cross-lingual dynamic feature paradigm that requires neither target language labels nor translation alignment. Specifically, SERE constructs an emotion-semantic structure using a small number of labeled samples. It learns human emotional experiences through an Instantaneous Resonance Field (IRF), enabling unlabeled samples to self-organize into this structure. This achieves semi-supervised semantic guidance and structural discovery. Additionally, we design a Triple-Resonance Interaction Chain (TRIC) loss to enable the model to reinforce the interaction and embedding capabilities between labeled and unlabeled samples during emotional highlights. Extensive experiments across multiple languages demonstrate the effectiveness of our method, requiring only 5-shot labeling in the source language.

Method: Uses latent diffusion model for accompaniment generation with sliding-window look-ahead protocol. Combines MAX/MSP front-end for real-time audio handling with Python inference server via OSC/UDP. Applies consistency distillation to reduce sampling time by 5.4x.

Result: Both models achieve real-time operation with strong performance in musical coherence, beat alignment, and audio quality in Retrospective regime. Performance degrades gracefully as look-ahead increases.

Conclusion: Demonstrates feasibility of diffusion-based real-time accompaniment and reveals fundamental trade-off between model latency, look-ahead depth, and generation quality.

Abstract: We present a framework for real-time human-AI musical co-performance, in which a latent diffusion model generates instrumental accompaniment in response to a live stream of context audio. The system combines a MAX/MSP front-end-handling real-time audio input, buffering, and playback-with a Python inference server running the generative model, communicating via OSC/UDP messages. This allows musicians to perform in MAX/MSP - a well-established, real-time capable environment - while interacting with a large-scale Python-based generative model, overcoming the fundamental disconnect between real-time music tools and state-of-the-art AI models. We formulate accompaniment generation as a sliding-window look-ahead protocol, training the model to predict future audio from partial context, where system latency is a critical constraint. To reduce latency, we apply consistency distillation to our diffusion model, achieving a 5.4x reduction in sampling time, with both models achieving real-time operation. Evaluated on musical coherence, beat alignment, and audio quality, both models achieve strong performance in the Retrospective regime and degrade gracefully as look-ahead increases. These results demonstrate the feasibility of diffusion-based real-time accompaniment and expose the fundamental trade-off between model latency, look-ahead depth, and generation quality that any such system must navigate.

Method: DFS uses a semi-supervised pipeline: clustering of unannotated recordings with manual validation, followed by supervised fine-tuning of an Audio Spectrogram Transformer (AST) using low-rank adaptation (LoRA). This is compared to a frozen-backbone linear baseline (DFS-Linear). The model detects multiple taxonomic groups from long-term acoustic recordings.

Result: DFS outperforms existing automatic detection tools for 8 out of 12 taxons, particularly for non-avian taxa, achieving average AP values of 0.964 for primates and 0.961 for elephants. LoRA-based fine-tuning substantially outperforms linear probing across taxa.

Conclusion: Task-oriented, region-specific training substantially improves detection performance in acoustically complex tropical environments. DFS shows potential as a practical tool for biodiversity monitoring and conservation in African rainforests.

Abstract: Passive Acoustic Monitoring (PAM) is widely used for biodiversity assessment. Its application in African tropical forests is limited by scarce annotated data, reducing the performance of general-purpose ecoacoustic models on underrepresented taxa. In this study, we introduce DeepForestSound (DFS), a multi-species automatic detection model designed for PAM in African tropical forests. DFS relies on a semi-supervised pipeline combining clustering of unannotated recordings with manual validation, followed by supervised fine-tuning of an Audio Spectrogram Transformer (AST) using low-rank adaptation, which is compared to a frozen-backbone linear baseline (DFS-Linear). The framework supports the detection of multiple taxonomic groups, including birds, primates, and elephants, from long-term acoustic recordings. DFS was trained on acoustic data collected in the Sebitoli area, in Kibale National Park, Uganda, and evaluated on an independent dataset recorded two years later at different locations within the same forest. This evaluation therefore assesses generalization across time and recording sites within a single tropical forest ecosystem. Across 8 out of 12 taxons, DFS outperforms existing automatic detection tools, particularly for non-avian taxa, achieving average AP values of 0.964 for primates and 0.961 for elephants. Results further show that LoRA-based fine-tuning substantially outperforms linear probing across taxa. Overall, these results demonstrate that task-oriented, region-specific training substantially improves detection performance in acoustically complex tropical environments, and highlight the potential of DFS as a practical tool for biodiversity monitoring and conservation in African rainforests.

Method: Formalized as Sequential Device-Addressed Routing (SDAR) and implemented as Selective Attention System (SAS) - an on-device system using sequential routing over interaction history rather than utterance-local classification, with optional audio+video fusion.

Result: Audio-only: F1=0.86; Audio+video: F1=0.95; Removing interaction history reduced F1 from 0.95 to 0.57; Runs on ARM Cortex-A hardware with <150ms latency and <20MB footprint.

Conclusion: Sequential routing using interaction history is crucial for device-addressed speech detection in multi-speaker environments, with audio+video fusion providing significant performance gains while maintaining edge deployment feasibility.

Abstract: We study device-addressed speech detection under pre-ASR edge deployment constraints, where systems must decide whether to forward audio before transcription under strict latency and compute limits. We show that, in multi-speaker environments with temporally ambiguous utterances, this task is more effectively modelled as a sequential routing problem over interaction history than as an utterance-local classification task. We formalize this as Sequential Device-Addressed Routing (SDAR) and present the Selective Attention System (SAS), an on-device implementation that instantiates this formulation. On a held-out 60-hour multi-speaker English test set, the primary audio-only configuration achieves F1=0.86 (precision=0.89, recall=0.83); with an optional camera, audio+video fusion raises F1 to 0.95 (precision=0.97, recall=0.93). Removing causal interaction history (Stage~3) reduced F1 from 0.95 to 0.57+/-0.03 in the audio+video configuration under our evaluation protocol. Among the tested components, this was the largest observed ablation effect, indicating that short-horizon interaction history carries substantial decision-relevant information in the evaluated setting. SAS runs fully on-device on ARM Cortex-A class hardware (<150 ms latency, <20 MB footprint). All results are from internal evaluation on a proprietary dataset evaluated primarily in English; a 5-hour evaluation subset may be shared for independent verification (Section 8.8).

Method: TG-DP decouples reconstruction and alignment into separate optimization paths with disentangled masking regimes. A teacher model provides auxiliary guidance for organizing visible tokens in the contrastive branch, reducing interference and stabilizing cross-modal representation learning.

Result: Achieves state-of-the-art performance in zero-shot retrieval: improves R@1 from 35.2% to 37.4% for video-to-audio retrieval and from 27.9% to 37.1% for audio-to-video retrieval on AudioSet. Also achieves SOTA linear-probe performance on AS20K and VGGSound.

Conclusion: Decoupling multimodal objectives and introducing teacher-guided structure into the contrastive pathway provides an effective framework for improving large-scale audio-visual pretraining.

Abstract: Recent advances in audio-visual representation learning have shown the value of combining contrastive alignment with masked reconstruction. However, jointly optimizing these objectives in a single forward pass forces the contrastive branch to rely on randomly visible patches designed for reconstruction rather than cross-modal alignment, introducing semantic noise and optimization interference. We propose TG-DP, a Teacher-Guided Dual-Path framework that decouples reconstruction and alignment into separate optimization paths. By disentangling the masking regimes of the two branches, TG-DP enables the contrastive pathway to use a visibility pattern better suited to cross-modal alignment. A teacher model further provides auxiliary guidance for organizing visible tokens in this branch, helping reduce interference and stabilize cross-modal representation learning. TG-DP achieves state-of-the-art performance in zero-shot retrieval. On AudioSet, it improves R@1 from 35.2% to 37.4% for video-to-audio retrieval and from 27.9% to 37.1% for audio-to-video retrieval. The learned representations also remain semantically robust, achieving state-of-the-art linear-probe performance on AS20K and VGGSound. Taken together, our results suggest that decoupling multimodal objectives and introducing teacher-guided structure into the contrastive pathway provide an effective framework for improving large-scale audio-visual pretraining. Code is available at https://github.com/wanglg20/TG-DP.

Method: Developed DeepFense, an open-source PyTorch toolkit integrating latest architectures, loss functions, and augmentation pipelines with over 100 recipes. Conducted large-scale evaluation of 400+ models to analyze factors affecting performance and biases.

Result: While curated training data improves cross-domain generalization, the choice of pre-trained front-end feature extractor dominates overall performance variance. Found severe biases in high-performing models regarding audio quality, speaker gender, and language.

Conclusion: DeepFense provides necessary tools for equitable training data selection and front-end fine-tuning to facilitate real-world deployment of speech deepfake detection systems while addressing identified biases.

Abstract: Speech deepfake detection is a well-established research field with different models, datasets, and training strategies. However, the lack of standardized implementations and evaluation protocols limits reproducibility, benchmarking, and comparison across studies. In this work, we present DeepFense, a comprehensive, open-source PyTorch toolkit integrating the latest architectures, loss functions, and augmentation pipelines, alongside over 100 recipes. Using DeepFense, we conducted a large-scale evaluation of more than 400 models. Our findings reveal that while carefully curated training data improves cross-domain generalization, the choice of pre-trained front-end feature extractor dominates overall performance variance. Crucially, we show severe biases in high-performing models regarding audio quality, speaker gender, and language. DeepFense is expected to facilitate real-world deployment with the necessary tools to address equitable training data selection and front-end fine-tuning.

Method: Proposes AT-ADD Grand Challenge with two tracks: 1) Robust Speech Deepfake Detection evaluating detectors in real-world scenarios against unseen speech generation methods, and 2) All-Type Audio Deepfake Detection extending detection to diverse unknown audio types (speech, sound, singing, music) with type-agnostic generalization.

Result: The challenge provides standardized datasets, rigorous evaluation protocols, and reproducible baselines to accelerate development of robust and generalizable audio forensic technologies.

Conclusion: AT-ADD aims to bridge academic evaluation with practical multimedia forensics, supporting secure communication, reliable media verification, and responsible governance in the era of pervasive synthetic audio.

Abstract: The rapid advancement of Audio Large Language Models (ALLMs) has enabled cost-effective, high-fidelity generation and manipulation of both speech and non-speech audio, including sound effects, singing voices, and music. While these capabilities foster creativity and content production, they also introduce significant security and trust challenges, as realistic audio deepfakes can now be generated and disseminated at scale. Existing audio deepfake detection (ADD) countermeasures (CMs) and benchmarks, however, remain largely speech-centric, often relying on speech-specific artifacts and exhibiting limited robustness to real-world distortions, as well as restricted generalization to heterogeneous audio types and emerging spoofing techniques. To address these gaps, we propose the All-Type Audio Deepfake Detection (AT-ADD) Grand Challenge for ACM Multimedia 2026, designed to bridge controlled academic evaluation with practical multimedia forensics. AT-ADD comprises two tracks: (1) Robust Speech Deepfake Detection, which evaluates detectors under real-world scenarios and against unseen, state-of-the-art speech generation methods; and (2) All-Type Audio Deepfake Detection, which extends detection beyond speech to diverse, unknown audio types and promotes type-agnostic generalization across speech, sound, singing, and music. By providing standardized datasets, rigorous evaluation protocols, and reproducible baselines, AT-ADD aims to accelerate the development of robust and generalizable audio forensic technologies, supporting secure communication, reliable media verification, and responsible governance in an era of pervasive synthetic audio.

Method: Proposes CapTalk, a unified caption-conditioned text-audio autoregressive framework using utterance-level captions for single-utterance voice design and speaker-level captions for dialogue speaker modeling. Introduces CoT control sequence for turn-level dynamic attributes planning and hierarchical variational conditioning module with utterance-level speaker encoder to balance timbre preservation and context-adaptive expression.

Result: CapTalk achieves state-of-the-art performance on single-utterance voice design benchmark and delivers better expression controllability and contextual appropriateness in multi-turn dialogue compared to existing methods.

Conclusion: The work successfully extends voice design to dialogue settings, enabling better speaker modeling and turn-level expressive control through a unified framework that balances stable timbre preservation with context-adaptive expression.

Abstract: Voice design from natural language descriptions is emerging as a new task in text-to-speech multimodal generation, aiming to synthesize speech with target timbre and speaking style without relying on reference audio. However, existing methods mainly focus on single-utterance generation, leaving conversational voice design largely unexplored. In this work, we extend voice design to dialogue, enabling better target speaker modeling and turn-level expressive control in natural conversational settings. We propose CapTalk, a unified caption-conditioned text-audio autoregressive framework for both single-utterance and dialogue voice design. CapTalk uses utterance-level captions for single-utterance voice design and speaker-level captions for dialogue speaker modeling, and further introduces a CoT control sequence in dialogue to explicitly plan turn-level dynamic attributes. To resolve the conflict between stable timbre preservation and context-adaptive expression, we propose a hierarchical variational conditioning module with an utterance-level speaker encoder to better balance stable timbre preservation and context-adaptive expression. This enables timbre reuse while keeping expression adaptive to the current utterance and, in dialogue, the surrounding context. We also build a comprehensive evaluation protocol for both single-utterance and dialogue settings. Experiments show that CapTalk achieves state-of-the-art performance on a single-utterance voice design benchmark and delivers better expression controllability and contextual appropriateness in multi-turn dialogue. Audio samples are available at: https://anonymous.4open.science/api/repo/Captalk-D601/file/index.html.

Method: AudioMoG analyzes CFG and AG to understand their advantages/limitations, then proposes a mixture-of-guidance framework that integrates diverse guidance signals with their interaction terms (including unconditional bad versions of the model) to maximize cumulative advantages. This improves sampling quality without additional training.

Result: Experiments show AudioMoG consistently outperforms single guidance methods in text-to-audio generation across different sampling steps, while also showing advantages in video-to-audio, text-to-music, and image generation tasks, all at the same inference speed.

Conclusion: AudioMoG provides an effective sampling enhancement for diffusion-based audio generation that combines the benefits of different guidance approaches, improving quality without compromising diversity or requiring retraining resources.

Abstract: The design of diffusion-based audio generation systems has been investigated from diverse perspectives, such as data space, network architecture, and conditioning techniques, while most of these innovations require model re-training. In sampling, classifier-free guidance (CFG) has been uniformly adopted to enhance generation quality by strengthening condition alignment. However, CFG often compromises diversity, resulting in suboptimal performance. Although the recent autoguidance (AG) method proposes another direction of guidance that maintains diversity, its direct application in audio generation has so far underperformed CFG. In this work, we introduce AudioMoG, an improved sampling method that enhances text-to-audio (T2A) and video-to-audio (V2A) generation quality without requiring extensive training resources. We start with an analysis of both CFG and AG, examining their respective advantages and limitations for guiding diffusion models. Building upon our insights, we introduce a mixture-of-guidance framework that integrates diverse guidance signals with their interaction terms (e.g., the unconditional bad version of the model) to maximize cumulative advantages. Experiments show that, given the same inference speed, our approach consistently outperforms single guidance in T2A generation across sampling steps, concurrently showing advantages in V2A, text-to-music, and image generation. Demo samples are available at: https://audiomog.github.io.

Method: Proposes a novel framework with: 1) A sampling mechanism to retrieve proxy targets balancing diversity and similarity, 2) A lightweight translation model trained with a joint objective that encourages mood transformation while preserving other attributes, 3) Embedding-based transformation approach using mood labels as guidance.

Result: Extensive experiments on two datasets show strong mood transformation performance while retaining genre and instrumentation far better than training-free baselines. The framework establishes controllable embedding transformation as effective for personalized music retrieval.

Conclusion: Controllable embedding transformation is a promising paradigm for personalized music retrieval, enabling fine-grained attribute adjustment while preserving other musical characteristics. The proposed mood-guided framework successfully addresses the limitation of traditional embeddings.

Abstract: Music representations are the backbone of modern recommendation systems, powering playlist generation, similarity search, and personalized discovery. Yet most embeddings offer little control for adjusting a single musical attribute, e.g., changing only the mood of a track while preserving its genre or instrumentation. In this work, we address the problem of controllable music retrieval through embedding-based transformation, where the objective is to retrieve songs that remain similar to a seed track but are modified along one chosen dimension. We propose a novel framework for mood-guided music embedding transformation, which learns a mapping from a seed audio embedding to a target embedding guided by mood labels, while preserving other musical attributes. Because mood cannot be directly altered in the seed audio, we introduce a sampling mechanism that retrieves proxy targets to balance diversity with similarity to the seed. We train a lightweight translation model using this sampling strategy and introduce a novel joint objective that encourages transformation and information preservation. Extensive experiments on two datasets show strong mood transformation performance while retaining genre and instrumentation far better than training-free baselines, establishing controllable embedding transformation as a promising paradigm for personalized music retrieval.

Method: Proposes Neural Spectral Transport Representation (NSTR) with learnable local spectrum field S(x) and frequency transport network F_θ enforcing ∇S(x) ≈ F_θ(x, S(x)), reconstructing signals by spatially modulating global sinusoidal bases.

Result: NSTR achieves significantly better accuracy-parameter trade-offs than SIREN, Fourier-feature MLPs, and Instant-NGP on 2D image regression, audio reconstruction, and implicit 3D geometry tasks, requiring fewer global frequencies and converging faster.

Conclusion: NSTR introduces explicit modeling of space-varying spectrum in INRs, offering better performance and interpretability through spectral transport fields visualization.

Abstract: Implicit Neural Representations (INRs) have emerged as a powerful paradigm for representing signals such as images, audio, and 3D scenes. However, existing INR frameworks – including MLPs with Fourier features, SIREN, and multiresolution hash grids – implicitly assume a \textit{global and stationary} spectral basis. This assumption is fundamentally misaligned with real-world signals whose frequency characteristics vary significantly across space, exhibiting local high-frequency textures, smooth regions, and frequency drift phenomena. We propose \textbf{Neural Spectral Transport Representation (NSTR)}, the first INR framework that \textbf{explicitly models a spatially varying local frequency field}. NSTR introduces a learnable \emph{frequency transport equation}, a PDE that governs how local spectral compositions evolve across space. Given a learnable local spectrum field $S(x)$ and a frequency transport network $F_θ$ enforcing $\nabla S(x) \approx F_θ(x, S(x))$, NSTR reconstructs signals by spatially modulating a compact set of global sinusoidal bases. This formulation enables strong local adaptivity and offers a new level of interpretability via visualizing frequency flows. Experiments on 2D image regression, audio reconstruction, and implicit 3D geometry show that NSTR achieves significantly better accuracy-parameter trade-offs than SIREN, Fourier-feature MLPs, and Instant-NGP. NSTR requires fewer global frequencies, converges faster, and naturally explains signal structure through spectral transport fields. We believe NSTR opens a new direction in INR research by introducing explicit modeling of space-varying spectrum.

Method: Models operate as independent agents with $10,000 each, making autonomous trading decisions every 15-45 minutes on live prediction markets (Kalshi and Polymarket). Two cohorts were evaluated: Cohort 1 (6 frontier models in live trading over 57 days) and Cohort 2 (4 next-generation models in paper trading over 3 days).

Result: Cohort 1 models showed negative returns on Kalshi (-16.0% to -30.8%) but performed better on Polymarket (-1.1% average). grok-4-20-checkpoint achieved 71.4% settlement win rate on Polymarket. gemini-3.1-pro-preview achieved +6.02% return on Polymarket in 3 days - the best performance across all models. Performance hierarchy was driven by initial prediction accuracy and ability to capitalize on correct predictions, not research volume.

Conclusion: Platform design significantly impacts which AI models succeed in prediction markets, with different models excelling on different platforms. The benchmark provides comprehensive insights into how frontier models behave under real financial pressure, revealing that prediction accuracy and execution capability are key drivers of success.

Abstract: We introduce Prediction Arena, a benchmark for evaluating AI models’ predictive accuracy and decision-making by enabling them to trade autonomously on live prediction markets with real capital. Unlike synthetic benchmarks, Prediction Arena tests models in environments where trades execute on actual exchanges (Kalshi and Polymarket), providing objective ground truth that cannot be gamed or overfitted. Each model operates as an independent agent starting with $10,000, making autonomous decisions every 15-45 minutes. Over a 57-day longitudinal evaluation (January 12 to March 9, 2026), we track two cohorts: six frontier models in live trading (Cohort 1, full period) and four next-generation models in paper trading (Cohort 2, 3-day preliminary). For Cohort 1, final Kalshi returns range from -16.0% to -30.8%. Our analysis identifies a clear performance hierarchy: initial prediction accuracy and the ability to capitalize on correct predictions are the main drivers, while research volume shows no correlation with outcomes. A striking cross-platform contrast emerges from parallel Polymarket live trading: Cohort 1 models averaged only -1.1% on Polymarket vs. -22.6% on Kalshi, with grok-4-20-checkpoint achieving a 71.4% settlement win rate - the highest across any platform or cohort. gemini-3.1-pro-preview (Cohort 2), which executed zero trades on Kalshi, achieved +6.02% on Polymarket in 3 days - the best return of any model across either cohort - demonstrating that platform design has a profound effect on which models succeed. Beyond performance, we analyze computational efficiency (token usage, cycle time), settlement accuracy, exit patterns, and market preferences, providing a comprehensive view of how frontier models behave under real financial pressure.

Method: BLEG uses a three-stage approach: 1) Prompt LLMs to generate augmented texts for fMRI graph data, 2) Design LLM-LM instruction tuning for enhanced textual representations at lower cost, with GNN trained for coarsened alignment, 3) Finetune an adapter after GNN for downstream tasks with alignment loss between LM and GNN logits.

Result: Extensive experiments on different datasets confirmed BLEG’s superiority over existing methods, demonstrating improved performance in brain network analysis tasks.

Conclusion: BLEG successfully integrates LLMs with GNNs for brain network analysis, leveraging LLMs’ representation power to enhance GNN performance without expensive direct tuning, offering a promising direction for multimodal neuroscience research.

Abstract: Graph Neural Networks (GNNs) have been widely used in diverse brain network analysis tasks based on preprocessed functional magnetic resonance imaging (fMRI) data. However, their performances are constrained due to high feature sparsity and inherent limitations of domain knowledge within uni-modal neurographs. Meanwhile, large language models (LLMs) have demonstrated powerful representation capabilities. Combining LLMs with GNNs presents a promising direction for brain network analysis. While LLMs and MLLMs have emerged in neuroscience, integration of LLMs with graph-based data remains unexplored. In this work, we deal with these issues by incorporating LLM’s powerful representation and generalization capabilities. Considering great cost for directly tuning LLMs, we instead function LLM as enhancer to boost GNN’s performance on downstream tasks. Our method, namely BLEG, can be divided into three stages. We firstly prompt LLM to get augmented texts for fMRI graph data, then we design a LLM-LM instruction tuning method to get enhanced textual representations at a relatively lower cost. GNN is trained together for coarsened alignment. Finally we finetune an adapter after GNN for given downstream tasks. Alignment loss between LM and GNN logits is designed to further enhance GNN’s representation. Extensive experiments on different datasets confirmed BLEG’s superiority.

Method: A two-phase framework: 1) Offline Phase uses LLMs to semantically generate structured fault scenarios and LDMs to synthesize high-fidelity sensor degradations, then distills these complex fault dynamics into a pre-computed lookup table. 2) Online Phase enables edge devices to perform real-time fault-aware inference using the lightweight lookup table without running heavy AI models locally.

Result: Validated on ResNet18 lane-following model across 460 fault scenarios. While baseline R^2 was ~0.85 on clean data, generated faults exposed significant robustness degradation: RMSE increased by up to 99% and within-0.10 localization accuracy dropped to as low as 31.0% under fog conditions.

Conclusion: The framework demonstrates that normal-data evaluation is inadequate for real-world edge AI deployment, and the decoupled approach enables comprehensive safety validation while maintaining real-time performance on resource-constrained edge devices.

Abstract: Deploying autonomous vision systems on edge devices faces a critical challenge: resource constraints prevent real-time and predictable execution of comprehensive safety tests. Existing validation methods depend on static datasets or manual fault injection, failing to capture the diverse environmental hazards encountered in real-world deployment. To address this, we introduce a decoupled offline-online fault injection framework. This architecture separates the validation process into two distinct phases: a computationally intensive Offline Phase and a lightweight Online Phase. In the offline phase, we employ Large Language Models (LLMs) to semantically generate structured fault scenarios and Latent Diffusion Models (LDMs) to synthesize high-fidelity sensor degradations. These complex fault dynamics are distilled into a pre-computed lookup table, enabling the edge device to perform real-time fault-aware inference without running heavy AI models locally. We extensively validated this framework on a ResNet18 lane-following model across 460 fault scenarios. Results show that while the model achieves a baseline R^2 of approximately 0.85 on clean data, our generated faults expose significant robustness degradation, with RMSE increasing by up to 99% and within-0.10 localization accuracy dropping to as low as 31.0% under fog conditions, demonstrating the inadequacy of normal-data evaluation for real-world edge AI deployment.

Method: Researchers designed controlled data interventions to isolate distributional effects under fixed training settings, using parameter-space diagnostics based on spectral and rank analyses to reveal structural signatures of different learning regimes.

Result: Benchmark-aligned data improves narrow evaluation metrics but limits broader representational development, while coverage-expanding data leads to more distributed parameter adaptation and better generalization. These patterns were observed across diverse model families including multimodal models.

Conclusion: Benchmark performance alone is insufficient to characterize model capability; data distribution plays a crucial role in shaping learning dynamics and generalization ability.

Abstract: Large language models often achieve strong benchmark gains without corresponding improvements in broader capability. We hypothesize that this discrepancy arises from differences in training regimes induced by data distribution. To investigate this, we design controlled data interventions that isolate distributional effects under fixed training settings. We find that benchmark-aligned data improves narrow evaluation metrics while limiting broader representational development, whereas coverage-expanding data leads to more distributed parameter adaptation and better generalization. We further introduce parameter-space diagnostics based on spectral and rank analyses, which reveal distinct structural signatures of these regimes. Similar patterns are observed across diverse open-source model families, including multimodal models as a key case study, suggesting that these effects extend beyond controlled settings. A case study on prompt repetition shows that not all data artifacts induce regime shifts. These results indicate that benchmark performance alone is insufficient to characterize model capability, and highlight the importance of data distribution in shaping learning dynamics.

Method: 1) Local Optimization (Local Opt.): Optimizes tokens within localized windows while leveraging contextual information to reduce error propagation. 2) Representation Continuity (ReCo): Inspired by Lipschitz continuity, uses continuity loss to constrain representation changes, improving model robustness and reducing error accumulation.

Result: Extensive experiments on class- and text-to-video datasets show the approach achieves superior performance to baseline while halving training cost without sacrificing quality.

Conclusion: The proposed methods effectively accelerate autoregressive video generation training while maintaining or improving video quality and consistency.

Abstract: Autoregressive models have shown superior performance and efficiency in image generation, but remain constrained by high computational costs and prolonged training times in video generation. In this study, we explore methods to accelerate training for autoregressive video generation models through empirical analyses. Our results reveal that while training on fewer video frames significantly reduces training time, it also exacerbates error accumulation and introduces inconsistencies in the generated videos. To address these issues, we propose a Local Optimization (Local Opt.) method, which optimizes tokens within localized windows while leveraging contextual information to reduce error propagation. Inspired by Lipschitz continuity, we propose a Representation Continuity (ReCo) strategy to improve the consistency of generated videos. ReCo utilizes continuity loss to constrain representation changes, improving model robustness and reducing error accumulation. Extensive experiments on class- and text-to-video datasets demonstrate that our approach achieves superior performance to the baseline while halving the training cost without sacrificing quality.

Method: Proposes “flow learners” that parameterize transport vector fields and generate trajectories through integration, aligning with continuous dynamics that define PDE evolution. This physics-to-physics approach supports continuous-time prediction and native uncertainty quantification.

Result: The paper presents a conceptual framework arguing that transport-based learning offers a stronger organizing principle for learned PDE solving, outlining a research agenda for this paradigm shift.

Conclusion: Transport-based learning (flow learners) provides better alignment with PDE physics, enabling continuous-time prediction, uncertainty quantification, and new opportunities for physics-aware solver design compared to state-prediction approaches.

Abstract: Partial differential equations (PDEs) govern nearly every physical process in science and engineering, yet solving them at scale remains prohibitively expensive. Generative AI has transformed language, vision, and protein science, but learned PDE solvers have not undergone a comparable shift. Existing paradigms each capture part of the problem. Physics-informed neural networks embed residual structure, yet they are often difficult to optimize in stiff, multiscale, or large-domain regimes. Neural operators amortize across instances, yet they commonly inherit a snapshot-prediction view of solving and can degrade over long rollouts. Diffusion-based solvers model uncertainty, yet they are often built on a solver template that still centers on state regression. We argue that the core issue is the abstraction used to train learned solvers. Many models are asked to predict states, while many scientific settings require modeling how uncertainty moves through constrained dynamics. The relevant object is transport over physically admissible futures. This motivates \emph{flow learners}: models that parameterize transport vector fields and generate trajectories through integration, echoing the continuous dynamics that define PDE evolution. This physics-to-physics alignment supports continuous-time prediction, native uncertainty quantification, and new opportunities for physics-aware solver design. We explain why transport-based learning offers a stronger organizing principle for learned PDE solving and outline the research agenda that follows from this shift.

Method: Two GAN-based models: 1) CGL-GAN uses Gaussian blur on inpainted regions, and 2) PDA-GAN combines unsupervised domain adaptation with a pixel-level discriminator connected to shallow-level feature maps. Also introduces the CGL-Dataset with 60,548 paired inpainted posters and 121,000 clean product images.

Result: PDA-GAN achieves state-of-the-art performance and generates high-quality image-aware layouts. The paper also proposes three novel content-aware metrics to evaluate the relationship between graphic elements and image content.

Conclusion: The PDA-GAN model effectively bridges the domain gap between inpainted posters and clean images, generating layouts that are aware of image content and visual texture, with proposed metrics providing better evaluation of content-aware layout generation.

Abstract: Layout plays a crucial role in graphic design and poster generation. Recently, the application of deep learning models for layout generation has gained significant attention. This paper focuses on using a GAN-based model conditioned on images to generate advertising poster graphic layouts, requiring a dataset of paired product images and layouts. To address this task, we introduce the Content-aware Graphic Layout Dataset (CGL-Dataset), consisting of 60,548 paired inpainted posters with annotations and 121,000 clean product images. The inpainting artifacts introduce a domain gap between the inpainted posters and clean images. To bridge this gap, we design two GAN-based models. The first model, CGL-GAN, uses Gaussian blur on the inpainted regions to generate layouts. The second model combines unsupervised domain adaptation by introducing a GAN with a pixel-level discriminator (PD), abbreviated as PDA-GAN, to generate image-aware layouts based on the visual texture of input images. The PD is connected to shallow-level feature maps and computes the GAN loss for each input-image pixel. Additionally, we propose three novel content-aware metrics to assess the model’s ability to capture the intricate relationships between graphic elements and image content. Quantitative and qualitative evaluations demonstrate that PDA-GAN achieves state-of-the-art performance and generates high-quality image-aware layouts.

Method: Developed a prompt-generation pipeline using GPT-4o mini to create a suite of LLM and human-generated prompts with varying intensities across four emotions (joy, encouragement, anger, insecurity). Compiled a “Gold Dataset” where human and model labels align, then evaluated LLM behavior on accuracy, sycophancy, and toxicity.

Result: Positive emotional stimuli (joy, encouragement) lead to more accurate and less toxic results, but also increase sycophantic behavior. The study provides empirical evidence on how different emotional prompts affect various aspects of LLM performance.

Conclusion: Emotional prompting significantly influences LLM behavior, with positive emotions improving some aspects of performance while introducing trade-offs like increased sycophancy. The findings highlight the importance of considering emotional dimensions in prompt engineering.

Abstract: Emotional prompting - the use of specific emotional diction in prompt engineering - has shown increasing promise in improving large language model (LLM) performance, truthfulness, and responsibility. However these studies have been limited to single types of positive emotional stimuli and have not considered varying degrees of emotion intensity in their analyses. In this paper, we explore the effects of four distinct emotions - joy, encouragement, anger, and insecurity - in emotional prompting and evaluate them on accuracy, sycophancy, and toxicity. We develop a prompt-generation pipeline with GPT-4o mini to create a suite of LLM and human-generated prompts with varying intensities across the four emotions. Then, we compile a “Gold Dataset” of prompts where human and model labels align. Our empirical evaluation on LLM behavior suggests that positive emotional stimuli lead to more accurate and less toxic results, but also increase sycophantic behavior.

Method: Decompose the spectral edge (dominant direction of Gram matrix of parameter updates) into gradient and weight-decay components during grokking in two sequence tasks (Dyck-1 and SCAN). Analyze using gap flow equation and nonlinear probes to track information encoding.

Result: Found sharp two-phase lifecycle: pre-grokking edge is gradient-driven and functionally active; at grokking, gradient and weight decay align, creating compression axis that is perturbation-flat but ablation-critical (4000x more impactful). Three universality classes emerge predicted by gap flow equation. Nonlinear probes show information is re-encoded, not lost.

Conclusion: Weight decay drives compression during grokking while preserving algorithmic information through re-encoding rather than loss, with compression being reversible post-grokking while maintaining algorithm functionality.

Abstract: We decompose the spectral edge – the dominant direction of the Gram matrix of parameter updates – into its gradient and weight-decay components during grokking in two sequence tasks (Dyck-1 and SCAN). We find a sharp two-phase lifecycle: before grokking the edge is gradient-driven and functionally active; at grokking, gradient and weight decay align, and the edge becomes a compression axis that is perturbation-flat yet ablation-critical (>4000x more impactful than random directions). Three universality classes emerge (functional, mixed, compression), predicted by the gap flow equation. Nonlinear probes show information is re-encoded, not lost (MLP $R^2=0.99$ where linear $R^2=0.86$), and removing weight decay post-grok reverses compression while preserving the algorithm.

Method: 1) Create benchmark with comprehensive relation corpus covering all potential real-world relations, 2) Introduce new audio event corpus of commonly heard audios, 3) Propose new evaluation metrics, 4) Develop finetuning framework to enhance existing TTA models

Result: Established first systematic benchmark for audio event relation modeling in TTA generation with comprehensive datasets and evaluation metrics, plus framework to improve existing models

Conclusion: This work addresses a critical gap in TTA generation by systematically studying and enhancing audio event relation modeling, providing foundational resources for future research

Abstract: Despite significant advancements in Text-to-Audio (TTA) generation models achieving high-fidelity audio with fine-grained context understanding, they struggle to model the relations between audio events described in the input text. However, previous TTA methods have not systematically explored audio event relation modeling, nor have they proposed frameworks to enhance this capability. In this work, we systematically study audio event relation modeling in TTA generation models. We first establish a benchmark for this task by: 1. proposing a comprehensive relation corpus covering all potential relations in real-world scenarios; 2. introducing a new audio event corpus encompassing commonly heard audios; and 3. proposing new evaluation metrics to assess audio event relation modeling from various perspectives. Furthermore, we propose a finetuning framework to enhance existing TTA models ability to model audio events relation. Code is available at: https://github.com/yuhanghe01/RiTTA

Method: Using geometric data analysis tools to probe affective representations in LLMs, examining alignment with psychological valence-arousal models, testing linear representation hypotheses, and leveraging the learned latent space for uncertainty quantification in emotion processing tasks.

Result: 1) LLMs learn coherent latent representations of affective emotions aligning with psychological valence-arousal models. 2) These representations exhibit nonlinear geometric structure that can be well-approximated linearly, supporting linear representation hypotheses. 3) The learned latent space can quantify uncertainty in emotion processing tasks.

Conclusion: LLMs acquire affective representations with geometric structure paralleling established models of human emotion, with practical implications for model interpretability and safety.

Abstract: The geometric structure of latent representations in large language models (LLMs) is an active area of research, driven in part by its implications for model transparency and AI safety. Existing literature has focused mainly on general geometric and topological properties of the learnt representations, but due to a lack of ground-truth latent geometry, validating the findings of such approaches is challenging. Emotion processing provides an intriguing testbed for probing representational geometry, as emotions exhibit both categorical organization and continuous affective dimensions, which are well-established in the psychology literature. Moreover, understanding such representations carries safety relevance. In this work, we investigate the latent structure of affective representations in LLMs using geometric data analysis tools. We present three main findings. First, we show that LLMs learn coherent latent representations of affective emotions that align with widely used valence–arousal models from psychology. Second, we find that these representations exhibit nonlinear geometric structure that can nonetheless be well-approximated linearly, providing empirical support for the linear representation hypothesis commonly assumed in model transparency methods. Third, we demonstrate that the learned latent representation space can be leveraged to quantify uncertainty in emotion processing tasks. Our findings suggest that LLMs acquire affective representations with geometric structure paralleling established models of human emotion, with practical implications for model interpretability and safety.

Method: SCOT uses Sinkhorn-based entropic optimal transport to learn explicit soft correspondences between unequal region sets. It sharpens transferable structure with an OT-weighted contrastive objective and stabilizes optimization through a cycle-style reconstruction regularizer. For multi-source transfer, it aligns each source and target to a shared prototype hub using balanced entropic transport guided by a target-induced prototype prior.

Result: SCOT consistently improves transfer accuracy and robustness across real-world cities and tasks. The learned transport couplings and hub assignments provide interpretable diagnostics of alignment quality.

Conclusion: SCOT effectively addresses cross-city transfer challenges by learning explicit soft correspondences through optimal transport, providing both improved performance and interpretable alignment diagnostics.

Abstract: Cross-city transfer improves prediction in label-scarce cities by leveraging labeled data from other cities, but it becomes challenging when cities adopt incompatible partitions and no ground-truth region correspondences exist. Existing approaches either rely on heuristic region matching, which is often sensitive to anchor choices, or perform distribution-level alignment that leaves correspondences implicit and can be unstable under strong heterogeneity. We propose SCOT, a cross-city representation learning framework that learns explicit soft correspondences between unequal region sets via Sinkhorn-based entropic optimal transport. SCOT further sharpens transferable structure with an OT-weighted contrastive objective and stabilizes optimization through a cycle-style reconstruction regularizer. For multi-source transfer, SCOT aligns each source and the target to a shared prototype hub using balanced entropic transport guided by a target-induced prototype prior. Across real-world cities and tasks, SCOT consistently improves transfer accuracy and robustness, while the learned transport couplings and hub assignments provide interpretable diagnostics of alignment quality.

Method: Uses Restless Multi-Armed Bandit (RMAB) framework with Decision-Focused Learning (DFL) approach that directly aligns learning with engagement maximization, moving from traditional Two-Stage “predict-then-optimize” to end-to-end optimization.

Result: DFL policy reduced cumulative engagement drops by 31% relative to standard care, significantly outperforming Two-Stage model, with improvements translating to statistically significant real-world health behavior changes (continued consumption of iron/calcium supplements).

Conclusion: SAHELI project provides scalable blueprint for applying sequential decision-making AI to optimize resource allocation in health programs, demonstrating practical impact on maternal health outcomes.

Abstract: Maternal and child health is a critical concern around the world. In many global health programs disseminating preventive care and health information, limited healthcare worker resources prevent continuous, personalised engagement with vulnerable beneficiaries. In such scenarios, it becomes crucial to optimally schedule limited live-service resources to maximise long-term engagement. To address this fundamental challenge, the multi-year SAHELI project (2020-2025), in collaboration with partner NGO ARMMAN, leverages AI to allocate scarce resources in a maternal and child health program in India. The SAHELI system solves this sequential resource allocation problem using a Restless Multi-Armed Bandit (RMAB) framework. A key methodological innovation is the transition from a traditional Two-Stage “predict-then-optimize” approach to Decision-Focused Learning (DFL), which directly aligns the framework’s learning method with the ultimate goal of maximizing beneficiary engagement. Empirical evaluation through large-scale randomized controlled trials demonstrates that the DFL policy reduced cumulative engagement drops by 31% relative to the current standard of care, significantly outperforming the Two-Stage model. Crucially, the studies also confirmed that this increased program engagement translates directly into statistically significant improvements in real-world health behaviors, notably the continued consumption of vital iron and calcium supplements by new mothers. Ultimately, the SAHELI project provides a scalable blueprint for applying sequential decision-making AI to optimize resource allocation in health programs.

Method: Combines ModernBERT encoder with hash embeddings, depth-aware token representations, and attention pooling classification head. Processes ASCII frame representations and depth maps to select game actions at 31ms per decision. Trained on only 31,000 human gameplay demonstrations.

Result: Achieves 178 frags in 10 episodes (17.8 per episode) in defend_the_center scenario, outperforming all tested LLMs combined (13 frags total). Only agent that actively engages enemies rather than evading. Runs at 31ms per decision, 92,000x smaller than Nemotron-120B.

Conclusion: Small, domain-specific models trained on appropriate data can decisively outperform general-purpose LLMs at real-time control tasks with dramatically lower inference costs and practical deployment on consumer hardware.

Abstract: We present SauerkrautLM-Doom-MultiVec, a 1.3 million parameter model that plays the classic first-person shooter DOOM in real time, outperforming large language models up to 92,000x its size, including Nemotron-120B, Qwen3.5-27B, and GPT-4o-mini. Our model combines a ModernBERT encoder with hash embeddings, depth-aware token representations, and an attention pooling classification head to select game actions from ASCII frame representations at 31ms per decision. Trained on just 31,000 human gameplay demonstrations, it achieves 178 frags in 10 episodes (17.8 per episode) in the defend_the_center scenario, more than all tested LLMs combined (13 frags total). All agents receive equivalent input: ASCII frames and depth maps. Despite having 92,000x fewer parameters than Nemotron-120B, our model is the only agent that actively engages enemies rather than purely evading them. These results demonstrate that small, task-specific models trained on domain-appropriate data can decisively outperform general-purpose LLMs at real-time control tasks, at a fraction of the inference cost, with deployment capability on consumer hardware.

Method: Proposes a quantum-classical comparison framework evaluating four paradigms: quantum models, classical ML baselines, and two hybrid quantum-classical architectures using 16-year Bangladesh crime statistics with rigorous cross-validation.

Result: Quantum-inspired approaches (particularly QAOA) achieve up to 84.6% accuracy with fewer trainable parameters than classical baselines, and hybrid approaches show competitive training efficiency for resource-constrained environments.

Conclusion: The framework demonstrates practical advantages for memory-constrained edge deployment and smart city surveillance systems, providing preliminary empirical assessment of quantum-enhanced ML for structured crime data.

Abstract: Crime pattern analysis is critical for law enforcement and predictive policing, yet the surge in criminal activities from rapid urbanization creates high-dimensional, imbalanced datasets that challenge traditional classification methods. This study presents a quantum-classical comparison framework for crime analytics, evaluating four computational paradigms: quantum models, classical baseline machine learning models, and two hybrid quantum-classical architectures. Using 16-year Bangladesh crime statistics, we systematically assess classification performance and computational efficiency under rigorous cross-validation methods. Experimental results show that quantum-inspired approaches, particularly QAOA, achieve up to 84.6% accuracy, while requiring fewer trainable parameters than classical baselines, suggesting practical advantages for memory-constrained edge deployment. The proposed correlation-aware circuit design demonstrates the potential of incorporating domain-specific feature relationships into quantum models. Furthermore, hybrid approaches exhibit competitive training efficiency, making them suitable candidates for resource-constrained environments. The framework’s low computational overhead and compact parameter footprint suggest potential advantages for wireless sensor network deployments in smart city surveillance systems, where distributed nodes perform localized crime analytics with minimal communication costs. Our findings provide a preliminary empirical assessment of quantum-enhanced machine learning for structured crime data and motivate further investigation with larger datasets and realistic quantum hardware considerations.

Method: Proposes layer-wise parameterization where each layer satisfies a linear matrix inequality (LMI) ensuring dissipativity. LipKernel directly parameterizes dissipative convolution kernels using 2-D Roesser-type state space models, accommodating various CNN layers (1-D/2-D convolutions, pooling, strided/dilated convolutions, zero padding).

Result: Run-time is orders of magnitude faster than state-of-the-art Lipschitz-bounded networks that use Fourier domain parameterization. The approach maintains computational efficiency without overhead during evaluation while providing stronger robustness guarantees.

Conclusion: LipKernel offers an efficient, expressive method for building robust CNNs with guaranteed Lipschitz bounds, particularly suitable for real-time perception and control applications in robotics and autonomous systems.

Abstract: We propose a novel layer-wise parameterization for convolutional neural networks (CNNs) that includes built-in robustness guarantees by enforcing a prescribed Lipschitz bound. Each layer in our parameterization is designed to satisfy a linear matrix inequality (LMI), which in turn implies dissipativity with respect to a specific supply rate. Collectively, these layer-wise LMIs ensure Lipschitz boundedness for the input-output mapping of the neural network, yielding a more expressive parameterization than through spectral bounds or orthogonal layers. Our new method LipKernel directly parameterizes dissipative convolution kernels using a 2-D Roesser-type state space model. This means that the convolutional layers are given in standard form after training and can be evaluated without computational overhead. In numerical experiments, we show that the run-time using our method is orders of magnitude faster than state-of-the-art Lipschitz-bounded networks that parameterize convolutions in the Fourier domain, making our approach particularly attractive for improving the robustness of learning-based real-time perception or control in robotics, autonomous vehicles, or automation systems. We focus on CNNs, and in contrast to previous works, our approach accommodates a wide variety of layers typically used in CNNs, including 1-D and 2-D convolutional layers, maximum and average pooling layers, as well as strided and dilated convolutions and zero padding. However, our approach naturally extends beyond CNNs as we can incorporate any layer that is incrementally dissipative.

Method: Proposes GFM-RA with interference-aware Transformer architecture using bias projector to inject interference topologies into attention mechanisms. Uses hybrid self-supervised pre-training combining masked edge prediction with negative-free Teacher-Student contrastive learning.

Result: Achieves state-of-the-art performance, scales effectively with model capacity, and exhibits exceptional sample efficiency for few-shot adaptation to diverse downstream objectives in out-of-distribution scenarios.

Conclusion: Demonstrates promise of pre-trained foundation models for adaptable wireless resource allocation and provides foundation for future research on generalizable learning-based wireless optimization.

Abstract: The aggressive densification of modern wireless networks necessitates judicious resource allocation to mitigate severe mutual interference. However, classical iterative algorithms remain computationally prohibitive for real-time applications requiring rapid responsiveness. While recent deep learning-based methods show promise, they typically function as task-specific solvers lacking the flexibility to adapt to different objectives and scenarios without expensive retraining. To address these limitations, we propose a graph foundation model for resource allocation (GFM-RA) based on a pre-training and fine-tuning paradigm to extract unified representations, thereby enabling rapid adaptation to different objectives and scenarios. Specifically, we introduce an interference-aware Transformer architecture with a bias projector that injects interference topologies into global attention mechanisms. Furthermore, we develop a hybrid self-supervised pre-training strategy that synergizes masked edge prediction with negative-free Teacher-Student contrastive learning, enabling the model to capture transferable structural representations from massive unlabeled datasets. Extensive experiments demonstrate that the proposed framework achieves state-of-the-art performance and scales effectively with increased model capacity. Crucially, leveraging its unified representations, the foundation model exhibits exceptional sample efficiency, enabling robust few-shot adaptation to diverse and unsupervised downstream objectives in out-of-distribution (OOD) scenarios. These results demonstrate the promise of pre-trained foundation models for adaptable wireless resource allocation and provide a strong foundation for future research on generalizable learning-based wireless optimization.

Method: Proposes a framework representing environments as structured semantic events encoded into permutation-invariant latent representations. Decision-making via retrieval over knowledge bank of prior experiences, with final action computed as weighted combination of retrieved solutions. Incorporates physics-informed knowledge into retrieval process.

Result: Experimental evaluation in UAV flight scenarios shows the framework operates within real-time control constraints while maintaining interpretable and consistent behavior.

Conclusion: The event-centric world modeling with memory-augmented retrieval enables structured abstraction of dynamic environments and supports interpretable decision-making through case-based reasoning while ensuring physical consistency.

Abstract: Autonomous agents operating in dynamic and safety-critical environments require decision-making frameworks that are both computationally efficient and physically grounded. However, many existing approaches rely on end-to-end learning, which often lacks interpretability and explicit mechanisms for ensuring consistency with physical constraints. In this work, we propose an event-centric world modeling framework with memory-augmented retrieval for embodied decision-making. The framework represents the environment as a structured set of semantic events, which are encoded into a permutation-invariant latent representation. Decision-making is performed via retrieval over a knowledge bank of prior experiences, where each entry associates an event representation with a corresponding maneuver. The final action is computed as a weighted combination of retrieved solutions, providing a transparent link between decision and stored experiences. The proposed design enables structured abstraction of dynamic environments and supports interpretable decision-making through case-based reasoning. In addition, incorporating physics-informed knowledge into the retrieval process encourages the selection of maneuvers that are consistent with observed system dynamics. Experimental evaluation in UAV flight scenarios demonstrates that the framework operates within real-time control constraints while maintaining interpretable and consistent behavior.

Method: Dual-stream architecture: first stream models statistical temporal evolution of individual variables; second stream focuses on residual dynamics with two key mechanisms: Adaptive Window module for flow-dependent transport delays, and Physics-Guided Dynamic Graph that incorporates physical priors to learn time-varying interaction structures while suppressing spurious correlations.

Result: Experiments on four industrial benchmarks show DSPR consistently improves forecasting accuracy and robustness under regime shifts while maintaining strong physical plausibility, achieving state-of-the-art performance with Mean Conservation Accuracy exceeding 99% and Total Variation Ratio up to 97.2%.

Conclusion: Architectural decoupling with physics-consistent inductive biases offers an effective path toward trustworthy industrial time-series forecasting, and DSPR’s robust performance in long-term deployment bridges the gap between advanced forecasting models and trustworthy autonomous control systems.

Abstract: Accurate forecasting of industrial time series requires balancing predictive accuracy with physical plausibility under non-stationary operating conditions. Existing data-driven models often achieve strong statistical performance but struggle to respect regime-dependent interaction structures and transport delays inherent in real-world systems. To address this challenge, we propose DSPR (Dual-Stream Physics-Residual Networks), a forecasting framework that explicitly decouples stable temporal patterns from regime-dependent residual dynamics. The first stream models the statistical temporal evolution of individual variables. The second stream focuses on residual dynamics through two key mechanisms: an Adaptive Window module that estimates flow-dependent transport delays, and a Physics-Guided Dynamic Graph that incorporates physical priors to learn time-varying interaction structures while suppressing spurious correlations. Experiments on four industrial benchmarks spanning heterogeneous regimes demonstrate that DSPR consistently improves forecasting accuracy and robustness under regime shifts while maintaining strong physical plausibility. It achieves state-of-the-art predictive performance, with Mean Conservation Accuracy exceeding 99% and Total Variation Ratio reaching up to 97.2%. Beyond forecasting, the learned interaction structures and adaptive lags provide interpretable insights that are consistent with known domain mechanisms, such as flow-dependent transport delays and wind-to-power scaling behaviors. These results suggest that architectural decoupling with physics-consistent inductive biases offers an effective path toward trustworthy industrial time-series forecasting. Furthermore, DSPR’s demonstrated robust performance in long-term industrial deployment bridges the gap between advanced forecasting models and trustworthy autonomous control systems.

Method: Introduces Flux Attention with a lightweight Layer Router integrated into frozen pretrained LLMs. The router dynamically routes each layer to either Full Attention (FA) or Sparse Attention (SA) based on input context, enabling layer-wise adaptive computation while maintaining contiguous memory access.

Result: Achieves up to 2.8× speedup in prefill stage and 2.0× speedup in decode stage across multiple long-context and mathematical reasoning benchmarks, with superior performance-speed trade-offs compared to baselines. Requires only 12 hours of training on 8×A800 GPUs.

Conclusion: Flux Attention provides an efficient parameter-efficient solution for accelerating LLMs in long-context scenarios by dynamically optimizing attention computation at layer level, translating theoretical computational reductions into practical wall-clock speedups.

Abstract: The quadratic computational complexity of standard attention mechanisms presents a severe scalability bottleneck for LLMs in long-context scenarios. While hybrid attention mechanisms combining Full Attention (FA) and Sparse Attention (SA) offer a potential solution, existing methods typically rely on static allocation ratios that fail to accommodate the variable retrieval demands of different tasks. Furthermore, head-level dynamic sparsity often introduces severe computational load imbalance and synchronization long-tails, which hinder hardware acceleration during autoregressive decoding. To bridge this gap, we introduce Flux Attention, a context-aware framework that dynamically optimizes attention computation at the layer level. By integrating a lightweight Layer Router into frozen pretrained LLMs, the proposed method adaptively routes each layer to FA or SA based on the input context. This layer-wise routing preserves high-fidelity information retrieval while ensuring contiguous memory access, translating theoretical computational reductions into practical wall-clock speedups. As a parameter-efficient approach, our framework requires only 12 hours of training on 8$\times$A800 GPUs. Extensive experiments across multiple long-context and mathematical reasoning benchmarks demonstrate that Flux Attention achieves a superior trade-off between performance and inference speed compared with baseline models, with speed improvements of up to $2.8\times$ and $2.0\times$ in the prefill and decode stages.

Method: Proposes Data Warmup with two key components: 1) Semantic-aware complexity metric combining foreground dominance (salient object coverage) and foreground typicality (match to learned visual prototypes), 2) Temperature-controlled sampler that prioritizes low-complexity images early and anneals toward uniform sampling. Requires only ~10 minutes of one-time preprocessing.

Result: On ImageNet 256x256 with SiT backbones (S/2 to XL/2), improves IS by up to 6.11 and FID by up to 3.41, reaching baseline quality tens of thousands of iterations earlier. Reversing the curriculum degrades performance below uniform baseline, confirming simple-to-complex ordering drives gains.

Conclusion: Data Warmup provides an effective curriculum learning strategy for diffusion models that improves training efficiency and quality without modifying model architecture or loss function, and can combine with other acceleration techniques.

Abstract: A key inefficiency in diffusion training occurs when a randomly initialized network, lacking visual priors, encounters gradients from the full complexity spectrum–most of which it lacks the capacity to resolve. We propose Data Warmup, a curriculum strategy that schedules training images from simple to complex without modifying the model or loss. Each image is scored offline by a semantic-aware complexity metric combining foreground dominance (how much of the image salient objects occupy) and foreground typicality (how closely the salient content matches learned visual prototypes). A temperature-controlled sampler then prioritizes low-complexity images early and anneals toward uniform sampling. On ImageNet 256x256 with SiT backbones (S/2 to XL/2), Data Warmup improves IS by up to 6.11 and FID by up to 3.41, reaching baseline quality tens of thousands of iterations earlier. Reversing the curriculum (exposing hard images first) degrades performance below the uniform baseline, confirming that the simple-to-complex ordering itself drives the gains. The method combines with orthogonal accelerators such as REPA and requires only ~10 minutes of one-time preprocessing with zero per-iteration overhead.

Method: Integrates critical patch sampling (CPS) for task-aware token reduction and decoupled prompt and classifier training (DPCT) to reduce backpropagation overhead, targeting training-time memory usage and computational cost.

Result: Improves peak memory, training time, and energy efficiency by about 1.6x over CODA-Prompt baseline while maintaining accuracy within 2% of state-of-the-art C-Prompt on average, competitive with CODA-Prompt in accuracy.

Conclusion: CPS-Prompt provides an efficient solution for on-device continual learning by balancing accuracy with training-time efficiency through sparse prompting techniques.

Abstract: Continual learning (CL) on edge devices requires not only high accuracy but also training-time efficiency to support on-device adaptation under strict memory and computational constraints. While prompt-based continual learning (PCL) is parameter-efficient and achieves competitive accuracy, prior work has focused mainly on accuracy or inference-time performance, often overlooking the memory and computational costs of on-device training. In this paper, we propose CPS-Prompt, a critical patch-aware sparse prompting framework that explicitly targets training-time memory usage and computational cost by integrating critical patch sampling (CPS) for task-aware token reduction and decoupled prompt and classifier training (DPCT) to reduce backpropagation overhead. Experiments on three public benchmarks and real edge hardware show that CPS-Prompt improves peak memory, training time, and energy efficiency by about 1.6x over the balanced CODA-Prompt baseline, while maintaining accuracy within 2% of the state-of-the-art C-Prompt on average and remaining competitive with CODA-Prompt in accuracy. The code is available at https://github.com/laymond1/cps-prompt.

Method: Analyze gradient flow on L-layer ReLU networks without bias, showing preservation of L-1 conservation laws. Study discrete gradient descent breaking these laws with drift scaling as eta^alpha. Derive exact decomposition of drift as eta^2 * S(eta) with closed-form spectral crossover formula. Examine cross-entropy loss dynamics and identify two dynamical regimes separated by width-dependent transitions.

Result: Gradient flow preserves conservation laws confining trajectories to lower-dimensional manifolds. Discrete gradient descent breaks laws with drift scaling eta^alpha (alpha≈1.1-1.6). Derived spectral formula validated for linear (R=0.85) and ReLU (R>0.80) networks. Cross-entropy loss drives exponential Hessian spectral compression with timescale independent of training set size, self-regularizing drift exponent near alpha=1.0. Identified two dynamical regimes: perturbative sub-Edge-of-Stability and non-perturbative with extensive mode coupling. All predictions validated across 23 experiments.

Conclusion: The success of gradient descent in non-convex neural network optimization can be explained by conservation laws in continuous gradient flow and controlled breaking of these laws in discrete gradient descent, with cross-entropy loss providing natural self-regularization through Hessian spectral compression.

Abstract: Why does gradient descent reliably find good solutions in non-convex neural network optimization, despite the landscape being NP-hard in the worst case? We show that gradient flow on L-layer ReLU networks without bias preserves L-1 conservation laws C_l = ||W_{l+1}||_F^2 - ||W_l||F^2, confining trajectories to lower-dimensional manifolds. Under discrete gradient descent, these laws break with total drift scaling as eta^alpha where alpha is approximately 1.1-1.6 depending on architecture, loss function, and width. We decompose this drift exactly as eta^2 * S(eta), where the gradient imbalance sum S(eta) admits a closed-form spectral crossover formula with mode coefficients c_k proportional to e_k(0)^2 * lambda{x,k}^2, derived from first principles and validated for both linear (R=0.85) and ReLU (R>0.80) networks. For cross-entropy loss, softmax probability concentration drives exponential Hessian spectral compression with timescale tau = Theta(1/eta) independent of training set size, explaining why cross-entropy self-regularizes the drift exponent near alpha=1.0. We identify two dynamical regimes separated by a width-dependent transition: a perturbative sub-Edge-of-Stability regime where the spectral formula applies, and a non-perturbative regime with extensive mode coupling. All predictions are validated across 23 experiments.

Method: Uses reinforcement learning with reward machines (RMs) to make sleep-control decisions, where RMs maintain abstract states to track QoS constraint violations over time to handle non-Markovian rewards

Result: Provides a principled, scalable approach to energy management for next-generation mobile networks under diverse traffic patterns and QoS requirements

Conclusion: The framework effectively addresses the challenge of time-averaged QoS constraints in energy-efficient mobile network management using RL with reward machines

Abstract: Energy efficiency in mobile networks is crucial for sustainable telecommunications infrastructure, particularly as network densification continues to increase power consumption. Sleep mechanisms for the components in mobile networks can reduce energy use, but deciding which components to put to sleep, when, and for how long while preserving quality of service (QoS) remains a difficult optimisation problem. In this paper, we utilise reinforcement learning with reward machines (RMs) to make sleep-control decisions that balance immediate energy savings and long-term QoS impact, i.e. time-averaged packet drop rates for deadline-constrained traffic and time-averaged minimum-throughput guarantees for constant-rate users. A challenge is that time-averaged constraints depend on cumulative performance over time rather than immediate performance. As a result, the effective reward is non-Markovian, and optimal actions depend on operational history rather than the instantaneous system state. RMs account for the history dependence by maintaining an abstract state that explicitly tracks the QoS constraint violations over time. Our framework provides a principled, scalable approach to energy management for next-generation mobile networks under diverse traffic patterns and QoS requirements.

Method: Uses a physics-informed neural operator that learns mapping from measurement data, spatial coordinates, and frequency to acoustic field quantities while inferring globally consistent surface admittance. Embeds acoustic governing equations (Helmholtz equation, linearized momentum equation, Robin boundary conditions) as physics-based regularization during training.

Result: Accurate reconstruction of both real and imaginary admittance components and reliable prediction of acoustic field quantities. Shows improved robustness to noise and sparse sampling compared to purely data-driven approaches.

Conclusion: Physics-informed neural operators offer promising potential for in situ acoustic material characterization by combining data-driven learning with physical constraints for more robust and accurate admittance estimation.

Abstract: Accurate knowledge of acoustic surface admittance or impedance is essential for reliable wave-based simulations, yet its in situ estimation remains challenging due to noise, model inaccuracies, and restrictive assumptions of conventional methods. This work presents a physics-informed neural operator approach for estimating frequency-dependent surface admittance directly from near-field measurements of sound pressure and particle velocity. A deep operator network is employed to learn the mapping from measurement data, spatial coordinates, and frequency to acoustic field quantities, while simultaneously inferring a globally consistent surface admittance spectrum without requiring an explicit forward model. The governing acoustic relations, including the Helmholtz equation, the linearized momentum equation, and Robin boundary conditions, are embedded into the training process as physics-based regularization, enabling physically consistent and noise-robust predictions while avoiding frequency-wise inversion. The method is validated using synthetically generated data from a simulation model for two planar porous absorbers under semi free-field conditions across a broad frequency range. Results demonstrate accurate reconstruction of both real and imaginary admittance components and reliable prediction of acoustic field quantities. Parameter studies confirm improved robustness to noise and sparse sampling compared to purely data-driven approaches, highlighting the potential of physics-informed neural operators for in situ acoustic material characterization.

Method: Generalize probabilistic reparameterization approach to handle non-equidistant discrete variables, enabling gradient-based optimization in fully mixed-variable settings with Gaussian process surrogates. Conduct systematic benchmarks to optimize kernel formulations and demonstrate robustness.

Result: The generalized PR method effectively optimizes highly discontinuous and discretized objective landscapes when combined with a modified BO workflow, establishing a practical framework for mixed optimization problems.

Conclusion: This work provides a practical Bayesian optimization framework for fully mixed optimization problems in natural sciences, particularly suited for autonomous laboratory settings with noise, discretization, and limited data.

Abstract: Optimizing expensive black-box objectives over mixed search spaces is a common challenge across the natural sciences. Bayesian optimization (BO) offers sample-efficient strategies through probabilistic surrogate models and acquisition functions. However, its effectiveness diminishes in mixed or high-cardinality discrete spaces, where gradients are unavailable and optimizing the acquisition function becomes computationally demanding. In this work, we generalize the probabilistic reparameterization (PR) approach of Daulton et al. to handle non-equidistant discrete variables, enabling gradient-based optimization in fully mixed-variable settings with Gaussian process (GP) surrogates. With real-world scientific optimization tasks in mind, we conduct systematic benchmarks on synthetic and experimental objectives to obtain an optimized kernel formulations and demonstrate the robustness of our generalized PR method. We additionally show that, when combined with a modified BO workflow, our approach can efficiently optimize highly discontinuous and discretized objective landscapes. This work establishes a practical BO framework for addressing fully mixed optimization problems in the natural sciences, and is particularly well suited to autonomous laboratory settings where noise, discretization, and limited data are inherent.

Method: Proposes SPAMoE with two key components: 1) Spectral-Preserving DINO Encoder that enforces a lower bound on high-to-low frequency energy ratio to prevent high-frequency collapse, and 2) Spectral Decomposition and Routing mechanism that dynamically assigns frequency bands to a Mixture-of-Experts ensemble (FNO, MNO, LNO).

Result: On ten OpenFWI sub-datasets, SPAMoE reduces average MAE by 54.1% relative to the best officially reported OpenFWI baseline, establishing a new architectural framework for learning-based full-waveform inversion.

Conclusion: SPAMoE successfully addresses frequency entanglement in multi-scale geological features through spectral-preserving encoding and adaptive frequency routing to specialized experts, significantly improving full-waveform inversion performance.

Abstract: Full-waveform inversion (FWI) is pivotal for reconstructing high-resolution subsurface velocity models but remains computationally intensive and ill-posed. While deep learning approaches promise efficiency, existing Convolutional Neural Networks (CNNs) and single-paradigm Neural Operators (NOs) struggle with one fundamental issue: frequency entanglement of multi-scale geological features. To address this challenge, we propose Spectral-Preserving Adaptive MoE (SPAMoE), a novel spectrum-aware framework for solving inverse problems with complex multi-scale structures. Our approach introduces a Spectral-Preserving DINO Encoder that enforces a lower bound on the high-to-low frequency energy ratio of the encoded representation, mitigating high-frequency collapse and stabilizing subsequent frequency-domain modeling. Furthermore, we design a novel Spectral Decomposition and Routing mechanism that dynamically assigns frequency bands to a Mixture-of-Experts (MoE) ensemble comprising FNO, MNO, and LNO. On the ten OpenFWI sub-datasets, experiments show that SPAMoE reduces the average MAE by 54.1% relative to the best officially reported OpenFWI baseline, thereby establishing a new architectural framework for learning-based full-waveform inversion.

Method: Proposes MUSIC with: 1) automatic scalable data generation pipeline, 2) vision chain-of-thought mechanism for multi-subject semantic understanding, 3) semantics-driven spatial layout planning to prevent identity entanglement, 4) training with complex subject images for chained reasoning, and 5) MSIC benchmark for evaluation.

Result: MUSIC significantly surpasses other methods in both multi- and single-subject scenarios, demonstrating superior performance on the new MSIC benchmark.

Conclusion: MUSIC effectively addresses multi-subject image generation challenges through innovative data generation, reasoning mechanisms, and layout planning, establishing a new state-of-the-art approach.

Abstract: Recent advances in text-to-image (T2I) generation have enabled visually coherent image synthesis from descriptions, but generating images containing multiple given subjects remains challenging. As the number of reference identities increases, existing methods often suffer from subject missing and semantic drift. To address this problem, we propose MUSIC, the first MLLM specifically designed for \textbf{MU}lti-\textbf{S}ubject \textbf{I}n-\textbf{C}ontext image generation. To overcome the data scarcity, we introduce an automatic and scalable data generation pipeline that eliminates the need for manual annotation. Furthermore, we enhance the model’s understanding of multi-subject semantic relationships through a vision chain-of-thought (CoT) mechanism, guiding step-by-step reasoning from subject images to semantics and generation. To mitigate identity entanglement and manage visual complexity, we develop a novel semantics-driven spatial layout planning method and demonstrate its test-time scalability. By incorporating complex subject images during training, we improve the model’s capacity for chained reasoning. In addition, we curate MSIC, a new benchmark tailored for multi-subject in-context generation. Experimental results demonstrate that MUSIC significantly surpasses other methods in both multi- and single-subject scenarios.

Method: Two key components: 1) Cross-modal grounding using frozen DINOv2 foundation model to anchor latent transitions in semantically consistent embedding space, 2) Uncertainty-adaptive trust-region bottleneck using KL regularizer as Lagrange multiplier with calibration by Expected Information Gain and Relative Performance Loss.

Result: Reduces latent rollout drift by 38-61% vs DreamerV3, improves asymptotic return, requires fewer environment interactions on long-horizon tasks, and outperforms TD-MPC2 on sparse-reward/high-contact settings.

Conclusion: GIRL effectively addresses imagination drift in MBRL through semantic grounding and adaptive trust regions, enabling more reliable long-horizon planning with improved sample efficiency.

Abstract: Model-based reinforcement learning (MBRL) improves sample efficiency by optimizing policies inside imagined rollouts, but long-horizon planning degrades when model errors compound and imagined trajectories drift off the training manifold. We introduce GIRL (Generative Imagination Reinforcement Learning), a latent world-model framework that addresses this failure mode with two key components. First, a cross-modal grounding signal derived from a frozen foundation model (DINOv2) anchors the latent transition prior to a semantically consistent embedding space, penalizing inconsistent or implausible predictions. Second, an uncertainty-adaptive trust-region bottleneck interprets the KL regularizer as the Lagrange multiplier of a constrained optimization problem, restricting imagination drift within a learned region calibrated by Expected Information Gain and a Relative Performance Loss signal. We re-derive a value-gap bound using the Performance Difference Lemma and Integral Probability Metrics, yielding a bound that remains informative as the discount factor approaches one and connects the objective to real-environment regret. Experiments across three benchmark suites, including DeepMind Control, Adroit Hand Manipulation, and Meta-World with visual distractors, show that GIRL reduces latent rollout drift by 38 to 61 percent across tasks relative to DreamerV3, improves asymptotic return, and requires fewer environment interactions on long-horizon tasks. GIRL also outperforms TD-MPC2 on sparse-reward and high-contact settings under standard evaluation metrics. A distilled-prior variant reduces inference overhead and improves computational efficiency relative to the full model.

Method: Introduces Regret-Aware Policy Optimization (RAPO) which augments environments with persistent harm-trace and scar fields, applying bounded mass-preserving transition reweighting to reduce reachability of historically harmful regions.

Result: On graph diffusion tasks (50-1000 nodes), RAPO suppresses replay, reducing re-amplification gain from 0.98 to 0.33 on 250-node graphs while retaining 82% of task return. Disabling transition deformation during replay restores re-amplification (RAG 0.91), confirming environment-level deformation as the causal mechanism.

Conclusion: The paper demonstrates that replay suppression in RL safety requires persistent environment modifications rather than just policy changes, and RAPO provides an effective approach for preventing delayed harm recurrence while maintaining task performance.

Abstract: Safety in reinforcement learning (RL) is typically enforced through objective shaping while keeping environment dynamics stationary with respect to observable state-action pairs. Under delayed harm, this can lead to replay: after a washout period, reintroducing the same stimulus under matched observable conditions reproduces a similar harmful cascade. We introduce the Replay Suppression Diagnostic (RSD), a controlled exposure-decay-replay protocol that isolates this failure mode under frozen-policy evaluation. We show that, under stationary observable transition kernels, replay cannot be structurally suppressed without inducing a persistent shift in replay-time action distributions. Motivated by platform-mediated systems, we propose Regret-Aware Policy Optimization (RAPO), which augments the environment with persistent harm-trace and scar fields and applies a bounded, mass-preserving transition reweighting to reduce reachability of historically harmful regions. On graph diffusion tasks (50-1000 nodes), RAPO suppresses replay, reducing re-amplification gain (RAG) from 0.98 to 0.33 on 250-node graphs while retaining 82% of task return. Disabling transition deformation only during replay restores re-amplification (RAG 0.91), isolating environment-level deformation as the causal mechanism.

Method: Two constraint-aware heuristics: Greedy Heuristic (GH) for single-pass allocation, and Adaptive Greedy Heuristic (AGH) with multi-start construction, relocate-based local search, and GPU consolidation. Three constraint-aware mechanisms ensure feasibility under memory, delay, error, and budget constraints.

Result: Both heuristics produce feasible solutions in under one second on Azure LLM Inference Trace workloads. AGH closely approaches optimal cost with over 260x speedup on large-scale instances. Under stress tests with parameter inflation, AGH maintains controlled SLO violations and stable cost while exact solver placement degrades sharply.

Conclusion: The proposed constraint-aware heuristics provide efficient, scalable solutions for LLM inference deployment that maintain performance under constraints while significantly outperforming exact optimization approaches in scalability and robustness.

Abstract: Deploying large language model (LLM) inference at scale requires jointly selecting base models, provisioning heterogeneous GPUs, configuring parallelism, and distributing workloads under tight latency, accuracy, and budget constraints. Exact mixed-integer linear programming (MILP) approaches guarantee optimality but scale poorly. We propose two constraint-aware heuristics: a Greedy Heuristic (GH) for single-pass allocation, and an Adaptive Greedy Heuristic (AGH) that enhances GH via multi-start construction, relocate-based local search, and GPU consolidation. Three constraint-aware mechanisms – TP-aware feasibility selection, cost-per-effective-coverage ranking, and TP upgrade – ensure feasibility under tightly coupled memory, delay, error, and budget constraints. On workloads calibrated with the Azure LLM Inference Trace (2025), both heuristics produce feasible solutions in under one second, with AGH closely approaching optimal cost while achieving over 260x speedup on large-scale instances. Under out-of-sample stress tests with up to 1.5x parameter inflation, AGH maintains controlled SLO violations and stable cost, whereas the exact solver’s placement degrades sharply.

Method: Proposes CLATT (Cluster Attention) which: 1) Uses off-the-shelf graph community detection algorithms to divide graph nodes into clusters, 2) Allows each node to attend to all other nodes within each cluster, 3) Can be used to augment either Message Passing Neural Networks or Graph Transformers.

Result: Augmenting Message Passing Neural Networks or Graph Transformers with CLATT significantly improves performance on a wide range of graph datasets, including datasets from the GraphLand benchmark representing real-world applications of graph machine learning.

Conclusion: CLATT provides an effective alternative that combines the benefits of large receptive fields from global attention with the important graph-structure-based inductive biases, leading to improved performance across diverse graph machine learning tasks.

Abstract: Message Passing Neural Networks have recently become the most popular approach to graph machine learning tasks; however, their receptive field is limited by the number of message passing layers. To increase the receptive field, Graph Transformers with global attention have been proposed; however, global attention does not take into account the graph topology and thus lacks graph-structure-based inductive biases, which are typically very important for graph machine learning tasks. In this work, we propose an alternative approach: cluster attention (CLATT). We divide graph nodes into clusters with off-the-shelf graph community detection algorithms and let each node attend to all other nodes in each cluster. CLATT provides large receptive fields while still having strong graph-structure-based inductive biases. We show that augmenting Message Passing Neural Networks or Graph Transformers with CLATT significantly improves their performance on a wide range of graph datasets including datasets from the recently introduced GraphLand benchmark representing real-world applications of graph machine learning.

Method: Proposes SYN-DIGITS, a model-agnostic calibration framework inspired by synthetic control methods. Uses a latent factor model to learn structure from digital-twin responses and transfer it to align predictions with human ground truth. Operates as a post-processing layer on top of any LLM-based simulator.

Result: Achieves up to 50% relative improvements in individual-level correlation and 50-90% relative reductions in distributional discrepancy compared to uncalibrated baselines. Supports simulation for unseen questions and unobserved populations with provable error guarantees.

Conclusion: SYN-DIGITS provides a principled, lightweight calibration framework that significantly improves the reliability of LLM-based persona simulations while maintaining flexibility across different models and applications.

Abstract: AI-based persona simulation – often referred to as digital twin simulation – is increasingly used for market research, recommender systems, and social sciences. Despite their flexibility, large language models (LLMs) often exhibit systematic bias and miscalibration relative to real human behavior, limiting their reliability. Inspired by synthetic control methods from causal inference, we propose SYN-DIGITS (SYNthetic Control Framework for Calibrated DIGItal Twin Simulation), a principled and lightweight calibration framework that learns latent structure from digital-twin responses and transfers it to align predictions with human ground truth. SYN-DIGITS operates as a post-processing layer on top of any LLM-based simulator and thus is model-agnostic. We develop a latent factor model that formalizes when and why calibration succeeds through latent space alignment conditions, and we systematically evaluate ten calibration methods across thirteen persona constructions, three LLMs, and two datasets. SYN-DIGITS supports both individual-level and distributional simulation for previously unseen questions and unobserved populations, with provable error guarantees. Experiments show that SYN-DIGITS achieves up to 50% relative improvements in individual-level correlation and 50–90% relative reductions in distributional discrepancy compared to uncalibrated baselines.

Method: Proposes a functional modeling framework leveraging sparse Sum-of-Squares forms for conditional density estimation. Introduces a novel functional form to address limitations of SoS modeling, enabling simultaneous learning of basis functions and coefficients while preserving analytical belief propagation. Includes a training method for exact adherence to normalization and non-negativity constraints.

Result: Achieves accuracy comparable to state-of-the-art approaches with significantly less memory in low-dimensional spaces. Scales to 12D systems where existing methods fail beyond 2D.

Conclusion: The proposed framework enables efficient and scalable analytical belief propagation for Markov process models while maintaining valid density estimation properties.

Abstract: Harnessing the predictive capability of Markov process models requires propagating probability density functions (beliefs) through the model. For many existing models however, belief propagation is analytically infeasible, requiring approximation or sampling to generate predictions. This paper proposes a functional modeling framework leveraging sparse Sum-of-Squares (SoS) forms for valid (conditional) density estimation. We study the theoretical restrictions of modeling conditional densities using the SoS form, and propose a novel functional form for addressing such limitations. The proposed architecture enables generalized simultaneous learning of basis functions and coefficients, while preserving analytical belief propagation. In addition, we propose a training method that allows for exact adherence to the normalization and non-negativity constraints. Our results show that the proposed method achieves accuracy comparable to state-of-the-art approaches while requiring significantly less memory in low-dimensional spaces, and it further scales to 12D systems when existing methods fail beyond 2D.

Method: Multiplicity-weighted Stochastic Attention (SA) based on modern Hopfield network theory embeds real patient profiles as memory patterns in a continuous energy landscape and generates synthetic patients via Langevin dynamics that interpolate between stored patterns while preserving original cohort geometry. Per-pattern multiplicity weights enable targeted amplification of rare clinical subgroups.

Result: Applied to longitudinal coagulation data from 23 pregnant patients (72 biochemical features across 3 visits), SA generated synthetic patients statistically, structurally, and mechanistically indistinguishable from real counterparts. A mechanistic model calibrated entirely on synthetic patients predicted held-out real patient outcomes as well as one calibrated on real data.

Conclusion: SA can produce clinically useful synthetic cohorts from very small longitudinal datasets, enabling data-augmented modeling in small-cohort settings without requiring retraining for targeted subgroup amplification.

Abstract: Small longitudinal clinical cohorts, common in maternal health, rare diseases, and early-phase trials, limit computational modeling: too few patients to train reliable models, yet too costly and slow to expand through additional enrollment. We present multiplicity-weighted Stochastic Attention (SA), a generative framework based on modern Hopfield network theory that addresses this gap. SA embeds real patient profiles as memory patterns in a continuous energy landscape and generates novel synthetic patients via Langevin dynamics that interpolate between stored patterns while preserving the geometry of the original cohort. Per-pattern multiplicity weights enable targeted amplification of rare clinical subgroups at inference time without retraining. We applied SA to a longitudinal coagulation dataset from 23 pregnant patients spanning 72 biochemical features across 3 visits (pre-pregnancy baseline, first trimester, and third trimester), including rare subgroups such as polycystic ovary syndrome and preeclampsia. Synthetic patients generated by SA were statistically, structurally, and mechanistically indistinguishable from their real counterparts across multiple independent validation tests, including an ordinary differential equation model of the coagulation cascade. A downstream utility test further showed that a mechanistic model calibrated entirely on synthetic patients predicted held-out real patient outcomes as well as one calibrated on real data. These results demonstrate that SA can produce clinically useful synthetic cohorts from very small longitudinal datasets, enabling data-augmented modeling in small-cohort settings.

Method: Analyzes LLMs as lossy compression systems, measuring how they approach the Information Bottleneck bound on compression during pre-training. Examines multiple open-weight models to study compression differences based on data and training recipes.

Result: Pre-training produces models optimally compressed for next-sequence prediction. Compression optimality and information content predict downstream performance across diverse benchmarks. Different LLM families compress differently but compression quality consistently correlates with performance.

Conclusion: LLMs learn by retaining only task-relevant information from training data, acting as optimal compressors. Compression quality serves as a predictor of model performance, providing an information-theoretic framework for understanding LLM learning.

Abstract: Despite the increasing prevalence of large language models (LLMs), we still have a limited understanding of how their representational spaces are structured. This limits our ability to interpret how and what they learn or relate them to learning in humans. We argue LLMs are best seen as an instance of lossy compression, where over training they learn by retaining only information in their training data relevant to their objective(s). We show pre-training results in models that are optimally compressed for next-sequence prediction, approaching the Information Bottleneck bound on compression. Across an array of open weights models, each compresses differently, likely due to differences in the data and training recipes used. However even across different families of LLMs the optimality of a model’s compression, and the information present in it, can predict downstream performance on across a wide array of benchmarks, letting us directly link representational structure to actionable insights about model performance. In the general case the work presented here offers a unified Information-Theoretic framing for how these models learn that is deployable at scale.

Method: The study uses controlled experiments with PyTorch on CIFAR-10 and MNIST datasets with multiple random seeds. It examines: 1) stochastic gradient descent across batch sizes, 2) geometry of flat vs sharp minima via Hessian eigenvalue estimation and weight perturbation analysis, 3) Neural Tangent Kernel regime through wide-network experiments, 4) double descent across model scales, and 5) Lottery Ticket Hypothesis through iterative magnitude pruning.

Result: Key findings: 1) Smaller batch sizes consistently produced lower test error and flatter minima, with an 11.8x difference in top Hessian eigenvalue between small-batch and large-batch solutions corresponding to 1.61 percentage points higher test accuracy. 2) Sparse subnetworks retaining only 10% of parameters achieved within 1.15 percentage points of full model performance when retrained from their original initialization.

Conclusion: Generalization in overparameterized neural networks is strongly influenced by the interaction between network architecture, optimization algorithms, and loss landscape geometry. The results highlight the need for revised learning-theoretic frameworks capable of explaining generalization in high-dimensional model regimes.

Abstract: Classical statistical learning theory predicts that overparameterized models should exhibit severe overfitting, yet modern deep neural networks with far more parameters than training samples consistently generalize well. This contradiction has become a central theoretical question in machine learning. This study investigates the role of optimization dynamics and implicit regularization in enabling generalization in overparameterized neural networks through controlled experiments. We examine stochastic gradient descent (SGD) across batch sizes, the geometry of flat versus sharp minima via Hessian eigenvalue estimation and weight perturbation analysis, the Neural Tangent Kernel (NTK) regime through wide-network experiments, double descent across model scales, and the Lottery Ticket Hypothesis through iterative magnitude pruning. All experiments use PyTorch on CIFAR-10 and MNIST with multiple random seeds. Our findings demonstrate that generalization is strongly influenced by the interaction between network architecture, optimization algorithms, and loss landscape geometry. Smaller batch sizes consistently produced lower test error and flatter minima, with an 11.8x difference in top Hessian eigenvalue between small-batch and large-batch solutions corresponding to 1.61 percentage points higher test accuracy. Sparse subnetworks retaining only 10% of parameters achieved within 1.15 percentage points of full model performance when retrained from their original initialization. These results highlight the need for revised learning-theoretic frameworks capable of explaining generalization in high-dimensional model regimes.

Method: Proposes MS-BCNN (Multi-Scale Bi-Branch Convolutional Neural Network) with short- and long-kernel branches for local fluctuations and long-term trends. Uses hierarchical-conditional mixed configuration space for alignment operators, architectural choices, and hyperparameters, and applies PHMOEA (Player-based Hybrid Multi-Objective Evolutionary Algorithm) to approximate error-complexity Pareto frontier within computational budget.

Result: Experiments on hierarchical synthetic benchmarks and real-world sintering dataset show framework outperforms competitive baselines under same budget and offers flexible deployment choices.

Conclusion: The auto-configuration framework successfully addresses the need for systematic co-design in industrial forecasting, providing deployable Pareto-optimal models that balance error and complexity within computational constraints.

Abstract: Industrial forecasting often involves multi-source asynchronous signals and multi-output targets, while deployment requires explicit trade-offs between prediction error and model complexity. Current practices typically fix alignment strategies or network designs, making it difficult to systematically co-design preprocessing, architecture, and hyperparameters in budget-limited training-based evaluations. To address this issue, we propose an auto-configuration framework that outputs a deployable Pareto set of forecasting models balancing error and complexity. At the model level, a Multi-Scale Bi-Branch Convolutional Neural Network (MS–BCNN) is developed, where short- and long-kernel branches capture local fluctuations and long-term trends, respectively, for multi-output regression. At the search level, we unify alignment operators, architectural choices, and training hyperparameters into a hierarchical-conditional mixed configuration space, and apply Player-based Hybrid Multi-Objective Evolutionary Algorithm (PHMOEA) to approximate the error–complexity Pareto frontier within a limited computational budget. Experiments on hierarchical synthetic benchmarks and a real-world sintering dataset demonstrate that our framework outperforms competitive baselines under the same budget and offers flexible deployment choices.

Method: Uses cellular sheaf theory with modality-independent neighborhood structures on sample indices. Defines two incompatibility measures: projection hardness (complexity of global alignment) and sheaf-Laplacian obstruction (spatial variation needed for local projections to achieve global consistency). Links spectral gap to alignment stability and provides explicit constructions.

Result: Shows compatibility is generally non-transitive, demonstrates that intermediate modalities can reduce effective hardness even when direct alignment is infeasible, and provides bounds relating obstruction energy to excess global-map error under Lipschitz assumptions.

Conclusion: The framework provides a principled way to analyze cross-modal compatibility failures, separating hardness failures (no low-complexity global projection) from obstruction failures (local projections exist but lack global consistency), with practical implications for multimodal representation learning.

Abstract: We develop a unified framework for analyzing cross-modal compatibility in learned representations. The core object is a modality-independent neighborhood site on sample indices, equipped with a cellular sheaf of finite-dimensional real inner-product spaces. For a directed modality pair $(a\to b)$, we formalize two complementary incompatibility mechanisms: projection hardness, the minimal complexity within a nested Lipschitz-controlled projection family needed for a single global map to align whitened embeddings; and sheaf-Laplacian obstruction, the minimal spatial variation required by a locally fit field of projection parameters to achieve a target alignment error. The obstruction invariant is implemented via a projection-parameter sheaf whose 0-Laplacian energy exactly matches the smoothness penalty used in sheaf-regularized regression, making the theory directly operational. This separates two distinct failure modes: hardness failure, where no low-complexity global projection exists, and obstruction failure, where local projections exist but cannot be made globally consistent over the semantic neighborhood graph without large parameter variation. We link the sheaf spectral gap to stability of global alignment, derive bounds relating obstruction energy to excess global-map error under mild Lipschitz assumptions, and give explicit constructions showing that compatibility is generally non-transitive. We further define bridging via composed projection families and show, in a concrete ReLU setting, that an intermediate modality can strictly reduce effective hardness even when direct alignment remains infeasible.

Method: Proposes CauPsi with two key mechanisms: 1) Causal Task Chain propagates upstream task predictions to downstream tasks via learnable prototype embeddings, modeling cognitive cascade from environmental perception to behavioral regulation; 2) Cross-Task Psychological Conditioning estimates psychological state from driver facial expressions/body posture and injects it as conditioning input to all tasks.

Result: Achieves 82.71% mean accuracy with only 5.05M parameters on AIDE dataset, surpassing prior work by +1.0% overall, with notable improvements on Driver Emotion Recognition (+3.65%) and Driver Behavior Recognition (+7.53%). Psychological state signal acquires systematic task-label-dependent patterns without explicit annotations.

Conclusion: CauPsi successfully models the hierarchical cognitive causal structure in driving behavior, demonstrating that explicitly modeling psychological state modulation and task dependencies improves multi-task learning for driver assistance systems.

Abstract: Multi-task learning for advanced driver assistance systems requires modeling the complex interplay between driver internal states and external traffic environments. However, existing methods treat recognition tasks as flat and independent objectives, failing to exploit the cognitive causal structure underlying driving behavior. In this paper, we propose CauPsi, a cognitive science-grounded causal multi-task learning framework that explicitly models the hierarchical dependencies among Traffic Context Recognition (TCR), Vehicle Context Recognition (VCR), Driver Emotion Recognition (DER), and Driver Behavior Recognition (DBR). The proposed framework introduces two key mechanisms. First, a Causal Task Chain propagates upstream task predictions to downstream tasks via learnable prototype embeddings, realizing the cognitive cascade from environmental perception to behavioral regulation in a differentiable manner. Second, Cross-Task Psychological Conditioning (CTPC) estimates a psychological state signal from driver facial expressions and body posture and injects it as a conditioning input to all tasks including environmental recognition, thereby modeling the modulatory effect of driver internal states on cognitive and decision-making processes. Evaluated on the AIDE dataset, CauPsi achieves a mean accuracy of 82.71% with only 5.05M parameters, surpassing prior work by +1.0% overall, with notable improvements on DER (+3.65%) and DBR (+7.53%). Ablation studies validate the independent contribution of each component, and analysis of the psychological state signal confirms that it acquires systematic task-label-dependent patterns in a self-supervised manner without explicit psychological annotations.

Method: Introduces Guardian-as-an-Advisor (GaaA) pipeline with GuardAdvisor model trained via supervised fine-tuning followed by reinforcement learning for label-explanation consistency. Uses GuardSet dataset (208k+ multi-domain cases) with robustness and honesty slices. The guardian predicts binary risk labels with explanations and prepends this advice to original queries for re-inference.

Result: GuardAdvisor achieves competitive detection accuracy while enabling advisory workflow; when used to augment inputs, responses improve over unaugmented prompts. Latency study shows advisor inference uses below 5% of base-model compute with only 2-10% end-to-end overhead under realistic harmful-input rates.

Conclusion: GaaA steers models to comply with model specifications, maintaining safety while reducing over-refusal through a soft-gating approach that keeps base models operating under their original specs.

Abstract: Hard-gated safety checkers often over-refuse and misalign with a vendor’s model spec; prevailing taxonomies also neglect robustness and honesty, yielding safer-on-paper yet less useful systems. This work introduces Guardian-as-an-Advisor (GaaA), a soft-gating pipeline where a guardian predicts a binary risk label plus a concise explanation and prepends this advice to the original query for re-inference, keeping the base model operating under its original spec. To support training and evaluation, GuardSet is constructed, a 208k+ multi-domain dataset unifying harmful and harmless cases with targeted robustness and honesty slices. GuardAdvisor is trained via SFT followed by RL to enforce label-explanation consistency. GuardAdvisor attains competitive detection accuracy while enabling the advisory workflow; when used to augment inputs, responses improve over unaugmented prompts. A latency study shows advisor inference uses below 5% of base-model compute and adds only 2-10% end-to-end overhead under realistic harmful-input rates. Overall, GaaA steers models to comply with the model spec, maintaining safety while reducing over-refusal.

Method: Introduces Position-Adaptive Spectral Tapering (PoST) with two mechanisms: 1) Spectral Reparameterization enforces geometrically spaced log-decay rates, and 2) Position-Adaptive Scaling eliminates scale mismatch by stretching the spectrum to actual dependency range, making impulse response scale-free.

Result: Proven minimax optimal at rate O(exp(-cN/log T)), achieves consistent zero-shot language modeling improvements at 180M-440M scales, significant long-context retrieval gains for Mamba-2, and competitive performance across other architectures.

Conclusion: PoST provides an architecture-agnostic framework that improves linear recurrent models’ long-range memory capabilities without computational overhead, demonstrating practical benefits across multiple state-space and attention-like architectures.

Abstract: Linear recurrent models offer linear-time sequence processing but often suffer from suboptimal long-range memory. We trace this to the decay spectrum: for $N$ channels, random initialization collapses the minimum spectral gap to $O(N^{-2})$, yielding sub-exponential error $\exp(-Ω(N/\log N))$; linear spacing avoids collapse but degrades to $\exp(-O(N/\sqrt{T}))$, practically algebraic over long contexts. We introduce Position-Adaptive Spectral Tapering (PoST), an architecture-agnostic framework combining two mechanisms: (1) Spectral Reparameterization, which structurally enforces geometrically spaced log-decay rates, proven minimax optimal at rate $O(\exp(-cN/\log T))$; and (2) Position-Adaptive Scaling, the provably unique mechanism that eliminates the scale mismatch of static spectra (where only $N\log t/\log T$ of $N$ channels are effective at position $t$) by stretching the spectrum to the actual dependency range, sharpening the rate to $O(\exp(-cN/\log t))$. This scaling natively induces fractional invariance: the impulse response becomes scale-free, with channels interpolating between relative and absolute temporal coordinates. PoST integrates into any diagonal linear recurrence without overhead. We instantiate it across Mamba-2, RWKV-7, Gated DeltaNet, Gated Linear Attention, and RetNet. Pre-training at 180M-440M scales shows consistent zero-shot language modeling improvements, significant long-context retrieval gains for Mamba-2 (MQAR and NIAH), and competitive or improved performance across other architectures. Code: https://github.com/SiLifen/PoST.

Method: Proposes SAGE (Sign Adaptive GradiEnt) optimizer combining Lion-style update direction with novel O(d) adaptive scale that acts as “safe damper” bounded by 1.0 to tame high-variance dimensions more effectively.

Result: On Llama models up to 1.3B parameters, SAGE-based hybrid achieves new SOTA perplexity, outperforming all baselines including SinkGD hybrid while significantly reducing optimizer state memory.

Conclusion: SAGE resolves embedding layer dilemma in hybrid training, enabling memory-efficient optimization without sacrificing convergence quality for large language models.

Abstract: The AdamW optimizer, while standard for LLM pretraining, is a critical memory bottleneck, consuming optimizer states equivalent to twice the model’s size. Although light-state optimizers like SinkGD attempt to address this issue, we identify the embedding layer dilemma: these methods fail to handle the sparse, high-variance gradients inherent to embeddings, forcing a hybrid design that reverts to AdamW and partially negates the memory gains. We propose SAGE (Sign Adaptive GradiEnt), a novel optimizer that resolves this dilemma by replacing AdamW in this hybrid structure. SAGE combines a Lion-style update direction with a new, memory-efficient $O(d)$ adaptive scale. This scale acts as a “safe damper,” provably bounded by 1.0, which tames high-variance dimensions more effectively than existing methods. This superior stability allows SAGE to achieve better convergence. On Llama models up to 1.3B parameters, our SAGE-based hybrid achieves new state-of-the-art perplexity, outperforming all baselines, including SinkGD hybrid, while significantly reducing optimizer state memory.

Method: Investigated RLVR robustness by introducing controlled noise into RL training across code generation and scientific reasoning domains, testing various noise types, model families (Qwen3, GLM4, Llama 3.1), and model sizes (4B-9B).

Result: Noise rates up to 15% yield peak validation accuracy within 2 percentage points of clean baseline, consistent across noise types, model families, and sizes. Imperfect verification doesn’t fundamentally hinder RLVR.

Conclusion: RLVR is robust to imperfect verifiers; practitioners should prioritize moderate accuracy with high precision over perfect verification for effective training.

Abstract: Reinforcement Learning with Verifiable Rewards (RLVR) has become a prominent method for post-training Large Language Models (LLMs). However, verifiers are rarely error-free; even deterministic checks can be inaccurate, and the growing dependence on model-based judges exacerbates the issue. The extent to which RLVR is robust to such noise and the verifier accuracy required for effective training remain unresolved questions. We investigate these questions in the domains of code generation and scientific reasoning by introducing noise into RL training. Noise rates up to 15% yield peak validation accuracy within 2 percentage points of the clean baseline. These findings are consistent across controlled and model-based noise types, three model families (Qwen3, GLM4, Llama 3.1), and model sizes from 4B to 9B. Overall, the results indicate that imperfect verification does not constitute a fundamental barrier to RLVR. Furthermore, our findings suggest that practitioners should prioritize moderate accuracy with high precision over perfect verification.

Method: Formulates lead optimization as a Markov Decision Process over synthesis-constrained action spaces defined by validated reaction templates. Uses tool-augmented LLM agents as dynamic reaction environments that invoke chemical analysis tools to identify reactive sites and propose chemically grounded transformations. Trains policy models via Group Relative Policy Optimization (GRPO) to select constrained actions maximizing long-term oracle reward across multi-step reaction trajectories, with SMILES-based caching for efficiency.

Result: Achieves average Top-10 score of 0.563 across 13 property optimization tasks from Therapeutic Data Commons and one structure-based docking task, outperforming strongest synthesizable baseline by 10.4% relative improvement. Attains best sample efficiency on 10 of 14 tasks. SMILES-based caching reduces end-to-end optimization time by ~43%.

Conclusion: MolReAct successfully grounds lead optimization in validated reaction templates, producing property-improved molecules with explicit synthetic pathways. Both tool-augmented reaction proposals and trajectory-level policy optimization contribute complementary gains to the framework’s effectiveness.

Abstract: Lead optimization in drug discovery requires improving therapeutic properties while ensuring that proposed molecular modifications correspond to feasible synthetic routes. Existing approaches either prioritize property scores without enforcing synthesizability, or rely on expensive enumeration over large reaction networks, while direct application of Large Language Models (LLMs) frequently produces chemically invalid structures. We introduce MolReAct, a framework that formulates lead optimization as a Markov Decision Process over a synthesis-constrained action space defined by validated reaction templates. A tool-augmented LLM agent serves as a dynamic reaction environment that invokes specialized chemical analysis tools to identify reactive sites and propose chemically grounded transformations from matched templates. A policy model trained via Group Relative Policy Optimization (GRPO) selects among these constrained actions to maximize long-term oracle reward across multi-step reaction trajectories. A SMILES-based caching mechanism further reduces end-to-end optimization time by approximately 43%. Across 13 property optimization tasks from the Therapeutic Data Commons and one structure-based docking task, MolReAct achieves an average Top-10 score of 0.563, outperforming the strongest synthesizable baseline by 10.4% in relative improvement, and attains the best sample efficiency on 10 of 14 tasks. Ablations confirm that both tool-augmented reaction proposals and trajectory-level policy optimization contribute complementary gains. By grounding every step in validated reaction templates, MolReAct produces molecules that are property-improved and each accompanied by an explicit synthetic pathway.

Method: Proposes a data-driven method to compute the Koopman generator in low-rank tensor train format by taking logarithms of Koopman eigenvalues while preserving the TT format, enabling scalability to higher dimensions.

Result: Experiments on 4D Lotka-Volterra and 10D Lorenz-96 systems show accurate recovery of vector field coefficients and demonstrate scalability to higher-dimensional systems.

Conclusion: The tensor train format enables efficient computation of Koopman generators for high-dimensional dynamical systems, overcoming the curse of dimensionality in system identification.

Abstract: Identifying governing equations of nonlinear dynamical systems from data is challenging. While sparse identification of nonlinear dynamics (SINDy) and its extensions are widely used for system identification, operator-logarithm approaches use the logarithm to avoid time differentiation, enabling larger sampling intervals. However, they still suffer from the curse of dimensionality. Then, we propose a data-driven method to compute the Koopman generator in a low-rank tensor train (TT) format by taking logarithms of Koopman eigenvalues while preserving the TT format. Experiments on 4-dimensional Lotka-Volterra and 10-dimensional Lorenz-96 systems show accurate recovery of vector field coefficients and scalability to higher-dimensional systems.

Method: Proposes a joint optimization problem for DMI task offloading, inference optimization, and UAV trajectory planning, modeled as heterogeneous-agent Markov decision process, solved with SU-HATD3 algorithm.

Result: Numerical results show the proposed algorithm outperforms baselines in improving system utility and convergence rate.

Conclusion: The GAI-empowered ITDT with the proposed optimization approach effectively balances fidelity and delay for intelligent transportation applications.

Abstract: To implement the intelligent transportation digital twin (ITDT), unmanned aerial vehicles (UAVs) are scheduled to process the sensing data from the roadside sensors. At this time, generative artificial intelligence (GAI) technologies such as diffusion models are deployed on the UAVs to transform the raw sensing data into the high-quality and valuable. Therefore, we propose the GAI-empowered ITDT. The dynamic processing of a set of diffusion model inference (DMI) tasks on the UAVs with dynamic mobility simultaneously influences the DT updating fidelity and delay. In this paper, we investigate a joint optimization problem of DMI task offloading, inference optimization and UAV trajectory planning as the system utility maximization (SUM) problem to address the fidelity-delay tradeoff for the GAI-empowered ITDT. To seek a solution to the problem under the network dynamics, we model the SUM problem as the heterogeneous-agent Markov decision process, and propose the sequential update-based heterogeneous-agent twin delayed deep deterministic policy gradient (SU-HATD3) algorithm, which can quickly learn a near-optimal solution. Numerical results demonstrate that compared with several baseline algorithms, the proposed algorithm has great advantages in improving the system utility and convergence rate.

Method: Tree-of-Evidence (ToE) frames interpretability as discrete optimization using lightweight Evidence Bottlenecks to score data units and beam search to find compact evidence sets that reproduce predictions.

Result: ToE maintains predictive performance (over 0.98 of full-model AUROC with ≤5 evidence units), achieves higher decision agreement and lower fidelity error than alternatives, and adapts search strategy based on data ambiguity.

Conclusion: ToE provides practical auditing for multimodal models by revealing discrete evidence units supporting predictions, addressing opacity in high-stakes domains like healthcare.

Abstract: Large Multimodal Models (LMMs) achieve state-of-the-art performance in high-stakes domains like healthcare, yet their reasoning remains opaque. Current interpretability methods, such as attention mechanisms or post-hoc saliency, often fail to faithfully represent the model’s decision-making process, particularly when integrating heterogeneous modalities like time-series and text. We introduce Tree-of-Evidence (ToE), an inference-time search algorithm that frames interpretability as a discrete optimization problem. Rather than relying on soft attention weights, ToE employs lightweight Evidence Bottlenecks that score coarse groups or units of data (e.g., vital-sign windows, report sentences) and performs a beam search to identify the compact evidence set required to reproduce the model’s prediction. We evaluate ToE across six tasks spanning three datasets and two domains: four clinical prediction tasks on MIMIC-IV, cross-center validation on eICU, and non-clinical fault detection on LEMMA-RCA. ToE produces auditable evidence traces while maintaining predictive performance, retaining over 0.98 of full-model AUROC with as few as five evidence units across all settings. Under sparse evidence budgets, ToE achieves higher decision agreement and lower probability fidelity error than other approaches. Qualitative analyses show that ToE adapts its search strategy: it often resolves straightforward cases using only vitals, while selectively incorporating text when physiological signals are ambiguous. ToE therefore provides a practical mechanism for auditing multimodal models by revealing which discrete evidence units support each prediction.

Method: CausalVAE is designed as a plug-in module that can be attached to diverse encoder-transition backbones in latent world models. It introduces causal structure into the latent space to enable intervention-aware counterfactual retrieval while preserving factual prediction performance.

Result: Significant improvements in counterfactual retrieval metrics, particularly on Physics benchmarks: +102.5% average improvement in CF-H@1 across 8 baselines, and from 11.0 to 41.0 (+272.7%) in a GNN-NLL setting. The learned causal structures recover meaningful physical interaction patterns.

Conclusion: CausalVAE effectively enhances latent world models with causal structure, improving robustness to distribution shifts and interventions while maintaining factual prediction performance and providing interpretable causal representations.

Abstract: In this work, CausalVAE is introduced as a plug-in structural module for latent world models and is attached to diverse encoder-transition backbones. Across the reported benchmarks, competitive factual prediction is preserved and intervention-aware counterfactual retrieval is improved after the plug-in is added, suggesting stronger robustness under distribution shift and interventions. The largest gains are observed on the Physics benchmark: when averaged over 8 paired baselines, CF-H@1 is improved by +102.5%. In a representative GNN-NLL setting on Physics, CF-H@1 is increased from 11.0 to 41.0 (+272.7%). Through causal analysis, learned structural dependencies are shown to recover meaningful first-order physical interaction trends, supporting the interpretability of the learned latent causal structure.

Method: Theoretical analysis of one-hidden-layer neural network with ReLU activations and fixed biases, studying both continuous and discrete versions. Uses L² squared loss function and gradient descent procedure, proving convergence and spectral bias properties mathematically.

Result: Proved rigorous convergence of learning process with gradient descent, demonstrated spectral bias property, and derived insights about activation function properties. Proposed new activation function called FReX (full-wave rectified exponential function) and discussed its convergence properties.

Conclusion: Theoretical analysis reveals important relationships between operator spectra and learning dynamics, provides guidelines for activation function design, and introduces FReX as a promising alternative activation function with proven convergence properties.

Abstract: We analyze a simple one-hidden-layer neural network with ReLU activation functions and fixed biases, with one-dimensional input and output. We study both continuous and discrete versions of the model, and we rigorously prove the convergence of the learning process with the $L^2$ squared loss function and the gradient descent procedure. We also prove the spectral bias property for this learning process. Several conclusions of this analysis are discussed; in particular, regarding the structure and properties that activation functions should possess, as well as the relationships between the spectrum of certain operators and the learning process. Based on this, we also propose an alternative activation function, the full-wave rectified exponential function (FReX), and we discuss the convergence of the gradient descent with this alternative activation function.

Method: Uses a learned measurement rate p_t ∈ (0,1) that dynamically routes computation between wave phase (p_t→0) for distributed complex-phase interference and particle phase (p_t→1) for precise local token storage, with a predicted phase transition at critical sequence length N*≈1024.

Result: Achieves 0.905 accuracy vs Transformer’s 0.736 on AG News (+16.6%), 42.8x memory reduction (810 MB vs 34,651 MB at N=8192), 0.968 accuracy on exact-recall tasks, and 99.8% sparsity rate with autonomous critical token storage. On WikiText-103, reaches PPL 92.1 vs Transformer’s 90.5 with O(64) vs O(N) KV cache.

Conclusion: MIPT-SSM provides a physics-inspired approach to sequence modeling that fundamentally addresses the wave-particle duality problem, achieving significant improvements in accuracy, memory efficiency, and autonomous information filtering.

Abstract: We present MIPT-SSM, a neural sequence architecture built on the physics of Measurement-Induced Phase Transitions (MIPT). The central idea is a learned measurement rate $p_{t}\in(0,1)$ that routes computation between two regimes: wave phase $(p_{t}\rightarrow0)$, where information propagates as distributed complex-phase interference; and particle phase $(p_{t}\rightarrow1)$ where the state collapses onto the current token, enabling precise local storage. These two regimes are provably incompatible in a single linear operator one of the few “no-go theorems” in sequence modeling and $p_{t}$ is our way around it. The model is predicted to exhibit a phase transition at critical sequence length $N^{*}\approx1024$, where the information density ratio $N/D$ crosses unity, consistent with our memory scaling observations. On AG News (four-class classification), MIPT achieves 0.905 accuracy versus Transformer’s 0.736 (+16.6%), stable across 3 seeds. At $N=8192$ MIPT requires 810 MB versus Transformer’s 34,651 MB a 42.8x memory reduction. On exact-recall (“needle-in-a-haystack”), our causal sparse KV cache achieves 0.968 accuracy. Remarkably, under unbounded cache capacity, the $p_{t}$ gate autonomously learns to store only the single critical token (averaging $1.0/512$ slots used), filtering out all noise and achieving a 99.8% sparsity rate. On language modeling (WikiText-103, 31M parameters), MIPT-LM with $K=64$ cache reaches PPL 92.1 versus Transformer’s 90.5 (gap: 1.8%) while inference KV cache shrinks from $O(N)$ to $O(64)$.

Method: Proposes physics augmented finite element model updating (paFEMU) as a transfer learning approach combining: 1) AI-enabled constitutive modeling, 2) sparsification for interpretable model discovery, and 3) finite element-based adjoint optimization using multi-modal data. Combines simple mechanical testing data with digital image correlation full-field data for constitutive model discovery.

Result: The sparse representation enables easy integration of neural constitutive models in existing finite element workflows and allows low-dimensional updating during transfer learning. The approach facilitates rapid constitutive modeling discovery by leveraging diverse data sources.

Conclusion: paFEMU provides a framework for combining multi-modal and multi-fidelity data in constitutive modeling while maintaining interpretability through sparsification and compatibility with existing finite element software through low-dimensional representations.

Abstract: Recent progress in AI-enabled constitutive modeling has concentrated on moving from a purely data-driven paradigm to the enforcement of physical constraints and mechanistic principles, a concept referred to as physics augmentation. Classical phenomenological approaches rely on selecting a pre-defined model and calibrating its parameters, while machine learning methods often focus on discovery of the model itself. Sparse regression approaches lie in between, where large libraries of pre-defined models are probed during calibration. Sparsification in the aforementioned paradigm, but also in the context of neural network architecture, has been shown to enable interpretability, uncertainty quantification, but also heterogeneous software integration due to the low-dimensional nature of the resulting models. Most works in AI-enabled constitutive modeling have also focused on data from a single source, but in reality, materials modeling workflows can contain data from many different sources (multi-modal data), and also from testing other materials within the same materials class (multi-fidelity data). In this work, we introduce physics augmented finite element model updating (paFEMU), as a transfer learning approach that combines AI-enabled constitutive modeling, sparsification for interpretable model discovery, and finite element-based adjoint optimization utilizing multi-modal data. This is achieved by combining simple mechanical testing data, potentially from a distinct material, with digital image correlation-type full-field data acquisition to ultimately enable rapid constitutive modeling discovery. The simplicity of the sparse representation enables easy integration of neural constitutive models in existing finite element workflows, and also enables low-dimensional updating during transfer learning.

Method: Proposes Agent-as-Annotators framework that structures synthetic trajectory generation by analogy to human annotation roles, replacing Task Designer, Annotator, and Supervisor with modular LLM components. Uses Gemini 3 Pro as teacher to generate 3,000 trajectories across six web environments, then fine-tunes a 9B-parameter student with supervised learning on 2,322 quality-filtered trajectories.

Result: The resulting model achieves 41.5% on WebArena, surpassing closed-source models like Claude 3.5 Sonnet (36.0%) and GPT-4o (31.5%), and nearly doubling previous best open-weight result (Go-Browse, 21.7%). Shows strong transfer to unseen environments with 18.2 percentage point gain on WorkArena L1 and consistent improvements across three additional benchmarks.

Conclusion: Structured trajectory synthesis from a single frontier teacher is sufficient to produce competitive, locally deployable web agents. The framework demonstrates that modular LLM components can effectively replace human annotation roles for synthetic data generation.

Abstract: Frontier LLMs can navigate complex websites, but their cost and reliance on third-party APIs make local deployment impractical. We introduce Agent-as-Annotators, a framework that structures synthetic trajectory generation for web agents by analogy to human annotation roles, replacing the Task Designer, Annotator, and Supervisor with modular LLM components. Using Gemini 3 Pro as teacher, we generate 3,000 trajectories across six web environments and fine-tune a 9B-parameter student with pure supervised learning on the 2,322 that pass quality filtering. The resulting model achieves 41.5% on WebArena, surpassing closed-source models such as Claude 3.5 Sonnet (36.0%) and GPT-4o (31.5%) under the same evaluation protocol, and nearly doubling the previous best open-weight result (Go-Browse, 21.7%). Capabilities transfer to unseen environments, with an 18.2 percentage point gain on WorkArena L1 (an enterprise platform never seen during training) and consistent improvements across three additional benchmarks. Ablations confirm that each pipeline component contributes meaningfully, with Judge filtering, evaluation hints, and reasoning traces each accounting for measurable gains. These results demonstrate that structured trajectory synthesis from a single frontier teacher is sufficient to produce competitive, locally deployable web agents. Project page: https://agent-as-annotators.github.io

Method: PolicyLong shifts to dynamic on-policy paradigm by iteratively re-executing data screening (entropy computation, retrieval, and verification) using the current model, ensuring training distribution tracks evolving capabilities and creates emergent self-curriculum.

Result: Experiments on RULER, HELMET, and LongBench-v2 show PolicyLong consistently outperforms EntropyLong and NExtLong, with gains growing at longer contexts (e.g., +2.54 at 128K on RULER).

Conclusion: On-policy data evolution is valuable for extending LLM context windows, with dynamic screening aligning training distribution with model capabilities and creating both positive and hard negative contexts from the current model’s entropy landscape.

Abstract: Extending LLM context windows is hindered by scarce high-quality long-context data. Recent methods synthesize data with genuine long-range dependencies via information-theoretic verification, selecting contexts that reduce a base model’s predictive entropy. However, their single-pass offline construction with a fixed model creates a fundamental off-policy gap: the static screening landscape misaligns with the model’s evolving capabilities, causing the training distribution to drift. We propose PolicyLong, shifting data construction towards a dynamic on-policy paradigm. By iteratively re-executing data screening (entropy computation, retrieval, and verification) using the current model, PolicyLong ensures the training distribution tracks evolving capabilities, yielding an emergent self-curriculum. Crucially, both positive and hard negative contexts derive from the current model’s entropy landscape, co-evolving what the model learns to exploit and resist. Experiments on RULER, HELMET, and LongBench-v2 (Qwen2.5-3B) show PolicyLong consistently outperforms EntropyLong and NExtLong, with gains growing at longer contexts (e.g., +2.54 at 128K on RULER), confirming the value of on-policy data evolution.

Method: The authors identify that tasks must share training instances for gradient conflicts to reveal genuine relationships. They discover a sample overlap requirement with a sharp phase transition: below 30% overlap, gradient-task correlations are noise; above 40%, they reliably recover structure. They validate this across multiple datasets.

Result: Comprehensive validation shows strong correlations and recovers biological pathway organization when overlap thresholds are met. Standard benchmarks like MoleculeNet (<5% overlap) and TDC (8-14% overlap) operate far below the meaningful threshold.

Conclusion: The sample overlap requirement provides the first principled explanation for seven years of inconsistent multi-task learning results, revealing that standard benchmarks systematically violate the fundamental requirement for meaningful gradient analysis.

Abstract: Multi-task learning shows strikingly inconsistent results – sometimes joint training helps substantially, sometimes it actively harms performance – yet the field lacks a principled framework for predicting these outcomes. We identify a fundamental but unstated assumption underlying gradient-based task analysis: tasks must share training instances for gradient conflicts to reveal genuine relationships. When tasks are measured on the same inputs, gradient alignment reflects shared mechanistic structure; when measured on disjoint inputs, any apparent signal conflates task relationships with distributional shift. We discover this sample overlap requirement exhibits a sharp phase transition: below 30% overlap, gradient-task correlations are statistically indistinguishable from noise; above 40%, they reliably recover known biological structure. Comprehensive validation across multiple datasets achieves strong correlations and recovers biological pathway organization. Standard benchmarks systematically violate this requirement – MoleculeNet operates at <5% overlap, TDC at 8-14% – far below the threshold where gradient analysis becomes meaningful. This provides the first principled explanation for seven years of inconsistent MTL results.

Method: Proposes QaRL (Rollout Alignment Quantization-Aware RL) to align training-side forward passes with quantized rollouts, and introduces TBPO (Trust-Band Policy Optimization) - a sequence-level objective with dual clipping for negative samples to mitigate error tokens in long-form responses.

Result: On Qwen3-30B-A3B MoE for math problems, QaRL outperforms quantized-rollout training by +5.5 while improving stability and preserving low-bit throughput benefits.

Conclusion: QaRL successfully addresses the training-inference gap in quantized LLM RL pipelines, improving both performance and stability while maintaining computational efficiency benefits.

Abstract: Large language model (LLM) reinforcement learning (RL) pipelines are often bottlenecked by rollout generation, making end-to-end training slow. Recent work mitigates this by running rollouts with quantization to accelerate decoding, which is the most expensive stage of the RL loop. However, these setups destabilize optimization by amplifying the training-inference gap: rollouts are operated at low precision, while learning updates are computed at full precision. To address this challenge, we propose QaRL (Rollout Alignment Quantization-Aware RL), which aligns training-side forward with the quantized rollout to minimize mismatch. We further identify a failure mode in quantized rollouts: long-form responses tend to produce repetitive, garbled tokens (error tokens). To mitigate these problems, we introduce TBPO (Trust-Band Policy Optimization), a sequence-level objective with dual clipping for negative samples, aimed at keeping updates within the trust region. On Qwen3-30B-A3B MoE for math problems, QaRL outperforms quantized-rollout training by +5.5 while improving stability and preserving low-bit throughput benefits.

Method: Three key components: 1) Block-wise progressive training reducing precision stage by stage, 2) Nested integer quantization grids for “train once, deploy any precision”, 3) Rounding-aware outlier channel splitting to mitigate quantization error while preserving outputs. Uses microscaling groups with E4M3 scales and custom 2-bit kernels.

Result: Achieves up to 11× speedup over BF16, significantly outperforms baselines like BitDistiller and EfficientQAT. Under W2A2 settings, achieves only 2.25 WikiText2 PPL loss compared to full-precision models on Llama2/3.

Conclusion: Bit-by-Bit enables stable and efficient ultra-low precision LLM training with minimal performance degradation, addressing key challenges in quantization-aware training.

Abstract: Training LLMs at ultra-low precision remains a formidable challenge. Direct low-bit QAT often suffers from convergence instability and substantial training costs, exacerbated by quantization noise from heavy-tailed outlier channels and error accumulation across layers. To address these issues, we present Bit-by-Bit, a progressive QAT framework with outlier channel splitting. Our approach integrates three key components: (1) block-wise progressive training that reduces precision stage by stage, ensuring stable initialization for low-bit optimization; (2) nested structure of integer quantization grids to enable a “train once, deploy any precision” paradigm, allowing a single model to support multiple bit-widths without retraining; (3) rounding-aware outlier channel splitting, which mitigates quantization error while acting as an identity transform that preserves the quantized outputs. Furthermore, we follow microscaling groups with E4M3 scales, capturing dynamic activation ranges in alignment with OCP/NVIDIA standards. To address the lack of efficient 2-bit kernels, we developed custom operators for both W2A2 and W2A16 configurations, achieving up to 11$\times$ speedup over BF16. Under W2A2 settings, Bit-by-Bit significantly outperforms baselines like BitDistiller and EfficientQAT on both Llama2/3, achieving a loss of only 2.25 WikiText2 PPL compared to full-precision models.

Method: Introduces Kuramoto oscillatory Phase Encoding (KoPE) as an additional evolving phase state to Vision Transformers, incorporating neuro-inspired synchronization mechanisms to enhance learning efficiency through synchronization-enhanced structure learning.

Result: KoPE improves training, parameter, and data efficiency of vision models; benefits tasks requiring structured understanding including semantic/panoptic segmentation, representation alignment with language, and few-shot abstract visual reasoning (ARC-AGI); accelerates attention concentration for learning efficiency.

Conclusion: Synchronization can serve as a scalable, neuro-inspired mechanism for advancing state-of-the-art neural network models, particularly for vision understanding tasks requiring structured reasoning.

Abstract: Spatiotemporal neural dynamics and oscillatory synchronization are widely implicated in biological information processing and have been hypothesized to support flexible coordination such as feature binding. By contrast, most deep learning architectures represent and propagate information through activation values, neglecting the joint dynamics of rate and phase. In this work, we introduce Kuramoto oscillatory Phase Encoding (KoPE) as an additional, evolving phase state to Vision Transformers, incorporating a neuro-inspired synchronization mechanism to advance learning efficiency. We show that KoPE can improve training, parameter, and data efficiency of vision models through synchronization-enhanced structure learning. Moreover, KoPE benefits tasks requiring structured understanding, including semantic and panoptic segmentation, representation alignment with language, and few-shot abstract visual reasoning (ARC-AGI). Theoretical analysis and empirical verification further suggest that KoPE can accelerate attention concentration for learning efficiency. These results indicate that synchronization can serve as a scalable, neuro-inspired mechanism for advancing state-of-the-art neural network models.

Method: Theoretical analysis of rank collapse with Sinkhorn-normalized doubly stochastic attention matrices, comparing to Softmax row-stochastic attention. Empirical validation on sentiment analysis and image classification tasks.

Result: Doubly stochastic attention preserves rank better than Softmax attention. Skip connections remain crucial for mitigating rank collapse. Theoretical bound shows rank decays to one doubly exponentially with depth for Sinkhorn normalization, similar to Softmax.

Conclusion: Doubly stochastic attention via Sinkhorn normalization is theoretically sound and empirically effective for mitigating rank collapse in Transformers, though fundamental rank decay patterns persist.

Abstract: The self-attention mechanism is central to the success of Transformer architectures. However, standard row-stochastic attention has been shown to suffer from significant signal degradation across layers. In particular, it can induce rank collapse, resulting in increasingly uniform token representations, as well as entropy collapse, characterized by highly concentrated attention distributions. Recent work has highlighted the benefits of doubly stochastic attention as a form of entropy regularization, promoting a more balanced attention distribution and leading to improved empirical performance. In this paper, we study rank collapse across network depth and show that doubly stochastic attention matrices normalized with Sinkhorn algorithm preserve rank more effectively than standard Softmax row-stochastic ones. As previously shown for Softmax, skip connections are crucial to mitigate rank collapse. We empirically validate this phenomenon on both sentiment analysis and image classification tasks. Moreover, we derive a theoretical bound for the pure self-attention rank decay when using Sinkhorn normalization and find that rank decays to one doubly exponentially with depth, a phenomenon that has already been shown for Softmax.

Method: ProD methods cast length prediction as robust estimation from heavy-tailed prompt-conditioned length distributions. Two variants: ProD-M uses median-based targets from multiple independent generations for robust point prediction, and ProD-D uses distributional targets that preserve prompt-conditioned uncertainty. Both reuse served LLM’s hidden states.

Result: Experiments across diverse scenarios show consistent gains in prediction quality compared to existing methods that use one-shot sampled lengths as labels.

Conclusion: Modeling prompt-conditioned output length distributions is crucial for accurate LLM serving optimization, and ProD methods provide effective solutions for robust length prediction by leveraging multiple generations per prompt.

Abstract: Output-length prediction is important for efficient LLM serving, as it directly affects batching, memory reservation, and scheduling. For prompt-only length prediction, most existing methods use a one-shot sampled length as the label, implicitly treating each prompt as if it had one true target length. We show that this is unreliable: even under a fixed model and decoding setup, the same prompt induces a \emph{prompt-conditioned output length distribution}, not a deterministic scalar, and this distribution is consistent with \emph{heavy-tailed} behavior. Motivated by this, we cast length prediction as robust estimation from heavy-tailed prompt-conditioned length distributions. We propose prompt-conditioned length distribution (ProD) methods, which construct training targets from multiple independent generations of the same prompt. Two variants are developed to reuse the served LLM’s hidden states: \mbox{ProD-M}, which uses a median-based target for robust point prediction, and ProD-D, which uses a distributional target that preserves prompt-conditioned uncertainty. We provide theoretical justifications by analyzing the estimation error under a surrogate model. Experiments across diverse scenarios show consistent gains in prediction quality.

Method: Proposes a framework that modularizes tabular data disentanglement into four core components: data extraction, data modeling, model analysis, and latent representation extrapolation. Demonstrates the framework through a case study on synthetic tabular data generation.

Result: The framework provides a systematic view of tabular data disentanglement and demonstrates applicability through a synthetic data generation case study, showing potential for downstream tasks like data synthesis.

Conclusion: This work provides deeper understanding of tabular data disentanglement, lays foundation for future research in robust, efficient, scalable disentanglement techniques, and shows framework’s potential for practical applications like synthetic data generation.

Abstract: Tabular data, widely used in various applications such as industrial control systems, finance, and supply chain, often contains complex interrelationships among its attributes. Data disentanglement seeks to transform such data into latent variables with reduced interdependencies, facilitating more effective and efficient processing. Despite the extensive studies on data disentanglement over image, text, or audio data, tabular data disentanglement may require further investigation due to the more intricate attribute interactions typically found in tabular data. Moreover, due to the highly complex interrelationships, direct translation from other data domains results in suboptimal data disentanglement. Existing tabular data disentanglement methods, such as factor analysis, CT-GAN, and VAE face limitations including scalability issues, mode collapse, and poor extrapolation. In this paper, we propose the use of a framework to provide a systematic view on tabular data disentanglement that modularizes the process into four core components: data extraction, data modeling, model analysis, and latent representation extrapolation. We believe this work provides a deeper understanding of tabular data disentanglement and existing methods, and lays the foundation for potential future research in developing robust, efficient, and scalable data disentanglement techniques. Finally, we demonstrate the framework’s applicability through a case study on synthetic tabular data generation, showcasing its potential in the particular downstream task of data synthesis.

Method: CRISP-DM methodology with hypothesis-driven exploratory analysis, feature refinement, comparative assessment of Logistic Regression, Random Forest, XGBoost, and Decision Tree, using SMOTE for class imbalance and GridSearchCV for hyperparameter optimization.

Result: The proposed framework provides a robust and scalable solution for fraud prevention in FinTech systems, though specific performance metrics are not detailed in the abstract.

Conclusion: Machine learning with proper handling of class imbalance and hyperparameter optimization offers an effective approach to financial fraud detection in digital payment systems.

Abstract: The expansion of digital payment systems has heightened both the scale and intricacy of online financial transactions, thereby increasing vulnerability to fraudulent activities. Detecting fraud effectively is complicated by the changing nature of attack strategies and the significant disparity between genuine and fraudulent transactions. This research introduces a machine learning-based fraud detection framework utilizing the PaySim synthetic financial transaction dataset. Following the CRISP-DM methodology, the study includes hypothesis-driven exploratory analysis, feature refinement, and a comparative assessment of baseline models such as Logistic Regression and tree-based classifiers like Random Forest, XGBoost, and Decision Tree. To tackle class imbalance, SMOTE is employed, and model performance is enhanced through hyperparameter tuning with GridSearchCV. The proposed framework provides a robust and scalable solution to enhance fraud prevention capabilities in FinTech transaction systems. Keywords: fraud detection, imbalanced data, HPO, SMOTE

Method: Introduces holistic evaluation framework exploring predictive performance vs. resource consumption. Applies theoretically bounded pruning to leading hybrid classifiers Hydra and Quant, and presents Hydrant - a novel prunable combination of both. Conducts over 4000 experiments across 20 MONSTER datasets, 13 methods, and three compute setups.

Result: Pruning can significantly reduce energy consumption by up to 80% while maintaining competitive predictive quality, usually costing less than 5% accuracy. Systematic analysis shows how model design, hyperparameters, and hardware choices affect practical TSC performance.

Conclusion: The proposed methodology, experimental results, and accompanying software advance TSC toward sustainable and reproducible practice by providing comprehensive understanding of performance-resource trade-offs.

Abstract: Time series classification (TSC) enables important use cases, however lacks a unified understanding of performance trade-offs across models, datasets, and hardware. While resource awareness has grown in the field, TSC methods have not yet been rigorously evaluated for energy efficiency. This paper introduces a holistic evaluation framework that explicitly explores the balance of predictive performance and resource consumption in TSC. To boost efficiency, we apply a theoretically bounded pruning strategy to leading hybrid classifiers - Hydra and Quant - and present Hydrant, a novel, prunable combination of both. With over 4000 experimental configurations across 20 MONSTER datasets, 13 methods, and three compute setups, we systematically analyze how model design, hyperparameters, and hardware choices affect practical TSC performance. Our results showcase that pruning can significantly reduce energy consumption by up to 80% while maintaining competitive predictive quality, usually costing the model less than 5% of accuracy. The proposed methodology, experimental results, and accompanying software advance TSC toward sustainable and reproducible practice.

Method: The authors redefine the calibration objective to precisely align quantized model outputs with original full-precision outputs at each step. They identify that residual error comes from both preceding layer output differences and compensation-aware error (discrepancy between compensated and original weights within each layer). Using neuron decomposition from GPTAQ, they efficiently incorporate compensation-aware error into weight updates.

Result: Extensive experiments on various LLMs and quantization settings show that the proposed enhancements integrate seamlessly with both GPTQ and GPTAQ, significantly improving their quantization performance.

Conclusion: The paper presents ResComp, which improves weight compensation methods for LLM quantization by addressing fundamental issues in calibration objectives and residual error formulation, leading to better quantization performance across different models and settings.

Abstract: Methods based on weight compensation, which iteratively apply quantization and weight compensation to minimize the output error, have recently demonstrated remarkable success in quantizing Large Language Models (LLMs). The representative work, GPTQ, introduces several key techniques that make such iterative methods practical for LLMs with billions of parameters. GPTAQ extends this approach by introducing an asymmetric calibration process that aligns the output of each quantized layer with its full-precision counterpart, incorporating a residual error into the weight compensation framework. In this work, we revisit the formulation of the residual error. We identify a sub-optimal calibration objective in existing methods: during the intra-layer calibration process, they align the quantized output with the output from compensated weights, rather than the true output from the original full-precision model. Therefore, we redefine the objective to precisely align the quantized model’s output with the original output of the full-precision model at each step. We then reveal that the residual error originates not only from the output difference of the preceding layer but also from the discrepancy between the compensated and original weights within each layer, which we name the ‘compensation-aware error’. By inheriting the neuron decomposition technique from GPTAQ, we can efficiently incorporate this compensation-aware error into the weight update process. Extensive experiments on various LLMs and quantization settings demonstrate that our proposed enhancements integrate seamlessly with both GPTQ and GPTAQ, significantly improving their quantization performance. Our code is publicly available at https://github.com/list0830/ResComp.

Method: The authors establish a fundamental distinction between “untraining” and “unlearning” through conceptual analysis and technical definitions, mapping existing literature to each problem formulation.

Result: The paper provides clear definitions: “untraining” reverses the effect of training on specific forget set examples, while “unlearning” aims to remove entire underlying distributions/concepts represented by those examples.

Conclusion: Disambiguating these technical definitions is crucial for accelerating progress in machine unlearning research, enabling proper algorithm evaluation and identifying overlooked research directions.

Abstract: As models are getting larger and are trained on increasing amounts of data, there has been an explosion of interest into how we can delete'' specific data points or behaviours from a trained model, after the fact. This goal has been referred to as machine unlearning’’. In this note, we argue that the term unlearning'' has been overloaded, with different research efforts spanning two distinct problem formulations, but without that distinction having been observed or acknowledged in the literature. This causes various issues, including ambiguity around when an algorithm is expected to work, use of inappropriate metrics and baselines when comparing different algorithms to one another, difficulty in interpreting results, as well as missed opportunities for pursuing critical research directions. In this note, we address this issue by establishing a fundamental distinction between two notions that we identify as \unlearning and \untraining, illustrated in Figure 1. In short, \untraining aims to reverse the effect of having trained on a given forget set, i.e. to remove the influence that that specific forget set examples had on the model during training. On the other hand, the goal of \unlearning is not just to remove the influence of those given examples, but to use those examples for the purpose of more broadly removing the entire underlying distribution from which those examples were sampled (e.g. the concept or behaviour that those examples represent). We discuss technical definitions of these problems and map problem settings studied in the literature to each. We hope to initiate discussions on disambiguating technical definitions and identify a set of overlooked research questions, as we believe that this a key missing step for accelerating progress in the field of unlearning’’.

Method: Proposes an ordered three-stage pipeline: 1) unstructured pruning for capacity reduction, 2) INT8 quantization-aware training (QAT) for dominant runtime benefits, and 3) knowledge distillation (KD) to recover accuracy within the constrained sparse INT8 regime. The ordering is critical, with pruning acting as a preconditioner for subsequent low-precision optimization.

Result: The pipeline achieves 0.99-1.42 ms CPU latency with competitive accuracy and compact checkpoints across CIFAR-10/100 using ResNet-18, WRN-28-10, and VGG-16-BN backbones. Ordered ablations confirm the proposed ordering (pruning → QAT → KD) generally performs best among permutations, creating a stronger accuracy-size-latency frontier than any single technique alone.

Conclusion: Provides a simple guideline for edge deployment: evaluate compression choices in the joint accuracy-size-latency space using measured runtime rather than proxy metrics alone. The ordered combination of pruning, quantization, and distillation yields practical benefits for CPU-constrained environments.

Abstract: Modern deployment often requires trading accuracy for efficiency under tight CPU and memory constraints, yet common compression proxies such as parameter count or FLOPs do not reliably predict wall-clock inference time. In particular, unstructured sparsity can reduce model storage while failing to accelerate (and sometimes slightly slowing down) standard CPU execution due to irregular memory access and sparse kernel overhead. Motivated by this gap between compression and acceleration, we study a practical, ordered pipeline that targets measured latency by combining three widely used techniques: unstructured pruning, INT8 quantization-aware training (QAT), and knowledge distillation (KD). Empirically, INT8 QAT provides the dominant runtime benefit, while pruning mainly acts as a capacity-reduction pre-conditioner that improves the robustness of subsequent low-precision optimization; KD, applied last, recovers accuracy within the already constrained sparse INT8 regime without changing the deployment form. We evaluate on CIFAR-10/100 using three backbones (ResNet-18, WRN-28-10, and VGG-16-BN). Across all settings, the ordered pipeline achieves a stronger accuracy-size-latency frontier than any single technique alone, reaching 0.99-1.42 ms CPU latency with competitive accuracy and compact checkpoints. Controlled ordering ablations with a fixed 20/40/40 epoch allocation further confirm that stage order is consequential, with the proposed ordering generally performing best among the tested permutations. Overall, our results provide a simple guideline for edge deployment: evaluate compression choices in the joint accuracy-size-latency space using measured runtime, rather than proxy metrics alone.

Method: Manually refined all 197 volumes from CRLM-CT-Seg dataset to create validated benchmark, then compared cascaded (Liver->CRLM->FLR) and end-to-end strategies using nnU-Net, SwinUNETR, and STU-Net architectures.

Result: Cascaded nnU-Net achieved best FLR segmentation Dice (0.767), while pretrained STU-Net provided superior CRLM segmentation (0.620 Dice) and was more robust to cascaded errors.

Conclusion: Provides first validated benchmark and reproducible framework to accelerate research in AI-assisted surgical planning for liver resection procedures.

Abstract: Accurate segmentation of the future liver remnant (FLR) is critical for surgical planning in colorectal liver metastases (CRLM) to prevent fatal post-hepatectomy liver failure. However, this segmentation task is technically challenging due to complex resection boundaries, convoluted hepatic vasculature and diffuse metastatic lesions. A primary bottleneck in developing automated AI tools has been the lack of high-fidelity, validated data. We address this gap by manually refining all 197 volumes from the public CRLM-CT-Seg dataset, creating the first open-source, validated benchmark for this task. We then establish the first segmentation baselines, comparing cascaded (Liver->CRLM->FLR) and end-to-end (E2E) strategies using nnU-Net, SwinUNETR, and STU-Net. We find a cascaded nnU-Net achieves the best final FLR segmentation Dice (0.767), while the pretrained STU-Net provides superior CRLM segmentation (0.620 Dice) and is significantly more robust to cascaded errors. This work provides the first validated benchmark and a reproducible framework to accelerate research in AI-assisted surgical planning.

Method: Combines literature synthesis with platform-level data analysis of major competition platforms (Kaggle, Zindi), examining workflows, evaluation methodologies, reward structures, competition quality, participant expertise, and demographic trends, supplemented by practitioner insights.

Result: MLCs function at the intersection of academic research and industrial application, fostering knowledge exchange, data sharing, and practical methodologies. They have strong ties to open-source communities, promote collaboration and reproducibility, shape research priorities, inform industry standards, and enable large-scale crowdsourced problem-solving.

Conclusion: Machine learning competitions play a pivotal role in AI development by advancing innovation, promoting collaboration, and driving technological progress, with significant implications for researchers, practitioners, and competition organizers in shaping the future trajectory of AI.

Abstract: Machine learning competitions (MLCs) play a pivotal role in advancing artificial intelligence (AI) by fostering innovation, skill development, and practical problem-solving. This study provides a comprehensive analysis of major competition platforms such as Kaggle and Zindi, examining their workflows, evaluation methodologies, and reward structures. It further assesses competition quality, participant expertise, and global reach, with particular attention to demographic trends among top-performing competitors. By exploring the motivations of competition hosts, this paper underscores the significant role of MLCs in shaping AI development, promoting collaboration, and driving impactful technological progress. Furthermore, by combining literature synthesis with platform-level data analysis and practitioner insights a comprehensive understanding of the MLC ecosystem is provided. Moreover, the paper demonstrates that MLCs function at the intersection of academic research and industrial application, fostering the exchange of knowledge, data, and practical methodologies across domains. Their strong ties to open-source communities further promote collaboration, reproducibility, and continuous innovation within the broader ML ecosystem. By shaping research priorities, informing industry standards, and enabling large-scale crowdsourced problem-solving, these competitions play a key role in the ongoing evolution of AI. The study provides insights relevant to researchers, practitioners, and competition organizers, and includes an examination of the future trajectory and sustained influence of MLCs on AI development.

Method: PRAC manipulates the model’s internal preferences by redirecting its attention toward a stealthy adversarial patch. Unlike prior attacks targeting VLM output directly, this approach works by influencing the model’s attention mechanisms. The attack requires white-box access for creation but generalizes to fine-tuned versions of the same model.

Result: PRAC successfully manipulates the selection process of a Computer Use Agent on an online shopping platform toward a chosen target product. The attack demonstrates generalization capability to fine-tuned versions of the same base model.

Conclusion: The attack presents a critical security threat as multiple companies build specific Computer Use Agents based on open weights models, highlighting the need for better security measures in vision-language models used for GUI interaction.

Abstract: Advancements in multimodal foundation models have enabled the development of Computer Use Agents (CUAs) capable of autonomously interacting with GUI environments. As CUAs are not restricted to certain tools, they allow to automate more complex agentic tasks but at the same time open up new security vulnerabilities. While prior work has concentrated on the language modality, the vulnerability of the vision modality has received less attention. In this paper, we introduce PRAC, a novel attack that, unlike prior work targeting the VLM output directly, manipulates the model’s internal preferences by redirecting its attention toward a stealthy adversarial patch. We show that PRAC is able to manipulate the selection process of a CUA on an online shopping platform towards a chosen target product. While we require white-box access to the model for the creation of the attack, we show that our attack generalizes to fine-tuned versions of the same model, presenting a critical threat as multiple companies build specific CUAs based on open weights models.

Method: Formalizes personalization as individual actionability with hard constraints (specifying actionable features) and soft constraints (capturing preferences over action values/costs). Operationalizes within causal algorithmic recourse framework using pre-hoc user-prompting where individuals express preferences via rankings or scores before recourse generation.

Result: Individual actionability constraints, especially hard ones, substantially degrade plausibility and validity of recourse recommendations across amortized and non-amortized approaches. Personalization reveals disparities in cost and plausibility of recourse actions across socio-demographic groups.

Conclusion: Highlights need for principled definitions, careful operationalization, and rigorous evaluation of personalization in algorithmic recourse due to important trade-offs and revealed disparities.

Abstract: Algorithmic recourse aims to provide actionable recommendations that enable individuals to change unfavorable model outcomes, and prior work has extensively studied properties such as efficiency, robustness, and fairness. However, the role of personalization in recourse remains largely implicit and underexplored. While existing approaches incorporate elements of personalization through user interactions, they typically lack an explicit definition of personalization and do not systematically analyze its downstream effects on other recourse desiderata. In this paper, we formalize personalization as individual actionability, characterized along two dimensions: hard constraints that specify which features are individually actionable, and soft, individualized constraints that capture preferences over action values and costs. We operationalize these dimensions within the causal algorithmic recourse framework, adopting a pre-hoc user-prompting approach in which individuals express preferences via rankings or scores prior to the generation of any recourse recommendation. Through extensive empirical evaluation, we investigate how personalization interacts with key recourse desiderata, including validity, cost, and plausibility. Our results highlight important trade-offs: individual actionability constraints, particularly hard ones, can substantially degrade the plausibility and validity of recourse recommendations across amortized and non-amortized approaches. Notably, we also find that incorporating individual actionability can reveal disparities in the cost and plausibility of recourse actions across socio-demographic groups. These findings underscore the need for principled definitions, careful operationalization, and rigorous evaluation of personalization in algorithmic recourse.

Method: Formalizes as POMDP with privileged planner agent using anytime-feasible MPC algorithm. Proposes P2P-SAC (Planner-to-Policy Soft Actor-Critic) that distills planner’s privileged knowledge to the learning agent to mitigate partial observability.

Result: The approach improves both sample efficiency and final policy performance, validated in simulation using NVIDIA Isaac Lab and successfully deployed on real-world Unitree Go2 quadruped navigating complex, obstacle-rich environments.

Conclusion: The framework effectively leverages privileged information available only during training to overcome partial observability challenges, with theoretical guarantees and practical real-world robotic applications.

Abstract: This paper addresses the problem of training a reinforcement learning (RL) policy under partial observability by exploiting a privileged, anytime-feasible planner agent available exclusively during training. We formalize this as a Partially Observable Markov Decision Process (POMDP) in which a planner agent with access to an approximate dynamical model and privileged state information guides a learning agent that observes only a lossy projection of the true state. To realize this framework, we introduce an anytime-feasible Model Predictive Control (MPC) algorithm that serves as the planner agent. For the learning agent, we propose Planner-to-Policy Soft Actor-Critic (P2P-SAC), a method that distills the planner agent’s privileged knowledge to mitigate partial observability and thereby improve both sample efficiency and final policy performance. We support this framework with rigorous theoretical analysis. Finally, we validate our approach in simulation using NVIDIA Isaac Lab and successfully deploy it on a real-world Unitree Go2 quadruped navigating complex, obstacle-rich environments.

Method: Two-mode framework: 1) Single-trial mode uses large language models to infer suitable strategies from user-provided or automatically detected data characteristics; 2) Multi-trial mode employs lightweight genetic search to efficiently explore alternatives under constrained budgets.

Result: Extensive experiments across diverse datasets show the approach enhances robustness and generalization under non-IID conditions while reducing the need for manual intervention.

Conclusion: The work advances accessible and adaptive federated learning by automating one of its most critical design decisions - aggregation strategy selection.

Abstract: Federated Learning enables collaborative model training without centralising data, but its effectiveness varies with the selection of the aggregation strategy. This choice is non-trivial, as performance varies widely across datasets, heterogeneity levels, and compute constraints. We present an end-to-end framework that automates, streamlines, and adapts aggregation strategy selection for federated learning. The framework operates in two modes: a single-trial mode, where large language models infer suitable strategies from user-provided or automatically detected data characteristics, and a multi-trial mode, where a lightweight genetic search efficiently explores alternatives under constrained budgets. Extensive experiments across diverse datasets show that our approach enhances robustness and generalisation under non-IID conditions while reducing the need for manual intervention. Overall, this work advances towards accessible and adaptive federated learning by automating one of its most critical design decisions, the choice of an aggregation strategy.

Method: Pearl learns predictive embeddings from expert tool-use trajectories entirely in latent space using a JEPA-inspired framework. Unlike reconstruction-based methods that autoregressively generate latent tokens, Pearl directly learns predictive embeddings while preserving the standard vision-language generation pipeline. It’s model-agnostic, simple to train, and supports multi-step tool use trajectories.

Result: Experiments across multiple perception benchmarks show Pearl matches or outperforms standard supervised fine-tuning and reconstruction-based latent reasoning approaches. The paper also provides empirical evidence that reconstruction-based methods primarily learn embeddings rather than image edits in latent space.

Conclusion: Predictive embedding learning in latent space is a more principled alternative to reconstruction-based methods for tool-augmented multimodal reasoning, offering efficiency gains without explicit tool invocation while maintaining or improving performance.

Abstract: Tool-augmented multimodal reasoning enables visual language models (VLMs) to improve perception by interacting with external tools (e.g., cropping, depth estimation). However, such approaches incur substantial inference overhead, require specialized supervision, and are prone to erroneous tool calls. We propose Pearl (Predictive Embedding Alignment for Reasoning in Latent space), a JEPA-inspired framework that learns from expert tool-use trajectories entirely in the latent space, eliminating the need for explicit tool invocation at inference time. Unlike reconstruction-based latent reasoning methods, which autoregressively generate latent tokens and suffer from training-inference mismatch and limited support for multi-step tool use, Pearl directly learns predictive embeddings from multimodal trajectories while preserving the standard vision-language generation pipeline: it is model-agnostic, simple to train, and naturally supports trajectories with multiple tool calls. Experiments across multiple perception benchmarks show that Pearl matches or outperforms standard supervised fine-tuning and reconstruction-based latent reasoning approaches. Furthermore, we provide empirical evidence that reconstruction-based methods primarily learn embeddings rather than image edits in latent space, motivating predictive embedding learning as a more principled alternative.

Method: Proposes Alloc-MoE with two components: Alloc-L for layer-level allocation using sensitivity profiling and dynamic programming, and Alloc-T for token-level allocation that dynamically redistributes activations based on routing scores without increasing latency.

Result: Achieves 1.15× prefill and 1.34× decode speedups on DeepSeek-V2-Lite at half the original budget while maintaining model performance across multiple MoE models.

Conclusion: Alloc-MoE effectively optimizes expert activation allocation at both layer and token levels, enabling efficient MoE inference under constrained activation budgets without performance degradation.

Abstract: Mixture-of-Experts (MoE) has become a dominant architecture for scaling large language models due to their sparse activation mechanism. However, the substantial number of expert activations creates a critical latency bottleneck during inference, especially in resource-constrained deployment scenarios. Existing approaches that reduce expert activations potentially lead to severe model performance degradation. In this work, we introduce the concept of \emph{activation budget} as a constraint on the number of expert activations and propose Alloc-MoE, a unified framework that optimizes budget allocation coordinately at both the layer and token levels to minimize performance degradation. At the layer level, we introduce Alloc-L, which leverages sensitivity profiling and dynamic programming to determine the optimal allocation of expert activations across layers. At the token level, we propose Alloc-T, which dynamically redistributes activations based on routing scores, optimizing budget allocation without increasing latency. Extensive experiments across multiple MoE models demonstrate that Alloc-MoE maintains model performance under a constrained activation budget. Especially, Alloc-MoE achieves $1.15\times$ prefill and $1.34\times$ decode speedups on DeepSeek-V2-Lite at half of the original budget.

Method: Evaluated three unlearning methods (Prompt Erasure, Prompt Reweighting, Refusal Vector) on CLIP models (ViT-B-32, ViT-L/14, ViT-B/16) using CelebA dataset in zero-shot classification. Analyzed intersectional groups by age and gender with metrics including per-group accuracy shifts, demographic parity gaps, and redistribution score.

Result: Unlearning redistributes bias primarily along gender rather than age boundaries. Removing dominant Young Female group consistently transfers performance to Old Female across all model scales. Refusal Vector reduces redistribution but fails to achieve complete forgetting and degrades retained performance.

Conclusion: Current unlearning methods have fundamental limitations: without accounting for embedding geometry, they risk amplifying bias in retained groups rather than eliminating bias.

Abstract: Machine unlearning enables models to selectively forget training data, driven by privacy regulations such as GDPR and CCPA. However, its fairness implications remain underexplored: when a model forgets a demographic group, does it neutralize that concept or redistribute it to correlated groups, potentially amplifying bias? We investigate this bias redistribution phenomenon on CelebA using CLIP models (ViT/B-32, ViT-L/14, ViT-B/16) under a zero-shot classification setting across intersectional groups defined by age and gender. We evaluate three unlearning methods, Prompt Erasure, Prompt Reweighting, and Refusal Vector using per-group accuracy shifts, demographic parity gaps, and a redistribution score. Our results show that unlearning does not eliminate bias but redistributes it primarily along gender rather than age boundaries. In particular, removing the dominant Young Female group consistently transfers performance to Old Female across all model scales, revealing a gender-dominant structure in CLIP’s embedding space. While the Refusal Vector method reduces redistribution, it fails to achieve complete forgetting and significantly degrades retained performance. These findings highlight a fundamental limitation of current unlearning methods: without accounting for embedding geometry, they risk amplifying bias in retained groups.

Method: The authors study a more natural model where rewards have direct dependencies on hidden states (in addition to observed contexts). They develop a fully adaptive strategy that estimates HMM parameters online and analyze it to obtain high-probability regret bounds.

Result: The paper obtains stronger, high-probability regret bounds that do not depend on reward functions and only depend on the model through HMM parameter estimation, eliminating unnecessary complex dependencies like reward gaps present in prior work.

Conclusion: The work provides improved theoretical guarantees for linear bandits with HMM contexts by considering a more natural model and developing adaptive online estimation, resulting in cleaner, more practical regret bounds.

Abstract: We revisit the finite-armed linear bandit model by Nelson et al. (2022), where contexts and rewards are governed by a finite hidden Markov chain. Nelson et al. (2022) approach this model by a reduction to linear contextual bandits; but to do so, they actually introduce a simplification in which rewards are linear functions of the posterior probabilities over the hidden states given the observed contexts, rather than functions of the hidden states themselves. Their analysis (but not their algorithm) also does not take into account the estimation of the HMM parameters, and only tackles expected, not high-probability, bounds, which suffer in addition from unnecessary complex dependencies on the model (like reward gaps). We instead study the more natural model incorporating direct dependencies in the hidden states (on top of dependencies on the observed contexts, as is natural for contextual bandits) and also obtain stronger, high-probability, regret bounds for a fully adaptive strategy that estimates HMM parameters online. These bounds do not depend on the reward functions and only depend on the model through the estimation of the HMM parameters.

Method: SOLAR expresses each PEFT update as a linear combination of basis vectors formed from the foundation model’s singular vectors with controlled random perturbations. It exploits subspace similarity (alignment of principal directions) between foundation models and task-specific updates, decoupling adapter size from PEFT structure.

Result: Experiments on language (LLaMA, GPT) and vision (ViT) tasks show SOLAR preserves task performance while significantly reducing model representation sizes, offering effective communication-efficient solutions for distributed systems and edge devices.

Conclusion: SOLAR provides a model-agnostic, theoretically-grounded compression framework that reduces PEFT adapter costs while maintaining performance, making it suitable for deployment in resource-constrained environments.

Abstract: Parameter-efficient fine-tuning (PEFT) methods, such as LoRA, enable scalable adaptation of foundation models by injecting low-rank adapters. However, their communication and storage costs remain a major bottleneck in resource-constrained settings. We propose SOLAR (Subspace-Oriented Latent Adapter Reparameterization), a post-training compression framework that substantially reduces the communication cost (i.e., the number of parameters to transmit or store) of PEFT adapters. SOLAR expresses each PEFT update as a linear combination of basis vectors formed from the foundation model’s singular vectors with controlled random perturbations. By exploiting the subspace similarity (the alignment of principal directions) between the foundation model and task-specific fine-tuned updates, SOLAR decouples the adapter size from PEFT structure and ensures compact yet expressive representations. It is model-agnostic and compatible with existing PEFT methods, including LoRA, AdaLoRA, and other adapter modules. We theoretically establish a bound on the reconstruction error. Experiments on language and vision tasks using LLaMA, GPT, and ViT models demonstrate that SOLAR preserves task performance while significantly reducing model representation sizes, offering an effective and communication-efficient solution for deployment in distributed systems and edge devices.

Method: Shift- and stretch-invariant non-negative matrix factorization framework that estimates both integer and non-integer temporal shifts and temporal stretching, implemented in frequency domain where shifts correspond to phase modifications and stretching handled via zero-padding/truncation.

Result: Demonstrated on synthetic data and brain emission tomography data that the model accounts for stretching to provide more detailed characterization of brain tissue structure.

Conclusion: The framework effectively addresses temporal distortions in dynamic neuroimaging data, improving analysis of diffusion-like processes in emission tomography measurements.

Abstract: Dynamic neuroimaging data, such as emission tomography measurements of radiotracer transport in blood or cerebrospinal fluid, often exhibit diffusion-like properties. These introduce distance-dependent temporal delays, scale-differences, and stretching effects that limit the effectiveness of conventional linear modeling and decomposition methods. To address this, we present the shift- and stretch-invariant non-negative matrix factorization framework. Our approach estimates both integer and non-integer temporal shifts as well as temporal stretching, all implemented in the frequency domain, where shifts correspond to phase modifications, and where stretching is handled via zero-padding or truncation. The model is implemented in PyTorch (https://github.com/anders-s-olsen/shiftstretchNMF). We demonstrate on synthetic data and brain emission tomography data that the model is able to account for stretching to provide more detailed characterization of brain tissue structure.

Method: Created Text2JSON benchmark for context-intensive tasks requiring structured knowledge extraction from raw text. Evaluated modern KV offloading on Text2JSON and other context-intensive tasks using Llama 3 and Qwen 3 models. Analyzed performance degradation causes and proposed alternative strategies.

Result: Found significant performance degradation with KV-cache offloading on context-intensive tasks. Identified two key issues: low-rank projection of keys and unreliable landmarks. Proposed simpler alternative strategy that significantly improves accuracy across multiple LLM families and benchmarks.

Conclusion: KV-cache offloading techniques need comprehensive evaluation on context-intensive tasks. The proposed improvements address accuracy degradation, highlighting the importance of rigorous testing for long-context compression methods.

Abstract: With the growing demand for long-context LLMs across a wide range of applications, the key-value (KV) cache has become a critical bottleneck for both latency and memory usage. Recently, KV-cache offloading has emerged as a promising approach to reduce memory footprint and inference latency while preserving accuracy. Prior evaluations have largely focused on tasks that do not require extracting large amounts of information from the context. In this work, we study KV-cache offloading on context-intensive tasks: problems where the solution requires looking up a lot of information from the input prompt. We create and release the Text2JSON benchmark, a highly context-intensive task that requires extracting structured knowledge from raw text. We evaluate modern KV offloading on Text2JSON and other context-intensive tasks and find significant performance degradation on both Llama 3 and Qwen 3 models. Our analysis identifies two key reasons for poor accuracy: low-rank projection of keys and unreliable landmarks, and proposes a simpler alternative strategy that significantly improves accuracy across multiple LLM families and benchmarks. These findings highlight the need for a comprehensive and rigorous evaluation of long-context compression techniques.

Method: Proposes VGM²P (Value Guidance Multi-agent MeanFlow Policy) that uses global advantage values to guide agent collaboration, treats optimal policy learning as conditional behavior cloning, and leverages classifier-free guidance MeanFlow for policy training and execution to improve expressiveness and inference efficiency.

Result: Experiments on tasks with both discrete and continuous action spaces show VGM²P achieves performance comparable to state-of-the-art methods while being trained solely via conditional behavior cloning, demonstrating efficient and robust performance.

Conclusion: VGM²P provides a simple yet effective flow-based policy learning framework for offline MARL that enables efficient action generation with coefficient-insensitive conditional behavior cloning, addressing efficiency and robustness issues in existing methods.

Abstract: Offline multi-agent reinforcement learning (MARL) aims to learn the optimal joint policy from pre-collected datasets, requiring a trade-off between maximizing global returns and mitigating distribution shift from offline data. Recent studies use diffusion or flow generative models to capture complex joint policy behaviors among agents; however, they typically rely on multi-step iterative sampling, thereby reducing training and inference efficiency. Although further research improves sampling efficiency through methods like distillation, it remains sensitive to the behavior regularization coefficient. To address the above-mentioned issues, we propose Value Guidance Multi-agent MeanFlow Policy (VGM$^2$P), a simple yet effective flow-based policy learning framework that enables efficient action generation with coefficient-insensitive conditional behavior cloning. Specifically, VGM$^2$P uses global advantage values to guide agent collaboration, treating optimal policy learning as conditional behavior cloning. Additionally, to improve policy expressiveness and inference efficiency in multi-agent scenarios, it leverages classifier-free guidance MeanFlow for both policy training and execution. Experiments on tasks with both discrete and continuous action spaces demonstrate that, even when trained solely via conditional behavior cloning, VGM$^2$P efficiently achieves performance comparable to state-of-the-art methods.

Method: Proposes Long-Term Embeddings (LTE) framework that constrains embeddings to fixed semantic basis of content-based item representations. Investigates integration strategies for causal language modeling, addressing data leakage issues. Evaluates two representations: heuristic average and asymmetric autoencoder with fixed decoder grounded in semantic basis.

Result: Online A/B tests on Zalando demonstrate significant uplifts in both user engagement and financial metrics when integrating LTE as contextual prefix token using lagged window approach.

Conclusion: LTE framework effectively bridges the gap between short-term intent and long-term preferences in sequential recommenders while solving production challenges related to feature consistency and offline-online mismatches.

Abstract: Modern transformer-based sequential recommenders excel at capturing short-term intent but often suffer from recency bias, overlooking stable long-term preferences. While extending sequence lengths is an intuitive fix, it is computationally inefficient, and recent interactions tend to dominate the model’s attention. We propose Long-Term Embeddings (LTE) as a high-inertia contextual anchor to bridge this gap. We address a critical production challenge: the point-in-time consistency problem caused by infrastructure constraints, as feature stores typically host only a single “live” version of features. This leads to an offline-online mismatch during model deployments and rollbacks, as models are forced to process evolved representations they never saw during training. To resolve this, we introduce an LTE framework that constrains embeddings to a fixed semantic basis of content-based item representations, ensuring cross-version compatibility. Furthermore, we investigate integration strategies for causal language modeling, considering the data leakage issue that occurs when the LTE and the transformer’s short-term sequence share a temporal horizon. We evaluate two representations: a heuristic average and an asymmetric autoencoder with a fixed decoder grounded in the semantic basis to enable behavioral fine-tuning while maintaining stability. Online A/B tests on Zalando demonstrate that integrating LTE as a contextual prefix token using a lagged window yields significant uplifts in both user engagement and financial metrics.

Method: Proposes Equivariant MeanFlow (EQUIMF) with unified time bridge and average-velocity updates with mutual conditioning between structure and geometry, plus novel discrete MeanFlow formulation for discrete graph structures.

Result: EQUIMF consistently outperforms prior diffusion and flow-matching methods in generation quality, physical validity, and sampling efficiency.

Conclusion: EQUIMF provides a unified SE(3)-equivariant framework for joint generation of graph structure and geometry with improved physical consistency and efficiency.

Abstract: Graph-structured data jointly contain discrete topology and continuous geometry, which poses fundamental challenges for generative modeling due to heterogeneous distributions, incompatible noise dynamics, and the need for equivariant inductive biases. Existing flow-matching approaches for graph generation typically decouple structure from geometry, lack synchronized cross-domain dynamics, and rely on iterative sampling, often resulting in physically inconsistent molecular conformations and slow sampling. To address these limitations, we propose Equivariant MeanFlow (EQUIMF), a unified SE(3)-equivariant generative framework that jointly models discrete and continuous components through synchronized MeanFlow dynamics. EQUIMF introduces a unified time bridge and average-velocity updates with mutual conditioning between structure and geometry, enabling efficient few-step generation while preserving physical consistency. Moreover, we develop a novel discrete MeanFlow formulation with a simple yet effective parameterization to support efficient generation over discrete graph structures. Extensive experiments demonstrate that EQUIMF consistently outperforms prior diffusion and flow-matching methods in generation quality, physical validity, and sampling efficiency.

Method: Uses circuit discovery to extract causal interactions between internal representations as circuits, then derives two metrics: (1) Dependency Depth Bias for model selection before deployment, and (2) Circuit Shift Score for monitoring performance after deployment under distribution shifts.

Result: Both metrics demonstrate significantly improved correlation with generalization performance across various tasks, outperforming existing proxies by an average of 13.4% (Dependency Depth Bias) and 34.1% (Circuit Shift Score).

Conclusion: Model circuits provide a reliable and label-free approach to predict generalization performance under distribution shift, addressing practical needs for model selection and performance monitoring in real-world applications.

Abstract: Reliable generalization metrics are fundamental to the evaluation of machine learning models. Especially in high-stakes applications where labeled target data are scarce, evaluation of models’ generalization performance under distribution shift is a pressing need. We focus on two practical scenarios: (1) Before deployment, how to select the best model for unlabeled target data? (2) After deployment, how to monitor model performance under distribution shift? The central need in both cases is a reliable and label-free proxy metric. Yet existing proxy metrics, such as model confidence or accuracy-on-the-line, are often unreliable as they only assess model output while ignoring the internal mechanisms that produce them. We address this limitation by introducing a new perspective: using the inner workings of a model, i.e., circuits, as a predictive metric of generalization performance. Leveraging circuit discovery, we extract the causal interactions between internal representations as a circuit, from which we derive two metrics tailored to the two practical scenarios. (1) Before deployment, we introduce Dependency Depth Bias, which measures different models’ generalization capability on target data. (2) After deployment, we propose Circuit Shift Score, which predicts a model’s generalization under different distribution shifts. Across various tasks, both metrics demonstrate significantly improved correlation with generalization performance, outperforming existing proxies by an average of 13.4% and 34.1%, respectively. Our code is available at https://github.com/deep-real/GenCircuit.

Method: Use universal differential equations concept to approximate history term via neural networks, converting MaRGE to system of ordinary differential equations solvable with standard methods like Runge-Kutta

Result: Proposed neural network approximation enables solving MaRGE with standard ODE solvers while preserving history effects

Conclusion: Neural networks provide effective approximation of Basset history term, making MaRGE computationally tractable without neglecting important physics

Abstract: The Maxey-Riley-Gatignol equations (MaRGE) model the motion of spherical inertial particles in a fluid. They contain the Basset force, an integral term which models history effects due to the formation of wakes and boundary layer effects. This causes the force that acts on a particle to depend on its past trajectory and complicates the numerical solution of MaRGE. Therefore, the Basset force is often neglected, despite substantial evidence that it has both quantitative and qualitative impact on the movement patterns of modelled particles. Using the concept of universal differential equations, we propose an approximation of the history term via neural networks which approximates MaRGE by a system of ordinary differential equations that can be solved with standard numerical solvers like Runge-Kutta methods.

Method: Introduces Echo Networks - recurrent networks represented as single connection matrices (rows=source neurons, columns=destination neurons, entries=weights), no layers, bidirectional connections, arbitrary input/output neuron assignment, with optional activation functions

Result: Successfully evaluated on electrocardiography (ECG) signal classification, demonstrating practical application for time series analysis

Conclusion: Echo Networks show promise primarily due to their matrix-based genome representation enabling systematic mutation/recombination through matrix computations and factorizations for evolutionary optimization

Abstract: For applications on the extreme edge, minimal networks of only a few dozen artificial neurons for event detection and classification in discrete time signals would be highly desirable. Feed-forward networks, RNNs, and CNNs evolved through evolutionary algorithms can all be successful in this respect but pose the problem of allowing little systematicity in mutation and recombination if the standard direct genetic encoding of the weights is used (as for instance in the classic NEAT algorithm). We therefore introduce Echo Networks, a type of recurrent network that consists of the connection matrix only, with the source neurons of the synapses represented as rows, destination neurons as columns and weights as entries. There are no layers, and connections between neurons can be bidirectional but are technically all recurrent. Input and output can be arbitrarily assigned to any of the neurons and only use an additional (optional) function in their computational path, e.g., a sigmoid to obtain a binary classification output. We evaluated Echo Networks successfully on the classification of electrocardiography signals but see the most promising potential in their genome representation as a single matrix, allowing matrix computations and factorisations as mutation and recombination operators.

Method: Analyzes internal representations of unlearned models, identifies feature-classifier misalignment phenomenon, proposes using class-mean features (CMF) classifier to explicitly enforce alignment between features and classifiers, and validates with classifier-only fine-tuning experiments.

Result: Shows that hidden features remain highly discriminative despite apparent unlearning success, simple linear probing can recover near-original accuracy, and CMF-based unlearning reduces forgotten information in representations while maintaining high retain accuracy.

Conclusion: Highlights the need for faithful representation-level evaluation of machine unlearning methods and demonstrates that explicit feature-classifier alignment through CMF-based approaches improves robustness against reintroduction of erased concepts.

Abstract: While numerous machine unlearning (MU) methods have recently been developed with promising results in erasing the influence of forgotten data, classes, or concepts, they are also highly vulnerable-for example, simple fine-tuning can inadvertently reintroduce erased concepts. In this paper, we address this contradiction by examining the internal representations of unlearned models, in contrast to prior work that focuses primarily on output-level behavior. Our analysis shows that many state-of-the-art MU methods appear successful mainly due to a misalignment between last-layer features and the classifier, a phenomenon we call feature-classifier misalignment. In fact, hidden features remain highly discriminative, and simple linear probing can recover near-original accuracy. Assuming neural collapse in the original model, we further demonstrate that adjusting only the classifier can achieve negligible forget accuracy while preserving retain accuracy, and we corroborate this with experiments using classifier-only fine-tuning. Motivated by these findings, we propose MU methods based on a class-mean features (CMF) classifier, which explicitly enforces alignment between features and classifiers. Experiments on standard benchmarks show that CMF-based unlearning reduces forgotten information in representations while maintaining high retain accuracy, highlighting the need for faithful representation-level evaluation of MU.

Method: Multi-token activation patching framework to analyze steering vectors, mathematical decomposition of steered OV circuits, and investigation of different steering methodologies’ effects on attention mechanisms.

Result: Steering vectors primarily interact with attention OV circuits while largely ignoring QK circuits (freezing attention scores only drops performance by 8.75%); steering vectors can be sparsified by 90-99% while retaining most performance; different steering methods agree on important dimensions; interpretable semantic concepts emerge from decomposition.

Conclusion: Steering vectors work through specific, interpretable attention mechanisms, enabling efficient sparsification and revealing that different alignment methods converge on similar internal circuits, providing mechanistic understanding of LLM steering.

Abstract: Applying steering vectors to large language models (LLMs) is an efficient and effective model alignment technique, but we lack an interpretable explanation for how it works– specifically, what internal mechanisms steering vectors affect and how this results in different model outputs. To investigate the causal mechanisms underlying the effectiveness of steering vectors, we conduct a comprehensive case study on refusal. We propose a multi-token activation patching framework and discover that different steering methodologies leverage functionally interchangeable circuits when applied at the same layer. These circuits reveal that steering vectors primarily interact with the attention mechanism through the OV circuit while largely ignoring the QK circuit– freezing all attention scores during steering drops performance by only 8.75% across two model families. A mathematical decomposition of the steered OV circuit further reveals semantically interpretable concepts, even in cases where the steering vector itself does not. Leveraging the activation patching results, we show that steering vectors can be sparsified by up to 90-99% while retaining most performance, and that different steering methodologies agree on a subset of important dimensions.

Method: Reformulates decoding as progressive self-refinement from mask embeddings to token embeddings using On-Policy Uniform Training and Soft Parallel Decoding with intermediate state interpolation.

Result: Significantly improves tokens per forward (TPF) from 2.04 to 5.47 on GSM8K and 2.71 to 5.86 on MBPP while preserving accuracy, achieving 1,338 TPS on H200 GPUs.

Conclusion: DMax effectively mitigates error accumulation in parallel decoding, enabling efficient diffusion language models with aggressive parallelism while maintaining generation quality.

Abstract: We present DMax, a new paradigm for efficient diffusion language models (dLLMs). It mitigates error accumulation in parallel decoding, enabling aggressive decoding parallelism while preserving generation quality. Unlike conventional masked dLLMs that decode through a binary mask-to-token transition, DMax reformulates decoding as a progressive self-refinement from mask embeddings to token embeddings. At the core of our approach is On-Policy Uniform Training, a novel training strategy that efficiently unifies masked and uniform dLLMs, equipping the model to recover clean tokens from both masked inputs and its own erroneous predictions. Building on this foundation, we further propose Soft Parallel Decoding. We represent each intermediate decoding state as an interpolation between the predicted token embedding and the mask embedding, enabling iterative self-revising in embedding space. Extensive experiments across a variety of benchmarks demonstrate the effectiveness of DMax. Compared with the original LLaDA-2.0-mini, our method improves TPF on GSM8K from 2.04 to 5.47 while preserving accuracy. On MBPP, it increases TPF from 2.71 to 5.86 while maintaining comparable performance. On two H200 GPUs, our model achieves an average of 1,338 TPS at batch size 1. Code is available at: https://github.com/czg1225/DMax

Method: Uses small frozen LLMs (Llama-3.2-1B, Qwen2.5-1.5B, Gemma-2-2B) to encode input into shared latent space, then injects aggregated signals into larger frozen models (Phi-3-mini, Mistral-7B), with a lightweight cross-attention output node. Only 17.6M parameters are trainable while ~12B remain frozen.

Result: Achieves 87.3% on ARC-Challenge, 82.8% on OpenBookQA, and 67.2% on MMLU, outperforming best single constituent model by 11.4, 6.2, and 1.2 percentage points respectively, and outperforming parameter-matched learned classifiers by 9.1, 5.2, and 6.7 points.

Conclusion: Demonstrates tractable gradient flow through multiple frozen model boundaries and develops selective routing behavior without explicit supervision, showing potential for efficient multimodal architectures using heterogeneous frozen LLMs.

Abstract: We present a feedforward graph architecture in which heterogeneous frozen large language models serve as computational nodes, communicating through a shared continuous latent space via learned linear projections. Building on recent work demonstrating geometric compatibility between independently trained LLM latent spaces~\cite{armstrong2026thinking}, we extend this finding from static two-model steering to end-to-end trainable multi-node graphs, where projection matrices are optimized jointly via backpropagation through residual stream injection hooks. Three small frozen models (Llama-3.2-1B, Qwen2.5-1.5B, Gemma-2-2B) encode the input into a shared latent space whose aggregate signal is injected into two larger frozen models (Phi-3-mini, Mistral-7B), whose representations feed a lightweight cross-attention output node. With only 17.6M trainable parameters against approximately 12B frozen, the architecture achieves 87.3% on ARC-Challenge, 82.8% on OpenBookQA, and 67.2% on MMLU, outperforming the best single constituent model by 11.4, 6.2, and 1.2 percentage points respectively, and outperforming parameter-matched learned classifiers on frozen single models by 9.1, 5.2, and 6.7 points. Gradient flow through multiple frozen model boundaries is empirically verified to be tractable, and the output node develops selective routing behavior across layer-2 nodes without explicit supervision.

Method: MERS (Multiple Embedding Replay Selection) replaces buffer selection with a graph-based approach that integrates both supervised and self-supervised embeddings to better capture diverse semantic information for sample selection.

Result: Consistent improvements over state-of-the-art selection strategies across various continual learning algorithms, with particularly strong gains in low-memory regimes. Outperforms single-embedding baselines on CIFAR-100 and TinyImageNet without adding parameters or increasing replay volume.

Conclusion: MERS is a practical, drop-in enhancement for replay-based continual learning that leverages multiple embeddings to improve sample selection, addressing catastrophic forgetting more effectively than single-embedding approaches.

Abstract: Catastrophic forgetting remains a key challenge in Continual Learning (CL). In replay-based CL with severe memory constraints, performance critically depends on the sample selection strategy for the replay buffer. Most existing approaches construct memory buffers using embeddings learned under supervised objectives. However, class-agnostic, self-supervised representations often encode rich, class-relevant semantics that are overlooked. We propose a new method, Multiple Embedding Replay Selection, MERS, which replaces the buffer selection module with a graph-based approach that integrates both supervised and self-supervised embeddings. Empirical results show consistent improvements over SOTA selection strategies across a range of continual learning algorithms, with particularly strong gains in low-memory regimes. On CIFAR-100 and TinyImageNet, MERS outperforms single-embedding baselines without adding model parameters or increasing replay volume, making it a practical, drop-in enhancement for replay-based continual learning.

Method: The benchmark leverages human attention signals abstracted from gaze data during question generation, creating questions that reflect natural human behavior. It includes over 5,000 multiple choice question-answer pairs spanning more than 100 hours of video.

Result: EgoEverything provides a more realistic evaluation setting for long-context egocentric video understanding by integrating human attention patterns, offering a benchmark with 5,000+ MCQs across extensive video content specifically designed for AR applications.

Conclusion: By explicitly considering human behavior through attention signals, EgoEverything offers a more faithful and realistic benchmark for evaluating long-context egocentric video understanding systems, particularly for augmented reality applications.

Abstract: Long context egocentric video understanding has recently attracted significant research attention, with augmented reality (AR) highlighted as one of its most important application domains. Nevertheless, the task remains highly challenging due to the need for reasoning over extended temporal contexts and diverse, unstructured activities. Although several benchmarks exist, most egocentric datasets rely on human worn cameras and focus mainly on visual content, with limited consideration of underlying user behavior when forming video-related queries. EgoEverything is a benchmark that explicitly considers human behavior by leveraging human attention signals, abstracted from gaze data, when generating questions. It comprises over 5,000 multiple choice question answer pairs, spanning more than 100 hours of video. By integrating human attention signals during question generation, it more faithfully captures natural human behavior and offers a realistic evaluation setting for long-context egocentric video understanding in AR.

Method: 1) Characterized relationship between noise schedule, reconstruction error, and diffusion exposure bias; 2) Proposed Adaptive Noise Schedule framework to minimize inference reconstruction error by dynamically constraining exposure bias; 3) Developed fast Proxy Unrolled Training method that stabilizes long-term rollouts without full Markov Chain sampling.

Result: Significant improvements in short-term accuracy and long-term stability over both diffusion and deterministic baselines across diverse benchmarks including forced Navier-Stokes, Kuramoto-Sivashinsky, and Transonic Flow.

Conclusion: The proposed adaptive noise scheduling and proxy unrolled training methods effectively address key limitations of autoregressive PDE diffusion models, enabling more accurate and stable emulation of complex spatiotemporal dynamics.

Abstract: Conditional Diffusion Models are powerful surrogates for emulating complex spatiotemporal dynamics, yet they often fail to match the accuracy of deterministic neural emulators for high-precision tasks. In this work, we address two critical limitations of autoregressive PDE diffusion models: their sub-optimal single-step accuracy and the prohibitive computational cost of unrolled training. First, we characterize the relationship between the noise schedule, the reconstruction error reduction rate and the diffusion exposure bias, demonstrating that standard schedules lead to suboptimal reconstruction error. Leveraging this insight, we propose an \textit{Adaptive Noise Schedule} framework that minimizes inference reconstruction error by dynamically constraining the model’s exposure bias. We further show that this optimized schedule enables a fast \textit{Proxy Unrolled Training} method to stabilize long-term rollouts without the cost of full Markov Chain sampling. Both proposed methods enable significant improvements in short-term accuracy and long-term stability over diffusion and deterministic baselines on diverse benchmarks, including forced Navier-Stokes, Kuramoto-Sivashinsky and Transonic Flow.

Method: 1) Partition dataset into domains, 2) Fit neural scaling laws from each domain to evaluation metrics, 3) Optimize data mixture by iteratively adding data from domains that maximize metric changes.

Result: Applied to autonomous driving, MOSAIC outperforms diverse baselines on the Extended Predictive Driver Model Score (EPDMS) with up to 80% less data.

Conclusion: MOSAIC provides an effective framework for data selection that accounts for metric ambiguity and enables more efficient training of physical AI models.

Abstract: Large-scale deep learning models for physical AI applications depend on diverse training data collection efforts. These models and correspondingly, the training data, must address different evaluation criteria necessary for the models to be deployable in real-world environments. Data selection policies can guide the development of the training set, but current frameworks do not account for the ambiguity in how data points affect different metrics. In this work, we propose Mixture Optimization via Scaling-Aware Iterative Collection (MOSAIC), a general data selection framework that operates by: (i) partitioning the dataset into domains; (ii) fitting neural scaling laws from each data domain to the evaluation metrics; and (iii) optimizing a data mixture by iteratively adding data from domains that maximize the change in metrics. We apply MOSAIC to autonomous driving (AD), where an End-to-End (E2E) planner model is evaluated on the Extended Predictive Driver Model Score (EPDMS), an aggregate of driving rule compliance metrics. Here, MOSAIC outperforms a diverse set of baselines on EPDMS with up to 80% less data.

Method: ADAPT introduces a pre-training paradigm that efficiently aligns physical properties of time-series data, enabling mixed-batch training despite extreme discrepancies in input sizes and channel dimensions across different datasets.

Result: Achieved state-of-the-art performance on 162 time-series classification datasets, successfully training a model on a wide range of datasets simultaneously.

Conclusion: ADAPT represents a major building block for developing generalist foundation models in time-series domains by enabling effective pre-training across diverse datasets.

Abstract: Recent work on time-series models has leveraged self-supervised training to learn meaningful features and patterns in order to improve performance on downstream tasks and generalize to unseen modalities. While these pretraining methods have shown great promise in one-to-many scenarios, where a model is pre-trained on one dataset and fine-tuned on a downstream dataset, they have struggled to generalize to new datasets when more datasets are added during pre-training. This is a fundamental challenge in building foundation models for time-series data, as it limits the ability to develop models that can learn from a large variety of diverse datasets available. To address this challenge, we present a new pre-training paradigm for time-series data called ADAPT, which can efficiently align the physical properties of data in the time-series domain, enabling mixed-batch pre-training despite the extreme discrepancies in the input sizes and channel dimensions of pre-training data. We trained on 162 time-series classification datasets and set new state-of-the-art performance for classification benchmarks. We successfully train a model within the time-series domain on a wide range of datasets simultaneously, which is a major building block for building generalist foundation models in time-series domains.

Method: The paper introduces a framework that recasts multivariate time series forecasting as a series of scalar regression problems, which can then be solved zero-shot by any tabular foundation model with regression capabilities. The TabPFN-TS backbone is used to implement this approach.

Result: The method shows positive results compared to current state-of-the-art tabular methods, demonstrating that tabular foundation models can effectively handle multivariate time series forecasting while capturing inter-channel dependencies.

Conclusion: Tabular foundation models can be successfully applied to multivariate time series forecasting through appropriate problem reformulation, overcoming the limitation of treating channels independently and enabling zero-shot forecasting capabilities.

Abstract: Tabular foundation models, particularly Prior-data Fitted Networks like TabPFN have emerged as the leading contender in a myriad of tasks ranging from data imputation to label prediction on the tabular data format surpassing the historical successes of tree-based models. This has led to investigations on their applicability to forecasting time series data which can be formulated as a tabular problem. While recent work to this end has displayed positive results, most works have limited their treatment of multivariate time series problems to several independent univariate time series forecasting subproblems, thus ignoring any inter-channel interactions. Overcoming this limitation, we introduce a generally applicable framework for multivariate time series forecasting using tabular foundation models. We achieve this by recasting the multivariate time series forecasting problem as a series of scalar regression problems which can then be solved zero-shot by any tabular foundation model with regression capabilities. We present results of our method using the TabPFN-TS backbone and compare performance with the current state of the art tabular methods.

Method: Proposes RIA (Regularization for Invariance with Adversarial training) that performs adversarial exploration for training data environments using adversarial label-invariant data augmentations, formulated as constrained optimization problems solved via alternating gradient descent-ascent algorithm.

Result: Extensive experiments on OoD graph classification show the method achieves high accuracy compared with OoD baselines across various synthetic and natural distribution shifts.

Conclusion: RIA effectively addresses covariate shift in OoD generalization through adversarial training with label-invariant augmentations, working with existing OoD methods and demonstrating strong performance on graph classification tasks.

Abstract: Out-of-distribution (OoD) generalization occurs when representation learning encounters a distribution shift. This occurs frequently in practice when training and testing data come from different environments. Covariate shift is a type of distribution shift that occurs only in the input data, while the concept distribution stays invariant. We propose RIA - Regularization for Invariance with Adversarial training, a new method for OoD generalization under convariate shift. Motivated by an analogy to $Q$-learning, it performs an adversarial exploration for training data environments. These new environments are induced by adversarial label invariant data augmentations that prevent a collapse to an in-distribution trained learner. It works with many existing OoD generalization methods for covariate shift that can be formulated as constrained optimization problems. We develop an alternating gradient descent-ascent algorithm to solve the problem, and perform extensive experiments on OoD graph classification for various kinds of synthetic and natural distribution shifts. We demonstrate that our method can achieve high accuracy compared with OoD baselines.

Method: Develops a theoretical framework using vector concentration inequalities to analyze query complexities. Systematically creates a linear-space algorithm requiring O(n/ε² log(1/δ)) utility queries. Bridges existing methods (OFA, unbiased kernelSHAP, SHAP-IQ, regression-adjusted) and introduces Adalina - an adaptive, linear-time, linear-space randomized algorithm that minimizes mean square error.

Result: Establishes sharper query complexity bounds for semi-value approximation under space constraints. Adalina theoretically achieves improved mean square error compared to existing methods. All theoretical findings are experimentally validated.

Conclusion: The paper provides an efficient framework for approximating Shapley values and semi-values with linear space complexity, bridging multiple existing methods and introducing an adaptive algorithm that optimizes mean square error for practical large-scale applications.

Abstract: The Shapley value, and its broader family of semi-values, has received much attention in various attribution problems. A fundamental and long-standing challenge is their efficient approximation, since exact computation generally requires an exponential number of utility queries in the number of players $n$. To meet the challenges of large-scale applications, we explore the limits of efficiently approximating semi-values under a $Θ(n)$ space constraint. Building upon a vector concentration inequality, we establish a theoretical framework that enables sharper query complexities for existing unbiased randomized algorithms. Within this framework, we systematically develop a linear-space algorithm that requires $O(\frac{n}{ε^{2}}\log\frac{1}δ)$ utility queries to ensure $P(|\hat{\boldsymbolφ}-\boldsymbolφ|_{2}\geqε)\leq δ$ for all commonly used semi-values. In particular, our framework naturally bridges OFA, unbiased kernelSHAP, SHAP-IQ and the regression-adjusted approach, and definitively characterizes when paired sampling is beneficial. Moreover, our algorithm allows explicit minimization of the mean square error for each specific utility function. Accordingly, we introduce the first adaptive, linear-time, linear-space randomized algorithm, Adalina, that theoretically achieves improved mean square error. All of our theoretical findings are experimentally validated.

Method: Proposes Progressive Reasoning Gain (PRG) metric to measure whether reasoning steps progressively strengthen support for final answers. Introduces HyTuning framework that adaptively reweights Reasoning Distillation (RD) and Reinforcement Learning from Internal Feedback (RLIF) using PRG-style metrics, using scarce supervised reasoning traces as anchor while exploiting abundant unlabeled queries for scalability.

Result: Experiments on several domain-specific and general benchmarks show HyTuning improves accuracy while achieving confidence faithfulness under limited supervision, supporting “Less Approximates More” effect.

Conclusion: HyTuning effectively addresses confidence faithfulness in LLMs through adaptive hybrid training that balances supervised reasoning distillation with unsupervised reinforcement learning, achieving better calibration between confidence and correctness.

Abstract: Large language models are increasingly deployed in high-stakes tasks, where confident yet incorrect inferences may cause severe real-world harm, bringing the previously overlooked issue of confidence faithfulness back to the forefront. A promising solution is to jointly optimize unsupervised Reinforcement Learning from Internal Feedback (RLIF) with reasoning-trace-guided Reasoning Distillation (RD), which may face three persistent challenges: scarcity of high-quality training corpora, factually unwarranted overconfidence and indiscriminate fusion that amplifies erroneous updates. Inspired by the human confidence accumulation from uncertainty to certainty, we propose Progressive Reasoning Gain (PRG) to measure whether reasoning steps progressively strengthen support for the final answer. Furthermore, we introduce HyTuning, a hybrid post-training framework that adaptively reweights RD and RLIF via a PRG-style metric, using scarce supervised reasoning traces as a stable anchor while exploiting abundant unlabeled queries for scalability. Experiments on several domain-specific and general benchmarks demonstrate that HyTuning improves accuracy while achieving confidence faithfulness under limited supervision, supporting a practical “Less Approximates More” effect.

Method: Introduces a new industry-oriented dataset with maintenance events and usage changes. Establishes benchmark comparing steady-state/nonstationary data-driven models and Bayesian filters. Proposes self-supervised learning approaches that learn latent representations without true health labels.

Result: Traditional Bayesian filters remain strong baselines. SSL methods reveal intrinsic complexity of health estimation and highlight need for more advanced interpretable inference strategies. Provides practical lower bound on difficulty of solving this inverse problem.

Conclusion: The work establishes a realistic benchmark for turbofan health estimation, shows traditional filters are still competitive, and demonstrates SSL’s potential while revealing the problem’s inherent complexity.

Abstract: Estimating the health state of turbofan engines is a challenging ill-posed inverse problem, hindered by sparse sensing and complex nonlinear thermodynamics. Research in this area remains fragmented, with comparisons limited by the use of unrealistic datasets and insufficient exploration of the exploitation of temporal information. This work investigates how to recover component-level health indicators from operational sensor data under realistic degradation and maintenance patterns. To support this study, we introduce a new dataset that incorporates industry-oriented complexities such as maintenance events and usage changes. Using this dataset, we establish an initial benchmark that compares steady-state and nonstationary data-driven models, and Bayesian filters, classic families of methods used to solve this problem. In addition to this benchmark, we introduce self-supervised learning (SSL) approaches that learn latent representations without access to true health labels, a scenario reflective of real-world operational constraints. By comparing the downstream estimation performance of these unsupervised representations against the direct prediction baselines, we establish a practical lower bound on the difficulty of solving this inverse problem. Our results reveal that traditional filters remain strong baselines, while SSL methods reveal the intrinsic complexity of health estimation and highlight the need for more advanced and interpretable inference strategies. For reproducibility, both the generated dataset and the implementation used in this work are made accessible.

Method: TTVS consists of two modules: (1) Online Variational Synthesis transforms static test queries into dynamic streams of semantically-equivalent variations, and (2) Test-time Hybrid Exploration balances accuracy-driven exploitation with consistency-driven exploration across synthetic variants.

Result: Extensive experiments across eight model architectures show TTVS yields superior performance, surpassing other test-time adaptation methods and even outperforming state-of-the-art supervised RL-based techniques trained on vast labeled data.

Conclusion: TTVS provides an effective framework for test-time adaptation of Large Reasoning Models without requiring labeled supervision, enabling models to self-evolve by learning underlying problem logic from dynamically synthesized query variations.

Abstract: Despite significant advances in Large Reasoning Models (LRMs) driven by reinforcement learning with verifiable rewards (RLVR), this paradigm is fundamentally limited in specialized or novel domains where such supervision is prohibitively expensive or unavailable, posing a key challenge for test-time adaptation. While existing test-time methods offer a potential solution, they are constrained by learning from static query sets, risking overfitting to textual patterns. To address this gap, we introduce Test-Time Variational Synthesis (TTVS), a novel framework that enables LRMs to self-evolve by dynamically augmenting the training stream from unlabeled test queries. TTVS comprises two synergistic modules: (1) Online Variational Synthesis, which transforms static test queries into a dynamic stream of diverse, semantically-equivalent variations, enforcing the model to learn underlying problem logic rather than superficial patterns; (2) Test-time Hybrid Exploration, which balances accuracy-driven exploitation with consistency-driven exploration across synthetic variants. Extensive experiments show TTVS yields superior performance across eight model architectures. Notably, using only unlabeled test-time data, TTVS not only surpasses other test-time adaptation methods but also outperforms state-of-the-art supervised RL-based techniques trained on vast, high-quality labeled data.

Method: Propose a persistence-based data augmentation framework that encodes local gradient flow regions and their hierarchical evolution using the Morse-Smale complex. This representation is compatible with both convolutional and graph neural networks and retains spatially localized topological information across multiple scales. The augmentation procedure has computational complexity O(n log n).

Result: The method outperforms baselines and global TDA descriptors (persistence images and landscapes) on histopathology image classification and 3D porous material regression. Pruning the base level of the hierarchy reduces memory usage while maintaining competitive performance.

Conclusion: The results highlight the potential of local, structured topological augmentation for scalable and interpretable learning across data modalities.

Abstract: Topological Data Analysis (TDA) provides tools to describe the shape of data, but integrating topological features into deep learning pipelines remains challenging, especially when preserving local geometric structure rather than summarizing it globally. We propose a persistence-based data augmentation framework that encodes local gradient flow regions and their hierarchical evolution using the Morse-Smale complex. This representation, compatible with both convolutional and graph neural networks, retains spatially localized topological information across multiple scales. Importantly, the augmentation procedure itself is efficient, with computational complexity $O(n \log n)$, making it practical for large datasets. We evaluate our method on histopathology image classification and 3D porous material regression, where it consistently outperforms baselines and global TDA descriptors such as persistence images and landscapes. We also show that pruning the base level of the hierarchy reduces memory usage while maintaining competitive performance. These results highlight the potential of local, structured topological augmentation for scalable and interpretable learning across data modalities.

Method: Investigates symmetric uniform quantization (32, 8, 4, 2 bits) on a lightweight 1-D convolutional model (AeroConv1D) trained via FL on NASA C-MAPSS benchmark under realistic Non-IID client partition using multi-seed evaluation (N=10 seeds).

Result: INT4 achieves accuracy statistically indistinguishable from FP32 while delivering 8× reduction in gradient communication cost. INT2 shows catastrophic instability under heterogeneous conditions despite lower MAE on some metrics.

Conclusion: INT4 quantization is viable for FL deployment on resource-constrained aerospace IoT nodes, while INT2 is unsuitable due to operational instability. Non-IID evaluation is crucial for realistic assessment.

Abstract: Federated learning (FL) enables privacy-preserving predictive maintenance across distributed aerospace fleets, but gradient communication overhead constrains deployment on bandwidth-limited IoT nodes. This paper investigates the impact of symmetric uniform quantization ($b \in {32,8,4,2}$ bits) on the accuracy–efficiency trade-off of a custom-designed lightweight 1-D convolutional model (AeroConv1D, 9,697 parameters) trained via FL on the NASA C-MAPSS benchmark under a realistic Non-IID client partition. Using a rigorous multi-seed evaluation ($N=10$ seeds), we show that INT4 achieves accuracy \emph{statistically indistinguishable} from FP32 on both FD001 ($p=0.341$) and FD002 ($p=0.264$ MAE, $p=0.534$ NASA score) while delivering an $8\times$ reduction in gradient communication cost (37.88KiB $\to$ 4.73KiB per round). A key methodological finding is that naïve IID client partitioning artificially suppresses variance; correct Non-IID evaluation reveals the true operational instability of extreme quantization, demonstrated via a direct empirical IID vs.\ Non-IID comparison. INT2 is empirically characterized as unsuitable: while it achieves lower MAE on FD002 through extreme quantization-induced over-regularization, this apparent gain is accompanied by catastrophic NASA score instability (CV,=,45.8% vs.\ 22.3% for FP32), confirming non-reproducibility under heterogeneous operating conditions. Analytical FPGA resource projections on the Xilinx ZCU102 confirm that INT4 fits within hardware constraints (85.5% DSP utilization), potentially enabling a complete FL pipeline on a single SoC. The full simulation codebase and FPGA estimation scripts are publicly available at https://github.com/therealdeadbeef/aerospace-fl-quantization.

Method: Systematically evaluated five methods (ASNE, DGI, GraphSAGE, node2vec, VERSE) across multiple datasets and embedding dimensions, assessing stability from both representational and functional perspectives alongside performance evaluation.

Result: Embedding stability varies significantly with dimensionality with different patterns across methods: node2vec and ASNE become more stable with higher dimensionality, while others don’t follow this trend. Maximum stability doesn’t necessarily align with optimal task performance.

Conclusion: Careful selection of embedding dimension is important, highlighting trade-offs between stability, performance, and computational effectiveness in graph representation learning.

Abstract: Previous work has established that neural network-based node embeddings return different outcomes when trained with identical parameters on the same dataset, just from using different training seeds. Yet, it has not been thoroughly analyzed how key hyperparameters such as embedding dimension could impact this instability. In this work, we investigate how varying the dimensionality of node embeddings influences both their stability and downstream performance. We systematically evaluate five widely used methods – ASNE, DGI, GraphSAGE, node2vec, and VERSE – across multiple datasets and embedding dimensions. We assess stability from both a representational perspective and a functional perspective, alongside performance evaluation. Our results show that embedding stability varies significantly with dimensionality, but we observe different patterns across the methods we consider: while some approaches, such as node2vec and ASNE, tend to become more stable with higher dimensionality, other methods do not exhibit the same trend. Moreover, we find that maximum stability does not necessarily align with optimal task performance. These findings highlight the importance of carefully selecting embedding dimension, and provide new insights into the trade-offs between stability, performance, and computational effectiveness in graph representation learning.

Method: Meta-optimized approach that conditions on a few image-brain activation examples from new subjects, performs hierarchical inference: 1) estimates per-voxel visual response encoder parameters using stimulus-response context, 2) aggregates functional inversion using encoder parameters and response values across voxels

Result: Demonstrates strong cross-subject and cross-scanner generalization across diverse visual backbones without retraining or fine-tuning, works without anatomical alignment or stimulus overlap

Conclusion: Critical step towards a generalizable foundation model for non-invasive brain decoding that can adapt to new subjects through in-context learning

Abstract: Visual decoding from brain signals is a key challenge at the intersection of computer vision and neuroscience, requiring methods that bridge neural representations and computational models of vision. A field-wide goal is to achieve generalizable, cross-subject models. A major obstacle towards this goal is the substantial variability in neural representations across individuals, which has so far required training bespoke models or fine-tuning separately for each subject. To address this challenge, we introduce a meta-optimized approach for semantic visual decoding from fMRI that generalizes to novel subjects without any fine-tuning. By simply conditioning on a small set of image-brain activation examples from the new individual, our model rapidly infers their unique neural encoding patterns to facilitate robust and efficient visual decoding. Our approach is explicitly optimized for in-context learning of the new subject’s encoding model and performs decoding by hierarchical inference, inverting the encoder. First, for multiple brain regions, we estimate the per-voxel visual response encoder parameters by constructing a context over multiple stimuli and responses. Second, we construct a context consisting of encoder parameters and response values over multiple voxels to perform aggregated functional inversion. We demonstrate strong cross-subject and cross-scanner generalization across diverse visual backbones without retraining or fine-tuning. Moreover, our approach requires neither anatomical alignment nor stimulus overlap. This work is a critical step towards a generalizable foundation model for non-invasive brain decoding.

Method: Proposes a modular framework with data preprocessing, model-agnostic interface, standardized training/inference pipelines, and comprehensive evaluation module; validates on 7 benchmark datasets

Result: Framework successfully supports all three paradigms; GAN-based augmentation improves downstream performance under distribution shift; implementation available as open source